2 * Copyright (C) 2015 Shaohua Li <shli@fb.com>
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms and conditions of the GNU General Public License,
6 * version 2, as published by the Free Software Foundation.
8 * This program is distributed in the hope it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 #include <linux/kernel.h>
15 #include <linux/wait.h>
16 #include <linux/blkdev.h>
17 #include <linux/slab.h>
18 #include <linux/raid/md_p.h>
19 #include <linux/crc32c.h>
20 #include <linux/random.h>
25 * metadata/data stored in disk with 4k size unit (a block) regardless
26 * underneath hardware sector size. only works with PAGE_SIZE == 4096
28 #define BLOCK_SECTORS (8)
31 * reclaim runs every 1/4 disk size or 10G reclaimable space. This can prevent
32 * recovery scans a very long log
34 #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
35 #define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
38 * We only need 2 bios per I/O unit to make progress, but ensure we
39 * have a few more available to not get too tight.
41 #define R5L_POOL_SIZE 4
48 sector_t device_size
; /* log device size, round to
50 sector_t max_free_space
; /* reclaim run if free space is at
53 sector_t last_checkpoint
; /* log tail. where recovery scan
55 u64 last_cp_seq
; /* log tail sequence */
57 sector_t log_start
; /* log head. where new data appends */
58 u64 seq
; /* log head sequence */
60 sector_t next_checkpoint
;
63 struct mutex io_mutex
;
64 struct r5l_io_unit
*current_io
; /* current io_unit accepting new data */
66 spinlock_t io_list_lock
;
67 struct list_head running_ios
; /* io_units which are still running,
68 * and have not yet been completely
69 * written to the log */
70 struct list_head io_end_ios
; /* io_units which have been completely
71 * written to the log but not yet written
73 struct list_head flushing_ios
; /* io_units which are waiting for log
75 struct list_head finished_ios
; /* io_units which settle down in log disk */
78 struct list_head no_mem_stripes
; /* pending stripes, -ENOMEM */
80 struct kmem_cache
*io_kc
;
85 struct md_thread
*reclaim_thread
;
86 unsigned long reclaim_target
; /* number of space that need to be
87 * reclaimed. if it's 0, reclaim spaces
88 * used by io_units which are in
89 * IO_UNIT_STRIPE_END state (eg, reclaim
90 * dones't wait for specific io_unit
91 * switching to IO_UNIT_STRIPE_END
93 wait_queue_head_t iounit_wait
;
95 struct list_head no_space_stripes
; /* pending stripes, log has no space */
96 spinlock_t no_space_stripes_lock
;
98 bool need_cache_flush
;
103 * an IO range starts from a meta data block and end at the next meta data
104 * block. The io unit's the meta data block tracks data/parity followed it. io
105 * unit is written to log disk with normal write, as we always flush log disk
106 * first and then start move data to raid disks, there is no requirement to
107 * write io unit with FLUSH/FUA
112 struct page
*meta_page
; /* store meta block */
113 int meta_offset
; /* current offset in meta_page */
115 struct bio
*current_bio
;/* current_bio accepting new data */
117 atomic_t pending_stripe
;/* how many stripes not flushed to raid */
118 u64 seq
; /* seq number of the metablock */
119 sector_t log_start
; /* where the io_unit starts */
120 sector_t log_end
; /* where the io_unit ends */
121 struct list_head log_sibling
; /* log->running_ios */
122 struct list_head stripe_list
; /* stripes added to the io_unit */
128 /* r5l_io_unit state */
129 enum r5l_io_unit_state
{
130 IO_UNIT_RUNNING
= 0, /* accepting new IO */
131 IO_UNIT_IO_START
= 1, /* io_unit bio start writing to log,
132 * don't accepting new bio */
133 IO_UNIT_IO_END
= 2, /* io_unit bio finish writing to log */
134 IO_UNIT_STRIPE_END
= 3, /* stripes data finished writing to raid */
137 static sector_t
r5l_ring_add(struct r5l_log
*log
, sector_t start
, sector_t inc
)
140 if (start
>= log
->device_size
)
141 start
= start
- log
->device_size
;
145 static sector_t
r5l_ring_distance(struct r5l_log
*log
, sector_t start
,
151 return end
+ log
->device_size
- start
;
154 static bool r5l_has_free_space(struct r5l_log
*log
, sector_t size
)
158 used_size
= r5l_ring_distance(log
, log
->last_checkpoint
,
161 return log
->device_size
> used_size
+ size
;
164 static void __r5l_set_io_unit_state(struct r5l_io_unit
*io
,
165 enum r5l_io_unit_state state
)
167 if (WARN_ON(io
->state
>= state
))
172 static void r5l_io_run_stripes(struct r5l_io_unit
*io
)
174 struct stripe_head
*sh
, *next
;
176 list_for_each_entry_safe(sh
, next
, &io
->stripe_list
, log_list
) {
177 list_del_init(&sh
->log_list
);
178 set_bit(STRIPE_HANDLE
, &sh
->state
);
179 raid5_release_stripe(sh
);
183 static void r5l_log_run_stripes(struct r5l_log
*log
)
185 struct r5l_io_unit
*io
, *next
;
187 assert_spin_locked(&log
->io_list_lock
);
189 list_for_each_entry_safe(io
, next
, &log
->running_ios
, log_sibling
) {
190 /* don't change list order */
191 if (io
->state
< IO_UNIT_IO_END
)
194 list_move_tail(&io
->log_sibling
, &log
->finished_ios
);
195 r5l_io_run_stripes(io
);
199 static void r5l_move_to_end_ios(struct r5l_log
*log
)
201 struct r5l_io_unit
*io
, *next
;
203 assert_spin_locked(&log
->io_list_lock
);
205 list_for_each_entry_safe(io
, next
, &log
->running_ios
, log_sibling
) {
206 /* don't change list order */
207 if (io
->state
< IO_UNIT_IO_END
)
209 list_move_tail(&io
->log_sibling
, &log
->io_end_ios
);
213 static void r5l_log_endio(struct bio
*bio
)
215 struct r5l_io_unit
*io
= bio
->bi_private
;
216 struct r5l_log
*log
= io
->log
;
220 md_error(log
->rdev
->mddev
, log
->rdev
);
223 mempool_free(io
->meta_page
, log
->meta_pool
);
225 spin_lock_irqsave(&log
->io_list_lock
, flags
);
226 __r5l_set_io_unit_state(io
, IO_UNIT_IO_END
);
227 if (log
->need_cache_flush
)
228 r5l_move_to_end_ios(log
);
230 r5l_log_run_stripes(log
);
231 spin_unlock_irqrestore(&log
->io_list_lock
, flags
);
233 if (log
->need_cache_flush
)
234 md_wakeup_thread(log
->rdev
->mddev
->thread
);
237 static void r5l_submit_current_io(struct r5l_log
*log
)
239 struct r5l_io_unit
*io
= log
->current_io
;
240 struct r5l_meta_block
*block
;
247 block
= page_address(io
->meta_page
);
248 block
->meta_size
= cpu_to_le32(io
->meta_offset
);
249 crc
= crc32c_le(log
->uuid_checksum
, block
, PAGE_SIZE
);
250 block
->checksum
= cpu_to_le32(crc
);
252 log
->current_io
= NULL
;
253 spin_lock_irqsave(&log
->io_list_lock
, flags
);
254 __r5l_set_io_unit_state(io
, IO_UNIT_IO_START
);
255 spin_unlock_irqrestore(&log
->io_list_lock
, flags
);
257 submit_bio(io
->current_bio
);
260 static struct bio
*r5l_bio_alloc(struct r5l_log
*log
)
262 struct bio
*bio
= bio_alloc_bioset(GFP_NOIO
, BIO_MAX_PAGES
, log
->bs
);
264 bio_set_op_attrs(bio
, REQ_OP_WRITE
, 0);
265 bio
->bi_bdev
= log
->rdev
->bdev
;
266 bio
->bi_iter
.bi_sector
= log
->rdev
->data_offset
+ log
->log_start
;
271 static void r5_reserve_log_entry(struct r5l_log
*log
, struct r5l_io_unit
*io
)
273 log
->log_start
= r5l_ring_add(log
, log
->log_start
, BLOCK_SECTORS
);
276 * If we filled up the log device start from the beginning again,
277 * which will require a new bio.
279 * Note: for this to work properly the log size needs to me a multiple
282 if (log
->log_start
== 0)
283 io
->need_split_bio
= true;
285 io
->log_end
= log
->log_start
;
288 static struct r5l_io_unit
*r5l_new_meta(struct r5l_log
*log
)
290 struct r5l_io_unit
*io
;
291 struct r5l_meta_block
*block
;
293 io
= mempool_alloc(log
->io_pool
, GFP_ATOMIC
);
296 memset(io
, 0, sizeof(*io
));
299 INIT_LIST_HEAD(&io
->log_sibling
);
300 INIT_LIST_HEAD(&io
->stripe_list
);
301 io
->state
= IO_UNIT_RUNNING
;
303 io
->meta_page
= mempool_alloc(log
->meta_pool
, GFP_NOIO
);
304 block
= page_address(io
->meta_page
);
306 block
->magic
= cpu_to_le32(R5LOG_MAGIC
);
307 block
->version
= R5LOG_VERSION
;
308 block
->seq
= cpu_to_le64(log
->seq
);
309 block
->position
= cpu_to_le64(log
->log_start
);
311 io
->log_start
= log
->log_start
;
312 io
->meta_offset
= sizeof(struct r5l_meta_block
);
313 io
->seq
= log
->seq
++;
315 io
->current_bio
= r5l_bio_alloc(log
);
316 io
->current_bio
->bi_end_io
= r5l_log_endio
;
317 io
->current_bio
->bi_private
= io
;
318 bio_add_page(io
->current_bio
, io
->meta_page
, PAGE_SIZE
, 0);
320 r5_reserve_log_entry(log
, io
);
322 spin_lock_irq(&log
->io_list_lock
);
323 list_add_tail(&io
->log_sibling
, &log
->running_ios
);
324 spin_unlock_irq(&log
->io_list_lock
);
329 static int r5l_get_meta(struct r5l_log
*log
, unsigned int payload_size
)
331 if (log
->current_io
&&
332 log
->current_io
->meta_offset
+ payload_size
> PAGE_SIZE
)
333 r5l_submit_current_io(log
);
335 if (!log
->current_io
) {
336 log
->current_io
= r5l_new_meta(log
);
337 if (!log
->current_io
)
344 static void r5l_append_payload_meta(struct r5l_log
*log
, u16 type
,
346 u32 checksum1
, u32 checksum2
,
347 bool checksum2_valid
)
349 struct r5l_io_unit
*io
= log
->current_io
;
350 struct r5l_payload_data_parity
*payload
;
352 payload
= page_address(io
->meta_page
) + io
->meta_offset
;
353 payload
->header
.type
= cpu_to_le16(type
);
354 payload
->header
.flags
= cpu_to_le16(0);
355 payload
->size
= cpu_to_le32((1 + !!checksum2_valid
) <<
357 payload
->location
= cpu_to_le64(location
);
358 payload
->checksum
[0] = cpu_to_le32(checksum1
);
360 payload
->checksum
[1] = cpu_to_le32(checksum2
);
362 io
->meta_offset
+= sizeof(struct r5l_payload_data_parity
) +
363 sizeof(__le32
) * (1 + !!checksum2_valid
);
366 static void r5l_append_payload_page(struct r5l_log
*log
, struct page
*page
)
368 struct r5l_io_unit
*io
= log
->current_io
;
370 if (io
->need_split_bio
) {
371 struct bio
*prev
= io
->current_bio
;
373 io
->current_bio
= r5l_bio_alloc(log
);
374 bio_chain(io
->current_bio
, prev
);
379 if (!bio_add_page(io
->current_bio
, page
, PAGE_SIZE
, 0))
382 r5_reserve_log_entry(log
, io
);
385 static int r5l_log_stripe(struct r5l_log
*log
, struct stripe_head
*sh
,
386 int data_pages
, int parity_pages
)
391 struct r5l_io_unit
*io
;
394 ((sizeof(struct r5l_payload_data_parity
) + sizeof(__le32
))
396 sizeof(struct r5l_payload_data_parity
) +
397 sizeof(__le32
) * parity_pages
;
399 ret
= r5l_get_meta(log
, meta_size
);
403 io
= log
->current_io
;
405 for (i
= 0; i
< sh
->disks
; i
++) {
406 if (!test_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
))
408 if (i
== sh
->pd_idx
|| i
== sh
->qd_idx
)
410 r5l_append_payload_meta(log
, R5LOG_PAYLOAD_DATA
,
411 raid5_compute_blocknr(sh
, i
, 0),
412 sh
->dev
[i
].log_checksum
, 0, false);
413 r5l_append_payload_page(log
, sh
->dev
[i
].page
);
416 if (sh
->qd_idx
>= 0) {
417 r5l_append_payload_meta(log
, R5LOG_PAYLOAD_PARITY
,
418 sh
->sector
, sh
->dev
[sh
->pd_idx
].log_checksum
,
419 sh
->dev
[sh
->qd_idx
].log_checksum
, true);
420 r5l_append_payload_page(log
, sh
->dev
[sh
->pd_idx
].page
);
421 r5l_append_payload_page(log
, sh
->dev
[sh
->qd_idx
].page
);
423 r5l_append_payload_meta(log
, R5LOG_PAYLOAD_PARITY
,
424 sh
->sector
, sh
->dev
[sh
->pd_idx
].log_checksum
,
426 r5l_append_payload_page(log
, sh
->dev
[sh
->pd_idx
].page
);
429 list_add_tail(&sh
->log_list
, &io
->stripe_list
);
430 atomic_inc(&io
->pending_stripe
);
436 static void r5l_wake_reclaim(struct r5l_log
*log
, sector_t space
);
438 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
439 * data from log to raid disks), so we shouldn't wait for reclaim here
441 int r5l_write_stripe(struct r5l_log
*log
, struct stripe_head
*sh
)
444 int data_pages
, parity_pages
;
452 /* Don't support stripe batch */
453 if (sh
->log_io
|| !test_bit(R5_Wantwrite
, &sh
->dev
[sh
->pd_idx
].flags
) ||
454 test_bit(STRIPE_SYNCING
, &sh
->state
)) {
455 /* the stripe is written to log, we start writing it to raid */
456 clear_bit(STRIPE_LOG_TRAPPED
, &sh
->state
);
460 for (i
= 0; i
< sh
->disks
; i
++) {
463 if (!test_bit(R5_Wantwrite
, &sh
->dev
[i
].flags
))
466 /* checksum is already calculated in last run */
467 if (test_bit(STRIPE_LOG_TRAPPED
, &sh
->state
))
469 addr
= kmap_atomic(sh
->dev
[i
].page
);
470 sh
->dev
[i
].log_checksum
= crc32c_le(log
->uuid_checksum
,
474 parity_pages
= 1 + !!(sh
->qd_idx
>= 0);
475 data_pages
= write_disks
- parity_pages
;
478 ((sizeof(struct r5l_payload_data_parity
) + sizeof(__le32
))
480 sizeof(struct r5l_payload_data_parity
) +
481 sizeof(__le32
) * parity_pages
;
482 /* Doesn't work with very big raid array */
483 if (meta_size
+ sizeof(struct r5l_meta_block
) > PAGE_SIZE
)
486 set_bit(STRIPE_LOG_TRAPPED
, &sh
->state
);
488 * The stripe must enter state machine again to finish the write, so
491 clear_bit(STRIPE_DELAYED
, &sh
->state
);
492 atomic_inc(&sh
->count
);
494 mutex_lock(&log
->io_mutex
);
496 reserve
= (1 + write_disks
) << (PAGE_SHIFT
- 9);
497 if (!r5l_has_free_space(log
, reserve
)) {
498 spin_lock(&log
->no_space_stripes_lock
);
499 list_add_tail(&sh
->log_list
, &log
->no_space_stripes
);
500 spin_unlock(&log
->no_space_stripes_lock
);
502 r5l_wake_reclaim(log
, reserve
);
504 ret
= r5l_log_stripe(log
, sh
, data_pages
, parity_pages
);
506 spin_lock_irq(&log
->io_list_lock
);
507 list_add_tail(&sh
->log_list
, &log
->no_mem_stripes
);
508 spin_unlock_irq(&log
->io_list_lock
);
512 mutex_unlock(&log
->io_mutex
);
516 void r5l_write_stripe_run(struct r5l_log
*log
)
520 mutex_lock(&log
->io_mutex
);
521 r5l_submit_current_io(log
);
522 mutex_unlock(&log
->io_mutex
);
525 int r5l_handle_flush_request(struct r5l_log
*log
, struct bio
*bio
)
530 * we flush log disk cache first, then write stripe data to raid disks.
531 * So if bio is finished, the log disk cache is flushed already. The
532 * recovery guarantees we can recovery the bio from log disk, so we
533 * don't need to flush again
535 if (bio
->bi_iter
.bi_size
== 0) {
539 bio
->bi_rw
&= ~REQ_PREFLUSH
;
543 /* This will run after log space is reclaimed */
544 static void r5l_run_no_space_stripes(struct r5l_log
*log
)
546 struct stripe_head
*sh
;
548 spin_lock(&log
->no_space_stripes_lock
);
549 while (!list_empty(&log
->no_space_stripes
)) {
550 sh
= list_first_entry(&log
->no_space_stripes
,
551 struct stripe_head
, log_list
);
552 list_del_init(&sh
->log_list
);
553 set_bit(STRIPE_HANDLE
, &sh
->state
);
554 raid5_release_stripe(sh
);
556 spin_unlock(&log
->no_space_stripes_lock
);
559 static sector_t
r5l_reclaimable_space(struct r5l_log
*log
)
561 return r5l_ring_distance(log
, log
->last_checkpoint
,
562 log
->next_checkpoint
);
565 static void r5l_run_no_mem_stripe(struct r5l_log
*log
)
567 struct stripe_head
*sh
;
569 assert_spin_locked(&log
->io_list_lock
);
571 if (!list_empty(&log
->no_mem_stripes
)) {
572 sh
= list_first_entry(&log
->no_mem_stripes
,
573 struct stripe_head
, log_list
);
574 list_del_init(&sh
->log_list
);
575 set_bit(STRIPE_HANDLE
, &sh
->state
);
576 raid5_release_stripe(sh
);
580 static bool r5l_complete_finished_ios(struct r5l_log
*log
)
582 struct r5l_io_unit
*io
, *next
;
585 assert_spin_locked(&log
->io_list_lock
);
587 list_for_each_entry_safe(io
, next
, &log
->finished_ios
, log_sibling
) {
588 /* don't change list order */
589 if (io
->state
< IO_UNIT_STRIPE_END
)
592 log
->next_checkpoint
= io
->log_start
;
593 log
->next_cp_seq
= io
->seq
;
595 list_del(&io
->log_sibling
);
596 mempool_free(io
, log
->io_pool
);
597 r5l_run_no_mem_stripe(log
);
605 static void __r5l_stripe_write_finished(struct r5l_io_unit
*io
)
607 struct r5l_log
*log
= io
->log
;
610 spin_lock_irqsave(&log
->io_list_lock
, flags
);
611 __r5l_set_io_unit_state(io
, IO_UNIT_STRIPE_END
);
613 if (!r5l_complete_finished_ios(log
)) {
614 spin_unlock_irqrestore(&log
->io_list_lock
, flags
);
618 if (r5l_reclaimable_space(log
) > log
->max_free_space
)
619 r5l_wake_reclaim(log
, 0);
621 spin_unlock_irqrestore(&log
->io_list_lock
, flags
);
622 wake_up(&log
->iounit_wait
);
625 void r5l_stripe_write_finished(struct stripe_head
*sh
)
627 struct r5l_io_unit
*io
;
632 if (io
&& atomic_dec_and_test(&io
->pending_stripe
))
633 __r5l_stripe_write_finished(io
);
636 static void r5l_log_flush_endio(struct bio
*bio
)
638 struct r5l_log
*log
= container_of(bio
, struct r5l_log
,
641 struct r5l_io_unit
*io
;
644 md_error(log
->rdev
->mddev
, log
->rdev
);
646 spin_lock_irqsave(&log
->io_list_lock
, flags
);
647 list_for_each_entry(io
, &log
->flushing_ios
, log_sibling
)
648 r5l_io_run_stripes(io
);
649 list_splice_tail_init(&log
->flushing_ios
, &log
->finished_ios
);
650 spin_unlock_irqrestore(&log
->io_list_lock
, flags
);
654 * Starting dispatch IO to raid.
655 * io_unit(meta) consists of a log. There is one situation we want to avoid. A
656 * broken meta in the middle of a log causes recovery can't find meta at the
657 * head of log. If operations require meta at the head persistent in log, we
658 * must make sure meta before it persistent in log too. A case is:
660 * stripe data/parity is in log, we start write stripe to raid disks. stripe
661 * data/parity must be persistent in log before we do the write to raid disks.
663 * The solution is we restrictly maintain io_unit list order. In this case, we
664 * only write stripes of an io_unit to raid disks till the io_unit is the first
665 * one whose data/parity is in log.
667 void r5l_flush_stripe_to_raid(struct r5l_log
*log
)
671 if (!log
|| !log
->need_cache_flush
)
674 spin_lock_irq(&log
->io_list_lock
);
675 /* flush bio is running */
676 if (!list_empty(&log
->flushing_ios
)) {
677 spin_unlock_irq(&log
->io_list_lock
);
680 list_splice_tail_init(&log
->io_end_ios
, &log
->flushing_ios
);
681 do_flush
= !list_empty(&log
->flushing_ios
);
682 spin_unlock_irq(&log
->io_list_lock
);
686 bio_reset(&log
->flush_bio
);
687 log
->flush_bio
.bi_bdev
= log
->rdev
->bdev
;
688 log
->flush_bio
.bi_end_io
= r5l_log_flush_endio
;
689 bio_set_op_attrs(&log
->flush_bio
, REQ_OP_WRITE
, WRITE_FLUSH
);
690 submit_bio(&log
->flush_bio
);
693 static void r5l_write_super(struct r5l_log
*log
, sector_t cp
);
694 static void r5l_write_super_and_discard_space(struct r5l_log
*log
,
697 struct block_device
*bdev
= log
->rdev
->bdev
;
700 r5l_write_super(log
, end
);
702 if (!blk_queue_discard(bdev_get_queue(bdev
)))
705 mddev
= log
->rdev
->mddev
;
707 * This is to avoid a deadlock. r5l_quiesce holds reconfig_mutex and
708 * wait for this thread to finish. This thread waits for
709 * MD_CHANGE_PENDING clear, which is supposed to be done in
710 * md_check_recovery(). md_check_recovery() tries to get
711 * reconfig_mutex. Since r5l_quiesce already holds the mutex,
712 * md_check_recovery() fails, so the PENDING never get cleared. The
713 * in_teardown check workaround this issue.
715 if (!log
->in_teardown
) {
716 set_mask_bits(&mddev
->flags
, 0,
717 BIT(MD_CHANGE_DEVS
) | BIT(MD_CHANGE_PENDING
));
718 md_wakeup_thread(mddev
->thread
);
719 wait_event(mddev
->sb_wait
,
720 !test_bit(MD_CHANGE_PENDING
, &mddev
->flags
) ||
723 * r5l_quiesce could run after in_teardown check and hold
724 * mutex first. Superblock might get updated twice.
726 if (log
->in_teardown
)
727 md_update_sb(mddev
, 1);
729 WARN_ON(!mddev_is_locked(mddev
));
730 md_update_sb(mddev
, 1);
733 /* discard IO error really doesn't matter, ignore it */
734 if (log
->last_checkpoint
< end
) {
735 blkdev_issue_discard(bdev
,
736 log
->last_checkpoint
+ log
->rdev
->data_offset
,
737 end
- log
->last_checkpoint
, GFP_NOIO
, 0);
739 blkdev_issue_discard(bdev
,
740 log
->last_checkpoint
+ log
->rdev
->data_offset
,
741 log
->device_size
- log
->last_checkpoint
,
743 blkdev_issue_discard(bdev
, log
->rdev
->data_offset
, end
,
749 static void r5l_do_reclaim(struct r5l_log
*log
)
751 sector_t reclaim_target
= xchg(&log
->reclaim_target
, 0);
752 sector_t reclaimable
;
753 sector_t next_checkpoint
;
756 spin_lock_irq(&log
->io_list_lock
);
758 * move proper io_unit to reclaim list. We should not change the order.
759 * reclaimable/unreclaimable io_unit can be mixed in the list, we
760 * shouldn't reuse space of an unreclaimable io_unit
763 reclaimable
= r5l_reclaimable_space(log
);
764 if (reclaimable
>= reclaim_target
||
765 (list_empty(&log
->running_ios
) &&
766 list_empty(&log
->io_end_ios
) &&
767 list_empty(&log
->flushing_ios
) &&
768 list_empty(&log
->finished_ios
)))
771 md_wakeup_thread(log
->rdev
->mddev
->thread
);
772 wait_event_lock_irq(log
->iounit_wait
,
773 r5l_reclaimable_space(log
) > reclaimable
,
777 next_checkpoint
= log
->next_checkpoint
;
778 next_cp_seq
= log
->next_cp_seq
;
779 spin_unlock_irq(&log
->io_list_lock
);
781 BUG_ON(reclaimable
< 0);
782 if (reclaimable
== 0)
786 * write_super will flush cache of each raid disk. We must write super
787 * here, because the log area might be reused soon and we don't want to
790 r5l_write_super_and_discard_space(log
, next_checkpoint
);
792 mutex_lock(&log
->io_mutex
);
793 log
->last_checkpoint
= next_checkpoint
;
794 log
->last_cp_seq
= next_cp_seq
;
795 mutex_unlock(&log
->io_mutex
);
797 r5l_run_no_space_stripes(log
);
800 static void r5l_reclaim_thread(struct md_thread
*thread
)
802 struct mddev
*mddev
= thread
->mddev
;
803 struct r5conf
*conf
= mddev
->private;
804 struct r5l_log
*log
= conf
->log
;
811 static void r5l_wake_reclaim(struct r5l_log
*log
, sector_t space
)
813 unsigned long target
;
814 unsigned long new = (unsigned long)space
; /* overflow in theory */
817 target
= log
->reclaim_target
;
820 } while (cmpxchg(&log
->reclaim_target
, target
, new) != target
);
821 md_wakeup_thread(log
->reclaim_thread
);
824 void r5l_quiesce(struct r5l_log
*log
, int state
)
827 if (!log
|| state
== 2)
830 log
->in_teardown
= 0;
832 * This is a special case for hotadd. In suspend, the array has
833 * no journal. In resume, journal is initialized as well as the
836 if (log
->reclaim_thread
)
838 log
->reclaim_thread
= md_register_thread(r5l_reclaim_thread
,
839 log
->rdev
->mddev
, "reclaim");
840 } else if (state
== 1) {
842 * at this point all stripes are finished, so io_unit is at
843 * least in STRIPE_END state
845 log
->in_teardown
= 1;
846 /* make sure r5l_write_super_and_discard_space exits */
847 mddev
= log
->rdev
->mddev
;
848 wake_up(&mddev
->sb_wait
);
849 r5l_wake_reclaim(log
, -1L);
850 md_unregister_thread(&log
->reclaim_thread
);
855 bool r5l_log_disk_error(struct r5conf
*conf
)
859 /* don't allow write if journal disk is missing */
861 log
= rcu_dereference(conf
->log
);
864 ret
= test_bit(MD_HAS_JOURNAL
, &conf
->mddev
->flags
);
866 ret
= test_bit(Faulty
, &log
->rdev
->flags
);
871 struct r5l_recovery_ctx
{
872 struct page
*meta_page
; /* current meta */
873 sector_t meta_total_blocks
; /* total size of current meta and data */
874 sector_t pos
; /* recovery position */
875 u64 seq
; /* recovery position seq */
878 static int r5l_read_meta_block(struct r5l_log
*log
,
879 struct r5l_recovery_ctx
*ctx
)
881 struct page
*page
= ctx
->meta_page
;
882 struct r5l_meta_block
*mb
;
885 if (!sync_page_io(log
->rdev
, ctx
->pos
, PAGE_SIZE
, page
, REQ_OP_READ
, 0,
889 mb
= page_address(page
);
890 stored_crc
= le32_to_cpu(mb
->checksum
);
893 if (le32_to_cpu(mb
->magic
) != R5LOG_MAGIC
||
894 le64_to_cpu(mb
->seq
) != ctx
->seq
||
895 mb
->version
!= R5LOG_VERSION
||
896 le64_to_cpu(mb
->position
) != ctx
->pos
)
899 crc
= crc32c_le(log
->uuid_checksum
, mb
, PAGE_SIZE
);
900 if (stored_crc
!= crc
)
903 if (le32_to_cpu(mb
->meta_size
) > PAGE_SIZE
)
906 ctx
->meta_total_blocks
= BLOCK_SECTORS
;
911 static int r5l_recovery_flush_one_stripe(struct r5l_log
*log
,
912 struct r5l_recovery_ctx
*ctx
,
913 sector_t stripe_sect
,
914 int *offset
, sector_t
*log_offset
)
916 struct r5conf
*conf
= log
->rdev
->mddev
->private;
917 struct stripe_head
*sh
;
918 struct r5l_payload_data_parity
*payload
;
921 sh
= raid5_get_active_stripe(conf
, stripe_sect
, 0, 0, 0);
923 payload
= page_address(ctx
->meta_page
) + *offset
;
925 if (le16_to_cpu(payload
->header
.type
) == R5LOG_PAYLOAD_DATA
) {
926 raid5_compute_sector(conf
,
927 le64_to_cpu(payload
->location
), 0,
930 sync_page_io(log
->rdev
, *log_offset
, PAGE_SIZE
,
931 sh
->dev
[disk_index
].page
, REQ_OP_READ
, 0,
933 sh
->dev
[disk_index
].log_checksum
=
934 le32_to_cpu(payload
->checksum
[0]);
935 set_bit(R5_Wantwrite
, &sh
->dev
[disk_index
].flags
);
936 ctx
->meta_total_blocks
+= BLOCK_SECTORS
;
938 disk_index
= sh
->pd_idx
;
939 sync_page_io(log
->rdev
, *log_offset
, PAGE_SIZE
,
940 sh
->dev
[disk_index
].page
, REQ_OP_READ
, 0,
942 sh
->dev
[disk_index
].log_checksum
=
943 le32_to_cpu(payload
->checksum
[0]);
944 set_bit(R5_Wantwrite
, &sh
->dev
[disk_index
].flags
);
946 if (sh
->qd_idx
>= 0) {
947 disk_index
= sh
->qd_idx
;
948 sync_page_io(log
->rdev
,
949 r5l_ring_add(log
, *log_offset
, BLOCK_SECTORS
),
950 PAGE_SIZE
, sh
->dev
[disk_index
].page
,
951 REQ_OP_READ
, 0, false);
952 sh
->dev
[disk_index
].log_checksum
=
953 le32_to_cpu(payload
->checksum
[1]);
954 set_bit(R5_Wantwrite
,
955 &sh
->dev
[disk_index
].flags
);
957 ctx
->meta_total_blocks
+= BLOCK_SECTORS
* conf
->max_degraded
;
960 *log_offset
= r5l_ring_add(log
, *log_offset
,
961 le32_to_cpu(payload
->size
));
962 *offset
+= sizeof(struct r5l_payload_data_parity
) +
964 (le32_to_cpu(payload
->size
) >> (PAGE_SHIFT
- 9));
965 if (le16_to_cpu(payload
->header
.type
) == R5LOG_PAYLOAD_PARITY
)
969 for (disk_index
= 0; disk_index
< sh
->disks
; disk_index
++) {
973 if (!test_bit(R5_Wantwrite
, &sh
->dev
[disk_index
].flags
))
975 addr
= kmap_atomic(sh
->dev
[disk_index
].page
);
976 checksum
= crc32c_le(log
->uuid_checksum
, addr
, PAGE_SIZE
);
978 if (checksum
!= sh
->dev
[disk_index
].log_checksum
)
982 for (disk_index
= 0; disk_index
< sh
->disks
; disk_index
++) {
983 struct md_rdev
*rdev
, *rrdev
;
985 if (!test_and_clear_bit(R5_Wantwrite
,
986 &sh
->dev
[disk_index
].flags
))
989 /* in case device is broken */
990 rdev
= rcu_dereference(conf
->disks
[disk_index
].rdev
);
992 sync_page_io(rdev
, stripe_sect
, PAGE_SIZE
,
993 sh
->dev
[disk_index
].page
, REQ_OP_WRITE
, 0,
995 rrdev
= rcu_dereference(conf
->disks
[disk_index
].replacement
);
997 sync_page_io(rrdev
, stripe_sect
, PAGE_SIZE
,
998 sh
->dev
[disk_index
].page
, REQ_OP_WRITE
, 0,
1001 raid5_release_stripe(sh
);
1005 for (disk_index
= 0; disk_index
< sh
->disks
; disk_index
++)
1006 sh
->dev
[disk_index
].flags
= 0;
1007 raid5_release_stripe(sh
);
1011 static int r5l_recovery_flush_one_meta(struct r5l_log
*log
,
1012 struct r5l_recovery_ctx
*ctx
)
1014 struct r5conf
*conf
= log
->rdev
->mddev
->private;
1015 struct r5l_payload_data_parity
*payload
;
1016 struct r5l_meta_block
*mb
;
1018 sector_t log_offset
;
1019 sector_t stripe_sector
;
1021 mb
= page_address(ctx
->meta_page
);
1022 offset
= sizeof(struct r5l_meta_block
);
1023 log_offset
= r5l_ring_add(log
, ctx
->pos
, BLOCK_SECTORS
);
1025 while (offset
< le32_to_cpu(mb
->meta_size
)) {
1028 payload
= (void *)mb
+ offset
;
1029 stripe_sector
= raid5_compute_sector(conf
,
1030 le64_to_cpu(payload
->location
), 0, &dd
, NULL
);
1031 if (r5l_recovery_flush_one_stripe(log
, ctx
, stripe_sector
,
1032 &offset
, &log_offset
))
1038 /* copy data/parity from log to raid disks */
1039 static void r5l_recovery_flush_log(struct r5l_log
*log
,
1040 struct r5l_recovery_ctx
*ctx
)
1043 if (r5l_read_meta_block(log
, ctx
))
1045 if (r5l_recovery_flush_one_meta(log
, ctx
))
1048 ctx
->pos
= r5l_ring_add(log
, ctx
->pos
, ctx
->meta_total_blocks
);
1052 static int r5l_log_write_empty_meta_block(struct r5l_log
*log
, sector_t pos
,
1056 struct r5l_meta_block
*mb
;
1059 page
= alloc_page(GFP_KERNEL
| __GFP_ZERO
);
1062 mb
= page_address(page
);
1063 mb
->magic
= cpu_to_le32(R5LOG_MAGIC
);
1064 mb
->version
= R5LOG_VERSION
;
1065 mb
->meta_size
= cpu_to_le32(sizeof(struct r5l_meta_block
));
1066 mb
->seq
= cpu_to_le64(seq
);
1067 mb
->position
= cpu_to_le64(pos
);
1068 crc
= crc32c_le(log
->uuid_checksum
, mb
, PAGE_SIZE
);
1069 mb
->checksum
= cpu_to_le32(crc
);
1071 if (!sync_page_io(log
->rdev
, pos
, PAGE_SIZE
, page
, REQ_OP_WRITE
,
1072 WRITE_FUA
, false)) {
1080 static int r5l_recovery_log(struct r5l_log
*log
)
1082 struct r5l_recovery_ctx ctx
;
1084 ctx
.pos
= log
->last_checkpoint
;
1085 ctx
.seq
= log
->last_cp_seq
;
1086 ctx
.meta_page
= alloc_page(GFP_KERNEL
);
1090 r5l_recovery_flush_log(log
, &ctx
);
1091 __free_page(ctx
.meta_page
);
1094 * we did a recovery. Now ctx.pos points to an invalid meta block. New
1095 * log will start here. but we can't let superblock point to last valid
1096 * meta block. The log might looks like:
1097 * | meta 1| meta 2| meta 3|
1098 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
1099 * superblock points to meta 1, we write a new valid meta 2n. if crash
1100 * happens again, new recovery will start from meta 1. Since meta 2n is
1101 * valid now, recovery will think meta 3 is valid, which is wrong.
1102 * The solution is we create a new meta in meta2 with its seq == meta
1103 * 1's seq + 10 and let superblock points to meta2. The same recovery will
1104 * not think meta 3 is a valid meta, because its seq doesn't match
1106 if (ctx
.seq
> log
->last_cp_seq
+ 1) {
1109 ret
= r5l_log_write_empty_meta_block(log
, ctx
.pos
, ctx
.seq
+ 10);
1112 log
->seq
= ctx
.seq
+ 11;
1113 log
->log_start
= r5l_ring_add(log
, ctx
.pos
, BLOCK_SECTORS
);
1114 r5l_write_super(log
, ctx
.pos
);
1116 log
->log_start
= ctx
.pos
;
1122 static void r5l_write_super(struct r5l_log
*log
, sector_t cp
)
1124 struct mddev
*mddev
= log
->rdev
->mddev
;
1126 log
->rdev
->journal_tail
= cp
;
1127 set_bit(MD_CHANGE_DEVS
, &mddev
->flags
);
1130 static int r5l_load_log(struct r5l_log
*log
)
1132 struct md_rdev
*rdev
= log
->rdev
;
1134 struct r5l_meta_block
*mb
;
1135 sector_t cp
= log
->rdev
->journal_tail
;
1136 u32 stored_crc
, expected_crc
;
1137 bool create_super
= false;
1140 /* Make sure it's valid */
1141 if (cp
>= rdev
->sectors
|| round_down(cp
, BLOCK_SECTORS
) != cp
)
1143 page
= alloc_page(GFP_KERNEL
);
1147 if (!sync_page_io(rdev
, cp
, PAGE_SIZE
, page
, REQ_OP_READ
, 0, false)) {
1151 mb
= page_address(page
);
1153 if (le32_to_cpu(mb
->magic
) != R5LOG_MAGIC
||
1154 mb
->version
!= R5LOG_VERSION
) {
1155 create_super
= true;
1158 stored_crc
= le32_to_cpu(mb
->checksum
);
1160 expected_crc
= crc32c_le(log
->uuid_checksum
, mb
, PAGE_SIZE
);
1161 if (stored_crc
!= expected_crc
) {
1162 create_super
= true;
1165 if (le64_to_cpu(mb
->position
) != cp
) {
1166 create_super
= true;
1171 log
->last_cp_seq
= prandom_u32();
1174 * Make sure super points to correct address. Log might have
1175 * data very soon. If super hasn't correct log tail address,
1176 * recovery can't find the log
1178 r5l_write_super(log
, cp
);
1180 log
->last_cp_seq
= le64_to_cpu(mb
->seq
);
1182 log
->device_size
= round_down(rdev
->sectors
, BLOCK_SECTORS
);
1183 log
->max_free_space
= log
->device_size
>> RECLAIM_MAX_FREE_SPACE_SHIFT
;
1184 if (log
->max_free_space
> RECLAIM_MAX_FREE_SPACE
)
1185 log
->max_free_space
= RECLAIM_MAX_FREE_SPACE
;
1186 log
->last_checkpoint
= cp
;
1190 return r5l_recovery_log(log
);
1196 int r5l_init_log(struct r5conf
*conf
, struct md_rdev
*rdev
)
1198 struct request_queue
*q
= bdev_get_queue(rdev
->bdev
);
1199 struct r5l_log
*log
;
1201 if (PAGE_SIZE
!= 4096)
1203 log
= kzalloc(sizeof(*log
), GFP_KERNEL
);
1208 log
->need_cache_flush
= test_bit(QUEUE_FLAG_WC
, &q
->queue_flags
) != 0;
1210 log
->uuid_checksum
= crc32c_le(~0, rdev
->mddev
->uuid
,
1211 sizeof(rdev
->mddev
->uuid
));
1213 mutex_init(&log
->io_mutex
);
1215 spin_lock_init(&log
->io_list_lock
);
1216 INIT_LIST_HEAD(&log
->running_ios
);
1217 INIT_LIST_HEAD(&log
->io_end_ios
);
1218 INIT_LIST_HEAD(&log
->flushing_ios
);
1219 INIT_LIST_HEAD(&log
->finished_ios
);
1220 bio_init(&log
->flush_bio
);
1222 log
->io_kc
= KMEM_CACHE(r5l_io_unit
, 0);
1226 log
->io_pool
= mempool_create_slab_pool(R5L_POOL_SIZE
, log
->io_kc
);
1230 log
->bs
= bioset_create(R5L_POOL_SIZE
, 0);
1234 log
->meta_pool
= mempool_create_page_pool(R5L_POOL_SIZE
, 0);
1235 if (!log
->meta_pool
)
1238 log
->reclaim_thread
= md_register_thread(r5l_reclaim_thread
,
1239 log
->rdev
->mddev
, "reclaim");
1240 if (!log
->reclaim_thread
)
1241 goto reclaim_thread
;
1242 init_waitqueue_head(&log
->iounit_wait
);
1244 INIT_LIST_HEAD(&log
->no_mem_stripes
);
1246 INIT_LIST_HEAD(&log
->no_space_stripes
);
1247 spin_lock_init(&log
->no_space_stripes_lock
);
1249 if (r5l_load_log(log
))
1252 rcu_assign_pointer(conf
->log
, log
);
1253 set_bit(MD_HAS_JOURNAL
, &conf
->mddev
->flags
);
1257 md_unregister_thread(&log
->reclaim_thread
);
1259 mempool_destroy(log
->meta_pool
);
1261 bioset_free(log
->bs
);
1263 mempool_destroy(log
->io_pool
);
1265 kmem_cache_destroy(log
->io_kc
);
1271 void r5l_exit_log(struct r5l_log
*log
)
1273 md_unregister_thread(&log
->reclaim_thread
);
1274 mempool_destroy(log
->meta_pool
);
1275 bioset_free(log
->bs
);
1276 mempool_destroy(log
->io_pool
);
1277 kmem_cache_destroy(log
->io_kc
);