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
)
399 && !test_bit(STRIPE_EXPANDING
, &sh
->state
));
401 if (!test_bit(STRIPE_HANDLE
, &sh
->state
))
402 atomic_inc(&conf
->active_stripes
);
403 if (list_empty(&sh
->lru
) &&
404 !test_bit(STRIPE_EXPANDING
, &sh
->state
))
406 list_del_init(&sh
->lru
);
409 } while (sh
== NULL
);
412 atomic_inc(&sh
->count
);
414 spin_unlock_irq(&conf
->device_lock
);
419 raid5_end_read_request(struct bio
*bi
, int error
);
421 raid5_end_write_request(struct bio
*bi
, int error
);
423 static void ops_run_io(struct stripe_head
*sh
, struct stripe_head_state
*s
)
425 raid5_conf_t
*conf
= sh
->raid_conf
;
426 int i
, disks
= sh
->disks
;
430 for (i
= disks
; i
--; ) {
434 if (test_and_clear_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
))
436 else if (test_and_clear_bit(R5_Wantread
, &sh
->dev
[i
].flags
))
441 bi
= &sh
->dev
[i
].req
;
445 bi
->bi_end_io
= raid5_end_write_request
;
447 bi
->bi_end_io
= raid5_end_read_request
;
450 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
451 if (rdev
&& test_bit(Faulty
, &rdev
->flags
))
454 atomic_inc(&rdev
->nr_pending
);
458 if (s
->syncing
|| s
->expanding
|| s
->expanded
)
459 md_sync_acct(rdev
->bdev
, STRIPE_SECTORS
);
461 set_bit(STRIPE_IO_STARTED
, &sh
->state
);
463 bi
->bi_bdev
= rdev
->bdev
;
464 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
465 __func__
, (unsigned long long)sh
->sector
,
467 atomic_inc(&sh
->count
);
468 bi
->bi_sector
= sh
->sector
+ rdev
->data_offset
;
469 bi
->bi_flags
= 1 << BIO_UPTODATE
;
473 bi
->bi_io_vec
= &sh
->dev
[i
].vec
;
474 bi
->bi_io_vec
[0].bv_len
= STRIPE_SIZE
;
475 bi
->bi_io_vec
[0].bv_offset
= 0;
476 bi
->bi_size
= STRIPE_SIZE
;
479 test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
480 atomic_add(STRIPE_SECTORS
,
481 &rdev
->corrected_errors
);
482 generic_make_request(bi
);
485 set_bit(STRIPE_DEGRADED
, &sh
->state
);
486 pr_debug("skip op %ld on disc %d for sector %llu\n",
487 bi
->bi_rw
, i
, (unsigned long long)sh
->sector
);
488 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
489 set_bit(STRIPE_HANDLE
, &sh
->state
);
494 static struct dma_async_tx_descriptor
*
495 async_copy_data(int frombio
, struct bio
*bio
, struct page
*page
,
496 sector_t sector
, struct dma_async_tx_descriptor
*tx
)
499 struct page
*bio_page
;
503 if (bio
->bi_sector
>= sector
)
504 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
506 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
507 bio_for_each_segment(bvl
, bio
, i
) {
508 int len
= bio_iovec_idx(bio
, i
)->bv_len
;
512 if (page_offset
< 0) {
513 b_offset
= -page_offset
;
514 page_offset
+= b_offset
;
518 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
519 clen
= STRIPE_SIZE
- page_offset
;
524 b_offset
+= bio_iovec_idx(bio
, i
)->bv_offset
;
525 bio_page
= bio_iovec_idx(bio
, i
)->bv_page
;
527 tx
= async_memcpy(page
, bio_page
, page_offset
,
531 tx
= async_memcpy(bio_page
, page
, b_offset
,
532 page_offset
, clen
, 0,
535 if (clen
< len
) /* hit end of page */
543 static void ops_complete_biofill(void *stripe_head_ref
)
545 struct stripe_head
*sh
= stripe_head_ref
;
546 struct bio
*return_bi
= NULL
;
547 raid5_conf_t
*conf
= sh
->raid_conf
;
550 pr_debug("%s: stripe %llu\n", __func__
,
551 (unsigned long long)sh
->sector
);
553 /* clear completed biofills */
554 spin_lock_irq(&conf
->device_lock
);
555 for (i
= sh
->disks
; i
--; ) {
556 struct r5dev
*dev
= &sh
->dev
[i
];
558 /* acknowledge completion of a biofill operation */
559 /* and check if we need to reply to a read request,
560 * new R5_Wantfill requests are held off until
561 * !STRIPE_BIOFILL_RUN
563 if (test_and_clear_bit(R5_Wantfill
, &dev
->flags
)) {
564 struct bio
*rbi
, *rbi2
;
569 while (rbi
&& rbi
->bi_sector
<
570 dev
->sector
+ STRIPE_SECTORS
) {
571 rbi2
= r5_next_bio(rbi
, dev
->sector
);
572 if (!raid5_dec_bi_phys_segments(rbi
)) {
573 rbi
->bi_next
= return_bi
;
580 spin_unlock_irq(&conf
->device_lock
);
581 clear_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
583 return_io(return_bi
);
585 set_bit(STRIPE_HANDLE
, &sh
->state
);
589 static void ops_run_biofill(struct stripe_head
*sh
)
591 struct dma_async_tx_descriptor
*tx
= NULL
;
592 raid5_conf_t
*conf
= sh
->raid_conf
;
595 pr_debug("%s: stripe %llu\n", __func__
,
596 (unsigned long long)sh
->sector
);
598 for (i
= sh
->disks
; i
--; ) {
599 struct r5dev
*dev
= &sh
->dev
[i
];
600 if (test_bit(R5_Wantfill
, &dev
->flags
)) {
602 spin_lock_irq(&conf
->device_lock
);
603 dev
->read
= rbi
= dev
->toread
;
605 spin_unlock_irq(&conf
->device_lock
);
606 while (rbi
&& rbi
->bi_sector
<
607 dev
->sector
+ STRIPE_SECTORS
) {
608 tx
= async_copy_data(0, rbi
, dev
->page
,
610 rbi
= r5_next_bio(rbi
, dev
->sector
);
615 atomic_inc(&sh
->count
);
616 async_trigger_callback(ASYNC_TX_ACK
, tx
, ops_complete_biofill
, sh
);
619 static void ops_complete_compute5(void *stripe_head_ref
)
621 struct stripe_head
*sh
= stripe_head_ref
;
622 int target
= sh
->ops
.target
;
623 struct r5dev
*tgt
= &sh
->dev
[target
];
625 pr_debug("%s: stripe %llu\n", __func__
,
626 (unsigned long long)sh
->sector
);
628 set_bit(R5_UPTODATE
, &tgt
->flags
);
629 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
630 clear_bit(R5_Wantcompute
, &tgt
->flags
);
631 clear_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
632 if (sh
->check_state
== check_state_compute_run
)
633 sh
->check_state
= check_state_compute_result
;
634 set_bit(STRIPE_HANDLE
, &sh
->state
);
638 static struct dma_async_tx_descriptor
*ops_run_compute5(struct stripe_head
*sh
)
640 /* kernel stack size limits the total number of disks */
641 int disks
= sh
->disks
;
642 struct page
*xor_srcs
[disks
];
643 int target
= sh
->ops
.target
;
644 struct r5dev
*tgt
= &sh
->dev
[target
];
645 struct page
*xor_dest
= tgt
->page
;
647 struct dma_async_tx_descriptor
*tx
;
650 pr_debug("%s: stripe %llu block: %d\n",
651 __func__
, (unsigned long long)sh
->sector
, target
);
652 BUG_ON(!test_bit(R5_Wantcompute
, &tgt
->flags
));
654 for (i
= disks
; i
--; )
656 xor_srcs
[count
++] = sh
->dev
[i
].page
;
658 atomic_inc(&sh
->count
);
660 if (unlikely(count
== 1))
661 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
,
662 0, NULL
, ops_complete_compute5
, sh
);
664 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
665 ASYNC_TX_XOR_ZERO_DST
, NULL
,
666 ops_complete_compute5
, sh
);
671 static void ops_complete_prexor(void *stripe_head_ref
)
673 struct stripe_head
*sh
= stripe_head_ref
;
675 pr_debug("%s: stripe %llu\n", __func__
,
676 (unsigned long long)sh
->sector
);
679 static struct dma_async_tx_descriptor
*
680 ops_run_prexor(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
682 /* kernel stack size limits the total number of disks */
683 int disks
= sh
->disks
;
684 struct page
*xor_srcs
[disks
];
685 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
687 /* existing parity data subtracted */
688 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
690 pr_debug("%s: stripe %llu\n", __func__
,
691 (unsigned long long)sh
->sector
);
693 for (i
= disks
; i
--; ) {
694 struct r5dev
*dev
= &sh
->dev
[i
];
695 /* Only process blocks that are known to be uptodate */
696 if (test_bit(R5_Wantdrain
, &dev
->flags
))
697 xor_srcs
[count
++] = dev
->page
;
700 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
701 ASYNC_TX_XOR_DROP_DST
, tx
,
702 ops_complete_prexor
, sh
);
707 static struct dma_async_tx_descriptor
*
708 ops_run_biodrain(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
710 int disks
= sh
->disks
;
713 pr_debug("%s: stripe %llu\n", __func__
,
714 (unsigned long long)sh
->sector
);
716 for (i
= disks
; i
--; ) {
717 struct r5dev
*dev
= &sh
->dev
[i
];
720 if (test_and_clear_bit(R5_Wantdrain
, &dev
->flags
)) {
723 spin_lock(&sh
->lock
);
724 chosen
= dev
->towrite
;
726 BUG_ON(dev
->written
);
727 wbi
= dev
->written
= chosen
;
728 spin_unlock(&sh
->lock
);
730 while (wbi
&& wbi
->bi_sector
<
731 dev
->sector
+ STRIPE_SECTORS
) {
732 tx
= async_copy_data(1, wbi
, dev
->page
,
734 wbi
= r5_next_bio(wbi
, dev
->sector
);
742 static void ops_complete_postxor(void *stripe_head_ref
)
744 struct stripe_head
*sh
= stripe_head_ref
;
745 int disks
= sh
->disks
, i
, pd_idx
= sh
->pd_idx
;
747 pr_debug("%s: stripe %llu\n", __func__
,
748 (unsigned long long)sh
->sector
);
750 for (i
= disks
; i
--; ) {
751 struct r5dev
*dev
= &sh
->dev
[i
];
752 if (dev
->written
|| i
== pd_idx
)
753 set_bit(R5_UPTODATE
, &dev
->flags
);
756 if (sh
->reconstruct_state
== reconstruct_state_drain_run
)
757 sh
->reconstruct_state
= reconstruct_state_drain_result
;
758 else if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
)
759 sh
->reconstruct_state
= reconstruct_state_prexor_drain_result
;
761 BUG_ON(sh
->reconstruct_state
!= reconstruct_state_run
);
762 sh
->reconstruct_state
= reconstruct_state_result
;
765 set_bit(STRIPE_HANDLE
, &sh
->state
);
770 ops_run_postxor(struct stripe_head
*sh
, struct dma_async_tx_descriptor
*tx
)
772 /* kernel stack size limits the total number of disks */
773 int disks
= sh
->disks
;
774 struct page
*xor_srcs
[disks
];
776 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
777 struct page
*xor_dest
;
781 pr_debug("%s: stripe %llu\n", __func__
,
782 (unsigned long long)sh
->sector
);
784 /* check if prexor is active which means only process blocks
785 * that are part of a read-modify-write (written)
787 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_run
) {
789 xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
790 for (i
= disks
; i
--; ) {
791 struct r5dev
*dev
= &sh
->dev
[i
];
793 xor_srcs
[count
++] = dev
->page
;
796 xor_dest
= sh
->dev
[pd_idx
].page
;
797 for (i
= disks
; i
--; ) {
798 struct r5dev
*dev
= &sh
->dev
[i
];
800 xor_srcs
[count
++] = dev
->page
;
804 /* 1/ if we prexor'd then the dest is reused as a source
805 * 2/ if we did not prexor then we are redoing the parity
806 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
807 * for the synchronous xor case
809 flags
= ASYNC_TX_ACK
|
810 (prexor
? ASYNC_TX_XOR_DROP_DST
: ASYNC_TX_XOR_ZERO_DST
);
812 atomic_inc(&sh
->count
);
814 if (unlikely(count
== 1)) {
815 flags
&= ~(ASYNC_TX_XOR_DROP_DST
| ASYNC_TX_XOR_ZERO_DST
);
816 tx
= async_memcpy(xor_dest
, xor_srcs
[0], 0, 0, STRIPE_SIZE
,
817 flags
, tx
, ops_complete_postxor
, sh
);
819 tx
= async_xor(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
820 flags
, tx
, ops_complete_postxor
, sh
);
823 static void ops_complete_check(void *stripe_head_ref
)
825 struct stripe_head
*sh
= stripe_head_ref
;
827 pr_debug("%s: stripe %llu\n", __func__
,
828 (unsigned long long)sh
->sector
);
830 sh
->check_state
= check_state_check_result
;
831 set_bit(STRIPE_HANDLE
, &sh
->state
);
835 static void ops_run_check(struct stripe_head
*sh
)
837 /* kernel stack size limits the total number of disks */
838 int disks
= sh
->disks
;
839 struct page
*xor_srcs
[disks
];
840 struct dma_async_tx_descriptor
*tx
;
842 int count
= 0, pd_idx
= sh
->pd_idx
, i
;
843 struct page
*xor_dest
= xor_srcs
[count
++] = sh
->dev
[pd_idx
].page
;
845 pr_debug("%s: stripe %llu\n", __func__
,
846 (unsigned long long)sh
->sector
);
848 for (i
= disks
; i
--; ) {
849 struct r5dev
*dev
= &sh
->dev
[i
];
851 xor_srcs
[count
++] = dev
->page
;
854 tx
= async_xor_val(xor_dest
, xor_srcs
, 0, count
, STRIPE_SIZE
,
855 &sh
->ops
.zero_sum_result
, 0, NULL
, NULL
, NULL
);
857 atomic_inc(&sh
->count
);
858 tx
= async_trigger_callback(ASYNC_TX_ACK
, tx
,
859 ops_complete_check
, sh
);
862 static void raid5_run_ops(struct stripe_head
*sh
, unsigned long ops_request
)
864 int overlap_clear
= 0, i
, disks
= sh
->disks
;
865 struct dma_async_tx_descriptor
*tx
= NULL
;
867 if (test_bit(STRIPE_OP_BIOFILL
, &ops_request
)) {
872 if (test_bit(STRIPE_OP_COMPUTE_BLK
, &ops_request
)) {
873 tx
= ops_run_compute5(sh
);
874 /* terminate the chain if postxor is not set to be run */
875 if (tx
&& !test_bit(STRIPE_OP_POSTXOR
, &ops_request
))
879 if (test_bit(STRIPE_OP_PREXOR
, &ops_request
))
880 tx
= ops_run_prexor(sh
, tx
);
882 if (test_bit(STRIPE_OP_BIODRAIN
, &ops_request
)) {
883 tx
= ops_run_biodrain(sh
, tx
);
887 if (test_bit(STRIPE_OP_POSTXOR
, &ops_request
))
888 ops_run_postxor(sh
, tx
);
890 if (test_bit(STRIPE_OP_CHECK
, &ops_request
))
894 for (i
= disks
; i
--; ) {
895 struct r5dev
*dev
= &sh
->dev
[i
];
896 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
897 wake_up(&sh
->raid_conf
->wait_for_overlap
);
901 static int grow_one_stripe(raid5_conf_t
*conf
)
903 struct stripe_head
*sh
;
904 sh
= kmem_cache_alloc(conf
->slab_cache
, GFP_KERNEL
);
907 memset(sh
, 0, sizeof(*sh
) + (conf
->raid_disks
-1)*sizeof(struct r5dev
));
908 sh
->raid_conf
= conf
;
909 spin_lock_init(&sh
->lock
);
911 if (grow_buffers(sh
, conf
->raid_disks
)) {
912 shrink_buffers(sh
, conf
->raid_disks
);
913 kmem_cache_free(conf
->slab_cache
, sh
);
916 sh
->disks
= conf
->raid_disks
;
917 /* we just created an active stripe so... */
918 atomic_set(&sh
->count
, 1);
919 atomic_inc(&conf
->active_stripes
);
920 INIT_LIST_HEAD(&sh
->lru
);
925 static int grow_stripes(raid5_conf_t
*conf
, int num
)
927 struct kmem_cache
*sc
;
928 int devs
= conf
->raid_disks
;
930 sprintf(conf
->cache_name
[0],
931 "raid%d-%s", conf
->level
, mdname(conf
->mddev
));
932 sprintf(conf
->cache_name
[1],
933 "raid%d-%s-alt", conf
->level
, mdname(conf
->mddev
));
934 conf
->active_name
= 0;
935 sc
= kmem_cache_create(conf
->cache_name
[conf
->active_name
],
936 sizeof(struct stripe_head
)+(devs
-1)*sizeof(struct r5dev
),
940 conf
->slab_cache
= sc
;
941 conf
->pool_size
= devs
;
943 if (!grow_one_stripe(conf
))
948 static int resize_stripes(raid5_conf_t
*conf
, int newsize
)
950 /* Make all the stripes able to hold 'newsize' devices.
951 * New slots in each stripe get 'page' set to a new page.
953 * This happens in stages:
954 * 1/ create a new kmem_cache and allocate the required number of
956 * 2/ gather all the old stripe_heads and tranfer the pages across
957 * to the new stripe_heads. This will have the side effect of
958 * freezing the array as once all stripe_heads have been collected,
959 * no IO will be possible. Old stripe heads are freed once their
960 * pages have been transferred over, and the old kmem_cache is
961 * freed when all stripes are done.
962 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
963 * we simple return a failre status - no need to clean anything up.
964 * 4/ allocate new pages for the new slots in the new stripe_heads.
965 * If this fails, we don't bother trying the shrink the
966 * stripe_heads down again, we just leave them as they are.
967 * As each stripe_head is processed the new one is released into
970 * Once step2 is started, we cannot afford to wait for a write,
971 * so we use GFP_NOIO allocations.
973 struct stripe_head
*osh
, *nsh
;
974 LIST_HEAD(newstripes
);
975 struct disk_info
*ndisks
;
977 struct kmem_cache
*sc
;
980 if (newsize
<= conf
->pool_size
)
981 return 0; /* never bother to shrink */
983 err
= md_allow_write(conf
->mddev
);
988 sc
= kmem_cache_create(conf
->cache_name
[1-conf
->active_name
],
989 sizeof(struct stripe_head
)+(newsize
-1)*sizeof(struct r5dev
),
994 for (i
= conf
->max_nr_stripes
; i
; i
--) {
995 nsh
= kmem_cache_alloc(sc
, GFP_KERNEL
);
999 memset(nsh
, 0, sizeof(*nsh
) + (newsize
-1)*sizeof(struct r5dev
));
1001 nsh
->raid_conf
= conf
;
1002 spin_lock_init(&nsh
->lock
);
1004 list_add(&nsh
->lru
, &newstripes
);
1007 /* didn't get enough, give up */
1008 while (!list_empty(&newstripes
)) {
1009 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1010 list_del(&nsh
->lru
);
1011 kmem_cache_free(sc
, nsh
);
1013 kmem_cache_destroy(sc
);
1016 /* Step 2 - Must use GFP_NOIO now.
1017 * OK, we have enough stripes, start collecting inactive
1018 * stripes and copying them over
1020 list_for_each_entry(nsh
, &newstripes
, lru
) {
1021 spin_lock_irq(&conf
->device_lock
);
1022 wait_event_lock_irq(conf
->wait_for_stripe
,
1023 !list_empty(&conf
->inactive_list
),
1025 unplug_slaves(conf
->mddev
)
1027 osh
= get_free_stripe(conf
);
1028 spin_unlock_irq(&conf
->device_lock
);
1029 atomic_set(&nsh
->count
, 1);
1030 for(i
=0; i
<conf
->pool_size
; i
++)
1031 nsh
->dev
[i
].page
= osh
->dev
[i
].page
;
1032 for( ; i
<newsize
; i
++)
1033 nsh
->dev
[i
].page
= NULL
;
1034 kmem_cache_free(conf
->slab_cache
, osh
);
1036 kmem_cache_destroy(conf
->slab_cache
);
1039 * At this point, we are holding all the stripes so the array
1040 * is completely stalled, so now is a good time to resize
1043 ndisks
= kzalloc(newsize
* sizeof(struct disk_info
), GFP_NOIO
);
1045 for (i
=0; i
<conf
->raid_disks
; i
++)
1046 ndisks
[i
] = conf
->disks
[i
];
1048 conf
->disks
= ndisks
;
1052 /* Step 4, return new stripes to service */
1053 while(!list_empty(&newstripes
)) {
1054 nsh
= list_entry(newstripes
.next
, struct stripe_head
, lru
);
1055 list_del_init(&nsh
->lru
);
1056 for (i
=conf
->raid_disks
; i
< newsize
; i
++)
1057 if (nsh
->dev
[i
].page
== NULL
) {
1058 struct page
*p
= alloc_page(GFP_NOIO
);
1059 nsh
->dev
[i
].page
= p
;
1063 release_stripe(nsh
);
1065 /* critical section pass, GFP_NOIO no longer needed */
1067 conf
->slab_cache
= sc
;
1068 conf
->active_name
= 1-conf
->active_name
;
1069 conf
->pool_size
= newsize
;
1073 static int drop_one_stripe(raid5_conf_t
*conf
)
1075 struct stripe_head
*sh
;
1077 spin_lock_irq(&conf
->device_lock
);
1078 sh
= get_free_stripe(conf
);
1079 spin_unlock_irq(&conf
->device_lock
);
1082 BUG_ON(atomic_read(&sh
->count
));
1083 shrink_buffers(sh
, conf
->pool_size
);
1084 kmem_cache_free(conf
->slab_cache
, sh
);
1085 atomic_dec(&conf
->active_stripes
);
1089 static void shrink_stripes(raid5_conf_t
*conf
)
1091 while (drop_one_stripe(conf
))
1094 if (conf
->slab_cache
)
1095 kmem_cache_destroy(conf
->slab_cache
);
1096 conf
->slab_cache
= NULL
;
1099 static void raid5_end_read_request(struct bio
* bi
, int error
)
1101 struct stripe_head
*sh
= bi
->bi_private
;
1102 raid5_conf_t
*conf
= sh
->raid_conf
;
1103 int disks
= sh
->disks
, i
;
1104 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1105 char b
[BDEVNAME_SIZE
];
1109 for (i
=0 ; i
<disks
; i
++)
1110 if (bi
== &sh
->dev
[i
].req
)
1113 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1114 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1122 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1123 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
1124 rdev
= conf
->disks
[i
].rdev
;
1125 printk_rl(KERN_INFO
"raid5:%s: read error corrected"
1126 " (%lu sectors at %llu on %s)\n",
1127 mdname(conf
->mddev
), STRIPE_SECTORS
,
1128 (unsigned long long)(sh
->sector
1129 + rdev
->data_offset
),
1130 bdevname(rdev
->bdev
, b
));
1131 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1132 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1134 if (atomic_read(&conf
->disks
[i
].rdev
->read_errors
))
1135 atomic_set(&conf
->disks
[i
].rdev
->read_errors
, 0);
1137 const char *bdn
= bdevname(conf
->disks
[i
].rdev
->bdev
, b
);
1139 rdev
= conf
->disks
[i
].rdev
;
1141 clear_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1142 atomic_inc(&rdev
->read_errors
);
1143 if (conf
->mddev
->degraded
)
1144 printk_rl(KERN_WARNING
1145 "raid5:%s: read error not correctable "
1146 "(sector %llu on %s).\n",
1147 mdname(conf
->mddev
),
1148 (unsigned long long)(sh
->sector
1149 + rdev
->data_offset
),
1151 else if (test_bit(R5_ReWrite
, &sh
->dev
[i
].flags
))
1153 printk_rl(KERN_WARNING
1154 "raid5:%s: read error NOT corrected!! "
1155 "(sector %llu on %s).\n",
1156 mdname(conf
->mddev
),
1157 (unsigned long long)(sh
->sector
1158 + rdev
->data_offset
),
1160 else if (atomic_read(&rdev
->read_errors
)
1161 > conf
->max_nr_stripes
)
1163 "raid5:%s: Too many read errors, failing device %s.\n",
1164 mdname(conf
->mddev
), bdn
);
1168 set_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1170 clear_bit(R5_ReadError
, &sh
->dev
[i
].flags
);
1171 clear_bit(R5_ReWrite
, &sh
->dev
[i
].flags
);
1172 md_error(conf
->mddev
, rdev
);
1175 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1176 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1177 set_bit(STRIPE_HANDLE
, &sh
->state
);
1181 static void raid5_end_write_request(struct bio
*bi
, int error
)
1183 struct stripe_head
*sh
= bi
->bi_private
;
1184 raid5_conf_t
*conf
= sh
->raid_conf
;
1185 int disks
= sh
->disks
, i
;
1186 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
1188 for (i
=0 ; i
<disks
; i
++)
1189 if (bi
== &sh
->dev
[i
].req
)
1192 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1193 (unsigned long long)sh
->sector
, i
, atomic_read(&sh
->count
),
1201 md_error(conf
->mddev
, conf
->disks
[i
].rdev
);
1203 rdev_dec_pending(conf
->disks
[i
].rdev
, conf
->mddev
);
1205 clear_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1206 set_bit(STRIPE_HANDLE
, &sh
->state
);
1211 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
);
1213 static void raid5_build_block(struct stripe_head
*sh
, int i
, int previous
)
1215 struct r5dev
*dev
= &sh
->dev
[i
];
1217 bio_init(&dev
->req
);
1218 dev
->req
.bi_io_vec
= &dev
->vec
;
1220 dev
->req
.bi_max_vecs
++;
1221 dev
->vec
.bv_page
= dev
->page
;
1222 dev
->vec
.bv_len
= STRIPE_SIZE
;
1223 dev
->vec
.bv_offset
= 0;
1225 dev
->req
.bi_sector
= sh
->sector
;
1226 dev
->req
.bi_private
= sh
;
1229 dev
->sector
= compute_blocknr(sh
, i
, previous
);
1232 static void error(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
1234 char b
[BDEVNAME_SIZE
];
1235 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
1236 pr_debug("raid5: error called\n");
1238 if (!test_bit(Faulty
, &rdev
->flags
)) {
1239 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1240 if (test_and_clear_bit(In_sync
, &rdev
->flags
)) {
1241 unsigned long flags
;
1242 spin_lock_irqsave(&conf
->device_lock
, flags
);
1244 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
1246 * if recovery was running, make sure it aborts.
1248 set_bit(MD_RECOVERY_INTR
, &mddev
->recovery
);
1250 set_bit(Faulty
, &rdev
->flags
);
1252 "raid5: Disk failure on %s, disabling device.\n"
1253 "raid5: Operation continuing on %d devices.\n",
1254 bdevname(rdev
->bdev
,b
), conf
->raid_disks
- mddev
->degraded
);
1259 * Input: a 'big' sector number,
1260 * Output: index of the data and parity disk, and the sector # in them.
1262 static sector_t
raid5_compute_sector(raid5_conf_t
*conf
, sector_t r_sector
,
1263 int previous
, int *dd_idx
,
1264 struct stripe_head
*sh
)
1267 unsigned long chunk_number
;
1268 unsigned int chunk_offset
;
1271 sector_t new_sector
;
1272 int algorithm
= previous
? conf
->prev_algo
1274 int sectors_per_chunk
= previous
? (conf
->prev_chunk
>> 9)
1275 : (conf
->chunk_size
>> 9);
1276 int raid_disks
= previous
? conf
->previous_raid_disks
1278 int data_disks
= raid_disks
- conf
->max_degraded
;
1280 /* First compute the information on this sector */
1283 * Compute the chunk number and the sector offset inside the chunk
1285 chunk_offset
= sector_div(r_sector
, sectors_per_chunk
);
1286 chunk_number
= r_sector
;
1287 BUG_ON(r_sector
!= chunk_number
);
1290 * Compute the stripe number
1292 stripe
= chunk_number
/ data_disks
;
1295 * Compute the data disk and parity disk indexes inside the stripe
1297 *dd_idx
= chunk_number
% data_disks
;
1300 * Select the parity disk based on the user selected algorithm.
1302 pd_idx
= qd_idx
= ~0;
1303 switch(conf
->level
) {
1305 pd_idx
= data_disks
;
1308 switch (algorithm
) {
1309 case ALGORITHM_LEFT_ASYMMETRIC
:
1310 pd_idx
= data_disks
- stripe
% raid_disks
;
1311 if (*dd_idx
>= pd_idx
)
1314 case ALGORITHM_RIGHT_ASYMMETRIC
:
1315 pd_idx
= stripe
% raid_disks
;
1316 if (*dd_idx
>= pd_idx
)
1319 case ALGORITHM_LEFT_SYMMETRIC
:
1320 pd_idx
= data_disks
- stripe
% raid_disks
;
1321 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1323 case ALGORITHM_RIGHT_SYMMETRIC
:
1324 pd_idx
= stripe
% raid_disks
;
1325 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1327 case ALGORITHM_PARITY_0
:
1331 case ALGORITHM_PARITY_N
:
1332 pd_idx
= data_disks
;
1335 printk(KERN_ERR
"raid5: unsupported algorithm %d\n",
1342 switch (algorithm
) {
1343 case ALGORITHM_LEFT_ASYMMETRIC
:
1344 pd_idx
= raid_disks
- 1 - (stripe
% raid_disks
);
1345 qd_idx
= pd_idx
+ 1;
1346 if (pd_idx
== raid_disks
-1) {
1347 (*dd_idx
)++; /* Q D D D P */
1349 } else if (*dd_idx
>= pd_idx
)
1350 (*dd_idx
) += 2; /* D D P Q D */
1352 case ALGORITHM_RIGHT_ASYMMETRIC
:
1353 pd_idx
= stripe
% raid_disks
;
1354 qd_idx
= pd_idx
+ 1;
1355 if (pd_idx
== raid_disks
-1) {
1356 (*dd_idx
)++; /* Q D D D P */
1358 } else if (*dd_idx
>= pd_idx
)
1359 (*dd_idx
) += 2; /* D D P Q D */
1361 case ALGORITHM_LEFT_SYMMETRIC
:
1362 pd_idx
= raid_disks
- 1 - (stripe
% raid_disks
);
1363 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1364 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1366 case ALGORITHM_RIGHT_SYMMETRIC
:
1367 pd_idx
= stripe
% raid_disks
;
1368 qd_idx
= (pd_idx
+ 1) % raid_disks
;
1369 *dd_idx
= (pd_idx
+ 2 + *dd_idx
) % raid_disks
;
1372 case ALGORITHM_PARITY_0
:
1377 case ALGORITHM_PARITY_N
:
1378 pd_idx
= data_disks
;
1379 qd_idx
= data_disks
+ 1;
1382 case ALGORITHM_ROTATING_ZERO_RESTART
:
1383 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1384 * of blocks for computing Q is different.
1386 pd_idx
= stripe
% raid_disks
;
1387 qd_idx
= pd_idx
+ 1;
1388 if (pd_idx
== raid_disks
-1) {
1389 (*dd_idx
)++; /* Q D D D P */
1391 } else if (*dd_idx
>= pd_idx
)
1392 (*dd_idx
) += 2; /* D D P Q D */
1396 case ALGORITHM_ROTATING_N_RESTART
:
1397 /* Same a left_asymmetric, by first stripe is
1398 * D D D P Q rather than
1401 pd_idx
= raid_disks
- 1 - ((stripe
+ 1) % raid_disks
);
1402 qd_idx
= pd_idx
+ 1;
1403 if (pd_idx
== raid_disks
-1) {
1404 (*dd_idx
)++; /* Q D D D P */
1406 } else if (*dd_idx
>= pd_idx
)
1407 (*dd_idx
) += 2; /* D D P Q D */
1411 case ALGORITHM_ROTATING_N_CONTINUE
:
1412 /* Same as left_symmetric but Q is before P */
1413 pd_idx
= raid_disks
- 1 - (stripe
% raid_disks
);
1414 qd_idx
= (pd_idx
+ raid_disks
- 1) % raid_disks
;
1415 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % raid_disks
;
1419 case ALGORITHM_LEFT_ASYMMETRIC_6
:
1420 /* RAID5 left_asymmetric, with Q on last device */
1421 pd_idx
= data_disks
- stripe
% (raid_disks
-1);
1422 if (*dd_idx
>= pd_idx
)
1424 qd_idx
= raid_disks
- 1;
1427 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
1428 pd_idx
= stripe
% (raid_disks
-1);
1429 if (*dd_idx
>= pd_idx
)
1431 qd_idx
= raid_disks
- 1;
1434 case ALGORITHM_LEFT_SYMMETRIC_6
:
1435 pd_idx
= data_disks
- stripe
% (raid_disks
-1);
1436 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1437 qd_idx
= raid_disks
- 1;
1440 case ALGORITHM_RIGHT_SYMMETRIC_6
:
1441 pd_idx
= stripe
% (raid_disks
-1);
1442 *dd_idx
= (pd_idx
+ 1 + *dd_idx
) % (raid_disks
-1);
1443 qd_idx
= raid_disks
- 1;
1446 case ALGORITHM_PARITY_0_6
:
1449 qd_idx
= raid_disks
- 1;
1454 printk(KERN_CRIT
"raid6: unsupported algorithm %d\n",
1462 sh
->pd_idx
= pd_idx
;
1463 sh
->qd_idx
= qd_idx
;
1464 sh
->ddf_layout
= ddf_layout
;
1467 * Finally, compute the new sector number
1469 new_sector
= (sector_t
)stripe
* sectors_per_chunk
+ chunk_offset
;
1474 static sector_t
compute_blocknr(struct stripe_head
*sh
, int i
, int previous
)
1476 raid5_conf_t
*conf
= sh
->raid_conf
;
1477 int raid_disks
= sh
->disks
;
1478 int data_disks
= raid_disks
- conf
->max_degraded
;
1479 sector_t new_sector
= sh
->sector
, check
;
1480 int sectors_per_chunk
= previous
? (conf
->prev_chunk
>> 9)
1481 : (conf
->chunk_size
>> 9);
1482 int algorithm
= previous
? conf
->prev_algo
1486 int chunk_number
, dummy1
, dd_idx
= i
;
1488 struct stripe_head sh2
;
1491 chunk_offset
= sector_div(new_sector
, sectors_per_chunk
);
1492 stripe
= new_sector
;
1493 BUG_ON(new_sector
!= stripe
);
1495 if (i
== sh
->pd_idx
)
1497 switch(conf
->level
) {
1500 switch (algorithm
) {
1501 case ALGORITHM_LEFT_ASYMMETRIC
:
1502 case ALGORITHM_RIGHT_ASYMMETRIC
:
1506 case ALGORITHM_LEFT_SYMMETRIC
:
1507 case ALGORITHM_RIGHT_SYMMETRIC
:
1510 i
-= (sh
->pd_idx
+ 1);
1512 case ALGORITHM_PARITY_0
:
1515 case ALGORITHM_PARITY_N
:
1518 printk(KERN_ERR
"raid5: unsupported algorithm %d\n",
1524 if (i
== sh
->qd_idx
)
1525 return 0; /* It is the Q disk */
1526 switch (algorithm
) {
1527 case ALGORITHM_LEFT_ASYMMETRIC
:
1528 case ALGORITHM_RIGHT_ASYMMETRIC
:
1529 case ALGORITHM_ROTATING_ZERO_RESTART
:
1530 case ALGORITHM_ROTATING_N_RESTART
:
1531 if (sh
->pd_idx
== raid_disks
-1)
1532 i
--; /* Q D D D P */
1533 else if (i
> sh
->pd_idx
)
1534 i
-= 2; /* D D P Q D */
1536 case ALGORITHM_LEFT_SYMMETRIC
:
1537 case ALGORITHM_RIGHT_SYMMETRIC
:
1538 if (sh
->pd_idx
== raid_disks
-1)
1539 i
--; /* Q D D D P */
1544 i
-= (sh
->pd_idx
+ 2);
1547 case ALGORITHM_PARITY_0
:
1550 case ALGORITHM_PARITY_N
:
1552 case ALGORITHM_ROTATING_N_CONTINUE
:
1553 if (sh
->pd_idx
== 0)
1554 i
--; /* P D D D Q */
1555 else if (i
> sh
->pd_idx
)
1556 i
-= 2; /* D D Q P D */
1558 case ALGORITHM_LEFT_ASYMMETRIC_6
:
1559 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
1563 case ALGORITHM_LEFT_SYMMETRIC_6
:
1564 case ALGORITHM_RIGHT_SYMMETRIC_6
:
1566 i
+= data_disks
+ 1;
1567 i
-= (sh
->pd_idx
+ 1);
1569 case ALGORITHM_PARITY_0_6
:
1573 printk(KERN_CRIT
"raid6: unsupported algorithm %d\n",
1580 chunk_number
= stripe
* data_disks
+ i
;
1581 r_sector
= (sector_t
)chunk_number
* sectors_per_chunk
+ chunk_offset
;
1583 check
= raid5_compute_sector(conf
, r_sector
,
1584 previous
, &dummy1
, &sh2
);
1585 if (check
!= sh
->sector
|| dummy1
!= dd_idx
|| sh2
.pd_idx
!= sh
->pd_idx
1586 || sh2
.qd_idx
!= sh
->qd_idx
) {
1587 printk(KERN_ERR
"compute_blocknr: map not correct\n");
1596 * Copy data between a page in the stripe cache, and one or more bion
1597 * The page could align with the middle of the bio, or there could be
1598 * several bion, each with several bio_vecs, which cover part of the page
1599 * Multiple bion are linked together on bi_next. There may be extras
1600 * at the end of this list. We ignore them.
1602 static void copy_data(int frombio
, struct bio
*bio
,
1606 char *pa
= page_address(page
);
1607 struct bio_vec
*bvl
;
1611 if (bio
->bi_sector
>= sector
)
1612 page_offset
= (signed)(bio
->bi_sector
- sector
) * 512;
1614 page_offset
= (signed)(sector
- bio
->bi_sector
) * -512;
1615 bio_for_each_segment(bvl
, bio
, i
) {
1616 int len
= bio_iovec_idx(bio
,i
)->bv_len
;
1620 if (page_offset
< 0) {
1621 b_offset
= -page_offset
;
1622 page_offset
+= b_offset
;
1626 if (len
> 0 && page_offset
+ len
> STRIPE_SIZE
)
1627 clen
= STRIPE_SIZE
- page_offset
;
1631 char *ba
= __bio_kmap_atomic(bio
, i
, KM_USER0
);
1633 memcpy(pa
+page_offset
, ba
+b_offset
, clen
);
1635 memcpy(ba
+b_offset
, pa
+page_offset
, clen
);
1636 __bio_kunmap_atomic(ba
, KM_USER0
);
1638 if (clen
< len
) /* hit end of page */
1644 #define check_xor() do { \
1645 if (count == MAX_XOR_BLOCKS) { \
1646 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1651 static void compute_parity6(struct stripe_head
*sh
, int method
)
1653 raid5_conf_t
*conf
= sh
->raid_conf
;
1654 int i
, pd_idx
, qd_idx
, d0_idx
, disks
= sh
->disks
, count
;
1655 int syndrome_disks
= sh
->ddf_layout
? disks
: (disks
- 2);
1657 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1658 void *ptrs
[syndrome_disks
+2];
1660 pd_idx
= sh
->pd_idx
;
1661 qd_idx
= sh
->qd_idx
;
1662 d0_idx
= raid6_d0(sh
);
1664 pr_debug("compute_parity, stripe %llu, method %d\n",
1665 (unsigned long long)sh
->sector
, method
);
1668 case READ_MODIFY_WRITE
:
1669 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1670 case RECONSTRUCT_WRITE
:
1671 for (i
= disks
; i
-- ;)
1672 if ( i
!= pd_idx
&& i
!= qd_idx
&& sh
->dev
[i
].towrite
) {
1673 chosen
= sh
->dev
[i
].towrite
;
1674 sh
->dev
[i
].towrite
= NULL
;
1676 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
1677 wake_up(&conf
->wait_for_overlap
);
1679 BUG_ON(sh
->dev
[i
].written
);
1680 sh
->dev
[i
].written
= chosen
;
1684 BUG(); /* Not implemented yet */
1687 for (i
= disks
; i
--;)
1688 if (sh
->dev
[i
].written
) {
1689 sector_t sector
= sh
->dev
[i
].sector
;
1690 struct bio
*wbi
= sh
->dev
[i
].written
;
1691 while (wbi
&& wbi
->bi_sector
< sector
+ STRIPE_SECTORS
) {
1692 copy_data(1, wbi
, sh
->dev
[i
].page
, sector
);
1693 wbi
= r5_next_bio(wbi
, sector
);
1696 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
1697 set_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
);
1700 /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
1702 for (i
= 0; i
< disks
; i
++)
1703 ptrs
[i
] = (void *)raid6_empty_zero_page
;
1708 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1710 ptrs
[slot
] = page_address(sh
->dev
[i
].page
);
1711 if (slot
< syndrome_disks
&&
1712 !test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
)) {
1713 printk(KERN_ERR
"block %d/%d not uptodate "
1714 "on parity calc\n", i
, count
);
1718 i
= raid6_next_disk(i
, disks
);
1719 } while (i
!= d0_idx
);
1720 BUG_ON(count
!= syndrome_disks
);
1722 raid6_call
.gen_syndrome(syndrome_disks
+2, STRIPE_SIZE
, ptrs
);
1725 case RECONSTRUCT_WRITE
:
1726 set_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
1727 set_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
);
1728 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
1729 set_bit(R5_LOCKED
, &sh
->dev
[qd_idx
].flags
);
1732 set_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
1733 set_bit(R5_UPTODATE
, &sh
->dev
[qd_idx
].flags
);
1739 /* Compute one missing block */
1740 static void compute_block_1(struct stripe_head
*sh
, int dd_idx
, int nozero
)
1742 int i
, count
, disks
= sh
->disks
;
1743 void *ptr
[MAX_XOR_BLOCKS
], *dest
, *p
;
1744 int qd_idx
= sh
->qd_idx
;
1746 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1747 (unsigned long long)sh
->sector
, dd_idx
);
1749 if ( dd_idx
== qd_idx
) {
1750 /* We're actually computing the Q drive */
1751 compute_parity6(sh
, UPDATE_PARITY
);
1753 dest
= page_address(sh
->dev
[dd_idx
].page
);
1754 if (!nozero
) memset(dest
, 0, STRIPE_SIZE
);
1756 for (i
= disks
; i
--; ) {
1757 if (i
== dd_idx
|| i
== qd_idx
)
1759 p
= page_address(sh
->dev
[i
].page
);
1760 if (test_bit(R5_UPTODATE
, &sh
->dev
[i
].flags
))
1763 printk("compute_block() %d, stripe %llu, %d"
1764 " not present\n", dd_idx
,
1765 (unsigned long long)sh
->sector
, i
);
1770 xor_blocks(count
, STRIPE_SIZE
, dest
, ptr
);
1771 if (!nozero
) set_bit(R5_UPTODATE
, &sh
->dev
[dd_idx
].flags
);
1772 else clear_bit(R5_UPTODATE
, &sh
->dev
[dd_idx
].flags
);
1776 /* Compute two missing blocks */
1777 static void compute_block_2(struct stripe_head
*sh
, int dd_idx1
, int dd_idx2
)
1779 int i
, count
, disks
= sh
->disks
;
1780 int syndrome_disks
= sh
->ddf_layout
? disks
: disks
-2;
1781 int d0_idx
= raid6_d0(sh
);
1782 int faila
= -1, failb
= -1;
1783 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1784 void *ptrs
[syndrome_disks
+2];
1786 for (i
= 0; i
< disks
; i
++)
1787 ptrs
[i
] = (void *)raid6_empty_zero_page
;
1791 int slot
= raid6_idx_to_slot(i
, sh
, &count
, syndrome_disks
);
1793 ptrs
[slot
] = page_address(sh
->dev
[i
].page
);
1799 i
= raid6_next_disk(i
, disks
);
1800 } while (i
!= d0_idx
);
1801 BUG_ON(count
!= syndrome_disks
);
1803 BUG_ON(faila
== failb
);
1804 if ( failb
< faila
) { int tmp
= faila
; faila
= failb
; failb
= tmp
; }
1806 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1807 (unsigned long long)sh
->sector
, dd_idx1
, dd_idx2
,
1810 if (failb
== syndrome_disks
+1) {
1811 /* Q disk is one of the missing disks */
1812 if (faila
== syndrome_disks
) {
1813 /* Missing P+Q, just recompute */
1814 compute_parity6(sh
, UPDATE_PARITY
);
1817 /* We're missing D+Q; recompute D from P */
1818 compute_block_1(sh
, ((dd_idx1
== sh
->qd_idx
) ?
1821 compute_parity6(sh
, UPDATE_PARITY
); /* Is this necessary? */
1826 /* We're missing D+P or D+D; */
1827 if (failb
== syndrome_disks
) {
1828 /* We're missing D+P. */
1829 raid6_datap_recov(syndrome_disks
+2, STRIPE_SIZE
, faila
, ptrs
);
1831 /* We're missing D+D. */
1832 raid6_2data_recov(syndrome_disks
+2, STRIPE_SIZE
, faila
, failb
,
1836 /* Both the above update both missing blocks */
1837 set_bit(R5_UPTODATE
, &sh
->dev
[dd_idx1
].flags
);
1838 set_bit(R5_UPTODATE
, &sh
->dev
[dd_idx2
].flags
);
1842 schedule_reconstruction5(struct stripe_head
*sh
, struct stripe_head_state
*s
,
1843 int rcw
, int expand
)
1845 int i
, pd_idx
= sh
->pd_idx
, disks
= sh
->disks
;
1848 /* if we are not expanding this is a proper write request, and
1849 * there will be bios with new data to be drained into the
1853 sh
->reconstruct_state
= reconstruct_state_drain_run
;
1854 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
1856 sh
->reconstruct_state
= reconstruct_state_run
;
1858 set_bit(STRIPE_OP_POSTXOR
, &s
->ops_request
);
1860 for (i
= disks
; i
--; ) {
1861 struct r5dev
*dev
= &sh
->dev
[i
];
1864 set_bit(R5_LOCKED
, &dev
->flags
);
1865 set_bit(R5_Wantdrain
, &dev
->flags
);
1867 clear_bit(R5_UPTODATE
, &dev
->flags
);
1871 if (s
->locked
+ 1 == disks
)
1872 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
1873 atomic_inc(&sh
->raid_conf
->pending_full_writes
);
1875 BUG_ON(!(test_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
) ||
1876 test_bit(R5_Wantcompute
, &sh
->dev
[pd_idx
].flags
)));
1878 sh
->reconstruct_state
= reconstruct_state_prexor_drain_run
;
1879 set_bit(STRIPE_OP_PREXOR
, &s
->ops_request
);
1880 set_bit(STRIPE_OP_BIODRAIN
, &s
->ops_request
);
1881 set_bit(STRIPE_OP_POSTXOR
, &s
->ops_request
);
1883 for (i
= disks
; i
--; ) {
1884 struct r5dev
*dev
= &sh
->dev
[i
];
1889 (test_bit(R5_UPTODATE
, &dev
->flags
) ||
1890 test_bit(R5_Wantcompute
, &dev
->flags
))) {
1891 set_bit(R5_Wantdrain
, &dev
->flags
);
1892 set_bit(R5_LOCKED
, &dev
->flags
);
1893 clear_bit(R5_UPTODATE
, &dev
->flags
);
1899 /* keep the parity disk locked while asynchronous operations
1902 set_bit(R5_LOCKED
, &sh
->dev
[pd_idx
].flags
);
1903 clear_bit(R5_UPTODATE
, &sh
->dev
[pd_idx
].flags
);
1906 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1907 __func__
, (unsigned long long)sh
->sector
,
1908 s
->locked
, s
->ops_request
);
1912 * Each stripe/dev can have one or more bion attached.
1913 * toread/towrite point to the first in a chain.
1914 * The bi_next chain must be in order.
1916 static int add_stripe_bio(struct stripe_head
*sh
, struct bio
*bi
, int dd_idx
, int forwrite
)
1919 raid5_conf_t
*conf
= sh
->raid_conf
;
1922 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1923 (unsigned long long)bi
->bi_sector
,
1924 (unsigned long long)sh
->sector
);
1927 spin_lock(&sh
->lock
);
1928 spin_lock_irq(&conf
->device_lock
);
1930 bip
= &sh
->dev
[dd_idx
].towrite
;
1931 if (*bip
== NULL
&& sh
->dev
[dd_idx
].written
== NULL
)
1934 bip
= &sh
->dev
[dd_idx
].toread
;
1935 while (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
) {
1936 if ((*bip
)->bi_sector
+ ((*bip
)->bi_size
>> 9) > bi
->bi_sector
)
1938 bip
= & (*bip
)->bi_next
;
1940 if (*bip
&& (*bip
)->bi_sector
< bi
->bi_sector
+ ((bi
->bi_size
)>>9))
1943 BUG_ON(*bip
&& bi
->bi_next
&& (*bip
) != bi
->bi_next
);
1947 bi
->bi_phys_segments
++;
1948 spin_unlock_irq(&conf
->device_lock
);
1949 spin_unlock(&sh
->lock
);
1951 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1952 (unsigned long long)bi
->bi_sector
,
1953 (unsigned long long)sh
->sector
, dd_idx
);
1955 if (conf
->mddev
->bitmap
&& firstwrite
) {
1956 bitmap_startwrite(conf
->mddev
->bitmap
, sh
->sector
,
1958 sh
->bm_seq
= conf
->seq_flush
+1;
1959 set_bit(STRIPE_BIT_DELAY
, &sh
->state
);
1963 /* check if page is covered */
1964 sector_t sector
= sh
->dev
[dd_idx
].sector
;
1965 for (bi
=sh
->dev
[dd_idx
].towrite
;
1966 sector
< sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
&&
1967 bi
&& bi
->bi_sector
<= sector
;
1968 bi
= r5_next_bio(bi
, sh
->dev
[dd_idx
].sector
)) {
1969 if (bi
->bi_sector
+ (bi
->bi_size
>>9) >= sector
)
1970 sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
1972 if (sector
>= sh
->dev
[dd_idx
].sector
+ STRIPE_SECTORS
)
1973 set_bit(R5_OVERWRITE
, &sh
->dev
[dd_idx
].flags
);
1978 set_bit(R5_Overlap
, &sh
->dev
[dd_idx
].flags
);
1979 spin_unlock_irq(&conf
->device_lock
);
1980 spin_unlock(&sh
->lock
);
1984 static void end_reshape(raid5_conf_t
*conf
);
1986 static int page_is_zero(struct page
*p
)
1988 char *a
= page_address(p
);
1989 return ((*(u32
*)a
) == 0 &&
1990 memcmp(a
, a
+4, STRIPE_SIZE
-4)==0);
1993 static void stripe_set_idx(sector_t stripe
, raid5_conf_t
*conf
, int previous
,
1994 struct stripe_head
*sh
)
1996 int sectors_per_chunk
=
1997 previous
? (conf
->prev_chunk
>> 9)
1998 : (conf
->chunk_size
>> 9);
2000 int chunk_offset
= sector_div(stripe
, sectors_per_chunk
);
2001 int disks
= previous
? conf
->previous_raid_disks
: conf
->raid_disks
;
2003 raid5_compute_sector(conf
,
2004 stripe
* (disks
- conf
->max_degraded
)
2005 *sectors_per_chunk
+ chunk_offset
,
2011 handle_failed_stripe(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2012 struct stripe_head_state
*s
, int disks
,
2013 struct bio
**return_bi
)
2016 for (i
= disks
; i
--; ) {
2020 if (test_bit(R5_ReadError
, &sh
->dev
[i
].flags
)) {
2023 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2024 if (rdev
&& test_bit(In_sync
, &rdev
->flags
))
2025 /* multiple read failures in one stripe */
2026 md_error(conf
->mddev
, rdev
);
2029 spin_lock_irq(&conf
->device_lock
);
2030 /* fail all writes first */
2031 bi
= sh
->dev
[i
].towrite
;
2032 sh
->dev
[i
].towrite
= NULL
;
2038 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2039 wake_up(&conf
->wait_for_overlap
);
2041 while (bi
&& bi
->bi_sector
<
2042 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2043 struct bio
*nextbi
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2044 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2045 if (!raid5_dec_bi_phys_segments(bi
)) {
2046 md_write_end(conf
->mddev
);
2047 bi
->bi_next
= *return_bi
;
2052 /* and fail all 'written' */
2053 bi
= sh
->dev
[i
].written
;
2054 sh
->dev
[i
].written
= NULL
;
2055 if (bi
) bitmap_end
= 1;
2056 while (bi
&& bi
->bi_sector
<
2057 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2058 struct bio
*bi2
= r5_next_bio(bi
, sh
->dev
[i
].sector
);
2059 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2060 if (!raid5_dec_bi_phys_segments(bi
)) {
2061 md_write_end(conf
->mddev
);
2062 bi
->bi_next
= *return_bi
;
2068 /* fail any reads if this device is non-operational and
2069 * the data has not reached the cache yet.
2071 if (!test_bit(R5_Wantfill
, &sh
->dev
[i
].flags
) &&
2072 (!test_bit(R5_Insync
, &sh
->dev
[i
].flags
) ||
2073 test_bit(R5_ReadError
, &sh
->dev
[i
].flags
))) {
2074 bi
= sh
->dev
[i
].toread
;
2075 sh
->dev
[i
].toread
= NULL
;
2076 if (test_and_clear_bit(R5_Overlap
, &sh
->dev
[i
].flags
))
2077 wake_up(&conf
->wait_for_overlap
);
2078 if (bi
) s
->to_read
--;
2079 while (bi
&& bi
->bi_sector
<
2080 sh
->dev
[i
].sector
+ STRIPE_SECTORS
) {
2081 struct bio
*nextbi
=
2082 r5_next_bio(bi
, sh
->dev
[i
].sector
);
2083 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
2084 if (!raid5_dec_bi_phys_segments(bi
)) {
2085 bi
->bi_next
= *return_bi
;
2091 spin_unlock_irq(&conf
->device_lock
);
2093 bitmap_endwrite(conf
->mddev
->bitmap
, sh
->sector
,
2094 STRIPE_SECTORS
, 0, 0);
2097 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2098 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2099 md_wakeup_thread(conf
->mddev
->thread
);
2102 /* fetch_block5 - checks the given member device to see if its data needs
2103 * to be read or computed to satisfy a request.
2105 * Returns 1 when no more member devices need to be checked, otherwise returns
2106 * 0 to tell the loop in handle_stripe_fill5 to continue
2108 static int fetch_block5(struct stripe_head
*sh
, struct stripe_head_state
*s
,
2109 int disk_idx
, int disks
)
2111 struct r5dev
*dev
= &sh
->dev
[disk_idx
];
2112 struct r5dev
*failed_dev
= &sh
->dev
[s
->failed_num
];
2114 /* is the data in this block needed, and can we get it? */
2115 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2116 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2118 (dev
->towrite
&& !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2119 s
->syncing
|| s
->expanding
||
2121 (failed_dev
->toread
||
2122 (failed_dev
->towrite
&&
2123 !test_bit(R5_OVERWRITE
, &failed_dev
->flags
)))))) {
2124 /* We would like to get this block, possibly by computing it,
2125 * otherwise read it if the backing disk is insync
2127 if ((s
->uptodate
== disks
- 1) &&
2128 (s
->failed
&& disk_idx
== s
->failed_num
)) {
2129 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2130 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2131 set_bit(R5_Wantcompute
, &dev
->flags
);
2132 sh
->ops
.target
= disk_idx
;
2134 /* Careful: from this point on 'uptodate' is in the eye
2135 * of raid5_run_ops which services 'compute' operations
2136 * before writes. R5_Wantcompute flags a block that will
2137 * be R5_UPTODATE by the time it is needed for a
2138 * subsequent operation.
2141 return 1; /* uptodate + compute == disks */
2142 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2143 set_bit(R5_LOCKED
, &dev
->flags
);
2144 set_bit(R5_Wantread
, &dev
->flags
);
2146 pr_debug("Reading block %d (sync=%d)\n", disk_idx
,
2155 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2157 static void handle_stripe_fill5(struct stripe_head
*sh
,
2158 struct stripe_head_state
*s
, int disks
)
2162 /* look for blocks to read/compute, skip this if a compute
2163 * is already in flight, or if the stripe contents are in the
2164 * midst of changing due to a write
2166 if (!test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) && !sh
->check_state
&&
2167 !sh
->reconstruct_state
)
2168 for (i
= disks
; i
--; )
2169 if (fetch_block5(sh
, s
, i
, disks
))
2171 set_bit(STRIPE_HANDLE
, &sh
->state
);
2174 static void handle_stripe_fill6(struct stripe_head
*sh
,
2175 struct stripe_head_state
*s
, struct r6_state
*r6s
,
2179 for (i
= disks
; i
--; ) {
2180 struct r5dev
*dev
= &sh
->dev
[i
];
2181 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2182 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2183 (dev
->toread
|| (dev
->towrite
&&
2184 !test_bit(R5_OVERWRITE
, &dev
->flags
)) ||
2185 s
->syncing
|| s
->expanding
||
2187 (sh
->dev
[r6s
->failed_num
[0]].toread
||
2190 (sh
->dev
[r6s
->failed_num
[1]].toread
||
2192 /* we would like to get this block, possibly
2193 * by computing it, but we might not be able to
2195 if ((s
->uptodate
== disks
- 1) &&
2196 (s
->failed
&& (i
== r6s
->failed_num
[0] ||
2197 i
== r6s
->failed_num
[1]))) {
2198 pr_debug("Computing stripe %llu block %d\n",
2199 (unsigned long long)sh
->sector
, i
);
2200 compute_block_1(sh
, i
, 0);
2202 } else if ( s
->uptodate
== disks
-2 && s
->failed
>= 2 ) {
2203 /* Computing 2-failure is *very* expensive; only
2204 * do it if failed >= 2
2207 for (other
= disks
; other
--; ) {
2210 if (!test_bit(R5_UPTODATE
,
2211 &sh
->dev
[other
].flags
))
2215 pr_debug("Computing stripe %llu blocks %d,%d\n",
2216 (unsigned long long)sh
->sector
,
2218 compute_block_2(sh
, i
, other
);
2220 } else if (test_bit(R5_Insync
, &dev
->flags
)) {
2221 set_bit(R5_LOCKED
, &dev
->flags
);
2222 set_bit(R5_Wantread
, &dev
->flags
);
2224 pr_debug("Reading block %d (sync=%d)\n",
2229 set_bit(STRIPE_HANDLE
, &sh
->state
);
2233 /* handle_stripe_clean_event
2234 * any written block on an uptodate or failed drive can be returned.
2235 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2236 * never LOCKED, so we don't need to test 'failed' directly.
2238 static void handle_stripe_clean_event(raid5_conf_t
*conf
,
2239 struct stripe_head
*sh
, int disks
, struct bio
**return_bi
)
2244 for (i
= disks
; i
--; )
2245 if (sh
->dev
[i
].written
) {
2247 if (!test_bit(R5_LOCKED
, &dev
->flags
) &&
2248 test_bit(R5_UPTODATE
, &dev
->flags
)) {
2249 /* We can return any write requests */
2250 struct bio
*wbi
, *wbi2
;
2252 pr_debug("Return write for disc %d\n", i
);
2253 spin_lock_irq(&conf
->device_lock
);
2255 dev
->written
= NULL
;
2256 while (wbi
&& wbi
->bi_sector
<
2257 dev
->sector
+ STRIPE_SECTORS
) {
2258 wbi2
= r5_next_bio(wbi
, dev
->sector
);
2259 if (!raid5_dec_bi_phys_segments(wbi
)) {
2260 md_write_end(conf
->mddev
);
2261 wbi
->bi_next
= *return_bi
;
2266 if (dev
->towrite
== NULL
)
2268 spin_unlock_irq(&conf
->device_lock
);
2270 bitmap_endwrite(conf
->mddev
->bitmap
,
2273 !test_bit(STRIPE_DEGRADED
, &sh
->state
),
2278 if (test_and_clear_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2279 if (atomic_dec_and_test(&conf
->pending_full_writes
))
2280 md_wakeup_thread(conf
->mddev
->thread
);
2283 static void handle_stripe_dirtying5(raid5_conf_t
*conf
,
2284 struct stripe_head
*sh
, struct stripe_head_state
*s
, int disks
)
2286 int rmw
= 0, rcw
= 0, i
;
2287 for (i
= disks
; i
--; ) {
2288 /* would I have to read this buffer for read_modify_write */
2289 struct r5dev
*dev
= &sh
->dev
[i
];
2290 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2291 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2292 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2293 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2294 if (test_bit(R5_Insync
, &dev
->flags
))
2297 rmw
+= 2*disks
; /* cannot read it */
2299 /* Would I have to read this buffer for reconstruct_write */
2300 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) && i
!= sh
->pd_idx
&&
2301 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2302 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2303 test_bit(R5_Wantcompute
, &dev
->flags
))) {
2304 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2309 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2310 (unsigned long long)sh
->sector
, rmw
, rcw
);
2311 set_bit(STRIPE_HANDLE
, &sh
->state
);
2312 if (rmw
< rcw
&& rmw
> 0)
2313 /* prefer read-modify-write, but need to get some data */
2314 for (i
= disks
; i
--; ) {
2315 struct r5dev
*dev
= &sh
->dev
[i
];
2316 if ((dev
->towrite
|| i
== sh
->pd_idx
) &&
2317 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2318 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2319 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2320 test_bit(R5_Insync
, &dev
->flags
)) {
2322 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2323 pr_debug("Read_old block "
2324 "%d for r-m-w\n", i
);
2325 set_bit(R5_LOCKED
, &dev
->flags
);
2326 set_bit(R5_Wantread
, &dev
->flags
);
2329 set_bit(STRIPE_DELAYED
, &sh
->state
);
2330 set_bit(STRIPE_HANDLE
, &sh
->state
);
2334 if (rcw
<= rmw
&& rcw
> 0)
2335 /* want reconstruct write, but need to get some data */
2336 for (i
= disks
; i
--; ) {
2337 struct r5dev
*dev
= &sh
->dev
[i
];
2338 if (!test_bit(R5_OVERWRITE
, &dev
->flags
) &&
2340 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2341 !(test_bit(R5_UPTODATE
, &dev
->flags
) ||
2342 test_bit(R5_Wantcompute
, &dev
->flags
)) &&
2343 test_bit(R5_Insync
, &dev
->flags
)) {
2345 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2346 pr_debug("Read_old block "
2347 "%d for Reconstruct\n", i
);
2348 set_bit(R5_LOCKED
, &dev
->flags
);
2349 set_bit(R5_Wantread
, &dev
->flags
);
2352 set_bit(STRIPE_DELAYED
, &sh
->state
);
2353 set_bit(STRIPE_HANDLE
, &sh
->state
);
2357 /* now if nothing is locked, and if we have enough data,
2358 * we can start a write request
2360 /* since handle_stripe can be called at any time we need to handle the
2361 * case where a compute block operation has been submitted and then a
2362 * subsequent call wants to start a write request. raid5_run_ops only
2363 * handles the case where compute block and postxor are requested
2364 * simultaneously. If this is not the case then new writes need to be
2365 * held off until the compute completes.
2367 if ((s
->req_compute
|| !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
)) &&
2368 (s
->locked
== 0 && (rcw
== 0 || rmw
== 0) &&
2369 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)))
2370 schedule_reconstruction5(sh
, s
, rcw
== 0, 0);
2373 static void handle_stripe_dirtying6(raid5_conf_t
*conf
,
2374 struct stripe_head
*sh
, struct stripe_head_state
*s
,
2375 struct r6_state
*r6s
, int disks
)
2377 int rcw
= 0, must_compute
= 0, pd_idx
= sh
->pd_idx
, i
;
2378 int qd_idx
= sh
->qd_idx
;
2379 for (i
= disks
; i
--; ) {
2380 struct r5dev
*dev
= &sh
->dev
[i
];
2381 /* Would I have to read this buffer for reconstruct_write */
2382 if (!test_bit(R5_OVERWRITE
, &dev
->flags
)
2383 && i
!= pd_idx
&& i
!= qd_idx
2384 && (!test_bit(R5_LOCKED
, &dev
->flags
)
2386 !test_bit(R5_UPTODATE
, &dev
->flags
)) {
2387 if (test_bit(R5_Insync
, &dev
->flags
)) rcw
++;
2389 pr_debug("raid6: must_compute: "
2390 "disk %d flags=%#lx\n", i
, dev
->flags
);
2395 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2396 (unsigned long long)sh
->sector
, rcw
, must_compute
);
2397 set_bit(STRIPE_HANDLE
, &sh
->state
);
2400 /* want reconstruct write, but need to get some data */
2401 for (i
= disks
; i
--; ) {
2402 struct r5dev
*dev
= &sh
->dev
[i
];
2403 if (!test_bit(R5_OVERWRITE
, &dev
->flags
)
2404 && !(s
->failed
== 0 && (i
== pd_idx
|| i
== qd_idx
))
2405 && !test_bit(R5_LOCKED
, &dev
->flags
) &&
2406 !test_bit(R5_UPTODATE
, &dev
->flags
) &&
2407 test_bit(R5_Insync
, &dev
->flags
)) {
2409 test_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2410 pr_debug("Read_old stripe %llu "
2411 "block %d for Reconstruct\n",
2412 (unsigned long long)sh
->sector
, i
);
2413 set_bit(R5_LOCKED
, &dev
->flags
);
2414 set_bit(R5_Wantread
, &dev
->flags
);
2417 pr_debug("Request delayed stripe %llu "
2418 "block %d for Reconstruct\n",
2419 (unsigned long long)sh
->sector
, i
);
2420 set_bit(STRIPE_DELAYED
, &sh
->state
);
2421 set_bit(STRIPE_HANDLE
, &sh
->state
);
2425 /* now if nothing is locked, and if we have enough data, we can start a
2428 if (s
->locked
== 0 && rcw
== 0 &&
2429 !test_bit(STRIPE_BIT_DELAY
, &sh
->state
)) {
2430 if (must_compute
> 0) {
2431 /* We have failed blocks and need to compute them */
2432 switch (s
->failed
) {
2436 compute_block_1(sh
, r6s
->failed_num
[0], 0);
2439 compute_block_2(sh
, r6s
->failed_num
[0],
2440 r6s
->failed_num
[1]);
2442 default: /* This request should have been failed? */
2447 pr_debug("Computing parity for stripe %llu\n",
2448 (unsigned long long)sh
->sector
);
2449 compute_parity6(sh
, RECONSTRUCT_WRITE
);
2450 /* now every locked buffer is ready to be written */
2451 for (i
= disks
; i
--; )
2452 if (test_bit(R5_LOCKED
, &sh
->dev
[i
].flags
)) {
2453 pr_debug("Writing stripe %llu block %d\n",
2454 (unsigned long long)sh
->sector
, i
);
2456 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
2458 if (s
->locked
== disks
)
2459 if (!test_and_set_bit(STRIPE_FULL_WRITE
, &sh
->state
))
2460 atomic_inc(&conf
->pending_full_writes
);
2461 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2462 set_bit(STRIPE_INSYNC
, &sh
->state
);
2464 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2465 atomic_dec(&conf
->preread_active_stripes
);
2466 if (atomic_read(&conf
->preread_active_stripes
) <
2468 md_wakeup_thread(conf
->mddev
->thread
);
2473 static void handle_parity_checks5(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2474 struct stripe_head_state
*s
, int disks
)
2476 struct r5dev
*dev
= NULL
;
2478 set_bit(STRIPE_HANDLE
, &sh
->state
);
2480 switch (sh
->check_state
) {
2481 case check_state_idle
:
2482 /* start a new check operation if there are no failures */
2483 if (s
->failed
== 0) {
2484 BUG_ON(s
->uptodate
!= disks
);
2485 sh
->check_state
= check_state_run
;
2486 set_bit(STRIPE_OP_CHECK
, &s
->ops_request
);
2487 clear_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
);
2491 dev
= &sh
->dev
[s
->failed_num
];
2493 case check_state_compute_result
:
2494 sh
->check_state
= check_state_idle
;
2496 dev
= &sh
->dev
[sh
->pd_idx
];
2498 /* check that a write has not made the stripe insync */
2499 if (test_bit(STRIPE_INSYNC
, &sh
->state
))
2502 /* either failed parity check, or recovery is happening */
2503 BUG_ON(!test_bit(R5_UPTODATE
, &dev
->flags
));
2504 BUG_ON(s
->uptodate
!= disks
);
2506 set_bit(R5_LOCKED
, &dev
->flags
);
2508 set_bit(R5_Wantwrite
, &dev
->flags
);
2510 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2511 set_bit(STRIPE_INSYNC
, &sh
->state
);
2513 case check_state_run
:
2514 break; /* we will be called again upon completion */
2515 case check_state_check_result
:
2516 sh
->check_state
= check_state_idle
;
2518 /* if a failure occurred during the check operation, leave
2519 * STRIPE_INSYNC not set and let the stripe be handled again
2524 /* handle a successful check operation, if parity is correct
2525 * we are done. Otherwise update the mismatch count and repair
2526 * parity if !MD_RECOVERY_CHECK
2528 if (sh
->ops
.zero_sum_result
== 0)
2529 /* parity is correct (on disc,
2530 * not in buffer any more)
2532 set_bit(STRIPE_INSYNC
, &sh
->state
);
2534 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2535 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2536 /* don't try to repair!! */
2537 set_bit(STRIPE_INSYNC
, &sh
->state
);
2539 sh
->check_state
= check_state_compute_run
;
2540 set_bit(STRIPE_COMPUTE_RUN
, &sh
->state
);
2541 set_bit(STRIPE_OP_COMPUTE_BLK
, &s
->ops_request
);
2542 set_bit(R5_Wantcompute
,
2543 &sh
->dev
[sh
->pd_idx
].flags
);
2544 sh
->ops
.target
= sh
->pd_idx
;
2549 case check_state_compute_run
:
2552 printk(KERN_ERR
"%s: unknown check_state: %d sector: %llu\n",
2553 __func__
, sh
->check_state
,
2554 (unsigned long long) sh
->sector
);
2560 static void handle_parity_checks6(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2561 struct stripe_head_state
*s
,
2562 struct r6_state
*r6s
, struct page
*tmp_page
,
2565 int update_p
= 0, update_q
= 0;
2567 int pd_idx
= sh
->pd_idx
;
2568 int qd_idx
= sh
->qd_idx
;
2570 set_bit(STRIPE_HANDLE
, &sh
->state
);
2572 BUG_ON(s
->failed
> 2);
2573 BUG_ON(s
->uptodate
< disks
);
2574 /* Want to check and possibly repair P and Q.
2575 * However there could be one 'failed' device, in which
2576 * case we can only check one of them, possibly using the
2577 * other to generate missing data
2580 /* If !tmp_page, we cannot do the calculations,
2581 * but as we have set STRIPE_HANDLE, we will soon be called
2582 * by stripe_handle with a tmp_page - just wait until then.
2585 if (s
->failed
== r6s
->q_failed
) {
2586 /* The only possible failed device holds 'Q', so it
2587 * makes sense to check P (If anything else were failed,
2588 * we would have used P to recreate it).
2590 compute_block_1(sh
, pd_idx
, 1);
2591 if (!page_is_zero(sh
->dev
[pd_idx
].page
)) {
2592 compute_block_1(sh
, pd_idx
, 0);
2596 if (!r6s
->q_failed
&& s
->failed
< 2) {
2597 /* q is not failed, and we didn't use it to generate
2598 * anything, so it makes sense to check it
2600 memcpy(page_address(tmp_page
),
2601 page_address(sh
->dev
[qd_idx
].page
),
2603 compute_parity6(sh
, UPDATE_PARITY
);
2604 if (memcmp(page_address(tmp_page
),
2605 page_address(sh
->dev
[qd_idx
].page
),
2606 STRIPE_SIZE
) != 0) {
2607 clear_bit(STRIPE_INSYNC
, &sh
->state
);
2611 if (update_p
|| update_q
) {
2612 conf
->mddev
->resync_mismatches
+= STRIPE_SECTORS
;
2613 if (test_bit(MD_RECOVERY_CHECK
, &conf
->mddev
->recovery
))
2614 /* don't try to repair!! */
2615 update_p
= update_q
= 0;
2618 /* now write out any block on a failed drive,
2619 * or P or Q if they need it
2622 if (s
->failed
== 2) {
2623 dev
= &sh
->dev
[r6s
->failed_num
[1]];
2625 set_bit(R5_LOCKED
, &dev
->flags
);
2626 set_bit(R5_Wantwrite
, &dev
->flags
);
2628 if (s
->failed
>= 1) {
2629 dev
= &sh
->dev
[r6s
->failed_num
[0]];
2631 set_bit(R5_LOCKED
, &dev
->flags
);
2632 set_bit(R5_Wantwrite
, &dev
->flags
);
2636 dev
= &sh
->dev
[pd_idx
];
2638 set_bit(R5_LOCKED
, &dev
->flags
);
2639 set_bit(R5_Wantwrite
, &dev
->flags
);
2642 dev
= &sh
->dev
[qd_idx
];
2644 set_bit(R5_LOCKED
, &dev
->flags
);
2645 set_bit(R5_Wantwrite
, &dev
->flags
);
2647 clear_bit(STRIPE_DEGRADED
, &sh
->state
);
2649 set_bit(STRIPE_INSYNC
, &sh
->state
);
2653 static void handle_stripe_expansion(raid5_conf_t
*conf
, struct stripe_head
*sh
,
2654 struct r6_state
*r6s
)
2658 /* We have read all the blocks in this stripe and now we need to
2659 * copy some of them into a target stripe for expand.
2661 struct dma_async_tx_descriptor
*tx
= NULL
;
2662 clear_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2663 for (i
= 0; i
< sh
->disks
; i
++)
2664 if (i
!= sh
->pd_idx
&& i
!= sh
->qd_idx
) {
2666 struct stripe_head
*sh2
;
2668 sector_t bn
= compute_blocknr(sh
, i
, 1);
2669 sector_t s
= raid5_compute_sector(conf
, bn
, 0,
2671 sh2
= get_active_stripe(conf
, s
, 0, 1);
2673 /* so far only the early blocks of this stripe
2674 * have been requested. When later blocks
2675 * get requested, we will try again
2678 if (!test_bit(STRIPE_EXPANDING
, &sh2
->state
) ||
2679 test_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
)) {
2680 /* must have already done this block */
2681 release_stripe(sh2
);
2685 /* place all the copies on one channel */
2686 tx
= async_memcpy(sh2
->dev
[dd_idx
].page
,
2687 sh
->dev
[i
].page
, 0, 0, STRIPE_SIZE
,
2690 set_bit(R5_Expanded
, &sh2
->dev
[dd_idx
].flags
);
2691 set_bit(R5_UPTODATE
, &sh2
->dev
[dd_idx
].flags
);
2692 for (j
= 0; j
< conf
->raid_disks
; j
++)
2693 if (j
!= sh2
->pd_idx
&&
2694 (!r6s
|| j
!= sh2
->qd_idx
) &&
2695 !test_bit(R5_Expanded
, &sh2
->dev
[j
].flags
))
2697 if (j
== conf
->raid_disks
) {
2698 set_bit(STRIPE_EXPAND_READY
, &sh2
->state
);
2699 set_bit(STRIPE_HANDLE
, &sh2
->state
);
2701 release_stripe(sh2
);
2704 /* done submitting copies, wait for them to complete */
2707 dma_wait_for_async_tx(tx
);
2713 * handle_stripe - do things to a stripe.
2715 * We lock the stripe and then examine the state of various bits
2716 * to see what needs to be done.
2718 * return some read request which now have data
2719 * return some write requests which are safely on disc
2720 * schedule a read on some buffers
2721 * schedule a write of some buffers
2722 * return confirmation of parity correctness
2724 * buffers are taken off read_list or write_list, and bh_cache buffers
2725 * get BH_Lock set before the stripe lock is released.
2729 static bool handle_stripe5(struct stripe_head
*sh
)
2731 raid5_conf_t
*conf
= sh
->raid_conf
;
2732 int disks
= sh
->disks
, i
;
2733 struct bio
*return_bi
= NULL
;
2734 struct stripe_head_state s
;
2736 mdk_rdev_t
*blocked_rdev
= NULL
;
2739 memset(&s
, 0, sizeof(s
));
2740 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2741 "reconstruct:%d\n", (unsigned long long)sh
->sector
, sh
->state
,
2742 atomic_read(&sh
->count
), sh
->pd_idx
, sh
->check_state
,
2743 sh
->reconstruct_state
);
2745 spin_lock(&sh
->lock
);
2746 clear_bit(STRIPE_HANDLE
, &sh
->state
);
2747 clear_bit(STRIPE_DELAYED
, &sh
->state
);
2749 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
2750 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
2751 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
2753 /* Now to look around and see what can be done */
2755 for (i
=disks
; i
--; ) {
2757 struct r5dev
*dev
= &sh
->dev
[i
];
2758 clear_bit(R5_Insync
, &dev
->flags
);
2760 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2761 "written %p\n", i
, dev
->flags
, dev
->toread
, dev
->read
,
2762 dev
->towrite
, dev
->written
);
2764 /* maybe we can request a biofill operation
2766 * new wantfill requests are only permitted while
2767 * ops_complete_biofill is guaranteed to be inactive
2769 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
&&
2770 !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
))
2771 set_bit(R5_Wantfill
, &dev
->flags
);
2773 /* now count some things */
2774 if (test_bit(R5_LOCKED
, &dev
->flags
)) s
.locked
++;
2775 if (test_bit(R5_UPTODATE
, &dev
->flags
)) s
.uptodate
++;
2776 if (test_bit(R5_Wantcompute
, &dev
->flags
)) s
.compute
++;
2778 if (test_bit(R5_Wantfill
, &dev
->flags
))
2780 else if (dev
->toread
)
2784 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
2789 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
2790 if (blocked_rdev
== NULL
&&
2791 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
2792 blocked_rdev
= rdev
;
2793 atomic_inc(&rdev
->nr_pending
);
2795 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)) {
2796 /* The ReadError flag will just be confusing now */
2797 clear_bit(R5_ReadError
, &dev
->flags
);
2798 clear_bit(R5_ReWrite
, &dev
->flags
);
2800 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)
2801 || test_bit(R5_ReadError
, &dev
->flags
)) {
2805 set_bit(R5_Insync
, &dev
->flags
);
2809 if (unlikely(blocked_rdev
)) {
2810 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
2811 s
.to_write
|| s
.written
) {
2812 set_bit(STRIPE_HANDLE
, &sh
->state
);
2815 /* There is nothing for the blocked_rdev to block */
2816 rdev_dec_pending(blocked_rdev
, conf
->mddev
);
2817 blocked_rdev
= NULL
;
2820 if (s
.to_fill
&& !test_bit(STRIPE_BIOFILL_RUN
, &sh
->state
)) {
2821 set_bit(STRIPE_OP_BIOFILL
, &s
.ops_request
);
2822 set_bit(STRIPE_BIOFILL_RUN
, &sh
->state
);
2825 pr_debug("locked=%d uptodate=%d to_read=%d"
2826 " to_write=%d failed=%d failed_num=%d\n",
2827 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
,
2828 s
.failed
, s
.failed_num
);
2829 /* check if the array has lost two devices and, if so, some requests might
2832 if (s
.failed
> 1 && s
.to_read
+s
.to_write
+s
.written
)
2833 handle_failed_stripe(conf
, sh
, &s
, disks
, &return_bi
);
2834 if (s
.failed
> 1 && s
.syncing
) {
2835 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
2836 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2840 /* might be able to return some write requests if the parity block
2841 * is safe, or on a failed drive
2843 dev
= &sh
->dev
[sh
->pd_idx
];
2845 ((test_bit(R5_Insync
, &dev
->flags
) &&
2846 !test_bit(R5_LOCKED
, &dev
->flags
) &&
2847 test_bit(R5_UPTODATE
, &dev
->flags
)) ||
2848 (s
.failed
== 1 && s
.failed_num
== sh
->pd_idx
)))
2849 handle_stripe_clean_event(conf
, sh
, disks
, &return_bi
);
2851 /* Now we might consider reading some blocks, either to check/generate
2852 * parity, or to satisfy requests
2853 * or to load a block that is being partially written.
2855 if (s
.to_read
|| s
.non_overwrite
||
2856 (s
.syncing
&& (s
.uptodate
+ s
.compute
< disks
)) || s
.expanding
)
2857 handle_stripe_fill5(sh
, &s
, disks
);
2859 /* Now we check to see if any write operations have recently
2863 if (sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
)
2865 if (sh
->reconstruct_state
== reconstruct_state_drain_result
||
2866 sh
->reconstruct_state
== reconstruct_state_prexor_drain_result
) {
2867 sh
->reconstruct_state
= reconstruct_state_idle
;
2869 /* All the 'written' buffers and the parity block are ready to
2870 * be written back to disk
2872 BUG_ON(!test_bit(R5_UPTODATE
, &sh
->dev
[sh
->pd_idx
].flags
));
2873 for (i
= disks
; i
--; ) {
2875 if (test_bit(R5_LOCKED
, &dev
->flags
) &&
2876 (i
== sh
->pd_idx
|| dev
->written
)) {
2877 pr_debug("Writing block %d\n", i
);
2878 set_bit(R5_Wantwrite
, &dev
->flags
);
2881 if (!test_bit(R5_Insync
, &dev
->flags
) ||
2882 (i
== sh
->pd_idx
&& s
.failed
== 0))
2883 set_bit(STRIPE_INSYNC
, &sh
->state
);
2886 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
)) {
2887 atomic_dec(&conf
->preread_active_stripes
);
2888 if (atomic_read(&conf
->preread_active_stripes
) <
2890 md_wakeup_thread(conf
->mddev
->thread
);
2894 /* Now to consider new write requests and what else, if anything
2895 * should be read. We do not handle new writes when:
2896 * 1/ A 'write' operation (copy+xor) is already in flight.
2897 * 2/ A 'check' operation is in flight, as it may clobber the parity
2900 if (s
.to_write
&& !sh
->reconstruct_state
&& !sh
->check_state
)
2901 handle_stripe_dirtying5(conf
, sh
, &s
, disks
);
2903 /* maybe we need to check and possibly fix the parity for this stripe
2904 * Any reads will already have been scheduled, so we just see if enough
2905 * data is available. The parity check is held off while parity
2906 * dependent operations are in flight.
2908 if (sh
->check_state
||
2909 (s
.syncing
&& s
.locked
== 0 &&
2910 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
) &&
2911 !test_bit(STRIPE_INSYNC
, &sh
->state
)))
2912 handle_parity_checks5(conf
, sh
, &s
, disks
);
2914 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
2915 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,1);
2916 clear_bit(STRIPE_SYNCING
, &sh
->state
);
2919 /* If the failed drive is just a ReadError, then we might need to progress
2920 * the repair/check process
2922 if (s
.failed
== 1 && !conf
->mddev
->ro
&&
2923 test_bit(R5_ReadError
, &sh
->dev
[s
.failed_num
].flags
)
2924 && !test_bit(R5_LOCKED
, &sh
->dev
[s
.failed_num
].flags
)
2925 && test_bit(R5_UPTODATE
, &sh
->dev
[s
.failed_num
].flags
)
2927 dev
= &sh
->dev
[s
.failed_num
];
2928 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
2929 set_bit(R5_Wantwrite
, &dev
->flags
);
2930 set_bit(R5_ReWrite
, &dev
->flags
);
2931 set_bit(R5_LOCKED
, &dev
->flags
);
2934 /* let's read it back */
2935 set_bit(R5_Wantread
, &dev
->flags
);
2936 set_bit(R5_LOCKED
, &dev
->flags
);
2941 /* Finish reconstruct operations initiated by the expansion process */
2942 if (sh
->reconstruct_state
== reconstruct_state_result
) {
2943 struct stripe_head
*sh2
2944 = get_active_stripe(conf
, sh
->sector
, 1, 1);
2945 if (sh2
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh2
->state
)) {
2946 /* sh cannot be written until sh2 has been read.
2947 * so arrange for sh to be delayed a little
2949 set_bit(STRIPE_DELAYED
, &sh
->state
);
2950 set_bit(STRIPE_HANDLE
, &sh
->state
);
2951 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
2953 atomic_inc(&conf
->preread_active_stripes
);
2954 release_stripe(sh2
);
2958 release_stripe(sh2
);
2960 sh
->reconstruct_state
= reconstruct_state_idle
;
2961 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
2962 for (i
= conf
->raid_disks
; i
--; ) {
2963 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
2964 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
2969 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
) &&
2970 !sh
->reconstruct_state
) {
2971 /* Need to write out all blocks after computing parity */
2972 sh
->disks
= conf
->raid_disks
;
2973 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
2974 schedule_reconstruction5(sh
, &s
, 1, 1);
2975 } else if (s
.expanded
&& !sh
->reconstruct_state
&& s
.locked
== 0) {
2976 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
2977 atomic_dec(&conf
->reshape_stripes
);
2978 wake_up(&conf
->wait_for_overlap
);
2979 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
2982 if (s
.expanding
&& s
.locked
== 0 &&
2983 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
2984 handle_stripe_expansion(conf
, sh
, NULL
);
2987 spin_unlock(&sh
->lock
);
2989 /* wait for this device to become unblocked */
2990 if (unlikely(blocked_rdev
))
2991 md_wait_for_blocked_rdev(blocked_rdev
, conf
->mddev
);
2994 raid5_run_ops(sh
, s
.ops_request
);
2998 return_io(return_bi
);
3000 return blocked_rdev
== NULL
;
3003 static bool handle_stripe6(struct stripe_head
*sh
, struct page
*tmp_page
)
3005 raid5_conf_t
*conf
= sh
->raid_conf
;
3006 int disks
= sh
->disks
;
3007 struct bio
*return_bi
= NULL
;
3008 int i
, pd_idx
= sh
->pd_idx
, qd_idx
= sh
->qd_idx
;
3009 struct stripe_head_state s
;
3010 struct r6_state r6s
;
3011 struct r5dev
*dev
, *pdev
, *qdev
;
3012 mdk_rdev_t
*blocked_rdev
= NULL
;
3014 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3015 "pd_idx=%d, qd_idx=%d\n",
3016 (unsigned long long)sh
->sector
, sh
->state
,
3017 atomic_read(&sh
->count
), pd_idx
, qd_idx
);
3018 memset(&s
, 0, sizeof(s
));
3020 spin_lock(&sh
->lock
);
3021 clear_bit(STRIPE_HANDLE
, &sh
->state
);
3022 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3024 s
.syncing
= test_bit(STRIPE_SYNCING
, &sh
->state
);
3025 s
.expanding
= test_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3026 s
.expanded
= test_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3027 /* Now to look around and see what can be done */
3030 for (i
=disks
; i
--; ) {
3033 clear_bit(R5_Insync
, &dev
->flags
);
3035 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3036 i
, dev
->flags
, dev
->toread
, dev
->towrite
, dev
->written
);
3037 /* maybe we can reply to a read */
3038 if (test_bit(R5_UPTODATE
, &dev
->flags
) && dev
->toread
) {
3039 struct bio
*rbi
, *rbi2
;
3040 pr_debug("Return read for disc %d\n", i
);
3041 spin_lock_irq(&conf
->device_lock
);
3044 if (test_and_clear_bit(R5_Overlap
, &dev
->flags
))
3045 wake_up(&conf
->wait_for_overlap
);
3046 spin_unlock_irq(&conf
->device_lock
);
3047 while (rbi
&& rbi
->bi_sector
< dev
->sector
+ STRIPE_SECTORS
) {
3048 copy_data(0, rbi
, dev
->page
, dev
->sector
);
3049 rbi2
= r5_next_bio(rbi
, dev
->sector
);
3050 spin_lock_irq(&conf
->device_lock
);
3051 if (!raid5_dec_bi_phys_segments(rbi
)) {
3052 rbi
->bi_next
= return_bi
;
3055 spin_unlock_irq(&conf
->device_lock
);
3060 /* now count some things */
3061 if (test_bit(R5_LOCKED
, &dev
->flags
)) s
.locked
++;
3062 if (test_bit(R5_UPTODATE
, &dev
->flags
)) s
.uptodate
++;
3069 if (!test_bit(R5_OVERWRITE
, &dev
->flags
))
3074 rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3075 if (blocked_rdev
== NULL
&&
3076 rdev
&& unlikely(test_bit(Blocked
, &rdev
->flags
))) {
3077 blocked_rdev
= rdev
;
3078 atomic_inc(&rdev
->nr_pending
);
3080 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)) {
3081 /* The ReadError flag will just be confusing now */
3082 clear_bit(R5_ReadError
, &dev
->flags
);
3083 clear_bit(R5_ReWrite
, &dev
->flags
);
3085 if (!rdev
|| !test_bit(In_sync
, &rdev
->flags
)
3086 || test_bit(R5_ReadError
, &dev
->flags
)) {
3088 r6s
.failed_num
[s
.failed
] = i
;
3091 set_bit(R5_Insync
, &dev
->flags
);
3095 if (unlikely(blocked_rdev
)) {
3096 if (s
.syncing
|| s
.expanding
|| s
.expanded
||
3097 s
.to_write
|| s
.written
) {
3098 set_bit(STRIPE_HANDLE
, &sh
->state
);
3101 /* There is nothing for the blocked_rdev to block */
3102 rdev_dec_pending(blocked_rdev
, conf
->mddev
);
3103 blocked_rdev
= NULL
;
3106 pr_debug("locked=%d uptodate=%d to_read=%d"
3107 " to_write=%d failed=%d failed_num=%d,%d\n",
3108 s
.locked
, s
.uptodate
, s
.to_read
, s
.to_write
, s
.failed
,
3109 r6s
.failed_num
[0], r6s
.failed_num
[1]);
3110 /* check if the array has lost >2 devices and, if so, some requests
3111 * might need to be failed
3113 if (s
.failed
> 2 && s
.to_read
+s
.to_write
+s
.written
)
3114 handle_failed_stripe(conf
, sh
, &s
, disks
, &return_bi
);
3115 if (s
.failed
> 2 && s
.syncing
) {
3116 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,0);
3117 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3122 * might be able to return some write requests if the parity blocks
3123 * are safe, or on a failed drive
3125 pdev
= &sh
->dev
[pd_idx
];
3126 r6s
.p_failed
= (s
.failed
>= 1 && r6s
.failed_num
[0] == pd_idx
)
3127 || (s
.failed
>= 2 && r6s
.failed_num
[1] == pd_idx
);
3128 qdev
= &sh
->dev
[qd_idx
];
3129 r6s
.q_failed
= (s
.failed
>= 1 && r6s
.failed_num
[0] == qd_idx
)
3130 || (s
.failed
>= 2 && r6s
.failed_num
[1] == qd_idx
);
3133 ( r6s
.p_failed
|| ((test_bit(R5_Insync
, &pdev
->flags
)
3134 && !test_bit(R5_LOCKED
, &pdev
->flags
)
3135 && test_bit(R5_UPTODATE
, &pdev
->flags
)))) &&
3136 ( r6s
.q_failed
|| ((test_bit(R5_Insync
, &qdev
->flags
)
3137 && !test_bit(R5_LOCKED
, &qdev
->flags
)
3138 && test_bit(R5_UPTODATE
, &qdev
->flags
)))))
3139 handle_stripe_clean_event(conf
, sh
, disks
, &return_bi
);
3141 /* Now we might consider reading some blocks, either to check/generate
3142 * parity, or to satisfy requests
3143 * or to load a block that is being partially written.
3145 if (s
.to_read
|| s
.non_overwrite
|| (s
.to_write
&& s
.failed
) ||
3146 (s
.syncing
&& (s
.uptodate
< disks
)) || s
.expanding
)
3147 handle_stripe_fill6(sh
, &s
, &r6s
, disks
);
3149 /* now to consider writing and what else, if anything should be read */
3151 handle_stripe_dirtying6(conf
, sh
, &s
, &r6s
, disks
);
3153 /* maybe we need to check and possibly fix the parity for this stripe
3154 * Any reads will already have been scheduled, so we just see if enough
3157 if (s
.syncing
&& s
.locked
== 0 && !test_bit(STRIPE_INSYNC
, &sh
->state
))
3158 handle_parity_checks6(conf
, sh
, &s
, &r6s
, tmp_page
, disks
);
3160 if (s
.syncing
&& s
.locked
== 0 && test_bit(STRIPE_INSYNC
, &sh
->state
)) {
3161 md_done_sync(conf
->mddev
, STRIPE_SECTORS
,1);
3162 clear_bit(STRIPE_SYNCING
, &sh
->state
);
3165 /* If the failed drives are just a ReadError, then we might need
3166 * to progress the repair/check process
3168 if (s
.failed
<= 2 && !conf
->mddev
->ro
)
3169 for (i
= 0; i
< s
.failed
; i
++) {
3170 dev
= &sh
->dev
[r6s
.failed_num
[i
]];
3171 if (test_bit(R5_ReadError
, &dev
->flags
)
3172 && !test_bit(R5_LOCKED
, &dev
->flags
)
3173 && test_bit(R5_UPTODATE
, &dev
->flags
)
3175 if (!test_bit(R5_ReWrite
, &dev
->flags
)) {
3176 set_bit(R5_Wantwrite
, &dev
->flags
);
3177 set_bit(R5_ReWrite
, &dev
->flags
);
3178 set_bit(R5_LOCKED
, &dev
->flags
);
3180 /* let's read it back */
3181 set_bit(R5_Wantread
, &dev
->flags
);
3182 set_bit(R5_LOCKED
, &dev
->flags
);
3187 if (s
.expanded
&& test_bit(STRIPE_EXPANDING
, &sh
->state
)) {
3188 struct stripe_head
*sh2
3189 = get_active_stripe(conf
, sh
->sector
, 1, 1);
3190 if (sh2
&& test_bit(STRIPE_EXPAND_SOURCE
, &sh2
->state
)) {
3191 /* sh cannot be written until sh2 has been read.
3192 * so arrange for sh to be delayed a little
3194 set_bit(STRIPE_DELAYED
, &sh
->state
);
3195 set_bit(STRIPE_HANDLE
, &sh
->state
);
3196 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
,
3198 atomic_inc(&conf
->preread_active_stripes
);
3199 release_stripe(sh2
);
3203 release_stripe(sh2
);
3205 /* Need to write out all blocks after computing P&Q */
3206 sh
->disks
= conf
->raid_disks
;
3207 stripe_set_idx(sh
->sector
, conf
, 0, sh
);
3208 compute_parity6(sh
, RECONSTRUCT_WRITE
);
3209 for (i
= conf
->raid_disks
; i
-- ; ) {
3210 set_bit(R5_LOCKED
, &sh
->dev
[i
].flags
);
3212 set_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
);
3214 clear_bit(STRIPE_EXPANDING
, &sh
->state
);
3215 } else if (s
.expanded
) {
3216 clear_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3217 atomic_dec(&conf
->reshape_stripes
);
3218 wake_up(&conf
->wait_for_overlap
);
3219 md_done_sync(conf
->mddev
, STRIPE_SECTORS
, 1);
3222 if (s
.expanding
&& s
.locked
== 0 &&
3223 !test_bit(STRIPE_COMPUTE_RUN
, &sh
->state
))
3224 handle_stripe_expansion(conf
, sh
, &r6s
);
3227 spin_unlock(&sh
->lock
);
3229 /* wait for this device to become unblocked */
3230 if (unlikely(blocked_rdev
))
3231 md_wait_for_blocked_rdev(blocked_rdev
, conf
->mddev
);
3235 return_io(return_bi
);
3237 return blocked_rdev
== NULL
;
3240 /* returns true if the stripe was handled */
3241 static bool handle_stripe(struct stripe_head
*sh
, struct page
*tmp_page
)
3243 if (sh
->raid_conf
->level
== 6)
3244 return handle_stripe6(sh
, tmp_page
);
3246 return handle_stripe5(sh
);
3251 static void raid5_activate_delayed(raid5_conf_t
*conf
)
3253 if (atomic_read(&conf
->preread_active_stripes
) < IO_THRESHOLD
) {
3254 while (!list_empty(&conf
->delayed_list
)) {
3255 struct list_head
*l
= conf
->delayed_list
.next
;
3256 struct stripe_head
*sh
;
3257 sh
= list_entry(l
, struct stripe_head
, lru
);
3259 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3260 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE
, &sh
->state
))
3261 atomic_inc(&conf
->preread_active_stripes
);
3262 list_add_tail(&sh
->lru
, &conf
->hold_list
);
3265 blk_plug_device(conf
->mddev
->queue
);
3268 static void activate_bit_delay(raid5_conf_t
*conf
)
3270 /* device_lock is held */
3271 struct list_head head
;
3272 list_add(&head
, &conf
->bitmap_list
);
3273 list_del_init(&conf
->bitmap_list
);
3274 while (!list_empty(&head
)) {
3275 struct stripe_head
*sh
= list_entry(head
.next
, struct stripe_head
, lru
);
3276 list_del_init(&sh
->lru
);
3277 atomic_inc(&sh
->count
);
3278 __release_stripe(conf
, sh
);
3282 static void unplug_slaves(mddev_t
*mddev
)
3284 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3288 for (i
=0; i
<mddev
->raid_disks
; i
++) {
3289 mdk_rdev_t
*rdev
= rcu_dereference(conf
->disks
[i
].rdev
);
3290 if (rdev
&& !test_bit(Faulty
, &rdev
->flags
) && atomic_read(&rdev
->nr_pending
)) {
3291 struct request_queue
*r_queue
= bdev_get_queue(rdev
->bdev
);
3293 atomic_inc(&rdev
->nr_pending
);
3296 blk_unplug(r_queue
);
3298 rdev_dec_pending(rdev
, mddev
);
3305 static void raid5_unplug_device(struct request_queue
*q
)
3307 mddev_t
*mddev
= q
->queuedata
;
3308 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3309 unsigned long flags
;
3311 spin_lock_irqsave(&conf
->device_lock
, flags
);
3313 if (blk_remove_plug(q
)) {
3315 raid5_activate_delayed(conf
);
3317 md_wakeup_thread(mddev
->thread
);
3319 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3321 unplug_slaves(mddev
);
3324 static int raid5_congested(void *data
, int bits
)
3326 mddev_t
*mddev
= data
;
3327 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3329 /* No difference between reads and writes. Just check
3330 * how busy the stripe_cache is
3332 if (conf
->inactive_blocked
)
3336 if (list_empty_careful(&conf
->inactive_list
))
3342 /* We want read requests to align with chunks where possible,
3343 * but write requests don't need to.
3345 static int raid5_mergeable_bvec(struct request_queue
*q
,
3346 struct bvec_merge_data
*bvm
,
3347 struct bio_vec
*biovec
)
3349 mddev_t
*mddev
= q
->queuedata
;
3350 sector_t sector
= bvm
->bi_sector
+ get_start_sect(bvm
->bi_bdev
);
3352 unsigned int chunk_sectors
= mddev
->chunk_size
>> 9;
3353 unsigned int bio_sectors
= bvm
->bi_size
>> 9;
3355 if ((bvm
->bi_rw
& 1) == WRITE
)
3356 return biovec
->bv_len
; /* always allow writes to be mergeable */
3358 if (mddev
->new_chunk
< mddev
->chunk_size
)
3359 chunk_sectors
= mddev
->new_chunk
>> 9;
3360 max
= (chunk_sectors
- ((sector
& (chunk_sectors
- 1)) + bio_sectors
)) << 9;
3361 if (max
< 0) max
= 0;
3362 if (max
<= biovec
->bv_len
&& bio_sectors
== 0)
3363 return biovec
->bv_len
;
3369 static int in_chunk_boundary(mddev_t
*mddev
, struct bio
*bio
)
3371 sector_t sector
= bio
->bi_sector
+ get_start_sect(bio
->bi_bdev
);
3372 unsigned int chunk_sectors
= mddev
->chunk_size
>> 9;
3373 unsigned int bio_sectors
= bio
->bi_size
>> 9;
3375 if (mddev
->new_chunk
< mddev
->chunk_size
)
3376 chunk_sectors
= mddev
->new_chunk
>> 9;
3377 return chunk_sectors
>=
3378 ((sector
& (chunk_sectors
- 1)) + bio_sectors
);
3382 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3383 * later sampled by raid5d.
3385 static void add_bio_to_retry(struct bio
*bi
,raid5_conf_t
*conf
)
3387 unsigned long flags
;
3389 spin_lock_irqsave(&conf
->device_lock
, flags
);
3391 bi
->bi_next
= conf
->retry_read_aligned_list
;
3392 conf
->retry_read_aligned_list
= bi
;
3394 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
3395 md_wakeup_thread(conf
->mddev
->thread
);
3399 static struct bio
*remove_bio_from_retry(raid5_conf_t
*conf
)
3403 bi
= conf
->retry_read_aligned
;
3405 conf
->retry_read_aligned
= NULL
;
3408 bi
= conf
->retry_read_aligned_list
;
3410 conf
->retry_read_aligned_list
= bi
->bi_next
;
3413 * this sets the active strip count to 1 and the processed
3414 * strip count to zero (upper 8 bits)
3416 bi
->bi_phys_segments
= 1; /* biased count of active stripes */
3424 * The "raid5_align_endio" should check if the read succeeded and if it
3425 * did, call bio_endio on the original bio (having bio_put the new bio
3427 * If the read failed..
3429 static void raid5_align_endio(struct bio
*bi
, int error
)
3431 struct bio
* raid_bi
= bi
->bi_private
;
3434 int uptodate
= test_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3439 mddev
= raid_bi
->bi_bdev
->bd_disk
->queue
->queuedata
;
3440 conf
= mddev_to_conf(mddev
);
3441 rdev
= (void*)raid_bi
->bi_next
;
3442 raid_bi
->bi_next
= NULL
;
3444 rdev_dec_pending(rdev
, conf
->mddev
);
3446 if (!error
&& uptodate
) {
3447 bio_endio(raid_bi
, 0);
3448 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
3449 wake_up(&conf
->wait_for_stripe
);
3454 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3456 add_bio_to_retry(raid_bi
, conf
);
3459 static int bio_fits_rdev(struct bio
*bi
)
3461 struct request_queue
*q
= bdev_get_queue(bi
->bi_bdev
);
3463 if ((bi
->bi_size
>>9) > q
->max_sectors
)
3465 blk_recount_segments(q
, bi
);
3466 if (bi
->bi_phys_segments
> q
->max_phys_segments
)
3469 if (q
->merge_bvec_fn
)
3470 /* it's too hard to apply the merge_bvec_fn at this stage,
3479 static int chunk_aligned_read(struct request_queue
*q
, struct bio
* raid_bio
)
3481 mddev_t
*mddev
= q
->queuedata
;
3482 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3483 unsigned int dd_idx
;
3484 struct bio
* align_bi
;
3487 if (!in_chunk_boundary(mddev
, raid_bio
)) {
3488 pr_debug("chunk_aligned_read : non aligned\n");
3492 * use bio_clone to make a copy of the bio
3494 align_bi
= bio_clone(raid_bio
, GFP_NOIO
);
3498 * set bi_end_io to a new function, and set bi_private to the
3501 align_bi
->bi_end_io
= raid5_align_endio
;
3502 align_bi
->bi_private
= raid_bio
;
3506 align_bi
->bi_sector
= raid5_compute_sector(conf
, raid_bio
->bi_sector
,
3511 rdev
= rcu_dereference(conf
->disks
[dd_idx
].rdev
);
3512 if (rdev
&& test_bit(In_sync
, &rdev
->flags
)) {
3513 atomic_inc(&rdev
->nr_pending
);
3515 raid_bio
->bi_next
= (void*)rdev
;
3516 align_bi
->bi_bdev
= rdev
->bdev
;
3517 align_bi
->bi_flags
&= ~(1 << BIO_SEG_VALID
);
3518 align_bi
->bi_sector
+= rdev
->data_offset
;
3520 if (!bio_fits_rdev(align_bi
)) {
3521 /* too big in some way */
3523 rdev_dec_pending(rdev
, mddev
);
3527 spin_lock_irq(&conf
->device_lock
);
3528 wait_event_lock_irq(conf
->wait_for_stripe
,
3530 conf
->device_lock
, /* nothing */);
3531 atomic_inc(&conf
->active_aligned_reads
);
3532 spin_unlock_irq(&conf
->device_lock
);
3534 generic_make_request(align_bi
);
3543 /* __get_priority_stripe - get the next stripe to process
3545 * Full stripe writes are allowed to pass preread active stripes up until
3546 * the bypass_threshold is exceeded. In general the bypass_count
3547 * increments when the handle_list is handled before the hold_list; however, it
3548 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3549 * stripe with in flight i/o. The bypass_count will be reset when the
3550 * head of the hold_list has changed, i.e. the head was promoted to the
3553 static struct stripe_head
*__get_priority_stripe(raid5_conf_t
*conf
)
3555 struct stripe_head
*sh
;
3557 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3559 list_empty(&conf
->handle_list
) ? "empty" : "busy",
3560 list_empty(&conf
->hold_list
) ? "empty" : "busy",
3561 atomic_read(&conf
->pending_full_writes
), conf
->bypass_count
);
3563 if (!list_empty(&conf
->handle_list
)) {
3564 sh
= list_entry(conf
->handle_list
.next
, typeof(*sh
), lru
);
3566 if (list_empty(&conf
->hold_list
))
3567 conf
->bypass_count
= 0;
3568 else if (!test_bit(STRIPE_IO_STARTED
, &sh
->state
)) {
3569 if (conf
->hold_list
.next
== conf
->last_hold
)
3570 conf
->bypass_count
++;
3572 conf
->last_hold
= conf
->hold_list
.next
;
3573 conf
->bypass_count
-= conf
->bypass_threshold
;
3574 if (conf
->bypass_count
< 0)
3575 conf
->bypass_count
= 0;
3578 } else if (!list_empty(&conf
->hold_list
) &&
3579 ((conf
->bypass_threshold
&&
3580 conf
->bypass_count
> conf
->bypass_threshold
) ||
3581 atomic_read(&conf
->pending_full_writes
) == 0)) {
3582 sh
= list_entry(conf
->hold_list
.next
,
3584 conf
->bypass_count
-= conf
->bypass_threshold
;
3585 if (conf
->bypass_count
< 0)
3586 conf
->bypass_count
= 0;
3590 list_del_init(&sh
->lru
);
3591 atomic_inc(&sh
->count
);
3592 BUG_ON(atomic_read(&sh
->count
) != 1);
3596 static int make_request(struct request_queue
*q
, struct bio
* bi
)
3598 mddev_t
*mddev
= q
->queuedata
;
3599 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
3601 sector_t new_sector
;
3602 sector_t logical_sector
, last_sector
;
3603 struct stripe_head
*sh
;
3604 const int rw
= bio_data_dir(bi
);
3607 if (unlikely(bio_barrier(bi
))) {
3608 bio_endio(bi
, -EOPNOTSUPP
);
3612 md_write_start(mddev
, bi
);
3614 cpu
= part_stat_lock();
3615 part_stat_inc(cpu
, &mddev
->gendisk
->part0
, ios
[rw
]);
3616 part_stat_add(cpu
, &mddev
->gendisk
->part0
, sectors
[rw
],
3621 mddev
->reshape_position
== MaxSector
&&
3622 chunk_aligned_read(q
,bi
))
3625 logical_sector
= bi
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
3626 last_sector
= bi
->bi_sector
+ (bi
->bi_size
>>9);
3628 bi
->bi_phys_segments
= 1; /* over-loaded to count active stripes */
3630 for (;logical_sector
< last_sector
; logical_sector
+= STRIPE_SECTORS
) {
3632 int disks
, data_disks
;
3637 disks
= conf
->raid_disks
;
3638 prepare_to_wait(&conf
->wait_for_overlap
, &w
, TASK_UNINTERRUPTIBLE
);
3639 if (unlikely(conf
->reshape_progress
!= MaxSector
)) {
3640 /* spinlock is needed as reshape_progress may be
3641 * 64bit on a 32bit platform, and so it might be
3642 * possible to see a half-updated value
3643 * Ofcourse reshape_progress could change after
3644 * the lock is dropped, so once we get a reference
3645 * to the stripe that we think it is, we will have
3648 spin_lock_irq(&conf
->device_lock
);
3649 if (mddev
->delta_disks
< 0
3650 ? logical_sector
< conf
->reshape_progress
3651 : logical_sector
>= conf
->reshape_progress
) {
3652 disks
= conf
->previous_raid_disks
;
3655 if (mddev
->delta_disks
< 0
3656 ? logical_sector
< conf
->reshape_safe
3657 : logical_sector
>= conf
->reshape_safe
) {
3658 spin_unlock_irq(&conf
->device_lock
);
3663 spin_unlock_irq(&conf
->device_lock
);
3665 data_disks
= disks
- conf
->max_degraded
;
3667 new_sector
= raid5_compute_sector(conf
, logical_sector
,
3670 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3671 (unsigned long long)new_sector
,
3672 (unsigned long long)logical_sector
);
3674 sh
= get_active_stripe(conf
, new_sector
, previous
,
3675 (bi
->bi_rw
&RWA_MASK
));
3677 if (unlikely(previous
)) {
3678 /* expansion might have moved on while waiting for a
3679 * stripe, so we must do the range check again.
3680 * Expansion could still move past after this
3681 * test, but as we are holding a reference to
3682 * 'sh', we know that if that happens,
3683 * STRIPE_EXPANDING will get set and the expansion
3684 * won't proceed until we finish with the stripe.
3687 spin_lock_irq(&conf
->device_lock
);
3688 if (mddev
->delta_disks
< 0
3689 ? logical_sector
>= conf
->reshape_progress
3690 : logical_sector
< conf
->reshape_progress
)
3691 /* mismatch, need to try again */
3693 spin_unlock_irq(&conf
->device_lock
);
3699 /* FIXME what if we get a false positive because these
3700 * are being updated.
3702 if (logical_sector
>= mddev
->suspend_lo
&&
3703 logical_sector
< mddev
->suspend_hi
) {
3709 if (test_bit(STRIPE_EXPANDING
, &sh
->state
) ||
3710 !add_stripe_bio(sh
, bi
, dd_idx
, (bi
->bi_rw
&RW_MASK
))) {
3711 /* Stripe is busy expanding or
3712 * add failed due to overlap. Flush everything
3715 raid5_unplug_device(mddev
->queue
);
3720 finish_wait(&conf
->wait_for_overlap
, &w
);
3721 set_bit(STRIPE_HANDLE
, &sh
->state
);
3722 clear_bit(STRIPE_DELAYED
, &sh
->state
);
3725 /* cannot get stripe for read-ahead, just give-up */
3726 clear_bit(BIO_UPTODATE
, &bi
->bi_flags
);
3727 finish_wait(&conf
->wait_for_overlap
, &w
);
3732 spin_lock_irq(&conf
->device_lock
);
3733 remaining
= raid5_dec_bi_phys_segments(bi
);
3734 spin_unlock_irq(&conf
->device_lock
);
3735 if (remaining
== 0) {
3738 md_write_end(mddev
);
3745 static sector_t
raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
);
3747 static sector_t
reshape_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
)
3749 /* reshaping is quite different to recovery/resync so it is
3750 * handled quite separately ... here.
3752 * On each call to sync_request, we gather one chunk worth of
3753 * destination stripes and flag them as expanding.
3754 * Then we find all the source stripes and request reads.
3755 * As the reads complete, handle_stripe will copy the data
3756 * into the destination stripe and release that stripe.
3758 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
3759 struct stripe_head
*sh
;
3760 sector_t first_sector
, last_sector
;
3761 int raid_disks
= conf
->previous_raid_disks
;
3762 int data_disks
= raid_disks
- conf
->max_degraded
;
3763 int new_data_disks
= conf
->raid_disks
- conf
->max_degraded
;
3766 sector_t writepos
, readpos
, safepos
;
3767 sector_t stripe_addr
;
3768 int reshape_sectors
;
3769 struct list_head stripes
;
3771 if (sector_nr
== 0) {
3772 /* If restarting in the middle, skip the initial sectors */
3773 if (mddev
->delta_disks
< 0 &&
3774 conf
->reshape_progress
< raid5_size(mddev
, 0, 0)) {
3775 sector_nr
= raid5_size(mddev
, 0, 0)
3776 - conf
->reshape_progress
;
3777 } else if (mddev
->delta_disks
> 0 &&
3778 conf
->reshape_progress
> 0)
3779 sector_nr
= conf
->reshape_progress
;
3780 sector_div(sector_nr
, new_data_disks
);
3787 /* We need to process a full chunk at a time.
3788 * If old and new chunk sizes differ, we need to process the
3791 if (mddev
->new_chunk
> mddev
->chunk_size
)
3792 reshape_sectors
= mddev
->new_chunk
/ 512;
3794 reshape_sectors
= mddev
->chunk_size
/ 512;
3796 /* we update the metadata when there is more than 3Meg
3797 * in the block range (that is rather arbitrary, should
3798 * probably be time based) or when the data about to be
3799 * copied would over-write the source of the data at
3800 * the front of the range.
3801 * i.e. one new_stripe along from reshape_progress new_maps
3802 * to after where reshape_safe old_maps to
3804 writepos
= conf
->reshape_progress
;
3805 sector_div(writepos
, new_data_disks
);
3806 readpos
= conf
->reshape_progress
;
3807 sector_div(readpos
, data_disks
);
3808 safepos
= conf
->reshape_safe
;
3809 sector_div(safepos
, data_disks
);
3810 if (mddev
->delta_disks
< 0) {
3811 writepos
-= reshape_sectors
;
3812 readpos
+= reshape_sectors
;
3813 safepos
+= reshape_sectors
;
3815 writepos
+= reshape_sectors
;
3816 readpos
-= reshape_sectors
;
3817 safepos
-= reshape_sectors
;
3820 /* 'writepos' is the most advanced device address we might write.
3821 * 'readpos' is the least advanced device address we might read.
3822 * 'safepos' is the least address recorded in the metadata as having
3824 * If 'readpos' is behind 'writepos', then there is no way that we can
3825 * ensure safety in the face of a crash - that must be done by userspace
3826 * making a backup of the data. So in that case there is no particular
3827 * rush to update metadata.
3828 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3829 * update the metadata to advance 'safepos' to match 'readpos' so that
3830 * we can be safe in the event of a crash.
3831 * So we insist on updating metadata if safepos is behind writepos and
3832 * readpos is beyond writepos.
3833 * In any case, update the metadata every 10 seconds.
3834 * Maybe that number should be configurable, but I'm not sure it is
3835 * worth it.... maybe it could be a multiple of safemode_delay???
3837 if ((mddev
->delta_disks
< 0
3838 ? (safepos
> writepos
&& readpos
< writepos
)
3839 : (safepos
< writepos
&& readpos
> writepos
)) ||
3840 time_after(jiffies
, conf
->reshape_checkpoint
+ 10*HZ
)) {
3841 /* Cannot proceed until we've updated the superblock... */
3842 wait_event(conf
->wait_for_overlap
,
3843 atomic_read(&conf
->reshape_stripes
)==0);
3844 mddev
->reshape_position
= conf
->reshape_progress
;
3845 conf
->reshape_checkpoint
= jiffies
;
3846 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
3847 md_wakeup_thread(mddev
->thread
);
3848 wait_event(mddev
->sb_wait
, mddev
->flags
== 0 ||
3849 kthread_should_stop());
3850 spin_lock_irq(&conf
->device_lock
);
3851 conf
->reshape_safe
= mddev
->reshape_position
;
3852 spin_unlock_irq(&conf
->device_lock
);
3853 wake_up(&conf
->wait_for_overlap
);
3856 if (mddev
->delta_disks
< 0) {
3857 BUG_ON(conf
->reshape_progress
== 0);
3858 stripe_addr
= writepos
;
3859 BUG_ON((mddev
->dev_sectors
&
3860 ~((sector_t
)reshape_sectors
- 1))
3861 - reshape_sectors
- stripe_addr
3864 BUG_ON(writepos
!= sector_nr
+ reshape_sectors
);
3865 stripe_addr
= sector_nr
;
3867 INIT_LIST_HEAD(&stripes
);
3868 for (i
= 0; i
< reshape_sectors
; i
+= STRIPE_SECTORS
) {
3871 sh
= get_active_stripe(conf
, stripe_addr
+i
, 0, 0);
3872 set_bit(STRIPE_EXPANDING
, &sh
->state
);
3873 atomic_inc(&conf
->reshape_stripes
);
3874 /* If any of this stripe is beyond the end of the old
3875 * array, then we need to zero those blocks
3877 for (j
=sh
->disks
; j
--;) {
3879 if (j
== sh
->pd_idx
)
3881 if (conf
->level
== 6 &&
3884 s
= compute_blocknr(sh
, j
, 0);
3885 if (s
< raid5_size(mddev
, 0, 0)) {
3889 memset(page_address(sh
->dev
[j
].page
), 0, STRIPE_SIZE
);
3890 set_bit(R5_Expanded
, &sh
->dev
[j
].flags
);
3891 set_bit(R5_UPTODATE
, &sh
->dev
[j
].flags
);
3894 set_bit(STRIPE_EXPAND_READY
, &sh
->state
);
3895 set_bit(STRIPE_HANDLE
, &sh
->state
);
3897 list_add(&sh
->lru
, &stripes
);
3899 spin_lock_irq(&conf
->device_lock
);
3900 if (mddev
->delta_disks
< 0)
3901 conf
->reshape_progress
-= reshape_sectors
* new_data_disks
;
3903 conf
->reshape_progress
+= reshape_sectors
* new_data_disks
;
3904 spin_unlock_irq(&conf
->device_lock
);
3905 /* Ok, those stripe are ready. We can start scheduling
3906 * reads on the source stripes.
3907 * The source stripes are determined by mapping the first and last
3908 * block on the destination stripes.
3911 raid5_compute_sector(conf
, stripe_addr
*(new_data_disks
),
3914 raid5_compute_sector(conf
, ((stripe_addr
+conf
->chunk_size
/512)
3915 *(new_data_disks
) - 1),
3917 if (last_sector
>= mddev
->dev_sectors
)
3918 last_sector
= mddev
->dev_sectors
- 1;
3919 while (first_sector
<= last_sector
) {
3920 sh
= get_active_stripe(conf
, first_sector
, 1, 0);
3921 set_bit(STRIPE_EXPAND_SOURCE
, &sh
->state
);
3922 set_bit(STRIPE_HANDLE
, &sh
->state
);
3924 first_sector
+= STRIPE_SECTORS
;
3926 /* Now that the sources are clearly marked, we can release
3927 * the destination stripes
3929 while (!list_empty(&stripes
)) {
3930 sh
= list_entry(stripes
.next
, struct stripe_head
, lru
);
3931 list_del_init(&sh
->lru
);
3934 /* If this takes us to the resync_max point where we have to pause,
3935 * then we need to write out the superblock.
3937 sector_nr
+= reshape_sectors
;
3938 if (sector_nr
>= mddev
->resync_max
) {
3939 /* Cannot proceed until we've updated the superblock... */
3940 wait_event(conf
->wait_for_overlap
,
3941 atomic_read(&conf
->reshape_stripes
) == 0);
3942 mddev
->reshape_position
= conf
->reshape_progress
;
3943 conf
->reshape_checkpoint
= jiffies
;
3944 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
3945 md_wakeup_thread(mddev
->thread
);
3946 wait_event(mddev
->sb_wait
,
3947 !test_bit(MD_CHANGE_DEVS
, &mddev
->flags
)
3948 || kthread_should_stop());
3949 spin_lock_irq(&conf
->device_lock
);
3950 conf
->reshape_safe
= mddev
->reshape_position
;
3951 spin_unlock_irq(&conf
->device_lock
);
3952 wake_up(&conf
->wait_for_overlap
);
3954 return reshape_sectors
;
3957 /* FIXME go_faster isn't used */
3958 static inline sector_t
sync_request(mddev_t
*mddev
, sector_t sector_nr
, int *skipped
, int go_faster
)
3960 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
3961 struct stripe_head
*sh
;
3962 sector_t max_sector
= mddev
->dev_sectors
;
3964 int still_degraded
= 0;
3967 if (sector_nr
>= max_sector
) {
3968 /* just being told to finish up .. nothing much to do */
3969 unplug_slaves(mddev
);
3971 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
)) {
3976 if (mddev
->curr_resync
< max_sector
) /* aborted */
3977 bitmap_end_sync(mddev
->bitmap
, mddev
->curr_resync
,
3979 else /* completed sync */
3981 bitmap_close_sync(mddev
->bitmap
);
3986 if (test_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
))
3987 return reshape_request(mddev
, sector_nr
, skipped
);
3989 /* No need to check resync_max as we never do more than one
3990 * stripe, and as resync_max will always be on a chunk boundary,
3991 * if the check in md_do_sync didn't fire, there is no chance
3992 * of overstepping resync_max here
3995 /* if there is too many failed drives and we are trying
3996 * to resync, then assert that we are finished, because there is
3997 * nothing we can do.
3999 if (mddev
->degraded
>= conf
->max_degraded
&&
4000 test_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
)) {
4001 sector_t rv
= mddev
->dev_sectors
- sector_nr
;
4005 if (!bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, 1) &&
4006 !test_bit(MD_RECOVERY_REQUESTED
, &mddev
->recovery
) &&
4007 !conf
->fullsync
&& sync_blocks
>= STRIPE_SECTORS
) {
4008 /* we can skip this block, and probably more */
4009 sync_blocks
/= STRIPE_SECTORS
;
4011 return sync_blocks
* STRIPE_SECTORS
; /* keep things rounded to whole stripes */
4015 bitmap_cond_end_sync(mddev
->bitmap
, sector_nr
);
4017 sh
= get_active_stripe(conf
, sector_nr
, 0, 1);
4019 sh
= get_active_stripe(conf
, sector_nr
, 0, 0);
4020 /* make sure we don't swamp the stripe cache if someone else
4021 * is trying to get access
4023 schedule_timeout_uninterruptible(1);
4025 /* Need to check if array will still be degraded after recovery/resync
4026 * We don't need to check the 'failed' flag as when that gets set,
4029 for (i
=0; i
<mddev
->raid_disks
; i
++)
4030 if (conf
->disks
[i
].rdev
== NULL
)
4033 bitmap_start_sync(mddev
->bitmap
, sector_nr
, &sync_blocks
, still_degraded
);
4035 spin_lock(&sh
->lock
);
4036 set_bit(STRIPE_SYNCING
, &sh
->state
);
4037 clear_bit(STRIPE_INSYNC
, &sh
->state
);
4038 spin_unlock(&sh
->lock
);
4040 /* wait for any blocked device to be handled */
4041 while(unlikely(!handle_stripe(sh
, NULL
)))
4045 return STRIPE_SECTORS
;
4048 static int retry_aligned_read(raid5_conf_t
*conf
, struct bio
*raid_bio
)
4050 /* We may not be able to submit a whole bio at once as there
4051 * may not be enough stripe_heads available.
4052 * We cannot pre-allocate enough stripe_heads as we may need
4053 * more than exist in the cache (if we allow ever large chunks).
4054 * So we do one stripe head at a time and record in
4055 * ->bi_hw_segments how many have been done.
4057 * We *know* that this entire raid_bio is in one chunk, so
4058 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4060 struct stripe_head
*sh
;
4062 sector_t sector
, logical_sector
, last_sector
;
4067 logical_sector
= raid_bio
->bi_sector
& ~((sector_t
)STRIPE_SECTORS
-1);
4068 sector
= raid5_compute_sector(conf
, logical_sector
,
4070 last_sector
= raid_bio
->bi_sector
+ (raid_bio
->bi_size
>>9);
4072 for (; logical_sector
< last_sector
;
4073 logical_sector
+= STRIPE_SECTORS
,
4074 sector
+= STRIPE_SECTORS
,
4077 if (scnt
< raid5_bi_hw_segments(raid_bio
))
4078 /* already done this stripe */
4081 sh
= get_active_stripe(conf
, sector
, 0, 1);
4084 /* failed to get a stripe - must wait */
4085 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4086 conf
->retry_read_aligned
= raid_bio
;
4090 set_bit(R5_ReadError
, &sh
->dev
[dd_idx
].flags
);
4091 if (!add_stripe_bio(sh
, raid_bio
, dd_idx
, 0)) {
4093 raid5_set_bi_hw_segments(raid_bio
, scnt
);
4094 conf
->retry_read_aligned
= raid_bio
;
4098 handle_stripe(sh
, NULL
);
4102 spin_lock_irq(&conf
->device_lock
);
4103 remaining
= raid5_dec_bi_phys_segments(raid_bio
);
4104 spin_unlock_irq(&conf
->device_lock
);
4106 bio_endio(raid_bio
, 0);
4107 if (atomic_dec_and_test(&conf
->active_aligned_reads
))
4108 wake_up(&conf
->wait_for_stripe
);
4115 * This is our raid5 kernel thread.
4117 * We scan the hash table for stripes which can be handled now.
4118 * During the scan, completed stripes are saved for us by the interrupt
4119 * handler, so that they will not have to wait for our next wakeup.
4121 static void raid5d(mddev_t
*mddev
)
4123 struct stripe_head
*sh
;
4124 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4127 pr_debug("+++ raid5d active\n");
4129 md_check_recovery(mddev
);
4132 spin_lock_irq(&conf
->device_lock
);
4136 if (conf
->seq_flush
!= conf
->seq_write
) {
4137 int seq
= conf
->seq_flush
;
4138 spin_unlock_irq(&conf
->device_lock
);
4139 bitmap_unplug(mddev
->bitmap
);
4140 spin_lock_irq(&conf
->device_lock
);
4141 conf
->seq_write
= seq
;
4142 activate_bit_delay(conf
);
4145 while ((bio
= remove_bio_from_retry(conf
))) {
4147 spin_unlock_irq(&conf
->device_lock
);
4148 ok
= retry_aligned_read(conf
, bio
);
4149 spin_lock_irq(&conf
->device_lock
);
4155 sh
= __get_priority_stripe(conf
);
4159 spin_unlock_irq(&conf
->device_lock
);
4162 handle_stripe(sh
, conf
->spare_page
);
4165 spin_lock_irq(&conf
->device_lock
);
4167 pr_debug("%d stripes handled\n", handled
);
4169 spin_unlock_irq(&conf
->device_lock
);
4171 async_tx_issue_pending_all();
4172 unplug_slaves(mddev
);
4174 pr_debug("--- raid5d inactive\n");
4178 raid5_show_stripe_cache_size(mddev_t
*mddev
, char *page
)
4180 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4182 return sprintf(page
, "%d\n", conf
->max_nr_stripes
);
4188 raid5_store_stripe_cache_size(mddev_t
*mddev
, const char *page
, size_t len
)
4190 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4194 if (len
>= PAGE_SIZE
)
4199 if (strict_strtoul(page
, 10, &new))
4201 if (new <= 16 || new > 32768)
4203 while (new < conf
->max_nr_stripes
) {
4204 if (drop_one_stripe(conf
))
4205 conf
->max_nr_stripes
--;
4209 err
= md_allow_write(mddev
);
4212 while (new > conf
->max_nr_stripes
) {
4213 if (grow_one_stripe(conf
))
4214 conf
->max_nr_stripes
++;
4220 static struct md_sysfs_entry
4221 raid5_stripecache_size
= __ATTR(stripe_cache_size
, S_IRUGO
| S_IWUSR
,
4222 raid5_show_stripe_cache_size
,
4223 raid5_store_stripe_cache_size
);
4226 raid5_show_preread_threshold(mddev_t
*mddev
, char *page
)
4228 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4230 return sprintf(page
, "%d\n", conf
->bypass_threshold
);
4236 raid5_store_preread_threshold(mddev_t
*mddev
, const char *page
, size_t len
)
4238 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4240 if (len
>= PAGE_SIZE
)
4245 if (strict_strtoul(page
, 10, &new))
4247 if (new > conf
->max_nr_stripes
)
4249 conf
->bypass_threshold
= new;
4253 static struct md_sysfs_entry
4254 raid5_preread_bypass_threshold
= __ATTR(preread_bypass_threshold
,
4256 raid5_show_preread_threshold
,
4257 raid5_store_preread_threshold
);
4260 stripe_cache_active_show(mddev_t
*mddev
, char *page
)
4262 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4264 return sprintf(page
, "%d\n", atomic_read(&conf
->active_stripes
));
4269 static struct md_sysfs_entry
4270 raid5_stripecache_active
= __ATTR_RO(stripe_cache_active
);
4272 static struct attribute
*raid5_attrs
[] = {
4273 &raid5_stripecache_size
.attr
,
4274 &raid5_stripecache_active
.attr
,
4275 &raid5_preread_bypass_threshold
.attr
,
4278 static struct attribute_group raid5_attrs_group
= {
4280 .attrs
= raid5_attrs
,
4284 raid5_size(mddev_t
*mddev
, sector_t sectors
, int raid_disks
)
4286 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4289 sectors
= mddev
->dev_sectors
;
4291 /* size is defined by the smallest of previous and new size */
4292 if (conf
->raid_disks
< conf
->previous_raid_disks
)
4293 raid_disks
= conf
->raid_disks
;
4295 raid_disks
= conf
->previous_raid_disks
;
4298 sectors
&= ~((sector_t
)mddev
->chunk_size
/512 - 1);
4299 sectors
&= ~((sector_t
)mddev
->new_chunk
/512 - 1);
4300 return sectors
* (raid_disks
- conf
->max_degraded
);
4303 static raid5_conf_t
*setup_conf(mddev_t
*mddev
)
4306 int raid_disk
, memory
;
4308 struct disk_info
*disk
;
4310 if (mddev
->new_level
!= 5
4311 && mddev
->new_level
!= 4
4312 && mddev
->new_level
!= 6) {
4313 printk(KERN_ERR
"raid5: %s: raid level not set to 4/5/6 (%d)\n",
4314 mdname(mddev
), mddev
->new_level
);
4315 return ERR_PTR(-EIO
);
4317 if ((mddev
->new_level
== 5
4318 && !algorithm_valid_raid5(mddev
->new_layout
)) ||
4319 (mddev
->new_level
== 6
4320 && !algorithm_valid_raid6(mddev
->new_layout
))) {
4321 printk(KERN_ERR
"raid5: %s: layout %d not supported\n",
4322 mdname(mddev
), mddev
->new_layout
);
4323 return ERR_PTR(-EIO
);
4325 if (mddev
->new_level
== 6 && mddev
->raid_disks
< 4) {
4326 printk(KERN_ERR
"raid6: not enough configured devices for %s (%d, minimum 4)\n",
4327 mdname(mddev
), mddev
->raid_disks
);
4328 return ERR_PTR(-EINVAL
);
4331 if (!mddev
->new_chunk
|| mddev
->new_chunk
% PAGE_SIZE
) {
4332 printk(KERN_ERR
"raid5: invalid chunk size %d for %s\n",
4333 mddev
->new_chunk
, mdname(mddev
));
4334 return ERR_PTR(-EINVAL
);
4337 conf
= kzalloc(sizeof(raid5_conf_t
), GFP_KERNEL
);
4341 conf
->raid_disks
= mddev
->raid_disks
;
4342 if (mddev
->reshape_position
== MaxSector
)
4343 conf
->previous_raid_disks
= mddev
->raid_disks
;
4345 conf
->previous_raid_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4347 conf
->disks
= kzalloc(conf
->raid_disks
* sizeof(struct disk_info
),
4352 conf
->mddev
= mddev
;
4354 if ((conf
->stripe_hashtbl
= kzalloc(PAGE_SIZE
, GFP_KERNEL
)) == NULL
)
4357 if (mddev
->new_level
== 6) {
4358 conf
->spare_page
= alloc_page(GFP_KERNEL
);
4359 if (!conf
->spare_page
)
4362 spin_lock_init(&conf
->device_lock
);
4363 init_waitqueue_head(&conf
->wait_for_stripe
);
4364 init_waitqueue_head(&conf
->wait_for_overlap
);
4365 INIT_LIST_HEAD(&conf
->handle_list
);
4366 INIT_LIST_HEAD(&conf
->hold_list
);
4367 INIT_LIST_HEAD(&conf
->delayed_list
);
4368 INIT_LIST_HEAD(&conf
->bitmap_list
);
4369 INIT_LIST_HEAD(&conf
->inactive_list
);
4370 atomic_set(&conf
->active_stripes
, 0);
4371 atomic_set(&conf
->preread_active_stripes
, 0);
4372 atomic_set(&conf
->active_aligned_reads
, 0);
4373 conf
->bypass_threshold
= BYPASS_THRESHOLD
;
4375 pr_debug("raid5: run(%s) called.\n", mdname(mddev
));
4377 list_for_each_entry(rdev
, &mddev
->disks
, same_set
) {
4378 raid_disk
= rdev
->raid_disk
;
4379 if (raid_disk
>= conf
->raid_disks
4382 disk
= conf
->disks
+ raid_disk
;
4386 if (test_bit(In_sync
, &rdev
->flags
)) {
4387 char b
[BDEVNAME_SIZE
];
4388 printk(KERN_INFO
"raid5: device %s operational as raid"
4389 " disk %d\n", bdevname(rdev
->bdev
,b
),
4392 /* Cannot rely on bitmap to complete recovery */
4396 conf
->chunk_size
= mddev
->new_chunk
;
4397 conf
->level
= mddev
->new_level
;
4398 if (conf
->level
== 6)
4399 conf
->max_degraded
= 2;
4401 conf
->max_degraded
= 1;
4402 conf
->algorithm
= mddev
->new_layout
;
4403 conf
->max_nr_stripes
= NR_STRIPES
;
4404 conf
->reshape_progress
= mddev
->reshape_position
;
4405 if (conf
->reshape_progress
!= MaxSector
) {
4406 conf
->prev_chunk
= mddev
->chunk_size
;
4407 conf
->prev_algo
= mddev
->layout
;
4410 memory
= conf
->max_nr_stripes
* (sizeof(struct stripe_head
) +
4411 conf
->raid_disks
* ((sizeof(struct bio
) + PAGE_SIZE
))) / 1024;
4412 if (grow_stripes(conf
, conf
->max_nr_stripes
)) {
4414 "raid5: couldn't allocate %dkB for buffers\n", memory
);
4417 printk(KERN_INFO
"raid5: allocated %dkB for %s\n",
4418 memory
, mdname(mddev
));
4420 conf
->thread
= md_register_thread(raid5d
, mddev
, "%s_raid5");
4421 if (!conf
->thread
) {
4423 "raid5: couldn't allocate thread for %s\n",
4432 shrink_stripes(conf
);
4433 safe_put_page(conf
->spare_page
);
4435 kfree(conf
->stripe_hashtbl
);
4437 return ERR_PTR(-EIO
);
4439 return ERR_PTR(-ENOMEM
);
4442 static int run(mddev_t
*mddev
)
4445 int working_disks
= 0;
4448 if (mddev
->reshape_position
!= MaxSector
) {
4449 /* Check that we can continue the reshape.
4450 * Currently only disks can change, it must
4451 * increase, and we must be past the point where
4452 * a stripe over-writes itself
4454 sector_t here_new
, here_old
;
4456 int max_degraded
= (mddev
->level
== 6 ? 2 : 1);
4458 if (mddev
->new_level
!= mddev
->level
) {
4459 printk(KERN_ERR
"raid5: %s: unsupported reshape "
4460 "required - aborting.\n",
4464 old_disks
= mddev
->raid_disks
- mddev
->delta_disks
;
4465 /* reshape_position must be on a new-stripe boundary, and one
4466 * further up in new geometry must map after here in old
4469 here_new
= mddev
->reshape_position
;
4470 if (sector_div(here_new
, (mddev
->new_chunk
>>9)*
4471 (mddev
->raid_disks
- max_degraded
))) {
4472 printk(KERN_ERR
"raid5: reshape_position not "
4473 "on a stripe boundary\n");
4476 /* here_new is the stripe we will write to */
4477 here_old
= mddev
->reshape_position
;
4478 sector_div(here_old
, (mddev
->chunk_size
>>9)*
4479 (old_disks
-max_degraded
));
4480 /* here_old is the first stripe that we might need to read
4482 if (here_new
>= here_old
) {
4483 /* Reading from the same stripe as writing to - bad */
4484 printk(KERN_ERR
"raid5: reshape_position too early for "
4485 "auto-recovery - aborting.\n");
4488 printk(KERN_INFO
"raid5: reshape will continue\n");
4489 /* OK, we should be able to continue; */
4491 BUG_ON(mddev
->level
!= mddev
->new_level
);
4492 BUG_ON(mddev
->layout
!= mddev
->new_layout
);
4493 BUG_ON(mddev
->chunk_size
!= mddev
->new_chunk
);
4494 BUG_ON(mddev
->delta_disks
!= 0);
4497 if (mddev
->private == NULL
)
4498 conf
= setup_conf(mddev
);
4500 conf
= mddev
->private;
4503 return PTR_ERR(conf
);
4505 mddev
->thread
= conf
->thread
;
4506 conf
->thread
= NULL
;
4507 mddev
->private = conf
;
4510 * 0 for a fully functional array, 1 or 2 for a degraded array.
4512 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
4513 if (rdev
->raid_disk
>= 0 &&
4514 test_bit(In_sync
, &rdev
->flags
))
4517 mddev
->degraded
= conf
->raid_disks
- working_disks
;
4519 if (mddev
->degraded
> conf
->max_degraded
) {
4520 printk(KERN_ERR
"raid5: not enough operational devices for %s"
4521 " (%d/%d failed)\n",
4522 mdname(mddev
), mddev
->degraded
, conf
->raid_disks
);
4526 /* device size must be a multiple of chunk size */
4527 mddev
->dev_sectors
&= ~(mddev
->chunk_size
/ 512 - 1);
4528 mddev
->resync_max_sectors
= mddev
->dev_sectors
;
4530 if (mddev
->degraded
> 0 &&
4531 mddev
->recovery_cp
!= MaxSector
) {
4532 if (mddev
->ok_start_degraded
)
4534 "raid5: starting dirty degraded array: %s"
4535 "- data corruption possible.\n",
4539 "raid5: cannot start dirty degraded array for %s\n",
4545 if (mddev
->degraded
== 0)
4546 printk("raid5: raid level %d set %s active with %d out of %d"
4547 " devices, algorithm %d\n", conf
->level
, mdname(mddev
),
4548 mddev
->raid_disks
-mddev
->degraded
, mddev
->raid_disks
,
4551 printk(KERN_ALERT
"raid5: raid level %d set %s active with %d"
4552 " out of %d devices, algorithm %d\n", conf
->level
,
4553 mdname(mddev
), mddev
->raid_disks
- mddev
->degraded
,
4554 mddev
->raid_disks
, mddev
->new_layout
);
4556 print_raid5_conf(conf
);
4558 if (conf
->reshape_progress
!= MaxSector
) {
4559 printk("...ok start reshape thread\n");
4560 conf
->reshape_safe
= conf
->reshape_progress
;
4561 atomic_set(&conf
->reshape_stripes
, 0);
4562 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4563 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4564 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4565 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4566 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4570 /* read-ahead size must cover two whole stripes, which is
4571 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4574 int data_disks
= conf
->previous_raid_disks
- conf
->max_degraded
;
4575 int stripe
= data_disks
*
4576 (mddev
->chunk_size
/ PAGE_SIZE
);
4577 if (mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
4578 mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
4581 /* Ok, everything is just fine now */
4582 if (sysfs_create_group(&mddev
->kobj
, &raid5_attrs_group
))
4584 "raid5: failed to create sysfs attributes for %s\n",
4587 mddev
->queue
->queue_lock
= &conf
->device_lock
;
4589 mddev
->queue
->unplug_fn
= raid5_unplug_device
;
4590 mddev
->queue
->backing_dev_info
.congested_data
= mddev
;
4591 mddev
->queue
->backing_dev_info
.congested_fn
= raid5_congested
;
4593 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
4595 blk_queue_merge_bvec(mddev
->queue
, raid5_mergeable_bvec
);
4599 md_unregister_thread(mddev
->thread
);
4600 mddev
->thread
= NULL
;
4602 shrink_stripes(conf
);
4603 print_raid5_conf(conf
);
4604 safe_put_page(conf
->spare_page
);
4606 kfree(conf
->stripe_hashtbl
);
4609 mddev
->private = NULL
;
4610 printk(KERN_ALERT
"raid5: failed to run raid set %s\n", mdname(mddev
));
4616 static int stop(mddev_t
*mddev
)
4618 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
4620 md_unregister_thread(mddev
->thread
);
4621 mddev
->thread
= NULL
;
4622 shrink_stripes(conf
);
4623 kfree(conf
->stripe_hashtbl
);
4624 mddev
->queue
->backing_dev_info
.congested_fn
= NULL
;
4625 blk_sync_queue(mddev
->queue
); /* the unplug fn references 'conf'*/
4626 sysfs_remove_group(&mddev
->kobj
, &raid5_attrs_group
);
4629 mddev
->private = NULL
;
4634 static void print_sh(struct seq_file
*seq
, struct stripe_head
*sh
)
4638 seq_printf(seq
, "sh %llu, pd_idx %d, state %ld.\n",
4639 (unsigned long long)sh
->sector
, sh
->pd_idx
, sh
->state
);
4640 seq_printf(seq
, "sh %llu, count %d.\n",
4641 (unsigned long long)sh
->sector
, atomic_read(&sh
->count
));
4642 seq_printf(seq
, "sh %llu, ", (unsigned long long)sh
->sector
);
4643 for (i
= 0; i
< sh
->disks
; i
++) {
4644 seq_printf(seq
, "(cache%d: %p %ld) ",
4645 i
, sh
->dev
[i
].page
, sh
->dev
[i
].flags
);
4647 seq_printf(seq
, "\n");
4650 static void printall(struct seq_file
*seq
, raid5_conf_t
*conf
)
4652 struct stripe_head
*sh
;
4653 struct hlist_node
*hn
;
4656 spin_lock_irq(&conf
->device_lock
);
4657 for (i
= 0; i
< NR_HASH
; i
++) {
4658 hlist_for_each_entry(sh
, hn
, &conf
->stripe_hashtbl
[i
], hash
) {
4659 if (sh
->raid_conf
!= conf
)
4664 spin_unlock_irq(&conf
->device_lock
);
4668 static void status(struct seq_file
*seq
, mddev_t
*mddev
)
4670 raid5_conf_t
*conf
= (raid5_conf_t
*) mddev
->private;
4673 seq_printf (seq
, " level %d, %dk chunk, algorithm %d", mddev
->level
, mddev
->chunk_size
>> 10, mddev
->layout
);
4674 seq_printf (seq
, " [%d/%d] [", conf
->raid_disks
, conf
->raid_disks
- mddev
->degraded
);
4675 for (i
= 0; i
< conf
->raid_disks
; i
++)
4676 seq_printf (seq
, "%s",
4677 conf
->disks
[i
].rdev
&&
4678 test_bit(In_sync
, &conf
->disks
[i
].rdev
->flags
) ? "U" : "_");
4679 seq_printf (seq
, "]");
4681 seq_printf (seq
, "\n");
4682 printall(seq
, conf
);
4686 static void print_raid5_conf (raid5_conf_t
*conf
)
4689 struct disk_info
*tmp
;
4691 printk("RAID5 conf printout:\n");
4693 printk("(conf==NULL)\n");
4696 printk(" --- rd:%d wd:%d\n", conf
->raid_disks
,
4697 conf
->raid_disks
- conf
->mddev
->degraded
);
4699 for (i
= 0; i
< conf
->raid_disks
; i
++) {
4700 char b
[BDEVNAME_SIZE
];
4701 tmp
= conf
->disks
+ i
;
4703 printk(" disk %d, o:%d, dev:%s\n",
4704 i
, !test_bit(Faulty
, &tmp
->rdev
->flags
),
4705 bdevname(tmp
->rdev
->bdev
,b
));
4709 static int raid5_spare_active(mddev_t
*mddev
)
4712 raid5_conf_t
*conf
= mddev
->private;
4713 struct disk_info
*tmp
;
4715 for (i
= 0; i
< conf
->raid_disks
; i
++) {
4716 tmp
= conf
->disks
+ i
;
4718 && !test_bit(Faulty
, &tmp
->rdev
->flags
)
4719 && !test_and_set_bit(In_sync
, &tmp
->rdev
->flags
)) {
4720 unsigned long flags
;
4721 spin_lock_irqsave(&conf
->device_lock
, flags
);
4723 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4726 print_raid5_conf(conf
);
4730 static int raid5_remove_disk(mddev_t
*mddev
, int number
)
4732 raid5_conf_t
*conf
= mddev
->private;
4735 struct disk_info
*p
= conf
->disks
+ number
;
4737 print_raid5_conf(conf
);
4740 if (number
>= conf
->raid_disks
&&
4741 conf
->reshape_progress
== MaxSector
)
4742 clear_bit(In_sync
, &rdev
->flags
);
4744 if (test_bit(In_sync
, &rdev
->flags
) ||
4745 atomic_read(&rdev
->nr_pending
)) {
4749 /* Only remove non-faulty devices if recovery
4752 if (!test_bit(Faulty
, &rdev
->flags
) &&
4753 mddev
->degraded
<= conf
->max_degraded
&&
4754 number
< conf
->raid_disks
) {
4760 if (atomic_read(&rdev
->nr_pending
)) {
4761 /* lost the race, try later */
4768 print_raid5_conf(conf
);
4772 static int raid5_add_disk(mddev_t
*mddev
, mdk_rdev_t
*rdev
)
4774 raid5_conf_t
*conf
= mddev
->private;
4777 struct disk_info
*p
;
4779 int last
= conf
->raid_disks
- 1;
4781 if (mddev
->degraded
> conf
->max_degraded
)
4782 /* no point adding a device */
4785 if (rdev
->raid_disk
>= 0)
4786 first
= last
= rdev
->raid_disk
;
4789 * find the disk ... but prefer rdev->saved_raid_disk
4792 if (rdev
->saved_raid_disk
>= 0 &&
4793 rdev
->saved_raid_disk
>= first
&&
4794 conf
->disks
[rdev
->saved_raid_disk
].rdev
== NULL
)
4795 disk
= rdev
->saved_raid_disk
;
4798 for ( ; disk
<= last
; disk
++)
4799 if ((p
=conf
->disks
+ disk
)->rdev
== NULL
) {
4800 clear_bit(In_sync
, &rdev
->flags
);
4801 rdev
->raid_disk
= disk
;
4803 if (rdev
->saved_raid_disk
!= disk
)
4805 rcu_assign_pointer(p
->rdev
, rdev
);
4808 print_raid5_conf(conf
);
4812 static int raid5_resize(mddev_t
*mddev
, sector_t sectors
)
4814 /* no resync is happening, and there is enough space
4815 * on all devices, so we can resize.
4816 * We need to make sure resync covers any new space.
4817 * If the array is shrinking we should possibly wait until
4818 * any io in the removed space completes, but it hardly seems
4821 sectors
&= ~((sector_t
)mddev
->chunk_size
/512 - 1);
4822 md_set_array_sectors(mddev
, raid5_size(mddev
, sectors
,
4823 mddev
->raid_disks
));
4824 if (mddev
->array_sectors
>
4825 raid5_size(mddev
, sectors
, mddev
->raid_disks
))
4827 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
4829 if (sectors
> mddev
->dev_sectors
&& mddev
->recovery_cp
== MaxSector
) {
4830 mddev
->recovery_cp
= mddev
->dev_sectors
;
4831 set_bit(MD_RECOVERY_NEEDED
, &mddev
->recovery
);
4833 mddev
->dev_sectors
= sectors
;
4834 mddev
->resync_max_sectors
= sectors
;
4838 static int raid5_check_reshape(mddev_t
*mddev
)
4840 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4842 if (mddev
->delta_disks
== 0 &&
4843 mddev
->new_layout
== mddev
->layout
&&
4844 mddev
->new_chunk
== mddev
->chunk_size
)
4845 return -EINVAL
; /* nothing to do */
4847 /* Cannot grow a bitmap yet */
4849 if (mddev
->degraded
> conf
->max_degraded
)
4851 if (mddev
->delta_disks
< 0) {
4852 /* We might be able to shrink, but the devices must
4853 * be made bigger first.
4854 * For raid6, 4 is the minimum size.
4855 * Otherwise 2 is the minimum
4858 if (mddev
->level
== 6)
4860 if (mddev
->raid_disks
+ mddev
->delta_disks
< min
)
4864 /* Can only proceed if there are plenty of stripe_heads.
4865 * We need a minimum of one full stripe,, and for sensible progress
4866 * it is best to have about 4 times that.
4867 * If we require 4 times, then the default 256 4K stripe_heads will
4868 * allow for chunk sizes up to 256K, which is probably OK.
4869 * If the chunk size is greater, user-space should request more
4870 * stripe_heads first.
4872 if ((mddev
->chunk_size
/ STRIPE_SIZE
) * 4 > conf
->max_nr_stripes
||
4873 (mddev
->new_chunk
/ STRIPE_SIZE
) * 4 > conf
->max_nr_stripes
) {
4874 printk(KERN_WARNING
"raid5: reshape: not enough stripes. Needed %lu\n",
4875 (max(mddev
->chunk_size
, mddev
->new_chunk
)
4880 return resize_stripes(conf
, conf
->raid_disks
+ mddev
->delta_disks
);
4883 static int raid5_start_reshape(mddev_t
*mddev
)
4885 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
4888 int added_devices
= 0;
4889 unsigned long flags
;
4891 if (test_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
))
4894 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
4895 if (rdev
->raid_disk
< 0 &&
4896 !test_bit(Faulty
, &rdev
->flags
))
4899 if (spares
- mddev
->degraded
< mddev
->delta_disks
- conf
->max_degraded
)
4900 /* Not enough devices even to make a degraded array
4905 /* Refuse to reduce size of the array. Any reductions in
4906 * array size must be through explicit setting of array_size
4909 if (raid5_size(mddev
, 0, conf
->raid_disks
+ mddev
->delta_disks
)
4910 < mddev
->array_sectors
) {
4911 printk(KERN_ERR
"md: %s: array size must be reduced "
4912 "before number of disks\n", mdname(mddev
));
4916 atomic_set(&conf
->reshape_stripes
, 0);
4917 spin_lock_irq(&conf
->device_lock
);
4918 conf
->previous_raid_disks
= conf
->raid_disks
;
4919 conf
->raid_disks
+= mddev
->delta_disks
;
4920 conf
->prev_chunk
= conf
->chunk_size
;
4921 conf
->chunk_size
= mddev
->new_chunk
;
4922 conf
->prev_algo
= conf
->algorithm
;
4923 conf
->algorithm
= mddev
->new_layout
;
4924 if (mddev
->delta_disks
< 0)
4925 conf
->reshape_progress
= raid5_size(mddev
, 0, 0);
4927 conf
->reshape_progress
= 0;
4928 conf
->reshape_safe
= conf
->reshape_progress
;
4930 spin_unlock_irq(&conf
->device_lock
);
4932 /* Add some new drives, as many as will fit.
4933 * We know there are enough to make the newly sized array work.
4935 list_for_each_entry(rdev
, &mddev
->disks
, same_set
)
4936 if (rdev
->raid_disk
< 0 &&
4937 !test_bit(Faulty
, &rdev
->flags
)) {
4938 if (raid5_add_disk(mddev
, rdev
) == 0) {
4940 set_bit(In_sync
, &rdev
->flags
);
4942 rdev
->recovery_offset
= 0;
4943 sprintf(nm
, "rd%d", rdev
->raid_disk
);
4944 if (sysfs_create_link(&mddev
->kobj
,
4947 "raid5: failed to create "
4948 " link %s for %s\n",
4954 if (mddev
->delta_disks
> 0) {
4955 spin_lock_irqsave(&conf
->device_lock
, flags
);
4956 mddev
->degraded
= (conf
->raid_disks
- conf
->previous_raid_disks
)
4958 spin_unlock_irqrestore(&conf
->device_lock
, flags
);
4960 mddev
->raid_disks
= conf
->raid_disks
;
4961 mddev
->reshape_position
= 0;
4962 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
4964 clear_bit(MD_RECOVERY_SYNC
, &mddev
->recovery
);
4965 clear_bit(MD_RECOVERY_CHECK
, &mddev
->recovery
);
4966 set_bit(MD_RECOVERY_RESHAPE
, &mddev
->recovery
);
4967 set_bit(MD_RECOVERY_RUNNING
, &mddev
->recovery
);
4968 mddev
->sync_thread
= md_register_thread(md_do_sync
, mddev
,
4970 if (!mddev
->sync_thread
) {
4971 mddev
->recovery
= 0;
4972 spin_lock_irq(&conf
->device_lock
);
4973 mddev
->raid_disks
= conf
->raid_disks
= conf
->previous_raid_disks
;
4974 conf
->reshape_progress
= MaxSector
;
4975 spin_unlock_irq(&conf
->device_lock
);
4978 conf
->reshape_checkpoint
= jiffies
;
4979 md_wakeup_thread(mddev
->sync_thread
);
4980 md_new_event(mddev
);
4984 /* This is called from the reshape thread and should make any
4985 * changes needed in 'conf'
4987 static void end_reshape(raid5_conf_t
*conf
)
4990 if (!test_bit(MD_RECOVERY_INTR
, &conf
->mddev
->recovery
)) {
4992 spin_lock_irq(&conf
->device_lock
);
4993 conf
->previous_raid_disks
= conf
->raid_disks
;
4994 conf
->reshape_progress
= MaxSector
;
4995 spin_unlock_irq(&conf
->device_lock
);
4996 wake_up(&conf
->wait_for_overlap
);
4998 /* read-ahead size must cover two whole stripes, which is
4999 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5002 int data_disks
= conf
->raid_disks
- conf
->max_degraded
;
5003 int stripe
= data_disks
* (conf
->chunk_size
5005 if (conf
->mddev
->queue
->backing_dev_info
.ra_pages
< 2 * stripe
)
5006 conf
->mddev
->queue
->backing_dev_info
.ra_pages
= 2 * stripe
;
5011 /* This is called from the raid5d thread with mddev_lock held.
5012 * It makes config changes to the device.
5014 static void raid5_finish_reshape(mddev_t
*mddev
)
5016 struct block_device
*bdev
;
5017 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
5019 if (!test_bit(MD_RECOVERY_INTR
, &mddev
->recovery
)) {
5021 if (mddev
->delta_disks
> 0) {
5022 md_set_array_sectors(mddev
, raid5_size(mddev
, 0, 0));
5023 set_capacity(mddev
->gendisk
, mddev
->array_sectors
);
5026 bdev
= bdget_disk(mddev
->gendisk
, 0);
5028 mutex_lock(&bdev
->bd_inode
->i_mutex
);
5029 i_size_write(bdev
->bd_inode
,
5030 (loff_t
)mddev
->array_sectors
<< 9);
5031 mutex_unlock(&bdev
->bd_inode
->i_mutex
);
5036 mddev
->degraded
= conf
->raid_disks
;
5037 for (d
= 0; d
< conf
->raid_disks
; d
++)
5038 if (conf
->disks
[d
].rdev
&&
5040 &conf
->disks
[d
].rdev
->flags
))
5042 for (d
= conf
->raid_disks
;
5043 d
< conf
->raid_disks
- mddev
->delta_disks
;
5045 raid5_remove_disk(mddev
, d
);
5047 mddev
->layout
= conf
->algorithm
;
5048 mddev
->chunk_size
= conf
->chunk_size
;
5049 mddev
->reshape_position
= MaxSector
;
5050 mddev
->delta_disks
= 0;
5054 static void raid5_quiesce(mddev_t
*mddev
, int state
)
5056 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
5059 case 2: /* resume for a suspend */
5060 wake_up(&conf
->wait_for_overlap
);
5063 case 1: /* stop all writes */
5064 spin_lock_irq(&conf
->device_lock
);
5066 wait_event_lock_irq(conf
->wait_for_stripe
,
5067 atomic_read(&conf
->active_stripes
) == 0 &&
5068 atomic_read(&conf
->active_aligned_reads
) == 0,
5069 conf
->device_lock
, /* nothing */);
5070 spin_unlock_irq(&conf
->device_lock
);
5073 case 0: /* re-enable writes */
5074 spin_lock_irq(&conf
->device_lock
);
5076 wake_up(&conf
->wait_for_stripe
);
5077 wake_up(&conf
->wait_for_overlap
);
5078 spin_unlock_irq(&conf
->device_lock
);
5084 static void *raid5_takeover_raid1(mddev_t
*mddev
)
5088 if (mddev
->raid_disks
!= 2 ||
5089 mddev
->degraded
> 1)
5090 return ERR_PTR(-EINVAL
);
5092 /* Should check if there are write-behind devices? */
5094 chunksect
= 64*2; /* 64K by default */
5096 /* The array must be an exact multiple of chunksize */
5097 while (chunksect
&& (mddev
->array_sectors
& (chunksect
-1)))
5100 if ((chunksect
<<9) < STRIPE_SIZE
)
5101 /* array size does not allow a suitable chunk size */
5102 return ERR_PTR(-EINVAL
);
5104 mddev
->new_level
= 5;
5105 mddev
->new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5106 mddev
->new_chunk
= chunksect
<< 9;
5108 return setup_conf(mddev
);
5111 static void *raid5_takeover_raid6(mddev_t
*mddev
)
5115 switch (mddev
->layout
) {
5116 case ALGORITHM_LEFT_ASYMMETRIC_6
:
5117 new_layout
= ALGORITHM_LEFT_ASYMMETRIC
;
5119 case ALGORITHM_RIGHT_ASYMMETRIC_6
:
5120 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC
;
5122 case ALGORITHM_LEFT_SYMMETRIC_6
:
5123 new_layout
= ALGORITHM_LEFT_SYMMETRIC
;
5125 case ALGORITHM_RIGHT_SYMMETRIC_6
:
5126 new_layout
= ALGORITHM_RIGHT_SYMMETRIC
;
5128 case ALGORITHM_PARITY_0_6
:
5129 new_layout
= ALGORITHM_PARITY_0
;
5131 case ALGORITHM_PARITY_N
:
5132 new_layout
= ALGORITHM_PARITY_N
;
5135 return ERR_PTR(-EINVAL
);
5137 mddev
->new_level
= 5;
5138 mddev
->new_layout
= new_layout
;
5139 mddev
->delta_disks
= -1;
5140 mddev
->raid_disks
-= 1;
5141 return setup_conf(mddev
);
5145 static int raid5_reconfig(mddev_t
*mddev
, int new_layout
, int new_chunk
)
5147 /* For a 2-drive array, the layout and chunk size can be changed
5148 * immediately as not restriping is needed.
5149 * For larger arrays we record the new value - after validation
5150 * to be used by a reshape pass.
5152 raid5_conf_t
*conf
= mddev_to_conf(mddev
);
5154 if (new_layout
>= 0 && !algorithm_valid_raid5(new_layout
))
5156 if (new_chunk
> 0) {
5157 if (new_chunk
& (new_chunk
-1))
5158 /* not a power of 2 */
5160 if (new_chunk
< PAGE_SIZE
)
5162 if (mddev
->array_sectors
& ((new_chunk
>>9)-1))
5163 /* not factor of array size */
5167 /* They look valid */
5169 if (mddev
->raid_disks
== 2) {
5171 if (new_layout
>= 0) {
5172 conf
->algorithm
= new_layout
;
5173 mddev
->layout
= mddev
->new_layout
= new_layout
;
5175 if (new_chunk
> 0) {
5176 conf
->chunk_size
= new_chunk
;
5177 mddev
->chunk_size
= mddev
->new_chunk
= new_chunk
;
5179 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
5180 md_wakeup_thread(mddev
->thread
);
5182 if (new_layout
>= 0)
5183 mddev
->new_layout
= new_layout
;
5185 mddev
->new_chunk
= new_chunk
;
5190 static int raid6_reconfig(mddev_t
*mddev
, int new_layout
, int new_chunk
)
5192 if (new_layout
>= 0 && !algorithm_valid_raid6(new_layout
))
5194 if (new_chunk
> 0) {
5195 if (new_chunk
& (new_chunk
-1))
5196 /* not a power of 2 */
5198 if (new_chunk
< PAGE_SIZE
)
5200 if (mddev
->array_sectors
& ((new_chunk
>>9)-1))
5201 /* not factor of array size */
5205 /* They look valid */
5207 if (new_layout
>= 0)
5208 mddev
->new_layout
= new_layout
;
5210 mddev
->new_chunk
= new_chunk
;
5215 static void *raid5_takeover(mddev_t
*mddev
)
5217 /* raid5 can take over:
5218 * raid0 - if all devices are the same - make it a raid4 layout
5219 * raid1 - if there are two drives. We need to know the chunk size
5220 * raid4 - trivial - just use a raid4 layout.
5221 * raid6 - Providing it is a *_6 layout
5223 * For now, just do raid1
5226 if (mddev
->level
== 1)
5227 return raid5_takeover_raid1(mddev
);
5228 if (mddev
->level
== 4) {
5229 mddev
->new_layout
= ALGORITHM_PARITY_N
;
5230 mddev
->new_level
= 5;
5231 return setup_conf(mddev
);
5233 if (mddev
->level
== 6)
5234 return raid5_takeover_raid6(mddev
);
5236 return ERR_PTR(-EINVAL
);
5240 static struct mdk_personality raid5_personality
;
5242 static void *raid6_takeover(mddev_t
*mddev
)
5244 /* Currently can only take over a raid5. We map the
5245 * personality to an equivalent raid6 personality
5246 * with the Q block at the end.
5250 if (mddev
->pers
!= &raid5_personality
)
5251 return ERR_PTR(-EINVAL
);
5252 if (mddev
->degraded
> 1)
5253 return ERR_PTR(-EINVAL
);
5254 if (mddev
->raid_disks
> 253)
5255 return ERR_PTR(-EINVAL
);
5256 if (mddev
->raid_disks
< 3)
5257 return ERR_PTR(-EINVAL
);
5259 switch (mddev
->layout
) {
5260 case ALGORITHM_LEFT_ASYMMETRIC
:
5261 new_layout
= ALGORITHM_LEFT_ASYMMETRIC_6
;
5263 case ALGORITHM_RIGHT_ASYMMETRIC
:
5264 new_layout
= ALGORITHM_RIGHT_ASYMMETRIC_6
;
5266 case ALGORITHM_LEFT_SYMMETRIC
:
5267 new_layout
= ALGORITHM_LEFT_SYMMETRIC_6
;
5269 case ALGORITHM_RIGHT_SYMMETRIC
:
5270 new_layout
= ALGORITHM_RIGHT_SYMMETRIC_6
;
5272 case ALGORITHM_PARITY_0
:
5273 new_layout
= ALGORITHM_PARITY_0_6
;
5275 case ALGORITHM_PARITY_N
:
5276 new_layout
= ALGORITHM_PARITY_N
;
5279 return ERR_PTR(-EINVAL
);
5281 mddev
->new_level
= 6;
5282 mddev
->new_layout
= new_layout
;
5283 mddev
->delta_disks
= 1;
5284 mddev
->raid_disks
+= 1;
5285 return setup_conf(mddev
);
5289 static struct mdk_personality raid6_personality
=
5293 .owner
= THIS_MODULE
,
5294 .make_request
= make_request
,
5298 .error_handler
= error
,
5299 .hot_add_disk
= raid5_add_disk
,
5300 .hot_remove_disk
= raid5_remove_disk
,
5301 .spare_active
= raid5_spare_active
,
5302 .sync_request
= sync_request
,
5303 .resize
= raid5_resize
,
5305 .check_reshape
= raid5_check_reshape
,
5306 .start_reshape
= raid5_start_reshape
,
5307 .finish_reshape
= raid5_finish_reshape
,
5308 .quiesce
= raid5_quiesce
,
5309 .takeover
= raid6_takeover
,
5310 .reconfig
= raid6_reconfig
,
5312 static struct mdk_personality raid5_personality
=
5316 .owner
= THIS_MODULE
,
5317 .make_request
= make_request
,
5321 .error_handler
= error
,
5322 .hot_add_disk
= raid5_add_disk
,
5323 .hot_remove_disk
= raid5_remove_disk
,
5324 .spare_active
= raid5_spare_active
,
5325 .sync_request
= sync_request
,
5326 .resize
= raid5_resize
,
5328 .check_reshape
= raid5_check_reshape
,
5329 .start_reshape
= raid5_start_reshape
,
5330 .finish_reshape
= raid5_finish_reshape
,
5331 .quiesce
= raid5_quiesce
,
5332 .takeover
= raid5_takeover
,
5333 .reconfig
= raid5_reconfig
,
5336 static struct mdk_personality raid4_personality
=
5340 .owner
= THIS_MODULE
,
5341 .make_request
= make_request
,
5345 .error_handler
= error
,
5346 .hot_add_disk
= raid5_add_disk
,
5347 .hot_remove_disk
= raid5_remove_disk
,
5348 .spare_active
= raid5_spare_active
,
5349 .sync_request
= sync_request
,
5350 .resize
= raid5_resize
,
5352 .check_reshape
= raid5_check_reshape
,
5353 .start_reshape
= raid5_start_reshape
,
5354 .finish_reshape
= raid5_finish_reshape
,
5355 .quiesce
= raid5_quiesce
,
5358 static int __init
raid5_init(void)
5360 register_md_personality(&raid6_personality
);
5361 register_md_personality(&raid5_personality
);
5362 register_md_personality(&raid4_personality
);
5366 static void raid5_exit(void)
5368 unregister_md_personality(&raid6_personality
);
5369 unregister_md_personality(&raid5_personality
);
5370 unregister_md_personality(&raid4_personality
);
5373 module_init(raid5_init
);
5374 module_exit(raid5_exit
);
5375 MODULE_LICENSE("GPL");
5376 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5377 MODULE_ALIAS("md-raid5");
5378 MODULE_ALIAS("md-raid4");
5379 MODULE_ALIAS("md-level-5");
5380 MODULE_ALIAS("md-level-4");
5381 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5382 MODULE_ALIAS("md-raid6");
5383 MODULE_ALIAS("md-level-6");
5385 /* This used to be two separate modules, they were: */
5386 MODULE_ALIAS("raid5");
5387 MODULE_ALIAS("raid6");