2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->bm_write is the number of the last batch successfully written.
31 * conf->bm_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is bm_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
54 #include <linux/raid/bitmap.h>
55 #include <linux/async_tx.h>
61 #define NR_STRIPES 256
62 #define STRIPE_SIZE PAGE_SIZE
63 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
65 #define IO_THRESHOLD 1
66 #define BYPASS_THRESHOLD 1
67 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
68 #define HASH_MASK (NR_HASH - 1)
70 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
72 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
73 * order without overlap. There may be several bio's per stripe+device, and
74 * a bio could span several devices.
75 * When walking this list for a particular stripe+device, we must never proceed
76 * beyond a bio that extends past this device, as the next bio might no longer
78 * This macro is used to determine the 'next' bio in the list, given the sector
79 * of the current stripe+device
81 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
83 * The following can be used to debug the driver
85 #define RAID5_PARANOIA 1
86 #if RAID5_PARANOIA && defined(CONFIG_SMP)
87 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
89 # define CHECK_DEVLOCK()
97 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
99 #if !RAID6_USE_EMPTY_ZERO_PAGE
100 /* In .bss so it's zeroed */
101 const char raid6_empty_zero_page
[PAGE_SIZE
] __attribute__((aligned(256)));
104 static inline int raid6_next_disk(int disk
, int raid_disks
)
107 return (disk
< raid_disks
) ? disk
: 0;
110 static void return_io(struct bio
*return_bi
)
112 struct bio
*bi
= return_bi
;
115 return_bi
= bi
->bi_next
;
123 static void print_raid5_conf (raid5_conf_t
*conf
);
125 static void __release_stripe(raid5_conf_t
*conf
, struct stripe_head
*sh
)
127 if (atomic_dec_and_test(&sh
->count
)) {
128 BUG_ON(!list_empty(&sh
->lru
));
129 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
130 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
131 if (test_bit(STRIPE_DELAYED
, &sh
->state
)) {
132 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
133 blk_plug_device(conf
->mddev
->queue
);
134 } else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
135 sh
->bm_seq
- conf
->seq_write
> 0) {
136 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
137 blk_plug_device(conf
->mddev
->queue
);
139 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
140 list_add_tail(&sh
->lru
, &conf
->handle_list
);
142 md_wakeup_thread(conf
->mddev
->thread
);
144 BUG_ON(sh
->ops
.pending
);
145 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
146 atomic_dec(&conf
->preread_active_stripes
);
147 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
)
148 md_wakeup_thread(conf
->mddev
->thread
);
150 atomic_dec(&conf
->active_stripes
);
151 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
152 list_add_tail(&sh
->lru
, &conf
->inactive_list
);
153 wake_up(&conf
->wait_for_stripe
);
154 if (conf
->retry_read_aligned
)
155 md_wakeup_thread(conf
->mddev
->thread
);
160 static void release_stripe(struct stripe_head
*sh
)
162 raid5_conf_t
*conf
= sh
->raid_conf
;
165 spin_lock_irqsave(&conf
->device_lock
, flags
);
166 __release_stripe(conf
, sh
);
167 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
170 static inline void remove_hash(struct stripe_head
*sh
)
172 pr_debug("remove_hash(), stripe %llu\n",
173 (unsigned long long)sh
->sector
);
175 hlist_del_init(&sh
->hash
);
178 static inline void insert_hash(raid5_conf_t
*conf
, struct stripe_head
*sh
)
180 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
182 pr_debug("insert_hash(), stripe %llu\n",
183 (unsigned long long)sh
->sector
);
186 hlist_add_head(&sh
->hash
, hp
);
190 /* find an idle stripe, make sure it is unhashed, and return it. */
191 static struct stripe_head
*get_free_stripe(raid5_conf_t
*conf
)
193 struct stripe_head
*sh
= NULL
;
194 struct list_head
*first
;
197 if (list_empty(&conf
->inactive_list
))
199 first
= conf
->inactive_list
.next
;
200 sh
= list_entry(first
, struct stripe_head
, lru
);
201 list_del_init(first
);
203 atomic_inc(&conf
->active_stripes
);
208 static void shrink_buffers(struct stripe_head
*sh
, int num
)
213 for (i
=0; i
<num
; i
++) {
217 sh
->dev
[i
].page
= NULL
;
222 static int grow_buffers(struct stripe_head
*sh
, int num
)
226 for (i
=0; i
<num
; i
++) {
229 if (!(page
= alloc_page(GFP_KERNEL
))) {
232 sh
->dev
[i
].page
= page
;
237 static void raid5_build_block (struct stripe_head
*sh
, int i
);
239 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int pd_idx
, int disks
)
241 raid5_conf_t
*conf
= sh
->raid_conf
;
244 BUG_ON(atomic_read(&sh
->count
) != 0);
245 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
246 BUG_ON(sh
->ops
.pending
|| sh
->ops
.ack
|| sh
->ops
.complete
);
249 pr_debug("init_stripe called, stripe %llu\n",
250 (unsigned long long)sh
->sector
);
260 for (i
= sh
->disks
; i
--; ) {
261 struct r5dev
*dev
= &sh
->dev
[i
];
263 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
264 test_bit(R5_LOCKED
, &dev
->flags
)) {
265 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
266 (unsigned long long)sh
->sector
, i
, dev
->toread
,
267 dev
->read
, dev
->towrite
, dev
->written
,
268 test_bit(R5_LOCKED
, &dev
->flags
));
272 raid5_build_block(sh
, i
);
274 insert_hash(conf
, sh
);
277 static struct stripe_head
*__find_stripe(raid5_conf_t
*conf
, sector_t sector
, int disks
)
279 struct stripe_head
*sh
;
280 struct hlist_node
*hn
;
283 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
284 hlist_for_each_entry(sh
, hn
, stripe_hash(conf
, sector
), hash
)
285 if (sh
->sector
== sector
&& sh
->disks
== disks
)
287 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
291 static void unplug_slaves(mddev_t
*mddev
);
292 static void raid5_unplug_device(struct request_queue
*q
);
294 static struct stripe_head
*get_active_stripe(raid5_conf_t
*conf
, sector_t sector
, int disks
,
295 int pd_idx
, int noblock
)
297 struct stripe_head
*sh
;
299 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
301 spin_lock_irq(&conf
->device_lock
);
304 wait_event_lock_irq(conf
->wait_for_stripe
,
306 conf
->device_lock
, /* nothing */);
307 sh
= __find_stripe(conf
, sector
, disks
);
309 if (!conf
->inactive_blocked
)
310 sh
= get_free_stripe(conf
);
311 if (noblock
&& sh
== NULL
)
314 conf
->inactive_blocked
= 1;
315 wait_event_lock_irq(conf
->wait_for_stripe
,
316 !list_empty(&conf
->inactive_list
) &&
317 (atomic_read(&conf
->active_stripes
)
318 < (conf
->max_nr_stripes
*3/4)
319 || !conf
->inactive_blocked
),
321 raid5_unplug_device(conf
->mddev
->queue
)
323 conf
->inactive_blocked
= 0;
325 init_stripe(sh
, sector
, pd_idx
, disks
);
327 if (atomic_read(&sh
->count
)) {
328 BUG_ON(!list_empty(&sh
->lru
));
330 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
331 atomic_inc(&conf
->active_stripes
);
332 if (list_empty(&sh
->lru
) &&
333 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
335 list_del_init(&sh
->lru
);
338 } while (sh
== NULL
);
341 atomic_inc(&sh
->count
);
343 spin_unlock_irq(&conf
->device_lock
);
347 /* test_and_ack_op() ensures that we only dequeue an operation once */
348 #define test_and_ack_op(op, pend) \
350 if (test_bit(op, &sh->ops.pending) && \
351 !test_bit(op, &sh->ops.complete)) { \
352 if (test_and_set_bit(op, &sh->ops.ack)) \
353 clear_bit(op, &pend); \
357 clear_bit(op, &pend); \
360 /* find new work to run, do not resubmit work that is already
363 static unsigned long get_stripe_work(struct stripe_head
*sh
)
365 unsigned long pending
;
368 pending
= sh
->ops
.pending
;
370 test_and_ack_op(STRIPE_OP_BIOFILL
, pending
);
371 test_and_ack_op(STRIPE_OP_COMPUTE_BLK
, pending
);
372 test_and_ack_op(STRIPE_OP_PREXOR
, pending
);
373 test_and_ack_op(STRIPE_OP_BIODRAIN
, pending
);
374 test_and_ack_op(STRIPE_OP_POSTXOR
, pending
);
375 test_and_ack_op(STRIPE_OP_CHECK
, pending
);
376 if (test_and_clear_bit(STRIPE_OP_IO
, &sh
->ops
.pending
))
379 sh
->ops
.count
-= ack
;
380 if (unlikely(sh
->ops
.count
< 0)) {
381 printk(KERN_ERR
"pending: %#lx ops.pending: %#lx ops.ack: %#lx "
382 "ops.complete: %#lx\n", pending
, sh
->ops
.pending
,
383 sh
->ops
.ack
, sh
->ops
.complete
);
391 raid5_end_read_request(struct bio
*bi
, int error
);
393 raid5_end_write_request(struct bio
*bi
, int error
);
395 static void ops_run_io(struct stripe_head
*sh
)
397 raid5_conf_t
*conf
= sh
->raid_conf
;
398 int i
, disks
= sh
->disks
;
402 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
403 for (i
= disks
; i
--; ) {
407 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
))
409 else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
414 bi
= &sh
->dev
[i
].req
;
418 bi
->bi_end_io
= raid5_end_write_request
;
420 bi
->bi_end_io
= raid5_end_read_request
;
423 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
424 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
427 atomic_inc(&rdev
->nr_pending
);
431 if (test_bit(STRIPE_SYNCING
, &sh
->state
) ||
432 test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
) ||
433 test_bit(STRIPE_EXPAND_READY
, &sh
->state
))
434 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
436 bi
->bi_bdev
= rdev
->bdev
;
437 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
438 __func__
, (unsigned long long)sh
->sector
,
440 atomic_inc(&sh
->count
);
441 bi
->bi_sector
= sh
->sector
+ rdev
->data_offset
;
442 bi
->bi_flags
= 1 << BIO_UPTODATE
;
446 bi
->bi_io_vec
= &sh
->dev
[i
].vec
;
447 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
448 bi
->bi_io_vec
[0].bv_offset
= 0;
449 bi
->bi_size
= STRIPE_SIZE
;
452 test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
453 atomic_add(STRIPE_SECTORS
,
454 &rdev
->corrected_errors
);
455 generic_make_request(bi
);
458 set_bit(STRIPE_DEGRADED
, &sh
->state
);
459 pr_debug("skip op %ld on disc %d for sector %llu\n",
460 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
461 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
462 set_bit(STRIPE_HANDLE
, &sh
->state
);
467 static struct dma_async_tx_descriptor
*
468 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
469 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
472 struct page
*bio_page
;
476 if (bio
->bi_sector
>= sector
)
477 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
479 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
480 bio_for_each_segment(bvl
, bio
, i
) {
481 int len
= bio_iovec_idx(bio
, i
)->bv_len
;
485 if (page_offset
< 0) {
486 b_offset
= -page_offset
;
487 page_offset
+= b_offset
;
491 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
492 clen
= STRIPE_SIZE
- page_offset
;
497 b_offset
+= bio_iovec_idx(bio
, i
)->bv_offset
;
498 bio_page
= bio_iovec_idx(bio
, i
)->bv_page
;
500 tx
= async_memcpy(page
, bio_page
, page_offset
,
505 tx
= async_memcpy(bio_page
, page
, b_offset
,
510 if (clen
< len
) /* hit end of page */
518 static void ops_complete_biofill(void *stripe_head_ref
)
520 struct stripe_head
*sh
= stripe_head_ref
;
521 struct bio
*return_bi
= NULL
;
522 raid5_conf_t
*conf
= sh
->raid_conf
;
525 pr_debug("%s: stripe %llu\n", __func__
,
526 (unsigned long long)sh
->sector
);
528 /* clear completed biofills */
529 for (i
= sh
->disks
; i
--; ) {
530 struct r5dev
*dev
= &sh
->dev
[i
];
532 /* acknowledge completion of a biofill operation */
533 /* and check if we need to reply to a read request,
534 * new R5_Wantfill requests are held off until
535 * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)
537 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
538 struct bio
*rbi
, *rbi2
;
540 /* The access to dev->read is outside of the
541 * spin_lock_irq(&conf->device_lock), but is protected
542 * by the STRIPE_OP_BIOFILL pending bit
547 while (rbi
&& rbi
->bi_sector
<
548 dev
->sector
+ STRIPE_SECTORS
) {
549 rbi2
= r5_next_bio(rbi
, dev
->sector
);
550 spin_lock_irq(&conf
->device_lock
);
551 if (--rbi
->bi_phys_segments
== 0) {
552 rbi
->bi_next
= return_bi
;
555 spin_unlock_irq(&conf
->device_lock
);
560 set_bit(STRIPE_OP_BIOFILL
, &sh
->ops
.complete
);
562 return_io(return_bi
);
564 set_bit(STRIPE_HANDLE
, &sh
->state
);
568 static void ops_run_biofill(struct stripe_head
*sh
)
570 struct dma_async_tx_descriptor
*tx
= NULL
;
571 raid5_conf_t
*conf
= sh
->raid_conf
;
574 pr_debug("%s: stripe %llu\n", __func__
,
575 (unsigned long long)sh
->sector
);
577 for (i
= sh
->disks
; i
--; ) {
578 struct r5dev
*dev
= &sh
->dev
[i
];
579 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
581 spin_lock_irq(&conf
->device_lock
);
582 dev
->read
= rbi
= dev
->toread
;
584 spin_unlock_irq(&conf
->device_lock
);
585 while (rbi
&& rbi
->bi_sector
<
586 dev
->sector
+ STRIPE_SECTORS
) {
587 tx
= async_copy_data(0, rbi
, dev
->page
,
589 rbi
= r5_next_bio(rbi
, dev
->sector
);
594 atomic_inc(&sh
->count
);
595 async_trigger_callback(ASYNC_TX_DEP_ACK
| ASYNC_TX_ACK
, tx
,
596 ops_complete_biofill
, sh
);
599 static void ops_complete_compute5(void *stripe_head_ref
)
601 struct stripe_head
*sh
= stripe_head_ref
;
602 int target
= sh
->ops
.target
;
603 struct r5dev
*tgt
= &sh
->dev
[target
];
605 pr_debug("%s: stripe %llu\n", __func__
,
606 (unsigned long long)sh
->sector
);
608 set_bit(R5_UPTODATE
, &tgt
->flags
);
609 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
610 clear_bit(R5_Wantcompute
, &tgt
->flags
);
611 set_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.complete
);
612 set_bit(STRIPE_HANDLE
, &sh
->state
);
616 static struct dma_async_tx_descriptor
*
617 ops_run_compute5(struct stripe_head
*sh
, unsigned long pending
)
619 /* kernel stack size limits the total number of disks */
620 int disks
= sh
->disks
;
621 struct page
*xor_srcs
[disks
];
622 int target
= sh
->ops
.target
;
623 struct r5dev
*tgt
= &sh
->dev
[target
];
624 struct page
*xor_dest
= tgt
->page
;
626 struct dma_async_tx_descriptor
*tx
;
629 pr_debug("%s: stripe %llu block: %d\n",
630 __func__
, (unsigned long long)sh
->sector
, target
);
631 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
633 for (i
= disks
; i
--; )
635 xor_srcs
[count
++] = sh
->dev
[i
].page
;
637 atomic_inc(&sh
->count
);
639 if (unlikely(count
== 1))
640 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
,
641 0, NULL
, ops_complete_compute5
, sh
);
643 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
644 ASYNC_TX_XOR_ZERO_DST
, NULL
,
645 ops_complete_compute5
, sh
);
647 /* ack now if postxor is not set to be run */
648 if (tx
&& !test_bit(STRIPE_OP_POSTXOR
, &pending
))
654 static void ops_complete_prexor(void *stripe_head_ref
)
656 struct stripe_head
*sh
= stripe_head_ref
;
658 pr_debug("%s: stripe %llu\n", __func__
,
659 (unsigned long long)sh
->sector
);
661 set_bit(STRIPE_OP_PREXOR
, &sh
->ops
.complete
);
664 static struct dma_async_tx_descriptor
*
665 ops_run_prexor(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
667 /* kernel stack size limits the total number of disks */
668 int disks
= sh
->disks
;
669 struct page
*xor_srcs
[disks
];
670 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
672 /* existing parity data subtracted */
673 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
675 pr_debug("%s: stripe %llu\n", __func__
,
676 (unsigned long long)sh
->sector
);
678 for (i
= disks
; i
--; ) {
679 struct r5dev
*dev
= &sh
->dev
[i
];
680 /* Only process blocks that are known to be uptodate */
681 if (dev
->towrite
&& test_bit(R5_Wantprexor
, &dev
->flags
))
682 xor_srcs
[count
++] = dev
->page
;
685 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
686 ASYNC_TX_DEP_ACK
| ASYNC_TX_XOR_DROP_DST
, tx
,
687 ops_complete_prexor
, sh
);
692 static struct dma_async_tx_descriptor
*
693 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
,
694 unsigned long pending
)
696 int disks
= sh
->disks
;
697 int pd_idx
= sh
->pd_idx
, i
;
699 /* check if prexor is active which means only process blocks
700 * that are part of a read-modify-write (Wantprexor)
702 int prexor
= test_bit(STRIPE_OP_PREXOR
, &pending
);
704 pr_debug("%s: stripe %llu\n", __func__
,
705 (unsigned long long)sh
->sector
);
707 for (i
= disks
; i
--; ) {
708 struct r5dev
*dev
= &sh
->dev
[i
];
713 if (prexor
) { /* rmw */
715 test_bit(R5_Wantprexor
, &dev
->flags
))
718 if (i
!= pd_idx
&& dev
->towrite
&&
719 test_bit(R5_LOCKED
, &dev
->flags
))
726 spin_lock(&sh
->lock
);
727 chosen
= dev
->towrite
;
729 BUG_ON(dev
->written
);
730 wbi
= dev
->written
= chosen
;
731 spin_unlock(&sh
->lock
);
733 while (wbi
&& wbi
->bi_sector
<
734 dev
->sector
+ STRIPE_SECTORS
) {
735 tx
= async_copy_data(1, wbi
, dev
->page
,
737 wbi
= r5_next_bio(wbi
, dev
->sector
);
745 static void ops_complete_postxor(void *stripe_head_ref
)
747 struct stripe_head
*sh
= stripe_head_ref
;
749 pr_debug("%s: stripe %llu\n", __func__
,
750 (unsigned long long)sh
->sector
);
752 set_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.complete
);
753 set_bit(STRIPE_HANDLE
, &sh
->state
);
757 static void ops_complete_write(void *stripe_head_ref
)
759 struct stripe_head
*sh
= stripe_head_ref
;
760 int disks
= sh
->disks
, i
, pd_idx
= sh
->pd_idx
;
762 pr_debug("%s: stripe %llu\n", __func__
,
763 (unsigned long long)sh
->sector
);
765 for (i
= disks
; i
--; ) {
766 struct r5dev
*dev
= &sh
->dev
[i
];
767 if (dev
->written
|| i
== pd_idx
)
768 set_bit(R5_UPTODATE
, &dev
->flags
);
771 set_bit(STRIPE_OP_BIODRAIN
, &sh
->ops
.complete
);
772 set_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.complete
);
774 set_bit(STRIPE_HANDLE
, &sh
->state
);
779 ops_run_postxor(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
,
780 unsigned long pending
)
782 /* kernel stack size limits the total number of disks */
783 int disks
= sh
->disks
;
784 struct page
*xor_srcs
[disks
];
786 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
787 struct page
*xor_dest
;
788 int prexor
= test_bit(STRIPE_OP_PREXOR
, &pending
);
790 dma_async_tx_callback callback
;
792 pr_debug("%s: stripe %llu\n", __func__
,
793 (unsigned long long)sh
->sector
);
795 /* check if prexor is active which means only process blocks
796 * that are part of a read-modify-write (written)
799 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
800 for (i
= disks
; i
--; ) {
801 struct r5dev
*dev
= &sh
->dev
[i
];
803 xor_srcs
[count
++] = dev
->page
;
806 xor_dest
= sh
->dev
[pd_idx
].page
;
807 for (i
= disks
; i
--; ) {
808 struct r5dev
*dev
= &sh
->dev
[i
];
810 xor_srcs
[count
++] = dev
->page
;
814 /* check whether this postxor is part of a write */
815 callback
= test_bit(STRIPE_OP_BIODRAIN
, &pending
) ?
816 ops_complete_write
: ops_complete_postxor
;
818 /* 1/ if we prexor'd then the dest is reused as a source
819 * 2/ if we did not prexor then we are redoing the parity
820 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
821 * for the synchronous xor case
823 flags
= ASYNC_TX_DEP_ACK
| ASYNC_TX_ACK
|
824 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
826 atomic_inc(&sh
->count
);
828 if (unlikely(count
== 1)) {
829 flags
&= ~(ASYNC_TX_XOR_DROP_DST
| ASYNC_TX_XOR_ZERO_DST
);
830 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
,
831 flags
, tx
, callback
, sh
);
833 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
834 flags
, tx
, callback
, sh
);
837 static void ops_complete_check(void *stripe_head_ref
)
839 struct stripe_head
*sh
= stripe_head_ref
;
840 int pd_idx
= sh
->pd_idx
;
842 pr_debug("%s: stripe %llu\n", __func__
,
843 (unsigned long long)sh
->sector
);
845 if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK
, &sh
->ops
.pending
) &&
846 sh
->ops
.zero_sum_result
== 0)
847 set_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
849 set_bit(STRIPE_OP_CHECK
, &sh
->ops
.complete
);
850 set_bit(STRIPE_HANDLE
, &sh
->state
);
854 static void ops_run_check(struct stripe_head
*sh
)
856 /* kernel stack size limits the total number of disks */
857 int disks
= sh
->disks
;
858 struct page
*xor_srcs
[disks
];
859 struct dma_async_tx_descriptor
*tx
;
861 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
862 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
864 pr_debug("%s: stripe %llu\n", __func__
,
865 (unsigned long long)sh
->sector
);
867 for (i
= disks
; i
--; ) {
868 struct r5dev
*dev
= &sh
->dev
[i
];
870 xor_srcs
[count
++] = dev
->page
;
873 tx
= async_xor_zero_sum(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
874 &sh
->ops
.zero_sum_result
, 0, NULL
, NULL
, NULL
);
877 set_bit(STRIPE_OP_MOD_DMA_CHECK
, &sh
->ops
.pending
);
879 clear_bit(STRIPE_OP_MOD_DMA_CHECK
, &sh
->ops
.pending
);
881 atomic_inc(&sh
->count
);
882 tx
= async_trigger_callback(ASYNC_TX_DEP_ACK
| ASYNC_TX_ACK
, tx
,
883 ops_complete_check
, sh
);
886 static void raid5_run_ops(struct stripe_head
*sh
, unsigned long pending
)
888 int overlap_clear
= 0, i
, disks
= sh
->disks
;
889 struct dma_async_tx_descriptor
*tx
= NULL
;
891 if (test_bit(STRIPE_OP_BIOFILL
, &pending
)) {
896 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &pending
))
897 tx
= ops_run_compute5(sh
, pending
);
899 if (test_bit(STRIPE_OP_PREXOR
, &pending
))
900 tx
= ops_run_prexor(sh
, tx
);
902 if (test_bit(STRIPE_OP_BIODRAIN
, &pending
)) {
903 tx
= ops_run_biodrain(sh
, tx
, pending
);
907 if (test_bit(STRIPE_OP_POSTXOR
, &pending
))
908 ops_run_postxor(sh
, tx
, pending
);
910 if (test_bit(STRIPE_OP_CHECK
, &pending
))
913 if (test_bit(STRIPE_OP_IO
, &pending
))
917 for (i
= disks
; i
--; ) {
918 struct r5dev
*dev
= &sh
->dev
[i
];
919 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
920 wake_up(&sh
->raid_conf
->wait_for_overlap
);
924 static int grow_one_stripe(raid5_conf_t
*conf
)
926 struct stripe_head
*sh
;
927 sh
= kmem_cache_alloc(conf
->slab_cache
, GFP_KERNEL
);
930 memset(sh
, 0, sizeof(*sh
) + (conf
->raid_disks
-1)*sizeof(struct r5dev
));
931 sh
->raid_conf
= conf
;
932 spin_lock_init(&sh
->lock
);
934 if (grow_buffers(sh
, conf
->raid_disks
)) {
935 shrink_buffers(sh
, conf
->raid_disks
);
936 kmem_cache_free(conf
->slab_cache
, sh
);
939 sh
->disks
= conf
->raid_disks
;
940 /* we just created an active stripe so... */
941 atomic_set(&sh
->count
, 1);
942 atomic_inc(&conf
->active_stripes
);
943 INIT_LIST_HEAD(&sh
->lru
);
948 static int grow_stripes(raid5_conf_t
*conf
, int num
)
950 struct kmem_cache
*sc
;
951 int devs
= conf
->raid_disks
;
953 sprintf(conf
->cache_name
[0], "raid5-%s", mdname(conf
->mddev
));
954 sprintf(conf
->cache_name
[1], "raid5-%s-alt", mdname(conf
->mddev
));
955 conf
->active_name
= 0;
956 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
957 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
961 conf
->slab_cache
= sc
;
962 conf
->pool_size
= devs
;
964 if (!grow_one_stripe(conf
))
969 #ifdef CONFIG_MD_RAID5_RESHAPE
970 static int resize_stripes(raid5_conf_t
*conf
, int newsize
)
972 /* Make all the stripes able to hold 'newsize' devices.
973 * New slots in each stripe get 'page' set to a new page.
975 * This happens in stages:
976 * 1/ create a new kmem_cache and allocate the required number of
978 * 2/ gather all the old stripe_heads and tranfer the pages across
979 * to the new stripe_heads. This will have the side effect of
980 * freezing the array as once all stripe_heads have been collected,
981 * no IO will be possible. Old stripe heads are freed once their
982 * pages have been transferred over, and the old kmem_cache is
983 * freed when all stripes are done.
984 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
985 * we simple return a failre status - no need to clean anything up.
986 * 4/ allocate new pages for the new slots in the new stripe_heads.
987 * If this fails, we don't bother trying the shrink the
988 * stripe_heads down again, we just leave them as they are.
989 * As each stripe_head is processed the new one is released into
992 * Once step2 is started, we cannot afford to wait for a write,
993 * so we use GFP_NOIO allocations.
995 struct stripe_head
*osh
, *nsh
;
996 LIST_HEAD(newstripes
);
997 struct disk_info
*ndisks
;
999 struct kmem_cache
*sc
;
1002 if (newsize
<= conf
->pool_size
)
1003 return 0; /* never bother to shrink */
1005 md_allow_write(conf
->mddev
);
1008 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
1009 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
1014 for (i
= conf
->max_nr_stripes
; i
; i
--) {
1015 nsh
= kmem_cache_alloc(sc
, GFP_KERNEL
);
1019 memset(nsh
, 0, sizeof(*nsh
) + (newsize
-1)*sizeof(struct r5dev
));
1021 nsh
->raid_conf
= conf
;
1022 spin_lock_init(&nsh
->lock
);
1024 list_add(&nsh
->lru
, &newstripes
);
1027 /* didn't get enough, give up */
1028 while (!list_empty(&newstripes
)) {
1029 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1030 list_del(&nsh
->lru
);
1031 kmem_cache_free(sc
, nsh
);
1033 kmem_cache_destroy(sc
);
1036 /* Step 2 - Must use GFP_NOIO now.
1037 * OK, we have enough stripes, start collecting inactive
1038 * stripes and copying them over
1040 list_for_each_entry(nsh
, &newstripes
, lru
) {
1041 spin_lock_irq(&conf
->device_lock
);
1042 wait_event_lock_irq(conf
->wait_for_stripe
,
1043 !list_empty(&conf
->inactive_list
),
1045 unplug_slaves(conf
->mddev
)
1047 osh
= get_free_stripe(conf
);
1048 spin_unlock_irq(&conf
->device_lock
);
1049 atomic_set(&nsh
->count
, 1);
1050 for(i
=0; i
<conf
->pool_size
; i
++)
1051 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1052 for( ; i
<newsize
; i
++)
1053 nsh
->dev
[i
].page
= NULL
;
1054 kmem_cache_free(conf
->slab_cache
, osh
);
1056 kmem_cache_destroy(conf
->slab_cache
);
1059 * At this point, we are holding all the stripes so the array
1060 * is completely stalled, so now is a good time to resize
1063 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1065 for (i
=0; i
<conf
->raid_disks
; i
++)
1066 ndisks
[i
] = conf
->disks
[i
];
1068 conf
->disks
= ndisks
;
1072 /* Step 4, return new stripes to service */
1073 while(!list_empty(&newstripes
)) {
1074 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1075 list_del_init(&nsh
->lru
);
1076 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1077 if (nsh
->dev
[i
].page
== NULL
) {
1078 struct page
*p
= alloc_page(GFP_NOIO
);
1079 nsh
->dev
[i
].page
= p
;
1083 release_stripe(nsh
);
1085 /* critical section pass, GFP_NOIO no longer needed */
1087 conf
->slab_cache
= sc
;
1088 conf
->active_name
= 1-conf
->active_name
;
1089 conf
->pool_size
= newsize
;
1094 static int drop_one_stripe(raid5_conf_t
*conf
)
1096 struct stripe_head
*sh
;
1098 spin_lock_irq(&conf
->device_lock
);
1099 sh
= get_free_stripe(conf
);
1100 spin_unlock_irq(&conf
->device_lock
);
1103 BUG_ON(atomic_read(&sh
->count
));
1104 shrink_buffers(sh
, conf
->pool_size
);
1105 kmem_cache_free(conf
->slab_cache
, sh
);
1106 atomic_dec(&conf
->active_stripes
);
1110 static void shrink_stripes(raid5_conf_t
*conf
)
1112 while (drop_one_stripe(conf
))
1115 if (conf
->slab_cache
)
1116 kmem_cache_destroy(conf
->slab_cache
);
1117 conf
->slab_cache
= NULL
;
1120 static void raid5_end_read_request(struct bio
* bi
, int error
)
1122 struct stripe_head
*sh
= bi
->bi_private
;
1123 raid5_conf_t
*conf
= sh
->raid_conf
;
1124 int disks
= sh
->disks
, i
;
1125 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1126 char b
[BDEVNAME_SIZE
];
1130 for (i
=0 ; i
<disks
; i
++)
1131 if (bi
== &sh
->dev
[i
].req
)
1134 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1135 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1143 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1144 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1145 rdev
= conf
->disks
[i
].rdev
;
1146 printk_rl(KERN_INFO
"raid5:%s: read error corrected"
1147 " (%lu sectors at %llu on %s)\n",
1148 mdname(conf
->mddev
), STRIPE_SECTORS
,
1149 (unsigned long long)(sh
->sector
1150 + rdev
->data_offset
),
1151 bdevname(rdev
->bdev
, b
));
1152 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1153 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1155 if (atomic_read(&conf
->disks
[i
].rdev
->read_errors
))
1156 atomic_set(&conf
->disks
[i
].rdev
->read_errors
, 0);
1158 const char *bdn
= bdevname(conf
->disks
[i
].rdev
->bdev
, b
);
1160 rdev
= conf
->disks
[i
].rdev
;
1162 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1163 atomic_inc(&rdev
->read_errors
);
1164 if (conf
->mddev
->degraded
)
1165 printk_rl(KERN_WARNING
1166 "raid5:%s: read error not correctable "
1167 "(sector %llu on %s).\n",
1168 mdname(conf
->mddev
),
1169 (unsigned long long)(sh
->sector
1170 + rdev
->data_offset
),
1172 else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
1174 printk_rl(KERN_WARNING
1175 "raid5:%s: read error NOT corrected!! "
1176 "(sector %llu on %s).\n",
1177 mdname(conf
->mddev
),
1178 (unsigned long long)(sh
->sector
1179 + rdev
->data_offset
),
1181 else if (atomic_read(&rdev
->read_errors
)
1182 > conf
->max_nr_stripes
)
1184 "raid5:%s: Too many read errors, failing device %s.\n",
1185 mdname(conf
->mddev
), bdn
);
1189 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1191 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1192 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1193 md_error(conf
->mddev
, rdev
);
1196 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1197 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1198 set_bit(STRIPE_HANDLE
, &sh
->state
);
1202 static void raid5_end_write_request (struct bio
*bi
, int error
)
1204 struct stripe_head
*sh
= bi
->bi_private
;
1205 raid5_conf_t
*conf
= sh
->raid_conf
;
1206 int disks
= sh
->disks
, i
;
1207 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1209 for (i
=0 ; i
<disks
; i
++)
1210 if (bi
== &sh
->dev
[i
].req
)
1213 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1214 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1222 md_error(conf
->mddev
, conf
->disks
[i
].rdev
);
1224 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1226 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1227 set_bit(STRIPE_HANDLE
, &sh
->state
);
1232 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
);
1234 static void raid5_build_block (struct stripe_head
*sh
, int i
)
1236 struct r5dev
*dev
= &sh
->dev
[i
];
1238 bio_init(&dev
->req
);
1239 dev
->req
.bi_io_vec
= &dev
->vec
;
1241 dev
->req
.bi_max_vecs
++;
1242 dev
->vec
.bv_page
= dev
->page
;
1243 dev
->vec
.bv_len
= STRIPE_SIZE
;
1244 dev
->vec
.bv_offset
= 0;
1246 dev
->req
.bi_sector
= sh
->sector
;
1247 dev
->req
.bi_private
= sh
;
1250 dev
->sector
= compute_blocknr(sh
, i
);
1253 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1255 char b
[BDEVNAME_SIZE
];
1256 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
1257 pr_debug("raid5: error called\n");
1259 if (!test_bit(Faulty
, &rdev
->flags
)) {
1260 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1261 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1262 unsigned long flags
;
1263 spin_lock_irqsave(&conf
->device_lock
, flags
);
1265 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1267 * if recovery was running, make sure it aborts.
1269 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1271 set_bit(Faulty
, &rdev
->flags
);
1273 "raid5: Disk failure on %s, disabling device.\n"
1274 "raid5: Operation continuing on %d devices.\n",
1275 bdevname(rdev
->bdev
,b
), conf
->raid_disks
- mddev
->degraded
);
1280 * Input: a 'big' sector number,
1281 * Output: index of the data and parity disk, and the sector # in them.
1283 static sector_t
raid5_compute_sector(sector_t r_sector
, unsigned int raid_disks
,
1284 unsigned int data_disks
, unsigned int * dd_idx
,
1285 unsigned int * pd_idx
, raid5_conf_t
*conf
)
1288 unsigned long chunk_number
;
1289 unsigned int chunk_offset
;
1290 sector_t new_sector
;
1291 int sectors_per_chunk
= conf
->chunk_size
>> 9;
1293 /* First compute the information on this sector */
1296 * Compute the chunk number and the sector offset inside the chunk
1298 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1299 chunk_number
= r_sector
;
1300 BUG_ON(r_sector
!= chunk_number
);
1303 * Compute the stripe number
1305 stripe
= chunk_number
/ data_disks
;
1308 * Compute the data disk and parity disk indexes inside the stripe
1310 *dd_idx
= chunk_number
% data_disks
;
1313 * Select the parity disk based on the user selected algorithm.
1315 switch(conf
->level
) {
1317 *pd_idx
= data_disks
;
1320 switch (conf
->algorithm
) {
1321 case ALGORITHM_LEFT_ASYMMETRIC
:
1322 *pd_idx
= data_disks
- stripe
% raid_disks
;
1323 if (*dd_idx
>= *pd_idx
)
1326 case ALGORITHM_RIGHT_ASYMMETRIC
:
1327 *pd_idx
= stripe
% raid_disks
;
1328 if (*dd_idx
>= *pd_idx
)
1331 case ALGORITHM_LEFT_SYMMETRIC
:
1332 *pd_idx
= data_disks
- stripe
% raid_disks
;
1333 *dd_idx
= (*pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1335 case ALGORITHM_RIGHT_SYMMETRIC
:
1336 *pd_idx
= stripe
% raid_disks
;
1337 *dd_idx
= (*pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1340 printk(KERN_ERR
"raid5: unsupported algorithm %d\n",
1346 /**** FIX THIS ****/
1347 switch (conf
->algorithm
) {
1348 case ALGORITHM_LEFT_ASYMMETRIC
:
1349 *pd_idx
= raid_disks
- 1 - (stripe
% raid_disks
);
1350 if (*pd_idx
== raid_disks
-1)
1351 (*dd_idx
)++; /* Q D D D P */
1352 else if (*dd_idx
>= *pd_idx
)
1353 (*dd_idx
) += 2; /* D D P Q D */
1355 case ALGORITHM_RIGHT_ASYMMETRIC
:
1356 *pd_idx
= stripe
% raid_disks
;
1357 if (*pd_idx
== raid_disks
-1)
1358 (*dd_idx
)++; /* Q D D D P */
1359 else if (*dd_idx
>= *pd_idx
)
1360 (*dd_idx
) += 2; /* D D P Q D */
1362 case ALGORITHM_LEFT_SYMMETRIC
:
1363 *pd_idx
= raid_disks
- 1 - (stripe
% raid_disks
);
1364 *dd_idx
= (*pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1366 case ALGORITHM_RIGHT_SYMMETRIC
:
1367 *pd_idx
= stripe
% raid_disks
;
1368 *dd_idx
= (*pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1371 printk (KERN_CRIT
"raid6: unsupported algorithm %d\n",
1378 * Finally, compute the new sector number
1380 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
1385 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
)
1387 raid5_conf_t
*conf
= sh
->raid_conf
;
1388 int raid_disks
= sh
->disks
;
1389 int data_disks
= raid_disks
- conf
->max_degraded
;
1390 sector_t new_sector
= sh
->sector
, check
;
1391 int sectors_per_chunk
= conf
->chunk_size
>> 9;
1394 int chunk_number
, dummy1
, dummy2
, dd_idx
= i
;
1398 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
1399 stripe
= new_sector
;
1400 BUG_ON(new_sector
!= stripe
);
1402 if (i
== sh
->pd_idx
)
1404 switch(conf
->level
) {
1407 switch (conf
->algorithm
) {
1408 case ALGORITHM_LEFT_ASYMMETRIC
:
1409 case ALGORITHM_RIGHT_ASYMMETRIC
:
1413 case ALGORITHM_LEFT_SYMMETRIC
:
1414 case ALGORITHM_RIGHT_SYMMETRIC
:
1417 i
-= (sh
->pd_idx
+ 1);
1420 printk(KERN_ERR
"raid5: unsupported algorithm %d\n",
1425 if (i
== raid6_next_disk(sh
->pd_idx
, raid_disks
))
1426 return 0; /* It is the Q disk */
1427 switch (conf
->algorithm
) {
1428 case ALGORITHM_LEFT_ASYMMETRIC
:
1429 case ALGORITHM_RIGHT_ASYMMETRIC
:
1430 if (sh
->pd_idx
== raid_disks
-1)
1431 i
--; /* Q D D D P */
1432 else if (i
> sh
->pd_idx
)
1433 i
-= 2; /* D D P Q D */
1435 case ALGORITHM_LEFT_SYMMETRIC
:
1436 case ALGORITHM_RIGHT_SYMMETRIC
:
1437 if (sh
->pd_idx
== raid_disks
-1)
1438 i
--; /* Q D D D P */
1443 i
-= (sh
->pd_idx
+ 2);
1447 printk (KERN_CRIT
"raid6: unsupported algorithm %d\n",
1453 chunk_number
= stripe
* data_disks
+ i
;
1454 r_sector
= (sector_t
)chunk_number
* sectors_per_chunk
+ chunk_offset
;
1456 check
= raid5_compute_sector (r_sector
, raid_disks
, data_disks
, &dummy1
, &dummy2
, conf
);
1457 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| dummy2
!= sh
->pd_idx
) {
1458 printk(KERN_ERR
"compute_blocknr: map not correct\n");
1467 * Copy data between a page in the stripe cache, and one or more bion
1468 * The page could align with the middle of the bio, or there could be
1469 * several bion, each with several bio_vecs, which cover part of the page
1470 * Multiple bion are linked together on bi_next. There may be extras
1471 * at the end of this list. We ignore them.
1473 static void copy_data(int frombio
, struct bio
*bio
,
1477 char *pa
= page_address(page
);
1478 struct bio_vec
*bvl
;
1482 if (bio
->bi_sector
>= sector
)
1483 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
1485 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
1486 bio_for_each_segment(bvl
, bio
, i
) {
1487 int len
= bio_iovec_idx(bio
,i
)->bv_len
;
1491 if (page_offset
< 0) {
1492 b_offset
= -page_offset
;
1493 page_offset
+= b_offset
;
1497 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1498 clen
= STRIPE_SIZE
- page_offset
;
1502 char *ba
= __bio_kmap_atomic(bio
, i
, KM_USER0
);
1504 memcpy(pa
+page_offset
, ba
+b_offset
, clen
);
1506 memcpy(ba
+b_offset
, pa
+page_offset
, clen
);
1507 __bio_kunmap_atomic(ba
, KM_USER0
);
1509 if (clen
< len
) /* hit end of page */
1515 #define check_xor() do { \
1516 if (count == MAX_XOR_BLOCKS) { \
1517 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1522 static void compute_parity6(struct stripe_head
*sh
, int method
)
1524 raid6_conf_t
*conf
= sh
->raid_conf
;
1525 int i
, pd_idx
= sh
->pd_idx
, qd_idx
, d0_idx
, disks
= sh
->disks
, count
;
1527 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1530 qd_idx
= raid6_next_disk(pd_idx
, disks
);
1531 d0_idx
= raid6_next_disk(qd_idx
, disks
);
1533 pr_debug("compute_parity, stripe %llu, method %d\n",
1534 (unsigned long long)sh
->sector
, method
);
1537 case READ_MODIFY_WRITE
:
1538 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1539 case RECONSTRUCT_WRITE
:
1540 for (i
= disks
; i
-- ;)
1541 if ( i
!= pd_idx
&& i
!= qd_idx
&& sh
->dev
[i
].towrite
) {
1542 chosen
= sh
->dev
[i
].towrite
;
1543 sh
->dev
[i
].towrite
= NULL
;
1545 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
1546 wake_up(&conf
->wait_for_overlap
);
1548 BUG_ON(sh
->dev
[i
].written
);
1549 sh
->dev
[i
].written
= chosen
;
1553 BUG(); /* Not implemented yet */
1556 for (i
= disks
; i
--;)
1557 if (sh
->dev
[i
].written
) {
1558 sector_t sector
= sh
->dev
[i
].sector
;
1559 struct bio
*wbi
= sh
->dev
[i
].written
;
1560 while (wbi
&& wbi
->bi_sector
< sector
+ STRIPE_SECTORS
) {
1561 copy_data(1, wbi
, sh
->dev
[i
].page
, sector
);
1562 wbi
= r5_next_bio(wbi
, sector
);
1565 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1566 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1570 // case RECONSTRUCT_WRITE:
1571 // case CHECK_PARITY:
1572 // case UPDATE_PARITY:
1573 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1574 /* FIX: Is this ordering of drives even remotely optimal? */
1578 ptrs
[count
++] = page_address(sh
->dev
[i
].page
);
1579 if (count
<= disks
-2 && !test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
))
1580 printk("block %d/%d not uptodate on parity calc\n", i
,count
);
1581 i
= raid6_next_disk(i
, disks
);
1582 } while ( i
!= d0_idx
);
1586 raid6_call
.gen_syndrome(disks
, STRIPE_SIZE
, ptrs
);
1589 case RECONSTRUCT_WRITE
:
1590 set_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
1591 set_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
);
1592 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
1593 set_bit(R5_LOCKED
, &sh
->dev
[qd_idx
].flags
);
1596 set_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
1597 set_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
);
1603 /* Compute one missing block */
1604 static void compute_block_1(struct stripe_head
*sh
, int dd_idx
, int nozero
)
1606 int i
, count
, disks
= sh
->disks
;
1607 void *ptr
[MAX_XOR_BLOCKS
], *dest
, *p
;
1608 int pd_idx
= sh
->pd_idx
;
1609 int qd_idx
= raid6_next_disk(pd_idx
, disks
);
1611 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1612 (unsigned long long)sh
->sector
, dd_idx
);
1614 if ( dd_idx
== qd_idx
) {
1615 /* We're actually computing the Q drive */
1616 compute_parity6(sh
, UPDATE_PARITY
);
1618 dest
= page_address(sh
->dev
[dd_idx
].page
);
1619 if (!nozero
) memset(dest
, 0, STRIPE_SIZE
);
1621 for (i
= disks
; i
--; ) {
1622 if (i
== dd_idx
|| i
== qd_idx
)
1624 p
= page_address(sh
->dev
[i
].page
);
1625 if (test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
))
1628 printk("compute_block() %d, stripe %llu, %d"
1629 " not present\n", dd_idx
,
1630 (unsigned long long)sh
->sector
, i
);
1635 xor_blocks(count
, STRIPE_SIZE
, dest
, ptr
);
1636 if (!nozero
) set_bit(R5_UPTODATE
, &sh
->dev
[dd_idx
].flags
);
1637 else clear_bit(R5_UPTODATE
, &sh
->dev
[dd_idx
].flags
);
1641 /* Compute two missing blocks */
1642 static void compute_block_2(struct stripe_head
*sh
, int dd_idx1
, int dd_idx2
)
1644 int i
, count
, disks
= sh
->disks
;
1645 int pd_idx
= sh
->pd_idx
;
1646 int qd_idx
= raid6_next_disk(pd_idx
, disks
);
1647 int d0_idx
= raid6_next_disk(qd_idx
, disks
);
1650 /* faila and failb are disk numbers relative to d0_idx */
1651 /* pd_idx become disks-2 and qd_idx become disks-1 */
1652 faila
= (dd_idx1
< d0_idx
) ? dd_idx1
+(disks
-d0_idx
) : dd_idx1
-d0_idx
;
1653 failb
= (dd_idx2
< d0_idx
) ? dd_idx2
+(disks
-d0_idx
) : dd_idx2
-d0_idx
;
1655 BUG_ON(faila
== failb
);
1656 if ( failb
< faila
) { int tmp
= faila
; faila
= failb
; failb
= tmp
; }
1658 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1659 (unsigned long long)sh
->sector
, dd_idx1
, dd_idx2
, faila
, failb
);
1661 if ( failb
== disks
-1 ) {
1662 /* Q disk is one of the missing disks */
1663 if ( faila
== disks
-2 ) {
1664 /* Missing P+Q, just recompute */
1665 compute_parity6(sh
, UPDATE_PARITY
);
1668 /* We're missing D+Q; recompute D from P */
1669 compute_block_1(sh
, (dd_idx1
== qd_idx
) ? dd_idx2
: dd_idx1
, 0);
1670 compute_parity6(sh
, UPDATE_PARITY
); /* Is this necessary? */
1675 /* We're missing D+P or D+D; build pointer table */
1677 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1683 ptrs
[count
++] = page_address(sh
->dev
[i
].page
);
1684 i
= raid6_next_disk(i
, disks
);
1685 if (i
!= dd_idx1
&& i
!= dd_idx2
&&
1686 !test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
))
1687 printk("compute_2 with missing block %d/%d\n", count
, i
);
1688 } while ( i
!= d0_idx
);
1690 if ( failb
== disks
-2 ) {
1691 /* We're missing D+P. */
1692 raid6_datap_recov(disks
, STRIPE_SIZE
, faila
, ptrs
);
1694 /* We're missing D+D. */
1695 raid6_2data_recov(disks
, STRIPE_SIZE
, faila
, failb
, ptrs
);
1698 /* Both the above update both missing blocks */
1699 set_bit(R5_UPTODATE
, &sh
->dev
[dd_idx1
].flags
);
1700 set_bit(R5_UPTODATE
, &sh
->dev
[dd_idx2
].flags
);
1705 handle_write_operations5(struct stripe_head
*sh
, int rcw
, int expand
)
1707 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
1711 /* if we are not expanding this is a proper write request, and
1712 * there will be bios with new data to be drained into the
1716 set_bit(STRIPE_OP_BIODRAIN
, &sh
->ops
.pending
);
1720 set_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.pending
);
1723 for (i
= disks
; i
--; ) {
1724 struct r5dev
*dev
= &sh
->dev
[i
];
1727 set_bit(R5_LOCKED
, &dev
->flags
);
1729 clear_bit(R5_UPTODATE
, &dev
->flags
);
1733 if (locked
+ 1 == disks
)
1734 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
1735 atomic_inc(&sh
->raid_conf
->pending_full_writes
);
1737 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
1738 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
1740 set_bit(STRIPE_OP_PREXOR
, &sh
->ops
.pending
);
1741 set_bit(STRIPE_OP_BIODRAIN
, &sh
->ops
.pending
);
1742 set_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.pending
);
1746 for (i
= disks
; i
--; ) {
1747 struct r5dev
*dev
= &sh
->dev
[i
];
1751 /* For a read-modify write there may be blocks that are
1752 * locked for reading while others are ready to be
1753 * written so we distinguish these blocks by the
1757 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
1758 test_bit(R5_Wantcompute
, &dev
->flags
))) {
1759 set_bit(R5_Wantprexor
, &dev
->flags
);
1760 set_bit(R5_LOCKED
, &dev
->flags
);
1761 clear_bit(R5_UPTODATE
, &dev
->flags
);
1767 /* keep the parity disk locked while asynchronous operations
1770 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
1771 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
1774 pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1775 __func__
, (unsigned long long)sh
->sector
,
1776 locked
, sh
->ops
.pending
);
1782 * Each stripe/dev can have one or more bion attached.
1783 * toread/towrite point to the first in a chain.
1784 * The bi_next chain must be in order.
1786 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
1789 raid5_conf_t
*conf
= sh
->raid_conf
;
1792 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1793 (unsigned long long)bi
->bi_sector
,
1794 (unsigned long long)sh
->sector
);
1797 spin_lock(&sh
->lock
);
1798 spin_lock_irq(&conf
->device_lock
);
1800 bip
= &sh
->dev
[dd_idx
].towrite
;
1801 if (*bip
== NULL
&& sh
->dev
[dd_idx
].written
== NULL
)
1804 bip
= &sh
->dev
[dd_idx
].toread
;
1805 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
1806 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
1808 bip
= & (*bip
)->bi_next
;
1810 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
1813 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
1817 bi
->bi_phys_segments
++;
1818 spin_unlock_irq(&conf
->device_lock
);
1819 spin_unlock(&sh
->lock
);
1821 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1822 (unsigned long long)bi
->bi_sector
,
1823 (unsigned long long)sh
->sector
, dd_idx
);
1825 if (conf
->mddev
->bitmap
&& firstwrite
) {
1826 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
1828 sh
->bm_seq
= conf
->seq_flush
+1;
1829 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
1833 /* check if page is covered */
1834 sector_t sector
= sh
->dev
[dd_idx
].sector
;
1835 for (bi
=sh
->dev
[dd_idx
].towrite
;
1836 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
1837 bi
&& bi
->bi_sector
<= sector
;
1838 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
1839 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
1840 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
1842 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
1843 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
1848 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
1849 spin_unlock_irq(&conf
->device_lock
);
1850 spin_unlock(&sh
->lock
);
1854 static void end_reshape(raid5_conf_t
*conf
);
1856 static int page_is_zero(struct page
*p
)
1858 char *a
= page_address(p
);
1859 return ((*(u32
*)a
) == 0 &&
1860 memcmp(a
, a
+4, STRIPE_SIZE
-4)==0);
1863 static int stripe_to_pdidx(sector_t stripe
, raid5_conf_t
*conf
, int disks
)
1865 int sectors_per_chunk
= conf
->chunk_size
>> 9;
1867 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
1869 raid5_compute_sector(stripe
* (disks
- conf
->max_degraded
)
1870 *sectors_per_chunk
+ chunk_offset
,
1871 disks
, disks
- conf
->max_degraded
,
1872 &dd_idx
, &pd_idx
, conf
);
1877 handle_requests_to_failed_array(raid5_conf_t
*conf
, struct stripe_head
*sh
,
1878 struct stripe_head_state
*s
, int disks
,
1879 struct bio
**return_bi
)
1882 for (i
= disks
; i
--; ) {
1886 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1889 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
1890 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
1891 /* multiple read failures in one stripe */
1892 md_error(conf
->mddev
, rdev
);
1895 spin_lock_irq(&conf
->device_lock
);
1896 /* fail all writes first */
1897 bi
= sh
->dev
[i
].towrite
;
1898 sh
->dev
[i
].towrite
= NULL
;
1904 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
1905 wake_up(&conf
->wait_for_overlap
);
1907 while (bi
&& bi
->bi_sector
<
1908 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
1909 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
1910 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1911 if (--bi
->bi_phys_segments
== 0) {
1912 md_write_end(conf
->mddev
);
1913 bi
->bi_next
= *return_bi
;
1918 /* and fail all 'written' */
1919 bi
= sh
->dev
[i
].written
;
1920 sh
->dev
[i
].written
= NULL
;
1921 if (bi
) bitmap_end
= 1;
1922 while (bi
&& bi
->bi_sector
<
1923 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
1924 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
1925 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1926 if (--bi
->bi_phys_segments
== 0) {
1927 md_write_end(conf
->mddev
);
1928 bi
->bi_next
= *return_bi
;
1934 /* fail any reads if this device is non-operational and
1935 * the data has not reached the cache yet.
1937 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
1938 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
1939 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
1940 bi
= sh
->dev
[i
].toread
;
1941 sh
->dev
[i
].toread
= NULL
;
1942 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
1943 wake_up(&conf
->wait_for_overlap
);
1944 if (bi
) s
->to_read
--;
1945 while (bi
&& bi
->bi_sector
<
1946 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
1947 struct bio
*nextbi
=
1948 r5_next_bio(bi
, sh
->dev
[i
].sector
);
1949 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1950 if (--bi
->bi_phys_segments
== 0) {
1951 bi
->bi_next
= *return_bi
;
1957 spin_unlock_irq(&conf
->device_lock
);
1959 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
1960 STRIPE_SECTORS
, 0, 0);
1963 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
1964 if (atomic_dec_and_test(&conf
->pending_full_writes
))
1965 md_wakeup_thread(conf
->mddev
->thread
);
1968 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1971 static int __handle_issuing_new_read_requests5(struct stripe_head
*sh
,
1972 struct stripe_head_state
*s
, int disk_idx
, int disks
)
1974 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
1975 struct r5dev
*failed_dev
= &sh
->dev
[s
->failed_num
];
1977 /* don't schedule compute operations or reads on the parity block while
1978 * a check is in flight
1980 if ((disk_idx
== sh
->pd_idx
) &&
1981 test_bit(STRIPE_OP_CHECK
, &sh
->ops
.pending
))
1984 /* is the data in this block needed, and can we get it? */
1985 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
1986 !test_bit(R5_UPTODATE
, &dev
->flags
) && (dev
->toread
||
1987 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
1988 s
->syncing
|| s
->expanding
|| (s
->failed
&&
1989 (failed_dev
->toread
|| (failed_dev
->towrite
&&
1990 !test_bit(R5_OVERWRITE
, &failed_dev
->flags
)
1992 /* 1/ We would like to get this block, possibly by computing it,
1993 * but we might not be able to.
1995 * 2/ Since parity check operations potentially make the parity
1996 * block !uptodate it will need to be refreshed before any
1997 * compute operations on data disks are scheduled.
1999 * 3/ We hold off parity block re-reads until check operations
2002 if ((s
->uptodate
== disks
- 1) &&
2003 (s
->failed
&& disk_idx
== s
->failed_num
) &&
2004 !test_bit(STRIPE_OP_CHECK
, &sh
->ops
.pending
)) {
2005 set_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.pending
);
2006 set_bit(R5_Wantcompute
, &dev
->flags
);
2007 sh
->ops
.target
= disk_idx
;
2010 /* Careful: from this point on 'uptodate' is in the eye
2011 * of raid5_run_ops which services 'compute' operations
2012 * before writes. R5_Wantcompute flags a block that will
2013 * be R5_UPTODATE by the time it is needed for a
2014 * subsequent operation.
2017 return 0; /* uptodate + compute == disks */
2018 } else if ((s
->uptodate
< disks
- 1) &&
2019 test_bit(R5_Insync
, &dev
->flags
)) {
2020 /* Note: we hold off compute operations while checks are
2021 * in flight, but we still prefer 'compute' over 'read'
2022 * hence we only read if (uptodate < * disks-1)
2024 set_bit(R5_LOCKED
, &dev
->flags
);
2025 set_bit(R5_Wantread
, &dev
->flags
);
2026 if (!test_and_set_bit(STRIPE_OP_IO
, &sh
->ops
.pending
))
2029 pr_debug("Reading block %d (sync=%d)\n", disk_idx
,
2037 static void handle_issuing_new_read_requests5(struct stripe_head
*sh
,
2038 struct stripe_head_state
*s
, int disks
)
2042 /* Clear completed compute operations. Parity recovery
2043 * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2044 * later on in this routine
2046 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.complete
) &&
2047 !test_bit(STRIPE_OP_MOD_REPAIR_PD
, &sh
->ops
.pending
)) {
2048 clear_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.complete
);
2049 clear_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.ack
);
2050 clear_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.pending
);
2053 /* look for blocks to read/compute, skip this if a compute
2054 * is already in flight, or if the stripe contents are in the
2055 * midst of changing due to a write
2057 if (!test_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.pending
) &&
2058 !test_bit(STRIPE_OP_PREXOR
, &sh
->ops
.pending
) &&
2059 !test_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.pending
)) {
2060 for (i
= disks
; i
--; )
2061 if (__handle_issuing_new_read_requests5(
2062 sh
, s
, i
, disks
) == 0)
2065 set_bit(STRIPE_HANDLE
, &sh
->state
);
2068 static void handle_issuing_new_read_requests6(struct stripe_head
*sh
,
2069 struct stripe_head_state
*s
, struct r6_state
*r6s
,
2073 for (i
= disks
; i
--; ) {
2074 struct r5dev
*dev
= &sh
->dev
[i
];
2075 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2076 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2077 (dev
->toread
|| (dev
->towrite
&&
2078 !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2079 s
->syncing
|| s
->expanding
||
2081 (sh
->dev
[r6s
->failed_num
[0]].toread
||
2084 (sh
->dev
[r6s
->failed_num
[1]].toread
||
2086 /* we would like to get this block, possibly
2087 * by computing it, but we might not be able to
2089 if ((s
->uptodate
== disks
- 1) &&
2090 (s
->failed
&& (i
== r6s
->failed_num
[0] ||
2091 i
== r6s
->failed_num
[1]))) {
2092 pr_debug("Computing stripe %llu block %d\n",
2093 (unsigned long long)sh
->sector
, i
);
2094 compute_block_1(sh
, i
, 0);
2096 } else if ( s
->uptodate
== disks
-2 && s
->failed
>= 2 ) {
2097 /* Computing 2-failure is *very* expensive; only
2098 * do it if failed >= 2
2101 for (other
= disks
; other
--; ) {
2104 if (!test_bit(R5_UPTODATE
,
2105 &sh
->dev
[other
].flags
))
2109 pr_debug("Computing stripe %llu blocks %d,%d\n",
2110 (unsigned long long)sh
->sector
,
2112 compute_block_2(sh
, i
, other
);
2114 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2115 set_bit(R5_LOCKED
, &dev
->flags
);
2116 set_bit(R5_Wantread
, &dev
->flags
);
2118 pr_debug("Reading block %d (sync=%d)\n",
2123 set_bit(STRIPE_HANDLE
, &sh
->state
);
2127 /* handle_completed_write_requests
2128 * any written block on an uptodate or failed drive can be returned.
2129 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2130 * never LOCKED, so we don't need to test 'failed' directly.
2132 static void handle_completed_write_requests(raid5_conf_t
*conf
,
2133 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2138 for (i
= disks
; i
--; )
2139 if (sh
->dev
[i
].written
) {
2141 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2142 test_bit(R5_UPTODATE
, &dev
->flags
)) {
2143 /* We can return any write requests */
2144 struct bio
*wbi
, *wbi2
;
2146 pr_debug("Return write for disc %d\n", i
);
2147 spin_lock_irq(&conf
->device_lock
);
2149 dev
->written
= NULL
;
2150 while (wbi
&& wbi
->bi_sector
<
2151 dev
->sector
+ STRIPE_SECTORS
) {
2152 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2153 if (--wbi
->bi_phys_segments
== 0) {
2154 md_write_end(conf
->mddev
);
2155 wbi
->bi_next
= *return_bi
;
2160 if (dev
->towrite
== NULL
)
2162 spin_unlock_irq(&conf
->device_lock
);
2164 bitmap_endwrite(conf
->mddev
->bitmap
,
2167 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2172 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2173 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2174 md_wakeup_thread(conf
->mddev
->thread
);
2177 static void handle_issuing_new_write_requests5(raid5_conf_t
*conf
,
2178 struct stripe_head
*sh
, struct stripe_head_state
*s
, int disks
)
2180 int rmw
= 0, rcw
= 0, i
;
2181 for (i
= disks
; i
--; ) {
2182 /* would I have to read this buffer for read_modify_write */
2183 struct r5dev
*dev
= &sh
->dev
[i
];
2184 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2185 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2186 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2187 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2188 if (test_bit(R5_Insync
, &dev
->flags
))
2191 rmw
+= 2*disks
; /* cannot read it */
2193 /* Would I have to read this buffer for reconstruct_write */
2194 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2195 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2196 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2197 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2198 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2203 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2204 (unsigned long long)sh
->sector
, rmw
, rcw
);
2205 set_bit(STRIPE_HANDLE
, &sh
->state
);
2206 if (rmw
< rcw
&& rmw
> 0)
2207 /* prefer read-modify-write, but need to get some data */
2208 for (i
= disks
; i
--; ) {
2209 struct r5dev
*dev
= &sh
->dev
[i
];
2210 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2211 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2212 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2213 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2214 test_bit(R5_Insync
, &dev
->flags
)) {
2216 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2217 pr_debug("Read_old block "
2218 "%d for r-m-w\n", i
);
2219 set_bit(R5_LOCKED
, &dev
->flags
);
2220 set_bit(R5_Wantread
, &dev
->flags
);
2221 if (!test_and_set_bit(
2222 STRIPE_OP_IO
, &sh
->ops
.pending
))
2226 set_bit(STRIPE_DELAYED
, &sh
->state
);
2227 set_bit(STRIPE_HANDLE
, &sh
->state
);
2231 if (rcw
<= rmw
&& rcw
> 0)
2232 /* want reconstruct write, but need to get some data */
2233 for (i
= disks
; i
--; ) {
2234 struct r5dev
*dev
= &sh
->dev
[i
];
2235 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2237 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2238 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2239 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2240 test_bit(R5_Insync
, &dev
->flags
)) {
2242 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2243 pr_debug("Read_old block "
2244 "%d for Reconstruct\n", i
);
2245 set_bit(R5_LOCKED
, &dev
->flags
);
2246 set_bit(R5_Wantread
, &dev
->flags
);
2247 if (!test_and_set_bit(
2248 STRIPE_OP_IO
, &sh
->ops
.pending
))
2252 set_bit(STRIPE_DELAYED
, &sh
->state
);
2253 set_bit(STRIPE_HANDLE
, &sh
->state
);
2257 /* now if nothing is locked, and if we have enough data,
2258 * we can start a write request
2260 /* since handle_stripe can be called at any time we need to handle the
2261 * case where a compute block operation has been submitted and then a
2262 * subsequent call wants to start a write request. raid5_run_ops only
2263 * handles the case where compute block and postxor are requested
2264 * simultaneously. If this is not the case then new writes need to be
2265 * held off until the compute completes.
2267 if ((s
->req_compute
||
2268 !test_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.pending
)) &&
2269 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2270 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2271 s
->locked
+= handle_write_operations5(sh
, rcw
== 0, 0);
2274 static void handle_issuing_new_write_requests6(raid5_conf_t
*conf
,
2275 struct stripe_head
*sh
, struct stripe_head_state
*s
,
2276 struct r6_state
*r6s
, int disks
)
2278 int rcw
= 0, must_compute
= 0, pd_idx
= sh
->pd_idx
, i
;
2279 int qd_idx
= r6s
->qd_idx
;
2280 for (i
= disks
; i
--; ) {
2281 struct r5dev
*dev
= &sh
->dev
[i
];
2282 /* Would I have to read this buffer for reconstruct_write */
2283 if (!test_bit(R5_OVERWRITE
, &dev
->flags
)
2284 && i
!= pd_idx
&& i
!= qd_idx
2285 && (!test_bit(R5_LOCKED
, &dev
->flags
)
2287 !test_bit(R5_UPTODATE
, &dev
->flags
)) {
2288 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2290 pr_debug("raid6: must_compute: "
2291 "disk %d flags=%#lx\n", i
, dev
->flags
);
2296 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2297 (unsigned long long)sh
->sector
, rcw
, must_compute
);
2298 set_bit(STRIPE_HANDLE
, &sh
->state
);
2301 /* want reconstruct write, but need to get some data */
2302 for (i
= disks
; i
--; ) {
2303 struct r5dev
*dev
= &sh
->dev
[i
];
2304 if (!test_bit(R5_OVERWRITE
, &dev
->flags
)
2305 && !(s
->failed
== 0 && (i
== pd_idx
|| i
== qd_idx
))
2306 && !test_bit(R5_LOCKED
, &dev
->flags
) &&
2307 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2308 test_bit(R5_Insync
, &dev
->flags
)) {
2310 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2311 pr_debug("Read_old stripe %llu "
2312 "block %d for Reconstruct\n",
2313 (unsigned long long)sh
->sector
, i
);
2314 set_bit(R5_LOCKED
, &dev
->flags
);
2315 set_bit(R5_Wantread
, &dev
->flags
);
2318 pr_debug("Request delayed stripe %llu "
2319 "block %d for Reconstruct\n",
2320 (unsigned long long)sh
->sector
, i
);
2321 set_bit(STRIPE_DELAYED
, &sh
->state
);
2322 set_bit(STRIPE_HANDLE
, &sh
->state
);
2326 /* now if nothing is locked, and if we have enough data, we can start a
2329 if (s
->locked
== 0 && rcw
== 0 &&
2330 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
2331 if (must_compute
> 0) {
2332 /* We have failed blocks and need to compute them */
2333 switch (s
->failed
) {
2337 compute_block_1(sh
, r6s
->failed_num
[0], 0);
2340 compute_block_2(sh
, r6s
->failed_num
[0],
2341 r6s
->failed_num
[1]);
2343 default: /* This request should have been failed? */
2348 pr_debug("Computing parity for stripe %llu\n",
2349 (unsigned long long)sh
->sector
);
2350 compute_parity6(sh
, RECONSTRUCT_WRITE
);
2351 /* now every locked buffer is ready to be written */
2352 for (i
= disks
; i
--; )
2353 if (test_bit(R5_LOCKED
, &sh
->dev
[i
].flags
)) {
2354 pr_debug("Writing stripe %llu block %d\n",
2355 (unsigned long long)sh
->sector
, i
);
2357 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
2359 if (s
->locked
== disks
)
2360 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2361 atomic_inc(&conf
->pending_full_writes
);
2362 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2363 set_bit(STRIPE_INSYNC
, &sh
->state
);
2365 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2366 atomic_dec(&conf
->preread_active_stripes
);
2367 if (atomic_read(&conf
->preread_active_stripes
) <
2369 md_wakeup_thread(conf
->mddev
->thread
);
2374 static void handle_parity_checks5(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2375 struct stripe_head_state
*s
, int disks
)
2377 int canceled_check
= 0;
2379 set_bit(STRIPE_HANDLE
, &sh
->state
);
2381 /* complete a check operation */
2382 if (test_and_clear_bit(STRIPE_OP_CHECK
, &sh
->ops
.complete
)) {
2383 clear_bit(STRIPE_OP_CHECK
, &sh
->ops
.ack
);
2384 clear_bit(STRIPE_OP_CHECK
, &sh
->ops
.pending
);
2385 if (s
->failed
== 0) {
2386 if (sh
->ops
.zero_sum_result
== 0)
2387 /* parity is correct (on disc,
2388 * not in buffer any more)
2390 set_bit(STRIPE_INSYNC
, &sh
->state
);
2392 conf
->mddev
->resync_mismatches
+=
2395 MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2396 /* don't try to repair!! */
2397 set_bit(STRIPE_INSYNC
, &sh
->state
);
2399 set_bit(STRIPE_OP_COMPUTE_BLK
,
2401 set_bit(STRIPE_OP_MOD_REPAIR_PD
,
2403 set_bit(R5_Wantcompute
,
2404 &sh
->dev
[sh
->pd_idx
].flags
);
2405 sh
->ops
.target
= sh
->pd_idx
;
2411 canceled_check
= 1; /* STRIPE_INSYNC is not set */
2414 /* start a new check operation if there are no failures, the stripe is
2415 * not insync, and a repair is not in flight
2417 if (s
->failed
== 0 &&
2418 !test_bit(STRIPE_INSYNC
, &sh
->state
) &&
2419 !test_bit(STRIPE_OP_MOD_REPAIR_PD
, &sh
->ops
.pending
)) {
2420 if (!test_and_set_bit(STRIPE_OP_CHECK
, &sh
->ops
.pending
)) {
2421 BUG_ON(s
->uptodate
!= disks
);
2422 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2428 /* check if we can clear a parity disk reconstruct */
2429 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.complete
) &&
2430 test_bit(STRIPE_OP_MOD_REPAIR_PD
, &sh
->ops
.pending
)) {
2432 clear_bit(STRIPE_OP_MOD_REPAIR_PD
, &sh
->ops
.pending
);
2433 clear_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.complete
);
2434 clear_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.ack
);
2435 clear_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.pending
);
2439 /* Wait for check parity and compute block operations to complete
2440 * before write-back. If a failure occurred while the check operation
2441 * was in flight we need to cycle this stripe through handle_stripe
2442 * since the parity block may not be uptodate
2444 if (!canceled_check
&& !test_bit(STRIPE_INSYNC
, &sh
->state
) &&
2445 !test_bit(STRIPE_OP_CHECK
, &sh
->ops
.pending
) &&
2446 !test_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.pending
)) {
2448 /* either failed parity check, or recovery is happening */
2450 s
->failed_num
= sh
->pd_idx
;
2451 dev
= &sh
->dev
[s
->failed_num
];
2452 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2453 BUG_ON(s
->uptodate
!= disks
);
2455 set_bit(R5_LOCKED
, &dev
->flags
);
2456 set_bit(R5_Wantwrite
, &dev
->flags
);
2457 if (!test_and_set_bit(STRIPE_OP_IO
, &sh
->ops
.pending
))
2460 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2462 set_bit(STRIPE_INSYNC
, &sh
->state
);
2467 static void handle_parity_checks6(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2468 struct stripe_head_state
*s
,
2469 struct r6_state
*r6s
, struct page
*tmp_page
,
2472 int update_p
= 0, update_q
= 0;
2474 int pd_idx
= sh
->pd_idx
;
2475 int qd_idx
= r6s
->qd_idx
;
2477 set_bit(STRIPE_HANDLE
, &sh
->state
);
2479 BUG_ON(s
->failed
> 2);
2480 BUG_ON(s
->uptodate
< disks
);
2481 /* Want to check and possibly repair P and Q.
2482 * However there could be one 'failed' device, in which
2483 * case we can only check one of them, possibly using the
2484 * other to generate missing data
2487 /* If !tmp_page, we cannot do the calculations,
2488 * but as we have set STRIPE_HANDLE, we will soon be called
2489 * by stripe_handle with a tmp_page - just wait until then.
2492 if (s
->failed
== r6s
->q_failed
) {
2493 /* The only possible failed device holds 'Q', so it
2494 * makes sense to check P (If anything else were failed,
2495 * we would have used P to recreate it).
2497 compute_block_1(sh
, pd_idx
, 1);
2498 if (!page_is_zero(sh
->dev
[pd_idx
].page
)) {
2499 compute_block_1(sh
, pd_idx
, 0);
2503 if (!r6s
->q_failed
&& s
->failed
< 2) {
2504 /* q is not failed, and we didn't use it to generate
2505 * anything, so it makes sense to check it
2507 memcpy(page_address(tmp_page
),
2508 page_address(sh
->dev
[qd_idx
].page
),
2510 compute_parity6(sh
, UPDATE_PARITY
);
2511 if (memcmp(page_address(tmp_page
),
2512 page_address(sh
->dev
[qd_idx
].page
),
2513 STRIPE_SIZE
) != 0) {
2514 clear_bit(STRIPE_INSYNC
, &sh
->state
);
2518 if (update_p
|| update_q
) {
2519 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2520 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2521 /* don't try to repair!! */
2522 update_p
= update_q
= 0;
2525 /* now write out any block on a failed drive,
2526 * or P or Q if they need it
2529 if (s
->failed
== 2) {
2530 dev
= &sh
->dev
[r6s
->failed_num
[1]];
2532 set_bit(R5_LOCKED
, &dev
->flags
);
2533 set_bit(R5_Wantwrite
, &dev
->flags
);
2535 if (s
->failed
>= 1) {
2536 dev
= &sh
->dev
[r6s
->failed_num
[0]];
2538 set_bit(R5_LOCKED
, &dev
->flags
);
2539 set_bit(R5_Wantwrite
, &dev
->flags
);
2543 dev
= &sh
->dev
[pd_idx
];
2545 set_bit(R5_LOCKED
, &dev
->flags
);
2546 set_bit(R5_Wantwrite
, &dev
->flags
);
2549 dev
= &sh
->dev
[qd_idx
];
2551 set_bit(R5_LOCKED
, &dev
->flags
);
2552 set_bit(R5_Wantwrite
, &dev
->flags
);
2554 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2556 set_bit(STRIPE_INSYNC
, &sh
->state
);
2560 static void handle_stripe_expansion(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2561 struct r6_state
*r6s
)
2565 /* We have read all the blocks in this stripe and now we need to
2566 * copy some of them into a target stripe for expand.
2568 struct dma_async_tx_descriptor
*tx
= NULL
;
2569 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2570 for (i
= 0; i
< sh
->disks
; i
++)
2571 if (i
!= sh
->pd_idx
&& (!r6s
|| i
!= r6s
->qd_idx
)) {
2572 int dd_idx
, pd_idx
, j
;
2573 struct stripe_head
*sh2
;
2575 sector_t bn
= compute_blocknr(sh
, i
);
2576 sector_t s
= raid5_compute_sector(bn
, conf
->raid_disks
,
2578 conf
->max_degraded
, &dd_idx
,
2580 sh2
= get_active_stripe(conf
, s
, conf
->raid_disks
,
2583 /* so far only the early blocks of this stripe
2584 * have been requested. When later blocks
2585 * get requested, we will try again
2588 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
2589 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
2590 /* must have already done this block */
2591 release_stripe(sh2
);
2595 /* place all the copies on one channel */
2596 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
2597 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
2598 ASYNC_TX_DEP_ACK
, tx
, NULL
, NULL
);
2600 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
2601 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
2602 for (j
= 0; j
< conf
->raid_disks
; j
++)
2603 if (j
!= sh2
->pd_idx
&&
2604 (!r6s
|| j
!= raid6_next_disk(sh2
->pd_idx
,
2606 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
2608 if (j
== conf
->raid_disks
) {
2609 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
2610 set_bit(STRIPE_HANDLE
, &sh2
->state
);
2612 release_stripe(sh2
);
2615 /* done submitting copies, wait for them to complete */
2618 dma_wait_for_async_tx(tx
);
2624 * handle_stripe - do things to a stripe.
2626 * We lock the stripe and then examine the state of various bits
2627 * to see what needs to be done.
2629 * return some read request which now have data
2630 * return some write requests which are safely on disc
2631 * schedule a read on some buffers
2632 * schedule a write of some buffers
2633 * return confirmation of parity correctness
2635 * buffers are taken off read_list or write_list, and bh_cache buffers
2636 * get BH_Lock set before the stripe lock is released.
2640 static void handle_stripe5(struct stripe_head
*sh
)
2642 raid5_conf_t
*conf
= sh
->raid_conf
;
2643 int disks
= sh
->disks
, i
;
2644 struct bio
*return_bi
= NULL
;
2645 struct stripe_head_state s
;
2647 unsigned long pending
= 0;
2648 mdk_rdev_t
*blocked_rdev
= NULL
;
2651 memset(&s
, 0, sizeof(s
));
2652 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2653 "ops=%lx:%lx:%lx\n", (unsigned long long)sh
->sector
, sh
->state
,
2654 atomic_read(&sh
->count
), sh
->pd_idx
,
2655 sh
->ops
.pending
, sh
->ops
.ack
, sh
->ops
.complete
);
2657 spin_lock(&sh
->lock
);
2658 clear_bit(STRIPE_HANDLE
, &sh
->state
);
2659 clear_bit(STRIPE_DELAYED
, &sh
->state
);
2661 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
2662 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2663 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
2664 /* Now to look around and see what can be done */
2666 /* clean-up completed biofill operations */
2667 if (test_bit(STRIPE_OP_BIOFILL
, &sh
->ops
.complete
)) {
2668 clear_bit(STRIPE_OP_BIOFILL
, &sh
->ops
.pending
);
2669 clear_bit(STRIPE_OP_BIOFILL
, &sh
->ops
.ack
);
2670 clear_bit(STRIPE_OP_BIOFILL
, &sh
->ops
.complete
);
2674 for (i
=disks
; i
--; ) {
2676 struct r5dev
*dev
= &sh
->dev
[i
];
2677 clear_bit(R5_Insync
, &dev
->flags
);
2679 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2680 "written %p\n", i
, dev
->flags
, dev
->toread
, dev
->read
,
2681 dev
->towrite
, dev
->written
);
2683 /* maybe we can request a biofill operation
2685 * new wantfill requests are only permitted while
2686 * STRIPE_OP_BIOFILL is clear
2688 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
2689 !test_bit(STRIPE_OP_BIOFILL
, &sh
->ops
.pending
))
2690 set_bit(R5_Wantfill
, &dev
->flags
);
2692 /* now count some things */
2693 if (test_bit(R5_LOCKED
, &dev
->flags
)) s
.locked
++;
2694 if (test_bit(R5_UPTODATE
, &dev
->flags
)) s
.uptodate
++;
2695 if (test_bit(R5_Wantcompute
, &dev
->flags
)) s
.compute
++;
2697 if (test_bit(R5_Wantfill
, &dev
->flags
))
2699 else if (dev
->toread
)
2703 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
2708 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2709 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
2710 blocked_rdev
= rdev
;
2711 atomic_inc(&rdev
->nr_pending
);
2714 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)) {
2715 /* The ReadError flag will just be confusing now */
2716 clear_bit(R5_ReadError
, &dev
->flags
);
2717 clear_bit(R5_ReWrite
, &dev
->flags
);
2719 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)
2720 || test_bit(R5_ReadError
, &dev
->flags
)) {
2724 set_bit(R5_Insync
, &dev
->flags
);
2728 if (unlikely(blocked_rdev
)) {
2729 set_bit(STRIPE_HANDLE
, &sh
->state
);
2733 if (s
.to_fill
&& !test_and_set_bit(STRIPE_OP_BIOFILL
, &sh
->ops
.pending
))
2736 pr_debug("locked=%d uptodate=%d to_read=%d"
2737 " to_write=%d failed=%d failed_num=%d\n",
2738 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
,
2739 s
.failed
, s
.failed_num
);
2740 /* check if the array has lost two devices and, if so, some requests might
2743 if (s
.failed
> 1 && s
.to_read
+s
.to_write
+s
.written
)
2744 handle_requests_to_failed_array(conf
, sh
, &s
, disks
,
2746 if (s
.failed
> 1 && s
.syncing
) {
2747 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
2748 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2752 /* might be able to return some write requests if the parity block
2753 * is safe, or on a failed drive
2755 dev
= &sh
->dev
[sh
->pd_idx
];
2757 ((test_bit(R5_Insync
, &dev
->flags
) &&
2758 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2759 test_bit(R5_UPTODATE
, &dev
->flags
)) ||
2760 (s
.failed
== 1 && s
.failed_num
== sh
->pd_idx
)))
2761 handle_completed_write_requests(conf
, sh
, disks
, &return_bi
);
2763 /* Now we might consider reading some blocks, either to check/generate
2764 * parity, or to satisfy requests
2765 * or to load a block that is being partially written.
2767 if (s
.to_read
|| s
.non_overwrite
||
2768 (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
)) || s
.expanding
||
2769 test_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.pending
))
2770 handle_issuing_new_read_requests5(sh
, &s
, disks
);
2772 /* Now we check to see if any write operations have recently
2776 /* leave prexor set until postxor is done, allows us to distinguish
2777 * a rmw from a rcw during biodrain
2780 if (test_bit(STRIPE_OP_PREXOR
, &sh
->ops
.complete
) &&
2781 test_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.complete
)) {
2784 clear_bit(STRIPE_OP_PREXOR
, &sh
->ops
.complete
);
2785 clear_bit(STRIPE_OP_PREXOR
, &sh
->ops
.ack
);
2786 clear_bit(STRIPE_OP_PREXOR
, &sh
->ops
.pending
);
2788 for (i
= disks
; i
--; )
2789 clear_bit(R5_Wantprexor
, &sh
->dev
[i
].flags
);
2792 /* if only POSTXOR is set then this is an 'expand' postxor */
2793 if (test_bit(STRIPE_OP_BIODRAIN
, &sh
->ops
.complete
) &&
2794 test_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.complete
)) {
2796 clear_bit(STRIPE_OP_BIODRAIN
, &sh
->ops
.complete
);
2797 clear_bit(STRIPE_OP_BIODRAIN
, &sh
->ops
.ack
);
2798 clear_bit(STRIPE_OP_BIODRAIN
, &sh
->ops
.pending
);
2800 clear_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.complete
);
2801 clear_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.ack
);
2802 clear_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.pending
);
2804 /* All the 'written' buffers and the parity block are ready to
2805 * be written back to disk
2807 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
2808 for (i
= disks
; i
--; ) {
2810 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
2811 (i
== sh
->pd_idx
|| dev
->written
)) {
2812 pr_debug("Writing block %d\n", i
);
2813 set_bit(R5_Wantwrite
, &dev
->flags
);
2814 if (!test_and_set_bit(
2815 STRIPE_OP_IO
, &sh
->ops
.pending
))
2819 if (!test_bit(R5_Insync
, &dev
->flags
) ||
2820 (i
== sh
->pd_idx
&& s
.failed
== 0))
2821 set_bit(STRIPE_INSYNC
, &sh
->state
);
2824 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2825 atomic_dec(&conf
->preread_active_stripes
);
2826 if (atomic_read(&conf
->preread_active_stripes
) <
2828 md_wakeup_thread(conf
->mddev
->thread
);
2832 /* Now to consider new write requests and what else, if anything
2833 * should be read. We do not handle new writes when:
2834 * 1/ A 'write' operation (copy+xor) is already in flight.
2835 * 2/ A 'check' operation is in flight, as it may clobber the parity
2838 if (s
.to_write
&& !test_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.pending
) &&
2839 !test_bit(STRIPE_OP_CHECK
, &sh
->ops
.pending
))
2840 handle_issuing_new_write_requests5(conf
, sh
, &s
, disks
);
2842 /* maybe we need to check and possibly fix the parity for this stripe
2843 * Any reads will already have been scheduled, so we just see if enough
2844 * data is available. The parity check is held off while parity
2845 * dependent operations are in flight.
2847 if ((s
.syncing
&& s
.locked
== 0 &&
2848 !test_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.pending
) &&
2849 !test_bit(STRIPE_INSYNC
, &sh
->state
)) ||
2850 test_bit(STRIPE_OP_CHECK
, &sh
->ops
.pending
) ||
2851 test_bit(STRIPE_OP_MOD_REPAIR_PD
, &sh
->ops
.pending
))
2852 handle_parity_checks5(conf
, sh
, &s
, disks
);
2854 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
2855 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,1);
2856 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2859 /* If the failed drive is just a ReadError, then we might need to progress
2860 * the repair/check process
2862 if (s
.failed
== 1 && !conf
->mddev
->ro
&&
2863 test_bit(R5_ReadError
, &sh
->dev
[s
.failed_num
].flags
)
2864 && !test_bit(R5_LOCKED
, &sh
->dev
[s
.failed_num
].flags
)
2865 && test_bit(R5_UPTODATE
, &sh
->dev
[s
.failed_num
].flags
)
2867 dev
= &sh
->dev
[s
.failed_num
];
2868 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
2869 set_bit(R5_Wantwrite
, &dev
->flags
);
2870 if (!test_and_set_bit(STRIPE_OP_IO
, &sh
->ops
.pending
))
2872 set_bit(R5_ReWrite
, &dev
->flags
);
2873 set_bit(R5_LOCKED
, &dev
->flags
);
2876 /* let's read it back */
2877 set_bit(R5_Wantread
, &dev
->flags
);
2878 if (!test_and_set_bit(STRIPE_OP_IO
, &sh
->ops
.pending
))
2880 set_bit(R5_LOCKED
, &dev
->flags
);
2885 /* Finish postxor operations initiated by the expansion
2888 if (test_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.complete
) &&
2889 !test_bit(STRIPE_OP_BIODRAIN
, &sh
->ops
.pending
)) {
2891 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
2893 clear_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.pending
);
2894 clear_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.ack
);
2895 clear_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.complete
);
2897 for (i
= conf
->raid_disks
; i
--; ) {
2898 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
2899 set_bit(R5_LOCKED
, &dev
->flags
);
2901 if (!test_and_set_bit(STRIPE_OP_IO
, &sh
->ops
.pending
))
2906 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
2907 !test_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.pending
)) {
2908 /* Need to write out all blocks after computing parity */
2909 sh
->disks
= conf
->raid_disks
;
2910 sh
->pd_idx
= stripe_to_pdidx(sh
->sector
, conf
,
2912 s
.locked
+= handle_write_operations5(sh
, 1, 1);
2913 } else if (s
.expanded
&&
2915 !test_bit(STRIPE_OP_POSTXOR
, &sh
->ops
.pending
)) {
2916 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
2917 atomic_dec(&conf
->reshape_stripes
);
2918 wake_up(&conf
->wait_for_overlap
);
2919 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
2922 if (s
.expanding
&& s
.locked
== 0 &&
2923 !test_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.pending
))
2924 handle_stripe_expansion(conf
, sh
, NULL
);
2927 pending
= get_stripe_work(sh
);
2930 spin_unlock(&sh
->lock
);
2932 /* wait for this device to become unblocked */
2933 if (unlikely(blocked_rdev
))
2934 md_wait_for_blocked_rdev(blocked_rdev
, conf
->mddev
);
2937 raid5_run_ops(sh
, pending
);
2939 return_io(return_bi
);
2943 static void handle_stripe6(struct stripe_head
*sh
, struct page
*tmp_page
)
2945 raid6_conf_t
*conf
= sh
->raid_conf
;
2946 int disks
= sh
->disks
;
2947 struct bio
*return_bi
= NULL
;
2948 int i
, pd_idx
= sh
->pd_idx
;
2949 struct stripe_head_state s
;
2950 struct r6_state r6s
;
2951 struct r5dev
*dev
, *pdev
, *qdev
;
2952 mdk_rdev_t
*blocked_rdev
= NULL
;
2954 r6s
.qd_idx
= raid6_next_disk(pd_idx
, disks
);
2955 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2956 "pd_idx=%d, qd_idx=%d\n",
2957 (unsigned long long)sh
->sector
, sh
->state
,
2958 atomic_read(&sh
->count
), pd_idx
, r6s
.qd_idx
);
2959 memset(&s
, 0, sizeof(s
));
2961 spin_lock(&sh
->lock
);
2962 clear_bit(STRIPE_HANDLE
, &sh
->state
);
2963 clear_bit(STRIPE_DELAYED
, &sh
->state
);
2965 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
2966 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2967 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
2968 /* Now to look around and see what can be done */
2971 for (i
=disks
; i
--; ) {
2974 clear_bit(R5_Insync
, &dev
->flags
);
2976 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2977 i
, dev
->flags
, dev
->toread
, dev
->towrite
, dev
->written
);
2978 /* maybe we can reply to a read */
2979 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
) {
2980 struct bio
*rbi
, *rbi2
;
2981 pr_debug("Return read for disc %d\n", i
);
2982 spin_lock_irq(&conf
->device_lock
);
2985 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
2986 wake_up(&conf
->wait_for_overlap
);
2987 spin_unlock_irq(&conf
->device_lock
);
2988 while (rbi
&& rbi
->bi_sector
< dev
->sector
+ STRIPE_SECTORS
) {
2989 copy_data(0, rbi
, dev
->page
, dev
->sector
);
2990 rbi2
= r5_next_bio(rbi
, dev
->sector
);
2991 spin_lock_irq(&conf
->device_lock
);
2992 if (--rbi
->bi_phys_segments
== 0) {
2993 rbi
->bi_next
= return_bi
;
2996 spin_unlock_irq(&conf
->device_lock
);
3001 /* now count some things */
3002 if (test_bit(R5_LOCKED
, &dev
->flags
)) s
.locked
++;
3003 if (test_bit(R5_UPTODATE
, &dev
->flags
)) s
.uptodate
++;
3010 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3015 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3016 if (rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
3017 blocked_rdev
= rdev
;
3018 atomic_inc(&rdev
->nr_pending
);
3021 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)) {
3022 /* The ReadError flag will just be confusing now */
3023 clear_bit(R5_ReadError
, &dev
->flags
);
3024 clear_bit(R5_ReWrite
, &dev
->flags
);
3026 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)
3027 || test_bit(R5_ReadError
, &dev
->flags
)) {
3029 r6s
.failed_num
[s
.failed
] = i
;
3032 set_bit(R5_Insync
, &dev
->flags
);
3036 if (unlikely(blocked_rdev
)) {
3037 set_bit(STRIPE_HANDLE
, &sh
->state
);
3040 pr_debug("locked=%d uptodate=%d to_read=%d"
3041 " to_write=%d failed=%d failed_num=%d,%d\n",
3042 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3043 r6s
.failed_num
[0], r6s
.failed_num
[1]);
3044 /* check if the array has lost >2 devices and, if so, some requests
3045 * might need to be failed
3047 if (s
.failed
> 2 && s
.to_read
+s
.to_write
+s
.written
)
3048 handle_requests_to_failed_array(conf
, sh
, &s
, disks
,
3050 if (s
.failed
> 2 && s
.syncing
) {
3051 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
3052 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3057 * might be able to return some write requests if the parity blocks
3058 * are safe, or on a failed drive
3060 pdev
= &sh
->dev
[pd_idx
];
3061 r6s
.p_failed
= (s
.failed
>= 1 && r6s
.failed_num
[0] == pd_idx
)
3062 || (s
.failed
>= 2 && r6s
.failed_num
[1] == pd_idx
);
3063 qdev
= &sh
->dev
[r6s
.qd_idx
];
3064 r6s
.q_failed
= (s
.failed
>= 1 && r6s
.failed_num
[0] == r6s
.qd_idx
)
3065 || (s
.failed
>= 2 && r6s
.failed_num
[1] == r6s
.qd_idx
);
3068 ( r6s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3069 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3070 && test_bit(R5_UPTODATE
, &pdev
->flags
)))) &&
3071 ( r6s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3072 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3073 && test_bit(R5_UPTODATE
, &qdev
->flags
)))))
3074 handle_completed_write_requests(conf
, sh
, disks
, &return_bi
);
3076 /* Now we might consider reading some blocks, either to check/generate
3077 * parity, or to satisfy requests
3078 * or to load a block that is being partially written.
3080 if (s
.to_read
|| s
.non_overwrite
|| (s
.to_write
&& s
.failed
) ||
3081 (s
.syncing
&& (s
.uptodate
< disks
)) || s
.expanding
)
3082 handle_issuing_new_read_requests6(sh
, &s
, &r6s
, disks
);
3084 /* now to consider writing and what else, if anything should be read */
3086 handle_issuing_new_write_requests6(conf
, sh
, &s
, &r6s
, disks
);
3088 /* maybe we need to check and possibly fix the parity for this stripe
3089 * Any reads will already have been scheduled, so we just see if enough
3092 if (s
.syncing
&& s
.locked
== 0 && !test_bit(STRIPE_INSYNC
, &sh
->state
))
3093 handle_parity_checks6(conf
, sh
, &s
, &r6s
, tmp_page
, disks
);
3095 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3096 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,1);
3097 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3100 /* If the failed drives are just a ReadError, then we might need
3101 * to progress the repair/check process
3103 if (s
.failed
<= 2 && !conf
->mddev
->ro
)
3104 for (i
= 0; i
< s
.failed
; i
++) {
3105 dev
= &sh
->dev
[r6s
.failed_num
[i
]];
3106 if (test_bit(R5_ReadError
, &dev
->flags
)
3107 && !test_bit(R5_LOCKED
, &dev
->flags
)
3108 && test_bit(R5_UPTODATE
, &dev
->flags
)
3110 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3111 set_bit(R5_Wantwrite
, &dev
->flags
);
3112 set_bit(R5_ReWrite
, &dev
->flags
);
3113 set_bit(R5_LOCKED
, &dev
->flags
);
3115 /* let's read it back */
3116 set_bit(R5_Wantread
, &dev
->flags
);
3117 set_bit(R5_LOCKED
, &dev
->flags
);
3122 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
3123 /* Need to write out all blocks after computing P&Q */
3124 sh
->disks
= conf
->raid_disks
;
3125 sh
->pd_idx
= stripe_to_pdidx(sh
->sector
, conf
,
3127 compute_parity6(sh
, RECONSTRUCT_WRITE
);
3128 for (i
= conf
->raid_disks
; i
-- ; ) {
3129 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3131 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3133 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3134 } else if (s
.expanded
) {
3135 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3136 atomic_dec(&conf
->reshape_stripes
);
3137 wake_up(&conf
->wait_for_overlap
);
3138 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3141 if (s
.expanding
&& s
.locked
== 0 &&
3142 !test_bit(STRIPE_OP_COMPUTE_BLK
, &sh
->ops
.pending
))
3143 handle_stripe_expansion(conf
, sh
, &r6s
);
3146 spin_unlock(&sh
->lock
);
3148 /* wait for this device to become unblocked */
3149 if (unlikely(blocked_rdev
))
3150 md_wait_for_blocked_rdev(blocked_rdev
, conf
->mddev
);
3152 return_io(return_bi
);
3154 for (i
=disks
; i
-- ;) {
3158 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
))
3160 else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
3165 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
3167 bi
= &sh
->dev
[i
].req
;
3171 bi
->bi_end_io
= raid5_end_write_request
;
3173 bi
->bi_end_io
= raid5_end_read_request
;
3176 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3177 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
3180 atomic_inc(&rdev
->nr_pending
);
3184 if (s
.syncing
|| s
.expanding
|| s
.expanded
)
3185 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
3187 bi
->bi_bdev
= rdev
->bdev
;
3188 pr_debug("for %llu schedule op %ld on disc %d\n",
3189 (unsigned long long)sh
->sector
, bi
->bi_rw
, i
);
3190 atomic_inc(&sh
->count
);
3191 bi
->bi_sector
= sh
->sector
+ rdev
->data_offset
;
3192 bi
->bi_flags
= 1 << BIO_UPTODATE
;
3194 bi
->bi_max_vecs
= 1;
3196 bi
->bi_io_vec
= &sh
->dev
[i
].vec
;
3197 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
3198 bi
->bi_io_vec
[0].bv_offset
= 0;
3199 bi
->bi_size
= STRIPE_SIZE
;
3202 test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
3203 atomic_add(STRIPE_SECTORS
, &rdev
->corrected_errors
);
3204 generic_make_request(bi
);
3207 set_bit(STRIPE_DEGRADED
, &sh
->state
);
3208 pr_debug("skip op %ld on disc %d for sector %llu\n",
3209 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
3210 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3211 set_bit(STRIPE_HANDLE
, &sh
->state
);
3216 static void handle_stripe(struct stripe_head
*sh
, struct page
*tmp_page
)
3218 if (sh
->raid_conf
->level
== 6)
3219 handle_stripe6(sh
, tmp_page
);
3226 static void raid5_activate_delayed(raid5_conf_t
*conf
)
3228 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3229 while (!list_empty(&conf
->delayed_list
)) {
3230 struct list_head
*l
= conf
->delayed_list
.next
;
3231 struct stripe_head
*sh
;
3232 sh
= list_entry(l
, struct stripe_head
, lru
);
3234 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3235 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3236 atomic_inc(&conf
->preread_active_stripes
);
3237 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3240 blk_plug_device(conf
->mddev
->queue
);
3243 static void activate_bit_delay(raid5_conf_t
*conf
)
3245 /* device_lock is held */
3246 struct list_head head
;
3247 list_add(&head
, &conf
->bitmap_list
);
3248 list_del_init(&conf
->bitmap_list
);
3249 while (!list_empty(&head
)) {
3250 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3251 list_del_init(&sh
->lru
);
3252 atomic_inc(&sh
->count
);
3253 __release_stripe(conf
, sh
);
3257 static void unplug_slaves(mddev_t
*mddev
)
3259 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3263 for (i
=0; i
<mddev
->raid_disks
; i
++) {
3264 mdk_rdev_t
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3265 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) && atomic_read(&rdev
->nr_pending
)) {
3266 struct request_queue
*r_queue
= bdev_get_queue(rdev
->bdev
);
3268 atomic_inc(&rdev
->nr_pending
);
3271 blk_unplug(r_queue
);
3273 rdev_dec_pending(rdev
, mddev
);
3280 static void raid5_unplug_device(struct request_queue
*q
)
3282 mddev_t
*mddev
= q
->queuedata
;
3283 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3284 unsigned long flags
;
3286 spin_lock_irqsave(&conf
->device_lock
, flags
);
3288 if (blk_remove_plug(q
)) {
3290 raid5_activate_delayed(conf
);
3292 md_wakeup_thread(mddev
->thread
);
3294 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3296 unplug_slaves(mddev
);
3299 static int raid5_congested(void *data
, int bits
)
3301 mddev_t
*mddev
= data
;
3302 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3304 /* No difference between reads and writes. Just check
3305 * how busy the stripe_cache is
3307 if (conf
->inactive_blocked
)
3311 if (list_empty_careful(&conf
->inactive_list
))
3317 /* We want read requests to align with chunks where possible,
3318 * but write requests don't need to.
3320 static int raid5_mergeable_bvec(struct request_queue
*q
, struct bio
*bio
, struct bio_vec
*biovec
)
3322 mddev_t
*mddev
= q
->queuedata
;
3323 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3325 unsigned int chunk_sectors
= mddev
->chunk_size
>> 9;
3326 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3328 if (bio_data_dir(bio
) == WRITE
)
3329 return biovec
->bv_len
; /* always allow writes to be mergeable */
3331 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3332 if (max
< 0) max
= 0;
3333 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3334 return biovec
->bv_len
;
3340 static int in_chunk_boundary(mddev_t
*mddev
, struct bio
*bio
)
3342 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3343 unsigned int chunk_sectors
= mddev
->chunk_size
>> 9;
3344 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3346 return chunk_sectors
>=
3347 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3351 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3352 * later sampled by raid5d.
3354 static void add_bio_to_retry(struct bio
*bi
,raid5_conf_t
*conf
)
3356 unsigned long flags
;
3358 spin_lock_irqsave(&conf
->device_lock
, flags
);
3360 bi
->bi_next
= conf
->retry_read_aligned_list
;
3361 conf
->retry_read_aligned_list
= bi
;
3363 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3364 md_wakeup_thread(conf
->mddev
->thread
);
3368 static struct bio
*remove_bio_from_retry(raid5_conf_t
*conf
)
3372 bi
= conf
->retry_read_aligned
;
3374 conf
->retry_read_aligned
= NULL
;
3377 bi
= conf
->retry_read_aligned_list
;
3379 conf
->retry_read_aligned_list
= bi
->bi_next
;
3381 bi
->bi_phys_segments
= 1; /* biased count of active stripes */
3382 bi
->bi_hw_segments
= 0; /* count of processed stripes */
3390 * The "raid5_align_endio" should check if the read succeeded and if it
3391 * did, call bio_endio on the original bio (having bio_put the new bio
3393 * If the read failed..
3395 static void raid5_align_endio(struct bio
*bi
, int error
)
3397 struct bio
* raid_bi
= bi
->bi_private
;
3400 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3405 mddev
= raid_bi
->bi_bdev
->bd_disk
->queue
->queuedata
;
3406 conf
= mddev_to_conf(mddev
);
3407 rdev
= (void*)raid_bi
->bi_next
;
3408 raid_bi
->bi_next
= NULL
;
3410 rdev_dec_pending(rdev
, conf
->mddev
);
3412 if (!error
&& uptodate
) {
3413 bio_endio(raid_bi
, 0);
3414 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3415 wake_up(&conf
->wait_for_stripe
);
3420 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3422 add_bio_to_retry(raid_bi
, conf
);
3425 static int bio_fits_rdev(struct bio
*bi
)
3427 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3429 if ((bi
->bi_size
>>9) > q
->max_sectors
)
3431 blk_recount_segments(q
, bi
);
3432 if (bi
->bi_phys_segments
> q
->max_phys_segments
||
3433 bi
->bi_hw_segments
> q
->max_hw_segments
)
3436 if (q
->merge_bvec_fn
)
3437 /* it's too hard to apply the merge_bvec_fn at this stage,
3446 static int chunk_aligned_read(struct request_queue
*q
, struct bio
* raid_bio
)
3448 mddev_t
*mddev
= q
->queuedata
;
3449 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3450 const unsigned int raid_disks
= conf
->raid_disks
;
3451 const unsigned int data_disks
= raid_disks
- conf
->max_degraded
;
3452 unsigned int dd_idx
, pd_idx
;
3453 struct bio
* align_bi
;
3456 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3457 pr_debug("chunk_aligned_read : non aligned\n");
3461 * use bio_clone to make a copy of the bio
3463 align_bi
= bio_clone(raid_bio
, GFP_NOIO
);
3467 * set bi_end_io to a new function, and set bi_private to the
3470 align_bi
->bi_end_io
= raid5_align_endio
;
3471 align_bi
->bi_private
= raid_bio
;
3475 align_bi
->bi_sector
= raid5_compute_sector(raid_bio
->bi_sector
,
3483 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3484 if (rdev
&& test_bit(In_sync
, &rdev
->flags
)) {
3485 atomic_inc(&rdev
->nr_pending
);
3487 raid_bio
->bi_next
= (void*)rdev
;
3488 align_bi
->bi_bdev
= rdev
->bdev
;
3489 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3490 align_bi
->bi_sector
+= rdev
->data_offset
;
3492 if (!bio_fits_rdev(align_bi
)) {
3493 /* too big in some way */
3495 rdev_dec_pending(rdev
, mddev
);
3499 spin_lock_irq(&conf
->device_lock
);
3500 wait_event_lock_irq(conf
->wait_for_stripe
,
3502 conf
->device_lock
, /* nothing */);
3503 atomic_inc(&conf
->active_aligned_reads
);
3504 spin_unlock_irq(&conf
->device_lock
);
3506 generic_make_request(align_bi
);
3515 /* __get_priority_stripe - get the next stripe to process
3517 * Full stripe writes are allowed to pass preread active stripes up until
3518 * the bypass_threshold is exceeded. In general the bypass_count
3519 * increments when the handle_list is handled before the hold_list; however, it
3520 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3521 * stripe with in flight i/o. The bypass_count will be reset when the
3522 * head of the hold_list has changed, i.e. the head was promoted to the
3525 static struct stripe_head
*__get_priority_stripe(raid5_conf_t
*conf
)
3527 struct stripe_head
*sh
;
3529 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3531 list_empty(&conf
->handle_list
) ? "empty" : "busy",
3532 list_empty(&conf
->hold_list
) ? "empty" : "busy",
3533 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
3535 if (!list_empty(&conf
->handle_list
)) {
3536 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
3538 if (list_empty(&conf
->hold_list
))
3539 conf
->bypass_count
= 0;
3540 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
3541 if (conf
->hold_list
.next
== conf
->last_hold
)
3542 conf
->bypass_count
++;
3544 conf
->last_hold
= conf
->hold_list
.next
;
3545 conf
->bypass_count
-= conf
->bypass_threshold
;
3546 if (conf
->bypass_count
< 0)
3547 conf
->bypass_count
= 0;
3550 } else if (!list_empty(&conf
->hold_list
) &&
3551 ((conf
->bypass_threshold
&&
3552 conf
->bypass_count
> conf
->bypass_threshold
) ||
3553 atomic_read(&conf
->pending_full_writes
) == 0)) {
3554 sh
= list_entry(conf
->hold_list
.next
,
3556 conf
->bypass_count
-= conf
->bypass_threshold
;
3557 if (conf
->bypass_count
< 0)
3558 conf
->bypass_count
= 0;
3562 list_del_init(&sh
->lru
);
3563 atomic_inc(&sh
->count
);
3564 BUG_ON(atomic_read(&sh
->count
) != 1);
3568 static int make_request(struct request_queue
*q
, struct bio
* bi
)
3570 mddev_t
*mddev
= q
->queuedata
;
3571 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3572 unsigned int dd_idx
, pd_idx
;
3573 sector_t new_sector
;
3574 sector_t logical_sector
, last_sector
;
3575 struct stripe_head
*sh
;
3576 const int rw
= bio_data_dir(bi
);
3579 if (unlikely(bio_barrier(bi
))) {
3580 bio_endio(bi
, -EOPNOTSUPP
);
3584 md_write_start(mddev
, bi
);
3586 disk_stat_inc(mddev
->gendisk
, ios
[rw
]);
3587 disk_stat_add(mddev
->gendisk
, sectors
[rw
], bio_sectors(bi
));
3590 mddev
->reshape_position
== MaxSector
&&
3591 chunk_aligned_read(q
,bi
))
3594 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
3595 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
3597 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
3599 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
3601 int disks
, data_disks
;
3604 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
3605 if (likely(conf
->expand_progress
== MaxSector
))
3606 disks
= conf
->raid_disks
;
3608 /* spinlock is needed as expand_progress may be
3609 * 64bit on a 32bit platform, and so it might be
3610 * possible to see a half-updated value
3611 * Ofcourse expand_progress could change after
3612 * the lock is dropped, so once we get a reference
3613 * to the stripe that we think it is, we will have
3616 spin_lock_irq(&conf
->device_lock
);
3617 disks
= conf
->raid_disks
;
3618 if (logical_sector
>= conf
->expand_progress
)
3619 disks
= conf
->previous_raid_disks
;
3621 if (logical_sector
>= conf
->expand_lo
) {
3622 spin_unlock_irq(&conf
->device_lock
);
3627 spin_unlock_irq(&conf
->device_lock
);
3629 data_disks
= disks
- conf
->max_degraded
;
3631 new_sector
= raid5_compute_sector(logical_sector
, disks
, data_disks
,
3632 &dd_idx
, &pd_idx
, conf
);
3633 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3634 (unsigned long long)new_sector
,
3635 (unsigned long long)logical_sector
);
3637 sh
= get_active_stripe(conf
, new_sector
, disks
, pd_idx
, (bi
->bi_rw
&RWA_MASK
));
3639 if (unlikely(conf
->expand_progress
!= MaxSector
)) {
3640 /* expansion might have moved on while waiting for a
3641 * stripe, so we must do the range check again.
3642 * Expansion could still move past after this
3643 * test, but as we are holding a reference to
3644 * 'sh', we know that if that happens,
3645 * STRIPE_EXPANDING will get set and the expansion
3646 * won't proceed until we finish with the stripe.
3649 spin_lock_irq(&conf
->device_lock
);
3650 if (logical_sector
< conf
->expand_progress
&&
3651 disks
== conf
->previous_raid_disks
)
3652 /* mismatch, need to try again */
3654 spin_unlock_irq(&conf
->device_lock
);
3660 /* FIXME what if we get a false positive because these
3661 * are being updated.
3663 if (logical_sector
>= mddev
->suspend_lo
&&
3664 logical_sector
< mddev
->suspend_hi
) {
3670 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
3671 !add_stripe_bio(sh
, bi
, dd_idx
, (bi
->bi_rw
&RW_MASK
))) {
3672 /* Stripe is busy expanding or
3673 * add failed due to overlap. Flush everything
3676 raid5_unplug_device(mddev
->queue
);
3681 finish_wait(&conf
->wait_for_overlap
, &w
);
3682 set_bit(STRIPE_HANDLE
, &sh
->state
);
3683 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3686 /* cannot get stripe for read-ahead, just give-up */
3687 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3688 finish_wait(&conf
->wait_for_overlap
, &w
);
3693 spin_lock_irq(&conf
->device_lock
);
3694 remaining
= --bi
->bi_phys_segments
;
3695 spin_unlock_irq(&conf
->device_lock
);
3696 if (remaining
== 0) {
3699 md_write_end(mddev
);
3706 static sector_t
reshape_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
)
3708 /* reshaping is quite different to recovery/resync so it is
3709 * handled quite separately ... here.
3711 * On each call to sync_request, we gather one chunk worth of
3712 * destination stripes and flag them as expanding.
3713 * Then we find all the source stripes and request reads.
3714 * As the reads complete, handle_stripe will copy the data
3715 * into the destination stripe and release that stripe.
3717 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
3718 struct stripe_head
*sh
;
3720 sector_t first_sector
, last_sector
;
3721 int raid_disks
= conf
->previous_raid_disks
;
3722 int data_disks
= raid_disks
- conf
->max_degraded
;
3723 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
3726 sector_t writepos
, safepos
, gap
;
3728 if (sector_nr
== 0 &&
3729 conf
->expand_progress
!= 0) {
3730 /* restarting in the middle, skip the initial sectors */
3731 sector_nr
= conf
->expand_progress
;
3732 sector_div(sector_nr
, new_data_disks
);
3737 /* we update the metadata when there is more than 3Meg
3738 * in the block range (that is rather arbitrary, should
3739 * probably be time based) or when the data about to be
3740 * copied would over-write the source of the data at
3741 * the front of the range.
3742 * i.e. one new_stripe forward from expand_progress new_maps
3743 * to after where expand_lo old_maps to
3745 writepos
= conf
->expand_progress
+
3746 conf
->chunk_size
/512*(new_data_disks
);
3747 sector_div(writepos
, new_data_disks
);
3748 safepos
= conf
->expand_lo
;
3749 sector_div(safepos
, data_disks
);
3750 gap
= conf
->expand_progress
- conf
->expand_lo
;
3752 if (writepos
>= safepos
||
3753 gap
> (new_data_disks
)*3000*2 /*3Meg*/) {
3754 /* Cannot proceed until we've updated the superblock... */
3755 wait_event(conf
->wait_for_overlap
,
3756 atomic_read(&conf
->reshape_stripes
)==0);
3757 mddev
->reshape_position
= conf
->expand_progress
;
3758 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
3759 md_wakeup_thread(mddev
->thread
);
3760 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
3761 kthread_should_stop());
3762 spin_lock_irq(&conf
->device_lock
);
3763 conf
->expand_lo
= mddev
->reshape_position
;
3764 spin_unlock_irq(&conf
->device_lock
);
3765 wake_up(&conf
->wait_for_overlap
);
3768 for (i
=0; i
< conf
->chunk_size
/512; i
+= STRIPE_SECTORS
) {
3771 pd_idx
= stripe_to_pdidx(sector_nr
+i
, conf
, conf
->raid_disks
);
3772 sh
= get_active_stripe(conf
, sector_nr
+i
,
3773 conf
->raid_disks
, pd_idx
, 0);
3774 set_bit(STRIPE_EXPANDING
, &sh
->state
);
3775 atomic_inc(&conf
->reshape_stripes
);
3776 /* If any of this stripe is beyond the end of the old
3777 * array, then we need to zero those blocks
3779 for (j
=sh
->disks
; j
--;) {
3781 if (j
== sh
->pd_idx
)
3783 if (conf
->level
== 6 &&
3784 j
== raid6_next_disk(sh
->pd_idx
, sh
->disks
))
3786 s
= compute_blocknr(sh
, j
);
3787 if (s
< (mddev
->array_size
<<1)) {
3791 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
3792 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
3793 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
3796 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3797 set_bit(STRIPE_HANDLE
, &sh
->state
);
3801 spin_lock_irq(&conf
->device_lock
);
3802 conf
->expand_progress
= (sector_nr
+ i
) * new_data_disks
;
3803 spin_unlock_irq(&conf
->device_lock
);
3804 /* Ok, those stripe are ready. We can start scheduling
3805 * reads on the source stripes.
3806 * The source stripes are determined by mapping the first and last
3807 * block on the destination stripes.
3810 raid5_compute_sector(sector_nr
*(new_data_disks
),
3811 raid_disks
, data_disks
,
3812 &dd_idx
, &pd_idx
, conf
);
3814 raid5_compute_sector((sector_nr
+conf
->chunk_size
/512)
3815 *(new_data_disks
) -1,
3816 raid_disks
, data_disks
,
3817 &dd_idx
, &pd_idx
, conf
);
3818 if (last_sector
>= (mddev
->size
<<1))
3819 last_sector
= (mddev
->size
<<1)-1;
3820 while (first_sector
<= last_sector
) {
3821 pd_idx
= stripe_to_pdidx(first_sector
, conf
,
3822 conf
->previous_raid_disks
);
3823 sh
= get_active_stripe(conf
, first_sector
,
3824 conf
->previous_raid_disks
, pd_idx
, 0);
3825 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3826 set_bit(STRIPE_HANDLE
, &sh
->state
);
3828 first_sector
+= STRIPE_SECTORS
;
3830 /* If this takes us to the resync_max point where we have to pause,
3831 * then we need to write out the superblock.
3833 sector_nr
+= conf
->chunk_size
>>9;
3834 if (sector_nr
>= mddev
->resync_max
) {
3835 /* Cannot proceed until we've updated the superblock... */
3836 wait_event(conf
->wait_for_overlap
,
3837 atomic_read(&conf
->reshape_stripes
) == 0);
3838 mddev
->reshape_position
= conf
->expand_progress
;
3839 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
3840 md_wakeup_thread(mddev
->thread
);
3841 wait_event(mddev
->sb_wait
,
3842 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
3843 || kthread_should_stop());
3844 spin_lock_irq(&conf
->device_lock
);
3845 conf
->expand_lo
= mddev
->reshape_position
;
3846 spin_unlock_irq(&conf
->device_lock
);
3847 wake_up(&conf
->wait_for_overlap
);
3849 return conf
->chunk_size
>>9;
3852 /* FIXME go_faster isn't used */
3853 static inline sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
3855 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
3856 struct stripe_head
*sh
;
3858 int raid_disks
= conf
->raid_disks
;
3859 sector_t max_sector
= mddev
->size
<< 1;
3861 int still_degraded
= 0;
3864 if (sector_nr
>= max_sector
) {
3865 /* just being told to finish up .. nothing much to do */
3866 unplug_slaves(mddev
);
3867 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
3872 if (mddev
->curr_resync
< max_sector
) /* aborted */
3873 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
3875 else /* completed sync */
3877 bitmap_close_sync(mddev
->bitmap
);
3882 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
3883 return reshape_request(mddev
, sector_nr
, skipped
);
3885 /* No need to check resync_max as we never do more than one
3886 * stripe, and as resync_max will always be on a chunk boundary,
3887 * if the check in md_do_sync didn't fire, there is no chance
3888 * of overstepping resync_max here
3891 /* if there is too many failed drives and we are trying
3892 * to resync, then assert that we are finished, because there is
3893 * nothing we can do.
3895 if (mddev
->degraded
>= conf
->max_degraded
&&
3896 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3897 sector_t rv
= (mddev
->size
<< 1) - sector_nr
;
3901 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
3902 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
3903 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
3904 /* we can skip this block, and probably more */
3905 sync_blocks
/= STRIPE_SECTORS
;
3907 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
3911 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3913 pd_idx
= stripe_to_pdidx(sector_nr
, conf
, raid_disks
);
3914 sh
= get_active_stripe(conf
, sector_nr
, raid_disks
, pd_idx
, 1);
3916 sh
= get_active_stripe(conf
, sector_nr
, raid_disks
, pd_idx
, 0);
3917 /* make sure we don't swamp the stripe cache if someone else
3918 * is trying to get access
3920 schedule_timeout_uninterruptible(1);
3922 /* Need to check if array will still be degraded after recovery/resync
3923 * We don't need to check the 'failed' flag as when that gets set,
3926 for (i
=0; i
<mddev
->raid_disks
; i
++)
3927 if (conf
->disks
[i
].rdev
== NULL
)
3930 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
3932 spin_lock(&sh
->lock
);
3933 set_bit(STRIPE_SYNCING
, &sh
->state
);
3934 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3935 spin_unlock(&sh
->lock
);
3937 handle_stripe(sh
, NULL
);
3940 return STRIPE_SECTORS
;
3943 static int retry_aligned_read(raid5_conf_t
*conf
, struct bio
*raid_bio
)
3945 /* We may not be able to submit a whole bio at once as there
3946 * may not be enough stripe_heads available.
3947 * We cannot pre-allocate enough stripe_heads as we may need
3948 * more than exist in the cache (if we allow ever large chunks).
3949 * So we do one stripe head at a time and record in
3950 * ->bi_hw_segments how many have been done.
3952 * We *know* that this entire raid_bio is in one chunk, so
3953 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3955 struct stripe_head
*sh
;
3957 sector_t sector
, logical_sector
, last_sector
;
3962 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
3963 sector
= raid5_compute_sector( logical_sector
,
3965 conf
->raid_disks
- conf
->max_degraded
,
3969 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
3971 for (; logical_sector
< last_sector
;
3972 logical_sector
+= STRIPE_SECTORS
,
3973 sector
+= STRIPE_SECTORS
,
3976 if (scnt
< raid_bio
->bi_hw_segments
)
3977 /* already done this stripe */
3980 sh
= get_active_stripe(conf
, sector
, conf
->raid_disks
, pd_idx
, 1);
3983 /* failed to get a stripe - must wait */
3984 raid_bio
->bi_hw_segments
= scnt
;
3985 conf
->retry_read_aligned
= raid_bio
;
3989 set_bit(R5_ReadError
, &sh
->dev
[dd_idx
].flags
);
3990 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
3992 raid_bio
->bi_hw_segments
= scnt
;
3993 conf
->retry_read_aligned
= raid_bio
;
3997 handle_stripe(sh
, NULL
);
4001 spin_lock_irq(&conf
->device_lock
);
4002 remaining
= --raid_bio
->bi_phys_segments
;
4003 spin_unlock_irq(&conf
->device_lock
);
4005 bio_endio(raid_bio
, 0);
4006 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4007 wake_up(&conf
->wait_for_stripe
);
4014 * This is our raid5 kernel thread.
4016 * We scan the hash table for stripes which can be handled now.
4017 * During the scan, completed stripes are saved for us by the interrupt
4018 * handler, so that they will not have to wait for our next wakeup.
4020 static void raid5d(mddev_t
*mddev
)
4022 struct stripe_head
*sh
;
4023 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4026 pr_debug("+++ raid5d active\n");
4028 md_check_recovery(mddev
);
4031 spin_lock_irq(&conf
->device_lock
);
4035 if (conf
->seq_flush
!= conf
->seq_write
) {
4036 int seq
= conf
->seq_flush
;
4037 spin_unlock_irq(&conf
->device_lock
);
4038 bitmap_unplug(mddev
->bitmap
);
4039 spin_lock_irq(&conf
->device_lock
);
4040 conf
->seq_write
= seq
;
4041 activate_bit_delay(conf
);
4044 while ((bio
= remove_bio_from_retry(conf
))) {
4046 spin_unlock_irq(&conf
->device_lock
);
4047 ok
= retry_aligned_read(conf
, bio
);
4048 spin_lock_irq(&conf
->device_lock
);
4054 sh
= __get_priority_stripe(conf
);
4057 async_tx_issue_pending_all();
4060 spin_unlock_irq(&conf
->device_lock
);
4063 handle_stripe(sh
, conf
->spare_page
);
4066 spin_lock_irq(&conf
->device_lock
);
4068 pr_debug("%d stripes handled\n", handled
);
4070 spin_unlock_irq(&conf
->device_lock
);
4072 unplug_slaves(mddev
);
4074 pr_debug("--- raid5d inactive\n");
4078 raid5_show_stripe_cache_size(mddev_t
*mddev
, char *page
)
4080 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4082 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4088 raid5_store_stripe_cache_size(mddev_t
*mddev
, const char *page
, size_t len
)
4090 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4092 if (len
>= PAGE_SIZE
)
4097 if (strict_strtoul(page
, 10, &new))
4099 if (new <= 16 || new > 32768)
4101 while (new < conf
->max_nr_stripes
) {
4102 if (drop_one_stripe(conf
))
4103 conf
->max_nr_stripes
--;
4107 md_allow_write(mddev
);
4108 while (new > conf
->max_nr_stripes
) {
4109 if (grow_one_stripe(conf
))
4110 conf
->max_nr_stripes
++;
4116 static struct md_sysfs_entry
4117 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4118 raid5_show_stripe_cache_size
,
4119 raid5_store_stripe_cache_size
);
4122 raid5_show_preread_threshold(mddev_t
*mddev
, char *page
)
4124 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4126 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4132 raid5_store_preread_threshold(mddev_t
*mddev
, const char *page
, size_t len
)
4134 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4136 if (len
>= PAGE_SIZE
)
4141 if (strict_strtoul(page
, 10, &new))
4143 if (new > conf
->max_nr_stripes
)
4145 conf
->bypass_threshold
= new;
4149 static struct md_sysfs_entry
4150 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4152 raid5_show_preread_threshold
,
4153 raid5_store_preread_threshold
);
4156 stripe_cache_active_show(mddev_t
*mddev
, char *page
)
4158 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4160 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4165 static struct md_sysfs_entry
4166 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4168 static struct attribute
*raid5_attrs
[] = {
4169 &raid5_stripecache_size
.attr
,
4170 &raid5_stripecache_active
.attr
,
4171 &raid5_preread_bypass_threshold
.attr
,
4174 static struct attribute_group raid5_attrs_group
= {
4176 .attrs
= raid5_attrs
,
4179 static int run(mddev_t
*mddev
)
4182 int raid_disk
, memory
;
4184 struct disk_info
*disk
;
4185 struct list_head
*tmp
;
4186 int working_disks
= 0;
4188 if (mddev
->level
!= 5 && mddev
->level
!= 4 && mddev
->level
!= 6) {
4189 printk(KERN_ERR
"raid5: %s: raid level not set to 4/5/6 (%d)\n",
4190 mdname(mddev
), mddev
->level
);
4194 if (mddev
->reshape_position
!= MaxSector
) {
4195 /* Check that we can continue the reshape.
4196 * Currently only disks can change, it must
4197 * increase, and we must be past the point where
4198 * a stripe over-writes itself
4200 sector_t here_new
, here_old
;
4202 int max_degraded
= (mddev
->level
== 5 ? 1 : 2);
4204 if (mddev
->new_level
!= mddev
->level
||
4205 mddev
->new_layout
!= mddev
->layout
||
4206 mddev
->new_chunk
!= mddev
->chunk_size
) {
4207 printk(KERN_ERR
"raid5: %s: unsupported reshape "
4208 "required - aborting.\n",
4212 if (mddev
->delta_disks
<= 0) {
4213 printk(KERN_ERR
"raid5: %s: unsupported reshape "
4214 "(reduce disks) required - aborting.\n",
4218 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4219 /* reshape_position must be on a new-stripe boundary, and one
4220 * further up in new geometry must map after here in old
4223 here_new
= mddev
->reshape_position
;
4224 if (sector_div(here_new
, (mddev
->chunk_size
>>9)*
4225 (mddev
->raid_disks
- max_degraded
))) {
4226 printk(KERN_ERR
"raid5: reshape_position not "
4227 "on a stripe boundary\n");
4230 /* here_new is the stripe we will write to */
4231 here_old
= mddev
->reshape_position
;
4232 sector_div(here_old
, (mddev
->chunk_size
>>9)*
4233 (old_disks
-max_degraded
));
4234 /* here_old is the first stripe that we might need to read
4236 if (here_new
>= here_old
) {
4237 /* Reading from the same stripe as writing to - bad */
4238 printk(KERN_ERR
"raid5: reshape_position too early for "
4239 "auto-recovery - aborting.\n");
4242 printk(KERN_INFO
"raid5: reshape will continue\n");
4243 /* OK, we should be able to continue; */
4247 mddev
->private = kzalloc(sizeof (raid5_conf_t
), GFP_KERNEL
);
4248 if ((conf
= mddev
->private) == NULL
)
4250 if (mddev
->reshape_position
== MaxSector
) {
4251 conf
->previous_raid_disks
= conf
->raid_disks
= mddev
->raid_disks
;
4253 conf
->raid_disks
= mddev
->raid_disks
;
4254 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4257 conf
->disks
= kzalloc(conf
->raid_disks
* sizeof(struct disk_info
),
4262 conf
->mddev
= mddev
;
4264 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
4267 if (mddev
->level
== 6) {
4268 conf
->spare_page
= alloc_page(GFP_KERNEL
);
4269 if (!conf
->spare_page
)
4272 spin_lock_init(&conf
->device_lock
);
4273 mddev
->queue
->queue_lock
= &conf
->device_lock
;
4274 init_waitqueue_head(&conf
->wait_for_stripe
);
4275 init_waitqueue_head(&conf
->wait_for_overlap
);
4276 INIT_LIST_HEAD(&conf
->handle_list
);
4277 INIT_LIST_HEAD(&conf
->hold_list
);
4278 INIT_LIST_HEAD(&conf
->delayed_list
);
4279 INIT_LIST_HEAD(&conf
->bitmap_list
);
4280 INIT_LIST_HEAD(&conf
->inactive_list
);
4281 atomic_set(&conf
->active_stripes
, 0);
4282 atomic_set(&conf
->preread_active_stripes
, 0);
4283 atomic_set(&conf
->active_aligned_reads
, 0);
4284 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
4286 pr_debug("raid5: run(%s) called.\n", mdname(mddev
));
4288 rdev_for_each(rdev
, tmp
, mddev
) {
4289 raid_disk
= rdev
->raid_disk
;
4290 if (raid_disk
>= conf
->raid_disks
4293 disk
= conf
->disks
+ raid_disk
;
4297 if (test_bit(In_sync
, &rdev
->flags
)) {
4298 char b
[BDEVNAME_SIZE
];
4299 printk(KERN_INFO
"raid5: device %s operational as raid"
4300 " disk %d\n", bdevname(rdev
->bdev
,b
),
4304 /* Cannot rely on bitmap to complete recovery */
4309 * 0 for a fully functional array, 1 or 2 for a degraded array.
4311 mddev
->degraded
= conf
->raid_disks
- working_disks
;
4312 conf
->mddev
= mddev
;
4313 conf
->chunk_size
= mddev
->chunk_size
;
4314 conf
->level
= mddev
->level
;
4315 if (conf
->level
== 6)
4316 conf
->max_degraded
= 2;
4318 conf
->max_degraded
= 1;
4319 conf
->algorithm
= mddev
->layout
;
4320 conf
->max_nr_stripes
= NR_STRIPES
;
4321 conf
->expand_progress
= mddev
->reshape_position
;
4323 /* device size must be a multiple of chunk size */
4324 mddev
->size
&= ~(mddev
->chunk_size
/1024 -1);
4325 mddev
->resync_max_sectors
= mddev
->size
<< 1;
4327 if (conf
->level
== 6 && conf
->raid_disks
< 4) {
4328 printk(KERN_ERR
"raid6: not enough configured devices for %s (%d, minimum 4)\n",
4329 mdname(mddev
), conf
->raid_disks
);
4332 if (!conf
->chunk_size
|| conf
->chunk_size
% 4) {
4333 printk(KERN_ERR
"raid5: invalid chunk size %d for %s\n",
4334 conf
->chunk_size
, mdname(mddev
));
4337 if (conf
->algorithm
> ALGORITHM_RIGHT_SYMMETRIC
) {
4339 "raid5: unsupported parity algorithm %d for %s\n",
4340 conf
->algorithm
, mdname(mddev
));
4343 if (mddev
->degraded
> conf
->max_degraded
) {
4344 printk(KERN_ERR
"raid5: not enough operational devices for %s"
4345 " (%d/%d failed)\n",
4346 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
4350 if (mddev
->degraded
> 0 &&
4351 mddev
->recovery_cp
!= MaxSector
) {
4352 if (mddev
->ok_start_degraded
)
4354 "raid5: starting dirty degraded array: %s"
4355 "- data corruption possible.\n",
4359 "raid5: cannot start dirty degraded array for %s\n",
4366 mddev
->thread
= md_register_thread(raid5d
, mddev
, "%s_raid5");
4367 if (!mddev
->thread
) {
4369 "raid5: couldn't allocate thread for %s\n",
4374 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
4375 conf
->raid_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
4376 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
4378 "raid5: couldn't allocate %dkB for buffers\n", memory
);
4379 shrink_stripes(conf
);
4380 md_unregister_thread(mddev
->thread
);
4383 printk(KERN_INFO
"raid5: allocated %dkB for %s\n",
4384 memory
, mdname(mddev
));
4386 if (mddev
->degraded
== 0)
4387 printk("raid5: raid level %d set %s active with %d out of %d"
4388 " devices, algorithm %d\n", conf
->level
, mdname(mddev
),
4389 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
4392 printk(KERN_ALERT
"raid5: raid level %d set %s active with %d"
4393 " out of %d devices, algorithm %d\n", conf
->level
,
4394 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
4395 mddev
->raid_disks
, conf
->algorithm
);
4397 print_raid5_conf(conf
);
4399 if (conf
->expand_progress
!= MaxSector
) {
4400 printk("...ok start reshape thread\n");
4401 conf
->expand_lo
= conf
->expand_progress
;
4402 atomic_set(&conf
->reshape_stripes
, 0);
4403 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4404 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4405 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4406 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4407 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4411 /* read-ahead size must cover two whole stripes, which is
4412 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4415 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
4416 int stripe
= data_disks
*
4417 (mddev
->chunk_size
/ PAGE_SIZE
);
4418 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4419 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4422 /* Ok, everything is just fine now */
4423 if (sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
4425 "raid5: failed to create sysfs attributes for %s\n",
4428 mddev
->queue
->unplug_fn
= raid5_unplug_device
;
4429 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
4430 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
4432 mddev
->array_size
= mddev
->size
* (conf
->previous_raid_disks
-
4433 conf
->max_degraded
);
4435 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
4440 print_raid5_conf(conf
);
4441 safe_put_page(conf
->spare_page
);
4443 kfree(conf
->stripe_hashtbl
);
4446 mddev
->private = NULL
;
4447 printk(KERN_ALERT
"raid5: failed to run raid set %s\n", mdname(mddev
));
4453 static int stop(mddev_t
*mddev
)
4455 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
4457 md_unregister_thread(mddev
->thread
);
4458 mddev
->thread
= NULL
;
4459 shrink_stripes(conf
);
4460 kfree(conf
->stripe_hashtbl
);
4461 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
4462 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
4463 sysfs_remove_group(&mddev
->kobj
, &raid5_attrs_group
);
4466 mddev
->private = NULL
;
4471 static void print_sh (struct seq_file
*seq
, struct stripe_head
*sh
)
4475 seq_printf(seq
, "sh %llu, pd_idx %d, state %ld.\n",
4476 (unsigned long long)sh
->sector
, sh
->pd_idx
, sh
->state
);
4477 seq_printf(seq
, "sh %llu, count %d.\n",
4478 (unsigned long long)sh
->sector
, atomic_read(&sh
->count
));
4479 seq_printf(seq
, "sh %llu, ", (unsigned long long)sh
->sector
);
4480 for (i
= 0; i
< sh
->disks
; i
++) {
4481 seq_printf(seq
, "(cache%d: %p %ld) ",
4482 i
, sh
->dev
[i
].page
, sh
->dev
[i
].flags
);
4484 seq_printf(seq
, "\n");
4487 static void printall (struct seq_file
*seq
, raid5_conf_t
*conf
)
4489 struct stripe_head
*sh
;
4490 struct hlist_node
*hn
;
4493 spin_lock_irq(&conf
->device_lock
);
4494 for (i
= 0; i
< NR_HASH
; i
++) {
4495 hlist_for_each_entry(sh
, hn
, &conf
->stripe_hashtbl
[i
], hash
) {
4496 if (sh
->raid_conf
!= conf
)
4501 spin_unlock_irq(&conf
->device_lock
);
4505 static void status (struct seq_file
*seq
, mddev_t
*mddev
)
4507 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
4510 seq_printf (seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
, mddev
->chunk_size
>> 10, mddev
->layout
);
4511 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
4512 for (i
= 0; i
< conf
->raid_disks
; i
++)
4513 seq_printf (seq
, "%s",
4514 conf
->disks
[i
].rdev
&&
4515 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
4516 seq_printf (seq
, "]");
4518 seq_printf (seq
, "\n");
4519 printall(seq
, conf
);
4523 static void print_raid5_conf (raid5_conf_t
*conf
)
4526 struct disk_info
*tmp
;
4528 printk("RAID5 conf printout:\n");
4530 printk("(conf==NULL)\n");
4533 printk(" --- rd:%d wd:%d\n", conf
->raid_disks
,
4534 conf
->raid_disks
- conf
->mddev
->degraded
);
4536 for (i
= 0; i
< conf
->raid_disks
; i
++) {
4537 char b
[BDEVNAME_SIZE
];
4538 tmp
= conf
->disks
+ i
;
4540 printk(" disk %d, o:%d, dev:%s\n",
4541 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
4542 bdevname(tmp
->rdev
->bdev
,b
));
4546 static int raid5_spare_active(mddev_t
*mddev
)
4549 raid5_conf_t
*conf
= mddev
->private;
4550 struct disk_info
*tmp
;
4552 for (i
= 0; i
< conf
->raid_disks
; i
++) {
4553 tmp
= conf
->disks
+ i
;
4555 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
4556 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
4557 unsigned long flags
;
4558 spin_lock_irqsave(&conf
->device_lock
, flags
);
4560 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4563 print_raid5_conf(conf
);
4567 static int raid5_remove_disk(mddev_t
*mddev
, int number
)
4569 raid5_conf_t
*conf
= mddev
->private;
4572 struct disk_info
*p
= conf
->disks
+ number
;
4574 print_raid5_conf(conf
);
4577 if (test_bit(In_sync
, &rdev
->flags
) ||
4578 atomic_read(&rdev
->nr_pending
)) {
4582 /* Only remove non-faulty devices if recovery
4585 if (!test_bit(Faulty
, &rdev
->flags
) &&
4586 mddev
->degraded
<= conf
->max_degraded
) {
4592 if (atomic_read(&rdev
->nr_pending
)) {
4593 /* lost the race, try later */
4600 print_raid5_conf(conf
);
4604 static int raid5_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
4606 raid5_conf_t
*conf
= mddev
->private;
4609 struct disk_info
*p
;
4611 int last
= conf
->raid_disks
- 1;
4613 if (mddev
->degraded
> conf
->max_degraded
)
4614 /* no point adding a device */
4617 if (rdev
->raid_disk
>= 0)
4618 first
= last
= rdev
->raid_disk
;
4621 * find the disk ... but prefer rdev->saved_raid_disk
4624 if (rdev
->saved_raid_disk
>= 0 &&
4625 rdev
->saved_raid_disk
>= first
&&
4626 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
4627 disk
= rdev
->saved_raid_disk
;
4630 for ( ; disk
<= last
; disk
++)
4631 if ((p
=conf
->disks
+ disk
)->rdev
== NULL
) {
4632 clear_bit(In_sync
, &rdev
->flags
);
4633 rdev
->raid_disk
= disk
;
4635 if (rdev
->saved_raid_disk
!= disk
)
4637 rcu_assign_pointer(p
->rdev
, rdev
);
4640 print_raid5_conf(conf
);
4644 static int raid5_resize(mddev_t
*mddev
, sector_t sectors
)
4646 /* no resync is happening, and there is enough space
4647 * on all devices, so we can resize.
4648 * We need to make sure resync covers any new space.
4649 * If the array is shrinking we should possibly wait until
4650 * any io in the removed space completes, but it hardly seems
4653 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4655 sectors
&= ~((sector_t
)mddev
->chunk_size
/512 - 1);
4656 mddev
->array_size
= (sectors
* (mddev
->raid_disks
-conf
->max_degraded
))>>1;
4657 set_capacity(mddev
->gendisk
, mddev
->array_size
<< 1);
4659 if (sectors
/2 > mddev
->size
&& mddev
->recovery_cp
== MaxSector
) {
4660 mddev
->recovery_cp
= mddev
->size
<< 1;
4661 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4663 mddev
->size
= sectors
/2;
4664 mddev
->resync_max_sectors
= sectors
;
4668 #ifdef CONFIG_MD_RAID5_RESHAPE
4669 static int raid5_check_reshape(mddev_t
*mddev
)
4671 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4674 if (mddev
->delta_disks
< 0 ||
4675 mddev
->new_level
!= mddev
->level
)
4676 return -EINVAL
; /* Cannot shrink array or change level yet */
4677 if (mddev
->delta_disks
== 0)
4678 return 0; /* nothing to do */
4680 /* Can only proceed if there are plenty of stripe_heads.
4681 * We need a minimum of one full stripe,, and for sensible progress
4682 * it is best to have about 4 times that.
4683 * If we require 4 times, then the default 256 4K stripe_heads will
4684 * allow for chunk sizes up to 256K, which is probably OK.
4685 * If the chunk size is greater, user-space should request more
4686 * stripe_heads first.
4688 if ((mddev
->chunk_size
/ STRIPE_SIZE
) * 4 > conf
->max_nr_stripes
||
4689 (mddev
->new_chunk
/ STRIPE_SIZE
) * 4 > conf
->max_nr_stripes
) {
4690 printk(KERN_WARNING
"raid5: reshape: not enough stripes. Needed %lu\n",
4691 (mddev
->chunk_size
/ STRIPE_SIZE
)*4);
4695 err
= resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
4699 if (mddev
->degraded
> conf
->max_degraded
)
4701 /* looks like we might be able to manage this */
4705 static int raid5_start_reshape(mddev_t
*mddev
)
4707 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4709 struct list_head
*rtmp
;
4711 int added_devices
= 0;
4712 unsigned long flags
;
4714 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4717 rdev_for_each(rdev
, rtmp
, mddev
)
4718 if (rdev
->raid_disk
< 0 &&
4719 !test_bit(Faulty
, &rdev
->flags
))
4722 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
4723 /* Not enough devices even to make a degraded array
4728 atomic_set(&conf
->reshape_stripes
, 0);
4729 spin_lock_irq(&conf
->device_lock
);
4730 conf
->previous_raid_disks
= conf
->raid_disks
;
4731 conf
->raid_disks
+= mddev
->delta_disks
;
4732 conf
->expand_progress
= 0;
4733 conf
->expand_lo
= 0;
4734 spin_unlock_irq(&conf
->device_lock
);
4736 /* Add some new drives, as many as will fit.
4737 * We know there are enough to make the newly sized array work.
4739 rdev_for_each(rdev
, rtmp
, mddev
)
4740 if (rdev
->raid_disk
< 0 &&
4741 !test_bit(Faulty
, &rdev
->flags
)) {
4742 if (raid5_add_disk(mddev
, rdev
)) {
4744 set_bit(In_sync
, &rdev
->flags
);
4746 rdev
->recovery_offset
= 0;
4747 sprintf(nm
, "rd%d", rdev
->raid_disk
);
4748 if (sysfs_create_link(&mddev
->kobj
,
4751 "raid5: failed to create "
4752 " link %s for %s\n",
4758 spin_lock_irqsave(&conf
->device_lock
, flags
);
4759 mddev
->degraded
= (conf
->raid_disks
- conf
->previous_raid_disks
) - added_devices
;
4760 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4761 mddev
->raid_disks
= conf
->raid_disks
;
4762 mddev
->reshape_position
= 0;
4763 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4765 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4766 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4767 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4768 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4769 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4771 if (!mddev
->sync_thread
) {
4772 mddev
->recovery
= 0;
4773 spin_lock_irq(&conf
->device_lock
);
4774 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
4775 conf
->expand_progress
= MaxSector
;
4776 spin_unlock_irq(&conf
->device_lock
);
4779 md_wakeup_thread(mddev
->sync_thread
);
4780 md_new_event(mddev
);
4785 static void end_reshape(raid5_conf_t
*conf
)
4787 struct block_device
*bdev
;
4789 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
4790 conf
->mddev
->array_size
= conf
->mddev
->size
*
4791 (conf
->raid_disks
- conf
->max_degraded
);
4792 set_capacity(conf
->mddev
->gendisk
, conf
->mddev
->array_size
<< 1);
4793 conf
->mddev
->changed
= 1;
4795 bdev
= bdget_disk(conf
->mddev
->gendisk
, 0);
4797 mutex_lock(&bdev
->bd_inode
->i_mutex
);
4798 i_size_write(bdev
->bd_inode
, (loff_t
)conf
->mddev
->array_size
<< 10);
4799 mutex_unlock(&bdev
->bd_inode
->i_mutex
);
4802 spin_lock_irq(&conf
->device_lock
);
4803 conf
->expand_progress
= MaxSector
;
4804 spin_unlock_irq(&conf
->device_lock
);
4805 conf
->mddev
->reshape_position
= MaxSector
;
4807 /* read-ahead size must cover two whole stripes, which is
4808 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4811 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
4812 int stripe
= data_disks
*
4813 (conf
->mddev
->chunk_size
/ PAGE_SIZE
);
4814 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4815 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4820 static void raid5_quiesce(mddev_t
*mddev
, int state
)
4822 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4825 case 2: /* resume for a suspend */
4826 wake_up(&conf
->wait_for_overlap
);
4829 case 1: /* stop all writes */
4830 spin_lock_irq(&conf
->device_lock
);
4832 wait_event_lock_irq(conf
->wait_for_stripe
,
4833 atomic_read(&conf
->active_stripes
) == 0 &&
4834 atomic_read(&conf
->active_aligned_reads
) == 0,
4835 conf
->device_lock
, /* nothing */);
4836 spin_unlock_irq(&conf
->device_lock
);
4839 case 0: /* re-enable writes */
4840 spin_lock_irq(&conf
->device_lock
);
4842 wake_up(&conf
->wait_for_stripe
);
4843 wake_up(&conf
->wait_for_overlap
);
4844 spin_unlock_irq(&conf
->device_lock
);
4849 static struct mdk_personality raid6_personality
=
4853 .owner
= THIS_MODULE
,
4854 .make_request
= make_request
,
4858 .error_handler
= error
,
4859 .hot_add_disk
= raid5_add_disk
,
4860 .hot_remove_disk
= raid5_remove_disk
,
4861 .spare_active
= raid5_spare_active
,
4862 .sync_request
= sync_request
,
4863 .resize
= raid5_resize
,
4864 #ifdef CONFIG_MD_RAID5_RESHAPE
4865 .check_reshape
= raid5_check_reshape
,
4866 .start_reshape
= raid5_start_reshape
,
4868 .quiesce
= raid5_quiesce
,
4870 static struct mdk_personality raid5_personality
=
4874 .owner
= THIS_MODULE
,
4875 .make_request
= make_request
,
4879 .error_handler
= error
,
4880 .hot_add_disk
= raid5_add_disk
,
4881 .hot_remove_disk
= raid5_remove_disk
,
4882 .spare_active
= raid5_spare_active
,
4883 .sync_request
= sync_request
,
4884 .resize
= raid5_resize
,
4885 #ifdef CONFIG_MD_RAID5_RESHAPE
4886 .check_reshape
= raid5_check_reshape
,
4887 .start_reshape
= raid5_start_reshape
,
4889 .quiesce
= raid5_quiesce
,
4892 static struct mdk_personality raid4_personality
=
4896 .owner
= THIS_MODULE
,
4897 .make_request
= make_request
,
4901 .error_handler
= error
,
4902 .hot_add_disk
= raid5_add_disk
,
4903 .hot_remove_disk
= raid5_remove_disk
,
4904 .spare_active
= raid5_spare_active
,
4905 .sync_request
= sync_request
,
4906 .resize
= raid5_resize
,
4907 #ifdef CONFIG_MD_RAID5_RESHAPE
4908 .check_reshape
= raid5_check_reshape
,
4909 .start_reshape
= raid5_start_reshape
,
4911 .quiesce
= raid5_quiesce
,
4914 static int __init
raid5_init(void)
4918 e
= raid6_select_algo();
4921 register_md_personality(&raid6_personality
);
4922 register_md_personality(&raid5_personality
);
4923 register_md_personality(&raid4_personality
);
4927 static void raid5_exit(void)
4929 unregister_md_personality(&raid6_personality
);
4930 unregister_md_personality(&raid5_personality
);
4931 unregister_md_personality(&raid4_personality
);
4934 module_init(raid5_init
);
4935 module_exit(raid5_exit
);
4936 MODULE_LICENSE("GPL");
4937 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4938 MODULE_ALIAS("md-raid5");
4939 MODULE_ALIAS("md-raid4");
4940 MODULE_ALIAS("md-level-5");
4941 MODULE_ALIAS("md-level-4");
4942 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4943 MODULE_ALIAS("md-raid6");
4944 MODULE_ALIAS("md-level-6");
4946 /* This used to be two separate modules, they were: */
4947 MODULE_ALIAS("raid5");
4948 MODULE_ALIAS("raid6");