util/fossilize_db: add basic fossilize db util to read/write shader caches

My benchmarking shows no significant change in cache load times with a
single shader cache file vs the existing cache implementation of many
small files (tested with my spinning rust HDD).

However this new single file cache implementation does reduce the total
size on disk used by the shader cache. We have a problem with the existing
cache where writing tiny files to disk causes more disk space to be used than
is actually needed for the files due to the minimum size required for a file.
In pratice this tends to inflate the size of the cache on disk to over 3x
larger.

There are other advantages of using a single file for shader cache entries
such as allowing better removal of cache entries once we hit the max cache
size limit (although we don't implement any max cache size handling in this
initial implementation).

The primary reason for implementing a single file cache for now is to allow
better performance and handling by third party applications such as steam
that collect and distribute precompiled cache entries.

For this reason we also implement a new environment variable
MESA_DISK_CACHE_READ_ONLY_FOZ_DBS which allows a user to pass in a path
to a number of external read only shader cache dbs. There is an initial
limit of 8 dbs that can be passed to mesa like so:

MESA_DISK_CACHE_READ_ONLY_FOZ_DBS=/full_path/filename1, ... ,/full_path/filename8

Where the filename represents the cache db and its index file e.g.
filename1.foz and filename1_idx.foz

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7725>
This commit is contained in:
Timothy Arceri 2020-10-19 14:33:44 +11:00 committed by Marge Bot
parent 1fabc4ecbe
commit eca6bb9540
3 changed files with 642 additions and 0 deletions

540
src/util/fossilize_db.c Normal file
View File

@ -0,0 +1,540 @@
/*
* Copyright © 2020 Valve Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
/* This is a basic c implementation of a fossilize db like format intended for
* use with the Mesa shader cache.
*
* The format is compatible enough to allow the fossilize db tools to be used
* to do things like merge db collections, but unlike fossilize db which uses
* a zlib implementation for compression of data entries, we use zstd for
* compression.
*/
#include "fossilize_db.h"
#ifdef FOZ_DB_UTIL
#include <assert.h>
#include <stddef.h>
#include <stdlib.h>
#include <string.h>
#include <sys/file.h>
#include <sys/types.h>
#include <unistd.h>
#include "zstd.h"
#include "crc32.h"
#include "hash_table.h"
#include "mesa-sha1.h"
#include "ralloc.h"
/* 3 is the recomended level, with 22 as the absolute maximum */
#define ZSTD_COMPRESSION_LEVEL 3
#define FOZ_REF_MAGIC_SIZE 16
static const uint8_t stream_reference_magic_and_version[FOZ_REF_MAGIC_SIZE] = {
0x81, 'F', 'O', 'S',
'S', 'I', 'L', 'I',
'Z', 'E', 'D', 'B',
0, 0, 0, FOSSILIZE_FORMAT_VERSION, /* 4 bytes to use for versioning. */
};
/* Mesa uses 160bit hashes to identify cache entries, a hash of this size
* makes collisions virtually impossible for our use case. However the foz db
* format uses a 64bit hash table to lookup file offsets for reading cache
* entries so we must shorten our hash.
*/
static uint64_t
truncate_hash_to_64bits(const uint8_t *cache_key)
{
uint64_t hash = 0;
unsigned shift = 7;
for (unsigned i = 0; i < 8; i++) {
hash |= ((uint64_t)cache_key[i]) << shift * 8;
shift--;
}
return hash;
}
static bool
check_files_opened_successfully(FILE *file, FILE *db_idx)
{
if (!file) {
if (db_idx)
fclose(db_idx);
return false;
}
if (!db_idx) {
if (file)
fclose(file);
return false;
}
return true;
}
static bool
create_foz_db_filenames(char *base_filename, char **filename,
char **idx_filename)
{
if (asprintf(filename, "%s.foz", base_filename) == -1)
return false;
if (asprintf(idx_filename, "%s_idx.foz", base_filename) == -1) {
free(*filename);
return false;
}
return true;
}
static bool
load_foz_dbs(struct foz_db *foz_db, FILE *db_idx, uint8_t file_idx,
bool read_only)
{
int err = flock(fileno(foz_db->file[file_idx]), LOCK_EX | LOCK_NB);
if (err == -1)
goto fail;
err = flock(fileno(db_idx), LOCK_EX | LOCK_NB);
if (err == -1)
goto fail;
/* Scan through the archive and get the list of cache entries. */
fseek(db_idx, 0, SEEK_END);
size_t len = ftell(db_idx);
rewind(db_idx);
if (!read_only)
fseek(foz_db->file[file_idx], 0, SEEK_END);
if (len != 0) {
uint8_t magic[FOZ_REF_MAGIC_SIZE];
if (fread(magic, 1, FOZ_REF_MAGIC_SIZE, db_idx) != FOZ_REF_MAGIC_SIZE)
goto fail;
if (memcmp(magic, stream_reference_magic_and_version,
FOZ_REF_MAGIC_SIZE - 1))
goto fail;
int version = magic[FOZ_REF_MAGIC_SIZE - 1];
if (version > FOSSILIZE_FORMAT_VERSION ||
version < FOSSILIZE_FORMAT_MIN_COMPAT_VERSION)
goto fail;
size_t offset = FOZ_REF_MAGIC_SIZE;
size_t begin_append_offset = len;
while (offset < len) {
begin_append_offset = offset;
char bytes_to_read[FOSSILIZE_BLOB_HASH_LENGTH + sizeof(struct foz_payload_header)];
struct foz_payload_header *header;
/* Corrupt entry. Our process might have been killed before we
* could write all data.
*/
if (offset + sizeof(bytes_to_read) > len)
break;
/* NAME + HEADER in one read */
if (fread(bytes_to_read, 1, sizeof(bytes_to_read), db_idx) !=
sizeof(bytes_to_read))
goto fail;
offset += sizeof(bytes_to_read);
header = (struct foz_payload_header*)&bytes_to_read[FOSSILIZE_BLOB_HASH_LENGTH];
/* Corrupt entry. Our process might have been killed before we
* could write all data.
*/
if (offset + header->payload_size > len ||
header->payload_size != sizeof(uint64_t))
break;
char hash_str[FOSSILIZE_BLOB_HASH_LENGTH + 1] = {};
memcpy(hash_str, bytes_to_read, FOSSILIZE_BLOB_HASH_LENGTH);
struct foz_db_entry *entry = ralloc(foz_db->mem_ctx,
struct foz_db_entry);
entry->header = *header;
entry->file_idx = file_idx;
_mesa_sha1_hex_to_sha1(entry->key, hash_str);
/* read cache item offset from index file */
uint64_t cache_offset;
if (fread(&cache_offset, 1, sizeof(cache_offset), db_idx) !=
sizeof(cache_offset))
return false;
entry->offset = cache_offset;
/* Truncate the entry's hash string to a 64bit hash for use with a
* 64bit hash table for looking up file offsets.
*/
hash_str[16] = '\0';
uint64_t key = strtoull(hash_str, NULL, 16);
_mesa_hash_table_u64_insert(foz_db->index_db, key, entry);
offset += header->payload_size;
}
if (!read_only && offset != len) {
if (fseek(db_idx, begin_append_offset, SEEK_SET) < 0)
goto fail;
}
} else {
/* Appending to a fresh file. Make sure we have the magic. */
if (fwrite(stream_reference_magic_and_version, 1,
sizeof(stream_reference_magic_and_version), foz_db->file[file_idx]) !=
sizeof(stream_reference_magic_and_version))
goto fail;
if (fwrite(stream_reference_magic_and_version, 1,
sizeof(stream_reference_magic_and_version), db_idx) !=
sizeof(stream_reference_magic_and_version))
goto fail;
}
foz_db->alive = true;
return true;
fail:
foz_destroy(foz_db);
return false;
}
/* Here we open mesa cache foz dbs files. If the files exist we load the index
* db into a hash table. The index db contains the offsets needed to later
* read cache entries from the foz db containing the actual cache entries.
*/
bool
foz_prepare(struct foz_db *foz_db, char *base_filename)
{
char *filename = NULL;
char *idx_filename = NULL;
if (!create_foz_db_filenames(base_filename, &filename, &idx_filename))
return false;
/* Open the default foz dbs for read/write. If the files didn't already exist
* create them.
*/
foz_db->file[0] = fopen(filename, "a+b");
foz_db->db_idx = fopen(idx_filename, "a+b");
free(filename);
free(idx_filename);
if (!check_files_opened_successfully(foz_db->file[0], foz_db->db_idx))
return false;
simple_mtx_init(&foz_db->mtx, mtx_plain);
foz_db->mem_ctx = ralloc_context(NULL);
foz_db->index_db = _mesa_hash_table_u64_create(NULL);
if (!load_foz_dbs(foz_db, foz_db->db_idx, 0, false))
return false;
uint8_t file_idx = 1;
char *foz_dbs = getenv("MESA_DISK_CACHE_READ_ONLY_FOZ_DBS");
if (!foz_dbs)
return true;
for (unsigned n; n = strcspn(foz_dbs, ","), *foz_dbs;
foz_dbs += MAX2(1, n)) {
char *foz_db_filename = strndup(foz_dbs, n);
filename = NULL;
idx_filename = NULL;
if (!create_foz_db_filenames(foz_db_filename, &filename, &idx_filename)) {
free(foz_db_filename);
continue; /* Ignore invalid user provided filename and continue */
}
free(foz_db_filename);
/* Open files as read only */
foz_db->file[file_idx] = fopen(filename, "rb");
FILE *db_idx = fopen(idx_filename, "rb");
free(filename);
free(idx_filename);
if (!check_files_opened_successfully(foz_db->file[file_idx], db_idx)) {
fclose(db_idx);
continue; /* Ignore invalid user provided filename and continue */
}
if (!load_foz_dbs(foz_db, db_idx, file_idx, true)) {
fclose(db_idx);
return false;
}
fclose(db_idx);
file_idx++;
if (file_idx >= FOZ_MAX_DBS)
break;
}
return true;
}
void
foz_destroy(struct foz_db *foz_db)
{
fclose(foz_db->db_idx);
for (unsigned i = 0; i < FOZ_MAX_DBS; i++) {
if (foz_db->file[i])
fclose(foz_db->file[i]);
}
if (foz_db->mem_ctx) {
_mesa_hash_table_u64_destroy(foz_db->index_db, NULL);
ralloc_free(foz_db->mem_ctx);
simple_mtx_destroy(&foz_db->mtx);
}
}
/* Here we lookup a cache entry in the index hash table. If an entry is found
* we use the retrieved offset to read the cache entry from disk.
*/
void *
foz_read_entry(struct foz_db *foz_db, const uint8_t *cache_key_160bit,
size_t *size)
{
uint64_t hash = truncate_hash_to_64bits(cache_key_160bit);
uint8_t *compressed_data = NULL;
void *blob = NULL;
if (!foz_db->alive)
return NULL;
simple_mtx_lock(&foz_db->mtx);
struct foz_db_entry *entry =
_mesa_hash_table_u64_search(foz_db->index_db, hash);
if (!entry) {
simple_mtx_unlock(&foz_db->mtx);
return NULL;
}
uint8_t file_idx = entry->file_idx;
off_t offset = ftell(foz_db->file[file_idx]);
if (fseek(foz_db->file[file_idx], entry->offset, SEEK_SET) < 0)
goto fail;
uint32_t header_size = sizeof(struct foz_payload_header);
if (fread(&entry->header, 1, header_size, foz_db->file[file_idx]) !=
header_size)
goto fail;
/* Check for collision using full 160bit hash for increased assurance
* against potential collisions.
*/
for (int i = 0; i < 20; i++) {
if (cache_key_160bit[i] != entry->key[i])
goto fail;
}
uint32_t compressed_data_sz = entry->header.payload_size;
compressed_data = malloc(compressed_data_sz);
if (!compressed_data ||
(fread(compressed_data, 1, compressed_data_sz, foz_db->file[file_idx]) !=
compressed_data_sz))
goto fail;
/* verify checksum */
if (entry->header.crc != 0) {
if (util_hash_crc32(compressed_data, compressed_data_sz) !=
entry->header.crc)
goto fail;
}
/* Uncompress the db entry */
uint32_t out_size = entry->header.uncompressed_size;
blob = malloc(out_size);
if (!blob)
goto fail;
size_t ret = ZSTD_decompress(blob, out_size, compressed_data,
compressed_data_sz);
if (ZSTD_isError(ret))
goto fail;
free(compressed_data);
simple_mtx_unlock(&foz_db->mtx);
if (size)
*size = out_size;
/* Reset file offset to the end of the file ready for writing */
fseek(foz_db->file[file_idx], offset, SEEK_SET);
return blob;
fail:
if (compressed_data)
free(compressed_data);
if (blob)
free(blob);
/* reading db entry failed. reset the file offset */
fseek(foz_db->file[file_idx], offset, SEEK_SET);
simple_mtx_unlock(&foz_db->mtx);
return NULL;
}
/* Here we write the cache entry to disk and store its offset in the index db.
*/
bool
foz_write_entry(struct foz_db *foz_db, const uint8_t *cache_key_160bit,
const void *blob, size_t size)
{
uint64_t hash = truncate_hash_to_64bits(cache_key_160bit);
if (!foz_db->alive)
return false;
simple_mtx_lock(&foz_db->mtx);
struct foz_db_entry *entry =
_mesa_hash_table_u64_search(foz_db->index_db, hash);
if (entry) {
simple_mtx_unlock(&foz_db->mtx);
return NULL;
}
/* Prepare db entry header and blob ready for writing */
struct foz_payload_header header;
header.uncompressed_size = size;
header.format = FOSSILIZE_COMPRESSION_DEFLATE;
/* 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(size);
void *out = malloc(out_size);
size_t blob_size = ZSTD_compress(out, out_size, blob, size,
ZSTD_COMPRESSION_LEVEL);
if (ZSTD_isError(blob_size))
goto fail;
header.payload_size = blob_size;
header.crc = util_hash_crc32(out, blob_size);
/* Write hash header to db */
char hash_str[FOSSILIZE_BLOB_HASH_LENGTH + 1]; /* 40 digits + null */
_mesa_sha1_format(hash_str, cache_key_160bit);
if (fwrite(hash_str, 1, FOSSILIZE_BLOB_HASH_LENGTH, foz_db->file[0]) !=
FOSSILIZE_BLOB_HASH_LENGTH)
goto fail;
off_t offset = ftell(foz_db->file[0]);
/* Write db entry header */
if (fwrite(&header, 1, sizeof(header), foz_db->file[0]) != sizeof(header))
goto fail;
/* Now write the db entry blob */
if (fwrite(out, 1, blob_size, foz_db->file[0]) != blob_size)
goto fail;
/* Flush everything to file to reduce chance of cache corruption */
fflush(foz_db->file[0]);
/* Write hash header to index db */
if (fwrite(hash_str, 1, FOSSILIZE_BLOB_HASH_LENGTH, foz_db->db_idx) !=
FOSSILIZE_BLOB_HASH_LENGTH)
goto fail;
header.uncompressed_size = sizeof(uint64_t);
header.format = FOSSILIZE_COMPRESSION_NONE;
header.payload_size = sizeof(uint64_t);
header.crc = 0;
if (fwrite(&header, 1, sizeof(header), foz_db->db_idx) !=
sizeof(header))
goto fail;
if (fwrite(&offset, 1, sizeof(uint64_t), foz_db->db_idx) !=
sizeof(uint64_t))
goto fail;
/* Flush everything to file to reduce chance of cache corruption */
fflush(foz_db->db_idx);
entry = ralloc(foz_db->mem_ctx, struct foz_db_entry);
entry->header = header;
entry->offset = offset;
entry->file_idx = 0;
_mesa_sha1_hex_to_sha1(entry->key, hash_str);
_mesa_hash_table_u64_insert(foz_db->index_db, hash, entry);
simple_mtx_unlock(&foz_db->mtx);
free(out);
return true;
fail:
simple_mtx_unlock(&foz_db->mtx);
free(out);
return false;
}
#else
bool
foz_prepare(struct foz_db *foz_db, char *filename)
{
fprintf(stderr, "Warning: Mesa single file cache selected but Mesa wasn't "
"built with single cache file support. Shader cache will be disabled"
"!\n");
return false;
}
void
foz_destroy(struct foz_db *foz_db)
{
}
void *
foz_read_entry(struct foz_db *foz_db, const uint8_t *cache_key_160bit,
size_t *size)
{
return false;
}
bool
foz_write_entry(struct foz_db *foz_db, const uint8_t *cache_key_160bit,
const void *blob, size_t size)
{
return false;
}
#endif

100
src/util/fossilize_db.h Normal file
View File

@ -0,0 +1,100 @@
/*
* Copyright © 2020 Valve Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
/* This is a basic c implementation of a fossilize db like format intended for
* use with the Mesa shader cache.
*
* The format is compatible enough to allow the fossilize db tools to be used
* to do things like merge db collections, but unlike fossilize db which uses
* a zlib implementation for compression of data entries, we use zstd for
* compression.
*/
#ifndef FOSSILIZE_DB_H
#define FOSSILIZE_DB_H
#ifdef HAVE_ZSTD
#ifdef HAVE_FLOCK
#define FOZ_DB_UTIL 1
#endif
#endif
#include <stdbool.h>
#include <stdint.h>
#include <stdio.h>
#include "simple_mtx.h"
/* Max number of DBs our implementation can read from at once */
#define FOZ_MAX_DBS 9 /* Default DB + 8 Read only DBs */
#define FOSSILIZE_BLOB_HASH_LENGTH 40
enum {
FOSSILIZE_COMPRESSION_NONE = 1,
FOSSILIZE_COMPRESSION_DEFLATE = 2
};
enum {
FOSSILIZE_FORMAT_VERSION = 6,
FOSSILIZE_FORMAT_MIN_COMPAT_VERSION = 5
};
struct foz_payload_header {
uint32_t payload_size;
uint32_t format;
uint32_t crc;
uint32_t uncompressed_size;
};
struct foz_db_entry {
uint8_t file_idx;
uint8_t key[20];
uint64_t offset;
struct foz_payload_header header;
};
struct foz_db {
FILE *file[FOZ_MAX_DBS]; /* An array of all foz dbs */
FILE *db_idx; /* The default writable foz db idx */
simple_mtx_t mtx; /* Mutex for file/hash table read/writes */
void *mem_ctx;
struct hash_table_u64 *index_db; /* Hash table of all foz db entries */
bool alive;
};
bool
foz_prepare(struct foz_db *foz_db, char *filename);
void
foz_destroy(struct foz_db *foz_db);
void *
foz_read_entry(struct foz_db *foz_db, const uint8_t *cache_key_160bit,
size_t *size);
bool
foz_write_entry(struct foz_db *foz_db, const uint8_t *cache_key_160bit,
const void *blob, size_t size);
#endif /* FOSSILIZE_DB_H */

View File

@ -52,6 +52,8 @@ files_mesa_util = files(
'format_r11g11b10f.h',
'format_rgb9e5.h',
'format_srgb.h',
'fossilize_db.c',
'fossilize_db.h',
'futex.h',
'half_float.c',
'half_float.h',