1 /*
2  * Copyright © 2014 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  */
23 
24 #ifdef ENABLE_SHADER_CACHE
25 
26 #include <ctype.h>
27 #include <string.h>
28 #include <stdlib.h>
29 #include <stdio.h>
30 #include <sys/file.h>
31 #include <sys/types.h>
32 #include <sys/stat.h>
33 #include <sys/mman.h>
34 #include <unistd.h>
35 #include <fcntl.h>
36 #include <pwd.h>
37 #include <errno.h>
38 #include <dirent.h>
39 
40 #include "util/u_atomic.h"
41 #include "util/mesa-sha1.h"
42 #include "util/ralloc.h"
43 #include "main/errors.h"
44 
45 #include "disk_cache.h"
46 
47 /* Number of bits to mask off from a cache key to get an index. */
48 #define CACHE_INDEX_KEY_BITS 16
49 
50 /* Mask for computing an index from a key. */
51 #define CACHE_INDEX_KEY_MASK ((1 << CACHE_INDEX_KEY_BITS) - 1)
52 
53 /* The number of keys that can be stored in the index. */
54 #define CACHE_INDEX_MAX_KEYS (1 << CACHE_INDEX_KEY_BITS)
55 
56 struct disk_cache {
57    /* The path to the cache directory. */
58    char *path;
59 
60    /* A pointer to the mmapped index file within the cache directory. */
61    uint8_t *index_mmap;
62    size_t index_mmap_size;
63 
64    /* Pointer to total size of all objects in cache (within index_mmap) */
65    uint64_t *size;
66 
67    /* Pointer to stored keys, (within index_mmap). */
68    uint8_t *stored_keys;
69 
70    /* Maximum size of all cached objects (in bytes). */
71    uint64_t max_size;
72 };
73 
74 /* Create a directory named 'path' if it does not already exist.
75  *
76  * Returns: 0 if path already exists as a directory or if created.
77  *         -1 in all other cases.
78  */
79 static int
mkdir_if_needed(char * path)80 mkdir_if_needed(char *path)
81 {
82    struct stat sb;
83 
84    /* If the path exists already, then our work is done if it's a
85     * directory, but it's an error if it is not.
86     */
87    if (stat(path, &sb) == 0) {
88       if (S_ISDIR(sb.st_mode)) {
89          return 0;
90       } else {
91          fprintf(stderr, "Cannot use %s for shader cache (not a directory)"
92                          "---disabling.\n", path);
93          return -1;
94       }
95    }
96 
97    int ret = mkdir(path, 0755);
98    if (ret == 0 || (ret == -1 && errno == EEXIST))
99      return 0;
100 
101    fprintf(stderr, "Failed to create %s for shader cache (%s)---disabling.\n",
102            path, strerror(errno));
103 
104    return -1;
105 }
106 
107 /* Concatenate an existing path and a new name to form a new path.  If the new
108  * path does not exist as a directory, create it then return the resulting
109  * name of the new path (ralloc'ed off of 'ctx').
110  *
111  * Returns NULL on any error, such as:
112  *
113  *      <path> does not exist or is not a directory
114  *      <path>/<name> exists but is not a directory
115  *      <path>/<name> cannot be created as a directory
116  */
117 static char *
concatenate_and_mkdir(void * ctx,char * path,char * name)118 concatenate_and_mkdir(void *ctx, char *path, char *name)
119 {
120    char *new_path;
121    struct stat sb;
122 
123    if (stat(path, &sb) != 0 || ! S_ISDIR(sb.st_mode))
124       return NULL;
125 
126    new_path = ralloc_asprintf(ctx, "%s/%s", path, name);
127 
128    if (mkdir_if_needed(new_path) == 0)
129       return new_path;
130    else
131       return NULL;
132 }
133 
134 struct disk_cache *
disk_cache_create(void)135 disk_cache_create(void)
136 {
137    void *local;
138    struct disk_cache *cache = NULL;
139    char *path, *max_size_str;
140    uint64_t max_size;
141    int fd = -1;
142    struct stat sb;
143    size_t size;
144 
145    /* A ralloc context for transient data during this invocation. */
146    local = ralloc_context(NULL);
147    if (local == NULL)
148       goto fail;
149 
150    /* At user request, disable shader cache entirely. */
151    if (getenv("MESA_GLSL_CACHE_DISABLE"))
152       goto fail;
153 
154    /* Determine path for cache based on the first defined name as follows:
155     *
156     *   $MESA_GLSL_CACHE_DIR
157     *   $XDG_CACHE_HOME/mesa
158     *   <pwd.pw_dir>/.cache/mesa
159     */
160    path = getenv("MESA_GLSL_CACHE_DIR");
161    if (path && mkdir_if_needed(path) == -1) {
162       goto fail;
163    }
164 
165    if (path == NULL) {
166       char *xdg_cache_home = getenv("XDG_CACHE_HOME");
167 
168       if (xdg_cache_home) {
169          if (mkdir_if_needed(xdg_cache_home) == -1)
170             goto fail;
171 
172          path = concatenate_and_mkdir(local, xdg_cache_home, "mesa");
173          if (path == NULL)
174             goto fail;
175       }
176    }
177 
178    if (path == NULL) {
179       char *buf;
180       size_t buf_size;
181       struct passwd pwd, *result;
182 
183       buf_size = sysconf(_SC_GETPW_R_SIZE_MAX);
184       if (buf_size == -1)
185          buf_size = 512;
186 
187       /* Loop until buf_size is large enough to query the directory */
188       while (1) {
189          buf = ralloc_size(local, buf_size);
190 
191          getpwuid_r(getuid(), &pwd, buf, buf_size, &result);
192          if (result)
193             break;
194 
195          if (errno == ERANGE) {
196             ralloc_free(buf);
197             buf = NULL;
198             buf_size *= 2;
199          } else {
200             goto fail;
201          }
202       }
203 
204       path = concatenate_and_mkdir(local, pwd.pw_dir, ".cache");
205       if (path == NULL)
206          goto fail;
207 
208       path = concatenate_and_mkdir(local, path, "mesa");
209       if (path == NULL)
210          goto fail;
211    }
212 
213    cache = ralloc(NULL, struct disk_cache);
214    if (cache == NULL)
215       goto fail;
216 
217    cache->path = ralloc_strdup(cache, path);
218    if (cache->path == NULL)
219       goto fail;
220 
221    path = ralloc_asprintf(local, "%s/index", cache->path);
222    if (path == NULL)
223       goto fail;
224 
225    fd = open(path, O_RDWR | O_CREAT | O_CLOEXEC, 0644);
226    if (fd == -1)
227       goto fail;
228 
229    if (fstat(fd, &sb) == -1)
230       goto fail;
231 
232    /* Force the index file to be the expected size. */
233    size = sizeof(*cache->size) + CACHE_INDEX_MAX_KEYS * CACHE_KEY_SIZE;
234    if (sb.st_size != size) {
235       if (ftruncate(fd, size) == -1)
236          goto fail;
237    }
238 
239    /* We map this shared so that other processes see updates that we
240     * make.
241     *
242     * Note: We do use atomic addition to ensure that multiple
243     * processes don't scramble the cache size recorded in the
244     * index. But we don't use any locking to prevent multiple
245     * processes from updating the same entry simultaneously. The idea
246     * is that if either result lands entirely in the index, then
247     * that's equivalent to a well-ordered write followed by an
248     * eviction and a write. On the other hand, if the simultaneous
249     * writes result in a corrupt entry, that's not really any
250     * different than both entries being evicted, (since within the
251     * guarantees of the cryptographic hash, a corrupt entry is
252     * unlikely to ever match a real cache key).
253     */
254    cache->index_mmap = mmap(NULL, size, PROT_READ | PROT_WRITE,
255                             MAP_SHARED, fd, 0);
256    if (cache->index_mmap == MAP_FAILED)
257       goto fail;
258    cache->index_mmap_size = size;
259 
260    close(fd);
261 
262    cache->size = (uint64_t *) cache->index_mmap;
263    cache->stored_keys = cache->index_mmap + sizeof(uint64_t);
264 
265    max_size = 0;
266 
267    max_size_str = getenv("MESA_GLSL_CACHE_MAX_SIZE");
268    if (max_size_str) {
269       char *end;
270       max_size = strtoul(max_size_str, &end, 10);
271       if (end == max_size_str) {
272          max_size = 0;
273       } else {
274          while (*end && isspace(*end))
275             end++;
276          switch (*end) {
277          case 'K':
278          case 'k':
279             max_size *= 1024;
280             break;
281          case 'M':
282          case 'm':
283             max_size *= 1024*1024;
284             break;
285          case '\0':
286          case 'G':
287          case 'g':
288          default:
289             max_size *= 1024*1024*1024;
290             break;
291          }
292       }
293    }
294 
295    /* Default to 1GB for maximum cache size. */
296    if (max_size == 0)
297       max_size = 1024*1024*1024;
298 
299    cache->max_size = max_size;
300 
301    ralloc_free(local);
302 
303    return cache;
304 
305  fail:
306    if (fd != -1)
307       close(fd);
308    if (cache)
309       ralloc_free(cache);
310    ralloc_free(local);
311 
312    return NULL;
313 }
314 
315 void
disk_cache_destroy(struct disk_cache * cache)316 disk_cache_destroy(struct disk_cache *cache)
317 {
318    munmap(cache->index_mmap, cache->index_mmap_size);
319 
320    ralloc_free(cache);
321 }
322 
323 /* Return a filename within the cache's directory corresponding to 'key'. The
324  * returned filename is ralloced with 'cache' as the parent context.
325  *
326  * Returns NULL if out of memory.
327  */
328 static char *
get_cache_file(struct disk_cache * cache,cache_key key)329 get_cache_file(struct disk_cache *cache, cache_key key)
330 {
331    char buf[41];
332 
333    _mesa_sha1_format(buf, key);
334 
335    return ralloc_asprintf(cache, "%s/%c%c/%s",
336                           cache->path, buf[0], buf[1], buf + 2);
337 }
338 
339 /* Create the directory that will be needed for the cache file for \key.
340  *
341  * Obviously, the implementation here must closely match
342  * _get_cache_file above.
343 */
344 static void
make_cache_file_directory(struct disk_cache * cache,cache_key key)345 make_cache_file_directory(struct disk_cache *cache, cache_key key)
346 {
347    char *dir;
348    char buf[41];
349 
350    _mesa_sha1_format(buf, key);
351 
352    dir = ralloc_asprintf(cache, "%s/%c%c", cache->path, buf[0], buf[1]);
353 
354    mkdir_if_needed(dir);
355 
356    ralloc_free(dir);
357 }
358 
359 /* Given a directory path and predicate function, count all entries in
360  * that directory for which the predicate returns true. Then choose a
361  * random entry from among those counted.
362  *
363  * Returns: A malloc'ed string for the path to the chosen file, (or
364  * NULL on any error). The caller should free the string when
365  * finished.
366  */
367 static char *
choose_random_file_matching(const char * dir_path,bool (* predicate)(struct dirent *,const char * dir_path))368 choose_random_file_matching(const char *dir_path,
369                             bool (*predicate)(struct dirent *,
370                                               const char *dir_path))
371 {
372    DIR *dir;
373    struct dirent *entry;
374    unsigned int count, victim;
375    char *filename;
376 
377    dir = opendir(dir_path);
378    if (dir == NULL)
379       return NULL;
380 
381    count = 0;
382 
383    while (1) {
384       entry = readdir(dir);
385       if (entry == NULL)
386          break;
387       if (!predicate(entry, dir_path))
388          continue;
389 
390       count++;
391    }
392 
393    if (count == 0) {
394       closedir(dir);
395       return NULL;
396    }
397 
398    victim = rand() % count;
399 
400    rewinddir(dir);
401    count = 0;
402 
403    while (1) {
404       entry = readdir(dir);
405       if (entry == NULL)
406          break;
407       if (!predicate(entry, dir_path))
408          continue;
409       if (count == victim)
410          break;
411 
412       count++;
413    }
414 
415    if (entry == NULL) {
416       closedir(dir);
417       return NULL;
418    }
419 
420    if (asprintf(&filename, "%s/%s", dir_path, entry->d_name) < 0)
421       filename = NULL;
422 
423    closedir(dir);
424 
425    return filename;
426 }
427 
428 /* Is entry a regular file, and not having a name with a trailing
429  * ".tmp"
430  */
431 static bool
is_regular_non_tmp_file(struct dirent * entry,const char * path)432 is_regular_non_tmp_file(struct dirent *entry, const char *path)
433 {
434    char *filename;
435    if (asprintf(&filename, "%s/%s", path, entry->d_name) == -1)
436       return false;
437 
438    struct stat sb;
439    int res = stat(filename, &sb);
440    free(filename);
441 
442    if (res == -1 || !S_ISREG(sb.st_mode))
443       return false;
444 
445    size_t len = strlen (entry->d_name);
446    if (len >= 4 && strcmp(&entry->d_name[len-4], ".tmp") == 0)
447       return false;
448 
449    return true;
450 }
451 
452 /* Returns the size of the deleted file, (or 0 on any error). */
453 static size_t
unlink_random_file_from_directory(const char * path)454 unlink_random_file_from_directory(const char *path)
455 {
456    struct stat sb;
457    char *filename;
458 
459    filename = choose_random_file_matching(path, is_regular_non_tmp_file);
460    if (filename == NULL)
461       return 0;
462 
463    if (stat(filename, &sb) == -1) {
464       free (filename);
465       return 0;
466    }
467 
468    unlink(filename);
469 
470    free (filename);
471 
472    return sb.st_size;
473 }
474 
475 /* Is entry a directory with a two-character name, (and not the
476  * special name of "..")
477  */
478 static bool
is_two_character_sub_directory(struct dirent * entry,const char * path)479 is_two_character_sub_directory(struct dirent *entry, const char *path)
480 {
481    char *subdir;
482    if (asprintf(&subdir, "%s/%s", path, entry->d_name) == -1)
483       return false;
484 
485    struct stat sb;
486    int res = stat(subdir, &sb);
487    free(subdir);
488 
489    if (res == -1 || !S_ISDIR(sb.st_mode))
490       return false;
491 
492    if (strlen(entry->d_name) != 2)
493       return false;
494 
495    if (strcmp(entry->d_name, "..") == 0)
496       return false;
497 
498    return true;
499 }
500 
501 static void
evict_random_item(struct disk_cache * cache)502 evict_random_item(struct disk_cache *cache)
503 {
504    const char hex[] = "0123456789abcde";
505    char *dir_path;
506    int a, b;
507    size_t size;
508 
509    /* With a reasonably-sized, full cache, (and with keys generated
510     * from a cryptographic hash), we can choose two random hex digits
511     * and reasonably expect the directory to exist with a file in it.
512     */
513    a = rand() % 16;
514    b = rand() % 16;
515 
516    if (asprintf(&dir_path, "%s/%c%c", cache->path, hex[a], hex[b]) < 0)
517       return;
518 
519    size = unlink_random_file_from_directory(dir_path);
520 
521    free(dir_path);
522 
523    if (size) {
524       p_atomic_add(cache->size, - size);
525       return;
526    }
527 
528    /* In the case where the random choice of directory didn't find
529     * something, we choose randomly from the existing directories.
530     *
531     * Really, the only reason this code exists is to allow the unit
532     * tests to work, (which use an artificially-small cache to be able
533     * to force a single cached item to be evicted).
534     */
535    dir_path = choose_random_file_matching(cache->path,
536                                           is_two_character_sub_directory);
537    if (dir_path == NULL)
538       return;
539 
540    size = unlink_random_file_from_directory(dir_path);
541 
542    free(dir_path);
543 
544    if (size)
545       p_atomic_add(cache->size, - size);
546 }
547 
548 void
disk_cache_put(struct disk_cache * cache,cache_key key,const void * data,size_t size)549 disk_cache_put(struct disk_cache *cache,
550           cache_key key,
551           const void *data,
552           size_t size)
553 {
554    int fd = -1, fd_final = -1, err, ret;
555    size_t len;
556    char *filename = NULL, *filename_tmp = NULL;
557    const char *p = data;
558 
559    filename = get_cache_file(cache, key);
560    if (filename == NULL)
561       goto done;
562 
563    /* Write to a temporary file to allow for an atomic rename to the
564     * final destination filename, (to prevent any readers from seeing
565     * a partially written file).
566     */
567    filename_tmp = ralloc_asprintf(cache, "%s.tmp", filename);
568    if (filename_tmp == NULL)
569       goto done;
570 
571    fd = open(filename_tmp, O_WRONLY | O_CLOEXEC | O_CREAT, 0644);
572 
573    /* Make the two-character subdirectory within the cache as needed. */
574    if (fd == -1) {
575       if (errno != ENOENT)
576          goto done;
577 
578       make_cache_file_directory(cache, key);
579 
580       fd = open(filename_tmp, O_WRONLY | O_CLOEXEC | O_CREAT, 0644);
581       if (fd == -1)
582          goto done;
583    }
584 
585    /* With the temporary file open, we take an exclusive flock on
586     * it. If the flock fails, then another process still has the file
587     * open with the flock held. So just let that file be responsible
588     * for writing the file.
589     */
590    err = flock(fd, LOCK_EX | LOCK_NB);
591    if (err == -1)
592       goto done;
593 
594    /* Now that we have the lock on the open temporary file, we can
595     * check to see if the destination file already exists. If so,
596     * another process won the race between when we saw that the file
597     * didn't exist and now. In this case, we don't do anything more,
598     * (to ensure the size accounting of the cache doesn't get off).
599     */
600    fd_final = open(filename, O_RDONLY | O_CLOEXEC);
601    if (fd_final != -1)
602       goto done;
603 
604    /* OK, we're now on the hook to write out a file that we know is
605     * not in the cache, and is also not being written out to the cache
606     * by some other process.
607     *
608     * Before we do that, if the cache is too large, evict something
609     * else first.
610     */
611    if (*cache->size + size > cache->max_size)
612       evict_random_item(cache);
613 
614    /* Now, finally, write out the contents to the temporary file, then
615     * rename them atomically to the destination filename, and also
616     * perform an atomic increment of the total cache size.
617     */
618    for (len = 0; len < size; len += ret) {
619       ret = write(fd, p + len, size - len);
620       if (ret == -1) {
621          unlink(filename_tmp);
622          goto done;
623       }
624    }
625 
626    rename(filename_tmp, filename);
627 
628    p_atomic_add(cache->size, size);
629 
630  done:
631    if (fd_final != -1)
632       close(fd_final);
633    /* This close finally releases the flock, (now that the final dile
634     * has been renamed into place and the size has been added).
635     */
636    if (fd != -1)
637       close(fd);
638    if (filename_tmp)
639       ralloc_free(filename_tmp);
640    if (filename)
641       ralloc_free(filename);
642 }
643 
644 void *
disk_cache_get(struct disk_cache * cache,cache_key key,size_t * size)645 disk_cache_get(struct disk_cache *cache, cache_key key, size_t *size)
646 {
647    int fd = -1, ret, len;
648    struct stat sb;
649    char *filename = NULL;
650    uint8_t *data = NULL;
651 
652    if (size)
653       *size = 0;
654 
655    filename = get_cache_file(cache, key);
656    if (filename == NULL)
657       goto fail;
658 
659    fd = open(filename, O_RDONLY | O_CLOEXEC);
660    if (fd == -1)
661       goto fail;
662 
663    if (fstat(fd, &sb) == -1)
664       goto fail;
665 
666    data = malloc(sb.st_size);
667    if (data == NULL)
668       goto fail;
669 
670    for (len = 0; len < sb.st_size; len += ret) {
671       ret = read(fd, data + len, sb.st_size - len);
672       if (ret == -1)
673          goto fail;
674    }
675 
676    ralloc_free(filename);
677    close(fd);
678 
679    if (size)
680       *size = sb.st_size;
681 
682    return data;
683 
684  fail:
685    if (data)
686       free(data);
687    if (filename)
688       ralloc_free(filename);
689    if (fd != -1)
690       close(fd);
691 
692    return NULL;
693 }
694 
695 void
disk_cache_put_key(struct disk_cache * cache,cache_key key)696 disk_cache_put_key(struct disk_cache *cache, cache_key key)
697 {
698    uint32_t *key_chunk = (uint32_t *) key;
699    int i = *key_chunk & CACHE_INDEX_KEY_MASK;
700    unsigned char *entry;
701 
702    entry = &cache->stored_keys[i + CACHE_KEY_SIZE];
703 
704    memcpy(entry, key, CACHE_KEY_SIZE);
705 }
706 
707 /* This function lets us test whether a given key was previously
708  * stored in the cache with disk_cache_put_key(). The implement is
709  * efficient by not using syscalls or hitting the disk. It's not
710  * race-free, but the races are benign. If we race with someone else
711  * calling disk_cache_put_key, then that's just an extra cache miss and an
712  * extra recompile.
713  */
714 bool
disk_cache_has_key(struct disk_cache * cache,cache_key key)715 disk_cache_has_key(struct disk_cache *cache, cache_key key)
716 {
717    uint32_t *key_chunk = (uint32_t *) key;
718    int i = *key_chunk & CACHE_INDEX_KEY_MASK;
719    unsigned char *entry;
720 
721    entry = &cache->stored_keys[i + CACHE_KEY_SIZE];
722 
723    return memcmp(entry, key, CACHE_KEY_SIZE) == 0;
724 }
725 
726 #endif /* ENABLE_SHADER_CACHE */
727