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 <asm/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/dm-kcopyd.h>
29 /*-----------------------------------------------------------------
30 * Each kcopyd client has its own little pool of preallocated
31 * pages for kcopyd io.
32 *---------------------------------------------------------------*/
33 struct dm_kcopyd_client
{
35 struct page_list
*pages
;
36 unsigned int nr_pages
;
37 unsigned int nr_free_pages
;
39 struct dm_io_client
*io_client
;
41 wait_queue_head_t destroyq
;
46 struct workqueue_struct
*kcopyd_wq
;
47 struct work_struct kcopyd_work
;
50 * We maintain three lists of jobs:
52 * i) jobs waiting for pages
53 * ii) jobs that have pages, and are waiting for the io to be issued.
54 * iii) jobs that have completed.
56 * All three of these are protected by job_lock.
59 struct list_head complete_jobs
;
60 struct list_head io_jobs
;
61 struct list_head pages_jobs
;
64 static void wake(struct dm_kcopyd_client
*kc
)
66 queue_work(kc
->kcopyd_wq
, &kc
->kcopyd_work
);
69 static struct page_list
*alloc_pl(void)
73 pl
= kmalloc(sizeof(*pl
), GFP_KERNEL
);
77 pl
->page
= alloc_page(GFP_KERNEL
);
86 static void free_pl(struct page_list
*pl
)
88 __free_page(pl
->page
);
92 static int kcopyd_get_pages(struct dm_kcopyd_client
*kc
,
93 unsigned int nr
, struct page_list
**pages
)
98 if (kc
->nr_free_pages
< nr
) {
99 spin_unlock(&kc
->lock
);
103 kc
->nr_free_pages
-= nr
;
104 for (*pages
= pl
= kc
->pages
; --nr
; pl
= pl
->next
)
107 kc
->pages
= pl
->next
;
110 spin_unlock(&kc
->lock
);
115 static void kcopyd_put_pages(struct dm_kcopyd_client
*kc
, struct page_list
*pl
)
117 struct page_list
*cursor
;
119 spin_lock(&kc
->lock
);
120 for (cursor
= pl
; cursor
->next
; cursor
= cursor
->next
)
124 cursor
->next
= kc
->pages
;
126 spin_unlock(&kc
->lock
);
130 * These three functions resize the page pool.
132 static void drop_pages(struct page_list
*pl
)
134 struct page_list
*next
;
143 static int client_alloc_pages(struct dm_kcopyd_client
*kc
, unsigned int nr
)
146 struct page_list
*pl
= NULL
, *next
;
148 for (i
= 0; i
< nr
; i
++) {
159 kcopyd_put_pages(kc
, pl
);
164 static void client_free_pages(struct dm_kcopyd_client
*kc
)
166 BUG_ON(kc
->nr_free_pages
!= kc
->nr_pages
);
167 drop_pages(kc
->pages
);
169 kc
->nr_free_pages
= kc
->nr_pages
= 0;
172 /*-----------------------------------------------------------------
173 * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
174 * for this reason we use a mempool to prevent the client from
175 * ever having to do io (which could cause a deadlock).
176 *---------------------------------------------------------------*/
178 struct dm_kcopyd_client
*kc
;
179 struct list_head list
;
183 * Error state of the job.
186 unsigned long write_err
;
189 * Either READ or WRITE
192 struct dm_io_region source
;
195 * The destinations for the transfer.
197 unsigned int num_dests
;
198 struct dm_io_region dests
[DM_KCOPYD_MAX_REGIONS
];
201 unsigned int nr_pages
;
202 struct page_list
*pages
;
205 * Set this to ensure you are notified when the job has
206 * completed. 'context' is for callback to use.
208 dm_kcopyd_notify_fn fn
;
212 * These fields are only used if the job has been split
213 * into more manageable parts.
220 /* FIXME: this should scale with the number of pages */
223 static struct kmem_cache
*_job_cache
;
225 int __init
dm_kcopyd_init(void)
227 _job_cache
= KMEM_CACHE(kcopyd_job
, 0);
234 void dm_kcopyd_exit(void)
236 kmem_cache_destroy(_job_cache
);
241 * Functions to push and pop a job onto the head of a given job
244 static struct kcopyd_job
*pop(struct list_head
*jobs
,
245 struct dm_kcopyd_client
*kc
)
247 struct kcopyd_job
*job
= NULL
;
250 spin_lock_irqsave(&kc
->job_lock
, flags
);
252 if (!list_empty(jobs
)) {
253 job
= list_entry(jobs
->next
, struct kcopyd_job
, list
);
254 list_del(&job
->list
);
256 spin_unlock_irqrestore(&kc
->job_lock
, flags
);
261 static void push(struct list_head
*jobs
, struct kcopyd_job
*job
)
264 struct dm_kcopyd_client
*kc
= job
->kc
;
266 spin_lock_irqsave(&kc
->job_lock
, flags
);
267 list_add_tail(&job
->list
, jobs
);
268 spin_unlock_irqrestore(&kc
->job_lock
, flags
);
272 static void push_head(struct list_head
*jobs
, struct kcopyd_job
*job
)
275 struct dm_kcopyd_client
*kc
= job
->kc
;
277 spin_lock_irqsave(&kc
->job_lock
, flags
);
278 list_add(&job
->list
, jobs
);
279 spin_unlock_irqrestore(&kc
->job_lock
, flags
);
283 * These three functions process 1 item from the corresponding
289 * > 0: can't process yet.
291 static int run_complete_job(struct kcopyd_job
*job
)
293 void *context
= job
->context
;
294 int read_err
= job
->read_err
;
295 unsigned long write_err
= job
->write_err
;
296 dm_kcopyd_notify_fn fn
= job
->fn
;
297 struct dm_kcopyd_client
*kc
= job
->kc
;
299 kcopyd_put_pages(kc
, job
->pages
);
300 mempool_free(job
, kc
->job_pool
);
301 fn(read_err
, write_err
, context
);
303 if (atomic_dec_and_test(&kc
->nr_jobs
))
304 wake_up(&kc
->destroyq
);
309 static void complete_io(unsigned long error
, void *context
)
311 struct kcopyd_job
*job
= (struct kcopyd_job
*) context
;
312 struct dm_kcopyd_client
*kc
= job
->kc
;
315 if (job
->rw
== WRITE
)
316 job
->write_err
|= error
;
320 if (!test_bit(DM_KCOPYD_IGNORE_ERROR
, &job
->flags
)) {
321 push(&kc
->complete_jobs
, job
);
327 if (job
->rw
== WRITE
)
328 push(&kc
->complete_jobs
, job
);
332 push(&kc
->io_jobs
, job
);
339 * Request io on as many buffer heads as we can currently get for
342 static int run_io_job(struct kcopyd_job
*job
)
345 struct dm_io_request io_req
= {
346 .bi_rw
= job
->rw
| (1 << BIO_RW_SYNC
),
347 .mem
.type
= DM_IO_PAGE_LIST
,
348 .mem
.ptr
.pl
= job
->pages
,
349 .mem
.offset
= job
->offset
,
350 .notify
.fn
= complete_io
,
351 .notify
.context
= job
,
352 .client
= job
->kc
->io_client
,
356 r
= dm_io(&io_req
, 1, &job
->source
, NULL
);
358 r
= dm_io(&io_req
, job
->num_dests
, job
->dests
, NULL
);
363 static int run_pages_job(struct kcopyd_job
*job
)
367 job
->nr_pages
= dm_div_up(job
->dests
[0].count
+ job
->offset
,
369 r
= kcopyd_get_pages(job
->kc
, job
->nr_pages
, &job
->pages
);
371 /* this job is ready for io */
372 push(&job
->kc
->io_jobs
, job
);
377 /* can't complete now */
384 * Run through a list for as long as possible. Returns the count
385 * of successful jobs.
387 static int process_jobs(struct list_head
*jobs
, struct dm_kcopyd_client
*kc
,
388 int (*fn
) (struct kcopyd_job
*))
390 struct kcopyd_job
*job
;
393 while ((job
= pop(jobs
, kc
))) {
398 /* error this rogue job */
399 if (job
->rw
== WRITE
)
400 job
->write_err
= (unsigned long) -1L;
403 push(&kc
->complete_jobs
, job
);
409 * We couldn't service this job ATM, so
410 * push this job back onto the list.
412 push_head(jobs
, job
);
423 * kcopyd does this every time it's woken up.
425 static void do_work(struct work_struct
*work
)
427 struct dm_kcopyd_client
*kc
= container_of(work
,
428 struct dm_kcopyd_client
, kcopyd_work
);
431 * The order that these are called is *very* important.
432 * complete jobs can free some pages for pages jobs.
433 * Pages jobs when successful will jump onto the io jobs
434 * list. io jobs call wake when they complete and it all
437 process_jobs(&kc
->complete_jobs
, kc
, run_complete_job
);
438 process_jobs(&kc
->pages_jobs
, kc
, run_pages_job
);
439 process_jobs(&kc
->io_jobs
, kc
, run_io_job
);
443 * If we are copying a small region we just dispatch a single job
444 * to do the copy, otherwise the io has to be split up into many
447 static void dispatch_job(struct kcopyd_job
*job
)
449 struct dm_kcopyd_client
*kc
= job
->kc
;
450 atomic_inc(&kc
->nr_jobs
);
451 push(&kc
->pages_jobs
, job
);
455 #define SUB_JOB_SIZE 128
456 static void segment_complete(int read_err
, unsigned long write_err
,
459 /* FIXME: tidy this function */
460 sector_t progress
= 0;
462 struct kcopyd_job
*job
= (struct kcopyd_job
*) context
;
464 mutex_lock(&job
->lock
);
466 /* update the error */
471 job
->write_err
|= write_err
;
474 * Only dispatch more work if there hasn't been an error.
476 if ((!job
->read_err
&& !job
->write_err
) ||
477 test_bit(DM_KCOPYD_IGNORE_ERROR
, &job
->flags
)) {
478 /* get the next chunk of work */
479 progress
= job
->progress
;
480 count
= job
->source
.count
- progress
;
482 if (count
> SUB_JOB_SIZE
)
483 count
= SUB_JOB_SIZE
;
485 job
->progress
+= count
;
488 mutex_unlock(&job
->lock
);
492 struct kcopyd_job
*sub_job
= mempool_alloc(job
->kc
->job_pool
,
496 sub_job
->source
.sector
+= progress
;
497 sub_job
->source
.count
= count
;
499 for (i
= 0; i
< job
->num_dests
; i
++) {
500 sub_job
->dests
[i
].sector
+= progress
;
501 sub_job
->dests
[i
].count
= count
;
504 sub_job
->fn
= segment_complete
;
505 sub_job
->context
= job
;
506 dispatch_job(sub_job
);
508 } else if (atomic_dec_and_test(&job
->sub_jobs
)) {
511 * To avoid a race we must keep the job around
512 * until after the notify function has completed.
513 * Otherwise the client may try and stop the job
514 * after we've completed.
516 job
->fn(read_err
, write_err
, job
->context
);
517 mempool_free(job
, job
->kc
->job_pool
);
522 * Create some little jobs that will do the move between
525 #define SPLIT_COUNT 8
526 static void split_job(struct kcopyd_job
*job
)
530 atomic_set(&job
->sub_jobs
, SPLIT_COUNT
);
531 for (i
= 0; i
< SPLIT_COUNT
; i
++)
532 segment_complete(0, 0u, job
);
535 int dm_kcopyd_copy(struct dm_kcopyd_client
*kc
, struct dm_io_region
*from
,
536 unsigned int num_dests
, struct dm_io_region
*dests
,
537 unsigned int flags
, dm_kcopyd_notify_fn fn
, void *context
)
539 struct kcopyd_job
*job
;
542 * Allocate a new job.
544 job
= mempool_alloc(kc
->job_pool
, GFP_NOIO
);
547 * set up for the read.
557 job
->num_dests
= num_dests
;
558 memcpy(&job
->dests
, dests
, sizeof(*dests
) * num_dests
);
565 job
->context
= context
;
567 if (job
->source
.count
< SUB_JOB_SIZE
)
571 mutex_init(&job
->lock
);
578 EXPORT_SYMBOL(dm_kcopyd_copy
);
581 * Cancels a kcopyd job, eg. someone might be deactivating a
585 int kcopyd_cancel(struct kcopyd_job
*job
, int block
)
592 /*-----------------------------------------------------------------
594 *---------------------------------------------------------------*/
595 int dm_kcopyd_client_create(unsigned int nr_pages
,
596 struct dm_kcopyd_client
**result
)
599 struct dm_kcopyd_client
*kc
;
601 kc
= kmalloc(sizeof(*kc
), GFP_KERNEL
);
605 spin_lock_init(&kc
->lock
);
606 spin_lock_init(&kc
->job_lock
);
607 INIT_LIST_HEAD(&kc
->complete_jobs
);
608 INIT_LIST_HEAD(&kc
->io_jobs
);
609 INIT_LIST_HEAD(&kc
->pages_jobs
);
611 kc
->job_pool
= mempool_create_slab_pool(MIN_JOBS
, _job_cache
);
615 INIT_WORK(&kc
->kcopyd_work
, do_work
);
616 kc
->kcopyd_wq
= create_singlethread_workqueue("kcopyd");
621 kc
->nr_pages
= kc
->nr_free_pages
= 0;
622 r
= client_alloc_pages(kc
, nr_pages
);
624 goto bad_client_pages
;
626 kc
->io_client
= dm_io_client_create(nr_pages
);
627 if (IS_ERR(kc
->io_client
)) {
628 r
= PTR_ERR(kc
->io_client
);
632 init_waitqueue_head(&kc
->destroyq
);
633 atomic_set(&kc
->nr_jobs
, 0);
639 client_free_pages(kc
);
641 destroy_workqueue(kc
->kcopyd_wq
);
643 mempool_destroy(kc
->job_pool
);
649 EXPORT_SYMBOL(dm_kcopyd_client_create
);
651 void dm_kcopyd_client_destroy(struct dm_kcopyd_client
*kc
)
653 /* Wait for completion of all jobs submitted by this client. */
654 wait_event(kc
->destroyq
, !atomic_read(&kc
->nr_jobs
));
656 BUG_ON(!list_empty(&kc
->complete_jobs
));
657 BUG_ON(!list_empty(&kc
->io_jobs
));
658 BUG_ON(!list_empty(&kc
->pages_jobs
));
659 destroy_workqueue(kc
->kcopyd_wq
);
660 dm_io_client_destroy(kc
->io_client
);
661 client_free_pages(kc
);
662 mempool_destroy(kc
->job_pool
);
665 EXPORT_SYMBOL(dm_kcopyd_client_destroy
);