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 <asm/types.h>
13 #include <asm/atomic.h>
15 #include <linux/blkdev.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>
29 static struct workqueue_struct *_kcopyd_wq;
30 static struct work_struct _kcopyd_work;
32 static void wake(void)
34 queue_work(_kcopyd_wq, &_kcopyd_work);
37 /*-----------------------------------------------------------------
38 * Each kcopyd client has its own little pool of preallocated
39 * pages for kcopyd io.
40 *---------------------------------------------------------------*/
41 struct kcopyd_client {
42 struct list_head list;
45 struct page_list *pages;
46 unsigned int nr_pages;
47 unsigned int nr_free_pages;
49 struct dm_io_client *io_client;
51 wait_queue_head_t destroyq;
55 static struct page_list *alloc_pl(void)
59 pl = kmalloc(sizeof(*pl), GFP_KERNEL);
63 pl->page = alloc_page(GFP_KERNEL);
72 static void free_pl(struct page_list *pl)
74 __free_page(pl->page);
78 static int kcopyd_get_pages(struct kcopyd_client *kc,
79 unsigned int nr, struct page_list **pages)
84 if (kc->nr_free_pages < nr) {
85 spin_unlock(&kc->lock);
89 kc->nr_free_pages -= nr;
90 for (*pages = pl = kc->pages; --nr; pl = pl->next)
96 spin_unlock(&kc->lock);
101 static void kcopyd_put_pages(struct kcopyd_client *kc, struct page_list *pl)
103 struct page_list *cursor;
105 spin_lock(&kc->lock);
106 for (cursor = pl; cursor->next; cursor = cursor->next)
110 cursor->next = kc->pages;
112 spin_unlock(&kc->lock);
116 * These three functions resize the page pool.
118 static void drop_pages(struct page_list *pl)
120 struct page_list *next;
129 static int client_alloc_pages(struct kcopyd_client *kc, unsigned int nr)
132 struct page_list *pl = NULL, *next;
134 for (i = 0; i < nr; i++) {
145 kcopyd_put_pages(kc, pl);
150 static void client_free_pages(struct kcopyd_client *kc)
152 BUG_ON(kc->nr_free_pages != kc->nr_pages);
153 drop_pages(kc->pages);
155 kc->nr_free_pages = kc->nr_pages = 0;
158 /*-----------------------------------------------------------------
159 * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
160 * for this reason we use a mempool to prevent the client from
161 * ever having to do io (which could cause a deadlock).
162 *---------------------------------------------------------------*/
164 struct kcopyd_client *kc;
165 struct list_head list;
169 * Error state of the job.
172 unsigned long write_err;
175 * Either READ or WRITE
178 struct io_region source;
181 * The destinations for the transfer.
183 unsigned int num_dests;
184 struct io_region dests[KCOPYD_MAX_REGIONS];
187 unsigned int nr_pages;
188 struct page_list *pages;
191 * Set this to ensure you are notified when the job has
192 * completed. 'context' is for callback to use.
198 * These fields are only used if the job has been split
199 * into more manageable parts.
206 /* FIXME: this should scale with the number of pages */
209 static struct kmem_cache *_job_cache;
210 static mempool_t *_job_pool;
213 * We maintain three lists of jobs:
215 * i) jobs waiting for pages
216 * ii) jobs that have pages, and are waiting for the io to be issued.
217 * iii) jobs that have completed.
219 * All three of these are protected by job_lock.
221 static DEFINE_SPINLOCK(_job_lock);
223 static LIST_HEAD(_complete_jobs);
224 static LIST_HEAD(_io_jobs);
225 static LIST_HEAD(_pages_jobs);
227 static int jobs_init(void)
229 _job_cache = KMEM_CACHE(kcopyd_job, 0);
233 _job_pool = mempool_create_slab_pool(MIN_JOBS, _job_cache);
235 kmem_cache_destroy(_job_cache);
242 static void jobs_exit(void)
244 BUG_ON(!list_empty(&_complete_jobs));
245 BUG_ON(!list_empty(&_io_jobs));
246 BUG_ON(!list_empty(&_pages_jobs));
248 mempool_destroy(_job_pool);
249 kmem_cache_destroy(_job_cache);
255 * Functions to push and pop a job onto the head of a given job
258 static struct kcopyd_job *pop(struct list_head *jobs)
260 struct kcopyd_job *job = NULL;
263 spin_lock_irqsave(&_job_lock, flags);
265 if (!list_empty(jobs)) {
266 job = list_entry(jobs->next, struct kcopyd_job, list);
267 list_del(&job->list);
269 spin_unlock_irqrestore(&_job_lock, flags);
274 static void push(struct list_head *jobs, struct kcopyd_job *job)
278 spin_lock_irqsave(&_job_lock, flags);
279 list_add_tail(&job->list, jobs);
280 spin_unlock_irqrestore(&_job_lock, flags);
284 * These three functions process 1 item from the corresponding
290 * > 0: can't process yet.
292 static int run_complete_job(struct kcopyd_job *job)
294 void *context = job->context;
295 int read_err = job->read_err;
296 unsigned long write_err = job->write_err;
297 kcopyd_notify_fn fn = job->fn;
298 struct kcopyd_client *kc = job->kc;
300 kcopyd_put_pages(kc, job->pages);
301 mempool_free(job, _job_pool);
302 fn(read_err, write_err, context);
304 if (atomic_dec_and_test(&kc->nr_jobs))
305 wake_up(&kc->destroyq);
310 static void complete_io(unsigned long error, void *context)
312 struct kcopyd_job *job = (struct kcopyd_job *) context;
315 if (job->rw == WRITE)
316 job->write_err |= error;
320 if (!test_bit(KCOPYD_IGNORE_ERROR, &job->flags)) {
321 push(&_complete_jobs, job);
327 if (job->rw == WRITE)
328 push(&_complete_jobs, job);
332 push(&_io_jobs, job);
339 * Request io on as many buffer heads as we can currently get for
342 static int run_io_job(struct kcopyd_job *job)
345 struct dm_io_request io_req = {
347 .mem.type = DM_IO_PAGE_LIST,
348 .mem.ptr.pl = job->pages,
349 .mem.offset = job->offset,
350 .notify.fn = complete_io,
351 .notify.context = job,
352 .client = job->kc->io_client,
356 r = dm_io(&io_req, 1, &job->source, NULL);
358 r = dm_io(&io_req, job->num_dests, job->dests, NULL);
363 static int run_pages_job(struct kcopyd_job *job)
367 job->nr_pages = dm_div_up(job->dests[0].count + job->offset,
369 r = kcopyd_get_pages(job->kc, job->nr_pages, &job->pages);
371 /* this job is ready for io */
372 push(&_io_jobs, job);
377 /* can't complete now */
384 * Run through a list for as long as possible. Returns the count
385 * of successful jobs.
387 static int process_jobs(struct list_head *jobs, int (*fn) (struct kcopyd_job *))
389 struct kcopyd_job *job;
392 while ((job = pop(jobs))) {
397 /* error this rogue job */
398 if (job->rw == WRITE)
399 job->write_err = (unsigned long) -1L;
402 push(&_complete_jobs, job);
408 * We couldn't service this job ATM, so
409 * push this job back onto the list.
422 * kcopyd does this every time it's woken up.
424 static void do_work(struct work_struct *ignored)
427 * The order that these are called is *very* important.
428 * complete jobs can free some pages for pages jobs.
429 * Pages jobs when successful will jump onto the io jobs
430 * list. io jobs call wake when they complete and it all
433 process_jobs(&_complete_jobs, run_complete_job);
434 process_jobs(&_pages_jobs, run_pages_job);
435 process_jobs(&_io_jobs, run_io_job);
439 * If we are copying a small region we just dispatch a single job
440 * to do the copy, otherwise the io has to be split up into many
443 static void dispatch_job(struct kcopyd_job *job)
445 atomic_inc(&job->kc->nr_jobs);
446 push(&_pages_jobs, job);
450 #define SUB_JOB_SIZE 128
451 static void segment_complete(int read_err, unsigned long write_err,
454 /* FIXME: tidy this function */
455 sector_t progress = 0;
457 struct kcopyd_job *job = (struct kcopyd_job *) context;
459 mutex_lock(&job->lock);
461 /* update the error */
466 job->write_err |= write_err;
469 * Only dispatch more work if there hasn't been an error.
471 if ((!job->read_err && !job->write_err) ||
472 test_bit(KCOPYD_IGNORE_ERROR, &job->flags)) {
473 /* get the next chunk of work */
474 progress = job->progress;
475 count = job->source.count - progress;
477 if (count > SUB_JOB_SIZE)
478 count = SUB_JOB_SIZE;
480 job->progress += count;
483 mutex_unlock(&job->lock);
487 struct kcopyd_job *sub_job = mempool_alloc(_job_pool, GFP_NOIO);
490 sub_job->source.sector += progress;
491 sub_job->source.count = count;
493 for (i = 0; i < job->num_dests; i++) {
494 sub_job->dests[i].sector += progress;
495 sub_job->dests[i].count = count;
498 sub_job->fn = segment_complete;
499 sub_job->context = job;
500 dispatch_job(sub_job);
502 } else if (atomic_dec_and_test(&job->sub_jobs)) {
505 * To avoid a race we must keep the job around
506 * until after the notify function has completed.
507 * Otherwise the client may try and stop the job
508 * after we've completed.
510 job->fn(read_err, write_err, job->context);
511 mempool_free(job, _job_pool);
516 * Create some little jobs that will do the move between
519 #define SPLIT_COUNT 8
520 static void split_job(struct kcopyd_job *job)
524 atomic_set(&job->sub_jobs, SPLIT_COUNT);
525 for (i = 0; i < SPLIT_COUNT; i++)
526 segment_complete(0, 0u, job);
529 int kcopyd_copy(struct kcopyd_client *kc, struct io_region *from,
530 unsigned int num_dests, struct io_region *dests,
531 unsigned int flags, kcopyd_notify_fn fn, void *context)
533 struct kcopyd_job *job;
536 * Allocate a new job.
538 job = mempool_alloc(_job_pool, GFP_NOIO);
541 * set up for the read.
551 job->num_dests = num_dests;
552 memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
559 job->context = context;
561 if (job->source.count < SUB_JOB_SIZE)
565 mutex_init(&job->lock);
574 * Cancels a kcopyd job, eg. someone might be deactivating a
578 int kcopyd_cancel(struct kcopyd_job *job, int block)
585 /*-----------------------------------------------------------------
587 *---------------------------------------------------------------*/
588 static DEFINE_MUTEX(_client_lock);
589 static LIST_HEAD(_clients);
591 static void client_add(struct kcopyd_client *kc)
593 mutex_lock(&_client_lock);
594 list_add(&kc->list, &_clients);
595 mutex_unlock(&_client_lock);
598 static void client_del(struct kcopyd_client *kc)
600 mutex_lock(&_client_lock);
602 mutex_unlock(&_client_lock);
605 static DEFINE_MUTEX(kcopyd_init_lock);
606 static int kcopyd_clients = 0;
608 static int kcopyd_init(void)
612 mutex_lock(&kcopyd_init_lock);
614 if (kcopyd_clients) {
615 /* Already initialized. */
617 mutex_unlock(&kcopyd_init_lock);
623 mutex_unlock(&kcopyd_init_lock);
627 _kcopyd_wq = create_singlethread_workqueue("kcopyd");
630 mutex_unlock(&kcopyd_init_lock);
635 INIT_WORK(&_kcopyd_work, do_work);
636 mutex_unlock(&kcopyd_init_lock);
640 static void kcopyd_exit(void)
642 mutex_lock(&kcopyd_init_lock);
644 if (!kcopyd_clients) {
646 destroy_workqueue(_kcopyd_wq);
649 mutex_unlock(&kcopyd_init_lock);
652 int kcopyd_client_create(unsigned int nr_pages, struct kcopyd_client **result)
655 struct kcopyd_client *kc;
661 kc = kmalloc(sizeof(*kc), GFP_KERNEL);
667 spin_lock_init(&kc->lock);
669 kc->nr_pages = kc->nr_free_pages = 0;
670 r = client_alloc_pages(kc, nr_pages);
677 kc->io_client = dm_io_client_create(nr_pages);
678 if (IS_ERR(kc->io_client)) {
679 r = PTR_ERR(kc->io_client);
680 client_free_pages(kc);
686 init_waitqueue_head(&kc->destroyq);
687 atomic_set(&kc->nr_jobs, 0);
694 void kcopyd_client_destroy(struct kcopyd_client *kc)
696 /* Wait for completion of all jobs submitted by this client. */
697 wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
699 dm_io_client_destroy(kc->io_client);
700 client_free_pages(kc);
706 EXPORT_SYMBOL(kcopyd_client_create);
707 EXPORT_SYMBOL(kcopyd_client_destroy);
708 EXPORT_SYMBOL(kcopyd_copy);