blob: 936b34e0959fdd5cdeaa6eddc51dee888fb05447 [file] [log] [blame]
Alasdair G Kergon4db6bfe2009-01-06 03:05:17 +00001/*
2 * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
3 * Copyright (C) 2006-2008 Red Hat GmbH
4 *
5 * This file is released under the GPL.
6 */
7
8#include "dm-exception-store.h"
9#include "dm-snap.h"
10
11#include <linux/mm.h>
12#include <linux/pagemap.h>
13#include <linux/vmalloc.h>
14#include <linux/slab.h>
15#include <linux/dm-io.h>
16
17#define DM_MSG_PREFIX "persistent snapshot"
18#define DM_CHUNK_SIZE_DEFAULT_SECTORS 32 /* 16KB */
19
20/*-----------------------------------------------------------------
21 * Persistent snapshots, by persistent we mean that the snapshot
22 * will survive a reboot.
23 *---------------------------------------------------------------*/
24
25/*
26 * We need to store a record of which parts of the origin have
27 * been copied to the snapshot device. The snapshot code
28 * requires that we copy exception chunks to chunk aligned areas
29 * of the COW store. It makes sense therefore, to store the
30 * metadata in chunk size blocks.
31 *
32 * There is no backward or forward compatibility implemented,
33 * snapshots with different disk versions than the kernel will
34 * not be usable. It is expected that "lvcreate" will blank out
35 * the start of a fresh COW device before calling the snapshot
36 * constructor.
37 *
38 * The first chunk of the COW device just contains the header.
39 * After this there is a chunk filled with exception metadata,
40 * followed by as many exception chunks as can fit in the
41 * metadata areas.
42 *
43 * All on disk structures are in little-endian format. The end
44 * of the exceptions info is indicated by an exception with a
45 * new_chunk of 0, which is invalid since it would point to the
46 * header chunk.
47 */
48
49/*
50 * Magic for persistent snapshots: "SnAp" - Feeble isn't it.
51 */
52#define SNAP_MAGIC 0x70416e53
53
54/*
55 * The on-disk version of the metadata.
56 */
57#define SNAPSHOT_DISK_VERSION 1
58
59struct disk_header {
60 uint32_t magic;
61
62 /*
63 * Is this snapshot valid. There is no way of recovering
64 * an invalid snapshot.
65 */
66 uint32_t valid;
67
68 /*
69 * Simple, incrementing version. no backward
70 * compatibility.
71 */
72 uint32_t version;
73
74 /* In sectors */
75 uint32_t chunk_size;
76};
77
78struct disk_exception {
79 uint64_t old_chunk;
80 uint64_t new_chunk;
81};
82
83struct commit_callback {
84 void (*callback)(void *, int success);
85 void *context;
86};
87
88/*
89 * The top level structure for a persistent exception store.
90 */
91struct pstore {
92 struct dm_snapshot *snap; /* up pointer to my snapshot */
93 int version;
94 int valid;
95 uint32_t exceptions_per_area;
96
97 /*
98 * Now that we have an asynchronous kcopyd there is no
99 * need for large chunk sizes, so it wont hurt to have a
100 * whole chunks worth of metadata in memory at once.
101 */
102 void *area;
103
104 /*
105 * An area of zeros used to clear the next area.
106 */
107 void *zero_area;
108
109 /*
110 * Used to keep track of which metadata area the data in
111 * 'chunk' refers to.
112 */
113 chunk_t current_area;
114
115 /*
116 * The next free chunk for an exception.
117 */
118 chunk_t next_free;
119
120 /*
121 * The index of next free exception in the current
122 * metadata area.
123 */
124 uint32_t current_committed;
125
126 atomic_t pending_count;
127 uint32_t callback_count;
128 struct commit_callback *callbacks;
129 struct dm_io_client *io_client;
130
131 struct workqueue_struct *metadata_wq;
132};
133
134static unsigned sectors_to_pages(unsigned sectors)
135{
136 return DIV_ROUND_UP(sectors, PAGE_SIZE >> 9);
137}
138
139static int alloc_area(struct pstore *ps)
140{
141 int r = -ENOMEM;
142 size_t len;
143
144 len = ps->snap->chunk_size << SECTOR_SHIFT;
145
146 /*
147 * Allocate the chunk_size block of memory that will hold
148 * a single metadata area.
149 */
150 ps->area = vmalloc(len);
151 if (!ps->area)
152 return r;
153
154 ps->zero_area = vmalloc(len);
155 if (!ps->zero_area) {
156 vfree(ps->area);
157 return r;
158 }
159 memset(ps->zero_area, 0, len);
160
161 return 0;
162}
163
164static void free_area(struct pstore *ps)
165{
166 vfree(ps->area);
167 ps->area = NULL;
168 vfree(ps->zero_area);
169 ps->zero_area = NULL;
170}
171
172struct mdata_req {
173 struct dm_io_region *where;
174 struct dm_io_request *io_req;
175 struct work_struct work;
176 int result;
177};
178
179static void do_metadata(struct work_struct *work)
180{
181 struct mdata_req *req = container_of(work, struct mdata_req, work);
182
183 req->result = dm_io(req->io_req, 1, req->where, NULL);
184}
185
186/*
187 * Read or write a chunk aligned and sized block of data from a device.
188 */
189static int chunk_io(struct pstore *ps, chunk_t chunk, int rw, int metadata)
190{
191 struct dm_io_region where = {
192 .bdev = ps->snap->cow->bdev,
193 .sector = ps->snap->chunk_size * chunk,
194 .count = ps->snap->chunk_size,
195 };
196 struct dm_io_request io_req = {
197 .bi_rw = rw,
198 .mem.type = DM_IO_VMA,
199 .mem.ptr.vma = ps->area,
200 .client = ps->io_client,
201 .notify.fn = NULL,
202 };
203 struct mdata_req req;
204
205 if (!metadata)
206 return dm_io(&io_req, 1, &where, NULL);
207
208 req.where = &where;
209 req.io_req = &io_req;
210
211 /*
212 * Issue the synchronous I/O from a different thread
213 * to avoid generic_make_request recursion.
214 */
215 INIT_WORK(&req.work, do_metadata);
216 queue_work(ps->metadata_wq, &req.work);
217 flush_workqueue(ps->metadata_wq);
218
219 return req.result;
220}
221
222/*
223 * Convert a metadata area index to a chunk index.
224 */
225static chunk_t area_location(struct pstore *ps, chunk_t area)
226{
227 return 1 + ((ps->exceptions_per_area + 1) * area);
228}
229
230/*
231 * Read or write a metadata area. Remembering to skip the first
232 * chunk which holds the header.
233 */
234static int area_io(struct pstore *ps, int rw)
235{
236 int r;
237 chunk_t chunk;
238
239 chunk = area_location(ps, ps->current_area);
240
241 r = chunk_io(ps, chunk, rw, 0);
242 if (r)
243 return r;
244
245 return 0;
246}
247
248static void zero_memory_area(struct pstore *ps)
249{
250 memset(ps->area, 0, ps->snap->chunk_size << SECTOR_SHIFT);
251}
252
253static int zero_disk_area(struct pstore *ps, chunk_t area)
254{
255 struct dm_io_region where = {
256 .bdev = ps->snap->cow->bdev,
257 .sector = ps->snap->chunk_size * area_location(ps, area),
258 .count = ps->snap->chunk_size,
259 };
260 struct dm_io_request io_req = {
261 .bi_rw = WRITE,
262 .mem.type = DM_IO_VMA,
263 .mem.ptr.vma = ps->zero_area,
264 .client = ps->io_client,
265 .notify.fn = NULL,
266 };
267
268 return dm_io(&io_req, 1, &where, NULL);
269}
270
271static int read_header(struct pstore *ps, int *new_snapshot)
272{
273 int r;
274 struct disk_header *dh;
275 chunk_t chunk_size;
276 int chunk_size_supplied = 1;
277
278 /*
279 * Use default chunk size (or hardsect_size, if larger) if none supplied
280 */
281 if (!ps->snap->chunk_size) {
282 ps->snap->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
283 bdev_hardsect_size(ps->snap->cow->bdev) >> 9);
284 ps->snap->chunk_mask = ps->snap->chunk_size - 1;
285 ps->snap->chunk_shift = ffs(ps->snap->chunk_size) - 1;
286 chunk_size_supplied = 0;
287 }
288
289 ps->io_client = dm_io_client_create(sectors_to_pages(ps->snap->
290 chunk_size));
291 if (IS_ERR(ps->io_client))
292 return PTR_ERR(ps->io_client);
293
294 r = alloc_area(ps);
295 if (r)
296 return r;
297
298 r = chunk_io(ps, 0, READ, 1);
299 if (r)
300 goto bad;
301
302 dh = (struct disk_header *) ps->area;
303
304 if (le32_to_cpu(dh->magic) == 0) {
305 *new_snapshot = 1;
306 return 0;
307 }
308
309 if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
310 DMWARN("Invalid or corrupt snapshot");
311 r = -ENXIO;
312 goto bad;
313 }
314
315 *new_snapshot = 0;
316 ps->valid = le32_to_cpu(dh->valid);
317 ps->version = le32_to_cpu(dh->version);
318 chunk_size = le32_to_cpu(dh->chunk_size);
319
320 if (!chunk_size_supplied || ps->snap->chunk_size == chunk_size)
321 return 0;
322
323 DMWARN("chunk size %llu in device metadata overrides "
324 "table chunk size of %llu.",
325 (unsigned long long)chunk_size,
326 (unsigned long long)ps->snap->chunk_size);
327
328 /* We had a bogus chunk_size. Fix stuff up. */
329 free_area(ps);
330
331 ps->snap->chunk_size = chunk_size;
332 ps->snap->chunk_mask = chunk_size - 1;
333 ps->snap->chunk_shift = ffs(chunk_size) - 1;
334
335 r = dm_io_client_resize(sectors_to_pages(ps->snap->chunk_size),
336 ps->io_client);
337 if (r)
338 return r;
339
340 r = alloc_area(ps);
341 return r;
342
343bad:
344 free_area(ps);
345 return r;
346}
347
348static int write_header(struct pstore *ps)
349{
350 struct disk_header *dh;
351
352 memset(ps->area, 0, ps->snap->chunk_size << SECTOR_SHIFT);
353
354 dh = (struct disk_header *) ps->area;
355 dh->magic = cpu_to_le32(SNAP_MAGIC);
356 dh->valid = cpu_to_le32(ps->valid);
357 dh->version = cpu_to_le32(ps->version);
358 dh->chunk_size = cpu_to_le32(ps->snap->chunk_size);
359
360 return chunk_io(ps, 0, WRITE, 1);
361}
362
363/*
364 * Access functions for the disk exceptions, these do the endian conversions.
365 */
366static struct disk_exception *get_exception(struct pstore *ps, uint32_t index)
367{
368 BUG_ON(index >= ps->exceptions_per_area);
369
370 return ((struct disk_exception *) ps->area) + index;
371}
372
373static void read_exception(struct pstore *ps,
374 uint32_t index, struct disk_exception *result)
375{
376 struct disk_exception *e = get_exception(ps, index);
377
378 /* copy it */
379 result->old_chunk = le64_to_cpu(e->old_chunk);
380 result->new_chunk = le64_to_cpu(e->new_chunk);
381}
382
383static void write_exception(struct pstore *ps,
384 uint32_t index, struct disk_exception *de)
385{
386 struct disk_exception *e = get_exception(ps, index);
387
388 /* copy it */
389 e->old_chunk = cpu_to_le64(de->old_chunk);
390 e->new_chunk = cpu_to_le64(de->new_chunk);
391}
392
393/*
394 * Registers the exceptions that are present in the current area.
395 * 'full' is filled in to indicate if the area has been
396 * filled.
397 */
Jonathan Brassowa159c1a2009-01-06 03:05:19 +0000398static int insert_exceptions(struct pstore *ps,
399 int (*callback)(void *callback_context,
400 chunk_t old, chunk_t new),
401 void *callback_context,
402 int *full)
Alasdair G Kergon4db6bfe2009-01-06 03:05:17 +0000403{
404 int r;
405 unsigned int i;
406 struct disk_exception de;
407
408 /* presume the area is full */
409 *full = 1;
410
411 for (i = 0; i < ps->exceptions_per_area; i++) {
412 read_exception(ps, i, &de);
413
414 /*
415 * If the new_chunk is pointing at the start of
416 * the COW device, where the first metadata area
417 * is we know that we've hit the end of the
418 * exceptions. Therefore the area is not full.
419 */
420 if (de.new_chunk == 0LL) {
421 ps->current_committed = i;
422 *full = 0;
423 break;
424 }
425
426 /*
427 * Keep track of the start of the free chunks.
428 */
429 if (ps->next_free <= de.new_chunk)
430 ps->next_free = de.new_chunk + 1;
431
432 /*
433 * Otherwise we add the exception to the snapshot.
434 */
Jonathan Brassowa159c1a2009-01-06 03:05:19 +0000435 r = callback(callback_context, de.old_chunk, de.new_chunk);
Alasdair G Kergon4db6bfe2009-01-06 03:05:17 +0000436 if (r)
437 return r;
438 }
439
440 return 0;
441}
442
Jonathan Brassowa159c1a2009-01-06 03:05:19 +0000443static int read_exceptions(struct pstore *ps,
444 int (*callback)(void *callback_context, chunk_t old,
445 chunk_t new),
446 void *callback_context)
Alasdair G Kergon4db6bfe2009-01-06 03:05:17 +0000447{
448 int r, full = 1;
449
450 /*
451 * Keeping reading chunks and inserting exceptions until
452 * we find a partially full area.
453 */
454 for (ps->current_area = 0; full; ps->current_area++) {
455 r = area_io(ps, READ);
456 if (r)
457 return r;
458
Jonathan Brassowa159c1a2009-01-06 03:05:19 +0000459 r = insert_exceptions(ps, callback, callback_context, &full);
Alasdair G Kergon4db6bfe2009-01-06 03:05:17 +0000460 if (r)
461 return r;
462 }
463
464 ps->current_area--;
465
466 return 0;
467}
468
469static struct pstore *get_info(struct dm_exception_store *store)
470{
471 return (struct pstore *) store->context;
472}
473
474static void persistent_fraction_full(struct dm_exception_store *store,
475 sector_t *numerator, sector_t *denominator)
476{
477 *numerator = get_info(store)->next_free * store->snap->chunk_size;
478 *denominator = get_dev_size(store->snap->cow->bdev);
479}
480
481static void persistent_destroy(struct dm_exception_store *store)
482{
483 struct pstore *ps = get_info(store);
484
485 destroy_workqueue(ps->metadata_wq);
486 dm_io_client_destroy(ps->io_client);
487 vfree(ps->callbacks);
488 free_area(ps);
489 kfree(ps);
490}
491
Jonathan Brassowa159c1a2009-01-06 03:05:19 +0000492static int persistent_read_metadata(struct dm_exception_store *store,
493 int (*callback)(void *callback_context,
494 chunk_t old, chunk_t new),
495 void *callback_context)
Alasdair G Kergon4db6bfe2009-01-06 03:05:17 +0000496{
497 int r, uninitialized_var(new_snapshot);
498 struct pstore *ps = get_info(store);
499
500 /*
501 * Read the snapshot header.
502 */
503 r = read_header(ps, &new_snapshot);
504 if (r)
505 return r;
506
507 /*
508 * Now we know correct chunk_size, complete the initialisation.
509 */
510 ps->exceptions_per_area = (ps->snap->chunk_size << SECTOR_SHIFT) /
511 sizeof(struct disk_exception);
512 ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
513 sizeof(*ps->callbacks));
514 if (!ps->callbacks)
515 return -ENOMEM;
516
517 /*
518 * Do we need to setup a new snapshot ?
519 */
520 if (new_snapshot) {
521 r = write_header(ps);
522 if (r) {
523 DMWARN("write_header failed");
524 return r;
525 }
526
527 ps->current_area = 0;
528 zero_memory_area(ps);
529 r = zero_disk_area(ps, 0);
530 if (r) {
531 DMWARN("zero_disk_area(0) failed");
532 return r;
533 }
534 } else {
535 /*
536 * Sanity checks.
537 */
538 if (ps->version != SNAPSHOT_DISK_VERSION) {
539 DMWARN("unable to handle snapshot disk version %d",
540 ps->version);
541 return -EINVAL;
542 }
543
544 /*
545 * Metadata are valid, but snapshot is invalidated
546 */
547 if (!ps->valid)
548 return 1;
549
550 /*
551 * Read the metadata.
552 */
Jonathan Brassowa159c1a2009-01-06 03:05:19 +0000553 r = read_exceptions(ps, callback, callback_context);
Alasdair G Kergon4db6bfe2009-01-06 03:05:17 +0000554 if (r)
555 return r;
556 }
557
558 return 0;
559}
560
Jonathan Brassowa159c1a2009-01-06 03:05:19 +0000561static int persistent_prepare_exception(struct dm_exception_store *store,
562 struct dm_snap_exception *e)
Alasdair G Kergon4db6bfe2009-01-06 03:05:17 +0000563{
564 struct pstore *ps = get_info(store);
565 uint32_t stride;
566 chunk_t next_free;
567 sector_t size = get_dev_size(store->snap->cow->bdev);
568
569 /* Is there enough room ? */
570 if (size < ((ps->next_free + 1) * store->snap->chunk_size))
571 return -ENOSPC;
572
573 e->new_chunk = ps->next_free;
574
575 /*
576 * Move onto the next free pending, making sure to take
577 * into account the location of the metadata chunks.
578 */
579 stride = (ps->exceptions_per_area + 1);
580 next_free = ++ps->next_free;
581 if (sector_div(next_free, stride) == 1)
582 ps->next_free++;
583
584 atomic_inc(&ps->pending_count);
585 return 0;
586}
587
Jonathan Brassowa159c1a2009-01-06 03:05:19 +0000588static void persistent_commit_exception(struct dm_exception_store *store,
589 struct dm_snap_exception *e,
590 void (*callback) (void *, int success),
591 void *callback_context)
Alasdair G Kergon4db6bfe2009-01-06 03:05:17 +0000592{
593 unsigned int i;
594 struct pstore *ps = get_info(store);
595 struct disk_exception de;
596 struct commit_callback *cb;
597
598 de.old_chunk = e->old_chunk;
599 de.new_chunk = e->new_chunk;
600 write_exception(ps, ps->current_committed++, &de);
601
602 /*
603 * Add the callback to the back of the array. This code
604 * is the only place where the callback array is
605 * manipulated, and we know that it will never be called
606 * multiple times concurrently.
607 */
608 cb = ps->callbacks + ps->callback_count++;
609 cb->callback = callback;
610 cb->context = callback_context;
611
612 /*
613 * If there are exceptions in flight and we have not yet
614 * filled this metadata area there's nothing more to do.
615 */
616 if (!atomic_dec_and_test(&ps->pending_count) &&
617 (ps->current_committed != ps->exceptions_per_area))
618 return;
619
620 /*
621 * If we completely filled the current area, then wipe the next one.
622 */
623 if ((ps->current_committed == ps->exceptions_per_area) &&
624 zero_disk_area(ps, ps->current_area + 1))
625 ps->valid = 0;
626
627 /*
628 * Commit exceptions to disk.
629 */
630 if (ps->valid && area_io(ps, WRITE))
631 ps->valid = 0;
632
633 /*
634 * Advance to the next area if this one is full.
635 */
636 if (ps->current_committed == ps->exceptions_per_area) {
637 ps->current_committed = 0;
638 ps->current_area++;
639 zero_memory_area(ps);
640 }
641
642 for (i = 0; i < ps->callback_count; i++) {
643 cb = ps->callbacks + i;
644 cb->callback(cb->context, ps->valid);
645 }
646
647 ps->callback_count = 0;
648}
649
Jonathan Brassowa159c1a2009-01-06 03:05:19 +0000650static void persistent_drop_snapshot(struct dm_exception_store *store)
Alasdair G Kergon4db6bfe2009-01-06 03:05:17 +0000651{
652 struct pstore *ps = get_info(store);
653
654 ps->valid = 0;
655 if (write_header(ps))
656 DMWARN("write header failed");
657}
658
659int dm_create_persistent(struct dm_exception_store *store)
660{
661 struct pstore *ps;
662
663 /* allocate the pstore */
664 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
665 if (!ps)
666 return -ENOMEM;
667
668 ps->snap = store->snap;
669 ps->valid = 1;
670 ps->version = SNAPSHOT_DISK_VERSION;
671 ps->area = NULL;
672 ps->next_free = 2; /* skipping the header and first area */
673 ps->current_committed = 0;
674
675 ps->callback_count = 0;
676 atomic_set(&ps->pending_count, 0);
677 ps->callbacks = NULL;
678
679 ps->metadata_wq = create_singlethread_workqueue("ksnaphd");
680 if (!ps->metadata_wq) {
681 kfree(ps);
682 DMERR("couldn't start header metadata update thread");
683 return -ENOMEM;
684 }
685
686 store->destroy = persistent_destroy;
687 store->read_metadata = persistent_read_metadata;
Jonathan Brassowa159c1a2009-01-06 03:05:19 +0000688 store->prepare_exception = persistent_prepare_exception;
689 store->commit_exception = persistent_commit_exception;
690 store->drop_snapshot = persistent_drop_snapshot;
Alasdair G Kergon4db6bfe2009-01-06 03:05:17 +0000691 store->fraction_full = persistent_fraction_full;
692 store->context = ps;
693
694 return 0;
695}
696
697int dm_persistent_snapshot_init(void)
698{
699 return 0;
700}
701
702void dm_persistent_snapshot_exit(void)
703{
704}