xref: /openbmc/linux/drivers/md/dm-kcopyd.c (revision 65417d9f)
1 /*
2  * Copyright (C) 2002 Sistina Software (UK) Limited.
3  * Copyright (C) 2006 Red Hat GmbH
4  *
5  * This file is released under the GPL.
6  *
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.
10  */
11 
12 #include <linux/types.h>
13 #include <linux/atomic.h>
14 #include <linux/blkdev.h>
15 #include <linux/fs.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/delay.h>
26 #include <linux/device-mapper.h>
27 #include <linux/dm-kcopyd.h>
28 
29 #include "dm-core.h"
30 
31 #define SUB_JOB_SIZE	128
32 #define SPLIT_COUNT	8
33 #define MIN_JOBS	8
34 #define RESERVE_PAGES	(DIV_ROUND_UP(SUB_JOB_SIZE << SECTOR_SHIFT, PAGE_SIZE))
35 
36 /*-----------------------------------------------------------------
37  * Each kcopyd client has its own little pool of preallocated
38  * pages for kcopyd io.
39  *---------------------------------------------------------------*/
40 struct dm_kcopyd_client {
41 	struct page_list *pages;
42 	unsigned nr_reserved_pages;
43 	unsigned nr_free_pages;
44 
45 	struct dm_io_client *io_client;
46 
47 	wait_queue_head_t destroyq;
48 	atomic_t nr_jobs;
49 
50 	mempool_t *job_pool;
51 
52 	struct workqueue_struct *kcopyd_wq;
53 	struct work_struct kcopyd_work;
54 
55 	struct dm_kcopyd_throttle *throttle;
56 
57 /*
58  * We maintain three lists of jobs:
59  *
60  * i)   jobs waiting for pages
61  * ii)  jobs that have pages, and are waiting for the io to be issued.
62  * iii) jobs that have completed.
63  *
64  * All three of these are protected by job_lock.
65  */
66 	spinlock_t job_lock;
67 	struct list_head complete_jobs;
68 	struct list_head io_jobs;
69 	struct list_head pages_jobs;
70 };
71 
72 static struct page_list zero_page_list;
73 
74 static DEFINE_SPINLOCK(throttle_spinlock);
75 
76 /*
77  * IO/IDLE accounting slowly decays after (1 << ACCOUNT_INTERVAL_SHIFT) period.
78  * When total_period >= (1 << ACCOUNT_INTERVAL_SHIFT) the counters are divided
79  * by 2.
80  */
81 #define ACCOUNT_INTERVAL_SHIFT		SHIFT_HZ
82 
83 /*
84  * Sleep this number of milliseconds.
85  *
86  * The value was decided experimentally.
87  * Smaller values seem to cause an increased copy rate above the limit.
88  * The reason for this is unknown but possibly due to jiffies rounding errors
89  * or read/write cache inside the disk.
90  */
91 #define SLEEP_MSEC			100
92 
93 /*
94  * Maximum number of sleep events. There is a theoretical livelock if more
95  * kcopyd clients do work simultaneously which this limit avoids.
96  */
97 #define MAX_SLEEPS			10
98 
99 static void io_job_start(struct dm_kcopyd_throttle *t)
100 {
101 	unsigned throttle, now, difference;
102 	int slept = 0, skew;
103 
104 	if (unlikely(!t))
105 		return;
106 
107 try_again:
108 	spin_lock_irq(&throttle_spinlock);
109 
110 	throttle = READ_ONCE(t->throttle);
111 
112 	if (likely(throttle >= 100))
113 		goto skip_limit;
114 
115 	now = jiffies;
116 	difference = now - t->last_jiffies;
117 	t->last_jiffies = now;
118 	if (t->num_io_jobs)
119 		t->io_period += difference;
120 	t->total_period += difference;
121 
122 	/*
123 	 * Maintain sane values if we got a temporary overflow.
124 	 */
125 	if (unlikely(t->io_period > t->total_period))
126 		t->io_period = t->total_period;
127 
128 	if (unlikely(t->total_period >= (1 << ACCOUNT_INTERVAL_SHIFT))) {
129 		int shift = fls(t->total_period >> ACCOUNT_INTERVAL_SHIFT);
130 		t->total_period >>= shift;
131 		t->io_period >>= shift;
132 	}
133 
134 	skew = t->io_period - throttle * t->total_period / 100;
135 
136 	if (unlikely(skew > 0) && slept < MAX_SLEEPS) {
137 		slept++;
138 		spin_unlock_irq(&throttle_spinlock);
139 		msleep(SLEEP_MSEC);
140 		goto try_again;
141 	}
142 
143 skip_limit:
144 	t->num_io_jobs++;
145 
146 	spin_unlock_irq(&throttle_spinlock);
147 }
148 
149 static void io_job_finish(struct dm_kcopyd_throttle *t)
150 {
151 	unsigned long flags;
152 
153 	if (unlikely(!t))
154 		return;
155 
156 	spin_lock_irqsave(&throttle_spinlock, flags);
157 
158 	t->num_io_jobs--;
159 
160 	if (likely(READ_ONCE(t->throttle) >= 100))
161 		goto skip_limit;
162 
163 	if (!t->num_io_jobs) {
164 		unsigned now, difference;
165 
166 		now = jiffies;
167 		difference = now - t->last_jiffies;
168 		t->last_jiffies = now;
169 
170 		t->io_period += difference;
171 		t->total_period += difference;
172 
173 		/*
174 		 * Maintain sane values if we got a temporary overflow.
175 		 */
176 		if (unlikely(t->io_period > t->total_period))
177 			t->io_period = t->total_period;
178 	}
179 
180 skip_limit:
181 	spin_unlock_irqrestore(&throttle_spinlock, flags);
182 }
183 
184 
185 static void wake(struct dm_kcopyd_client *kc)
186 {
187 	queue_work(kc->kcopyd_wq, &kc->kcopyd_work);
188 }
189 
190 /*
191  * Obtain one page for the use of kcopyd.
192  */
193 static struct page_list *alloc_pl(gfp_t gfp)
194 {
195 	struct page_list *pl;
196 
197 	pl = kmalloc(sizeof(*pl), gfp);
198 	if (!pl)
199 		return NULL;
200 
201 	pl->page = alloc_page(gfp);
202 	if (!pl->page) {
203 		kfree(pl);
204 		return NULL;
205 	}
206 
207 	return pl;
208 }
209 
210 static void free_pl(struct page_list *pl)
211 {
212 	__free_page(pl->page);
213 	kfree(pl);
214 }
215 
216 /*
217  * Add the provided pages to a client's free page list, releasing
218  * back to the system any beyond the reserved_pages limit.
219  */
220 static void kcopyd_put_pages(struct dm_kcopyd_client *kc, struct page_list *pl)
221 {
222 	struct page_list *next;
223 
224 	do {
225 		next = pl->next;
226 
227 		if (kc->nr_free_pages >= kc->nr_reserved_pages)
228 			free_pl(pl);
229 		else {
230 			pl->next = kc->pages;
231 			kc->pages = pl;
232 			kc->nr_free_pages++;
233 		}
234 
235 		pl = next;
236 	} while (pl);
237 }
238 
239 static int kcopyd_get_pages(struct dm_kcopyd_client *kc,
240 			    unsigned int nr, struct page_list **pages)
241 {
242 	struct page_list *pl;
243 
244 	*pages = NULL;
245 
246 	do {
247 		pl = alloc_pl(__GFP_NOWARN | __GFP_NORETRY | __GFP_KSWAPD_RECLAIM);
248 		if (unlikely(!pl)) {
249 			/* Use reserved pages */
250 			pl = kc->pages;
251 			if (unlikely(!pl))
252 				goto out_of_memory;
253 			kc->pages = pl->next;
254 			kc->nr_free_pages--;
255 		}
256 		pl->next = *pages;
257 		*pages = pl;
258 	} while (--nr);
259 
260 	return 0;
261 
262 out_of_memory:
263 	if (*pages)
264 		kcopyd_put_pages(kc, *pages);
265 	return -ENOMEM;
266 }
267 
268 /*
269  * These three functions resize the page pool.
270  */
271 static void drop_pages(struct page_list *pl)
272 {
273 	struct page_list *next;
274 
275 	while (pl) {
276 		next = pl->next;
277 		free_pl(pl);
278 		pl = next;
279 	}
280 }
281 
282 /*
283  * Allocate and reserve nr_pages for the use of a specific client.
284  */
285 static int client_reserve_pages(struct dm_kcopyd_client *kc, unsigned nr_pages)
286 {
287 	unsigned i;
288 	struct page_list *pl = NULL, *next;
289 
290 	for (i = 0; i < nr_pages; i++) {
291 		next = alloc_pl(GFP_KERNEL);
292 		if (!next) {
293 			if (pl)
294 				drop_pages(pl);
295 			return -ENOMEM;
296 		}
297 		next->next = pl;
298 		pl = next;
299 	}
300 
301 	kc->nr_reserved_pages += nr_pages;
302 	kcopyd_put_pages(kc, pl);
303 
304 	return 0;
305 }
306 
307 static void client_free_pages(struct dm_kcopyd_client *kc)
308 {
309 	BUG_ON(kc->nr_free_pages != kc->nr_reserved_pages);
310 	drop_pages(kc->pages);
311 	kc->pages = NULL;
312 	kc->nr_free_pages = kc->nr_reserved_pages = 0;
313 }
314 
315 /*-----------------------------------------------------------------
316  * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
317  * for this reason we use a mempool to prevent the client from
318  * ever having to do io (which could cause a deadlock).
319  *---------------------------------------------------------------*/
320 struct kcopyd_job {
321 	struct dm_kcopyd_client *kc;
322 	struct list_head list;
323 	unsigned long flags;
324 
325 	/*
326 	 * Error state of the job.
327 	 */
328 	int read_err;
329 	unsigned long write_err;
330 
331 	/*
332 	 * Either READ or WRITE
333 	 */
334 	int rw;
335 	struct dm_io_region source;
336 
337 	/*
338 	 * The destinations for the transfer.
339 	 */
340 	unsigned int num_dests;
341 	struct dm_io_region dests[DM_KCOPYD_MAX_REGIONS];
342 
343 	struct page_list *pages;
344 
345 	/*
346 	 * Set this to ensure you are notified when the job has
347 	 * completed.  'context' is for callback to use.
348 	 */
349 	dm_kcopyd_notify_fn fn;
350 	void *context;
351 
352 	/*
353 	 * These fields are only used if the job has been split
354 	 * into more manageable parts.
355 	 */
356 	struct mutex lock;
357 	atomic_t sub_jobs;
358 	sector_t progress;
359 	sector_t write_offset;
360 
361 	struct kcopyd_job *master_job;
362 };
363 
364 static struct kmem_cache *_job_cache;
365 
366 int __init dm_kcopyd_init(void)
367 {
368 	_job_cache = kmem_cache_create("kcopyd_job",
369 				sizeof(struct kcopyd_job) * (SPLIT_COUNT + 1),
370 				__alignof__(struct kcopyd_job), 0, NULL);
371 	if (!_job_cache)
372 		return -ENOMEM;
373 
374 	zero_page_list.next = &zero_page_list;
375 	zero_page_list.page = ZERO_PAGE(0);
376 
377 	return 0;
378 }
379 
380 void dm_kcopyd_exit(void)
381 {
382 	kmem_cache_destroy(_job_cache);
383 	_job_cache = NULL;
384 }
385 
386 /*
387  * Functions to push and pop a job onto the head of a given job
388  * list.
389  */
390 static struct kcopyd_job *pop_io_job(struct list_head *jobs,
391 				     struct dm_kcopyd_client *kc)
392 {
393 	struct kcopyd_job *job;
394 
395 	/*
396 	 * For I/O jobs, pop any read, any write without sequential write
397 	 * constraint and sequential writes that are at the right position.
398 	 */
399 	list_for_each_entry(job, jobs, list) {
400 		if (job->rw == READ || !test_bit(DM_KCOPYD_WRITE_SEQ, &job->flags)) {
401 			list_del(&job->list);
402 			return job;
403 		}
404 
405 		if (job->write_offset == job->master_job->write_offset) {
406 			job->master_job->write_offset += job->source.count;
407 			list_del(&job->list);
408 			return job;
409 		}
410 	}
411 
412 	return NULL;
413 }
414 
415 static struct kcopyd_job *pop(struct list_head *jobs,
416 			      struct dm_kcopyd_client *kc)
417 {
418 	struct kcopyd_job *job = NULL;
419 	unsigned long flags;
420 
421 	spin_lock_irqsave(&kc->job_lock, flags);
422 
423 	if (!list_empty(jobs)) {
424 		if (jobs == &kc->io_jobs)
425 			job = pop_io_job(jobs, kc);
426 		else {
427 			job = list_entry(jobs->next, struct kcopyd_job, list);
428 			list_del(&job->list);
429 		}
430 	}
431 	spin_unlock_irqrestore(&kc->job_lock, flags);
432 
433 	return job;
434 }
435 
436 static void push(struct list_head *jobs, struct kcopyd_job *job)
437 {
438 	unsigned long flags;
439 	struct dm_kcopyd_client *kc = job->kc;
440 
441 	spin_lock_irqsave(&kc->job_lock, flags);
442 	list_add_tail(&job->list, jobs);
443 	spin_unlock_irqrestore(&kc->job_lock, flags);
444 }
445 
446 
447 static void push_head(struct list_head *jobs, struct kcopyd_job *job)
448 {
449 	unsigned long flags;
450 	struct dm_kcopyd_client *kc = job->kc;
451 
452 	spin_lock_irqsave(&kc->job_lock, flags);
453 	list_add(&job->list, jobs);
454 	spin_unlock_irqrestore(&kc->job_lock, flags);
455 }
456 
457 /*
458  * These three functions process 1 item from the corresponding
459  * job list.
460  *
461  * They return:
462  * < 0: error
463  *   0: success
464  * > 0: can't process yet.
465  */
466 static int run_complete_job(struct kcopyd_job *job)
467 {
468 	void *context = job->context;
469 	int read_err = job->read_err;
470 	unsigned long write_err = job->write_err;
471 	dm_kcopyd_notify_fn fn = job->fn;
472 	struct dm_kcopyd_client *kc = job->kc;
473 
474 	if (job->pages && job->pages != &zero_page_list)
475 		kcopyd_put_pages(kc, job->pages);
476 	/*
477 	 * If this is the master job, the sub jobs have already
478 	 * completed so we can free everything.
479 	 */
480 	if (job->master_job == job)
481 		mempool_free(job, kc->job_pool);
482 	fn(read_err, write_err, context);
483 
484 	if (atomic_dec_and_test(&kc->nr_jobs))
485 		wake_up(&kc->destroyq);
486 
487 	return 0;
488 }
489 
490 static void complete_io(unsigned long error, void *context)
491 {
492 	struct kcopyd_job *job = (struct kcopyd_job *) context;
493 	struct dm_kcopyd_client *kc = job->kc;
494 
495 	io_job_finish(kc->throttle);
496 
497 	if (error) {
498 		if (op_is_write(job->rw))
499 			job->write_err |= error;
500 		else
501 			job->read_err = 1;
502 
503 		if (!test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
504 			push(&kc->complete_jobs, job);
505 			wake(kc);
506 			return;
507 		}
508 	}
509 
510 	if (op_is_write(job->rw))
511 		push(&kc->complete_jobs, job);
512 
513 	else {
514 		job->rw = WRITE;
515 		push(&kc->io_jobs, job);
516 	}
517 
518 	wake(kc);
519 }
520 
521 /*
522  * Request io on as many buffer heads as we can currently get for
523  * a particular job.
524  */
525 static int run_io_job(struct kcopyd_job *job)
526 {
527 	int r;
528 	struct dm_io_request io_req = {
529 		.bi_op = job->rw,
530 		.bi_op_flags = 0,
531 		.mem.type = DM_IO_PAGE_LIST,
532 		.mem.ptr.pl = job->pages,
533 		.mem.offset = 0,
534 		.notify.fn = complete_io,
535 		.notify.context = job,
536 		.client = job->kc->io_client,
537 	};
538 
539 	/*
540 	 * If we need to write sequentially and some reads or writes failed,
541 	 * no point in continuing.
542 	 */
543 	if (test_bit(DM_KCOPYD_WRITE_SEQ, &job->flags) &&
544 	    job->master_job->write_err)
545 		return -EIO;
546 
547 	io_job_start(job->kc->throttle);
548 
549 	if (job->rw == READ)
550 		r = dm_io(&io_req, 1, &job->source, NULL);
551 	else
552 		r = dm_io(&io_req, job->num_dests, job->dests, NULL);
553 
554 	return r;
555 }
556 
557 static int run_pages_job(struct kcopyd_job *job)
558 {
559 	int r;
560 	unsigned nr_pages = dm_div_up(job->dests[0].count, PAGE_SIZE >> 9);
561 
562 	r = kcopyd_get_pages(job->kc, nr_pages, &job->pages);
563 	if (!r) {
564 		/* this job is ready for io */
565 		push(&job->kc->io_jobs, job);
566 		return 0;
567 	}
568 
569 	if (r == -ENOMEM)
570 		/* can't complete now */
571 		return 1;
572 
573 	return r;
574 }
575 
576 /*
577  * Run through a list for as long as possible.  Returns the count
578  * of successful jobs.
579  */
580 static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
581 			int (*fn) (struct kcopyd_job *))
582 {
583 	struct kcopyd_job *job;
584 	int r, count = 0;
585 
586 	while ((job = pop(jobs, kc))) {
587 
588 		r = fn(job);
589 
590 		if (r < 0) {
591 			/* error this rogue job */
592 			if (op_is_write(job->rw))
593 				job->write_err = (unsigned long) -1L;
594 			else
595 				job->read_err = 1;
596 			push(&kc->complete_jobs, job);
597 			break;
598 		}
599 
600 		if (r > 0) {
601 			/*
602 			 * We couldn't service this job ATM, so
603 			 * push this job back onto the list.
604 			 */
605 			push_head(jobs, job);
606 			break;
607 		}
608 
609 		count++;
610 	}
611 
612 	return count;
613 }
614 
615 /*
616  * kcopyd does this every time it's woken up.
617  */
618 static void do_work(struct work_struct *work)
619 {
620 	struct dm_kcopyd_client *kc = container_of(work,
621 					struct dm_kcopyd_client, kcopyd_work);
622 	struct blk_plug plug;
623 
624 	/*
625 	 * The order that these are called is *very* important.
626 	 * complete jobs can free some pages for pages jobs.
627 	 * Pages jobs when successful will jump onto the io jobs
628 	 * list.  io jobs call wake when they complete and it all
629 	 * starts again.
630 	 */
631 	blk_start_plug(&plug);
632 	process_jobs(&kc->complete_jobs, kc, run_complete_job);
633 	process_jobs(&kc->pages_jobs, kc, run_pages_job);
634 	process_jobs(&kc->io_jobs, kc, run_io_job);
635 	blk_finish_plug(&plug);
636 }
637 
638 /*
639  * If we are copying a small region we just dispatch a single job
640  * to do the copy, otherwise the io has to be split up into many
641  * jobs.
642  */
643 static void dispatch_job(struct kcopyd_job *job)
644 {
645 	struct dm_kcopyd_client *kc = job->kc;
646 	atomic_inc(&kc->nr_jobs);
647 	if (unlikely(!job->source.count))
648 		push(&kc->complete_jobs, job);
649 	else if (job->pages == &zero_page_list)
650 		push(&kc->io_jobs, job);
651 	else
652 		push(&kc->pages_jobs, job);
653 	wake(kc);
654 }
655 
656 static void segment_complete(int read_err, unsigned long write_err,
657 			     void *context)
658 {
659 	/* FIXME: tidy this function */
660 	sector_t progress = 0;
661 	sector_t count = 0;
662 	struct kcopyd_job *sub_job = (struct kcopyd_job *) context;
663 	struct kcopyd_job *job = sub_job->master_job;
664 	struct dm_kcopyd_client *kc = job->kc;
665 
666 	mutex_lock(&job->lock);
667 
668 	/* update the error */
669 	if (read_err)
670 		job->read_err = 1;
671 
672 	if (write_err)
673 		job->write_err |= write_err;
674 
675 	/*
676 	 * Only dispatch more work if there hasn't been an error.
677 	 */
678 	if ((!job->read_err && !job->write_err) ||
679 	    test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
680 		/* get the next chunk of work */
681 		progress = job->progress;
682 		count = job->source.count - progress;
683 		if (count) {
684 			if (count > SUB_JOB_SIZE)
685 				count = SUB_JOB_SIZE;
686 
687 			job->progress += count;
688 		}
689 	}
690 	mutex_unlock(&job->lock);
691 
692 	if (count) {
693 		int i;
694 
695 		*sub_job = *job;
696 		sub_job->write_offset = progress;
697 		sub_job->source.sector += progress;
698 		sub_job->source.count = count;
699 
700 		for (i = 0; i < job->num_dests; i++) {
701 			sub_job->dests[i].sector += progress;
702 			sub_job->dests[i].count = count;
703 		}
704 
705 		sub_job->fn = segment_complete;
706 		sub_job->context = sub_job;
707 		dispatch_job(sub_job);
708 
709 	} else if (atomic_dec_and_test(&job->sub_jobs)) {
710 
711 		/*
712 		 * Queue the completion callback to the kcopyd thread.
713 		 *
714 		 * Some callers assume that all the completions are called
715 		 * from a single thread and don't race with each other.
716 		 *
717 		 * We must not call the callback directly here because this
718 		 * code may not be executing in the thread.
719 		 */
720 		push(&kc->complete_jobs, job);
721 		wake(kc);
722 	}
723 }
724 
725 /*
726  * Create some sub jobs to share the work between them.
727  */
728 static void split_job(struct kcopyd_job *master_job)
729 {
730 	int i;
731 
732 	atomic_inc(&master_job->kc->nr_jobs);
733 
734 	atomic_set(&master_job->sub_jobs, SPLIT_COUNT);
735 	for (i = 0; i < SPLIT_COUNT; i++) {
736 		master_job[i + 1].master_job = master_job;
737 		segment_complete(0, 0u, &master_job[i + 1]);
738 	}
739 }
740 
741 int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
742 		   unsigned int num_dests, struct dm_io_region *dests,
743 		   unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
744 {
745 	struct kcopyd_job *job;
746 	int i;
747 
748 	/*
749 	 * Allocate an array of jobs consisting of one master job
750 	 * followed by SPLIT_COUNT sub jobs.
751 	 */
752 	job = mempool_alloc(kc->job_pool, GFP_NOIO);
753 
754 	/*
755 	 * set up for the read.
756 	 */
757 	job->kc = kc;
758 	job->flags = flags;
759 	job->read_err = 0;
760 	job->write_err = 0;
761 
762 	job->num_dests = num_dests;
763 	memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
764 
765 	/*
766 	 * If one of the destination is a host-managed zoned block device,
767 	 * we need to write sequentially. If one of the destination is a
768 	 * host-aware device, then leave it to the caller to choose what to do.
769 	 */
770 	if (!test_bit(DM_KCOPYD_WRITE_SEQ, &job->flags)) {
771 		for (i = 0; i < job->num_dests; i++) {
772 			if (bdev_zoned_model(dests[i].bdev) == BLK_ZONED_HM) {
773 				set_bit(DM_KCOPYD_WRITE_SEQ, &job->flags);
774 				break;
775 			}
776 		}
777 	}
778 
779 	/*
780 	 * If we need to write sequentially, errors cannot be ignored.
781 	 */
782 	if (test_bit(DM_KCOPYD_WRITE_SEQ, &job->flags) &&
783 	    test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags))
784 		clear_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags);
785 
786 	if (from) {
787 		job->source = *from;
788 		job->pages = NULL;
789 		job->rw = READ;
790 	} else {
791 		memset(&job->source, 0, sizeof job->source);
792 		job->source.count = job->dests[0].count;
793 		job->pages = &zero_page_list;
794 
795 		/*
796 		 * Use WRITE ZEROES to optimize zeroing if all dests support it.
797 		 */
798 		job->rw = REQ_OP_WRITE_ZEROES;
799 		for (i = 0; i < job->num_dests; i++)
800 			if (!bdev_write_zeroes_sectors(job->dests[i].bdev)) {
801 				job->rw = WRITE;
802 				break;
803 			}
804 	}
805 
806 	job->fn = fn;
807 	job->context = context;
808 	job->master_job = job;
809 	job->write_offset = 0;
810 
811 	if (job->source.count <= SUB_JOB_SIZE)
812 		dispatch_job(job);
813 	else {
814 		mutex_init(&job->lock);
815 		job->progress = 0;
816 		split_job(job);
817 	}
818 
819 	return 0;
820 }
821 EXPORT_SYMBOL(dm_kcopyd_copy);
822 
823 int dm_kcopyd_zero(struct dm_kcopyd_client *kc,
824 		   unsigned num_dests, struct dm_io_region *dests,
825 		   unsigned flags, dm_kcopyd_notify_fn fn, void *context)
826 {
827 	return dm_kcopyd_copy(kc, NULL, num_dests, dests, flags, fn, context);
828 }
829 EXPORT_SYMBOL(dm_kcopyd_zero);
830 
831 void *dm_kcopyd_prepare_callback(struct dm_kcopyd_client *kc,
832 				 dm_kcopyd_notify_fn fn, void *context)
833 {
834 	struct kcopyd_job *job;
835 
836 	job = mempool_alloc(kc->job_pool, GFP_NOIO);
837 
838 	memset(job, 0, sizeof(struct kcopyd_job));
839 	job->kc = kc;
840 	job->fn = fn;
841 	job->context = context;
842 	job->master_job = job;
843 
844 	atomic_inc(&kc->nr_jobs);
845 
846 	return job;
847 }
848 EXPORT_SYMBOL(dm_kcopyd_prepare_callback);
849 
850 void dm_kcopyd_do_callback(void *j, int read_err, unsigned long write_err)
851 {
852 	struct kcopyd_job *job = j;
853 	struct dm_kcopyd_client *kc = job->kc;
854 
855 	job->read_err = read_err;
856 	job->write_err = write_err;
857 
858 	push(&kc->complete_jobs, job);
859 	wake(kc);
860 }
861 EXPORT_SYMBOL(dm_kcopyd_do_callback);
862 
863 /*
864  * Cancels a kcopyd job, eg. someone might be deactivating a
865  * mirror.
866  */
867 #if 0
868 int kcopyd_cancel(struct kcopyd_job *job, int block)
869 {
870 	/* FIXME: finish */
871 	return -1;
872 }
873 #endif  /*  0  */
874 
875 /*-----------------------------------------------------------------
876  * Client setup
877  *---------------------------------------------------------------*/
878 struct dm_kcopyd_client *dm_kcopyd_client_create(struct dm_kcopyd_throttle *throttle)
879 {
880 	int r = -ENOMEM;
881 	struct dm_kcopyd_client *kc;
882 
883 	kc = kmalloc(sizeof(*kc), GFP_KERNEL);
884 	if (!kc)
885 		return ERR_PTR(-ENOMEM);
886 
887 	spin_lock_init(&kc->job_lock);
888 	INIT_LIST_HEAD(&kc->complete_jobs);
889 	INIT_LIST_HEAD(&kc->io_jobs);
890 	INIT_LIST_HEAD(&kc->pages_jobs);
891 	kc->throttle = throttle;
892 
893 	kc->job_pool = mempool_create_slab_pool(MIN_JOBS, _job_cache);
894 	if (!kc->job_pool)
895 		goto bad_slab;
896 
897 	INIT_WORK(&kc->kcopyd_work, do_work);
898 	kc->kcopyd_wq = alloc_workqueue("kcopyd", WQ_MEM_RECLAIM, 0);
899 	if (!kc->kcopyd_wq)
900 		goto bad_workqueue;
901 
902 	kc->pages = NULL;
903 	kc->nr_reserved_pages = kc->nr_free_pages = 0;
904 	r = client_reserve_pages(kc, RESERVE_PAGES);
905 	if (r)
906 		goto bad_client_pages;
907 
908 	kc->io_client = dm_io_client_create();
909 	if (IS_ERR(kc->io_client)) {
910 		r = PTR_ERR(kc->io_client);
911 		goto bad_io_client;
912 	}
913 
914 	init_waitqueue_head(&kc->destroyq);
915 	atomic_set(&kc->nr_jobs, 0);
916 
917 	return kc;
918 
919 bad_io_client:
920 	client_free_pages(kc);
921 bad_client_pages:
922 	destroy_workqueue(kc->kcopyd_wq);
923 bad_workqueue:
924 	mempool_destroy(kc->job_pool);
925 bad_slab:
926 	kfree(kc);
927 
928 	return ERR_PTR(r);
929 }
930 EXPORT_SYMBOL(dm_kcopyd_client_create);
931 
932 void dm_kcopyd_client_destroy(struct dm_kcopyd_client *kc)
933 {
934 	/* Wait for completion of all jobs submitted by this client. */
935 	wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
936 
937 	BUG_ON(!list_empty(&kc->complete_jobs));
938 	BUG_ON(!list_empty(&kc->io_jobs));
939 	BUG_ON(!list_empty(&kc->pages_jobs));
940 	destroy_workqueue(kc->kcopyd_wq);
941 	dm_io_client_destroy(kc->io_client);
942 	client_free_pages(kc);
943 	mempool_destroy(kc->job_pool);
944 	kfree(kc);
945 }
946 EXPORT_SYMBOL(dm_kcopyd_client_destroy);
947