2 * Copyright (C) 2002 Sistina Software (UK) Limited.
4 * This file is released under the GPL.
6 * Kcopyd provides a simple interface for copying an area of one
7 * block-device to one or more other block-devices, with an asynchronous
8 * completion notification.
11 #include <asm/types.h>
12 #include <asm/atomic.h>
14 #include <linux/blkdev.h>
15 #include <linux/config.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>
28 static struct workqueue_struct *_kcopyd_wq;
29 static struct work_struct _kcopyd_work;
31 static inline void wake(void)
33 queue_work(_kcopyd_wq, &_kcopyd_work);
36 /*-----------------------------------------------------------------
37 * Each kcopyd client has its own little pool of preallocated
38 * pages for kcopyd io.
39 *---------------------------------------------------------------*/
40 struct kcopyd_client {
41 struct list_head list;
44 struct page_list *pages;
45 unsigned int nr_pages;
46 unsigned int nr_free_pages;
49 static struct page_list *alloc_pl(void)
53 pl = kmalloc(sizeof(*pl), GFP_KERNEL);
57 pl->page = alloc_page(GFP_KERNEL);
66 static void free_pl(struct page_list *pl)
68 __free_page(pl->page);
72 static int kcopyd_get_pages(struct kcopyd_client *kc,
73 unsigned int nr, struct page_list **pages)
78 if (kc->nr_free_pages < nr) {
79 spin_unlock(&kc->lock);
83 kc->nr_free_pages -= nr;
84 for (*pages = pl = kc->pages; --nr; pl = pl->next)
90 spin_unlock(&kc->lock);
95 static void kcopyd_put_pages(struct kcopyd_client *kc, struct page_list *pl)
97 struct page_list *cursor;
100 for (cursor = pl; cursor->next; cursor = cursor->next)
104 cursor->next = kc->pages;
106 spin_unlock(&kc->lock);
110 * These three functions resize the page pool.
112 static void drop_pages(struct page_list *pl)
114 struct page_list *next;
123 static int client_alloc_pages(struct kcopyd_client *kc, unsigned int nr)
126 struct page_list *pl = NULL, *next;
128 for (i = 0; i < nr; i++) {
139 kcopyd_put_pages(kc, pl);
144 static void client_free_pages(struct kcopyd_client *kc)
146 BUG_ON(kc->nr_free_pages != kc->nr_pages);
147 drop_pages(kc->pages);
149 kc->nr_free_pages = kc->nr_pages = 0;
152 /*-----------------------------------------------------------------
153 * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
154 * for this reason we use a mempool to prevent the client from
155 * ever having to do io (which could cause a deadlock).
156 *---------------------------------------------------------------*/
158 struct kcopyd_client *kc;
159 struct list_head list;
163 * Error state of the job.
166 unsigned int write_err;
169 * Either READ or WRITE
172 struct io_region source;
175 * The destinations for the transfer.
177 unsigned int num_dests;
178 struct io_region dests[KCOPYD_MAX_REGIONS];
181 unsigned int nr_pages;
182 struct page_list *pages;
185 * Set this to ensure you are notified when the job has
186 * completed. 'context' is for callback to use.
192 * These fields are only used if the job has been split
193 * into more manageable parts.
195 struct semaphore lock;
200 /* FIXME: this should scale with the number of pages */
203 static kmem_cache_t *_job_cache;
204 static mempool_t *_job_pool;
207 * We maintain three lists of jobs:
209 * i) jobs waiting for pages
210 * ii) jobs that have pages, and are waiting for the io to be issued.
211 * iii) jobs that have completed.
213 * All three of these are protected by job_lock.
215 static DEFINE_SPINLOCK(_job_lock);
217 static LIST_HEAD(_complete_jobs);
218 static LIST_HEAD(_io_jobs);
219 static LIST_HEAD(_pages_jobs);
221 static int jobs_init(void)
223 _job_cache = kmem_cache_create("kcopyd-jobs",
224 sizeof(struct kcopyd_job),
225 __alignof__(struct kcopyd_job),
230 _job_pool = mempool_create(MIN_JOBS, mempool_alloc_slab,
231 mempool_free_slab, _job_cache);
233 kmem_cache_destroy(_job_cache);
240 static void jobs_exit(void)
242 BUG_ON(!list_empty(&_complete_jobs));
243 BUG_ON(!list_empty(&_io_jobs));
244 BUG_ON(!list_empty(&_pages_jobs));
246 mempool_destroy(_job_pool);
247 kmem_cache_destroy(_job_cache);
253 * Functions to push and pop a job onto the head of a given job
256 static inline struct kcopyd_job *pop(struct list_head *jobs)
258 struct kcopyd_job *job = NULL;
261 spin_lock_irqsave(&_job_lock, flags);
263 if (!list_empty(jobs)) {
264 job = list_entry(jobs->next, struct kcopyd_job, list);
265 list_del(&job->list);
267 spin_unlock_irqrestore(&_job_lock, flags);
272 static inline void push(struct list_head *jobs, struct kcopyd_job *job)
276 spin_lock_irqsave(&_job_lock, flags);
277 list_add_tail(&job->list, jobs);
278 spin_unlock_irqrestore(&_job_lock, flags);
282 * These three functions process 1 item from the corresponding
288 * > 0: can't process yet.
290 static int run_complete_job(struct kcopyd_job *job)
292 void *context = job->context;
293 int read_err = job->read_err;
294 unsigned int write_err = job->write_err;
295 kcopyd_notify_fn fn = job->fn;
297 kcopyd_put_pages(job->kc, job->pages);
298 mempool_free(job, _job_pool);
299 fn(read_err, write_err, context);
303 static void complete_io(unsigned long error, void *context)
305 struct kcopyd_job *job = (struct kcopyd_job *) context;
308 if (job->rw == WRITE)
309 job->write_err &= error;
313 if (!test_bit(KCOPYD_IGNORE_ERROR, &job->flags)) {
314 push(&_complete_jobs, job);
320 if (job->rw == WRITE)
321 push(&_complete_jobs, job);
325 push(&_io_jobs, job);
332 * Request io on as many buffer heads as we can currently get for
335 static int run_io_job(struct kcopyd_job *job)
340 r = dm_io_async(1, &job->source, job->rw,
342 job->offset, complete_io, job);
345 r = dm_io_async(job->num_dests, job->dests, job->rw,
347 job->offset, complete_io, job);
352 static int run_pages_job(struct kcopyd_job *job)
356 job->nr_pages = dm_div_up(job->dests[0].count + job->offset,
358 r = kcopyd_get_pages(job->kc, job->nr_pages, &job->pages);
360 /* this job is ready for io */
361 push(&_io_jobs, job);
366 /* can't complete now */
373 * Run through a list for as long as possible. Returns the count
374 * of successful jobs.
376 static int process_jobs(struct list_head *jobs, int (*fn) (struct kcopyd_job *))
378 struct kcopyd_job *job;
381 while ((job = pop(jobs))) {
386 /* error this rogue job */
387 if (job->rw == WRITE)
388 job->write_err = (unsigned int) -1;
391 push(&_complete_jobs, job);
397 * We couldn't service this job ATM, so
398 * push this job back onto the list.
411 * kcopyd does this every time it's woken up.
413 static void do_work(void *ignored)
416 * The order that these are called is *very* important.
417 * complete jobs can free some pages for pages jobs.
418 * Pages jobs when successful will jump onto the io jobs
419 * list. io jobs call wake when they complete and it all
422 process_jobs(&_complete_jobs, run_complete_job);
423 process_jobs(&_pages_jobs, run_pages_job);
424 process_jobs(&_io_jobs, run_io_job);
428 * If we are copying a small region we just dispatch a single job
429 * to do the copy, otherwise the io has to be split up into many
432 static void dispatch_job(struct kcopyd_job *job)
434 push(&_pages_jobs, job);
438 #define SUB_JOB_SIZE 128
439 static void segment_complete(int read_err,
440 unsigned int write_err, void *context)
442 /* FIXME: tidy this function */
443 sector_t progress = 0;
445 struct kcopyd_job *job = (struct kcopyd_job *) context;
449 /* update the error */
454 job->write_err &= write_err;
457 * Only dispatch more work if there hasn't been an error.
459 if ((!job->read_err && !job->write_err) ||
460 test_bit(KCOPYD_IGNORE_ERROR, &job->flags)) {
461 /* get the next chunk of work */
462 progress = job->progress;
463 count = job->source.count - progress;
465 if (count > SUB_JOB_SIZE)
466 count = SUB_JOB_SIZE;
468 job->progress += count;
475 struct kcopyd_job *sub_job = mempool_alloc(_job_pool, GFP_NOIO);
478 sub_job->source.sector += progress;
479 sub_job->source.count = count;
481 for (i = 0; i < job->num_dests; i++) {
482 sub_job->dests[i].sector += progress;
483 sub_job->dests[i].count = count;
486 sub_job->fn = segment_complete;
487 sub_job->context = job;
488 dispatch_job(sub_job);
490 } else if (atomic_dec_and_test(&job->sub_jobs)) {
493 * To avoid a race we must keep the job around
494 * until after the notify function has completed.
495 * Otherwise the client may try and stop the job
496 * after we've completed.
498 job->fn(read_err, write_err, job->context);
499 mempool_free(job, _job_pool);
504 * Create some little jobs that will do the move between
507 #define SPLIT_COUNT 8
508 static void split_job(struct kcopyd_job *job)
512 atomic_set(&job->sub_jobs, SPLIT_COUNT);
513 for (i = 0; i < SPLIT_COUNT; i++)
514 segment_complete(0, 0u, job);
517 int kcopyd_copy(struct kcopyd_client *kc, struct io_region *from,
518 unsigned int num_dests, struct io_region *dests,
519 unsigned int flags, kcopyd_notify_fn fn, void *context)
521 struct kcopyd_job *job;
524 * Allocate a new job.
526 job = mempool_alloc(_job_pool, GFP_NOIO);
529 * set up for the read.
539 job->num_dests = num_dests;
540 memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
547 job->context = context;
549 if (job->source.count < SUB_JOB_SIZE)
553 init_MUTEX(&job->lock);
562 * Cancels a kcopyd job, eg. someone might be deactivating a
566 int kcopyd_cancel(struct kcopyd_job *job, int block)
573 /*-----------------------------------------------------------------
575 *---------------------------------------------------------------*/
576 static DECLARE_MUTEX(_client_lock);
577 static LIST_HEAD(_clients);
579 static void client_add(struct kcopyd_client *kc)
582 list_add(&kc->list, &_clients);
586 static void client_del(struct kcopyd_client *kc)
593 static DECLARE_MUTEX(kcopyd_init_lock);
594 static int kcopyd_clients = 0;
596 static int kcopyd_init(void)
600 down(&kcopyd_init_lock);
602 if (kcopyd_clients) {
603 /* Already initialized. */
605 up(&kcopyd_init_lock);
611 up(&kcopyd_init_lock);
615 _kcopyd_wq = create_singlethread_workqueue("kcopyd");
618 up(&kcopyd_init_lock);
623 INIT_WORK(&_kcopyd_work, do_work, NULL);
624 up(&kcopyd_init_lock);
628 static void kcopyd_exit(void)
630 down(&kcopyd_init_lock);
632 if (!kcopyd_clients) {
634 destroy_workqueue(_kcopyd_wq);
637 up(&kcopyd_init_lock);
640 int kcopyd_client_create(unsigned int nr_pages, struct kcopyd_client **result)
643 struct kcopyd_client *kc;
649 kc = kmalloc(sizeof(*kc), GFP_KERNEL);
655 spin_lock_init(&kc->lock);
657 kc->nr_pages = kc->nr_free_pages = 0;
658 r = client_alloc_pages(kc, nr_pages);
665 r = dm_io_get(nr_pages);
667 client_free_pages(kc);
678 void kcopyd_client_destroy(struct kcopyd_client *kc)
680 dm_io_put(kc->nr_pages);
681 client_free_pages(kc);
687 EXPORT_SYMBOL(kcopyd_client_create);
688 EXPORT_SYMBOL(kcopyd_client_destroy);
689 EXPORT_SYMBOL(kcopyd_copy);