2 * Copyright (C) 2002 Sistina Software (UK) Limited.
3 * Copyright (C) 2006 Red Hat GmbH
5 * This file is released under the GPL.
7 * Kcopyd provides a simple interface for copying an area of one
8 * block-device to one or more other block-devices, with an asynchronous
9 * completion notification.
12 #include <linux/types.h>
13 #include <linux/atomic.h>
14 #include <linux/blkdev.h>
16 #include <linux/init.h>
17 #include <linux/list.h>
18 #include <linux/mempool.h>
19 #include <linux/module.h>
20 #include <linux/pagemap.h>
21 #include <linux/slab.h>
22 #include <linux/vmalloc.h>
23 #include <linux/workqueue.h>
24 #include <linux/mutex.h>
25 #include <linux/delay.h>
26 #include <linux/device-mapper.h>
27 #include <linux/dm-kcopyd.h>
31 #define SUB_JOB_SIZE 128
34 #define RESERVE_PAGES (DIV_ROUND_UP(SUB_JOB_SIZE << SECTOR_SHIFT, PAGE_SIZE))
36 /*-----------------------------------------------------------------
37 * Each kcopyd client has its own little pool of preallocated
38 * pages for kcopyd io.
39 *---------------------------------------------------------------*/
40 struct dm_kcopyd_client
{
41 struct page_list
*pages
;
42 unsigned nr_reserved_pages
;
43 unsigned nr_free_pages
;
45 struct dm_io_client
*io_client
;
47 wait_queue_head_t destroyq
;
52 struct workqueue_struct
*kcopyd_wq
;
53 struct work_struct kcopyd_work
;
55 struct dm_kcopyd_throttle
*throttle
;
58 * We maintain three lists of jobs:
60 * i) jobs waiting for pages
61 * ii) jobs that have pages, and are waiting for the io to be issued.
62 * iii) jobs that have completed.
64 * All three of these are protected by job_lock.
67 struct list_head complete_jobs
;
68 struct list_head io_jobs
;
69 struct list_head pages_jobs
;
72 static struct page_list zero_page_list
;
74 static DEFINE_SPINLOCK(throttle_spinlock
);
77 * IO/IDLE accounting slowly decays after (1 << ACCOUNT_INTERVAL_SHIFT) period.
78 * When total_period >= (1 << ACCOUNT_INTERVAL_SHIFT) the counters are divided
81 #define ACCOUNT_INTERVAL_SHIFT SHIFT_HZ
84 * Sleep this number of milliseconds.
86 * The value was decided experimentally.
87 * Smaller values seem to cause an increased copy rate above the limit.
88 * The reason for this is unknown but possibly due to jiffies rounding errors
89 * or read/write cache inside the disk.
91 #define SLEEP_MSEC 100
94 * Maximum number of sleep events. There is a theoretical livelock if more
95 * kcopyd clients do work simultaneously which this limit avoids.
99 static void io_job_start(struct dm_kcopyd_throttle
*t
)
101 unsigned throttle
, now
, difference
;
108 spin_lock_irq(&throttle_spinlock
);
110 throttle
= ACCESS_ONCE(t
->throttle
);
112 if (likely(throttle
>= 100))
116 difference
= now
- t
->last_jiffies
;
117 t
->last_jiffies
= now
;
119 t
->io_period
+= difference
;
120 t
->total_period
+= difference
;
123 * Maintain sane values if we got a temporary overflow.
125 if (unlikely(t
->io_period
> t
->total_period
))
126 t
->io_period
= t
->total_period
;
128 if (unlikely(t
->total_period
>= (1 << ACCOUNT_INTERVAL_SHIFT
))) {
129 int shift
= fls(t
->total_period
>> ACCOUNT_INTERVAL_SHIFT
);
130 t
->total_period
>>= shift
;
131 t
->io_period
>>= shift
;
134 skew
= t
->io_period
- throttle
* t
->total_period
/ 100;
136 if (unlikely(skew
> 0) && slept
< MAX_SLEEPS
) {
138 spin_unlock_irq(&throttle_spinlock
);
146 spin_unlock_irq(&throttle_spinlock
);
149 static void io_job_finish(struct dm_kcopyd_throttle
*t
)
156 spin_lock_irqsave(&throttle_spinlock
, flags
);
160 if (likely(ACCESS_ONCE(t
->throttle
) >= 100))
163 if (!t
->num_io_jobs
) {
164 unsigned now
, difference
;
167 difference
= now
- t
->last_jiffies
;
168 t
->last_jiffies
= now
;
170 t
->io_period
+= difference
;
171 t
->total_period
+= difference
;
174 * Maintain sane values if we got a temporary overflow.
176 if (unlikely(t
->io_period
> t
->total_period
))
177 t
->io_period
= t
->total_period
;
181 spin_unlock_irqrestore(&throttle_spinlock
, flags
);
185 static void wake(struct dm_kcopyd_client
*kc
)
187 queue_work(kc
->kcopyd_wq
, &kc
->kcopyd_work
);
191 * Obtain one page for the use of kcopyd.
193 static struct page_list
*alloc_pl(gfp_t gfp
)
195 struct page_list
*pl
;
197 pl
= kmalloc(sizeof(*pl
), gfp
);
201 pl
->page
= alloc_page(gfp
);
210 static void free_pl(struct page_list
*pl
)
212 __free_page(pl
->page
);
217 * Add the provided pages to a client's free page list, releasing
218 * back to the system any beyond the reserved_pages limit.
220 static void kcopyd_put_pages(struct dm_kcopyd_client
*kc
, struct page_list
*pl
)
222 struct page_list
*next
;
227 if (kc
->nr_free_pages
>= kc
->nr_reserved_pages
)
230 pl
->next
= kc
->pages
;
239 static int kcopyd_get_pages(struct dm_kcopyd_client
*kc
,
240 unsigned int nr
, struct page_list
**pages
)
242 struct page_list
*pl
;
247 pl
= alloc_pl(__GFP_NOWARN
| __GFP_NORETRY
| __GFP_KSWAPD_RECLAIM
);
249 /* Use reserved pages */
253 kc
->pages
= pl
->next
;
264 kcopyd_put_pages(kc
, *pages
);
269 * These three functions resize the page pool.
271 static void drop_pages(struct page_list
*pl
)
273 struct page_list
*next
;
283 * Allocate and reserve nr_pages for the use of a specific client.
285 static int client_reserve_pages(struct dm_kcopyd_client
*kc
, unsigned nr_pages
)
288 struct page_list
*pl
= NULL
, *next
;
290 for (i
= 0; i
< nr_pages
; i
++) {
291 next
= alloc_pl(GFP_KERNEL
);
301 kc
->nr_reserved_pages
+= nr_pages
;
302 kcopyd_put_pages(kc
, pl
);
307 static void client_free_pages(struct dm_kcopyd_client
*kc
)
309 BUG_ON(kc
->nr_free_pages
!= kc
->nr_reserved_pages
);
310 drop_pages(kc
->pages
);
312 kc
->nr_free_pages
= kc
->nr_reserved_pages
= 0;
315 /*-----------------------------------------------------------------
316 * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
317 * for this reason we use a mempool to prevent the client from
318 * ever having to do io (which could cause a deadlock).
319 *---------------------------------------------------------------*/
321 struct dm_kcopyd_client
*kc
;
322 struct list_head list
;
326 * Error state of the job.
329 unsigned long write_err
;
332 * Either READ or WRITE
335 struct dm_io_region source
;
338 * The destinations for the transfer.
340 unsigned int num_dests
;
341 struct dm_io_region dests
[DM_KCOPYD_MAX_REGIONS
];
343 struct page_list
*pages
;
346 * Set this to ensure you are notified when the job has
347 * completed. 'context' is for callback to use.
349 dm_kcopyd_notify_fn fn
;
353 * These fields are only used if the job has been split
354 * into more manageable parts.
359 sector_t write_offset
;
361 struct kcopyd_job
*master_job
;
364 static struct kmem_cache
*_job_cache
;
366 int __init
dm_kcopyd_init(void)
368 _job_cache
= kmem_cache_create("kcopyd_job",
369 sizeof(struct kcopyd_job
) * (SPLIT_COUNT
+ 1),
370 __alignof__(struct kcopyd_job
), 0, NULL
);
374 zero_page_list
.next
= &zero_page_list
;
375 zero_page_list
.page
= ZERO_PAGE(0);
380 void dm_kcopyd_exit(void)
382 kmem_cache_destroy(_job_cache
);
387 * Functions to push and pop a job onto the head of a given job
390 static struct kcopyd_job
*pop_io_job(struct list_head
*jobs
,
391 struct dm_kcopyd_client
*kc
)
393 struct kcopyd_job
*job
;
396 * For I/O jobs, pop any read, any write without sequential write
397 * constraint and sequential writes that are at the right position.
399 list_for_each_entry(job
, jobs
, list
) {
400 if (job
->rw
== READ
|| !test_bit(DM_KCOPYD_WRITE_SEQ
, &job
->flags
)) {
401 list_del(&job
->list
);
405 if (job
->write_offset
== job
->master_job
->write_offset
) {
406 job
->master_job
->write_offset
+= job
->source
.count
;
407 list_del(&job
->list
);
415 static struct kcopyd_job
*pop(struct list_head
*jobs
,
416 struct dm_kcopyd_client
*kc
)
418 struct kcopyd_job
*job
= NULL
;
421 spin_lock_irqsave(&kc
->job_lock
, flags
);
423 if (!list_empty(jobs
)) {
424 if (jobs
== &kc
->io_jobs
)
425 job
= pop_io_job(jobs
, kc
);
427 job
= list_entry(jobs
->next
, struct kcopyd_job
, list
);
428 list_del(&job
->list
);
431 spin_unlock_irqrestore(&kc
->job_lock
, flags
);
436 static void push(struct list_head
*jobs
, struct kcopyd_job
*job
)
439 struct dm_kcopyd_client
*kc
= job
->kc
;
441 spin_lock_irqsave(&kc
->job_lock
, flags
);
442 list_add_tail(&job
->list
, jobs
);
443 spin_unlock_irqrestore(&kc
->job_lock
, flags
);
447 static void push_head(struct list_head
*jobs
, struct kcopyd_job
*job
)
450 struct dm_kcopyd_client
*kc
= job
->kc
;
452 spin_lock_irqsave(&kc
->job_lock
, flags
);
453 list_add(&job
->list
, jobs
);
454 spin_unlock_irqrestore(&kc
->job_lock
, flags
);
458 * These three functions process 1 item from the corresponding
464 * > 0: can't process yet.
466 static int run_complete_job(struct kcopyd_job
*job
)
468 void *context
= job
->context
;
469 int read_err
= job
->read_err
;
470 unsigned long write_err
= job
->write_err
;
471 dm_kcopyd_notify_fn fn
= job
->fn
;
472 struct dm_kcopyd_client
*kc
= job
->kc
;
474 if (job
->pages
&& job
->pages
!= &zero_page_list
)
475 kcopyd_put_pages(kc
, job
->pages
);
477 * If this is the master job, the sub jobs have already
478 * completed so we can free everything.
480 if (job
->master_job
== job
)
481 mempool_free(job
, kc
->job_pool
);
482 fn(read_err
, write_err
, context
);
484 if (atomic_dec_and_test(&kc
->nr_jobs
))
485 wake_up(&kc
->destroyq
);
490 static void complete_io(unsigned long error
, void *context
)
492 struct kcopyd_job
*job
= (struct kcopyd_job
*) context
;
493 struct dm_kcopyd_client
*kc
= job
->kc
;
495 io_job_finish(kc
->throttle
);
498 if (op_is_write(job
->rw
))
499 job
->write_err
|= error
;
503 if (!test_bit(DM_KCOPYD_IGNORE_ERROR
, &job
->flags
)) {
504 push(&kc
->complete_jobs
, job
);
510 if (op_is_write(job
->rw
))
511 push(&kc
->complete_jobs
, job
);
515 push(&kc
->io_jobs
, job
);
522 * Request io on as many buffer heads as we can currently get for
525 static int run_io_job(struct kcopyd_job
*job
)
528 struct dm_io_request io_req
= {
531 .mem
.type
= DM_IO_PAGE_LIST
,
532 .mem
.ptr
.pl
= job
->pages
,
534 .notify
.fn
= complete_io
,
535 .notify
.context
= job
,
536 .client
= job
->kc
->io_client
,
540 * If we need to write sequentially and some reads or writes failed,
541 * no point in continuing.
543 if (test_bit(DM_KCOPYD_WRITE_SEQ
, &job
->flags
) &&
544 job
->master_job
->write_err
)
547 io_job_start(job
->kc
->throttle
);
550 r
= dm_io(&io_req
, 1, &job
->source
, NULL
);
552 r
= dm_io(&io_req
, job
->num_dests
, job
->dests
, NULL
);
557 static int run_pages_job(struct kcopyd_job
*job
)
560 unsigned nr_pages
= dm_div_up(job
->dests
[0].count
, PAGE_SIZE
>> 9);
562 r
= kcopyd_get_pages(job
->kc
, nr_pages
, &job
->pages
);
564 /* this job is ready for io */
565 push(&job
->kc
->io_jobs
, job
);
570 /* can't complete now */
577 * Run through a list for as long as possible. Returns the count
578 * of successful jobs.
580 static int process_jobs(struct list_head
*jobs
, struct dm_kcopyd_client
*kc
,
581 int (*fn
) (struct kcopyd_job
*))
583 struct kcopyd_job
*job
;
586 while ((job
= pop(jobs
, kc
))) {
591 /* error this rogue job */
592 if (op_is_write(job
->rw
))
593 job
->write_err
= (unsigned long) -1L;
596 push(&kc
->complete_jobs
, job
);
602 * We couldn't service this job ATM, so
603 * push this job back onto the list.
605 push_head(jobs
, job
);
616 * kcopyd does this every time it's woken up.
618 static void do_work(struct work_struct
*work
)
620 struct dm_kcopyd_client
*kc
= container_of(work
,
621 struct dm_kcopyd_client
, kcopyd_work
);
622 struct blk_plug plug
;
625 * The order that these are called is *very* important.
626 * complete jobs can free some pages for pages jobs.
627 * Pages jobs when successful will jump onto the io jobs
628 * list. io jobs call wake when they complete and it all
631 blk_start_plug(&plug
);
632 process_jobs(&kc
->complete_jobs
, kc
, run_complete_job
);
633 process_jobs(&kc
->pages_jobs
, kc
, run_pages_job
);
634 process_jobs(&kc
->io_jobs
, kc
, run_io_job
);
635 blk_finish_plug(&plug
);
639 * If we are copying a small region we just dispatch a single job
640 * to do the copy, otherwise the io has to be split up into many
643 static void dispatch_job(struct kcopyd_job
*job
)
645 struct dm_kcopyd_client
*kc
= job
->kc
;
646 atomic_inc(&kc
->nr_jobs
);
647 if (unlikely(!job
->source
.count
))
648 push(&kc
->complete_jobs
, job
);
649 else if (job
->pages
== &zero_page_list
)
650 push(&kc
->io_jobs
, job
);
652 push(&kc
->pages_jobs
, job
);
656 static void segment_complete(int read_err
, unsigned long write_err
,
659 /* FIXME: tidy this function */
660 sector_t progress
= 0;
662 struct kcopyd_job
*sub_job
= (struct kcopyd_job
*) context
;
663 struct kcopyd_job
*job
= sub_job
->master_job
;
664 struct dm_kcopyd_client
*kc
= job
->kc
;
666 mutex_lock(&job
->lock
);
668 /* update the error */
673 job
->write_err
|= write_err
;
676 * Only dispatch more work if there hasn't been an error.
678 if ((!job
->read_err
&& !job
->write_err
) ||
679 test_bit(DM_KCOPYD_IGNORE_ERROR
, &job
->flags
)) {
680 /* get the next chunk of work */
681 progress
= job
->progress
;
682 count
= job
->source
.count
- progress
;
684 if (count
> SUB_JOB_SIZE
)
685 count
= SUB_JOB_SIZE
;
687 job
->progress
+= count
;
690 mutex_unlock(&job
->lock
);
696 sub_job
->write_offset
= progress
;
697 sub_job
->source
.sector
+= progress
;
698 sub_job
->source
.count
= count
;
700 for (i
= 0; i
< job
->num_dests
; i
++) {
701 sub_job
->dests
[i
].sector
+= progress
;
702 sub_job
->dests
[i
].count
= count
;
705 sub_job
->fn
= segment_complete
;
706 sub_job
->context
= sub_job
;
707 dispatch_job(sub_job
);
709 } else if (atomic_dec_and_test(&job
->sub_jobs
)) {
712 * Queue the completion callback to the kcopyd thread.
714 * Some callers assume that all the completions are called
715 * from a single thread and don't race with each other.
717 * We must not call the callback directly here because this
718 * code may not be executing in the thread.
720 push(&kc
->complete_jobs
, job
);
726 * Create some sub jobs to share the work between them.
728 static void split_job(struct kcopyd_job
*master_job
)
732 atomic_inc(&master_job
->kc
->nr_jobs
);
734 atomic_set(&master_job
->sub_jobs
, SPLIT_COUNT
);
735 for (i
= 0; i
< SPLIT_COUNT
; i
++) {
736 master_job
[i
+ 1].master_job
= master_job
;
737 segment_complete(0, 0u, &master_job
[i
+ 1]);
741 int dm_kcopyd_copy(struct dm_kcopyd_client
*kc
, struct dm_io_region
*from
,
742 unsigned int num_dests
, struct dm_io_region
*dests
,
743 unsigned int flags
, dm_kcopyd_notify_fn fn
, void *context
)
745 struct kcopyd_job
*job
;
749 * Allocate an array of jobs consisting of one master job
750 * followed by SPLIT_COUNT sub jobs.
752 job
= mempool_alloc(kc
->job_pool
, GFP_NOIO
);
755 * set up for the read.
762 job
->num_dests
= num_dests
;
763 memcpy(&job
->dests
, dests
, sizeof(*dests
) * num_dests
);
766 * If one of the destination is a host-managed zoned block device,
767 * we need to write sequentially. If one of the destination is a
768 * host-aware device, then leave it to the caller to choose what to do.
770 if (!test_bit(DM_KCOPYD_WRITE_SEQ
, &job
->flags
)) {
771 for (i
= 0; i
< job
->num_dests
; i
++) {
772 if (bdev_zoned_model(dests
[i
].bdev
) == BLK_ZONED_HM
) {
773 set_bit(DM_KCOPYD_WRITE_SEQ
, &job
->flags
);
780 * If we need to write sequentially, errors cannot be ignored.
782 if (test_bit(DM_KCOPYD_WRITE_SEQ
, &job
->flags
) &&
783 test_bit(DM_KCOPYD_IGNORE_ERROR
, &job
->flags
))
784 clear_bit(DM_KCOPYD_IGNORE_ERROR
, &job
->flags
);
791 memset(&job
->source
, 0, sizeof job
->source
);
792 job
->source
.count
= job
->dests
[0].count
;
793 job
->pages
= &zero_page_list
;
796 * Use WRITE ZEROES to optimize zeroing if all dests support it.
798 job
->rw
= REQ_OP_WRITE_ZEROES
;
799 for (i
= 0; i
< job
->num_dests
; i
++)
800 if (!bdev_write_zeroes_sectors(job
->dests
[i
].bdev
)) {
807 job
->context
= context
;
808 job
->master_job
= job
;
809 job
->write_offset
= 0;
811 if (job
->source
.count
<= SUB_JOB_SIZE
)
814 mutex_init(&job
->lock
);
821 EXPORT_SYMBOL(dm_kcopyd_copy
);
823 int dm_kcopyd_zero(struct dm_kcopyd_client
*kc
,
824 unsigned num_dests
, struct dm_io_region
*dests
,
825 unsigned flags
, dm_kcopyd_notify_fn fn
, void *context
)
827 return dm_kcopyd_copy(kc
, NULL
, num_dests
, dests
, flags
, fn
, context
);
829 EXPORT_SYMBOL(dm_kcopyd_zero
);
831 void *dm_kcopyd_prepare_callback(struct dm_kcopyd_client
*kc
,
832 dm_kcopyd_notify_fn fn
, void *context
)
834 struct kcopyd_job
*job
;
836 job
= mempool_alloc(kc
->job_pool
, GFP_NOIO
);
838 memset(job
, 0, sizeof(struct kcopyd_job
));
841 job
->context
= context
;
842 job
->master_job
= job
;
844 atomic_inc(&kc
->nr_jobs
);
848 EXPORT_SYMBOL(dm_kcopyd_prepare_callback
);
850 void dm_kcopyd_do_callback(void *j
, int read_err
, unsigned long write_err
)
852 struct kcopyd_job
*job
= j
;
853 struct dm_kcopyd_client
*kc
= job
->kc
;
855 job
->read_err
= read_err
;
856 job
->write_err
= write_err
;
858 push(&kc
->complete_jobs
, job
);
861 EXPORT_SYMBOL(dm_kcopyd_do_callback
);
864 * Cancels a kcopyd job, eg. someone might be deactivating a
868 int kcopyd_cancel(struct kcopyd_job
*job
, int block
)
875 /*-----------------------------------------------------------------
877 *---------------------------------------------------------------*/
878 struct dm_kcopyd_client
*dm_kcopyd_client_create(struct dm_kcopyd_throttle
*throttle
)
881 struct dm_kcopyd_client
*kc
;
883 kc
= kmalloc(sizeof(*kc
), GFP_KERNEL
);
885 return ERR_PTR(-ENOMEM
);
887 spin_lock_init(&kc
->job_lock
);
888 INIT_LIST_HEAD(&kc
->complete_jobs
);
889 INIT_LIST_HEAD(&kc
->io_jobs
);
890 INIT_LIST_HEAD(&kc
->pages_jobs
);
891 kc
->throttle
= throttle
;
893 kc
->job_pool
= mempool_create_slab_pool(MIN_JOBS
, _job_cache
);
897 INIT_WORK(&kc
->kcopyd_work
, do_work
);
898 kc
->kcopyd_wq
= alloc_workqueue("kcopyd", WQ_MEM_RECLAIM
, 0);
903 kc
->nr_reserved_pages
= kc
->nr_free_pages
= 0;
904 r
= client_reserve_pages(kc
, RESERVE_PAGES
);
906 goto bad_client_pages
;
908 kc
->io_client
= dm_io_client_create();
909 if (IS_ERR(kc
->io_client
)) {
910 r
= PTR_ERR(kc
->io_client
);
914 init_waitqueue_head(&kc
->destroyq
);
915 atomic_set(&kc
->nr_jobs
, 0);
920 client_free_pages(kc
);
922 destroy_workqueue(kc
->kcopyd_wq
);
924 mempool_destroy(kc
->job_pool
);
930 EXPORT_SYMBOL(dm_kcopyd_client_create
);
932 void dm_kcopyd_client_destroy(struct dm_kcopyd_client
*kc
)
934 /* Wait for completion of all jobs submitted by this client. */
935 wait_event(kc
->destroyq
, !atomic_read(&kc
->nr_jobs
));
937 BUG_ON(!list_empty(&kc
->complete_jobs
));
938 BUG_ON(!list_empty(&kc
->io_jobs
));
939 BUG_ON(!list_empty(&kc
->pages_jobs
));
940 destroy_workqueue(kc
->kcopyd_wq
);
941 dm_io_client_destroy(kc
->io_client
);
942 client_free_pages(kc
);
943 mempool_destroy(kc
->job_pool
);
946 EXPORT_SYMBOL(dm_kcopyd_client_destroy
);