2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->bm_write is the number of the last batch successfully written.
31 * conf->bm_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is bm_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/seq_file.h>
59 #define NR_STRIPES 256
60 #define STRIPE_SIZE PAGE_SIZE
61 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
62 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
63 #define IO_THRESHOLD 1
64 #define BYPASS_THRESHOLD 1
65 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
66 #define HASH_MASK (NR_HASH - 1)
68 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
70 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
71 * order without overlap. There may be several bio's per stripe+device, and
72 * a bio could span several devices.
73 * When walking this list for a particular stripe+device, we must never proceed
74 * beyond a bio that extends past this device, as the next bio might no longer
76 * This macro is used to determine the 'next' bio in the list, given the sector
77 * of the current stripe+device
79 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
81 * The following can be used to debug the driver
83 #define RAID5_PARANOIA 1
84 #if RAID5_PARANOIA && defined(CONFIG_SMP)
85 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
87 # define CHECK_DEVLOCK()
95 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
98 * We maintain a biased count of active stripes in the bottom 16 bits of
99 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
101 static inline int raid5_bi_phys_segments(struct bio
*bio
)
103 return bio
->bi_phys_segments
& 0xffff;
106 static inline int raid5_bi_hw_segments(struct bio
*bio
)
108 return (bio
->bi_phys_segments
>> 16) & 0xffff;
111 static inline int raid5_dec_bi_phys_segments(struct bio
*bio
)
113 --bio
->bi_phys_segments
;
114 return raid5_bi_phys_segments(bio
);
117 static inline int raid5_dec_bi_hw_segments(struct bio
*bio
)
119 unsigned short val
= raid5_bi_hw_segments(bio
);
122 bio
->bi_phys_segments
= (val
<< 16) | raid5_bi_phys_segments(bio
);
126 static inline void raid5_set_bi_hw_segments(struct bio
*bio
, unsigned int cnt
)
128 bio
->bi_phys_segments
= raid5_bi_phys_segments(bio
) || (cnt
<< 16);
131 /* Find first data disk in a raid6 stripe */
132 static inline int raid6_d0(struct stripe_head
*sh
)
135 /* ddf always start from first device */
137 /* md starts just after Q block */
138 if (sh
->qd_idx
== sh
->disks
- 1)
141 return sh
->qd_idx
+ 1;
143 static inline int raid6_next_disk(int disk
, int raid_disks
)
146 return (disk
< raid_disks
) ? disk
: 0;
149 /* When walking through the disks in a raid5, starting at raid6_d0,
150 * We need to map each disk to a 'slot', where the data disks are slot
151 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
152 * is raid_disks-1. This help does that mapping.
154 static int raid6_idx_to_slot(int idx
, struct stripe_head
*sh
,
155 int *count
, int syndrome_disks
)
159 if (idx
== sh
->pd_idx
)
160 return syndrome_disks
;
161 if (idx
== sh
->qd_idx
)
162 return syndrome_disks
+ 1;
167 static void return_io(struct bio
*return_bi
)
169 struct bio
*bi
= return_bi
;
172 return_bi
= bi
->bi_next
;
180 static void print_raid5_conf (raid5_conf_t
*conf
);
182 static int stripe_operations_active(struct stripe_head
*sh
)
184 return sh
->check_state
|| sh
->reconstruct_state
||
185 test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
) ||
186 test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
189 static void __release_stripe(raid5_conf_t
*conf
, struct stripe_head
*sh
)
191 if (atomic_dec_and_test(&sh
->count
)) {
192 BUG_ON(!list_empty(&sh
->lru
));
193 BUG_ON(atomic_read(&conf
->active_stripes
)==0);
194 if (test_bit(STRIPE_HANDLE
, &sh
->state
)) {
195 if (test_bit(STRIPE_DELAYED
, &sh
->state
)) {
196 list_add_tail(&sh
->lru
, &conf
->delayed_list
);
197 blk_plug_device(conf
->mddev
->queue
);
198 } else if (test_bit(STRIPE_BIT_DELAY
, &sh
->state
) &&
199 sh
->bm_seq
- conf
->seq_write
> 0) {
200 list_add_tail(&sh
->lru
, &conf
->bitmap_list
);
201 blk_plug_device(conf
->mddev
->queue
);
203 clear_bit(STRIPE_BIT_DELAY
, &sh
->state
);
204 list_add_tail(&sh
->lru
, &conf
->handle_list
);
206 md_wakeup_thread(conf
->mddev
->thread
);
208 BUG_ON(stripe_operations_active(sh
));
209 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
210 atomic_dec(&conf
->preread_active_stripes
);
211 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
)
212 md_wakeup_thread(conf
->mddev
->thread
);
214 atomic_dec(&conf
->active_stripes
);
215 if (!test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
216 list_add_tail(&sh
->lru
, &conf
->inactive_list
);
217 wake_up(&conf
->wait_for_stripe
);
218 if (conf
->retry_read_aligned
)
219 md_wakeup_thread(conf
->mddev
->thread
);
225 static void release_stripe(struct stripe_head
*sh
)
227 raid5_conf_t
*conf
= sh
->raid_conf
;
230 spin_lock_irqsave(&conf
->device_lock
, flags
);
231 __release_stripe(conf
, sh
);
232 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
235 static inline void remove_hash(struct stripe_head
*sh
)
237 pr_debug("remove_hash(), stripe %llu\n",
238 (unsigned long long)sh
->sector
);
240 hlist_del_init(&sh
->hash
);
243 static inline void insert_hash(raid5_conf_t
*conf
, struct stripe_head
*sh
)
245 struct hlist_head
*hp
= stripe_hash(conf
, sh
->sector
);
247 pr_debug("insert_hash(), stripe %llu\n",
248 (unsigned long long)sh
->sector
);
251 hlist_add_head(&sh
->hash
, hp
);
255 /* find an idle stripe, make sure it is unhashed, and return it. */
256 static struct stripe_head
*get_free_stripe(raid5_conf_t
*conf
)
258 struct stripe_head
*sh
= NULL
;
259 struct list_head
*first
;
262 if (list_empty(&conf
->inactive_list
))
264 first
= conf
->inactive_list
.next
;
265 sh
= list_entry(first
, struct stripe_head
, lru
);
266 list_del_init(first
);
268 atomic_inc(&conf
->active_stripes
);
273 static void shrink_buffers(struct stripe_head
*sh
, int num
)
278 for (i
=0; i
<num
; i
++) {
282 sh
->dev
[i
].page
= NULL
;
287 static int grow_buffers(struct stripe_head
*sh
, int num
)
291 for (i
=0; i
<num
; i
++) {
294 if (!(page
= alloc_page(GFP_KERNEL
))) {
297 sh
->dev
[i
].page
= page
;
302 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
);
303 static void stripe_set_idx(sector_t stripe
, raid5_conf_t
*conf
, int previous
,
304 struct stripe_head
*sh
);
306 static void init_stripe(struct stripe_head
*sh
, sector_t sector
, int previous
)
308 raid5_conf_t
*conf
= sh
->raid_conf
;
311 BUG_ON(atomic_read(&sh
->count
) != 0);
312 BUG_ON(test_bit(STRIPE_HANDLE
, &sh
->state
));
313 BUG_ON(stripe_operations_active(sh
));
316 pr_debug("init_stripe called, stripe %llu\n",
317 (unsigned long long)sh
->sector
);
321 sh
->generation
= conf
->generation
- previous
;
322 sh
->disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
324 stripe_set_idx(sector
, conf
, previous
, sh
);
328 for (i
= sh
->disks
; i
--; ) {
329 struct r5dev
*dev
= &sh
->dev
[i
];
331 if (dev
->toread
|| dev
->read
|| dev
->towrite
|| dev
->written
||
332 test_bit(R5_LOCKED
, &dev
->flags
)) {
333 printk(KERN_ERR
"sector=%llx i=%d %p %p %p %p %d\n",
334 (unsigned long long)sh
->sector
, i
, dev
->toread
,
335 dev
->read
, dev
->towrite
, dev
->written
,
336 test_bit(R5_LOCKED
, &dev
->flags
));
340 raid5_build_block(sh
, i
, previous
);
342 insert_hash(conf
, sh
);
345 static struct stripe_head
*__find_stripe(raid5_conf_t
*conf
, sector_t sector
,
348 struct stripe_head
*sh
;
349 struct hlist_node
*hn
;
352 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector
);
353 hlist_for_each_entry(sh
, hn
, stripe_hash(conf
, sector
), hash
)
354 if (sh
->sector
== sector
&& sh
->generation
== generation
)
356 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector
);
360 static void unplug_slaves(mddev_t
*mddev
);
361 static void raid5_unplug_device(struct request_queue
*q
);
363 static struct stripe_head
*
364 get_active_stripe(raid5_conf_t
*conf
, sector_t sector
,
365 int previous
, int noblock
)
367 struct stripe_head
*sh
;
369 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector
);
371 spin_lock_irq(&conf
->device_lock
);
374 wait_event_lock_irq(conf
->wait_for_stripe
,
376 conf
->device_lock
, /* nothing */);
377 sh
= __find_stripe(conf
, sector
, conf
->generation
- previous
);
379 if (!conf
->inactive_blocked
)
380 sh
= get_free_stripe(conf
);
381 if (noblock
&& sh
== NULL
)
384 conf
->inactive_blocked
= 1;
385 wait_event_lock_irq(conf
->wait_for_stripe
,
386 !list_empty(&conf
->inactive_list
) &&
387 (atomic_read(&conf
->active_stripes
)
388 < (conf
->max_nr_stripes
*3/4)
389 || !conf
->inactive_blocked
),
391 raid5_unplug_device(conf
->mddev
->queue
)
393 conf
->inactive_blocked
= 0;
395 init_stripe(sh
, sector
, previous
);
397 if (atomic_read(&sh
->count
)) {
398 BUG_ON(!list_empty(&sh
->lru
));
400 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
401 atomic_inc(&conf
->active_stripes
);
402 if (list_empty(&sh
->lru
) &&
403 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
405 list_del_init(&sh
->lru
);
408 } while (sh
== NULL
);
411 atomic_inc(&sh
->count
);
413 spin_unlock_irq(&conf
->device_lock
);
418 raid5_end_read_request(struct bio
*bi
, int error
);
420 raid5_end_write_request(struct bio
*bi
, int error
);
422 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
424 raid5_conf_t
*conf
= sh
->raid_conf
;
425 int i
, disks
= sh
->disks
;
429 for (i
= disks
; i
--; ) {
433 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
))
435 else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
440 bi
= &sh
->dev
[i
].req
;
444 bi
->bi_end_io
= raid5_end_write_request
;
446 bi
->bi_end_io
= raid5_end_read_request
;
449 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
450 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
453 atomic_inc(&rdev
->nr_pending
);
457 if (s
->syncing
|| s
->expanding
|| s
->expanded
)
458 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
460 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
462 bi
->bi_bdev
= rdev
->bdev
;
463 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
464 __func__
, (unsigned long long)sh
->sector
,
466 atomic_inc(&sh
->count
);
467 bi
->bi_sector
= sh
->sector
+ rdev
->data_offset
;
468 bi
->bi_flags
= 1 << BIO_UPTODATE
;
472 bi
->bi_io_vec
= &sh
->dev
[i
].vec
;
473 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
474 bi
->bi_io_vec
[0].bv_offset
= 0;
475 bi
->bi_size
= STRIPE_SIZE
;
478 test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
479 atomic_add(STRIPE_SECTORS
,
480 &rdev
->corrected_errors
);
481 generic_make_request(bi
);
484 set_bit(STRIPE_DEGRADED
, &sh
->state
);
485 pr_debug("skip op %ld on disc %d for sector %llu\n",
486 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
487 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
488 set_bit(STRIPE_HANDLE
, &sh
->state
);
493 static struct dma_async_tx_descriptor
*
494 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
495 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
498 struct page
*bio_page
;
502 if (bio
->bi_sector
>= sector
)
503 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
505 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
506 bio_for_each_segment(bvl
, bio
, i
) {
507 int len
= bio_iovec_idx(bio
, i
)->bv_len
;
511 if (page_offset
< 0) {
512 b_offset
= -page_offset
;
513 page_offset
+= b_offset
;
517 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
518 clen
= STRIPE_SIZE
- page_offset
;
523 b_offset
+= bio_iovec_idx(bio
, i
)->bv_offset
;
524 bio_page
= bio_iovec_idx(bio
, i
)->bv_page
;
526 tx
= async_memcpy(page
, bio_page
, page_offset
,
531 tx
= async_memcpy(bio_page
, page
, b_offset
,
536 if (clen
< len
) /* hit end of page */
544 static void ops_complete_biofill(void *stripe_head_ref
)
546 struct stripe_head
*sh
= stripe_head_ref
;
547 struct bio
*return_bi
= NULL
;
548 raid5_conf_t
*conf
= sh
->raid_conf
;
551 pr_debug("%s: stripe %llu\n", __func__
,
552 (unsigned long long)sh
->sector
);
554 /* clear completed biofills */
555 spin_lock_irq(&conf
->device_lock
);
556 for (i
= sh
->disks
; i
--; ) {
557 struct r5dev
*dev
= &sh
->dev
[i
];
559 /* acknowledge completion of a biofill operation */
560 /* and check if we need to reply to a read request,
561 * new R5_Wantfill requests are held off until
562 * !STRIPE_BIOFILL_RUN
564 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
565 struct bio
*rbi
, *rbi2
;
570 while (rbi
&& rbi
->bi_sector
<
571 dev
->sector
+ STRIPE_SECTORS
) {
572 rbi2
= r5_next_bio(rbi
, dev
->sector
);
573 if (!raid5_dec_bi_phys_segments(rbi
)) {
574 rbi
->bi_next
= return_bi
;
581 spin_unlock_irq(&conf
->device_lock
);
582 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
584 return_io(return_bi
);
586 set_bit(STRIPE_HANDLE
, &sh
->state
);
590 static void ops_run_biofill(struct stripe_head
*sh
)
592 struct dma_async_tx_descriptor
*tx
= NULL
;
593 raid5_conf_t
*conf
= sh
->raid_conf
;
596 pr_debug("%s: stripe %llu\n", __func__
,
597 (unsigned long long)sh
->sector
);
599 for (i
= sh
->disks
; i
--; ) {
600 struct r5dev
*dev
= &sh
->dev
[i
];
601 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
603 spin_lock_irq(&conf
->device_lock
);
604 dev
->read
= rbi
= dev
->toread
;
606 spin_unlock_irq(&conf
->device_lock
);
607 while (rbi
&& rbi
->bi_sector
<
608 dev
->sector
+ STRIPE_SECTORS
) {
609 tx
= async_copy_data(0, rbi
, dev
->page
,
611 rbi
= r5_next_bio(rbi
, dev
->sector
);
616 atomic_inc(&sh
->count
);
617 async_trigger_callback(ASYNC_TX_DEP_ACK
| ASYNC_TX_ACK
, tx
,
618 ops_complete_biofill
, sh
);
621 static void ops_complete_compute5(void *stripe_head_ref
)
623 struct stripe_head
*sh
= stripe_head_ref
;
624 int target
= sh
->ops
.target
;
625 struct r5dev
*tgt
= &sh
->dev
[target
];
627 pr_debug("%s: stripe %llu\n", __func__
,
628 (unsigned long long)sh
->sector
);
630 set_bit(R5_UPTODATE
, &tgt
->flags
);
631 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
632 clear_bit(R5_Wantcompute
, &tgt
->flags
);
633 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
634 if (sh
->check_state
== check_state_compute_run
)
635 sh
->check_state
= check_state_compute_result
;
636 set_bit(STRIPE_HANDLE
, &sh
->state
);
640 static struct dma_async_tx_descriptor
*ops_run_compute5(struct stripe_head
*sh
)
642 /* kernel stack size limits the total number of disks */
643 int disks
= sh
->disks
;
644 struct page
*xor_srcs
[disks
];
645 int target
= sh
->ops
.target
;
646 struct r5dev
*tgt
= &sh
->dev
[target
];
647 struct page
*xor_dest
= tgt
->page
;
649 struct dma_async_tx_descriptor
*tx
;
652 pr_debug("%s: stripe %llu block: %d\n",
653 __func__
, (unsigned long long)sh
->sector
, target
);
654 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
656 for (i
= disks
; i
--; )
658 xor_srcs
[count
++] = sh
->dev
[i
].page
;
660 atomic_inc(&sh
->count
);
662 if (unlikely(count
== 1))
663 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
,
664 0, NULL
, ops_complete_compute5
, sh
);
666 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
667 ASYNC_TX_XOR_ZERO_DST
, NULL
,
668 ops_complete_compute5
, sh
);
673 static void ops_complete_prexor(void *stripe_head_ref
)
675 struct stripe_head
*sh
= stripe_head_ref
;
677 pr_debug("%s: stripe %llu\n", __func__
,
678 (unsigned long long)sh
->sector
);
681 static struct dma_async_tx_descriptor
*
682 ops_run_prexor(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
684 /* kernel stack size limits the total number of disks */
685 int disks
= sh
->disks
;
686 struct page
*xor_srcs
[disks
];
687 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
689 /* existing parity data subtracted */
690 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
692 pr_debug("%s: stripe %llu\n", __func__
,
693 (unsigned long long)sh
->sector
);
695 for (i
= disks
; i
--; ) {
696 struct r5dev
*dev
= &sh
->dev
[i
];
697 /* Only process blocks that are known to be uptodate */
698 if (test_bit(R5_Wantdrain
, &dev
->flags
))
699 xor_srcs
[count
++] = dev
->page
;
702 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
703 ASYNC_TX_DEP_ACK
| ASYNC_TX_XOR_DROP_DST
, tx
,
704 ops_complete_prexor
, sh
);
709 static struct dma_async_tx_descriptor
*
710 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
712 int disks
= sh
->disks
;
715 pr_debug("%s: stripe %llu\n", __func__
,
716 (unsigned long long)sh
->sector
);
718 for (i
= disks
; i
--; ) {
719 struct r5dev
*dev
= &sh
->dev
[i
];
722 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
725 spin_lock(&sh
->lock
);
726 chosen
= dev
->towrite
;
728 BUG_ON(dev
->written
);
729 wbi
= dev
->written
= chosen
;
730 spin_unlock(&sh
->lock
);
732 while (wbi
&& wbi
->bi_sector
<
733 dev
->sector
+ STRIPE_SECTORS
) {
734 tx
= async_copy_data(1, wbi
, dev
->page
,
736 wbi
= r5_next_bio(wbi
, dev
->sector
);
744 static void ops_complete_postxor(void *stripe_head_ref
)
746 struct stripe_head
*sh
= stripe_head_ref
;
747 int disks
= sh
->disks
, i
, pd_idx
= sh
->pd_idx
;
749 pr_debug("%s: stripe %llu\n", __func__
,
750 (unsigned long long)sh
->sector
);
752 for (i
= disks
; i
--; ) {
753 struct r5dev
*dev
= &sh
->dev
[i
];
754 if (dev
->written
|| i
== pd_idx
)
755 set_bit(R5_UPTODATE
, &dev
->flags
);
758 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
759 sh
->reconstruct_state
= reconstruct_state_drain_result
;
760 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
761 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
763 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
764 sh
->reconstruct_state
= reconstruct_state_result
;
767 set_bit(STRIPE_HANDLE
, &sh
->state
);
772 ops_run_postxor(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
774 /* kernel stack size limits the total number of disks */
775 int disks
= sh
->disks
;
776 struct page
*xor_srcs
[disks
];
778 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
779 struct page
*xor_dest
;
783 pr_debug("%s: stripe %llu\n", __func__
,
784 (unsigned long long)sh
->sector
);
786 /* check if prexor is active which means only process blocks
787 * that are part of a read-modify-write (written)
789 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
791 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
792 for (i
= disks
; i
--; ) {
793 struct r5dev
*dev
= &sh
->dev
[i
];
795 xor_srcs
[count
++] = dev
->page
;
798 xor_dest
= sh
->dev
[pd_idx
].page
;
799 for (i
= disks
; i
--; ) {
800 struct r5dev
*dev
= &sh
->dev
[i
];
802 xor_srcs
[count
++] = dev
->page
;
806 /* 1/ if we prexor'd then the dest is reused as a source
807 * 2/ if we did not prexor then we are redoing the parity
808 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
809 * for the synchronous xor case
811 flags
= ASYNC_TX_DEP_ACK
| ASYNC_TX_ACK
|
812 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
814 atomic_inc(&sh
->count
);
816 if (unlikely(count
== 1)) {
817 flags
&= ~(ASYNC_TX_XOR_DROP_DST
| ASYNC_TX_XOR_ZERO_DST
);
818 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
,
819 flags
, tx
, ops_complete_postxor
, sh
);
821 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
822 flags
, tx
, ops_complete_postxor
, sh
);
825 static void ops_complete_check(void *stripe_head_ref
)
827 struct stripe_head
*sh
= stripe_head_ref
;
829 pr_debug("%s: stripe %llu\n", __func__
,
830 (unsigned long long)sh
->sector
);
832 sh
->check_state
= check_state_check_result
;
833 set_bit(STRIPE_HANDLE
, &sh
->state
);
837 static void ops_run_check(struct stripe_head
*sh
)
839 /* kernel stack size limits the total number of disks */
840 int disks
= sh
->disks
;
841 struct page
*xor_srcs
[disks
];
842 struct dma_async_tx_descriptor
*tx
;
844 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
845 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
847 pr_debug("%s: stripe %llu\n", __func__
,
848 (unsigned long long)sh
->sector
);
850 for (i
= disks
; i
--; ) {
851 struct r5dev
*dev
= &sh
->dev
[i
];
853 xor_srcs
[count
++] = dev
->page
;
856 tx
= async_xor_zero_sum(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
857 &sh
->ops
.zero_sum_result
, 0, NULL
, NULL
, NULL
);
859 atomic_inc(&sh
->count
);
860 tx
= async_trigger_callback(ASYNC_TX_DEP_ACK
| ASYNC_TX_ACK
, tx
,
861 ops_complete_check
, sh
);
864 static void raid5_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
866 int overlap_clear
= 0, i
, disks
= sh
->disks
;
867 struct dma_async_tx_descriptor
*tx
= NULL
;
869 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
874 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
875 tx
= ops_run_compute5(sh
);
876 /* terminate the chain if postxor is not set to be run */
877 if (tx
&& !test_bit(STRIPE_OP_POSTXOR
, &ops_request
))
881 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
882 tx
= ops_run_prexor(sh
, tx
);
884 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
885 tx
= ops_run_biodrain(sh
, tx
);
889 if (test_bit(STRIPE_OP_POSTXOR
, &ops_request
))
890 ops_run_postxor(sh
, tx
);
892 if (test_bit(STRIPE_OP_CHECK
, &ops_request
))
896 for (i
= disks
; i
--; ) {
897 struct r5dev
*dev
= &sh
->dev
[i
];
898 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
899 wake_up(&sh
->raid_conf
->wait_for_overlap
);
903 static int grow_one_stripe(raid5_conf_t
*conf
)
905 struct stripe_head
*sh
;
906 sh
= kmem_cache_alloc(conf
->slab_cache
, GFP_KERNEL
);
909 memset(sh
, 0, sizeof(*sh
) + (conf
->raid_disks
-1)*sizeof(struct r5dev
));
910 sh
->raid_conf
= conf
;
911 spin_lock_init(&sh
->lock
);
913 if (grow_buffers(sh
, conf
->raid_disks
)) {
914 shrink_buffers(sh
, conf
->raid_disks
);
915 kmem_cache_free(conf
->slab_cache
, sh
);
918 sh
->disks
= conf
->raid_disks
;
919 /* we just created an active stripe so... */
920 atomic_set(&sh
->count
, 1);
921 atomic_inc(&conf
->active_stripes
);
922 INIT_LIST_HEAD(&sh
->lru
);
927 static int grow_stripes(raid5_conf_t
*conf
, int num
)
929 struct kmem_cache
*sc
;
930 int devs
= conf
->raid_disks
;
932 sprintf(conf
->cache_name
[0],
933 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
934 sprintf(conf
->cache_name
[1],
935 "raid%d-%s-alt", conf
->level
, mdname(conf
->mddev
));
936 conf
->active_name
= 0;
937 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
938 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
942 conf
->slab_cache
= sc
;
943 conf
->pool_size
= devs
;
945 if (!grow_one_stripe(conf
))
950 #ifdef CONFIG_MD_RAID5_RESHAPE
951 static int resize_stripes(raid5_conf_t
*conf
, int newsize
)
953 /* Make all the stripes able to hold 'newsize' devices.
954 * New slots in each stripe get 'page' set to a new page.
956 * This happens in stages:
957 * 1/ create a new kmem_cache and allocate the required number of
959 * 2/ gather all the old stripe_heads and tranfer the pages across
960 * to the new stripe_heads. This will have the side effect of
961 * freezing the array as once all stripe_heads have been collected,
962 * no IO will be possible. Old stripe heads are freed once their
963 * pages have been transferred over, and the old kmem_cache is
964 * freed when all stripes are done.
965 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
966 * we simple return a failre status - no need to clean anything up.
967 * 4/ allocate new pages for the new slots in the new stripe_heads.
968 * If this fails, we don't bother trying the shrink the
969 * stripe_heads down again, we just leave them as they are.
970 * As each stripe_head is processed the new one is released into
973 * Once step2 is started, we cannot afford to wait for a write,
974 * so we use GFP_NOIO allocations.
976 struct stripe_head
*osh
, *nsh
;
977 LIST_HEAD(newstripes
);
978 struct disk_info
*ndisks
;
980 struct kmem_cache
*sc
;
983 if (newsize
<= conf
->pool_size
)
984 return 0; /* never bother to shrink */
986 err
= md_allow_write(conf
->mddev
);
991 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
992 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
997 for (i
= conf
->max_nr_stripes
; i
; i
--) {
998 nsh
= kmem_cache_alloc(sc
, GFP_KERNEL
);
1002 memset(nsh
, 0, sizeof(*nsh
) + (newsize
-1)*sizeof(struct r5dev
));
1004 nsh
->raid_conf
= conf
;
1005 spin_lock_init(&nsh
->lock
);
1007 list_add(&nsh
->lru
, &newstripes
);
1010 /* didn't get enough, give up */
1011 while (!list_empty(&newstripes
)) {
1012 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1013 list_del(&nsh
->lru
);
1014 kmem_cache_free(sc
, nsh
);
1016 kmem_cache_destroy(sc
);
1019 /* Step 2 - Must use GFP_NOIO now.
1020 * OK, we have enough stripes, start collecting inactive
1021 * stripes and copying them over
1023 list_for_each_entry(nsh
, &newstripes
, lru
) {
1024 spin_lock_irq(&conf
->device_lock
);
1025 wait_event_lock_irq(conf
->wait_for_stripe
,
1026 !list_empty(&conf
->inactive_list
),
1028 unplug_slaves(conf
->mddev
)
1030 osh
= get_free_stripe(conf
);
1031 spin_unlock_irq(&conf
->device_lock
);
1032 atomic_set(&nsh
->count
, 1);
1033 for(i
=0; i
<conf
->pool_size
; i
++)
1034 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1035 for( ; i
<newsize
; i
++)
1036 nsh
->dev
[i
].page
= NULL
;
1037 kmem_cache_free(conf
->slab_cache
, osh
);
1039 kmem_cache_destroy(conf
->slab_cache
);
1042 * At this point, we are holding all the stripes so the array
1043 * is completely stalled, so now is a good time to resize
1046 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1048 for (i
=0; i
<conf
->raid_disks
; i
++)
1049 ndisks
[i
] = conf
->disks
[i
];
1051 conf
->disks
= ndisks
;
1055 /* Step 4, return new stripes to service */
1056 while(!list_empty(&newstripes
)) {
1057 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1058 list_del_init(&nsh
->lru
);
1059 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1060 if (nsh
->dev
[i
].page
== NULL
) {
1061 struct page
*p
= alloc_page(GFP_NOIO
);
1062 nsh
->dev
[i
].page
= p
;
1066 release_stripe(nsh
);
1068 /* critical section pass, GFP_NOIO no longer needed */
1070 conf
->slab_cache
= sc
;
1071 conf
->active_name
= 1-conf
->active_name
;
1072 conf
->pool_size
= newsize
;
1077 static int drop_one_stripe(raid5_conf_t
*conf
)
1079 struct stripe_head
*sh
;
1081 spin_lock_irq(&conf
->device_lock
);
1082 sh
= get_free_stripe(conf
);
1083 spin_unlock_irq(&conf
->device_lock
);
1086 BUG_ON(atomic_read(&sh
->count
));
1087 shrink_buffers(sh
, conf
->pool_size
);
1088 kmem_cache_free(conf
->slab_cache
, sh
);
1089 atomic_dec(&conf
->active_stripes
);
1093 static void shrink_stripes(raid5_conf_t
*conf
)
1095 while (drop_one_stripe(conf
))
1098 if (conf
->slab_cache
)
1099 kmem_cache_destroy(conf
->slab_cache
);
1100 conf
->slab_cache
= NULL
;
1103 static void raid5_end_read_request(struct bio
* bi
, int error
)
1105 struct stripe_head
*sh
= bi
->bi_private
;
1106 raid5_conf_t
*conf
= sh
->raid_conf
;
1107 int disks
= sh
->disks
, i
;
1108 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1109 char b
[BDEVNAME_SIZE
];
1113 for (i
=0 ; i
<disks
; i
++)
1114 if (bi
== &sh
->dev
[i
].req
)
1117 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1118 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1126 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1127 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1128 rdev
= conf
->disks
[i
].rdev
;
1129 printk_rl(KERN_INFO
"raid5:%s: read error corrected"
1130 " (%lu sectors at %llu on %s)\n",
1131 mdname(conf
->mddev
), STRIPE_SECTORS
,
1132 (unsigned long long)(sh
->sector
1133 + rdev
->data_offset
),
1134 bdevname(rdev
->bdev
, b
));
1135 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1136 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1138 if (atomic_read(&conf
->disks
[i
].rdev
->read_errors
))
1139 atomic_set(&conf
->disks
[i
].rdev
->read_errors
, 0);
1141 const char *bdn
= bdevname(conf
->disks
[i
].rdev
->bdev
, b
);
1143 rdev
= conf
->disks
[i
].rdev
;
1145 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1146 atomic_inc(&rdev
->read_errors
);
1147 if (conf
->mddev
->degraded
)
1148 printk_rl(KERN_WARNING
1149 "raid5:%s: read error not correctable "
1150 "(sector %llu on %s).\n",
1151 mdname(conf
->mddev
),
1152 (unsigned long long)(sh
->sector
1153 + rdev
->data_offset
),
1155 else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
1157 printk_rl(KERN_WARNING
1158 "raid5:%s: read error NOT corrected!! "
1159 "(sector %llu on %s).\n",
1160 mdname(conf
->mddev
),
1161 (unsigned long long)(sh
->sector
1162 + rdev
->data_offset
),
1164 else if (atomic_read(&rdev
->read_errors
)
1165 > conf
->max_nr_stripes
)
1167 "raid5:%s: Too many read errors, failing device %s.\n",
1168 mdname(conf
->mddev
), bdn
);
1172 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1174 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1175 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1176 md_error(conf
->mddev
, rdev
);
1179 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1180 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1181 set_bit(STRIPE_HANDLE
, &sh
->state
);
1185 static void raid5_end_write_request(struct bio
*bi
, int error
)
1187 struct stripe_head
*sh
= bi
->bi_private
;
1188 raid5_conf_t
*conf
= sh
->raid_conf
;
1189 int disks
= sh
->disks
, i
;
1190 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1192 for (i
=0 ; i
<disks
; i
++)
1193 if (bi
== &sh
->dev
[i
].req
)
1196 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1197 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1205 md_error(conf
->mddev
, conf
->disks
[i
].rdev
);
1207 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1209 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1210 set_bit(STRIPE_HANDLE
, &sh
->state
);
1215 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1217 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1219 struct r5dev
*dev
= &sh
->dev
[i
];
1221 bio_init(&dev
->req
);
1222 dev
->req
.bi_io_vec
= &dev
->vec
;
1224 dev
->req
.bi_max_vecs
++;
1225 dev
->vec
.bv_page
= dev
->page
;
1226 dev
->vec
.bv_len
= STRIPE_SIZE
;
1227 dev
->vec
.bv_offset
= 0;
1229 dev
->req
.bi_sector
= sh
->sector
;
1230 dev
->req
.bi_private
= sh
;
1233 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1236 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1238 char b
[BDEVNAME_SIZE
];
1239 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
1240 pr_debug("raid5: error called\n");
1242 if (!test_bit(Faulty
, &rdev
->flags
)) {
1243 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1244 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1245 unsigned long flags
;
1246 spin_lock_irqsave(&conf
->device_lock
, flags
);
1248 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1250 * if recovery was running, make sure it aborts.
1252 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1254 set_bit(Faulty
, &rdev
->flags
);
1256 "raid5: Disk failure on %s, disabling device.\n"
1257 "raid5: Operation continuing on %d devices.\n",
1258 bdevname(rdev
->bdev
,b
), conf
->raid_disks
- mddev
->degraded
);
1263 * Input: a 'big' sector number,
1264 * Output: index of the data and parity disk, and the sector # in them.
1266 static sector_t
raid5_compute_sector(raid5_conf_t
*conf
, sector_t r_sector
,
1267 int previous
, int *dd_idx
,
1268 struct stripe_head
*sh
)
1271 unsigned long chunk_number
;
1272 unsigned int chunk_offset
;
1275 sector_t new_sector
;
1276 int sectors_per_chunk
= previous
? (conf
->prev_chunk
>> 9)
1277 : (conf
->chunk_size
>> 9);
1278 int raid_disks
= previous
? conf
->previous_raid_disks
1280 int data_disks
= raid_disks
- conf
->max_degraded
;
1282 /* First compute the information on this sector */
1285 * Compute the chunk number and the sector offset inside the chunk
1287 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1288 chunk_number
= r_sector
;
1289 BUG_ON(r_sector
!= chunk_number
);
1292 * Compute the stripe number
1294 stripe
= chunk_number
/ data_disks
;
1297 * Compute the data disk and parity disk indexes inside the stripe
1299 *dd_idx
= chunk_number
% data_disks
;
1302 * Select the parity disk based on the user selected algorithm.
1304 pd_idx
= qd_idx
= ~0;
1305 switch(conf
->level
) {
1307 pd_idx
= data_disks
;
1310 switch (conf
->algorithm
) {
1311 case ALGORITHM_LEFT_ASYMMETRIC
:
1312 pd_idx
= data_disks
- stripe
% raid_disks
;
1313 if (*dd_idx
>= pd_idx
)
1316 case ALGORITHM_RIGHT_ASYMMETRIC
:
1317 pd_idx
= stripe
% raid_disks
;
1318 if (*dd_idx
>= pd_idx
)
1321 case ALGORITHM_LEFT_SYMMETRIC
:
1322 pd_idx
= data_disks
- stripe
% raid_disks
;
1323 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1325 case ALGORITHM_RIGHT_SYMMETRIC
:
1326 pd_idx
= stripe
% raid_disks
;
1327 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1329 case ALGORITHM_PARITY_0
:
1333 case ALGORITHM_PARITY_N
:
1334 pd_idx
= data_disks
;
1337 printk(KERN_ERR
"raid5: unsupported algorithm %d\n",
1344 switch (conf
->algorithm
) {
1345 case ALGORITHM_LEFT_ASYMMETRIC
:
1346 pd_idx
= raid_disks
- 1 - (stripe
% raid_disks
);
1347 qd_idx
= pd_idx
+ 1;
1348 if (pd_idx
== raid_disks
-1) {
1349 (*dd_idx
)++; /* Q D D D P */
1351 } else if (*dd_idx
>= pd_idx
)
1352 (*dd_idx
) += 2; /* D D P Q D */
1354 case ALGORITHM_RIGHT_ASYMMETRIC
:
1355 pd_idx
= stripe
% raid_disks
;
1356 qd_idx
= pd_idx
+ 1;
1357 if (pd_idx
== raid_disks
-1) {
1358 (*dd_idx
)++; /* Q D D D P */
1360 } else if (*dd_idx
>= pd_idx
)
1361 (*dd_idx
) += 2; /* D D P Q D */
1363 case ALGORITHM_LEFT_SYMMETRIC
:
1364 pd_idx
= raid_disks
- 1 - (stripe
% raid_disks
);
1365 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1366 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1368 case ALGORITHM_RIGHT_SYMMETRIC
:
1369 pd_idx
= stripe
% raid_disks
;
1370 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1371 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1374 case ALGORITHM_PARITY_0
:
1379 case ALGORITHM_PARITY_N
:
1380 pd_idx
= data_disks
;
1381 qd_idx
= data_disks
+ 1;
1384 case ALGORITHM_ROTATING_ZERO_RESTART
:
1385 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1386 * of blocks for computing Q is different.
1388 pd_idx
= stripe
% raid_disks
;
1389 qd_idx
= pd_idx
+ 1;
1390 if (pd_idx
== raid_disks
-1) {
1391 (*dd_idx
)++; /* Q D D D P */
1393 } else if (*dd_idx
>= pd_idx
)
1394 (*dd_idx
) += 2; /* D D P Q D */
1398 case ALGORITHM_ROTATING_N_RESTART
:
1399 /* Same a left_asymmetric, by first stripe is
1400 * D D D P Q rather than
1403 pd_idx
= raid_disks
- 1 - ((stripe
+ 1) % raid_disks
);
1404 qd_idx
= pd_idx
+ 1;
1405 if (pd_idx
== raid_disks
-1) {
1406 (*dd_idx
)++; /* Q D D D P */
1408 } else if (*dd_idx
>= pd_idx
)
1409 (*dd_idx
) += 2; /* D D P Q D */
1413 case ALGORITHM_ROTATING_N_CONTINUE
:
1414 /* Same as left_symmetric but Q is before P */
1415 pd_idx
= raid_disks
- 1 - (stripe
% raid_disks
);
1416 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
1417 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1421 case ALGORITHM_LEFT_ASYMMETRIC_6
:
1422 /* RAID5 left_asymmetric, with Q on last device */
1423 pd_idx
= data_disks
- stripe
% (raid_disks
-1);
1424 if (*dd_idx
>= pd_idx
)
1426 qd_idx
= raid_disks
- 1;
1429 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
1430 pd_idx
= stripe
% (raid_disks
-1);
1431 if (*dd_idx
>= pd_idx
)
1433 qd_idx
= raid_disks
- 1;
1436 case ALGORITHM_LEFT_SYMMETRIC_6
:
1437 pd_idx
= data_disks
- stripe
% (raid_disks
-1);
1438 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1439 qd_idx
= raid_disks
- 1;
1442 case ALGORITHM_RIGHT_SYMMETRIC_6
:
1443 pd_idx
= stripe
% (raid_disks
-1);
1444 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1445 qd_idx
= raid_disks
- 1;
1448 case ALGORITHM_PARITY_0_6
:
1451 qd_idx
= raid_disks
- 1;
1456 printk(KERN_CRIT
"raid6: unsupported algorithm %d\n",
1464 sh
->pd_idx
= pd_idx
;
1465 sh
->qd_idx
= qd_idx
;
1466 sh
->ddf_layout
= ddf_layout
;
1469 * Finally, compute the new sector number
1471 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
1476 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
1478 raid5_conf_t
*conf
= sh
->raid_conf
;
1479 int raid_disks
= sh
->disks
;
1480 int data_disks
= raid_disks
- conf
->max_degraded
;
1481 sector_t new_sector
= sh
->sector
, check
;
1482 int sectors_per_chunk
= previous
? (conf
->prev_chunk
>> 9)
1483 : (conf
->chunk_size
>> 9);
1486 int chunk_number
, dummy1
, dd_idx
= i
;
1488 struct stripe_head sh2
;
1491 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
1492 stripe
= new_sector
;
1493 BUG_ON(new_sector
!= stripe
);
1495 if (i
== sh
->pd_idx
)
1497 switch(conf
->level
) {
1500 switch (conf
->algorithm
) {
1501 case ALGORITHM_LEFT_ASYMMETRIC
:
1502 case ALGORITHM_RIGHT_ASYMMETRIC
:
1506 case ALGORITHM_LEFT_SYMMETRIC
:
1507 case ALGORITHM_RIGHT_SYMMETRIC
:
1510 i
-= (sh
->pd_idx
+ 1);
1512 case ALGORITHM_PARITY_0
:
1515 case ALGORITHM_PARITY_N
:
1518 printk(KERN_ERR
"raid5: unsupported algorithm %d\n",
1524 if (i
== sh
->qd_idx
)
1525 return 0; /* It is the Q disk */
1526 switch (conf
->algorithm
) {
1527 case ALGORITHM_LEFT_ASYMMETRIC
:
1528 case ALGORITHM_RIGHT_ASYMMETRIC
:
1529 case ALGORITHM_ROTATING_ZERO_RESTART
:
1530 case ALGORITHM_ROTATING_N_RESTART
:
1531 if (sh
->pd_idx
== raid_disks
-1)
1532 i
--; /* Q D D D P */
1533 else if (i
> sh
->pd_idx
)
1534 i
-= 2; /* D D P Q D */
1536 case ALGORITHM_LEFT_SYMMETRIC
:
1537 case ALGORITHM_RIGHT_SYMMETRIC
:
1538 if (sh
->pd_idx
== raid_disks
-1)
1539 i
--; /* Q D D D P */
1544 i
-= (sh
->pd_idx
+ 2);
1547 case ALGORITHM_PARITY_0
:
1550 case ALGORITHM_PARITY_N
:
1552 case ALGORITHM_ROTATING_N_CONTINUE
:
1553 if (sh
->pd_idx
== 0)
1554 i
--; /* P D D D Q */
1555 else if (i
> sh
->pd_idx
)
1556 i
-= 2; /* D D Q P D */
1558 case ALGORITHM_LEFT_ASYMMETRIC_6
:
1559 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
1563 case ALGORITHM_LEFT_SYMMETRIC_6
:
1564 case ALGORITHM_RIGHT_SYMMETRIC_6
:
1566 i
+= data_disks
+ 1;
1567 i
-= (sh
->pd_idx
+ 1);
1569 case ALGORITHM_PARITY_0_6
:
1573 printk(KERN_CRIT
"raid6: unsupported algorithm %d\n",
1580 chunk_number
= stripe
* data_disks
+ i
;
1581 r_sector
= (sector_t
)chunk_number
* sectors_per_chunk
+ chunk_offset
;
1583 check
= raid5_compute_sector(conf
, r_sector
,
1584 previous
, &dummy1
, &sh2
);
1585 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
1586 || sh2
.qd_idx
!= sh
->qd_idx
) {
1587 printk(KERN_ERR
"compute_blocknr: map not correct\n");
1596 * Copy data between a page in the stripe cache, and one or more bion
1597 * The page could align with the middle of the bio, or there could be
1598 * several bion, each with several bio_vecs, which cover part of the page
1599 * Multiple bion are linked together on bi_next. There may be extras
1600 * at the end of this list. We ignore them.
1602 static void copy_data(int frombio
, struct bio
*bio
,
1606 char *pa
= page_address(page
);
1607 struct bio_vec
*bvl
;
1611 if (bio
->bi_sector
>= sector
)
1612 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
1614 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
1615 bio_for_each_segment(bvl
, bio
, i
) {
1616 int len
= bio_iovec_idx(bio
,i
)->bv_len
;
1620 if (page_offset
< 0) {
1621 b_offset
= -page_offset
;
1622 page_offset
+= b_offset
;
1626 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1627 clen
= STRIPE_SIZE
- page_offset
;
1631 char *ba
= __bio_kmap_atomic(bio
, i
, KM_USER0
);
1633 memcpy(pa
+page_offset
, ba
+b_offset
, clen
);
1635 memcpy(ba
+b_offset
, pa
+page_offset
, clen
);
1636 __bio_kunmap_atomic(ba
, KM_USER0
);
1638 if (clen
< len
) /* hit end of page */
1644 #define check_xor() do { \
1645 if (count == MAX_XOR_BLOCKS) { \
1646 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1651 static void compute_parity6(struct stripe_head
*sh
, int method
)
1653 raid5_conf_t
*conf
= sh
->raid_conf
;
1654 int i
, pd_idx
, qd_idx
, d0_idx
, disks
= sh
->disks
, count
;
1655 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1657 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1658 void *ptrs
[syndrome_disks
+2];
1660 pd_idx
= sh
->pd_idx
;
1661 qd_idx
= sh
->qd_idx
;
1662 d0_idx
= raid6_d0(sh
);
1664 pr_debug("compute_parity, stripe %llu, method %d\n",
1665 (unsigned long long)sh
->sector
, method
);
1668 case READ_MODIFY_WRITE
:
1669 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1670 case RECONSTRUCT_WRITE
:
1671 for (i
= disks
; i
-- ;)
1672 if ( i
!= pd_idx
&& i
!= qd_idx
&& sh
->dev
[i
].towrite
) {
1673 chosen
= sh
->dev
[i
].towrite
;
1674 sh
->dev
[i
].towrite
= NULL
;
1676 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
1677 wake_up(&conf
->wait_for_overlap
);
1679 BUG_ON(sh
->dev
[i
].written
);
1680 sh
->dev
[i
].written
= chosen
;
1684 BUG(); /* Not implemented yet */
1687 for (i
= disks
; i
--;)
1688 if (sh
->dev
[i
].written
) {
1689 sector_t sector
= sh
->dev
[i
].sector
;
1690 struct bio
*wbi
= sh
->dev
[i
].written
;
1691 while (wbi
&& wbi
->bi_sector
< sector
+ STRIPE_SECTORS
) {
1692 copy_data(1, wbi
, sh
->dev
[i
].page
, sector
);
1693 wbi
= r5_next_bio(wbi
, sector
);
1696 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1697 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1700 /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
1702 for (i
= 0; i
< disks
; i
++)
1703 ptrs
[i
] = (void *)raid6_empty_zero_page
;
1708 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1710 ptrs
[slot
] = page_address(sh
->dev
[i
].page
);
1711 if (slot
< syndrome_disks
&&
1712 !test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
)) {
1713 printk(KERN_ERR
"block %d/%d not uptodate "
1714 "on parity calc\n", i
, count
);
1718 i
= raid6_next_disk(i
, disks
);
1719 } while (i
!= d0_idx
);
1720 BUG_ON(count
!= syndrome_disks
);
1722 raid6_call
.gen_syndrome(syndrome_disks
+2, STRIPE_SIZE
, ptrs
);
1725 case RECONSTRUCT_WRITE
:
1726 set_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
1727 set_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
);
1728 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
1729 set_bit(R5_LOCKED
, &sh
->dev
[qd_idx
].flags
);
1732 set_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
1733 set_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
);
1739 /* Compute one missing block */
1740 static void compute_block_1(struct stripe_head
*sh
, int dd_idx
, int nozero
)
1742 int i
, count
, disks
= sh
->disks
;
1743 void *ptr
[MAX_XOR_BLOCKS
], *dest
, *p
;
1744 int qd_idx
= sh
->qd_idx
;
1746 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1747 (unsigned long long)sh
->sector
, dd_idx
);
1749 if ( dd_idx
== qd_idx
) {
1750 /* We're actually computing the Q drive */
1751 compute_parity6(sh
, UPDATE_PARITY
);
1753 dest
= page_address(sh
->dev
[dd_idx
].page
);
1754 if (!nozero
) memset(dest
, 0, STRIPE_SIZE
);
1756 for (i
= disks
; i
--; ) {
1757 if (i
== dd_idx
|| i
== qd_idx
)
1759 p
= page_address(sh
->dev
[i
].page
);
1760 if (test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
))
1763 printk("compute_block() %d, stripe %llu, %d"
1764 " not present\n", dd_idx
,
1765 (unsigned long long)sh
->sector
, i
);
1770 xor_blocks(count
, STRIPE_SIZE
, dest
, ptr
);
1771 if (!nozero
) set_bit(R5_UPTODATE
, &sh
->dev
[dd_idx
].flags
);
1772 else clear_bit(R5_UPTODATE
, &sh
->dev
[dd_idx
].flags
);
1776 /* Compute two missing blocks */
1777 static void compute_block_2(struct stripe_head
*sh
, int dd_idx1
, int dd_idx2
)
1779 int i
, count
, disks
= sh
->disks
;
1780 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1781 int d0_idx
= raid6_d0(sh
);
1782 int faila
= -1, failb
= -1;
1783 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1784 void *ptrs
[syndrome_disks
+2];
1786 for (i
= 0; i
< disks
; i
++)
1787 ptrs
[i
] = (void *)raid6_empty_zero_page
;
1791 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1793 ptrs
[slot
] = page_address(sh
->dev
[i
].page
);
1799 i
= raid6_next_disk(i
, disks
);
1800 } while (i
!= d0_idx
);
1801 BUG_ON(count
!= syndrome_disks
);
1803 BUG_ON(faila
== failb
);
1804 if ( failb
< faila
) { int tmp
= faila
; faila
= failb
; failb
= tmp
; }
1806 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1807 (unsigned long long)sh
->sector
, dd_idx1
, dd_idx2
,
1810 if (failb
== syndrome_disks
+1) {
1811 /* Q disk is one of the missing disks */
1812 if (faila
== syndrome_disks
) {
1813 /* Missing P+Q, just recompute */
1814 compute_parity6(sh
, UPDATE_PARITY
);
1817 /* We're missing D+Q; recompute D from P */
1818 compute_block_1(sh
, ((dd_idx1
== sh
->qd_idx
) ?
1821 compute_parity6(sh
, UPDATE_PARITY
); /* Is this necessary? */
1826 /* We're missing D+P or D+D; */
1827 if (failb
== syndrome_disks
) {
1828 /* We're missing D+P. */
1829 raid6_datap_recov(syndrome_disks
+2, STRIPE_SIZE
, faila
, ptrs
);
1831 /* We're missing D+D. */
1832 raid6_2data_recov(syndrome_disks
+2, STRIPE_SIZE
, faila
, failb
,
1836 /* Both the above update both missing blocks */
1837 set_bit(R5_UPTODATE
, &sh
->dev
[dd_idx1
].flags
);
1838 set_bit(R5_UPTODATE
, &sh
->dev
[dd_idx2
].flags
);
1842 schedule_reconstruction5(struct stripe_head
*sh
, struct stripe_head_state
*s
,
1843 int rcw
, int expand
)
1845 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
1848 /* if we are not expanding this is a proper write request, and
1849 * there will be bios with new data to be drained into the
1853 sh
->reconstruct_state
= reconstruct_state_drain_run
;
1854 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
1856 sh
->reconstruct_state
= reconstruct_state_run
;
1858 set_bit(STRIPE_OP_POSTXOR
, &s
->ops_request
);
1860 for (i
= disks
; i
--; ) {
1861 struct r5dev
*dev
= &sh
->dev
[i
];
1864 set_bit(R5_LOCKED
, &dev
->flags
);
1865 set_bit(R5_Wantdrain
, &dev
->flags
);
1867 clear_bit(R5_UPTODATE
, &dev
->flags
);
1871 if (s
->locked
+ 1 == disks
)
1872 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
1873 atomic_inc(&sh
->raid_conf
->pending_full_writes
);
1875 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
1876 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
1878 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
1879 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
1880 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
1881 set_bit(STRIPE_OP_POSTXOR
, &s
->ops_request
);
1883 for (i
= disks
; i
--; ) {
1884 struct r5dev
*dev
= &sh
->dev
[i
];
1889 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
1890 test_bit(R5_Wantcompute
, &dev
->flags
))) {
1891 set_bit(R5_Wantdrain
, &dev
->flags
);
1892 set_bit(R5_LOCKED
, &dev
->flags
);
1893 clear_bit(R5_UPTODATE
, &dev
->flags
);
1899 /* keep the parity disk locked while asynchronous operations
1902 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
1903 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
1906 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1907 __func__
, (unsigned long long)sh
->sector
,
1908 s
->locked
, s
->ops_request
);
1912 * Each stripe/dev can have one or more bion attached.
1913 * toread/towrite point to the first in a chain.
1914 * The bi_next chain must be in order.
1916 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
1919 raid5_conf_t
*conf
= sh
->raid_conf
;
1922 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1923 (unsigned long long)bi
->bi_sector
,
1924 (unsigned long long)sh
->sector
);
1927 spin_lock(&sh
->lock
);
1928 spin_lock_irq(&conf
->device_lock
);
1930 bip
= &sh
->dev
[dd_idx
].towrite
;
1931 if (*bip
== NULL
&& sh
->dev
[dd_idx
].written
== NULL
)
1934 bip
= &sh
->dev
[dd_idx
].toread
;
1935 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
1936 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
1938 bip
= & (*bip
)->bi_next
;
1940 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
1943 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
1947 bi
->bi_phys_segments
++;
1948 spin_unlock_irq(&conf
->device_lock
);
1949 spin_unlock(&sh
->lock
);
1951 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1952 (unsigned long long)bi
->bi_sector
,
1953 (unsigned long long)sh
->sector
, dd_idx
);
1955 if (conf
->mddev
->bitmap
&& firstwrite
) {
1956 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
1958 sh
->bm_seq
= conf
->seq_flush
+1;
1959 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
1963 /* check if page is covered */
1964 sector_t sector
= sh
->dev
[dd_idx
].sector
;
1965 for (bi
=sh
->dev
[dd_idx
].towrite
;
1966 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
1967 bi
&& bi
->bi_sector
<= sector
;
1968 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
1969 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
1970 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
1972 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
1973 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
1978 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
1979 spin_unlock_irq(&conf
->device_lock
);
1980 spin_unlock(&sh
->lock
);
1984 static void end_reshape(raid5_conf_t
*conf
);
1986 static int page_is_zero(struct page
*p
)
1988 char *a
= page_address(p
);
1989 return ((*(u32
*)a
) == 0 &&
1990 memcmp(a
, a
+4, STRIPE_SIZE
-4)==0);
1993 static void stripe_set_idx(sector_t stripe
, raid5_conf_t
*conf
, int previous
,
1994 struct stripe_head
*sh
)
1996 int sectors_per_chunk
=
1997 previous
? (conf
->prev_chunk
>> 9)
1998 : (conf
->chunk_size
>> 9);
2000 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2001 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2003 raid5_compute_sector(conf
,
2004 stripe
* (disks
- conf
->max_degraded
)
2005 *sectors_per_chunk
+ chunk_offset
,
2011 handle_failed_stripe(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2012 struct stripe_head_state
*s
, int disks
,
2013 struct bio
**return_bi
)
2016 for (i
= disks
; i
--; ) {
2020 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2023 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2024 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2025 /* multiple read failures in one stripe */
2026 md_error(conf
->mddev
, rdev
);
2029 spin_lock_irq(&conf
->device_lock
);
2030 /* fail all writes first */
2031 bi
= sh
->dev
[i
].towrite
;
2032 sh
->dev
[i
].towrite
= NULL
;
2038 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2039 wake_up(&conf
->wait_for_overlap
);
2041 while (bi
&& bi
->bi_sector
<
2042 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2043 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2044 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2045 if (!raid5_dec_bi_phys_segments(bi
)) {
2046 md_write_end(conf
->mddev
);
2047 bi
->bi_next
= *return_bi
;
2052 /* and fail all 'written' */
2053 bi
= sh
->dev
[i
].written
;
2054 sh
->dev
[i
].written
= NULL
;
2055 if (bi
) bitmap_end
= 1;
2056 while (bi
&& bi
->bi_sector
<
2057 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2058 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2059 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2060 if (!raid5_dec_bi_phys_segments(bi
)) {
2061 md_write_end(conf
->mddev
);
2062 bi
->bi_next
= *return_bi
;
2068 /* fail any reads if this device is non-operational and
2069 * the data has not reached the cache yet.
2071 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2072 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2073 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2074 bi
= sh
->dev
[i
].toread
;
2075 sh
->dev
[i
].toread
= NULL
;
2076 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2077 wake_up(&conf
->wait_for_overlap
);
2078 if (bi
) s
->to_read
--;
2079 while (bi
&& bi
->bi_sector
<
2080 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2081 struct bio
*nextbi
=
2082 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2083 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2084 if (!raid5_dec_bi_phys_segments(bi
)) {
2085 bi
->bi_next
= *return_bi
;
2091 spin_unlock_irq(&conf
->device_lock
);
2093 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2094 STRIPE_SECTORS
, 0, 0);
2097 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2098 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2099 md_wakeup_thread(conf
->mddev
->thread
);
2102 /* fetch_block5 - checks the given member device to see if its data needs
2103 * to be read or computed to satisfy a request.
2105 * Returns 1 when no more member devices need to be checked, otherwise returns
2106 * 0 to tell the loop in handle_stripe_fill5 to continue
2108 static int fetch_block5(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2109 int disk_idx
, int disks
)
2111 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2112 struct r5dev
*failed_dev
= &sh
->dev
[s
->failed_num
];
2114 /* is the data in this block needed, and can we get it? */
2115 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2116 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2118 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2119 s
->syncing
|| s
->expanding
||
2121 (failed_dev
->toread
||
2122 (failed_dev
->towrite
&&
2123 !test_bit(R5_OVERWRITE
, &failed_dev
->flags
)))))) {
2124 /* We would like to get this block, possibly by computing it,
2125 * otherwise read it if the backing disk is insync
2127 if ((s
->uptodate
== disks
- 1) &&
2128 (s
->failed
&& disk_idx
== s
->failed_num
)) {
2129 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2130 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2131 set_bit(R5_Wantcompute
, &dev
->flags
);
2132 sh
->ops
.target
= disk_idx
;
2134 /* Careful: from this point on 'uptodate' is in the eye
2135 * of raid5_run_ops which services 'compute' operations
2136 * before writes. R5_Wantcompute flags a block that will
2137 * be R5_UPTODATE by the time it is needed for a
2138 * subsequent operation.
2141 return 1; /* uptodate + compute == disks */
2142 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2143 set_bit(R5_LOCKED
, &dev
->flags
);
2144 set_bit(R5_Wantread
, &dev
->flags
);
2146 pr_debug("Reading block %d (sync=%d)\n", disk_idx
,
2155 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2157 static void handle_stripe_fill5(struct stripe_head
*sh
,
2158 struct stripe_head_state
*s
, int disks
)
2162 /* look for blocks to read/compute, skip this if a compute
2163 * is already in flight, or if the stripe contents are in the
2164 * midst of changing due to a write
2166 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2167 !sh
->reconstruct_state
)
2168 for (i
= disks
; i
--; )
2169 if (fetch_block5(sh
, s
, i
, disks
))
2171 set_bit(STRIPE_HANDLE
, &sh
->state
);
2174 static void handle_stripe_fill6(struct stripe_head
*sh
,
2175 struct stripe_head_state
*s
, struct r6_state
*r6s
,
2179 for (i
= disks
; i
--; ) {
2180 struct r5dev
*dev
= &sh
->dev
[i
];
2181 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2182 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2183 (dev
->toread
|| (dev
->towrite
&&
2184 !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2185 s
->syncing
|| s
->expanding
||
2187 (sh
->dev
[r6s
->failed_num
[0]].toread
||
2190 (sh
->dev
[r6s
->failed_num
[1]].toread
||
2192 /* we would like to get this block, possibly
2193 * by computing it, but we might not be able to
2195 if ((s
->uptodate
== disks
- 1) &&
2196 (s
->failed
&& (i
== r6s
->failed_num
[0] ||
2197 i
== r6s
->failed_num
[1]))) {
2198 pr_debug("Computing stripe %llu block %d\n",
2199 (unsigned long long)sh
->sector
, i
);
2200 compute_block_1(sh
, i
, 0);
2202 } else if ( s
->uptodate
== disks
-2 && s
->failed
>= 2 ) {
2203 /* Computing 2-failure is *very* expensive; only
2204 * do it if failed >= 2
2207 for (other
= disks
; other
--; ) {
2210 if (!test_bit(R5_UPTODATE
,
2211 &sh
->dev
[other
].flags
))
2215 pr_debug("Computing stripe %llu blocks %d,%d\n",
2216 (unsigned long long)sh
->sector
,
2218 compute_block_2(sh
, i
, other
);
2220 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2221 set_bit(R5_LOCKED
, &dev
->flags
);
2222 set_bit(R5_Wantread
, &dev
->flags
);
2224 pr_debug("Reading block %d (sync=%d)\n",
2229 set_bit(STRIPE_HANDLE
, &sh
->state
);
2233 /* handle_stripe_clean_event
2234 * any written block on an uptodate or failed drive can be returned.
2235 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2236 * never LOCKED, so we don't need to test 'failed' directly.
2238 static void handle_stripe_clean_event(raid5_conf_t
*conf
,
2239 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2244 for (i
= disks
; i
--; )
2245 if (sh
->dev
[i
].written
) {
2247 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2248 test_bit(R5_UPTODATE
, &dev
->flags
)) {
2249 /* We can return any write requests */
2250 struct bio
*wbi
, *wbi2
;
2252 pr_debug("Return write for disc %d\n", i
);
2253 spin_lock_irq(&conf
->device_lock
);
2255 dev
->written
= NULL
;
2256 while (wbi
&& wbi
->bi_sector
<
2257 dev
->sector
+ STRIPE_SECTORS
) {
2258 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2259 if (!raid5_dec_bi_phys_segments(wbi
)) {
2260 md_write_end(conf
->mddev
);
2261 wbi
->bi_next
= *return_bi
;
2266 if (dev
->towrite
== NULL
)
2268 spin_unlock_irq(&conf
->device_lock
);
2270 bitmap_endwrite(conf
->mddev
->bitmap
,
2273 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2278 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2279 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2280 md_wakeup_thread(conf
->mddev
->thread
);
2283 static void handle_stripe_dirtying5(raid5_conf_t
*conf
,
2284 struct stripe_head
*sh
, struct stripe_head_state
*s
, int disks
)
2286 int rmw
= 0, rcw
= 0, i
;
2287 for (i
= disks
; i
--; ) {
2288 /* would I have to read this buffer for read_modify_write */
2289 struct r5dev
*dev
= &sh
->dev
[i
];
2290 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2291 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2292 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2293 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2294 if (test_bit(R5_Insync
, &dev
->flags
))
2297 rmw
+= 2*disks
; /* cannot read it */
2299 /* Would I have to read this buffer for reconstruct_write */
2300 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2301 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2302 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2303 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2304 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2309 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2310 (unsigned long long)sh
->sector
, rmw
, rcw
);
2311 set_bit(STRIPE_HANDLE
, &sh
->state
);
2312 if (rmw
< rcw
&& rmw
> 0)
2313 /* prefer read-modify-write, but need to get some data */
2314 for (i
= disks
; i
--; ) {
2315 struct r5dev
*dev
= &sh
->dev
[i
];
2316 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2317 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2318 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2319 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2320 test_bit(R5_Insync
, &dev
->flags
)) {
2322 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2323 pr_debug("Read_old block "
2324 "%d for r-m-w\n", i
);
2325 set_bit(R5_LOCKED
, &dev
->flags
);
2326 set_bit(R5_Wantread
, &dev
->flags
);
2329 set_bit(STRIPE_DELAYED
, &sh
->state
);
2330 set_bit(STRIPE_HANDLE
, &sh
->state
);
2334 if (rcw
<= rmw
&& rcw
> 0)
2335 /* want reconstruct write, but need to get some data */
2336 for (i
= disks
; i
--; ) {
2337 struct r5dev
*dev
= &sh
->dev
[i
];
2338 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2340 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2341 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2342 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2343 test_bit(R5_Insync
, &dev
->flags
)) {
2345 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2346 pr_debug("Read_old block "
2347 "%d for Reconstruct\n", i
);
2348 set_bit(R5_LOCKED
, &dev
->flags
);
2349 set_bit(R5_Wantread
, &dev
->flags
);
2352 set_bit(STRIPE_DELAYED
, &sh
->state
);
2353 set_bit(STRIPE_HANDLE
, &sh
->state
);
2357 /* now if nothing is locked, and if we have enough data,
2358 * we can start a write request
2360 /* since handle_stripe can be called at any time we need to handle the
2361 * case where a compute block operation has been submitted and then a
2362 * subsequent call wants to start a write request. raid5_run_ops only
2363 * handles the case where compute block and postxor are requested
2364 * simultaneously. If this is not the case then new writes need to be
2365 * held off until the compute completes.
2367 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2368 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2369 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2370 schedule_reconstruction5(sh
, s
, rcw
== 0, 0);
2373 static void handle_stripe_dirtying6(raid5_conf_t
*conf
,
2374 struct stripe_head
*sh
, struct stripe_head_state
*s
,
2375 struct r6_state
*r6s
, int disks
)
2377 int rcw
= 0, must_compute
= 0, pd_idx
= sh
->pd_idx
, i
;
2378 int qd_idx
= sh
->qd_idx
;
2379 for (i
= disks
; i
--; ) {
2380 struct r5dev
*dev
= &sh
->dev
[i
];
2381 /* Would I have to read this buffer for reconstruct_write */
2382 if (!test_bit(R5_OVERWRITE
, &dev
->flags
)
2383 && i
!= pd_idx
&& i
!= qd_idx
2384 && (!test_bit(R5_LOCKED
, &dev
->flags
)
2386 !test_bit(R5_UPTODATE
, &dev
->flags
)) {
2387 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2389 pr_debug("raid6: must_compute: "
2390 "disk %d flags=%#lx\n", i
, dev
->flags
);
2395 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2396 (unsigned long long)sh
->sector
, rcw
, must_compute
);
2397 set_bit(STRIPE_HANDLE
, &sh
->state
);
2400 /* want reconstruct write, but need to get some data */
2401 for (i
= disks
; i
--; ) {
2402 struct r5dev
*dev
= &sh
->dev
[i
];
2403 if (!test_bit(R5_OVERWRITE
, &dev
->flags
)
2404 && !(s
->failed
== 0 && (i
== pd_idx
|| i
== qd_idx
))
2405 && !test_bit(R5_LOCKED
, &dev
->flags
) &&
2406 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2407 test_bit(R5_Insync
, &dev
->flags
)) {
2409 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2410 pr_debug("Read_old stripe %llu "
2411 "block %d for Reconstruct\n",
2412 (unsigned long long)sh
->sector
, i
);
2413 set_bit(R5_LOCKED
, &dev
->flags
);
2414 set_bit(R5_Wantread
, &dev
->flags
);
2417 pr_debug("Request delayed stripe %llu "
2418 "block %d for Reconstruct\n",
2419 (unsigned long long)sh
->sector
, i
);
2420 set_bit(STRIPE_DELAYED
, &sh
->state
);
2421 set_bit(STRIPE_HANDLE
, &sh
->state
);
2425 /* now if nothing is locked, and if we have enough data, we can start a
2428 if (s
->locked
== 0 && rcw
== 0 &&
2429 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
2430 if (must_compute
> 0) {
2431 /* We have failed blocks and need to compute them */
2432 switch (s
->failed
) {
2436 compute_block_1(sh
, r6s
->failed_num
[0], 0);
2439 compute_block_2(sh
, r6s
->failed_num
[0],
2440 r6s
->failed_num
[1]);
2442 default: /* This request should have been failed? */
2447 pr_debug("Computing parity for stripe %llu\n",
2448 (unsigned long long)sh
->sector
);
2449 compute_parity6(sh
, RECONSTRUCT_WRITE
);
2450 /* now every locked buffer is ready to be written */
2451 for (i
= disks
; i
--; )
2452 if (test_bit(R5_LOCKED
, &sh
->dev
[i
].flags
)) {
2453 pr_debug("Writing stripe %llu block %d\n",
2454 (unsigned long long)sh
->sector
, i
);
2456 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
2458 if (s
->locked
== disks
)
2459 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2460 atomic_inc(&conf
->pending_full_writes
);
2461 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2462 set_bit(STRIPE_INSYNC
, &sh
->state
);
2464 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2465 atomic_dec(&conf
->preread_active_stripes
);
2466 if (atomic_read(&conf
->preread_active_stripes
) <
2468 md_wakeup_thread(conf
->mddev
->thread
);
2473 static void handle_parity_checks5(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2474 struct stripe_head_state
*s
, int disks
)
2476 struct r5dev
*dev
= NULL
;
2478 set_bit(STRIPE_HANDLE
, &sh
->state
);
2480 switch (sh
->check_state
) {
2481 case check_state_idle
:
2482 /* start a new check operation if there are no failures */
2483 if (s
->failed
== 0) {
2484 BUG_ON(s
->uptodate
!= disks
);
2485 sh
->check_state
= check_state_run
;
2486 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2487 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2491 dev
= &sh
->dev
[s
->failed_num
];
2493 case check_state_compute_result
:
2494 sh
->check_state
= check_state_idle
;
2496 dev
= &sh
->dev
[sh
->pd_idx
];
2498 /* check that a write has not made the stripe insync */
2499 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2502 /* either failed parity check, or recovery is happening */
2503 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2504 BUG_ON(s
->uptodate
!= disks
);
2506 set_bit(R5_LOCKED
, &dev
->flags
);
2508 set_bit(R5_Wantwrite
, &dev
->flags
);
2510 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2511 set_bit(STRIPE_INSYNC
, &sh
->state
);
2513 case check_state_run
:
2514 break; /* we will be called again upon completion */
2515 case check_state_check_result
:
2516 sh
->check_state
= check_state_idle
;
2518 /* if a failure occurred during the check operation, leave
2519 * STRIPE_INSYNC not set and let the stripe be handled again
2524 /* handle a successful check operation, if parity is correct
2525 * we are done. Otherwise update the mismatch count and repair
2526 * parity if !MD_RECOVERY_CHECK
2528 if (sh
->ops
.zero_sum_result
== 0)
2529 /* parity is correct (on disc,
2530 * not in buffer any more)
2532 set_bit(STRIPE_INSYNC
, &sh
->state
);
2534 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2535 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2536 /* don't try to repair!! */
2537 set_bit(STRIPE_INSYNC
, &sh
->state
);
2539 sh
->check_state
= check_state_compute_run
;
2540 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2541 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2542 set_bit(R5_Wantcompute
,
2543 &sh
->dev
[sh
->pd_idx
].flags
);
2544 sh
->ops
.target
= sh
->pd_idx
;
2549 case check_state_compute_run
:
2552 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2553 __func__
, sh
->check_state
,
2554 (unsigned long long) sh
->sector
);
2560 static void handle_parity_checks6(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2561 struct stripe_head_state
*s
,
2562 struct r6_state
*r6s
, struct page
*tmp_page
,
2565 int update_p
= 0, update_q
= 0;
2567 int pd_idx
= sh
->pd_idx
;
2568 int qd_idx
= sh
->qd_idx
;
2570 set_bit(STRIPE_HANDLE
, &sh
->state
);
2572 BUG_ON(s
->failed
> 2);
2573 BUG_ON(s
->uptodate
< disks
);
2574 /* Want to check and possibly repair P and Q.
2575 * However there could be one 'failed' device, in which
2576 * case we can only check one of them, possibly using the
2577 * other to generate missing data
2580 /* If !tmp_page, we cannot do the calculations,
2581 * but as we have set STRIPE_HANDLE, we will soon be called
2582 * by stripe_handle with a tmp_page - just wait until then.
2585 if (s
->failed
== r6s
->q_failed
) {
2586 /* The only possible failed device holds 'Q', so it
2587 * makes sense to check P (If anything else were failed,
2588 * we would have used P to recreate it).
2590 compute_block_1(sh
, pd_idx
, 1);
2591 if (!page_is_zero(sh
->dev
[pd_idx
].page
)) {
2592 compute_block_1(sh
, pd_idx
, 0);
2596 if (!r6s
->q_failed
&& s
->failed
< 2) {
2597 /* q is not failed, and we didn't use it to generate
2598 * anything, so it makes sense to check it
2600 memcpy(page_address(tmp_page
),
2601 page_address(sh
->dev
[qd_idx
].page
),
2603 compute_parity6(sh
, UPDATE_PARITY
);
2604 if (memcmp(page_address(tmp_page
),
2605 page_address(sh
->dev
[qd_idx
].page
),
2606 STRIPE_SIZE
) != 0) {
2607 clear_bit(STRIPE_INSYNC
, &sh
->state
);
2611 if (update_p
|| update_q
) {
2612 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2613 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2614 /* don't try to repair!! */
2615 update_p
= update_q
= 0;
2618 /* now write out any block on a failed drive,
2619 * or P or Q if they need it
2622 if (s
->failed
== 2) {
2623 dev
= &sh
->dev
[r6s
->failed_num
[1]];
2625 set_bit(R5_LOCKED
, &dev
->flags
);
2626 set_bit(R5_Wantwrite
, &dev
->flags
);
2628 if (s
->failed
>= 1) {
2629 dev
= &sh
->dev
[r6s
->failed_num
[0]];
2631 set_bit(R5_LOCKED
, &dev
->flags
);
2632 set_bit(R5_Wantwrite
, &dev
->flags
);
2636 dev
= &sh
->dev
[pd_idx
];
2638 set_bit(R5_LOCKED
, &dev
->flags
);
2639 set_bit(R5_Wantwrite
, &dev
->flags
);
2642 dev
= &sh
->dev
[qd_idx
];
2644 set_bit(R5_LOCKED
, &dev
->flags
);
2645 set_bit(R5_Wantwrite
, &dev
->flags
);
2647 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2649 set_bit(STRIPE_INSYNC
, &sh
->state
);
2653 static void handle_stripe_expansion(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2654 struct r6_state
*r6s
)
2658 /* We have read all the blocks in this stripe and now we need to
2659 * copy some of them into a target stripe for expand.
2661 struct dma_async_tx_descriptor
*tx
= NULL
;
2662 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2663 for (i
= 0; i
< sh
->disks
; i
++)
2664 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
2666 struct stripe_head
*sh2
;
2668 sector_t bn
= compute_blocknr(sh
, i
, 1);
2669 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
2671 sh2
= get_active_stripe(conf
, s
, 0, 1);
2673 /* so far only the early blocks of this stripe
2674 * have been requested. When later blocks
2675 * get requested, we will try again
2678 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
2679 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
2680 /* must have already done this block */
2681 release_stripe(sh2
);
2685 /* place all the copies on one channel */
2686 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
2687 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
2688 ASYNC_TX_DEP_ACK
, tx
, NULL
, NULL
);
2690 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
2691 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
2692 for (j
= 0; j
< conf
->raid_disks
; j
++)
2693 if (j
!= sh2
->pd_idx
&&
2694 (!r6s
|| j
!= sh2
->qd_idx
) &&
2695 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
2697 if (j
== conf
->raid_disks
) {
2698 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
2699 set_bit(STRIPE_HANDLE
, &sh2
->state
);
2701 release_stripe(sh2
);
2704 /* done submitting copies, wait for them to complete */
2707 dma_wait_for_async_tx(tx
);
2713 * handle_stripe - do things to a stripe.
2715 * We lock the stripe and then examine the state of various bits
2716 * to see what needs to be done.
2718 * return some read request which now have data
2719 * return some write requests which are safely on disc
2720 * schedule a read on some buffers
2721 * schedule a write of some buffers
2722 * return confirmation of parity correctness
2724 * buffers are taken off read_list or write_list, and bh_cache buffers
2725 * get BH_Lock set before the stripe lock is released.
2729 static bool handle_stripe5(struct stripe_head
*sh
)
2731 raid5_conf_t
*conf
= sh
->raid_conf
;
2732 int disks
= sh
->disks
, i
;
2733 struct bio
*return_bi
= NULL
;
2734 struct stripe_head_state s
;
2736 mdk_rdev_t
*blocked_rdev
= NULL
;
2739 memset(&s
, 0, sizeof(s
));
2740 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2741 "reconstruct:%d\n", (unsigned long long)sh
->sector
, sh
->state
,
2742 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->check_state
,
2743 sh
->reconstruct_state
);
2745 spin_lock(&sh
->lock
);
2746 clear_bit(STRIPE_HANDLE
, &sh
->state
);
2747 clear_bit(STRIPE_DELAYED
, &sh
->state
);
2749 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
2750 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2751 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
2753 /* Now to look around and see what can be done */
2755 for (i
=disks
; i
--; ) {
2757 struct r5dev
*dev
= &sh
->dev
[i
];
2758 clear_bit(R5_Insync
, &dev
->flags
);
2760 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2761 "written %p\n", i
, dev
->flags
, dev
->toread
, dev
->read
,
2762 dev
->towrite
, dev
->written
);
2764 /* maybe we can request a biofill operation
2766 * new wantfill requests are only permitted while
2767 * ops_complete_biofill is guaranteed to be inactive
2769 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
2770 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
2771 set_bit(R5_Wantfill
, &dev
->flags
);
2773 /* now count some things */
2774 if (test_bit(R5_LOCKED
, &dev
->flags
)) s
.locked
++;
2775 if (test_bit(R5_UPTODATE
, &dev
->flags
)) s
.uptodate
++;
2776 if (test_bit(R5_Wantcompute
, &dev
->flags
)) s
.compute
++;
2778 if (test_bit(R5_Wantfill
, &dev
->flags
))
2780 else if (dev
->toread
)
2784 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
2789 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2790 if (blocked_rdev
== NULL
&&
2791 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
2792 blocked_rdev
= rdev
;
2793 atomic_inc(&rdev
->nr_pending
);
2795 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)) {
2796 /* The ReadError flag will just be confusing now */
2797 clear_bit(R5_ReadError
, &dev
->flags
);
2798 clear_bit(R5_ReWrite
, &dev
->flags
);
2800 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)
2801 || test_bit(R5_ReadError
, &dev
->flags
)) {
2805 set_bit(R5_Insync
, &dev
->flags
);
2809 if (unlikely(blocked_rdev
)) {
2810 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
2811 s
.to_write
|| s
.written
) {
2812 set_bit(STRIPE_HANDLE
, &sh
->state
);
2815 /* There is nothing for the blocked_rdev to block */
2816 rdev_dec_pending(blocked_rdev
, conf
->mddev
);
2817 blocked_rdev
= NULL
;
2820 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
2821 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
2822 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
2825 pr_debug("locked=%d uptodate=%d to_read=%d"
2826 " to_write=%d failed=%d failed_num=%d\n",
2827 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
,
2828 s
.failed
, s
.failed_num
);
2829 /* check if the array has lost two devices and, if so, some requests might
2832 if (s
.failed
> 1 && s
.to_read
+s
.to_write
+s
.written
)
2833 handle_failed_stripe(conf
, sh
, &s
, disks
, &return_bi
);
2834 if (s
.failed
> 1 && s
.syncing
) {
2835 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
2836 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2840 /* might be able to return some write requests if the parity block
2841 * is safe, or on a failed drive
2843 dev
= &sh
->dev
[sh
->pd_idx
];
2845 ((test_bit(R5_Insync
, &dev
->flags
) &&
2846 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2847 test_bit(R5_UPTODATE
, &dev
->flags
)) ||
2848 (s
.failed
== 1 && s
.failed_num
== sh
->pd_idx
)))
2849 handle_stripe_clean_event(conf
, sh
, disks
, &return_bi
);
2851 /* Now we might consider reading some blocks, either to check/generate
2852 * parity, or to satisfy requests
2853 * or to load a block that is being partially written.
2855 if (s
.to_read
|| s
.non_overwrite
||
2856 (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
)) || s
.expanding
)
2857 handle_stripe_fill5(sh
, &s
, disks
);
2859 /* Now we check to see if any write operations have recently
2863 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
2865 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
2866 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
2867 sh
->reconstruct_state
= reconstruct_state_idle
;
2869 /* All the 'written' buffers and the parity block are ready to
2870 * be written back to disk
2872 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
2873 for (i
= disks
; i
--; ) {
2875 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
2876 (i
== sh
->pd_idx
|| dev
->written
)) {
2877 pr_debug("Writing block %d\n", i
);
2878 set_bit(R5_Wantwrite
, &dev
->flags
);
2881 if (!test_bit(R5_Insync
, &dev
->flags
) ||
2882 (i
== sh
->pd_idx
&& s
.failed
== 0))
2883 set_bit(STRIPE_INSYNC
, &sh
->state
);
2886 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2887 atomic_dec(&conf
->preread_active_stripes
);
2888 if (atomic_read(&conf
->preread_active_stripes
) <
2890 md_wakeup_thread(conf
->mddev
->thread
);
2894 /* Now to consider new write requests and what else, if anything
2895 * should be read. We do not handle new writes when:
2896 * 1/ A 'write' operation (copy+xor) is already in flight.
2897 * 2/ A 'check' operation is in flight, as it may clobber the parity
2900 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
2901 handle_stripe_dirtying5(conf
, sh
, &s
, disks
);
2903 /* maybe we need to check and possibly fix the parity for this stripe
2904 * Any reads will already have been scheduled, so we just see if enough
2905 * data is available. The parity check is held off while parity
2906 * dependent operations are in flight.
2908 if (sh
->check_state
||
2909 (s
.syncing
&& s
.locked
== 0 &&
2910 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
2911 !test_bit(STRIPE_INSYNC
, &sh
->state
)))
2912 handle_parity_checks5(conf
, sh
, &s
, disks
);
2914 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
2915 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,1);
2916 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2919 /* If the failed drive is just a ReadError, then we might need to progress
2920 * the repair/check process
2922 if (s
.failed
== 1 && !conf
->mddev
->ro
&&
2923 test_bit(R5_ReadError
, &sh
->dev
[s
.failed_num
].flags
)
2924 && !test_bit(R5_LOCKED
, &sh
->dev
[s
.failed_num
].flags
)
2925 && test_bit(R5_UPTODATE
, &sh
->dev
[s
.failed_num
].flags
)
2927 dev
= &sh
->dev
[s
.failed_num
];
2928 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
2929 set_bit(R5_Wantwrite
, &dev
->flags
);
2930 set_bit(R5_ReWrite
, &dev
->flags
);
2931 set_bit(R5_LOCKED
, &dev
->flags
);
2934 /* let's read it back */
2935 set_bit(R5_Wantread
, &dev
->flags
);
2936 set_bit(R5_LOCKED
, &dev
->flags
);
2941 /* Finish reconstruct operations initiated by the expansion process */
2942 if (sh
->reconstruct_state
== reconstruct_state_result
) {
2943 sh
->reconstruct_state
= reconstruct_state_idle
;
2944 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
2945 for (i
= conf
->raid_disks
; i
--; ) {
2946 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
2947 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2952 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
2953 !sh
->reconstruct_state
) {
2954 /* Need to write out all blocks after computing parity */
2955 sh
->disks
= conf
->raid_disks
;
2956 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
2957 schedule_reconstruction5(sh
, &s
, 1, 1);
2958 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
2959 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
2960 atomic_dec(&conf
->reshape_stripes
);
2961 wake_up(&conf
->wait_for_overlap
);
2962 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
2965 if (s
.expanding
&& s
.locked
== 0 &&
2966 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
2967 handle_stripe_expansion(conf
, sh
, NULL
);
2970 spin_unlock(&sh
->lock
);
2972 /* wait for this device to become unblocked */
2973 if (unlikely(blocked_rdev
))
2974 md_wait_for_blocked_rdev(blocked_rdev
, conf
->mddev
);
2977 raid5_run_ops(sh
, s
.ops_request
);
2981 return_io(return_bi
);
2983 return blocked_rdev
== NULL
;
2986 static bool handle_stripe6(struct stripe_head
*sh
, struct page
*tmp_page
)
2988 raid5_conf_t
*conf
= sh
->raid_conf
;
2989 int disks
= sh
->disks
;
2990 struct bio
*return_bi
= NULL
;
2991 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
;
2992 struct stripe_head_state s
;
2993 struct r6_state r6s
;
2994 struct r5dev
*dev
, *pdev
, *qdev
;
2995 mdk_rdev_t
*blocked_rdev
= NULL
;
2997 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2998 "pd_idx=%d, qd_idx=%d\n",
2999 (unsigned long long)sh
->sector
, sh
->state
,
3000 atomic_read(&sh
->count
), pd_idx
, qd_idx
);
3001 memset(&s
, 0, sizeof(s
));
3003 spin_lock(&sh
->lock
);
3004 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3005 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3007 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
3008 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3009 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3010 /* Now to look around and see what can be done */
3013 for (i
=disks
; i
--; ) {
3016 clear_bit(R5_Insync
, &dev
->flags
);
3018 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3019 i
, dev
->flags
, dev
->toread
, dev
->towrite
, dev
->written
);
3020 /* maybe we can reply to a read */
3021 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
) {
3022 struct bio
*rbi
, *rbi2
;
3023 pr_debug("Return read for disc %d\n", i
);
3024 spin_lock_irq(&conf
->device_lock
);
3027 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
3028 wake_up(&conf
->wait_for_overlap
);
3029 spin_unlock_irq(&conf
->device_lock
);
3030 while (rbi
&& rbi
->bi_sector
< dev
->sector
+ STRIPE_SECTORS
) {
3031 copy_data(0, rbi
, dev
->page
, dev
->sector
);
3032 rbi2
= r5_next_bio(rbi
, dev
->sector
);
3033 spin_lock_irq(&conf
->device_lock
);
3034 if (!raid5_dec_bi_phys_segments(rbi
)) {
3035 rbi
->bi_next
= return_bi
;
3038 spin_unlock_irq(&conf
->device_lock
);
3043 /* now count some things */
3044 if (test_bit(R5_LOCKED
, &dev
->flags
)) s
.locked
++;
3045 if (test_bit(R5_UPTODATE
, &dev
->flags
)) s
.uptodate
++;
3052 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3057 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3058 if (blocked_rdev
== NULL
&&
3059 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
3060 blocked_rdev
= rdev
;
3061 atomic_inc(&rdev
->nr_pending
);
3063 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)) {
3064 /* The ReadError flag will just be confusing now */
3065 clear_bit(R5_ReadError
, &dev
->flags
);
3066 clear_bit(R5_ReWrite
, &dev
->flags
);
3068 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)
3069 || test_bit(R5_ReadError
, &dev
->flags
)) {
3071 r6s
.failed_num
[s
.failed
] = i
;
3074 set_bit(R5_Insync
, &dev
->flags
);
3078 if (unlikely(blocked_rdev
)) {
3079 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3080 s
.to_write
|| s
.written
) {
3081 set_bit(STRIPE_HANDLE
, &sh
->state
);
3084 /* There is nothing for the blocked_rdev to block */
3085 rdev_dec_pending(blocked_rdev
, conf
->mddev
);
3086 blocked_rdev
= NULL
;
3089 pr_debug("locked=%d uptodate=%d to_read=%d"
3090 " to_write=%d failed=%d failed_num=%d,%d\n",
3091 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3092 r6s
.failed_num
[0], r6s
.failed_num
[1]);
3093 /* check if the array has lost >2 devices and, if so, some requests
3094 * might need to be failed
3096 if (s
.failed
> 2 && s
.to_read
+s
.to_write
+s
.written
)
3097 handle_failed_stripe(conf
, sh
, &s
, disks
, &return_bi
);
3098 if (s
.failed
> 2 && s
.syncing
) {
3099 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
3100 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3105 * might be able to return some write requests if the parity blocks
3106 * are safe, or on a failed drive
3108 pdev
= &sh
->dev
[pd_idx
];
3109 r6s
.p_failed
= (s
.failed
>= 1 && r6s
.failed_num
[0] == pd_idx
)
3110 || (s
.failed
>= 2 && r6s
.failed_num
[1] == pd_idx
);
3111 qdev
= &sh
->dev
[qd_idx
];
3112 r6s
.q_failed
= (s
.failed
>= 1 && r6s
.failed_num
[0] == qd_idx
)
3113 || (s
.failed
>= 2 && r6s
.failed_num
[1] == qd_idx
);
3116 ( r6s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3117 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3118 && test_bit(R5_UPTODATE
, &pdev
->flags
)))) &&
3119 ( r6s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3120 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3121 && test_bit(R5_UPTODATE
, &qdev
->flags
)))))
3122 handle_stripe_clean_event(conf
, sh
, disks
, &return_bi
);
3124 /* Now we might consider reading some blocks, either to check/generate
3125 * parity, or to satisfy requests
3126 * or to load a block that is being partially written.
3128 if (s
.to_read
|| s
.non_overwrite
|| (s
.to_write
&& s
.failed
) ||
3129 (s
.syncing
&& (s
.uptodate
< disks
)) || s
.expanding
)
3130 handle_stripe_fill6(sh
, &s
, &r6s
, disks
);
3132 /* now to consider writing and what else, if anything should be read */
3134 handle_stripe_dirtying6(conf
, sh
, &s
, &r6s
, disks
);
3136 /* maybe we need to check and possibly fix the parity for this stripe
3137 * Any reads will already have been scheduled, so we just see if enough
3140 if (s
.syncing
&& s
.locked
== 0 && !test_bit(STRIPE_INSYNC
, &sh
->state
))
3141 handle_parity_checks6(conf
, sh
, &s
, &r6s
, tmp_page
, disks
);
3143 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3144 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,1);
3145 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3148 /* If the failed drives are just a ReadError, then we might need
3149 * to progress the repair/check process
3151 if (s
.failed
<= 2 && !conf
->mddev
->ro
)
3152 for (i
= 0; i
< s
.failed
; i
++) {
3153 dev
= &sh
->dev
[r6s
.failed_num
[i
]];
3154 if (test_bit(R5_ReadError
, &dev
->flags
)
3155 && !test_bit(R5_LOCKED
, &dev
->flags
)
3156 && test_bit(R5_UPTODATE
, &dev
->flags
)
3158 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3159 set_bit(R5_Wantwrite
, &dev
->flags
);
3160 set_bit(R5_ReWrite
, &dev
->flags
);
3161 set_bit(R5_LOCKED
, &dev
->flags
);
3163 /* let's read it back */
3164 set_bit(R5_Wantread
, &dev
->flags
);
3165 set_bit(R5_LOCKED
, &dev
->flags
);
3170 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
3171 /* Need to write out all blocks after computing P&Q */
3172 sh
->disks
= conf
->raid_disks
;
3173 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3174 compute_parity6(sh
, RECONSTRUCT_WRITE
);
3175 for (i
= conf
->raid_disks
; i
-- ; ) {
3176 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3178 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3180 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3181 } else if (s
.expanded
) {
3182 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3183 atomic_dec(&conf
->reshape_stripes
);
3184 wake_up(&conf
->wait_for_overlap
);
3185 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3188 if (s
.expanding
&& s
.locked
== 0 &&
3189 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3190 handle_stripe_expansion(conf
, sh
, &r6s
);
3193 spin_unlock(&sh
->lock
);
3195 /* wait for this device to become unblocked */
3196 if (unlikely(blocked_rdev
))
3197 md_wait_for_blocked_rdev(blocked_rdev
, conf
->mddev
);
3201 return_io(return_bi
);
3203 return blocked_rdev
== NULL
;
3206 /* returns true if the stripe was handled */
3207 static bool handle_stripe(struct stripe_head
*sh
, struct page
*tmp_page
)
3209 if (sh
->raid_conf
->level
== 6)
3210 return handle_stripe6(sh
, tmp_page
);
3212 return handle_stripe5(sh
);
3217 static void raid5_activate_delayed(raid5_conf_t
*conf
)
3219 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3220 while (!list_empty(&conf
->delayed_list
)) {
3221 struct list_head
*l
= conf
->delayed_list
.next
;
3222 struct stripe_head
*sh
;
3223 sh
= list_entry(l
, struct stripe_head
, lru
);
3225 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3226 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3227 atomic_inc(&conf
->preread_active_stripes
);
3228 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3231 blk_plug_device(conf
->mddev
->queue
);
3234 static void activate_bit_delay(raid5_conf_t
*conf
)
3236 /* device_lock is held */
3237 struct list_head head
;
3238 list_add(&head
, &conf
->bitmap_list
);
3239 list_del_init(&conf
->bitmap_list
);
3240 while (!list_empty(&head
)) {
3241 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3242 list_del_init(&sh
->lru
);
3243 atomic_inc(&sh
->count
);
3244 __release_stripe(conf
, sh
);
3248 static void unplug_slaves(mddev_t
*mddev
)
3250 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3254 for (i
=0; i
<mddev
->raid_disks
; i
++) {
3255 mdk_rdev_t
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3256 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) && atomic_read(&rdev
->nr_pending
)) {
3257 struct request_queue
*r_queue
= bdev_get_queue(rdev
->bdev
);
3259 atomic_inc(&rdev
->nr_pending
);
3262 blk_unplug(r_queue
);
3264 rdev_dec_pending(rdev
, mddev
);
3271 static void raid5_unplug_device(struct request_queue
*q
)
3273 mddev_t
*mddev
= q
->queuedata
;
3274 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3275 unsigned long flags
;
3277 spin_lock_irqsave(&conf
->device_lock
, flags
);
3279 if (blk_remove_plug(q
)) {
3281 raid5_activate_delayed(conf
);
3283 md_wakeup_thread(mddev
->thread
);
3285 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3287 unplug_slaves(mddev
);
3290 static int raid5_congested(void *data
, int bits
)
3292 mddev_t
*mddev
= data
;
3293 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3295 /* No difference between reads and writes. Just check
3296 * how busy the stripe_cache is
3298 if (conf
->inactive_blocked
)
3302 if (list_empty_careful(&conf
->inactive_list
))
3308 /* We want read requests to align with chunks where possible,
3309 * but write requests don't need to.
3311 static int raid5_mergeable_bvec(struct request_queue
*q
,
3312 struct bvec_merge_data
*bvm
,
3313 struct bio_vec
*biovec
)
3315 mddev_t
*mddev
= q
->queuedata
;
3316 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3318 unsigned int chunk_sectors
= mddev
->chunk_size
>> 9;
3319 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3321 if ((bvm
->bi_rw
& 1) == WRITE
)
3322 return biovec
->bv_len
; /* always allow writes to be mergeable */
3324 if (mddev
->new_chunk
< mddev
->chunk_size
)
3325 chunk_sectors
= mddev
->new_chunk
>> 9;
3326 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3327 if (max
< 0) max
= 0;
3328 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3329 return biovec
->bv_len
;
3335 static int in_chunk_boundary(mddev_t
*mddev
, struct bio
*bio
)
3337 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3338 unsigned int chunk_sectors
= mddev
->chunk_size
>> 9;
3339 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3341 if (mddev
->new_chunk
< mddev
->chunk_size
)
3342 chunk_sectors
= mddev
->new_chunk
>> 9;
3343 return chunk_sectors
>=
3344 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3348 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3349 * later sampled by raid5d.
3351 static void add_bio_to_retry(struct bio
*bi
,raid5_conf_t
*conf
)
3353 unsigned long flags
;
3355 spin_lock_irqsave(&conf
->device_lock
, flags
);
3357 bi
->bi_next
= conf
->retry_read_aligned_list
;
3358 conf
->retry_read_aligned_list
= bi
;
3360 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3361 md_wakeup_thread(conf
->mddev
->thread
);
3365 static struct bio
*remove_bio_from_retry(raid5_conf_t
*conf
)
3369 bi
= conf
->retry_read_aligned
;
3371 conf
->retry_read_aligned
= NULL
;
3374 bi
= conf
->retry_read_aligned_list
;
3376 conf
->retry_read_aligned_list
= bi
->bi_next
;
3379 * this sets the active strip count to 1 and the processed
3380 * strip count to zero (upper 8 bits)
3382 bi
->bi_phys_segments
= 1; /* biased count of active 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
)
3435 if (q
->merge_bvec_fn
)
3436 /* it's too hard to apply the merge_bvec_fn at this stage,
3445 static int chunk_aligned_read(struct request_queue
*q
, struct bio
* raid_bio
)
3447 mddev_t
*mddev
= q
->queuedata
;
3448 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3449 unsigned int dd_idx
;
3450 struct bio
* align_bi
;
3453 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3454 pr_debug("chunk_aligned_read : non aligned\n");
3458 * use bio_clone to make a copy of the bio
3460 align_bi
= bio_clone(raid_bio
, GFP_NOIO
);
3464 * set bi_end_io to a new function, and set bi_private to the
3467 align_bi
->bi_end_io
= raid5_align_endio
;
3468 align_bi
->bi_private
= raid_bio
;
3472 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3477 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3478 if (rdev
&& test_bit(In_sync
, &rdev
->flags
)) {
3479 atomic_inc(&rdev
->nr_pending
);
3481 raid_bio
->bi_next
= (void*)rdev
;
3482 align_bi
->bi_bdev
= rdev
->bdev
;
3483 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3484 align_bi
->bi_sector
+= rdev
->data_offset
;
3486 if (!bio_fits_rdev(align_bi
)) {
3487 /* too big in some way */
3489 rdev_dec_pending(rdev
, mddev
);
3493 spin_lock_irq(&conf
->device_lock
);
3494 wait_event_lock_irq(conf
->wait_for_stripe
,
3496 conf
->device_lock
, /* nothing */);
3497 atomic_inc(&conf
->active_aligned_reads
);
3498 spin_unlock_irq(&conf
->device_lock
);
3500 generic_make_request(align_bi
);
3509 /* __get_priority_stripe - get the next stripe to process
3511 * Full stripe writes are allowed to pass preread active stripes up until
3512 * the bypass_threshold is exceeded. In general the bypass_count
3513 * increments when the handle_list is handled before the hold_list; however, it
3514 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3515 * stripe with in flight i/o. The bypass_count will be reset when the
3516 * head of the hold_list has changed, i.e. the head was promoted to the
3519 static struct stripe_head
*__get_priority_stripe(raid5_conf_t
*conf
)
3521 struct stripe_head
*sh
;
3523 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3525 list_empty(&conf
->handle_list
) ? "empty" : "busy",
3526 list_empty(&conf
->hold_list
) ? "empty" : "busy",
3527 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
3529 if (!list_empty(&conf
->handle_list
)) {
3530 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
3532 if (list_empty(&conf
->hold_list
))
3533 conf
->bypass_count
= 0;
3534 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
3535 if (conf
->hold_list
.next
== conf
->last_hold
)
3536 conf
->bypass_count
++;
3538 conf
->last_hold
= conf
->hold_list
.next
;
3539 conf
->bypass_count
-= conf
->bypass_threshold
;
3540 if (conf
->bypass_count
< 0)
3541 conf
->bypass_count
= 0;
3544 } else if (!list_empty(&conf
->hold_list
) &&
3545 ((conf
->bypass_threshold
&&
3546 conf
->bypass_count
> conf
->bypass_threshold
) ||
3547 atomic_read(&conf
->pending_full_writes
) == 0)) {
3548 sh
= list_entry(conf
->hold_list
.next
,
3550 conf
->bypass_count
-= conf
->bypass_threshold
;
3551 if (conf
->bypass_count
< 0)
3552 conf
->bypass_count
= 0;
3556 list_del_init(&sh
->lru
);
3557 atomic_inc(&sh
->count
);
3558 BUG_ON(atomic_read(&sh
->count
) != 1);
3562 static int make_request(struct request_queue
*q
, struct bio
* bi
)
3564 mddev_t
*mddev
= q
->queuedata
;
3565 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3567 sector_t new_sector
;
3568 sector_t logical_sector
, last_sector
;
3569 struct stripe_head
*sh
;
3570 const int rw
= bio_data_dir(bi
);
3573 if (unlikely(bio_barrier(bi
))) {
3574 bio_endio(bi
, -EOPNOTSUPP
);
3578 md_write_start(mddev
, bi
);
3580 cpu
= part_stat_lock();
3581 part_stat_inc(cpu
, &mddev
->gendisk
->part0
, ios
[rw
]);
3582 part_stat_add(cpu
, &mddev
->gendisk
->part0
, sectors
[rw
],
3587 mddev
->reshape_position
== MaxSector
&&
3588 chunk_aligned_read(q
,bi
))
3591 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
3592 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
3594 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
3596 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
3598 int disks
, data_disks
;
3603 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
3604 if (likely(conf
->reshape_progress
== MaxSector
))
3605 disks
= conf
->raid_disks
;
3607 /* spinlock is needed as reshape_progress may be
3608 * 64bit on a 32bit platform, and so it might be
3609 * possible to see a half-updated value
3610 * Ofcourse reshape_progress could change after
3611 * the lock is dropped, so once we get a reference
3612 * to the stripe that we think it is, we will have
3615 spin_lock_irq(&conf
->device_lock
);
3616 disks
= conf
->raid_disks
;
3617 if (mddev
->delta_disks
< 0
3618 ? logical_sector
< conf
->reshape_progress
3619 : logical_sector
>= conf
->reshape_progress
) {
3620 disks
= conf
->previous_raid_disks
;
3623 if (mddev
->delta_disks
< 0
3624 ? logical_sector
< conf
->reshape_safe
3625 : logical_sector
>= conf
->reshape_safe
) {
3626 spin_unlock_irq(&conf
->device_lock
);
3631 spin_unlock_irq(&conf
->device_lock
);
3633 data_disks
= disks
- conf
->max_degraded
;
3635 new_sector
= raid5_compute_sector(conf
, logical_sector
,
3638 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3639 (unsigned long long)new_sector
,
3640 (unsigned long long)logical_sector
);
3642 sh
= get_active_stripe(conf
, new_sector
, previous
,
3643 (bi
->bi_rw
&RWA_MASK
));
3645 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
3646 /* expansion might have moved on while waiting for a
3647 * stripe, so we must do the range check again.
3648 * Expansion could still move past after this
3649 * test, but as we are holding a reference to
3650 * 'sh', we know that if that happens,
3651 * STRIPE_EXPANDING will get set and the expansion
3652 * won't proceed until we finish with the stripe.
3655 spin_lock_irq(&conf
->device_lock
);
3656 if ((mddev
->delta_disks
< 0
3657 ? logical_sector
>= conf
->reshape_progress
3658 : logical_sector
< conf
->reshape_progress
)
3660 /* mismatch, need to try again */
3662 spin_unlock_irq(&conf
->device_lock
);
3668 /* FIXME what if we get a false positive because these
3669 * are being updated.
3671 if (logical_sector
>= mddev
->suspend_lo
&&
3672 logical_sector
< mddev
->suspend_hi
) {
3678 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
3679 !add_stripe_bio(sh
, bi
, dd_idx
, (bi
->bi_rw
&RW_MASK
))) {
3680 /* Stripe is busy expanding or
3681 * add failed due to overlap. Flush everything
3684 raid5_unplug_device(mddev
->queue
);
3689 finish_wait(&conf
->wait_for_overlap
, &w
);
3690 set_bit(STRIPE_HANDLE
, &sh
->state
);
3691 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3694 /* cannot get stripe for read-ahead, just give-up */
3695 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3696 finish_wait(&conf
->wait_for_overlap
, &w
);
3701 spin_lock_irq(&conf
->device_lock
);
3702 remaining
= raid5_dec_bi_phys_segments(bi
);
3703 spin_unlock_irq(&conf
->device_lock
);
3704 if (remaining
== 0) {
3707 md_write_end(mddev
);
3714 static sector_t
raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
);
3716 static sector_t
reshape_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
)
3718 /* reshaping is quite different to recovery/resync so it is
3719 * handled quite separately ... here.
3721 * On each call to sync_request, we gather one chunk worth of
3722 * destination stripes and flag them as expanding.
3723 * Then we find all the source stripes and request reads.
3724 * As the reads complete, handle_stripe will copy the data
3725 * into the destination stripe and release that stripe.
3727 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
3728 struct stripe_head
*sh
;
3729 sector_t first_sector
, last_sector
;
3730 int raid_disks
= conf
->previous_raid_disks
;
3731 int data_disks
= raid_disks
- conf
->max_degraded
;
3732 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
3735 sector_t writepos
, safepos
, gap
;
3736 sector_t stripe_addr
;
3738 if (sector_nr
== 0) {
3739 /* If restarting in the middle, skip the initial sectors */
3740 if (mddev
->delta_disks
< 0 &&
3741 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
3742 sector_nr
= raid5_size(mddev
, 0, 0)
3743 - conf
->reshape_progress
;
3744 } else if (mddev
->delta_disks
> 0 &&
3745 conf
->reshape_progress
> 0)
3746 sector_nr
= conf
->reshape_progress
;
3747 sector_div(sector_nr
, new_data_disks
);
3754 /* we update the metadata when there is more than 3Meg
3755 * in the block range (that is rather arbitrary, should
3756 * probably be time based) or when the data about to be
3757 * copied would over-write the source of the data at
3758 * the front of the range.
3759 * i.e. one new_stripe along from reshape_progress new_maps
3760 * to after where reshape_safe old_maps to
3762 writepos
= conf
->reshape_progress
;
3763 sector_div(writepos
, new_data_disks
);
3764 safepos
= conf
->reshape_safe
;
3765 sector_div(safepos
, data_disks
);
3766 if (mddev
->delta_disks
< 0) {
3767 writepos
-= conf
->chunk_size
/512;
3768 safepos
+= conf
->chunk_size
/512;
3769 gap
= conf
->reshape_safe
- conf
->reshape_progress
;
3771 writepos
+= conf
->chunk_size
/512;
3772 safepos
-= conf
->chunk_size
/512;
3773 gap
= conf
->reshape_progress
- conf
->reshape_safe
;
3776 if ((mddev
->delta_disks
< 0
3777 ? writepos
< safepos
3778 : writepos
> safepos
) ||
3779 gap
> (new_data_disks
)*3000*2 /*3Meg*/) {
3780 /* Cannot proceed until we've updated the superblock... */
3781 wait_event(conf
->wait_for_overlap
,
3782 atomic_read(&conf
->reshape_stripes
)==0);
3783 mddev
->reshape_position
= conf
->reshape_progress
;
3784 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
3785 md_wakeup_thread(mddev
->thread
);
3786 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
3787 kthread_should_stop());
3788 spin_lock_irq(&conf
->device_lock
);
3789 conf
->reshape_safe
= mddev
->reshape_position
;
3790 spin_unlock_irq(&conf
->device_lock
);
3791 wake_up(&conf
->wait_for_overlap
);
3794 if (mddev
->delta_disks
< 0) {
3795 BUG_ON(conf
->reshape_progress
== 0);
3796 stripe_addr
= writepos
;
3797 BUG_ON((mddev
->dev_sectors
&
3798 ~((sector_t
)conf
->chunk_size
/ 512 - 1))
3799 - (conf
->chunk_size
/ 512) - stripe_addr
3802 BUG_ON(writepos
!= sector_nr
+ conf
->chunk_size
/ 512);
3803 stripe_addr
= sector_nr
;
3805 for (i
=0; i
< conf
->chunk_size
/512; i
+= STRIPE_SECTORS
) {
3808 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0);
3809 set_bit(STRIPE_EXPANDING
, &sh
->state
);
3810 atomic_inc(&conf
->reshape_stripes
);
3811 /* If any of this stripe is beyond the end of the old
3812 * array, then we need to zero those blocks
3814 for (j
=sh
->disks
; j
--;) {
3816 if (j
== sh
->pd_idx
)
3818 if (conf
->level
== 6 &&
3821 s
= compute_blocknr(sh
, j
, 0);
3822 if (s
< raid5_size(mddev
, 0, 0)) {
3826 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
3827 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
3828 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
3831 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3832 set_bit(STRIPE_HANDLE
, &sh
->state
);
3836 spin_lock_irq(&conf
->device_lock
);
3837 if (mddev
->delta_disks
< 0)
3838 conf
->reshape_progress
-= i
* new_data_disks
;
3840 conf
->reshape_progress
+= i
* new_data_disks
;
3841 spin_unlock_irq(&conf
->device_lock
);
3842 /* Ok, those stripe are ready. We can start scheduling
3843 * reads on the source stripes.
3844 * The source stripes are determined by mapping the first and last
3845 * block on the destination stripes.
3848 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
3851 raid5_compute_sector(conf
, ((stripe_addr
+conf
->chunk_size
/512)
3852 *(new_data_disks
) - 1),
3854 if (last_sector
>= mddev
->dev_sectors
)
3855 last_sector
= mddev
->dev_sectors
- 1;
3856 while (first_sector
<= last_sector
) {
3857 sh
= get_active_stripe(conf
, first_sector
, 1, 0);
3858 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3859 set_bit(STRIPE_HANDLE
, &sh
->state
);
3861 first_sector
+= STRIPE_SECTORS
;
3863 /* If this takes us to the resync_max point where we have to pause,
3864 * then we need to write out the superblock.
3866 sector_nr
+= conf
->chunk_size
>>9;
3867 if (sector_nr
>= mddev
->resync_max
) {
3868 /* Cannot proceed until we've updated the superblock... */
3869 wait_event(conf
->wait_for_overlap
,
3870 atomic_read(&conf
->reshape_stripes
) == 0);
3871 mddev
->reshape_position
= conf
->reshape_progress
;
3872 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
3873 md_wakeup_thread(mddev
->thread
);
3874 wait_event(mddev
->sb_wait
,
3875 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
3876 || kthread_should_stop());
3877 spin_lock_irq(&conf
->device_lock
);
3878 conf
->reshape_safe
= mddev
->reshape_position
;
3879 spin_unlock_irq(&conf
->device_lock
);
3880 wake_up(&conf
->wait_for_overlap
);
3882 return conf
->chunk_size
>>9;
3885 /* FIXME go_faster isn't used */
3886 static inline sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
3888 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
3889 struct stripe_head
*sh
;
3890 sector_t max_sector
= mddev
->dev_sectors
;
3892 int still_degraded
= 0;
3895 if (sector_nr
>= max_sector
) {
3896 /* just being told to finish up .. nothing much to do */
3897 unplug_slaves(mddev
);
3899 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
3904 if (mddev
->curr_resync
< max_sector
) /* aborted */
3905 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
3907 else /* completed sync */
3909 bitmap_close_sync(mddev
->bitmap
);
3914 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
3915 return reshape_request(mddev
, sector_nr
, skipped
);
3917 /* No need to check resync_max as we never do more than one
3918 * stripe, and as resync_max will always be on a chunk boundary,
3919 * if the check in md_do_sync didn't fire, there is no chance
3920 * of overstepping resync_max here
3923 /* if there is too many failed drives and we are trying
3924 * to resync, then assert that we are finished, because there is
3925 * nothing we can do.
3927 if (mddev
->degraded
>= conf
->max_degraded
&&
3928 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3929 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
3933 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
3934 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
3935 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
3936 /* we can skip this block, and probably more */
3937 sync_blocks
/= STRIPE_SECTORS
;
3939 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
3943 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3945 sh
= get_active_stripe(conf
, sector_nr
, 0, 1);
3947 sh
= get_active_stripe(conf
, sector_nr
, 0, 0);
3948 /* make sure we don't swamp the stripe cache if someone else
3949 * is trying to get access
3951 schedule_timeout_uninterruptible(1);
3953 /* Need to check if array will still be degraded after recovery/resync
3954 * We don't need to check the 'failed' flag as when that gets set,
3957 for (i
=0; i
<mddev
->raid_disks
; i
++)
3958 if (conf
->disks
[i
].rdev
== NULL
)
3961 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
3963 spin_lock(&sh
->lock
);
3964 set_bit(STRIPE_SYNCING
, &sh
->state
);
3965 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3966 spin_unlock(&sh
->lock
);
3968 /* wait for any blocked device to be handled */
3969 while(unlikely(!handle_stripe(sh
, NULL
)))
3973 return STRIPE_SECTORS
;
3976 static int retry_aligned_read(raid5_conf_t
*conf
, struct bio
*raid_bio
)
3978 /* We may not be able to submit a whole bio at once as there
3979 * may not be enough stripe_heads available.
3980 * We cannot pre-allocate enough stripe_heads as we may need
3981 * more than exist in the cache (if we allow ever large chunks).
3982 * So we do one stripe head at a time and record in
3983 * ->bi_hw_segments how many have been done.
3985 * We *know* that this entire raid_bio is in one chunk, so
3986 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3988 struct stripe_head
*sh
;
3990 sector_t sector
, logical_sector
, last_sector
;
3995 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
3996 sector
= raid5_compute_sector(conf
, logical_sector
,
3998 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4000 for (; logical_sector
< last_sector
;
4001 logical_sector
+= STRIPE_SECTORS
,
4002 sector
+= STRIPE_SECTORS
,
4005 if (scnt
< raid5_bi_hw_segments(raid_bio
))
4006 /* already done this stripe */
4009 sh
= get_active_stripe(conf
, sector
, 0, 1);
4012 /* failed to get a stripe - must wait */
4013 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4014 conf
->retry_read_aligned
= raid_bio
;
4018 set_bit(R5_ReadError
, &sh
->dev
[dd_idx
].flags
);
4019 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4021 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4022 conf
->retry_read_aligned
= raid_bio
;
4026 handle_stripe(sh
, NULL
);
4030 spin_lock_irq(&conf
->device_lock
);
4031 remaining
= raid5_dec_bi_phys_segments(raid_bio
);
4032 spin_unlock_irq(&conf
->device_lock
);
4034 bio_endio(raid_bio
, 0);
4035 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4036 wake_up(&conf
->wait_for_stripe
);
4043 * This is our raid5 kernel thread.
4045 * We scan the hash table for stripes which can be handled now.
4046 * During the scan, completed stripes are saved for us by the interrupt
4047 * handler, so that they will not have to wait for our next wakeup.
4049 static void raid5d(mddev_t
*mddev
)
4051 struct stripe_head
*sh
;
4052 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4055 pr_debug("+++ raid5d active\n");
4057 md_check_recovery(mddev
);
4060 spin_lock_irq(&conf
->device_lock
);
4064 if (conf
->seq_flush
!= conf
->seq_write
) {
4065 int seq
= conf
->seq_flush
;
4066 spin_unlock_irq(&conf
->device_lock
);
4067 bitmap_unplug(mddev
->bitmap
);
4068 spin_lock_irq(&conf
->device_lock
);
4069 conf
->seq_write
= seq
;
4070 activate_bit_delay(conf
);
4073 while ((bio
= remove_bio_from_retry(conf
))) {
4075 spin_unlock_irq(&conf
->device_lock
);
4076 ok
= retry_aligned_read(conf
, bio
);
4077 spin_lock_irq(&conf
->device_lock
);
4083 sh
= __get_priority_stripe(conf
);
4087 spin_unlock_irq(&conf
->device_lock
);
4090 handle_stripe(sh
, conf
->spare_page
);
4093 spin_lock_irq(&conf
->device_lock
);
4095 pr_debug("%d stripes handled\n", handled
);
4097 spin_unlock_irq(&conf
->device_lock
);
4099 async_tx_issue_pending_all();
4100 unplug_slaves(mddev
);
4102 pr_debug("--- raid5d inactive\n");
4106 raid5_show_stripe_cache_size(mddev_t
*mddev
, char *page
)
4108 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4110 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4116 raid5_store_stripe_cache_size(mddev_t
*mddev
, const char *page
, size_t len
)
4118 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4122 if (len
>= PAGE_SIZE
)
4127 if (strict_strtoul(page
, 10, &new))
4129 if (new <= 16 || new > 32768)
4131 while (new < conf
->max_nr_stripes
) {
4132 if (drop_one_stripe(conf
))
4133 conf
->max_nr_stripes
--;
4137 err
= md_allow_write(mddev
);
4140 while (new > conf
->max_nr_stripes
) {
4141 if (grow_one_stripe(conf
))
4142 conf
->max_nr_stripes
++;
4148 static struct md_sysfs_entry
4149 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4150 raid5_show_stripe_cache_size
,
4151 raid5_store_stripe_cache_size
);
4154 raid5_show_preread_threshold(mddev_t
*mddev
, char *page
)
4156 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4158 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4164 raid5_store_preread_threshold(mddev_t
*mddev
, const char *page
, size_t len
)
4166 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4168 if (len
>= PAGE_SIZE
)
4173 if (strict_strtoul(page
, 10, &new))
4175 if (new > conf
->max_nr_stripes
)
4177 conf
->bypass_threshold
= new;
4181 static struct md_sysfs_entry
4182 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4184 raid5_show_preread_threshold
,
4185 raid5_store_preread_threshold
);
4188 stripe_cache_active_show(mddev_t
*mddev
, char *page
)
4190 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4192 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4197 static struct md_sysfs_entry
4198 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4200 static struct attribute
*raid5_attrs
[] = {
4201 &raid5_stripecache_size
.attr
,
4202 &raid5_stripecache_active
.attr
,
4203 &raid5_preread_bypass_threshold
.attr
,
4206 static struct attribute_group raid5_attrs_group
= {
4208 .attrs
= raid5_attrs
,
4212 raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
4214 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4217 sectors
= mddev
->dev_sectors
;
4219 /* size is defined by the smallest of previous and new size */
4220 if (conf
->raid_disks
< conf
->previous_raid_disks
)
4221 raid_disks
= conf
->raid_disks
;
4223 raid_disks
= conf
->previous_raid_disks
;
4226 sectors
&= ~((sector_t
)mddev
->chunk_size
/512 - 1);
4227 sectors
&= ~((sector_t
)mddev
->new_chunk
/512 - 1);
4228 return sectors
* (raid_disks
- conf
->max_degraded
);
4231 static raid5_conf_t
*setup_conf(mddev_t
*mddev
)
4234 int raid_disk
, memory
;
4236 struct disk_info
*disk
;
4238 if (mddev
->new_level
!= 5
4239 && mddev
->new_level
!= 4
4240 && mddev
->new_level
!= 6) {
4241 printk(KERN_ERR
"raid5: %s: raid level not set to 4/5/6 (%d)\n",
4242 mdname(mddev
), mddev
->new_level
);
4243 return ERR_PTR(-EIO
);
4245 if ((mddev
->new_level
== 5
4246 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
4247 (mddev
->new_level
== 6
4248 && !algorithm_valid_raid6(mddev
->new_layout
))) {
4249 printk(KERN_ERR
"raid5: %s: layout %d not supported\n",
4250 mdname(mddev
), mddev
->new_layout
);
4251 return ERR_PTR(-EIO
);
4253 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
4254 printk(KERN_ERR
"raid6: not enough configured devices for %s (%d, minimum 4)\n",
4255 mdname(mddev
), mddev
->raid_disks
);
4256 return ERR_PTR(-EINVAL
);
4259 if (!mddev
->new_chunk
|| mddev
->new_chunk
% PAGE_SIZE
) {
4260 printk(KERN_ERR
"raid5: invalid chunk size %d for %s\n",
4261 mddev
->new_chunk
, mdname(mddev
));
4262 return ERR_PTR(-EINVAL
);
4265 conf
= kzalloc(sizeof(raid5_conf_t
), GFP_KERNEL
);
4269 conf
->raid_disks
= mddev
->raid_disks
;
4270 if (mddev
->reshape_position
== MaxSector
)
4271 conf
->previous_raid_disks
= mddev
->raid_disks
;
4273 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4275 conf
->disks
= kzalloc(conf
->raid_disks
* sizeof(struct disk_info
),
4280 conf
->mddev
= mddev
;
4282 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
4285 if (mddev
->new_level
== 6) {
4286 conf
->spare_page
= alloc_page(GFP_KERNEL
);
4287 if (!conf
->spare_page
)
4290 spin_lock_init(&conf
->device_lock
);
4291 init_waitqueue_head(&conf
->wait_for_stripe
);
4292 init_waitqueue_head(&conf
->wait_for_overlap
);
4293 INIT_LIST_HEAD(&conf
->handle_list
);
4294 INIT_LIST_HEAD(&conf
->hold_list
);
4295 INIT_LIST_HEAD(&conf
->delayed_list
);
4296 INIT_LIST_HEAD(&conf
->bitmap_list
);
4297 INIT_LIST_HEAD(&conf
->inactive_list
);
4298 atomic_set(&conf
->active_stripes
, 0);
4299 atomic_set(&conf
->preread_active_stripes
, 0);
4300 atomic_set(&conf
->active_aligned_reads
, 0);
4301 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
4303 pr_debug("raid5: run(%s) called.\n", mdname(mddev
));
4305 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
4306 raid_disk
= rdev
->raid_disk
;
4307 if (raid_disk
>= conf
->raid_disks
4310 disk
= conf
->disks
+ raid_disk
;
4314 if (test_bit(In_sync
, &rdev
->flags
)) {
4315 char b
[BDEVNAME_SIZE
];
4316 printk(KERN_INFO
"raid5: device %s operational as raid"
4317 " disk %d\n", bdevname(rdev
->bdev
,b
),
4320 /* Cannot rely on bitmap to complete recovery */
4324 conf
->chunk_size
= mddev
->new_chunk
;
4325 conf
->level
= mddev
->new_level
;
4326 if (conf
->level
== 6)
4327 conf
->max_degraded
= 2;
4329 conf
->max_degraded
= 1;
4330 conf
->algorithm
= mddev
->new_layout
;
4331 conf
->max_nr_stripes
= NR_STRIPES
;
4332 conf
->reshape_progress
= mddev
->reshape_position
;
4333 if (conf
->reshape_progress
!= MaxSector
)
4334 conf
->prev_chunk
= mddev
->chunk_size
;
4336 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
4337 conf
->raid_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
4338 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
4340 "raid5: couldn't allocate %dkB for buffers\n", memory
);
4343 printk(KERN_INFO
"raid5: allocated %dkB for %s\n",
4344 memory
, mdname(mddev
));
4346 conf
->thread
= md_register_thread(raid5d
, mddev
, "%s_raid5");
4347 if (!conf
->thread
) {
4349 "raid5: couldn't allocate thread for %s\n",
4358 shrink_stripes(conf
);
4359 safe_put_page(conf
->spare_page
);
4361 kfree(conf
->stripe_hashtbl
);
4363 return ERR_PTR(-EIO
);
4365 return ERR_PTR(-ENOMEM
);
4368 static int run(mddev_t
*mddev
)
4371 int working_disks
= 0;
4374 if (mddev
->reshape_position
!= MaxSector
) {
4375 /* Check that we can continue the reshape.
4376 * Currently only disks can change, it must
4377 * increase, and we must be past the point where
4378 * a stripe over-writes itself
4380 sector_t here_new
, here_old
;
4382 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
4384 if (mddev
->new_level
!= mddev
->level
||
4385 mddev
->new_layout
!= mddev
->layout
||
4386 mddev
->new_chunk
!= mddev
->chunk_size
) {
4387 printk(KERN_ERR
"raid5: %s: unsupported reshape "
4388 "required - aborting.\n",
4392 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4393 /* reshape_position must be on a new-stripe boundary, and one
4394 * further up in new geometry must map after here in old
4397 here_new
= mddev
->reshape_position
;
4398 if (sector_div(here_new
, (mddev
->new_chunk
>>9)*
4399 (mddev
->raid_disks
- max_degraded
))) {
4400 printk(KERN_ERR
"raid5: reshape_position not "
4401 "on a stripe boundary\n");
4404 /* here_new is the stripe we will write to */
4405 here_old
= mddev
->reshape_position
;
4406 sector_div(here_old
, (mddev
->chunk_size
>>9)*
4407 (old_disks
-max_degraded
));
4408 /* here_old is the first stripe that we might need to read
4410 if (here_new
>= here_old
) {
4411 /* Reading from the same stripe as writing to - bad */
4412 printk(KERN_ERR
"raid5: reshape_position too early for "
4413 "auto-recovery - aborting.\n");
4416 printk(KERN_INFO
"raid5: reshape will continue\n");
4417 /* OK, we should be able to continue; */
4419 BUG_ON(mddev
->level
!= mddev
->new_level
);
4420 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
4421 BUG_ON(mddev
->chunk_size
!= mddev
->new_chunk
);
4422 BUG_ON(mddev
->delta_disks
!= 0);
4425 if (mddev
->private == NULL
)
4426 conf
= setup_conf(mddev
);
4428 conf
= mddev
->private;
4431 return PTR_ERR(conf
);
4433 mddev
->thread
= conf
->thread
;
4434 conf
->thread
= NULL
;
4435 mddev
->private = conf
;
4438 * 0 for a fully functional array, 1 or 2 for a degraded array.
4440 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
4441 if (rdev
->raid_disk
>= 0 &&
4442 test_bit(In_sync
, &rdev
->flags
))
4445 mddev
->degraded
= conf
->raid_disks
- working_disks
;
4447 if (mddev
->degraded
> conf
->max_degraded
) {
4448 printk(KERN_ERR
"raid5: not enough operational devices for %s"
4449 " (%d/%d failed)\n",
4450 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
4454 /* device size must be a multiple of chunk size */
4455 mddev
->dev_sectors
&= ~(mddev
->chunk_size
/ 512 - 1);
4456 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
4458 if (mddev
->degraded
> 0 &&
4459 mddev
->recovery_cp
!= MaxSector
) {
4460 if (mddev
->ok_start_degraded
)
4462 "raid5: starting dirty degraded array: %s"
4463 "- data corruption possible.\n",
4467 "raid5: cannot start dirty degraded array for %s\n",
4473 if (mddev
->degraded
== 0)
4474 printk("raid5: raid level %d set %s active with %d out of %d"
4475 " devices, algorithm %d\n", conf
->level
, mdname(mddev
),
4476 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
4479 printk(KERN_ALERT
"raid5: raid level %d set %s active with %d"
4480 " out of %d devices, algorithm %d\n", conf
->level
,
4481 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
4482 mddev
->raid_disks
, conf
->algorithm
);
4484 print_raid5_conf(conf
);
4486 if (conf
->reshape_progress
!= MaxSector
) {
4487 printk("...ok start reshape thread\n");
4488 conf
->reshape_safe
= conf
->reshape_progress
;
4489 atomic_set(&conf
->reshape_stripes
, 0);
4490 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4491 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4492 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4493 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4494 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4498 /* read-ahead size must cover two whole stripes, which is
4499 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4502 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
4503 int stripe
= data_disks
*
4504 (mddev
->chunk_size
/ PAGE_SIZE
);
4505 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4506 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4509 /* Ok, everything is just fine now */
4510 if (sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
4512 "raid5: failed to create sysfs attributes for %s\n",
4515 mddev
->queue
->queue_lock
= &conf
->device_lock
;
4517 mddev
->queue
->unplug_fn
= raid5_unplug_device
;
4518 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
4519 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
4521 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
4523 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
4527 md_unregister_thread(mddev
->thread
);
4528 mddev
->thread
= NULL
;
4530 shrink_stripes(conf
);
4531 print_raid5_conf(conf
);
4532 safe_put_page(conf
->spare_page
);
4534 kfree(conf
->stripe_hashtbl
);
4537 mddev
->private = NULL
;
4538 printk(KERN_ALERT
"raid5: failed to run raid set %s\n", mdname(mddev
));
4544 static int stop(mddev_t
*mddev
)
4546 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
4548 md_unregister_thread(mddev
->thread
);
4549 mddev
->thread
= NULL
;
4550 shrink_stripes(conf
);
4551 kfree(conf
->stripe_hashtbl
);
4552 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
4553 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
4554 sysfs_remove_group(&mddev
->kobj
, &raid5_attrs_group
);
4557 mddev
->private = NULL
;
4562 static void print_sh(struct seq_file
*seq
, struct stripe_head
*sh
)
4566 seq_printf(seq
, "sh %llu, pd_idx %d, state %ld.\n",
4567 (unsigned long long)sh
->sector
, sh
->pd_idx
, sh
->state
);
4568 seq_printf(seq
, "sh %llu, count %d.\n",
4569 (unsigned long long)sh
->sector
, atomic_read(&sh
->count
));
4570 seq_printf(seq
, "sh %llu, ", (unsigned long long)sh
->sector
);
4571 for (i
= 0; i
< sh
->disks
; i
++) {
4572 seq_printf(seq
, "(cache%d: %p %ld) ",
4573 i
, sh
->dev
[i
].page
, sh
->dev
[i
].flags
);
4575 seq_printf(seq
, "\n");
4578 static void printall(struct seq_file
*seq
, raid5_conf_t
*conf
)
4580 struct stripe_head
*sh
;
4581 struct hlist_node
*hn
;
4584 spin_lock_irq(&conf
->device_lock
);
4585 for (i
= 0; i
< NR_HASH
; i
++) {
4586 hlist_for_each_entry(sh
, hn
, &conf
->stripe_hashtbl
[i
], hash
) {
4587 if (sh
->raid_conf
!= conf
)
4592 spin_unlock_irq(&conf
->device_lock
);
4596 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
4598 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
4601 seq_printf (seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
, mddev
->chunk_size
>> 10, mddev
->layout
);
4602 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
4603 for (i
= 0; i
< conf
->raid_disks
; i
++)
4604 seq_printf (seq
, "%s",
4605 conf
->disks
[i
].rdev
&&
4606 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
4607 seq_printf (seq
, "]");
4609 seq_printf (seq
, "\n");
4610 printall(seq
, conf
);
4614 static void print_raid5_conf (raid5_conf_t
*conf
)
4617 struct disk_info
*tmp
;
4619 printk("RAID5 conf printout:\n");
4621 printk("(conf==NULL)\n");
4624 printk(" --- rd:%d wd:%d\n", conf
->raid_disks
,
4625 conf
->raid_disks
- conf
->mddev
->degraded
);
4627 for (i
= 0; i
< conf
->raid_disks
; i
++) {
4628 char b
[BDEVNAME_SIZE
];
4629 tmp
= conf
->disks
+ i
;
4631 printk(" disk %d, o:%d, dev:%s\n",
4632 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
4633 bdevname(tmp
->rdev
->bdev
,b
));
4637 static int raid5_spare_active(mddev_t
*mddev
)
4640 raid5_conf_t
*conf
= mddev
->private;
4641 struct disk_info
*tmp
;
4643 for (i
= 0; i
< conf
->raid_disks
; i
++) {
4644 tmp
= conf
->disks
+ i
;
4646 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
4647 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
4648 unsigned long flags
;
4649 spin_lock_irqsave(&conf
->device_lock
, flags
);
4651 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4654 print_raid5_conf(conf
);
4658 static int raid5_remove_disk(mddev_t
*mddev
, int number
)
4660 raid5_conf_t
*conf
= mddev
->private;
4663 struct disk_info
*p
= conf
->disks
+ number
;
4665 print_raid5_conf(conf
);
4668 if (number
>= conf
->raid_disks
&&
4669 conf
->reshape_progress
== MaxSector
)
4670 clear_bit(In_sync
, &rdev
->flags
);
4672 if (test_bit(In_sync
, &rdev
->flags
) ||
4673 atomic_read(&rdev
->nr_pending
)) {
4677 /* Only remove non-faulty devices if recovery
4680 if (!test_bit(Faulty
, &rdev
->flags
) &&
4681 mddev
->degraded
<= conf
->max_degraded
&&
4682 number
< conf
->raid_disks
) {
4688 if (atomic_read(&rdev
->nr_pending
)) {
4689 /* lost the race, try later */
4696 print_raid5_conf(conf
);
4700 static int raid5_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
4702 raid5_conf_t
*conf
= mddev
->private;
4705 struct disk_info
*p
;
4707 int last
= conf
->raid_disks
- 1;
4709 if (mddev
->degraded
> conf
->max_degraded
)
4710 /* no point adding a device */
4713 if (rdev
->raid_disk
>= 0)
4714 first
= last
= rdev
->raid_disk
;
4717 * find the disk ... but prefer rdev->saved_raid_disk
4720 if (rdev
->saved_raid_disk
>= 0 &&
4721 rdev
->saved_raid_disk
>= first
&&
4722 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
4723 disk
= rdev
->saved_raid_disk
;
4726 for ( ; disk
<= last
; disk
++)
4727 if ((p
=conf
->disks
+ disk
)->rdev
== NULL
) {
4728 clear_bit(In_sync
, &rdev
->flags
);
4729 rdev
->raid_disk
= disk
;
4731 if (rdev
->saved_raid_disk
!= disk
)
4733 rcu_assign_pointer(p
->rdev
, rdev
);
4736 print_raid5_conf(conf
);
4740 static int raid5_resize(mddev_t
*mddev
, sector_t sectors
)
4742 /* no resync is happening, and there is enough space
4743 * on all devices, so we can resize.
4744 * We need to make sure resync covers any new space.
4745 * If the array is shrinking we should possibly wait until
4746 * any io in the removed space completes, but it hardly seems
4749 sectors
&= ~((sector_t
)mddev
->chunk_size
/512 - 1);
4750 md_set_array_sectors(mddev
, raid5_size(mddev
, sectors
,
4751 mddev
->raid_disks
));
4752 if (mddev
->array_sectors
>
4753 raid5_size(mddev
, sectors
, mddev
->raid_disks
))
4755 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4757 if (sectors
> mddev
->dev_sectors
&& mddev
->recovery_cp
== MaxSector
) {
4758 mddev
->recovery_cp
= mddev
->dev_sectors
;
4759 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4761 mddev
->dev_sectors
= sectors
;
4762 mddev
->resync_max_sectors
= sectors
;
4766 #ifdef CONFIG_MD_RAID5_RESHAPE
4767 static int raid5_check_reshape(mddev_t
*mddev
)
4769 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4771 if (mddev
->delta_disks
== 0)
4772 return 0; /* nothing to do */
4774 /* Cannot grow a bitmap yet */
4776 if (mddev
->degraded
> conf
->max_degraded
)
4778 if (mddev
->delta_disks
< 0) {
4779 /* We might be able to shrink, but the devices must
4780 * be made bigger first.
4781 * For raid6, 4 is the minimum size.
4782 * Otherwise 2 is the minimum
4785 if (mddev
->level
== 6)
4787 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
4791 /* Can only proceed if there are plenty of stripe_heads.
4792 * We need a minimum of one full stripe,, and for sensible progress
4793 * it is best to have about 4 times that.
4794 * If we require 4 times, then the default 256 4K stripe_heads will
4795 * allow for chunk sizes up to 256K, which is probably OK.
4796 * If the chunk size is greater, user-space should request more
4797 * stripe_heads first.
4799 if ((mddev
->chunk_size
/ STRIPE_SIZE
) * 4 > conf
->max_nr_stripes
||
4800 (mddev
->new_chunk
/ STRIPE_SIZE
) * 4 > conf
->max_nr_stripes
) {
4801 printk(KERN_WARNING
"raid5: reshape: not enough stripes. Needed %lu\n",
4802 (max(mddev
->chunk_size
, mddev
->new_chunk
)
4807 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
4810 static int raid5_start_reshape(mddev_t
*mddev
)
4812 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4815 int added_devices
= 0;
4816 unsigned long flags
;
4818 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4821 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
4822 if (rdev
->raid_disk
< 0 &&
4823 !test_bit(Faulty
, &rdev
->flags
))
4826 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
4827 /* Not enough devices even to make a degraded array
4832 /* Refuse to reduce size of the array. Any reductions in
4833 * array size must be through explicit setting of array_size
4836 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
4837 < mddev
->array_sectors
) {
4838 printk(KERN_ERR
"md: %s: array size must be reduced "
4839 "before number of disks\n", mdname(mddev
));
4843 atomic_set(&conf
->reshape_stripes
, 0);
4844 spin_lock_irq(&conf
->device_lock
);
4845 conf
->previous_raid_disks
= conf
->raid_disks
;
4846 conf
->raid_disks
+= mddev
->delta_disks
;
4847 if (mddev
->delta_disks
< 0)
4848 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
4850 conf
->reshape_progress
= 0;
4851 conf
->reshape_safe
= conf
->reshape_progress
;
4853 spin_unlock_irq(&conf
->device_lock
);
4855 /* Add some new drives, as many as will fit.
4856 * We know there are enough to make the newly sized array work.
4858 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
4859 if (rdev
->raid_disk
< 0 &&
4860 !test_bit(Faulty
, &rdev
->flags
)) {
4861 if (raid5_add_disk(mddev
, rdev
) == 0) {
4863 set_bit(In_sync
, &rdev
->flags
);
4865 rdev
->recovery_offset
= 0;
4866 sprintf(nm
, "rd%d", rdev
->raid_disk
);
4867 if (sysfs_create_link(&mddev
->kobj
,
4870 "raid5: failed to create "
4871 " link %s for %s\n",
4877 if (mddev
->delta_disks
> 0) {
4878 spin_lock_irqsave(&conf
->device_lock
, flags
);
4879 mddev
->degraded
= (conf
->raid_disks
- conf
->previous_raid_disks
)
4881 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4883 mddev
->raid_disks
= conf
->raid_disks
;
4884 mddev
->reshape_position
= 0;
4885 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4887 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4888 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4889 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4890 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4891 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4893 if (!mddev
->sync_thread
) {
4894 mddev
->recovery
= 0;
4895 spin_lock_irq(&conf
->device_lock
);
4896 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
4897 conf
->reshape_progress
= MaxSector
;
4898 spin_unlock_irq(&conf
->device_lock
);
4901 md_wakeup_thread(mddev
->sync_thread
);
4902 md_new_event(mddev
);
4907 /* This is called from the reshape thread and should make any
4908 * changes needed in 'conf'
4910 static void end_reshape(raid5_conf_t
*conf
)
4913 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
4915 spin_lock_irq(&conf
->device_lock
);
4916 conf
->previous_raid_disks
= conf
->raid_disks
;
4917 conf
->reshape_progress
= MaxSector
;
4918 spin_unlock_irq(&conf
->device_lock
);
4920 /* read-ahead size must cover two whole stripes, which is
4921 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4924 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4925 int stripe
= data_disks
* (conf
->chunk_size
4927 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4928 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4933 /* This is called from the raid5d thread with mddev_lock held.
4934 * It makes config changes to the device.
4936 static void raid5_finish_reshape(mddev_t
*mddev
)
4938 struct block_device
*bdev
;
4940 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
4942 if (mddev
->delta_disks
> 0) {
4943 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
4944 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4947 bdev
= bdget_disk(mddev
->gendisk
, 0);
4949 mutex_lock(&bdev
->bd_inode
->i_mutex
);
4950 i_size_write(bdev
->bd_inode
,
4951 (loff_t
)mddev
->array_sectors
<< 9);
4952 mutex_unlock(&bdev
->bd_inode
->i_mutex
);
4957 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4958 mddev
->degraded
= conf
->raid_disks
;
4959 for (d
= 0; d
< conf
->raid_disks
; d
++)
4960 if (conf
->disks
[d
].rdev
&&
4962 &conf
->disks
[d
].rdev
->flags
))
4964 for (d
= conf
->raid_disks
;
4965 d
< conf
->raid_disks
- mddev
->delta_disks
;
4967 raid5_remove_disk(mddev
, d
);
4969 mddev
->reshape_position
= MaxSector
;
4970 mddev
->delta_disks
= 0;
4974 static void raid5_quiesce(mddev_t
*mddev
, int state
)
4976 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4979 case 2: /* resume for a suspend */
4980 wake_up(&conf
->wait_for_overlap
);
4983 case 1: /* stop all writes */
4984 spin_lock_irq(&conf
->device_lock
);
4986 wait_event_lock_irq(conf
->wait_for_stripe
,
4987 atomic_read(&conf
->active_stripes
) == 0 &&
4988 atomic_read(&conf
->active_aligned_reads
) == 0,
4989 conf
->device_lock
, /* nothing */);
4990 spin_unlock_irq(&conf
->device_lock
);
4993 case 0: /* re-enable writes */
4994 spin_lock_irq(&conf
->device_lock
);
4996 wake_up(&conf
->wait_for_stripe
);
4997 wake_up(&conf
->wait_for_overlap
);
4998 spin_unlock_irq(&conf
->device_lock
);
5004 static void *raid5_takeover_raid1(mddev_t
*mddev
)
5008 if (mddev
->raid_disks
!= 2 ||
5009 mddev
->degraded
> 1)
5010 return ERR_PTR(-EINVAL
);
5012 /* Should check if there are write-behind devices? */
5014 chunksect
= 64*2; /* 64K by default */
5016 /* The array must be an exact multiple of chunksize */
5017 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
5020 if ((chunksect
<<9) < STRIPE_SIZE
)
5021 /* array size does not allow a suitable chunk size */
5022 return ERR_PTR(-EINVAL
);
5024 mddev
->new_level
= 5;
5025 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5026 mddev
->new_chunk
= chunksect
<< 9;
5028 return setup_conf(mddev
);
5031 static void *raid5_takeover_raid6(mddev_t
*mddev
)
5035 switch (mddev
->layout
) {
5036 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5037 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
5039 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5040 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
5042 case ALGORITHM_LEFT_SYMMETRIC_6
:
5043 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5045 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5046 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
5048 case ALGORITHM_PARITY_0_6
:
5049 new_layout
= ALGORITHM_PARITY_0
;
5051 case ALGORITHM_PARITY_N
:
5052 new_layout
= ALGORITHM_PARITY_N
;
5055 return ERR_PTR(-EINVAL
);
5057 mddev
->new_level
= 5;
5058 mddev
->new_layout
= new_layout
;
5059 mddev
->delta_disks
= -1;
5060 mddev
->raid_disks
-= 1;
5061 return setup_conf(mddev
);
5065 static int raid5_reconfig(mddev_t
*mddev
, int new_layout
, int new_chunk
)
5067 /* Currently the layout and chunk size can only be changed
5068 * for a 2-drive raid array, as in that case no data shuffling
5070 * Later we might validate these and set new_* so a reshape
5071 * can complete the change.
5073 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
5075 if (new_layout
>= 0 && !algorithm_valid_raid5(new_layout
))
5077 if (new_chunk
> 0) {
5078 if (new_chunk
& (new_chunk
-1))
5079 /* not a power of 2 */
5081 if (new_chunk
< PAGE_SIZE
)
5083 if (mddev
->array_sectors
& ((new_chunk
>>9)-1))
5084 /* not factor of array size */
5088 /* They look valid */
5090 if (mddev
->raid_disks
!= 2)
5093 if (new_layout
>= 0) {
5094 conf
->algorithm
= new_layout
;
5095 mddev
->layout
= mddev
->new_layout
= new_layout
;
5097 if (new_chunk
> 0) {
5098 conf
->chunk_size
= new_chunk
;
5099 mddev
->chunk_size
= mddev
->new_chunk
= new_chunk
;
5101 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5102 md_wakeup_thread(mddev
->thread
);
5106 static void *raid5_takeover(mddev_t
*mddev
)
5108 /* raid5 can take over:
5109 * raid0 - if all devices are the same - make it a raid4 layout
5110 * raid1 - if there are two drives. We need to know the chunk size
5111 * raid4 - trivial - just use a raid4 layout.
5112 * raid6 - Providing it is a *_6 layout
5114 * For now, just do raid1
5117 if (mddev
->level
== 1)
5118 return raid5_takeover_raid1(mddev
);
5119 if (mddev
->level
== 4) {
5120 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5121 mddev
->new_level
= 5;
5122 return setup_conf(mddev
);
5124 if (mddev
->level
== 6)
5125 return raid5_takeover_raid6(mddev
);
5127 return ERR_PTR(-EINVAL
);
5131 static struct mdk_personality raid5_personality
;
5133 static void *raid6_takeover(mddev_t
*mddev
)
5135 /* Currently can only take over a raid5. We map the
5136 * personality to an equivalent raid6 personality
5137 * with the Q block at the end.
5141 if (mddev
->pers
!= &raid5_personality
)
5142 return ERR_PTR(-EINVAL
);
5143 if (mddev
->degraded
> 1)
5144 return ERR_PTR(-EINVAL
);
5145 if (mddev
->raid_disks
> 253)
5146 return ERR_PTR(-EINVAL
);
5147 if (mddev
->raid_disks
< 3)
5148 return ERR_PTR(-EINVAL
);
5150 switch (mddev
->layout
) {
5151 case ALGORITHM_LEFT_ASYMMETRIC
:
5152 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
5154 case ALGORITHM_RIGHT_ASYMMETRIC
:
5155 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
5157 case ALGORITHM_LEFT_SYMMETRIC
:
5158 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
5160 case ALGORITHM_RIGHT_SYMMETRIC
:
5161 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
5163 case ALGORITHM_PARITY_0
:
5164 new_layout
= ALGORITHM_PARITY_0_6
;
5166 case ALGORITHM_PARITY_N
:
5167 new_layout
= ALGORITHM_PARITY_N
;
5170 return ERR_PTR(-EINVAL
);
5172 mddev
->new_level
= 6;
5173 mddev
->new_layout
= new_layout
;
5174 mddev
->delta_disks
= 1;
5175 mddev
->raid_disks
+= 1;
5176 return setup_conf(mddev
);
5180 static struct mdk_personality raid6_personality
=
5184 .owner
= THIS_MODULE
,
5185 .make_request
= make_request
,
5189 .error_handler
= error
,
5190 .hot_add_disk
= raid5_add_disk
,
5191 .hot_remove_disk
= raid5_remove_disk
,
5192 .spare_active
= raid5_spare_active
,
5193 .sync_request
= sync_request
,
5194 .resize
= raid5_resize
,
5196 #ifdef CONFIG_MD_RAID5_RESHAPE
5197 .check_reshape
= raid5_check_reshape
,
5198 .start_reshape
= raid5_start_reshape
,
5199 .finish_reshape
= raid5_finish_reshape
,
5201 .quiesce
= raid5_quiesce
,
5202 .takeover
= raid6_takeover
,
5204 static struct mdk_personality raid5_personality
=
5208 .owner
= THIS_MODULE
,
5209 .make_request
= make_request
,
5213 .error_handler
= error
,
5214 .hot_add_disk
= raid5_add_disk
,
5215 .hot_remove_disk
= raid5_remove_disk
,
5216 .spare_active
= raid5_spare_active
,
5217 .sync_request
= sync_request
,
5218 .resize
= raid5_resize
,
5220 #ifdef CONFIG_MD_RAID5_RESHAPE
5221 .check_reshape
= raid5_check_reshape
,
5222 .start_reshape
= raid5_start_reshape
,
5223 .finish_reshape
= raid5_finish_reshape
,
5225 .quiesce
= raid5_quiesce
,
5226 .takeover
= raid5_takeover
,
5227 .reconfig
= raid5_reconfig
,
5230 static struct mdk_personality raid4_personality
=
5234 .owner
= THIS_MODULE
,
5235 .make_request
= make_request
,
5239 .error_handler
= error
,
5240 .hot_add_disk
= raid5_add_disk
,
5241 .hot_remove_disk
= raid5_remove_disk
,
5242 .spare_active
= raid5_spare_active
,
5243 .sync_request
= sync_request
,
5244 .resize
= raid5_resize
,
5246 #ifdef CONFIG_MD_RAID5_RESHAPE
5247 .check_reshape
= raid5_check_reshape
,
5248 .start_reshape
= raid5_start_reshape
,
5249 .finish_reshape
= raid5_finish_reshape
,
5251 .quiesce
= raid5_quiesce
,
5254 static int __init
raid5_init(void)
5256 register_md_personality(&raid6_personality
);
5257 register_md_personality(&raid5_personality
);
5258 register_md_personality(&raid4_personality
);
5262 static void raid5_exit(void)
5264 unregister_md_personality(&raid6_personality
);
5265 unregister_md_personality(&raid5_personality
);
5266 unregister_md_personality(&raid4_personality
);
5269 module_init(raid5_init
);
5270 module_exit(raid5_exit
);
5271 MODULE_LICENSE("GPL");
5272 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5273 MODULE_ALIAS("md-raid5");
5274 MODULE_ALIAS("md-raid4");
5275 MODULE_ALIAS("md-level-5");
5276 MODULE_ALIAS("md-level-4");
5277 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5278 MODULE_ALIAS("md-raid6");
5279 MODULE_ALIAS("md-level-6");
5281 /* This used to be two separate modules, they were: */
5282 MODULE_ALIAS("raid5");
5283 MODULE_ALIAS("raid6");