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/device-mapper.h>
26 #include <linux/dm-kcopyd.h>
30 #define SUB_JOB_SIZE 128
34 /*-----------------------------------------------------------------
35 * Each kcopyd client has its own little pool of preallocated
36 * pages for kcopyd io.
37 *---------------------------------------------------------------*/
38 struct dm_kcopyd_client {
39 struct page_list *pages;
40 unsigned int nr_pages;
41 unsigned int nr_free_pages;
43 struct dm_io_client *io_client;
45 wait_queue_head_t destroyq;
50 struct workqueue_struct *kcopyd_wq;
51 struct work_struct kcopyd_work;
54 * We maintain three lists of jobs:
56 * i) jobs waiting for pages
57 * ii) jobs that have pages, and are waiting for the io to be issued.
58 * iii) jobs that have completed.
60 * All three of these are protected by job_lock.
63 struct list_head complete_jobs;
64 struct list_head io_jobs;
65 struct list_head pages_jobs;
68 static void wake(struct dm_kcopyd_client *kc)
70 queue_work(kc->kcopyd_wq, &kc->kcopyd_work);
73 static struct page_list *alloc_pl(void)
77 pl = kmalloc(sizeof(*pl), GFP_KERNEL);
81 pl->page = alloc_page(GFP_KERNEL);
90 static void free_pl(struct page_list *pl)
92 __free_page(pl->page);
96 static int kcopyd_get_pages(struct dm_kcopyd_client *kc,
97 unsigned int nr, struct page_list **pages)
101 if (kc->nr_free_pages < nr)
104 kc->nr_free_pages -= nr;
105 for (*pages = pl = kc->pages; --nr; pl = pl->next)
108 kc->pages = pl->next;
114 static void kcopyd_put_pages(struct dm_kcopyd_client *kc, struct page_list *pl)
116 struct page_list *cursor;
118 for (cursor = pl; cursor->next; cursor = cursor->next)
122 cursor->next = kc->pages;
127 * These three functions resize the page pool.
129 static void drop_pages(struct page_list *pl)
131 struct page_list *next;
140 static int client_alloc_pages(struct dm_kcopyd_client *kc, unsigned int nr)
143 struct page_list *pl = NULL, *next;
145 for (i = 0; i < nr; i++) {
156 kcopyd_put_pages(kc, pl);
161 static void client_free_pages(struct dm_kcopyd_client *kc)
163 BUG_ON(kc->nr_free_pages != kc->nr_pages);
164 drop_pages(kc->pages);
166 kc->nr_free_pages = kc->nr_pages = 0;
169 /*-----------------------------------------------------------------
170 * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
171 * for this reason we use a mempool to prevent the client from
172 * ever having to do io (which could cause a deadlock).
173 *---------------------------------------------------------------*/
175 struct dm_kcopyd_client *kc;
176 struct list_head list;
180 * Error state of the job.
183 unsigned long write_err;
186 * Either READ or WRITE
189 struct dm_io_region source;
192 * The destinations for the transfer.
194 unsigned int num_dests;
195 struct dm_io_region dests[DM_KCOPYD_MAX_REGIONS];
198 unsigned int nr_pages;
199 struct page_list *pages;
202 * Set this to ensure you are notified when the job has
203 * completed. 'context' is for callback to use.
205 dm_kcopyd_notify_fn fn;
209 * These fields are only used if the job has been split
210 * into more manageable parts.
216 struct kcopyd_job *master_job;
219 static struct kmem_cache *_job_cache;
221 int __init dm_kcopyd_init(void)
223 _job_cache = kmem_cache_create("kcopyd_job",
224 sizeof(struct kcopyd_job) * (SPLIT_COUNT + 1),
225 __alignof__(struct kcopyd_job), 0, NULL);
232 void dm_kcopyd_exit(void)
234 kmem_cache_destroy(_job_cache);
239 * Functions to push and pop a job onto the head of a given job
242 static struct kcopyd_job *pop(struct list_head *jobs,
243 struct dm_kcopyd_client *kc)
245 struct kcopyd_job *job = NULL;
248 spin_lock_irqsave(&kc->job_lock, flags);
250 if (!list_empty(jobs)) {
251 job = list_entry(jobs->next, struct kcopyd_job, list);
252 list_del(&job->list);
254 spin_unlock_irqrestore(&kc->job_lock, flags);
259 static void push(struct list_head *jobs, struct kcopyd_job *job)
262 struct dm_kcopyd_client *kc = job->kc;
264 spin_lock_irqsave(&kc->job_lock, flags);
265 list_add_tail(&job->list, jobs);
266 spin_unlock_irqrestore(&kc->job_lock, flags);
270 static void push_head(struct list_head *jobs, struct kcopyd_job *job)
273 struct dm_kcopyd_client *kc = job->kc;
275 spin_lock_irqsave(&kc->job_lock, flags);
276 list_add(&job->list, jobs);
277 spin_unlock_irqrestore(&kc->job_lock, flags);
281 * These three functions process 1 item from the corresponding
287 * > 0: can't process yet.
289 static int run_complete_job(struct kcopyd_job *job)
291 void *context = job->context;
292 int read_err = job->read_err;
293 unsigned long write_err = job->write_err;
294 dm_kcopyd_notify_fn fn = job->fn;
295 struct dm_kcopyd_client *kc = job->kc;
298 kcopyd_put_pages(kc, job->pages);
300 * If this is the master job, the sub jobs have already
301 * completed so we can free everything.
303 if (job->master_job == job)
304 mempool_free(job, kc->job_pool);
305 fn(read_err, write_err, context);
307 if (atomic_dec_and_test(&kc->nr_jobs))
308 wake_up(&kc->destroyq);
313 static void complete_io(unsigned long error, void *context)
315 struct kcopyd_job *job = (struct kcopyd_job *) context;
316 struct dm_kcopyd_client *kc = job->kc;
319 if (job->rw == WRITE)
320 job->write_err |= error;
324 if (!test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
325 push(&kc->complete_jobs, job);
331 if (job->rw == WRITE)
332 push(&kc->complete_jobs, job);
336 push(&kc->io_jobs, job);
343 * Request io on as many buffer heads as we can currently get for
346 static int run_io_job(struct kcopyd_job *job)
349 struct dm_io_request io_req = {
351 .mem.type = DM_IO_PAGE_LIST,
352 .mem.ptr.pl = job->pages,
353 .mem.offset = job->offset,
354 .notify.fn = complete_io,
355 .notify.context = job,
356 .client = job->kc->io_client,
360 r = dm_io(&io_req, 1, &job->source, NULL);
362 r = dm_io(&io_req, job->num_dests, job->dests, NULL);
367 static int run_pages_job(struct kcopyd_job *job)
371 job->nr_pages = dm_div_up(job->dests[0].count + job->offset,
373 r = kcopyd_get_pages(job->kc, job->nr_pages, &job->pages);
375 /* this job is ready for io */
376 push(&job->kc->io_jobs, job);
381 /* can't complete now */
388 * Run through a list for as long as possible. Returns the count
389 * of successful jobs.
391 static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
392 int (*fn) (struct kcopyd_job *))
394 struct kcopyd_job *job;
397 while ((job = pop(jobs, kc))) {
402 /* error this rogue job */
403 if (job->rw == WRITE)
404 job->write_err = (unsigned long) -1L;
407 push(&kc->complete_jobs, job);
413 * We couldn't service this job ATM, so
414 * push this job back onto the list.
416 push_head(jobs, job);
427 * kcopyd does this every time it's woken up.
429 static void do_work(struct work_struct *work)
431 struct dm_kcopyd_client *kc = container_of(work,
432 struct dm_kcopyd_client, kcopyd_work);
433 struct blk_plug plug;
436 * The order that these are called is *very* important.
437 * complete jobs can free some pages for pages jobs.
438 * Pages jobs when successful will jump onto the io jobs
439 * list. io jobs call wake when they complete and it all
442 blk_start_plug(&plug);
443 process_jobs(&kc->complete_jobs, kc, run_complete_job);
444 process_jobs(&kc->pages_jobs, kc, run_pages_job);
445 process_jobs(&kc->io_jobs, kc, run_io_job);
446 blk_finish_plug(&plug);
450 * If we are copying a small region we just dispatch a single job
451 * to do the copy, otherwise the io has to be split up into many
454 static void dispatch_job(struct kcopyd_job *job)
456 struct dm_kcopyd_client *kc = job->kc;
457 atomic_inc(&kc->nr_jobs);
458 if (unlikely(!job->source.count))
459 push(&kc->complete_jobs, job);
461 push(&kc->pages_jobs, job);
465 static void segment_complete(int read_err, unsigned long write_err,
468 /* FIXME: tidy this function */
469 sector_t progress = 0;
471 struct kcopyd_job *sub_job = (struct kcopyd_job *) context;
472 struct kcopyd_job *job = sub_job->master_job;
473 struct dm_kcopyd_client *kc = job->kc;
475 mutex_lock(&job->lock);
477 /* update the error */
482 job->write_err |= write_err;
485 * Only dispatch more work if there hasn't been an error.
487 if ((!job->read_err && !job->write_err) ||
488 test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
489 /* get the next chunk of work */
490 progress = job->progress;
491 count = job->source.count - progress;
493 if (count > SUB_JOB_SIZE)
494 count = SUB_JOB_SIZE;
496 job->progress += count;
499 mutex_unlock(&job->lock);
505 sub_job->source.sector += progress;
506 sub_job->source.count = count;
508 for (i = 0; i < job->num_dests; i++) {
509 sub_job->dests[i].sector += progress;
510 sub_job->dests[i].count = count;
513 sub_job->fn = segment_complete;
514 sub_job->context = sub_job;
515 dispatch_job(sub_job);
517 } else if (atomic_dec_and_test(&job->sub_jobs)) {
520 * Queue the completion callback to the kcopyd thread.
522 * Some callers assume that all the completions are called
523 * from a single thread and don't race with each other.
525 * We must not call the callback directly here because this
526 * code may not be executing in the thread.
528 push(&kc->complete_jobs, job);
534 * Create some sub jobs to share the work between them.
536 static void split_job(struct kcopyd_job *master_job)
540 atomic_inc(&master_job->kc->nr_jobs);
542 atomic_set(&master_job->sub_jobs, SPLIT_COUNT);
543 for (i = 0; i < SPLIT_COUNT; i++) {
544 master_job[i + 1].master_job = master_job;
545 segment_complete(0, 0u, &master_job[i + 1]);
549 int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
550 unsigned int num_dests, struct dm_io_region *dests,
551 unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
553 struct kcopyd_job *job;
556 * Allocate an array of jobs consisting of one master job
557 * followed by SPLIT_COUNT sub jobs.
559 job = mempool_alloc(kc->job_pool, GFP_NOIO);
562 * set up for the read.
572 job->num_dests = num_dests;
573 memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
580 job->context = context;
581 job->master_job = job;
583 if (job->source.count <= SUB_JOB_SIZE)
586 mutex_init(&job->lock);
593 EXPORT_SYMBOL(dm_kcopyd_copy);
596 * Cancels a kcopyd job, eg. someone might be deactivating a
600 int kcopyd_cancel(struct kcopyd_job *job, int block)
607 /*-----------------------------------------------------------------
609 *---------------------------------------------------------------*/
610 int dm_kcopyd_client_create(unsigned int nr_pages,
611 struct dm_kcopyd_client **result)
614 struct dm_kcopyd_client *kc;
616 kc = kmalloc(sizeof(*kc), GFP_KERNEL);
620 spin_lock_init(&kc->job_lock);
621 INIT_LIST_HEAD(&kc->complete_jobs);
622 INIT_LIST_HEAD(&kc->io_jobs);
623 INIT_LIST_HEAD(&kc->pages_jobs);
625 kc->job_pool = mempool_create_slab_pool(MIN_JOBS, _job_cache);
629 INIT_WORK(&kc->kcopyd_work, do_work);
630 kc->kcopyd_wq = alloc_workqueue("kcopyd",
631 WQ_NON_REENTRANT | WQ_MEM_RECLAIM, 0);
636 kc->nr_pages = kc->nr_free_pages = 0;
637 r = client_alloc_pages(kc, nr_pages);
639 goto bad_client_pages;
641 kc->io_client = dm_io_client_create(nr_pages);
642 if (IS_ERR(kc->io_client)) {
643 r = PTR_ERR(kc->io_client);
647 init_waitqueue_head(&kc->destroyq);
648 atomic_set(&kc->nr_jobs, 0);
654 client_free_pages(kc);
656 destroy_workqueue(kc->kcopyd_wq);
658 mempool_destroy(kc->job_pool);
664 EXPORT_SYMBOL(dm_kcopyd_client_create);
666 void dm_kcopyd_client_destroy(struct dm_kcopyd_client *kc)
668 /* Wait for completion of all jobs submitted by this client. */
669 wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
671 BUG_ON(!list_empty(&kc->complete_jobs));
672 BUG_ON(!list_empty(&kc->io_jobs));
673 BUG_ON(!list_empty(&kc->pages_jobs));
674 destroy_workqueue(kc->kcopyd_wq);
675 dm_io_client_destroy(kc->io_client);
676 client_free_pages(kc);
677 mempool_destroy(kc->job_pool);
680 EXPORT_SYMBOL(dm_kcopyd_client_destroy);