Imported Upstream version 2.28.0
[platform/upstream/git.git] / bloom.c
1 #include "git-compat-util.h"
2 #include "bloom.h"
3 #include "diff.h"
4 #include "diffcore.h"
5 #include "revision.h"
6 #include "hashmap.h"
7 #include "commit-graph.h"
8 #include "commit.h"
9
10 define_commit_slab(bloom_filter_slab, struct bloom_filter);
11
12 static struct bloom_filter_slab bloom_filters;
13
14 struct pathmap_hash_entry {
15     struct hashmap_entry entry;
16     const char path[FLEX_ARRAY];
17 };
18
19 static uint32_t rotate_left(uint32_t value, int32_t count)
20 {
21         uint32_t mask = 8 * sizeof(uint32_t) - 1;
22         count &= mask;
23         return ((value << count) | (value >> ((-count) & mask)));
24 }
25
26 static inline unsigned char get_bitmask(uint32_t pos)
27 {
28         return ((unsigned char)1) << (pos & (BITS_PER_WORD - 1));
29 }
30
31 static int load_bloom_filter_from_graph(struct commit_graph *g,
32                                         struct bloom_filter *filter,
33                                         struct commit *c)
34 {
35         uint32_t lex_pos, start_index, end_index;
36         uint32_t graph_pos = commit_graph_position(c);
37
38         while (graph_pos < g->num_commits_in_base)
39                 g = g->base_graph;
40
41         /* The commit graph commit 'c' lives in doesn't carry bloom filters. */
42         if (!g->chunk_bloom_indexes)
43                 return 0;
44
45         lex_pos = graph_pos - g->num_commits_in_base;
46
47         end_index = get_be32(g->chunk_bloom_indexes + 4 * lex_pos);
48
49         if (lex_pos > 0)
50                 start_index = get_be32(g->chunk_bloom_indexes + 4 * (lex_pos - 1));
51         else
52                 start_index = 0;
53
54         filter->len = end_index - start_index;
55         filter->data = (unsigned char *)(g->chunk_bloom_data +
56                                         sizeof(unsigned char) * start_index +
57                                         BLOOMDATA_CHUNK_HEADER_SIZE);
58
59         return 1;
60 }
61
62 /*
63  * Calculate the murmur3 32-bit hash value for the given data
64  * using the given seed.
65  * Produces a uniformly distributed hash value.
66  * Not considered to be cryptographically secure.
67  * Implemented as described in https://en.wikipedia.org/wiki/MurmurHash#Algorithm
68  */
69 uint32_t murmur3_seeded(uint32_t seed, const char *data, size_t len)
70 {
71         const uint32_t c1 = 0xcc9e2d51;
72         const uint32_t c2 = 0x1b873593;
73         const uint32_t r1 = 15;
74         const uint32_t r2 = 13;
75         const uint32_t m = 5;
76         const uint32_t n = 0xe6546b64;
77         int i;
78         uint32_t k1 = 0;
79         const char *tail;
80
81         int len4 = len / sizeof(uint32_t);
82
83         uint32_t k;
84         for (i = 0; i < len4; i++) {
85                 uint32_t byte1 = (uint32_t)data[4*i];
86                 uint32_t byte2 = ((uint32_t)data[4*i + 1]) << 8;
87                 uint32_t byte3 = ((uint32_t)data[4*i + 2]) << 16;
88                 uint32_t byte4 = ((uint32_t)data[4*i + 3]) << 24;
89                 k = byte1 | byte2 | byte3 | byte4;
90                 k *= c1;
91                 k = rotate_left(k, r1);
92                 k *= c2;
93
94                 seed ^= k;
95                 seed = rotate_left(seed, r2) * m + n;
96         }
97
98         tail = (data + len4 * sizeof(uint32_t));
99
100         switch (len & (sizeof(uint32_t) - 1)) {
101         case 3:
102                 k1 ^= ((uint32_t)tail[2]) << 16;
103                 /*-fallthrough*/
104         case 2:
105                 k1 ^= ((uint32_t)tail[1]) << 8;
106                 /*-fallthrough*/
107         case 1:
108                 k1 ^= ((uint32_t)tail[0]) << 0;
109                 k1 *= c1;
110                 k1 = rotate_left(k1, r1);
111                 k1 *= c2;
112                 seed ^= k1;
113                 break;
114         }
115
116         seed ^= (uint32_t)len;
117         seed ^= (seed >> 16);
118         seed *= 0x85ebca6b;
119         seed ^= (seed >> 13);
120         seed *= 0xc2b2ae35;
121         seed ^= (seed >> 16);
122
123         return seed;
124 }
125
126 void fill_bloom_key(const char *data,
127                     size_t len,
128                     struct bloom_key *key,
129                     const struct bloom_filter_settings *settings)
130 {
131         int i;
132         const uint32_t seed0 = 0x293ae76f;
133         const uint32_t seed1 = 0x7e646e2c;
134         const uint32_t hash0 = murmur3_seeded(seed0, data, len);
135         const uint32_t hash1 = murmur3_seeded(seed1, data, len);
136
137         key->hashes = (uint32_t *)xcalloc(settings->num_hashes, sizeof(uint32_t));
138         for (i = 0; i < settings->num_hashes; i++)
139                 key->hashes[i] = hash0 + i * hash1;
140 }
141
142 void clear_bloom_key(struct bloom_key *key)
143 {
144         FREE_AND_NULL(key->hashes);
145 }
146
147 void add_key_to_filter(const struct bloom_key *key,
148                        struct bloom_filter *filter,
149                        const struct bloom_filter_settings *settings)
150 {
151         int i;
152         uint64_t mod = filter->len * BITS_PER_WORD;
153
154         for (i = 0; i < settings->num_hashes; i++) {
155                 uint64_t hash_mod = key->hashes[i] % mod;
156                 uint64_t block_pos = hash_mod / BITS_PER_WORD;
157
158                 filter->data[block_pos] |= get_bitmask(hash_mod);
159         }
160 }
161
162 void init_bloom_filters(void)
163 {
164         init_bloom_filter_slab(&bloom_filters);
165 }
166
167 static int pathmap_cmp(const void *hashmap_cmp_fn_data,
168                        const struct hashmap_entry *eptr,
169                        const struct hashmap_entry *entry_or_key,
170                        const void *keydata)
171 {
172         const struct pathmap_hash_entry *e1, *e2;
173
174         e1 = container_of(eptr, const struct pathmap_hash_entry, entry);
175         e2 = container_of(entry_or_key, const struct pathmap_hash_entry, entry);
176
177         return strcmp(e1->path, e2->path);
178 }
179
180 struct bloom_filter *get_bloom_filter(struct repository *r,
181                                       struct commit *c,
182                                       int compute_if_not_present)
183 {
184         struct bloom_filter *filter;
185         struct bloom_filter_settings settings = DEFAULT_BLOOM_FILTER_SETTINGS;
186         int i;
187         struct diff_options diffopt;
188         int max_changes = 512;
189
190         if (bloom_filters.slab_size == 0)
191                 return NULL;
192
193         filter = bloom_filter_slab_at(&bloom_filters, c);
194
195         if (!filter->data) {
196                 load_commit_graph_info(r, c);
197                 if (commit_graph_position(c) != COMMIT_NOT_FROM_GRAPH &&
198                         r->objects->commit_graph->chunk_bloom_indexes) {
199                         if (load_bloom_filter_from_graph(r->objects->commit_graph, filter, c))
200                                 return filter;
201                         else
202                                 return NULL;
203                 }
204         }
205
206         if (filter->data || !compute_if_not_present)
207                 return filter;
208
209         repo_diff_setup(r, &diffopt);
210         diffopt.flags.recursive = 1;
211         diffopt.detect_rename = 0;
212         diffopt.max_changes = max_changes;
213         diff_setup_done(&diffopt);
214
215         /* ensure commit is parsed so we have parent information */
216         repo_parse_commit(r, c);
217
218         if (c->parents)
219                 diff_tree_oid(&c->parents->item->object.oid, &c->object.oid, "", &diffopt);
220         else
221                 diff_tree_oid(NULL, &c->object.oid, "", &diffopt);
222         diffcore_std(&diffopt);
223
224         if (diffopt.num_changes <= max_changes) {
225                 struct hashmap pathmap;
226                 struct pathmap_hash_entry *e;
227                 struct hashmap_iter iter;
228                 hashmap_init(&pathmap, pathmap_cmp, NULL, 0);
229
230                 for (i = 0; i < diff_queued_diff.nr; i++) {
231                         const char *path = diff_queued_diff.queue[i]->two->path;
232
233                         /*
234                          * Add each leading directory of the changed file, i.e. for
235                          * 'dir/subdir/file' add 'dir' and 'dir/subdir' as well, so
236                          * the Bloom filter could be used to speed up commands like
237                          * 'git log dir/subdir', too.
238                          *
239                          * Note that directories are added without the trailing '/'.
240                          */
241                         do {
242                                 char *last_slash = strrchr(path, '/');
243
244                                 FLEX_ALLOC_STR(e, path, path);
245                                 hashmap_entry_init(&e->entry, strhash(path));
246
247                                 if (!hashmap_get(&pathmap, &e->entry, NULL))
248                                         hashmap_add(&pathmap, &e->entry);
249                                 else
250                                         free(e);
251
252                                 if (!last_slash)
253                                         last_slash = (char*)path;
254                                 *last_slash = '\0';
255
256                         } while (*path);
257
258                         diff_free_filepair(diff_queued_diff.queue[i]);
259                 }
260
261                 filter->len = (hashmap_get_size(&pathmap) * settings.bits_per_entry + BITS_PER_WORD - 1) / BITS_PER_WORD;
262                 filter->data = xcalloc(filter->len, sizeof(unsigned char));
263
264                 hashmap_for_each_entry(&pathmap, &iter, e, entry) {
265                         struct bloom_key key;
266                         fill_bloom_key(e->path, strlen(e->path), &key, &settings);
267                         add_key_to_filter(&key, filter, &settings);
268                 }
269
270                 hashmap_free_entries(&pathmap, struct pathmap_hash_entry, entry);
271         } else {
272                 for (i = 0; i < diff_queued_diff.nr; i++)
273                         diff_free_filepair(diff_queued_diff.queue[i]);
274                 filter->data = NULL;
275                 filter->len = 0;
276         }
277
278         free(diff_queued_diff.queue);
279         DIFF_QUEUE_CLEAR(&diff_queued_diff);
280
281         return filter;
282 }
283
284 int bloom_filter_contains(const struct bloom_filter *filter,
285                           const struct bloom_key *key,
286                           const struct bloom_filter_settings *settings)
287 {
288         int i;
289         uint64_t mod = filter->len * BITS_PER_WORD;
290
291         if (!mod)
292                 return -1;
293
294         for (i = 0; i < settings->num_hashes; i++) {
295                 uint64_t hash_mod = key->hashes[i] % mod;
296                 uint64_t block_pos = hash_mod / BITS_PER_WORD;
297                 if (!(filter->data[block_pos] & get_bitmask(hash_mod)))
298                         return 0;
299         }
300
301         return 1;
302 }