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.
362 struct kcopyd_job *master_job;
365 static struct kmem_cache *_job_cache;
367 int __init dm_kcopyd_init(void)
369 _job_cache = kmem_cache_create("kcopyd_job",
370 sizeof(struct kcopyd_job) * (SPLIT_COUNT + 1),
371 __alignof__(struct kcopyd_job), 0, NULL);
375 zero_page_list.next = &zero_page_list;
376 zero_page_list.page = ZERO_PAGE(0);
381 void dm_kcopyd_exit(void)
383 kmem_cache_destroy(_job_cache);
388 * Functions to push and pop a job onto the head of a given job
391 static struct kcopyd_job *pop(struct list_head *jobs,
392 struct dm_kcopyd_client *kc)
394 struct kcopyd_job *job = NULL;
397 spin_lock_irqsave(&kc->job_lock, flags);
399 if (!list_empty(jobs)) {
400 job = list_entry(jobs->next, struct kcopyd_job, list);
401 list_del(&job->list);
403 spin_unlock_irqrestore(&kc->job_lock, flags);
408 static void push(struct list_head *jobs, struct kcopyd_job *job)
411 struct dm_kcopyd_client *kc = job->kc;
413 spin_lock_irqsave(&kc->job_lock, flags);
414 list_add_tail(&job->list, jobs);
415 spin_unlock_irqrestore(&kc->job_lock, flags);
419 static void push_head(struct list_head *jobs, struct kcopyd_job *job)
422 struct dm_kcopyd_client *kc = job->kc;
424 spin_lock_irqsave(&kc->job_lock, flags);
425 list_add(&job->list, jobs);
426 spin_unlock_irqrestore(&kc->job_lock, flags);
430 * These three functions process 1 item from the corresponding
436 * > 0: can't process yet.
438 static int run_complete_job(struct kcopyd_job *job)
440 void *context = job->context;
441 int read_err = job->read_err;
442 unsigned long write_err = job->write_err;
443 dm_kcopyd_notify_fn fn = job->fn;
444 struct dm_kcopyd_client *kc = job->kc;
446 if (job->pages && job->pages != &zero_page_list)
447 kcopyd_put_pages(kc, job->pages);
449 * If this is the master job, the sub jobs have already
450 * completed so we can free everything.
452 if (job->master_job == job)
453 mempool_free(job, kc->job_pool);
454 fn(read_err, write_err, context);
456 if (atomic_dec_and_test(&kc->nr_jobs))
457 wake_up(&kc->destroyq);
464 static void complete_io(unsigned long error, void *context)
466 struct kcopyd_job *job = (struct kcopyd_job *) context;
467 struct dm_kcopyd_client *kc = job->kc;
469 io_job_finish(kc->throttle);
473 job->write_err |= error;
477 if (!test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
478 push(&kc->complete_jobs, job);
485 push(&kc->complete_jobs, job);
489 push(&kc->io_jobs, job);
496 * Request io on as many buffer heads as we can currently get for
499 static int run_io_job(struct kcopyd_job *job)
502 struct dm_io_request io_req = {
504 .mem.type = DM_IO_PAGE_LIST,
505 .mem.ptr.pl = job->pages,
507 .notify.fn = complete_io,
508 .notify.context = job,
509 .client = job->kc->io_client,
512 io_job_start(job->kc->throttle);
515 r = dm_io(&io_req, 1, &job->source, NULL);
517 r = dm_io(&io_req, job->num_dests, job->dests, NULL);
522 static int run_pages_job(struct kcopyd_job *job)
525 unsigned nr_pages = dm_div_up(job->dests[0].count, PAGE_SIZE >> 9);
527 r = kcopyd_get_pages(job->kc, nr_pages, &job->pages);
529 /* this job is ready for io */
530 push(&job->kc->io_jobs, job);
535 /* can't complete now */
542 * Run through a list for as long as possible. Returns the count
543 * of successful jobs.
545 static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
546 int (*fn) (struct kcopyd_job *))
548 struct kcopyd_job *job;
551 while ((job = pop(jobs, kc))) {
556 /* error this rogue job */
558 job->write_err = (unsigned long) -1L;
561 push(&kc->complete_jobs, job);
567 * We couldn't service this job ATM, so
568 * push this job back onto the list.
570 push_head(jobs, job);
581 * kcopyd does this every time it's woken up.
583 static void do_work(struct work_struct *work)
585 struct dm_kcopyd_client *kc = container_of(work,
586 struct dm_kcopyd_client, kcopyd_work);
587 struct blk_plug plug;
591 * The order that these are called is *very* important.
592 * complete jobs can free some pages for pages jobs.
593 * Pages jobs when successful will jump onto the io jobs
594 * list. io jobs call wake when they complete and it all
597 spin_lock_irqsave(&kc->job_lock, flags);
598 list_splice_tail_init(&kc->callback_jobs, &kc->complete_jobs);
599 spin_unlock_irqrestore(&kc->job_lock, flags);
601 blk_start_plug(&plug);
602 process_jobs(&kc->complete_jobs, kc, run_complete_job);
603 process_jobs(&kc->pages_jobs, kc, run_pages_job);
604 process_jobs(&kc->io_jobs, kc, run_io_job);
605 blk_finish_plug(&plug);
609 * If we are copying a small region we just dispatch a single job
610 * to do the copy, otherwise the io has to be split up into many
613 static void dispatch_job(struct kcopyd_job *job)
615 struct dm_kcopyd_client *kc = job->kc;
616 atomic_inc(&kc->nr_jobs);
617 if (unlikely(!job->source.count))
618 push(&kc->callback_jobs, job);
619 else if (job->pages == &zero_page_list)
620 push(&kc->io_jobs, job);
622 push(&kc->pages_jobs, job);
626 static void segment_complete(int read_err, unsigned long write_err,
629 /* FIXME: tidy this function */
630 sector_t progress = 0;
632 struct kcopyd_job *sub_job = (struct kcopyd_job *) context;
633 struct kcopyd_job *job = sub_job->master_job;
634 struct dm_kcopyd_client *kc = job->kc;
636 mutex_lock(&job->lock);
638 /* update the error */
643 job->write_err |= write_err;
646 * Only dispatch more work if there hasn't been an error.
648 if ((!job->read_err && !job->write_err) ||
649 test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
650 /* get the next chunk of work */
651 progress = job->progress;
652 count = job->source.count - progress;
654 if (count > SUB_JOB_SIZE)
655 count = SUB_JOB_SIZE;
657 job->progress += count;
660 mutex_unlock(&job->lock);
666 sub_job->source.sector += progress;
667 sub_job->source.count = count;
669 for (i = 0; i < job->num_dests; i++) {
670 sub_job->dests[i].sector += progress;
671 sub_job->dests[i].count = count;
674 sub_job->fn = segment_complete;
675 sub_job->context = sub_job;
676 dispatch_job(sub_job);
678 } else if (atomic_dec_and_test(&job->sub_jobs)) {
681 * Queue the completion callback to the kcopyd thread.
683 * Some callers assume that all the completions are called
684 * from a single thread and don't race with each other.
686 * We must not call the callback directly here because this
687 * code may not be executing in the thread.
689 push(&kc->complete_jobs, job);
695 * Create some sub jobs to share the work between them.
697 static void split_job(struct kcopyd_job *master_job)
701 atomic_inc(&master_job->kc->nr_jobs);
703 atomic_set(&master_job->sub_jobs, SPLIT_COUNT);
704 for (i = 0; i < SPLIT_COUNT; i++) {
705 master_job[i + 1].master_job = master_job;
706 segment_complete(0, 0u, &master_job[i + 1]);
710 int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
711 unsigned int num_dests, struct dm_io_region *dests,
712 unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
714 struct kcopyd_job *job;
718 * Allocate an array of jobs consisting of one master job
719 * followed by SPLIT_COUNT sub jobs.
721 job = mempool_alloc(kc->job_pool, GFP_NOIO);
724 * set up for the read.
731 job->num_dests = num_dests;
732 memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
739 memset(&job->source, 0, sizeof job->source);
740 job->source.count = job->dests[0].count;
741 job->pages = &zero_page_list;
744 * Use WRITE SAME to optimize zeroing if all dests support it.
746 job->rw = WRITE | REQ_WRITE_SAME;
747 for (i = 0; i < job->num_dests; i++)
748 if (!bdev_write_same(job->dests[i].bdev)) {
755 job->context = context;
756 job->master_job = job;
758 if (job->source.count <= SUB_JOB_SIZE)
761 mutex_init(&job->lock);
768 EXPORT_SYMBOL(dm_kcopyd_copy);
770 int dm_kcopyd_zero(struct dm_kcopyd_client *kc,
771 unsigned num_dests, struct dm_io_region *dests,
772 unsigned flags, dm_kcopyd_notify_fn fn, void *context)
774 return dm_kcopyd_copy(kc, NULL, num_dests, dests, flags, fn, context);
776 EXPORT_SYMBOL(dm_kcopyd_zero);
778 void *dm_kcopyd_prepare_callback(struct dm_kcopyd_client *kc,
779 dm_kcopyd_notify_fn fn, void *context)
781 struct kcopyd_job *job;
783 job = mempool_alloc(kc->job_pool, GFP_NOIO);
785 memset(job, 0, sizeof(struct kcopyd_job));
788 job->context = context;
789 job->master_job = job;
791 atomic_inc(&kc->nr_jobs);
795 EXPORT_SYMBOL(dm_kcopyd_prepare_callback);
797 void dm_kcopyd_do_callback(void *j, int read_err, unsigned long write_err)
799 struct kcopyd_job *job = j;
800 struct dm_kcopyd_client *kc = job->kc;
802 job->read_err = read_err;
803 job->write_err = write_err;
805 push(&kc->callback_jobs, job);
808 EXPORT_SYMBOL(dm_kcopyd_do_callback);
811 * Cancels a kcopyd job, eg. someone might be deactivating a
815 int kcopyd_cancel(struct kcopyd_job *job, int block)
822 /*-----------------------------------------------------------------
824 *---------------------------------------------------------------*/
825 struct dm_kcopyd_client *dm_kcopyd_client_create(struct dm_kcopyd_throttle *throttle)
828 struct dm_kcopyd_client *kc;
830 kc = kzalloc(sizeof(*kc), GFP_KERNEL);
832 return ERR_PTR(-ENOMEM);
834 spin_lock_init(&kc->job_lock);
835 INIT_LIST_HEAD(&kc->callback_jobs);
836 INIT_LIST_HEAD(&kc->complete_jobs);
837 INIT_LIST_HEAD(&kc->io_jobs);
838 INIT_LIST_HEAD(&kc->pages_jobs);
839 kc->throttle = throttle;
841 kc->job_pool = mempool_create_slab_pool(MIN_JOBS, _job_cache);
845 INIT_WORK(&kc->kcopyd_work, do_work);
846 kc->kcopyd_wq = alloc_workqueue("kcopyd", WQ_MEM_RECLAIM, 0);
851 kc->nr_reserved_pages = kc->nr_free_pages = 0;
852 r = client_reserve_pages(kc, RESERVE_PAGES);
854 goto bad_client_pages;
856 kc->io_client = dm_io_client_create();
857 if (IS_ERR(kc->io_client)) {
858 r = PTR_ERR(kc->io_client);
862 init_waitqueue_head(&kc->destroyq);
863 atomic_set(&kc->nr_jobs, 0);
868 client_free_pages(kc);
870 destroy_workqueue(kc->kcopyd_wq);
872 mempool_destroy(kc->job_pool);
878 EXPORT_SYMBOL(dm_kcopyd_client_create);
880 void dm_kcopyd_client_destroy(struct dm_kcopyd_client *kc)
882 /* Wait for completion of all jobs submitted by this client. */
883 wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
885 BUG_ON(!list_empty(&kc->callback_jobs));
886 BUG_ON(!list_empty(&kc->complete_jobs));
887 BUG_ON(!list_empty(&kc->io_jobs));
888 BUG_ON(!list_empty(&kc->pages_jobs));
889 destroy_workqueue(kc->kcopyd_wq);
890 dm_io_client_destroy(kc->io_client);
891 client_free_pages(kc);
892 mempool_destroy(kc->job_pool);
895 EXPORT_SYMBOL(dm_kcopyd_client_destroy);