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 spin_lock_irq(&sh
->stripe_lock
);
2556 bi
= sh
->dev
[i
].toread
;
2557 sh
->dev
[i
].toread
= NULL
;
2558 spin_unlock_irq(&sh
->stripe_lock
);
2559 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2560 wake_up(&conf
->wait_for_overlap
);
2561 if (bi
) s
->to_read
--;
2562 while (bi
&& bi
->bi_sector
<
2563 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2564 struct bio
*nextbi
=
2565 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2566 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2567 if (!raid5_dec_bi_active_stripes(bi
)) {
2568 bi
->bi_next
= *return_bi
;
2575 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2576 STRIPE_SECTORS
, 0, 0);
2577 /* If we were in the middle of a write the parity block might
2578 * still be locked - so just clear all R5_LOCKED flags
2580 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2583 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2584 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2585 md_wakeup_thread(conf
->mddev
->thread
);
2589 handle_failed_sync(struct r5conf
*conf
, struct stripe_head
*sh
,
2590 struct stripe_head_state
*s
)
2595 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2598 /* There is nothing more to do for sync/check/repair.
2599 * Don't even need to abort as that is handled elsewhere
2600 * if needed, and not always wanted e.g. if there is a known
2602 * For recover/replace we need to record a bad block on all
2603 * non-sync devices, or abort the recovery
2605 if (test_bit(MD_RECOVERY_RECOVER
, &conf
->mddev
->recovery
)) {
2606 /* During recovery devices cannot be removed, so
2607 * locking and refcounting of rdevs is not needed
2609 for (i
= 0; i
< conf
->raid_disks
; i
++) {
2610 struct md_rdev
*rdev
= conf
->disks
[i
].rdev
;
2612 && !test_bit(Faulty
, &rdev
->flags
)
2613 && !test_bit(In_sync
, &rdev
->flags
)
2614 && !rdev_set_badblocks(rdev
, sh
->sector
,
2617 rdev
= conf
->disks
[i
].replacement
;
2619 && !test_bit(Faulty
, &rdev
->flags
)
2620 && !test_bit(In_sync
, &rdev
->flags
)
2621 && !rdev_set_badblocks(rdev
, sh
->sector
,
2626 conf
->recovery_disabled
=
2627 conf
->mddev
->recovery_disabled
;
2629 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, !abort
);
2632 static int want_replace(struct stripe_head
*sh
, int disk_idx
)
2634 struct md_rdev
*rdev
;
2636 /* Doing recovery so rcu locking not required */
2637 rdev
= sh
->raid_conf
->disks
[disk_idx
].replacement
;
2639 && !test_bit(Faulty
, &rdev
->flags
)
2640 && !test_bit(In_sync
, &rdev
->flags
)
2641 && (rdev
->recovery_offset
<= sh
->sector
2642 || rdev
->mddev
->recovery_cp
<= sh
->sector
))
2648 /* fetch_block - checks the given member device to see if its data needs
2649 * to be read or computed to satisfy a request.
2651 * Returns 1 when no more member devices need to be checked, otherwise returns
2652 * 0 to tell the loop in handle_stripe_fill to continue
2654 static int fetch_block(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2655 int disk_idx
, int disks
)
2657 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2658 struct r5dev
*fdev
[2] = { &sh
->dev
[s
->failed_num
[0]],
2659 &sh
->dev
[s
->failed_num
[1]] };
2661 /* is the data in this block needed, and can we get it? */
2662 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2663 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2665 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2666 s
->syncing
|| s
->expanding
||
2667 (s
->replacing
&& want_replace(sh
, disk_idx
)) ||
2668 (s
->failed
>= 1 && fdev
[0]->toread
) ||
2669 (s
->failed
>= 2 && fdev
[1]->toread
) ||
2670 (sh
->raid_conf
->level
<= 5 && s
->failed
&& fdev
[0]->towrite
&&
2671 !test_bit(R5_OVERWRITE
, &fdev
[0]->flags
)) ||
2672 (sh
->raid_conf
->level
== 6 && s
->failed
&& s
->to_write
))) {
2673 /* we would like to get this block, possibly by computing it,
2674 * otherwise read it if the backing disk is insync
2676 BUG_ON(test_bit(R5_Wantcompute
, &dev
->flags
));
2677 BUG_ON(test_bit(R5_Wantread
, &dev
->flags
));
2678 if ((s
->uptodate
== disks
- 1) &&
2679 (s
->failed
&& (disk_idx
== s
->failed_num
[0] ||
2680 disk_idx
== s
->failed_num
[1]))) {
2681 /* have disk failed, and we're requested to fetch it;
2684 pr_debug("Computing stripe %llu block %d\n",
2685 (unsigned long long)sh
->sector
, disk_idx
);
2686 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2687 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2688 set_bit(R5_Wantcompute
, &dev
->flags
);
2689 sh
->ops
.target
= disk_idx
;
2690 sh
->ops
.target2
= -1; /* no 2nd target */
2692 /* Careful: from this point on 'uptodate' is in the eye
2693 * of raid_run_ops which services 'compute' operations
2694 * before writes. R5_Wantcompute flags a block that will
2695 * be R5_UPTODATE by the time it is needed for a
2696 * subsequent operation.
2700 } else if (s
->uptodate
== disks
-2 && s
->failed
>= 2) {
2701 /* Computing 2-failure is *very* expensive; only
2702 * do it if failed >= 2
2705 for (other
= disks
; other
--; ) {
2706 if (other
== disk_idx
)
2708 if (!test_bit(R5_UPTODATE
,
2709 &sh
->dev
[other
].flags
))
2713 pr_debug("Computing stripe %llu blocks %d,%d\n",
2714 (unsigned long long)sh
->sector
,
2716 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2717 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2718 set_bit(R5_Wantcompute
, &sh
->dev
[disk_idx
].flags
);
2719 set_bit(R5_Wantcompute
, &sh
->dev
[other
].flags
);
2720 sh
->ops
.target
= disk_idx
;
2721 sh
->ops
.target2
= other
;
2725 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2726 set_bit(R5_LOCKED
, &dev
->flags
);
2727 set_bit(R5_Wantread
, &dev
->flags
);
2729 pr_debug("Reading block %d (sync=%d)\n",
2730 disk_idx
, s
->syncing
);
2738 * handle_stripe_fill - read or compute data to satisfy pending requests.
2740 static void handle_stripe_fill(struct stripe_head
*sh
,
2741 struct stripe_head_state
*s
,
2746 /* look for blocks to read/compute, skip this if a compute
2747 * is already in flight, or if the stripe contents are in the
2748 * midst of changing due to a write
2750 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2751 !sh
->reconstruct_state
)
2752 for (i
= disks
; i
--; )
2753 if (fetch_block(sh
, s
, i
, disks
))
2755 set_bit(STRIPE_HANDLE
, &sh
->state
);
2759 /* handle_stripe_clean_event
2760 * any written block on an uptodate or failed drive can be returned.
2761 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2762 * never LOCKED, so we don't need to test 'failed' directly.
2764 static void handle_stripe_clean_event(struct r5conf
*conf
,
2765 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2770 for (i
= disks
; i
--; )
2771 if (sh
->dev
[i
].written
) {
2773 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2774 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
2775 test_and_clear_bit(R5_Discard
, &dev
->flags
))) {
2776 /* We can return any write requests */
2777 struct bio
*wbi
, *wbi2
;
2778 pr_debug("Return write for disc %d\n", i
);
2780 dev
->written
= NULL
;
2781 while (wbi
&& wbi
->bi_sector
<
2782 dev
->sector
+ STRIPE_SECTORS
) {
2783 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2784 if (!raid5_dec_bi_active_stripes(wbi
)) {
2785 md_write_end(conf
->mddev
);
2786 wbi
->bi_next
= *return_bi
;
2791 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2793 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2798 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2799 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2800 md_wakeup_thread(conf
->mddev
->thread
);
2803 static void handle_stripe_dirtying(struct r5conf
*conf
,
2804 struct stripe_head
*sh
,
2805 struct stripe_head_state
*s
,
2808 int rmw
= 0, rcw
= 0, i
;
2809 if (conf
->max_degraded
== 2) {
2810 /* RAID6 requires 'rcw' in current implementation
2811 * Calculate the real rcw later - for now fake it
2812 * look like rcw is cheaper
2815 } else for (i
= disks
; i
--; ) {
2816 /* would I have to read this buffer for read_modify_write */
2817 struct r5dev
*dev
= &sh
->dev
[i
];
2818 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2819 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2820 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2821 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2822 if (test_bit(R5_Insync
, &dev
->flags
))
2825 rmw
+= 2*disks
; /* cannot read it */
2827 /* Would I have to read this buffer for reconstruct_write */
2828 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2829 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2830 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2831 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2832 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2837 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2838 (unsigned long long)sh
->sector
, rmw
, rcw
);
2839 set_bit(STRIPE_HANDLE
, &sh
->state
);
2840 if (rmw
< rcw
&& rmw
> 0)
2841 /* prefer read-modify-write, but need to get some data */
2842 for (i
= disks
; i
--; ) {
2843 struct r5dev
*dev
= &sh
->dev
[i
];
2844 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2845 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2846 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2847 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2848 test_bit(R5_Insync
, &dev
->flags
)) {
2850 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2851 pr_debug("Read_old block "
2852 "%d for r-m-w\n", i
);
2853 set_bit(R5_LOCKED
, &dev
->flags
);
2854 set_bit(R5_Wantread
, &dev
->flags
);
2857 set_bit(STRIPE_DELAYED
, &sh
->state
);
2858 set_bit(STRIPE_HANDLE
, &sh
->state
);
2862 if (rcw
<= rmw
&& rcw
> 0) {
2863 /* want reconstruct write, but need to get some data */
2865 for (i
= disks
; i
--; ) {
2866 struct r5dev
*dev
= &sh
->dev
[i
];
2867 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2868 i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
&&
2869 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2870 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2871 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2873 if (!test_bit(R5_Insync
, &dev
->flags
))
2874 continue; /* it's a failed drive */
2876 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2877 pr_debug("Read_old block "
2878 "%d for Reconstruct\n", i
);
2879 set_bit(R5_LOCKED
, &dev
->flags
);
2880 set_bit(R5_Wantread
, &dev
->flags
);
2883 set_bit(STRIPE_DELAYED
, &sh
->state
);
2884 set_bit(STRIPE_HANDLE
, &sh
->state
);
2889 /* now if nothing is locked, and if we have enough data,
2890 * we can start a write request
2892 /* since handle_stripe can be called at any time we need to handle the
2893 * case where a compute block operation has been submitted and then a
2894 * subsequent call wants to start a write request. raid_run_ops only
2895 * handles the case where compute block and reconstruct are requested
2896 * simultaneously. If this is not the case then new writes need to be
2897 * held off until the compute completes.
2899 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2900 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2901 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2902 schedule_reconstruction(sh
, s
, rcw
== 0, 0);
2905 static void handle_parity_checks5(struct r5conf
*conf
, struct stripe_head
*sh
,
2906 struct stripe_head_state
*s
, int disks
)
2908 struct r5dev
*dev
= NULL
;
2910 set_bit(STRIPE_HANDLE
, &sh
->state
);
2912 switch (sh
->check_state
) {
2913 case check_state_idle
:
2914 /* start a new check operation if there are no failures */
2915 if (s
->failed
== 0) {
2916 BUG_ON(s
->uptodate
!= disks
);
2917 sh
->check_state
= check_state_run
;
2918 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2919 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2923 dev
= &sh
->dev
[s
->failed_num
[0]];
2925 case check_state_compute_result
:
2926 sh
->check_state
= check_state_idle
;
2928 dev
= &sh
->dev
[sh
->pd_idx
];
2930 /* check that a write has not made the stripe insync */
2931 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2934 /* either failed parity check, or recovery is happening */
2935 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2936 BUG_ON(s
->uptodate
!= disks
);
2938 set_bit(R5_LOCKED
, &dev
->flags
);
2940 set_bit(R5_Wantwrite
, &dev
->flags
);
2942 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2943 set_bit(STRIPE_INSYNC
, &sh
->state
);
2945 case check_state_run
:
2946 break; /* we will be called again upon completion */
2947 case check_state_check_result
:
2948 sh
->check_state
= check_state_idle
;
2950 /* if a failure occurred during the check operation, leave
2951 * STRIPE_INSYNC not set and let the stripe be handled again
2956 /* handle a successful check operation, if parity is correct
2957 * we are done. Otherwise update the mismatch count and repair
2958 * parity if !MD_RECOVERY_CHECK
2960 if ((sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) == 0)
2961 /* parity is correct (on disc,
2962 * not in buffer any more)
2964 set_bit(STRIPE_INSYNC
, &sh
->state
);
2966 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2967 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2968 /* don't try to repair!! */
2969 set_bit(STRIPE_INSYNC
, &sh
->state
);
2971 sh
->check_state
= check_state_compute_run
;
2972 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2973 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2974 set_bit(R5_Wantcompute
,
2975 &sh
->dev
[sh
->pd_idx
].flags
);
2976 sh
->ops
.target
= sh
->pd_idx
;
2977 sh
->ops
.target2
= -1;
2982 case check_state_compute_run
:
2985 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2986 __func__
, sh
->check_state
,
2987 (unsigned long long) sh
->sector
);
2993 static void handle_parity_checks6(struct r5conf
*conf
, struct stripe_head
*sh
,
2994 struct stripe_head_state
*s
,
2997 int pd_idx
= sh
->pd_idx
;
2998 int qd_idx
= sh
->qd_idx
;
3001 set_bit(STRIPE_HANDLE
, &sh
->state
);
3003 BUG_ON(s
->failed
> 2);
3005 /* Want to check and possibly repair P and Q.
3006 * However there could be one 'failed' device, in which
3007 * case we can only check one of them, possibly using the
3008 * other to generate missing data
3011 switch (sh
->check_state
) {
3012 case check_state_idle
:
3013 /* start a new check operation if there are < 2 failures */
3014 if (s
->failed
== s
->q_failed
) {
3015 /* The only possible failed device holds Q, so it
3016 * makes sense to check P (If anything else were failed,
3017 * we would have used P to recreate it).
3019 sh
->check_state
= check_state_run
;
3021 if (!s
->q_failed
&& s
->failed
< 2) {
3022 /* Q is not failed, and we didn't use it to generate
3023 * anything, so it makes sense to check it
3025 if (sh
->check_state
== check_state_run
)
3026 sh
->check_state
= check_state_run_pq
;
3028 sh
->check_state
= check_state_run_q
;
3031 /* discard potentially stale zero_sum_result */
3032 sh
->ops
.zero_sum_result
= 0;
3034 if (sh
->check_state
== check_state_run
) {
3035 /* async_xor_zero_sum destroys the contents of P */
3036 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
3039 if (sh
->check_state
>= check_state_run
&&
3040 sh
->check_state
<= check_state_run_pq
) {
3041 /* async_syndrome_zero_sum preserves P and Q, so
3042 * no need to mark them !uptodate here
3044 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
3048 /* we have 2-disk failure */
3049 BUG_ON(s
->failed
!= 2);
3051 case check_state_compute_result
:
3052 sh
->check_state
= check_state_idle
;
3054 /* check that a write has not made the stripe insync */
3055 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
3058 /* now write out any block on a failed drive,
3059 * or P or Q if they were recomputed
3061 BUG_ON(s
->uptodate
< disks
- 1); /* We don't need Q to recover */
3062 if (s
->failed
== 2) {
3063 dev
= &sh
->dev
[s
->failed_num
[1]];
3065 set_bit(R5_LOCKED
, &dev
->flags
);
3066 set_bit(R5_Wantwrite
, &dev
->flags
);
3068 if (s
->failed
>= 1) {
3069 dev
= &sh
->dev
[s
->failed_num
[0]];
3071 set_bit(R5_LOCKED
, &dev
->flags
);
3072 set_bit(R5_Wantwrite
, &dev
->flags
);
3074 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3075 dev
= &sh
->dev
[pd_idx
];
3077 set_bit(R5_LOCKED
, &dev
->flags
);
3078 set_bit(R5_Wantwrite
, &dev
->flags
);
3080 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3081 dev
= &sh
->dev
[qd_idx
];
3083 set_bit(R5_LOCKED
, &dev
->flags
);
3084 set_bit(R5_Wantwrite
, &dev
->flags
);
3086 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
3088 set_bit(STRIPE_INSYNC
, &sh
->state
);
3090 case check_state_run
:
3091 case check_state_run_q
:
3092 case check_state_run_pq
:
3093 break; /* we will be called again upon completion */
3094 case check_state_check_result
:
3095 sh
->check_state
= check_state_idle
;
3097 /* handle a successful check operation, if parity is correct
3098 * we are done. Otherwise update the mismatch count and repair
3099 * parity if !MD_RECOVERY_CHECK
3101 if (sh
->ops
.zero_sum_result
== 0) {
3102 /* both parities are correct */
3104 set_bit(STRIPE_INSYNC
, &sh
->state
);
3106 /* in contrast to the raid5 case we can validate
3107 * parity, but still have a failure to write
3110 sh
->check_state
= check_state_compute_result
;
3111 /* Returning at this point means that we may go
3112 * off and bring p and/or q uptodate again so
3113 * we make sure to check zero_sum_result again
3114 * to verify if p or q need writeback
3118 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
3119 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
3120 /* don't try to repair!! */
3121 set_bit(STRIPE_INSYNC
, &sh
->state
);
3123 int *target
= &sh
->ops
.target
;
3125 sh
->ops
.target
= -1;
3126 sh
->ops
.target2
= -1;
3127 sh
->check_state
= check_state_compute_run
;
3128 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
3129 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
3130 if (sh
->ops
.zero_sum_result
& SUM_CHECK_P_RESULT
) {
3131 set_bit(R5_Wantcompute
,
3132 &sh
->dev
[pd_idx
].flags
);
3134 target
= &sh
->ops
.target2
;
3137 if (sh
->ops
.zero_sum_result
& SUM_CHECK_Q_RESULT
) {
3138 set_bit(R5_Wantcompute
,
3139 &sh
->dev
[qd_idx
].flags
);
3146 case check_state_compute_run
:
3149 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
3150 __func__
, sh
->check_state
,
3151 (unsigned long long) sh
->sector
);
3156 static void handle_stripe_expansion(struct r5conf
*conf
, struct stripe_head
*sh
)
3160 /* We have read all the blocks in this stripe and now we need to
3161 * copy some of them into a target stripe for expand.
3163 struct dma_async_tx_descriptor
*tx
= NULL
;
3164 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3165 for (i
= 0; i
< sh
->disks
; i
++)
3166 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
3168 struct stripe_head
*sh2
;
3169 struct async_submit_ctl submit
;
3171 sector_t bn
= compute_blocknr(sh
, i
, 1);
3172 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
3174 sh2
= get_active_stripe(conf
, s
, 0, 1, 1);
3176 /* so far only the early blocks of this stripe
3177 * have been requested. When later blocks
3178 * get requested, we will try again
3181 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
3182 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
3183 /* must have already done this block */
3184 release_stripe(sh2
);
3188 /* place all the copies on one channel */
3189 init_async_submit(&submit
, 0, tx
, NULL
, NULL
, NULL
);
3190 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
3191 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
3194 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
3195 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
3196 for (j
= 0; j
< conf
->raid_disks
; j
++)
3197 if (j
!= sh2
->pd_idx
&&
3199 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
3201 if (j
== conf
->raid_disks
) {
3202 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
3203 set_bit(STRIPE_HANDLE
, &sh2
->state
);
3205 release_stripe(sh2
);
3208 /* done submitting copies, wait for them to complete */
3211 dma_wait_for_async_tx(tx
);
3216 * handle_stripe - do things to a stripe.
3218 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3219 * state of various bits to see what needs to be done.
3221 * return some read requests which now have data
3222 * return some write requests which are safely on storage
3223 * schedule a read on some buffers
3224 * schedule a write of some buffers
3225 * return confirmation of parity correctness
3229 static void analyse_stripe(struct stripe_head
*sh
, struct stripe_head_state
*s
)
3231 struct r5conf
*conf
= sh
->raid_conf
;
3232 int disks
= sh
->disks
;
3235 int do_recovery
= 0;
3237 memset(s
, 0, sizeof(*s
));
3239 s
->expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3240 s
->expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3241 s
->failed_num
[0] = -1;
3242 s
->failed_num
[1] = -1;
3244 /* Now to look around and see what can be done */
3246 for (i
=disks
; i
--; ) {
3247 struct md_rdev
*rdev
;
3254 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3256 dev
->toread
, dev
->towrite
, dev
->written
);
3257 /* maybe we can reply to a read
3259 * new wantfill requests are only permitted while
3260 * ops_complete_biofill is guaranteed to be inactive
3262 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
3263 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
3264 set_bit(R5_Wantfill
, &dev
->flags
);
3266 /* now count some things */
3267 if (test_bit(R5_LOCKED
, &dev
->flags
))
3269 if (test_bit(R5_UPTODATE
, &dev
->flags
))
3271 if (test_bit(R5_Wantcompute
, &dev
->flags
)) {
3273 BUG_ON(s
->compute
> 2);
3276 if (test_bit(R5_Wantfill
, &dev
->flags
))
3278 else if (dev
->toread
)
3282 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3287 /* Prefer to use the replacement for reads, but only
3288 * if it is recovered enough and has no bad blocks.
3290 rdev
= rcu_dereference(conf
->disks
[i
].replacement
);
3291 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) &&
3292 rdev
->recovery_offset
>= sh
->sector
+ STRIPE_SECTORS
&&
3293 !is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3294 &first_bad
, &bad_sectors
))
3295 set_bit(R5_ReadRepl
, &dev
->flags
);
3298 set_bit(R5_NeedReplace
, &dev
->flags
);
3299 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3300 clear_bit(R5_ReadRepl
, &dev
->flags
);
3302 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
3305 is_bad
= is_badblock(rdev
, sh
->sector
, STRIPE_SECTORS
,
3306 &first_bad
, &bad_sectors
);
3307 if (s
->blocked_rdev
== NULL
3308 && (test_bit(Blocked
, &rdev
->flags
)
3311 set_bit(BlockedBadBlocks
,
3313 s
->blocked_rdev
= rdev
;
3314 atomic_inc(&rdev
->nr_pending
);
3317 clear_bit(R5_Insync
, &dev
->flags
);
3321 /* also not in-sync */
3322 if (!test_bit(WriteErrorSeen
, &rdev
->flags
) &&
3323 test_bit(R5_UPTODATE
, &dev
->flags
)) {
3324 /* treat as in-sync, but with a read error
3325 * which we can now try to correct
3327 set_bit(R5_Insync
, &dev
->flags
);
3328 set_bit(R5_ReadError
, &dev
->flags
);
3330 } else if (test_bit(In_sync
, &rdev
->flags
))
3331 set_bit(R5_Insync
, &dev
->flags
);
3332 else if (sh
->sector
+ STRIPE_SECTORS
<= rdev
->recovery_offset
)
3333 /* in sync if before recovery_offset */
3334 set_bit(R5_Insync
, &dev
->flags
);
3335 else if (test_bit(R5_UPTODATE
, &dev
->flags
) &&
3336 test_bit(R5_Expanded
, &dev
->flags
))
3337 /* If we've reshaped into here, we assume it is Insync.
3338 * We will shortly update recovery_offset to make
3341 set_bit(R5_Insync
, &dev
->flags
);
3343 if (rdev
&& test_bit(R5_WriteError
, &dev
->flags
)) {
3344 /* This flag does not apply to '.replacement'
3345 * only to .rdev, so make sure to check that*/
3346 struct md_rdev
*rdev2
= rcu_dereference(
3347 conf
->disks
[i
].rdev
);
3349 clear_bit(R5_Insync
, &dev
->flags
);
3350 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3351 s
->handle_bad_blocks
= 1;
3352 atomic_inc(&rdev2
->nr_pending
);
3354 clear_bit(R5_WriteError
, &dev
->flags
);
3356 if (rdev
&& test_bit(R5_MadeGood
, &dev
->flags
)) {
3357 /* This flag does not apply to '.replacement'
3358 * only to .rdev, so make sure to check that*/
3359 struct md_rdev
*rdev2
= rcu_dereference(
3360 conf
->disks
[i
].rdev
);
3361 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3362 s
->handle_bad_blocks
= 1;
3363 atomic_inc(&rdev2
->nr_pending
);
3365 clear_bit(R5_MadeGood
, &dev
->flags
);
3367 if (test_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3368 struct md_rdev
*rdev2
= rcu_dereference(
3369 conf
->disks
[i
].replacement
);
3370 if (rdev2
&& !test_bit(Faulty
, &rdev2
->flags
)) {
3371 s
->handle_bad_blocks
= 1;
3372 atomic_inc(&rdev2
->nr_pending
);
3374 clear_bit(R5_MadeGoodRepl
, &dev
->flags
);
3376 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3377 /* The ReadError flag will just be confusing now */
3378 clear_bit(R5_ReadError
, &dev
->flags
);
3379 clear_bit(R5_ReWrite
, &dev
->flags
);
3381 if (test_bit(R5_ReadError
, &dev
->flags
))
3382 clear_bit(R5_Insync
, &dev
->flags
);
3383 if (!test_bit(R5_Insync
, &dev
->flags
)) {
3385 s
->failed_num
[s
->failed
] = i
;
3387 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
))
3391 if (test_bit(STRIPE_SYNCING
, &sh
->state
)) {
3392 /* If there is a failed device being replaced,
3393 * we must be recovering.
3394 * else if we are after recovery_cp, we must be syncing
3395 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3396 * else we can only be replacing
3397 * sync and recovery both need to read all devices, and so
3398 * use the same flag.
3401 sh
->sector
>= conf
->mddev
->recovery_cp
||
3402 test_bit(MD_RECOVERY_REQUESTED
, &(conf
->mddev
->recovery
)))
3410 static void handle_stripe(struct stripe_head
*sh
)
3412 struct stripe_head_state s
;
3413 struct r5conf
*conf
= sh
->raid_conf
;
3416 int disks
= sh
->disks
;
3417 struct r5dev
*pdev
, *qdev
;
3419 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3420 if (test_and_set_bit_lock(STRIPE_ACTIVE
, &sh
->state
)) {
3421 /* already being handled, ensure it gets handled
3422 * again when current action finishes */
3423 set_bit(STRIPE_HANDLE
, &sh
->state
);
3427 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
)) {
3428 set_bit(STRIPE_SYNCING
, &sh
->state
);
3429 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3431 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3433 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3434 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3435 (unsigned long long)sh
->sector
, sh
->state
,
3436 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->qd_idx
,
3437 sh
->check_state
, sh
->reconstruct_state
);
3439 analyse_stripe(sh
, &s
);
3441 if (s
.handle_bad_blocks
) {
3442 set_bit(STRIPE_HANDLE
, &sh
->state
);
3446 if (unlikely(s
.blocked_rdev
)) {
3447 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3448 s
.replacing
|| s
.to_write
|| s
.written
) {
3449 set_bit(STRIPE_HANDLE
, &sh
->state
);
3452 /* There is nothing for the blocked_rdev to block */
3453 rdev_dec_pending(s
.blocked_rdev
, conf
->mddev
);
3454 s
.blocked_rdev
= NULL
;
3457 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
3458 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
3459 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
3462 pr_debug("locked=%d uptodate=%d to_read=%d"
3463 " to_write=%d failed=%d failed_num=%d,%d\n",
3464 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3465 s
.failed_num
[0], s
.failed_num
[1]);
3466 /* check if the array has lost more than max_degraded devices and,
3467 * if so, some requests might need to be failed.
3469 if (s
.failed
> conf
->max_degraded
) {
3470 sh
->check_state
= 0;
3471 sh
->reconstruct_state
= 0;
3472 if (s
.to_read
+s
.to_write
+s
.written
)
3473 handle_failed_stripe(conf
, sh
, &s
, disks
, &s
.return_bi
);
3474 if (s
.syncing
+ s
.replacing
)
3475 handle_failed_sync(conf
, sh
, &s
);
3479 * might be able to return some write requests if the parity blocks
3480 * are safe, or on a failed drive
3482 pdev
= &sh
->dev
[sh
->pd_idx
];
3483 s
.p_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->pd_idx
)
3484 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->pd_idx
);
3485 qdev
= &sh
->dev
[sh
->qd_idx
];
3486 s
.q_failed
= (s
.failed
>= 1 && s
.failed_num
[0] == sh
->qd_idx
)
3487 || (s
.failed
>= 2 && s
.failed_num
[1] == sh
->qd_idx
)
3491 (s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3492 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3493 && (test_bit(R5_UPTODATE
, &pdev
->flags
) ||
3494 test_bit(R5_Discard
, &pdev
->flags
))))) &&
3495 (s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3496 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3497 && (test_bit(R5_UPTODATE
, &qdev
->flags
) ||
3498 test_bit(R5_Discard
, &qdev
->flags
))))))
3499 handle_stripe_clean_event(conf
, sh
, disks
, &s
.return_bi
);
3501 /* Now we might consider reading some blocks, either to check/generate
3502 * parity, or to satisfy requests
3503 * or to load a block that is being partially written.
3505 if (s
.to_read
|| s
.non_overwrite
3506 || (conf
->level
== 6 && s
.to_write
&& s
.failed
)
3507 || (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
))
3510 handle_stripe_fill(sh
, &s
, disks
);
3512 /* Now we check to see if any write operations have recently
3516 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
3518 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
3519 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
3520 sh
->reconstruct_state
= reconstruct_state_idle
;
3522 /* All the 'written' buffers and the parity block are ready to
3523 * be written back to disk
3525 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
) &&
3526 !test_bit(R5_Discard
, &sh
->dev
[sh
->pd_idx
].flags
));
3527 BUG_ON(sh
->qd_idx
>= 0 &&
3528 !test_bit(R5_UPTODATE
, &sh
->dev
[sh
->qd_idx
].flags
) &&
3529 !test_bit(R5_Discard
, &sh
->dev
[sh
->qd_idx
].flags
));
3530 for (i
= disks
; i
--; ) {
3531 struct r5dev
*dev
= &sh
->dev
[i
];
3532 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
3533 (i
== sh
->pd_idx
|| i
== sh
->qd_idx
||
3535 pr_debug("Writing block %d\n", i
);
3536 set_bit(R5_Wantwrite
, &dev
->flags
);
3539 if (!test_bit(R5_Insync
, &dev
->flags
) ||
3540 ((i
== sh
->pd_idx
|| i
== sh
->qd_idx
) &&
3542 set_bit(STRIPE_INSYNC
, &sh
->state
);
3545 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3546 s
.dec_preread_active
= 1;
3549 /* Now to consider new write requests and what else, if anything
3550 * should be read. We do not handle new writes when:
3551 * 1/ A 'write' operation (copy+xor) is already in flight.
3552 * 2/ A 'check' operation is in flight, as it may clobber the parity
3555 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
3556 handle_stripe_dirtying(conf
, sh
, &s
, disks
);
3558 /* maybe we need to check and possibly fix the parity for this stripe
3559 * Any reads will already have been scheduled, so we just see if enough
3560 * data is available. The parity check is held off while parity
3561 * dependent operations are in flight.
3563 if (sh
->check_state
||
3564 (s
.syncing
&& s
.locked
== 0 &&
3565 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
3566 !test_bit(STRIPE_INSYNC
, &sh
->state
))) {
3567 if (conf
->level
== 6)
3568 handle_parity_checks6(conf
, sh
, &s
, disks
);
3570 handle_parity_checks5(conf
, sh
, &s
, disks
);
3573 if (s
.replacing
&& s
.locked
== 0
3574 && !test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3575 /* Write out to replacement devices where possible */
3576 for (i
= 0; i
< conf
->raid_disks
; i
++)
3577 if (test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
) &&
3578 test_bit(R5_NeedReplace
, &sh
->dev
[i
].flags
)) {
3579 set_bit(R5_WantReplace
, &sh
->dev
[i
].flags
);
3580 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3583 set_bit(STRIPE_INSYNC
, &sh
->state
);
3585 if ((s
.syncing
|| s
.replacing
) && s
.locked
== 0 &&
3586 test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3587 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3588 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3591 /* If the failed drives are just a ReadError, then we might need
3592 * to progress the repair/check process
3594 if (s
.failed
<= conf
->max_degraded
&& !conf
->mddev
->ro
)
3595 for (i
= 0; i
< s
.failed
; i
++) {
3596 struct r5dev
*dev
= &sh
->dev
[s
.failed_num
[i
]];
3597 if (test_bit(R5_ReadError
, &dev
->flags
)
3598 && !test_bit(R5_LOCKED
, &dev
->flags
)
3599 && test_bit(R5_UPTODATE
, &dev
->flags
)
3601 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3602 set_bit(R5_Wantwrite
, &dev
->flags
);
3603 set_bit(R5_ReWrite
, &dev
->flags
);
3604 set_bit(R5_LOCKED
, &dev
->flags
);
3607 /* let's read it back */
3608 set_bit(R5_Wantread
, &dev
->flags
);
3609 set_bit(R5_LOCKED
, &dev
->flags
);
3616 /* Finish reconstruct operations initiated by the expansion process */
3617 if (sh
->reconstruct_state
== reconstruct_state_result
) {
3618 struct stripe_head
*sh_src
3619 = get_active_stripe(conf
, sh
->sector
, 1, 1, 1);
3620 if (sh_src
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh_src
->state
)) {
3621 /* sh cannot be written until sh_src has been read.
3622 * so arrange for sh to be delayed a little
3624 set_bit(STRIPE_DELAYED
, &sh
->state
);
3625 set_bit(STRIPE_HANDLE
, &sh
->state
);
3626 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3628 atomic_inc(&conf
->preread_active_stripes
);
3629 release_stripe(sh_src
);
3633 release_stripe(sh_src
);
3635 sh
->reconstruct_state
= reconstruct_state_idle
;
3636 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3637 for (i
= conf
->raid_disks
; i
--; ) {
3638 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3639 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3644 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
3645 !sh
->reconstruct_state
) {
3646 /* Need to write out all blocks after computing parity */
3647 sh
->disks
= conf
->raid_disks
;
3648 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3649 schedule_reconstruction(sh
, &s
, 1, 1);
3650 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
3651 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3652 atomic_dec(&conf
->reshape_stripes
);
3653 wake_up(&conf
->wait_for_overlap
);
3654 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3657 if (s
.expanding
&& s
.locked
== 0 &&
3658 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3659 handle_stripe_expansion(conf
, sh
);
3662 /* wait for this device to become unblocked */
3663 if (unlikely(s
.blocked_rdev
)) {
3664 if (conf
->mddev
->external
)
3665 md_wait_for_blocked_rdev(s
.blocked_rdev
,
3668 /* Internal metadata will immediately
3669 * be written by raid5d, so we don't
3670 * need to wait here.
3672 rdev_dec_pending(s
.blocked_rdev
,
3676 if (s
.handle_bad_blocks
)
3677 for (i
= disks
; i
--; ) {
3678 struct md_rdev
*rdev
;
3679 struct r5dev
*dev
= &sh
->dev
[i
];
3680 if (test_and_clear_bit(R5_WriteError
, &dev
->flags
)) {
3681 /* We own a safe reference to the rdev */
3682 rdev
= conf
->disks
[i
].rdev
;
3683 if (!rdev_set_badblocks(rdev
, sh
->sector
,
3685 md_error(conf
->mddev
, rdev
);
3686 rdev_dec_pending(rdev
, conf
->mddev
);
3688 if (test_and_clear_bit(R5_MadeGood
, &dev
->flags
)) {
3689 rdev
= conf
->disks
[i
].rdev
;
3690 rdev_clear_badblocks(rdev
, sh
->sector
,
3692 rdev_dec_pending(rdev
, conf
->mddev
);
3694 if (test_and_clear_bit(R5_MadeGoodRepl
, &dev
->flags
)) {
3695 rdev
= conf
->disks
[i
].replacement
;
3697 /* rdev have been moved down */
3698 rdev
= conf
->disks
[i
].rdev
;
3699 rdev_clear_badblocks(rdev
, sh
->sector
,
3701 rdev_dec_pending(rdev
, conf
->mddev
);
3706 raid_run_ops(sh
, s
.ops_request
);
3710 if (s
.dec_preread_active
) {
3711 /* We delay this until after ops_run_io so that if make_request
3712 * is waiting on a flush, it won't continue until the writes
3713 * have actually been submitted.
3715 atomic_dec(&conf
->preread_active_stripes
);
3716 if (atomic_read(&conf
->preread_active_stripes
) <
3718 md_wakeup_thread(conf
->mddev
->thread
);
3721 return_io(s
.return_bi
);
3723 clear_bit_unlock(STRIPE_ACTIVE
, &sh
->state
);
3726 static void raid5_activate_delayed(struct r5conf
*conf
)
3728 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3729 while (!list_empty(&conf
->delayed_list
)) {
3730 struct list_head
*l
= conf
->delayed_list
.next
;
3731 struct stripe_head
*sh
;
3732 sh
= list_entry(l
, struct stripe_head
, lru
);
3734 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3735 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3736 atomic_inc(&conf
->preread_active_stripes
);
3737 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3742 static void activate_bit_delay(struct r5conf
*conf
)
3744 /* device_lock is held */
3745 struct list_head head
;
3746 list_add(&head
, &conf
->bitmap_list
);
3747 list_del_init(&conf
->bitmap_list
);
3748 while (!list_empty(&head
)) {
3749 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3750 list_del_init(&sh
->lru
);
3751 atomic_inc(&sh
->count
);
3752 __release_stripe(conf
, sh
);
3756 int md_raid5_congested(struct mddev
*mddev
, int bits
)
3758 struct r5conf
*conf
= mddev
->private;
3760 /* No difference between reads and writes. Just check
3761 * how busy the stripe_cache is
3764 if (conf
->inactive_blocked
)
3768 if (list_empty_careful(&conf
->inactive_list
))
3773 EXPORT_SYMBOL_GPL(md_raid5_congested
);
3775 static int raid5_congested(void *data
, int bits
)
3777 struct mddev
*mddev
= data
;
3779 return mddev_congested(mddev
, bits
) ||
3780 md_raid5_congested(mddev
, bits
);
3783 /* We want read requests to align with chunks where possible,
3784 * but write requests don't need to.
3786 static int raid5_mergeable_bvec(struct request_queue
*q
,
3787 struct bvec_merge_data
*bvm
,
3788 struct bio_vec
*biovec
)
3790 struct mddev
*mddev
= q
->queuedata
;
3791 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3793 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3794 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3796 if ((bvm
->bi_rw
& 1) == WRITE
)
3797 return biovec
->bv_len
; /* always allow writes to be mergeable */
3799 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3800 chunk_sectors
= mddev
->new_chunk_sectors
;
3801 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3802 if (max
< 0) max
= 0;
3803 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3804 return biovec
->bv_len
;
3810 static int in_chunk_boundary(struct mddev
*mddev
, struct bio
*bio
)
3812 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3813 unsigned int chunk_sectors
= mddev
->chunk_sectors
;
3814 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3816 if (mddev
->new_chunk_sectors
< mddev
->chunk_sectors
)
3817 chunk_sectors
= mddev
->new_chunk_sectors
;
3818 return chunk_sectors
>=
3819 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3823 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3824 * later sampled by raid5d.
3826 static void add_bio_to_retry(struct bio
*bi
,struct r5conf
*conf
)
3828 unsigned long flags
;
3830 spin_lock_irqsave(&conf
->device_lock
, flags
);
3832 bi
->bi_next
= conf
->retry_read_aligned_list
;
3833 conf
->retry_read_aligned_list
= bi
;
3835 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3836 md_wakeup_thread(conf
->mddev
->thread
);
3840 static struct bio
*remove_bio_from_retry(struct r5conf
*conf
)
3844 bi
= conf
->retry_read_aligned
;
3846 conf
->retry_read_aligned
= NULL
;
3849 bi
= conf
->retry_read_aligned_list
;
3851 conf
->retry_read_aligned_list
= bi
->bi_next
;
3854 * this sets the active strip count to 1 and the processed
3855 * strip count to zero (upper 8 bits)
3857 raid5_set_bi_stripes(bi
, 1); /* biased count of active stripes */
3865 * The "raid5_align_endio" should check if the read succeeded and if it
3866 * did, call bio_endio on the original bio (having bio_put the new bio
3868 * If the read failed..
3870 static void raid5_align_endio(struct bio
*bi
, int error
)
3872 struct bio
* raid_bi
= bi
->bi_private
;
3873 struct mddev
*mddev
;
3874 struct r5conf
*conf
;
3875 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3876 struct md_rdev
*rdev
;
3880 rdev
= (void*)raid_bi
->bi_next
;
3881 raid_bi
->bi_next
= NULL
;
3882 mddev
= rdev
->mddev
;
3883 conf
= mddev
->private;
3885 rdev_dec_pending(rdev
, conf
->mddev
);
3887 if (!error
&& uptodate
) {
3888 bio_endio(raid_bi
, 0);
3889 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3890 wake_up(&conf
->wait_for_stripe
);
3895 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3897 add_bio_to_retry(raid_bi
, conf
);
3900 static int bio_fits_rdev(struct bio
*bi
)
3902 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3904 if ((bi
->bi_size
>>9) > queue_max_sectors(q
))
3906 blk_recount_segments(q
, bi
);
3907 if (bi
->bi_phys_segments
> queue_max_segments(q
))
3910 if (q
->merge_bvec_fn
)
3911 /* it's too hard to apply the merge_bvec_fn at this stage,
3920 static int chunk_aligned_read(struct mddev
*mddev
, struct bio
* raid_bio
)
3922 struct r5conf
*conf
= mddev
->private;
3924 struct bio
* align_bi
;
3925 struct md_rdev
*rdev
;
3926 sector_t end_sector
;
3928 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3929 pr_debug("chunk_aligned_read : non aligned\n");
3933 * use bio_clone_mddev to make a copy of the bio
3935 align_bi
= bio_clone_mddev(raid_bio
, GFP_NOIO
, mddev
);
3939 * set bi_end_io to a new function, and set bi_private to the
3942 align_bi
->bi_end_io
= raid5_align_endio
;
3943 align_bi
->bi_private
= raid_bio
;
3947 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3951 end_sector
= align_bi
->bi_sector
+ (align_bi
->bi_size
>> 9);
3953 rdev
= rcu_dereference(conf
->disks
[dd_idx
].replacement
);
3954 if (!rdev
|| test_bit(Faulty
, &rdev
->flags
) ||
3955 rdev
->recovery_offset
< end_sector
) {
3956 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3958 (test_bit(Faulty
, &rdev
->flags
) ||
3959 !(test_bit(In_sync
, &rdev
->flags
) ||
3960 rdev
->recovery_offset
>= end_sector
)))
3967 atomic_inc(&rdev
->nr_pending
);
3969 raid_bio
->bi_next
= (void*)rdev
;
3970 align_bi
->bi_bdev
= rdev
->bdev
;
3971 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3973 if (!bio_fits_rdev(align_bi
) ||
3974 is_badblock(rdev
, align_bi
->bi_sector
, align_bi
->bi_size
>>9,
3975 &first_bad
, &bad_sectors
)) {
3976 /* too big in some way, or has a known bad block */
3978 rdev_dec_pending(rdev
, mddev
);
3982 /* No reshape active, so we can trust rdev->data_offset */
3983 align_bi
->bi_sector
+= rdev
->data_offset
;
3985 spin_lock_irq(&conf
->device_lock
);
3986 wait_event_lock_irq(conf
->wait_for_stripe
,
3988 conf
->device_lock
, /* nothing */);
3989 atomic_inc(&conf
->active_aligned_reads
);
3990 spin_unlock_irq(&conf
->device_lock
);
3992 generic_make_request(align_bi
);
4001 /* __get_priority_stripe - get the next stripe to process
4003 * Full stripe writes are allowed to pass preread active stripes up until
4004 * the bypass_threshold is exceeded. In general the bypass_count
4005 * increments when the handle_list is handled before the hold_list; however, it
4006 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4007 * stripe with in flight i/o. The bypass_count will be reset when the
4008 * head of the hold_list has changed, i.e. the head was promoted to the
4011 static struct stripe_head
*__get_priority_stripe(struct r5conf
*conf
)
4013 struct stripe_head
*sh
;
4015 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4017 list_empty(&conf
->handle_list
) ? "empty" : "busy",
4018 list_empty(&conf
->hold_list
) ? "empty" : "busy",
4019 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
4021 if (!list_empty(&conf
->handle_list
)) {
4022 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
4024 if (list_empty(&conf
->hold_list
))
4025 conf
->bypass_count
= 0;
4026 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
4027 if (conf
->hold_list
.next
== conf
->last_hold
)
4028 conf
->bypass_count
++;
4030 conf
->last_hold
= conf
->hold_list
.next
;
4031 conf
->bypass_count
-= conf
->bypass_threshold
;
4032 if (conf
->bypass_count
< 0)
4033 conf
->bypass_count
= 0;
4036 } else if (!list_empty(&conf
->hold_list
) &&
4037 ((conf
->bypass_threshold
&&
4038 conf
->bypass_count
> conf
->bypass_threshold
) ||
4039 atomic_read(&conf
->pending_full_writes
) == 0)) {
4040 sh
= list_entry(conf
->hold_list
.next
,
4042 conf
->bypass_count
-= conf
->bypass_threshold
;
4043 if (conf
->bypass_count
< 0)
4044 conf
->bypass_count
= 0;
4048 list_del_init(&sh
->lru
);
4049 atomic_inc(&sh
->count
);
4050 BUG_ON(atomic_read(&sh
->count
) != 1);
4054 struct raid5_plug_cb
{
4055 struct blk_plug_cb cb
;
4056 struct list_head list
;
4059 static void raid5_unplug(struct blk_plug_cb
*blk_cb
, bool from_schedule
)
4061 struct raid5_plug_cb
*cb
= container_of(
4062 blk_cb
, struct raid5_plug_cb
, cb
);
4063 struct stripe_head
*sh
;
4064 struct mddev
*mddev
= cb
->cb
.data
;
4065 struct r5conf
*conf
= mddev
->private;
4067 if (cb
->list
.next
&& !list_empty(&cb
->list
)) {
4068 spin_lock_irq(&conf
->device_lock
);
4069 while (!list_empty(&cb
->list
)) {
4070 sh
= list_first_entry(&cb
->list
, struct stripe_head
, lru
);
4071 list_del_init(&sh
->lru
);
4073 * avoid race release_stripe_plug() sees
4074 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4075 * is still in our list
4077 smp_mb__before_clear_bit();
4078 clear_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
);
4079 __release_stripe(conf
, sh
);
4081 spin_unlock_irq(&conf
->device_lock
);
4086 static void release_stripe_plug(struct mddev
*mddev
,
4087 struct stripe_head
*sh
)
4089 struct blk_plug_cb
*blk_cb
= blk_check_plugged(
4090 raid5_unplug
, mddev
,
4091 sizeof(struct raid5_plug_cb
));
4092 struct raid5_plug_cb
*cb
;
4099 cb
= container_of(blk_cb
, struct raid5_plug_cb
, cb
);
4101 if (cb
->list
.next
== NULL
)
4102 INIT_LIST_HEAD(&cb
->list
);
4104 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST
, &sh
->state
))
4105 list_add_tail(&sh
->lru
, &cb
->list
);
4110 static void make_discard_request(struct mddev
*mddev
, struct bio
*bi
)
4112 struct r5conf
*conf
= mddev
->private;
4113 sector_t logical_sector
, last_sector
;
4114 struct stripe_head
*sh
;
4118 if (mddev
->reshape_position
!= MaxSector
)
4119 /* Skip discard while reshape is happening */
4122 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4123 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4126 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4128 stripe_sectors
= conf
->chunk_sectors
*
4129 (conf
->raid_disks
- conf
->max_degraded
);
4130 logical_sector
= DIV_ROUND_UP_SECTOR_T(logical_sector
,
4132 sector_div(last_sector
, stripe_sectors
);
4134 logical_sector
*= conf
->chunk_sectors
;
4135 last_sector
*= conf
->chunk_sectors
;
4137 for (; logical_sector
< last_sector
;
4138 logical_sector
+= STRIPE_SECTORS
) {
4142 sh
= get_active_stripe(conf
, logical_sector
, 0, 0, 0);
4143 prepare_to_wait(&conf
->wait_for_overlap
, &w
,
4144 TASK_UNINTERRUPTIBLE
);
4145 spin_lock_irq(&sh
->stripe_lock
);
4146 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4147 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4149 if (sh
->dev
[d
].towrite
|| sh
->dev
[d
].toread
) {
4150 set_bit(R5_Overlap
, &sh
->dev
[d
].flags
);
4151 spin_unlock_irq(&sh
->stripe_lock
);
4157 finish_wait(&conf
->wait_for_overlap
, &w
);
4158 for (d
= 0; d
< conf
->raid_disks
; d
++) {
4159 if (d
== sh
->pd_idx
|| d
== sh
->qd_idx
)
4161 sh
->dev
[d
].towrite
= bi
;
4162 set_bit(R5_OVERWRITE
, &sh
->dev
[d
].flags
);
4163 raid5_inc_bi_active_stripes(bi
);
4165 spin_unlock_irq(&sh
->stripe_lock
);
4166 if (conf
->mddev
->bitmap
) {
4168 d
< conf
->raid_disks
- conf
->max_degraded
;
4170 bitmap_startwrite(mddev
->bitmap
,
4174 sh
->bm_seq
= conf
->seq_flush
+ 1;
4175 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
4178 set_bit(STRIPE_HANDLE
, &sh
->state
);
4179 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4180 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4181 atomic_inc(&conf
->preread_active_stripes
);
4182 release_stripe_plug(mddev
, sh
);
4185 remaining
= raid5_dec_bi_active_stripes(bi
);
4186 if (remaining
== 0) {
4187 md_write_end(mddev
);
4192 static void make_request(struct mddev
*mddev
, struct bio
* bi
)
4194 struct r5conf
*conf
= mddev
->private;
4196 sector_t new_sector
;
4197 sector_t logical_sector
, last_sector
;
4198 struct stripe_head
*sh
;
4199 const int rw
= bio_data_dir(bi
);
4202 if (unlikely(bi
->bi_rw
& REQ_FLUSH
)) {
4203 md_flush_request(mddev
, bi
);
4207 md_write_start(mddev
, bi
);
4210 mddev
->reshape_position
== MaxSector
&&
4211 chunk_aligned_read(mddev
,bi
))
4214 if (unlikely(bi
->bi_rw
& REQ_DISCARD
)) {
4215 make_discard_request(mddev
, bi
);
4219 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4220 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
4222 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
4224 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
4230 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
4231 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
4232 /* spinlock is needed as reshape_progress may be
4233 * 64bit on a 32bit platform, and so it might be
4234 * possible to see a half-updated value
4235 * Of course reshape_progress could change after
4236 * the lock is dropped, so once we get a reference
4237 * to the stripe that we think it is, we will have
4240 spin_lock_irq(&conf
->device_lock
);
4241 if (mddev
->reshape_backwards
4242 ? logical_sector
< conf
->reshape_progress
4243 : logical_sector
>= conf
->reshape_progress
) {
4246 if (mddev
->reshape_backwards
4247 ? logical_sector
< conf
->reshape_safe
4248 : logical_sector
>= conf
->reshape_safe
) {
4249 spin_unlock_irq(&conf
->device_lock
);
4254 spin_unlock_irq(&conf
->device_lock
);
4257 new_sector
= raid5_compute_sector(conf
, logical_sector
,
4260 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4261 (unsigned long long)new_sector
,
4262 (unsigned long long)logical_sector
);
4264 sh
= get_active_stripe(conf
, new_sector
, previous
,
4265 (bi
->bi_rw
&RWA_MASK
), 0);
4267 if (unlikely(previous
)) {
4268 /* expansion might have moved on while waiting for a
4269 * stripe, so we must do the range check again.
4270 * Expansion could still move past after this
4271 * test, but as we are holding a reference to
4272 * 'sh', we know that if that happens,
4273 * STRIPE_EXPANDING will get set and the expansion
4274 * won't proceed until we finish with the stripe.
4277 spin_lock_irq(&conf
->device_lock
);
4278 if (mddev
->reshape_backwards
4279 ? logical_sector
>= conf
->reshape_progress
4280 : logical_sector
< conf
->reshape_progress
)
4281 /* mismatch, need to try again */
4283 spin_unlock_irq(&conf
->device_lock
);
4292 logical_sector
>= mddev
->suspend_lo
&&
4293 logical_sector
< mddev
->suspend_hi
) {
4295 /* As the suspend_* range is controlled by
4296 * userspace, we want an interruptible
4299 flush_signals(current
);
4300 prepare_to_wait(&conf
->wait_for_overlap
,
4301 &w
, TASK_INTERRUPTIBLE
);
4302 if (logical_sector
>= mddev
->suspend_lo
&&
4303 logical_sector
< mddev
->suspend_hi
)
4308 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
4309 !add_stripe_bio(sh
, bi
, dd_idx
, rw
)) {
4310 /* Stripe is busy expanding or
4311 * add failed due to overlap. Flush everything
4314 md_wakeup_thread(mddev
->thread
);
4319 finish_wait(&conf
->wait_for_overlap
, &w
);
4320 set_bit(STRIPE_HANDLE
, &sh
->state
);
4321 clear_bit(STRIPE_DELAYED
, &sh
->state
);
4322 if ((bi
->bi_rw
& REQ_NOIDLE
) &&
4323 !test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
4324 atomic_inc(&conf
->preread_active_stripes
);
4325 release_stripe_plug(mddev
, sh
);
4327 /* cannot get stripe for read-ahead, just give-up */
4328 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
4329 finish_wait(&conf
->wait_for_overlap
, &w
);
4334 remaining
= raid5_dec_bi_active_stripes(bi
);
4335 if (remaining
== 0) {
4338 md_write_end(mddev
);
4344 static sector_t
raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
);
4346 static sector_t
reshape_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
)
4348 /* reshaping is quite different to recovery/resync so it is
4349 * handled quite separately ... here.
4351 * On each call to sync_request, we gather one chunk worth of
4352 * destination stripes and flag them as expanding.
4353 * Then we find all the source stripes and request reads.
4354 * As the reads complete, handle_stripe will copy the data
4355 * into the destination stripe and release that stripe.
4357 struct r5conf
*conf
= mddev
->private;
4358 struct stripe_head
*sh
;
4359 sector_t first_sector
, last_sector
;
4360 int raid_disks
= conf
->previous_raid_disks
;
4361 int data_disks
= raid_disks
- conf
->max_degraded
;
4362 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4365 sector_t writepos
, readpos
, safepos
;
4366 sector_t stripe_addr
;
4367 int reshape_sectors
;
4368 struct list_head stripes
;
4370 if (sector_nr
== 0) {
4371 /* If restarting in the middle, skip the initial sectors */
4372 if (mddev
->reshape_backwards
&&
4373 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
4374 sector_nr
= raid5_size(mddev
, 0, 0)
4375 - conf
->reshape_progress
;
4376 } else if (!mddev
->reshape_backwards
&&
4377 conf
->reshape_progress
> 0)
4378 sector_nr
= conf
->reshape_progress
;
4379 sector_div(sector_nr
, new_data_disks
);
4381 mddev
->curr_resync_completed
= sector_nr
;
4382 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4388 /* We need to process a full chunk at a time.
4389 * If old and new chunk sizes differ, we need to process the
4392 if (mddev
->new_chunk_sectors
> mddev
->chunk_sectors
)
4393 reshape_sectors
= mddev
->new_chunk_sectors
;
4395 reshape_sectors
= mddev
->chunk_sectors
;
4397 /* We update the metadata at least every 10 seconds, or when
4398 * the data about to be copied would over-write the source of
4399 * the data at the front of the range. i.e. one new_stripe
4400 * along from reshape_progress new_maps to after where
4401 * reshape_safe old_maps to
4403 writepos
= conf
->reshape_progress
;
4404 sector_div(writepos
, new_data_disks
);
4405 readpos
= conf
->reshape_progress
;
4406 sector_div(readpos
, data_disks
);
4407 safepos
= conf
->reshape_safe
;
4408 sector_div(safepos
, data_disks
);
4409 if (mddev
->reshape_backwards
) {
4410 writepos
-= min_t(sector_t
, reshape_sectors
, writepos
);
4411 readpos
+= reshape_sectors
;
4412 safepos
+= reshape_sectors
;
4414 writepos
+= reshape_sectors
;
4415 readpos
-= min_t(sector_t
, reshape_sectors
, readpos
);
4416 safepos
-= min_t(sector_t
, reshape_sectors
, safepos
);
4419 /* Having calculated the 'writepos' possibly use it
4420 * to set 'stripe_addr' which is where we will write to.
4422 if (mddev
->reshape_backwards
) {
4423 BUG_ON(conf
->reshape_progress
== 0);
4424 stripe_addr
= writepos
;
4425 BUG_ON((mddev
->dev_sectors
&
4426 ~((sector_t
)reshape_sectors
- 1))
4427 - reshape_sectors
- stripe_addr
4430 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
4431 stripe_addr
= sector_nr
;
4434 /* 'writepos' is the most advanced device address we might write.
4435 * 'readpos' is the least advanced device address we might read.
4436 * 'safepos' is the least address recorded in the metadata as having
4438 * If there is a min_offset_diff, these are adjusted either by
4439 * increasing the safepos/readpos if diff is negative, or
4440 * increasing writepos if diff is positive.
4441 * If 'readpos' is then behind 'writepos', there is no way that we can
4442 * ensure safety in the face of a crash - that must be done by userspace
4443 * making a backup of the data. So in that case there is no particular
4444 * rush to update metadata.
4445 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4446 * update the metadata to advance 'safepos' to match 'readpos' so that
4447 * we can be safe in the event of a crash.
4448 * So we insist on updating metadata if safepos is behind writepos and
4449 * readpos is beyond writepos.
4450 * In any case, update the metadata every 10 seconds.
4451 * Maybe that number should be configurable, but I'm not sure it is
4452 * worth it.... maybe it could be a multiple of safemode_delay???
4454 if (conf
->min_offset_diff
< 0) {
4455 safepos
+= -conf
->min_offset_diff
;
4456 readpos
+= -conf
->min_offset_diff
;
4458 writepos
+= conf
->min_offset_diff
;
4460 if ((mddev
->reshape_backwards
4461 ? (safepos
> writepos
&& readpos
< writepos
)
4462 : (safepos
< writepos
&& readpos
> writepos
)) ||
4463 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
4464 /* Cannot proceed until we've updated the superblock... */
4465 wait_event(conf
->wait_for_overlap
,
4466 atomic_read(&conf
->reshape_stripes
)==0);
4467 mddev
->reshape_position
= conf
->reshape_progress
;
4468 mddev
->curr_resync_completed
= sector_nr
;
4469 conf
->reshape_checkpoint
= jiffies
;
4470 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4471 md_wakeup_thread(mddev
->thread
);
4472 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
4473 kthread_should_stop());
4474 spin_lock_irq(&conf
->device_lock
);
4475 conf
->reshape_safe
= mddev
->reshape_position
;
4476 spin_unlock_irq(&conf
->device_lock
);
4477 wake_up(&conf
->wait_for_overlap
);
4478 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4481 INIT_LIST_HEAD(&stripes
);
4482 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
4484 int skipped_disk
= 0;
4485 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0, 1);
4486 set_bit(STRIPE_EXPANDING
, &sh
->state
);
4487 atomic_inc(&conf
->reshape_stripes
);
4488 /* If any of this stripe is beyond the end of the old
4489 * array, then we need to zero those blocks
4491 for (j
=sh
->disks
; j
--;) {
4493 if (j
== sh
->pd_idx
)
4495 if (conf
->level
== 6 &&
4498 s
= compute_blocknr(sh
, j
, 0);
4499 if (s
< raid5_size(mddev
, 0, 0)) {
4503 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
4504 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
4505 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
4507 if (!skipped_disk
) {
4508 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
4509 set_bit(STRIPE_HANDLE
, &sh
->state
);
4511 list_add(&sh
->lru
, &stripes
);
4513 spin_lock_irq(&conf
->device_lock
);
4514 if (mddev
->reshape_backwards
)
4515 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
4517 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
4518 spin_unlock_irq(&conf
->device_lock
);
4519 /* Ok, those stripe are ready. We can start scheduling
4520 * reads on the source stripes.
4521 * The source stripes are determined by mapping the first and last
4522 * block on the destination stripes.
4525 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
4528 raid5_compute_sector(conf
, ((stripe_addr
+reshape_sectors
)
4529 * new_data_disks
- 1),
4531 if (last_sector
>= mddev
->dev_sectors
)
4532 last_sector
= mddev
->dev_sectors
- 1;
4533 while (first_sector
<= last_sector
) {
4534 sh
= get_active_stripe(conf
, first_sector
, 1, 0, 1);
4535 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
4536 set_bit(STRIPE_HANDLE
, &sh
->state
);
4538 first_sector
+= STRIPE_SECTORS
;
4540 /* Now that the sources are clearly marked, we can release
4541 * the destination stripes
4543 while (!list_empty(&stripes
)) {
4544 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
4545 list_del_init(&sh
->lru
);
4548 /* If this takes us to the resync_max point where we have to pause,
4549 * then we need to write out the superblock.
4551 sector_nr
+= reshape_sectors
;
4552 if ((sector_nr
- mddev
->curr_resync_completed
) * 2
4553 >= mddev
->resync_max
- mddev
->curr_resync_completed
) {
4554 /* Cannot proceed until we've updated the superblock... */
4555 wait_event(conf
->wait_for_overlap
,
4556 atomic_read(&conf
->reshape_stripes
) == 0);
4557 mddev
->reshape_position
= conf
->reshape_progress
;
4558 mddev
->curr_resync_completed
= sector_nr
;
4559 conf
->reshape_checkpoint
= jiffies
;
4560 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4561 md_wakeup_thread(mddev
->thread
);
4562 wait_event(mddev
->sb_wait
,
4563 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
4564 || kthread_should_stop());
4565 spin_lock_irq(&conf
->device_lock
);
4566 conf
->reshape_safe
= mddev
->reshape_position
;
4567 spin_unlock_irq(&conf
->device_lock
);
4568 wake_up(&conf
->wait_for_overlap
);
4569 sysfs_notify(&mddev
->kobj
, NULL
, "sync_completed");
4571 return reshape_sectors
;
4574 /* FIXME go_faster isn't used */
4575 static inline sector_t
sync_request(struct mddev
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
4577 struct r5conf
*conf
= mddev
->private;
4578 struct stripe_head
*sh
;
4579 sector_t max_sector
= mddev
->dev_sectors
;
4580 sector_t sync_blocks
;
4581 int still_degraded
= 0;
4584 if (sector_nr
>= max_sector
) {
4585 /* just being told to finish up .. nothing much to do */
4587 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
4592 if (mddev
->curr_resync
< max_sector
) /* aborted */
4593 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
4595 else /* completed sync */
4597 bitmap_close_sync(mddev
->bitmap
);
4602 /* Allow raid5_quiesce to complete */
4603 wait_event(conf
->wait_for_overlap
, conf
->quiesce
!= 2);
4605 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
4606 return reshape_request(mddev
, sector_nr
, skipped
);
4608 /* No need to check resync_max as we never do more than one
4609 * stripe, and as resync_max will always be on a chunk boundary,
4610 * if the check in md_do_sync didn't fire, there is no chance
4611 * of overstepping resync_max here
4614 /* if there is too many failed drives and we are trying
4615 * to resync, then assert that we are finished, because there is
4616 * nothing we can do.
4618 if (mddev
->degraded
>= conf
->max_degraded
&&
4619 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4620 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4624 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4625 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4626 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4627 /* we can skip this block, and probably more */
4628 sync_blocks
/= STRIPE_SECTORS
;
4630 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4633 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4635 sh
= get_active_stripe(conf
, sector_nr
, 0, 1, 0);
4637 sh
= get_active_stripe(conf
, sector_nr
, 0, 0, 0);
4638 /* make sure we don't swamp the stripe cache if someone else
4639 * is trying to get access
4641 schedule_timeout_uninterruptible(1);
4643 /* Need to check if array will still be degraded after recovery/resync
4644 * We don't need to check the 'failed' flag as when that gets set,
4647 for (i
= 0; i
< conf
->raid_disks
; i
++)
4648 if (conf
->disks
[i
].rdev
== NULL
)
4651 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4653 set_bit(STRIPE_SYNC_REQUESTED
, &sh
->state
);
4658 return STRIPE_SECTORS
;
4661 static int retry_aligned_read(struct r5conf
*conf
, struct bio
*raid_bio
)
4663 /* We may not be able to submit a whole bio at once as there
4664 * may not be enough stripe_heads available.
4665 * We cannot pre-allocate enough stripe_heads as we may need
4666 * more than exist in the cache (if we allow ever large chunks).
4667 * So we do one stripe head at a time and record in
4668 * ->bi_hw_segments how many have been done.
4670 * We *know* that this entire raid_bio is in one chunk, so
4671 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4673 struct stripe_head
*sh
;
4675 sector_t sector
, logical_sector
, last_sector
;
4680 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4681 sector
= raid5_compute_sector(conf
, logical_sector
,
4683 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4685 for (; logical_sector
< last_sector
;
4686 logical_sector
+= STRIPE_SECTORS
,
4687 sector
+= STRIPE_SECTORS
,
4690 if (scnt
< raid5_bi_processed_stripes(raid_bio
))
4691 /* already done this stripe */
4694 sh
= get_active_stripe(conf
, sector
, 0, 1, 0);
4697 /* failed to get a stripe - must wait */
4698 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4699 conf
->retry_read_aligned
= raid_bio
;
4703 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4705 raid5_set_bi_processed_stripes(raid_bio
, scnt
);
4706 conf
->retry_read_aligned
= raid_bio
;
4710 set_bit(R5_ReadNoMerge
, &sh
->dev
[dd_idx
].flags
);
4715 remaining
= raid5_dec_bi_active_stripes(raid_bio
);
4717 bio_endio(raid_bio
, 0);
4718 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4719 wake_up(&conf
->wait_for_stripe
);
4723 #define MAX_STRIPE_BATCH 8
4724 static int handle_active_stripes(struct r5conf
*conf
)
4726 struct stripe_head
*batch
[MAX_STRIPE_BATCH
], *sh
;
4727 int i
, batch_size
= 0;
4729 while (batch_size
< MAX_STRIPE_BATCH
&&
4730 (sh
= __get_priority_stripe(conf
)) != NULL
)
4731 batch
[batch_size
++] = sh
;
4733 if (batch_size
== 0)
4735 spin_unlock_irq(&conf
->device_lock
);
4737 for (i
= 0; i
< batch_size
; i
++)
4738 handle_stripe(batch
[i
]);
4742 spin_lock_irq(&conf
->device_lock
);
4743 for (i
= 0; i
< batch_size
; i
++)
4744 __release_stripe(conf
, batch
[i
]);
4749 * This is our raid5 kernel thread.
4751 * We scan the hash table for stripes which can be handled now.
4752 * During the scan, completed stripes are saved for us by the interrupt
4753 * handler, so that they will not have to wait for our next wakeup.
4755 static void raid5d(struct md_thread
*thread
)
4757 struct mddev
*mddev
= thread
->mddev
;
4758 struct r5conf
*conf
= mddev
->private;
4760 struct blk_plug plug
;
4762 pr_debug("+++ raid5d active\n");
4764 md_check_recovery(mddev
);
4766 blk_start_plug(&plug
);
4768 spin_lock_irq(&conf
->device_lock
);
4774 !list_empty(&conf
->bitmap_list
)) {
4775 /* Now is a good time to flush some bitmap updates */
4777 spin_unlock_irq(&conf
->device_lock
);
4778 bitmap_unplug(mddev
->bitmap
);
4779 spin_lock_irq(&conf
->device_lock
);
4780 conf
->seq_write
= conf
->seq_flush
;
4781 activate_bit_delay(conf
);
4783 raid5_activate_delayed(conf
);
4785 while ((bio
= remove_bio_from_retry(conf
))) {
4787 spin_unlock_irq(&conf
->device_lock
);
4788 ok
= retry_aligned_read(conf
, bio
);
4789 spin_lock_irq(&conf
->device_lock
);
4795 batch_size
= handle_active_stripes(conf
);
4798 handled
+= batch_size
;
4800 if (mddev
->flags
& ~(1<<MD_CHANGE_PENDING
)) {
4801 spin_unlock_irq(&conf
->device_lock
);
4802 md_check_recovery(mddev
);
4803 spin_lock_irq(&conf
->device_lock
);
4806 pr_debug("%d stripes handled\n", handled
);
4808 spin_unlock_irq(&conf
->device_lock
);
4810 async_tx_issue_pending_all();
4811 blk_finish_plug(&plug
);
4813 pr_debug("--- raid5d inactive\n");
4817 raid5_show_stripe_cache_size(struct mddev
*mddev
, char *page
)
4819 struct r5conf
*conf
= mddev
->private;
4821 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4827 raid5_set_cache_size(struct mddev
*mddev
, int size
)
4829 struct r5conf
*conf
= mddev
->private;
4832 if (size
<= 16 || size
> 32768)
4834 while (size
< conf
->max_nr_stripes
) {
4835 if (drop_one_stripe(conf
))
4836 conf
->max_nr_stripes
--;
4840 err
= md_allow_write(mddev
);
4843 while (size
> conf
->max_nr_stripes
) {
4844 if (grow_one_stripe(conf
))
4845 conf
->max_nr_stripes
++;
4850 EXPORT_SYMBOL(raid5_set_cache_size
);
4853 raid5_store_stripe_cache_size(struct mddev
*mddev
, const char *page
, size_t len
)
4855 struct r5conf
*conf
= mddev
->private;
4859 if (len
>= PAGE_SIZE
)
4864 if (strict_strtoul(page
, 10, &new))
4866 err
= raid5_set_cache_size(mddev
, new);
4872 static struct md_sysfs_entry
4873 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4874 raid5_show_stripe_cache_size
,
4875 raid5_store_stripe_cache_size
);
4878 raid5_show_preread_threshold(struct mddev
*mddev
, char *page
)
4880 struct r5conf
*conf
= mddev
->private;
4882 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4888 raid5_store_preread_threshold(struct mddev
*mddev
, const char *page
, size_t len
)
4890 struct r5conf
*conf
= mddev
->private;
4892 if (len
>= PAGE_SIZE
)
4897 if (strict_strtoul(page
, 10, &new))
4899 if (new > conf
->max_nr_stripes
)
4901 conf
->bypass_threshold
= new;
4905 static struct md_sysfs_entry
4906 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4908 raid5_show_preread_threshold
,
4909 raid5_store_preread_threshold
);
4912 stripe_cache_active_show(struct mddev
*mddev
, char *page
)
4914 struct r5conf
*conf
= mddev
->private;
4916 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4921 static struct md_sysfs_entry
4922 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4924 static struct attribute
*raid5_attrs
[] = {
4925 &raid5_stripecache_size
.attr
,
4926 &raid5_stripecache_active
.attr
,
4927 &raid5_preread_bypass_threshold
.attr
,
4930 static struct attribute_group raid5_attrs_group
= {
4932 .attrs
= raid5_attrs
,
4936 raid5_size(struct mddev
*mddev
, sector_t sectors
, int raid_disks
)
4938 struct r5conf
*conf
= mddev
->private;
4941 sectors
= mddev
->dev_sectors
;
4943 /* size is defined by the smallest of previous and new size */
4944 raid_disks
= min(conf
->raid_disks
, conf
->previous_raid_disks
);
4946 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
4947 sectors
&= ~((sector_t
)mddev
->new_chunk_sectors
- 1);
4948 return sectors
* (raid_disks
- conf
->max_degraded
);
4951 static void raid5_free_percpu(struct r5conf
*conf
)
4953 struct raid5_percpu
*percpu
;
4960 for_each_possible_cpu(cpu
) {
4961 percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4962 safe_put_page(percpu
->spare_page
);
4963 kfree(percpu
->scribble
);
4965 #ifdef CONFIG_HOTPLUG_CPU
4966 unregister_cpu_notifier(&conf
->cpu_notify
);
4970 free_percpu(conf
->percpu
);
4973 static void free_conf(struct r5conf
*conf
)
4975 shrink_stripes(conf
);
4976 raid5_free_percpu(conf
);
4978 kfree(conf
->stripe_hashtbl
);
4982 #ifdef CONFIG_HOTPLUG_CPU
4983 static int raid456_cpu_notify(struct notifier_block
*nfb
, unsigned long action
,
4986 struct r5conf
*conf
= container_of(nfb
, struct r5conf
, cpu_notify
);
4987 long cpu
= (long)hcpu
;
4988 struct raid5_percpu
*percpu
= per_cpu_ptr(conf
->percpu
, cpu
);
4991 case CPU_UP_PREPARE
:
4992 case CPU_UP_PREPARE_FROZEN
:
4993 if (conf
->level
== 6 && !percpu
->spare_page
)
4994 percpu
->spare_page
= alloc_page(GFP_KERNEL
);
4995 if (!percpu
->scribble
)
4996 percpu
->scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
4998 if (!percpu
->scribble
||
4999 (conf
->level
== 6 && !percpu
->spare_page
)) {
5000 safe_put_page(percpu
->spare_page
);
5001 kfree(percpu
->scribble
);
5002 pr_err("%s: failed memory allocation for cpu%ld\n",
5004 return notifier_from_errno(-ENOMEM
);
5008 case CPU_DEAD_FROZEN
:
5009 safe_put_page(percpu
->spare_page
);
5010 kfree(percpu
->scribble
);
5011 percpu
->spare_page
= NULL
;
5012 percpu
->scribble
= NULL
;
5021 static int raid5_alloc_percpu(struct r5conf
*conf
)
5024 struct page
*spare_page
;
5025 struct raid5_percpu __percpu
*allcpus
;
5029 allcpus
= alloc_percpu(struct raid5_percpu
);
5032 conf
->percpu
= allcpus
;
5036 for_each_present_cpu(cpu
) {
5037 if (conf
->level
== 6) {
5038 spare_page
= alloc_page(GFP_KERNEL
);
5043 per_cpu_ptr(conf
->percpu
, cpu
)->spare_page
= spare_page
;
5045 scribble
= kmalloc(conf
->scribble_len
, GFP_KERNEL
);
5050 per_cpu_ptr(conf
->percpu
, cpu
)->scribble
= scribble
;
5052 #ifdef CONFIG_HOTPLUG_CPU
5053 conf
->cpu_notify
.notifier_call
= raid456_cpu_notify
;
5054 conf
->cpu_notify
.priority
= 0;
5056 err
= register_cpu_notifier(&conf
->cpu_notify
);
5063 static struct r5conf
*setup_conf(struct mddev
*mddev
)
5065 struct r5conf
*conf
;
5066 int raid_disk
, memory
, max_disks
;
5067 struct md_rdev
*rdev
;
5068 struct disk_info
*disk
;
5071 if (mddev
->new_level
!= 5
5072 && mddev
->new_level
!= 4
5073 && mddev
->new_level
!= 6) {
5074 printk(KERN_ERR
"md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5075 mdname(mddev
), mddev
->new_level
);
5076 return ERR_PTR(-EIO
);
5078 if ((mddev
->new_level
== 5
5079 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
5080 (mddev
->new_level
== 6
5081 && !algorithm_valid_raid6(mddev
->new_layout
))) {
5082 printk(KERN_ERR
"md/raid:%s: layout %d not supported\n",
5083 mdname(mddev
), mddev
->new_layout
);
5084 return ERR_PTR(-EIO
);
5086 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
5087 printk(KERN_ERR
"md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5088 mdname(mddev
), mddev
->raid_disks
);
5089 return ERR_PTR(-EINVAL
);
5092 if (!mddev
->new_chunk_sectors
||
5093 (mddev
->new_chunk_sectors
<< 9) % PAGE_SIZE
||
5094 !is_power_of_2(mddev
->new_chunk_sectors
)) {
5095 printk(KERN_ERR
"md/raid:%s: invalid chunk size %d\n",
5096 mdname(mddev
), mddev
->new_chunk_sectors
<< 9);
5097 return ERR_PTR(-EINVAL
);
5100 conf
= kzalloc(sizeof(struct r5conf
), GFP_KERNEL
);
5103 spin_lock_init(&conf
->device_lock
);
5104 init_waitqueue_head(&conf
->wait_for_stripe
);
5105 init_waitqueue_head(&conf
->wait_for_overlap
);
5106 INIT_LIST_HEAD(&conf
->handle_list
);
5107 INIT_LIST_HEAD(&conf
->hold_list
);
5108 INIT_LIST_HEAD(&conf
->delayed_list
);
5109 INIT_LIST_HEAD(&conf
->bitmap_list
);
5110 INIT_LIST_HEAD(&conf
->inactive_list
);
5111 atomic_set(&conf
->active_stripes
, 0);
5112 atomic_set(&conf
->preread_active_stripes
, 0);
5113 atomic_set(&conf
->active_aligned_reads
, 0);
5114 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
5115 conf
->recovery_disabled
= mddev
->recovery_disabled
- 1;
5117 conf
->raid_disks
= mddev
->raid_disks
;
5118 if (mddev
->reshape_position
== MaxSector
)
5119 conf
->previous_raid_disks
= mddev
->raid_disks
;
5121 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5122 max_disks
= max(conf
->raid_disks
, conf
->previous_raid_disks
);
5123 conf
->scribble_len
= scribble_len(max_disks
);
5125 conf
->disks
= kzalloc(max_disks
* sizeof(struct disk_info
),
5130 conf
->mddev
= mddev
;
5132 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
5135 conf
->level
= mddev
->new_level
;
5136 if (raid5_alloc_percpu(conf
) != 0)
5139 pr_debug("raid456: run(%s) called.\n", mdname(mddev
));
5141 rdev_for_each(rdev
, mddev
) {
5142 raid_disk
= rdev
->raid_disk
;
5143 if (raid_disk
>= max_disks
5146 disk
= conf
->disks
+ raid_disk
;
5148 if (test_bit(Replacement
, &rdev
->flags
)) {
5149 if (disk
->replacement
)
5151 disk
->replacement
= rdev
;
5158 if (test_bit(In_sync
, &rdev
->flags
)) {
5159 char b
[BDEVNAME_SIZE
];
5160 printk(KERN_INFO
"md/raid:%s: device %s operational as raid"
5162 mdname(mddev
), bdevname(rdev
->bdev
, b
), raid_disk
);
5163 } else if (rdev
->saved_raid_disk
!= raid_disk
)
5164 /* Cannot rely on bitmap to complete recovery */
5168 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5169 conf
->level
= mddev
->new_level
;
5170 if (conf
->level
== 6)
5171 conf
->max_degraded
= 2;
5173 conf
->max_degraded
= 1;
5174 conf
->algorithm
= mddev
->new_layout
;
5175 conf
->max_nr_stripes
= NR_STRIPES
;
5176 conf
->reshape_progress
= mddev
->reshape_position
;
5177 if (conf
->reshape_progress
!= MaxSector
) {
5178 conf
->prev_chunk_sectors
= mddev
->chunk_sectors
;
5179 conf
->prev_algo
= mddev
->layout
;
5182 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
5183 max_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
5184 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
5186 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5187 mdname(mddev
), memory
);
5190 printk(KERN_INFO
"md/raid:%s: allocated %dkB\n",
5191 mdname(mddev
), memory
);
5193 sprintf(pers_name
, "raid%d", mddev
->new_level
);
5194 conf
->thread
= md_register_thread(raid5d
, mddev
, pers_name
);
5195 if (!conf
->thread
) {
5197 "md/raid:%s: couldn't allocate thread.\n",
5207 return ERR_PTR(-EIO
);
5209 return ERR_PTR(-ENOMEM
);
5213 static int only_parity(int raid_disk
, int algo
, int raid_disks
, int max_degraded
)
5216 case ALGORITHM_PARITY_0
:
5217 if (raid_disk
< max_degraded
)
5220 case ALGORITHM_PARITY_N
:
5221 if (raid_disk
>= raid_disks
- max_degraded
)
5224 case ALGORITHM_PARITY_0_6
:
5225 if (raid_disk
== 0 ||
5226 raid_disk
== raid_disks
- 1)
5229 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5230 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5231 case ALGORITHM_LEFT_SYMMETRIC_6
:
5232 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5233 if (raid_disk
== raid_disks
- 1)
5239 static int run(struct mddev
*mddev
)
5241 struct r5conf
*conf
;
5242 int working_disks
= 0;
5243 int dirty_parity_disks
= 0;
5244 struct md_rdev
*rdev
;
5245 sector_t reshape_offset
= 0;
5247 long long min_offset_diff
= 0;
5250 if (mddev
->recovery_cp
!= MaxSector
)
5251 printk(KERN_NOTICE
"md/raid:%s: not clean"
5252 " -- starting background reconstruction\n",
5255 rdev_for_each(rdev
, mddev
) {
5257 if (rdev
->raid_disk
< 0)
5259 diff
= (rdev
->new_data_offset
- rdev
->data_offset
);
5261 min_offset_diff
= diff
;
5263 } else if (mddev
->reshape_backwards
&&
5264 diff
< min_offset_diff
)
5265 min_offset_diff
= diff
;
5266 else if (!mddev
->reshape_backwards
&&
5267 diff
> min_offset_diff
)
5268 min_offset_diff
= diff
;
5271 if (mddev
->reshape_position
!= MaxSector
) {
5272 /* Check that we can continue the reshape.
5273 * Difficulties arise if the stripe we would write to
5274 * next is at or after the stripe we would read from next.
5275 * For a reshape that changes the number of devices, this
5276 * is only possible for a very short time, and mdadm makes
5277 * sure that time appears to have past before assembling
5278 * the array. So we fail if that time hasn't passed.
5279 * For a reshape that keeps the number of devices the same
5280 * mdadm must be monitoring the reshape can keeping the
5281 * critical areas read-only and backed up. It will start
5282 * the array in read-only mode, so we check for that.
5284 sector_t here_new
, here_old
;
5286 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
5288 if (mddev
->new_level
!= mddev
->level
) {
5289 printk(KERN_ERR
"md/raid:%s: unsupported reshape "
5290 "required - aborting.\n",
5294 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
5295 /* reshape_position must be on a new-stripe boundary, and one
5296 * further up in new geometry must map after here in old
5299 here_new
= mddev
->reshape_position
;
5300 if (sector_div(here_new
, mddev
->new_chunk_sectors
*
5301 (mddev
->raid_disks
- max_degraded
))) {
5302 printk(KERN_ERR
"md/raid:%s: reshape_position not "
5303 "on a stripe boundary\n", mdname(mddev
));
5306 reshape_offset
= here_new
* mddev
->new_chunk_sectors
;
5307 /* here_new is the stripe we will write to */
5308 here_old
= mddev
->reshape_position
;
5309 sector_div(here_old
, mddev
->chunk_sectors
*
5310 (old_disks
-max_degraded
));
5311 /* here_old is the first stripe that we might need to read
5313 if (mddev
->delta_disks
== 0) {
5314 if ((here_new
* mddev
->new_chunk_sectors
!=
5315 here_old
* mddev
->chunk_sectors
)) {
5316 printk(KERN_ERR
"md/raid:%s: reshape position is"
5317 " confused - aborting\n", mdname(mddev
));
5320 /* We cannot be sure it is safe to start an in-place
5321 * reshape. It is only safe if user-space is monitoring
5322 * and taking constant backups.
5323 * mdadm always starts a situation like this in
5324 * readonly mode so it can take control before
5325 * allowing any writes. So just check for that.
5327 if (abs(min_offset_diff
) >= mddev
->chunk_sectors
&&
5328 abs(min_offset_diff
) >= mddev
->new_chunk_sectors
)
5329 /* not really in-place - so OK */;
5330 else if (mddev
->ro
== 0) {
5331 printk(KERN_ERR
"md/raid:%s: in-place reshape "
5332 "must be started in read-only mode "
5337 } else if (mddev
->reshape_backwards
5338 ? (here_new
* mddev
->new_chunk_sectors
+ min_offset_diff
<=
5339 here_old
* mddev
->chunk_sectors
)
5340 : (here_new
* mddev
->new_chunk_sectors
>=
5341 here_old
* mddev
->chunk_sectors
+ (-min_offset_diff
))) {
5342 /* Reading from the same stripe as writing to - bad */
5343 printk(KERN_ERR
"md/raid:%s: reshape_position too early for "
5344 "auto-recovery - aborting.\n",
5348 printk(KERN_INFO
"md/raid:%s: reshape will continue\n",
5350 /* OK, we should be able to continue; */
5352 BUG_ON(mddev
->level
!= mddev
->new_level
);
5353 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
5354 BUG_ON(mddev
->chunk_sectors
!= mddev
->new_chunk_sectors
);
5355 BUG_ON(mddev
->delta_disks
!= 0);
5358 if (mddev
->private == NULL
)
5359 conf
= setup_conf(mddev
);
5361 conf
= mddev
->private;
5364 return PTR_ERR(conf
);
5366 conf
->min_offset_diff
= min_offset_diff
;
5367 mddev
->thread
= conf
->thread
;
5368 conf
->thread
= NULL
;
5369 mddev
->private = conf
;
5371 for (i
= 0; i
< conf
->raid_disks
&& conf
->previous_raid_disks
;
5373 rdev
= conf
->disks
[i
].rdev
;
5374 if (!rdev
&& conf
->disks
[i
].replacement
) {
5375 /* The replacement is all we have yet */
5376 rdev
= conf
->disks
[i
].replacement
;
5377 conf
->disks
[i
].replacement
= NULL
;
5378 clear_bit(Replacement
, &rdev
->flags
);
5379 conf
->disks
[i
].rdev
= rdev
;
5383 if (conf
->disks
[i
].replacement
&&
5384 conf
->reshape_progress
!= MaxSector
) {
5385 /* replacements and reshape simply do not mix. */
5386 printk(KERN_ERR
"md: cannot handle concurrent "
5387 "replacement and reshape.\n");
5390 if (test_bit(In_sync
, &rdev
->flags
)) {
5394 /* This disc is not fully in-sync. However if it
5395 * just stored parity (beyond the recovery_offset),
5396 * when we don't need to be concerned about the
5397 * array being dirty.
5398 * When reshape goes 'backwards', we never have
5399 * partially completed devices, so we only need
5400 * to worry about reshape going forwards.
5402 /* Hack because v0.91 doesn't store recovery_offset properly. */
5403 if (mddev
->major_version
== 0 &&
5404 mddev
->minor_version
> 90)
5405 rdev
->recovery_offset
= reshape_offset
;
5407 if (rdev
->recovery_offset
< reshape_offset
) {
5408 /* We need to check old and new layout */
5409 if (!only_parity(rdev
->raid_disk
,
5412 conf
->max_degraded
))
5415 if (!only_parity(rdev
->raid_disk
,
5417 conf
->previous_raid_disks
,
5418 conf
->max_degraded
))
5420 dirty_parity_disks
++;
5424 * 0 for a fully functional array, 1 or 2 for a degraded array.
5426 mddev
->degraded
= calc_degraded(conf
);
5428 if (has_failed(conf
)) {
5429 printk(KERN_ERR
"md/raid:%s: not enough operational devices"
5430 " (%d/%d failed)\n",
5431 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
5435 /* device size must be a multiple of chunk size */
5436 mddev
->dev_sectors
&= ~(mddev
->chunk_sectors
- 1);
5437 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
5439 if (mddev
->degraded
> dirty_parity_disks
&&
5440 mddev
->recovery_cp
!= MaxSector
) {
5441 if (mddev
->ok_start_degraded
)
5443 "md/raid:%s: starting dirty degraded array"
5444 " - data corruption possible.\n",
5448 "md/raid:%s: cannot start dirty degraded array.\n",
5454 if (mddev
->degraded
== 0)
5455 printk(KERN_INFO
"md/raid:%s: raid level %d active with %d out of %d"
5456 " devices, algorithm %d\n", mdname(mddev
), conf
->level
,
5457 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
5460 printk(KERN_ALERT
"md/raid:%s: raid level %d active with %d"
5461 " out of %d devices, algorithm %d\n",
5462 mdname(mddev
), conf
->level
,
5463 mddev
->raid_disks
- mddev
->degraded
,
5464 mddev
->raid_disks
, mddev
->new_layout
);
5466 print_raid5_conf(conf
);
5468 if (conf
->reshape_progress
!= MaxSector
) {
5469 conf
->reshape_safe
= conf
->reshape_progress
;
5470 atomic_set(&conf
->reshape_stripes
, 0);
5471 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5472 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5473 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5474 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5475 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5480 /* Ok, everything is just fine now */
5481 if (mddev
->to_remove
== &raid5_attrs_group
)
5482 mddev
->to_remove
= NULL
;
5483 else if (mddev
->kobj
.sd
&&
5484 sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
5486 "raid5: failed to create sysfs attributes for %s\n",
5488 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5492 bool discard_supported
= true;
5493 /* read-ahead size must cover two whole stripes, which
5494 * is 2 * (datadisks) * chunksize where 'n' is the
5495 * number of raid devices
5497 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
5498 int stripe
= data_disks
*
5499 ((mddev
->chunk_sectors
<< 9) / PAGE_SIZE
);
5500 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5501 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5503 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
5505 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
5506 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
5508 chunk_size
= mddev
->chunk_sectors
<< 9;
5509 blk_queue_io_min(mddev
->queue
, chunk_size
);
5510 blk_queue_io_opt(mddev
->queue
, chunk_size
*
5511 (conf
->raid_disks
- conf
->max_degraded
));
5513 * We can only discard a whole stripe. It doesn't make sense to
5514 * discard data disk but write parity disk
5516 stripe
= stripe
* PAGE_SIZE
;
5517 mddev
->queue
->limits
.discard_alignment
= stripe
;
5518 mddev
->queue
->limits
.discard_granularity
= stripe
;
5520 * unaligned part of discard request will be ignored, so can't
5521 * guarantee discard_zerors_data
5523 mddev
->queue
->limits
.discard_zeroes_data
= 0;
5525 rdev_for_each(rdev
, mddev
) {
5526 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5527 rdev
->data_offset
<< 9);
5528 disk_stack_limits(mddev
->gendisk
, rdev
->bdev
,
5529 rdev
->new_data_offset
<< 9);
5531 * discard_zeroes_data is required, otherwise data
5532 * could be lost. Consider a scenario: discard a stripe
5533 * (the stripe could be inconsistent if
5534 * discard_zeroes_data is 0); write one disk of the
5535 * stripe (the stripe could be inconsistent again
5536 * depending on which disks are used to calculate
5537 * parity); the disk is broken; The stripe data of this
5540 if (!blk_queue_discard(bdev_get_queue(rdev
->bdev
)) ||
5541 !bdev_get_queue(rdev
->bdev
)->
5542 limits
.discard_zeroes_data
)
5543 discard_supported
= false;
5546 if (discard_supported
&&
5547 mddev
->queue
->limits
.max_discard_sectors
>= stripe
&&
5548 mddev
->queue
->limits
.discard_granularity
>= stripe
)
5549 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD
,
5552 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD
,
5558 md_unregister_thread(&mddev
->thread
);
5559 print_raid5_conf(conf
);
5561 mddev
->private = NULL
;
5562 printk(KERN_ALERT
"md/raid:%s: failed to run raid set.\n", mdname(mddev
));
5566 static int stop(struct mddev
*mddev
)
5568 struct r5conf
*conf
= mddev
->private;
5570 md_unregister_thread(&mddev
->thread
);
5572 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
5574 mddev
->private = NULL
;
5575 mddev
->to_remove
= &raid5_attrs_group
;
5579 static void status(struct seq_file
*seq
, struct mddev
*mddev
)
5581 struct r5conf
*conf
= mddev
->private;
5584 seq_printf(seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
,
5585 mddev
->chunk_sectors
/ 2, mddev
->layout
);
5586 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
5587 for (i
= 0; i
< conf
->raid_disks
; i
++)
5588 seq_printf (seq
, "%s",
5589 conf
->disks
[i
].rdev
&&
5590 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
5591 seq_printf (seq
, "]");
5594 static void print_raid5_conf (struct r5conf
*conf
)
5597 struct disk_info
*tmp
;
5599 printk(KERN_DEBUG
"RAID conf printout:\n");
5601 printk("(conf==NULL)\n");
5604 printk(KERN_DEBUG
" --- level:%d rd:%d wd:%d\n", conf
->level
,
5606 conf
->raid_disks
- conf
->mddev
->degraded
);
5608 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5609 char b
[BDEVNAME_SIZE
];
5610 tmp
= conf
->disks
+ i
;
5612 printk(KERN_DEBUG
" disk %d, o:%d, dev:%s\n",
5613 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
5614 bdevname(tmp
->rdev
->bdev
, b
));
5618 static int raid5_spare_active(struct mddev
*mddev
)
5621 struct r5conf
*conf
= mddev
->private;
5622 struct disk_info
*tmp
;
5624 unsigned long flags
;
5626 for (i
= 0; i
< conf
->raid_disks
; i
++) {
5627 tmp
= conf
->disks
+ i
;
5628 if (tmp
->replacement
5629 && tmp
->replacement
->recovery_offset
== MaxSector
5630 && !test_bit(Faulty
, &tmp
->replacement
->flags
)
5631 && !test_and_set_bit(In_sync
, &tmp
->replacement
->flags
)) {
5632 /* Replacement has just become active. */
5634 || !test_and_clear_bit(In_sync
, &tmp
->rdev
->flags
))
5637 /* Replaced device not technically faulty,
5638 * but we need to be sure it gets removed
5639 * and never re-added.
5641 set_bit(Faulty
, &tmp
->rdev
->flags
);
5642 sysfs_notify_dirent_safe(
5643 tmp
->rdev
->sysfs_state
);
5645 sysfs_notify_dirent_safe(tmp
->replacement
->sysfs_state
);
5646 } else if (tmp
->rdev
5647 && tmp
->rdev
->recovery_offset
== MaxSector
5648 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
5649 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
5651 sysfs_notify_dirent_safe(tmp
->rdev
->sysfs_state
);
5654 spin_lock_irqsave(&conf
->device_lock
, flags
);
5655 mddev
->degraded
= calc_degraded(conf
);
5656 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5657 print_raid5_conf(conf
);
5661 static int raid5_remove_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5663 struct r5conf
*conf
= mddev
->private;
5665 int number
= rdev
->raid_disk
;
5666 struct md_rdev
**rdevp
;
5667 struct disk_info
*p
= conf
->disks
+ number
;
5669 print_raid5_conf(conf
);
5670 if (rdev
== p
->rdev
)
5672 else if (rdev
== p
->replacement
)
5673 rdevp
= &p
->replacement
;
5677 if (number
>= conf
->raid_disks
&&
5678 conf
->reshape_progress
== MaxSector
)
5679 clear_bit(In_sync
, &rdev
->flags
);
5681 if (test_bit(In_sync
, &rdev
->flags
) ||
5682 atomic_read(&rdev
->nr_pending
)) {
5686 /* Only remove non-faulty devices if recovery
5689 if (!test_bit(Faulty
, &rdev
->flags
) &&
5690 mddev
->recovery_disabled
!= conf
->recovery_disabled
&&
5691 !has_failed(conf
) &&
5692 (!p
->replacement
|| p
->replacement
== rdev
) &&
5693 number
< conf
->raid_disks
) {
5699 if (atomic_read(&rdev
->nr_pending
)) {
5700 /* lost the race, try later */
5703 } else if (p
->replacement
) {
5704 /* We must have just cleared 'rdev' */
5705 p
->rdev
= p
->replacement
;
5706 clear_bit(Replacement
, &p
->replacement
->flags
);
5707 smp_mb(); /* Make sure other CPUs may see both as identical
5708 * but will never see neither - if they are careful
5710 p
->replacement
= NULL
;
5711 clear_bit(WantReplacement
, &rdev
->flags
);
5713 /* We might have just removed the Replacement as faulty-
5714 * clear the bit just in case
5716 clear_bit(WantReplacement
, &rdev
->flags
);
5719 print_raid5_conf(conf
);
5723 static int raid5_add_disk(struct mddev
*mddev
, struct md_rdev
*rdev
)
5725 struct r5conf
*conf
= mddev
->private;
5728 struct disk_info
*p
;
5730 int last
= conf
->raid_disks
- 1;
5732 if (mddev
->recovery_disabled
== conf
->recovery_disabled
)
5735 if (rdev
->saved_raid_disk
< 0 && has_failed(conf
))
5736 /* no point adding a device */
5739 if (rdev
->raid_disk
>= 0)
5740 first
= last
= rdev
->raid_disk
;
5743 * find the disk ... but prefer rdev->saved_raid_disk
5746 if (rdev
->saved_raid_disk
>= 0 &&
5747 rdev
->saved_raid_disk
>= first
&&
5748 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
5749 first
= rdev
->saved_raid_disk
;
5751 for (disk
= first
; disk
<= last
; disk
++) {
5752 p
= conf
->disks
+ disk
;
5753 if (p
->rdev
== NULL
) {
5754 clear_bit(In_sync
, &rdev
->flags
);
5755 rdev
->raid_disk
= disk
;
5757 if (rdev
->saved_raid_disk
!= disk
)
5759 rcu_assign_pointer(p
->rdev
, rdev
);
5763 for (disk
= first
; disk
<= last
; disk
++) {
5764 p
= conf
->disks
+ disk
;
5765 if (test_bit(WantReplacement
, &p
->rdev
->flags
) &&
5766 p
->replacement
== NULL
) {
5767 clear_bit(In_sync
, &rdev
->flags
);
5768 set_bit(Replacement
, &rdev
->flags
);
5769 rdev
->raid_disk
= disk
;
5772 rcu_assign_pointer(p
->replacement
, rdev
);
5777 print_raid5_conf(conf
);
5781 static int raid5_resize(struct mddev
*mddev
, sector_t sectors
)
5783 /* no resync is happening, and there is enough space
5784 * on all devices, so we can resize.
5785 * We need to make sure resync covers any new space.
5786 * If the array is shrinking we should possibly wait until
5787 * any io in the removed space completes, but it hardly seems
5791 sectors
&= ~((sector_t
)mddev
->chunk_sectors
- 1);
5792 newsize
= raid5_size(mddev
, sectors
, mddev
->raid_disks
);
5793 if (mddev
->external_size
&&
5794 mddev
->array_sectors
> newsize
)
5796 if (mddev
->bitmap
) {
5797 int ret
= bitmap_resize(mddev
->bitmap
, sectors
, 0, 0);
5801 md_set_array_sectors(mddev
, newsize
);
5802 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5803 revalidate_disk(mddev
->gendisk
);
5804 if (sectors
> mddev
->dev_sectors
&&
5805 mddev
->recovery_cp
> mddev
->dev_sectors
) {
5806 mddev
->recovery_cp
= mddev
->dev_sectors
;
5807 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
5809 mddev
->dev_sectors
= sectors
;
5810 mddev
->resync_max_sectors
= sectors
;
5814 static int check_stripe_cache(struct mddev
*mddev
)
5816 /* Can only proceed if there are plenty of stripe_heads.
5817 * We need a minimum of one full stripe,, and for sensible progress
5818 * it is best to have about 4 times that.
5819 * If we require 4 times, then the default 256 4K stripe_heads will
5820 * allow for chunk sizes up to 256K, which is probably OK.
5821 * If the chunk size is greater, user-space should request more
5822 * stripe_heads first.
5824 struct r5conf
*conf
= mddev
->private;
5825 if (((mddev
->chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5826 > conf
->max_nr_stripes
||
5827 ((mddev
->new_chunk_sectors
<< 9) / STRIPE_SIZE
) * 4
5828 > conf
->max_nr_stripes
) {
5829 printk(KERN_WARNING
"md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5831 ((max(mddev
->chunk_sectors
, mddev
->new_chunk_sectors
) << 9)
5838 static int check_reshape(struct mddev
*mddev
)
5840 struct r5conf
*conf
= mddev
->private;
5842 if (mddev
->delta_disks
== 0 &&
5843 mddev
->new_layout
== mddev
->layout
&&
5844 mddev
->new_chunk_sectors
== mddev
->chunk_sectors
)
5845 return 0; /* nothing to do */
5846 if (has_failed(conf
))
5848 if (mddev
->delta_disks
< 0) {
5849 /* We might be able to shrink, but the devices must
5850 * be made bigger first.
5851 * For raid6, 4 is the minimum size.
5852 * Otherwise 2 is the minimum
5855 if (mddev
->level
== 6)
5857 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
5861 if (!check_stripe_cache(mddev
))
5864 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
5867 static int raid5_start_reshape(struct mddev
*mddev
)
5869 struct r5conf
*conf
= mddev
->private;
5870 struct md_rdev
*rdev
;
5872 unsigned long flags
;
5874 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
5877 if (!check_stripe_cache(mddev
))
5880 if (has_failed(conf
))
5883 rdev_for_each(rdev
, mddev
) {
5884 if (!test_bit(In_sync
, &rdev
->flags
)
5885 && !test_bit(Faulty
, &rdev
->flags
))
5889 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
5890 /* Not enough devices even to make a degraded array
5895 /* Refuse to reduce size of the array. Any reductions in
5896 * array size must be through explicit setting of array_size
5899 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
5900 < mddev
->array_sectors
) {
5901 printk(KERN_ERR
"md/raid:%s: array size must be reduced "
5902 "before number of disks\n", mdname(mddev
));
5906 atomic_set(&conf
->reshape_stripes
, 0);
5907 spin_lock_irq(&conf
->device_lock
);
5908 conf
->previous_raid_disks
= conf
->raid_disks
;
5909 conf
->raid_disks
+= mddev
->delta_disks
;
5910 conf
->prev_chunk_sectors
= conf
->chunk_sectors
;
5911 conf
->chunk_sectors
= mddev
->new_chunk_sectors
;
5912 conf
->prev_algo
= conf
->algorithm
;
5913 conf
->algorithm
= mddev
->new_layout
;
5915 /* Code that selects data_offset needs to see the generation update
5916 * if reshape_progress has been set - so a memory barrier needed.
5919 if (mddev
->reshape_backwards
)
5920 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
5922 conf
->reshape_progress
= 0;
5923 conf
->reshape_safe
= conf
->reshape_progress
;
5924 spin_unlock_irq(&conf
->device_lock
);
5926 /* Add some new drives, as many as will fit.
5927 * We know there are enough to make the newly sized array work.
5928 * Don't add devices if we are reducing the number of
5929 * devices in the array. This is because it is not possible
5930 * to correctly record the "partially reconstructed" state of
5931 * such devices during the reshape and confusion could result.
5933 if (mddev
->delta_disks
>= 0) {
5934 rdev_for_each(rdev
, mddev
)
5935 if (rdev
->raid_disk
< 0 &&
5936 !test_bit(Faulty
, &rdev
->flags
)) {
5937 if (raid5_add_disk(mddev
, rdev
) == 0) {
5939 >= conf
->previous_raid_disks
)
5940 set_bit(In_sync
, &rdev
->flags
);
5942 rdev
->recovery_offset
= 0;
5944 if (sysfs_link_rdev(mddev
, rdev
))
5945 /* Failure here is OK */;
5947 } else if (rdev
->raid_disk
>= conf
->previous_raid_disks
5948 && !test_bit(Faulty
, &rdev
->flags
)) {
5949 /* This is a spare that was manually added */
5950 set_bit(In_sync
, &rdev
->flags
);
5953 /* When a reshape changes the number of devices,
5954 * ->degraded is measured against the larger of the
5955 * pre and post number of devices.
5957 spin_lock_irqsave(&conf
->device_lock
, flags
);
5958 mddev
->degraded
= calc_degraded(conf
);
5959 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
5961 mddev
->raid_disks
= conf
->raid_disks
;
5962 mddev
->reshape_position
= conf
->reshape_progress
;
5963 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5965 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
5966 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
5967 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
5968 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
5969 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
5971 if (!mddev
->sync_thread
) {
5972 mddev
->recovery
= 0;
5973 spin_lock_irq(&conf
->device_lock
);
5974 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
5975 rdev_for_each(rdev
, mddev
)
5976 rdev
->new_data_offset
= rdev
->data_offset
;
5978 conf
->reshape_progress
= MaxSector
;
5979 mddev
->reshape_position
= MaxSector
;
5980 spin_unlock_irq(&conf
->device_lock
);
5983 conf
->reshape_checkpoint
= jiffies
;
5984 md_wakeup_thread(mddev
->sync_thread
);
5985 md_new_event(mddev
);
5989 /* This is called from the reshape thread and should make any
5990 * changes needed in 'conf'
5992 static void end_reshape(struct r5conf
*conf
)
5995 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
5996 struct md_rdev
*rdev
;
5998 spin_lock_irq(&conf
->device_lock
);
5999 conf
->previous_raid_disks
= conf
->raid_disks
;
6000 rdev_for_each(rdev
, conf
->mddev
)
6001 rdev
->data_offset
= rdev
->new_data_offset
;
6003 conf
->reshape_progress
= MaxSector
;
6004 spin_unlock_irq(&conf
->device_lock
);
6005 wake_up(&conf
->wait_for_overlap
);
6007 /* read-ahead size must cover two whole stripes, which is
6008 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6010 if (conf
->mddev
->queue
) {
6011 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
6012 int stripe
= data_disks
* ((conf
->chunk_sectors
<< 9)
6014 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
6015 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
6020 /* This is called from the raid5d thread with mddev_lock held.
6021 * It makes config changes to the device.
6023 static void raid5_finish_reshape(struct mddev
*mddev
)
6025 struct r5conf
*conf
= mddev
->private;
6027 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
6029 if (mddev
->delta_disks
> 0) {
6030 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
6031 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
6032 revalidate_disk(mddev
->gendisk
);
6035 spin_lock_irq(&conf
->device_lock
);
6036 mddev
->degraded
= calc_degraded(conf
);
6037 spin_unlock_irq(&conf
->device_lock
);
6038 for (d
= conf
->raid_disks
;
6039 d
< conf
->raid_disks
- mddev
->delta_disks
;
6041 struct md_rdev
*rdev
= conf
->disks
[d
].rdev
;
6043 clear_bit(In_sync
, &rdev
->flags
);
6044 rdev
= conf
->disks
[d
].replacement
;
6046 clear_bit(In_sync
, &rdev
->flags
);
6049 mddev
->layout
= conf
->algorithm
;
6050 mddev
->chunk_sectors
= conf
->chunk_sectors
;
6051 mddev
->reshape_position
= MaxSector
;
6052 mddev
->delta_disks
= 0;
6053 mddev
->reshape_backwards
= 0;
6057 static void raid5_quiesce(struct mddev
*mddev
, int state
)
6059 struct r5conf
*conf
= mddev
->private;
6062 case 2: /* resume for a suspend */
6063 wake_up(&conf
->wait_for_overlap
);
6066 case 1: /* stop all writes */
6067 spin_lock_irq(&conf
->device_lock
);
6068 /* '2' tells resync/reshape to pause so that all
6069 * active stripes can drain
6072 wait_event_lock_irq(conf
->wait_for_stripe
,
6073 atomic_read(&conf
->active_stripes
) == 0 &&
6074 atomic_read(&conf
->active_aligned_reads
) == 0,
6075 conf
->device_lock
, /* nothing */);
6077 spin_unlock_irq(&conf
->device_lock
);
6078 /* allow reshape to continue */
6079 wake_up(&conf
->wait_for_overlap
);
6082 case 0: /* re-enable writes */
6083 spin_lock_irq(&conf
->device_lock
);
6085 wake_up(&conf
->wait_for_stripe
);
6086 wake_up(&conf
->wait_for_overlap
);
6087 spin_unlock_irq(&conf
->device_lock
);
6093 static void *raid45_takeover_raid0(struct mddev
*mddev
, int level
)
6095 struct r0conf
*raid0_conf
= mddev
->private;
6098 /* for raid0 takeover only one zone is supported */
6099 if (raid0_conf
->nr_strip_zones
> 1) {
6100 printk(KERN_ERR
"md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6102 return ERR_PTR(-EINVAL
);
6105 sectors
= raid0_conf
->strip_zone
[0].zone_end
;
6106 sector_div(sectors
, raid0_conf
->strip_zone
[0].nb_dev
);
6107 mddev
->dev_sectors
= sectors
;
6108 mddev
->new_level
= level
;
6109 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6110 mddev
->new_chunk_sectors
= mddev
->chunk_sectors
;
6111 mddev
->raid_disks
+= 1;
6112 mddev
->delta_disks
= 1;
6113 /* make sure it will be not marked as dirty */
6114 mddev
->recovery_cp
= MaxSector
;
6116 return setup_conf(mddev
);
6120 static void *raid5_takeover_raid1(struct mddev
*mddev
)
6124 if (mddev
->raid_disks
!= 2 ||
6125 mddev
->degraded
> 1)
6126 return ERR_PTR(-EINVAL
);
6128 /* Should check if there are write-behind devices? */
6130 chunksect
= 64*2; /* 64K by default */
6132 /* The array must be an exact multiple of chunksize */
6133 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
6136 if ((chunksect
<<9) < STRIPE_SIZE
)
6137 /* array size does not allow a suitable chunk size */
6138 return ERR_PTR(-EINVAL
);
6140 mddev
->new_level
= 5;
6141 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6142 mddev
->new_chunk_sectors
= chunksect
;
6144 return setup_conf(mddev
);
6147 static void *raid5_takeover_raid6(struct mddev
*mddev
)
6151 switch (mddev
->layout
) {
6152 case ALGORITHM_LEFT_ASYMMETRIC_6
:
6153 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
6155 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
6156 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
6158 case ALGORITHM_LEFT_SYMMETRIC_6
:
6159 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
6161 case ALGORITHM_RIGHT_SYMMETRIC_6
:
6162 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
6164 case ALGORITHM_PARITY_0_6
:
6165 new_layout
= ALGORITHM_PARITY_0
;
6167 case ALGORITHM_PARITY_N
:
6168 new_layout
= ALGORITHM_PARITY_N
;
6171 return ERR_PTR(-EINVAL
);
6173 mddev
->new_level
= 5;
6174 mddev
->new_layout
= new_layout
;
6175 mddev
->delta_disks
= -1;
6176 mddev
->raid_disks
-= 1;
6177 return setup_conf(mddev
);
6181 static int raid5_check_reshape(struct mddev
*mddev
)
6183 /* For a 2-drive array, the layout and chunk size can be changed
6184 * immediately as not restriping is needed.
6185 * For larger arrays we record the new value - after validation
6186 * to be used by a reshape pass.
6188 struct r5conf
*conf
= mddev
->private;
6189 int new_chunk
= mddev
->new_chunk_sectors
;
6191 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid5(mddev
->new_layout
))
6193 if (new_chunk
> 0) {
6194 if (!is_power_of_2(new_chunk
))
6196 if (new_chunk
< (PAGE_SIZE
>>9))
6198 if (mddev
->array_sectors
& (new_chunk
-1))
6199 /* not factor of array size */
6203 /* They look valid */
6205 if (mddev
->raid_disks
== 2) {
6206 /* can make the change immediately */
6207 if (mddev
->new_layout
>= 0) {
6208 conf
->algorithm
= mddev
->new_layout
;
6209 mddev
->layout
= mddev
->new_layout
;
6211 if (new_chunk
> 0) {
6212 conf
->chunk_sectors
= new_chunk
;
6213 mddev
->chunk_sectors
= new_chunk
;
6215 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
6216 md_wakeup_thread(mddev
->thread
);
6218 return check_reshape(mddev
);
6221 static int raid6_check_reshape(struct mddev
*mddev
)
6223 int new_chunk
= mddev
->new_chunk_sectors
;
6225 if (mddev
->new_layout
>= 0 && !algorithm_valid_raid6(mddev
->new_layout
))
6227 if (new_chunk
> 0) {
6228 if (!is_power_of_2(new_chunk
))
6230 if (new_chunk
< (PAGE_SIZE
>> 9))
6232 if (mddev
->array_sectors
& (new_chunk
-1))
6233 /* not factor of array size */
6237 /* They look valid */
6238 return check_reshape(mddev
);
6241 static void *raid5_takeover(struct mddev
*mddev
)
6243 /* raid5 can take over:
6244 * raid0 - if there is only one strip zone - make it a raid4 layout
6245 * raid1 - if there are two drives. We need to know the chunk size
6246 * raid4 - trivial - just use a raid4 layout.
6247 * raid6 - Providing it is a *_6 layout
6249 if (mddev
->level
== 0)
6250 return raid45_takeover_raid0(mddev
, 5);
6251 if (mddev
->level
== 1)
6252 return raid5_takeover_raid1(mddev
);
6253 if (mddev
->level
== 4) {
6254 mddev
->new_layout
= ALGORITHM_PARITY_N
;
6255 mddev
->new_level
= 5;
6256 return setup_conf(mddev
);
6258 if (mddev
->level
== 6)
6259 return raid5_takeover_raid6(mddev
);
6261 return ERR_PTR(-EINVAL
);
6264 static void *raid4_takeover(struct mddev
*mddev
)
6266 /* raid4 can take over:
6267 * raid0 - if there is only one strip zone
6268 * raid5 - if layout is right
6270 if (mddev
->level
== 0)
6271 return raid45_takeover_raid0(mddev
, 4);
6272 if (mddev
->level
== 5 &&
6273 mddev
->layout
== ALGORITHM_PARITY_N
) {
6274 mddev
->new_layout
= 0;
6275 mddev
->new_level
= 4;
6276 return setup_conf(mddev
);
6278 return ERR_PTR(-EINVAL
);
6281 static struct md_personality raid5_personality
;
6283 static void *raid6_takeover(struct mddev
*mddev
)
6285 /* Currently can only take over a raid5. We map the
6286 * personality to an equivalent raid6 personality
6287 * with the Q block at the end.
6291 if (mddev
->pers
!= &raid5_personality
)
6292 return ERR_PTR(-EINVAL
);
6293 if (mddev
->degraded
> 1)
6294 return ERR_PTR(-EINVAL
);
6295 if (mddev
->raid_disks
> 253)
6296 return ERR_PTR(-EINVAL
);
6297 if (mddev
->raid_disks
< 3)
6298 return ERR_PTR(-EINVAL
);
6300 switch (mddev
->layout
) {
6301 case ALGORITHM_LEFT_ASYMMETRIC
:
6302 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
6304 case ALGORITHM_RIGHT_ASYMMETRIC
:
6305 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
6307 case ALGORITHM_LEFT_SYMMETRIC
:
6308 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
6310 case ALGORITHM_RIGHT_SYMMETRIC
:
6311 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
6313 case ALGORITHM_PARITY_0
:
6314 new_layout
= ALGORITHM_PARITY_0_6
;
6316 case ALGORITHM_PARITY_N
:
6317 new_layout
= ALGORITHM_PARITY_N
;
6320 return ERR_PTR(-EINVAL
);
6322 mddev
->new_level
= 6;
6323 mddev
->new_layout
= new_layout
;
6324 mddev
->delta_disks
= 1;
6325 mddev
->raid_disks
+= 1;
6326 return setup_conf(mddev
);
6330 static struct md_personality raid6_personality
=
6334 .owner
= THIS_MODULE
,
6335 .make_request
= make_request
,
6339 .error_handler
= error
,
6340 .hot_add_disk
= raid5_add_disk
,
6341 .hot_remove_disk
= raid5_remove_disk
,
6342 .spare_active
= raid5_spare_active
,
6343 .sync_request
= sync_request
,
6344 .resize
= raid5_resize
,
6346 .check_reshape
= raid6_check_reshape
,
6347 .start_reshape
= raid5_start_reshape
,
6348 .finish_reshape
= raid5_finish_reshape
,
6349 .quiesce
= raid5_quiesce
,
6350 .takeover
= raid6_takeover
,
6352 static struct md_personality raid5_personality
=
6356 .owner
= THIS_MODULE
,
6357 .make_request
= make_request
,
6361 .error_handler
= error
,
6362 .hot_add_disk
= raid5_add_disk
,
6363 .hot_remove_disk
= raid5_remove_disk
,
6364 .spare_active
= raid5_spare_active
,
6365 .sync_request
= sync_request
,
6366 .resize
= raid5_resize
,
6368 .check_reshape
= raid5_check_reshape
,
6369 .start_reshape
= raid5_start_reshape
,
6370 .finish_reshape
= raid5_finish_reshape
,
6371 .quiesce
= raid5_quiesce
,
6372 .takeover
= raid5_takeover
,
6375 static struct md_personality raid4_personality
=
6379 .owner
= THIS_MODULE
,
6380 .make_request
= make_request
,
6384 .error_handler
= error
,
6385 .hot_add_disk
= raid5_add_disk
,
6386 .hot_remove_disk
= raid5_remove_disk
,
6387 .spare_active
= raid5_spare_active
,
6388 .sync_request
= sync_request
,
6389 .resize
= raid5_resize
,
6391 .check_reshape
= raid5_check_reshape
,
6392 .start_reshape
= raid5_start_reshape
,
6393 .finish_reshape
= raid5_finish_reshape
,
6394 .quiesce
= raid5_quiesce
,
6395 .takeover
= raid4_takeover
,
6398 static int __init
raid5_init(void)
6400 register_md_personality(&raid6_personality
);
6401 register_md_personality(&raid5_personality
);
6402 register_md_personality(&raid4_personality
);
6406 static void raid5_exit(void)
6408 unregister_md_personality(&raid6_personality
);
6409 unregister_md_personality(&raid5_personality
);
6410 unregister_md_personality(&raid4_personality
);
6413 module_init(raid5_init
);
6414 module_exit(raid5_exit
);
6415 MODULE_LICENSE("GPL");
6416 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6417 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6418 MODULE_ALIAS("md-raid5");
6419 MODULE_ALIAS("md-raid4");
6420 MODULE_ALIAS("md-level-5");
6421 MODULE_ALIAS("md-level-4");
6422 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6423 MODULE_ALIAS("md-raid6");
6424 MODULE_ALIAS("md-level-6");
6426 /* This used to be two separate modules, they were: */
6427 MODULE_ALIAS("raid5");
6428 MODULE_ALIAS("raid6");