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 void wake(struct dm_kcopyd_client *kc)
71 queue_work(kc->kcopyd_wq, &kc->kcopyd_work);
75 * Obtain one page for the use of kcopyd.
77 static struct page_list *alloc_pl(gfp_t gfp)
81 pl = kmalloc(sizeof(*pl), gfp);
85 pl->page = alloc_page(gfp);
94 static void free_pl(struct page_list *pl)
96 __free_page(pl->page);
101 * Add the provided pages to a client's free page list, releasing
102 * back to the system any beyond the reserved_pages limit.
104 static void kcopyd_put_pages(struct dm_kcopyd_client *kc, struct page_list *pl)
106 struct page_list *next;
111 if (kc->nr_free_pages >= kc->nr_reserved_pages)
114 pl->next = kc->pages;
123 static int kcopyd_get_pages(struct dm_kcopyd_client *kc,
124 unsigned int nr, struct page_list **pages)
126 struct page_list *pl;
131 pl = alloc_pl(__GFP_NOWARN | __GFP_NORETRY);
133 /* Use reserved pages */
137 kc->pages = pl->next;
148 kcopyd_put_pages(kc, *pages);
153 * These three functions resize the page pool.
155 static void drop_pages(struct page_list *pl)
157 struct page_list *next;
167 * Allocate and reserve nr_pages for the use of a specific client.
169 static int client_reserve_pages(struct dm_kcopyd_client *kc, unsigned nr_pages)
172 struct page_list *pl = NULL, *next;
174 for (i = 0; i < nr_pages; i++) {
175 next = alloc_pl(GFP_KERNEL);
185 kc->nr_reserved_pages += nr_pages;
186 kcopyd_put_pages(kc, pl);
191 static void client_free_pages(struct dm_kcopyd_client *kc)
193 BUG_ON(kc->nr_free_pages != kc->nr_reserved_pages);
194 drop_pages(kc->pages);
196 kc->nr_free_pages = kc->nr_reserved_pages = 0;
199 /*-----------------------------------------------------------------
200 * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
201 * for this reason we use a mempool to prevent the client from
202 * ever having to do io (which could cause a deadlock).
203 *---------------------------------------------------------------*/
205 struct dm_kcopyd_client *kc;
206 struct list_head list;
210 * Error state of the job.
213 unsigned long write_err;
216 * Either READ or WRITE
219 struct dm_io_region source;
222 * The destinations for the transfer.
224 unsigned int num_dests;
225 struct dm_io_region dests[DM_KCOPYD_MAX_REGIONS];
227 unsigned int nr_pages;
228 struct page_list *pages;
231 * Set this to ensure you are notified when the job has
232 * completed. 'context' is for callback to use.
234 dm_kcopyd_notify_fn fn;
238 * These fields are only used if the job has been split
239 * into more manageable parts.
245 struct kcopyd_job *master_job;
248 static struct kmem_cache *_job_cache;
250 int __init dm_kcopyd_init(void)
252 _job_cache = kmem_cache_create("kcopyd_job",
253 sizeof(struct kcopyd_job) * (SPLIT_COUNT + 1),
254 __alignof__(struct kcopyd_job), 0, NULL);
261 void dm_kcopyd_exit(void)
263 kmem_cache_destroy(_job_cache);
268 * Functions to push and pop a job onto the head of a given job
271 static struct kcopyd_job *pop(struct list_head *jobs,
272 struct dm_kcopyd_client *kc)
274 struct kcopyd_job *job = NULL;
277 spin_lock_irqsave(&kc->job_lock, flags);
279 if (!list_empty(jobs)) {
280 job = list_entry(jobs->next, struct kcopyd_job, list);
281 list_del(&job->list);
283 spin_unlock_irqrestore(&kc->job_lock, flags);
288 static void push(struct list_head *jobs, struct kcopyd_job *job)
291 struct dm_kcopyd_client *kc = job->kc;
293 spin_lock_irqsave(&kc->job_lock, flags);
294 list_add_tail(&job->list, jobs);
295 spin_unlock_irqrestore(&kc->job_lock, flags);
299 static void push_head(struct list_head *jobs, struct kcopyd_job *job)
302 struct dm_kcopyd_client *kc = job->kc;
304 spin_lock_irqsave(&kc->job_lock, flags);
305 list_add(&job->list, jobs);
306 spin_unlock_irqrestore(&kc->job_lock, flags);
310 * These three functions process 1 item from the corresponding
316 * > 0: can't process yet.
318 static int run_complete_job(struct kcopyd_job *job)
320 void *context = job->context;
321 int read_err = job->read_err;
322 unsigned long write_err = job->write_err;
323 dm_kcopyd_notify_fn fn = job->fn;
324 struct dm_kcopyd_client *kc = job->kc;
327 kcopyd_put_pages(kc, job->pages);
329 * If this is the master job, the sub jobs have already
330 * completed so we can free everything.
332 if (job->master_job == job)
333 mempool_free(job, kc->job_pool);
334 fn(read_err, write_err, context);
336 if (atomic_dec_and_test(&kc->nr_jobs))
337 wake_up(&kc->destroyq);
342 static void complete_io(unsigned long error, void *context)
344 struct kcopyd_job *job = (struct kcopyd_job *) context;
345 struct dm_kcopyd_client *kc = job->kc;
348 if (job->rw == WRITE)
349 job->write_err |= error;
353 if (!test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
354 push(&kc->complete_jobs, job);
360 if (job->rw == WRITE)
361 push(&kc->complete_jobs, job);
365 push(&kc->io_jobs, job);
372 * Request io on as many buffer heads as we can currently get for
375 static int run_io_job(struct kcopyd_job *job)
378 struct dm_io_request io_req = {
380 .mem.type = DM_IO_PAGE_LIST,
381 .mem.ptr.pl = job->pages,
383 .notify.fn = complete_io,
384 .notify.context = job,
385 .client = job->kc->io_client,
389 r = dm_io(&io_req, 1, &job->source, NULL);
391 r = dm_io(&io_req, job->num_dests, job->dests, NULL);
396 static int run_pages_job(struct kcopyd_job *job)
400 job->nr_pages = dm_div_up(job->dests[0].count, PAGE_SIZE >> 9);
401 r = kcopyd_get_pages(job->kc, job->nr_pages, &job->pages);
403 /* this job is ready for io */
404 push(&job->kc->io_jobs, job);
409 /* can't complete now */
416 * Run through a list for as long as possible. Returns the count
417 * of successful jobs.
419 static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
420 int (*fn) (struct kcopyd_job *))
422 struct kcopyd_job *job;
425 while ((job = pop(jobs, kc))) {
430 /* error this rogue job */
431 if (job->rw == WRITE)
432 job->write_err = (unsigned long) -1L;
435 push(&kc->complete_jobs, job);
441 * We couldn't service this job ATM, so
442 * push this job back onto the list.
444 push_head(jobs, job);
455 * kcopyd does this every time it's woken up.
457 static void do_work(struct work_struct *work)
459 struct dm_kcopyd_client *kc = container_of(work,
460 struct dm_kcopyd_client, kcopyd_work);
461 struct blk_plug plug;
464 * The order that these are called is *very* important.
465 * complete jobs can free some pages for pages jobs.
466 * Pages jobs when successful will jump onto the io jobs
467 * list. io jobs call wake when they complete and it all
470 blk_start_plug(&plug);
471 process_jobs(&kc->complete_jobs, kc, run_complete_job);
472 process_jobs(&kc->pages_jobs, kc, run_pages_job);
473 process_jobs(&kc->io_jobs, kc, run_io_job);
474 blk_finish_plug(&plug);
478 * If we are copying a small region we just dispatch a single job
479 * to do the copy, otherwise the io has to be split up into many
482 static void dispatch_job(struct kcopyd_job *job)
484 struct dm_kcopyd_client *kc = job->kc;
485 atomic_inc(&kc->nr_jobs);
486 if (unlikely(!job->source.count))
487 push(&kc->complete_jobs, job);
489 push(&kc->pages_jobs, job);
493 static void segment_complete(int read_err, unsigned long write_err,
496 /* FIXME: tidy this function */
497 sector_t progress = 0;
499 struct kcopyd_job *sub_job = (struct kcopyd_job *) context;
500 struct kcopyd_job *job = sub_job->master_job;
501 struct dm_kcopyd_client *kc = job->kc;
503 mutex_lock(&job->lock);
505 /* update the error */
510 job->write_err |= write_err;
513 * Only dispatch more work if there hasn't been an error.
515 if ((!job->read_err && !job->write_err) ||
516 test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
517 /* get the next chunk of work */
518 progress = job->progress;
519 count = job->source.count - progress;
521 if (count > SUB_JOB_SIZE)
522 count = SUB_JOB_SIZE;
524 job->progress += count;
527 mutex_unlock(&job->lock);
533 sub_job->source.sector += progress;
534 sub_job->source.count = count;
536 for (i = 0; i < job->num_dests; i++) {
537 sub_job->dests[i].sector += progress;
538 sub_job->dests[i].count = count;
541 sub_job->fn = segment_complete;
542 sub_job->context = sub_job;
543 dispatch_job(sub_job);
545 } else if (atomic_dec_and_test(&job->sub_jobs)) {
548 * Queue the completion callback to the kcopyd thread.
550 * Some callers assume that all the completions are called
551 * from a single thread and don't race with each other.
553 * We must not call the callback directly here because this
554 * code may not be executing in the thread.
556 push(&kc->complete_jobs, job);
562 * Create some sub jobs to share the work between them.
564 static void split_job(struct kcopyd_job *master_job)
568 atomic_inc(&master_job->kc->nr_jobs);
570 atomic_set(&master_job->sub_jobs, SPLIT_COUNT);
571 for (i = 0; i < SPLIT_COUNT; i++) {
572 master_job[i + 1].master_job = master_job;
573 segment_complete(0, 0u, &master_job[i + 1]);
577 int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
578 unsigned int num_dests, struct dm_io_region *dests,
579 unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
581 struct kcopyd_job *job;
584 * Allocate an array of jobs consisting of one master job
585 * followed by SPLIT_COUNT sub jobs.
587 job = mempool_alloc(kc->job_pool, GFP_NOIO);
590 * set up for the read.
600 job->num_dests = num_dests;
601 memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
607 job->context = context;
608 job->master_job = job;
610 if (job->source.count <= SUB_JOB_SIZE)
613 mutex_init(&job->lock);
620 EXPORT_SYMBOL(dm_kcopyd_copy);
623 * Cancels a kcopyd job, eg. someone might be deactivating a
627 int kcopyd_cancel(struct kcopyd_job *job, int block)
634 /*-----------------------------------------------------------------
636 *---------------------------------------------------------------*/
637 struct dm_kcopyd_client *dm_kcopyd_client_create(void)
640 struct dm_kcopyd_client *kc;
642 kc = kmalloc(sizeof(*kc), GFP_KERNEL);
644 return ERR_PTR(-ENOMEM);
646 spin_lock_init(&kc->job_lock);
647 INIT_LIST_HEAD(&kc->complete_jobs);
648 INIT_LIST_HEAD(&kc->io_jobs);
649 INIT_LIST_HEAD(&kc->pages_jobs);
651 kc->job_pool = mempool_create_slab_pool(MIN_JOBS, _job_cache);
655 INIT_WORK(&kc->kcopyd_work, do_work);
656 kc->kcopyd_wq = alloc_workqueue("kcopyd",
657 WQ_NON_REENTRANT | WQ_MEM_RECLAIM, 0);
662 kc->nr_reserved_pages = kc->nr_free_pages = 0;
663 r = client_reserve_pages(kc, RESERVE_PAGES);
665 goto bad_client_pages;
667 kc->io_client = dm_io_client_create();
668 if (IS_ERR(kc->io_client)) {
669 r = PTR_ERR(kc->io_client);
673 init_waitqueue_head(&kc->destroyq);
674 atomic_set(&kc->nr_jobs, 0);
679 client_free_pages(kc);
681 destroy_workqueue(kc->kcopyd_wq);
683 mempool_destroy(kc->job_pool);
689 EXPORT_SYMBOL(dm_kcopyd_client_create);
691 void dm_kcopyd_client_destroy(struct dm_kcopyd_client *kc)
693 /* Wait for completion of all jobs submitted by this client. */
694 wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
696 BUG_ON(!list_empty(&kc->complete_jobs));
697 BUG_ON(!list_empty(&kc->io_jobs));
698 BUG_ON(!list_empty(&kc->pages_jobs));
699 destroy_workqueue(kc->kcopyd_wq);
700 dm_io_client_destroy(kc->io_client);
701 client_free_pages(kc);
702 mempool_destroy(kc->job_pool);
705 EXPORT_SYMBOL(dm_kcopyd_client_destroy);