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 algorithm
= previous
? conf
->prev_algo
1278 int sectors_per_chunk
= previous
? (conf
->prev_chunk
>> 9)
1279 : (conf
->chunk_size
>> 9);
1280 int raid_disks
= previous
? conf
->previous_raid_disks
1282 int data_disks
= raid_disks
- conf
->max_degraded
;
1284 /* First compute the information on this sector */
1287 * Compute the chunk number and the sector offset inside the chunk
1289 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1290 chunk_number
= r_sector
;
1291 BUG_ON(r_sector
!= chunk_number
);
1294 * Compute the stripe number
1296 stripe
= chunk_number
/ data_disks
;
1299 * Compute the data disk and parity disk indexes inside the stripe
1301 *dd_idx
= chunk_number
% data_disks
;
1304 * Select the parity disk based on the user selected algorithm.
1306 pd_idx
= qd_idx
= ~0;
1307 switch(conf
->level
) {
1309 pd_idx
= data_disks
;
1312 switch (algorithm
) {
1313 case ALGORITHM_LEFT_ASYMMETRIC
:
1314 pd_idx
= data_disks
- stripe
% raid_disks
;
1315 if (*dd_idx
>= pd_idx
)
1318 case ALGORITHM_RIGHT_ASYMMETRIC
:
1319 pd_idx
= stripe
% raid_disks
;
1320 if (*dd_idx
>= pd_idx
)
1323 case ALGORITHM_LEFT_SYMMETRIC
:
1324 pd_idx
= data_disks
- stripe
% raid_disks
;
1325 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1327 case ALGORITHM_RIGHT_SYMMETRIC
:
1328 pd_idx
= stripe
% raid_disks
;
1329 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1331 case ALGORITHM_PARITY_0
:
1335 case ALGORITHM_PARITY_N
:
1336 pd_idx
= data_disks
;
1339 printk(KERN_ERR
"raid5: unsupported algorithm %d\n",
1346 switch (algorithm
) {
1347 case ALGORITHM_LEFT_ASYMMETRIC
:
1348 pd_idx
= raid_disks
- 1 - (stripe
% raid_disks
);
1349 qd_idx
= pd_idx
+ 1;
1350 if (pd_idx
== raid_disks
-1) {
1351 (*dd_idx
)++; /* Q D D D P */
1353 } else if (*dd_idx
>= pd_idx
)
1354 (*dd_idx
) += 2; /* D D P Q D */
1356 case ALGORITHM_RIGHT_ASYMMETRIC
:
1357 pd_idx
= stripe
% raid_disks
;
1358 qd_idx
= pd_idx
+ 1;
1359 if (pd_idx
== raid_disks
-1) {
1360 (*dd_idx
)++; /* Q D D D P */
1362 } else if (*dd_idx
>= pd_idx
)
1363 (*dd_idx
) += 2; /* D D P Q D */
1365 case ALGORITHM_LEFT_SYMMETRIC
:
1366 pd_idx
= raid_disks
- 1 - (stripe
% raid_disks
);
1367 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1368 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1370 case ALGORITHM_RIGHT_SYMMETRIC
:
1371 pd_idx
= stripe
% raid_disks
;
1372 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1373 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1376 case ALGORITHM_PARITY_0
:
1381 case ALGORITHM_PARITY_N
:
1382 pd_idx
= data_disks
;
1383 qd_idx
= data_disks
+ 1;
1386 case ALGORITHM_ROTATING_ZERO_RESTART
:
1387 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1388 * of blocks for computing Q is different.
1390 pd_idx
= stripe
% raid_disks
;
1391 qd_idx
= pd_idx
+ 1;
1392 if (pd_idx
== raid_disks
-1) {
1393 (*dd_idx
)++; /* Q D D D P */
1395 } else if (*dd_idx
>= pd_idx
)
1396 (*dd_idx
) += 2; /* D D P Q D */
1400 case ALGORITHM_ROTATING_N_RESTART
:
1401 /* Same a left_asymmetric, by first stripe is
1402 * D D D P Q rather than
1405 pd_idx
= raid_disks
- 1 - ((stripe
+ 1) % raid_disks
);
1406 qd_idx
= pd_idx
+ 1;
1407 if (pd_idx
== raid_disks
-1) {
1408 (*dd_idx
)++; /* Q D D D P */
1410 } else if (*dd_idx
>= pd_idx
)
1411 (*dd_idx
) += 2; /* D D P Q D */
1415 case ALGORITHM_ROTATING_N_CONTINUE
:
1416 /* Same as left_symmetric but Q is before P */
1417 pd_idx
= raid_disks
- 1 - (stripe
% raid_disks
);
1418 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
1419 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1423 case ALGORITHM_LEFT_ASYMMETRIC_6
:
1424 /* RAID5 left_asymmetric, with Q on last device */
1425 pd_idx
= data_disks
- stripe
% (raid_disks
-1);
1426 if (*dd_idx
>= pd_idx
)
1428 qd_idx
= raid_disks
- 1;
1431 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
1432 pd_idx
= stripe
% (raid_disks
-1);
1433 if (*dd_idx
>= pd_idx
)
1435 qd_idx
= raid_disks
- 1;
1438 case ALGORITHM_LEFT_SYMMETRIC_6
:
1439 pd_idx
= data_disks
- stripe
% (raid_disks
-1);
1440 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1441 qd_idx
= raid_disks
- 1;
1444 case ALGORITHM_RIGHT_SYMMETRIC_6
:
1445 pd_idx
= stripe
% (raid_disks
-1);
1446 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1447 qd_idx
= raid_disks
- 1;
1450 case ALGORITHM_PARITY_0_6
:
1453 qd_idx
= raid_disks
- 1;
1458 printk(KERN_CRIT
"raid6: unsupported algorithm %d\n",
1466 sh
->pd_idx
= pd_idx
;
1467 sh
->qd_idx
= qd_idx
;
1468 sh
->ddf_layout
= ddf_layout
;
1471 * Finally, compute the new sector number
1473 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
1478 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
1480 raid5_conf_t
*conf
= sh
->raid_conf
;
1481 int raid_disks
= sh
->disks
;
1482 int data_disks
= raid_disks
- conf
->max_degraded
;
1483 sector_t new_sector
= sh
->sector
, check
;
1484 int sectors_per_chunk
= previous
? (conf
->prev_chunk
>> 9)
1485 : (conf
->chunk_size
>> 9);
1486 int algorithm
= previous
? conf
->prev_algo
1490 int chunk_number
, dummy1
, dd_idx
= i
;
1492 struct stripe_head sh2
;
1495 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
1496 stripe
= new_sector
;
1497 BUG_ON(new_sector
!= stripe
);
1499 if (i
== sh
->pd_idx
)
1501 switch(conf
->level
) {
1504 switch (algorithm
) {
1505 case ALGORITHM_LEFT_ASYMMETRIC
:
1506 case ALGORITHM_RIGHT_ASYMMETRIC
:
1510 case ALGORITHM_LEFT_SYMMETRIC
:
1511 case ALGORITHM_RIGHT_SYMMETRIC
:
1514 i
-= (sh
->pd_idx
+ 1);
1516 case ALGORITHM_PARITY_0
:
1519 case ALGORITHM_PARITY_N
:
1522 printk(KERN_ERR
"raid5: unsupported algorithm %d\n",
1528 if (i
== sh
->qd_idx
)
1529 return 0; /* It is the Q disk */
1530 switch (algorithm
) {
1531 case ALGORITHM_LEFT_ASYMMETRIC
:
1532 case ALGORITHM_RIGHT_ASYMMETRIC
:
1533 case ALGORITHM_ROTATING_ZERO_RESTART
:
1534 case ALGORITHM_ROTATING_N_RESTART
:
1535 if (sh
->pd_idx
== raid_disks
-1)
1536 i
--; /* Q D D D P */
1537 else if (i
> sh
->pd_idx
)
1538 i
-= 2; /* D D P Q D */
1540 case ALGORITHM_LEFT_SYMMETRIC
:
1541 case ALGORITHM_RIGHT_SYMMETRIC
:
1542 if (sh
->pd_idx
== raid_disks
-1)
1543 i
--; /* Q D D D P */
1548 i
-= (sh
->pd_idx
+ 2);
1551 case ALGORITHM_PARITY_0
:
1554 case ALGORITHM_PARITY_N
:
1556 case ALGORITHM_ROTATING_N_CONTINUE
:
1557 if (sh
->pd_idx
== 0)
1558 i
--; /* P D D D Q */
1559 else if (i
> sh
->pd_idx
)
1560 i
-= 2; /* D D Q P D */
1562 case ALGORITHM_LEFT_ASYMMETRIC_6
:
1563 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
1567 case ALGORITHM_LEFT_SYMMETRIC_6
:
1568 case ALGORITHM_RIGHT_SYMMETRIC_6
:
1570 i
+= data_disks
+ 1;
1571 i
-= (sh
->pd_idx
+ 1);
1573 case ALGORITHM_PARITY_0_6
:
1577 printk(KERN_CRIT
"raid6: unsupported algorithm %d\n",
1584 chunk_number
= stripe
* data_disks
+ i
;
1585 r_sector
= (sector_t
)chunk_number
* sectors_per_chunk
+ chunk_offset
;
1587 check
= raid5_compute_sector(conf
, r_sector
,
1588 previous
, &dummy1
, &sh2
);
1589 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
1590 || sh2
.qd_idx
!= sh
->qd_idx
) {
1591 printk(KERN_ERR
"compute_blocknr: map not correct\n");
1600 * Copy data between a page in the stripe cache, and one or more bion
1601 * The page could align with the middle of the bio, or there could be
1602 * several bion, each with several bio_vecs, which cover part of the page
1603 * Multiple bion are linked together on bi_next. There may be extras
1604 * at the end of this list. We ignore them.
1606 static void copy_data(int frombio
, struct bio
*bio
,
1610 char *pa
= page_address(page
);
1611 struct bio_vec
*bvl
;
1615 if (bio
->bi_sector
>= sector
)
1616 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
1618 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
1619 bio_for_each_segment(bvl
, bio
, i
) {
1620 int len
= bio_iovec_idx(bio
,i
)->bv_len
;
1624 if (page_offset
< 0) {
1625 b_offset
= -page_offset
;
1626 page_offset
+= b_offset
;
1630 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1631 clen
= STRIPE_SIZE
- page_offset
;
1635 char *ba
= __bio_kmap_atomic(bio
, i
, KM_USER0
);
1637 memcpy(pa
+page_offset
, ba
+b_offset
, clen
);
1639 memcpy(ba
+b_offset
, pa
+page_offset
, clen
);
1640 __bio_kunmap_atomic(ba
, KM_USER0
);
1642 if (clen
< len
) /* hit end of page */
1648 #define check_xor() do { \
1649 if (count == MAX_XOR_BLOCKS) { \
1650 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1655 static void compute_parity6(struct stripe_head
*sh
, int method
)
1657 raid5_conf_t
*conf
= sh
->raid_conf
;
1658 int i
, pd_idx
, qd_idx
, d0_idx
, disks
= sh
->disks
, count
;
1659 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1661 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1662 void *ptrs
[syndrome_disks
+2];
1664 pd_idx
= sh
->pd_idx
;
1665 qd_idx
= sh
->qd_idx
;
1666 d0_idx
= raid6_d0(sh
);
1668 pr_debug("compute_parity, stripe %llu, method %d\n",
1669 (unsigned long long)sh
->sector
, method
);
1672 case READ_MODIFY_WRITE
:
1673 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1674 case RECONSTRUCT_WRITE
:
1675 for (i
= disks
; i
-- ;)
1676 if ( i
!= pd_idx
&& i
!= qd_idx
&& sh
->dev
[i
].towrite
) {
1677 chosen
= sh
->dev
[i
].towrite
;
1678 sh
->dev
[i
].towrite
= NULL
;
1680 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
1681 wake_up(&conf
->wait_for_overlap
);
1683 BUG_ON(sh
->dev
[i
].written
);
1684 sh
->dev
[i
].written
= chosen
;
1688 BUG(); /* Not implemented yet */
1691 for (i
= disks
; i
--;)
1692 if (sh
->dev
[i
].written
) {
1693 sector_t sector
= sh
->dev
[i
].sector
;
1694 struct bio
*wbi
= sh
->dev
[i
].written
;
1695 while (wbi
&& wbi
->bi_sector
< sector
+ STRIPE_SECTORS
) {
1696 copy_data(1, wbi
, sh
->dev
[i
].page
, sector
);
1697 wbi
= r5_next_bio(wbi
, sector
);
1700 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1701 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1704 /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
1706 for (i
= 0; i
< disks
; i
++)
1707 ptrs
[i
] = (void *)raid6_empty_zero_page
;
1712 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1714 ptrs
[slot
] = page_address(sh
->dev
[i
].page
);
1715 if (slot
< syndrome_disks
&&
1716 !test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
)) {
1717 printk(KERN_ERR
"block %d/%d not uptodate "
1718 "on parity calc\n", i
, count
);
1722 i
= raid6_next_disk(i
, disks
);
1723 } while (i
!= d0_idx
);
1724 BUG_ON(count
!= syndrome_disks
);
1726 raid6_call
.gen_syndrome(syndrome_disks
+2, STRIPE_SIZE
, ptrs
);
1729 case RECONSTRUCT_WRITE
:
1730 set_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
1731 set_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
);
1732 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
1733 set_bit(R5_LOCKED
, &sh
->dev
[qd_idx
].flags
);
1736 set_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
1737 set_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
);
1743 /* Compute one missing block */
1744 static void compute_block_1(struct stripe_head
*sh
, int dd_idx
, int nozero
)
1746 int i
, count
, disks
= sh
->disks
;
1747 void *ptr
[MAX_XOR_BLOCKS
], *dest
, *p
;
1748 int qd_idx
= sh
->qd_idx
;
1750 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1751 (unsigned long long)sh
->sector
, dd_idx
);
1753 if ( dd_idx
== qd_idx
) {
1754 /* We're actually computing the Q drive */
1755 compute_parity6(sh
, UPDATE_PARITY
);
1757 dest
= page_address(sh
->dev
[dd_idx
].page
);
1758 if (!nozero
) memset(dest
, 0, STRIPE_SIZE
);
1760 for (i
= disks
; i
--; ) {
1761 if (i
== dd_idx
|| i
== qd_idx
)
1763 p
= page_address(sh
->dev
[i
].page
);
1764 if (test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
))
1767 printk("compute_block() %d, stripe %llu, %d"
1768 " not present\n", dd_idx
,
1769 (unsigned long long)sh
->sector
, i
);
1774 xor_blocks(count
, STRIPE_SIZE
, dest
, ptr
);
1775 if (!nozero
) set_bit(R5_UPTODATE
, &sh
->dev
[dd_idx
].flags
);
1776 else clear_bit(R5_UPTODATE
, &sh
->dev
[dd_idx
].flags
);
1780 /* Compute two missing blocks */
1781 static void compute_block_2(struct stripe_head
*sh
, int dd_idx1
, int dd_idx2
)
1783 int i
, count
, disks
= sh
->disks
;
1784 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1785 int d0_idx
= raid6_d0(sh
);
1786 int faila
= -1, failb
= -1;
1787 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1788 void *ptrs
[syndrome_disks
+2];
1790 for (i
= 0; i
< disks
; i
++)
1791 ptrs
[i
] = (void *)raid6_empty_zero_page
;
1795 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1797 ptrs
[slot
] = page_address(sh
->dev
[i
].page
);
1803 i
= raid6_next_disk(i
, disks
);
1804 } while (i
!= d0_idx
);
1805 BUG_ON(count
!= syndrome_disks
);
1807 BUG_ON(faila
== failb
);
1808 if ( failb
< faila
) { int tmp
= faila
; faila
= failb
; failb
= tmp
; }
1810 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1811 (unsigned long long)sh
->sector
, dd_idx1
, dd_idx2
,
1814 if (failb
== syndrome_disks
+1) {
1815 /* Q disk is one of the missing disks */
1816 if (faila
== syndrome_disks
) {
1817 /* Missing P+Q, just recompute */
1818 compute_parity6(sh
, UPDATE_PARITY
);
1821 /* We're missing D+Q; recompute D from P */
1822 compute_block_1(sh
, ((dd_idx1
== sh
->qd_idx
) ?
1825 compute_parity6(sh
, UPDATE_PARITY
); /* Is this necessary? */
1830 /* We're missing D+P or D+D; */
1831 if (failb
== syndrome_disks
) {
1832 /* We're missing D+P. */
1833 raid6_datap_recov(syndrome_disks
+2, STRIPE_SIZE
, faila
, ptrs
);
1835 /* We're missing D+D. */
1836 raid6_2data_recov(syndrome_disks
+2, STRIPE_SIZE
, faila
, failb
,
1840 /* Both the above update both missing blocks */
1841 set_bit(R5_UPTODATE
, &sh
->dev
[dd_idx1
].flags
);
1842 set_bit(R5_UPTODATE
, &sh
->dev
[dd_idx2
].flags
);
1846 schedule_reconstruction5(struct stripe_head
*sh
, struct stripe_head_state
*s
,
1847 int rcw
, int expand
)
1849 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
1852 /* if we are not expanding this is a proper write request, and
1853 * there will be bios with new data to be drained into the
1857 sh
->reconstruct_state
= reconstruct_state_drain_run
;
1858 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
1860 sh
->reconstruct_state
= reconstruct_state_run
;
1862 set_bit(STRIPE_OP_POSTXOR
, &s
->ops_request
);
1864 for (i
= disks
; i
--; ) {
1865 struct r5dev
*dev
= &sh
->dev
[i
];
1868 set_bit(R5_LOCKED
, &dev
->flags
);
1869 set_bit(R5_Wantdrain
, &dev
->flags
);
1871 clear_bit(R5_UPTODATE
, &dev
->flags
);
1875 if (s
->locked
+ 1 == disks
)
1876 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
1877 atomic_inc(&sh
->raid_conf
->pending_full_writes
);
1879 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
1880 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
1882 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
1883 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
1884 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
1885 set_bit(STRIPE_OP_POSTXOR
, &s
->ops_request
);
1887 for (i
= disks
; i
--; ) {
1888 struct r5dev
*dev
= &sh
->dev
[i
];
1893 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
1894 test_bit(R5_Wantcompute
, &dev
->flags
))) {
1895 set_bit(R5_Wantdrain
, &dev
->flags
);
1896 set_bit(R5_LOCKED
, &dev
->flags
);
1897 clear_bit(R5_UPTODATE
, &dev
->flags
);
1903 /* keep the parity disk locked while asynchronous operations
1906 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
1907 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
1910 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1911 __func__
, (unsigned long long)sh
->sector
,
1912 s
->locked
, s
->ops_request
);
1916 * Each stripe/dev can have one or more bion attached.
1917 * toread/towrite point to the first in a chain.
1918 * The bi_next chain must be in order.
1920 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
1923 raid5_conf_t
*conf
= sh
->raid_conf
;
1926 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1927 (unsigned long long)bi
->bi_sector
,
1928 (unsigned long long)sh
->sector
);
1931 spin_lock(&sh
->lock
);
1932 spin_lock_irq(&conf
->device_lock
);
1934 bip
= &sh
->dev
[dd_idx
].towrite
;
1935 if (*bip
== NULL
&& sh
->dev
[dd_idx
].written
== NULL
)
1938 bip
= &sh
->dev
[dd_idx
].toread
;
1939 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
1940 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
1942 bip
= & (*bip
)->bi_next
;
1944 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
1947 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
1951 bi
->bi_phys_segments
++;
1952 spin_unlock_irq(&conf
->device_lock
);
1953 spin_unlock(&sh
->lock
);
1955 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1956 (unsigned long long)bi
->bi_sector
,
1957 (unsigned long long)sh
->sector
, dd_idx
);
1959 if (conf
->mddev
->bitmap
&& firstwrite
) {
1960 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
1962 sh
->bm_seq
= conf
->seq_flush
+1;
1963 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
1967 /* check if page is covered */
1968 sector_t sector
= sh
->dev
[dd_idx
].sector
;
1969 for (bi
=sh
->dev
[dd_idx
].towrite
;
1970 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
1971 bi
&& bi
->bi_sector
<= sector
;
1972 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
1973 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
1974 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
1976 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
1977 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
1982 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
1983 spin_unlock_irq(&conf
->device_lock
);
1984 spin_unlock(&sh
->lock
);
1988 static void end_reshape(raid5_conf_t
*conf
);
1990 static int page_is_zero(struct page
*p
)
1992 char *a
= page_address(p
);
1993 return ((*(u32
*)a
) == 0 &&
1994 memcmp(a
, a
+4, STRIPE_SIZE
-4)==0);
1997 static void stripe_set_idx(sector_t stripe
, raid5_conf_t
*conf
, int previous
,
1998 struct stripe_head
*sh
)
2000 int sectors_per_chunk
=
2001 previous
? (conf
->prev_chunk
>> 9)
2002 : (conf
->chunk_size
>> 9);
2004 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2005 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2007 raid5_compute_sector(conf
,
2008 stripe
* (disks
- conf
->max_degraded
)
2009 *sectors_per_chunk
+ chunk_offset
,
2015 handle_failed_stripe(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2016 struct stripe_head_state
*s
, int disks
,
2017 struct bio
**return_bi
)
2020 for (i
= disks
; i
--; ) {
2024 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2027 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2028 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2029 /* multiple read failures in one stripe */
2030 md_error(conf
->mddev
, rdev
);
2033 spin_lock_irq(&conf
->device_lock
);
2034 /* fail all writes first */
2035 bi
= sh
->dev
[i
].towrite
;
2036 sh
->dev
[i
].towrite
= NULL
;
2042 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2043 wake_up(&conf
->wait_for_overlap
);
2045 while (bi
&& bi
->bi_sector
<
2046 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2047 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2048 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2049 if (!raid5_dec_bi_phys_segments(bi
)) {
2050 md_write_end(conf
->mddev
);
2051 bi
->bi_next
= *return_bi
;
2056 /* and fail all 'written' */
2057 bi
= sh
->dev
[i
].written
;
2058 sh
->dev
[i
].written
= NULL
;
2059 if (bi
) bitmap_end
= 1;
2060 while (bi
&& bi
->bi_sector
<
2061 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2062 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2063 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2064 if (!raid5_dec_bi_phys_segments(bi
)) {
2065 md_write_end(conf
->mddev
);
2066 bi
->bi_next
= *return_bi
;
2072 /* fail any reads if this device is non-operational and
2073 * the data has not reached the cache yet.
2075 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2076 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2077 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2078 bi
= sh
->dev
[i
].toread
;
2079 sh
->dev
[i
].toread
= NULL
;
2080 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2081 wake_up(&conf
->wait_for_overlap
);
2082 if (bi
) s
->to_read
--;
2083 while (bi
&& bi
->bi_sector
<
2084 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2085 struct bio
*nextbi
=
2086 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2087 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2088 if (!raid5_dec_bi_phys_segments(bi
)) {
2089 bi
->bi_next
= *return_bi
;
2095 spin_unlock_irq(&conf
->device_lock
);
2097 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2098 STRIPE_SECTORS
, 0, 0);
2101 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2102 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2103 md_wakeup_thread(conf
->mddev
->thread
);
2106 /* fetch_block5 - checks the given member device to see if its data needs
2107 * to be read or computed to satisfy a request.
2109 * Returns 1 when no more member devices need to be checked, otherwise returns
2110 * 0 to tell the loop in handle_stripe_fill5 to continue
2112 static int fetch_block5(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2113 int disk_idx
, int disks
)
2115 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2116 struct r5dev
*failed_dev
= &sh
->dev
[s
->failed_num
];
2118 /* is the data in this block needed, and can we get it? */
2119 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2120 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2122 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2123 s
->syncing
|| s
->expanding
||
2125 (failed_dev
->toread
||
2126 (failed_dev
->towrite
&&
2127 !test_bit(R5_OVERWRITE
, &failed_dev
->flags
)))))) {
2128 /* We would like to get this block, possibly by computing it,
2129 * otherwise read it if the backing disk is insync
2131 if ((s
->uptodate
== disks
- 1) &&
2132 (s
->failed
&& disk_idx
== s
->failed_num
)) {
2133 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2134 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2135 set_bit(R5_Wantcompute
, &dev
->flags
);
2136 sh
->ops
.target
= disk_idx
;
2138 /* Careful: from this point on 'uptodate' is in the eye
2139 * of raid5_run_ops which services 'compute' operations
2140 * before writes. R5_Wantcompute flags a block that will
2141 * be R5_UPTODATE by the time it is needed for a
2142 * subsequent operation.
2145 return 1; /* uptodate + compute == disks */
2146 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2147 set_bit(R5_LOCKED
, &dev
->flags
);
2148 set_bit(R5_Wantread
, &dev
->flags
);
2150 pr_debug("Reading block %d (sync=%d)\n", disk_idx
,
2159 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2161 static void handle_stripe_fill5(struct stripe_head
*sh
,
2162 struct stripe_head_state
*s
, int disks
)
2166 /* look for blocks to read/compute, skip this if a compute
2167 * is already in flight, or if the stripe contents are in the
2168 * midst of changing due to a write
2170 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2171 !sh
->reconstruct_state
)
2172 for (i
= disks
; i
--; )
2173 if (fetch_block5(sh
, s
, i
, disks
))
2175 set_bit(STRIPE_HANDLE
, &sh
->state
);
2178 static void handle_stripe_fill6(struct stripe_head
*sh
,
2179 struct stripe_head_state
*s
, struct r6_state
*r6s
,
2183 for (i
= disks
; i
--; ) {
2184 struct r5dev
*dev
= &sh
->dev
[i
];
2185 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2186 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2187 (dev
->toread
|| (dev
->towrite
&&
2188 !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2189 s
->syncing
|| s
->expanding
||
2191 (sh
->dev
[r6s
->failed_num
[0]].toread
||
2194 (sh
->dev
[r6s
->failed_num
[1]].toread
||
2196 /* we would like to get this block, possibly
2197 * by computing it, but we might not be able to
2199 if ((s
->uptodate
== disks
- 1) &&
2200 (s
->failed
&& (i
== r6s
->failed_num
[0] ||
2201 i
== r6s
->failed_num
[1]))) {
2202 pr_debug("Computing stripe %llu block %d\n",
2203 (unsigned long long)sh
->sector
, i
);
2204 compute_block_1(sh
, i
, 0);
2206 } else if ( s
->uptodate
== disks
-2 && s
->failed
>= 2 ) {
2207 /* Computing 2-failure is *very* expensive; only
2208 * do it if failed >= 2
2211 for (other
= disks
; other
--; ) {
2214 if (!test_bit(R5_UPTODATE
,
2215 &sh
->dev
[other
].flags
))
2219 pr_debug("Computing stripe %llu blocks %d,%d\n",
2220 (unsigned long long)sh
->sector
,
2222 compute_block_2(sh
, i
, other
);
2224 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2225 set_bit(R5_LOCKED
, &dev
->flags
);
2226 set_bit(R5_Wantread
, &dev
->flags
);
2228 pr_debug("Reading block %d (sync=%d)\n",
2233 set_bit(STRIPE_HANDLE
, &sh
->state
);
2237 /* handle_stripe_clean_event
2238 * any written block on an uptodate or failed drive can be returned.
2239 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2240 * never LOCKED, so we don't need to test 'failed' directly.
2242 static void handle_stripe_clean_event(raid5_conf_t
*conf
,
2243 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2248 for (i
= disks
; i
--; )
2249 if (sh
->dev
[i
].written
) {
2251 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2252 test_bit(R5_UPTODATE
, &dev
->flags
)) {
2253 /* We can return any write requests */
2254 struct bio
*wbi
, *wbi2
;
2256 pr_debug("Return write for disc %d\n", i
);
2257 spin_lock_irq(&conf
->device_lock
);
2259 dev
->written
= NULL
;
2260 while (wbi
&& wbi
->bi_sector
<
2261 dev
->sector
+ STRIPE_SECTORS
) {
2262 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2263 if (!raid5_dec_bi_phys_segments(wbi
)) {
2264 md_write_end(conf
->mddev
);
2265 wbi
->bi_next
= *return_bi
;
2270 if (dev
->towrite
== NULL
)
2272 spin_unlock_irq(&conf
->device_lock
);
2274 bitmap_endwrite(conf
->mddev
->bitmap
,
2277 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2282 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2283 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2284 md_wakeup_thread(conf
->mddev
->thread
);
2287 static void handle_stripe_dirtying5(raid5_conf_t
*conf
,
2288 struct stripe_head
*sh
, struct stripe_head_state
*s
, int disks
)
2290 int rmw
= 0, rcw
= 0, i
;
2291 for (i
= disks
; i
--; ) {
2292 /* would I have to read this buffer for read_modify_write */
2293 struct r5dev
*dev
= &sh
->dev
[i
];
2294 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2295 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2296 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2297 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2298 if (test_bit(R5_Insync
, &dev
->flags
))
2301 rmw
+= 2*disks
; /* cannot read it */
2303 /* Would I have to read this buffer for reconstruct_write */
2304 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2305 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2306 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2307 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2308 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2313 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2314 (unsigned long long)sh
->sector
, rmw
, rcw
);
2315 set_bit(STRIPE_HANDLE
, &sh
->state
);
2316 if (rmw
< rcw
&& rmw
> 0)
2317 /* prefer read-modify-write, but need to get some data */
2318 for (i
= disks
; i
--; ) {
2319 struct r5dev
*dev
= &sh
->dev
[i
];
2320 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2321 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2322 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2323 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2324 test_bit(R5_Insync
, &dev
->flags
)) {
2326 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2327 pr_debug("Read_old block "
2328 "%d for r-m-w\n", i
);
2329 set_bit(R5_LOCKED
, &dev
->flags
);
2330 set_bit(R5_Wantread
, &dev
->flags
);
2333 set_bit(STRIPE_DELAYED
, &sh
->state
);
2334 set_bit(STRIPE_HANDLE
, &sh
->state
);
2338 if (rcw
<= rmw
&& rcw
> 0)
2339 /* want reconstruct write, but need to get some data */
2340 for (i
= disks
; i
--; ) {
2341 struct r5dev
*dev
= &sh
->dev
[i
];
2342 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2344 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2345 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2346 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2347 test_bit(R5_Insync
, &dev
->flags
)) {
2349 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2350 pr_debug("Read_old block "
2351 "%d for Reconstruct\n", i
);
2352 set_bit(R5_LOCKED
, &dev
->flags
);
2353 set_bit(R5_Wantread
, &dev
->flags
);
2356 set_bit(STRIPE_DELAYED
, &sh
->state
);
2357 set_bit(STRIPE_HANDLE
, &sh
->state
);
2361 /* now if nothing is locked, and if we have enough data,
2362 * we can start a write request
2364 /* since handle_stripe can be called at any time we need to handle the
2365 * case where a compute block operation has been submitted and then a
2366 * subsequent call wants to start a write request. raid5_run_ops only
2367 * handles the case where compute block and postxor are requested
2368 * simultaneously. If this is not the case then new writes need to be
2369 * held off until the compute completes.
2371 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2372 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2373 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2374 schedule_reconstruction5(sh
, s
, rcw
== 0, 0);
2377 static void handle_stripe_dirtying6(raid5_conf_t
*conf
,
2378 struct stripe_head
*sh
, struct stripe_head_state
*s
,
2379 struct r6_state
*r6s
, int disks
)
2381 int rcw
= 0, must_compute
= 0, pd_idx
= sh
->pd_idx
, i
;
2382 int qd_idx
= sh
->qd_idx
;
2383 for (i
= disks
; i
--; ) {
2384 struct r5dev
*dev
= &sh
->dev
[i
];
2385 /* Would I have to read this buffer for reconstruct_write */
2386 if (!test_bit(R5_OVERWRITE
, &dev
->flags
)
2387 && i
!= pd_idx
&& i
!= qd_idx
2388 && (!test_bit(R5_LOCKED
, &dev
->flags
)
2390 !test_bit(R5_UPTODATE
, &dev
->flags
)) {
2391 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2393 pr_debug("raid6: must_compute: "
2394 "disk %d flags=%#lx\n", i
, dev
->flags
);
2399 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2400 (unsigned long long)sh
->sector
, rcw
, must_compute
);
2401 set_bit(STRIPE_HANDLE
, &sh
->state
);
2404 /* want reconstruct write, but need to get some data */
2405 for (i
= disks
; i
--; ) {
2406 struct r5dev
*dev
= &sh
->dev
[i
];
2407 if (!test_bit(R5_OVERWRITE
, &dev
->flags
)
2408 && !(s
->failed
== 0 && (i
== pd_idx
|| i
== qd_idx
))
2409 && !test_bit(R5_LOCKED
, &dev
->flags
) &&
2410 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2411 test_bit(R5_Insync
, &dev
->flags
)) {
2413 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2414 pr_debug("Read_old stripe %llu "
2415 "block %d for Reconstruct\n",
2416 (unsigned long long)sh
->sector
, i
);
2417 set_bit(R5_LOCKED
, &dev
->flags
);
2418 set_bit(R5_Wantread
, &dev
->flags
);
2421 pr_debug("Request delayed stripe %llu "
2422 "block %d for Reconstruct\n",
2423 (unsigned long long)sh
->sector
, i
);
2424 set_bit(STRIPE_DELAYED
, &sh
->state
);
2425 set_bit(STRIPE_HANDLE
, &sh
->state
);
2429 /* now if nothing is locked, and if we have enough data, we can start a
2432 if (s
->locked
== 0 && rcw
== 0 &&
2433 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
2434 if (must_compute
> 0) {
2435 /* We have failed blocks and need to compute them */
2436 switch (s
->failed
) {
2440 compute_block_1(sh
, r6s
->failed_num
[0], 0);
2443 compute_block_2(sh
, r6s
->failed_num
[0],
2444 r6s
->failed_num
[1]);
2446 default: /* This request should have been failed? */
2451 pr_debug("Computing parity for stripe %llu\n",
2452 (unsigned long long)sh
->sector
);
2453 compute_parity6(sh
, RECONSTRUCT_WRITE
);
2454 /* now every locked buffer is ready to be written */
2455 for (i
= disks
; i
--; )
2456 if (test_bit(R5_LOCKED
, &sh
->dev
[i
].flags
)) {
2457 pr_debug("Writing stripe %llu block %d\n",
2458 (unsigned long long)sh
->sector
, i
);
2460 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
2462 if (s
->locked
== disks
)
2463 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2464 atomic_inc(&conf
->pending_full_writes
);
2465 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2466 set_bit(STRIPE_INSYNC
, &sh
->state
);
2468 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2469 atomic_dec(&conf
->preread_active_stripes
);
2470 if (atomic_read(&conf
->preread_active_stripes
) <
2472 md_wakeup_thread(conf
->mddev
->thread
);
2477 static void handle_parity_checks5(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2478 struct stripe_head_state
*s
, int disks
)
2480 struct r5dev
*dev
= NULL
;
2482 set_bit(STRIPE_HANDLE
, &sh
->state
);
2484 switch (sh
->check_state
) {
2485 case check_state_idle
:
2486 /* start a new check operation if there are no failures */
2487 if (s
->failed
== 0) {
2488 BUG_ON(s
->uptodate
!= disks
);
2489 sh
->check_state
= check_state_run
;
2490 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2491 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2495 dev
= &sh
->dev
[s
->failed_num
];
2497 case check_state_compute_result
:
2498 sh
->check_state
= check_state_idle
;
2500 dev
= &sh
->dev
[sh
->pd_idx
];
2502 /* check that a write has not made the stripe insync */
2503 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2506 /* either failed parity check, or recovery is happening */
2507 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2508 BUG_ON(s
->uptodate
!= disks
);
2510 set_bit(R5_LOCKED
, &dev
->flags
);
2512 set_bit(R5_Wantwrite
, &dev
->flags
);
2514 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2515 set_bit(STRIPE_INSYNC
, &sh
->state
);
2517 case check_state_run
:
2518 break; /* we will be called again upon completion */
2519 case check_state_check_result
:
2520 sh
->check_state
= check_state_idle
;
2522 /* if a failure occurred during the check operation, leave
2523 * STRIPE_INSYNC not set and let the stripe be handled again
2528 /* handle a successful check operation, if parity is correct
2529 * we are done. Otherwise update the mismatch count and repair
2530 * parity if !MD_RECOVERY_CHECK
2532 if (sh
->ops
.zero_sum_result
== 0)
2533 /* parity is correct (on disc,
2534 * not in buffer any more)
2536 set_bit(STRIPE_INSYNC
, &sh
->state
);
2538 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2539 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2540 /* don't try to repair!! */
2541 set_bit(STRIPE_INSYNC
, &sh
->state
);
2543 sh
->check_state
= check_state_compute_run
;
2544 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2545 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2546 set_bit(R5_Wantcompute
,
2547 &sh
->dev
[sh
->pd_idx
].flags
);
2548 sh
->ops
.target
= sh
->pd_idx
;
2553 case check_state_compute_run
:
2556 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2557 __func__
, sh
->check_state
,
2558 (unsigned long long) sh
->sector
);
2564 static void handle_parity_checks6(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2565 struct stripe_head_state
*s
,
2566 struct r6_state
*r6s
, struct page
*tmp_page
,
2569 int update_p
= 0, update_q
= 0;
2571 int pd_idx
= sh
->pd_idx
;
2572 int qd_idx
= sh
->qd_idx
;
2574 set_bit(STRIPE_HANDLE
, &sh
->state
);
2576 BUG_ON(s
->failed
> 2);
2577 BUG_ON(s
->uptodate
< disks
);
2578 /* Want to check and possibly repair P and Q.
2579 * However there could be one 'failed' device, in which
2580 * case we can only check one of them, possibly using the
2581 * other to generate missing data
2584 /* If !tmp_page, we cannot do the calculations,
2585 * but as we have set STRIPE_HANDLE, we will soon be called
2586 * by stripe_handle with a tmp_page - just wait until then.
2589 if (s
->failed
== r6s
->q_failed
) {
2590 /* The only possible failed device holds 'Q', so it
2591 * makes sense to check P (If anything else were failed,
2592 * we would have used P to recreate it).
2594 compute_block_1(sh
, pd_idx
, 1);
2595 if (!page_is_zero(sh
->dev
[pd_idx
].page
)) {
2596 compute_block_1(sh
, pd_idx
, 0);
2600 if (!r6s
->q_failed
&& s
->failed
< 2) {
2601 /* q is not failed, and we didn't use it to generate
2602 * anything, so it makes sense to check it
2604 memcpy(page_address(tmp_page
),
2605 page_address(sh
->dev
[qd_idx
].page
),
2607 compute_parity6(sh
, UPDATE_PARITY
);
2608 if (memcmp(page_address(tmp_page
),
2609 page_address(sh
->dev
[qd_idx
].page
),
2610 STRIPE_SIZE
) != 0) {
2611 clear_bit(STRIPE_INSYNC
, &sh
->state
);
2615 if (update_p
|| update_q
) {
2616 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2617 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2618 /* don't try to repair!! */
2619 update_p
= update_q
= 0;
2622 /* now write out any block on a failed drive,
2623 * or P or Q if they need it
2626 if (s
->failed
== 2) {
2627 dev
= &sh
->dev
[r6s
->failed_num
[1]];
2629 set_bit(R5_LOCKED
, &dev
->flags
);
2630 set_bit(R5_Wantwrite
, &dev
->flags
);
2632 if (s
->failed
>= 1) {
2633 dev
= &sh
->dev
[r6s
->failed_num
[0]];
2635 set_bit(R5_LOCKED
, &dev
->flags
);
2636 set_bit(R5_Wantwrite
, &dev
->flags
);
2640 dev
= &sh
->dev
[pd_idx
];
2642 set_bit(R5_LOCKED
, &dev
->flags
);
2643 set_bit(R5_Wantwrite
, &dev
->flags
);
2646 dev
= &sh
->dev
[qd_idx
];
2648 set_bit(R5_LOCKED
, &dev
->flags
);
2649 set_bit(R5_Wantwrite
, &dev
->flags
);
2651 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2653 set_bit(STRIPE_INSYNC
, &sh
->state
);
2657 static void handle_stripe_expansion(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2658 struct r6_state
*r6s
)
2662 /* We have read all the blocks in this stripe and now we need to
2663 * copy some of them into a target stripe for expand.
2665 struct dma_async_tx_descriptor
*tx
= NULL
;
2666 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2667 for (i
= 0; i
< sh
->disks
; i
++)
2668 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
2670 struct stripe_head
*sh2
;
2672 sector_t bn
= compute_blocknr(sh
, i
, 1);
2673 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
2675 sh2
= get_active_stripe(conf
, s
, 0, 1);
2677 /* so far only the early blocks of this stripe
2678 * have been requested. When later blocks
2679 * get requested, we will try again
2682 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
2683 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
2684 /* must have already done this block */
2685 release_stripe(sh2
);
2689 /* place all the copies on one channel */
2690 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
2691 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
2692 ASYNC_TX_DEP_ACK
, tx
, NULL
, NULL
);
2694 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
2695 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
2696 for (j
= 0; j
< conf
->raid_disks
; j
++)
2697 if (j
!= sh2
->pd_idx
&&
2698 (!r6s
|| j
!= sh2
->qd_idx
) &&
2699 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
2701 if (j
== conf
->raid_disks
) {
2702 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
2703 set_bit(STRIPE_HANDLE
, &sh2
->state
);
2705 release_stripe(sh2
);
2708 /* done submitting copies, wait for them to complete */
2711 dma_wait_for_async_tx(tx
);
2717 * handle_stripe - do things to a stripe.
2719 * We lock the stripe and then examine the state of various bits
2720 * to see what needs to be done.
2722 * return some read request which now have data
2723 * return some write requests which are safely on disc
2724 * schedule a read on some buffers
2725 * schedule a write of some buffers
2726 * return confirmation of parity correctness
2728 * buffers are taken off read_list or write_list, and bh_cache buffers
2729 * get BH_Lock set before the stripe lock is released.
2733 static bool handle_stripe5(struct stripe_head
*sh
)
2735 raid5_conf_t
*conf
= sh
->raid_conf
;
2736 int disks
= sh
->disks
, i
;
2737 struct bio
*return_bi
= NULL
;
2738 struct stripe_head_state s
;
2740 mdk_rdev_t
*blocked_rdev
= NULL
;
2743 memset(&s
, 0, sizeof(s
));
2744 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2745 "reconstruct:%d\n", (unsigned long long)sh
->sector
, sh
->state
,
2746 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->check_state
,
2747 sh
->reconstruct_state
);
2749 spin_lock(&sh
->lock
);
2750 clear_bit(STRIPE_HANDLE
, &sh
->state
);
2751 clear_bit(STRIPE_DELAYED
, &sh
->state
);
2753 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
2754 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2755 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
2757 /* Now to look around and see what can be done */
2759 for (i
=disks
; i
--; ) {
2761 struct r5dev
*dev
= &sh
->dev
[i
];
2762 clear_bit(R5_Insync
, &dev
->flags
);
2764 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2765 "written %p\n", i
, dev
->flags
, dev
->toread
, dev
->read
,
2766 dev
->towrite
, dev
->written
);
2768 /* maybe we can request a biofill operation
2770 * new wantfill requests are only permitted while
2771 * ops_complete_biofill is guaranteed to be inactive
2773 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
2774 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
2775 set_bit(R5_Wantfill
, &dev
->flags
);
2777 /* now count some things */
2778 if (test_bit(R5_LOCKED
, &dev
->flags
)) s
.locked
++;
2779 if (test_bit(R5_UPTODATE
, &dev
->flags
)) s
.uptodate
++;
2780 if (test_bit(R5_Wantcompute
, &dev
->flags
)) s
.compute
++;
2782 if (test_bit(R5_Wantfill
, &dev
->flags
))
2784 else if (dev
->toread
)
2788 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
2793 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2794 if (blocked_rdev
== NULL
&&
2795 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
2796 blocked_rdev
= rdev
;
2797 atomic_inc(&rdev
->nr_pending
);
2799 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)) {
2800 /* The ReadError flag will just be confusing now */
2801 clear_bit(R5_ReadError
, &dev
->flags
);
2802 clear_bit(R5_ReWrite
, &dev
->flags
);
2804 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)
2805 || test_bit(R5_ReadError
, &dev
->flags
)) {
2809 set_bit(R5_Insync
, &dev
->flags
);
2813 if (unlikely(blocked_rdev
)) {
2814 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
2815 s
.to_write
|| s
.written
) {
2816 set_bit(STRIPE_HANDLE
, &sh
->state
);
2819 /* There is nothing for the blocked_rdev to block */
2820 rdev_dec_pending(blocked_rdev
, conf
->mddev
);
2821 blocked_rdev
= NULL
;
2824 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
2825 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
2826 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
2829 pr_debug("locked=%d uptodate=%d to_read=%d"
2830 " to_write=%d failed=%d failed_num=%d\n",
2831 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
,
2832 s
.failed
, s
.failed_num
);
2833 /* check if the array has lost two devices and, if so, some requests might
2836 if (s
.failed
> 1 && s
.to_read
+s
.to_write
+s
.written
)
2837 handle_failed_stripe(conf
, sh
, &s
, disks
, &return_bi
);
2838 if (s
.failed
> 1 && s
.syncing
) {
2839 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
2840 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2844 /* might be able to return some write requests if the parity block
2845 * is safe, or on a failed drive
2847 dev
= &sh
->dev
[sh
->pd_idx
];
2849 ((test_bit(R5_Insync
, &dev
->flags
) &&
2850 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2851 test_bit(R5_UPTODATE
, &dev
->flags
)) ||
2852 (s
.failed
== 1 && s
.failed_num
== sh
->pd_idx
)))
2853 handle_stripe_clean_event(conf
, sh
, disks
, &return_bi
);
2855 /* Now we might consider reading some blocks, either to check/generate
2856 * parity, or to satisfy requests
2857 * or to load a block that is being partially written.
2859 if (s
.to_read
|| s
.non_overwrite
||
2860 (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
)) || s
.expanding
)
2861 handle_stripe_fill5(sh
, &s
, disks
);
2863 /* Now we check to see if any write operations have recently
2867 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
2869 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
2870 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
2871 sh
->reconstruct_state
= reconstruct_state_idle
;
2873 /* All the 'written' buffers and the parity block are ready to
2874 * be written back to disk
2876 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
2877 for (i
= disks
; i
--; ) {
2879 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
2880 (i
== sh
->pd_idx
|| dev
->written
)) {
2881 pr_debug("Writing block %d\n", i
);
2882 set_bit(R5_Wantwrite
, &dev
->flags
);
2885 if (!test_bit(R5_Insync
, &dev
->flags
) ||
2886 (i
== sh
->pd_idx
&& s
.failed
== 0))
2887 set_bit(STRIPE_INSYNC
, &sh
->state
);
2890 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2891 atomic_dec(&conf
->preread_active_stripes
);
2892 if (atomic_read(&conf
->preread_active_stripes
) <
2894 md_wakeup_thread(conf
->mddev
->thread
);
2898 /* Now to consider new write requests and what else, if anything
2899 * should be read. We do not handle new writes when:
2900 * 1/ A 'write' operation (copy+xor) is already in flight.
2901 * 2/ A 'check' operation is in flight, as it may clobber the parity
2904 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
2905 handle_stripe_dirtying5(conf
, sh
, &s
, disks
);
2907 /* maybe we need to check and possibly fix the parity for this stripe
2908 * Any reads will already have been scheduled, so we just see if enough
2909 * data is available. The parity check is held off while parity
2910 * dependent operations are in flight.
2912 if (sh
->check_state
||
2913 (s
.syncing
&& s
.locked
== 0 &&
2914 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
2915 !test_bit(STRIPE_INSYNC
, &sh
->state
)))
2916 handle_parity_checks5(conf
, sh
, &s
, disks
);
2918 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
2919 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,1);
2920 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2923 /* If the failed drive is just a ReadError, then we might need to progress
2924 * the repair/check process
2926 if (s
.failed
== 1 && !conf
->mddev
->ro
&&
2927 test_bit(R5_ReadError
, &sh
->dev
[s
.failed_num
].flags
)
2928 && !test_bit(R5_LOCKED
, &sh
->dev
[s
.failed_num
].flags
)
2929 && test_bit(R5_UPTODATE
, &sh
->dev
[s
.failed_num
].flags
)
2931 dev
= &sh
->dev
[s
.failed_num
];
2932 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
2933 set_bit(R5_Wantwrite
, &dev
->flags
);
2934 set_bit(R5_ReWrite
, &dev
->flags
);
2935 set_bit(R5_LOCKED
, &dev
->flags
);
2938 /* let's read it back */
2939 set_bit(R5_Wantread
, &dev
->flags
);
2940 set_bit(R5_LOCKED
, &dev
->flags
);
2945 /* Finish reconstruct operations initiated by the expansion process */
2946 if (sh
->reconstruct_state
== reconstruct_state_result
) {
2947 sh
->reconstruct_state
= reconstruct_state_idle
;
2948 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
2949 for (i
= conf
->raid_disks
; i
--; ) {
2950 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
2951 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2956 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
2957 !sh
->reconstruct_state
) {
2958 /* Need to write out all blocks after computing parity */
2959 sh
->disks
= conf
->raid_disks
;
2960 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
2961 schedule_reconstruction5(sh
, &s
, 1, 1);
2962 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
2963 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
2964 atomic_dec(&conf
->reshape_stripes
);
2965 wake_up(&conf
->wait_for_overlap
);
2966 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
2969 if (s
.expanding
&& s
.locked
== 0 &&
2970 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
2971 handle_stripe_expansion(conf
, sh
, NULL
);
2974 spin_unlock(&sh
->lock
);
2976 /* wait for this device to become unblocked */
2977 if (unlikely(blocked_rdev
))
2978 md_wait_for_blocked_rdev(blocked_rdev
, conf
->mddev
);
2981 raid5_run_ops(sh
, s
.ops_request
);
2985 return_io(return_bi
);
2987 return blocked_rdev
== NULL
;
2990 static bool handle_stripe6(struct stripe_head
*sh
, struct page
*tmp_page
)
2992 raid5_conf_t
*conf
= sh
->raid_conf
;
2993 int disks
= sh
->disks
;
2994 struct bio
*return_bi
= NULL
;
2995 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
;
2996 struct stripe_head_state s
;
2997 struct r6_state r6s
;
2998 struct r5dev
*dev
, *pdev
, *qdev
;
2999 mdk_rdev_t
*blocked_rdev
= NULL
;
3001 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3002 "pd_idx=%d, qd_idx=%d\n",
3003 (unsigned long long)sh
->sector
, sh
->state
,
3004 atomic_read(&sh
->count
), pd_idx
, qd_idx
);
3005 memset(&s
, 0, sizeof(s
));
3007 spin_lock(&sh
->lock
);
3008 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3009 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3011 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
3012 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3013 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3014 /* Now to look around and see what can be done */
3017 for (i
=disks
; i
--; ) {
3020 clear_bit(R5_Insync
, &dev
->flags
);
3022 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3023 i
, dev
->flags
, dev
->toread
, dev
->towrite
, dev
->written
);
3024 /* maybe we can reply to a read */
3025 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
) {
3026 struct bio
*rbi
, *rbi2
;
3027 pr_debug("Return read for disc %d\n", i
);
3028 spin_lock_irq(&conf
->device_lock
);
3031 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
3032 wake_up(&conf
->wait_for_overlap
);
3033 spin_unlock_irq(&conf
->device_lock
);
3034 while (rbi
&& rbi
->bi_sector
< dev
->sector
+ STRIPE_SECTORS
) {
3035 copy_data(0, rbi
, dev
->page
, dev
->sector
);
3036 rbi2
= r5_next_bio(rbi
, dev
->sector
);
3037 spin_lock_irq(&conf
->device_lock
);
3038 if (!raid5_dec_bi_phys_segments(rbi
)) {
3039 rbi
->bi_next
= return_bi
;
3042 spin_unlock_irq(&conf
->device_lock
);
3047 /* now count some things */
3048 if (test_bit(R5_LOCKED
, &dev
->flags
)) s
.locked
++;
3049 if (test_bit(R5_UPTODATE
, &dev
->flags
)) s
.uptodate
++;
3056 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3061 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3062 if (blocked_rdev
== NULL
&&
3063 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
3064 blocked_rdev
= rdev
;
3065 atomic_inc(&rdev
->nr_pending
);
3067 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)) {
3068 /* The ReadError flag will just be confusing now */
3069 clear_bit(R5_ReadError
, &dev
->flags
);
3070 clear_bit(R5_ReWrite
, &dev
->flags
);
3072 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)
3073 || test_bit(R5_ReadError
, &dev
->flags
)) {
3075 r6s
.failed_num
[s
.failed
] = i
;
3078 set_bit(R5_Insync
, &dev
->flags
);
3082 if (unlikely(blocked_rdev
)) {
3083 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3084 s
.to_write
|| s
.written
) {
3085 set_bit(STRIPE_HANDLE
, &sh
->state
);
3088 /* There is nothing for the blocked_rdev to block */
3089 rdev_dec_pending(blocked_rdev
, conf
->mddev
);
3090 blocked_rdev
= NULL
;
3093 pr_debug("locked=%d uptodate=%d to_read=%d"
3094 " to_write=%d failed=%d failed_num=%d,%d\n",
3095 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3096 r6s
.failed_num
[0], r6s
.failed_num
[1]);
3097 /* check if the array has lost >2 devices and, if so, some requests
3098 * might need to be failed
3100 if (s
.failed
> 2 && s
.to_read
+s
.to_write
+s
.written
)
3101 handle_failed_stripe(conf
, sh
, &s
, disks
, &return_bi
);
3102 if (s
.failed
> 2 && s
.syncing
) {
3103 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
3104 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3109 * might be able to return some write requests if the parity blocks
3110 * are safe, or on a failed drive
3112 pdev
= &sh
->dev
[pd_idx
];
3113 r6s
.p_failed
= (s
.failed
>= 1 && r6s
.failed_num
[0] == pd_idx
)
3114 || (s
.failed
>= 2 && r6s
.failed_num
[1] == pd_idx
);
3115 qdev
= &sh
->dev
[qd_idx
];
3116 r6s
.q_failed
= (s
.failed
>= 1 && r6s
.failed_num
[0] == qd_idx
)
3117 || (s
.failed
>= 2 && r6s
.failed_num
[1] == qd_idx
);
3120 ( r6s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3121 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3122 && test_bit(R5_UPTODATE
, &pdev
->flags
)))) &&
3123 ( r6s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3124 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3125 && test_bit(R5_UPTODATE
, &qdev
->flags
)))))
3126 handle_stripe_clean_event(conf
, sh
, disks
, &return_bi
);
3128 /* Now we might consider reading some blocks, either to check/generate
3129 * parity, or to satisfy requests
3130 * or to load a block that is being partially written.
3132 if (s
.to_read
|| s
.non_overwrite
|| (s
.to_write
&& s
.failed
) ||
3133 (s
.syncing
&& (s
.uptodate
< disks
)) || s
.expanding
)
3134 handle_stripe_fill6(sh
, &s
, &r6s
, disks
);
3136 /* now to consider writing and what else, if anything should be read */
3138 handle_stripe_dirtying6(conf
, sh
, &s
, &r6s
, disks
);
3140 /* maybe we need to check and possibly fix the parity for this stripe
3141 * Any reads will already have been scheduled, so we just see if enough
3144 if (s
.syncing
&& s
.locked
== 0 && !test_bit(STRIPE_INSYNC
, &sh
->state
))
3145 handle_parity_checks6(conf
, sh
, &s
, &r6s
, tmp_page
, disks
);
3147 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3148 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,1);
3149 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3152 /* If the failed drives are just a ReadError, then we might need
3153 * to progress the repair/check process
3155 if (s
.failed
<= 2 && !conf
->mddev
->ro
)
3156 for (i
= 0; i
< s
.failed
; i
++) {
3157 dev
= &sh
->dev
[r6s
.failed_num
[i
]];
3158 if (test_bit(R5_ReadError
, &dev
->flags
)
3159 && !test_bit(R5_LOCKED
, &dev
->flags
)
3160 && test_bit(R5_UPTODATE
, &dev
->flags
)
3162 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3163 set_bit(R5_Wantwrite
, &dev
->flags
);
3164 set_bit(R5_ReWrite
, &dev
->flags
);
3165 set_bit(R5_LOCKED
, &dev
->flags
);
3167 /* let's read it back */
3168 set_bit(R5_Wantread
, &dev
->flags
);
3169 set_bit(R5_LOCKED
, &dev
->flags
);
3174 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
3175 /* Need to write out all blocks after computing P&Q */
3176 sh
->disks
= conf
->raid_disks
;
3177 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3178 compute_parity6(sh
, RECONSTRUCT_WRITE
);
3179 for (i
= conf
->raid_disks
; i
-- ; ) {
3180 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3182 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3184 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3185 } else if (s
.expanded
) {
3186 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3187 atomic_dec(&conf
->reshape_stripes
);
3188 wake_up(&conf
->wait_for_overlap
);
3189 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3192 if (s
.expanding
&& s
.locked
== 0 &&
3193 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3194 handle_stripe_expansion(conf
, sh
, &r6s
);
3197 spin_unlock(&sh
->lock
);
3199 /* wait for this device to become unblocked */
3200 if (unlikely(blocked_rdev
))
3201 md_wait_for_blocked_rdev(blocked_rdev
, conf
->mddev
);
3205 return_io(return_bi
);
3207 return blocked_rdev
== NULL
;
3210 /* returns true if the stripe was handled */
3211 static bool handle_stripe(struct stripe_head
*sh
, struct page
*tmp_page
)
3213 if (sh
->raid_conf
->level
== 6)
3214 return handle_stripe6(sh
, tmp_page
);
3216 return handle_stripe5(sh
);
3221 static void raid5_activate_delayed(raid5_conf_t
*conf
)
3223 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3224 while (!list_empty(&conf
->delayed_list
)) {
3225 struct list_head
*l
= conf
->delayed_list
.next
;
3226 struct stripe_head
*sh
;
3227 sh
= list_entry(l
, struct stripe_head
, lru
);
3229 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3230 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3231 atomic_inc(&conf
->preread_active_stripes
);
3232 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3235 blk_plug_device(conf
->mddev
->queue
);
3238 static void activate_bit_delay(raid5_conf_t
*conf
)
3240 /* device_lock is held */
3241 struct list_head head
;
3242 list_add(&head
, &conf
->bitmap_list
);
3243 list_del_init(&conf
->bitmap_list
);
3244 while (!list_empty(&head
)) {
3245 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3246 list_del_init(&sh
->lru
);
3247 atomic_inc(&sh
->count
);
3248 __release_stripe(conf
, sh
);
3252 static void unplug_slaves(mddev_t
*mddev
)
3254 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3258 for (i
=0; i
<mddev
->raid_disks
; i
++) {
3259 mdk_rdev_t
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3260 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) && atomic_read(&rdev
->nr_pending
)) {
3261 struct request_queue
*r_queue
= bdev_get_queue(rdev
->bdev
);
3263 atomic_inc(&rdev
->nr_pending
);
3266 blk_unplug(r_queue
);
3268 rdev_dec_pending(rdev
, mddev
);
3275 static void raid5_unplug_device(struct request_queue
*q
)
3277 mddev_t
*mddev
= q
->queuedata
;
3278 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3279 unsigned long flags
;
3281 spin_lock_irqsave(&conf
->device_lock
, flags
);
3283 if (blk_remove_plug(q
)) {
3285 raid5_activate_delayed(conf
);
3287 md_wakeup_thread(mddev
->thread
);
3289 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3291 unplug_slaves(mddev
);
3294 static int raid5_congested(void *data
, int bits
)
3296 mddev_t
*mddev
= data
;
3297 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3299 /* No difference between reads and writes. Just check
3300 * how busy the stripe_cache is
3302 if (conf
->inactive_blocked
)
3306 if (list_empty_careful(&conf
->inactive_list
))
3312 /* We want read requests to align with chunks where possible,
3313 * but write requests don't need to.
3315 static int raid5_mergeable_bvec(struct request_queue
*q
,
3316 struct bvec_merge_data
*bvm
,
3317 struct bio_vec
*biovec
)
3319 mddev_t
*mddev
= q
->queuedata
;
3320 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3322 unsigned int chunk_sectors
= mddev
->chunk_size
>> 9;
3323 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3325 if ((bvm
->bi_rw
& 1) == WRITE
)
3326 return biovec
->bv_len
; /* always allow writes to be mergeable */
3328 if (mddev
->new_chunk
< mddev
->chunk_size
)
3329 chunk_sectors
= mddev
->new_chunk
>> 9;
3330 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3331 if (max
< 0) max
= 0;
3332 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3333 return biovec
->bv_len
;
3339 static int in_chunk_boundary(mddev_t
*mddev
, struct bio
*bio
)
3341 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3342 unsigned int chunk_sectors
= mddev
->chunk_size
>> 9;
3343 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3345 if (mddev
->new_chunk
< mddev
->chunk_size
)
3346 chunk_sectors
= mddev
->new_chunk
>> 9;
3347 return chunk_sectors
>=
3348 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3352 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3353 * later sampled by raid5d.
3355 static void add_bio_to_retry(struct bio
*bi
,raid5_conf_t
*conf
)
3357 unsigned long flags
;
3359 spin_lock_irqsave(&conf
->device_lock
, flags
);
3361 bi
->bi_next
= conf
->retry_read_aligned_list
;
3362 conf
->retry_read_aligned_list
= bi
;
3364 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3365 md_wakeup_thread(conf
->mddev
->thread
);
3369 static struct bio
*remove_bio_from_retry(raid5_conf_t
*conf
)
3373 bi
= conf
->retry_read_aligned
;
3375 conf
->retry_read_aligned
= NULL
;
3378 bi
= conf
->retry_read_aligned_list
;
3380 conf
->retry_read_aligned_list
= bi
->bi_next
;
3383 * this sets the active strip count to 1 and the processed
3384 * strip count to zero (upper 8 bits)
3386 bi
->bi_phys_segments
= 1; /* biased count of active stripes */
3394 * The "raid5_align_endio" should check if the read succeeded and if it
3395 * did, call bio_endio on the original bio (having bio_put the new bio
3397 * If the read failed..
3399 static void raid5_align_endio(struct bio
*bi
, int error
)
3401 struct bio
* raid_bi
= bi
->bi_private
;
3404 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3409 mddev
= raid_bi
->bi_bdev
->bd_disk
->queue
->queuedata
;
3410 conf
= mddev_to_conf(mddev
);
3411 rdev
= (void*)raid_bi
->bi_next
;
3412 raid_bi
->bi_next
= NULL
;
3414 rdev_dec_pending(rdev
, conf
->mddev
);
3416 if (!error
&& uptodate
) {
3417 bio_endio(raid_bi
, 0);
3418 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3419 wake_up(&conf
->wait_for_stripe
);
3424 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3426 add_bio_to_retry(raid_bi
, conf
);
3429 static int bio_fits_rdev(struct bio
*bi
)
3431 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3433 if ((bi
->bi_size
>>9) > q
->max_sectors
)
3435 blk_recount_segments(q
, bi
);
3436 if (bi
->bi_phys_segments
> q
->max_phys_segments
)
3439 if (q
->merge_bvec_fn
)
3440 /* it's too hard to apply the merge_bvec_fn at this stage,
3449 static int chunk_aligned_read(struct request_queue
*q
, struct bio
* raid_bio
)
3451 mddev_t
*mddev
= q
->queuedata
;
3452 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3453 unsigned int dd_idx
;
3454 struct bio
* align_bi
;
3457 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3458 pr_debug("chunk_aligned_read : non aligned\n");
3462 * use bio_clone to make a copy of the bio
3464 align_bi
= bio_clone(raid_bio
, GFP_NOIO
);
3468 * set bi_end_io to a new function, and set bi_private to the
3471 align_bi
->bi_end_io
= raid5_align_endio
;
3472 align_bi
->bi_private
= raid_bio
;
3476 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3481 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3482 if (rdev
&& test_bit(In_sync
, &rdev
->flags
)) {
3483 atomic_inc(&rdev
->nr_pending
);
3485 raid_bio
->bi_next
= (void*)rdev
;
3486 align_bi
->bi_bdev
= rdev
->bdev
;
3487 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3488 align_bi
->bi_sector
+= rdev
->data_offset
;
3490 if (!bio_fits_rdev(align_bi
)) {
3491 /* too big in some way */
3493 rdev_dec_pending(rdev
, mddev
);
3497 spin_lock_irq(&conf
->device_lock
);
3498 wait_event_lock_irq(conf
->wait_for_stripe
,
3500 conf
->device_lock
, /* nothing */);
3501 atomic_inc(&conf
->active_aligned_reads
);
3502 spin_unlock_irq(&conf
->device_lock
);
3504 generic_make_request(align_bi
);
3513 /* __get_priority_stripe - get the next stripe to process
3515 * Full stripe writes are allowed to pass preread active stripes up until
3516 * the bypass_threshold is exceeded. In general the bypass_count
3517 * increments when the handle_list is handled before the hold_list; however, it
3518 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3519 * stripe with in flight i/o. The bypass_count will be reset when the
3520 * head of the hold_list has changed, i.e. the head was promoted to the
3523 static struct stripe_head
*__get_priority_stripe(raid5_conf_t
*conf
)
3525 struct stripe_head
*sh
;
3527 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3529 list_empty(&conf
->handle_list
) ? "empty" : "busy",
3530 list_empty(&conf
->hold_list
) ? "empty" : "busy",
3531 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
3533 if (!list_empty(&conf
->handle_list
)) {
3534 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
3536 if (list_empty(&conf
->hold_list
))
3537 conf
->bypass_count
= 0;
3538 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
3539 if (conf
->hold_list
.next
== conf
->last_hold
)
3540 conf
->bypass_count
++;
3542 conf
->last_hold
= conf
->hold_list
.next
;
3543 conf
->bypass_count
-= conf
->bypass_threshold
;
3544 if (conf
->bypass_count
< 0)
3545 conf
->bypass_count
= 0;
3548 } else if (!list_empty(&conf
->hold_list
) &&
3549 ((conf
->bypass_threshold
&&
3550 conf
->bypass_count
> conf
->bypass_threshold
) ||
3551 atomic_read(&conf
->pending_full_writes
) == 0)) {
3552 sh
= list_entry(conf
->hold_list
.next
,
3554 conf
->bypass_count
-= conf
->bypass_threshold
;
3555 if (conf
->bypass_count
< 0)
3556 conf
->bypass_count
= 0;
3560 list_del_init(&sh
->lru
);
3561 atomic_inc(&sh
->count
);
3562 BUG_ON(atomic_read(&sh
->count
) != 1);
3566 static int make_request(struct request_queue
*q
, struct bio
* bi
)
3568 mddev_t
*mddev
= q
->queuedata
;
3569 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3571 sector_t new_sector
;
3572 sector_t logical_sector
, last_sector
;
3573 struct stripe_head
*sh
;
3574 const int rw
= bio_data_dir(bi
);
3577 if (unlikely(bio_barrier(bi
))) {
3578 bio_endio(bi
, -EOPNOTSUPP
);
3582 md_write_start(mddev
, bi
);
3584 cpu
= part_stat_lock();
3585 part_stat_inc(cpu
, &mddev
->gendisk
->part0
, ios
[rw
]);
3586 part_stat_add(cpu
, &mddev
->gendisk
->part0
, sectors
[rw
],
3591 mddev
->reshape_position
== MaxSector
&&
3592 chunk_aligned_read(q
,bi
))
3595 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
3596 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
3598 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
3600 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
3602 int disks
, data_disks
;
3607 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
3608 if (likely(conf
->reshape_progress
== MaxSector
))
3609 disks
= conf
->raid_disks
;
3611 /* spinlock is needed as reshape_progress may be
3612 * 64bit on a 32bit platform, and so it might be
3613 * possible to see a half-updated value
3614 * Ofcourse reshape_progress could change after
3615 * the lock is dropped, so once we get a reference
3616 * to the stripe that we think it is, we will have
3619 spin_lock_irq(&conf
->device_lock
);
3620 disks
= conf
->raid_disks
;
3621 if (mddev
->delta_disks
< 0
3622 ? logical_sector
< conf
->reshape_progress
3623 : logical_sector
>= conf
->reshape_progress
) {
3624 disks
= conf
->previous_raid_disks
;
3627 if (mddev
->delta_disks
< 0
3628 ? logical_sector
< conf
->reshape_safe
3629 : logical_sector
>= conf
->reshape_safe
) {
3630 spin_unlock_irq(&conf
->device_lock
);
3635 spin_unlock_irq(&conf
->device_lock
);
3637 data_disks
= disks
- conf
->max_degraded
;
3639 new_sector
= raid5_compute_sector(conf
, logical_sector
,
3642 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3643 (unsigned long long)new_sector
,
3644 (unsigned long long)logical_sector
);
3646 sh
= get_active_stripe(conf
, new_sector
, previous
,
3647 (bi
->bi_rw
&RWA_MASK
));
3649 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
3650 /* expansion might have moved on while waiting for a
3651 * stripe, so we must do the range check again.
3652 * Expansion could still move past after this
3653 * test, but as we are holding a reference to
3654 * 'sh', we know that if that happens,
3655 * STRIPE_EXPANDING will get set and the expansion
3656 * won't proceed until we finish with the stripe.
3659 spin_lock_irq(&conf
->device_lock
);
3660 if ((mddev
->delta_disks
< 0
3661 ? logical_sector
>= conf
->reshape_progress
3662 : logical_sector
< conf
->reshape_progress
)
3664 /* mismatch, need to try again */
3666 spin_unlock_irq(&conf
->device_lock
);
3672 /* FIXME what if we get a false positive because these
3673 * are being updated.
3675 if (logical_sector
>= mddev
->suspend_lo
&&
3676 logical_sector
< mddev
->suspend_hi
) {
3682 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
3683 !add_stripe_bio(sh
, bi
, dd_idx
, (bi
->bi_rw
&RW_MASK
))) {
3684 /* Stripe is busy expanding or
3685 * add failed due to overlap. Flush everything
3688 raid5_unplug_device(mddev
->queue
);
3693 finish_wait(&conf
->wait_for_overlap
, &w
);
3694 set_bit(STRIPE_HANDLE
, &sh
->state
);
3695 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3698 /* cannot get stripe for read-ahead, just give-up */
3699 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3700 finish_wait(&conf
->wait_for_overlap
, &w
);
3705 spin_lock_irq(&conf
->device_lock
);
3706 remaining
= raid5_dec_bi_phys_segments(bi
);
3707 spin_unlock_irq(&conf
->device_lock
);
3708 if (remaining
== 0) {
3711 md_write_end(mddev
);
3718 static sector_t
raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
);
3720 static sector_t
reshape_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
)
3722 /* reshaping is quite different to recovery/resync so it is
3723 * handled quite separately ... here.
3725 * On each call to sync_request, we gather one chunk worth of
3726 * destination stripes and flag them as expanding.
3727 * Then we find all the source stripes and request reads.
3728 * As the reads complete, handle_stripe will copy the data
3729 * into the destination stripe and release that stripe.
3731 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
3732 struct stripe_head
*sh
;
3733 sector_t first_sector
, last_sector
;
3734 int raid_disks
= conf
->previous_raid_disks
;
3735 int data_disks
= raid_disks
- conf
->max_degraded
;
3736 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
3739 sector_t writepos
, safepos
, gap
;
3740 sector_t stripe_addr
;
3741 int reshape_sectors
;
3743 if (sector_nr
== 0) {
3744 /* If restarting in the middle, skip the initial sectors */
3745 if (mddev
->delta_disks
< 0 &&
3746 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
3747 sector_nr
= raid5_size(mddev
, 0, 0)
3748 - conf
->reshape_progress
;
3749 } else if (mddev
->delta_disks
> 0 &&
3750 conf
->reshape_progress
> 0)
3751 sector_nr
= conf
->reshape_progress
;
3752 sector_div(sector_nr
, new_data_disks
);
3759 /* We need to process a full chunk at a time.
3760 * If old and new chunk sizes differ, we need to process the
3763 if (mddev
->new_chunk
> mddev
->chunk_size
)
3764 reshape_sectors
= mddev
->new_chunk
/ 512;
3766 reshape_sectors
= mddev
->chunk_size
/ 512;
3768 /* we update the metadata when there is more than 3Meg
3769 * in the block range (that is rather arbitrary, should
3770 * probably be time based) or when the data about to be
3771 * copied would over-write the source of the data at
3772 * the front of the range.
3773 * i.e. one new_stripe along from reshape_progress new_maps
3774 * to after where reshape_safe old_maps to
3776 writepos
= conf
->reshape_progress
;
3777 sector_div(writepos
, new_data_disks
);
3778 safepos
= conf
->reshape_safe
;
3779 sector_div(safepos
, data_disks
);
3780 if (mddev
->delta_disks
< 0) {
3781 writepos
-= reshape_sectors
;
3782 safepos
+= reshape_sectors
;
3783 gap
= conf
->reshape_safe
- conf
->reshape_progress
;
3785 writepos
+= reshape_sectors
;
3786 safepos
-= reshape_sectors
;
3787 gap
= conf
->reshape_progress
- conf
->reshape_safe
;
3790 if ((mddev
->delta_disks
< 0
3791 ? writepos
< safepos
3792 : writepos
> safepos
) ||
3793 gap
> (new_data_disks
)*3000*2 /*3Meg*/) {
3794 /* Cannot proceed until we've updated the superblock... */
3795 wait_event(conf
->wait_for_overlap
,
3796 atomic_read(&conf
->reshape_stripes
)==0);
3797 mddev
->reshape_position
= conf
->reshape_progress
;
3798 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
3799 md_wakeup_thread(mddev
->thread
);
3800 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
3801 kthread_should_stop());
3802 spin_lock_irq(&conf
->device_lock
);
3803 conf
->reshape_safe
= mddev
->reshape_position
;
3804 spin_unlock_irq(&conf
->device_lock
);
3805 wake_up(&conf
->wait_for_overlap
);
3808 if (mddev
->delta_disks
< 0) {
3809 BUG_ON(conf
->reshape_progress
== 0);
3810 stripe_addr
= writepos
;
3811 BUG_ON((mddev
->dev_sectors
&
3812 ~((sector_t
)reshape_sectors
- 1))
3813 - reshape_sectors
- stripe_addr
3816 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
3817 stripe_addr
= sector_nr
;
3819 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
3822 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0);
3823 set_bit(STRIPE_EXPANDING
, &sh
->state
);
3824 atomic_inc(&conf
->reshape_stripes
);
3825 /* If any of this stripe is beyond the end of the old
3826 * array, then we need to zero those blocks
3828 for (j
=sh
->disks
; j
--;) {
3830 if (j
== sh
->pd_idx
)
3832 if (conf
->level
== 6 &&
3835 s
= compute_blocknr(sh
, j
, 0);
3836 if (s
< raid5_size(mddev
, 0, 0)) {
3840 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
3841 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
3842 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
3845 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3846 set_bit(STRIPE_HANDLE
, &sh
->state
);
3850 spin_lock_irq(&conf
->device_lock
);
3851 if (mddev
->delta_disks
< 0)
3852 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
3854 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
3855 spin_unlock_irq(&conf
->device_lock
);
3856 /* Ok, those stripe are ready. We can start scheduling
3857 * reads on the source stripes.
3858 * The source stripes are determined by mapping the first and last
3859 * block on the destination stripes.
3862 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
3865 raid5_compute_sector(conf
, ((stripe_addr
+conf
->chunk_size
/512)
3866 *(new_data_disks
) - 1),
3868 if (last_sector
>= mddev
->dev_sectors
)
3869 last_sector
= mddev
->dev_sectors
- 1;
3870 while (first_sector
<= last_sector
) {
3871 sh
= get_active_stripe(conf
, first_sector
, 1, 0);
3872 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3873 set_bit(STRIPE_HANDLE
, &sh
->state
);
3875 first_sector
+= STRIPE_SECTORS
;
3877 /* If this takes us to the resync_max point where we have to pause,
3878 * then we need to write out the superblock.
3880 sector_nr
+= reshape_sectors
;
3881 if (sector_nr
>= mddev
->resync_max
) {
3882 /* Cannot proceed until we've updated the superblock... */
3883 wait_event(conf
->wait_for_overlap
,
3884 atomic_read(&conf
->reshape_stripes
) == 0);
3885 mddev
->reshape_position
= conf
->reshape_progress
;
3886 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
3887 md_wakeup_thread(mddev
->thread
);
3888 wait_event(mddev
->sb_wait
,
3889 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
3890 || kthread_should_stop());
3891 spin_lock_irq(&conf
->device_lock
);
3892 conf
->reshape_safe
= mddev
->reshape_position
;
3893 spin_unlock_irq(&conf
->device_lock
);
3894 wake_up(&conf
->wait_for_overlap
);
3896 return reshape_sectors
;
3899 /* FIXME go_faster isn't used */
3900 static inline sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
3902 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
3903 struct stripe_head
*sh
;
3904 sector_t max_sector
= mddev
->dev_sectors
;
3906 int still_degraded
= 0;
3909 if (sector_nr
>= max_sector
) {
3910 /* just being told to finish up .. nothing much to do */
3911 unplug_slaves(mddev
);
3913 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
3918 if (mddev
->curr_resync
< max_sector
) /* aborted */
3919 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
3921 else /* completed sync */
3923 bitmap_close_sync(mddev
->bitmap
);
3928 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
3929 return reshape_request(mddev
, sector_nr
, skipped
);
3931 /* No need to check resync_max as we never do more than one
3932 * stripe, and as resync_max will always be on a chunk boundary,
3933 * if the check in md_do_sync didn't fire, there is no chance
3934 * of overstepping resync_max here
3937 /* if there is too many failed drives and we are trying
3938 * to resync, then assert that we are finished, because there is
3939 * nothing we can do.
3941 if (mddev
->degraded
>= conf
->max_degraded
&&
3942 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
3943 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
3947 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
3948 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
3949 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
3950 /* we can skip this block, and probably more */
3951 sync_blocks
/= STRIPE_SECTORS
;
3953 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
3957 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
3959 sh
= get_active_stripe(conf
, sector_nr
, 0, 1);
3961 sh
= get_active_stripe(conf
, sector_nr
, 0, 0);
3962 /* make sure we don't swamp the stripe cache if someone else
3963 * is trying to get access
3965 schedule_timeout_uninterruptible(1);
3967 /* Need to check if array will still be degraded after recovery/resync
3968 * We don't need to check the 'failed' flag as when that gets set,
3971 for (i
=0; i
<mddev
->raid_disks
; i
++)
3972 if (conf
->disks
[i
].rdev
== NULL
)
3975 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
3977 spin_lock(&sh
->lock
);
3978 set_bit(STRIPE_SYNCING
, &sh
->state
);
3979 clear_bit(STRIPE_INSYNC
, &sh
->state
);
3980 spin_unlock(&sh
->lock
);
3982 /* wait for any blocked device to be handled */
3983 while(unlikely(!handle_stripe(sh
, NULL
)))
3987 return STRIPE_SECTORS
;
3990 static int retry_aligned_read(raid5_conf_t
*conf
, struct bio
*raid_bio
)
3992 /* We may not be able to submit a whole bio at once as there
3993 * may not be enough stripe_heads available.
3994 * We cannot pre-allocate enough stripe_heads as we may need
3995 * more than exist in the cache (if we allow ever large chunks).
3996 * So we do one stripe head at a time and record in
3997 * ->bi_hw_segments how many have been done.
3999 * We *know* that this entire raid_bio is in one chunk, so
4000 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4002 struct stripe_head
*sh
;
4004 sector_t sector
, logical_sector
, last_sector
;
4009 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4010 sector
= raid5_compute_sector(conf
, logical_sector
,
4012 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4014 for (; logical_sector
< last_sector
;
4015 logical_sector
+= STRIPE_SECTORS
,
4016 sector
+= STRIPE_SECTORS
,
4019 if (scnt
< raid5_bi_hw_segments(raid_bio
))
4020 /* already done this stripe */
4023 sh
= get_active_stripe(conf
, sector
, 0, 1);
4026 /* failed to get a stripe - must wait */
4027 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4028 conf
->retry_read_aligned
= raid_bio
;
4032 set_bit(R5_ReadError
, &sh
->dev
[dd_idx
].flags
);
4033 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4035 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4036 conf
->retry_read_aligned
= raid_bio
;
4040 handle_stripe(sh
, NULL
);
4044 spin_lock_irq(&conf
->device_lock
);
4045 remaining
= raid5_dec_bi_phys_segments(raid_bio
);
4046 spin_unlock_irq(&conf
->device_lock
);
4048 bio_endio(raid_bio
, 0);
4049 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4050 wake_up(&conf
->wait_for_stripe
);
4057 * This is our raid5 kernel thread.
4059 * We scan the hash table for stripes which can be handled now.
4060 * During the scan, completed stripes are saved for us by the interrupt
4061 * handler, so that they will not have to wait for our next wakeup.
4063 static void raid5d(mddev_t
*mddev
)
4065 struct stripe_head
*sh
;
4066 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4069 pr_debug("+++ raid5d active\n");
4071 md_check_recovery(mddev
);
4074 spin_lock_irq(&conf
->device_lock
);
4078 if (conf
->seq_flush
!= conf
->seq_write
) {
4079 int seq
= conf
->seq_flush
;
4080 spin_unlock_irq(&conf
->device_lock
);
4081 bitmap_unplug(mddev
->bitmap
);
4082 spin_lock_irq(&conf
->device_lock
);
4083 conf
->seq_write
= seq
;
4084 activate_bit_delay(conf
);
4087 while ((bio
= remove_bio_from_retry(conf
))) {
4089 spin_unlock_irq(&conf
->device_lock
);
4090 ok
= retry_aligned_read(conf
, bio
);
4091 spin_lock_irq(&conf
->device_lock
);
4097 sh
= __get_priority_stripe(conf
);
4101 spin_unlock_irq(&conf
->device_lock
);
4104 handle_stripe(sh
, conf
->spare_page
);
4107 spin_lock_irq(&conf
->device_lock
);
4109 pr_debug("%d stripes handled\n", handled
);
4111 spin_unlock_irq(&conf
->device_lock
);
4113 async_tx_issue_pending_all();
4114 unplug_slaves(mddev
);
4116 pr_debug("--- raid5d inactive\n");
4120 raid5_show_stripe_cache_size(mddev_t
*mddev
, char *page
)
4122 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4124 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4130 raid5_store_stripe_cache_size(mddev_t
*mddev
, const char *page
, size_t len
)
4132 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4136 if (len
>= PAGE_SIZE
)
4141 if (strict_strtoul(page
, 10, &new))
4143 if (new <= 16 || new > 32768)
4145 while (new < conf
->max_nr_stripes
) {
4146 if (drop_one_stripe(conf
))
4147 conf
->max_nr_stripes
--;
4151 err
= md_allow_write(mddev
);
4154 while (new > conf
->max_nr_stripes
) {
4155 if (grow_one_stripe(conf
))
4156 conf
->max_nr_stripes
++;
4162 static struct md_sysfs_entry
4163 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4164 raid5_show_stripe_cache_size
,
4165 raid5_store_stripe_cache_size
);
4168 raid5_show_preread_threshold(mddev_t
*mddev
, char *page
)
4170 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4172 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4178 raid5_store_preread_threshold(mddev_t
*mddev
, const char *page
, size_t len
)
4180 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4182 if (len
>= PAGE_SIZE
)
4187 if (strict_strtoul(page
, 10, &new))
4189 if (new > conf
->max_nr_stripes
)
4191 conf
->bypass_threshold
= new;
4195 static struct md_sysfs_entry
4196 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4198 raid5_show_preread_threshold
,
4199 raid5_store_preread_threshold
);
4202 stripe_cache_active_show(mddev_t
*mddev
, char *page
)
4204 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4206 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4211 static struct md_sysfs_entry
4212 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4214 static struct attribute
*raid5_attrs
[] = {
4215 &raid5_stripecache_size
.attr
,
4216 &raid5_stripecache_active
.attr
,
4217 &raid5_preread_bypass_threshold
.attr
,
4220 static struct attribute_group raid5_attrs_group
= {
4222 .attrs
= raid5_attrs
,
4226 raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
4228 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4231 sectors
= mddev
->dev_sectors
;
4233 /* size is defined by the smallest of previous and new size */
4234 if (conf
->raid_disks
< conf
->previous_raid_disks
)
4235 raid_disks
= conf
->raid_disks
;
4237 raid_disks
= conf
->previous_raid_disks
;
4240 sectors
&= ~((sector_t
)mddev
->chunk_size
/512 - 1);
4241 sectors
&= ~((sector_t
)mddev
->new_chunk
/512 - 1);
4242 return sectors
* (raid_disks
- conf
->max_degraded
);
4245 static raid5_conf_t
*setup_conf(mddev_t
*mddev
)
4248 int raid_disk
, memory
;
4250 struct disk_info
*disk
;
4252 if (mddev
->new_level
!= 5
4253 && mddev
->new_level
!= 4
4254 && mddev
->new_level
!= 6) {
4255 printk(KERN_ERR
"raid5: %s: raid level not set to 4/5/6 (%d)\n",
4256 mdname(mddev
), mddev
->new_level
);
4257 return ERR_PTR(-EIO
);
4259 if ((mddev
->new_level
== 5
4260 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
4261 (mddev
->new_level
== 6
4262 && !algorithm_valid_raid6(mddev
->new_layout
))) {
4263 printk(KERN_ERR
"raid5: %s: layout %d not supported\n",
4264 mdname(mddev
), mddev
->new_layout
);
4265 return ERR_PTR(-EIO
);
4267 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
4268 printk(KERN_ERR
"raid6: not enough configured devices for %s (%d, minimum 4)\n",
4269 mdname(mddev
), mddev
->raid_disks
);
4270 return ERR_PTR(-EINVAL
);
4273 if (!mddev
->new_chunk
|| mddev
->new_chunk
% PAGE_SIZE
) {
4274 printk(KERN_ERR
"raid5: invalid chunk size %d for %s\n",
4275 mddev
->new_chunk
, mdname(mddev
));
4276 return ERR_PTR(-EINVAL
);
4279 conf
= kzalloc(sizeof(raid5_conf_t
), GFP_KERNEL
);
4283 conf
->raid_disks
= mddev
->raid_disks
;
4284 if (mddev
->reshape_position
== MaxSector
)
4285 conf
->previous_raid_disks
= mddev
->raid_disks
;
4287 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4289 conf
->disks
= kzalloc(conf
->raid_disks
* sizeof(struct disk_info
),
4294 conf
->mddev
= mddev
;
4296 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
4299 if (mddev
->new_level
== 6) {
4300 conf
->spare_page
= alloc_page(GFP_KERNEL
);
4301 if (!conf
->spare_page
)
4304 spin_lock_init(&conf
->device_lock
);
4305 init_waitqueue_head(&conf
->wait_for_stripe
);
4306 init_waitqueue_head(&conf
->wait_for_overlap
);
4307 INIT_LIST_HEAD(&conf
->handle_list
);
4308 INIT_LIST_HEAD(&conf
->hold_list
);
4309 INIT_LIST_HEAD(&conf
->delayed_list
);
4310 INIT_LIST_HEAD(&conf
->bitmap_list
);
4311 INIT_LIST_HEAD(&conf
->inactive_list
);
4312 atomic_set(&conf
->active_stripes
, 0);
4313 atomic_set(&conf
->preread_active_stripes
, 0);
4314 atomic_set(&conf
->active_aligned_reads
, 0);
4315 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
4317 pr_debug("raid5: run(%s) called.\n", mdname(mddev
));
4319 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
4320 raid_disk
= rdev
->raid_disk
;
4321 if (raid_disk
>= conf
->raid_disks
4324 disk
= conf
->disks
+ raid_disk
;
4328 if (test_bit(In_sync
, &rdev
->flags
)) {
4329 char b
[BDEVNAME_SIZE
];
4330 printk(KERN_INFO
"raid5: device %s operational as raid"
4331 " disk %d\n", bdevname(rdev
->bdev
,b
),
4334 /* Cannot rely on bitmap to complete recovery */
4338 conf
->chunk_size
= mddev
->new_chunk
;
4339 conf
->level
= mddev
->new_level
;
4340 if (conf
->level
== 6)
4341 conf
->max_degraded
= 2;
4343 conf
->max_degraded
= 1;
4344 conf
->algorithm
= mddev
->new_layout
;
4345 conf
->max_nr_stripes
= NR_STRIPES
;
4346 conf
->reshape_progress
= mddev
->reshape_position
;
4347 if (conf
->reshape_progress
!= MaxSector
) {
4348 conf
->prev_chunk
= mddev
->chunk_size
;
4349 conf
->prev_algo
= mddev
->layout
;
4352 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
4353 conf
->raid_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
4354 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
4356 "raid5: couldn't allocate %dkB for buffers\n", memory
);
4359 printk(KERN_INFO
"raid5: allocated %dkB for %s\n",
4360 memory
, mdname(mddev
));
4362 conf
->thread
= md_register_thread(raid5d
, mddev
, "%s_raid5");
4363 if (!conf
->thread
) {
4365 "raid5: couldn't allocate thread for %s\n",
4374 shrink_stripes(conf
);
4375 safe_put_page(conf
->spare_page
);
4377 kfree(conf
->stripe_hashtbl
);
4379 return ERR_PTR(-EIO
);
4381 return ERR_PTR(-ENOMEM
);
4384 static int run(mddev_t
*mddev
)
4387 int working_disks
= 0;
4390 if (mddev
->reshape_position
!= MaxSector
) {
4391 /* Check that we can continue the reshape.
4392 * Currently only disks can change, it must
4393 * increase, and we must be past the point where
4394 * a stripe over-writes itself
4396 sector_t here_new
, here_old
;
4398 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
4400 if (mddev
->new_level
!= mddev
->level
) {
4401 printk(KERN_ERR
"raid5: %s: unsupported reshape "
4402 "required - aborting.\n",
4406 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4407 /* reshape_position must be on a new-stripe boundary, and one
4408 * further up in new geometry must map after here in old
4411 here_new
= mddev
->reshape_position
;
4412 if (sector_div(here_new
, (mddev
->new_chunk
>>9)*
4413 (mddev
->raid_disks
- max_degraded
))) {
4414 printk(KERN_ERR
"raid5: reshape_position not "
4415 "on a stripe boundary\n");
4418 /* here_new is the stripe we will write to */
4419 here_old
= mddev
->reshape_position
;
4420 sector_div(here_old
, (mddev
->chunk_size
>>9)*
4421 (old_disks
-max_degraded
));
4422 /* here_old is the first stripe that we might need to read
4424 if (here_new
>= here_old
) {
4425 /* Reading from the same stripe as writing to - bad */
4426 printk(KERN_ERR
"raid5: reshape_position too early for "
4427 "auto-recovery - aborting.\n");
4430 printk(KERN_INFO
"raid5: reshape will continue\n");
4431 /* OK, we should be able to continue; */
4433 BUG_ON(mddev
->level
!= mddev
->new_level
);
4434 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
4435 BUG_ON(mddev
->chunk_size
!= mddev
->new_chunk
);
4436 BUG_ON(mddev
->delta_disks
!= 0);
4439 if (mddev
->private == NULL
)
4440 conf
= setup_conf(mddev
);
4442 conf
= mddev
->private;
4445 return PTR_ERR(conf
);
4447 mddev
->thread
= conf
->thread
;
4448 conf
->thread
= NULL
;
4449 mddev
->private = conf
;
4452 * 0 for a fully functional array, 1 or 2 for a degraded array.
4454 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
4455 if (rdev
->raid_disk
>= 0 &&
4456 test_bit(In_sync
, &rdev
->flags
))
4459 mddev
->degraded
= conf
->raid_disks
- working_disks
;
4461 if (mddev
->degraded
> conf
->max_degraded
) {
4462 printk(KERN_ERR
"raid5: not enough operational devices for %s"
4463 " (%d/%d failed)\n",
4464 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
4468 /* device size must be a multiple of chunk size */
4469 mddev
->dev_sectors
&= ~(mddev
->chunk_size
/ 512 - 1);
4470 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
4472 if (mddev
->degraded
> 0 &&
4473 mddev
->recovery_cp
!= MaxSector
) {
4474 if (mddev
->ok_start_degraded
)
4476 "raid5: starting dirty degraded array: %s"
4477 "- data corruption possible.\n",
4481 "raid5: cannot start dirty degraded array for %s\n",
4487 if (mddev
->degraded
== 0)
4488 printk("raid5: raid level %d set %s active with %d out of %d"
4489 " devices, algorithm %d\n", conf
->level
, mdname(mddev
),
4490 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
4493 printk(KERN_ALERT
"raid5: raid level %d set %s active with %d"
4494 " out of %d devices, algorithm %d\n", conf
->level
,
4495 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
4496 mddev
->raid_disks
, mddev
->new_layout
);
4498 print_raid5_conf(conf
);
4500 if (conf
->reshape_progress
!= MaxSector
) {
4501 printk("...ok start reshape thread\n");
4502 conf
->reshape_safe
= conf
->reshape_progress
;
4503 atomic_set(&conf
->reshape_stripes
, 0);
4504 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4505 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4506 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4507 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4508 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4512 /* read-ahead size must cover two whole stripes, which is
4513 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4516 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
4517 int stripe
= data_disks
*
4518 (mddev
->chunk_size
/ PAGE_SIZE
);
4519 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4520 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4523 /* Ok, everything is just fine now */
4524 if (sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
4526 "raid5: failed to create sysfs attributes for %s\n",
4529 mddev
->queue
->queue_lock
= &conf
->device_lock
;
4531 mddev
->queue
->unplug_fn
= raid5_unplug_device
;
4532 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
4533 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
4535 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
4537 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
4541 md_unregister_thread(mddev
->thread
);
4542 mddev
->thread
= NULL
;
4544 shrink_stripes(conf
);
4545 print_raid5_conf(conf
);
4546 safe_put_page(conf
->spare_page
);
4548 kfree(conf
->stripe_hashtbl
);
4551 mddev
->private = NULL
;
4552 printk(KERN_ALERT
"raid5: failed to run raid set %s\n", mdname(mddev
));
4558 static int stop(mddev_t
*mddev
)
4560 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
4562 md_unregister_thread(mddev
->thread
);
4563 mddev
->thread
= NULL
;
4564 shrink_stripes(conf
);
4565 kfree(conf
->stripe_hashtbl
);
4566 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
4567 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
4568 sysfs_remove_group(&mddev
->kobj
, &raid5_attrs_group
);
4571 mddev
->private = NULL
;
4576 static void print_sh(struct seq_file
*seq
, struct stripe_head
*sh
)
4580 seq_printf(seq
, "sh %llu, pd_idx %d, state %ld.\n",
4581 (unsigned long long)sh
->sector
, sh
->pd_idx
, sh
->state
);
4582 seq_printf(seq
, "sh %llu, count %d.\n",
4583 (unsigned long long)sh
->sector
, atomic_read(&sh
->count
));
4584 seq_printf(seq
, "sh %llu, ", (unsigned long long)sh
->sector
);
4585 for (i
= 0; i
< sh
->disks
; i
++) {
4586 seq_printf(seq
, "(cache%d: %p %ld) ",
4587 i
, sh
->dev
[i
].page
, sh
->dev
[i
].flags
);
4589 seq_printf(seq
, "\n");
4592 static void printall(struct seq_file
*seq
, raid5_conf_t
*conf
)
4594 struct stripe_head
*sh
;
4595 struct hlist_node
*hn
;
4598 spin_lock_irq(&conf
->device_lock
);
4599 for (i
= 0; i
< NR_HASH
; i
++) {
4600 hlist_for_each_entry(sh
, hn
, &conf
->stripe_hashtbl
[i
], hash
) {
4601 if (sh
->raid_conf
!= conf
)
4606 spin_unlock_irq(&conf
->device_lock
);
4610 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
4612 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
4615 seq_printf (seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
, mddev
->chunk_size
>> 10, mddev
->layout
);
4616 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
4617 for (i
= 0; i
< conf
->raid_disks
; i
++)
4618 seq_printf (seq
, "%s",
4619 conf
->disks
[i
].rdev
&&
4620 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
4621 seq_printf (seq
, "]");
4623 seq_printf (seq
, "\n");
4624 printall(seq
, conf
);
4628 static void print_raid5_conf (raid5_conf_t
*conf
)
4631 struct disk_info
*tmp
;
4633 printk("RAID5 conf printout:\n");
4635 printk("(conf==NULL)\n");
4638 printk(" --- rd:%d wd:%d\n", conf
->raid_disks
,
4639 conf
->raid_disks
- conf
->mddev
->degraded
);
4641 for (i
= 0; i
< conf
->raid_disks
; i
++) {
4642 char b
[BDEVNAME_SIZE
];
4643 tmp
= conf
->disks
+ i
;
4645 printk(" disk %d, o:%d, dev:%s\n",
4646 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
4647 bdevname(tmp
->rdev
->bdev
,b
));
4651 static int raid5_spare_active(mddev_t
*mddev
)
4654 raid5_conf_t
*conf
= mddev
->private;
4655 struct disk_info
*tmp
;
4657 for (i
= 0; i
< conf
->raid_disks
; i
++) {
4658 tmp
= conf
->disks
+ i
;
4660 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
4661 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
4662 unsigned long flags
;
4663 spin_lock_irqsave(&conf
->device_lock
, flags
);
4665 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4668 print_raid5_conf(conf
);
4672 static int raid5_remove_disk(mddev_t
*mddev
, int number
)
4674 raid5_conf_t
*conf
= mddev
->private;
4677 struct disk_info
*p
= conf
->disks
+ number
;
4679 print_raid5_conf(conf
);
4682 if (number
>= conf
->raid_disks
&&
4683 conf
->reshape_progress
== MaxSector
)
4684 clear_bit(In_sync
, &rdev
->flags
);
4686 if (test_bit(In_sync
, &rdev
->flags
) ||
4687 atomic_read(&rdev
->nr_pending
)) {
4691 /* Only remove non-faulty devices if recovery
4694 if (!test_bit(Faulty
, &rdev
->flags
) &&
4695 mddev
->degraded
<= conf
->max_degraded
&&
4696 number
< conf
->raid_disks
) {
4702 if (atomic_read(&rdev
->nr_pending
)) {
4703 /* lost the race, try later */
4710 print_raid5_conf(conf
);
4714 static int raid5_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
4716 raid5_conf_t
*conf
= mddev
->private;
4719 struct disk_info
*p
;
4721 int last
= conf
->raid_disks
- 1;
4723 if (mddev
->degraded
> conf
->max_degraded
)
4724 /* no point adding a device */
4727 if (rdev
->raid_disk
>= 0)
4728 first
= last
= rdev
->raid_disk
;
4731 * find the disk ... but prefer rdev->saved_raid_disk
4734 if (rdev
->saved_raid_disk
>= 0 &&
4735 rdev
->saved_raid_disk
>= first
&&
4736 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
4737 disk
= rdev
->saved_raid_disk
;
4740 for ( ; disk
<= last
; disk
++)
4741 if ((p
=conf
->disks
+ disk
)->rdev
== NULL
) {
4742 clear_bit(In_sync
, &rdev
->flags
);
4743 rdev
->raid_disk
= disk
;
4745 if (rdev
->saved_raid_disk
!= disk
)
4747 rcu_assign_pointer(p
->rdev
, rdev
);
4750 print_raid5_conf(conf
);
4754 static int raid5_resize(mddev_t
*mddev
, sector_t sectors
)
4756 /* no resync is happening, and there is enough space
4757 * on all devices, so we can resize.
4758 * We need to make sure resync covers any new space.
4759 * If the array is shrinking we should possibly wait until
4760 * any io in the removed space completes, but it hardly seems
4763 sectors
&= ~((sector_t
)mddev
->chunk_size
/512 - 1);
4764 md_set_array_sectors(mddev
, raid5_size(mddev
, sectors
,
4765 mddev
->raid_disks
));
4766 if (mddev
->array_sectors
>
4767 raid5_size(mddev
, sectors
, mddev
->raid_disks
))
4769 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4771 if (sectors
> mddev
->dev_sectors
&& mddev
->recovery_cp
== MaxSector
) {
4772 mddev
->recovery_cp
= mddev
->dev_sectors
;
4773 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4775 mddev
->dev_sectors
= sectors
;
4776 mddev
->resync_max_sectors
= sectors
;
4780 #ifdef CONFIG_MD_RAID5_RESHAPE
4781 static int raid5_check_reshape(mddev_t
*mddev
)
4783 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4785 if (mddev
->delta_disks
== 0 &&
4786 mddev
->new_layout
== mddev
->layout
&&
4787 mddev
->new_chunk
== mddev
->chunk_size
)
4788 return -EINVAL
; /* nothing to do */
4790 /* Cannot grow a bitmap yet */
4792 if (mddev
->degraded
> conf
->max_degraded
)
4794 if (mddev
->delta_disks
< 0) {
4795 /* We might be able to shrink, but the devices must
4796 * be made bigger first.
4797 * For raid6, 4 is the minimum size.
4798 * Otherwise 2 is the minimum
4801 if (mddev
->level
== 6)
4803 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
4807 /* Can only proceed if there are plenty of stripe_heads.
4808 * We need a minimum of one full stripe,, and for sensible progress
4809 * it is best to have about 4 times that.
4810 * If we require 4 times, then the default 256 4K stripe_heads will
4811 * allow for chunk sizes up to 256K, which is probably OK.
4812 * If the chunk size is greater, user-space should request more
4813 * stripe_heads first.
4815 if ((mddev
->chunk_size
/ STRIPE_SIZE
) * 4 > conf
->max_nr_stripes
||
4816 (mddev
->new_chunk
/ STRIPE_SIZE
) * 4 > conf
->max_nr_stripes
) {
4817 printk(KERN_WARNING
"raid5: reshape: not enough stripes. Needed %lu\n",
4818 (max(mddev
->chunk_size
, mddev
->new_chunk
)
4823 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
4826 static int raid5_start_reshape(mddev_t
*mddev
)
4828 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4831 int added_devices
= 0;
4832 unsigned long flags
;
4834 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4837 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
4838 if (rdev
->raid_disk
< 0 &&
4839 !test_bit(Faulty
, &rdev
->flags
))
4842 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
4843 /* Not enough devices even to make a degraded array
4848 /* Refuse to reduce size of the array. Any reductions in
4849 * array size must be through explicit setting of array_size
4852 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
4853 < mddev
->array_sectors
) {
4854 printk(KERN_ERR
"md: %s: array size must be reduced "
4855 "before number of disks\n", mdname(mddev
));
4859 atomic_set(&conf
->reshape_stripes
, 0);
4860 spin_lock_irq(&conf
->device_lock
);
4861 conf
->previous_raid_disks
= conf
->raid_disks
;
4862 conf
->raid_disks
+= mddev
->delta_disks
;
4863 conf
->prev_chunk
= conf
->chunk_size
;
4864 conf
->chunk_size
= mddev
->new_chunk
;
4865 conf
->prev_algo
= conf
->algorithm
;
4866 conf
->algorithm
= mddev
->new_layout
;
4867 if (mddev
->delta_disks
< 0)
4868 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
4870 conf
->reshape_progress
= 0;
4871 conf
->reshape_safe
= conf
->reshape_progress
;
4873 spin_unlock_irq(&conf
->device_lock
);
4875 /* Add some new drives, as many as will fit.
4876 * We know there are enough to make the newly sized array work.
4878 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
4879 if (rdev
->raid_disk
< 0 &&
4880 !test_bit(Faulty
, &rdev
->flags
)) {
4881 if (raid5_add_disk(mddev
, rdev
) == 0) {
4883 set_bit(In_sync
, &rdev
->flags
);
4885 rdev
->recovery_offset
= 0;
4886 sprintf(nm
, "rd%d", rdev
->raid_disk
);
4887 if (sysfs_create_link(&mddev
->kobj
,
4890 "raid5: failed to create "
4891 " link %s for %s\n",
4897 if (mddev
->delta_disks
> 0) {
4898 spin_lock_irqsave(&conf
->device_lock
, flags
);
4899 mddev
->degraded
= (conf
->raid_disks
- conf
->previous_raid_disks
)
4901 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4903 mddev
->raid_disks
= conf
->raid_disks
;
4904 mddev
->reshape_position
= 0;
4905 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4907 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4908 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4909 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4910 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4911 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4913 if (!mddev
->sync_thread
) {
4914 mddev
->recovery
= 0;
4915 spin_lock_irq(&conf
->device_lock
);
4916 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
4917 conf
->reshape_progress
= MaxSector
;
4918 spin_unlock_irq(&conf
->device_lock
);
4921 md_wakeup_thread(mddev
->sync_thread
);
4922 md_new_event(mddev
);
4927 /* This is called from the reshape thread and should make any
4928 * changes needed in 'conf'
4930 static void end_reshape(raid5_conf_t
*conf
)
4933 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
4935 spin_lock_irq(&conf
->device_lock
);
4936 conf
->previous_raid_disks
= conf
->raid_disks
;
4937 conf
->reshape_progress
= MaxSector
;
4938 spin_unlock_irq(&conf
->device_lock
);
4940 /* read-ahead size must cover two whole stripes, which is
4941 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4944 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
4945 int stripe
= data_disks
* (conf
->chunk_size
4947 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4948 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4953 /* This is called from the raid5d thread with mddev_lock held.
4954 * It makes config changes to the device.
4956 static void raid5_finish_reshape(mddev_t
*mddev
)
4958 struct block_device
*bdev
;
4959 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4961 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
4963 if (mddev
->delta_disks
> 0) {
4964 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
4965 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4968 bdev
= bdget_disk(mddev
->gendisk
, 0);
4970 mutex_lock(&bdev
->bd_inode
->i_mutex
);
4971 i_size_write(bdev
->bd_inode
,
4972 (loff_t
)mddev
->array_sectors
<< 9);
4973 mutex_unlock(&bdev
->bd_inode
->i_mutex
);
4978 mddev
->degraded
= conf
->raid_disks
;
4979 for (d
= 0; d
< conf
->raid_disks
; d
++)
4980 if (conf
->disks
[d
].rdev
&&
4982 &conf
->disks
[d
].rdev
->flags
))
4984 for (d
= conf
->raid_disks
;
4985 d
< conf
->raid_disks
- mddev
->delta_disks
;
4987 raid5_remove_disk(mddev
, d
);
4989 mddev
->layout
= conf
->algorithm
;
4990 mddev
->chunk_size
= conf
->chunk_size
;
4991 mddev
->reshape_position
= MaxSector
;
4992 mddev
->delta_disks
= 0;
4996 static void raid5_quiesce(mddev_t
*mddev
, int state
)
4998 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
5001 case 2: /* resume for a suspend */
5002 wake_up(&conf
->wait_for_overlap
);
5005 case 1: /* stop all writes */
5006 spin_lock_irq(&conf
->device_lock
);
5008 wait_event_lock_irq(conf
->wait_for_stripe
,
5009 atomic_read(&conf
->active_stripes
) == 0 &&
5010 atomic_read(&conf
->active_aligned_reads
) == 0,
5011 conf
->device_lock
, /* nothing */);
5012 spin_unlock_irq(&conf
->device_lock
);
5015 case 0: /* re-enable writes */
5016 spin_lock_irq(&conf
->device_lock
);
5018 wake_up(&conf
->wait_for_stripe
);
5019 wake_up(&conf
->wait_for_overlap
);
5020 spin_unlock_irq(&conf
->device_lock
);
5026 static void *raid5_takeover_raid1(mddev_t
*mddev
)
5030 if (mddev
->raid_disks
!= 2 ||
5031 mddev
->degraded
> 1)
5032 return ERR_PTR(-EINVAL
);
5034 /* Should check if there are write-behind devices? */
5036 chunksect
= 64*2; /* 64K by default */
5038 /* The array must be an exact multiple of chunksize */
5039 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
5042 if ((chunksect
<<9) < STRIPE_SIZE
)
5043 /* array size does not allow a suitable chunk size */
5044 return ERR_PTR(-EINVAL
);
5046 mddev
->new_level
= 5;
5047 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5048 mddev
->new_chunk
= chunksect
<< 9;
5050 return setup_conf(mddev
);
5053 static void *raid5_takeover_raid6(mddev_t
*mddev
)
5057 switch (mddev
->layout
) {
5058 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5059 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
5061 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5062 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
5064 case ALGORITHM_LEFT_SYMMETRIC_6
:
5065 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5067 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5068 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
5070 case ALGORITHM_PARITY_0_6
:
5071 new_layout
= ALGORITHM_PARITY_0
;
5073 case ALGORITHM_PARITY_N
:
5074 new_layout
= ALGORITHM_PARITY_N
;
5077 return ERR_PTR(-EINVAL
);
5079 mddev
->new_level
= 5;
5080 mddev
->new_layout
= new_layout
;
5081 mddev
->delta_disks
= -1;
5082 mddev
->raid_disks
-= 1;
5083 return setup_conf(mddev
);
5087 static int raid5_reconfig(mddev_t
*mddev
, int new_layout
, int new_chunk
)
5089 /* For a 2-drive array, the layout and chunk size can be changed
5090 * immediately as not restriping is needed.
5091 * For larger arrays we record the new value - after validation
5092 * to be used by a reshape pass.
5094 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
5096 if (new_layout
>= 0 && !algorithm_valid_raid5(new_layout
))
5098 if (new_chunk
> 0) {
5099 if (new_chunk
& (new_chunk
-1))
5100 /* not a power of 2 */
5102 if (new_chunk
< PAGE_SIZE
)
5104 if (mddev
->array_sectors
& ((new_chunk
>>9)-1))
5105 /* not factor of array size */
5109 /* They look valid */
5111 if (mddev
->raid_disks
== 2) {
5113 if (new_layout
>= 0) {
5114 conf
->algorithm
= new_layout
;
5115 mddev
->layout
= mddev
->new_layout
= new_layout
;
5117 if (new_chunk
> 0) {
5118 conf
->chunk_size
= new_chunk
;
5119 mddev
->chunk_size
= mddev
->new_chunk
= new_chunk
;
5121 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5122 md_wakeup_thread(mddev
->thread
);
5124 if (new_layout
>= 0)
5125 mddev
->new_layout
= new_layout
;
5127 mddev
->new_chunk
= new_chunk
;
5132 static int raid6_reconfig(mddev_t
*mddev
, int new_layout
, int new_chunk
)
5134 if (new_layout
>= 0 && !algorithm_valid_raid6(new_layout
))
5136 if (new_chunk
> 0) {
5137 if (new_chunk
& (new_chunk
-1))
5138 /* not a power of 2 */
5140 if (new_chunk
< PAGE_SIZE
)
5142 if (mddev
->array_sectors
& ((new_chunk
>>9)-1))
5143 /* not factor of array size */
5147 /* They look valid */
5149 if (new_layout
>= 0)
5150 mddev
->new_layout
= new_layout
;
5152 mddev
->new_chunk
= new_chunk
;
5157 static void *raid5_takeover(mddev_t
*mddev
)
5159 /* raid5 can take over:
5160 * raid0 - if all devices are the same - make it a raid4 layout
5161 * raid1 - if there are two drives. We need to know the chunk size
5162 * raid4 - trivial - just use a raid4 layout.
5163 * raid6 - Providing it is a *_6 layout
5165 * For now, just do raid1
5168 if (mddev
->level
== 1)
5169 return raid5_takeover_raid1(mddev
);
5170 if (mddev
->level
== 4) {
5171 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5172 mddev
->new_level
= 5;
5173 return setup_conf(mddev
);
5175 if (mddev
->level
== 6)
5176 return raid5_takeover_raid6(mddev
);
5178 return ERR_PTR(-EINVAL
);
5182 static struct mdk_personality raid5_personality
;
5184 static void *raid6_takeover(mddev_t
*mddev
)
5186 /* Currently can only take over a raid5. We map the
5187 * personality to an equivalent raid6 personality
5188 * with the Q block at the end.
5192 if (mddev
->pers
!= &raid5_personality
)
5193 return ERR_PTR(-EINVAL
);
5194 if (mddev
->degraded
> 1)
5195 return ERR_PTR(-EINVAL
);
5196 if (mddev
->raid_disks
> 253)
5197 return ERR_PTR(-EINVAL
);
5198 if (mddev
->raid_disks
< 3)
5199 return ERR_PTR(-EINVAL
);
5201 switch (mddev
->layout
) {
5202 case ALGORITHM_LEFT_ASYMMETRIC
:
5203 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
5205 case ALGORITHM_RIGHT_ASYMMETRIC
:
5206 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
5208 case ALGORITHM_LEFT_SYMMETRIC
:
5209 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
5211 case ALGORITHM_RIGHT_SYMMETRIC
:
5212 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
5214 case ALGORITHM_PARITY_0
:
5215 new_layout
= ALGORITHM_PARITY_0_6
;
5217 case ALGORITHM_PARITY_N
:
5218 new_layout
= ALGORITHM_PARITY_N
;
5221 return ERR_PTR(-EINVAL
);
5223 mddev
->new_level
= 6;
5224 mddev
->new_layout
= new_layout
;
5225 mddev
->delta_disks
= 1;
5226 mddev
->raid_disks
+= 1;
5227 return setup_conf(mddev
);
5231 static struct mdk_personality raid6_personality
=
5235 .owner
= THIS_MODULE
,
5236 .make_request
= make_request
,
5240 .error_handler
= error
,
5241 .hot_add_disk
= raid5_add_disk
,
5242 .hot_remove_disk
= raid5_remove_disk
,
5243 .spare_active
= raid5_spare_active
,
5244 .sync_request
= sync_request
,
5245 .resize
= raid5_resize
,
5247 #ifdef CONFIG_MD_RAID5_RESHAPE
5248 .check_reshape
= raid5_check_reshape
,
5249 .start_reshape
= raid5_start_reshape
,
5250 .finish_reshape
= raid5_finish_reshape
,
5252 .quiesce
= raid5_quiesce
,
5253 .takeover
= raid6_takeover
,
5254 .reconfig
= raid6_reconfig
,
5256 static struct mdk_personality raid5_personality
=
5260 .owner
= THIS_MODULE
,
5261 .make_request
= make_request
,
5265 .error_handler
= error
,
5266 .hot_add_disk
= raid5_add_disk
,
5267 .hot_remove_disk
= raid5_remove_disk
,
5268 .spare_active
= raid5_spare_active
,
5269 .sync_request
= sync_request
,
5270 .resize
= raid5_resize
,
5272 #ifdef CONFIG_MD_RAID5_RESHAPE
5273 .check_reshape
= raid5_check_reshape
,
5274 .start_reshape
= raid5_start_reshape
,
5275 .finish_reshape
= raid5_finish_reshape
,
5277 .quiesce
= raid5_quiesce
,
5278 .takeover
= raid5_takeover
,
5279 .reconfig
= raid5_reconfig
,
5282 static struct mdk_personality raid4_personality
=
5286 .owner
= THIS_MODULE
,
5287 .make_request
= make_request
,
5291 .error_handler
= error
,
5292 .hot_add_disk
= raid5_add_disk
,
5293 .hot_remove_disk
= raid5_remove_disk
,
5294 .spare_active
= raid5_spare_active
,
5295 .sync_request
= sync_request
,
5296 .resize
= raid5_resize
,
5298 #ifdef CONFIG_MD_RAID5_RESHAPE
5299 .check_reshape
= raid5_check_reshape
,
5300 .start_reshape
= raid5_start_reshape
,
5301 .finish_reshape
= raid5_finish_reshape
,
5303 .quiesce
= raid5_quiesce
,
5306 static int __init
raid5_init(void)
5308 register_md_personality(&raid6_personality
);
5309 register_md_personality(&raid5_personality
);
5310 register_md_personality(&raid4_personality
);
5314 static void raid5_exit(void)
5316 unregister_md_personality(&raid6_personality
);
5317 unregister_md_personality(&raid5_personality
);
5318 unregister_md_personality(&raid4_personality
);
5321 module_init(raid5_init
);
5322 module_exit(raid5_exit
);
5323 MODULE_LICENSE("GPL");
5324 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5325 MODULE_ALIAS("md-raid5");
5326 MODULE_ALIAS("md-raid4");
5327 MODULE_ALIAS("md-level-5");
5328 MODULE_ALIAS("md-level-4");
5329 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5330 MODULE_ALIAS("md-raid6");
5331 MODULE_ALIAS("md-level-6");
5333 /* This used to be two separate modules, they were: */
5334 MODULE_ALIAS("raid5");
5335 MODULE_ALIAS("raid6");