dm: send just one event on resize, not two
[linux-2.6-block.git] / drivers / md / persistent-data / dm-transaction-manager.c
CommitLineData
3241b1d3
JT
1/*
2 * Copyright (C) 2011 Red Hat, Inc.
3 *
4 * This file is released under the GPL.
5 */
6#include "dm-transaction-manager.h"
7#include "dm-space-map.h"
3241b1d3
JT
8#include "dm-space-map-disk.h"
9#include "dm-space-map-metadata.h"
10#include "dm-persistent-data-internal.h"
11
1944ce60 12#include <linux/export.h>
4646015d
JT
13#include <linux/mutex.h>
14#include <linux/hash.h>
3241b1d3
JT
15#include <linux/slab.h>
16#include <linux/device-mapper.h>
17
18#define DM_MSG_PREFIX "transaction manager"
19
20/*----------------------------------------------------------------*/
21
4646015d
JT
22#define PREFETCH_SIZE 128
23#define PREFETCH_BITS 7
24#define PREFETCH_SENTINEL ((dm_block_t) -1ULL)
25
26struct prefetch_set {
27 struct mutex lock;
28 dm_block_t blocks[PREFETCH_SIZE];
29};
30
31static unsigned prefetch_hash(dm_block_t b)
32{
33 return hash_64(b, PREFETCH_BITS);
34}
35
36static void prefetch_wipe(struct prefetch_set *p)
37{
38 unsigned i;
39 for (i = 0; i < PREFETCH_SIZE; i++)
40 p->blocks[i] = PREFETCH_SENTINEL;
41}
42
43static void prefetch_init(struct prefetch_set *p)
44{
45 mutex_init(&p->lock);
46 prefetch_wipe(p);
47}
48
49static void prefetch_add(struct prefetch_set *p, dm_block_t b)
50{
51 unsigned h = prefetch_hash(b);
52
53 mutex_lock(&p->lock);
54 if (p->blocks[h] == PREFETCH_SENTINEL)
55 p->blocks[h] = b;
56
57 mutex_unlock(&p->lock);
58}
59
60static void prefetch_issue(struct prefetch_set *p, struct dm_block_manager *bm)
61{
62 unsigned i;
63
64 mutex_lock(&p->lock);
65
66 for (i = 0; i < PREFETCH_SIZE; i++)
67 if (p->blocks[i] != PREFETCH_SENTINEL) {
68 dm_bm_prefetch(bm, p->blocks[i]);
69 p->blocks[i] = PREFETCH_SENTINEL;
70 }
71
72 mutex_unlock(&p->lock);
73}
74
75/*----------------------------------------------------------------*/
76
3241b1d3
JT
77struct shadow_info {
78 struct hlist_node hlist;
79 dm_block_t where;
80};
81
82/*
83 * It would be nice if we scaled with the size of transaction.
84 */
df855798
AM
85#define DM_HASH_SIZE 256
86#define DM_HASH_MASK (DM_HASH_SIZE - 1)
3241b1d3
JT
87
88struct dm_transaction_manager {
89 int is_clone;
90 struct dm_transaction_manager *real;
91
92 struct dm_block_manager *bm;
93 struct dm_space_map *sm;
94
95 spinlock_t lock;
df855798 96 struct hlist_head buckets[DM_HASH_SIZE];
4646015d
JT
97
98 struct prefetch_set prefetches;
3241b1d3
JT
99};
100
101/*----------------------------------------------------------------*/
102
103static int is_shadow(struct dm_transaction_manager *tm, dm_block_t b)
104{
105 int r = 0;
df855798 106 unsigned bucket = dm_hash_block(b, DM_HASH_MASK);
3241b1d3 107 struct shadow_info *si;
3241b1d3
JT
108
109 spin_lock(&tm->lock);
b67bfe0d 110 hlist_for_each_entry(si, tm->buckets + bucket, hlist)
3241b1d3
JT
111 if (si->where == b) {
112 r = 1;
113 break;
114 }
115 spin_unlock(&tm->lock);
116
117 return r;
118}
119
120/*
121 * This can silently fail if there's no memory. We're ok with this since
122 * creating redundant shadows causes no harm.
123 */
124static void insert_shadow(struct dm_transaction_manager *tm, dm_block_t b)
125{
126 unsigned bucket;
127 struct shadow_info *si;
128
129 si = kmalloc(sizeof(*si), GFP_NOIO);
130 if (si) {
131 si->where = b;
df855798 132 bucket = dm_hash_block(b, DM_HASH_MASK);
3241b1d3
JT
133 spin_lock(&tm->lock);
134 hlist_add_head(&si->hlist, tm->buckets + bucket);
135 spin_unlock(&tm->lock);
136 }
137}
138
139static void wipe_shadow_table(struct dm_transaction_manager *tm)
140{
141 struct shadow_info *si;
b67bfe0d 142 struct hlist_node *tmp;
3241b1d3
JT
143 struct hlist_head *bucket;
144 int i;
145
146 spin_lock(&tm->lock);
df855798 147 for (i = 0; i < DM_HASH_SIZE; i++) {
3241b1d3 148 bucket = tm->buckets + i;
b67bfe0d 149 hlist_for_each_entry_safe(si, tmp, bucket, hlist)
3241b1d3
JT
150 kfree(si);
151
152 INIT_HLIST_HEAD(bucket);
153 }
154
155 spin_unlock(&tm->lock);
156}
157
158/*----------------------------------------------------------------*/
159
160static struct dm_transaction_manager *dm_tm_create(struct dm_block_manager *bm,
161 struct dm_space_map *sm)
162{
163 int i;
164 struct dm_transaction_manager *tm;
165
166 tm = kmalloc(sizeof(*tm), GFP_KERNEL);
167 if (!tm)
168 return ERR_PTR(-ENOMEM);
169
170 tm->is_clone = 0;
171 tm->real = NULL;
172 tm->bm = bm;
173 tm->sm = sm;
174
175 spin_lock_init(&tm->lock);
df855798 176 for (i = 0; i < DM_HASH_SIZE; i++)
3241b1d3
JT
177 INIT_HLIST_HEAD(tm->buckets + i);
178
4646015d
JT
179 prefetch_init(&tm->prefetches);
180
3241b1d3
JT
181 return tm;
182}
183
184struct dm_transaction_manager *dm_tm_create_non_blocking_clone(struct dm_transaction_manager *real)
185{
186 struct dm_transaction_manager *tm;
187
188 tm = kmalloc(sizeof(*tm), GFP_KERNEL);
189 if (tm) {
190 tm->is_clone = 1;
191 tm->real = real;
192 }
193
194 return tm;
195}
196EXPORT_SYMBOL_GPL(dm_tm_create_non_blocking_clone);
197
198void dm_tm_destroy(struct dm_transaction_manager *tm)
199{
25d7cd6f
MS
200 if (!tm->is_clone)
201 wipe_shadow_table(tm);
202
3241b1d3
JT
203 kfree(tm);
204}
205EXPORT_SYMBOL_GPL(dm_tm_destroy);
206
207int dm_tm_pre_commit(struct dm_transaction_manager *tm)
208{
209 int r;
210
211 if (tm->is_clone)
212 return -EWOULDBLOCK;
213
214 r = dm_sm_commit(tm->sm);
215 if (r < 0)
216 return r;
217
a9d45396 218 return dm_bm_flush(tm->bm);
3241b1d3
JT
219}
220EXPORT_SYMBOL_GPL(dm_tm_pre_commit);
221
222int dm_tm_commit(struct dm_transaction_manager *tm, struct dm_block *root)
223{
224 if (tm->is_clone)
225 return -EWOULDBLOCK;
226
227 wipe_shadow_table(tm);
a9d45396 228 dm_bm_unlock(root);
3241b1d3 229
a9d45396 230 return dm_bm_flush(tm->bm);
3241b1d3
JT
231}
232EXPORT_SYMBOL_GPL(dm_tm_commit);
233
234int dm_tm_new_block(struct dm_transaction_manager *tm,
235 struct dm_block_validator *v,
236 struct dm_block **result)
237{
238 int r;
239 dm_block_t new_block;
240
241 if (tm->is_clone)
242 return -EWOULDBLOCK;
243
244 r = dm_sm_new_block(tm->sm, &new_block);
245 if (r < 0)
246 return r;
247
248 r = dm_bm_write_lock_zero(tm->bm, new_block, v, result);
249 if (r < 0) {
250 dm_sm_dec_block(tm->sm, new_block);
251 return r;
252 }
253
254 /*
255 * New blocks count as shadows in that they don't need to be
256 * shadowed again.
257 */
258 insert_shadow(tm, new_block);
259
260 return 0;
261}
262
263static int __shadow_block(struct dm_transaction_manager *tm, dm_block_t orig,
264 struct dm_block_validator *v,
265 struct dm_block **result)
266{
267 int r;
268 dm_block_t new;
269 struct dm_block *orig_block;
270
271 r = dm_sm_new_block(tm->sm, &new);
272 if (r < 0)
273 return r;
274
275 r = dm_sm_dec_block(tm->sm, orig);
276 if (r < 0)
277 return r;
278
279 r = dm_bm_read_lock(tm->bm, orig, v, &orig_block);
280 if (r < 0)
281 return r;
282
3c9ad9bd
JT
283 /*
284 * It would be tempting to use dm_bm_unlock_move here, but some
285 * code, such as the space maps, keeps using the old data structures
286 * secure in the knowledge they won't be changed until the next
287 * transaction. Using unlock_move would force a synchronous read
288 * since the old block would no longer be in the cache.
289 */
290 r = dm_bm_write_lock_zero(tm->bm, new, v, result);
291 if (r) {
3241b1d3
JT
292 dm_bm_unlock(orig_block);
293 return r;
294 }
295
3c9ad9bd
JT
296 memcpy(dm_block_data(*result), dm_block_data(orig_block),
297 dm_bm_block_size(tm->bm));
298
299 dm_bm_unlock(orig_block);
300 return r;
3241b1d3
JT
301}
302
303int dm_tm_shadow_block(struct dm_transaction_manager *tm, dm_block_t orig,
304 struct dm_block_validator *v, struct dm_block **result,
305 int *inc_children)
306{
307 int r;
308
309 if (tm->is_clone)
310 return -EWOULDBLOCK;
311
312 r = dm_sm_count_is_more_than_one(tm->sm, orig, inc_children);
313 if (r < 0)
314 return r;
315
316 if (is_shadow(tm, orig) && !*inc_children)
317 return dm_bm_write_lock(tm->bm, orig, v, result);
318
319 r = __shadow_block(tm, orig, v, result);
320 if (r < 0)
321 return r;
322 insert_shadow(tm, dm_block_location(*result));
323
324 return r;
325}
cc8394d8 326EXPORT_SYMBOL_GPL(dm_tm_shadow_block);
3241b1d3
JT
327
328int dm_tm_read_lock(struct dm_transaction_manager *tm, dm_block_t b,
329 struct dm_block_validator *v,
330 struct dm_block **blk)
331{
4646015d
JT
332 if (tm->is_clone) {
333 int r = dm_bm_read_try_lock(tm->real->bm, b, v, blk);
334
335 if (r == -EWOULDBLOCK)
336 prefetch_add(&tm->real->prefetches, b);
337
338 return r;
339 }
3241b1d3
JT
340
341 return dm_bm_read_lock(tm->bm, b, v, blk);
342}
cc8394d8 343EXPORT_SYMBOL_GPL(dm_tm_read_lock);
3241b1d3 344
4c7da06f 345void dm_tm_unlock(struct dm_transaction_manager *tm, struct dm_block *b)
3241b1d3 346{
4c7da06f 347 dm_bm_unlock(b);
3241b1d3
JT
348}
349EXPORT_SYMBOL_GPL(dm_tm_unlock);
350
351void dm_tm_inc(struct dm_transaction_manager *tm, dm_block_t b)
352{
353 /*
354 * The non-blocking clone doesn't support this.
355 */
356 BUG_ON(tm->is_clone);
357
358 dm_sm_inc_block(tm->sm, b);
359}
360EXPORT_SYMBOL_GPL(dm_tm_inc);
361
be500ed7
JT
362void dm_tm_inc_range(struct dm_transaction_manager *tm, dm_block_t b, dm_block_t e)
363{
364 /*
365 * The non-blocking clone doesn't support this.
366 */
367 BUG_ON(tm->is_clone);
368
369 dm_sm_inc_blocks(tm->sm, b, e);
370}
371EXPORT_SYMBOL_GPL(dm_tm_inc_range);
372
3241b1d3
JT
373void dm_tm_dec(struct dm_transaction_manager *tm, dm_block_t b)
374{
375 /*
376 * The non-blocking clone doesn't support this.
377 */
378 BUG_ON(tm->is_clone);
379
380 dm_sm_dec_block(tm->sm, b);
381}
382EXPORT_SYMBOL_GPL(dm_tm_dec);
383
be500ed7
JT
384void dm_tm_dec_range(struct dm_transaction_manager *tm, dm_block_t b, dm_block_t e)
385{
386 /*
387 * The non-blocking clone doesn't support this.
388 */
389 BUG_ON(tm->is_clone);
390
391 dm_sm_dec_blocks(tm->sm, b, e);
392}
393EXPORT_SYMBOL_GPL(dm_tm_dec_range);
394
395void dm_tm_with_runs(struct dm_transaction_manager *tm,
396 const __le64 *value_le, unsigned count, dm_tm_run_fn fn)
397{
398 uint64_t b, begin, end;
399 bool in_run = false;
400 unsigned i;
401
402 for (i = 0; i < count; i++, value_le++) {
403 b = le64_to_cpu(*value_le);
404
405 if (in_run) {
406 if (b == end)
407 end++;
408 else {
409 fn(tm, begin, end);
410 begin = b;
411 end = b + 1;
412 }
413 } else {
414 in_run = true;
415 begin = b;
416 end = b + 1;
417 }
418 }
419
420 if (in_run)
421 fn(tm, begin, end);
422}
423EXPORT_SYMBOL_GPL(dm_tm_with_runs);
424
3241b1d3
JT
425int dm_tm_ref(struct dm_transaction_manager *tm, dm_block_t b,
426 uint32_t *result)
427{
428 if (tm->is_clone)
429 return -EWOULDBLOCK;
430
431 return dm_sm_get_count(tm->sm, b, result);
432}
433
4eafdb15
JT
434int dm_tm_block_is_shared(struct dm_transaction_manager *tm, dm_block_t b,
435 int *result)
436{
437 if (tm->is_clone)
438 return -EWOULDBLOCK;
439
440 return dm_sm_count_is_more_than_one(tm->sm, b, result);
441}
442
3241b1d3
JT
443struct dm_block_manager *dm_tm_get_bm(struct dm_transaction_manager *tm)
444{
445 return tm->bm;
446}
447
4646015d
JT
448void dm_tm_issue_prefetches(struct dm_transaction_manager *tm)
449{
450 prefetch_issue(&tm->prefetches, tm->bm);
451}
452EXPORT_SYMBOL_GPL(dm_tm_issue_prefetches);
453
3241b1d3
JT
454/*----------------------------------------------------------------*/
455
456static int dm_tm_create_internal(struct dm_block_manager *bm,
457 dm_block_t sb_location,
3241b1d3
JT
458 struct dm_transaction_manager **tm,
459 struct dm_space_map **sm,
384ef0e6
JT
460 int create,
461 void *sm_root, size_t sm_len)
3241b1d3
JT
462{
463 int r;
3241b1d3 464
3caf6d73
JT
465 *sm = dm_sm_metadata_init();
466 if (IS_ERR(*sm))
467 return PTR_ERR(*sm);
3241b1d3 468
3caf6d73 469 *tm = dm_tm_create(bm, *sm);
3241b1d3 470 if (IS_ERR(*tm)) {
3caf6d73 471 dm_sm_destroy(*sm);
3241b1d3
JT
472 return PTR_ERR(*tm);
473 }
474
475 if (create) {
3caf6d73 476 r = dm_sm_metadata_create(*sm, *tm, dm_bm_nr_blocks(bm),
3241b1d3
JT
477 sb_location);
478 if (r) {
479 DMERR("couldn't create metadata space map");
384ef0e6 480 goto bad;
3241b1d3
JT
481 }
482
3241b1d3 483 } else {
384ef0e6 484 r = dm_sm_metadata_open(*sm, *tm, sm_root, sm_len);
3241b1d3
JT
485 if (r) {
486 DMERR("couldn't open metadata space map");
384ef0e6 487 goto bad;
3241b1d3 488 }
3241b1d3
JT
489 }
490
491 return 0;
492
384ef0e6 493bad:
3241b1d3 494 dm_tm_destroy(*tm);
384ef0e6 495 dm_sm_destroy(*sm);
3241b1d3
JT
496 return r;
497}
498
499int dm_tm_create_with_sm(struct dm_block_manager *bm, dm_block_t sb_location,
3241b1d3 500 struct dm_transaction_manager **tm,
384ef0e6 501 struct dm_space_map **sm)
3241b1d3 502{
384ef0e6 503 return dm_tm_create_internal(bm, sb_location, tm, sm, 1, NULL, 0);
3241b1d3
JT
504}
505EXPORT_SYMBOL_GPL(dm_tm_create_with_sm);
506
507int dm_tm_open_with_sm(struct dm_block_manager *bm, dm_block_t sb_location,
384ef0e6 508 void *sm_root, size_t root_len,
3241b1d3 509 struct dm_transaction_manager **tm,
384ef0e6 510 struct dm_space_map **sm)
3241b1d3 511{
384ef0e6 512 return dm_tm_create_internal(bm, sb_location, tm, sm, 0, sm_root, root_len);
3241b1d3
JT
513}
514EXPORT_SYMBOL_GPL(dm_tm_open_with_sm);
515
516/*----------------------------------------------------------------*/