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/device-mapper.h>
26 #include <linux/dm-kcopyd.h>
30 #define SUB_JOB_SIZE 128
33 #define RESERVE_PAGES (DIV_ROUND_UP(SUB_JOB_SIZE << SECTOR_SHIFT, PAGE_SIZE))
35 /*-----------------------------------------------------------------
36 * Each kcopyd client has its own little pool of preallocated
37 * pages for kcopyd io.
38 *---------------------------------------------------------------*/
39 struct dm_kcopyd_client {
40 struct page_list *pages;
41 unsigned nr_reserved_pages;
42 unsigned nr_free_pages;
44 struct dm_io_client *io_client;
46 wait_queue_head_t destroyq;
51 struct workqueue_struct *kcopyd_wq;
52 struct work_struct kcopyd_work;
55 * We maintain three lists of jobs:
57 * i) jobs waiting for pages
58 * ii) jobs that have pages, and are waiting for the io to be issued.
59 * iii) jobs that have completed.
61 * All three of these are protected by job_lock.
64 struct list_head complete_jobs;
65 struct list_head io_jobs;
66 struct list_head pages_jobs;
69 static struct page_list zero_page_list;
71 static void wake(struct dm_kcopyd_client *kc)
73 queue_work(kc->kcopyd_wq, &kc->kcopyd_work);
77 * Obtain one page for the use of kcopyd.
79 static struct page_list *alloc_pl(gfp_t gfp)
83 pl = kmalloc(sizeof(*pl), gfp);
87 pl->page = alloc_page(gfp);
96 static void free_pl(struct page_list *pl)
98 __free_page(pl->page);
103 * Add the provided pages to a client's free page list, releasing
104 * back to the system any beyond the reserved_pages limit.
106 static void kcopyd_put_pages(struct dm_kcopyd_client *kc, struct page_list *pl)
108 struct page_list *next;
113 if (kc->nr_free_pages >= kc->nr_reserved_pages)
116 pl->next = kc->pages;
125 static int kcopyd_get_pages(struct dm_kcopyd_client *kc,
126 unsigned int nr, struct page_list **pages)
128 struct page_list *pl;
133 pl = alloc_pl(__GFP_NOWARN | __GFP_NORETRY);
135 /* Use reserved pages */
139 kc->pages = pl->next;
150 kcopyd_put_pages(kc, *pages);
155 * These three functions resize the page pool.
157 static void drop_pages(struct page_list *pl)
159 struct page_list *next;
169 * Allocate and reserve nr_pages for the use of a specific client.
171 static int client_reserve_pages(struct dm_kcopyd_client *kc, unsigned nr_pages)
174 struct page_list *pl = NULL, *next;
176 for (i = 0; i < nr_pages; i++) {
177 next = alloc_pl(GFP_KERNEL);
187 kc->nr_reserved_pages += nr_pages;
188 kcopyd_put_pages(kc, pl);
193 static void client_free_pages(struct dm_kcopyd_client *kc)
195 BUG_ON(kc->nr_free_pages != kc->nr_reserved_pages);
196 drop_pages(kc->pages);
198 kc->nr_free_pages = kc->nr_reserved_pages = 0;
201 /*-----------------------------------------------------------------
202 * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
203 * for this reason we use a mempool to prevent the client from
204 * ever having to do io (which could cause a deadlock).
205 *---------------------------------------------------------------*/
207 struct dm_kcopyd_client *kc;
208 struct list_head list;
212 * Error state of the job.
215 unsigned long write_err;
218 * Either READ or WRITE
221 struct dm_io_region source;
224 * The destinations for the transfer.
226 unsigned int num_dests;
227 struct dm_io_region dests[DM_KCOPYD_MAX_REGIONS];
229 struct page_list *pages;
232 * Set this to ensure you are notified when the job has
233 * completed. 'context' is for callback to use.
235 dm_kcopyd_notify_fn fn;
239 * These fields are only used if the job has been split
240 * into more manageable parts.
246 struct kcopyd_job *master_job;
249 static struct kmem_cache *_job_cache;
251 int __init dm_kcopyd_init(void)
253 _job_cache = kmem_cache_create("kcopyd_job",
254 sizeof(struct kcopyd_job) * (SPLIT_COUNT + 1),
255 __alignof__(struct kcopyd_job), 0, NULL);
259 zero_page_list.next = &zero_page_list;
260 zero_page_list.page = ZERO_PAGE(0);
265 void dm_kcopyd_exit(void)
267 kmem_cache_destroy(_job_cache);
272 * Functions to push and pop a job onto the head of a given job
275 static struct kcopyd_job *pop(struct list_head *jobs,
276 struct dm_kcopyd_client *kc)
278 struct kcopyd_job *job = NULL;
281 spin_lock_irqsave(&kc->job_lock, flags);
283 if (!list_empty(jobs)) {
284 job = list_entry(jobs->next, struct kcopyd_job, list);
285 list_del(&job->list);
287 spin_unlock_irqrestore(&kc->job_lock, flags);
292 static void push(struct list_head *jobs, struct kcopyd_job *job)
295 struct dm_kcopyd_client *kc = job->kc;
297 spin_lock_irqsave(&kc->job_lock, flags);
298 list_add_tail(&job->list, jobs);
299 spin_unlock_irqrestore(&kc->job_lock, flags);
303 static void push_head(struct list_head *jobs, struct kcopyd_job *job)
306 struct dm_kcopyd_client *kc = job->kc;
308 spin_lock_irqsave(&kc->job_lock, flags);
309 list_add(&job->list, jobs);
310 spin_unlock_irqrestore(&kc->job_lock, flags);
314 * These three functions process 1 item from the corresponding
320 * > 0: can't process yet.
322 static int run_complete_job(struct kcopyd_job *job)
324 void *context = job->context;
325 int read_err = job->read_err;
326 unsigned long write_err = job->write_err;
327 dm_kcopyd_notify_fn fn = job->fn;
328 struct dm_kcopyd_client *kc = job->kc;
330 if (job->pages && job->pages != &zero_page_list)
331 kcopyd_put_pages(kc, job->pages);
333 * If this is the master job, the sub jobs have already
334 * completed so we can free everything.
336 if (job->master_job == job)
337 mempool_free(job, kc->job_pool);
338 fn(read_err, write_err, context);
340 if (atomic_dec_and_test(&kc->nr_jobs))
341 wake_up(&kc->destroyq);
346 static void complete_io(unsigned long error, void *context)
348 struct kcopyd_job *job = (struct kcopyd_job *) context;
349 struct dm_kcopyd_client *kc = job->kc;
353 job->write_err |= error;
357 if (!test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
358 push(&kc->complete_jobs, job);
365 push(&kc->complete_jobs, job);
369 push(&kc->io_jobs, job);
376 * Request io on as many buffer heads as we can currently get for
379 static int run_io_job(struct kcopyd_job *job)
382 struct dm_io_request io_req = {
384 .mem.type = DM_IO_PAGE_LIST,
385 .mem.ptr.pl = job->pages,
387 .notify.fn = complete_io,
388 .notify.context = job,
389 .client = job->kc->io_client,
393 r = dm_io(&io_req, 1, &job->source, NULL);
395 r = dm_io(&io_req, job->num_dests, job->dests, NULL);
400 static int run_pages_job(struct kcopyd_job *job)
403 unsigned nr_pages = dm_div_up(job->dests[0].count, PAGE_SIZE >> 9);
405 r = kcopyd_get_pages(job->kc, nr_pages, &job->pages);
407 /* this job is ready for io */
408 push(&job->kc->io_jobs, job);
413 /* can't complete now */
420 * Run through a list for as long as possible. Returns the count
421 * of successful jobs.
423 static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
424 int (*fn) (struct kcopyd_job *))
426 struct kcopyd_job *job;
429 while ((job = pop(jobs, kc))) {
434 /* error this rogue job */
436 job->write_err = (unsigned long) -1L;
439 push(&kc->complete_jobs, job);
445 * We couldn't service this job ATM, so
446 * push this job back onto the list.
448 push_head(jobs, job);
459 * kcopyd does this every time it's woken up.
461 static void do_work(struct work_struct *work)
463 struct dm_kcopyd_client *kc = container_of(work,
464 struct dm_kcopyd_client, kcopyd_work);
465 struct blk_plug plug;
468 * The order that these are called is *very* important.
469 * complete jobs can free some pages for pages jobs.
470 * Pages jobs when successful will jump onto the io jobs
471 * list. io jobs call wake when they complete and it all
474 blk_start_plug(&plug);
475 process_jobs(&kc->complete_jobs, kc, run_complete_job);
476 process_jobs(&kc->pages_jobs, kc, run_pages_job);
477 process_jobs(&kc->io_jobs, kc, run_io_job);
478 blk_finish_plug(&plug);
482 * If we are copying a small region we just dispatch a single job
483 * to do the copy, otherwise the io has to be split up into many
486 static void dispatch_job(struct kcopyd_job *job)
488 struct dm_kcopyd_client *kc = job->kc;
489 atomic_inc(&kc->nr_jobs);
490 if (unlikely(!job->source.count))
491 push(&kc->complete_jobs, job);
492 else if (job->pages == &zero_page_list)
493 push(&kc->io_jobs, job);
495 push(&kc->pages_jobs, job);
499 static void segment_complete(int read_err, unsigned long write_err,
502 /* FIXME: tidy this function */
503 sector_t progress = 0;
505 struct kcopyd_job *sub_job = (struct kcopyd_job *) context;
506 struct kcopyd_job *job = sub_job->master_job;
507 struct dm_kcopyd_client *kc = job->kc;
509 mutex_lock(&job->lock);
511 /* update the error */
516 job->write_err |= write_err;
519 * Only dispatch more work if there hasn't been an error.
521 if ((!job->read_err && !job->write_err) ||
522 test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
523 /* get the next chunk of work */
524 progress = job->progress;
525 count = job->source.count - progress;
527 if (count > SUB_JOB_SIZE)
528 count = SUB_JOB_SIZE;
530 job->progress += count;
533 mutex_unlock(&job->lock);
539 sub_job->source.sector += progress;
540 sub_job->source.count = count;
542 for (i = 0; i < job->num_dests; i++) {
543 sub_job->dests[i].sector += progress;
544 sub_job->dests[i].count = count;
547 sub_job->fn = segment_complete;
548 sub_job->context = sub_job;
549 dispatch_job(sub_job);
551 } else if (atomic_dec_and_test(&job->sub_jobs)) {
554 * Queue the completion callback to the kcopyd thread.
556 * Some callers assume that all the completions are called
557 * from a single thread and don't race with each other.
559 * We must not call the callback directly here because this
560 * code may not be executing in the thread.
562 push(&kc->complete_jobs, job);
568 * Create some sub jobs to share the work between them.
570 static void split_job(struct kcopyd_job *master_job)
574 atomic_inc(&master_job->kc->nr_jobs);
576 atomic_set(&master_job->sub_jobs, SPLIT_COUNT);
577 for (i = 0; i < SPLIT_COUNT; i++) {
578 master_job[i + 1].master_job = master_job;
579 segment_complete(0, 0u, &master_job[i + 1]);
583 int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
584 unsigned int num_dests, struct dm_io_region *dests,
585 unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
587 struct kcopyd_job *job;
591 * Allocate an array of jobs consisting of one master job
592 * followed by SPLIT_COUNT sub jobs.
594 job = mempool_alloc(kc->job_pool, GFP_NOIO);
597 * set up for the read.
604 job->num_dests = num_dests;
605 memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
612 memset(&job->source, 0, sizeof job->source);
613 job->source.count = job->dests[0].count;
614 job->pages = &zero_page_list;
617 * Use WRITE SAME to optimize zeroing if all dests support it.
619 job->rw = WRITE | REQ_WRITE_SAME;
620 for (i = 0; i < job->num_dests; i++)
621 if (!bdev_write_same(job->dests[i].bdev)) {
628 job->context = context;
629 job->master_job = job;
631 if (job->source.count <= SUB_JOB_SIZE)
634 mutex_init(&job->lock);
641 EXPORT_SYMBOL(dm_kcopyd_copy);
643 int dm_kcopyd_zero(struct dm_kcopyd_client *kc,
644 unsigned num_dests, struct dm_io_region *dests,
645 unsigned flags, dm_kcopyd_notify_fn fn, void *context)
647 return dm_kcopyd_copy(kc, NULL, num_dests, dests, flags, fn, context);
649 EXPORT_SYMBOL(dm_kcopyd_zero);
651 void *dm_kcopyd_prepare_callback(struct dm_kcopyd_client *kc,
652 dm_kcopyd_notify_fn fn, void *context)
654 struct kcopyd_job *job;
656 job = mempool_alloc(kc->job_pool, GFP_NOIO);
658 memset(job, 0, sizeof(struct kcopyd_job));
661 job->context = context;
662 job->master_job = job;
664 atomic_inc(&kc->nr_jobs);
668 EXPORT_SYMBOL(dm_kcopyd_prepare_callback);
670 void dm_kcopyd_do_callback(void *j, int read_err, unsigned long write_err)
672 struct kcopyd_job *job = j;
673 struct dm_kcopyd_client *kc = job->kc;
675 job->read_err = read_err;
676 job->write_err = write_err;
678 push(&kc->complete_jobs, job);
681 EXPORT_SYMBOL(dm_kcopyd_do_callback);
684 * Cancels a kcopyd job, eg. someone might be deactivating a
688 int kcopyd_cancel(struct kcopyd_job *job, int block)
695 /*-----------------------------------------------------------------
697 *---------------------------------------------------------------*/
698 struct dm_kcopyd_client *dm_kcopyd_client_create(void)
701 struct dm_kcopyd_client *kc;
703 kc = kmalloc(sizeof(*kc), GFP_KERNEL);
705 return ERR_PTR(-ENOMEM);
707 spin_lock_init(&kc->job_lock);
708 INIT_LIST_HEAD(&kc->complete_jobs);
709 INIT_LIST_HEAD(&kc->io_jobs);
710 INIT_LIST_HEAD(&kc->pages_jobs);
712 kc->job_pool = mempool_create_slab_pool(MIN_JOBS, _job_cache);
716 INIT_WORK(&kc->kcopyd_work, do_work);
717 kc->kcopyd_wq = alloc_workqueue("kcopyd",
718 WQ_NON_REENTRANT | WQ_MEM_RECLAIM, 0);
723 kc->nr_reserved_pages = kc->nr_free_pages = 0;
724 r = client_reserve_pages(kc, RESERVE_PAGES);
726 goto bad_client_pages;
728 kc->io_client = dm_io_client_create();
729 if (IS_ERR(kc->io_client)) {
730 r = PTR_ERR(kc->io_client);
734 init_waitqueue_head(&kc->destroyq);
735 atomic_set(&kc->nr_jobs, 0);
740 client_free_pages(kc);
742 destroy_workqueue(kc->kcopyd_wq);
744 mempool_destroy(kc->job_pool);
750 EXPORT_SYMBOL(dm_kcopyd_client_create);
752 void dm_kcopyd_client_destroy(struct dm_kcopyd_client *kc)
754 /* Wait for completion of all jobs submitted by this client. */
755 wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
757 BUG_ON(!list_empty(&kc->complete_jobs));
758 BUG_ON(!list_empty(&kc->io_jobs));
759 BUG_ON(!list_empty(&kc->pages_jobs));
760 destroy_workqueue(kc->kcopyd_wq);
761 dm_io_client_destroy(kc->io_client);
762 client_free_pages(kc);
763 mempool_destroy(kc->job_pool);
766 EXPORT_SYMBOL(dm_kcopyd_client_destroy);