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/delay.h>
26 #include <linux/device-mapper.h>
27 #include <linux/dm-kcopyd.h>
31 #define SUB_JOB_SIZE 128
34 #define RESERVE_PAGES (DIV_ROUND_UP(SUB_JOB_SIZE << SECTOR_SHIFT, PAGE_SIZE))
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;
45 struct dm_io_client *io_client;
47 wait_queue_head_t destroyq;
52 struct workqueue_struct *kcopyd_wq;
53 struct work_struct kcopyd_work;
55 struct dm_kcopyd_throttle *throttle;
58 * We maintain four lists of jobs:
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 don't need to do any IO and just run a callback
63 * iv) jobs that have completed.
65 * All four of these are protected by job_lock.
68 struct list_head callback_jobs;
69 struct list_head complete_jobs;
70 struct list_head io_jobs;
71 struct list_head pages_jobs;
74 static struct page_list zero_page_list;
76 static DEFINE_SPINLOCK(throttle_spinlock);
79 * IO/IDLE accounting slowly decays after (1 << ACCOUNT_INTERVAL_SHIFT) period.
80 * When total_period >= (1 << ACCOUNT_INTERVAL_SHIFT) the counters are divided
83 #define ACCOUNT_INTERVAL_SHIFT SHIFT_HZ
86 * Sleep this number of milliseconds.
88 * The value was decided experimentally.
89 * Smaller values seem to cause an increased copy rate above the limit.
90 * The reason for this is unknown but possibly due to jiffies rounding errors
91 * or read/write cache inside the disk.
93 #define SLEEP_MSEC 100
96 * Maximum number of sleep events. There is a theoretical livelock if more
97 * kcopyd clients do work simultaneously which this limit avoids.
101 static void io_job_start(struct dm_kcopyd_throttle *t)
103 unsigned throttle, now, difference;
110 spin_lock_irq(&throttle_spinlock);
112 throttle = ACCESS_ONCE(t->throttle);
114 if (likely(throttle >= 100))
118 difference = now - t->last_jiffies;
119 t->last_jiffies = now;
121 t->io_period += difference;
122 t->total_period += difference;
125 * Maintain sane values if we got a temporary overflow.
127 if (unlikely(t->io_period > t->total_period))
128 t->io_period = t->total_period;
130 if (unlikely(t->total_period >= (1 << ACCOUNT_INTERVAL_SHIFT))) {
131 int shift = fls(t->total_period >> ACCOUNT_INTERVAL_SHIFT);
132 t->total_period >>= shift;
133 t->io_period >>= shift;
136 skew = t->io_period - throttle * t->total_period / 100;
138 if (unlikely(skew > 0) && slept < MAX_SLEEPS) {
140 spin_unlock_irq(&throttle_spinlock);
148 spin_unlock_irq(&throttle_spinlock);
151 static void io_job_finish(struct dm_kcopyd_throttle *t)
158 spin_lock_irqsave(&throttle_spinlock, flags);
162 if (likely(ACCESS_ONCE(t->throttle) >= 100))
165 if (!t->num_io_jobs) {
166 unsigned now, difference;
169 difference = now - t->last_jiffies;
170 t->last_jiffies = now;
172 t->io_period += difference;
173 t->total_period += difference;
176 * Maintain sane values if we got a temporary overflow.
178 if (unlikely(t->io_period > t->total_period))
179 t->io_period = t->total_period;
183 spin_unlock_irqrestore(&throttle_spinlock, flags);
187 static void wake(struct dm_kcopyd_client *kc)
189 queue_work(kc->kcopyd_wq, &kc->kcopyd_work);
193 * Obtain one page for the use of kcopyd.
195 static struct page_list *alloc_pl(gfp_t gfp)
197 struct page_list *pl;
199 pl = kmalloc(sizeof(*pl), gfp);
203 pl->page = alloc_page(gfp);
212 static void free_pl(struct page_list *pl)
214 __free_page(pl->page);
219 * Add the provided pages to a client's free page list, releasing
220 * back to the system any beyond the reserved_pages limit.
222 static void kcopyd_put_pages(struct dm_kcopyd_client *kc, struct page_list *pl)
224 struct page_list *next;
229 if (kc->nr_free_pages >= kc->nr_reserved_pages)
232 pl->next = kc->pages;
241 static int kcopyd_get_pages(struct dm_kcopyd_client *kc,
242 unsigned int nr, struct page_list **pages)
244 struct page_list *pl;
249 pl = alloc_pl(__GFP_NOWARN | __GFP_NORETRY | __GFP_KSWAPD_RECLAIM);
251 /* Use reserved pages */
255 kc->pages = pl->next;
266 kcopyd_put_pages(kc, *pages);
271 * These three functions resize the page pool.
273 static void drop_pages(struct page_list *pl)
275 struct page_list *next;
285 * Allocate and reserve nr_pages for the use of a specific client.
287 static int client_reserve_pages(struct dm_kcopyd_client *kc, unsigned nr_pages)
290 struct page_list *pl = NULL, *next;
292 for (i = 0; i < nr_pages; i++) {
293 next = alloc_pl(GFP_KERNEL);
303 kc->nr_reserved_pages += nr_pages;
304 kcopyd_put_pages(kc, pl);
309 static void client_free_pages(struct dm_kcopyd_client *kc)
311 BUG_ON(kc->nr_free_pages != kc->nr_reserved_pages);
312 drop_pages(kc->pages);
314 kc->nr_free_pages = kc->nr_reserved_pages = 0;
317 /*-----------------------------------------------------------------
318 * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
319 * for this reason we use a mempool to prevent the client from
320 * ever having to do io (which could cause a deadlock).
321 *---------------------------------------------------------------*/
323 struct dm_kcopyd_client *kc;
324 struct list_head list;
328 * Error state of the job.
331 unsigned long write_err;
334 * Either READ or WRITE
337 struct dm_io_region source;
340 * The destinations for the transfer.
342 unsigned int num_dests;
343 struct dm_io_region dests[DM_KCOPYD_MAX_REGIONS];
345 struct page_list *pages;
348 * Set this to ensure you are notified when the job has
349 * completed. 'context' is for callback to use.
351 dm_kcopyd_notify_fn fn;
355 * These fields are only used if the job has been split
356 * into more manageable parts.
361 sector_t write_offset;
363 struct kcopyd_job *master_job;
366 static struct kmem_cache *_job_cache;
368 int __init dm_kcopyd_init(void)
370 _job_cache = kmem_cache_create("kcopyd_job",
371 sizeof(struct kcopyd_job) * (SPLIT_COUNT + 1),
372 __alignof__(struct kcopyd_job), 0, NULL);
376 zero_page_list.next = &zero_page_list;
377 zero_page_list.page = ZERO_PAGE(0);
382 void dm_kcopyd_exit(void)
384 kmem_cache_destroy(_job_cache);
389 * Functions to push and pop a job onto the head of a given job
392 static struct kcopyd_job *pop_io_job(struct list_head *jobs,
393 struct dm_kcopyd_client *kc)
395 struct kcopyd_job *job;
398 * For I/O jobs, pop any read, any write without sequential write
399 * constraint and sequential writes that are at the right position.
401 list_for_each_entry(job, jobs, list) {
402 if (job->rw == READ || !test_bit(DM_KCOPYD_WRITE_SEQ, &job->flags)) {
403 list_del(&job->list);
407 if (job->write_offset == job->master_job->write_offset) {
408 job->master_job->write_offset += job->source.count;
409 list_del(&job->list);
417 static struct kcopyd_job *pop(struct list_head *jobs,
418 struct dm_kcopyd_client *kc)
420 struct kcopyd_job *job = NULL;
423 spin_lock_irqsave(&kc->job_lock, flags);
425 if (!list_empty(jobs)) {
426 if (jobs == &kc->io_jobs)
427 job = pop_io_job(jobs, kc);
429 job = list_entry(jobs->next, struct kcopyd_job, list);
430 list_del(&job->list);
433 spin_unlock_irqrestore(&kc->job_lock, flags);
438 static void push(struct list_head *jobs, struct kcopyd_job *job)
441 struct dm_kcopyd_client *kc = job->kc;
443 spin_lock_irqsave(&kc->job_lock, flags);
444 list_add_tail(&job->list, jobs);
445 spin_unlock_irqrestore(&kc->job_lock, flags);
449 static void push_head(struct list_head *jobs, struct kcopyd_job *job)
452 struct dm_kcopyd_client *kc = job->kc;
454 spin_lock_irqsave(&kc->job_lock, flags);
455 list_add(&job->list, jobs);
456 spin_unlock_irqrestore(&kc->job_lock, flags);
460 * These three functions process 1 item from the corresponding
466 * > 0: can't process yet.
468 static int run_complete_job(struct kcopyd_job *job)
470 void *context = job->context;
471 int read_err = job->read_err;
472 unsigned long write_err = job->write_err;
473 dm_kcopyd_notify_fn fn = job->fn;
474 struct dm_kcopyd_client *kc = job->kc;
476 if (job->pages && job->pages != &zero_page_list)
477 kcopyd_put_pages(kc, job->pages);
479 * If this is the master job, the sub jobs have already
480 * completed so we can free everything.
482 if (job->master_job == job)
483 mempool_free(job, kc->job_pool);
484 fn(read_err, write_err, context);
486 if (atomic_dec_and_test(&kc->nr_jobs))
487 wake_up(&kc->destroyq);
494 static void complete_io(unsigned long error, void *context)
496 struct kcopyd_job *job = (struct kcopyd_job *) context;
497 struct dm_kcopyd_client *kc = job->kc;
499 io_job_finish(kc->throttle);
502 if (op_is_write(job->rw))
503 job->write_err |= error;
507 if (!test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
508 push(&kc->complete_jobs, job);
514 if (op_is_write(job->rw))
515 push(&kc->complete_jobs, job);
519 push(&kc->io_jobs, job);
526 * Request io on as many buffer heads as we can currently get for
529 static int run_io_job(struct kcopyd_job *job)
532 struct dm_io_request io_req = {
535 .mem.type = DM_IO_PAGE_LIST,
536 .mem.ptr.pl = job->pages,
538 .notify.fn = complete_io,
539 .notify.context = job,
540 .client = job->kc->io_client,
544 * If we need to write sequentially and some reads or writes failed,
545 * no point in continuing.
547 if (test_bit(DM_KCOPYD_WRITE_SEQ, &job->flags) &&
548 job->master_job->write_err) {
549 job->write_err = job->master_job->write_err;
553 io_job_start(job->kc->throttle);
556 r = dm_io(&io_req, 1, &job->source, NULL);
558 r = dm_io(&io_req, job->num_dests, job->dests, NULL);
563 static int run_pages_job(struct kcopyd_job *job)
566 unsigned nr_pages = dm_div_up(job->dests[0].count, PAGE_SIZE >> 9);
568 r = kcopyd_get_pages(job->kc, nr_pages, &job->pages);
570 /* this job is ready for io */
571 push(&job->kc->io_jobs, job);
576 /* can't complete now */
583 * Run through a list for as long as possible. Returns the count
584 * of successful jobs.
586 static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
587 int (*fn) (struct kcopyd_job *))
589 struct kcopyd_job *job;
592 while ((job = pop(jobs, kc))) {
597 /* error this rogue job */
598 if (op_is_write(job->rw))
599 job->write_err = (unsigned long) -1L;
602 push(&kc->complete_jobs, job);
609 * We couldn't service this job ATM, so
610 * push this job back onto the list.
612 push_head(jobs, job);
623 * kcopyd does this every time it's woken up.
625 static void do_work(struct work_struct *work)
627 struct dm_kcopyd_client *kc = container_of(work,
628 struct dm_kcopyd_client, kcopyd_work);
629 struct blk_plug plug;
633 * The order that these are called is *very* important.
634 * complete jobs can free some pages for pages jobs.
635 * Pages jobs when successful will jump onto the io jobs
636 * list. io jobs call wake when they complete and it all
639 spin_lock_irqsave(&kc->job_lock, flags);
640 list_splice_tail_init(&kc->callback_jobs, &kc->complete_jobs);
641 spin_unlock_irqrestore(&kc->job_lock, flags);
643 blk_start_plug(&plug);
644 process_jobs(&kc->complete_jobs, kc, run_complete_job);
645 process_jobs(&kc->pages_jobs, kc, run_pages_job);
646 process_jobs(&kc->io_jobs, kc, run_io_job);
647 blk_finish_plug(&plug);
651 * If we are copying a small region we just dispatch a single job
652 * to do the copy, otherwise the io has to be split up into many
655 static void dispatch_job(struct kcopyd_job *job)
657 struct dm_kcopyd_client *kc = job->kc;
658 atomic_inc(&kc->nr_jobs);
659 if (unlikely(!job->source.count))
660 push(&kc->callback_jobs, job);
661 else if (job->pages == &zero_page_list)
662 push(&kc->io_jobs, job);
664 push(&kc->pages_jobs, job);
668 static void segment_complete(int read_err, unsigned long write_err,
671 /* FIXME: tidy this function */
672 sector_t progress = 0;
674 struct kcopyd_job *sub_job = (struct kcopyd_job *) context;
675 struct kcopyd_job *job = sub_job->master_job;
676 struct dm_kcopyd_client *kc = job->kc;
678 mutex_lock(&job->lock);
680 /* update the error */
685 job->write_err |= write_err;
688 * Only dispatch more work if there hasn't been an error.
690 if ((!job->read_err && !job->write_err) ||
691 test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
692 /* get the next chunk of work */
693 progress = job->progress;
694 count = job->source.count - progress;
696 if (count > SUB_JOB_SIZE)
697 count = SUB_JOB_SIZE;
699 job->progress += count;
702 mutex_unlock(&job->lock);
708 sub_job->write_offset = progress;
709 sub_job->source.sector += progress;
710 sub_job->source.count = count;
712 for (i = 0; i < job->num_dests; i++) {
713 sub_job->dests[i].sector += progress;
714 sub_job->dests[i].count = count;
717 sub_job->fn = segment_complete;
718 sub_job->context = sub_job;
719 dispatch_job(sub_job);
721 } else if (atomic_dec_and_test(&job->sub_jobs)) {
724 * Queue the completion callback to the kcopyd thread.
726 * Some callers assume that all the completions are called
727 * from a single thread and don't race with each other.
729 * We must not call the callback directly here because this
730 * code may not be executing in the thread.
732 push(&kc->complete_jobs, job);
738 * Create some sub jobs to share the work between them.
740 static void split_job(struct kcopyd_job *master_job)
744 atomic_inc(&master_job->kc->nr_jobs);
746 atomic_set(&master_job->sub_jobs, SPLIT_COUNT);
747 for (i = 0; i < SPLIT_COUNT; i++) {
748 master_job[i + 1].master_job = master_job;
749 segment_complete(0, 0u, &master_job[i + 1]);
753 int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
754 unsigned int num_dests, struct dm_io_region *dests,
755 unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
757 struct kcopyd_job *job;
761 * Allocate an array of jobs consisting of one master job
762 * followed by SPLIT_COUNT sub jobs.
764 job = mempool_alloc(kc->job_pool, GFP_NOIO);
767 * set up for the read.
774 job->num_dests = num_dests;
775 memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
778 * If one of the destination is a host-managed zoned block device,
779 * we need to write sequentially. If one of the destination is a
780 * host-aware device, then leave it to the caller to choose what to do.
782 if (!test_bit(DM_KCOPYD_WRITE_SEQ, &job->flags)) {
783 for (i = 0; i < job->num_dests; i++) {
784 if (bdev_zoned_model(dests[i].bdev) == BLK_ZONED_HM) {
785 set_bit(DM_KCOPYD_WRITE_SEQ, &job->flags);
792 * If we need to write sequentially, errors cannot be ignored.
794 if (test_bit(DM_KCOPYD_WRITE_SEQ, &job->flags) &&
795 test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags))
796 clear_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags);
803 memset(&job->source, 0, sizeof job->source);
804 job->source.count = job->dests[0].count;
805 job->pages = &zero_page_list;
808 * Use WRITE ZEROES to optimize zeroing if all dests support it.
810 job->rw = REQ_OP_WRITE_ZEROES;
811 for (i = 0; i < job->num_dests; i++)
812 if (!bdev_write_zeroes_sectors(job->dests[i].bdev)) {
819 job->context = context;
820 job->master_job = job;
821 job->write_offset = 0;
823 if (job->source.count <= SUB_JOB_SIZE)
826 mutex_init(&job->lock);
833 EXPORT_SYMBOL(dm_kcopyd_copy);
835 int dm_kcopyd_zero(struct dm_kcopyd_client *kc,
836 unsigned num_dests, struct dm_io_region *dests,
837 unsigned flags, dm_kcopyd_notify_fn fn, void *context)
839 return dm_kcopyd_copy(kc, NULL, num_dests, dests, flags, fn, context);
841 EXPORT_SYMBOL(dm_kcopyd_zero);
843 void *dm_kcopyd_prepare_callback(struct dm_kcopyd_client *kc,
844 dm_kcopyd_notify_fn fn, void *context)
846 struct kcopyd_job *job;
848 job = mempool_alloc(kc->job_pool, GFP_NOIO);
850 memset(job, 0, sizeof(struct kcopyd_job));
853 job->context = context;
854 job->master_job = job;
856 atomic_inc(&kc->nr_jobs);
860 EXPORT_SYMBOL(dm_kcopyd_prepare_callback);
862 void dm_kcopyd_do_callback(void *j, int read_err, unsigned long write_err)
864 struct kcopyd_job *job = j;
865 struct dm_kcopyd_client *kc = job->kc;
867 job->read_err = read_err;
868 job->write_err = write_err;
870 push(&kc->callback_jobs, job);
873 EXPORT_SYMBOL(dm_kcopyd_do_callback);
876 * Cancels a kcopyd job, eg. someone might be deactivating a
880 int kcopyd_cancel(struct kcopyd_job *job, int block)
887 /*-----------------------------------------------------------------
889 *---------------------------------------------------------------*/
890 struct dm_kcopyd_client *dm_kcopyd_client_create(struct dm_kcopyd_throttle *throttle)
893 struct dm_kcopyd_client *kc;
895 kc = kzalloc(sizeof(*kc), GFP_KERNEL);
897 return ERR_PTR(-ENOMEM);
899 spin_lock_init(&kc->job_lock);
900 INIT_LIST_HEAD(&kc->callback_jobs);
901 INIT_LIST_HEAD(&kc->complete_jobs);
902 INIT_LIST_HEAD(&kc->io_jobs);
903 INIT_LIST_HEAD(&kc->pages_jobs);
904 kc->throttle = throttle;
906 kc->job_pool = mempool_create_slab_pool(MIN_JOBS, _job_cache);
910 INIT_WORK(&kc->kcopyd_work, do_work);
911 kc->kcopyd_wq = alloc_workqueue("kcopyd", WQ_MEM_RECLAIM, 0);
916 kc->nr_reserved_pages = kc->nr_free_pages = 0;
917 r = client_reserve_pages(kc, RESERVE_PAGES);
919 goto bad_client_pages;
921 kc->io_client = dm_io_client_create();
922 if (IS_ERR(kc->io_client)) {
923 r = PTR_ERR(kc->io_client);
927 init_waitqueue_head(&kc->destroyq);
928 atomic_set(&kc->nr_jobs, 0);
933 client_free_pages(kc);
935 destroy_workqueue(kc->kcopyd_wq);
937 mempool_destroy(kc->job_pool);
943 EXPORT_SYMBOL(dm_kcopyd_client_create);
945 void dm_kcopyd_client_destroy(struct dm_kcopyd_client *kc)
947 /* Wait for completion of all jobs submitted by this client. */
948 wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
950 BUG_ON(!list_empty(&kc->callback_jobs));
951 BUG_ON(!list_empty(&kc->complete_jobs));
952 BUG_ON(!list_empty(&kc->io_jobs));
953 BUG_ON(!list_empty(&kc->pages_jobs));
954 destroy_workqueue(kc->kcopyd_wq);
955 dm_io_client_destroy(kc->io_client);
956 client_free_pages(kc);
957 mempool_destroy(kc->job_pool);
960 EXPORT_SYMBOL(dm_kcopyd_client_destroy);