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>
47 #include "util/crc32.h"
48 #include "util/debug.h"
49 #include "util/rand_xor.h"
50 #include "util/u_atomic.h"
51 #include "util/u_queue.h"
52 #include "util/mesa-sha1.h"
53 #include "util/ralloc.h"
54 #include "main/compiler.h"
55 #include "main/errors.h"
57 #include "disk_cache.h"
59 /* Number of bits to mask off from a cache key to get an index. */
60 #define CACHE_INDEX_KEY_BITS 16
62 /* Mask for computing an index from a key. */
63 #define CACHE_INDEX_KEY_MASK ((1 << CACHE_INDEX_KEY_BITS) - 1)
65 /* The number of keys that can be stored in the index. */
66 #define CACHE_INDEX_MAX_KEYS (1 << CACHE_INDEX_KEY_BITS)
68 /* The cache version should be bumped whenever a change is made to the
69 * structure of cache entries or the index. This will give any 3rd party
70 * applications reading the cache entries a chance to adjust to the changes.
72 * - The cache version is checked internally when reading a cache entry. If we
73 * ever have a mismatch we are in big trouble as this means we had a cache
74 * collision. In case of such an event please check the skys for giant
75 * asteroids and that the entire Mesa team hasn't been eaten by wolves.
77 * - There is no strict requirement that cache versions be backwards
78 * compatible but effort should be taken to limit disruption where possible.
80 #define CACHE_VERSION 1
82 /* 3 is the recomended level, with 22 as the absolute maximum */
83 #define ZSTD_COMPRESSION_LEVEL 3
86 /* The path to the cache directory. */
88 bool path_init_failed
;
90 /* Thread queue for compressing and writing cache entries to disk */
91 struct util_queue cache_queue
;
93 /* Seed for rand, which is used to pick a random directory */
94 uint64_t seed_xorshift128plus
[2];
96 /* A pointer to the mmapped index file within the cache directory. */
98 size_t index_mmap_size
;
100 /* Pointer to total size of all objects in cache (within index_mmap) */
103 /* Pointer to stored keys, (within index_mmap). */
104 uint8_t *stored_keys
;
106 /* Maximum size of all cached objects (in bytes). */
109 /* Driver cache keys. */
110 uint8_t *driver_keys_blob
;
111 size_t driver_keys_blob_size
;
113 disk_cache_put_cb blob_put_cb
;
114 disk_cache_get_cb blob_get_cb
;
117 struct disk_cache_put_job
{
118 struct util_queue_fence fence
;
120 struct disk_cache
*cache
;
124 /* Copy of cache data to be compressed and written. */
127 /* Size of data to be compressed and written. */
130 struct cache_item_metadata cache_item_metadata
;
133 /* Create a directory named 'path' if it does not already exist.
135 * Returns: 0 if path already exists as a directory or if created.
136 * -1 in all other cases.
139 mkdir_if_needed(const char *path
)
143 /* If the path exists already, then our work is done if it's a
144 * directory, but it's an error if it is not.
146 if (stat(path
, &sb
) == 0) {
147 if (S_ISDIR(sb
.st_mode
)) {
150 fprintf(stderr
, "Cannot use %s for shader cache (not a directory)"
151 "---disabling.\n", path
);
156 int ret
= mkdir(path
, 0755);
157 if (ret
== 0 || (ret
== -1 && errno
== EEXIST
))
160 fprintf(stderr
, "Failed to create %s for shader cache (%s)---disabling.\n",
161 path
, strerror(errno
));
166 /* Concatenate an existing path and a new name to form a new path. If the new
167 * path does not exist as a directory, create it then return the resulting
168 * name of the new path (ralloc'ed off of 'ctx').
170 * Returns NULL on any error, such as:
172 * <path> does not exist or is not a directory
173 * <path>/<name> exists but is not a directory
174 * <path>/<name> cannot be created as a directory
177 concatenate_and_mkdir(void *ctx
, const char *path
, const char *name
)
182 if (stat(path
, &sb
) != 0 || ! S_ISDIR(sb
.st_mode
))
185 new_path
= ralloc_asprintf(ctx
, "%s/%s", path
, name
);
187 if (mkdir_if_needed(new_path
) == 0)
193 #define DRV_KEY_CPY(_dst, _src, _src_size) \
195 memcpy(_dst, _src, _src_size); \
200 disk_cache_create(const char *gpu_name
, const char *driver_id
,
201 uint64_t driver_flags
)
204 struct disk_cache
*cache
= NULL
;
205 char *path
, *max_size_str
;
211 uint8_t cache_version
= CACHE_VERSION
;
212 size_t cv_size
= sizeof(cache_version
);
214 /* If running as a users other than the real user disable cache */
215 if (geteuid() != getuid())
218 /* A ralloc context for transient data during this invocation. */
219 local
= ralloc_context(NULL
);
223 /* At user request, disable shader cache entirely. */
224 if (env_var_as_boolean("MESA_GLSL_CACHE_DISABLE", false))
227 cache
= rzalloc(NULL
, struct disk_cache
);
231 /* Assume failure. */
232 cache
->path_init_failed
= true;
234 /* Determine path for cache based on the first defined name as follows:
236 * $MESA_GLSL_CACHE_DIR
237 * $XDG_CACHE_HOME/mesa_shader_cache
238 * <pwd.pw_dir>/.cache/mesa_shader_cache
240 path
= getenv("MESA_GLSL_CACHE_DIR");
242 if (mkdir_if_needed(path
) == -1)
245 path
= concatenate_and_mkdir(local
, path
, CACHE_DIR_NAME
);
251 char *xdg_cache_home
= getenv("XDG_CACHE_HOME");
253 if (xdg_cache_home
) {
254 if (mkdir_if_needed(xdg_cache_home
) == -1)
257 path
= concatenate_and_mkdir(local
, xdg_cache_home
, CACHE_DIR_NAME
);
266 struct passwd pwd
, *result
;
268 buf_size
= sysconf(_SC_GETPW_R_SIZE_MAX
);
272 /* Loop until buf_size is large enough to query the directory */
274 buf
= ralloc_size(local
, buf_size
);
276 getpwuid_r(getuid(), &pwd
, buf
, buf_size
, &result
);
280 if (errno
== ERANGE
) {
289 path
= concatenate_and_mkdir(local
, pwd
.pw_dir
, ".cache");
293 path
= concatenate_and_mkdir(local
, path
, CACHE_DIR_NAME
);
298 cache
->path
= ralloc_strdup(cache
, path
);
299 if (cache
->path
== NULL
)
302 path
= ralloc_asprintf(local
, "%s/index", cache
->path
);
306 fd
= open(path
, O_RDWR
| O_CREAT
| O_CLOEXEC
, 0644);
310 if (fstat(fd
, &sb
) == -1)
313 /* Force the index file to be the expected size. */
314 size
= sizeof(*cache
->size
) + CACHE_INDEX_MAX_KEYS
* CACHE_KEY_SIZE
;
315 if (sb
.st_size
!= size
) {
316 if (ftruncate(fd
, size
) == -1)
320 /* We map this shared so that other processes see updates that we
323 * Note: We do use atomic addition to ensure that multiple
324 * processes don't scramble the cache size recorded in the
325 * index. But we don't use any locking to prevent multiple
326 * processes from updating the same entry simultaneously. The idea
327 * is that if either result lands entirely in the index, then
328 * that's equivalent to a well-ordered write followed by an
329 * eviction and a write. On the other hand, if the simultaneous
330 * writes result in a corrupt entry, that's not really any
331 * different than both entries being evicted, (since within the
332 * guarantees of the cryptographic hash, a corrupt entry is
333 * unlikely to ever match a real cache key).
335 cache
->index_mmap
= mmap(NULL
, size
, PROT_READ
| PROT_WRITE
,
337 if (cache
->index_mmap
== MAP_FAILED
)
339 cache
->index_mmap_size
= size
;
341 cache
->size
= (uint64_t *) cache
->index_mmap
;
342 cache
->stored_keys
= cache
->index_mmap
+ sizeof(uint64_t);
346 max_size_str
= getenv("MESA_GLSL_CACHE_MAX_SIZE");
349 max_size
= strtoul(max_size_str
, &end
, 10);
350 if (end
== max_size_str
) {
360 max_size
*= 1024*1024;
366 max_size
*= 1024*1024*1024;
372 /* Default to 1GB for maximum cache size. */
374 max_size
= 1024*1024*1024;
377 cache
->max_size
= max_size
;
379 /* 4 threads were chosen below because just about all modern CPUs currently
380 * available that run Mesa have *at least* 4 cores. For these CPUs allowing
381 * more threads can result in the queue being processed faster, thus
382 * avoiding excessive memory use due to a backlog of cache entrys building
383 * up in the queue. Since we set the UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY
384 * flag this should have little negative impact on low core systems.
386 * The queue will resize automatically when it's full, so adding new jobs
389 util_queue_init(&cache
->cache_queue
, "disk$", 32, 4,
390 UTIL_QUEUE_INIT_RESIZE_IF_FULL
|
391 UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY
|
392 UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY
);
394 cache
->path_init_failed
= false;
401 cache
->driver_keys_blob_size
= cv_size
;
403 /* Create driver id keys */
404 size_t id_size
= strlen(driver_id
) + 1;
405 size_t gpu_name_size
= strlen(gpu_name
) + 1;
406 cache
->driver_keys_blob_size
+= id_size
;
407 cache
->driver_keys_blob_size
+= gpu_name_size
;
409 /* We sometimes store entire structs that contains a pointers in the cache,
410 * use pointer size as a key to avoid hard to debug issues.
412 uint8_t ptr_size
= sizeof(void *);
413 size_t ptr_size_size
= sizeof(ptr_size
);
414 cache
->driver_keys_blob_size
+= ptr_size_size
;
416 size_t driver_flags_size
= sizeof(driver_flags
);
417 cache
->driver_keys_blob_size
+= driver_flags_size
;
419 cache
->driver_keys_blob
=
420 ralloc_size(cache
, cache
->driver_keys_blob_size
);
421 if (!cache
->driver_keys_blob
)
424 uint8_t *drv_key_blob
= cache
->driver_keys_blob
;
425 DRV_KEY_CPY(drv_key_blob
, &cache_version
, cv_size
)
426 DRV_KEY_CPY(drv_key_blob
, driver_id
, id_size
)
427 DRV_KEY_CPY(drv_key_blob
, gpu_name
, gpu_name_size
)
428 DRV_KEY_CPY(drv_key_blob
, &ptr_size
, ptr_size_size
)
429 DRV_KEY_CPY(drv_key_blob
, &driver_flags
, driver_flags_size
)
431 /* Seed our rand function */
432 s_rand_xorshift128plus(cache
->seed_xorshift128plus
, true);
447 disk_cache_destroy(struct disk_cache
*cache
)
449 if (cache
&& !cache
->path_init_failed
) {
450 util_queue_finish(&cache
->cache_queue
);
451 util_queue_destroy(&cache
->cache_queue
);
452 munmap(cache
->index_mmap
, cache
->index_mmap_size
);
458 /* Return a filename within the cache's directory corresponding to 'key'. The
459 * returned filename is ralloced with 'cache' as the parent context.
461 * Returns NULL if out of memory.
464 get_cache_file(struct disk_cache
*cache
, const cache_key key
)
469 if (cache
->path_init_failed
)
472 _mesa_sha1_format(buf
, key
);
473 if (asprintf(&filename
, "%s/%c%c/%s", cache
->path
, buf
[0],
474 buf
[1], buf
+ 2) == -1)
480 /* Create the directory that will be needed for the cache file for \key.
482 * Obviously, the implementation here must closely match
483 * _get_cache_file above.
486 make_cache_file_directory(struct disk_cache
*cache
, const cache_key key
)
491 _mesa_sha1_format(buf
, key
);
492 if (asprintf(&dir
, "%s/%c%c", cache
->path
, buf
[0], buf
[1]) == -1)
495 mkdir_if_needed(dir
);
499 /* Given a directory path and predicate function, find the entry with
500 * the oldest access time in that directory for which the predicate
503 * Returns: A malloc'ed string for the path to the chosen file, (or
504 * NULL on any error). The caller should free the string when
508 choose_lru_file_matching(const char *dir_path
,
509 bool (*predicate
)(const char *dir_path
,
511 const char *, const size_t))
514 struct dirent
*entry
;
516 char *lru_name
= NULL
;
517 time_t lru_atime
= 0;
519 dir
= opendir(dir_path
);
524 entry
= readdir(dir
);
529 if (fstatat(dirfd(dir
), entry
->d_name
, &sb
, 0) == 0) {
530 if (!lru_atime
|| (sb
.st_atime
< lru_atime
)) {
531 size_t len
= strlen(entry
->d_name
);
533 if (!predicate(dir_path
, &sb
, entry
->d_name
, len
))
536 char *tmp
= realloc(lru_name
, len
+ 1);
539 memcpy(lru_name
, entry
->d_name
, len
+ 1);
540 lru_atime
= sb
.st_atime
;
546 if (lru_name
== NULL
) {
551 if (asprintf(&filename
, "%s/%s", dir_path
, lru_name
) < 0)
560 /* Is entry a regular file, and not having a name with a trailing
564 is_regular_non_tmp_file(const char *path
, const struct stat
*sb
,
565 const char *d_name
, const size_t len
)
567 if (!S_ISREG(sb
->st_mode
))
570 if (len
>= 4 && strcmp(&d_name
[len
-4], ".tmp") == 0)
576 /* Returns the size of the deleted file, (or 0 on any error). */
578 unlink_lru_file_from_directory(const char *path
)
583 filename
= choose_lru_file_matching(path
, is_regular_non_tmp_file
);
584 if (filename
== NULL
)
587 if (stat(filename
, &sb
) == -1) {
595 return sb
.st_blocks
* 512;
598 /* Is entry a directory with a two-character name, (and not the
599 * special name of ".."). We also return false if the dir is empty.
602 is_two_character_sub_directory(const char *path
, const struct stat
*sb
,
603 const char *d_name
, const size_t len
)
605 if (!S_ISDIR(sb
->st_mode
))
611 if (strcmp(d_name
, "..") == 0)
615 if (asprintf(&subdir
, "%s/%s", path
, d_name
) == -1)
617 DIR *dir
= opendir(subdir
);
623 unsigned subdir_entries
= 0;
625 while ((d
= readdir(dir
)) != NULL
) {
626 if(++subdir_entries
> 2)
631 /* If dir only contains '.' and '..' it must be empty */
632 if (subdir_entries
<= 2)
639 evict_lru_item(struct disk_cache
*cache
)
643 /* With a reasonably-sized, full cache, (and with keys generated
644 * from a cryptographic hash), we can choose two random hex digits
645 * and reasonably expect the directory to exist with a file in it.
646 * Provides pseudo-LRU eviction to reduce checking all cache files.
648 uint64_t rand64
= rand_xorshift128plus(cache
->seed_xorshift128plus
);
649 if (asprintf(&dir_path
, "%s/%02" PRIx64
, cache
->path
, rand64
& 0xff) < 0)
652 size_t size
= unlink_lru_file_from_directory(dir_path
);
657 p_atomic_add(cache
->size
, - (uint64_t)size
);
661 /* In the case where the random choice of directory didn't find
662 * something, we choose the least recently accessed from the
663 * existing directories.
665 * Really, the only reason this code exists is to allow the unit
666 * tests to work, (which use an artificially-small cache to be able
667 * to force a single cached item to be evicted).
669 dir_path
= choose_lru_file_matching(cache
->path
,
670 is_two_character_sub_directory
);
671 if (dir_path
== NULL
)
674 size
= unlink_lru_file_from_directory(dir_path
);
679 p_atomic_add(cache
->size
, - (uint64_t)size
);
683 disk_cache_remove(struct disk_cache
*cache
, const cache_key key
)
687 char *filename
= get_cache_file(cache
, key
);
688 if (filename
== NULL
) {
692 if (stat(filename
, &sb
) == -1) {
701 p_atomic_add(cache
->size
, - (uint64_t)sb
.st_blocks
* 512);
705 read_all(int fd
, void *buf
, size_t count
)
711 for (done
= 0; done
< count
; done
+= read_ret
) {
712 read_ret
= read(fd
, in
+ done
, count
- done
);
713 if (read_ret
== -1 || read_ret
== 0)
720 write_all(int fd
, const void *buf
, size_t count
)
722 const char *out
= buf
;
726 for (done
= 0; done
< count
; done
+= written
) {
727 written
= write(fd
, out
+ done
, count
- done
);
734 /* From the zlib docs:
735 * "If the memory is available, buffers sizes on the order of 128K or 256K
736 * bytes should be used."
738 #define BUFSIZE 256 * 1024
741 * Compresses cache entry in memory and writes it to disk. Returns the size
742 * of the data written to disk.
745 deflate_and_write_to_disk(const void *in_data
, size_t in_data_size
, int dest
,
746 const char *filename
)
749 /* from the zstd docs (https://facebook.github.io/zstd/zstd_manual.html):
750 * compression runs faster if `dstCapacity` >= `ZSTD_compressBound(srcSize)`.
752 size_t out_size
= ZSTD_compressBound(in_data_size
);
753 void * out
= malloc(out_size
);
755 size_t ret
= ZSTD_compress(out
, out_size
, in_data
, in_data_size
,
756 ZSTD_COMPRESSION_LEVEL
);
757 if (ZSTD_isError(ret
)) {
761 write_all(dest
, out
, ret
);
767 /* allocate deflate state */
769 strm
.zalloc
= Z_NULL
;
771 strm
.opaque
= Z_NULL
;
772 strm
.next_in
= (uint8_t *) in_data
;
773 strm
.avail_in
= in_data_size
;
775 int ret
= deflateInit(&strm
, Z_BEST_COMPRESSION
);
779 /* compress until end of in_data */
780 size_t compressed_size
= 0;
783 out
= malloc(BUFSIZE
* sizeof(unsigned char));
788 int remaining
= in_data_size
- BUFSIZE
;
789 flush
= remaining
> 0 ? Z_NO_FLUSH
: Z_FINISH
;
790 in_data_size
-= BUFSIZE
;
792 /* Run deflate() on input until the output buffer is not full (which
793 * means there is no more data to deflate).
796 strm
.avail_out
= BUFSIZE
;
799 ret
= deflate(&strm
, flush
); /* no bad return value */
800 assert(ret
!= Z_STREAM_ERROR
); /* state not clobbered */
802 size_t have
= BUFSIZE
- strm
.avail_out
;
803 compressed_size
+= have
;
805 ssize_t written
= write_all(dest
, out
, have
);
807 (void)deflateEnd(&strm
);
811 } while (strm
.avail_out
== 0);
813 /* all input should be used */
814 assert(strm
.avail_in
== 0);
816 } while (flush
!= Z_FINISH
);
818 /* stream should be complete */
819 assert(ret
== Z_STREAM_END
);
821 /* clean up and return */
822 (void)deflateEnd(&strm
);
824 return compressed_size
;
828 static struct disk_cache_put_job
*
829 create_put_job(struct disk_cache
*cache
, const cache_key key
,
830 const void *data
, size_t size
,
831 struct cache_item_metadata
*cache_item_metadata
)
833 struct disk_cache_put_job
*dc_job
= (struct disk_cache_put_job
*)
834 malloc(sizeof(struct disk_cache_put_job
) + size
);
837 dc_job
->cache
= cache
;
838 memcpy(dc_job
->key
, key
, sizeof(cache_key
));
839 dc_job
->data
= dc_job
+ 1;
840 memcpy(dc_job
->data
, data
, size
);
843 /* Copy the cache item metadata */
844 if (cache_item_metadata
) {
845 dc_job
->cache_item_metadata
.type
= cache_item_metadata
->type
;
846 if (cache_item_metadata
->type
== CACHE_ITEM_TYPE_GLSL
) {
847 dc_job
->cache_item_metadata
.num_keys
=
848 cache_item_metadata
->num_keys
;
849 dc_job
->cache_item_metadata
.keys
= (cache_key
*)
850 malloc(cache_item_metadata
->num_keys
* sizeof(cache_key
));
852 if (!dc_job
->cache_item_metadata
.keys
)
855 memcpy(dc_job
->cache_item_metadata
.keys
,
856 cache_item_metadata
->keys
,
857 sizeof(cache_key
) * cache_item_metadata
->num_keys
);
860 dc_job
->cache_item_metadata
.type
= CACHE_ITEM_TYPE_UNKNOWN
;
861 dc_job
->cache_item_metadata
.keys
= NULL
;
874 destroy_put_job(void *job
, int thread_index
)
877 struct disk_cache_put_job
*dc_job
= (struct disk_cache_put_job
*) job
;
878 free(dc_job
->cache_item_metadata
.keys
);
884 struct cache_entry_file_data
{
886 uint32_t uncompressed_size
;
890 cache_put(void *job
, int thread_index
)
894 int fd
= -1, fd_final
= -1, err
, ret
;
896 char *filename
= NULL
, *filename_tmp
= NULL
;
897 struct disk_cache_put_job
*dc_job
= (struct disk_cache_put_job
*) job
;
899 filename
= get_cache_file(dc_job
->cache
, dc_job
->key
);
900 if (filename
== NULL
)
903 /* If the cache is too large, evict something else first. */
904 while (*dc_job
->cache
->size
+ dc_job
->size
> dc_job
->cache
->max_size
&&
906 evict_lru_item(dc_job
->cache
);
910 /* Write to a temporary file to allow for an atomic rename to the
911 * final destination filename, (to prevent any readers from seeing
912 * a partially written file).
914 if (asprintf(&filename_tmp
, "%s.tmp", filename
) == -1)
917 fd
= open(filename_tmp
, O_WRONLY
| O_CLOEXEC
| O_CREAT
, 0644);
919 /* Make the two-character subdirectory within the cache as needed. */
924 make_cache_file_directory(dc_job
->cache
, dc_job
->key
);
926 fd
= open(filename_tmp
, O_WRONLY
| O_CLOEXEC
| O_CREAT
, 0644);
931 /* With the temporary file open, we take an exclusive flock on
932 * it. If the flock fails, then another process still has the file
933 * open with the flock held. So just let that file be responsible
934 * for writing the file.
937 err
= flock(fd
, LOCK_EX
| LOCK_NB
);
939 struct flock lock
= {
941 .l_len
= 0, /* entire file */
945 err
= fcntl(fd
, F_SETLK
, &lock
);
950 /* Now that we have the lock on the open temporary file, we can
951 * check to see if the destination file already exists. If so,
952 * another process won the race between when we saw that the file
953 * didn't exist and now. In this case, we don't do anything more,
954 * (to ensure the size accounting of the cache doesn't get off).
956 fd_final
= open(filename
, O_RDONLY
| O_CLOEXEC
);
957 if (fd_final
!= -1) {
958 unlink(filename_tmp
);
962 /* OK, we're now on the hook to write out a file that we know is
963 * not in the cache, and is also not being written out to the cache
964 * by some other process.
967 /* Write the driver_keys_blob, this can be used find information about the
968 * mesa version that produced the entry or deal with hash collisions,
969 * should that ever become a real problem.
971 ret
= write_all(fd
, dc_job
->cache
->driver_keys_blob
,
972 dc_job
->cache
->driver_keys_blob_size
);
974 unlink(filename_tmp
);
978 /* Write the cache item metadata. This data can be used to deal with
979 * hash collisions, as well as providing useful information to 3rd party
980 * tools reading the cache files.
982 ret
= write_all(fd
, &dc_job
->cache_item_metadata
.type
,
985 unlink(filename_tmp
);
989 if (dc_job
->cache_item_metadata
.type
== CACHE_ITEM_TYPE_GLSL
) {
990 ret
= write_all(fd
, &dc_job
->cache_item_metadata
.num_keys
,
993 unlink(filename_tmp
);
997 ret
= write_all(fd
, dc_job
->cache_item_metadata
.keys
[0],
998 dc_job
->cache_item_metadata
.num_keys
*
1001 unlink(filename_tmp
);
1006 /* Create CRC of the data. We will read this when restoring the cache and
1007 * use it to check for corruption.
1009 struct cache_entry_file_data cf_data
;
1010 cf_data
.crc32
= util_hash_crc32(dc_job
->data
, dc_job
->size
);
1011 cf_data
.uncompressed_size
= dc_job
->size
;
1013 size_t cf_data_size
= sizeof(cf_data
);
1014 ret
= write_all(fd
, &cf_data
, cf_data_size
);
1016 unlink(filename_tmp
);
1020 /* Now, finally, write out the contents to the temporary file, then
1021 * rename them atomically to the destination filename, and also
1022 * perform an atomic increment of the total cache size.
1024 size_t file_size
= deflate_and_write_to_disk(dc_job
->data
, dc_job
->size
,
1026 if (file_size
== 0) {
1027 unlink(filename_tmp
);
1030 ret
= rename(filename_tmp
, filename
);
1032 unlink(filename_tmp
);
1037 if (stat(filename
, &sb
) == -1) {
1038 /* Something went wrong remove the file */
1043 p_atomic_add(dc_job
->cache
->size
, sb
.st_blocks
* 512);
1048 /* This close finally releases the flock, (now that the final file
1049 * has been renamed into place and the size has been added).
1058 disk_cache_put(struct disk_cache
*cache
, const cache_key key
,
1059 const void *data
, size_t size
,
1060 struct cache_item_metadata
*cache_item_metadata
)
1062 if (cache
->blob_put_cb
) {
1063 cache
->blob_put_cb(key
, CACHE_KEY_SIZE
, data
, size
);
1067 if (cache
->path_init_failed
)
1070 struct disk_cache_put_job
*dc_job
=
1071 create_put_job(cache
, key
, data
, size
, cache_item_metadata
);
1074 util_queue_fence_init(&dc_job
->fence
);
1075 util_queue_add_job(&cache
->cache_queue
, dc_job
, &dc_job
->fence
,
1076 cache_put
, destroy_put_job
, dc_job
->size
);
1081 * Decompresses cache entry, returns true if successful.
1084 inflate_cache_data(uint8_t *in_data
, size_t in_data_size
,
1085 uint8_t *out_data
, size_t out_data_size
)
1088 size_t ret
= ZSTD_decompress(out_data
, out_data_size
, in_data
, in_data_size
);
1089 return !ZSTD_isError(ret
);
1093 /* allocate inflate state */
1094 strm
.zalloc
= Z_NULL
;
1095 strm
.zfree
= Z_NULL
;
1096 strm
.opaque
= Z_NULL
;
1097 strm
.next_in
= in_data
;
1098 strm
.avail_in
= in_data_size
;
1099 strm
.next_out
= out_data
;
1100 strm
.avail_out
= out_data_size
;
1102 int ret
= inflateInit(&strm
);
1106 ret
= inflate(&strm
, Z_NO_FLUSH
);
1107 assert(ret
!= Z_STREAM_ERROR
); /* state not clobbered */
1109 /* Unless there was an error we should have decompressed everything in one
1110 * go as we know the uncompressed file size.
1112 if (ret
!= Z_STREAM_END
) {
1113 (void)inflateEnd(&strm
);
1116 assert(strm
.avail_out
== 0);
1118 /* clean up and return */
1119 (void)inflateEnd(&strm
);
1125 disk_cache_get(struct disk_cache
*cache
, const cache_key key
, size_t *size
)
1129 char *filename
= NULL
;
1130 uint8_t *data
= NULL
;
1131 uint8_t *uncompressed_data
= NULL
;
1132 uint8_t *file_header
= NULL
;
1137 if (cache
->blob_get_cb
) {
1138 /* This is what Android EGL defines as the maxValueSize in egl_cache_t
1139 * class implementation.
1141 const signed long max_blob_size
= 64 * 1024;
1142 void *blob
= malloc(max_blob_size
);
1147 cache
->blob_get_cb(key
, CACHE_KEY_SIZE
, blob
, max_blob_size
);
1159 filename
= get_cache_file(cache
, key
);
1160 if (filename
== NULL
)
1163 fd
= open(filename
, O_RDONLY
| O_CLOEXEC
);
1167 if (fstat(fd
, &sb
) == -1)
1170 data
= malloc(sb
.st_size
);
1174 size_t ck_size
= cache
->driver_keys_blob_size
;
1175 file_header
= malloc(ck_size
);
1179 if (sb
.st_size
< ck_size
)
1182 ret
= read_all(fd
, file_header
, ck_size
);
1186 /* Check for extremely unlikely hash collisions */
1187 if (memcmp(cache
->driver_keys_blob
, file_header
, ck_size
) != 0) {
1188 assert(!"Mesa cache keys mismatch!");
1192 size_t cache_item_md_size
= sizeof(uint32_t);
1194 ret
= read_all(fd
, &md_type
, cache_item_md_size
);
1198 if (md_type
== CACHE_ITEM_TYPE_GLSL
) {
1200 cache_item_md_size
+= sizeof(uint32_t);
1201 ret
= read_all(fd
, &num_keys
, sizeof(uint32_t));
1205 /* The cache item metadata is currently just used for distributing
1206 * precompiled shaders, they are not used by Mesa so just skip them for
1208 * TODO: pass the metadata back to the caller and do some basic
1211 cache_item_md_size
+= num_keys
* sizeof(cache_key
);
1212 ret
= lseek(fd
, num_keys
* sizeof(cache_key
), SEEK_CUR
);
1217 /* Load the CRC that was created when the file was written. */
1218 struct cache_entry_file_data cf_data
;
1219 size_t cf_data_size
= sizeof(cf_data
);
1220 ret
= read_all(fd
, &cf_data
, cf_data_size
);
1224 /* Load the actual cache data. */
1225 size_t cache_data_size
=
1226 sb
.st_size
- cf_data_size
- ck_size
- cache_item_md_size
;
1227 ret
= read_all(fd
, data
, cache_data_size
);
1231 /* Uncompress the cache data */
1232 uncompressed_data
= malloc(cf_data
.uncompressed_size
);
1233 if (!inflate_cache_data(data
, cache_data_size
, uncompressed_data
,
1234 cf_data
.uncompressed_size
))
1237 /* Check the data for corruption */
1238 if (cf_data
.crc32
!= util_hash_crc32(uncompressed_data
,
1239 cf_data
.uncompressed_size
))
1248 *size
= cf_data
.uncompressed_size
;
1250 return uncompressed_data
;
1255 if (uncompressed_data
)
1256 free(uncompressed_data
);
1268 disk_cache_put_key(struct disk_cache
*cache
, const cache_key key
)
1270 const uint32_t *key_chunk
= (const uint32_t *) key
;
1271 int i
= CPU_TO_LE32(*key_chunk
) & CACHE_INDEX_KEY_MASK
;
1272 unsigned char *entry
;
1274 if (cache
->blob_put_cb
) {
1275 cache
->blob_put_cb(key
, CACHE_KEY_SIZE
, key_chunk
, sizeof(uint32_t));
1279 if (cache
->path_init_failed
)
1282 entry
= &cache
->stored_keys
[i
* CACHE_KEY_SIZE
];
1284 memcpy(entry
, key
, CACHE_KEY_SIZE
);
1287 /* This function lets us test whether a given key was previously
1288 * stored in the cache with disk_cache_put_key(). The implement is
1289 * efficient by not using syscalls or hitting the disk. It's not
1290 * race-free, but the races are benign. If we race with someone else
1291 * calling disk_cache_put_key, then that's just an extra cache miss and an
1295 disk_cache_has_key(struct disk_cache
*cache
, const cache_key key
)
1297 const uint32_t *key_chunk
= (const uint32_t *) key
;
1298 int i
= CPU_TO_LE32(*key_chunk
) & CACHE_INDEX_KEY_MASK
;
1299 unsigned char *entry
;
1301 if (cache
->blob_get_cb
) {
1303 return cache
->blob_get_cb(key
, CACHE_KEY_SIZE
, &blob
, sizeof(uint32_t));
1306 if (cache
->path_init_failed
)
1309 entry
= &cache
->stored_keys
[i
* CACHE_KEY_SIZE
];
1311 return memcmp(entry
, key
, CACHE_KEY_SIZE
) == 0;
1315 disk_cache_compute_key(struct disk_cache
*cache
, const void *data
, size_t size
,
1318 struct mesa_sha1 ctx
;
1320 _mesa_sha1_init(&ctx
);
1321 _mesa_sha1_update(&ctx
, cache
->driver_keys_blob
,
1322 cache
->driver_keys_blob_size
);
1323 _mesa_sha1_update(&ctx
, data
, size
);
1324 _mesa_sha1_final(&ctx
, key
);
1328 disk_cache_set_callbacks(struct disk_cache
*cache
, disk_cache_put_cb put
,
1329 disk_cache_get_cb get
)
1331 cache
->blob_put_cb
= put
;
1332 cache
->blob_get_cb
= get
;
1335 #endif /* ENABLE_SHADER_CACHE */