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 struct page_list *pages;
230 * Set this to ensure you are notified when the job has
231 * completed. 'context' is for callback to use.
233 dm_kcopyd_notify_fn fn;
237 * These fields are only used if the job has been split
238 * into more manageable parts.
244 struct kcopyd_job *master_job;
247 static struct kmem_cache *_job_cache;
249 int __init dm_kcopyd_init(void)
251 _job_cache = kmem_cache_create("kcopyd_job",
252 sizeof(struct kcopyd_job) * (SPLIT_COUNT + 1),
253 __alignof__(struct kcopyd_job), 0, NULL);
260 void dm_kcopyd_exit(void)
262 kmem_cache_destroy(_job_cache);
267 * Functions to push and pop a job onto the head of a given job
270 static struct kcopyd_job *pop(struct list_head *jobs,
271 struct dm_kcopyd_client *kc)
273 struct kcopyd_job *job = NULL;
276 spin_lock_irqsave(&kc->job_lock, flags);
278 if (!list_empty(jobs)) {
279 job = list_entry(jobs->next, struct kcopyd_job, list);
280 list_del(&job->list);
282 spin_unlock_irqrestore(&kc->job_lock, flags);
287 static void push(struct list_head *jobs, struct kcopyd_job *job)
290 struct dm_kcopyd_client *kc = job->kc;
292 spin_lock_irqsave(&kc->job_lock, flags);
293 list_add_tail(&job->list, jobs);
294 spin_unlock_irqrestore(&kc->job_lock, flags);
298 static void push_head(struct list_head *jobs, struct kcopyd_job *job)
301 struct dm_kcopyd_client *kc = job->kc;
303 spin_lock_irqsave(&kc->job_lock, flags);
304 list_add(&job->list, jobs);
305 spin_unlock_irqrestore(&kc->job_lock, flags);
309 * These three functions process 1 item from the corresponding
315 * > 0: can't process yet.
317 static int run_complete_job(struct kcopyd_job *job)
319 void *context = job->context;
320 int read_err = job->read_err;
321 unsigned long write_err = job->write_err;
322 dm_kcopyd_notify_fn fn = job->fn;
323 struct dm_kcopyd_client *kc = job->kc;
326 kcopyd_put_pages(kc, job->pages);
328 * If this is the master job, the sub jobs have already
329 * completed so we can free everything.
331 if (job->master_job == job)
332 mempool_free(job, kc->job_pool);
333 fn(read_err, write_err, context);
335 if (atomic_dec_and_test(&kc->nr_jobs))
336 wake_up(&kc->destroyq);
341 static void complete_io(unsigned long error, void *context)
343 struct kcopyd_job *job = (struct kcopyd_job *) context;
344 struct dm_kcopyd_client *kc = job->kc;
347 if (job->rw == WRITE)
348 job->write_err |= error;
352 if (!test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
353 push(&kc->complete_jobs, job);
359 if (job->rw == WRITE)
360 push(&kc->complete_jobs, job);
364 push(&kc->io_jobs, job);
371 * Request io on as many buffer heads as we can currently get for
374 static int run_io_job(struct kcopyd_job *job)
377 struct dm_io_request io_req = {
379 .mem.type = DM_IO_PAGE_LIST,
380 .mem.ptr.pl = job->pages,
382 .notify.fn = complete_io,
383 .notify.context = job,
384 .client = job->kc->io_client,
388 r = dm_io(&io_req, 1, &job->source, NULL);
390 r = dm_io(&io_req, job->num_dests, job->dests, NULL);
395 static int run_pages_job(struct kcopyd_job *job)
398 unsigned nr_pages = dm_div_up(job->dests[0].count, PAGE_SIZE >> 9);
400 r = kcopyd_get_pages(job->kc, nr_pages, &job->pages);
402 /* this job is ready for io */
403 push(&job->kc->io_jobs, job);
408 /* can't complete now */
415 * Run through a list for as long as possible. Returns the count
416 * of successful jobs.
418 static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
419 int (*fn) (struct kcopyd_job *))
421 struct kcopyd_job *job;
424 while ((job = pop(jobs, kc))) {
429 /* error this rogue job */
430 if (job->rw == WRITE)
431 job->write_err = (unsigned long) -1L;
434 push(&kc->complete_jobs, job);
440 * We couldn't service this job ATM, so
441 * push this job back onto the list.
443 push_head(jobs, job);
454 * kcopyd does this every time it's woken up.
456 static void do_work(struct work_struct *work)
458 struct dm_kcopyd_client *kc = container_of(work,
459 struct dm_kcopyd_client, kcopyd_work);
460 struct blk_plug plug;
463 * The order that these are called is *very* important.
464 * complete jobs can free some pages for pages jobs.
465 * Pages jobs when successful will jump onto the io jobs
466 * list. io jobs call wake when they complete and it all
469 blk_start_plug(&plug);
470 process_jobs(&kc->complete_jobs, kc, run_complete_job);
471 process_jobs(&kc->pages_jobs, kc, run_pages_job);
472 process_jobs(&kc->io_jobs, kc, run_io_job);
473 blk_finish_plug(&plug);
477 * If we are copying a small region we just dispatch a single job
478 * to do the copy, otherwise the io has to be split up into many
481 static void dispatch_job(struct kcopyd_job *job)
483 struct dm_kcopyd_client *kc = job->kc;
484 atomic_inc(&kc->nr_jobs);
485 if (unlikely(!job->source.count))
486 push(&kc->complete_jobs, job);
488 push(&kc->pages_jobs, job);
492 static void segment_complete(int read_err, unsigned long write_err,
495 /* FIXME: tidy this function */
496 sector_t progress = 0;
498 struct kcopyd_job *sub_job = (struct kcopyd_job *) context;
499 struct kcopyd_job *job = sub_job->master_job;
500 struct dm_kcopyd_client *kc = job->kc;
502 mutex_lock(&job->lock);
504 /* update the error */
509 job->write_err |= write_err;
512 * Only dispatch more work if there hasn't been an error.
514 if ((!job->read_err && !job->write_err) ||
515 test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
516 /* get the next chunk of work */
517 progress = job->progress;
518 count = job->source.count - progress;
520 if (count > SUB_JOB_SIZE)
521 count = SUB_JOB_SIZE;
523 job->progress += count;
526 mutex_unlock(&job->lock);
532 sub_job->source.sector += progress;
533 sub_job->source.count = count;
535 for (i = 0; i < job->num_dests; i++) {
536 sub_job->dests[i].sector += progress;
537 sub_job->dests[i].count = count;
540 sub_job->fn = segment_complete;
541 sub_job->context = sub_job;
542 dispatch_job(sub_job);
544 } else if (atomic_dec_and_test(&job->sub_jobs)) {
547 * Queue the completion callback to the kcopyd thread.
549 * Some callers assume that all the completions are called
550 * from a single thread and don't race with each other.
552 * We must not call the callback directly here because this
553 * code may not be executing in the thread.
555 push(&kc->complete_jobs, job);
561 * Create some sub jobs to share the work between them.
563 static void split_job(struct kcopyd_job *master_job)
567 atomic_inc(&master_job->kc->nr_jobs);
569 atomic_set(&master_job->sub_jobs, SPLIT_COUNT);
570 for (i = 0; i < SPLIT_COUNT; i++) {
571 master_job[i + 1].master_job = master_job;
572 segment_complete(0, 0u, &master_job[i + 1]);
576 int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
577 unsigned int num_dests, struct dm_io_region *dests,
578 unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
580 struct kcopyd_job *job;
583 * Allocate an array of jobs consisting of one master job
584 * followed by SPLIT_COUNT sub jobs.
586 job = mempool_alloc(kc->job_pool, GFP_NOIO);
589 * set up for the read.
599 job->num_dests = num_dests;
600 memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
605 job->context = context;
606 job->master_job = job;
608 if (job->source.count <= SUB_JOB_SIZE)
611 mutex_init(&job->lock);
618 EXPORT_SYMBOL(dm_kcopyd_copy);
620 void *dm_kcopyd_prepare_callback(struct dm_kcopyd_client *kc,
621 dm_kcopyd_notify_fn fn, void *context)
623 struct kcopyd_job *job;
625 job = mempool_alloc(kc->job_pool, GFP_NOIO);
627 memset(job, 0, sizeof(struct kcopyd_job));
630 job->context = context;
632 atomic_inc(&kc->nr_jobs);
636 EXPORT_SYMBOL(dm_kcopyd_prepare_callback);
638 void dm_kcopyd_do_callback(void *j, int read_err, unsigned long write_err)
640 struct kcopyd_job *job = j;
641 struct dm_kcopyd_client *kc = job->kc;
643 job->read_err = read_err;
644 job->write_err = write_err;
646 push(&kc->complete_jobs, job);
649 EXPORT_SYMBOL(dm_kcopyd_do_callback);
652 * Cancels a kcopyd job, eg. someone might be deactivating a
656 int kcopyd_cancel(struct kcopyd_job *job, int block)
663 /*-----------------------------------------------------------------
665 *---------------------------------------------------------------*/
666 struct dm_kcopyd_client *dm_kcopyd_client_create(void)
669 struct dm_kcopyd_client *kc;
671 kc = kmalloc(sizeof(*kc), GFP_KERNEL);
673 return ERR_PTR(-ENOMEM);
675 spin_lock_init(&kc->job_lock);
676 INIT_LIST_HEAD(&kc->complete_jobs);
677 INIT_LIST_HEAD(&kc->io_jobs);
678 INIT_LIST_HEAD(&kc->pages_jobs);
680 kc->job_pool = mempool_create_slab_pool(MIN_JOBS, _job_cache);
684 INIT_WORK(&kc->kcopyd_work, do_work);
685 kc->kcopyd_wq = alloc_workqueue("kcopyd",
686 WQ_NON_REENTRANT | WQ_MEM_RECLAIM, 0);
691 kc->nr_reserved_pages = kc->nr_free_pages = 0;
692 r = client_reserve_pages(kc, RESERVE_PAGES);
694 goto bad_client_pages;
696 kc->io_client = dm_io_client_create();
697 if (IS_ERR(kc->io_client)) {
698 r = PTR_ERR(kc->io_client);
702 init_waitqueue_head(&kc->destroyq);
703 atomic_set(&kc->nr_jobs, 0);
708 client_free_pages(kc);
710 destroy_workqueue(kc->kcopyd_wq);
712 mempool_destroy(kc->job_pool);
718 EXPORT_SYMBOL(dm_kcopyd_client_create);
720 void dm_kcopyd_client_destroy(struct dm_kcopyd_client *kc)
722 /* Wait for completion of all jobs submitted by this client. */
723 wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
725 BUG_ON(!list_empty(&kc->complete_jobs));
726 BUG_ON(!list_empty(&kc->io_jobs));
727 BUG_ON(!list_empty(&kc->pages_jobs));
728 destroy_workqueue(kc->kcopyd_wq);
729 dm_io_client_destroy(kc->io_client);
730 client_free_pages(kc);
731 mempool_destroy(kc->job_pool);
734 EXPORT_SYMBOL(dm_kcopyd_client_destroy);