Loading...
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/device-mapper.h>
26#include <linux/dm-kcopyd.h>
27
28#include "dm.h"
29
30#define SUB_JOB_SIZE 128
31#define SPLIT_COUNT 8
32#define MIN_JOBS 8
33#define RESERVE_PAGES (DIV_ROUND_UP(SUB_JOB_SIZE << SECTOR_SHIFT, PAGE_SIZE))
34
35/*-----------------------------------------------------------------
36 * Each kcopyd client has its own little pool of preallocated
37 * pages for kcopyd io.
38 *---------------------------------------------------------------*/
39struct dm_kcopyd_client {
40 struct page_list *pages;
41 unsigned nr_reserved_pages;
42 unsigned nr_free_pages;
43
44 struct dm_io_client *io_client;
45
46 wait_queue_head_t destroyq;
47 atomic_t nr_jobs;
48
49 mempool_t *job_pool;
50
51 struct workqueue_struct *kcopyd_wq;
52 struct work_struct kcopyd_work;
53
54/*
55 * We maintain three lists of jobs:
56 *
57 * i) jobs waiting for pages
58 * ii) jobs that have pages, and are waiting for the io to be issued.
59 * iii) jobs that have completed.
60 *
61 * All three of these are protected by job_lock.
62 */
63 spinlock_t job_lock;
64 struct list_head complete_jobs;
65 struct list_head io_jobs;
66 struct list_head pages_jobs;
67};
68
69static struct page_list zero_page_list;
70
71static void wake(struct dm_kcopyd_client *kc)
72{
73 queue_work(kc->kcopyd_wq, &kc->kcopyd_work);
74}
75
76/*
77 * Obtain one page for the use of kcopyd.
78 */
79static struct page_list *alloc_pl(gfp_t gfp)
80{
81 struct page_list *pl;
82
83 pl = kmalloc(sizeof(*pl), gfp);
84 if (!pl)
85 return NULL;
86
87 pl->page = alloc_page(gfp);
88 if (!pl->page) {
89 kfree(pl);
90 return NULL;
91 }
92
93 return pl;
94}
95
96static void free_pl(struct page_list *pl)
97{
98 __free_page(pl->page);
99 kfree(pl);
100}
101
102/*
103 * Add the provided pages to a client's free page list, releasing
104 * back to the system any beyond the reserved_pages limit.
105 */
106static void kcopyd_put_pages(struct dm_kcopyd_client *kc, struct page_list *pl)
107{
108 struct page_list *next;
109
110 do {
111 next = pl->next;
112
113 if (kc->nr_free_pages >= kc->nr_reserved_pages)
114 free_pl(pl);
115 else {
116 pl->next = kc->pages;
117 kc->pages = pl;
118 kc->nr_free_pages++;
119 }
120
121 pl = next;
122 } while (pl);
123}
124
125static int kcopyd_get_pages(struct dm_kcopyd_client *kc,
126 unsigned int nr, struct page_list **pages)
127{
128 struct page_list *pl;
129
130 *pages = NULL;
131
132 do {
133 pl = alloc_pl(__GFP_NOWARN | __GFP_NORETRY);
134 if (unlikely(!pl)) {
135 /* Use reserved pages */
136 pl = kc->pages;
137 if (unlikely(!pl))
138 goto out_of_memory;
139 kc->pages = pl->next;
140 kc->nr_free_pages--;
141 }
142 pl->next = *pages;
143 *pages = pl;
144 } while (--nr);
145
146 return 0;
147
148out_of_memory:
149 if (*pages)
150 kcopyd_put_pages(kc, *pages);
151 return -ENOMEM;
152}
153
154/*
155 * These three functions resize the page pool.
156 */
157static void drop_pages(struct page_list *pl)
158{
159 struct page_list *next;
160
161 while (pl) {
162 next = pl->next;
163 free_pl(pl);
164 pl = next;
165 }
166}
167
168/*
169 * Allocate and reserve nr_pages for the use of a specific client.
170 */
171static int client_reserve_pages(struct dm_kcopyd_client *kc, unsigned nr_pages)
172{
173 unsigned i;
174 struct page_list *pl = NULL, *next;
175
176 for (i = 0; i < nr_pages; i++) {
177 next = alloc_pl(GFP_KERNEL);
178 if (!next) {
179 if (pl)
180 drop_pages(pl);
181 return -ENOMEM;
182 }
183 next->next = pl;
184 pl = next;
185 }
186
187 kc->nr_reserved_pages += nr_pages;
188 kcopyd_put_pages(kc, pl);
189
190 return 0;
191}
192
193static void client_free_pages(struct dm_kcopyd_client *kc)
194{
195 BUG_ON(kc->nr_free_pages != kc->nr_reserved_pages);
196 drop_pages(kc->pages);
197 kc->pages = NULL;
198 kc->nr_free_pages = kc->nr_reserved_pages = 0;
199}
200
201/*-----------------------------------------------------------------
202 * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
203 * for this reason we use a mempool to prevent the client from
204 * ever having to do io (which could cause a deadlock).
205 *---------------------------------------------------------------*/
206struct kcopyd_job {
207 struct dm_kcopyd_client *kc;
208 struct list_head list;
209 unsigned long flags;
210
211 /*
212 * Error state of the job.
213 */
214 int read_err;
215 unsigned long write_err;
216
217 /*
218 * Either READ or WRITE
219 */
220 int rw;
221 struct dm_io_region source;
222
223 /*
224 * The destinations for the transfer.
225 */
226 unsigned int num_dests;
227 struct dm_io_region dests[DM_KCOPYD_MAX_REGIONS];
228
229 struct page_list *pages;
230
231 /*
232 * Set this to ensure you are notified when the job has
233 * completed. 'context' is for callback to use.
234 */
235 dm_kcopyd_notify_fn fn;
236 void *context;
237
238 /*
239 * These fields are only used if the job has been split
240 * into more manageable parts.
241 */
242 struct mutex lock;
243 atomic_t sub_jobs;
244 sector_t progress;
245
246 struct kcopyd_job *master_job;
247};
248
249static struct kmem_cache *_job_cache;
250
251int __init dm_kcopyd_init(void)
252{
253 _job_cache = kmem_cache_create("kcopyd_job",
254 sizeof(struct kcopyd_job) * (SPLIT_COUNT + 1),
255 __alignof__(struct kcopyd_job), 0, NULL);
256 if (!_job_cache)
257 return -ENOMEM;
258
259 zero_page_list.next = &zero_page_list;
260 zero_page_list.page = ZERO_PAGE(0);
261
262 return 0;
263}
264
265void dm_kcopyd_exit(void)
266{
267 kmem_cache_destroy(_job_cache);
268 _job_cache = NULL;
269}
270
271/*
272 * Functions to push and pop a job onto the head of a given job
273 * list.
274 */
275static struct kcopyd_job *pop(struct list_head *jobs,
276 struct dm_kcopyd_client *kc)
277{
278 struct kcopyd_job *job = NULL;
279 unsigned long flags;
280
281 spin_lock_irqsave(&kc->job_lock, flags);
282
283 if (!list_empty(jobs)) {
284 job = list_entry(jobs->next, struct kcopyd_job, list);
285 list_del(&job->list);
286 }
287 spin_unlock_irqrestore(&kc->job_lock, flags);
288
289 return job;
290}
291
292static void push(struct list_head *jobs, struct kcopyd_job *job)
293{
294 unsigned long flags;
295 struct dm_kcopyd_client *kc = job->kc;
296
297 spin_lock_irqsave(&kc->job_lock, flags);
298 list_add_tail(&job->list, jobs);
299 spin_unlock_irqrestore(&kc->job_lock, flags);
300}
301
302
303static void push_head(struct list_head *jobs, struct kcopyd_job *job)
304{
305 unsigned long flags;
306 struct dm_kcopyd_client *kc = job->kc;
307
308 spin_lock_irqsave(&kc->job_lock, flags);
309 list_add(&job->list, jobs);
310 spin_unlock_irqrestore(&kc->job_lock, flags);
311}
312
313/*
314 * These three functions process 1 item from the corresponding
315 * job list.
316 *
317 * They return:
318 * < 0: error
319 * 0: success
320 * > 0: can't process yet.
321 */
322static int run_complete_job(struct kcopyd_job *job)
323{
324 void *context = job->context;
325 int read_err = job->read_err;
326 unsigned long write_err = job->write_err;
327 dm_kcopyd_notify_fn fn = job->fn;
328 struct dm_kcopyd_client *kc = job->kc;
329
330 if (job->pages && job->pages != &zero_page_list)
331 kcopyd_put_pages(kc, job->pages);
332 /*
333 * If this is the master job, the sub jobs have already
334 * completed so we can free everything.
335 */
336 if (job->master_job == job)
337 mempool_free(job, kc->job_pool);
338 fn(read_err, write_err, context);
339
340 if (atomic_dec_and_test(&kc->nr_jobs))
341 wake_up(&kc->destroyq);
342
343 return 0;
344}
345
346static void complete_io(unsigned long error, void *context)
347{
348 struct kcopyd_job *job = (struct kcopyd_job *) context;
349 struct dm_kcopyd_client *kc = job->kc;
350
351 if (error) {
352 if (job->rw == WRITE)
353 job->write_err |= error;
354 else
355 job->read_err = 1;
356
357 if (!test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
358 push(&kc->complete_jobs, job);
359 wake(kc);
360 return;
361 }
362 }
363
364 if (job->rw == WRITE)
365 push(&kc->complete_jobs, job);
366
367 else {
368 job->rw = WRITE;
369 push(&kc->io_jobs, job);
370 }
371
372 wake(kc);
373}
374
375/*
376 * Request io on as many buffer heads as we can currently get for
377 * a particular job.
378 */
379static int run_io_job(struct kcopyd_job *job)
380{
381 int r;
382 struct dm_io_request io_req = {
383 .bi_rw = job->rw,
384 .mem.type = DM_IO_PAGE_LIST,
385 .mem.ptr.pl = job->pages,
386 .mem.offset = 0,
387 .notify.fn = complete_io,
388 .notify.context = job,
389 .client = job->kc->io_client,
390 };
391
392 if (job->rw == READ)
393 r = dm_io(&io_req, 1, &job->source, NULL);
394 else
395 r = dm_io(&io_req, job->num_dests, job->dests, NULL);
396
397 return r;
398}
399
400static int run_pages_job(struct kcopyd_job *job)
401{
402 int r;
403 unsigned nr_pages = dm_div_up(job->dests[0].count, PAGE_SIZE >> 9);
404
405 r = kcopyd_get_pages(job->kc, nr_pages, &job->pages);
406 if (!r) {
407 /* this job is ready for io */
408 push(&job->kc->io_jobs, job);
409 return 0;
410 }
411
412 if (r == -ENOMEM)
413 /* can't complete now */
414 return 1;
415
416 return r;
417}
418
419/*
420 * Run through a list for as long as possible. Returns the count
421 * of successful jobs.
422 */
423static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
424 int (*fn) (struct kcopyd_job *))
425{
426 struct kcopyd_job *job;
427 int r, count = 0;
428
429 while ((job = pop(jobs, kc))) {
430
431 r = fn(job);
432
433 if (r < 0) {
434 /* error this rogue job */
435 if (job->rw == WRITE)
436 job->write_err = (unsigned long) -1L;
437 else
438 job->read_err = 1;
439 push(&kc->complete_jobs, job);
440 break;
441 }
442
443 if (r > 0) {
444 /*
445 * We couldn't service this job ATM, so
446 * push this job back onto the list.
447 */
448 push_head(jobs, job);
449 break;
450 }
451
452 count++;
453 }
454
455 return count;
456}
457
458/*
459 * kcopyd does this every time it's woken up.
460 */
461static void do_work(struct work_struct *work)
462{
463 struct dm_kcopyd_client *kc = container_of(work,
464 struct dm_kcopyd_client, kcopyd_work);
465 struct blk_plug plug;
466
467 /*
468 * The order that these are called is *very* important.
469 * complete jobs can free some pages for pages jobs.
470 * Pages jobs when successful will jump onto the io jobs
471 * list. io jobs call wake when they complete and it all
472 * starts again.
473 */
474 blk_start_plug(&plug);
475 process_jobs(&kc->complete_jobs, kc, run_complete_job);
476 process_jobs(&kc->pages_jobs, kc, run_pages_job);
477 process_jobs(&kc->io_jobs, kc, run_io_job);
478 blk_finish_plug(&plug);
479}
480
481/*
482 * If we are copying a small region we just dispatch a single job
483 * to do the copy, otherwise the io has to be split up into many
484 * jobs.
485 */
486static void dispatch_job(struct kcopyd_job *job)
487{
488 struct dm_kcopyd_client *kc = job->kc;
489 atomic_inc(&kc->nr_jobs);
490 if (unlikely(!job->source.count))
491 push(&kc->complete_jobs, job);
492 else if (job->pages == &zero_page_list)
493 push(&kc->io_jobs, job);
494 else
495 push(&kc->pages_jobs, job);
496 wake(kc);
497}
498
499static void segment_complete(int read_err, unsigned long write_err,
500 void *context)
501{
502 /* FIXME: tidy this function */
503 sector_t progress = 0;
504 sector_t count = 0;
505 struct kcopyd_job *sub_job = (struct kcopyd_job *) context;
506 struct kcopyd_job *job = sub_job->master_job;
507 struct dm_kcopyd_client *kc = job->kc;
508
509 mutex_lock(&job->lock);
510
511 /* update the error */
512 if (read_err)
513 job->read_err = 1;
514
515 if (write_err)
516 job->write_err |= write_err;
517
518 /*
519 * Only dispatch more work if there hasn't been an error.
520 */
521 if ((!job->read_err && !job->write_err) ||
522 test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
523 /* get the next chunk of work */
524 progress = job->progress;
525 count = job->source.count - progress;
526 if (count) {
527 if (count > SUB_JOB_SIZE)
528 count = SUB_JOB_SIZE;
529
530 job->progress += count;
531 }
532 }
533 mutex_unlock(&job->lock);
534
535 if (count) {
536 int i;
537
538 *sub_job = *job;
539 sub_job->source.sector += progress;
540 sub_job->source.count = count;
541
542 for (i = 0; i < job->num_dests; i++) {
543 sub_job->dests[i].sector += progress;
544 sub_job->dests[i].count = count;
545 }
546
547 sub_job->fn = segment_complete;
548 sub_job->context = sub_job;
549 dispatch_job(sub_job);
550
551 } else if (atomic_dec_and_test(&job->sub_jobs)) {
552
553 /*
554 * Queue the completion callback to the kcopyd thread.
555 *
556 * Some callers assume that all the completions are called
557 * from a single thread and don't race with each other.
558 *
559 * We must not call the callback directly here because this
560 * code may not be executing in the thread.
561 */
562 push(&kc->complete_jobs, job);
563 wake(kc);
564 }
565}
566
567/*
568 * Create some sub jobs to share the work between them.
569 */
570static void split_job(struct kcopyd_job *master_job)
571{
572 int i;
573
574 atomic_inc(&master_job->kc->nr_jobs);
575
576 atomic_set(&master_job->sub_jobs, SPLIT_COUNT);
577 for (i = 0; i < SPLIT_COUNT; i++) {
578 master_job[i + 1].master_job = master_job;
579 segment_complete(0, 0u, &master_job[i + 1]);
580 }
581}
582
583int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
584 unsigned int num_dests, struct dm_io_region *dests,
585 unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
586{
587 struct kcopyd_job *job;
588
589 /*
590 * Allocate an array of jobs consisting of one master job
591 * followed by SPLIT_COUNT sub jobs.
592 */
593 job = mempool_alloc(kc->job_pool, GFP_NOIO);
594
595 /*
596 * set up for the read.
597 */
598 job->kc = kc;
599 job->flags = flags;
600 job->read_err = 0;
601 job->write_err = 0;
602
603 job->num_dests = num_dests;
604 memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
605
606 if (from) {
607 job->source = *from;
608 job->pages = NULL;
609 job->rw = READ;
610 } else {
611 memset(&job->source, 0, sizeof job->source);
612 job->source.count = job->dests[0].count;
613 job->pages = &zero_page_list;
614 job->rw = WRITE;
615 }
616
617 job->fn = fn;
618 job->context = context;
619 job->master_job = job;
620
621 if (job->source.count <= SUB_JOB_SIZE)
622 dispatch_job(job);
623 else {
624 mutex_init(&job->lock);
625 job->progress = 0;
626 split_job(job);
627 }
628
629 return 0;
630}
631EXPORT_SYMBOL(dm_kcopyd_copy);
632
633int dm_kcopyd_zero(struct dm_kcopyd_client *kc,
634 unsigned num_dests, struct dm_io_region *dests,
635 unsigned flags, dm_kcopyd_notify_fn fn, void *context)
636{
637 return dm_kcopyd_copy(kc, NULL, num_dests, dests, flags, fn, context);
638}
639EXPORT_SYMBOL(dm_kcopyd_zero);
640
641void *dm_kcopyd_prepare_callback(struct dm_kcopyd_client *kc,
642 dm_kcopyd_notify_fn fn, void *context)
643{
644 struct kcopyd_job *job;
645
646 job = mempool_alloc(kc->job_pool, GFP_NOIO);
647
648 memset(job, 0, sizeof(struct kcopyd_job));
649 job->kc = kc;
650 job->fn = fn;
651 job->context = context;
652 job->master_job = job;
653
654 atomic_inc(&kc->nr_jobs);
655
656 return job;
657}
658EXPORT_SYMBOL(dm_kcopyd_prepare_callback);
659
660void dm_kcopyd_do_callback(void *j, int read_err, unsigned long write_err)
661{
662 struct kcopyd_job *job = j;
663 struct dm_kcopyd_client *kc = job->kc;
664
665 job->read_err = read_err;
666 job->write_err = write_err;
667
668 push(&kc->complete_jobs, job);
669 wake(kc);
670}
671EXPORT_SYMBOL(dm_kcopyd_do_callback);
672
673/*
674 * Cancels a kcopyd job, eg. someone might be deactivating a
675 * mirror.
676 */
677#if 0
678int kcopyd_cancel(struct kcopyd_job *job, int block)
679{
680 /* FIXME: finish */
681 return -1;
682}
683#endif /* 0 */
684
685/*-----------------------------------------------------------------
686 * Client setup
687 *---------------------------------------------------------------*/
688struct dm_kcopyd_client *dm_kcopyd_client_create(void)
689{
690 int r = -ENOMEM;
691 struct dm_kcopyd_client *kc;
692
693 kc = kmalloc(sizeof(*kc), GFP_KERNEL);
694 if (!kc)
695 return ERR_PTR(-ENOMEM);
696
697 spin_lock_init(&kc->job_lock);
698 INIT_LIST_HEAD(&kc->complete_jobs);
699 INIT_LIST_HEAD(&kc->io_jobs);
700 INIT_LIST_HEAD(&kc->pages_jobs);
701
702 kc->job_pool = mempool_create_slab_pool(MIN_JOBS, _job_cache);
703 if (!kc->job_pool)
704 goto bad_slab;
705
706 INIT_WORK(&kc->kcopyd_work, do_work);
707 kc->kcopyd_wq = alloc_workqueue("kcopyd",
708 WQ_NON_REENTRANT | WQ_MEM_RECLAIM, 0);
709 if (!kc->kcopyd_wq)
710 goto bad_workqueue;
711
712 kc->pages = NULL;
713 kc->nr_reserved_pages = kc->nr_free_pages = 0;
714 r = client_reserve_pages(kc, RESERVE_PAGES);
715 if (r)
716 goto bad_client_pages;
717
718 kc->io_client = dm_io_client_create();
719 if (IS_ERR(kc->io_client)) {
720 r = PTR_ERR(kc->io_client);
721 goto bad_io_client;
722 }
723
724 init_waitqueue_head(&kc->destroyq);
725 atomic_set(&kc->nr_jobs, 0);
726
727 return kc;
728
729bad_io_client:
730 client_free_pages(kc);
731bad_client_pages:
732 destroy_workqueue(kc->kcopyd_wq);
733bad_workqueue:
734 mempool_destroy(kc->job_pool);
735bad_slab:
736 kfree(kc);
737
738 return ERR_PTR(r);
739}
740EXPORT_SYMBOL(dm_kcopyd_client_create);
741
742void dm_kcopyd_client_destroy(struct dm_kcopyd_client *kc)
743{
744 /* Wait for completion of all jobs submitted by this client. */
745 wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
746
747 BUG_ON(!list_empty(&kc->complete_jobs));
748 BUG_ON(!list_empty(&kc->io_jobs));
749 BUG_ON(!list_empty(&kc->pages_jobs));
750 destroy_workqueue(kc->kcopyd_wq);
751 dm_io_client_destroy(kc->io_client);
752 client_free_pages(kc);
753 mempool_destroy(kc->job_pool);
754 kfree(kc);
755}
756EXPORT_SYMBOL(dm_kcopyd_client_destroy);
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/device-mapper.h>
26#include <linux/dm-kcopyd.h>
27
28#include "dm.h"
29
30#define SUB_JOB_SIZE 128
31#define SPLIT_COUNT 8
32#define MIN_JOBS 8
33#define RESERVE_PAGES (DIV_ROUND_UP(SUB_JOB_SIZE << SECTOR_SHIFT, PAGE_SIZE))
34
35/*-----------------------------------------------------------------
36 * Each kcopyd client has its own little pool of preallocated
37 * pages for kcopyd io.
38 *---------------------------------------------------------------*/
39struct dm_kcopyd_client {
40 struct page_list *pages;
41 unsigned nr_reserved_pages;
42 unsigned nr_free_pages;
43
44 struct dm_io_client *io_client;
45
46 wait_queue_head_t destroyq;
47 atomic_t nr_jobs;
48
49 mempool_t *job_pool;
50
51 struct workqueue_struct *kcopyd_wq;
52 struct work_struct kcopyd_work;
53
54/*
55 * We maintain three lists of jobs:
56 *
57 * i) jobs waiting for pages
58 * ii) jobs that have pages, and are waiting for the io to be issued.
59 * iii) jobs that have completed.
60 *
61 * All three of these are protected by job_lock.
62 */
63 spinlock_t job_lock;
64 struct list_head complete_jobs;
65 struct list_head io_jobs;
66 struct list_head pages_jobs;
67};
68
69static void wake(struct dm_kcopyd_client *kc)
70{
71 queue_work(kc->kcopyd_wq, &kc->kcopyd_work);
72}
73
74/*
75 * Obtain one page for the use of kcopyd.
76 */
77static struct page_list *alloc_pl(gfp_t gfp)
78{
79 struct page_list *pl;
80
81 pl = kmalloc(sizeof(*pl), gfp);
82 if (!pl)
83 return NULL;
84
85 pl->page = alloc_page(gfp);
86 if (!pl->page) {
87 kfree(pl);
88 return NULL;
89 }
90
91 return pl;
92}
93
94static void free_pl(struct page_list *pl)
95{
96 __free_page(pl->page);
97 kfree(pl);
98}
99
100/*
101 * Add the provided pages to a client's free page list, releasing
102 * back to the system any beyond the reserved_pages limit.
103 */
104static void kcopyd_put_pages(struct dm_kcopyd_client *kc, struct page_list *pl)
105{
106 struct page_list *next;
107
108 do {
109 next = pl->next;
110
111 if (kc->nr_free_pages >= kc->nr_reserved_pages)
112 free_pl(pl);
113 else {
114 pl->next = kc->pages;
115 kc->pages = pl;
116 kc->nr_free_pages++;
117 }
118
119 pl = next;
120 } while (pl);
121}
122
123static int kcopyd_get_pages(struct dm_kcopyd_client *kc,
124 unsigned int nr, struct page_list **pages)
125{
126 struct page_list *pl;
127
128 *pages = NULL;
129
130 do {
131 pl = alloc_pl(__GFP_NOWARN | __GFP_NORETRY);
132 if (unlikely(!pl)) {
133 /* Use reserved pages */
134 pl = kc->pages;
135 if (unlikely(!pl))
136 goto out_of_memory;
137 kc->pages = pl->next;
138 kc->nr_free_pages--;
139 }
140 pl->next = *pages;
141 *pages = pl;
142 } while (--nr);
143
144 return 0;
145
146out_of_memory:
147 if (*pages)
148 kcopyd_put_pages(kc, *pages);
149 return -ENOMEM;
150}
151
152/*
153 * These three functions resize the page pool.
154 */
155static void drop_pages(struct page_list *pl)
156{
157 struct page_list *next;
158
159 while (pl) {
160 next = pl->next;
161 free_pl(pl);
162 pl = next;
163 }
164}
165
166/*
167 * Allocate and reserve nr_pages for the use of a specific client.
168 */
169static int client_reserve_pages(struct dm_kcopyd_client *kc, unsigned nr_pages)
170{
171 unsigned i;
172 struct page_list *pl = NULL, *next;
173
174 for (i = 0; i < nr_pages; i++) {
175 next = alloc_pl(GFP_KERNEL);
176 if (!next) {
177 if (pl)
178 drop_pages(pl);
179 return -ENOMEM;
180 }
181 next->next = pl;
182 pl = next;
183 }
184
185 kc->nr_reserved_pages += nr_pages;
186 kcopyd_put_pages(kc, pl);
187
188 return 0;
189}
190
191static void client_free_pages(struct dm_kcopyd_client *kc)
192{
193 BUG_ON(kc->nr_free_pages != kc->nr_reserved_pages);
194 drop_pages(kc->pages);
195 kc->pages = NULL;
196 kc->nr_free_pages = kc->nr_reserved_pages = 0;
197}
198
199/*-----------------------------------------------------------------
200 * kcopyd_jobs need to be allocated by the *clients* of kcopyd,
201 * for this reason we use a mempool to prevent the client from
202 * ever having to do io (which could cause a deadlock).
203 *---------------------------------------------------------------*/
204struct kcopyd_job {
205 struct dm_kcopyd_client *kc;
206 struct list_head list;
207 unsigned long flags;
208
209 /*
210 * Error state of the job.
211 */
212 int read_err;
213 unsigned long write_err;
214
215 /*
216 * Either READ or WRITE
217 */
218 int rw;
219 struct dm_io_region source;
220
221 /*
222 * The destinations for the transfer.
223 */
224 unsigned int num_dests;
225 struct dm_io_region dests[DM_KCOPYD_MAX_REGIONS];
226
227 struct page_list *pages;
228
229 /*
230 * Set this to ensure you are notified when the job has
231 * completed. 'context' is for callback to use.
232 */
233 dm_kcopyd_notify_fn fn;
234 void *context;
235
236 /*
237 * These fields are only used if the job has been split
238 * into more manageable parts.
239 */
240 struct mutex lock;
241 atomic_t sub_jobs;
242 sector_t progress;
243
244 struct kcopyd_job *master_job;
245};
246
247static struct kmem_cache *_job_cache;
248
249int __init dm_kcopyd_init(void)
250{
251 _job_cache = kmem_cache_create("kcopyd_job",
252 sizeof(struct kcopyd_job) * (SPLIT_COUNT + 1),
253 __alignof__(struct kcopyd_job), 0, NULL);
254 if (!_job_cache)
255 return -ENOMEM;
256
257 return 0;
258}
259
260void dm_kcopyd_exit(void)
261{
262 kmem_cache_destroy(_job_cache);
263 _job_cache = NULL;
264}
265
266/*
267 * Functions to push and pop a job onto the head of a given job
268 * list.
269 */
270static struct kcopyd_job *pop(struct list_head *jobs,
271 struct dm_kcopyd_client *kc)
272{
273 struct kcopyd_job *job = NULL;
274 unsigned long flags;
275
276 spin_lock_irqsave(&kc->job_lock, flags);
277
278 if (!list_empty(jobs)) {
279 job = list_entry(jobs->next, struct kcopyd_job, list);
280 list_del(&job->list);
281 }
282 spin_unlock_irqrestore(&kc->job_lock, flags);
283
284 return job;
285}
286
287static void push(struct list_head *jobs, struct kcopyd_job *job)
288{
289 unsigned long flags;
290 struct dm_kcopyd_client *kc = job->kc;
291
292 spin_lock_irqsave(&kc->job_lock, flags);
293 list_add_tail(&job->list, jobs);
294 spin_unlock_irqrestore(&kc->job_lock, flags);
295}
296
297
298static void push_head(struct list_head *jobs, struct kcopyd_job *job)
299{
300 unsigned long flags;
301 struct dm_kcopyd_client *kc = job->kc;
302
303 spin_lock_irqsave(&kc->job_lock, flags);
304 list_add(&job->list, jobs);
305 spin_unlock_irqrestore(&kc->job_lock, flags);
306}
307
308/*
309 * These three functions process 1 item from the corresponding
310 * job list.
311 *
312 * They return:
313 * < 0: error
314 * 0: success
315 * > 0: can't process yet.
316 */
317static int run_complete_job(struct kcopyd_job *job)
318{
319 void *context = job->context;
320 int read_err = job->read_err;
321 unsigned long write_err = job->write_err;
322 dm_kcopyd_notify_fn fn = job->fn;
323 struct dm_kcopyd_client *kc = job->kc;
324
325 if (job->pages)
326 kcopyd_put_pages(kc, job->pages);
327 /*
328 * If this is the master job, the sub jobs have already
329 * completed so we can free everything.
330 */
331 if (job->master_job == job)
332 mempool_free(job, kc->job_pool);
333 fn(read_err, write_err, context);
334
335 if (atomic_dec_and_test(&kc->nr_jobs))
336 wake_up(&kc->destroyq);
337
338 return 0;
339}
340
341static void complete_io(unsigned long error, void *context)
342{
343 struct kcopyd_job *job = (struct kcopyd_job *) context;
344 struct dm_kcopyd_client *kc = job->kc;
345
346 if (error) {
347 if (job->rw == WRITE)
348 job->write_err |= error;
349 else
350 job->read_err = 1;
351
352 if (!test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
353 push(&kc->complete_jobs, job);
354 wake(kc);
355 return;
356 }
357 }
358
359 if (job->rw == WRITE)
360 push(&kc->complete_jobs, job);
361
362 else {
363 job->rw = WRITE;
364 push(&kc->io_jobs, job);
365 }
366
367 wake(kc);
368}
369
370/*
371 * Request io on as many buffer heads as we can currently get for
372 * a particular job.
373 */
374static int run_io_job(struct kcopyd_job *job)
375{
376 int r;
377 struct dm_io_request io_req = {
378 .bi_rw = job->rw,
379 .mem.type = DM_IO_PAGE_LIST,
380 .mem.ptr.pl = job->pages,
381 .mem.offset = 0,
382 .notify.fn = complete_io,
383 .notify.context = job,
384 .client = job->kc->io_client,
385 };
386
387 if (job->rw == READ)
388 r = dm_io(&io_req, 1, &job->source, NULL);
389 else
390 r = dm_io(&io_req, job->num_dests, job->dests, NULL);
391
392 return r;
393}
394
395static int run_pages_job(struct kcopyd_job *job)
396{
397 int r;
398 unsigned nr_pages = dm_div_up(job->dests[0].count, PAGE_SIZE >> 9);
399
400 r = kcopyd_get_pages(job->kc, nr_pages, &job->pages);
401 if (!r) {
402 /* this job is ready for io */
403 push(&job->kc->io_jobs, job);
404 return 0;
405 }
406
407 if (r == -ENOMEM)
408 /* can't complete now */
409 return 1;
410
411 return r;
412}
413
414/*
415 * Run through a list for as long as possible. Returns the count
416 * of successful jobs.
417 */
418static int process_jobs(struct list_head *jobs, struct dm_kcopyd_client *kc,
419 int (*fn) (struct kcopyd_job *))
420{
421 struct kcopyd_job *job;
422 int r, count = 0;
423
424 while ((job = pop(jobs, kc))) {
425
426 r = fn(job);
427
428 if (r < 0) {
429 /* error this rogue job */
430 if (job->rw == WRITE)
431 job->write_err = (unsigned long) -1L;
432 else
433 job->read_err = 1;
434 push(&kc->complete_jobs, job);
435 break;
436 }
437
438 if (r > 0) {
439 /*
440 * We couldn't service this job ATM, so
441 * push this job back onto the list.
442 */
443 push_head(jobs, job);
444 break;
445 }
446
447 count++;
448 }
449
450 return count;
451}
452
453/*
454 * kcopyd does this every time it's woken up.
455 */
456static void do_work(struct work_struct *work)
457{
458 struct dm_kcopyd_client *kc = container_of(work,
459 struct dm_kcopyd_client, kcopyd_work);
460 struct blk_plug plug;
461
462 /*
463 * The order that these are called is *very* important.
464 * complete jobs can free some pages for pages jobs.
465 * Pages jobs when successful will jump onto the io jobs
466 * list. io jobs call wake when they complete and it all
467 * starts again.
468 */
469 blk_start_plug(&plug);
470 process_jobs(&kc->complete_jobs, kc, run_complete_job);
471 process_jobs(&kc->pages_jobs, kc, run_pages_job);
472 process_jobs(&kc->io_jobs, kc, run_io_job);
473 blk_finish_plug(&plug);
474}
475
476/*
477 * If we are copying a small region we just dispatch a single job
478 * to do the copy, otherwise the io has to be split up into many
479 * jobs.
480 */
481static void dispatch_job(struct kcopyd_job *job)
482{
483 struct dm_kcopyd_client *kc = job->kc;
484 atomic_inc(&kc->nr_jobs);
485 if (unlikely(!job->source.count))
486 push(&kc->complete_jobs, job);
487 else
488 push(&kc->pages_jobs, job);
489 wake(kc);
490}
491
492static void segment_complete(int read_err, unsigned long write_err,
493 void *context)
494{
495 /* FIXME: tidy this function */
496 sector_t progress = 0;
497 sector_t count = 0;
498 struct kcopyd_job *sub_job = (struct kcopyd_job *) context;
499 struct kcopyd_job *job = sub_job->master_job;
500 struct dm_kcopyd_client *kc = job->kc;
501
502 mutex_lock(&job->lock);
503
504 /* update the error */
505 if (read_err)
506 job->read_err = 1;
507
508 if (write_err)
509 job->write_err |= write_err;
510
511 /*
512 * Only dispatch more work if there hasn't been an error.
513 */
514 if ((!job->read_err && !job->write_err) ||
515 test_bit(DM_KCOPYD_IGNORE_ERROR, &job->flags)) {
516 /* get the next chunk of work */
517 progress = job->progress;
518 count = job->source.count - progress;
519 if (count) {
520 if (count > SUB_JOB_SIZE)
521 count = SUB_JOB_SIZE;
522
523 job->progress += count;
524 }
525 }
526 mutex_unlock(&job->lock);
527
528 if (count) {
529 int i;
530
531 *sub_job = *job;
532 sub_job->source.sector += progress;
533 sub_job->source.count = count;
534
535 for (i = 0; i < job->num_dests; i++) {
536 sub_job->dests[i].sector += progress;
537 sub_job->dests[i].count = count;
538 }
539
540 sub_job->fn = segment_complete;
541 sub_job->context = sub_job;
542 dispatch_job(sub_job);
543
544 } else if (atomic_dec_and_test(&job->sub_jobs)) {
545
546 /*
547 * Queue the completion callback to the kcopyd thread.
548 *
549 * Some callers assume that all the completions are called
550 * from a single thread and don't race with each other.
551 *
552 * We must not call the callback directly here because this
553 * code may not be executing in the thread.
554 */
555 push(&kc->complete_jobs, job);
556 wake(kc);
557 }
558}
559
560/*
561 * Create some sub jobs to share the work between them.
562 */
563static void split_job(struct kcopyd_job *master_job)
564{
565 int i;
566
567 atomic_inc(&master_job->kc->nr_jobs);
568
569 atomic_set(&master_job->sub_jobs, SPLIT_COUNT);
570 for (i = 0; i < SPLIT_COUNT; i++) {
571 master_job[i + 1].master_job = master_job;
572 segment_complete(0, 0u, &master_job[i + 1]);
573 }
574}
575
576int dm_kcopyd_copy(struct dm_kcopyd_client *kc, struct dm_io_region *from,
577 unsigned int num_dests, struct dm_io_region *dests,
578 unsigned int flags, dm_kcopyd_notify_fn fn, void *context)
579{
580 struct kcopyd_job *job;
581
582 /*
583 * Allocate an array of jobs consisting of one master job
584 * followed by SPLIT_COUNT sub jobs.
585 */
586 job = mempool_alloc(kc->job_pool, GFP_NOIO);
587
588 /*
589 * set up for the read.
590 */
591 job->kc = kc;
592 job->flags = flags;
593 job->read_err = 0;
594 job->write_err = 0;
595 job->rw = READ;
596
597 job->source = *from;
598
599 job->num_dests = num_dests;
600 memcpy(&job->dests, dests, sizeof(*dests) * num_dests);
601
602 job->pages = NULL;
603
604 job->fn = fn;
605 job->context = context;
606 job->master_job = job;
607
608 if (job->source.count <= SUB_JOB_SIZE)
609 dispatch_job(job);
610 else {
611 mutex_init(&job->lock);
612 job->progress = 0;
613 split_job(job);
614 }
615
616 return 0;
617}
618EXPORT_SYMBOL(dm_kcopyd_copy);
619
620void *dm_kcopyd_prepare_callback(struct dm_kcopyd_client *kc,
621 dm_kcopyd_notify_fn fn, void *context)
622{
623 struct kcopyd_job *job;
624
625 job = mempool_alloc(kc->job_pool, GFP_NOIO);
626
627 memset(job, 0, sizeof(struct kcopyd_job));
628 job->kc = kc;
629 job->fn = fn;
630 job->context = context;
631 job->master_job = job;
632
633 atomic_inc(&kc->nr_jobs);
634
635 return job;
636}
637EXPORT_SYMBOL(dm_kcopyd_prepare_callback);
638
639void dm_kcopyd_do_callback(void *j, int read_err, unsigned long write_err)
640{
641 struct kcopyd_job *job = j;
642 struct dm_kcopyd_client *kc = job->kc;
643
644 job->read_err = read_err;
645 job->write_err = write_err;
646
647 push(&kc->complete_jobs, job);
648 wake(kc);
649}
650EXPORT_SYMBOL(dm_kcopyd_do_callback);
651
652/*
653 * Cancels a kcopyd job, eg. someone might be deactivating a
654 * mirror.
655 */
656#if 0
657int kcopyd_cancel(struct kcopyd_job *job, int block)
658{
659 /* FIXME: finish */
660 return -1;
661}
662#endif /* 0 */
663
664/*-----------------------------------------------------------------
665 * Client setup
666 *---------------------------------------------------------------*/
667struct dm_kcopyd_client *dm_kcopyd_client_create(void)
668{
669 int r = -ENOMEM;
670 struct dm_kcopyd_client *kc;
671
672 kc = kmalloc(sizeof(*kc), GFP_KERNEL);
673 if (!kc)
674 return ERR_PTR(-ENOMEM);
675
676 spin_lock_init(&kc->job_lock);
677 INIT_LIST_HEAD(&kc->complete_jobs);
678 INIT_LIST_HEAD(&kc->io_jobs);
679 INIT_LIST_HEAD(&kc->pages_jobs);
680
681 kc->job_pool = mempool_create_slab_pool(MIN_JOBS, _job_cache);
682 if (!kc->job_pool)
683 goto bad_slab;
684
685 INIT_WORK(&kc->kcopyd_work, do_work);
686 kc->kcopyd_wq = alloc_workqueue("kcopyd",
687 WQ_NON_REENTRANT | WQ_MEM_RECLAIM, 0);
688 if (!kc->kcopyd_wq)
689 goto bad_workqueue;
690
691 kc->pages = NULL;
692 kc->nr_reserved_pages = kc->nr_free_pages = 0;
693 r = client_reserve_pages(kc, RESERVE_PAGES);
694 if (r)
695 goto bad_client_pages;
696
697 kc->io_client = dm_io_client_create();
698 if (IS_ERR(kc->io_client)) {
699 r = PTR_ERR(kc->io_client);
700 goto bad_io_client;
701 }
702
703 init_waitqueue_head(&kc->destroyq);
704 atomic_set(&kc->nr_jobs, 0);
705
706 return kc;
707
708bad_io_client:
709 client_free_pages(kc);
710bad_client_pages:
711 destroy_workqueue(kc->kcopyd_wq);
712bad_workqueue:
713 mempool_destroy(kc->job_pool);
714bad_slab:
715 kfree(kc);
716
717 return ERR_PTR(r);
718}
719EXPORT_SYMBOL(dm_kcopyd_client_create);
720
721void dm_kcopyd_client_destroy(struct dm_kcopyd_client *kc)
722{
723 /* Wait for completion of all jobs submitted by this client. */
724 wait_event(kc->destroyq, !atomic_read(&kc->nr_jobs));
725
726 BUG_ON(!list_empty(&kc->complete_jobs));
727 BUG_ON(!list_empty(&kc->io_jobs));
728 BUG_ON(!list_empty(&kc->pages_jobs));
729 destroy_workqueue(kc->kcopyd_wq);
730 dm_io_client_destroy(kc->io_client);
731 client_free_pages(kc);
732 mempool_destroy(kc->job_pool);
733 kfree(kc);
734}
735EXPORT_SYMBOL(dm_kcopyd_client_destroy);