src/util/disk_cache.c

   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
  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 *
 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 *
 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
 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 *
 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
 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 *
 368 choose_random_file_matching(const char *dir_path,
 369                             bool (*predicate)(struct dirent *))
 370 {
 371    DIR *dir;
 372    struct dirent *entry;
 373    unsigned int count, victim;
 374    char *filename;
 375
 376    dir = opendir(dir_path);
 377    if (dir == NULL)
 378       return NULL;
 379
 380    count = 0;
 381
 382    while (1) {
 383       entry = readdir(dir);
 384       if (entry == NULL)
 385          break;
 386       if (! predicate(entry))
 387          continue;
 388
 389       count++;
 390    }
 391
 392    if (count == 0) {
 393       closedir(dir);
 394       return NULL;
 395    }
 396
 397    victim = rand() % count;
 398
 399    rewinddir(dir);
 400    count = 0;
 401
 402    while (1) {
 403       entry = readdir(dir);
 404       if (entry == NULL)
 405          break;
 406       if (! predicate(entry))
 407          continue;
 408       if (count == victim)
 409          break;
 410
 411       count++;
 412    }
 413
 414    if (entry == NULL) {
 415       closedir(dir);
 416       return NULL;
 417    }
 418
 419    if (asprintf(&filename, "%s/%s", dir_path, entry->d_name) < 0)
 420       filename = NULL;
 421
 422    closedir(dir);
 423
 424    return filename;
 425 }
 426
 427 /* Is entry a regular file, and not having a name with a trailing
 428  * ".tmp"
 429  */
 430 static bool
 431 is_regular_non_tmp_file(struct dirent *entry)
 432 {
 433    size_t len;
 434
 435    if (entry->d_type != DT_REG)
 436       return false;
 437
 438    len = strlen (entry->d_name);
 439    if (len >= 4 && strcmp(&entry->d_name[len-4], ".tmp") == 0)
 440       return false;
 441
 442    return true;
 443 }
 444
 445 /* Returns the size of the deleted file, (or 0 on any error). */
 446 static size_t
 447 unlink_random_file_from_directory(const char *path)
 448 {
 449    struct stat sb;
 450    char *filename;
 451
 452    filename = choose_random_file_matching(path, is_regular_non_tmp_file);
 453    if (filename == NULL)
 454       return 0;
 455
 456    if (stat(filename, &sb) == -1) {
 457       free (filename);
 458       return 0;
 459    }
 460
 461    unlink(filename);
 462
 463    free (filename);
 464
 465    return sb.st_size;
 466 }
 467
 468 /* Is entry a directory with a two-character name, (and not the
 469  * special name of "..")
 470  */
 471 static bool
 472 is_two_character_sub_directory(struct dirent *entry)
 473 {
 474    if (entry->d_type != DT_DIR)
 475       return false;
 476
 477    if (strlen(entry->d_name) != 2)
 478       return false;
 479
 480    if (strcmp(entry->d_name, "..") == 0)
 481       return false;
 482
 483    return true;
 484 }
 485
 486 static void
 487 evict_random_item(struct disk_cache *cache)
 488 {
 489    const char hex[] = "0123456789abcde";
 490    char *dir_path;
 491    int a, b;
 492    size_t size;
 493
 494    /* With a reasonably-sized, full cache, (and with keys generated
 495     * from a cryptographic hash), we can choose two random hex digits
 496     * and reasonably expect the directory to exist with a file in it.
 497     */
 498    a = rand() % 16;
 499    b = rand() % 16;
 500
 501    if (asprintf(&dir_path, "%s/%c%c", cache->path, hex[a], hex[b]) < 0)
 502       return;
 503
 504    size = unlink_random_file_from_directory(dir_path);
 505
 506    free(dir_path);
 507
 508    if (size) {
 509       p_atomic_add(cache->size, - size);
 510       return;
 511    }
 512
 513    /* In the case where the random choice of directory didn't find
 514     * something, we choose randomly from the existing directories.
 515     *
 516     * Really, the only reason this code exists is to allow the unit
 517     * tests to work, (which use an artificially-small cache to be able
 518     * to force a single cached item to be evicted).
 519     */
 520    dir_path = choose_random_file_matching(cache->path,
 521                                           is_two_character_sub_directory);
 522    if (dir_path == NULL)
 523       return;
 524
 525    size = unlink_random_file_from_directory(dir_path);
 526
 527    free(dir_path);
 528
 529    if (size)
 530       p_atomic_add(cache->size, - size);
 531 }
 532
 533 void
 534 disk_cache_put(struct disk_cache *cache,
 535           cache_key key,
 536           const void *data,
 537           size_t size)
 538 {
 539    int fd = -1, fd_final = -1, err, ret;
 540    size_t len;
 541    char *filename = NULL, *filename_tmp = NULL;
 542    const char *p = data;
 543
 544    filename = get_cache_file(cache, key);
 545    if (filename == NULL)
 546       goto done;
 547
 548    /* Write to a temporary file to allow for an atomic rename to the
 549     * final destination filename, (to prevent any readers from seeing
 550     * a partially written file).
 551     */
 552    filename_tmp = ralloc_asprintf(cache, "%s.tmp", filename);
 553    if (filename_tmp == NULL)
 554       goto done;
 555
 556    fd = open(filename_tmp, O_WRONLY | O_CLOEXEC | O_CREAT, 0644);
 557
 558    /* Make the two-character subdirectory within the cache as needed. */
 559    if (fd == -1) {
 560       if (errno != ENOENT)
 561          goto done;
 562
 563       make_cache_file_directory(cache, key);
 564
 565       fd = open(filename_tmp, O_WRONLY | O_CLOEXEC | O_CREAT, 0644);
 566       if (fd == -1)
 567          goto done;
 568    }
 569
 570    /* With the temporary file open, we take an exclusive flock on
 571     * it. If the flock fails, then another process still has the file
 572     * open with the flock held. So just let that file be responsible
 573     * for writing the file.
 574     */
 575    err = flock(fd, LOCK_EX | LOCK_NB);
 576    if (err == -1)
 577       goto done;
 578
 579    /* Now that we have the lock on the open temporary file, we can
 580     * check to see if the destination file already exists. If so,
 581     * another process won the race between when we saw that the file
 582     * didn't exist and now. In this case, we don't do anything more,
 583     * (to ensure the size accounting of the cache doesn't get off).
 584     */
 585    fd_final = open(filename, O_RDONLY | O_CLOEXEC);
 586    if (fd_final != -1)
 587       goto done;
 588
 589    /* OK, we're now on the hook to write out a file that we know is
 590     * not in the cache, and is also not being written out to the cache
 591     * by some other process.
 592     *
 593     * Before we do that, if the cache is too large, evict something
 594     * else first.
 595     */
 596    if (*cache->size + size > cache->max_size)
 597       evict_random_item(cache);
 598
 599    /* Now, finally, write out the contents to the temporary file, then
 600     * rename them atomically to the destination filename, and also
 601     * perform an atomic increment of the total cache size.
 602     */
 603    for (len = 0; len < size; len += ret) {
 604       ret = write(fd, p + len, size - len);
 605       if (ret == -1) {
 606          unlink(filename_tmp);
 607          goto done;
 608       }
 609    }
 610
 611    rename(filename_tmp, filename);
 612
 613    p_atomic_add(cache->size, size);
 614
 615  done:
 616    if (fd_final != -1)
 617       close(fd_final);
 618    /* This close finally releases the flock, (now that the final dile
 619     * has been renamed into place and the size has been added).
 620     */
 621    if (fd != -1)
 622       close(fd);
 623    if (filename_tmp)
 624       ralloc_free(filename_tmp);
 625    if (filename)
 626       ralloc_free(filename);
 627 }
 628
 629 void *
 630 disk_cache_get(struct disk_cache *cache, cache_key key, size_t *size)
 631 {
 632    int fd = -1, ret, len;
 633    struct stat sb;
 634    char *filename = NULL;
 635    uint8_t *data = NULL;
 636
 637    if (size)
 638       *size = 0;
 639
 640    filename = get_cache_file(cache, key);
 641    if (filename == NULL)
 642       goto fail;
 643
 644    fd = open(filename, O_RDONLY | O_CLOEXEC);
 645    if (fd == -1)
 646       goto fail;
 647
 648    if (fstat(fd, &sb) == -1)
 649       goto fail;
 650
 651    data = malloc(sb.st_size);
 652    if (data == NULL)
 653       goto fail;
 654
 655    for (len = 0; len < sb.st_size; len += ret) {
 656       ret = read(fd, data + len, sb.st_size - len);
 657       if (ret == -1)
 658          goto fail;
 659    }
 660
 661    ralloc_free(filename);
 662    close(fd);
 663
 664    if (size)
 665       *size = sb.st_size;
 666
 667    return data;
 668
 669  fail:
 670    if (data)
 671       free(data);
 672    if (filename)
 673       ralloc_free(filename);
 674    if (fd != -1)
 675       close(fd);
 676
 677    return NULL;
 678 }
 679
 680 void
 681 disk_cache_put_key(struct disk_cache *cache, cache_key key)
 682 {
 683    uint32_t *key_chunk = (uint32_t *) key;
 684    int i = *key_chunk & CACHE_INDEX_KEY_MASK;
 685    unsigned char *entry;
 686
 687    entry = &cache->stored_keys[i + CACHE_KEY_SIZE];
 688
 689    memcpy(entry, key, CACHE_KEY_SIZE);
 690 }
 691
 692 /* This function lets us test whether a given key was previously
 693  * stored in the cache with disk_cache_put_key(). The implement is
 694  * efficient by not using syscalls or hitting the disk. It's not
 695  * race-free, but the races are benign. If we race with someone else
 696  * calling disk_cache_put_key, then that's just an extra cache miss and an
 697  * extra recompile.
 698  */
 699 bool
 700 disk_cache_has_key(struct disk_cache *cache, cache_key key)
 701 {
 702    uint32_t *key_chunk = (uint32_t *) key;
 703    int i = *key_chunk & CACHE_INDEX_KEY_MASK;
 704    unsigned char *entry;
 705
 706    entry = &cache->stored_keys[i + CACHE_KEY_SIZE];
 707
 708    return memcmp(entry, key, CACHE_KEY_SIZE) == 0;
 709 }
 710
 711 #endif