2 * Copyright © 2014 Intel Corporation
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:
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
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
24 #ifdef ENABLE_SHADER_CACHE
32 #include <sys/types.h>
42 #include "util/crc32.h"
43 #include "util/rand_xor.h"
44 #include "util/u_atomic.h"
45 #include "util/u_queue.h"
46 #include "util/mesa-sha1.h"
47 #include "util/ralloc.h"
48 #include "main/errors.h"
50 #include "disk_cache.h"
52 /* Number of bits to mask off from a cache key to get an index. */
53 #define CACHE_INDEX_KEY_BITS 16
55 /* Mask for computing an index from a key. */
56 #define CACHE_INDEX_KEY_MASK ((1 << CACHE_INDEX_KEY_BITS) - 1)
58 /* The number of keys that can be stored in the index. */
59 #define CACHE_INDEX_MAX_KEYS (1 << CACHE_INDEX_KEY_BITS)
61 /* The cache version should be bumped whenever a change is made to the
62 * structure of cache entries or the index. This will give any 3rd party
63 * applications reading the cache entries a chance to adjust to the changes.
65 * - The cache version is checked internally when reading a cache entry. If we
66 * ever have a mismatch we are in big trouble as this means we had a cache
67 * collision. In case of such an event please check the skys for giant
68 * asteroids and that the entire Mesa team hasn't been eaten by wolves.
70 * - There is no strict requirement that cache versions be backwards
71 * compatible but effort should be taken to limit disruption where possible.
73 #define CACHE_VERSION 1
76 /* The path to the cache directory. */
79 /* Thread queue for compressing and writing cache entries to disk */
80 struct util_queue cache_queue
;
82 /* Seed for rand, which is used to pick a random directory */
83 uint64_t seed_xorshift128plus
[2];
85 /* A pointer to the mmapped index file within the cache directory. */
87 size_t index_mmap_size
;
89 /* Pointer to total size of all objects in cache (within index_mmap) */
92 /* Pointer to stored keys, (within index_mmap). */
95 /* Maximum size of all cached objects (in bytes). */
98 /* Driver cache keys. */
99 uint8_t *driver_keys_blob
;
100 size_t driver_keys_blob_size
;
103 struct disk_cache_put_job
{
104 struct util_queue_fence fence
;
106 struct disk_cache
*cache
;
110 /* Copy of cache data to be compressed and written. */
113 /* Size of data to be compressed and written. */
116 struct cache_item_metadata cache_item_metadata
;
119 /* Create a directory named 'path' if it does not already exist.
121 * Returns: 0 if path already exists as a directory or if created.
122 * -1 in all other cases.
125 mkdir_if_needed(const char *path
)
129 /* If the path exists already, then our work is done if it's a
130 * directory, but it's an error if it is not.
132 if (stat(path
, &sb
) == 0) {
133 if (S_ISDIR(sb
.st_mode
)) {
136 fprintf(stderr
, "Cannot use %s for shader cache (not a directory)"
137 "---disabling.\n", path
);
142 int ret
= mkdir(path
, 0755);
143 if (ret
== 0 || (ret
== -1 && errno
== EEXIST
))
146 fprintf(stderr
, "Failed to create %s for shader cache (%s)---disabling.\n",
147 path
, strerror(errno
));
152 /* Concatenate an existing path and a new name to form a new path. If the new
153 * path does not exist as a directory, create it then return the resulting
154 * name of the new path (ralloc'ed off of 'ctx').
156 * Returns NULL on any error, such as:
158 * <path> does not exist or is not a directory
159 * <path>/<name> exists but is not a directory
160 * <path>/<name> cannot be created as a directory
163 concatenate_and_mkdir(void *ctx
, const char *path
, const char *name
)
168 if (stat(path
, &sb
) != 0 || ! S_ISDIR(sb
.st_mode
))
171 new_path
= ralloc_asprintf(ctx
, "%s/%s", path
, name
);
173 if (mkdir_if_needed(new_path
) == 0)
179 #define DRV_KEY_CPY(_dst, _src, _src_size) \
181 memcpy(_dst, _src, _src_size); \
186 disk_cache_create(const char *gpu_name
, const char *timestamp
,
187 uint64_t driver_flags
)
190 struct disk_cache
*cache
= NULL
;
191 char *path
, *max_size_str
;
197 /* If running as a users other than the real user disable cache */
198 if (geteuid() != getuid())
201 /* A ralloc context for transient data during this invocation. */
202 local
= ralloc_context(NULL
);
206 /* At user request, disable shader cache entirely. */
207 if (getenv("MESA_GLSL_CACHE_DISABLE"))
210 /* Determine path for cache based on the first defined name as follows:
212 * $MESA_GLSL_CACHE_DIR
213 * $XDG_CACHE_HOME/mesa_shader_cache
214 * <pwd.pw_dir>/.cache/mesa_shader_cache
216 path
= getenv("MESA_GLSL_CACHE_DIR");
218 if (mkdir_if_needed(path
) == -1)
221 path
= concatenate_and_mkdir(local
, path
, CACHE_DIR_NAME
);
227 char *xdg_cache_home
= getenv("XDG_CACHE_HOME");
229 if (xdg_cache_home
) {
230 if (mkdir_if_needed(xdg_cache_home
) == -1)
233 path
= concatenate_and_mkdir(local
, xdg_cache_home
, CACHE_DIR_NAME
);
242 struct passwd pwd
, *result
;
244 buf_size
= sysconf(_SC_GETPW_R_SIZE_MAX
);
248 /* Loop until buf_size is large enough to query the directory */
250 buf
= ralloc_size(local
, buf_size
);
252 getpwuid_r(getuid(), &pwd
, buf
, buf_size
, &result
);
256 if (errno
== ERANGE
) {
265 path
= concatenate_and_mkdir(local
, pwd
.pw_dir
, ".cache");
269 path
= concatenate_and_mkdir(local
, path
, CACHE_DIR_NAME
);
274 cache
= ralloc(NULL
, struct disk_cache
);
278 cache
->path
= ralloc_strdup(cache
, path
);
279 if (cache
->path
== NULL
)
282 path
= ralloc_asprintf(local
, "%s/index", cache
->path
);
286 fd
= open(path
, O_RDWR
| O_CREAT
| O_CLOEXEC
, 0644);
290 if (fstat(fd
, &sb
) == -1)
293 /* Force the index file to be the expected size. */
294 size
= sizeof(*cache
->size
) + CACHE_INDEX_MAX_KEYS
* CACHE_KEY_SIZE
;
295 if (sb
.st_size
!= size
) {
296 if (ftruncate(fd
, size
) == -1)
300 /* We map this shared so that other processes see updates that we
303 * Note: We do use atomic addition to ensure that multiple
304 * processes don't scramble the cache size recorded in the
305 * index. But we don't use any locking to prevent multiple
306 * processes from updating the same entry simultaneously. The idea
307 * is that if either result lands entirely in the index, then
308 * that's equivalent to a well-ordered write followed by an
309 * eviction and a write. On the other hand, if the simultaneous
310 * writes result in a corrupt entry, that's not really any
311 * different than both entries being evicted, (since within the
312 * guarantees of the cryptographic hash, a corrupt entry is
313 * unlikely to ever match a real cache key).
315 cache
->index_mmap
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
317 if (cache
->index_mmap
== MAP_FAILED
)
319 cache
->index_mmap_size
= size
;
323 cache
->size
= (uint64_t *) cache
->index_mmap
;
324 cache
->stored_keys
= cache
->index_mmap
+ sizeof(uint64_t);
328 max_size_str
= getenv("MESA_GLSL_CACHE_MAX_SIZE");
331 max_size
= strtoul(max_size_str
, &end
, 10);
332 if (end
== max_size_str
) {
342 max_size
*= 1024*1024;
348 max_size
*= 1024*1024*1024;
354 /* Default to 1GB for maximum cache size. */
356 max_size
= 1024*1024*1024;
359 cache
->max_size
= max_size
;
361 /* A limit of 32 jobs was choosen as observations of Deus Ex start-up times
362 * showed that we reached at most 11 jobs on an Intel i5-6400 CPU@2.70GHz
363 * (a fairly modest desktop CPU). 1 thread was chosen because we don't
364 * really care about getting things to disk quickly just that it's not
365 * blocking other tasks.
367 util_queue_init(&cache
->cache_queue
, "disk_cache", 32, 1, 0);
369 uint8_t cache_version
= CACHE_VERSION
;
370 size_t cv_size
= sizeof(cache_version
);
371 cache
->driver_keys_blob_size
= cv_size
;
373 /* Create driver id keys */
374 size_t ts_size
= strlen(timestamp
) + 1;
375 size_t gpu_name_size
= strlen(gpu_name
) + 1;
376 cache
->driver_keys_blob_size
+= ts_size
;
377 cache
->driver_keys_blob_size
+= gpu_name_size
;
379 /* We sometimes store entire structs that contains a pointers in the cache,
380 * use pointer size as a key to avoid hard to debug issues.
382 uint8_t ptr_size
= sizeof(void *);
383 size_t ptr_size_size
= sizeof(ptr_size
);
384 cache
->driver_keys_blob_size
+= ptr_size_size
;
386 size_t driver_flags_size
= sizeof(driver_flags
);
387 cache
->driver_keys_blob_size
+= driver_flags_size
;
389 cache
->driver_keys_blob
=
390 ralloc_size(cache
, cache
->driver_keys_blob_size
);
391 if (!cache
->driver_keys_blob
)
394 uint8_t *drv_key_blob
= cache
->driver_keys_blob
;
395 DRV_KEY_CPY(drv_key_blob
, &cache_version
, cv_size
)
396 DRV_KEY_CPY(drv_key_blob
, timestamp
, ts_size
)
397 DRV_KEY_CPY(drv_key_blob
, gpu_name
, gpu_name_size
)
398 DRV_KEY_CPY(drv_key_blob
, &ptr_size
, ptr_size_size
)
399 DRV_KEY_CPY(drv_key_blob
, &driver_flags
, driver_flags_size
)
401 /* Seed our rand function */
402 s_rand_xorshift128plus(cache
->seed_xorshift128plus
, true);
419 disk_cache_destroy(struct disk_cache
*cache
)
422 util_queue_destroy(&cache
->cache_queue
);
423 munmap(cache
->index_mmap
, cache
->index_mmap_size
);
429 /* Return a filename within the cache's directory corresponding to 'key'. The
430 * returned filename is ralloced with 'cache' as the parent context.
432 * Returns NULL if out of memory.
435 get_cache_file(struct disk_cache
*cache
, const cache_key key
)
440 _mesa_sha1_format(buf
, key
);
441 if (asprintf(&filename
, "%s/%c%c/%s", cache
->path
, buf
[0],
442 buf
[1], buf
+ 2) == -1)
448 /* Create the directory that will be needed for the cache file for \key.
450 * Obviously, the implementation here must closely match
451 * _get_cache_file above.
454 make_cache_file_directory(struct disk_cache
*cache
, const cache_key key
)
459 _mesa_sha1_format(buf
, key
);
460 if (asprintf(&dir
, "%s/%c%c", cache
->path
, buf
[0], buf
[1]) == -1)
463 mkdir_if_needed(dir
);
467 /* Given a directory path and predicate function, find the entry with
468 * the oldest access time in that directory for which the predicate
471 * Returns: A malloc'ed string for the path to the chosen file, (or
472 * NULL on any error). The caller should free the string when
476 choose_lru_file_matching(const char *dir_path
,
477 bool (*predicate
)(const char *dir_path
,
479 const char *, const size_t))
482 struct dirent
*entry
;
484 char *lru_name
= NULL
;
485 time_t lru_atime
= 0;
487 dir
= opendir(dir_path
);
492 entry
= readdir(dir
);
497 if (fstatat(dirfd(dir
), entry
->d_name
, &sb
, 0) == 0) {
498 if (!lru_atime
|| (sb
.st_atime
< lru_atime
)) {
499 size_t len
= strlen(entry
->d_name
);
501 if (!predicate(dir_path
, &sb
, entry
->d_name
, len
))
504 char *tmp
= realloc(lru_name
, len
+ 1);
507 memcpy(lru_name
, entry
->d_name
, len
+ 1);
508 lru_atime
= sb
.st_atime
;
514 if (lru_name
== NULL
) {
519 if (asprintf(&filename
, "%s/%s", dir_path
, lru_name
) < 0)
528 /* Is entry a regular file, and not having a name with a trailing
532 is_regular_non_tmp_file(const char *path
, const struct stat
*sb
,
533 const char *d_name
, const size_t len
)
535 if (!S_ISREG(sb
->st_mode
))
538 if (len
>= 4 && strcmp(&d_name
[len
-4], ".tmp") == 0)
544 /* Returns the size of the deleted file, (or 0 on any error). */
546 unlink_lru_file_from_directory(const char *path
)
551 filename
= choose_lru_file_matching(path
, is_regular_non_tmp_file
);
552 if (filename
== NULL
)
555 if (stat(filename
, &sb
) == -1) {
563 return sb
.st_blocks
* 512;
566 /* Is entry a directory with a two-character name, (and not the
567 * special name of ".."). We also return false if the dir is empty.
570 is_two_character_sub_directory(const char *path
, const struct stat
*sb
,
571 const char *d_name
, const size_t len
)
573 if (!S_ISDIR(sb
->st_mode
))
579 if (strcmp(d_name
, "..") == 0)
583 if (asprintf(&subdir
, "%s/%s", path
, d_name
) == -1)
585 DIR *dir
= opendir(subdir
);
591 unsigned subdir_entries
= 0;
593 while ((d
= readdir(dir
)) != NULL
) {
594 if(++subdir_entries
> 2)
599 /* If dir only contains '.' and '..' it must be empty */
600 if (subdir_entries
<= 2)
607 evict_lru_item(struct disk_cache
*cache
)
611 /* With a reasonably-sized, full cache, (and with keys generated
612 * from a cryptographic hash), we can choose two random hex digits
613 * and reasonably expect the directory to exist with a file in it.
614 * Provides pseudo-LRU eviction to reduce checking all cache files.
616 uint64_t rand64
= rand_xorshift128plus(cache
->seed_xorshift128plus
);
617 if (asprintf(&dir_path
, "%s/%02" PRIx64
, cache
->path
, rand64
& 0xff) < 0)
620 size_t size
= unlink_lru_file_from_directory(dir_path
);
625 p_atomic_add(cache
->size
, - (uint64_t)size
);
629 /* In the case where the random choice of directory didn't find
630 * something, we choose the least recently accessed from the
631 * existing directories.
633 * Really, the only reason this code exists is to allow the unit
634 * tests to work, (which use an artificially-small cache to be able
635 * to force a single cached item to be evicted).
637 dir_path
= choose_lru_file_matching(cache
->path
,
638 is_two_character_sub_directory
);
639 if (dir_path
== NULL
)
642 size
= unlink_lru_file_from_directory(dir_path
);
647 p_atomic_add(cache
->size
, - (uint64_t)size
);
651 disk_cache_remove(struct disk_cache
*cache
, const cache_key key
)
655 char *filename
= get_cache_file(cache
, key
);
656 if (filename
== NULL
) {
660 if (stat(filename
, &sb
) == -1) {
669 p_atomic_add(cache
->size
, - (uint64_t)sb
.st_blocks
* 512);
673 read_all(int fd
, void *buf
, size_t count
)
679 for (done
= 0; done
< count
; done
+= read_ret
) {
680 read_ret
= read(fd
, in
+ done
, count
- done
);
681 if (read_ret
== -1 || read_ret
== 0)
688 write_all(int fd
, const void *buf
, size_t count
)
690 const char *out
= buf
;
694 for (done
= 0; done
< count
; done
+= written
) {
695 written
= write(fd
, out
+ done
, count
- done
);
702 /* From the zlib docs:
703 * "If the memory is available, buffers sizes on the order of 128K or 256K
704 * bytes should be used."
706 #define BUFSIZE 256 * 1024
709 * Compresses cache entry in memory and writes it to disk. Returns the size
710 * of the data written to disk.
713 deflate_and_write_to_disk(const void *in_data
, size_t in_data_size
, int dest
,
714 const char *filename
)
716 unsigned char out
[BUFSIZE
];
718 /* allocate deflate state */
720 strm
.zalloc
= Z_NULL
;
722 strm
.opaque
= Z_NULL
;
723 strm
.next_in
= (uint8_t *) in_data
;
724 strm
.avail_in
= in_data_size
;
726 int ret
= deflateInit(&strm
, Z_BEST_COMPRESSION
);
730 /* compress until end of in_data */
731 size_t compressed_size
= 0;
734 int remaining
= in_data_size
- BUFSIZE
;
735 flush
= remaining
> 0 ? Z_NO_FLUSH
: Z_FINISH
;
736 in_data_size
-= BUFSIZE
;
738 /* Run deflate() on input until the output buffer is not full (which
739 * means there is no more data to deflate).
742 strm
.avail_out
= BUFSIZE
;
745 ret
= deflate(&strm
, flush
); /* no bad return value */
746 assert(ret
!= Z_STREAM_ERROR
); /* state not clobbered */
748 size_t have
= BUFSIZE
- strm
.avail_out
;
749 compressed_size
+= have
;
751 ssize_t written
= write_all(dest
, out
, have
);
753 (void)deflateEnd(&strm
);
756 } while (strm
.avail_out
== 0);
758 /* all input should be used */
759 assert(strm
.avail_in
== 0);
761 } while (flush
!= Z_FINISH
);
763 /* stream should be complete */
764 assert(ret
== Z_STREAM_END
);
766 /* clean up and return */
767 (void)deflateEnd(&strm
);
768 return compressed_size
;
771 static struct disk_cache_put_job
*
772 create_put_job(struct disk_cache
*cache
, const cache_key key
,
773 const void *data
, size_t size
,
774 struct cache_item_metadata
*cache_item_metadata
)
776 struct disk_cache_put_job
*dc_job
= (struct disk_cache_put_job
*)
777 malloc(sizeof(struct disk_cache_put_job
) + size
);
780 dc_job
->cache
= cache
;
781 memcpy(dc_job
->key
, key
, sizeof(cache_key
));
782 dc_job
->data
= dc_job
+ 1;
783 memcpy(dc_job
->data
, data
, size
);
786 /* Copy the cache item metadata */
787 if (cache_item_metadata
) {
788 dc_job
->cache_item_metadata
.type
= cache_item_metadata
->type
;
789 if (cache_item_metadata
->type
== CACHE_ITEM_TYPE_GLSL
) {
790 dc_job
->cache_item_metadata
.num_keys
=
791 cache_item_metadata
->num_keys
;
792 dc_job
->cache_item_metadata
.keys
= (cache_key
*)
793 malloc(cache_item_metadata
->num_keys
* sizeof(cache_key
));
795 if (!dc_job
->cache_item_metadata
.keys
)
798 memcpy(dc_job
->cache_item_metadata
.keys
,
799 cache_item_metadata
->keys
,
800 sizeof(cache_key
) * cache_item_metadata
->num_keys
);
803 dc_job
->cache_item_metadata
.type
= CACHE_ITEM_TYPE_UNKNOWN
;
804 dc_job
->cache_item_metadata
.keys
= NULL
;
811 free(dc_job
->cache_item_metadata
.keys
);
818 destroy_put_job(void *job
, int thread_index
)
821 struct disk_cache_put_job
*dc_job
= (struct disk_cache_put_job
*) job
;
822 free(dc_job
->cache_item_metadata
.keys
);
828 struct cache_entry_file_data
{
830 uint32_t uncompressed_size
;
834 cache_put(void *job
, int thread_index
)
838 int fd
= -1, fd_final
= -1, err
, ret
;
840 char *filename
= NULL
, *filename_tmp
= NULL
;
841 struct disk_cache_put_job
*dc_job
= (struct disk_cache_put_job
*) job
;
843 filename
= get_cache_file(dc_job
->cache
, dc_job
->key
);
844 if (filename
== NULL
)
847 /* If the cache is too large, evict something else first. */
848 while (*dc_job
->cache
->size
+ dc_job
->size
> dc_job
->cache
->max_size
&&
850 evict_lru_item(dc_job
->cache
);
854 /* Write to a temporary file to allow for an atomic rename to the
855 * final destination filename, (to prevent any readers from seeing
856 * a partially written file).
858 if (asprintf(&filename_tmp
, "%s.tmp", filename
) == -1)
861 fd
= open(filename_tmp
, O_WRONLY
| O_CLOEXEC
| O_CREAT
, 0644);
863 /* Make the two-character subdirectory within the cache as needed. */
868 make_cache_file_directory(dc_job
->cache
, dc_job
->key
);
870 fd
= open(filename_tmp
, O_WRONLY
| O_CLOEXEC
| O_CREAT
, 0644);
875 /* With the temporary file open, we take an exclusive flock on
876 * it. If the flock fails, then another process still has the file
877 * open with the flock held. So just let that file be responsible
878 * for writing the file.
880 err
= flock(fd
, LOCK_EX
| LOCK_NB
);
884 /* Now that we have the lock on the open temporary file, we can
885 * check to see if the destination file already exists. If so,
886 * another process won the race between when we saw that the file
887 * didn't exist and now. In this case, we don't do anything more,
888 * (to ensure the size accounting of the cache doesn't get off).
890 fd_final
= open(filename
, O_RDONLY
| O_CLOEXEC
);
891 if (fd_final
!= -1) {
892 unlink(filename_tmp
);
896 /* OK, we're now on the hook to write out a file that we know is
897 * not in the cache, and is also not being written out to the cache
898 * by some other process.
901 /* Write the driver_keys_blob, this can be used find information about the
902 * mesa version that produced the entry or deal with hash collisions,
903 * should that ever become a real problem.
905 ret
= write_all(fd
, dc_job
->cache
->driver_keys_blob
,
906 dc_job
->cache
->driver_keys_blob_size
);
908 unlink(filename_tmp
);
912 /* Write the cache item metadata. This data can be used to deal with
913 * hash collisions, as well as providing useful information to 3rd party
914 * tools reading the cache files.
916 ret
= write_all(fd
, &dc_job
->cache_item_metadata
.type
,
919 unlink(filename_tmp
);
923 if (dc_job
->cache_item_metadata
.type
== CACHE_ITEM_TYPE_GLSL
) {
924 ret
= write_all(fd
, &dc_job
->cache_item_metadata
.num_keys
,
927 unlink(filename_tmp
);
931 ret
= write_all(fd
, dc_job
->cache_item_metadata
.keys
[0],
932 dc_job
->cache_item_metadata
.num_keys
*
935 unlink(filename_tmp
);
940 /* Create CRC of the data. We will read this when restoring the cache and
941 * use it to check for corruption.
943 struct cache_entry_file_data cf_data
;
944 cf_data
.crc32
= util_hash_crc32(dc_job
->data
, dc_job
->size
);
945 cf_data
.uncompressed_size
= dc_job
->size
;
947 size_t cf_data_size
= sizeof(cf_data
);
948 ret
= write_all(fd
, &cf_data
, cf_data_size
);
950 unlink(filename_tmp
);
954 /* Now, finally, write out the contents to the temporary file, then
955 * rename them atomically to the destination filename, and also
956 * perform an atomic increment of the total cache size.
958 size_t file_size
= deflate_and_write_to_disk(dc_job
->data
, dc_job
->size
,
960 if (file_size
== 0) {
961 unlink(filename_tmp
);
964 ret
= rename(filename_tmp
, filename
);
966 unlink(filename_tmp
);
971 if (stat(filename
, &sb
) == -1) {
972 /* Something went wrong remove the file */
977 p_atomic_add(dc_job
->cache
->size
, sb
.st_blocks
* 512);
982 /* This close finally releases the flock, (now that the final file
983 * has been renamed into place and the size has been added).
994 disk_cache_put(struct disk_cache
*cache
, const cache_key key
,
995 const void *data
, size_t size
,
996 struct cache_item_metadata
*cache_item_metadata
)
998 struct disk_cache_put_job
*dc_job
=
999 create_put_job(cache
, key
, data
, size
, cache_item_metadata
);
1002 util_queue_fence_init(&dc_job
->fence
);
1003 util_queue_add_job(&cache
->cache_queue
, dc_job
, &dc_job
->fence
,
1004 cache_put
, destroy_put_job
);
1009 * Decompresses cache entry, returns true if successful.
1012 inflate_cache_data(uint8_t *in_data
, size_t in_data_size
,
1013 uint8_t *out_data
, size_t out_data_size
)
1017 /* allocate inflate state */
1018 strm
.zalloc
= Z_NULL
;
1019 strm
.zfree
= Z_NULL
;
1020 strm
.opaque
= Z_NULL
;
1021 strm
.next_in
= in_data
;
1022 strm
.avail_in
= in_data_size
;
1023 strm
.next_out
= out_data
;
1024 strm
.avail_out
= out_data_size
;
1026 int ret
= inflateInit(&strm
);
1030 ret
= inflate(&strm
, Z_NO_FLUSH
);
1031 assert(ret
!= Z_STREAM_ERROR
); /* state not clobbered */
1033 /* Unless there was an error we should have decompressed everything in one
1034 * go as we know the uncompressed file size.
1036 if (ret
!= Z_STREAM_END
) {
1037 (void)inflateEnd(&strm
);
1040 assert(strm
.avail_out
== 0);
1042 /* clean up and return */
1043 (void)inflateEnd(&strm
);
1048 disk_cache_get(struct disk_cache
*cache
, const cache_key key
, size_t *size
)
1052 char *filename
= NULL
;
1053 uint8_t *data
= NULL
;
1054 uint8_t *uncompressed_data
= NULL
;
1055 uint8_t *file_header
= NULL
;
1060 filename
= get_cache_file(cache
, key
);
1061 if (filename
== NULL
)
1064 fd
= open(filename
, O_RDONLY
| O_CLOEXEC
);
1068 if (fstat(fd
, &sb
) == -1)
1071 data
= malloc(sb
.st_size
);
1075 size_t ck_size
= cache
->driver_keys_blob_size
;
1076 file_header
= malloc(ck_size
);
1080 if (sb
.st_size
< ck_size
)
1083 ret
= read_all(fd
, file_header
, ck_size
);
1087 /* Check for extremely unlikely hash collisions */
1088 if (memcmp(cache
->driver_keys_blob
, file_header
, ck_size
) != 0) {
1089 assert(!"Mesa cache keys mismatch!");
1093 size_t cache_item_md_size
= sizeof(uint32_t);
1095 ret
= read_all(fd
, &md_type
, cache_item_md_size
);
1099 if (md_type
== CACHE_ITEM_TYPE_GLSL
) {
1101 cache_item_md_size
+= sizeof(uint32_t);
1102 ret
= read_all(fd
, &num_keys
, sizeof(uint32_t));
1106 /* The cache item metadata is currently just used for distributing
1107 * precompiled shaders, they are not used by Mesa so just skip them for
1109 * TODO: pass the metadata back to the caller and do some basic
1112 cache_item_md_size
+= sizeof(cache_key
);
1113 ret
= lseek(fd
, num_keys
* sizeof(cache_key
), SEEK_CUR
);
1118 /* Load the CRC that was created when the file was written. */
1119 struct cache_entry_file_data cf_data
;
1120 size_t cf_data_size
= sizeof(cf_data
);
1121 ret
= read_all(fd
, &cf_data
, cf_data_size
);
1125 /* Load the actual cache data. */
1126 size_t cache_data_size
=
1127 sb
.st_size
- cf_data_size
- ck_size
- cache_item_md_size
;
1128 ret
= read_all(fd
, data
, cache_data_size
);
1132 /* Uncompress the cache data */
1133 uncompressed_data
= malloc(cf_data
.uncompressed_size
);
1134 if (!inflate_cache_data(data
, cache_data_size
, uncompressed_data
,
1135 cf_data
.uncompressed_size
))
1138 /* Check the data for corruption */
1139 if (cf_data
.crc32
!= util_hash_crc32(uncompressed_data
,
1140 cf_data
.uncompressed_size
))
1148 *size
= cf_data
.uncompressed_size
;
1150 return uncompressed_data
;
1155 if (uncompressed_data
)
1156 free(uncompressed_data
);
1168 disk_cache_put_key(struct disk_cache
*cache
, const cache_key key
)
1170 const uint32_t *key_chunk
= (const uint32_t *) key
;
1171 int i
= *key_chunk
& CACHE_INDEX_KEY_MASK
;
1172 unsigned char *entry
;
1174 entry
= &cache
->stored_keys
[i
* CACHE_KEY_SIZE
];
1176 memcpy(entry
, key
, CACHE_KEY_SIZE
);
1179 /* This function lets us test whether a given key was previously
1180 * stored in the cache with disk_cache_put_key(). The implement is
1181 * efficient by not using syscalls or hitting the disk. It's not
1182 * race-free, but the races are benign. If we race with someone else
1183 * calling disk_cache_put_key, then that's just an extra cache miss and an
1187 disk_cache_has_key(struct disk_cache
*cache
, const cache_key key
)
1189 const uint32_t *key_chunk
= (const uint32_t *) key
;
1190 int i
= *key_chunk
& CACHE_INDEX_KEY_MASK
;
1191 unsigned char *entry
;
1193 entry
= &cache
->stored_keys
[i
* CACHE_KEY_SIZE
];
1195 return memcmp(entry
, key
, CACHE_KEY_SIZE
) == 0;
1199 disk_cache_compute_key(struct disk_cache
*cache
, const void *data
, size_t size
,
1202 struct mesa_sha1 ctx
;
1204 _mesa_sha1_init(&ctx
);
1205 _mesa_sha1_update(&ctx
, cache
->driver_keys_blob
,
1206 cache
->driver_keys_blob_size
);
1207 _mesa_sha1_update(&ctx
, data
, size
);
1208 _mesa_sha1_final(&ctx
, key
);
1211 #endif /* ENABLE_SHADER_CACHE */