2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is seq_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
65 #define NR_STRIPES 256
66 #define STRIPE_SIZE PAGE_SIZE
67 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
68 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
69 #define IO_THRESHOLD 1
70 #define BYPASS_THRESHOLD 1
71 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
72 #define HASH_MASK (NR_HASH - 1)
74 static inline struct hlist_head
*stripe_hash(struct r5conf
*conf
, sector_t sect
)
76 int hash
= (sect
>> STRIPE_SHIFT
) & HASH_MASK
;
77 return &conf
->stripe_hashtbl
[hash
];
80 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
81 * order without overlap. There may be several bio's per stripe+device, and
82 * a bio could span several devices.
83 * When walking this list for a particular stripe+device, we must never proceed
84 * beyond a bio that extends past this device, as the next bio might no longer
86 * This function is used to determine the 'next' bio in the list, given the sector
87 * of the current stripe+device
89 static inline struct bio
*r5_next_bio(struct bio
*bio
, sector_t sector
)
91 int sectors
= bio
->bi_size
>> 9;
92 if (bio
->bi_sector
+ sectors
< sector
+ STRIPE_SECTORS
)
99 * We maintain a biased count of active stripes in the bottom 16 bits of
100 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
102 static inline int raid5_bi_processed_stripes(struct bio
*bio
)
104 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
105 return (atomic_read(segments
) >> 16) & 0xffff;
108 static inline int raid5_dec_bi_active_stripes(struct bio
*bio
)
110 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
111 return atomic_sub_return(1, segments
) & 0xffff;
114 static inline void raid5_inc_bi_active_stripes(struct bio
*bio
)
116 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
117 atomic_inc(segments
);
120 static inline void raid5_set_bi_processed_stripes(struct bio
*bio
,
123 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
127 old
= atomic_read(segments
);
128 new = (old
& 0xffff) | (cnt
<< 16);
129 } while (atomic_cmpxchg(segments
, old
, new) != old
);
132 static inline void raid5_set_bi_stripes(struct bio
*bio
, unsigned int cnt
)
134 atomic_t
*segments
= (atomic_t
*)&bio
->bi_phys_segments
;
135 atomic_set(segments
, cnt
);
138 /* Find first data disk in a raid6 stripe */
139 static inline int raid6_d0(struct stripe_head
*sh
)
142 /* ddf always start from first device */
144 /* md starts just after Q block */
145 if (sh
->qd_idx
== sh
->disks
- 1)
148 return sh
->qd_idx
+ 1;
150 static inline int raid6_next_disk(int disk
, int raid_disks
)
153 return (disk
< raid_disks
) ? disk
: 0;
156 /* When walking through the disks in a raid5, starting at raid6_d0,
157 * We need to map each disk to a 'slot', where the data disks are slot
158 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
159 * is raid_disks-1. This help does that mapping.
161 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
162 int *count
, int syndrome_disks
)
168 if (idx
== sh
->pd_idx
)
169 return syndrome_disks
;
170 if (idx
== sh
->qd_idx
)
171 return syndrome_disks
+ 1;
177 static void return_io(struct bio
*return_bi
)
179 struct bio
*bi
= return_bi
;
182 return_bi
= bi
->bi_next
;
190 static void print_raid5_conf (struct r5conf
*conf
);
192 static int stripe_operations_active(struct stripe_head
*sh
)
194 return sh
->check_state
|| sh
->reconstruct_state
||
195 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
196 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
199 static void do_release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
)
201 BUG_ON(!list_empty(&sh
->lru
));
202 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
203 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
204 if (test_bit(STRIPE_DELAYED
, &sh
->state
) &&
205 !test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
206 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
207 else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
208 sh
->bm_seq
- conf
->seq_write
> 0)
209 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
211 clear_bit(STRIPE_DELAYED
, &sh
->state
);
212 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
213 list_add_tail(&sh
->lru
, &conf
->handle_list
);
215 md_wakeup_thread(conf
->mddev
->thread
);
217 BUG_ON(stripe_operations_active(sh
));
218 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
219 if (atomic_dec_return(&conf
->preread_active_stripes
)
221 md_wakeup_thread(conf
->mddev
->thread
);
222 atomic_dec(&conf
->active_stripes
);
223 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
224 list_add_tail(&sh
->lru
, &conf
->inactive_list
);
225 wake_up(&conf
->wait_for_stripe
);
226 if (conf
->retry_read_aligned
)
227 md_wakeup_thread(conf
->mddev
->thread
);
232 static void __release_stripe(struct r5conf
*conf
, struct stripe_head
*sh
)
234 if (atomic_dec_and_test(&sh
->count
))
235 do_release_stripe(conf
, sh
);
238 static void release_stripe(struct stripe_head
*sh
)
240 struct r5conf
*conf
= sh
->raid_conf
;
243 local_irq_save(flags
);
244 if (atomic_dec_and_lock(&sh
->count
, &conf
->device_lock
)) {
245 do_release_stripe(conf
, sh
);
246 spin_unlock(&conf
->device_lock
);
248 local_irq_restore(flags
);
251 static inline void remove_hash(struct stripe_head
*sh
)
253 pr_debug("remove_hash(), stripe %llu\n",
254 (unsigned long long)sh
->sector
);
256 hlist_del_init(&sh
->hash
);
259 static inline void insert_hash(struct r5conf
*conf
, struct stripe_head
*sh
)
261 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
263 pr_debug("insert_hash(), stripe %llu\n",
264 (unsigned long long)sh
->sector
);
266 hlist_add_head(&sh
->hash
, hp
);
270 /* find an idle stripe, make sure it is unhashed, and return it. */
271 static struct stripe_head
*get_free_stripe(struct r5conf
*conf
)
273 struct stripe_head
*sh
= NULL
;
274 struct list_head
*first
;
276 if (list_empty(&conf
->inactive_list
))
278 first
= conf
->inactive_list
.next
;
279 sh
= list_entry(first
, struct stripe_head
, lru
);
280 list_del_init(first
);
282 atomic_inc(&conf
->active_stripes
);
287 static void shrink_buffers(struct stripe_head
*sh
)
291 int num
= sh
->raid_conf
->pool_size
;
293 for (i
= 0; i
< num
; i
++) {
297 sh
->dev
[i
].page
= NULL
;
302 static int grow_buffers(struct stripe_head
*sh
)
305 int num
= sh
->raid_conf
->pool_size
;
307 for (i
= 0; i
< num
; i
++) {
310 if (!(page
= alloc_page(GFP_KERNEL
))) {
313 sh
->dev
[i
].page
= page
;
318 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
319 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
320 struct stripe_head
*sh
);
322 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
324 struct r5conf
*conf
= sh
->raid_conf
;
327 BUG_ON(atomic_read(&sh
->count
) != 0);
328 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
329 BUG_ON(stripe_operations_active(sh
));
331 pr_debug("init_stripe called, stripe %llu\n",
332 (unsigned long long)sh
->sector
);
336 sh
->generation
= conf
->generation
- previous
;
337 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
339 stripe_set_idx(sector
, conf
, previous
, sh
);
343 for (i
= sh
->disks
; i
--; ) {
344 struct r5dev
*dev
= &sh
->dev
[i
];
346 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
347 test_bit(R5_LOCKED
, &dev
->flags
)) {
348 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
349 (unsigned long long)sh
->sector
, i
, dev
->toread
,
350 dev
->read
, dev
->towrite
, dev
->written
,
351 test_bit(R5_LOCKED
, &dev
->flags
));
355 raid5_build_block(sh
, i
, previous
);
357 insert_hash(conf
, sh
);
360 static struct stripe_head
*__find_stripe(struct r5conf
*conf
, sector_t sector
,
363 struct stripe_head
*sh
;
364 struct hlist_node
*hn
;
366 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
367 hlist_for_each_entry(sh
, hn
, stripe_hash(conf
, sector
), hash
)
368 if (sh
->sector
== sector
&& sh
->generation
== generation
)
370 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
375 * Need to check if array has failed when deciding whether to:
377 * - remove non-faulty devices
380 * This determination is simple when no reshape is happening.
381 * However if there is a reshape, we need to carefully check
382 * both the before and after sections.
383 * This is because some failed devices may only affect one
384 * of the two sections, and some non-in_sync devices may
385 * be insync in the section most affected by failed devices.
387 static int calc_degraded(struct r5conf
*conf
)
389 int degraded
, degraded2
;
394 for (i
= 0; i
< conf
->previous_raid_disks
; i
++) {
395 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
396 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
398 else if (test_bit(In_sync
, &rdev
->flags
))
401 /* not in-sync or faulty.
402 * If the reshape increases the number of devices,
403 * this is being recovered by the reshape, so
404 * this 'previous' section is not in_sync.
405 * If the number of devices is being reduced however,
406 * the device can only be part of the array if
407 * we are reverting a reshape, so this section will
410 if (conf
->raid_disks
>= conf
->previous_raid_disks
)
414 if (conf
->raid_disks
== conf
->previous_raid_disks
)
418 for (i
= 0; i
< conf
->raid_disks
; i
++) {
419 struct md_rdev
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
420 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
))
422 else if (test_bit(In_sync
, &rdev
->flags
))
425 /* not in-sync or faulty.
426 * If reshape increases the number of devices, this
427 * section has already been recovered, else it
428 * almost certainly hasn't.
430 if (conf
->raid_disks
<= conf
->previous_raid_disks
)
434 if (degraded2
> degraded
)
439 static int has_failed(struct r5conf
*conf
)
443 if (conf
->mddev
->reshape_position
== MaxSector
)
444 return conf
->mddev
->degraded
> conf
->max_degraded
;
446 degraded
= calc_degraded(conf
);
447 if (degraded
> conf
->max_degraded
)
452 static struct stripe_head
*
453 get_active_stripe(struct r5conf
*conf
, sector_t sector
,
454 int previous
, int noblock
, int noquiesce
)
456 struct stripe_head
*sh
;
458 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
460 spin_lock_irq(&conf
->device_lock
);
463 wait_event_lock_irq(conf
->wait_for_stripe
,
464 conf
->quiesce
== 0 || noquiesce
,
465 conf
->device_lock
, /* nothing */);
466 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
468 if (!conf
->inactive_blocked
)
469 sh
= get_free_stripe(conf
);
470 if (noblock
&& sh
== NULL
)
473 conf
->inactive_blocked
= 1;
474 wait_event_lock_irq(conf
->wait_for_stripe
,
475 !list_empty(&conf
->inactive_list
) &&
476 (atomic_read(&conf
->active_stripes
)
477 < (conf
->max_nr_stripes
*3/4)
478 || !conf
->inactive_blocked
),
481 conf
->inactive_blocked
= 0;
483 init_stripe(sh
, sector
, previous
);
485 if (atomic_read(&sh
->count
)) {
486 BUG_ON(!list_empty(&sh
->lru
)
487 && !test_bit(STRIPE_EXPANDING
, &sh
->state
)
488 && !test_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
));
490 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
491 atomic_inc(&conf
->active_stripes
);
492 if (list_empty(&sh
->lru
) &&
493 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
495 list_del_init(&sh
->lru
);
498 } while (sh
== NULL
);
501 atomic_inc(&sh
->count
);
503 spin_unlock_irq(&conf
->device_lock
);
507 /* Determine if 'data_offset' or 'new_data_offset' should be used
508 * in this stripe_head.
510 static int use_new_offset(struct r5conf
*conf
, struct stripe_head
*sh
)
512 sector_t progress
= conf
->reshape_progress
;
513 /* Need a memory barrier to make sure we see the value
514 * of conf->generation, or ->data_offset that was set before
515 * reshape_progress was updated.
518 if (progress
== MaxSector
)
520 if (sh
->generation
== conf
->generation
- 1)
522 /* We are in a reshape, and this is a new-generation stripe,
523 * so use new_data_offset.
529 raid5_end_read_request(struct bio
*bi
, int error
);
531 raid5_end_write_request(struct bio
*bi
, int error
);
533 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
535 struct r5conf
*conf
= sh
->raid_conf
;
536 int i
, disks
= sh
->disks
;
540 for (i
= disks
; i
--; ) {
542 int replace_only
= 0;
543 struct bio
*bi
, *rbi
;
544 struct md_rdev
*rdev
, *rrdev
= NULL
;
545 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
)) {
546 if (test_and_clear_bit(R5_WantFUA
, &sh
->dev
[i
].flags
))
550 if (test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
552 } else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
554 else if (test_and_clear_bit(R5_WantReplace
,
555 &sh
->dev
[i
].flags
)) {
560 if (test_and_clear_bit(R5_SyncIO
, &sh
->dev
[i
].flags
))
563 bi
= &sh
->dev
[i
].req
;
564 rbi
= &sh
->dev
[i
].rreq
; /* For writing to replacement */
569 bi
->bi_end_io
= raid5_end_write_request
;
570 rbi
->bi_end_io
= raid5_end_write_request
;
572 bi
->bi_end_io
= raid5_end_read_request
;
575 rrdev
= rcu_dereference(conf
->disks
[i
].replacement
);
576 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
577 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
586 /* We raced and saw duplicates */
589 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
) && rrdev
)
594 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
597 atomic_inc(&rdev
->nr_pending
);
598 if (rrdev
&& test_bit(Faulty
, &rrdev
->flags
))
601 atomic_inc(&rrdev
->nr_pending
);
604 /* We have already checked bad blocks for reads. Now
605 * need to check for writes. We never accept write errors
606 * on the replacement, so we don't to check rrdev.
608 while ((rw
& WRITE
) && rdev
&&
609 test_bit(WriteErrorSeen
, &rdev
->flags
)) {
612 int bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
613 &first_bad
, &bad_sectors
);
618 set_bit(BlockedBadBlocks
, &rdev
->flags
);
619 if (!conf
->mddev
->external
&&
620 conf
->mddev
->flags
) {
621 /* It is very unlikely, but we might
622 * still need to write out the
623 * bad block log - better give it
625 md_check_recovery(conf
->mddev
);
628 * Because md_wait_for_blocked_rdev
629 * will dec nr_pending, we must
630 * increment it first.
632 atomic_inc(&rdev
->nr_pending
);
633 md_wait_for_blocked_rdev(rdev
, conf
->mddev
);
635 /* Acknowledged bad block - skip the write */
636 rdev_dec_pending(rdev
, conf
->mddev
);
642 if (s
->syncing
|| s
->expanding
|| s
->expanded
644 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
646 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
648 bi
->bi_bdev
= rdev
->bdev
;
649 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
650 __func__
, (unsigned long long)sh
->sector
,
652 atomic_inc(&sh
->count
);
653 if (use_new_offset(conf
, sh
))
654 bi
->bi_sector
= (sh
->sector
655 + rdev
->new_data_offset
);
657 bi
->bi_sector
= (sh
->sector
658 + rdev
->data_offset
);
659 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
660 bi
->bi_rw
|= REQ_FLUSH
;
662 bi
->bi_flags
= 1 << BIO_UPTODATE
;
664 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
665 bi
->bi_io_vec
[0].bv_offset
= 0;
666 bi
->bi_size
= STRIPE_SIZE
;
669 set_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
);
670 generic_make_request(bi
);
673 if (s
->syncing
|| s
->expanding
|| s
->expanded
675 md_sync_acct(rrdev
->bdev
, STRIPE_SECTORS
);
677 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
679 rbi
->bi_bdev
= rrdev
->bdev
;
680 pr_debug("%s: for %llu schedule op %ld on "
681 "replacement disc %d\n",
682 __func__
, (unsigned long long)sh
->sector
,
684 atomic_inc(&sh
->count
);
685 if (use_new_offset(conf
, sh
))
686 rbi
->bi_sector
= (sh
->sector
687 + rrdev
->new_data_offset
);
689 rbi
->bi_sector
= (sh
->sector
690 + rrdev
->data_offset
);
691 rbi
->bi_flags
= 1 << BIO_UPTODATE
;
693 rbi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
694 rbi
->bi_io_vec
[0].bv_offset
= 0;
695 rbi
->bi_size
= STRIPE_SIZE
;
697 generic_make_request(rbi
);
699 if (!rdev
&& !rrdev
) {
701 set_bit(STRIPE_DEGRADED
, &sh
->state
);
702 pr_debug("skip op %ld on disc %d for sector %llu\n",
703 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
704 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
705 set_bit(STRIPE_HANDLE
, &sh
->state
);
710 static struct dma_async_tx_descriptor
*
711 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
712 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
715 struct page
*bio_page
;
718 struct async_submit_ctl submit
;
719 enum async_tx_flags flags
= 0;
721 if (bio
->bi_sector
>= sector
)
722 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
724 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
727 flags
|= ASYNC_TX_FENCE
;
728 init_async_submit(&submit
, flags
, tx
, NULL
, NULL
, NULL
);
730 bio_for_each_segment(bvl
, bio
, i
) {
731 int len
= bvl
->bv_len
;
735 if (page_offset
< 0) {
736 b_offset
= -page_offset
;
737 page_offset
+= b_offset
;
741 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
742 clen
= STRIPE_SIZE
- page_offset
;
747 b_offset
+= bvl
->bv_offset
;
748 bio_page
= bvl
->bv_page
;
750 tx
= async_memcpy(page
, bio_page
, page_offset
,
751 b_offset
, clen
, &submit
);
753 tx
= async_memcpy(bio_page
, page
, b_offset
,
754 page_offset
, clen
, &submit
);
756 /* chain the operations */
757 submit
.depend_tx
= tx
;
759 if (clen
< len
) /* hit end of page */
767 static void ops_complete_biofill(void *stripe_head_ref
)
769 struct stripe_head
*sh
= stripe_head_ref
;
770 struct bio
*return_bi
= NULL
;
773 pr_debug("%s: stripe %llu\n", __func__
,
774 (unsigned long long)sh
->sector
);
776 /* clear completed biofills */
777 for (i
= sh
->disks
; i
--; ) {
778 struct r5dev
*dev
= &sh
->dev
[i
];
780 /* acknowledge completion of a biofill operation */
781 /* and check if we need to reply to a read request,
782 * new R5_Wantfill requests are held off until
783 * !STRIPE_BIOFILL_RUN
785 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
786 struct bio
*rbi
, *rbi2
;
791 while (rbi
&& rbi
->bi_sector
<
792 dev
->sector
+ STRIPE_SECTORS
) {
793 rbi2
= r5_next_bio(rbi
, dev
->sector
);
794 if (!raid5_dec_bi_active_stripes(rbi
)) {
795 rbi
->bi_next
= return_bi
;
802 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
804 return_io(return_bi
);
806 set_bit(STRIPE_HANDLE
, &sh
->state
);
810 static void ops_run_biofill(struct stripe_head
*sh
)
812 struct dma_async_tx_descriptor
*tx
= NULL
;
813 struct async_submit_ctl submit
;
816 pr_debug("%s: stripe %llu\n", __func__
,
817 (unsigned long long)sh
->sector
);
819 for (i
= sh
->disks
; i
--; ) {
820 struct r5dev
*dev
= &sh
->dev
[i
];
821 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
823 spin_lock_irq(&sh
->stripe_lock
);
824 dev
->read
= rbi
= dev
->toread
;
826 spin_unlock_irq(&sh
->stripe_lock
);
827 while (rbi
&& rbi
->bi_sector
<
828 dev
->sector
+ STRIPE_SECTORS
) {
829 tx
= async_copy_data(0, rbi
, dev
->page
,
831 rbi
= r5_next_bio(rbi
, dev
->sector
);
836 atomic_inc(&sh
->count
);
837 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
, NULL
);
838 async_trigger_callback(&submit
);
841 static void mark_target_uptodate(struct stripe_head
*sh
, int target
)
848 tgt
= &sh
->dev
[target
];
849 set_bit(R5_UPTODATE
, &tgt
->flags
);
850 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
851 clear_bit(R5_Wantcompute
, &tgt
->flags
);
854 static void ops_complete_compute(void *stripe_head_ref
)
856 struct stripe_head
*sh
= stripe_head_ref
;
858 pr_debug("%s: stripe %llu\n", __func__
,
859 (unsigned long long)sh
->sector
);
861 /* mark the computed target(s) as uptodate */
862 mark_target_uptodate(sh
, sh
->ops
.target
);
863 mark_target_uptodate(sh
, sh
->ops
.target2
);
865 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
866 if (sh
->check_state
== check_state_compute_run
)
867 sh
->check_state
= check_state_compute_result
;
868 set_bit(STRIPE_HANDLE
, &sh
->state
);
872 /* return a pointer to the address conversion region of the scribble buffer */
873 static addr_conv_t
*to_addr_conv(struct stripe_head
*sh
,
874 struct raid5_percpu
*percpu
)
876 return percpu
->scribble
+ sizeof(struct page
*) * (sh
->disks
+ 2);
879 static struct dma_async_tx_descriptor
*
880 ops_run_compute5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
882 int disks
= sh
->disks
;
883 struct page
**xor_srcs
= percpu
->scribble
;
884 int target
= sh
->ops
.target
;
885 struct r5dev
*tgt
= &sh
->dev
[target
];
886 struct page
*xor_dest
= tgt
->page
;
888 struct dma_async_tx_descriptor
*tx
;
889 struct async_submit_ctl submit
;
892 pr_debug("%s: stripe %llu block: %d\n",
893 __func__
, (unsigned long long)sh
->sector
, target
);
894 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
896 for (i
= disks
; i
--; )
898 xor_srcs
[count
++] = sh
->dev
[i
].page
;
900 atomic_inc(&sh
->count
);
902 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
, NULL
,
903 ops_complete_compute
, sh
, to_addr_conv(sh
, percpu
));
904 if (unlikely(count
== 1))
905 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
907 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
912 /* set_syndrome_sources - populate source buffers for gen_syndrome
913 * @srcs - (struct page *) array of size sh->disks
914 * @sh - stripe_head to parse
916 * Populates srcs in proper layout order for the stripe and returns the
917 * 'count' of sources to be used in a call to async_gen_syndrome. The P
918 * destination buffer is recorded in srcs[count] and the Q destination
919 * is recorded in srcs[count+1]].
921 static int set_syndrome_sources(struct page
**srcs
, struct stripe_head
*sh
)
923 int disks
= sh
->disks
;
924 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
925 int d0_idx
= raid6_d0(sh
);
929 for (i
= 0; i
< disks
; i
++)
935 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
937 srcs
[slot
] = sh
->dev
[i
].page
;
938 i
= raid6_next_disk(i
, disks
);
939 } while (i
!= d0_idx
);
941 return syndrome_disks
;
944 static struct dma_async_tx_descriptor
*
945 ops_run_compute6_1(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
947 int disks
= sh
->disks
;
948 struct page
**blocks
= percpu
->scribble
;
950 int qd_idx
= sh
->qd_idx
;
951 struct dma_async_tx_descriptor
*tx
;
952 struct async_submit_ctl submit
;
958 if (sh
->ops
.target
< 0)
959 target
= sh
->ops
.target2
;
960 else if (sh
->ops
.target2
< 0)
961 target
= sh
->ops
.target
;
963 /* we should only have one valid target */
966 pr_debug("%s: stripe %llu block: %d\n",
967 __func__
, (unsigned long long)sh
->sector
, target
);
969 tgt
= &sh
->dev
[target
];
970 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
973 atomic_inc(&sh
->count
);
975 if (target
== qd_idx
) {
976 count
= set_syndrome_sources(blocks
, sh
);
977 blocks
[count
] = NULL
; /* regenerating p is not necessary */
978 BUG_ON(blocks
[count
+1] != dest
); /* q should already be set */
979 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
980 ops_complete_compute
, sh
,
981 to_addr_conv(sh
, percpu
));
982 tx
= async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
984 /* Compute any data- or p-drive using XOR */
986 for (i
= disks
; i
-- ; ) {
987 if (i
== target
|| i
== qd_idx
)
989 blocks
[count
++] = sh
->dev
[i
].page
;
992 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
993 NULL
, ops_complete_compute
, sh
,
994 to_addr_conv(sh
, percpu
));
995 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
, &submit
);
1001 static struct dma_async_tx_descriptor
*
1002 ops_run_compute6_2(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1004 int i
, count
, disks
= sh
->disks
;
1005 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1006 int d0_idx
= raid6_d0(sh
);
1007 int faila
= -1, failb
= -1;
1008 int target
= sh
->ops
.target
;
1009 int target2
= sh
->ops
.target2
;
1010 struct r5dev
*tgt
= &sh
->dev
[target
];
1011 struct r5dev
*tgt2
= &sh
->dev
[target2
];
1012 struct dma_async_tx_descriptor
*tx
;
1013 struct page
**blocks
= percpu
->scribble
;
1014 struct async_submit_ctl submit
;
1016 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1017 __func__
, (unsigned long long)sh
->sector
, target
, target2
);
1018 BUG_ON(target
< 0 || target2
< 0);
1019 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
1020 BUG_ON(!test_bit(R5_Wantcompute
, &tgt2
->flags
));
1022 /* we need to open-code set_syndrome_sources to handle the
1023 * slot number conversion for 'faila' and 'failb'
1025 for (i
= 0; i
< disks
; i
++)
1030 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1032 blocks
[slot
] = sh
->dev
[i
].page
;
1038 i
= raid6_next_disk(i
, disks
);
1039 } while (i
!= d0_idx
);
1041 BUG_ON(faila
== failb
);
1044 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1045 __func__
, (unsigned long long)sh
->sector
, faila
, failb
);
1047 atomic_inc(&sh
->count
);
1049 if (failb
== syndrome_disks
+1) {
1050 /* Q disk is one of the missing disks */
1051 if (faila
== syndrome_disks
) {
1052 /* Missing P+Q, just recompute */
1053 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1054 ops_complete_compute
, sh
,
1055 to_addr_conv(sh
, percpu
));
1056 return async_gen_syndrome(blocks
, 0, syndrome_disks
+2,
1057 STRIPE_SIZE
, &submit
);
1061 int qd_idx
= sh
->qd_idx
;
1063 /* Missing D+Q: recompute D from P, then recompute Q */
1064 if (target
== qd_idx
)
1065 data_target
= target2
;
1067 data_target
= target
;
1070 for (i
= disks
; i
-- ; ) {
1071 if (i
== data_target
|| i
== qd_idx
)
1073 blocks
[count
++] = sh
->dev
[i
].page
;
1075 dest
= sh
->dev
[data_target
].page
;
1076 init_async_submit(&submit
,
1077 ASYNC_TX_FENCE
|ASYNC_TX_XOR_ZERO_DST
,
1079 to_addr_conv(sh
, percpu
));
1080 tx
= async_xor(dest
, blocks
, 0, count
, STRIPE_SIZE
,
1083 count
= set_syndrome_sources(blocks
, sh
);
1084 init_async_submit(&submit
, ASYNC_TX_FENCE
, tx
,
1085 ops_complete_compute
, sh
,
1086 to_addr_conv(sh
, percpu
));
1087 return async_gen_syndrome(blocks
, 0, count
+2,
1088 STRIPE_SIZE
, &submit
);
1091 init_async_submit(&submit
, ASYNC_TX_FENCE
, NULL
,
1092 ops_complete_compute
, sh
,
1093 to_addr_conv(sh
, percpu
));
1094 if (failb
== syndrome_disks
) {
1095 /* We're missing D+P. */
1096 return async_raid6_datap_recov(syndrome_disks
+2,
1100 /* We're missing D+D. */
1101 return async_raid6_2data_recov(syndrome_disks
+2,
1102 STRIPE_SIZE
, faila
, failb
,
1109 static void ops_complete_prexor(void *stripe_head_ref
)
1111 struct stripe_head
*sh
= stripe_head_ref
;
1113 pr_debug("%s: stripe %llu\n", __func__
,
1114 (unsigned long long)sh
->sector
);
1117 static struct dma_async_tx_descriptor
*
1118 ops_run_prexor(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1119 struct dma_async_tx_descriptor
*tx
)
1121 int disks
= sh
->disks
;
1122 struct page
**xor_srcs
= percpu
->scribble
;
1123 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1124 struct async_submit_ctl submit
;
1126 /* existing parity data subtracted */
1127 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1129 pr_debug("%s: stripe %llu\n", __func__
,
1130 (unsigned long long)sh
->sector
);
1132 for (i
= disks
; i
--; ) {
1133 struct r5dev
*dev
= &sh
->dev
[i
];
1134 /* Only process blocks that are known to be uptodate */
1135 if (test_bit(R5_Wantdrain
, &dev
->flags
))
1136 xor_srcs
[count
++] = dev
->page
;
1139 init_async_submit(&submit
, ASYNC_TX_FENCE
|ASYNC_TX_XOR_DROP_DST
, tx
,
1140 ops_complete_prexor
, sh
, to_addr_conv(sh
, percpu
));
1141 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1146 static struct dma_async_tx_descriptor
*
1147 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
1149 int disks
= sh
->disks
;
1152 pr_debug("%s: stripe %llu\n", __func__
,
1153 (unsigned long long)sh
->sector
);
1155 for (i
= disks
; i
--; ) {
1156 struct r5dev
*dev
= &sh
->dev
[i
];
1159 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
1162 spin_lock_irq(&sh
->stripe_lock
);
1163 chosen
= dev
->towrite
;
1164 dev
->towrite
= NULL
;
1165 BUG_ON(dev
->written
);
1166 wbi
= dev
->written
= chosen
;
1167 spin_unlock_irq(&sh
->stripe_lock
);
1169 while (wbi
&& wbi
->bi_sector
<
1170 dev
->sector
+ STRIPE_SECTORS
) {
1171 if (wbi
->bi_rw
& REQ_FUA
)
1172 set_bit(R5_WantFUA
, &dev
->flags
);
1173 if (wbi
->bi_rw
& REQ_SYNC
)
1174 set_bit(R5_SyncIO
, &dev
->flags
);
1175 if (wbi
->bi_rw
& REQ_DISCARD
)
1176 set_bit(R5_Discard
, &dev
->flags
);
1178 tx
= async_copy_data(1, wbi
, dev
->page
,
1180 wbi
= r5_next_bio(wbi
, dev
->sector
);
1188 static void ops_complete_reconstruct(void *stripe_head_ref
)
1190 struct stripe_head
*sh
= stripe_head_ref
;
1191 int disks
= sh
->disks
;
1192 int pd_idx
= sh
->pd_idx
;
1193 int qd_idx
= sh
->qd_idx
;
1195 bool fua
= false, sync
= false, discard
= false;
1197 pr_debug("%s: stripe %llu\n", __func__
,
1198 (unsigned long long)sh
->sector
);
1200 for (i
= disks
; i
--; ) {
1201 fua
|= test_bit(R5_WantFUA
, &sh
->dev
[i
].flags
);
1202 sync
|= test_bit(R5_SyncIO
, &sh
->dev
[i
].flags
);
1203 discard
|= test_bit(R5_Discard
, &sh
->dev
[i
].flags
);
1206 for (i
= disks
; i
--; ) {
1207 struct r5dev
*dev
= &sh
->dev
[i
];
1209 if (dev
->written
|| i
== pd_idx
|| i
== qd_idx
) {
1211 set_bit(R5_UPTODATE
, &dev
->flags
);
1213 set_bit(R5_WantFUA
, &dev
->flags
);
1215 set_bit(R5_SyncIO
, &dev
->flags
);
1219 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
1220 sh
->reconstruct_state
= reconstruct_state_drain_result
;
1221 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
1222 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
1224 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
1225 sh
->reconstruct_state
= reconstruct_state_result
;
1228 set_bit(STRIPE_HANDLE
, &sh
->state
);
1233 ops_run_reconstruct5(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1234 struct dma_async_tx_descriptor
*tx
)
1236 int disks
= sh
->disks
;
1237 struct page
**xor_srcs
= percpu
->scribble
;
1238 struct async_submit_ctl submit
;
1239 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
1240 struct page
*xor_dest
;
1242 unsigned long flags
;
1244 pr_debug("%s: stripe %llu\n", __func__
,
1245 (unsigned long long)sh
->sector
);
1247 for (i
= 0; i
< sh
->disks
; i
++) {
1250 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1253 if (i
>= sh
->disks
) {
1254 atomic_inc(&sh
->count
);
1255 set_bit(R5_Discard
, &sh
->dev
[pd_idx
].flags
);
1256 ops_complete_reconstruct(sh
);
1259 /* check if prexor is active which means only process blocks
1260 * that are part of a read-modify-write (written)
1262 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
1264 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
1265 for (i
= disks
; i
--; ) {
1266 struct r5dev
*dev
= &sh
->dev
[i
];
1268 xor_srcs
[count
++] = dev
->page
;
1271 xor_dest
= sh
->dev
[pd_idx
].page
;
1272 for (i
= disks
; i
--; ) {
1273 struct r5dev
*dev
= &sh
->dev
[i
];
1275 xor_srcs
[count
++] = dev
->page
;
1279 /* 1/ if we prexor'd then the dest is reused as a source
1280 * 2/ if we did not prexor then we are redoing the parity
1281 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1282 * for the synchronous xor case
1284 flags
= ASYNC_TX_ACK
|
1285 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
1287 atomic_inc(&sh
->count
);
1289 init_async_submit(&submit
, flags
, tx
, ops_complete_reconstruct
, sh
,
1290 to_addr_conv(sh
, percpu
));
1291 if (unlikely(count
== 1))
1292 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
, &submit
);
1294 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
, &submit
);
1298 ops_run_reconstruct6(struct stripe_head
*sh
, struct raid5_percpu
*percpu
,
1299 struct dma_async_tx_descriptor
*tx
)
1301 struct async_submit_ctl submit
;
1302 struct page
**blocks
= percpu
->scribble
;
1305 pr_debug("%s: stripe %llu\n", __func__
, (unsigned long long)sh
->sector
);
1307 for (i
= 0; i
< sh
->disks
; i
++) {
1308 if (sh
->pd_idx
== i
|| sh
->qd_idx
== i
)
1310 if (!test_bit(R5_Discard
, &sh
->dev
[i
].flags
))
1313 if (i
>= sh
->disks
) {
1314 atomic_inc(&sh
->count
);
1315 set_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
);
1316 set_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
);
1317 ops_complete_reconstruct(sh
);
1321 count
= set_syndrome_sources(blocks
, sh
);
1323 atomic_inc(&sh
->count
);
1325 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_reconstruct
,
1326 sh
, to_addr_conv(sh
, percpu
));
1327 async_gen_syndrome(blocks
, 0, count
+2, STRIPE_SIZE
, &submit
);
1330 static void ops_complete_check(void *stripe_head_ref
)
1332 struct stripe_head
*sh
= stripe_head_ref
;
1334 pr_debug("%s: stripe %llu\n", __func__
,
1335 (unsigned long long)sh
->sector
);
1337 sh
->check_state
= check_state_check_result
;
1338 set_bit(STRIPE_HANDLE
, &sh
->state
);
1342 static void ops_run_check_p(struct stripe_head
*sh
, struct raid5_percpu
*percpu
)
1344 int disks
= sh
->disks
;
1345 int pd_idx
= sh
->pd_idx
;
1346 int qd_idx
= sh
->qd_idx
;
1347 struct page
*xor_dest
;
1348 struct page
**xor_srcs
= percpu
->scribble
;
1349 struct dma_async_tx_descriptor
*tx
;
1350 struct async_submit_ctl submit
;
1354 pr_debug("%s: stripe %llu\n", __func__
,
1355 (unsigned long long)sh
->sector
);
1358 xor_dest
= sh
->dev
[pd_idx
].page
;
1359 xor_srcs
[count
++] = xor_dest
;
1360 for (i
= disks
; i
--; ) {
1361 if (i
== pd_idx
|| i
== qd_idx
)
1363 xor_srcs
[count
++] = sh
->dev
[i
].page
;
1366 init_async_submit(&submit
, 0, NULL
, NULL
, NULL
,
1367 to_addr_conv(sh
, percpu
));
1368 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
1369 &sh
->ops
.zero_sum_result
, &submit
);
1371 atomic_inc(&sh
->count
);
1372 init_async_submit(&submit
, ASYNC_TX_ACK
, tx
, ops_complete_check
, sh
, NULL
);
1373 tx
= async_trigger_callback(&submit
);
1376 static void ops_run_check_pq(struct stripe_head
*sh
, struct raid5_percpu
*percpu
, int checkp
)
1378 struct page
**srcs
= percpu
->scribble
;
1379 struct async_submit_ctl submit
;
1382 pr_debug("%s: stripe %llu checkp: %d\n", __func__
,
1383 (unsigned long long)sh
->sector
, checkp
);
1385 count
= set_syndrome_sources(srcs
, sh
);
1389 atomic_inc(&sh
->count
);
1390 init_async_submit(&submit
, ASYNC_TX_ACK
, NULL
, ops_complete_check
,
1391 sh
, to_addr_conv(sh
, percpu
));
1392 async_syndrome_val(srcs
, 0, count
+2, STRIPE_SIZE
,
1393 &sh
->ops
.zero_sum_result
, percpu
->spare_page
, &submit
);
1396 static void __raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1398 int overlap_clear
= 0, i
, disks
= sh
->disks
;
1399 struct dma_async_tx_descriptor
*tx
= NULL
;
1400 struct r5conf
*conf
= sh
->raid_conf
;
1401 int level
= conf
->level
;
1402 struct raid5_percpu
*percpu
;
1406 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1407 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
1408 ops_run_biofill(sh
);
1412 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
1414 tx
= ops_run_compute5(sh
, percpu
);
1416 if (sh
->ops
.target2
< 0 || sh
->ops
.target
< 0)
1417 tx
= ops_run_compute6_1(sh
, percpu
);
1419 tx
= ops_run_compute6_2(sh
, percpu
);
1421 /* terminate the chain if reconstruct is not set to be run */
1422 if (tx
&& !test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
))
1426 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
1427 tx
= ops_run_prexor(sh
, percpu
, tx
);
1429 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
1430 tx
= ops_run_biodrain(sh
, tx
);
1434 if (test_bit(STRIPE_OP_RECONSTRUCT
, &ops_request
)) {
1436 ops_run_reconstruct5(sh
, percpu
, tx
);
1438 ops_run_reconstruct6(sh
, percpu
, tx
);
1441 if (test_bit(STRIPE_OP_CHECK
, &ops_request
)) {
1442 if (sh
->check_state
== check_state_run
)
1443 ops_run_check_p(sh
, percpu
);
1444 else if (sh
->check_state
== check_state_run_q
)
1445 ops_run_check_pq(sh
, percpu
, 0);
1446 else if (sh
->check_state
== check_state_run_pq
)
1447 ops_run_check_pq(sh
, percpu
, 1);
1453 for (i
= disks
; i
--; ) {
1454 struct r5dev
*dev
= &sh
->dev
[i
];
1455 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
1456 wake_up(&sh
->raid_conf
->wait_for_overlap
);
1461 #ifdef CONFIG_MULTICORE_RAID456
1462 static void async_run_ops(void *param
, async_cookie_t cookie
)
1464 struct stripe_head
*sh
= param
;
1465 unsigned long ops_request
= sh
->ops
.request
;
1467 clear_bit_unlock(STRIPE_OPS_REQ_PENDING
, &sh
->state
);
1468 wake_up(&sh
->ops
.wait_for_ops
);
1470 __raid_run_ops(sh
, ops_request
);
1474 static void raid_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
1476 /* since handle_stripe can be called outside of raid5d context
1477 * we need to ensure sh->ops.request is de-staged before another
1480 wait_event(sh
->ops
.wait_for_ops
,
1481 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING
, &sh
->state
));
1482 sh
->ops
.request
= ops_request
;
1484 atomic_inc(&sh
->count
);
1485 async_schedule(async_run_ops
, sh
);
1488 #define raid_run_ops __raid_run_ops
1491 static int grow_one_stripe(struct r5conf
*conf
)
1493 struct stripe_head
*sh
;
1494 sh
= kmem_cache_zalloc(conf
->slab_cache
, GFP_KERNEL
);
1498 sh
->raid_conf
= conf
;
1499 #ifdef CONFIG_MULTICORE_RAID456
1500 init_waitqueue_head(&sh
->ops
.wait_for_ops
);
1503 spin_lock_init(&sh
->stripe_lock
);
1505 if (grow_buffers(sh
)) {
1507 kmem_cache_free(conf
->slab_cache
, sh
);
1510 /* we just created an active stripe so... */
1511 atomic_set(&sh
->count
, 1);
1512 atomic_inc(&conf
->active_stripes
);
1513 INIT_LIST_HEAD(&sh
->lru
);
1518 static int grow_stripes(struct r5conf
*conf
, int num
)
1520 struct kmem_cache
*sc
;
1521 int devs
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
1523 if (conf
->mddev
->gendisk
)
1524 sprintf(conf
->cache_name
[0],
1525 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
1527 sprintf(conf
->cache_name
[0],
1528 "raid%d-%p", conf
->level
, conf
->mddev
);
1529 sprintf(conf
->cache_name
[1], "%s-alt", conf
->cache_name
[0]);
1531 conf
->active_name
= 0;
1532 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
1533 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
1537 conf
->slab_cache
= sc
;
1538 conf
->pool_size
= devs
;
1540 if (!grow_one_stripe(conf
))
1546 * scribble_len - return the required size of the scribble region
1547 * @num - total number of disks in the array
1549 * The size must be enough to contain:
1550 * 1/ a struct page pointer for each device in the array +2
1551 * 2/ room to convert each entry in (1) to its corresponding dma
1552 * (dma_map_page()) or page (page_address()) address.
1554 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1555 * calculate over all devices (not just the data blocks), using zeros in place
1556 * of the P and Q blocks.
1558 static size_t scribble_len(int num
)
1562 len
= sizeof(struct page
*) * (num
+2) + sizeof(addr_conv_t
) * (num
+2);
1567 static int resize_stripes(struct r5conf
*conf
, int newsize
)
1569 /* Make all the stripes able to hold 'newsize' devices.
1570 * New slots in each stripe get 'page' set to a new page.
1572 * This happens in stages:
1573 * 1/ create a new kmem_cache and allocate the required number of
1575 * 2/ gather all the old stripe_heads and tranfer the pages across
1576 * to the new stripe_heads. This will have the side effect of
1577 * freezing the array as once all stripe_heads have been collected,
1578 * no IO will be possible. Old stripe heads are freed once their
1579 * pages have been transferred over, and the old kmem_cache is
1580 * freed when all stripes are done.
1581 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1582 * we simple return a failre status - no need to clean anything up.
1583 * 4/ allocate new pages for the new slots in the new stripe_heads.
1584 * If this fails, we don't bother trying the shrink the
1585 * stripe_heads down again, we just leave them as they are.
1586 * As each stripe_head is processed the new one is released into
1589 * Once step2 is started, we cannot afford to wait for a write,
1590 * so we use GFP_NOIO allocations.
1592 struct stripe_head
*osh
, *nsh
;
1593 LIST_HEAD(newstripes
);
1594 struct disk_info
*ndisks
;
1597 struct kmem_cache
*sc
;
1600 if (newsize
<= conf
->pool_size
)
1601 return 0; /* never bother to shrink */
1603 err
= md_allow_write(conf
->mddev
);
1608 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1609 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1614 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1615 nsh
= kmem_cache_zalloc(sc
, GFP_KERNEL
);
1619 nsh
->raid_conf
= conf
;
1620 #ifdef CONFIG_MULTICORE_RAID456
1621 init_waitqueue_head(&nsh
->ops
.wait_for_ops
);
1624 list_add(&nsh
->lru
, &newstripes
);
1627 /* didn't get enough, give up */
1628 while (!list_empty(&newstripes
)) {
1629 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1630 list_del(&nsh
->lru
);
1631 kmem_cache_free(sc
, nsh
);
1633 kmem_cache_destroy(sc
);
1636 /* Step 2 - Must use GFP_NOIO now.
1637 * OK, we have enough stripes, start collecting inactive
1638 * stripes and copying them over
1640 list_for_each_entry(nsh
, &newstripes
, lru
) {
1641 spin_lock_irq(&conf
->device_lock
);
1642 wait_event_lock_irq(conf
->wait_for_stripe
,
1643 !list_empty(&conf
->inactive_list
),
1646 osh
= get_free_stripe(conf
);
1647 spin_unlock_irq(&conf
->device_lock
);
1648 atomic_set(&nsh
->count
, 1);
1649 for(i
=0; i
<conf
->pool_size
; i
++)
1650 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1651 for( ; i
<newsize
; i
++)
1652 nsh
->dev
[i
].page
= NULL
;
1653 kmem_cache_free(conf
->slab_cache
, osh
);
1655 kmem_cache_destroy(conf
->slab_cache
);
1658 * At this point, we are holding all the stripes so the array
1659 * is completely stalled, so now is a good time to resize
1660 * conf->disks and the scribble region
1662 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1664 for (i
=0; i
<conf
->raid_disks
; i
++)
1665 ndisks
[i
] = conf
->disks
[i
];
1667 conf
->disks
= ndisks
;
1672 conf
->scribble_len
= scribble_len(newsize
);
1673 for_each_present_cpu(cpu
) {
1674 struct raid5_percpu
*percpu
;
1677 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
1678 scribble
= kmalloc(conf
->scribble_len
, GFP_NOIO
);
1681 kfree(percpu
->scribble
);
1682 percpu
->scribble
= scribble
;
1690 /* Step 4, return new stripes to service */
1691 while(!list_empty(&newstripes
)) {
1692 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1693 list_del_init(&nsh
->lru
);
1695 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1696 if (nsh
->dev
[i
].page
== NULL
) {
1697 struct page
*p
= alloc_page(GFP_NOIO
);
1698 nsh
->dev
[i
].page
= p
;
1702 release_stripe(nsh
);
1704 /* critical section pass, GFP_NOIO no longer needed */
1706 conf
->slab_cache
= sc
;
1707 conf
->active_name
= 1-conf
->active_name
;
1708 conf
->pool_size
= newsize
;
1712 static int drop_one_stripe(struct r5conf
*conf
)
1714 struct stripe_head
*sh
;
1716 spin_lock_irq(&conf
->device_lock
);
1717 sh
= get_free_stripe(conf
);
1718 spin_unlock_irq(&conf
->device_lock
);
1721 BUG_ON(atomic_read(&sh
->count
));
1723 kmem_cache_free(conf
->slab_cache
, sh
);
1724 atomic_dec(&conf
->active_stripes
);
1728 static void shrink_stripes(struct r5conf
*conf
)
1730 while (drop_one_stripe(conf
))
1733 if (conf
->slab_cache
)
1734 kmem_cache_destroy(conf
->slab_cache
);
1735 conf
->slab_cache
= NULL
;
1738 static void raid5_end_read_request(struct bio
* bi
, int error
)
1740 struct stripe_head
*sh
= bi
->bi_private
;
1741 struct r5conf
*conf
= sh
->raid_conf
;
1742 int disks
= sh
->disks
, i
;
1743 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1744 char b
[BDEVNAME_SIZE
];
1745 struct md_rdev
*rdev
= NULL
;
1748 for (i
=0 ; i
<disks
; i
++)
1749 if (bi
== &sh
->dev
[i
].req
)
1752 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1753 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1759 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1760 /* If replacement finished while this request was outstanding,
1761 * 'replacement' might be NULL already.
1762 * In that case it moved down to 'rdev'.
1763 * rdev is not removed until all requests are finished.
1765 rdev
= conf
->disks
[i
].replacement
;
1767 rdev
= conf
->disks
[i
].rdev
;
1769 if (use_new_offset(conf
, sh
))
1770 s
= sh
->sector
+ rdev
->new_data_offset
;
1772 s
= sh
->sector
+ rdev
->data_offset
;
1774 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1775 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1776 /* Note that this cannot happen on a
1777 * replacement device. We just fail those on
1782 "md/raid:%s: read error corrected"
1783 " (%lu sectors at %llu on %s)\n",
1784 mdname(conf
->mddev
), STRIPE_SECTORS
,
1785 (unsigned long long)s
,
1786 bdevname(rdev
->bdev
, b
));
1787 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
1788 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1789 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1790 } else if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
))
1791 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1793 if (atomic_read(&rdev
->read_errors
))
1794 atomic_set(&rdev
->read_errors
, 0);
1796 const char *bdn
= bdevname(rdev
->bdev
, b
);
1800 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1801 atomic_inc(&rdev
->read_errors
);
1802 if (test_bit(R5_ReadRepl
, &sh
->dev
[i
].flags
))
1805 "md/raid:%s: read error on replacement device "
1806 "(sector %llu on %s).\n",
1807 mdname(conf
->mddev
),
1808 (unsigned long long)s
,
1810 else if (conf
->mddev
->degraded
>= conf
->max_degraded
) {
1814 "md/raid:%s: read error not correctable "
1815 "(sector %llu on %s).\n",
1816 mdname(conf
->mddev
),
1817 (unsigned long long)s
,
1819 } else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
)) {
1824 "md/raid:%s: read error NOT corrected!! "
1825 "(sector %llu on %s).\n",
1826 mdname(conf
->mddev
),
1827 (unsigned long long)s
,
1829 } else if (atomic_read(&rdev
->read_errors
)
1830 > conf
->max_nr_stripes
)
1832 "md/raid:%s: Too many read errors, failing device %s.\n",
1833 mdname(conf
->mddev
), bdn
);
1837 if (test_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
)) {
1838 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1839 clear_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1841 set_bit(R5_ReadNoMerge
, &sh
->dev
[i
].flags
);
1843 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1844 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1846 && test_bit(In_sync
, &rdev
->flags
)
1847 && rdev_set_badblocks(
1848 rdev
, sh
->sector
, STRIPE_SECTORS
, 0)))
1849 md_error(conf
->mddev
, rdev
);
1852 rdev_dec_pending(rdev
, conf
->mddev
);
1853 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1854 set_bit(STRIPE_HANDLE
, &sh
->state
);
1858 static void raid5_end_write_request(struct bio
*bi
, int error
)
1860 struct stripe_head
*sh
= bi
->bi_private
;
1861 struct r5conf
*conf
= sh
->raid_conf
;
1862 int disks
= sh
->disks
, i
;
1863 struct md_rdev
*uninitialized_var(rdev
);
1864 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1867 int replacement
= 0;
1869 for (i
= 0 ; i
< disks
; i
++) {
1870 if (bi
== &sh
->dev
[i
].req
) {
1871 rdev
= conf
->disks
[i
].rdev
;
1874 if (bi
== &sh
->dev
[i
].rreq
) {
1875 rdev
= conf
->disks
[i
].replacement
;
1879 /* rdev was removed and 'replacement'
1880 * replaced it. rdev is not removed
1881 * until all requests are finished.
1883 rdev
= conf
->disks
[i
].rdev
;
1887 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1888 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1897 md_error(conf
->mddev
, rdev
);
1898 else if (is_badblock(rdev
, sh
->sector
,
1900 &first_bad
, &bad_sectors
))
1901 set_bit(R5_MadeGoodRepl
, &sh
->dev
[i
].flags
);
1904 set_bit(WriteErrorSeen
, &rdev
->flags
);
1905 set_bit(R5_WriteError
, &sh
->dev
[i
].flags
);
1906 if (!test_and_set_bit(WantReplacement
, &rdev
->flags
))
1907 set_bit(MD_RECOVERY_NEEDED
,
1908 &rdev
->mddev
->recovery
);
1909 } else if (is_badblock(rdev
, sh
->sector
,
1911 &first_bad
, &bad_sectors
))
1912 set_bit(R5_MadeGood
, &sh
->dev
[i
].flags
);
1914 rdev_dec_pending(rdev
, conf
->mddev
);
1916 if (!test_and_clear_bit(R5_DOUBLE_LOCKED
, &sh
->dev
[i
].flags
))
1917 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1918 set_bit(STRIPE_HANDLE
, &sh
->state
);
1922 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1924 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1926 struct r5dev
*dev
= &sh
->dev
[i
];
1928 bio_init(&dev
->req
);
1929 dev
->req
.bi_io_vec
= &dev
->vec
;
1931 dev
->req
.bi_max_vecs
++;
1932 dev
->req
.bi_private
= sh
;
1933 dev
->vec
.bv_page
= dev
->page
;
1935 bio_init(&dev
->rreq
);
1936 dev
->rreq
.bi_io_vec
= &dev
->rvec
;
1937 dev
->rreq
.bi_vcnt
++;
1938 dev
->rreq
.bi_max_vecs
++;
1939 dev
->rreq
.bi_private
= sh
;
1940 dev
->rvec
.bv_page
= dev
->page
;
1943 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1946 static void error(struct mddev
*mddev
, struct md_rdev
*rdev
)
1948 char b
[BDEVNAME_SIZE
];
1949 struct r5conf
*conf
= mddev
->private;
1950 unsigned long flags
;
1951 pr_debug("raid456: error called\n");
1953 spin_lock_irqsave(&conf
->device_lock
, flags
);
1954 clear_bit(In_sync
, &rdev
->flags
);
1955 mddev
->degraded
= calc_degraded(conf
);
1956 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1957 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1959 set_bit(Blocked
, &rdev
->flags
);
1960 set_bit(Faulty
, &rdev
->flags
);
1961 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1963 "md/raid:%s: Disk failure on %s, disabling device.\n"
1964 "md/raid:%s: Operation continuing on %d devices.\n",
1966 bdevname(rdev
->bdev
, b
),
1968 conf
->raid_disks
- mddev
->degraded
);
1972 * Input: a 'big' sector number,
1973 * Output: index of the data and parity disk, and the sector # in them.
1975 static sector_t
raid5_compute_sector(struct r5conf
*conf
, sector_t r_sector
,
1976 int previous
, int *dd_idx
,
1977 struct stripe_head
*sh
)
1979 sector_t stripe
, stripe2
;
1980 sector_t chunk_number
;
1981 unsigned int chunk_offset
;
1984 sector_t new_sector
;
1985 int algorithm
= previous
? conf
->prev_algo
1987 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
1988 : conf
->chunk_sectors
;
1989 int raid_disks
= previous
? conf
->previous_raid_disks
1991 int data_disks
= raid_disks
- conf
->max_degraded
;
1993 /* First compute the information on this sector */
1996 * Compute the chunk number and the sector offset inside the chunk
1998 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1999 chunk_number
= r_sector
;
2002 * Compute the stripe number
2004 stripe
= chunk_number
;
2005 *dd_idx
= sector_div(stripe
, data_disks
);
2008 * Select the parity disk based on the user selected algorithm.
2010 pd_idx
= qd_idx
= -1;
2011 switch(conf
->level
) {
2013 pd_idx
= data_disks
;
2016 switch (algorithm
) {
2017 case ALGORITHM_LEFT_ASYMMETRIC
:
2018 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2019 if (*dd_idx
>= pd_idx
)
2022 case ALGORITHM_RIGHT_ASYMMETRIC
:
2023 pd_idx
= sector_div(stripe2
, raid_disks
);
2024 if (*dd_idx
>= pd_idx
)
2027 case ALGORITHM_LEFT_SYMMETRIC
:
2028 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
);
2029 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2031 case ALGORITHM_RIGHT_SYMMETRIC
:
2032 pd_idx
= sector_div(stripe2
, raid_disks
);
2033 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2035 case ALGORITHM_PARITY_0
:
2039 case ALGORITHM_PARITY_N
:
2040 pd_idx
= data_disks
;
2048 switch (algorithm
) {
2049 case ALGORITHM_LEFT_ASYMMETRIC
:
2050 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2051 qd_idx
= pd_idx
+ 1;
2052 if (pd_idx
== raid_disks
-1) {
2053 (*dd_idx
)++; /* Q D D D P */
2055 } else if (*dd_idx
>= pd_idx
)
2056 (*dd_idx
) += 2; /* D D P Q D */
2058 case ALGORITHM_RIGHT_ASYMMETRIC
:
2059 pd_idx
= sector_div(stripe2
, raid_disks
);
2060 qd_idx
= pd_idx
+ 1;
2061 if (pd_idx
== raid_disks
-1) {
2062 (*dd_idx
)++; /* Q D D D P */
2064 } else if (*dd_idx
>= pd_idx
)
2065 (*dd_idx
) += 2; /* D D P Q D */
2067 case ALGORITHM_LEFT_SYMMETRIC
:
2068 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2069 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2070 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2072 case ALGORITHM_RIGHT_SYMMETRIC
:
2073 pd_idx
= sector_div(stripe2
, raid_disks
);
2074 qd_idx
= (pd_idx
+ 1) % raid_disks
;
2075 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
2078 case ALGORITHM_PARITY_0
:
2083 case ALGORITHM_PARITY_N
:
2084 pd_idx
= data_disks
;
2085 qd_idx
= data_disks
+ 1;
2088 case ALGORITHM_ROTATING_ZERO_RESTART
:
2089 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2090 * of blocks for computing Q is different.
2092 pd_idx
= sector_div(stripe2
, raid_disks
);
2093 qd_idx
= pd_idx
+ 1;
2094 if (pd_idx
== raid_disks
-1) {
2095 (*dd_idx
)++; /* Q D D D P */
2097 } else if (*dd_idx
>= pd_idx
)
2098 (*dd_idx
) += 2; /* D D P Q D */
2102 case ALGORITHM_ROTATING_N_RESTART
:
2103 /* Same a left_asymmetric, by first stripe is
2104 * D D D P Q rather than
2108 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2109 qd_idx
= pd_idx
+ 1;
2110 if (pd_idx
== raid_disks
-1) {
2111 (*dd_idx
)++; /* Q D D D P */
2113 } else if (*dd_idx
>= pd_idx
)
2114 (*dd_idx
) += 2; /* D D P Q D */
2118 case ALGORITHM_ROTATING_N_CONTINUE
:
2119 /* Same as left_symmetric but Q is before P */
2120 pd_idx
= raid_disks
- 1 - sector_div(stripe2
, raid_disks
);
2121 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
2122 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
2126 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2127 /* RAID5 left_asymmetric, with Q on last device */
2128 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2129 if (*dd_idx
>= pd_idx
)
2131 qd_idx
= raid_disks
- 1;
2134 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2135 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2136 if (*dd_idx
>= pd_idx
)
2138 qd_idx
= raid_disks
- 1;
2141 case ALGORITHM_LEFT_SYMMETRIC_6
:
2142 pd_idx
= data_disks
- sector_div(stripe2
, raid_disks
-1);
2143 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2144 qd_idx
= raid_disks
- 1;
2147 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2148 pd_idx
= sector_div(stripe2
, raid_disks
-1);
2149 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
2150 qd_idx
= raid_disks
- 1;
2153 case ALGORITHM_PARITY_0_6
:
2156 qd_idx
= raid_disks
- 1;
2166 sh
->pd_idx
= pd_idx
;
2167 sh
->qd_idx
= qd_idx
;
2168 sh
->ddf_layout
= ddf_layout
;
2171 * Finally, compute the new sector number
2173 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
2178 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
2180 struct r5conf
*conf
= sh
->raid_conf
;
2181 int raid_disks
= sh
->disks
;
2182 int data_disks
= raid_disks
- conf
->max_degraded
;
2183 sector_t new_sector
= sh
->sector
, check
;
2184 int sectors_per_chunk
= previous
? conf
->prev_chunk_sectors
2185 : conf
->chunk_sectors
;
2186 int algorithm
= previous
? conf
->prev_algo
2190 sector_t chunk_number
;
2191 int dummy1
, dd_idx
= i
;
2193 struct stripe_head sh2
;
2196 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
2197 stripe
= new_sector
;
2199 if (i
== sh
->pd_idx
)
2201 switch(conf
->level
) {
2204 switch (algorithm
) {
2205 case ALGORITHM_LEFT_ASYMMETRIC
:
2206 case ALGORITHM_RIGHT_ASYMMETRIC
:
2210 case ALGORITHM_LEFT_SYMMETRIC
:
2211 case ALGORITHM_RIGHT_SYMMETRIC
:
2214 i
-= (sh
->pd_idx
+ 1);
2216 case ALGORITHM_PARITY_0
:
2219 case ALGORITHM_PARITY_N
:
2226 if (i
== sh
->qd_idx
)
2227 return 0; /* It is the Q disk */
2228 switch (algorithm
) {
2229 case ALGORITHM_LEFT_ASYMMETRIC
:
2230 case ALGORITHM_RIGHT_ASYMMETRIC
:
2231 case ALGORITHM_ROTATING_ZERO_RESTART
:
2232 case ALGORITHM_ROTATING_N_RESTART
:
2233 if (sh
->pd_idx
== raid_disks
-1)
2234 i
--; /* Q D D D P */
2235 else if (i
> sh
->pd_idx
)
2236 i
-= 2; /* D D P Q D */
2238 case ALGORITHM_LEFT_SYMMETRIC
:
2239 case ALGORITHM_RIGHT_SYMMETRIC
:
2240 if (sh
->pd_idx
== raid_disks
-1)
2241 i
--; /* Q D D D P */
2246 i
-= (sh
->pd_idx
+ 2);
2249 case ALGORITHM_PARITY_0
:
2252 case ALGORITHM_PARITY_N
:
2254 case ALGORITHM_ROTATING_N_CONTINUE
:
2255 /* Like left_symmetric, but P is before Q */
2256 if (sh
->pd_idx
== 0)
2257 i
--; /* P D D D Q */
2262 i
-= (sh
->pd_idx
+ 1);
2265 case ALGORITHM_LEFT_ASYMMETRIC_6
:
2266 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
2270 case ALGORITHM_LEFT_SYMMETRIC_6
:
2271 case ALGORITHM_RIGHT_SYMMETRIC_6
:
2273 i
+= data_disks
+ 1;
2274 i
-= (sh
->pd_idx
+ 1);
2276 case ALGORITHM_PARITY_0_6
:
2285 chunk_number
= stripe
* data_disks
+ i
;
2286 r_sector
= chunk_number
* sectors_per_chunk
+ chunk_offset
;
2288 check
= raid5_compute_sector(conf
, r_sector
,
2289 previous
, &dummy1
, &sh2
);
2290 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
2291 || sh2
.qd_idx
!= sh
->qd_idx
) {
2292 printk(KERN_ERR
"md/raid:%s: compute_blocknr: map not correct\n",
2293 mdname(conf
->mddev
));
2301 schedule_reconstruction(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2302 int rcw
, int expand
)
2304 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
2305 struct r5conf
*conf
= sh
->raid_conf
;
2306 int level
= conf
->level
;
2309 /* if we are not expanding this is a proper write request, and
2310 * there will be bios with new data to be drained into the
2314 sh
->reconstruct_state
= reconstruct_state_drain_run
;
2315 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2317 sh
->reconstruct_state
= reconstruct_state_run
;
2319 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2321 for (i
= disks
; i
--; ) {
2322 struct r5dev
*dev
= &sh
->dev
[i
];
2325 set_bit(R5_LOCKED
, &dev
->flags
);
2326 set_bit(R5_Wantdrain
, &dev
->flags
);
2328 clear_bit(R5_UPTODATE
, &dev
->flags
);
2332 if (s
->locked
+ conf
->max_degraded
== disks
)
2333 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2334 atomic_inc(&conf
->pending_full_writes
);
2337 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
2338 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
2340 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
2341 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
2342 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
2343 set_bit(STRIPE_OP_RECONSTRUCT
, &s
->ops_request
);
2345 for (i
= disks
; i
--; ) {
2346 struct r5dev
*dev
= &sh
->dev
[i
];
2351 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2352 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2353 set_bit(R5_Wantdrain
, &dev
->flags
);
2354 set_bit(R5_LOCKED
, &dev
->flags
);
2355 clear_bit(R5_UPTODATE
, &dev
->flags
);
2361 /* keep the parity disk(s) locked while asynchronous operations
2364 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
2365 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
2369 int qd_idx
= sh
->qd_idx
;
2370 struct r5dev
*dev
= &sh
->dev
[qd_idx
];
2372 set_bit(R5_LOCKED
, &dev
->flags
);
2373 clear_bit(R5_UPTODATE
, &dev
->flags
);
2377 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2378 __func__
, (unsigned long long)sh
->sector
,
2379 s
->locked
, s
->ops_request
);
2383 * Each stripe/dev can have one or more bion attached.
2384 * toread/towrite point to the first in a chain.
2385 * The bi_next chain must be in order.
2387 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
2390 struct r5conf
*conf
= sh
->raid_conf
;
2393 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2394 (unsigned long long)bi
->bi_sector
,
2395 (unsigned long long)sh
->sector
);
2398 * If several bio share a stripe. The bio bi_phys_segments acts as a
2399 * reference count to avoid race. The reference count should already be
2400 * increased before this function is called (for example, in
2401 * make_request()), so other bio sharing this stripe will not free the
2402 * stripe. If a stripe is owned by one stripe, the stripe lock will
2405 spin_lock_irq(&sh
->stripe_lock
);
2407 bip
= &sh
->dev
[dd_idx
].towrite
;
2411 bip
= &sh
->dev
[dd_idx
].toread
;
2412 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
2413 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
2415 bip
= & (*bip
)->bi_next
;
2417 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
2420 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
2424 raid5_inc_bi_active_stripes(bi
);
2427 /* check if page is covered */
2428 sector_t sector
= sh
->dev
[dd_idx
].sector
;
2429 for (bi
=sh
->dev
[dd_idx
].towrite
;
2430 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
2431 bi
&& bi
->bi_sector
<= sector
;
2432 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
2433 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
2434 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
2436 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
2437 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
2439 spin_unlock_irq(&sh
->stripe_lock
);
2441 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2442 (unsigned long long)(*bip
)->bi_sector
,
2443 (unsigned long long)sh
->sector
, dd_idx
);
2445 if (conf
->mddev
->bitmap
&& firstwrite
) {
2446 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
2448 sh
->bm_seq
= conf
->seq_flush
+1;
2449 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
2454 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
2455 spin_unlock_irq(&sh
->stripe_lock
);
2459 static void end_reshape(struct r5conf
*conf
);
2461 static void stripe_set_idx(sector_t stripe
, struct r5conf
*conf
, int previous
,
2462 struct stripe_head
*sh
)
2464 int sectors_per_chunk
=
2465 previous
? conf
->prev_chunk_sectors
: conf
->chunk_sectors
;
2467 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2468 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2470 raid5_compute_sector(conf
,
2471 stripe
* (disks
- conf
->max_degraded
)
2472 *sectors_per_chunk
+ chunk_offset
,
2478 handle_failed_stripe(struct r5conf
*conf
, struct stripe_head
*sh
,
2479 struct stripe_head_state
*s
, int disks
,
2480 struct bio
**return_bi
)
2483 for (i
= disks
; i
--; ) {
2487 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2488 struct md_rdev
*rdev
;
2490 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2491 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2492 atomic_inc(&rdev
->nr_pending
);
2497 if (!rdev_set_badblocks(
2501 md_error(conf
->mddev
, rdev
);
2502 rdev_dec_pending(rdev
, conf
->mddev
);
2505 spin_lock_irq(&sh
->stripe_lock
);
2506 /* fail all writes first */
2507 bi
= sh
->dev
[i
].towrite
;
2508 sh
->dev
[i
].towrite
= NULL
;
2509 spin_unlock_irq(&sh
->stripe_lock
);
2515 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2516 wake_up(&conf
->wait_for_overlap
);
2518 while (bi
&& bi
->bi_sector
<
2519 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2520 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2521 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2522 if (!raid5_dec_bi_active_stripes(bi
)) {
2523 md_write_end(conf
->mddev
);
2524 bi
->bi_next
= *return_bi
;
2530 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2531 STRIPE_SECTORS
, 0, 0);
2533 /* and fail all 'written' */
2534 bi
= sh
->dev
[i
].written
;
2535 sh
->dev
[i
].written
= NULL
;
2536 if (bi
) bitmap_end
= 1;
2537 while (bi
&& bi
->bi_sector
<
2538 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2539 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2540 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2541 if (!raid5_dec_bi_active_stripes(bi
)) {
2542 md_write_end(conf
->mddev
);
2543 bi
->bi_next
= *return_bi
;
2549 /* fail any reads if this device is non-operational and
2550 * the data has not reached the cache yet.
2552 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2553 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2554 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2555 bi
= sh
->dev
[i
].toread
;
2556 sh
->dev
[i
].toread
= NULL
;
2557 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2558 wake_up(&conf
->wait_for_overlap
);
2559 if (bi
) s
->to_read
--;
2560 while (bi
&& bi
->bi_sector
<
2561 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2562 struct bio
*nextbi
=
2563 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2564 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2565 if (!raid5_dec_bi_active_stripes(bi
)) {
2566 bi
->bi_next
= *return_bi
;
2573 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2574 STRIPE_SECTORS
, 0, 0);
2575 /* If we were in the middle of a write the parity block might
2576 * still be locked - so just clear all R5_LOCKED flags
2578 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2581 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2582 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2583 md_wakeup_thread(conf
->mddev
->thread
);
2587 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
2588 struct stripe_head_state
*s
)
2593 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2596 /* There is nothing more to do for sync/check/repair.
2597 * Don't even need to abort as that is handled elsewhere
2598 * if needed, and not always wanted e.g. if there is a known
2600 * For recover/replace we need to record a bad block on all
2601 * non-sync devices, or abort the recovery
2603 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
2604 /* During recovery devices cannot be removed, so
2605 * locking and refcounting of rdevs is not needed
2607 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2608 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
2610 && !test_bit(Faulty
, &rdev
->flags
)
2611 && !test_bit(In_sync
, &rdev
->flags
)
2612 && !rdev_set_badblocks(rdev
, sh
->sector
,
2615 rdev
= conf
->disks
[i
].replacement
;
2617 && !test_bit(Faulty
, &rdev
->flags
)
2618 && !test_bit(In_sync
, &rdev
->flags
)
2619 && !rdev_set_badblocks(rdev
, sh
->sector
,
2624 conf
->recovery_disabled
=
2625 conf
->mddev
->recovery_disabled
;
2627 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
2630 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
2632 struct md_rdev
*rdev
;
2634 /* Doing recovery so rcu locking not required */
2635 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
2637 && !test_bit(Faulty
, &rdev
->flags
)
2638 && !test_bit(In_sync
, &rdev
->flags
)
2639 && (rdev
->recovery_offset
<= sh
->sector
2640 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
2646 /* fetch_block - checks the given member device to see if its data needs
2647 * to be read or computed to satisfy a request.
2649 * Returns 1 when no more member devices need to be checked, otherwise returns
2650 * 0 to tell the loop in handle_stripe_fill to continue
2652 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2653 int disk_idx
, int disks
)
2655 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2656 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2657 &sh
->dev
[s
->failed_num
[1]] };
2659 /* is the data in this block needed, and can we get it? */
2660 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2661 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2663 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2664 s
->syncing
|| s
->expanding
||
2665 (s
->replacing
&& want_replace(sh
, disk_idx
)) ||
2666 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2667 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2668 (sh
->raid_conf
->level
<= 5 && s
->failed
&& fdev
[0]->towrite
&&
2669 !test_bit(R5_OVERWRITE
, &fdev
[0]->flags
)) ||
2670 (sh
->raid_conf
->level
== 6 && s
->failed
&& s
->to_write
))) {
2671 /* we would like to get this block, possibly by computing it,
2672 * otherwise read it if the backing disk is insync
2674 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2675 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2676 if ((s
->uptodate
== disks
- 1) &&
2677 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2678 disk_idx
== s
->failed_num
[1]))) {
2679 /* have disk failed, and we're requested to fetch it;
2682 pr_debug("Computing stripe %llu block %d\n",
2683 (unsigned long long)sh
->sector
, disk_idx
);
2684 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2685 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2686 set_bit(R5_Wantcompute
, &dev
->flags
);
2687 sh
->ops
.target
= disk_idx
;
2688 sh
->ops
.target2
= -1; /* no 2nd target */
2690 /* Careful: from this point on 'uptodate' is in the eye
2691 * of raid_run_ops which services 'compute' operations
2692 * before writes. R5_Wantcompute flags a block that will
2693 * be R5_UPTODATE by the time it is needed for a
2694 * subsequent operation.
2698 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2699 /* Computing 2-failure is *very* expensive; only
2700 * do it if failed >= 2
2703 for (other
= disks
; other
--; ) {
2704 if (other
== disk_idx
)
2706 if (!test_bit(R5_UPTODATE
,
2707 &sh
->dev
[other
].flags
))
2711 pr_debug("Computing stripe %llu blocks %d,%d\n",
2712 (unsigned long long)sh
->sector
,
2714 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2715 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2716 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2717 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2718 sh
->ops
.target
= disk_idx
;
2719 sh
->ops
.target2
= other
;
2723 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2724 set_bit(R5_LOCKED
, &dev
->flags
);
2725 set_bit(R5_Wantread
, &dev
->flags
);
2727 pr_debug("Reading block %d (sync=%d)\n",
2728 disk_idx
, s
->syncing
);
2736 * handle_stripe_fill - read or compute data to satisfy pending requests.
2738 static void handle_stripe_fill(struct stripe_head
*sh
,
2739 struct stripe_head_state
*s
,
2744 /* look for blocks to read/compute, skip this if a compute
2745 * is already in flight, or if the stripe contents are in the
2746 * midst of changing due to a write
2748 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2749 !sh
->reconstruct_state
)
2750 for (i
= disks
; i
--; )
2751 if (fetch_block(sh
, s
, i
, disks
))
2753 set_bit(STRIPE_HANDLE
, &sh
->state
);
2757 /* handle_stripe_clean_event
2758 * any written block on an uptodate or failed drive can be returned.
2759 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2760 * never LOCKED, so we don't need to test 'failed' directly.
2762 static void handle_stripe_clean_event(struct r5conf
*conf
,
2763 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2768 for (i
= disks
; i
--; )
2769 if (sh
->dev
[i
].written
) {
2771 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2772 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2773 test_and_clear_bit(R5_Discard
, &dev
->flags
))) {
2774 /* We can return any write requests */
2775 struct bio
*wbi
, *wbi2
;
2776 pr_debug("Return write for disc %d\n", i
);
2778 dev
->written
= NULL
;
2779 while (wbi
&& wbi
->bi_sector
<
2780 dev
->sector
+ STRIPE_SECTORS
) {
2781 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2782 if (!raid5_dec_bi_active_stripes(wbi
)) {
2783 md_write_end(conf
->mddev
);
2784 wbi
->bi_next
= *return_bi
;
2789 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2791 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2796 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2797 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2798 md_wakeup_thread(conf
->mddev
->thread
);
2801 static void handle_stripe_dirtying(struct r5conf
*conf
,
2802 struct stripe_head
*sh
,
2803 struct stripe_head_state
*s
,
2806 int rmw
= 0, rcw
= 0, i
;
2807 if (conf
->max_degraded
== 2) {
2808 /* RAID6 requires 'rcw' in current implementation
2809 * Calculate the real rcw later - for now fake it
2810 * look like rcw is cheaper
2813 } else for (i
= disks
; i
--; ) {
2814 /* would I have to read this buffer for read_modify_write */
2815 struct r5dev
*dev
= &sh
->dev
[i
];
2816 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2817 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2818 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2819 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2820 if (test_bit(R5_Insync
, &dev
->flags
))
2823 rmw
+= 2*disks
; /* cannot read it */
2825 /* Would I have to read this buffer for reconstruct_write */
2826 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2827 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2828 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2829 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2830 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2835 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2836 (unsigned long long)sh
->sector
, rmw
, rcw
);
2837 set_bit(STRIPE_HANDLE
, &sh
->state
);
2838 if (rmw
< rcw
&& rmw
> 0)
2839 /* prefer read-modify-write, but need to get some data */
2840 for (i
= disks
; i
--; ) {
2841 struct r5dev
*dev
= &sh
->dev
[i
];
2842 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2843 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2844 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2845 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2846 test_bit(R5_Insync
, &dev
->flags
)) {
2848 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2849 pr_debug("Read_old block "
2850 "%d for r-m-w\n", i
);
2851 set_bit(R5_LOCKED
, &dev
->flags
);
2852 set_bit(R5_Wantread
, &dev
->flags
);
2855 set_bit(STRIPE_DELAYED
, &sh
->state
);
2856 set_bit(STRIPE_HANDLE
, &sh
->state
);
2860 if (rcw
<= rmw
&& rcw
> 0) {
2861 /* want reconstruct write, but need to get some data */
2863 for (i
= disks
; i
--; ) {
2864 struct r5dev
*dev
= &sh
->dev
[i
];
2865 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2866 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
2867 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2868 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2869 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2871 if (!test_bit(R5_Insync
, &dev
->flags
))
2872 continue; /* it's a failed drive */
2874 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2875 pr_debug("Read_old block "
2876 "%d for Reconstruct\n", i
);
2877 set_bit(R5_LOCKED
, &dev
->flags
);
2878 set_bit(R5_Wantread
, &dev
->flags
);
2881 set_bit(STRIPE_DELAYED
, &sh
->state
);
2882 set_bit(STRIPE_HANDLE
, &sh
->state
);
2887 /* now if nothing is locked, and if we have enough data,
2888 * we can start a write request
2890 /* since handle_stripe can be called at any time we need to handle the
2891 * case where a compute block operation has been submitted and then a
2892 * subsequent call wants to start a write request. raid_run_ops only
2893 * handles the case where compute block and reconstruct are requested
2894 * simultaneously. If this is not the case then new writes need to be
2895 * held off until the compute completes.
2897 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2898 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2899 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2900 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2903 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
2904 struct stripe_head_state
*s
, int disks
)
2906 struct r5dev
*dev
= NULL
;
2908 set_bit(STRIPE_HANDLE
, &sh
->state
);
2910 switch (sh
->check_state
) {
2911 case check_state_idle
:
2912 /* start a new check operation if there are no failures */
2913 if (s
->failed
== 0) {
2914 BUG_ON(s
->uptodate
!= disks
);
2915 sh
->check_state
= check_state_run
;
2916 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2917 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2921 dev
= &sh
->dev
[s
->failed_num
[0]];
2923 case check_state_compute_result
:
2924 sh
->check_state
= check_state_idle
;
2926 dev
= &sh
->dev
[sh
->pd_idx
];
2928 /* check that a write has not made the stripe insync */
2929 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2932 /* either failed parity check, or recovery is happening */
2933 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2934 BUG_ON(s
->uptodate
!= disks
);
2936 set_bit(R5_LOCKED
, &dev
->flags
);
2938 set_bit(R5_Wantwrite
, &dev
->flags
);
2940 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2941 set_bit(STRIPE_INSYNC
, &sh
->state
);
2943 case check_state_run
:
2944 break; /* we will be called again upon completion */
2945 case check_state_check_result
:
2946 sh
->check_state
= check_state_idle
;
2948 /* if a failure occurred during the check operation, leave
2949 * STRIPE_INSYNC not set and let the stripe be handled again
2954 /* handle a successful check operation, if parity is correct
2955 * we are done. Otherwise update the mismatch count and repair
2956 * parity if !MD_RECOVERY_CHECK
2958 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2959 /* parity is correct (on disc,
2960 * not in buffer any more)
2962 set_bit(STRIPE_INSYNC
, &sh
->state
);
2964 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2965 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2966 /* don't try to repair!! */
2967 set_bit(STRIPE_INSYNC
, &sh
->state
);
2969 sh
->check_state
= check_state_compute_run
;
2970 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2971 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2972 set_bit(R5_Wantcompute
,
2973 &sh
->dev
[sh
->pd_idx
].flags
);
2974 sh
->ops
.target
= sh
->pd_idx
;
2975 sh
->ops
.target2
= -1;
2980 case check_state_compute_run
:
2983 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2984 __func__
, sh
->check_state
,
2985 (unsigned long long) sh
->sector
);
2991 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
2992 struct stripe_head_state
*s
,
2995 int pd_idx
= sh
->pd_idx
;
2996 int qd_idx
= sh
->qd_idx
;
2999 set_bit(STRIPE_HANDLE
, &sh
->state
);
3001 BUG_ON(s
->failed
> 2);
3003 /* Want to check and possibly repair P and Q.
3004 * However there could be one 'failed' device, in which
3005 * case we can only check one of them, possibly using the
3006 * other to generate missing data
3009 switch (sh
->check_state
) {
3010 case check_state_idle
:
3011 /* start a new check operation if there are < 2 failures */
3012 if (s
->failed
== s
->q_failed
) {
3013 /* The only possible failed device holds Q, so it
3014 * makes sense to check P (If anything else were failed,
3015 * we would have used P to recreate it).
3017 sh
->check_state
= check_state_run
;
3019 if (!s
->q_failed
&& s
->failed
< 2) {
3020 /* Q is not failed, and we didn't use it to generate
3021 * anything, so it makes sense to check it
3023 if (sh
->check_state
== check_state_run
)
3024 sh
->check_state
= check_state_run_pq
;
3026 sh
->check_state
= check_state_run_q
;
3029 /* discard potentially stale zero_sum_result */
3030 sh
->ops
.zero_sum_result
= 0;
3032 if (sh
->check_state
== check_state_run
) {
3033 /* async_xor_zero_sum destroys the contents of P */
3034 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3037 if (sh
->check_state
>= check_state_run
&&
3038 sh
->check_state
<= check_state_run_pq
) {
3039 /* async_syndrome_zero_sum preserves P and Q, so
3040 * no need to mark them !uptodate here
3042 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3046 /* we have 2-disk failure */
3047 BUG_ON(s
->failed
!= 2);
3049 case check_state_compute_result
:
3050 sh
->check_state
= check_state_idle
;
3052 /* check that a write has not made the stripe insync */
3053 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3056 /* now write out any block on a failed drive,
3057 * or P or Q if they were recomputed
3059 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3060 if (s
->failed
== 2) {
3061 dev
= &sh
->dev
[s
->failed_num
[1]];
3063 set_bit(R5_LOCKED
, &dev
->flags
);
3064 set_bit(R5_Wantwrite
, &dev
->flags
);
3066 if (s
->failed
>= 1) {
3067 dev
= &sh
->dev
[s
->failed_num
[0]];
3069 set_bit(R5_LOCKED
, &dev
->flags
);
3070 set_bit(R5_Wantwrite
, &dev
->flags
);
3072 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3073 dev
= &sh
->dev
[pd_idx
];
3075 set_bit(R5_LOCKED
, &dev
->flags
);
3076 set_bit(R5_Wantwrite
, &dev
->flags
);
3078 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3079 dev
= &sh
->dev
[qd_idx
];
3081 set_bit(R5_LOCKED
, &dev
->flags
);
3082 set_bit(R5_Wantwrite
, &dev
->flags
);
3084 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3086 set_bit(STRIPE_INSYNC
, &sh
->state
);
3088 case check_state_run
:
3089 case check_state_run_q
:
3090 case check_state_run_pq
:
3091 break; /* we will be called again upon completion */
3092 case check_state_check_result
:
3093 sh
->check_state
= check_state_idle
;
3095 /* handle a successful check operation, if parity is correct
3096 * we are done. Otherwise update the mismatch count and repair
3097 * parity if !MD_RECOVERY_CHECK
3099 if (sh
->ops
.zero_sum_result
== 0) {
3100 /* both parities are correct */
3102 set_bit(STRIPE_INSYNC
, &sh
->state
);
3104 /* in contrast to the raid5 case we can validate
3105 * parity, but still have a failure to write
3108 sh
->check_state
= check_state_compute_result
;
3109 /* Returning at this point means that we may go
3110 * off and bring p and/or q uptodate again so
3111 * we make sure to check zero_sum_result again
3112 * to verify if p or q need writeback
3116 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
3117 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3118 /* don't try to repair!! */
3119 set_bit(STRIPE_INSYNC
, &sh
->state
);
3121 int *target
= &sh
->ops
.target
;
3123 sh
->ops
.target
= -1;
3124 sh
->ops
.target2
= -1;
3125 sh
->check_state
= check_state_compute_run
;
3126 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3127 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3128 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3129 set_bit(R5_Wantcompute
,
3130 &sh
->dev
[pd_idx
].flags
);
3132 target
= &sh
->ops
.target2
;
3135 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3136 set_bit(R5_Wantcompute
,
3137 &sh
->dev
[qd_idx
].flags
);
3144 case check_state_compute_run
:
3147 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3148 __func__
, sh
->check_state
,
3149 (unsigned long long) sh
->sector
);
3154 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3158 /* We have read all the blocks in this stripe and now we need to
3159 * copy some of them into a target stripe for expand.
3161 struct dma_async_tx_descriptor
*tx
= NULL
;
3162 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3163 for (i
= 0; i
< sh
->disks
; i
++)
3164 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3166 struct stripe_head
*sh2
;
3167 struct async_submit_ctl submit
;
3169 sector_t bn
= compute_blocknr(sh
, i
, 1);
3170 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3172 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3174 /* so far only the early blocks of this stripe
3175 * have been requested. When later blocks
3176 * get requested, we will try again
3179 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3180 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3181 /* must have already done this block */
3182 release_stripe(sh2
);
3186 /* place all the copies on one channel */
3187 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3188 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3189 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3192 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3193 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3194 for (j
= 0; j
< conf
->raid_disks
; j
++)
3195 if (j
!= sh2
->pd_idx
&&
3197 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3199 if (j
== conf
->raid_disks
) {
3200 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3201 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3203 release_stripe(sh2
);
3206 /* done submitting copies, wait for them to complete */
3209 dma_wait_for_async_tx(tx
);
3214 * handle_stripe - do things to a stripe.
3216 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3217 * state of various bits to see what needs to be done.
3219 * return some read requests which now have data
3220 * return some write requests which are safely on storage
3221 * schedule a read on some buffers
3222 * schedule a write of some buffers
3223 * return confirmation of parity correctness
3227 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3229 struct r5conf
*conf
= sh
->raid_conf
;
3230 int disks
= sh
->disks
;
3233 int do_recovery
= 0;
3235 memset(s
, 0, sizeof(*s
));
3237 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3238 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3239 s
->failed_num
[0] = -1;
3240 s
->failed_num
[1] = -1;
3242 /* Now to look around and see what can be done */
3244 for (i
=disks
; i
--; ) {
3245 struct md_rdev
*rdev
;
3252 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3254 dev
->toread
, dev
->towrite
, dev
->written
);
3255 /* maybe we can reply to a read
3257 * new wantfill requests are only permitted while
3258 * ops_complete_biofill is guaranteed to be inactive
3260 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3261 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3262 set_bit(R5_Wantfill
, &dev
->flags
);
3264 /* now count some things */
3265 if (test_bit(R5_LOCKED
, &dev
->flags
))
3267 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3269 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3271 BUG_ON(s
->compute
> 2);
3274 if (test_bit(R5_Wantfill
, &dev
->flags
))
3276 else if (dev
->toread
)
3280 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3285 /* Prefer to use the replacement for reads, but only
3286 * if it is recovered enough and has no bad blocks.
3288 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3289 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
3290 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
3291 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3292 &first_bad
, &bad_sectors
))
3293 set_bit(R5_ReadRepl
, &dev
->flags
);
3296 set_bit(R5_NeedReplace
, &dev
->flags
);
3297 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3298 clear_bit(R5_ReadRepl
, &dev
->flags
);
3300 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
3303 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3304 &first_bad
, &bad_sectors
);
3305 if (s
->blocked_rdev
== NULL
3306 && (test_bit(Blocked
, &rdev
->flags
)
3309 set_bit(BlockedBadBlocks
,
3311 s
->blocked_rdev
= rdev
;
3312 atomic_inc(&rdev
->nr_pending
);
3315 clear_bit(R5_Insync
, &dev
->flags
);
3319 /* also not in-sync */
3320 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
3321 test_bit(R5_UPTODATE
, &dev
->flags
)) {
3322 /* treat as in-sync, but with a read error
3323 * which we can now try to correct
3325 set_bit(R5_Insync
, &dev
->flags
);
3326 set_bit(R5_ReadError
, &dev
->flags
);
3328 } else if (test_bit(In_sync
, &rdev
->flags
))
3329 set_bit(R5_Insync
, &dev
->flags
);
3330 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3331 /* in sync if before recovery_offset */
3332 set_bit(R5_Insync
, &dev
->flags
);
3333 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
3334 test_bit(R5_Expanded
, &dev
->flags
))
3335 /* If we've reshaped into here, we assume it is Insync.
3336 * We will shortly update recovery_offset to make
3339 set_bit(R5_Insync
, &dev
->flags
);
3341 if (rdev
&& test_bit(R5_WriteError
, &dev
->flags
)) {
3342 /* This flag does not apply to '.replacement'
3343 * only to .rdev, so make sure to check that*/
3344 struct md_rdev
*rdev2
= rcu_dereference(
3345 conf
->disks
[i
].rdev
);
3347 clear_bit(R5_Insync
, &dev
->flags
);
3348 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3349 s
->handle_bad_blocks
= 1;
3350 atomic_inc(&rdev2
->nr_pending
);
3352 clear_bit(R5_WriteError
, &dev
->flags
);
3354 if (rdev
&& test_bit(R5_MadeGood
, &dev
->flags
)) {
3355 /* This flag does not apply to '.replacement'
3356 * only to .rdev, so make sure to check that*/
3357 struct md_rdev
*rdev2
= rcu_dereference(
3358 conf
->disks
[i
].rdev
);
3359 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3360 s
->handle_bad_blocks
= 1;
3361 atomic_inc(&rdev2
->nr_pending
);
3363 clear_bit(R5_MadeGood
, &dev
->flags
);
3365 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3366 struct md_rdev
*rdev2
= rcu_dereference(
3367 conf
->disks
[i
].replacement
);
3368 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3369 s
->handle_bad_blocks
= 1;
3370 atomic_inc(&rdev2
->nr_pending
);
3372 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
3374 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3375 /* The ReadError flag will just be confusing now */
3376 clear_bit(R5_ReadError
, &dev
->flags
);
3377 clear_bit(R5_ReWrite
, &dev
->flags
);
3379 if (test_bit(R5_ReadError
, &dev
->flags
))
3380 clear_bit(R5_Insync
, &dev
->flags
);
3381 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3383 s
->failed_num
[s
->failed
] = i
;
3385 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
3389 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
3390 /* If there is a failed device being replaced,
3391 * we must be recovering.
3392 * else if we are after recovery_cp, we must be syncing
3393 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3394 * else we can only be replacing
3395 * sync and recovery both need to read all devices, and so
3396 * use the same flag.
3399 sh
->sector
>= conf
->mddev
->recovery_cp
||
3400 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
3408 static void handle_stripe(struct stripe_head
*sh
)
3410 struct stripe_head_state s
;
3411 struct r5conf
*conf
= sh
->raid_conf
;
3414 int disks
= sh
->disks
;
3415 struct r5dev
*pdev
, *qdev
;
3417 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3418 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
3419 /* already being handled, ensure it gets handled
3420 * again when current action finishes */
3421 set_bit(STRIPE_HANDLE
, &sh
->state
);
3425 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3426 set_bit(STRIPE_SYNCING
, &sh
->state
);
3427 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3429 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3431 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3432 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3433 (unsigned long long)sh
->sector
, sh
->state
,
3434 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3435 sh
->check_state
, sh
->reconstruct_state
);
3437 analyse_stripe(sh
, &s
);
3439 if (s
.handle_bad_blocks
) {
3440 set_bit(STRIPE_HANDLE
, &sh
->state
);
3444 if (unlikely(s
.blocked_rdev
)) {
3445 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3446 s
.replacing
|| s
.to_write
|| s
.written
) {
3447 set_bit(STRIPE_HANDLE
, &sh
->state
);
3450 /* There is nothing for the blocked_rdev to block */
3451 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
3452 s
.blocked_rdev
= NULL
;
3455 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3456 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3457 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3460 pr_debug("locked=%d uptodate=%d to_read=%d"
3461 " to_write=%d failed=%d failed_num=%d,%d\n",
3462 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3463 s
.failed_num
[0], s
.failed_num
[1]);
3464 /* check if the array has lost more than max_degraded devices and,
3465 * if so, some requests might need to be failed.
3467 if (s
.failed
> conf
->max_degraded
) {
3468 sh
->check_state
= 0;
3469 sh
->reconstruct_state
= 0;
3470 if (s
.to_read
+s
.to_write
+s
.written
)
3471 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
3472 if (s
.syncing
+ s
.replacing
)
3473 handle_failed_sync(conf
, sh
, &s
);
3477 * might be able to return some write requests if the parity blocks
3478 * are safe, or on a failed drive
3480 pdev
= &sh
->dev
[sh
->pd_idx
];
3481 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
3482 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
3483 qdev
= &sh
->dev
[sh
->qd_idx
];
3484 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
3485 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
3489 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3490 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3491 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
3492 test_bit(R5_Discard
, &pdev
->flags
))))) &&
3493 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3494 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3495 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
3496 test_bit(R5_Discard
, &qdev
->flags
))))))
3497 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
3499 /* Now we might consider reading some blocks, either to check/generate
3500 * parity, or to satisfy requests
3501 * or to load a block that is being partially written.
3503 if (s
.to_read
|| s
.non_overwrite
3504 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
3505 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
3508 handle_stripe_fill(sh
, &s
, disks
);
3510 /* Now we check to see if any write operations have recently
3514 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3516 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3517 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3518 sh
->reconstruct_state
= reconstruct_state_idle
;
3520 /* All the 'written' buffers and the parity block are ready to
3521 * be written back to disk
3523 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
3524 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
3525 BUG_ON(sh
->qd_idx
>= 0 &&
3526 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
3527 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
3528 for (i
= disks
; i
--; ) {
3529 struct r5dev
*dev
= &sh
->dev
[i
];
3530 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3531 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3533 pr_debug("Writing block %d\n", i
);
3534 set_bit(R5_Wantwrite
, &dev
->flags
);
3537 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3538 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3540 set_bit(STRIPE_INSYNC
, &sh
->state
);
3543 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3544 s
.dec_preread_active
= 1;
3547 /* Now to consider new write requests and what else, if anything
3548 * should be read. We do not handle new writes when:
3549 * 1/ A 'write' operation (copy+xor) is already in flight.
3550 * 2/ A 'check' operation is in flight, as it may clobber the parity
3553 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3554 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
3556 /* maybe we need to check and possibly fix the parity for this stripe
3557 * Any reads will already have been scheduled, so we just see if enough
3558 * data is available. The parity check is held off while parity
3559 * dependent operations are in flight.
3561 if (sh
->check_state
||
3562 (s
.syncing
&& s
.locked
== 0 &&
3563 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3564 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
3565 if (conf
->level
== 6)
3566 handle_parity_checks6(conf
, sh
, &s
, disks
);
3568 handle_parity_checks5(conf
, sh
, &s
, disks
);
3571 if (s
.replacing
&& s
.locked
== 0
3572 && !test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3573 /* Write out to replacement devices where possible */
3574 for (i
= 0; i
< conf
->raid_disks
; i
++)
3575 if (test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
) &&
3576 test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
3577 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
3578 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3581 set_bit(STRIPE_INSYNC
, &sh
->state
);
3583 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
3584 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3585 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3586 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3589 /* If the failed drives are just a ReadError, then we might need
3590 * to progress the repair/check process
3592 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3593 for (i
= 0; i
< s
.failed
; i
++) {
3594 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3595 if (test_bit(R5_ReadError
, &dev
->flags
)
3596 && !test_bit(R5_LOCKED
, &dev
->flags
)
3597 && test_bit(R5_UPTODATE
, &dev
->flags
)
3599 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3600 set_bit(R5_Wantwrite
, &dev
->flags
);
3601 set_bit(R5_ReWrite
, &dev
->flags
);
3602 set_bit(R5_LOCKED
, &dev
->flags
);
3605 /* let's read it back */
3606 set_bit(R5_Wantread
, &dev
->flags
);
3607 set_bit(R5_LOCKED
, &dev
->flags
);
3614 /* Finish reconstruct operations initiated by the expansion process */
3615 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3616 struct stripe_head
*sh_src
3617 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3618 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3619 /* sh cannot be written until sh_src has been read.
3620 * so arrange for sh to be delayed a little
3622 set_bit(STRIPE_DELAYED
, &sh
->state
);
3623 set_bit(STRIPE_HANDLE
, &sh
->state
);
3624 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3626 atomic_inc(&conf
->preread_active_stripes
);
3627 release_stripe(sh_src
);
3631 release_stripe(sh_src
);
3633 sh
->reconstruct_state
= reconstruct_state_idle
;
3634 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3635 for (i
= conf
->raid_disks
; i
--; ) {
3636 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3637 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3642 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3643 !sh
->reconstruct_state
) {
3644 /* Need to write out all blocks after computing parity */
3645 sh
->disks
= conf
->raid_disks
;
3646 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3647 schedule_reconstruction(sh
, &s
, 1, 1);
3648 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3649 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3650 atomic_dec(&conf
->reshape_stripes
);
3651 wake_up(&conf
->wait_for_overlap
);
3652 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3655 if (s
.expanding
&& s
.locked
== 0 &&
3656 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3657 handle_stripe_expansion(conf
, sh
);
3660 /* wait for this device to become unblocked */
3661 if (unlikely(s
.blocked_rdev
)) {
3662 if (conf
->mddev
->external
)
3663 md_wait_for_blocked_rdev(s
.blocked_rdev
,
3666 /* Internal metadata will immediately
3667 * be written by raid5d, so we don't
3668 * need to wait here.
3670 rdev_dec_pending(s
.blocked_rdev
,
3674 if (s
.handle_bad_blocks
)
3675 for (i
= disks
; i
--; ) {
3676 struct md_rdev
*rdev
;
3677 struct r5dev
*dev
= &sh
->dev
[i
];
3678 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
3679 /* We own a safe reference to the rdev */
3680 rdev
= conf
->disks
[i
].rdev
;
3681 if (!rdev_set_badblocks(rdev
, sh
->sector
,
3683 md_error(conf
->mddev
, rdev
);
3684 rdev_dec_pending(rdev
, conf
->mddev
);
3686 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
3687 rdev
= conf
->disks
[i
].rdev
;
3688 rdev_clear_badblocks(rdev
, sh
->sector
,
3690 rdev_dec_pending(rdev
, conf
->mddev
);
3692 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3693 rdev
= conf
->disks
[i
].replacement
;
3695 /* rdev have been moved down */
3696 rdev
= conf
->disks
[i
].rdev
;
3697 rdev_clear_badblocks(rdev
, sh
->sector
,
3699 rdev_dec_pending(rdev
, conf
->mddev
);
3704 raid_run_ops(sh
, s
.ops_request
);
3708 if (s
.dec_preread_active
) {
3709 /* We delay this until after ops_run_io so that if make_request
3710 * is waiting on a flush, it won't continue until the writes
3711 * have actually been submitted.
3713 atomic_dec(&conf
->preread_active_stripes
);
3714 if (atomic_read(&conf
->preread_active_stripes
) <
3716 md_wakeup_thread(conf
->mddev
->thread
);
3719 return_io(s
.return_bi
);
3721 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
3724 static void raid5_activate_delayed(struct r5conf
*conf
)
3726 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3727 while (!list_empty(&conf
->delayed_list
)) {
3728 struct list_head
*l
= conf
->delayed_list
.next
;
3729 struct stripe_head
*sh
;
3730 sh
= list_entry(l
, struct stripe_head
, lru
);
3732 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3733 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3734 atomic_inc(&conf
->preread_active_stripes
);
3735 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3740 static void activate_bit_delay(struct r5conf
*conf
)
3742 /* device_lock is held */
3743 struct list_head head
;
3744 list_add(&head
, &conf
->bitmap_list
);
3745 list_del_init(&conf
->bitmap_list
);
3746 while (!list_empty(&head
)) {
3747 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3748 list_del_init(&sh
->lru
);
3749 atomic_inc(&sh
->count
);
3750 __release_stripe(conf
, sh
);
3754 int md_raid5_congested(struct mddev
*mddev
, int bits
)
3756 struct r5conf
*conf
= mddev
->private;
3758 /* No difference between reads and writes. Just check
3759 * how busy the stripe_cache is
3762 if (conf
->inactive_blocked
)
3766 if (list_empty_careful(&conf
->inactive_list
))
3771 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3773 static int raid5_congested(void *data
, int bits
)
3775 struct mddev
*mddev
= data
;
3777 return mddev_congested(mddev
, bits
) ||
3778 md_raid5_congested(mddev
, bits
);
3781 /* We want read requests to align with chunks where possible,
3782 * but write requests don't need to.
3784 static int raid5_mergeable_bvec(struct request_queue
*q
,
3785 struct bvec_merge_data
*bvm
,
3786 struct bio_vec
*biovec
)
3788 struct mddev
*mddev
= q
->queuedata
;
3789 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3791 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3792 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3794 if ((bvm
->bi_rw
& 1) == WRITE
)
3795 return biovec
->bv_len
; /* always allow writes to be mergeable */
3797 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3798 chunk_sectors
= mddev
->new_chunk_sectors
;
3799 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3800 if (max
< 0) max
= 0;
3801 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3802 return biovec
->bv_len
;
3808 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
3810 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3811 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3812 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3814 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3815 chunk_sectors
= mddev
->new_chunk_sectors
;
3816 return chunk_sectors
>=
3817 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3821 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3822 * later sampled by raid5d.
3824 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
3826 unsigned long flags
;
3828 spin_lock_irqsave(&conf
->device_lock
, flags
);
3830 bi
->bi_next
= conf
->retry_read_aligned_list
;
3831 conf
->retry_read_aligned_list
= bi
;
3833 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3834 md_wakeup_thread(conf
->mddev
->thread
);
3838 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
3842 bi
= conf
->retry_read_aligned
;
3844 conf
->retry_read_aligned
= NULL
;
3847 bi
= conf
->retry_read_aligned_list
;
3849 conf
->retry_read_aligned_list
= bi
->bi_next
;
3852 * this sets the active strip count to 1 and the processed
3853 * strip count to zero (upper 8 bits)
3855 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
3863 * The "raid5_align_endio" should check if the read succeeded and if it
3864 * did, call bio_endio on the original bio (having bio_put the new bio
3866 * If the read failed..
3868 static void raid5_align_endio(struct bio
*bi
, int error
)
3870 struct bio
* raid_bi
= bi
->bi_private
;
3871 struct mddev
*mddev
;
3872 struct r5conf
*conf
;
3873 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3874 struct md_rdev
*rdev
;
3878 rdev
= (void*)raid_bi
->bi_next
;
3879 raid_bi
->bi_next
= NULL
;
3880 mddev
= rdev
->mddev
;
3881 conf
= mddev
->private;
3883 rdev_dec_pending(rdev
, conf
->mddev
);
3885 if (!error
&& uptodate
) {
3886 bio_endio(raid_bi
, 0);
3887 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3888 wake_up(&conf
->wait_for_stripe
);
3893 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3895 add_bio_to_retry(raid_bi
, conf
);
3898 static int bio_fits_rdev(struct bio
*bi
)
3900 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3902 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3904 blk_recount_segments(q
, bi
);
3905 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3908 if (q
->merge_bvec_fn
)
3909 /* it's too hard to apply the merge_bvec_fn at this stage,
3918 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
3920 struct r5conf
*conf
= mddev
->private;
3922 struct bio
* align_bi
;
3923 struct md_rdev
*rdev
;
3924 sector_t end_sector
;
3926 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3927 pr_debug("chunk_aligned_read : non aligned\n");
3931 * use bio_clone_mddev to make a copy of the bio
3933 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
3937 * set bi_end_io to a new function, and set bi_private to the
3940 align_bi
->bi_end_io
= raid5_align_endio
;
3941 align_bi
->bi_private
= raid_bio
;
3945 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3949 end_sector
= align_bi
->bi_sector
+ (align_bi
->bi_size
>> 9);
3951 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
3952 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
3953 rdev
->recovery_offset
< end_sector
) {
3954 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3956 (test_bit(Faulty
, &rdev
->flags
) ||
3957 !(test_bit(In_sync
, &rdev
->flags
) ||
3958 rdev
->recovery_offset
>= end_sector
)))
3965 atomic_inc(&rdev
->nr_pending
);
3967 raid_bio
->bi_next
= (void*)rdev
;
3968 align_bi
->bi_bdev
= rdev
->bdev
;
3969 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3971 if (!bio_fits_rdev(align_bi
) ||
3972 is_badblock(rdev
, align_bi
->bi_sector
, align_bi
->bi_size
>>9,
3973 &first_bad
, &bad_sectors
)) {
3974 /* too big in some way, or has a known bad block */
3976 rdev_dec_pending(rdev
, mddev
);
3980 /* No reshape active, so we can trust rdev->data_offset */
3981 align_bi
->bi_sector
+= rdev
->data_offset
;
3983 spin_lock_irq(&conf
->device_lock
);
3984 wait_event_lock_irq(conf
->wait_for_stripe
,
3986 conf
->device_lock
, /* nothing */);
3987 atomic_inc(&conf
->active_aligned_reads
);
3988 spin_unlock_irq(&conf
->device_lock
);
3990 generic_make_request(align_bi
);
3999 /* __get_priority_stripe - get the next stripe to process
4001 * Full stripe writes are allowed to pass preread active stripes up until
4002 * the bypass_threshold is exceeded. In general the bypass_count
4003 * increments when the handle_list is handled before the hold_list; however, it
4004 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4005 * stripe with in flight i/o. The bypass_count will be reset when the
4006 * head of the hold_list has changed, i.e. the head was promoted to the
4009 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
)
4011 struct stripe_head
*sh
;
4013 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4015 list_empty(&conf
->handle_list
) ? "empty" : "busy",
4016 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4017 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4019 if (!list_empty(&conf
->handle_list
)) {
4020 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
4022 if (list_empty(&conf
->hold_list
))
4023 conf
->bypass_count
= 0;
4024 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4025 if (conf
->hold_list
.next
== conf
->last_hold
)
4026 conf
->bypass_count
++;
4028 conf
->last_hold
= conf
->hold_list
.next
;
4029 conf
->bypass_count
-= conf
->bypass_threshold
;
4030 if (conf
->bypass_count
< 0)
4031 conf
->bypass_count
= 0;
4034 } else if (!list_empty(&conf
->hold_list
) &&
4035 ((conf
->bypass_threshold
&&
4036 conf
->bypass_count
> conf
->bypass_threshold
) ||
4037 atomic_read(&conf
->pending_full_writes
) == 0)) {
4038 sh
= list_entry(conf
->hold_list
.next
,
4040 conf
->bypass_count
-= conf
->bypass_threshold
;
4041 if (conf
->bypass_count
< 0)
4042 conf
->bypass_count
= 0;
4046 list_del_init(&sh
->lru
);
4047 atomic_inc(&sh
->count
);
4048 BUG_ON(atomic_read(&sh
->count
) != 1);
4052 struct raid5_plug_cb
{
4053 struct blk_plug_cb cb
;
4054 struct list_head list
;
4057 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
4059 struct raid5_plug_cb
*cb
= container_of(
4060 blk_cb
, struct raid5_plug_cb
, cb
);
4061 struct stripe_head
*sh
;
4062 struct mddev
*mddev
= cb
->cb
.data
;
4063 struct r5conf
*conf
= mddev
->private;
4065 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
4066 spin_lock_irq(&conf
->device_lock
);
4067 while (!list_empty(&cb
->list
)) {
4068 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
4069 list_del_init(&sh
->lru
);
4071 * avoid race release_stripe_plug() sees
4072 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4073 * is still in our list
4075 smp_mb__before_clear_bit();
4076 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
4077 __release_stripe(conf
, sh
);
4079 spin_unlock_irq(&conf
->device_lock
);
4084 static void release_stripe_plug(struct mddev
*mddev
,
4085 struct stripe_head
*sh
)
4087 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
4088 raid5_unplug
, mddev
,
4089 sizeof(struct raid5_plug_cb
));
4090 struct raid5_plug_cb
*cb
;
4097 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
4099 if (cb
->list
.next
== NULL
)
4100 INIT_LIST_HEAD(&cb
->list
);
4102 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
4103 list_add_tail(&sh
->lru
, &cb
->list
);
4108 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
4110 struct r5conf
*conf
= mddev
->private;
4111 sector_t logical_sector
, last_sector
;
4112 struct stripe_head
*sh
;
4116 if (mddev
->reshape_position
!= MaxSector
)
4117 /* Skip discard while reshape is happening */
4120 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4121 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4124 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4126 stripe_sectors
= conf
->chunk_sectors
*
4127 (conf
->raid_disks
- conf
->max_degraded
);
4128 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
4130 sector_div(last_sector
, stripe_sectors
);
4132 logical_sector
*= conf
->chunk_sectors
;
4133 last_sector
*= conf
->chunk_sectors
;
4135 for (; logical_sector
< last_sector
;
4136 logical_sector
+= STRIPE_SECTORS
) {
4140 sh
= get_active_stripe(conf
, logical_sector
, 0, 0, 0);
4141 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
4142 TASK_UNINTERRUPTIBLE
);
4143 spin_lock_irq(&sh
->stripe_lock
);
4144 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4145 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4147 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
4148 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
4149 spin_unlock_irq(&sh
->stripe_lock
);
4155 finish_wait(&conf
->wait_for_overlap
, &w
);
4156 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4157 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4159 sh
->dev
[d
].towrite
= bi
;
4160 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
4161 raid5_inc_bi_active_stripes(bi
);
4163 spin_unlock_irq(&sh
->stripe_lock
);
4164 if (conf
->mddev
->bitmap
) {
4166 d
< conf
->raid_disks
- conf
->max_degraded
;
4168 bitmap_startwrite(mddev
->bitmap
,
4172 sh
->bm_seq
= conf
->seq_flush
+ 1;
4173 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
4176 set_bit(STRIPE_HANDLE
, &sh
->state
);
4177 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4178 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4179 atomic_inc(&conf
->preread_active_stripes
);
4180 release_stripe_plug(mddev
, sh
);
4183 remaining
= raid5_dec_bi_active_stripes(bi
);
4184 if (remaining
== 0) {
4185 md_write_end(mddev
);
4190 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
4192 struct r5conf
*conf
= mddev
->private;
4194 sector_t new_sector
;
4195 sector_t logical_sector
, last_sector
;
4196 struct stripe_head
*sh
;
4197 const int rw
= bio_data_dir(bi
);
4200 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
4201 md_flush_request(mddev
, bi
);
4205 md_write_start(mddev
, bi
);
4208 mddev
->reshape_position
== MaxSector
&&
4209 chunk_aligned_read(mddev
,bi
))
4212 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
4213 make_discard_request(mddev
, bi
);
4217 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4218 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4220 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4222 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
4228 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
4229 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
4230 /* spinlock is needed as reshape_progress may be
4231 * 64bit on a 32bit platform, and so it might be
4232 * possible to see a half-updated value
4233 * Of course reshape_progress could change after
4234 * the lock is dropped, so once we get a reference
4235 * to the stripe that we think it is, we will have
4238 spin_lock_irq(&conf
->device_lock
);
4239 if (mddev
->reshape_backwards
4240 ? logical_sector
< conf
->reshape_progress
4241 : logical_sector
>= conf
->reshape_progress
) {
4244 if (mddev
->reshape_backwards
4245 ? logical_sector
< conf
->reshape_safe
4246 : logical_sector
>= conf
->reshape_safe
) {
4247 spin_unlock_irq(&conf
->device_lock
);
4252 spin_unlock_irq(&conf
->device_lock
);
4255 new_sector
= raid5_compute_sector(conf
, logical_sector
,
4258 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4259 (unsigned long long)new_sector
,
4260 (unsigned long long)logical_sector
);
4262 sh
= get_active_stripe(conf
, new_sector
, previous
,
4263 (bi
->bi_rw
&RWA_MASK
), 0);
4265 if (unlikely(previous
)) {
4266 /* expansion might have moved on while waiting for a
4267 * stripe, so we must do the range check again.
4268 * Expansion could still move past after this
4269 * test, but as we are holding a reference to
4270 * 'sh', we know that if that happens,
4271 * STRIPE_EXPANDING will get set and the expansion
4272 * won't proceed until we finish with the stripe.
4275 spin_lock_irq(&conf
->device_lock
);
4276 if (mddev
->reshape_backwards
4277 ? logical_sector
>= conf
->reshape_progress
4278 : logical_sector
< conf
->reshape_progress
)
4279 /* mismatch, need to try again */
4281 spin_unlock_irq(&conf
->device_lock
);
4290 logical_sector
>= mddev
->suspend_lo
&&
4291 logical_sector
< mddev
->suspend_hi
) {
4293 /* As the suspend_* range is controlled by
4294 * userspace, we want an interruptible
4297 flush_signals(current
);
4298 prepare_to_wait(&conf
->wait_for_overlap
,
4299 &w
, TASK_INTERRUPTIBLE
);
4300 if (logical_sector
>= mddev
->suspend_lo
&&
4301 logical_sector
< mddev
->suspend_hi
)
4306 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
4307 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
4308 /* Stripe is busy expanding or
4309 * add failed due to overlap. Flush everything
4312 md_wakeup_thread(mddev
->thread
);
4317 finish_wait(&conf
->wait_for_overlap
, &w
);
4318 set_bit(STRIPE_HANDLE
, &sh
->state
);
4319 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4320 if ((bi
->bi_rw
& REQ_NOIDLE
) &&
4321 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4322 atomic_inc(&conf
->preread_active_stripes
);
4323 release_stripe_plug(mddev
, sh
);
4325 /* cannot get stripe for read-ahead, just give-up */
4326 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4327 finish_wait(&conf
->wait_for_overlap
, &w
);
4332 remaining
= raid5_dec_bi_active_stripes(bi
);
4333 if (remaining
== 0) {
4336 md_write_end(mddev
);
4342 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
4344 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
4346 /* reshaping is quite different to recovery/resync so it is
4347 * handled quite separately ... here.
4349 * On each call to sync_request, we gather one chunk worth of
4350 * destination stripes and flag them as expanding.
4351 * Then we find all the source stripes and request reads.
4352 * As the reads complete, handle_stripe will copy the data
4353 * into the destination stripe and release that stripe.
4355 struct r5conf
*conf
= mddev
->private;
4356 struct stripe_head
*sh
;
4357 sector_t first_sector
, last_sector
;
4358 int raid_disks
= conf
->previous_raid_disks
;
4359 int data_disks
= raid_disks
- conf
->max_degraded
;
4360 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4363 sector_t writepos
, readpos
, safepos
;
4364 sector_t stripe_addr
;
4365 int reshape_sectors
;
4366 struct list_head stripes
;
4368 if (sector_nr
== 0) {
4369 /* If restarting in the middle, skip the initial sectors */
4370 if (mddev
->reshape_backwards
&&
4371 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
4372 sector_nr
= raid5_size(mddev
, 0, 0)
4373 - conf
->reshape_progress
;
4374 } else if (!mddev
->reshape_backwards
&&
4375 conf
->reshape_progress
> 0)
4376 sector_nr
= conf
->reshape_progress
;
4377 sector_div(sector_nr
, new_data_disks
);
4379 mddev
->curr_resync_completed
= sector_nr
;
4380 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4386 /* We need to process a full chunk at a time.
4387 * If old and new chunk sizes differ, we need to process the
4390 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
4391 reshape_sectors
= mddev
->new_chunk_sectors
;
4393 reshape_sectors
= mddev
->chunk_sectors
;
4395 /* We update the metadata at least every 10 seconds, or when
4396 * the data about to be copied would over-write the source of
4397 * the data at the front of the range. i.e. one new_stripe
4398 * along from reshape_progress new_maps to after where
4399 * reshape_safe old_maps to
4401 writepos
= conf
->reshape_progress
;
4402 sector_div(writepos
, new_data_disks
);
4403 readpos
= conf
->reshape_progress
;
4404 sector_div(readpos
, data_disks
);
4405 safepos
= conf
->reshape_safe
;
4406 sector_div(safepos
, data_disks
);
4407 if (mddev
->reshape_backwards
) {
4408 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
4409 readpos
+= reshape_sectors
;
4410 safepos
+= reshape_sectors
;
4412 writepos
+= reshape_sectors
;
4413 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
4414 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
4417 /* Having calculated the 'writepos' possibly use it
4418 * to set 'stripe_addr' which is where we will write to.
4420 if (mddev
->reshape_backwards
) {
4421 BUG_ON(conf
->reshape_progress
== 0);
4422 stripe_addr
= writepos
;
4423 BUG_ON((mddev
->dev_sectors
&
4424 ~((sector_t
)reshape_sectors
- 1))
4425 - reshape_sectors
- stripe_addr
4428 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
4429 stripe_addr
= sector_nr
;
4432 /* 'writepos' is the most advanced device address we might write.
4433 * 'readpos' is the least advanced device address we might read.
4434 * 'safepos' is the least address recorded in the metadata as having
4436 * If there is a min_offset_diff, these are adjusted either by
4437 * increasing the safepos/readpos if diff is negative, or
4438 * increasing writepos if diff is positive.
4439 * If 'readpos' is then behind 'writepos', there is no way that we can
4440 * ensure safety in the face of a crash - that must be done by userspace
4441 * making a backup of the data. So in that case there is no particular
4442 * rush to update metadata.
4443 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4444 * update the metadata to advance 'safepos' to match 'readpos' so that
4445 * we can be safe in the event of a crash.
4446 * So we insist on updating metadata if safepos is behind writepos and
4447 * readpos is beyond writepos.
4448 * In any case, update the metadata every 10 seconds.
4449 * Maybe that number should be configurable, but I'm not sure it is
4450 * worth it.... maybe it could be a multiple of safemode_delay???
4452 if (conf
->min_offset_diff
< 0) {
4453 safepos
+= -conf
->min_offset_diff
;
4454 readpos
+= -conf
->min_offset_diff
;
4456 writepos
+= conf
->min_offset_diff
;
4458 if ((mddev
->reshape_backwards
4459 ? (safepos
> writepos
&& readpos
< writepos
)
4460 : (safepos
< writepos
&& readpos
> writepos
)) ||
4461 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4462 /* Cannot proceed until we've updated the superblock... */
4463 wait_event(conf
->wait_for_overlap
,
4464 atomic_read(&conf
->reshape_stripes
)==0);
4465 mddev
->reshape_position
= conf
->reshape_progress
;
4466 mddev
->curr_resync_completed
= sector_nr
;
4467 conf
->reshape_checkpoint
= jiffies
;
4468 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4469 md_wakeup_thread(mddev
->thread
);
4470 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4471 kthread_should_stop());
4472 spin_lock_irq(&conf
->device_lock
);
4473 conf
->reshape_safe
= mddev
->reshape_position
;
4474 spin_unlock_irq(&conf
->device_lock
);
4475 wake_up(&conf
->wait_for_overlap
);
4476 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4479 INIT_LIST_HEAD(&stripes
);
4480 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
4482 int skipped_disk
= 0;
4483 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
4484 set_bit(STRIPE_EXPANDING
, &sh
->state
);
4485 atomic_inc(&conf
->reshape_stripes
);
4486 /* If any of this stripe is beyond the end of the old
4487 * array, then we need to zero those blocks
4489 for (j
=sh
->disks
; j
--;) {
4491 if (j
== sh
->pd_idx
)
4493 if (conf
->level
== 6 &&
4496 s
= compute_blocknr(sh
, j
, 0);
4497 if (s
< raid5_size(mddev
, 0, 0)) {
4501 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4502 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4503 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4505 if (!skipped_disk
) {
4506 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4507 set_bit(STRIPE_HANDLE
, &sh
->state
);
4509 list_add(&sh
->lru
, &stripes
);
4511 spin_lock_irq(&conf
->device_lock
);
4512 if (mddev
->reshape_backwards
)
4513 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4515 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4516 spin_unlock_irq(&conf
->device_lock
);
4517 /* Ok, those stripe are ready. We can start scheduling
4518 * reads on the source stripes.
4519 * The source stripes are determined by mapping the first and last
4520 * block on the destination stripes.
4523 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4526 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4527 * new_data_disks
- 1),
4529 if (last_sector
>= mddev
->dev_sectors
)
4530 last_sector
= mddev
->dev_sectors
- 1;
4531 while (first_sector
<= last_sector
) {
4532 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4533 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4534 set_bit(STRIPE_HANDLE
, &sh
->state
);
4536 first_sector
+= STRIPE_SECTORS
;
4538 /* Now that the sources are clearly marked, we can release
4539 * the destination stripes
4541 while (!list_empty(&stripes
)) {
4542 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4543 list_del_init(&sh
->lru
);
4546 /* If this takes us to the resync_max point where we have to pause,
4547 * then we need to write out the superblock.
4549 sector_nr
+= reshape_sectors
;
4550 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4551 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4552 /* Cannot proceed until we've updated the superblock... */
4553 wait_event(conf
->wait_for_overlap
,
4554 atomic_read(&conf
->reshape_stripes
) == 0);
4555 mddev
->reshape_position
= conf
->reshape_progress
;
4556 mddev
->curr_resync_completed
= sector_nr
;
4557 conf
->reshape_checkpoint
= jiffies
;
4558 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4559 md_wakeup_thread(mddev
->thread
);
4560 wait_event(mddev
->sb_wait
,
4561 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4562 || kthread_should_stop());
4563 spin_lock_irq(&conf
->device_lock
);
4564 conf
->reshape_safe
= mddev
->reshape_position
;
4565 spin_unlock_irq(&conf
->device_lock
);
4566 wake_up(&conf
->wait_for_overlap
);
4567 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4569 return reshape_sectors
;
4572 /* FIXME go_faster isn't used */
4573 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4575 struct r5conf
*conf
= mddev
->private;
4576 struct stripe_head
*sh
;
4577 sector_t max_sector
= mddev
->dev_sectors
;
4578 sector_t sync_blocks
;
4579 int still_degraded
= 0;
4582 if (sector_nr
>= max_sector
) {
4583 /* just being told to finish up .. nothing much to do */
4585 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4590 if (mddev
->curr_resync
< max_sector
) /* aborted */
4591 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4593 else /* completed sync */
4595 bitmap_close_sync(mddev
->bitmap
);
4600 /* Allow raid5_quiesce to complete */
4601 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4603 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4604 return reshape_request(mddev
, sector_nr
, skipped
);
4606 /* No need to check resync_max as we never do more than one
4607 * stripe, and as resync_max will always be on a chunk boundary,
4608 * if the check in md_do_sync didn't fire, there is no chance
4609 * of overstepping resync_max here
4612 /* if there is too many failed drives and we are trying
4613 * to resync, then assert that we are finished, because there is
4614 * nothing we can do.
4616 if (mddev
->degraded
>= conf
->max_degraded
&&
4617 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4618 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4622 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4623 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4624 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4625 /* we can skip this block, and probably more */
4626 sync_blocks
/= STRIPE_SECTORS
;
4628 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4631 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4633 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4635 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4636 /* make sure we don't swamp the stripe cache if someone else
4637 * is trying to get access
4639 schedule_timeout_uninterruptible(1);
4641 /* Need to check if array will still be degraded after recovery/resync
4642 * We don't need to check the 'failed' flag as when that gets set,
4645 for (i
= 0; i
< conf
->raid_disks
; i
++)
4646 if (conf
->disks
[i
].rdev
== NULL
)
4649 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4651 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
4656 return STRIPE_SECTORS
;
4659 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
4661 /* We may not be able to submit a whole bio at once as there
4662 * may not be enough stripe_heads available.
4663 * We cannot pre-allocate enough stripe_heads as we may need
4664 * more than exist in the cache (if we allow ever large chunks).
4665 * So we do one stripe head at a time and record in
4666 * ->bi_hw_segments how many have been done.
4668 * We *know* that this entire raid_bio is in one chunk, so
4669 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4671 struct stripe_head
*sh
;
4673 sector_t sector
, logical_sector
, last_sector
;
4678 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4679 sector
= raid5_compute_sector(conf
, logical_sector
,
4681 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4683 for (; logical_sector
< last_sector
;
4684 logical_sector
+= STRIPE_SECTORS
,
4685 sector
+= STRIPE_SECTORS
,
4688 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
4689 /* already done this stripe */
4692 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4695 /* failed to get a stripe - must wait */
4696 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4697 conf
->retry_read_aligned
= raid_bio
;
4701 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4703 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4704 conf
->retry_read_aligned
= raid_bio
;
4708 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
4713 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
4715 bio_endio(raid_bio
, 0);
4716 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4717 wake_up(&conf
->wait_for_stripe
);
4721 #define MAX_STRIPE_BATCH 8
4722 static int handle_active_stripes(struct r5conf
*conf
)
4724 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
4725 int i
, batch_size
= 0;
4727 while (batch_size
< MAX_STRIPE_BATCH
&&
4728 (sh
= __get_priority_stripe(conf
)) != NULL
)
4729 batch
[batch_size
++] = sh
;
4731 if (batch_size
== 0)
4733 spin_unlock_irq(&conf
->device_lock
);
4735 for (i
= 0; i
< batch_size
; i
++)
4736 handle_stripe(batch
[i
]);
4740 spin_lock_irq(&conf
->device_lock
);
4741 for (i
= 0; i
< batch_size
; i
++)
4742 __release_stripe(conf
, batch
[i
]);
4747 * This is our raid5 kernel thread.
4749 * We scan the hash table for stripes which can be handled now.
4750 * During the scan, completed stripes are saved for us by the interrupt
4751 * handler, so that they will not have to wait for our next wakeup.
4753 static void raid5d(struct md_thread
*thread
)
4755 struct mddev
*mddev
= thread
->mddev
;
4756 struct r5conf
*conf
= mddev
->private;
4758 struct blk_plug plug
;
4760 pr_debug("+++ raid5d active\n");
4762 md_check_recovery(mddev
);
4764 blk_start_plug(&plug
);
4766 spin_lock_irq(&conf
->device_lock
);
4772 !list_empty(&conf
->bitmap_list
)) {
4773 /* Now is a good time to flush some bitmap updates */
4775 spin_unlock_irq(&conf
->device_lock
);
4776 bitmap_unplug(mddev
->bitmap
);
4777 spin_lock_irq(&conf
->device_lock
);
4778 conf
->seq_write
= conf
->seq_flush
;
4779 activate_bit_delay(conf
);
4781 raid5_activate_delayed(conf
);
4783 while ((bio
= remove_bio_from_retry(conf
))) {
4785 spin_unlock_irq(&conf
->device_lock
);
4786 ok
= retry_aligned_read(conf
, bio
);
4787 spin_lock_irq(&conf
->device_lock
);
4793 batch_size
= handle_active_stripes(conf
);
4796 handled
+= batch_size
;
4798 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
4799 spin_unlock_irq(&conf
->device_lock
);
4800 md_check_recovery(mddev
);
4801 spin_lock_irq(&conf
->device_lock
);
4804 pr_debug("%d stripes handled\n", handled
);
4806 spin_unlock_irq(&conf
->device_lock
);
4808 async_tx_issue_pending_all();
4809 blk_finish_plug(&plug
);
4811 pr_debug("--- raid5d inactive\n");
4815 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
4817 struct r5conf
*conf
= mddev
->private;
4819 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4825 raid5_set_cache_size(struct mddev
*mddev
, int size
)
4827 struct r5conf
*conf
= mddev
->private;
4830 if (size
<= 16 || size
> 32768)
4832 while (size
< conf
->max_nr_stripes
) {
4833 if (drop_one_stripe(conf
))
4834 conf
->max_nr_stripes
--;
4838 err
= md_allow_write(mddev
);
4841 while (size
> conf
->max_nr_stripes
) {
4842 if (grow_one_stripe(conf
))
4843 conf
->max_nr_stripes
++;
4848 EXPORT_SYMBOL(raid5_set_cache_size
);
4851 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
4853 struct r5conf
*conf
= mddev
->private;
4857 if (len
>= PAGE_SIZE
)
4862 if (strict_strtoul(page
, 10, &new))
4864 err
= raid5_set_cache_size(mddev
, new);
4870 static struct md_sysfs_entry
4871 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4872 raid5_show_stripe_cache_size
,
4873 raid5_store_stripe_cache_size
);
4876 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
4878 struct r5conf
*conf
= mddev
->private;
4880 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4886 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
4888 struct r5conf
*conf
= mddev
->private;
4890 if (len
>= PAGE_SIZE
)
4895 if (strict_strtoul(page
, 10, &new))
4897 if (new > conf
->max_nr_stripes
)
4899 conf
->bypass_threshold
= new;
4903 static struct md_sysfs_entry
4904 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4906 raid5_show_preread_threshold
,
4907 raid5_store_preread_threshold
);
4910 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
4912 struct r5conf
*conf
= mddev
->private;
4914 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4919 static struct md_sysfs_entry
4920 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4922 static struct attribute
*raid5_attrs
[] = {
4923 &raid5_stripecache_size
.attr
,
4924 &raid5_stripecache_active
.attr
,
4925 &raid5_preread_bypass_threshold
.attr
,
4928 static struct attribute_group raid5_attrs_group
= {
4930 .attrs
= raid5_attrs
,
4934 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
4936 struct r5conf
*conf
= mddev
->private;
4939 sectors
= mddev
->dev_sectors
;
4941 /* size is defined by the smallest of previous and new size */
4942 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
4944 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
4945 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
4946 return sectors
* (raid_disks
- conf
->max_degraded
);
4949 static void raid5_free_percpu(struct r5conf
*conf
)
4951 struct raid5_percpu
*percpu
;
4958 for_each_possible_cpu(cpu
) {
4959 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4960 safe_put_page(percpu
->spare_page
);
4961 kfree(percpu
->scribble
);
4963 #ifdef CONFIG_HOTPLUG_CPU
4964 unregister_cpu_notifier(&conf
->cpu_notify
);
4968 free_percpu(conf
->percpu
);
4971 static void free_conf(struct r5conf
*conf
)
4973 shrink_stripes(conf
);
4974 raid5_free_percpu(conf
);
4976 kfree(conf
->stripe_hashtbl
);
4980 #ifdef CONFIG_HOTPLUG_CPU
4981 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
4984 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
4985 long cpu
= (long)hcpu
;
4986 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4989 case CPU_UP_PREPARE
:
4990 case CPU_UP_PREPARE_FROZEN
:
4991 if (conf
->level
== 6 && !percpu
->spare_page
)
4992 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
4993 if (!percpu
->scribble
)
4994 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4996 if (!percpu
->scribble
||
4997 (conf
->level
== 6 && !percpu
->spare_page
)) {
4998 safe_put_page(percpu
->spare_page
);
4999 kfree(percpu
->scribble
);
5000 pr_err("%s: failed memory allocation for cpu%ld\n",
5002 return notifier_from_errno(-ENOMEM
);
5006 case CPU_DEAD_FROZEN
:
5007 safe_put_page(percpu
->spare_page
);
5008 kfree(percpu
->scribble
);
5009 percpu
->spare_page
= NULL
;
5010 percpu
->scribble
= NULL
;
5019 static int raid5_alloc_percpu(struct r5conf
*conf
)
5022 struct page
*spare_page
;
5023 struct raid5_percpu __percpu
*allcpus
;
5027 allcpus
= alloc_percpu(struct raid5_percpu
);
5030 conf
->percpu
= allcpus
;
5034 for_each_present_cpu(cpu
) {
5035 if (conf
->level
== 6) {
5036 spare_page
= alloc_page(GFP_KERNEL
);
5041 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
5043 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
5048 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
5050 #ifdef CONFIG_HOTPLUG_CPU
5051 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
5052 conf
->cpu_notify
.priority
= 0;
5054 err
= register_cpu_notifier(&conf
->cpu_notify
);
5061 static struct r5conf
*setup_conf(struct mddev
*mddev
)
5063 struct r5conf
*conf
;
5064 int raid_disk
, memory
, max_disks
;
5065 struct md_rdev
*rdev
;
5066 struct disk_info
*disk
;
5069 if (mddev
->new_level
!= 5
5070 && mddev
->new_level
!= 4
5071 && mddev
->new_level
!= 6) {
5072 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5073 mdname(mddev
), mddev
->new_level
);
5074 return ERR_PTR(-EIO
);
5076 if ((mddev
->new_level
== 5
5077 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
5078 (mddev
->new_level
== 6
5079 && !algorithm_valid_raid6(mddev
->new_layout
))) {
5080 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
5081 mdname(mddev
), mddev
->new_layout
);
5082 return ERR_PTR(-EIO
);
5084 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
5085 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5086 mdname(mddev
), mddev
->raid_disks
);
5087 return ERR_PTR(-EINVAL
);
5090 if (!mddev
->new_chunk_sectors
||
5091 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
5092 !is_power_of_2(mddev
->new_chunk_sectors
)) {
5093 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
5094 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
5095 return ERR_PTR(-EINVAL
);
5098 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
5101 spin_lock_init(&conf
->device_lock
);
5102 init_waitqueue_head(&conf
->wait_for_stripe
);
5103 init_waitqueue_head(&conf
->wait_for_overlap
);
5104 INIT_LIST_HEAD(&conf
->handle_list
);
5105 INIT_LIST_HEAD(&conf
->hold_list
);
5106 INIT_LIST_HEAD(&conf
->delayed_list
);
5107 INIT_LIST_HEAD(&conf
->bitmap_list
);
5108 INIT_LIST_HEAD(&conf
->inactive_list
);
5109 atomic_set(&conf
->active_stripes
, 0);
5110 atomic_set(&conf
->preread_active_stripes
, 0);
5111 atomic_set(&conf
->active_aligned_reads
, 0);
5112 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
5113 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
5115 conf
->raid_disks
= mddev
->raid_disks
;
5116 if (mddev
->reshape_position
== MaxSector
)
5117 conf
->previous_raid_disks
= mddev
->raid_disks
;
5119 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5120 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
5121 conf
->scribble_len
= scribble_len(max_disks
);
5123 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
5128 conf
->mddev
= mddev
;
5130 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
5133 conf
->level
= mddev
->new_level
;
5134 if (raid5_alloc_percpu(conf
) != 0)
5137 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
5139 rdev_for_each(rdev
, mddev
) {
5140 raid_disk
= rdev
->raid_disk
;
5141 if (raid_disk
>= max_disks
5144 disk
= conf
->disks
+ raid_disk
;
5146 if (test_bit(Replacement
, &rdev
->flags
)) {
5147 if (disk
->replacement
)
5149 disk
->replacement
= rdev
;
5156 if (test_bit(In_sync
, &rdev
->flags
)) {
5157 char b
[BDEVNAME_SIZE
];
5158 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
5160 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
5161 } else if (rdev
->saved_raid_disk
!= raid_disk
)
5162 /* Cannot rely on bitmap to complete recovery */
5166 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5167 conf
->level
= mddev
->new_level
;
5168 if (conf
->level
== 6)
5169 conf
->max_degraded
= 2;
5171 conf
->max_degraded
= 1;
5172 conf
->algorithm
= mddev
->new_layout
;
5173 conf
->max_nr_stripes
= NR_STRIPES
;
5174 conf
->reshape_progress
= mddev
->reshape_position
;
5175 if (conf
->reshape_progress
!= MaxSector
) {
5176 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
5177 conf
->prev_algo
= mddev
->layout
;
5180 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
5181 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
5182 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
5184 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5185 mdname(mddev
), memory
);
5188 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
5189 mdname(mddev
), memory
);
5191 sprintf(pers_name
, "raid%d", mddev
->new_level
);
5192 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
5193 if (!conf
->thread
) {
5195 "md/raid:%s: couldn't allocate thread.\n",
5205 return ERR_PTR(-EIO
);
5207 return ERR_PTR(-ENOMEM
);
5211 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
5214 case ALGORITHM_PARITY_0
:
5215 if (raid_disk
< max_degraded
)
5218 case ALGORITHM_PARITY_N
:
5219 if (raid_disk
>= raid_disks
- max_degraded
)
5222 case ALGORITHM_PARITY_0_6
:
5223 if (raid_disk
== 0 ||
5224 raid_disk
== raid_disks
- 1)
5227 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5228 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5229 case ALGORITHM_LEFT_SYMMETRIC_6
:
5230 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5231 if (raid_disk
== raid_disks
- 1)
5237 static int run(struct mddev
*mddev
)
5239 struct r5conf
*conf
;
5240 int working_disks
= 0;
5241 int dirty_parity_disks
= 0;
5242 struct md_rdev
*rdev
;
5243 sector_t reshape_offset
= 0;
5245 long long min_offset_diff
= 0;
5248 if (mddev
->recovery_cp
!= MaxSector
)
5249 printk(KERN_NOTICE
"md/raid:%s: not clean"
5250 " -- starting background reconstruction\n",
5253 rdev_for_each(rdev
, mddev
) {
5255 if (rdev
->raid_disk
< 0)
5257 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
5259 min_offset_diff
= diff
;
5261 } else if (mddev
->reshape_backwards
&&
5262 diff
< min_offset_diff
)
5263 min_offset_diff
= diff
;
5264 else if (!mddev
->reshape_backwards
&&
5265 diff
> min_offset_diff
)
5266 min_offset_diff
= diff
;
5269 if (mddev
->reshape_position
!= MaxSector
) {
5270 /* Check that we can continue the reshape.
5271 * Difficulties arise if the stripe we would write to
5272 * next is at or after the stripe we would read from next.
5273 * For a reshape that changes the number of devices, this
5274 * is only possible for a very short time, and mdadm makes
5275 * sure that time appears to have past before assembling
5276 * the array. So we fail if that time hasn't passed.
5277 * For a reshape that keeps the number of devices the same
5278 * mdadm must be monitoring the reshape can keeping the
5279 * critical areas read-only and backed up. It will start
5280 * the array in read-only mode, so we check for that.
5282 sector_t here_new
, here_old
;
5284 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
5286 if (mddev
->new_level
!= mddev
->level
) {
5287 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
5288 "required - aborting.\n",
5292 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5293 /* reshape_position must be on a new-stripe boundary, and one
5294 * further up in new geometry must map after here in old
5297 here_new
= mddev
->reshape_position
;
5298 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
5299 (mddev
->raid_disks
- max_degraded
))) {
5300 printk(KERN_ERR
"md/raid:%s: reshape_position not "
5301 "on a stripe boundary\n", mdname(mddev
));
5304 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
5305 /* here_new is the stripe we will write to */
5306 here_old
= mddev
->reshape_position
;
5307 sector_div(here_old
, mddev
->chunk_sectors
*
5308 (old_disks
-max_degraded
));
5309 /* here_old is the first stripe that we might need to read
5311 if (mddev
->delta_disks
== 0) {
5312 if ((here_new
* mddev
->new_chunk_sectors
!=
5313 here_old
* mddev
->chunk_sectors
)) {
5314 printk(KERN_ERR
"md/raid:%s: reshape position is"
5315 " confused - aborting\n", mdname(mddev
));
5318 /* We cannot be sure it is safe to start an in-place
5319 * reshape. It is only safe if user-space is monitoring
5320 * and taking constant backups.
5321 * mdadm always starts a situation like this in
5322 * readonly mode so it can take control before
5323 * allowing any writes. So just check for that.
5325 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
5326 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
5327 /* not really in-place - so OK */;
5328 else if (mddev
->ro
== 0) {
5329 printk(KERN_ERR
"md/raid:%s: in-place reshape "
5330 "must be started in read-only mode "
5335 } else if (mddev
->reshape_backwards
5336 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
5337 here_old
* mddev
->chunk_sectors
)
5338 : (here_new
* mddev
->new_chunk_sectors
>=
5339 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
5340 /* Reading from the same stripe as writing to - bad */
5341 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
5342 "auto-recovery - aborting.\n",
5346 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
5348 /* OK, we should be able to continue; */
5350 BUG_ON(mddev
->level
!= mddev
->new_level
);
5351 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
5352 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
5353 BUG_ON(mddev
->delta_disks
!= 0);
5356 if (mddev
->private == NULL
)
5357 conf
= setup_conf(mddev
);
5359 conf
= mddev
->private;
5362 return PTR_ERR(conf
);
5364 conf
->min_offset_diff
= min_offset_diff
;
5365 mddev
->thread
= conf
->thread
;
5366 conf
->thread
= NULL
;
5367 mddev
->private = conf
;
5369 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
5371 rdev
= conf
->disks
[i
].rdev
;
5372 if (!rdev
&& conf
->disks
[i
].replacement
) {
5373 /* The replacement is all we have yet */
5374 rdev
= conf
->disks
[i
].replacement
;
5375 conf
->disks
[i
].replacement
= NULL
;
5376 clear_bit(Replacement
, &rdev
->flags
);
5377 conf
->disks
[i
].rdev
= rdev
;
5381 if (conf
->disks
[i
].replacement
&&
5382 conf
->reshape_progress
!= MaxSector
) {
5383 /* replacements and reshape simply do not mix. */
5384 printk(KERN_ERR
"md: cannot handle concurrent "
5385 "replacement and reshape.\n");
5388 if (test_bit(In_sync
, &rdev
->flags
)) {
5392 /* This disc is not fully in-sync. However if it
5393 * just stored parity (beyond the recovery_offset),
5394 * when we don't need to be concerned about the
5395 * array being dirty.
5396 * When reshape goes 'backwards', we never have
5397 * partially completed devices, so we only need
5398 * to worry about reshape going forwards.
5400 /* Hack because v0.91 doesn't store recovery_offset properly. */
5401 if (mddev
->major_version
== 0 &&
5402 mddev
->minor_version
> 90)
5403 rdev
->recovery_offset
= reshape_offset
;
5405 if (rdev
->recovery_offset
< reshape_offset
) {
5406 /* We need to check old and new layout */
5407 if (!only_parity(rdev
->raid_disk
,
5410 conf
->max_degraded
))
5413 if (!only_parity(rdev
->raid_disk
,
5415 conf
->previous_raid_disks
,
5416 conf
->max_degraded
))
5418 dirty_parity_disks
++;
5422 * 0 for a fully functional array, 1 or 2 for a degraded array.
5424 mddev
->degraded
= calc_degraded(conf
);
5426 if (has_failed(conf
)) {
5427 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
5428 " (%d/%d failed)\n",
5429 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
5433 /* device size must be a multiple of chunk size */
5434 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
5435 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
5437 if (mddev
->degraded
> dirty_parity_disks
&&
5438 mddev
->recovery_cp
!= MaxSector
) {
5439 if (mddev
->ok_start_degraded
)
5441 "md/raid:%s: starting dirty degraded array"
5442 " - data corruption possible.\n",
5446 "md/raid:%s: cannot start dirty degraded array.\n",
5452 if (mddev
->degraded
== 0)
5453 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
5454 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
5455 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
5458 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
5459 " out of %d devices, algorithm %d\n",
5460 mdname(mddev
), conf
->level
,
5461 mddev
->raid_disks
- mddev
->degraded
,
5462 mddev
->raid_disks
, mddev
->new_layout
);
5464 print_raid5_conf(conf
);
5466 if (conf
->reshape_progress
!= MaxSector
) {
5467 conf
->reshape_safe
= conf
->reshape_progress
;
5468 atomic_set(&conf
->reshape_stripes
, 0);
5469 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5470 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5471 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5472 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5473 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5478 /* Ok, everything is just fine now */
5479 if (mddev
->to_remove
== &raid5_attrs_group
)
5480 mddev
->to_remove
= NULL
;
5481 else if (mddev
->kobj
.sd
&&
5482 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
5484 "raid5: failed to create sysfs attributes for %s\n",
5486 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5490 bool discard_supported
= true;
5491 /* read-ahead size must cover two whole stripes, which
5492 * is 2 * (datadisks) * chunksize where 'n' is the
5493 * number of raid devices
5495 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
5496 int stripe
= data_disks
*
5497 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
5498 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5499 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5501 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
5503 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
5504 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
5506 chunk_size
= mddev
->chunk_sectors
<< 9;
5507 blk_queue_io_min(mddev
->queue
, chunk_size
);
5508 blk_queue_io_opt(mddev
->queue
, chunk_size
*
5509 (conf
->raid_disks
- conf
->max_degraded
));
5511 * We can only discard a whole stripe. It doesn't make sense to
5512 * discard data disk but write parity disk
5514 stripe
= stripe
* PAGE_SIZE
;
5515 mddev
->queue
->limits
.discard_alignment
= stripe
;
5516 mddev
->queue
->limits
.discard_granularity
= stripe
;
5518 * unaligned part of discard request will be ignored, so can't
5519 * guarantee discard_zerors_data
5521 mddev
->queue
->limits
.discard_zeroes_data
= 0;
5523 rdev_for_each(rdev
, mddev
) {
5524 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5525 rdev
->data_offset
<< 9);
5526 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5527 rdev
->new_data_offset
<< 9);
5529 * discard_zeroes_data is required, otherwise data
5530 * could be lost. Consider a scenario: discard a stripe
5531 * (the stripe could be inconsistent if
5532 * discard_zeroes_data is 0); write one disk of the
5533 * stripe (the stripe could be inconsistent again
5534 * depending on which disks are used to calculate
5535 * parity); the disk is broken; The stripe data of this
5538 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
5539 !bdev_get_queue(rdev
->bdev
)->
5540 limits
.discard_zeroes_data
)
5541 discard_supported
= false;
5544 if (discard_supported
&&
5545 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
5546 mddev
->queue
->limits
.discard_granularity
>= stripe
)
5547 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
5550 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
5556 md_unregister_thread(&mddev
->thread
);
5557 print_raid5_conf(conf
);
5559 mddev
->private = NULL
;
5560 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
5564 static int stop(struct mddev
*mddev
)
5566 struct r5conf
*conf
= mddev
->private;
5568 md_unregister_thread(&mddev
->thread
);
5570 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
5572 mddev
->private = NULL
;
5573 mddev
->to_remove
= &raid5_attrs_group
;
5577 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
5579 struct r5conf
*conf
= mddev
->private;
5582 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
5583 mddev
->chunk_sectors
/ 2, mddev
->layout
);
5584 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
5585 for (i
= 0; i
< conf
->raid_disks
; i
++)
5586 seq_printf (seq
, "%s",
5587 conf
->disks
[i
].rdev
&&
5588 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5589 seq_printf (seq
, "]");
5592 static void print_raid5_conf (struct r5conf
*conf
)
5595 struct disk_info
*tmp
;
5597 printk(KERN_DEBUG
"RAID conf printout:\n");
5599 printk("(conf==NULL)\n");
5602 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5604 conf
->raid_disks
- conf
->mddev
->degraded
);
5606 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5607 char b
[BDEVNAME_SIZE
];
5608 tmp
= conf
->disks
+ i
;
5610 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5611 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5612 bdevname(tmp
->rdev
->bdev
, b
));
5616 static int raid5_spare_active(struct mddev
*mddev
)
5619 struct r5conf
*conf
= mddev
->private;
5620 struct disk_info
*tmp
;
5622 unsigned long flags
;
5624 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5625 tmp
= conf
->disks
+ i
;
5626 if (tmp
->replacement
5627 && tmp
->replacement
->recovery_offset
== MaxSector
5628 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
5629 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
5630 /* Replacement has just become active. */
5632 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
5635 /* Replaced device not technically faulty,
5636 * but we need to be sure it gets removed
5637 * and never re-added.
5639 set_bit(Faulty
, &tmp
->rdev
->flags
);
5640 sysfs_notify_dirent_safe(
5641 tmp
->rdev
->sysfs_state
);
5643 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
5644 } else if (tmp
->rdev
5645 && tmp
->rdev
->recovery_offset
== MaxSector
5646 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
5647 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
5649 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
5652 spin_lock_irqsave(&conf
->device_lock
, flags
);
5653 mddev
->degraded
= calc_degraded(conf
);
5654 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5655 print_raid5_conf(conf
);
5659 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5661 struct r5conf
*conf
= mddev
->private;
5663 int number
= rdev
->raid_disk
;
5664 struct md_rdev
**rdevp
;
5665 struct disk_info
*p
= conf
->disks
+ number
;
5667 print_raid5_conf(conf
);
5668 if (rdev
== p
->rdev
)
5670 else if (rdev
== p
->replacement
)
5671 rdevp
= &p
->replacement
;
5675 if (number
>= conf
->raid_disks
&&
5676 conf
->reshape_progress
== MaxSector
)
5677 clear_bit(In_sync
, &rdev
->flags
);
5679 if (test_bit(In_sync
, &rdev
->flags
) ||
5680 atomic_read(&rdev
->nr_pending
)) {
5684 /* Only remove non-faulty devices if recovery
5687 if (!test_bit(Faulty
, &rdev
->flags
) &&
5688 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
5689 !has_failed(conf
) &&
5690 (!p
->replacement
|| p
->replacement
== rdev
) &&
5691 number
< conf
->raid_disks
) {
5697 if (atomic_read(&rdev
->nr_pending
)) {
5698 /* lost the race, try later */
5701 } else if (p
->replacement
) {
5702 /* We must have just cleared 'rdev' */
5703 p
->rdev
= p
->replacement
;
5704 clear_bit(Replacement
, &p
->replacement
->flags
);
5705 smp_mb(); /* Make sure other CPUs may see both as identical
5706 * but will never see neither - if they are careful
5708 p
->replacement
= NULL
;
5709 clear_bit(WantReplacement
, &rdev
->flags
);
5711 /* We might have just removed the Replacement as faulty-
5712 * clear the bit just in case
5714 clear_bit(WantReplacement
, &rdev
->flags
);
5717 print_raid5_conf(conf
);
5721 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5723 struct r5conf
*conf
= mddev
->private;
5726 struct disk_info
*p
;
5728 int last
= conf
->raid_disks
- 1;
5730 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
5733 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
5734 /* no point adding a device */
5737 if (rdev
->raid_disk
>= 0)
5738 first
= last
= rdev
->raid_disk
;
5741 * find the disk ... but prefer rdev->saved_raid_disk
5744 if (rdev
->saved_raid_disk
>= 0 &&
5745 rdev
->saved_raid_disk
>= first
&&
5746 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5747 first
= rdev
->saved_raid_disk
;
5749 for (disk
= first
; disk
<= last
; disk
++) {
5750 p
= conf
->disks
+ disk
;
5751 if (p
->rdev
== NULL
) {
5752 clear_bit(In_sync
, &rdev
->flags
);
5753 rdev
->raid_disk
= disk
;
5755 if (rdev
->saved_raid_disk
!= disk
)
5757 rcu_assign_pointer(p
->rdev
, rdev
);
5761 for (disk
= first
; disk
<= last
; disk
++) {
5762 p
= conf
->disks
+ disk
;
5763 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
5764 p
->replacement
== NULL
) {
5765 clear_bit(In_sync
, &rdev
->flags
);
5766 set_bit(Replacement
, &rdev
->flags
);
5767 rdev
->raid_disk
= disk
;
5770 rcu_assign_pointer(p
->replacement
, rdev
);
5775 print_raid5_conf(conf
);
5779 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
5781 /* no resync is happening, and there is enough space
5782 * on all devices, so we can resize.
5783 * We need to make sure resync covers any new space.
5784 * If the array is shrinking we should possibly wait until
5785 * any io in the removed space completes, but it hardly seems
5789 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5790 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
5791 if (mddev
->external_size
&&
5792 mddev
->array_sectors
> newsize
)
5794 if (mddev
->bitmap
) {
5795 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
5799 md_set_array_sectors(mddev
, newsize
);
5800 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5801 revalidate_disk(mddev
->gendisk
);
5802 if (sectors
> mddev
->dev_sectors
&&
5803 mddev
->recovery_cp
> mddev
->dev_sectors
) {
5804 mddev
->recovery_cp
= mddev
->dev_sectors
;
5805 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5807 mddev
->dev_sectors
= sectors
;
5808 mddev
->resync_max_sectors
= sectors
;
5812 static int check_stripe_cache(struct mddev
*mddev
)
5814 /* Can only proceed if there are plenty of stripe_heads.
5815 * We need a minimum of one full stripe,, and for sensible progress
5816 * it is best to have about 4 times that.
5817 * If we require 4 times, then the default 256 4K stripe_heads will
5818 * allow for chunk sizes up to 256K, which is probably OK.
5819 * If the chunk size is greater, user-space should request more
5820 * stripe_heads first.
5822 struct r5conf
*conf
= mddev
->private;
5823 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5824 > conf
->max_nr_stripes
||
5825 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5826 > conf
->max_nr_stripes
) {
5827 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5829 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5836 static int check_reshape(struct mddev
*mddev
)
5838 struct r5conf
*conf
= mddev
->private;
5840 if (mddev
->delta_disks
== 0 &&
5841 mddev
->new_layout
== mddev
->layout
&&
5842 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5843 return 0; /* nothing to do */
5844 if (has_failed(conf
))
5846 if (mddev
->delta_disks
< 0) {
5847 /* We might be able to shrink, but the devices must
5848 * be made bigger first.
5849 * For raid6, 4 is the minimum size.
5850 * Otherwise 2 is the minimum
5853 if (mddev
->level
== 6)
5855 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5859 if (!check_stripe_cache(mddev
))
5862 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
5865 static int raid5_start_reshape(struct mddev
*mddev
)
5867 struct r5conf
*conf
= mddev
->private;
5868 struct md_rdev
*rdev
;
5870 unsigned long flags
;
5872 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5875 if (!check_stripe_cache(mddev
))
5878 if (has_failed(conf
))
5881 rdev_for_each(rdev
, mddev
) {
5882 if (!test_bit(In_sync
, &rdev
->flags
)
5883 && !test_bit(Faulty
, &rdev
->flags
))
5887 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5888 /* Not enough devices even to make a degraded array
5893 /* Refuse to reduce size of the array. Any reductions in
5894 * array size must be through explicit setting of array_size
5897 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5898 < mddev
->array_sectors
) {
5899 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5900 "before number of disks\n", mdname(mddev
));
5904 atomic_set(&conf
->reshape_stripes
, 0);
5905 spin_lock_irq(&conf
->device_lock
);
5906 conf
->previous_raid_disks
= conf
->raid_disks
;
5907 conf
->raid_disks
+= mddev
->delta_disks
;
5908 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
5909 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5910 conf
->prev_algo
= conf
->algorithm
;
5911 conf
->algorithm
= mddev
->new_layout
;
5913 /* Code that selects data_offset needs to see the generation update
5914 * if reshape_progress has been set - so a memory barrier needed.
5917 if (mddev
->reshape_backwards
)
5918 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
5920 conf
->reshape_progress
= 0;
5921 conf
->reshape_safe
= conf
->reshape_progress
;
5922 spin_unlock_irq(&conf
->device_lock
);
5924 /* Add some new drives, as many as will fit.
5925 * We know there are enough to make the newly sized array work.
5926 * Don't add devices if we are reducing the number of
5927 * devices in the array. This is because it is not possible
5928 * to correctly record the "partially reconstructed" state of
5929 * such devices during the reshape and confusion could result.
5931 if (mddev
->delta_disks
>= 0) {
5932 rdev_for_each(rdev
, mddev
)
5933 if (rdev
->raid_disk
< 0 &&
5934 !test_bit(Faulty
, &rdev
->flags
)) {
5935 if (raid5_add_disk(mddev
, rdev
) == 0) {
5937 >= conf
->previous_raid_disks
)
5938 set_bit(In_sync
, &rdev
->flags
);
5940 rdev
->recovery_offset
= 0;
5942 if (sysfs_link_rdev(mddev
, rdev
))
5943 /* Failure here is OK */;
5945 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
5946 && !test_bit(Faulty
, &rdev
->flags
)) {
5947 /* This is a spare that was manually added */
5948 set_bit(In_sync
, &rdev
->flags
);
5951 /* When a reshape changes the number of devices,
5952 * ->degraded is measured against the larger of the
5953 * pre and post number of devices.
5955 spin_lock_irqsave(&conf
->device_lock
, flags
);
5956 mddev
->degraded
= calc_degraded(conf
);
5957 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5959 mddev
->raid_disks
= conf
->raid_disks
;
5960 mddev
->reshape_position
= conf
->reshape_progress
;
5961 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5963 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5964 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5965 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5966 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5967 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5969 if (!mddev
->sync_thread
) {
5970 mddev
->recovery
= 0;
5971 spin_lock_irq(&conf
->device_lock
);
5972 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
5973 rdev_for_each(rdev
, mddev
)
5974 rdev
->new_data_offset
= rdev
->data_offset
;
5976 conf
->reshape_progress
= MaxSector
;
5977 mddev
->reshape_position
= MaxSector
;
5978 spin_unlock_irq(&conf
->device_lock
);
5981 conf
->reshape_checkpoint
= jiffies
;
5982 md_wakeup_thread(mddev
->sync_thread
);
5983 md_new_event(mddev
);
5987 /* This is called from the reshape thread and should make any
5988 * changes needed in 'conf'
5990 static void end_reshape(struct r5conf
*conf
)
5993 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
5994 struct md_rdev
*rdev
;
5996 spin_lock_irq(&conf
->device_lock
);
5997 conf
->previous_raid_disks
= conf
->raid_disks
;
5998 rdev_for_each(rdev
, conf
->mddev
)
5999 rdev
->data_offset
= rdev
->new_data_offset
;
6001 conf
->reshape_progress
= MaxSector
;
6002 spin_unlock_irq(&conf
->device_lock
);
6003 wake_up(&conf
->wait_for_overlap
);
6005 /* read-ahead size must cover two whole stripes, which is
6006 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6008 if (conf
->mddev
->queue
) {
6009 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
6010 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
6012 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6013 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6018 /* This is called from the raid5d thread with mddev_lock held.
6019 * It makes config changes to the device.
6021 static void raid5_finish_reshape(struct mddev
*mddev
)
6023 struct r5conf
*conf
= mddev
->private;
6025 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
6027 if (mddev
->delta_disks
> 0) {
6028 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6029 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
6030 revalidate_disk(mddev
->gendisk
);
6033 spin_lock_irq(&conf
->device_lock
);
6034 mddev
->degraded
= calc_degraded(conf
);
6035 spin_unlock_irq(&conf
->device_lock
);
6036 for (d
= conf
->raid_disks
;
6037 d
< conf
->raid_disks
- mddev
->delta_disks
;
6039 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
6041 clear_bit(In_sync
, &rdev
->flags
);
6042 rdev
= conf
->disks
[d
].replacement
;
6044 clear_bit(In_sync
, &rdev
->flags
);
6047 mddev
->layout
= conf
->algorithm
;
6048 mddev
->chunk_sectors
= conf
->chunk_sectors
;
6049 mddev
->reshape_position
= MaxSector
;
6050 mddev
->delta_disks
= 0;
6051 mddev
->reshape_backwards
= 0;
6055 static void raid5_quiesce(struct mddev
*mddev
, int state
)
6057 struct r5conf
*conf
= mddev
->private;
6060 case 2: /* resume for a suspend */
6061 wake_up(&conf
->wait_for_overlap
);
6064 case 1: /* stop all writes */
6065 spin_lock_irq(&conf
->device_lock
);
6066 /* '2' tells resync/reshape to pause so that all
6067 * active stripes can drain
6070 wait_event_lock_irq(conf
->wait_for_stripe
,
6071 atomic_read(&conf
->active_stripes
) == 0 &&
6072 atomic_read(&conf
->active_aligned_reads
) == 0,
6073 conf
->device_lock
, /* nothing */);
6075 spin_unlock_irq(&conf
->device_lock
);
6076 /* allow reshape to continue */
6077 wake_up(&conf
->wait_for_overlap
);
6080 case 0: /* re-enable writes */
6081 spin_lock_irq(&conf
->device_lock
);
6083 wake_up(&conf
->wait_for_stripe
);
6084 wake_up(&conf
->wait_for_overlap
);
6085 spin_unlock_irq(&conf
->device_lock
);
6091 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
6093 struct r0conf
*raid0_conf
= mddev
->private;
6096 /* for raid0 takeover only one zone is supported */
6097 if (raid0_conf
->nr_strip_zones
> 1) {
6098 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6100 return ERR_PTR(-EINVAL
);
6103 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
6104 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
6105 mddev
->dev_sectors
= sectors
;
6106 mddev
->new_level
= level
;
6107 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6108 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
6109 mddev
->raid_disks
+= 1;
6110 mddev
->delta_disks
= 1;
6111 /* make sure it will be not marked as dirty */
6112 mddev
->recovery_cp
= MaxSector
;
6114 return setup_conf(mddev
);
6118 static void *raid5_takeover_raid1(struct mddev
*mddev
)
6122 if (mddev
->raid_disks
!= 2 ||
6123 mddev
->degraded
> 1)
6124 return ERR_PTR(-EINVAL
);
6126 /* Should check if there are write-behind devices? */
6128 chunksect
= 64*2; /* 64K by default */
6130 /* The array must be an exact multiple of chunksize */
6131 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
6134 if ((chunksect
<<9) < STRIPE_SIZE
)
6135 /* array size does not allow a suitable chunk size */
6136 return ERR_PTR(-EINVAL
);
6138 mddev
->new_level
= 5;
6139 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6140 mddev
->new_chunk_sectors
= chunksect
;
6142 return setup_conf(mddev
);
6145 static void *raid5_takeover_raid6(struct mddev
*mddev
)
6149 switch (mddev
->layout
) {
6150 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6151 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
6153 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6154 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
6156 case ALGORITHM_LEFT_SYMMETRIC_6
:
6157 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6159 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6160 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
6162 case ALGORITHM_PARITY_0_6
:
6163 new_layout
= ALGORITHM_PARITY_0
;
6165 case ALGORITHM_PARITY_N
:
6166 new_layout
= ALGORITHM_PARITY_N
;
6169 return ERR_PTR(-EINVAL
);
6171 mddev
->new_level
= 5;
6172 mddev
->new_layout
= new_layout
;
6173 mddev
->delta_disks
= -1;
6174 mddev
->raid_disks
-= 1;
6175 return setup_conf(mddev
);
6179 static int raid5_check_reshape(struct mddev
*mddev
)
6181 /* For a 2-drive array, the layout and chunk size can be changed
6182 * immediately as not restriping is needed.
6183 * For larger arrays we record the new value - after validation
6184 * to be used by a reshape pass.
6186 struct r5conf
*conf
= mddev
->private;
6187 int new_chunk
= mddev
->new_chunk_sectors
;
6189 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
6191 if (new_chunk
> 0) {
6192 if (!is_power_of_2(new_chunk
))
6194 if (new_chunk
< (PAGE_SIZE
>>9))
6196 if (mddev
->array_sectors
& (new_chunk
-1))
6197 /* not factor of array size */
6201 /* They look valid */
6203 if (mddev
->raid_disks
== 2) {
6204 /* can make the change immediately */
6205 if (mddev
->new_layout
>= 0) {
6206 conf
->algorithm
= mddev
->new_layout
;
6207 mddev
->layout
= mddev
->new_layout
;
6209 if (new_chunk
> 0) {
6210 conf
->chunk_sectors
= new_chunk
;
6211 mddev
->chunk_sectors
= new_chunk
;
6213 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
6214 md_wakeup_thread(mddev
->thread
);
6216 return check_reshape(mddev
);
6219 static int raid6_check_reshape(struct mddev
*mddev
)
6221 int new_chunk
= mddev
->new_chunk_sectors
;
6223 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
6225 if (new_chunk
> 0) {
6226 if (!is_power_of_2(new_chunk
))
6228 if (new_chunk
< (PAGE_SIZE
>> 9))
6230 if (mddev
->array_sectors
& (new_chunk
-1))
6231 /* not factor of array size */
6235 /* They look valid */
6236 return check_reshape(mddev
);
6239 static void *raid5_takeover(struct mddev
*mddev
)
6241 /* raid5 can take over:
6242 * raid0 - if there is only one strip zone - make it a raid4 layout
6243 * raid1 - if there are two drives. We need to know the chunk size
6244 * raid4 - trivial - just use a raid4 layout.
6245 * raid6 - Providing it is a *_6 layout
6247 if (mddev
->level
== 0)
6248 return raid45_takeover_raid0(mddev
, 5);
6249 if (mddev
->level
== 1)
6250 return raid5_takeover_raid1(mddev
);
6251 if (mddev
->level
== 4) {
6252 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6253 mddev
->new_level
= 5;
6254 return setup_conf(mddev
);
6256 if (mddev
->level
== 6)
6257 return raid5_takeover_raid6(mddev
);
6259 return ERR_PTR(-EINVAL
);
6262 static void *raid4_takeover(struct mddev
*mddev
)
6264 /* raid4 can take over:
6265 * raid0 - if there is only one strip zone
6266 * raid5 - if layout is right
6268 if (mddev
->level
== 0)
6269 return raid45_takeover_raid0(mddev
, 4);
6270 if (mddev
->level
== 5 &&
6271 mddev
->layout
== ALGORITHM_PARITY_N
) {
6272 mddev
->new_layout
= 0;
6273 mddev
->new_level
= 4;
6274 return setup_conf(mddev
);
6276 return ERR_PTR(-EINVAL
);
6279 static struct md_personality raid5_personality
;
6281 static void *raid6_takeover(struct mddev
*mddev
)
6283 /* Currently can only take over a raid5. We map the
6284 * personality to an equivalent raid6 personality
6285 * with the Q block at the end.
6289 if (mddev
->pers
!= &raid5_personality
)
6290 return ERR_PTR(-EINVAL
);
6291 if (mddev
->degraded
> 1)
6292 return ERR_PTR(-EINVAL
);
6293 if (mddev
->raid_disks
> 253)
6294 return ERR_PTR(-EINVAL
);
6295 if (mddev
->raid_disks
< 3)
6296 return ERR_PTR(-EINVAL
);
6298 switch (mddev
->layout
) {
6299 case ALGORITHM_LEFT_ASYMMETRIC
:
6300 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
6302 case ALGORITHM_RIGHT_ASYMMETRIC
:
6303 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
6305 case ALGORITHM_LEFT_SYMMETRIC
:
6306 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
6308 case ALGORITHM_RIGHT_SYMMETRIC
:
6309 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
6311 case ALGORITHM_PARITY_0
:
6312 new_layout
= ALGORITHM_PARITY_0_6
;
6314 case ALGORITHM_PARITY_N
:
6315 new_layout
= ALGORITHM_PARITY_N
;
6318 return ERR_PTR(-EINVAL
);
6320 mddev
->new_level
= 6;
6321 mddev
->new_layout
= new_layout
;
6322 mddev
->delta_disks
= 1;
6323 mddev
->raid_disks
+= 1;
6324 return setup_conf(mddev
);
6328 static struct md_personality raid6_personality
=
6332 .owner
= THIS_MODULE
,
6333 .make_request
= make_request
,
6337 .error_handler
= error
,
6338 .hot_add_disk
= raid5_add_disk
,
6339 .hot_remove_disk
= raid5_remove_disk
,
6340 .spare_active
= raid5_spare_active
,
6341 .sync_request
= sync_request
,
6342 .resize
= raid5_resize
,
6344 .check_reshape
= raid6_check_reshape
,
6345 .start_reshape
= raid5_start_reshape
,
6346 .finish_reshape
= raid5_finish_reshape
,
6347 .quiesce
= raid5_quiesce
,
6348 .takeover
= raid6_takeover
,
6350 static struct md_personality raid5_personality
=
6354 .owner
= THIS_MODULE
,
6355 .make_request
= make_request
,
6359 .error_handler
= error
,
6360 .hot_add_disk
= raid5_add_disk
,
6361 .hot_remove_disk
= raid5_remove_disk
,
6362 .spare_active
= raid5_spare_active
,
6363 .sync_request
= sync_request
,
6364 .resize
= raid5_resize
,
6366 .check_reshape
= raid5_check_reshape
,
6367 .start_reshape
= raid5_start_reshape
,
6368 .finish_reshape
= raid5_finish_reshape
,
6369 .quiesce
= raid5_quiesce
,
6370 .takeover
= raid5_takeover
,
6373 static struct md_personality raid4_personality
=
6377 .owner
= THIS_MODULE
,
6378 .make_request
= make_request
,
6382 .error_handler
= error
,
6383 .hot_add_disk
= raid5_add_disk
,
6384 .hot_remove_disk
= raid5_remove_disk
,
6385 .spare_active
= raid5_spare_active
,
6386 .sync_request
= sync_request
,
6387 .resize
= raid5_resize
,
6389 .check_reshape
= raid5_check_reshape
,
6390 .start_reshape
= raid5_start_reshape
,
6391 .finish_reshape
= raid5_finish_reshape
,
6392 .quiesce
= raid5_quiesce
,
6393 .takeover
= raid4_takeover
,
6396 static int __init
raid5_init(void)
6398 register_md_personality(&raid6_personality
);
6399 register_md_personality(&raid5_personality
);
6400 register_md_personality(&raid4_personality
);
6404 static void raid5_exit(void)
6406 unregister_md_personality(&raid6_personality
);
6407 unregister_md_personality(&raid5_personality
);
6408 unregister_md_personality(&raid4_personality
);
6411 module_init(raid5_init
);
6412 module_exit(raid5_exit
);
6413 MODULE_LICENSE("GPL");
6414 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6415 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6416 MODULE_ALIAS("md-raid5");
6417 MODULE_ALIAS("md-raid4");
6418 MODULE_ALIAS("md-level-5");
6419 MODULE_ALIAS("md-level-4");
6420 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6421 MODULE_ALIAS("md-raid6");
6422 MODULE_ALIAS("md-level-6");
6424 /* This used to be two separate modules, they were: */
6425 MODULE_ALIAS("raid5");
6426 MODULE_ALIAS("raid6");