Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / drivers / md / persistent-data / dm-transaction-manager.c
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"
8 #include "dm-space-map-disk.h"
9 #include "dm-space-map-metadata.h"
10 #include "dm-persistent-data-internal.h"
11
12 #include <linux/export.h>
13 #include <linux/mutex.h>
14 #include <linux/hash.h>
15 #include <linux/slab.h>
16 #include <linux/device-mapper.h>
17
18 #define DM_MSG_PREFIX "transaction manager"
19
20 /*----------------------------------------------------------------*/
21
22 #define PREFETCH_SIZE 128
23 #define PREFETCH_BITS 7
24 #define PREFETCH_SENTINEL ((dm_block_t) -1ULL)
25
26 struct prefetch_set {
27         struct mutex lock;
28         dm_block_t blocks[PREFETCH_SIZE];
29 };
30
31 static unsigned prefetch_hash(dm_block_t b)
32 {
33         return hash_64(b, PREFETCH_BITS);
34 }
35
36 static 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
43 static void prefetch_init(struct prefetch_set *p)
44 {
45         mutex_init(&p->lock);
46         prefetch_wipe(p);
47 }
48
49 static 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
60 static 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
77 struct 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  */
85 #define DM_HASH_SIZE 256
86 #define DM_HASH_MASK (DM_HASH_SIZE - 1)
87
88 struct 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;
96         struct hlist_head buckets[DM_HASH_SIZE];
97
98         struct prefetch_set prefetches;
99 };
100
101 /*----------------------------------------------------------------*/
102
103 static int is_shadow(struct dm_transaction_manager *tm, dm_block_t b)
104 {
105         int r = 0;
106         unsigned bucket = dm_hash_block(b, DM_HASH_MASK);
107         struct shadow_info *si;
108
109         spin_lock(&tm->lock);
110         hlist_for_each_entry(si, tm->buckets + bucket, hlist)
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  */
124 static 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;
132                 bucket = dm_hash_block(b, DM_HASH_MASK);
133                 spin_lock(&tm->lock);
134                 hlist_add_head(&si->hlist, tm->buckets + bucket);
135                 spin_unlock(&tm->lock);
136         }
137 }
138
139 static void wipe_shadow_table(struct dm_transaction_manager *tm)
140 {
141         struct shadow_info *si;
142         struct hlist_node *tmp;
143         struct hlist_head *bucket;
144         int i;
145
146         spin_lock(&tm->lock);
147         for (i = 0; i < DM_HASH_SIZE; i++) {
148                 bucket = tm->buckets + i;
149                 hlist_for_each_entry_safe(si, tmp, bucket, hlist)
150                         kfree(si);
151
152                 INIT_HLIST_HEAD(bucket);
153         }
154
155         spin_unlock(&tm->lock);
156 }
157
158 /*----------------------------------------------------------------*/
159
160 static 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);
176         for (i = 0; i < DM_HASH_SIZE; i++)
177                 INIT_HLIST_HEAD(tm->buckets + i);
178
179         prefetch_init(&tm->prefetches);
180
181         return tm;
182 }
183
184 struct 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 }
196 EXPORT_SYMBOL_GPL(dm_tm_create_non_blocking_clone);
197
198 void dm_tm_destroy(struct dm_transaction_manager *tm)
199 {
200         if (!tm->is_clone)
201                 wipe_shadow_table(tm);
202
203         kfree(tm);
204 }
205 EXPORT_SYMBOL_GPL(dm_tm_destroy);
206
207 int 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
218         return dm_bm_flush(tm->bm);
219 }
220 EXPORT_SYMBOL_GPL(dm_tm_pre_commit);
221
222 int 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);
228         dm_bm_unlock(root);
229
230         return dm_bm_flush(tm->bm);
231 }
232 EXPORT_SYMBOL_GPL(dm_tm_commit);
233
234 int 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
263 static 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
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) {
292                 dm_bm_unlock(orig_block);
293                 return r;
294         }
295
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;
301 }
302
303 int 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 }
326 EXPORT_SYMBOL_GPL(dm_tm_shadow_block);
327
328 int 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 {
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         }
340
341         return dm_bm_read_lock(tm->bm, b, v, blk);
342 }
343 EXPORT_SYMBOL_GPL(dm_tm_read_lock);
344
345 int dm_tm_unlock(struct dm_transaction_manager *tm, struct dm_block *b)
346 {
347         return dm_bm_unlock(b);
348 }
349 EXPORT_SYMBOL_GPL(dm_tm_unlock);
350
351 void 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 }
360 EXPORT_SYMBOL_GPL(dm_tm_inc);
361
362 void dm_tm_dec(struct dm_transaction_manager *tm, dm_block_t b)
363 {
364         /*
365          * The non-blocking clone doesn't support this.
366          */
367         BUG_ON(tm->is_clone);
368
369         dm_sm_dec_block(tm->sm, b);
370 }
371 EXPORT_SYMBOL_GPL(dm_tm_dec);
372
373 int dm_tm_ref(struct dm_transaction_manager *tm, dm_block_t b,
374               uint32_t *result)
375 {
376         if (tm->is_clone)
377                 return -EWOULDBLOCK;
378
379         return dm_sm_get_count(tm->sm, b, result);
380 }
381
382 struct dm_block_manager *dm_tm_get_bm(struct dm_transaction_manager *tm)
383 {
384         return tm->bm;
385 }
386
387 void dm_tm_issue_prefetches(struct dm_transaction_manager *tm)
388 {
389         prefetch_issue(&tm->prefetches, tm->bm);
390 }
391 EXPORT_SYMBOL_GPL(dm_tm_issue_prefetches);
392
393 /*----------------------------------------------------------------*/
394
395 static int dm_tm_create_internal(struct dm_block_manager *bm,
396                                  dm_block_t sb_location,
397                                  struct dm_transaction_manager **tm,
398                                  struct dm_space_map **sm,
399                                  int create,
400                                  void *sm_root, size_t sm_len)
401 {
402         int r;
403
404         *sm = dm_sm_metadata_init();
405         if (IS_ERR(*sm))
406                 return PTR_ERR(*sm);
407
408         *tm = dm_tm_create(bm, *sm);
409         if (IS_ERR(*tm)) {
410                 dm_sm_destroy(*sm);
411                 return PTR_ERR(*tm);
412         }
413
414         if (create) {
415                 r = dm_sm_metadata_create(*sm, *tm, dm_bm_nr_blocks(bm),
416                                           sb_location);
417                 if (r) {
418                         DMERR("couldn't create metadata space map");
419                         goto bad;
420                 }
421
422         } else {
423                 r = dm_sm_metadata_open(*sm, *tm, sm_root, sm_len);
424                 if (r) {
425                         DMERR("couldn't open metadata space map");
426                         goto bad;
427                 }
428         }
429
430         return 0;
431
432 bad:
433         dm_tm_destroy(*tm);
434         dm_sm_destroy(*sm);
435         return r;
436 }
437
438 int dm_tm_create_with_sm(struct dm_block_manager *bm, dm_block_t sb_location,
439                          struct dm_transaction_manager **tm,
440                          struct dm_space_map **sm)
441 {
442         return dm_tm_create_internal(bm, sb_location, tm, sm, 1, NULL, 0);
443 }
444 EXPORT_SYMBOL_GPL(dm_tm_create_with_sm);
445
446 int dm_tm_open_with_sm(struct dm_block_manager *bm, dm_block_t sb_location,
447                        void *sm_root, size_t root_len,
448                        struct dm_transaction_manager **tm,
449                        struct dm_space_map **sm)
450 {
451         return dm_tm_create_internal(bm, sb_location, tm, sm, 0, sm_root, root_len);
452 }
453 EXPORT_SYMBOL_GPL(dm_tm_open_with_sm);
454
455 /*----------------------------------------------------------------*/