Merge tag 'bcachefs-2024-10-05' of git://evilpiepirate.org/bcachefs
[linux-2.6-block.git] / drivers / md / persistent-data / dm-transaction-manager.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2011 Red Hat, Inc.
4  *
5  * This file is released under the GPL.
6  */
7 #include "dm-transaction-manager.h"
8 #include "dm-space-map.h"
9 #include "dm-space-map-disk.h"
10 #include "dm-space-map-metadata.h"
11 #include "dm-persistent-data-internal.h"
12
13 #include <linux/export.h>
14 #include <linux/mutex.h>
15 #include <linux/hash.h>
16 #include <linux/slab.h>
17 #include <linux/device-mapper.h>
18
19 #define DM_MSG_PREFIX "transaction manager"
20
21 /*----------------------------------------------------------------*/
22
23 #define PREFETCH_SIZE 128
24 #define PREFETCH_BITS 7
25 #define PREFETCH_SENTINEL ((dm_block_t) -1ULL)
26
27 struct prefetch_set {
28         struct mutex lock;
29         dm_block_t blocks[PREFETCH_SIZE];
30 };
31
32 static unsigned int prefetch_hash(dm_block_t b)
33 {
34         return hash_64(b, PREFETCH_BITS);
35 }
36
37 static void prefetch_wipe(struct prefetch_set *p)
38 {
39         unsigned int i;
40
41         for (i = 0; i < PREFETCH_SIZE; i++)
42                 p->blocks[i] = PREFETCH_SENTINEL;
43 }
44
45 static void prefetch_init(struct prefetch_set *p)
46 {
47         mutex_init(&p->lock);
48         prefetch_wipe(p);
49 }
50
51 static void prefetch_add(struct prefetch_set *p, dm_block_t b)
52 {
53         unsigned int h = prefetch_hash(b);
54
55         mutex_lock(&p->lock);
56         if (p->blocks[h] == PREFETCH_SENTINEL)
57                 p->blocks[h] = b;
58
59         mutex_unlock(&p->lock);
60 }
61
62 static void prefetch_issue(struct prefetch_set *p, struct dm_block_manager *bm)
63 {
64         unsigned int i;
65
66         mutex_lock(&p->lock);
67
68         for (i = 0; i < PREFETCH_SIZE; i++)
69                 if (p->blocks[i] != PREFETCH_SENTINEL) {
70                         dm_bm_prefetch(bm, p->blocks[i]);
71                         p->blocks[i] = PREFETCH_SENTINEL;
72                 }
73
74         mutex_unlock(&p->lock);
75 }
76
77 /*----------------------------------------------------------------*/
78
79 struct shadow_info {
80         struct hlist_node hlist;
81         dm_block_t where;
82 };
83
84 /*
85  * It would be nice if we scaled with the size of transaction.
86  */
87 #define DM_HASH_SIZE 256
88 #define DM_HASH_MASK (DM_HASH_SIZE - 1)
89
90 struct dm_transaction_manager {
91         int is_clone;
92         struct dm_transaction_manager *real;
93
94         struct dm_block_manager *bm;
95         struct dm_space_map *sm;
96
97         spinlock_t lock;
98         struct hlist_head buckets[DM_HASH_SIZE];
99
100         struct prefetch_set prefetches;
101 };
102
103 /*----------------------------------------------------------------*/
104
105 static int is_shadow(struct dm_transaction_manager *tm, dm_block_t b)
106 {
107         int r = 0;
108         unsigned int bucket = dm_hash_block(b, DM_HASH_MASK);
109         struct shadow_info *si;
110
111         spin_lock(&tm->lock);
112         hlist_for_each_entry(si, tm->buckets + bucket, hlist)
113                 if (si->where == b) {
114                         r = 1;
115                         break;
116                 }
117         spin_unlock(&tm->lock);
118
119         return r;
120 }
121
122 /*
123  * This can silently fail if there's no memory.  We're ok with this since
124  * creating redundant shadows causes no harm.
125  */
126 static void insert_shadow(struct dm_transaction_manager *tm, dm_block_t b)
127 {
128         unsigned int bucket;
129         struct shadow_info *si;
130
131         si = kmalloc(sizeof(*si), GFP_NOIO);
132         if (si) {
133                 si->where = b;
134                 bucket = dm_hash_block(b, DM_HASH_MASK);
135                 spin_lock(&tm->lock);
136                 hlist_add_head(&si->hlist, tm->buckets + bucket);
137                 spin_unlock(&tm->lock);
138         }
139 }
140
141 static void wipe_shadow_table(struct dm_transaction_manager *tm)
142 {
143         struct shadow_info *si;
144         struct hlist_node *tmp;
145         struct hlist_head *bucket;
146         int i;
147
148         spin_lock(&tm->lock);
149         for (i = 0; i < DM_HASH_SIZE; i++) {
150                 bucket = tm->buckets + i;
151                 hlist_for_each_entry_safe(si, tmp, bucket, hlist)
152                         kfree(si);
153
154                 INIT_HLIST_HEAD(bucket);
155         }
156
157         spin_unlock(&tm->lock);
158 }
159
160 /*----------------------------------------------------------------*/
161
162 static struct dm_transaction_manager *dm_tm_create(struct dm_block_manager *bm,
163                                                    struct dm_space_map *sm)
164 {
165         int i;
166         struct dm_transaction_manager *tm;
167
168         tm = kmalloc(sizeof(*tm), GFP_KERNEL);
169         if (!tm)
170                 return ERR_PTR(-ENOMEM);
171
172         tm->is_clone = 0;
173         tm->real = NULL;
174         tm->bm = bm;
175         tm->sm = sm;
176
177         spin_lock_init(&tm->lock);
178         for (i = 0; i < DM_HASH_SIZE; i++)
179                 INIT_HLIST_HEAD(tm->buckets + i);
180
181         prefetch_init(&tm->prefetches);
182
183         return tm;
184 }
185
186 struct dm_transaction_manager *dm_tm_create_non_blocking_clone(struct dm_transaction_manager *real)
187 {
188         struct dm_transaction_manager *tm;
189
190         tm = kmalloc(sizeof(*tm), GFP_KERNEL);
191         if (tm) {
192                 tm->is_clone = 1;
193                 tm->real = real;
194         }
195
196         return tm;
197 }
198 EXPORT_SYMBOL_GPL(dm_tm_create_non_blocking_clone);
199
200 void dm_tm_destroy(struct dm_transaction_manager *tm)
201 {
202         if (!tm)
203                 return;
204
205         if (!tm->is_clone)
206                 wipe_shadow_table(tm);
207
208         kfree(tm);
209 }
210 EXPORT_SYMBOL_GPL(dm_tm_destroy);
211
212 int dm_tm_pre_commit(struct dm_transaction_manager *tm)
213 {
214         int r;
215
216         if (tm->is_clone)
217                 return -EWOULDBLOCK;
218
219         r = dm_sm_commit(tm->sm);
220         if (r < 0)
221                 return r;
222
223         return dm_bm_flush(tm->bm);
224 }
225 EXPORT_SYMBOL_GPL(dm_tm_pre_commit);
226
227 int dm_tm_commit(struct dm_transaction_manager *tm, struct dm_block *root)
228 {
229         if (tm->is_clone)
230                 return -EWOULDBLOCK;
231
232         wipe_shadow_table(tm);
233         dm_bm_unlock(root);
234
235         return dm_bm_flush(tm->bm);
236 }
237 EXPORT_SYMBOL_GPL(dm_tm_commit);
238
239 int dm_tm_new_block(struct dm_transaction_manager *tm,
240                     const struct dm_block_validator *v,
241                     struct dm_block **result)
242 {
243         int r;
244         dm_block_t new_block;
245
246         if (tm->is_clone)
247                 return -EWOULDBLOCK;
248
249         r = dm_sm_new_block(tm->sm, &new_block);
250         if (r < 0)
251                 return r;
252
253         r = dm_bm_write_lock_zero(tm->bm, new_block, v, result);
254         if (r < 0) {
255                 dm_sm_dec_block(tm->sm, new_block);
256                 return r;
257         }
258
259         /*
260          * New blocks count as shadows in that they don't need to be
261          * shadowed again.
262          */
263         insert_shadow(tm, new_block);
264
265         return 0;
266 }
267
268 static int __shadow_block(struct dm_transaction_manager *tm, dm_block_t orig,
269                           const struct dm_block_validator *v,
270                           struct dm_block **result)
271 {
272         int r;
273         dm_block_t new;
274         struct dm_block *orig_block;
275
276         r = dm_sm_new_block(tm->sm, &new);
277         if (r < 0)
278                 return r;
279
280         r = dm_sm_dec_block(tm->sm, orig);
281         if (r < 0)
282                 return r;
283
284         r = dm_bm_read_lock(tm->bm, orig, v, &orig_block);
285         if (r < 0)
286                 return r;
287
288         /*
289          * It would be tempting to use dm_bm_unlock_move here, but some
290          * code, such as the space maps, keeps using the old data structures
291          * secure in the knowledge they won't be changed until the next
292          * transaction.  Using unlock_move would force a synchronous read
293          * since the old block would no longer be in the cache.
294          */
295         r = dm_bm_write_lock_zero(tm->bm, new, v, result);
296         if (r) {
297                 dm_bm_unlock(orig_block);
298                 return r;
299         }
300
301         memcpy(dm_block_data(*result), dm_block_data(orig_block),
302                dm_bm_block_size(tm->bm));
303
304         dm_bm_unlock(orig_block);
305         return r;
306 }
307
308 int dm_tm_shadow_block(struct dm_transaction_manager *tm, dm_block_t orig,
309                        const struct dm_block_validator *v, struct dm_block **result,
310                        int *inc_children)
311 {
312         int r;
313
314         if (tm->is_clone)
315                 return -EWOULDBLOCK;
316
317         r = dm_sm_count_is_more_than_one(tm->sm, orig, inc_children);
318         if (r < 0)
319                 return r;
320
321         if (is_shadow(tm, orig) && !*inc_children)
322                 return dm_bm_write_lock(tm->bm, orig, v, result);
323
324         r = __shadow_block(tm, orig, v, result);
325         if (r < 0)
326                 return r;
327         insert_shadow(tm, dm_block_location(*result));
328
329         return r;
330 }
331 EXPORT_SYMBOL_GPL(dm_tm_shadow_block);
332
333 int dm_tm_read_lock(struct dm_transaction_manager *tm, dm_block_t b,
334                     const struct dm_block_validator *v,
335                     struct dm_block **blk)
336 {
337         if (tm->is_clone) {
338                 int r = dm_bm_read_try_lock(tm->real->bm, b, v, blk);
339
340                 if (r == -EWOULDBLOCK)
341                         prefetch_add(&tm->real->prefetches, b);
342
343                 return r;
344         }
345
346         return dm_bm_read_lock(tm->bm, b, v, blk);
347 }
348 EXPORT_SYMBOL_GPL(dm_tm_read_lock);
349
350 void dm_tm_unlock(struct dm_transaction_manager *tm, struct dm_block *b)
351 {
352         dm_bm_unlock(b);
353 }
354 EXPORT_SYMBOL_GPL(dm_tm_unlock);
355
356 void dm_tm_inc(struct dm_transaction_manager *tm, dm_block_t b)
357 {
358         /*
359          * The non-blocking clone doesn't support this.
360          */
361         BUG_ON(tm->is_clone);
362
363         dm_sm_inc_block(tm->sm, b);
364 }
365 EXPORT_SYMBOL_GPL(dm_tm_inc);
366
367 void dm_tm_inc_range(struct dm_transaction_manager *tm, dm_block_t b, dm_block_t e)
368 {
369         /*
370          * The non-blocking clone doesn't support this.
371          */
372         BUG_ON(tm->is_clone);
373
374         dm_sm_inc_blocks(tm->sm, b, e);
375 }
376 EXPORT_SYMBOL_GPL(dm_tm_inc_range);
377
378 void dm_tm_dec(struct dm_transaction_manager *tm, dm_block_t b)
379 {
380         /*
381          * The non-blocking clone doesn't support this.
382          */
383         BUG_ON(tm->is_clone);
384
385         dm_sm_dec_block(tm->sm, b);
386 }
387 EXPORT_SYMBOL_GPL(dm_tm_dec);
388
389 void dm_tm_dec_range(struct dm_transaction_manager *tm, dm_block_t b, dm_block_t e)
390 {
391         /*
392          * The non-blocking clone doesn't support this.
393          */
394         BUG_ON(tm->is_clone);
395
396         dm_sm_dec_blocks(tm->sm, b, e);
397 }
398 EXPORT_SYMBOL_GPL(dm_tm_dec_range);
399
400 void dm_tm_with_runs(struct dm_transaction_manager *tm,
401                      const __le64 *value_le, unsigned int count, dm_tm_run_fn fn)
402 {
403         uint64_t b, begin, end;
404         bool in_run = false;
405         unsigned int i;
406
407         for (i = 0; i < count; i++, value_le++) {
408                 b = le64_to_cpu(*value_le);
409
410                 if (in_run) {
411                         if (b == end)
412                                 end++;
413                         else {
414                                 fn(tm, begin, end);
415                                 begin = b;
416                                 end = b + 1;
417                         }
418                 } else {
419                         in_run = true;
420                         begin = b;
421                         end = b + 1;
422                 }
423         }
424
425         if (in_run)
426                 fn(tm, begin, end);
427 }
428 EXPORT_SYMBOL_GPL(dm_tm_with_runs);
429
430 int dm_tm_ref(struct dm_transaction_manager *tm, dm_block_t b,
431               uint32_t *result)
432 {
433         if (tm->is_clone)
434                 return -EWOULDBLOCK;
435
436         return dm_sm_get_count(tm->sm, b, result);
437 }
438
439 int dm_tm_block_is_shared(struct dm_transaction_manager *tm, dm_block_t b,
440                           int *result)
441 {
442         if (tm->is_clone)
443                 return -EWOULDBLOCK;
444
445         return dm_sm_count_is_more_than_one(tm->sm, b, result);
446 }
447
448 struct dm_block_manager *dm_tm_get_bm(struct dm_transaction_manager *tm)
449 {
450         return tm->bm;
451 }
452
453 void dm_tm_issue_prefetches(struct dm_transaction_manager *tm)
454 {
455         prefetch_issue(&tm->prefetches, tm->bm);
456 }
457 EXPORT_SYMBOL_GPL(dm_tm_issue_prefetches);
458
459 /*----------------------------------------------------------------*/
460
461 static int dm_tm_create_internal(struct dm_block_manager *bm,
462                                  dm_block_t sb_location,
463                                  struct dm_transaction_manager **tm,
464                                  struct dm_space_map **sm,
465                                  int create,
466                                  void *sm_root, size_t sm_len)
467 {
468         int r;
469
470         *sm = dm_sm_metadata_init();
471         if (IS_ERR(*sm))
472                 return PTR_ERR(*sm);
473
474         *tm = dm_tm_create(bm, *sm);
475         if (IS_ERR(*tm)) {
476                 dm_sm_destroy(*sm);
477                 return PTR_ERR(*tm);
478         }
479
480         if (create) {
481                 r = dm_sm_metadata_create(*sm, *tm, dm_bm_nr_blocks(bm),
482                                           sb_location);
483                 if (r) {
484                         DMERR("couldn't create metadata space map");
485                         goto bad;
486                 }
487
488         } else {
489                 r = dm_sm_metadata_open(*sm, *tm, sm_root, sm_len);
490                 if (r) {
491                         DMERR("couldn't open metadata space map");
492                         goto bad;
493                 }
494         }
495
496         return 0;
497
498 bad:
499         dm_tm_destroy(*tm);
500         dm_sm_destroy(*sm);
501         return r;
502 }
503
504 int dm_tm_create_with_sm(struct dm_block_manager *bm, dm_block_t sb_location,
505                          struct dm_transaction_manager **tm,
506                          struct dm_space_map **sm)
507 {
508         return dm_tm_create_internal(bm, sb_location, tm, sm, 1, NULL, 0);
509 }
510 EXPORT_SYMBOL_GPL(dm_tm_create_with_sm);
511
512 int dm_tm_open_with_sm(struct dm_block_manager *bm, dm_block_t sb_location,
513                        void *sm_root, size_t root_len,
514                        struct dm_transaction_manager **tm,
515                        struct dm_space_map **sm)
516 {
517         return dm_tm_create_internal(bm, sb_location, tm, sm, 0, sm_root, root_len);
518 }
519 EXPORT_SYMBOL_GPL(dm_tm_open_with_sm);
520
521 /*----------------------------------------------------------------*/