xref: /openbmc/linux/drivers/md/dm-kcopyd.c (revision 8dce88fe)
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 SPLIT_COUNT	8
32 #define MIN_JOBS	8
33 
34 #define DEFAULT_SUB_JOB_SIZE_KB 512
35 #define MAX_SUB_JOB_SIZE_KB     1024
36 
37 static unsigned kcopyd_subjob_size_kb = DEFAULT_SUB_JOB_SIZE_KB;
38 
39 module_param(kcopyd_subjob_size_kb, uint, S_IRUGO | S_IWUSR);
40 MODULE_PARM_DESC(kcopyd_subjob_size_kb, "Sub-job size for dm-kcopyd clients");
41 
42 static unsigned dm_get_kcopyd_subjob_size(void)
43 {
44 	unsigned sub_job_size_kb;
45 
46 	sub_job_size_kb = __dm_get_module_param(&kcopyd_subjob_size_kb,
47 						DEFAULT_SUB_JOB_SIZE_KB,
48 						MAX_SUB_JOB_SIZE_KB);
49 
50 	return sub_job_size_kb << 1;
51 }
52 
53 /*-----------------------------------------------------------------
54  * Each kcopyd client has its own little pool of preallocated
55  * pages for kcopyd io.
56  *---------------------------------------------------------------*/
57 struct dm_kcopyd_client {
58 	struct page_list *pages;
59 	unsigned nr_reserved_pages;
60 	unsigned nr_free_pages;
61 	unsigned sub_job_size;
62 
63 	struct dm_io_client *io_client;
64 
65 	wait_queue_head_t destroyq;
66 
67 	mempool_t job_pool;
68 
69 	struct workqueue_struct *kcopyd_wq;
70 	struct work_struct kcopyd_work;
71 
72 	struct dm_kcopyd_throttle *throttle;
73 
74 	atomic_t nr_jobs;
75 
76 /*
77  * We maintain four lists of jobs:
78  *
79  * i)   jobs waiting for pages
80  * ii)  jobs that have pages, and are waiting for the io to be issued.
81  * iii) jobs that don't need to do any IO and just run a callback
82  * iv) jobs that have completed.
83  *
84  * All four of these are protected by job_lock.
85  */
86 	spinlock_t job_lock;
87 	struct list_head callback_jobs;
88 	struct list_head complete_jobs;
89 	struct list_head io_jobs;
90 	struct list_head pages_jobs;
91 };
92 
93 static struct page_list zero_page_list;
94 
95 static DEFINE_SPINLOCK(throttle_spinlock);
96 
97 /*
98  * IO/IDLE accounting slowly decays after (1 << ACCOUNT_INTERVAL_SHIFT) period.
99  * When total_period >= (1 << ACCOUNT_INTERVAL_SHIFT) the counters are divided
100  * by 2.
101  */
102 #define ACCOUNT_INTERVAL_SHIFT		SHIFT_HZ
103 
104 /*
105  * Sleep this number of milliseconds.
106  *
107  * The value was decided experimentally.
108  * Smaller values seem to cause an increased copy rate above the limit.
109  * The reason for this is unknown but possibly due to jiffies rounding errors
110  * or read/write cache inside the disk.
111  */
112 #define SLEEP_MSEC			100
113 
114 /*
115  * Maximum number of sleep events. There is a theoretical livelock if more
116  * kcopyd clients do work simultaneously which this limit avoids.
117  */
118 #define MAX_SLEEPS			10
119 
120 static void io_job_start(struct dm_kcopyd_throttle *t)
121 {
122 	unsigned throttle, now, difference;
123 	int slept = 0, skew;
124 
125 	if (unlikely(!t))
126 		return;
127 
128 try_again:
129 	spin_lock_irq(&throttle_spinlock);
130 
131 	throttle = READ_ONCE(t->throttle);
132 
133 	if (likely(throttle >= 100))
134 		goto skip_limit;
135 
136 	now = jiffies;
137 	difference = now - t->last_jiffies;
138 	t->last_jiffies = now;
139 	if (t->num_io_jobs)
140 		t->io_period += difference;
141 	t->total_period += difference;
142 
143 	/*
144 	 * Maintain sane values if we got a temporary overflow.
145 	 */
146 	if (unlikely(t->io_period > t->total_period))
147 		t->io_period = t->total_period;
148 
149 	if (unlikely(t->total_period >= (1 << ACCOUNT_INTERVAL_SHIFT))) {
150 		int shift = fls(t->total_period >> ACCOUNT_INTERVAL_SHIFT);
151 		t->total_period >>= shift;
152 		t->io_period >>= shift;
153 	}
154 
155 	skew = t->io_period - throttle * t->total_period / 100;
156 
157 	if (unlikely(skew > 0) && slept < MAX_SLEEPS) {
158 		slept++;
159 		spin_unlock_irq(&throttle_spinlock);
160 		msleep(SLEEP_MSEC);
161 		goto try_again;
162 	}
163 
164 skip_limit:
165 	t->num_io_jobs++;
166 
167 	spin_unlock_irq(&throttle_spinlock);
168 }
169 
170 static void io_job_finish(struct dm_kcopyd_throttle *t)
171 {
172 	unsigned long flags;
173 
174 	if (unlikely(!t))
175 		return;
176 
177 	spin_lock_irqsave(&throttle_spinlock, flags);
178 
179 	t->num_io_jobs--;
180 
181 	if (likely(READ_ONCE(t->throttle) >= 100))
182 		goto skip_limit;
183 
184 	if (!t->num_io_jobs) {
185 		unsigned now, difference;
186 
187 		now = jiffies;
188 		difference = now - t->last_jiffies;
189 		t->last_jiffies = now;
190 
191 		t->io_period += difference;
192 		t->total_period += difference;
193 
194 		/*
195 		 * Maintain sane values if we got a temporary overflow.
196 		 */
197 		if (unlikely(t->io_period > t->total_period))
198 			t->io_period = t->total_period;
199 	}
200 
201 skip_limit:
202 	spin_unlock_irqrestore(&throttle_spinlock, flags);
203 }
204 
205 
206 static void wake(struct dm_kcopyd_client *kc)
207 {
208 	queue_work(kc->kcopyd_wq, &kc->kcopyd_work);
209 }
210 
211 /*
212  * Obtain one page for the use of kcopyd.
213  */
214 static struct page_list *alloc_pl(gfp_t gfp)
215 {
216 	struct page_list *pl;
217 
218 	pl = kmalloc(sizeof(*pl), gfp);
219 	if (!pl)
220 		return NULL;
221 
222 	pl->page = alloc_page(gfp);
223 	if (!pl->page) {
224 		kfree(pl);
225 		return NULL;
226 	}
227 
228 	return pl;
229 }
230 
231 static void free_pl(struct page_list *pl)
232 {
233 	__free_page(pl->page);
234 	kfree(pl);
235 }
236 
237 /*
238  * Add the provided pages to a client's free page list, releasing
239  * back to the system any beyond the reserved_pages limit.
240  */
241 static void kcopyd_put_pages(struct dm_kcopyd_client *kc, struct page_list *pl)
242 {
243 	struct page_list *next;
244 
245 	do {
246 		next = pl->next;
247 
248 		if (kc->nr_free_pages >= kc->nr_reserved_pages)
249 			free_pl(pl);
250 		else {
251 			pl->next = kc->pages;
252 			kc->pages = pl;
253 			kc->nr_free_pages++;
254 		}
255 
256 		pl = next;
257 	} while (pl);
258 }
259 
260 static int kcopyd_get_pages(struct dm_kcopyd_client *kc,
261 			    unsigned int nr, struct page_list **pages)
262 {
263 	struct page_list *pl;
264 
265 	*pages = NULL;
266 
267 	do {
268 		pl = alloc_pl(__GFP_NOWARN | __GFP_NORETRY | __GFP_KSWAPD_RECLAIM);
269 		if (unlikely(!pl)) {
270 			/* Use reserved pages */
271 			pl = kc->pages;
272 			if (unlikely(!pl))
273 				goto out_of_memory;
274 			kc->pages = pl->next;
275 			kc->nr_free_pages--;
276 		}
277 		pl->next = *pages;
278 		*pages = pl;
279 	} while (--nr);
280 
281 	return 0;
282 
283 out_of_memory:
284 	if (*pages)
285 		kcopyd_put_pages(kc, *pages);
286 	return -ENOMEM;
287 }
288 
289 /*
290  * These three functions resize the page pool.
291  */
292 static void drop_pages(struct page_list *pl)
293 {
294 	struct page_list *next;
295 
296 	while (pl) {
297 		next = pl->next;
298 		free_pl(pl);
299 		pl = next;
300 	}
301 }
302 
303 /*
304  * Allocate and reserve nr_pages for the use of a specific client.
305  */
306 static int client_reserve_pages(struct dm_kcopyd_client *kc, unsigned nr_pages)
307 {
308 	unsigned i;
309 	struct page_list *pl = NULL, *next;
310 
311 	for (i = 0; i < nr_pages; i++) {
312 		next = alloc_pl(GFP_KERNEL);
313 		if (!next) {
314 			if (pl)
315 				drop_pages(pl);
316 			return -ENOMEM;
317 		}
318 		next->next = pl;
319 		pl = next;
320 	}
321 
322 	kc->nr_reserved_pages += nr_pages;
323 	kcopyd_put_pages(kc, pl);
324 
325 	return 0;
326 }
327 
328 static void client_free_pages(struct dm_kcopyd_client *kc)
329 {
330 	BUG_ON(kc->nr_free_pages != kc->nr_reserved_pages);
331 	drop_pages(kc->pages);
332 	kc->pages = NULL;
333 	kc->nr_free_pages = kc->nr_reserved_pages = 0;
334 }
335 
336 /*-----------------------------------------------------------------
337  * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
338  * for this reason we use a mempool to prevent the client from
339  * ever having to do io (which could cause a deadlock).
340  *---------------------------------------------------------------*/
341 struct kcopyd_job {
342 	struct dm_kcopyd_client *kc;
343 	struct list_head list;
344 	unsigned flags;
345 
346 	/*
347 	 * Error state of the job.
348 	 */
349 	int read_err;
350 	unsigned long write_err;
351 
352 	/*
353 	 * Either READ or WRITE
354 	 */
355 	int rw;
356 	struct dm_io_region source;
357 
358 	/*
359 	 * The destinations for the transfer.
360 	 */
361 	unsigned int num_dests;
362 	struct dm_io_region dests[DM_KCOPYD_MAX_REGIONS];
363 
364 	struct page_list *pages;
365 
366 	/*
367 	 * Set this to ensure you are notified when the job has
368 	 * completed.  'context' is for callback to use.
369 	 */
370 	dm_kcopyd_notify_fn fn;
371 	void *context;
372 
373 	/*
374 	 * These fields are only used if the job has been split
375 	 * into more manageable parts.
376 	 */
377 	struct mutex lock;
378 	atomic_t sub_jobs;
379 	sector_t progress;
380 	sector_t write_offset;
381 
382 	struct kcopyd_job *master_job;
383 };
384 
385 static struct kmem_cache *_job_cache;
386 
387 int __init dm_kcopyd_init(void)
388 {
389 	_job_cache = kmem_cache_create("kcopyd_job",
390 				sizeof(struct kcopyd_job) * (SPLIT_COUNT + 1),
391 				__alignof__(struct kcopyd_job), 0, NULL);
392 	if (!_job_cache)
393 		return -ENOMEM;
394 
395 	zero_page_list.next = &zero_page_list;
396 	zero_page_list.page = ZERO_PAGE(0);
397 
398 	return 0;
399 }
400 
401 void dm_kcopyd_exit(void)
402 {
403 	kmem_cache_destroy(_job_cache);
404 	_job_cache = NULL;
405 }
406 
407 /*
408  * Functions to push and pop a job onto the head of a given job
409  * list.
410  */
411 static struct kcopyd_job *pop_io_job(struct list_head *jobs,
412 				     struct dm_kcopyd_client *kc)
413 {
414 	struct kcopyd_job *job;
415 
416 	/*
417 	 * For I/O jobs, pop any read, any write without sequential write
418 	 * constraint and sequential writes that are at the right position.
419 	 */
420 	list_for_each_entry(job, jobs, list) {
421 		if (job->rw == READ || !(job->flags & BIT(DM_KCOPYD_WRITE_SEQ))) {
422 			list_del(&job->list);
423 			return job;
424 		}
425 
426 		if (job->write_offset == job->master_job->write_offset) {
427 			job->master_job->write_offset += job->source.count;
428 			list_del(&job->list);
429 			return job;
430 		}
431 	}
432 
433 	return NULL;
434 }
435 
436 static struct kcopyd_job *pop(struct list_head *jobs,
437 			      struct dm_kcopyd_client *kc)
438 {
439 	struct kcopyd_job *job = NULL;
440 
441 	spin_lock_irq(&kc->job_lock);
442 
443 	if (!list_empty(jobs)) {
444 		if (jobs == &kc->io_jobs)
445 			job = pop_io_job(jobs, kc);
446 		else {
447 			job = list_entry(jobs->next, struct kcopyd_job, list);
448 			list_del(&job->list);
449 		}
450 	}
451 	spin_unlock_irq(&kc->job_lock);
452 
453 	return job;
454 }
455 
456 static void push(struct list_head *jobs, struct kcopyd_job *job)
457 {
458 	unsigned long flags;
459 	struct dm_kcopyd_client *kc = job->kc;
460 
461 	spin_lock_irqsave(&kc->job_lock, flags);
462 	list_add_tail(&job->list, jobs);
463 	spin_unlock_irqrestore(&kc->job_lock, flags);
464 }
465 
466 
467 static void push_head(struct list_head *jobs, struct kcopyd_job *job)
468 {
469 	struct dm_kcopyd_client *kc = job->kc;
470 
471 	spin_lock_irq(&kc->job_lock);
472 	list_add(&job->list, jobs);
473 	spin_unlock_irq(&kc->job_lock);
474 }
475 
476 /*
477  * These three functions process 1 item from the corresponding
478  * job list.
479  *
480  * They return:
481  * < 0: error
482  *   0: success
483  * > 0: can't process yet.
484  */
485 static int run_complete_job(struct kcopyd_job *job)
486 {
487 	void *context = job->context;
488 	int read_err = job->read_err;
489 	unsigned long write_err = job->write_err;
490 	dm_kcopyd_notify_fn fn = job->fn;
491 	struct dm_kcopyd_client *kc = job->kc;
492 
493 	if (job->pages && job->pages != &zero_page_list)
494 		kcopyd_put_pages(kc, job->pages);
495 	/*
496 	 * If this is the master job, the sub jobs have already
497 	 * completed so we can free everything.
498 	 */
499 	if (job->master_job == job) {
500 		mutex_destroy(&job->lock);
501 		mempool_free(job, &kc->job_pool);
502 	}
503 	fn(read_err, write_err, context);
504 
505 	if (atomic_dec_and_test(&kc->nr_jobs))
506 		wake_up(&kc->destroyq);
507 
508 	cond_resched();
509 
510 	return 0;
511 }
512 
513 static void complete_io(unsigned long error, void *context)
514 {
515 	struct kcopyd_job *job = (struct kcopyd_job *) context;
516 	struct dm_kcopyd_client *kc = job->kc;
517 
518 	io_job_finish(kc->throttle);
519 
520 	if (error) {
521 		if (op_is_write(job->rw))
522 			job->write_err |= error;
523 		else
524 			job->read_err = 1;
525 
526 		if (!(job->flags & BIT(DM_KCOPYD_IGNORE_ERROR))) {
527 			push(&kc->complete_jobs, job);
528 			wake(kc);
529 			return;
530 		}
531 	}
532 
533 	if (op_is_write(job->rw))
534 		push(&kc->complete_jobs, job);
535 
536 	else {
537 		job->rw = WRITE;
538 		push(&kc->io_jobs, job);
539 	}
540 
541 	wake(kc);
542 }
543 
544 /*
545  * Request io on as many buffer heads as we can currently get for
546  * a particular job.
547  */
548 static int run_io_job(struct kcopyd_job *job)
549 {
550 	int r;
551 	struct dm_io_request io_req = {
552 		.bi_op = job->rw,
553 		.bi_op_flags = 0,
554 		.mem.type = DM_IO_PAGE_LIST,
555 		.mem.ptr.pl = job->pages,
556 		.mem.offset = 0,
557 		.notify.fn = complete_io,
558 		.notify.context = job,
559 		.client = job->kc->io_client,
560 	};
561 
562 	/*
563 	 * If we need to write sequentially and some reads or writes failed,
564 	 * no point in continuing.
565 	 */
566 	if (job->flags & BIT(DM_KCOPYD_WRITE_SEQ) &&
567 	    job->master_job->write_err) {
568 		job->write_err = job->master_job->write_err;
569 		return -EIO;
570 	}
571 
572 	io_job_start(job->kc->throttle);
573 
574 	if (job->rw == READ)
575 		r = dm_io(&io_req, 1, &job->source, NULL);
576 	else
577 		r = dm_io(&io_req, job->num_dests, job->dests, NULL);
578 
579 	return r;
580 }
581 
582 static int run_pages_job(struct kcopyd_job *job)
583 {
584 	int r;
585 	unsigned nr_pages = dm_div_up(job->dests[0].count, PAGE_SIZE >> 9);
586 
587 	r = kcopyd_get_pages(job->kc, nr_pages, &job->pages);
588 	if (!r) {
589 		/* this job is ready for io */
590 		push(&job->kc->io_jobs, job);
591 		return 0;
592 	}
593 
594 	if (r == -ENOMEM)
595 		/* can't complete now */
596 		return 1;
597 
598 	return r;
599 }
600 
601 /*
602  * Run through a list for as long as possible.  Returns the count
603  * of successful jobs.
604  */
605 static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
606 			int (*fn) (struct kcopyd_job *))
607 {
608 	struct kcopyd_job *job;
609 	int r, count = 0;
610 
611 	while ((job = pop(jobs, kc))) {
612 
613 		r = fn(job);
614 
615 		if (r < 0) {
616 			/* error this rogue job */
617 			if (op_is_write(job->rw))
618 				job->write_err = (unsigned long) -1L;
619 			else
620 				job->read_err = 1;
621 			push(&kc->complete_jobs, job);
622 			wake(kc);
623 			break;
624 		}
625 
626 		if (r > 0) {
627 			/*
628 			 * We couldn't service this job ATM, so
629 			 * push this job back onto the list.
630 			 */
631 			push_head(jobs, job);
632 			break;
633 		}
634 
635 		count++;
636 	}
637 
638 	return count;
639 }
640 
641 /*
642  * kcopyd does this every time it's woken up.
643  */
644 static void do_work(struct work_struct *work)
645 {
646 	struct dm_kcopyd_client *kc = container_of(work,
647 					struct dm_kcopyd_client, kcopyd_work);
648 	struct blk_plug plug;
649 
650 	/*
651 	 * The order that these are called is *very* important.
652 	 * complete jobs can free some pages for pages jobs.
653 	 * Pages jobs when successful will jump onto the io jobs
654 	 * list.  io jobs call wake when they complete and it all
655 	 * starts again.
656 	 */
657 	spin_lock_irq(&kc->job_lock);
658 	list_splice_tail_init(&kc->callback_jobs, &kc->complete_jobs);
659 	spin_unlock_irq(&kc->job_lock);
660 
661 	blk_start_plug(&plug);
662 	process_jobs(&kc->complete_jobs, kc, run_complete_job);
663 	process_jobs(&kc->pages_jobs, kc, run_pages_job);
664 	process_jobs(&kc->io_jobs, kc, run_io_job);
665 	blk_finish_plug(&plug);
666 }
667 
668 /*
669  * If we are copying a small region we just dispatch a single job
670  * to do the copy, otherwise the io has to be split up into many
671  * jobs.
672  */
673 static void dispatch_job(struct kcopyd_job *job)
674 {
675 	struct dm_kcopyd_client *kc = job->kc;
676 	atomic_inc(&kc->nr_jobs);
677 	if (unlikely(!job->source.count))
678 		push(&kc->callback_jobs, job);
679 	else if (job->pages == &zero_page_list)
680 		push(&kc->io_jobs, job);
681 	else
682 		push(&kc->pages_jobs, job);
683 	wake(kc);
684 }
685 
686 static void segment_complete(int read_err, unsigned long write_err,
687 			     void *context)
688 {
689 	/* FIXME: tidy this function */
690 	sector_t progress = 0;
691 	sector_t count = 0;
692 	struct kcopyd_job *sub_job = (struct kcopyd_job *) context;
693 	struct kcopyd_job *job = sub_job->master_job;
694 	struct dm_kcopyd_client *kc = job->kc;
695 
696 	mutex_lock(&job->lock);
697 
698 	/* update the error */
699 	if (read_err)
700 		job->read_err = 1;
701 
702 	if (write_err)
703 		job->write_err |= write_err;
704 
705 	/*
706 	 * Only dispatch more work if there hasn't been an error.
707 	 */
708 	if ((!job->read_err && !job->write_err) ||
709 	    job->flags & BIT(DM_KCOPYD_IGNORE_ERROR)) {
710 		/* get the next chunk of work */
711 		progress = job->progress;
712 		count = job->source.count - progress;
713 		if (count) {
714 			if (count > kc->sub_job_size)
715 				count = kc->sub_job_size;
716 
717 			job->progress += count;
718 		}
719 	}
720 	mutex_unlock(&job->lock);
721 
722 	if (count) {
723 		int i;
724 
725 		*sub_job = *job;
726 		sub_job->write_offset = progress;
727 		sub_job->source.sector += progress;
728 		sub_job->source.count = count;
729 
730 		for (i = 0; i < job->num_dests; i++) {
731 			sub_job->dests[i].sector += progress;
732 			sub_job->dests[i].count = count;
733 		}
734 
735 		sub_job->fn = segment_complete;
736 		sub_job->context = sub_job;
737 		dispatch_job(sub_job);
738 
739 	} else if (atomic_dec_and_test(&job->sub_jobs)) {
740 
741 		/*
742 		 * Queue the completion callback to the kcopyd thread.
743 		 *
744 		 * Some callers assume that all the completions are called
745 		 * from a single thread and don't race with each other.
746 		 *
747 		 * We must not call the callback directly here because this
748 		 * code may not be executing in the thread.
749 		 */
750 		push(&kc->complete_jobs, job);
751 		wake(kc);
752 	}
753 }
754 
755 /*
756  * Create some sub jobs to share the work between them.
757  */
758 static void split_job(struct kcopyd_job *master_job)
759 {
760 	int i;
761 
762 	atomic_inc(&master_job->kc->nr_jobs);
763 
764 	atomic_set(&master_job->sub_jobs, SPLIT_COUNT);
765 	for (i = 0; i < SPLIT_COUNT; i++) {
766 		master_job[i + 1].master_job = master_job;
767 		segment_complete(0, 0u, &master_job[i + 1]);
768 	}
769 }
770 
771 void dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
772 		    unsigned int num_dests, struct dm_io_region *dests,
773 		    unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
774 {
775 	struct kcopyd_job *job;
776 	int i;
777 
778 	/*
779 	 * Allocate an array of jobs consisting of one master job
780 	 * followed by SPLIT_COUNT sub jobs.
781 	 */
782 	job = mempool_alloc(&kc->job_pool, GFP_NOIO);
783 	mutex_init(&job->lock);
784 
785 	/*
786 	 * set up for the read.
787 	 */
788 	job->kc = kc;
789 	job->flags = flags;
790 	job->read_err = 0;
791 	job->write_err = 0;
792 
793 	job->num_dests = num_dests;
794 	memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
795 
796 	/*
797 	 * If one of the destination is a host-managed zoned block device,
798 	 * we need to write sequentially. If one of the destination is a
799 	 * host-aware device, then leave it to the caller to choose what to do.
800 	 */
801 	if (!(job->flags & BIT(DM_KCOPYD_WRITE_SEQ))) {
802 		for (i = 0; i < job->num_dests; i++) {
803 			if (bdev_zoned_model(dests[i].bdev) == BLK_ZONED_HM) {
804 				job->flags |= BIT(DM_KCOPYD_WRITE_SEQ);
805 				break;
806 			}
807 		}
808 	}
809 
810 	/*
811 	 * If we need to write sequentially, errors cannot be ignored.
812 	 */
813 	if (job->flags & BIT(DM_KCOPYD_WRITE_SEQ) &&
814 	    job->flags & BIT(DM_KCOPYD_IGNORE_ERROR))
815 		job->flags &= ~BIT(DM_KCOPYD_IGNORE_ERROR);
816 
817 	if (from) {
818 		job->source = *from;
819 		job->pages = NULL;
820 		job->rw = READ;
821 	} else {
822 		memset(&job->source, 0, sizeof job->source);
823 		job->source.count = job->dests[0].count;
824 		job->pages = &zero_page_list;
825 
826 		/*
827 		 * Use WRITE ZEROES to optimize zeroing if all dests support it.
828 		 */
829 		job->rw = REQ_OP_WRITE_ZEROES;
830 		for (i = 0; i < job->num_dests; i++)
831 			if (!bdev_write_zeroes_sectors(job->dests[i].bdev)) {
832 				job->rw = WRITE;
833 				break;
834 			}
835 	}
836 
837 	job->fn = fn;
838 	job->context = context;
839 	job->master_job = job;
840 	job->write_offset = 0;
841 
842 	if (job->source.count <= kc->sub_job_size)
843 		dispatch_job(job);
844 	else {
845 		job->progress = 0;
846 		split_job(job);
847 	}
848 }
849 EXPORT_SYMBOL(dm_kcopyd_copy);
850 
851 void dm_kcopyd_zero(struct dm_kcopyd_client *kc,
852 		    unsigned num_dests, struct dm_io_region *dests,
853 		    unsigned flags, dm_kcopyd_notify_fn fn, void *context)
854 {
855 	dm_kcopyd_copy(kc, NULL, num_dests, dests, flags, fn, context);
856 }
857 EXPORT_SYMBOL(dm_kcopyd_zero);
858 
859 void *dm_kcopyd_prepare_callback(struct dm_kcopyd_client *kc,
860 				 dm_kcopyd_notify_fn fn, void *context)
861 {
862 	struct kcopyd_job *job;
863 
864 	job = mempool_alloc(&kc->job_pool, GFP_NOIO);
865 
866 	memset(job, 0, sizeof(struct kcopyd_job));
867 	job->kc = kc;
868 	job->fn = fn;
869 	job->context = context;
870 	job->master_job = job;
871 
872 	atomic_inc(&kc->nr_jobs);
873 
874 	return job;
875 }
876 EXPORT_SYMBOL(dm_kcopyd_prepare_callback);
877 
878 void dm_kcopyd_do_callback(void *j, int read_err, unsigned long write_err)
879 {
880 	struct kcopyd_job *job = j;
881 	struct dm_kcopyd_client *kc = job->kc;
882 
883 	job->read_err = read_err;
884 	job->write_err = write_err;
885 
886 	push(&kc->callback_jobs, job);
887 	wake(kc);
888 }
889 EXPORT_SYMBOL(dm_kcopyd_do_callback);
890 
891 /*
892  * Cancels a kcopyd job, eg. someone might be deactivating a
893  * mirror.
894  */
895 #if 0
896 int kcopyd_cancel(struct kcopyd_job *job, int block)
897 {
898 	/* FIXME: finish */
899 	return -1;
900 }
901 #endif  /*  0  */
902 
903 /*-----------------------------------------------------------------
904  * Client setup
905  *---------------------------------------------------------------*/
906 struct dm_kcopyd_client *dm_kcopyd_client_create(struct dm_kcopyd_throttle *throttle)
907 {
908 	int r;
909 	unsigned reserve_pages;
910 	struct dm_kcopyd_client *kc;
911 
912 	kc = kzalloc(sizeof(*kc), GFP_KERNEL);
913 	if (!kc)
914 		return ERR_PTR(-ENOMEM);
915 
916 	spin_lock_init(&kc->job_lock);
917 	INIT_LIST_HEAD(&kc->callback_jobs);
918 	INIT_LIST_HEAD(&kc->complete_jobs);
919 	INIT_LIST_HEAD(&kc->io_jobs);
920 	INIT_LIST_HEAD(&kc->pages_jobs);
921 	kc->throttle = throttle;
922 
923 	r = mempool_init_slab_pool(&kc->job_pool, MIN_JOBS, _job_cache);
924 	if (r)
925 		goto bad_slab;
926 
927 	INIT_WORK(&kc->kcopyd_work, do_work);
928 	kc->kcopyd_wq = alloc_workqueue("kcopyd", WQ_MEM_RECLAIM, 0);
929 	if (!kc->kcopyd_wq) {
930 		r = -ENOMEM;
931 		goto bad_workqueue;
932 	}
933 
934 	kc->sub_job_size = dm_get_kcopyd_subjob_size();
935 	reserve_pages = DIV_ROUND_UP(kc->sub_job_size << SECTOR_SHIFT, PAGE_SIZE);
936 
937 	kc->pages = NULL;
938 	kc->nr_reserved_pages = kc->nr_free_pages = 0;
939 	r = client_reserve_pages(kc, reserve_pages);
940 	if (r)
941 		goto bad_client_pages;
942 
943 	kc->io_client = dm_io_client_create();
944 	if (IS_ERR(kc->io_client)) {
945 		r = PTR_ERR(kc->io_client);
946 		goto bad_io_client;
947 	}
948 
949 	init_waitqueue_head(&kc->destroyq);
950 	atomic_set(&kc->nr_jobs, 0);
951 
952 	return kc;
953 
954 bad_io_client:
955 	client_free_pages(kc);
956 bad_client_pages:
957 	destroy_workqueue(kc->kcopyd_wq);
958 bad_workqueue:
959 	mempool_exit(&kc->job_pool);
960 bad_slab:
961 	kfree(kc);
962 
963 	return ERR_PTR(r);
964 }
965 EXPORT_SYMBOL(dm_kcopyd_client_create);
966 
967 void dm_kcopyd_client_destroy(struct dm_kcopyd_client *kc)
968 {
969 	/* Wait for completion of all jobs submitted by this client. */
970 	wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
971 
972 	BUG_ON(!list_empty(&kc->callback_jobs));
973 	BUG_ON(!list_empty(&kc->complete_jobs));
974 	BUG_ON(!list_empty(&kc->io_jobs));
975 	BUG_ON(!list_empty(&kc->pages_jobs));
976 	destroy_workqueue(kc->kcopyd_wq);
977 	dm_io_client_destroy(kc->io_client);
978 	client_free_pages(kc);
979 	mempool_exit(&kc->job_pool);
980 	kfree(kc);
981 }
982 EXPORT_SYMBOL(dm_kcopyd_client_destroy);
983 
984 void dm_kcopyd_client_flush(struct dm_kcopyd_client *kc)
985 {
986 	flush_workqueue(kc->kcopyd_wq);
987 }
988 EXPORT_SYMBOL(dm_kcopyd_client_flush);
989