#include <pwd.h>
#include <errno.h>
#include <dirent.h>
+#include <inttypes.h>
#include "zlib.h"
+#ifdef HAVE_ZSTD
+#include "zstd.h"
+#endif
+
#include "util/crc32.h"
+#include "util/debug.h"
#include "util/rand_xor.h"
#include "util/u_atomic.h"
#include "util/u_queue.h"
#include "util/mesa-sha1.h"
#include "util/ralloc.h"
-#include "main/errors.h"
+#include "util/compiler.h"
#include "disk_cache.h"
*/
#define CACHE_VERSION 1
+/* 3 is the recomended level, with 22 as the absolute maximum */
+#define ZSTD_COMPRESSION_LEVEL 3
+
struct disk_cache {
/* The path to the cache directory. */
char *path;
+ bool path_init_failed;
/* Thread queue for compressing and writing cache entries to disk */
struct util_queue cache_queue;
/* Driver cache keys. */
uint8_t *driver_keys_blob;
size_t driver_keys_blob_size;
+
+ disk_cache_put_cb blob_put_cb;
+ disk_cache_get_cb blob_get_cb;
};
struct disk_cache_put_job {
} while (0);
struct disk_cache *
-disk_cache_create(const char *gpu_name, const char *timestamp,
+disk_cache_create(const char *gpu_name, const char *driver_id,
uint64_t driver_flags)
{
void *local;
struct stat sb;
size_t size;
+ uint8_t cache_version = CACHE_VERSION;
+ size_t cv_size = sizeof(cache_version);
+
/* If running as a users other than the real user disable cache */
if (geteuid() != getuid())
return NULL;
goto fail;
/* At user request, disable shader cache entirely. */
- if (getenv("MESA_GLSL_CACHE_DISABLE"))
+ if (env_var_as_boolean("MESA_GLSL_CACHE_DISABLE", false))
goto fail;
+ cache = rzalloc(NULL, struct disk_cache);
+ if (cache == NULL)
+ goto fail;
+
+ /* Assume failure. */
+ cache->path_init_failed = true;
+
/* Determine path for cache based on the first defined name as follows:
*
* $MESA_GLSL_CACHE_DIR
path = getenv("MESA_GLSL_CACHE_DIR");
if (path) {
if (mkdir_if_needed(path) == -1)
- goto fail;
+ goto path_fail;
path = concatenate_and_mkdir(local, path, CACHE_DIR_NAME);
if (path == NULL)
- goto fail;
+ goto path_fail;
}
if (path == NULL) {
if (xdg_cache_home) {
if (mkdir_if_needed(xdg_cache_home) == -1)
- goto fail;
+ goto path_fail;
path = concatenate_and_mkdir(local, xdg_cache_home, CACHE_DIR_NAME);
if (path == NULL)
- goto fail;
+ goto path_fail;
}
}
buf = NULL;
buf_size *= 2;
} else {
- goto fail;
+ goto path_fail;
}
}
path = concatenate_and_mkdir(local, pwd.pw_dir, ".cache");
if (path == NULL)
- goto fail;
+ goto path_fail;
path = concatenate_and_mkdir(local, path, CACHE_DIR_NAME);
if (path == NULL)
- goto fail;
+ goto path_fail;
}
- cache = ralloc(NULL, struct disk_cache);
- if (cache == NULL)
- goto fail;
-
cache->path = ralloc_strdup(cache, path);
if (cache->path == NULL)
- goto fail;
+ goto path_fail;
path = ralloc_asprintf(local, "%s/index", cache->path);
if (path == NULL)
- goto fail;
+ goto path_fail;
fd = open(path, O_RDWR | O_CREAT | O_CLOEXEC, 0644);
if (fd == -1)
- goto fail;
+ goto path_fail;
if (fstat(fd, &sb) == -1)
- goto fail;
+ goto path_fail;
/* Force the index file to be the expected size. */
size = sizeof(*cache->size) + CACHE_INDEX_MAX_KEYS * CACHE_KEY_SIZE;
if (sb.st_size != size) {
if (ftruncate(fd, size) == -1)
- goto fail;
+ goto path_fail;
}
/* We map this shared so that other processes see updates that we
cache->index_mmap = mmap(NULL, size, PROT_READ | PROT_WRITE,
MAP_SHARED, fd, 0);
if (cache->index_mmap == MAP_FAILED)
- goto fail;
+ goto path_fail;
cache->index_mmap_size = size;
- close(fd);
-
cache->size = (uint64_t *) cache->index_mmap;
cache->stored_keys = cache->index_mmap + sizeof(uint64_t);
cache->max_size = max_size;
- /* 1 thread was chosen because we don't really care about getting things
- * to disk quickly just that it's not blocking other tasks.
+ /* 4 threads were chosen below because just about all modern CPUs currently
+ * available that run Mesa have *at least* 4 cores. For these CPUs allowing
+ * more threads can result in the queue being processed faster, thus
+ * avoiding excessive memory use due to a backlog of cache entrys building
+ * up in the queue. Since we set the UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY
+ * flag this should have little negative impact on low core systems.
*
* The queue will resize automatically when it's full, so adding new jobs
* doesn't stall.
*/
- util_queue_init(&cache->cache_queue, "disk_cache", 32, 1,
+ util_queue_init(&cache->cache_queue, "disk$", 32, 4,
UTIL_QUEUE_INIT_RESIZE_IF_FULL |
- UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY);
+ UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY |
+ UTIL_QUEUE_INIT_SET_FULL_THREAD_AFFINITY);
+
+ cache->path_init_failed = false;
+
+ path_fail:
+
+ if (fd != -1)
+ close(fd);
- uint8_t cache_version = CACHE_VERSION;
- size_t cv_size = sizeof(cache_version);
cache->driver_keys_blob_size = cv_size;
/* Create driver id keys */
- size_t ts_size = strlen(timestamp) + 1;
+ size_t id_size = strlen(driver_id) + 1;
size_t gpu_name_size = strlen(gpu_name) + 1;
- cache->driver_keys_blob_size += ts_size;
+ cache->driver_keys_blob_size += id_size;
cache->driver_keys_blob_size += gpu_name_size;
/* We sometimes store entire structs that contains a pointers in the cache,
uint8_t *drv_key_blob = cache->driver_keys_blob;
DRV_KEY_CPY(drv_key_blob, &cache_version, cv_size)
- DRV_KEY_CPY(drv_key_blob, timestamp, ts_size)
+ DRV_KEY_CPY(drv_key_blob, driver_id, id_size)
DRV_KEY_CPY(drv_key_blob, gpu_name, gpu_name_size)
DRV_KEY_CPY(drv_key_blob, &ptr_size, ptr_size_size)
DRV_KEY_CPY(drv_key_blob, &driver_flags, driver_flags_size)
return cache;
fail:
- if (fd != -1)
- close(fd);
if (cache)
ralloc_free(cache);
ralloc_free(local);
void
disk_cache_destroy(struct disk_cache *cache)
{
- if (cache) {
+ if (cache && !cache->path_init_failed) {
+ util_queue_finish(&cache->cache_queue);
util_queue_destroy(&cache->cache_queue);
munmap(cache->index_mmap, cache->index_mmap_size);
}
ralloc_free(cache);
}
+void
+disk_cache_wait_for_idle(struct disk_cache *cache)
+{
+ util_queue_finish(&cache->cache_queue);
+}
+
/* Return a filename within the cache's directory corresponding to 'key'. The
* returned filename is ralloced with 'cache' as the parent context.
*
char buf[41];
char *filename;
+ if (cache->path_init_failed)
+ return NULL;
+
_mesa_sha1_format(buf, key);
if (asprintf(&filename, "%s/%c%c/%s", cache->path, buf[0],
buf[1], buf + 2) == -1)
deflate_and_write_to_disk(const void *in_data, size_t in_data_size, int dest,
const char *filename)
{
- unsigned char out[BUFSIZE];
+#ifdef HAVE_ZSTD
+ /* from the zstd docs (https://facebook.github.io/zstd/zstd_manual.html):
+ * compression runs faster if `dstCapacity` >= `ZSTD_compressBound(srcSize)`.
+ */
+ size_t out_size = ZSTD_compressBound(in_data_size);
+ void * out = malloc(out_size);
+
+ size_t ret = ZSTD_compress(out, out_size, in_data, in_data_size,
+ ZSTD_COMPRESSION_LEVEL);
+ if (ZSTD_isError(ret)) {
+ free(out);
+ return 0;
+ }
+ ssize_t written = write_all(dest, out, ret);
+ if (written == -1) {
+ free(out);
+ return 0;
+ }
+ free(out);
+ return ret;
+#else
+ unsigned char *out;
/* allocate deflate state */
z_stream strm;
/* compress until end of in_data */
size_t compressed_size = 0;
int flush;
+
+ out = malloc(BUFSIZE * sizeof(unsigned char));
+ if (out == NULL)
+ return 0;
+
do {
int remaining = in_data_size - BUFSIZE;
flush = remaining > 0 ? Z_NO_FLUSH : Z_FINISH;
ssize_t written = write_all(dest, out, have);
if (written == -1) {
(void)deflateEnd(&strm);
+ free(out);
return 0;
}
} while (strm.avail_out == 0);
/* clean up and return */
(void)deflateEnd(&strm);
+ free(out);
return compressed_size;
+# endif
}
static struct disk_cache_put_job *
return dc_job;
fail:
- free(dc_job->cache_item_metadata.keys);
free(dc_job);
return NULL;
* open with the flock held. So just let that file be responsible
* for writing the file.
*/
+#ifdef HAVE_FLOCK
err = flock(fd, LOCK_EX | LOCK_NB);
+#else
+ struct flock lock = {
+ .l_start = 0,
+ .l_len = 0, /* entire file */
+ .l_type = F_WRLCK,
+ .l_whence = SEEK_SET
+ };
+ err = fcntl(fd, F_SETLK, &lock);
+#endif
if (err == -1)
goto done;
*/
if (fd != -1)
close(fd);
- if (filename_tmp)
- free(filename_tmp);
- if (filename)
- free(filename);
+ free(filename_tmp);
+ free(filename);
}
void
const void *data, size_t size,
struct cache_item_metadata *cache_item_metadata)
{
+ if (cache->blob_put_cb) {
+ cache->blob_put_cb(key, CACHE_KEY_SIZE, data, size);
+ return;
+ }
+
+ if (cache->path_init_failed)
+ return;
+
struct disk_cache_put_job *dc_job =
create_put_job(cache, key, data, size, cache_item_metadata);
if (dc_job) {
util_queue_fence_init(&dc_job->fence);
util_queue_add_job(&cache->cache_queue, dc_job, &dc_job->fence,
- cache_put, destroy_put_job);
+ cache_put, destroy_put_job, dc_job->size);
}
}
inflate_cache_data(uint8_t *in_data, size_t in_data_size,
uint8_t *out_data, size_t out_data_size)
{
+#ifdef HAVE_ZSTD
+ size_t ret = ZSTD_decompress(out_data, out_data_size, in_data, in_data_size);
+ return !ZSTD_isError(ret);
+#else
z_stream strm;
/* allocate inflate state */
/* clean up and return */
(void)inflateEnd(&strm);
return true;
+#endif
}
void *
if (size)
*size = 0;
+ if (cache->blob_get_cb) {
+ /* This is what Android EGL defines as the maxValueSize in egl_cache_t
+ * class implementation.
+ */
+ const signed long max_blob_size = 64 * 1024;
+ void *blob = malloc(max_blob_size);
+ if (!blob)
+ return NULL;
+
+ signed long bytes =
+ cache->blob_get_cb(key, CACHE_KEY_SIZE, blob, max_blob_size);
+
+ if (!bytes) {
+ free(blob);
+ return NULL;
+ }
+
+ if (size)
+ *size = bytes;
+ return blob;
+ }
+
filename = get_cache_file(cache, key);
if (filename == NULL)
goto fail;
* TODO: pass the metadata back to the caller and do some basic
* validation.
*/
- cache_item_md_size += sizeof(cache_key);
+ cache_item_md_size += num_keys * sizeof(cache_key);
ret = lseek(fd, num_keys * sizeof(cache_key), SEEK_CUR);
if (ret == -1)
goto fail;
free(data);
free(filename);
+ free(file_header);
close(fd);
if (size)
disk_cache_put_key(struct disk_cache *cache, const cache_key key)
{
const uint32_t *key_chunk = (const uint32_t *) key;
- int i = *key_chunk & CACHE_INDEX_KEY_MASK;
+ int i = CPU_TO_LE32(*key_chunk) & CACHE_INDEX_KEY_MASK;
unsigned char *entry;
+ if (cache->blob_put_cb) {
+ cache->blob_put_cb(key, CACHE_KEY_SIZE, key_chunk, sizeof(uint32_t));
+ return;
+ }
+
+ if (cache->path_init_failed)
+ return;
+
entry = &cache->stored_keys[i * CACHE_KEY_SIZE];
memcpy(entry, key, CACHE_KEY_SIZE);
disk_cache_has_key(struct disk_cache *cache, const cache_key key)
{
const uint32_t *key_chunk = (const uint32_t *) key;
- int i = *key_chunk & CACHE_INDEX_KEY_MASK;
+ int i = CPU_TO_LE32(*key_chunk) & CACHE_INDEX_KEY_MASK;
unsigned char *entry;
+ if (cache->blob_get_cb) {
+ uint32_t blob;
+ return cache->blob_get_cb(key, CACHE_KEY_SIZE, &blob, sizeof(uint32_t));
+ }
+
+ if (cache->path_init_failed)
+ return false;
+
entry = &cache->stored_keys[i * CACHE_KEY_SIZE];
return memcmp(entry, key, CACHE_KEY_SIZE) == 0;
_mesa_sha1_final(&ctx, key);
}
+void
+disk_cache_set_callbacks(struct disk_cache *cache, disk_cache_put_cb put,
+ disk_cache_get_cb get)
+{
+ cache->blob_put_cb = put;
+ cache->blob_get_cb = get;
+}
+
#endif /* ENABLE_SHADER_CACHE */
projects / mesa.git / blobdiff