xref: /openbmc/linux/drivers/md/dm-kcopyd.c (revision c9dc580c)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2002 Sistina Software (UK) Limited.
4  * Copyright (C) 2006 Red Hat GmbH
5  *
6  * This file is released under the GPL.
7  *
8  * Kcopyd provides a simple interface for copying an area of one
9  * block-device to one or more other block-devices, with an asynchronous
10  * completion notification.
11  */
12 
13 #include <linux/types.h>
14 #include <linux/atomic.h>
15 #include <linux/blkdev.h>
16 #include <linux/fs.h>
17 #include <linux/init.h>
18 #include <linux/list.h>
19 #include <linux/mempool.h>
20 #include <linux/module.h>
21 #include <linux/pagemap.h>
22 #include <linux/slab.h>
23 #include <linux/vmalloc.h>
24 #include <linux/workqueue.h>
25 #include <linux/mutex.h>
26 #include <linux/delay.h>
27 #include <linux/device-mapper.h>
28 #include <linux/dm-kcopyd.h>
29 
30 #include "dm-core.h"
31 
32 #define SPLIT_COUNT	8
33 #define MIN_JOBS	8
34 
35 #define DEFAULT_SUB_JOB_SIZE_KB 512
36 #define MAX_SUB_JOB_SIZE_KB     1024
37 
38 static unsigned int kcopyd_subjob_size_kb = DEFAULT_SUB_JOB_SIZE_KB;
39 
40 module_param(kcopyd_subjob_size_kb, uint, 0644);
41 MODULE_PARM_DESC(kcopyd_subjob_size_kb, "Sub-job size for dm-kcopyd clients");
42 
43 static unsigned int dm_get_kcopyd_subjob_size(void)
44 {
45 	unsigned int sub_job_size_kb;
46 
47 	sub_job_size_kb = __dm_get_module_param(&kcopyd_subjob_size_kb,
48 						DEFAULT_SUB_JOB_SIZE_KB,
49 						MAX_SUB_JOB_SIZE_KB);
50 
51 	return sub_job_size_kb << 1;
52 }
53 
54 /*
55  *----------------------------------------------------------------
56  * Each kcopyd client has its own little pool of preallocated
57  * pages for kcopyd io.
58  *---------------------------------------------------------------
59  */
60 struct dm_kcopyd_client {
61 	struct page_list *pages;
62 	unsigned int nr_reserved_pages;
63 	unsigned int nr_free_pages;
64 	unsigned int sub_job_size;
65 
66 	struct dm_io_client *io_client;
67 
68 	wait_queue_head_t destroyq;
69 
70 	mempool_t job_pool;
71 
72 	struct workqueue_struct *kcopyd_wq;
73 	struct work_struct kcopyd_work;
74 
75 	struct dm_kcopyd_throttle *throttle;
76 
77 	atomic_t nr_jobs;
78 
79 /*
80  * We maintain four lists of jobs:
81  *
82  * i)   jobs waiting for pages
83  * ii)  jobs that have pages, and are waiting for the io to be issued.
84  * iii) jobs that don't need to do any IO and just run a callback
85  * iv) jobs that have completed.
86  *
87  * All four of these are protected by job_lock.
88  */
89 	spinlock_t job_lock;
90 	struct list_head callback_jobs;
91 	struct list_head complete_jobs;
92 	struct list_head io_jobs;
93 	struct list_head pages_jobs;
94 };
95 
96 static struct page_list zero_page_list;
97 
98 static DEFINE_SPINLOCK(throttle_spinlock);
99 
100 /*
101  * IO/IDLE accounting slowly decays after (1 << ACCOUNT_INTERVAL_SHIFT) period.
102  * When total_period >= (1 << ACCOUNT_INTERVAL_SHIFT) the counters are divided
103  * by 2.
104  */
105 #define ACCOUNT_INTERVAL_SHIFT		SHIFT_HZ
106 
107 /*
108  * Sleep this number of milliseconds.
109  *
110  * The value was decided experimentally.
111  * Smaller values seem to cause an increased copy rate above the limit.
112  * The reason for this is unknown but possibly due to jiffies rounding errors
113  * or read/write cache inside the disk.
114  */
115 #define SLEEP_USEC			100000
116 
117 /*
118  * Maximum number of sleep events. There is a theoretical livelock if more
119  * kcopyd clients do work simultaneously which this limit avoids.
120  */
121 #define MAX_SLEEPS			10
122 
123 static void io_job_start(struct dm_kcopyd_throttle *t)
124 {
125 	unsigned int throttle, now, difference;
126 	int slept = 0, skew;
127 
128 	if (unlikely(!t))
129 		return;
130 
131 try_again:
132 	spin_lock_irq(&throttle_spinlock);
133 
134 	throttle = READ_ONCE(t->throttle);
135 
136 	if (likely(throttle >= 100))
137 		goto skip_limit;
138 
139 	now = jiffies;
140 	difference = now - t->last_jiffies;
141 	t->last_jiffies = now;
142 	if (t->num_io_jobs)
143 		t->io_period += difference;
144 	t->total_period += difference;
145 
146 	/*
147 	 * Maintain sane values if we got a temporary overflow.
148 	 */
149 	if (unlikely(t->io_period > t->total_period))
150 		t->io_period = t->total_period;
151 
152 	if (unlikely(t->total_period >= (1 << ACCOUNT_INTERVAL_SHIFT))) {
153 		int shift = fls(t->total_period >> ACCOUNT_INTERVAL_SHIFT);
154 
155 		t->total_period >>= shift;
156 		t->io_period >>= shift;
157 	}
158 
159 	skew = t->io_period - throttle * t->total_period / 100;
160 
161 	if (unlikely(skew > 0) && slept < MAX_SLEEPS) {
162 		slept++;
163 		spin_unlock_irq(&throttle_spinlock);
164 		fsleep(SLEEP_USEC);
165 		goto try_again;
166 	}
167 
168 skip_limit:
169 	t->num_io_jobs++;
170 
171 	spin_unlock_irq(&throttle_spinlock);
172 }
173 
174 static void io_job_finish(struct dm_kcopyd_throttle *t)
175 {
176 	unsigned long flags;
177 
178 	if (unlikely(!t))
179 		return;
180 
181 	spin_lock_irqsave(&throttle_spinlock, flags);
182 
183 	t->num_io_jobs--;
184 
185 	if (likely(READ_ONCE(t->throttle) >= 100))
186 		goto skip_limit;
187 
188 	if (!t->num_io_jobs) {
189 		unsigned int now, difference;
190 
191 		now = jiffies;
192 		difference = now - t->last_jiffies;
193 		t->last_jiffies = now;
194 
195 		t->io_period += difference;
196 		t->total_period += difference;
197 
198 		/*
199 		 * Maintain sane values if we got a temporary overflow.
200 		 */
201 		if (unlikely(t->io_period > t->total_period))
202 			t->io_period = t->total_period;
203 	}
204 
205 skip_limit:
206 	spin_unlock_irqrestore(&throttle_spinlock, flags);
207 }
208 
209 
210 static void wake(struct dm_kcopyd_client *kc)
211 {
212 	queue_work(kc->kcopyd_wq, &kc->kcopyd_work);
213 }
214 
215 /*
216  * Obtain one page for the use of kcopyd.
217  */
218 static struct page_list *alloc_pl(gfp_t gfp)
219 {
220 	struct page_list *pl;
221 
222 	pl = kmalloc(sizeof(*pl), gfp);
223 	if (!pl)
224 		return NULL;
225 
226 	pl->page = alloc_page(gfp | __GFP_HIGHMEM);
227 	if (!pl->page) {
228 		kfree(pl);
229 		return NULL;
230 	}
231 
232 	return pl;
233 }
234 
235 static void free_pl(struct page_list *pl)
236 {
237 	__free_page(pl->page);
238 	kfree(pl);
239 }
240 
241 /*
242  * Add the provided pages to a client's free page list, releasing
243  * back to the system any beyond the reserved_pages limit.
244  */
245 static void kcopyd_put_pages(struct dm_kcopyd_client *kc, struct page_list *pl)
246 {
247 	struct page_list *next;
248 
249 	do {
250 		next = pl->next;
251 
252 		if (kc->nr_free_pages >= kc->nr_reserved_pages)
253 			free_pl(pl);
254 		else {
255 			pl->next = kc->pages;
256 			kc->pages = pl;
257 			kc->nr_free_pages++;
258 		}
259 
260 		pl = next;
261 	} while (pl);
262 }
263 
264 static int kcopyd_get_pages(struct dm_kcopyd_client *kc,
265 			    unsigned int nr, struct page_list **pages)
266 {
267 	struct page_list *pl;
268 
269 	*pages = NULL;
270 
271 	do {
272 		pl = alloc_pl(__GFP_NOWARN | __GFP_NORETRY | __GFP_KSWAPD_RECLAIM);
273 		if (unlikely(!pl)) {
274 			/* Use reserved pages */
275 			pl = kc->pages;
276 			if (unlikely(!pl))
277 				goto out_of_memory;
278 			kc->pages = pl->next;
279 			kc->nr_free_pages--;
280 		}
281 		pl->next = *pages;
282 		*pages = pl;
283 	} while (--nr);
284 
285 	return 0;
286 
287 out_of_memory:
288 	if (*pages)
289 		kcopyd_put_pages(kc, *pages);
290 	return -ENOMEM;
291 }
292 
293 /*
294  * These three functions resize the page pool.
295  */
296 static void drop_pages(struct page_list *pl)
297 {
298 	struct page_list *next;
299 
300 	while (pl) {
301 		next = pl->next;
302 		free_pl(pl);
303 		pl = next;
304 	}
305 }
306 
307 /*
308  * Allocate and reserve nr_pages for the use of a specific client.
309  */
310 static int client_reserve_pages(struct dm_kcopyd_client *kc, unsigned int nr_pages)
311 {
312 	unsigned int i;
313 	struct page_list *pl = NULL, *next;
314 
315 	for (i = 0; i < nr_pages; i++) {
316 		next = alloc_pl(GFP_KERNEL);
317 		if (!next) {
318 			if (pl)
319 				drop_pages(pl);
320 			return -ENOMEM;
321 		}
322 		next->next = pl;
323 		pl = next;
324 	}
325 
326 	kc->nr_reserved_pages += nr_pages;
327 	kcopyd_put_pages(kc, pl);
328 
329 	return 0;
330 }
331 
332 static void client_free_pages(struct dm_kcopyd_client *kc)
333 {
334 	BUG_ON(kc->nr_free_pages != kc->nr_reserved_pages);
335 	drop_pages(kc->pages);
336 	kc->pages = NULL;
337 	kc->nr_free_pages = kc->nr_reserved_pages = 0;
338 }
339 
340 /*
341  *---------------------------------------------------------------
342  * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
343  * for this reason we use a mempool to prevent the client from
344  * ever having to do io (which could cause a deadlock).
345  *---------------------------------------------------------------
346  */
347 struct kcopyd_job {
348 	struct dm_kcopyd_client *kc;
349 	struct list_head list;
350 	unsigned int flags;
351 
352 	/*
353 	 * Error state of the job.
354 	 */
355 	int read_err;
356 	unsigned long write_err;
357 
358 	/*
359 	 * REQ_OP_READ, REQ_OP_WRITE or REQ_OP_WRITE_ZEROES.
360 	 */
361 	enum req_op op;
362 	struct dm_io_region source;
363 
364 	/*
365 	 * The destinations for the transfer.
366 	 */
367 	unsigned int num_dests;
368 	struct dm_io_region dests[DM_KCOPYD_MAX_REGIONS];
369 
370 	struct page_list *pages;
371 
372 	/*
373 	 * Set this to ensure you are notified when the job has
374 	 * completed.  'context' is for callback to use.
375 	 */
376 	dm_kcopyd_notify_fn fn;
377 	void *context;
378 
379 	/*
380 	 * These fields are only used if the job has been split
381 	 * into more manageable parts.
382 	 */
383 	struct mutex lock;
384 	atomic_t sub_jobs;
385 	sector_t progress;
386 	sector_t write_offset;
387 
388 	struct kcopyd_job *master_job;
389 };
390 
391 static struct kmem_cache *_job_cache;
392 
393 int __init dm_kcopyd_init(void)
394 {
395 	_job_cache = kmem_cache_create("kcopyd_job",
396 				sizeof(struct kcopyd_job) * (SPLIT_COUNT + 1),
397 				__alignof__(struct kcopyd_job), 0, NULL);
398 	if (!_job_cache)
399 		return -ENOMEM;
400 
401 	zero_page_list.next = &zero_page_list;
402 	zero_page_list.page = ZERO_PAGE(0);
403 
404 	return 0;
405 }
406 
407 void dm_kcopyd_exit(void)
408 {
409 	kmem_cache_destroy(_job_cache);
410 	_job_cache = NULL;
411 }
412 
413 /*
414  * Functions to push and pop a job onto the head of a given job
415  * list.
416  */
417 static struct kcopyd_job *pop_io_job(struct list_head *jobs,
418 				     struct dm_kcopyd_client *kc)
419 {
420 	struct kcopyd_job *job;
421 
422 	/*
423 	 * For I/O jobs, pop any read, any write without sequential write
424 	 * constraint and sequential writes that are at the right position.
425 	 */
426 	list_for_each_entry(job, jobs, list) {
427 		if (job->op == REQ_OP_READ ||
428 		    !(job->flags & BIT(DM_KCOPYD_WRITE_SEQ))) {
429 			list_del(&job->list);
430 			return job;
431 		}
432 
433 		if (job->write_offset == job->master_job->write_offset) {
434 			job->master_job->write_offset += job->source.count;
435 			list_del(&job->list);
436 			return job;
437 		}
438 	}
439 
440 	return NULL;
441 }
442 
443 static struct kcopyd_job *pop(struct list_head *jobs,
444 			      struct dm_kcopyd_client *kc)
445 {
446 	struct kcopyd_job *job = NULL;
447 
448 	spin_lock_irq(&kc->job_lock);
449 
450 	if (!list_empty(jobs)) {
451 		if (jobs == &kc->io_jobs)
452 			job = pop_io_job(jobs, kc);
453 		else {
454 			job = list_entry(jobs->next, struct kcopyd_job, list);
455 			list_del(&job->list);
456 		}
457 	}
458 	spin_unlock_irq(&kc->job_lock);
459 
460 	return job;
461 }
462 
463 static void push(struct list_head *jobs, struct kcopyd_job *job)
464 {
465 	unsigned long flags;
466 	struct dm_kcopyd_client *kc = job->kc;
467 
468 	spin_lock_irqsave(&kc->job_lock, flags);
469 	list_add_tail(&job->list, jobs);
470 	spin_unlock_irqrestore(&kc->job_lock, flags);
471 }
472 
473 
474 static void push_head(struct list_head *jobs, struct kcopyd_job *job)
475 {
476 	struct dm_kcopyd_client *kc = job->kc;
477 
478 	spin_lock_irq(&kc->job_lock);
479 	list_add(&job->list, jobs);
480 	spin_unlock_irq(&kc->job_lock);
481 }
482 
483 /*
484  * These three functions process 1 item from the corresponding
485  * job list.
486  *
487  * They return:
488  * < 0: error
489  *   0: success
490  * > 0: can't process yet.
491  */
492 static int run_complete_job(struct kcopyd_job *job)
493 {
494 	void *context = job->context;
495 	int read_err = job->read_err;
496 	unsigned long write_err = job->write_err;
497 	dm_kcopyd_notify_fn fn = job->fn;
498 	struct dm_kcopyd_client *kc = job->kc;
499 
500 	if (job->pages && job->pages != &zero_page_list)
501 		kcopyd_put_pages(kc, job->pages);
502 	/*
503 	 * If this is the master job, the sub jobs have already
504 	 * completed so we can free everything.
505 	 */
506 	if (job->master_job == job) {
507 		mutex_destroy(&job->lock);
508 		mempool_free(job, &kc->job_pool);
509 	}
510 	fn(read_err, write_err, context);
511 
512 	if (atomic_dec_and_test(&kc->nr_jobs))
513 		wake_up(&kc->destroyq);
514 
515 	cond_resched();
516 
517 	return 0;
518 }
519 
520 static void complete_io(unsigned long error, void *context)
521 {
522 	struct kcopyd_job *job = (struct kcopyd_job *) context;
523 	struct dm_kcopyd_client *kc = job->kc;
524 
525 	io_job_finish(kc->throttle);
526 
527 	if (error) {
528 		if (op_is_write(job->op))
529 			job->write_err |= error;
530 		else
531 			job->read_err = 1;
532 
533 		if (!(job->flags & BIT(DM_KCOPYD_IGNORE_ERROR))) {
534 			push(&kc->complete_jobs, job);
535 			wake(kc);
536 			return;
537 		}
538 	}
539 
540 	if (op_is_write(job->op))
541 		push(&kc->complete_jobs, job);
542 
543 	else {
544 		job->op = REQ_OP_WRITE;
545 		push(&kc->io_jobs, job);
546 	}
547 
548 	wake(kc);
549 }
550 
551 /*
552  * Request io on as many buffer heads as we can currently get for
553  * a particular job.
554  */
555 static int run_io_job(struct kcopyd_job *job)
556 {
557 	int r;
558 	struct dm_io_request io_req = {
559 		.bi_opf = job->op,
560 		.mem.type = DM_IO_PAGE_LIST,
561 		.mem.ptr.pl = job->pages,
562 		.mem.offset = 0,
563 		.notify.fn = complete_io,
564 		.notify.context = job,
565 		.client = job->kc->io_client,
566 	};
567 
568 	/*
569 	 * If we need to write sequentially and some reads or writes failed,
570 	 * no point in continuing.
571 	 */
572 	if (job->flags & BIT(DM_KCOPYD_WRITE_SEQ) &&
573 	    job->master_job->write_err) {
574 		job->write_err = job->master_job->write_err;
575 		return -EIO;
576 	}
577 
578 	io_job_start(job->kc->throttle);
579 
580 	if (job->op == REQ_OP_READ)
581 		r = dm_io(&io_req, 1, &job->source, NULL);
582 	else
583 		r = dm_io(&io_req, job->num_dests, job->dests, NULL);
584 
585 	return r;
586 }
587 
588 static int run_pages_job(struct kcopyd_job *job)
589 {
590 	int r;
591 	unsigned int nr_pages = dm_div_up(job->dests[0].count, PAGE_SIZE >> 9);
592 
593 	r = kcopyd_get_pages(job->kc, nr_pages, &job->pages);
594 	if (!r) {
595 		/* this job is ready for io */
596 		push(&job->kc->io_jobs, job);
597 		return 0;
598 	}
599 
600 	if (r == -ENOMEM)
601 		/* can't complete now */
602 		return 1;
603 
604 	return r;
605 }
606 
607 /*
608  * Run through a list for as long as possible.  Returns the count
609  * of successful jobs.
610  */
611 static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
612 			int (*fn)(struct kcopyd_job *))
613 {
614 	struct kcopyd_job *job;
615 	int r, count = 0;
616 
617 	while ((job = pop(jobs, kc))) {
618 
619 		r = fn(job);
620 
621 		if (r < 0) {
622 			/* error this rogue job */
623 			if (op_is_write(job->op))
624 				job->write_err = (unsigned long) -1L;
625 			else
626 				job->read_err = 1;
627 			push(&kc->complete_jobs, job);
628 			wake(kc);
629 			break;
630 		}
631 
632 		if (r > 0) {
633 			/*
634 			 * We couldn't service this job ATM, so
635 			 * push this job back onto the list.
636 			 */
637 			push_head(jobs, job);
638 			break;
639 		}
640 
641 		count++;
642 	}
643 
644 	return count;
645 }
646 
647 /*
648  * kcopyd does this every time it's woken up.
649  */
650 static void do_work(struct work_struct *work)
651 {
652 	struct dm_kcopyd_client *kc = container_of(work,
653 					struct dm_kcopyd_client, kcopyd_work);
654 	struct blk_plug plug;
655 
656 	/*
657 	 * The order that these are called is *very* important.
658 	 * complete jobs can free some pages for pages jobs.
659 	 * Pages jobs when successful will jump onto the io jobs
660 	 * list.  io jobs call wake when they complete and it all
661 	 * starts again.
662 	 */
663 	spin_lock_irq(&kc->job_lock);
664 	list_splice_tail_init(&kc->callback_jobs, &kc->complete_jobs);
665 	spin_unlock_irq(&kc->job_lock);
666 
667 	blk_start_plug(&plug);
668 	process_jobs(&kc->complete_jobs, kc, run_complete_job);
669 	process_jobs(&kc->pages_jobs, kc, run_pages_job);
670 	process_jobs(&kc->io_jobs, kc, run_io_job);
671 	blk_finish_plug(&plug);
672 }
673 
674 /*
675  * If we are copying a small region we just dispatch a single job
676  * to do the copy, otherwise the io has to be split up into many
677  * jobs.
678  */
679 static void dispatch_job(struct kcopyd_job *job)
680 {
681 	struct dm_kcopyd_client *kc = job->kc;
682 
683 	atomic_inc(&kc->nr_jobs);
684 	if (unlikely(!job->source.count))
685 		push(&kc->callback_jobs, job);
686 	else if (job->pages == &zero_page_list)
687 		push(&kc->io_jobs, job);
688 	else
689 		push(&kc->pages_jobs, job);
690 	wake(kc);
691 }
692 
693 static void segment_complete(int read_err, unsigned long write_err,
694 			     void *context)
695 {
696 	/* FIXME: tidy this function */
697 	sector_t progress = 0;
698 	sector_t count = 0;
699 	struct kcopyd_job *sub_job = (struct kcopyd_job *) context;
700 	struct kcopyd_job *job = sub_job->master_job;
701 	struct dm_kcopyd_client *kc = job->kc;
702 
703 	mutex_lock(&job->lock);
704 
705 	/* update the error */
706 	if (read_err)
707 		job->read_err = 1;
708 
709 	if (write_err)
710 		job->write_err |= write_err;
711 
712 	/*
713 	 * Only dispatch more work if there hasn't been an error.
714 	 */
715 	if ((!job->read_err && !job->write_err) ||
716 	    job->flags & BIT(DM_KCOPYD_IGNORE_ERROR)) {
717 		/* get the next chunk of work */
718 		progress = job->progress;
719 		count = job->source.count - progress;
720 		if (count) {
721 			if (count > kc->sub_job_size)
722 				count = kc->sub_job_size;
723 
724 			job->progress += count;
725 		}
726 	}
727 	mutex_unlock(&job->lock);
728 
729 	if (count) {
730 		int i;
731 
732 		*sub_job = *job;
733 		sub_job->write_offset = progress;
734 		sub_job->source.sector += progress;
735 		sub_job->source.count = count;
736 
737 		for (i = 0; i < job->num_dests; i++) {
738 			sub_job->dests[i].sector += progress;
739 			sub_job->dests[i].count = count;
740 		}
741 
742 		sub_job->fn = segment_complete;
743 		sub_job->context = sub_job;
744 		dispatch_job(sub_job);
745 
746 	} else if (atomic_dec_and_test(&job->sub_jobs)) {
747 
748 		/*
749 		 * Queue the completion callback to the kcopyd thread.
750 		 *
751 		 * Some callers assume that all the completions are called
752 		 * from a single thread and don't race with each other.
753 		 *
754 		 * We must not call the callback directly here because this
755 		 * code may not be executing in the thread.
756 		 */
757 		push(&kc->complete_jobs, job);
758 		wake(kc);
759 	}
760 }
761 
762 /*
763  * Create some sub jobs to share the work between them.
764  */
765 static void split_job(struct kcopyd_job *master_job)
766 {
767 	int i;
768 
769 	atomic_inc(&master_job->kc->nr_jobs);
770 
771 	atomic_set(&master_job->sub_jobs, SPLIT_COUNT);
772 	for (i = 0; i < SPLIT_COUNT; i++) {
773 		master_job[i + 1].master_job = master_job;
774 		segment_complete(0, 0u, &master_job[i + 1]);
775 	}
776 }
777 
778 void dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
779 		    unsigned int num_dests, struct dm_io_region *dests,
780 		    unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
781 {
782 	struct kcopyd_job *job;
783 	int i;
784 
785 	/*
786 	 * Allocate an array of jobs consisting of one master job
787 	 * followed by SPLIT_COUNT sub jobs.
788 	 */
789 	job = mempool_alloc(&kc->job_pool, GFP_NOIO);
790 	mutex_init(&job->lock);
791 
792 	/*
793 	 * set up for the read.
794 	 */
795 	job->kc = kc;
796 	job->flags = flags;
797 	job->read_err = 0;
798 	job->write_err = 0;
799 
800 	job->num_dests = num_dests;
801 	memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
802 
803 	/*
804 	 * If one of the destination is a host-managed zoned block device,
805 	 * we need to write sequentially. If one of the destination is a
806 	 * host-aware device, then leave it to the caller to choose what to do.
807 	 */
808 	if (!(job->flags & BIT(DM_KCOPYD_WRITE_SEQ))) {
809 		for (i = 0; i < job->num_dests; i++) {
810 			if (bdev_zoned_model(dests[i].bdev) == BLK_ZONED_HM) {
811 				job->flags |= BIT(DM_KCOPYD_WRITE_SEQ);
812 				break;
813 			}
814 		}
815 	}
816 
817 	/*
818 	 * If we need to write sequentially, errors cannot be ignored.
819 	 */
820 	if (job->flags & BIT(DM_KCOPYD_WRITE_SEQ) &&
821 	    job->flags & BIT(DM_KCOPYD_IGNORE_ERROR))
822 		job->flags &= ~BIT(DM_KCOPYD_IGNORE_ERROR);
823 
824 	if (from) {
825 		job->source = *from;
826 		job->pages = NULL;
827 		job->op = REQ_OP_READ;
828 	} else {
829 		memset(&job->source, 0, sizeof(job->source));
830 		job->source.count = job->dests[0].count;
831 		job->pages = &zero_page_list;
832 
833 		/*
834 		 * Use WRITE ZEROES to optimize zeroing if all dests support it.
835 		 */
836 		job->op = REQ_OP_WRITE_ZEROES;
837 		for (i = 0; i < job->num_dests; i++)
838 			if (!bdev_write_zeroes_sectors(job->dests[i].bdev)) {
839 				job->op = REQ_OP_WRITE;
840 				break;
841 			}
842 	}
843 
844 	job->fn = fn;
845 	job->context = context;
846 	job->master_job = job;
847 	job->write_offset = 0;
848 
849 	if (job->source.count <= kc->sub_job_size)
850 		dispatch_job(job);
851 	else {
852 		job->progress = 0;
853 		split_job(job);
854 	}
855 }
856 EXPORT_SYMBOL(dm_kcopyd_copy);
857 
858 void dm_kcopyd_zero(struct dm_kcopyd_client *kc,
859 		    unsigned int num_dests, struct dm_io_region *dests,
860 		    unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
861 {
862 	dm_kcopyd_copy(kc, NULL, num_dests, dests, flags, fn, context);
863 }
864 EXPORT_SYMBOL(dm_kcopyd_zero);
865 
866 void *dm_kcopyd_prepare_callback(struct dm_kcopyd_client *kc,
867 				 dm_kcopyd_notify_fn fn, void *context)
868 {
869 	struct kcopyd_job *job;
870 
871 	job = mempool_alloc(&kc->job_pool, GFP_NOIO);
872 
873 	memset(job, 0, sizeof(struct kcopyd_job));
874 	job->kc = kc;
875 	job->fn = fn;
876 	job->context = context;
877 	job->master_job = job;
878 
879 	atomic_inc(&kc->nr_jobs);
880 
881 	return job;
882 }
883 EXPORT_SYMBOL(dm_kcopyd_prepare_callback);
884 
885 void dm_kcopyd_do_callback(void *j, int read_err, unsigned long write_err)
886 {
887 	struct kcopyd_job *job = j;
888 	struct dm_kcopyd_client *kc = job->kc;
889 
890 	job->read_err = read_err;
891 	job->write_err = write_err;
892 
893 	push(&kc->callback_jobs, job);
894 	wake(kc);
895 }
896 EXPORT_SYMBOL(dm_kcopyd_do_callback);
897 
898 /*
899  * Cancels a kcopyd job, eg. someone might be deactivating a
900  * mirror.
901  */
902 #if 0
903 int kcopyd_cancel(struct kcopyd_job *job, int block)
904 {
905 	/* FIXME: finish */
906 	return -1;
907 }
908 #endif  /*  0  */
909 
910 /*
911  *---------------------------------------------------------------
912  * Client setup
913  *---------------------------------------------------------------
914  */
915 struct dm_kcopyd_client *dm_kcopyd_client_create(struct dm_kcopyd_throttle *throttle)
916 {
917 	int r;
918 	unsigned int reserve_pages;
919 	struct dm_kcopyd_client *kc;
920 
921 	kc = kzalloc(sizeof(*kc), GFP_KERNEL);
922 	if (!kc)
923 		return ERR_PTR(-ENOMEM);
924 
925 	spin_lock_init(&kc->job_lock);
926 	INIT_LIST_HEAD(&kc->callback_jobs);
927 	INIT_LIST_HEAD(&kc->complete_jobs);
928 	INIT_LIST_HEAD(&kc->io_jobs);
929 	INIT_LIST_HEAD(&kc->pages_jobs);
930 	kc->throttle = throttle;
931 
932 	r = mempool_init_slab_pool(&kc->job_pool, MIN_JOBS, _job_cache);
933 	if (r)
934 		goto bad_slab;
935 
936 	INIT_WORK(&kc->kcopyd_work, do_work);
937 	kc->kcopyd_wq = alloc_workqueue("kcopyd", WQ_MEM_RECLAIM, 0);
938 	if (!kc->kcopyd_wq) {
939 		r = -ENOMEM;
940 		goto bad_workqueue;
941 	}
942 
943 	kc->sub_job_size = dm_get_kcopyd_subjob_size();
944 	reserve_pages = DIV_ROUND_UP(kc->sub_job_size << SECTOR_SHIFT, PAGE_SIZE);
945 
946 	kc->pages = NULL;
947 	kc->nr_reserved_pages = kc->nr_free_pages = 0;
948 	r = client_reserve_pages(kc, reserve_pages);
949 	if (r)
950 		goto bad_client_pages;
951 
952 	kc->io_client = dm_io_client_create();
953 	if (IS_ERR(kc->io_client)) {
954 		r = PTR_ERR(kc->io_client);
955 		goto bad_io_client;
956 	}
957 
958 	init_waitqueue_head(&kc->destroyq);
959 	atomic_set(&kc->nr_jobs, 0);
960 
961 	return kc;
962 
963 bad_io_client:
964 	client_free_pages(kc);
965 bad_client_pages:
966 	destroy_workqueue(kc->kcopyd_wq);
967 bad_workqueue:
968 	mempool_exit(&kc->job_pool);
969 bad_slab:
970 	kfree(kc);
971 
972 	return ERR_PTR(r);
973 }
974 EXPORT_SYMBOL(dm_kcopyd_client_create);
975 
976 void dm_kcopyd_client_destroy(struct dm_kcopyd_client *kc)
977 {
978 	/* Wait for completion of all jobs submitted by this client. */
979 	wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
980 
981 	BUG_ON(!list_empty(&kc->callback_jobs));
982 	BUG_ON(!list_empty(&kc->complete_jobs));
983 	BUG_ON(!list_empty(&kc->io_jobs));
984 	BUG_ON(!list_empty(&kc->pages_jobs));
985 	destroy_workqueue(kc->kcopyd_wq);
986 	dm_io_client_destroy(kc->io_client);
987 	client_free_pages(kc);
988 	mempool_exit(&kc->job_pool);
989 	kfree(kc);
990 }
991 EXPORT_SYMBOL(dm_kcopyd_client_destroy);
992 
993 void dm_kcopyd_client_flush(struct dm_kcopyd_client *kc)
994 {
995 	flush_workqueue(kc->kcopyd_wq);
996 }
997 EXPORT_SYMBOL(dm_kcopyd_client_flush);
998