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 #include <ctype.h>
  25 #include <string.h>
  26 #include <stdlib.h>
  27 #include <stdio.h>
  28 #include <sys/file.h>
  29 #include <sys/types.h>
  30 #include <sys/stat.h>
  31 #include <sys/mman.h>
  32 #include <unistd.h>
  33 #include <fcntl.h>
  34 #include <pwd.h>
  35 #include <errno.h>
  36 #include <dirent.h>
  37
  38 #include "util/u_atomic.h"
  39 #include "util/mesa-sha1.h"
  40 #include "util/ralloc.h"
  41 #include "main/errors.h"
  42
  43 #include "disk_cache.h"
  44
  45 /* Number of bits to mask off from a cache key to get an index. */
  46 #define CACHE_INDEX_KEY_BITS 16
  47
  48 /* Mask for computing an index from a key. */
  49 #define CACHE_INDEX_KEY_MASK ((1 << CACHE_INDEX_KEY_BITS) - 1)
  50
  51 /* The number of keys that can be stored in the index. */
  52 #define CACHE_INDEX_MAX_KEYS (1 << CACHE_INDEX_KEY_BITS)
  53
  54 struct disk_cache {
  55    /* The path to the cache directory. */
  56    char *path;
  57
  58    /* A pointer to the mmapped index file within the cache directory. */
  59    uint8_t *index_mmap;
  60    size_t index_mmap_size;
  61
  62    /* Pointer to total size of all objects in cache (within index_mmap) */
  63    uint64_t *size;
  64
  65    /* Pointer to stored keys, (within index_mmap). */
  66    uint8_t *stored_keys;
  67
  68    /* Maximum size of all cached objects (in bytes). */
  69    uint64_t max_size;
  70 };
  71
  72 /* Create a directory named 'path' if it does not already exist.
  73  *
  74  * Returns: 0 if path already exists as a directory or if created.
  75  *         -1 in all other cases.
  76  */
  77 static int
  78 mkdir_if_needed(char *path)
  79 {
  80    struct stat sb;
  81
  82    /* If the path exists already, then our work is done if it's a
  83     * directory, but it's an error if it is not.
  84     */
  85    if (stat(path, &sb) == 0) {
  86       if (S_ISDIR(sb.st_mode)) {
  87          return 0;
  88       } else {
  89          fprintf(stderr, "Cannot use %s for shader cache (not a directory)"
  90                          "---disabling.\n", path);
  91          return -1;
  92       }
  93    }
  94
  95    int ret = mkdir(path, 0755);
  96    if (ret == 0 || (ret == -1 && errno == EEXIST))
  97      return 0;
  98
  99    fprintf(stderr, "Failed to create %s for shader cache (%s)---disabling.\n",
 100            path, strerror(errno));
 101
 102    return -1;
 103 }
 104
 105 /* Concatenate an existing path and a new name to form a new path.  If the new
 106  * path does not exist as a directory, create it then return the resulting
 107  * name of the new path (ralloc'ed off of 'ctx').
 108  *
 109  * Returns NULL on any error, such as:
 110  *
 111  *      <path> does not exist or is not a directory
 112  *      <path>/<name> exists but is not a directory
 113  *      <path>/<name> cannot be created as a directory
 114  */
 115 static char *
 116 concatenate_and_mkdir(void *ctx, char *path, char *name)
 117 {
 118    char *new_path;
 119    struct stat sb;
 120
 121    if (stat(path, &sb) != 0 || ! S_ISDIR(sb.st_mode))
 122       return NULL;
 123
 124    new_path = ralloc_asprintf(ctx, "%s/%s", path, name);
 125
 126    if (mkdir_if_needed(new_path) == 0)
 127       return new_path;
 128    else
 129       return NULL;
 130 }
 131
 132 struct disk_cache *
 133 disk_cache_create(void)
 134 {
 135    void *local;
 136    struct disk_cache *cache = NULL;
 137    char *path, *max_size_str;
 138    uint64_t max_size;
 139    int fd = -1;
 140    struct stat sb;
 141    size_t size;
 142
 143    /* A ralloc context for transient data during this invocation. */
 144    local = ralloc_context(NULL);
 145    if (local == NULL)
 146       goto fail;
 147
 148    /* At user request, disable shader cache entirely. */
 149    if (getenv("MESA_GLSL_CACHE_DISABLE"))
 150       goto fail;
 151
 152    /* Determine path for cache based on the first defined name as follows:
 153     *
 154     *   $MESA_GLSL_CACHE_DIR
 155     *   $XDG_CACHE_HOME/mesa
 156     *   <pwd.pw_dir>/.cache/mesa
 157     */
 158    path = getenv("MESA_GLSL_CACHE_DIR");
 159    if (path && mkdir_if_needed(path) == -1) {
 160       goto fail;
 161    }
 162
 163    if (path == NULL) {
 164       char *xdg_cache_home = getenv("XDG_CACHE_HOME");
 165
 166       if (xdg_cache_home) {
 167          if (mkdir_if_needed(xdg_cache_home) == -1)
 168             goto fail;
 169
 170          path = concatenate_and_mkdir(local, xdg_cache_home, "mesa");
 171          if (path == NULL)
 172             goto fail;
 173       }
 174    }
 175
 176    if (path == NULL) {
 177       char *buf;
 178       size_t buf_size;
 179       struct passwd pwd, *result;
 180
 181       buf_size = sysconf(_SC_GETPW_R_SIZE_MAX);
 182       if (buf_size == -1)
 183          buf_size = 512;
 184
 185       /* Loop until buf_size is large enough to query the directory */
 186       while (1) {
 187          buf = ralloc_size(local, buf_size);
 188
 189          getpwuid_r(getuid(), &pwd, buf, buf_size, &result);
 190          if (result)
 191             break;
 192
 193          if (errno == ERANGE) {
 194             ralloc_free(buf);
 195             buf = NULL;
 196             buf_size *= 2;
 197          } else {
 198             goto fail;
 199          }
 200       }
 201
 202       path = concatenate_and_mkdir(local, pwd.pw_dir, ".cache");
 203       if (path == NULL)
 204          goto fail;
 205
 206       path = concatenate_and_mkdir(local, path, "mesa");
 207       if (path == NULL)
 208          goto fail;
 209    }
 210
 211    cache = ralloc(NULL, struct disk_cache);
 212    if (cache == NULL)
 213       goto fail;
 214
 215    cache->path = ralloc_strdup(cache, path);
 216    if (cache->path == NULL)
 217       goto fail;
 218
 219    path = ralloc_asprintf(local, "%s/index", cache->path);
 220    if (path == NULL)
 221       goto fail;
 222
 223    fd = open(path, O_RDWR | O_CREAT | O_CLOEXEC, 0644);
 224    if (fd == -1)
 225       goto fail;
 226
 227    if (fstat(fd, &sb) == -1)
 228       goto fail;
 229
 230    /* Force the index file to be the expected size. */
 231    size = sizeof(*cache->size) + CACHE_INDEX_MAX_KEYS * CACHE_KEY_SIZE;
 232    if (sb.st_size != size) {
 233       if (ftruncate(fd, size) == -1)
 234          goto fail;
 235    }
 236
 237    /* We map this shared so that other processes see updates that we
 238     * make.
 239     *
 240     * Note: We do use atomic addition to ensure that multiple
 241     * processes don't scramble the cache size recorded in the
 242     * index. But we don't use any locking to prevent multiple
 243     * processes from updating the same entry simultaneously. The idea
 244     * is that if either result lands entirely in the index, then
 245     * that's equivalent to a well-ordered write followed by an
 246     * eviction and a write. On the other hand, if the simultaneous
 247     * writes result in a corrupt entry, that's not really any
 248     * different than both entries being evicted, (since within the
 249     * guarantees of the cryptographic hash, a corrupt entry is
 250     * unlikely to ever match a real cache key).
 251     */
 252    cache->index_mmap = mmap(NULL, size, PROT_READ | PROT_WRITE,
 253                             MAP_SHARED, fd, 0);
 254    if (cache->index_mmap == MAP_FAILED)
 255       goto fail;
 256    cache->index_mmap_size = size;
 257
 258    close(fd);
 259
 260    cache->size = (uint64_t *) cache->index_mmap;
 261    cache->stored_keys = cache->index_mmap + sizeof(uint64_t);
 262
 263    max_size = 0;
 264
 265    max_size_str = getenv("MESA_GLSL_CACHE_MAX_SIZE");
 266    if (max_size_str) {
 267       char *end;
 268       max_size = strtoul(max_size_str, &end, 10);
 269       if (end == max_size_str) {
 270          max_size = 0;
 271       } else {
 272          while (*end && isspace(*end))
 273             end++;
 274          switch (*end) {
 275          case 'K':
 276          case 'k':
 277             max_size *= 1024;
 278             break;
 279          case 'M':
 280          case 'm':
 281             max_size *= 1024*1024;
 282             break;
 283          case '\0':
 284          case 'G':
 285          case 'g':
 286          default:
 287             max_size *= 1024*1024*1024;
 288             break;
 289          }
 290       }
 291    }
 292
 293    /* Default to 1GB for maximum cache size. */
 294    if (max_size == 0)
 295       max_size = 1024*1024*1024;
 296
 297    cache->max_size = max_size;
 298
 299    ralloc_free(local);
 300
 301    return cache;
 302
 303  fail:
 304    if (fd != -1)
 305       close(fd);
 306    if (cache)
 307       ralloc_free(cache);
 308    ralloc_free(local);
 309
 310    return NULL;
 311 }
 312
 313 void
 314 disk_cache_destroy(struct disk_cache *cache)
 315 {
 316    munmap(cache->index_mmap, cache->index_mmap_size);
 317
 318    ralloc_free(cache);
 319 }
 320
 321 /* Return a filename within the cache's directory corresponding to 'key'. The
 322  * returned filename is ralloced with 'cache' as the parent context.
 323  *
 324  * Returns NULL if out of memory.
 325  */
 326 static char *
 327 get_cache_file(struct disk_cache *cache, cache_key key)
 328 {
 329    char buf[41];
 330
 331    _mesa_sha1_format(buf, key);
 332
 333    return ralloc_asprintf(cache, "%s/%c%c/%s",
 334                           cache->path, buf[0], buf[1], buf + 2);
 335 }
 336
 337 /* Create the directory that will be needed for the cache file for \key.
 338  *
 339  * Obviously, the implementation here must closely match
 340  * _get_cache_file above.
 341 */
 342 static void
 343 make_cache_file_directory(struct disk_cache *cache, cache_key key)
 344 {
 345    char *dir;
 346    char buf[41];
 347
 348    _mesa_sha1_format(buf, key);
 349
 350    dir = ralloc_asprintf(cache, "%s/%c%c", cache->path, buf[0], buf[1]);
 351
 352    mkdir_if_needed(dir);
 353
 354    ralloc_free(dir);
 355 }
 356
 357 /* Given a directory path and predicate function, count all entries in
 358  * that directory for which the predicate returns true. Then choose a
 359  * random entry from among those counted.
 360  *
 361  * Returns: A malloc'ed string for the path to the chosen file, (or
 362  * NULL on any error). The caller should free the string when
 363  * finished.
 364  */
 365 static char *
 366 choose_random_file_matching(const char *dir_path,
 367                             bool (*predicate)(struct dirent *))
 368 {
 369    DIR *dir;
 370    struct dirent *entry;
 371    unsigned int count, victim;
 372    char *filename;
 373
 374    dir = opendir(dir_path);
 375    if (dir == NULL)
 376       return NULL;
 377
 378    count = 0;
 379
 380    while (1) {
 381       entry = readdir(dir);
 382       if (entry == NULL)
 383          break;
 384       if (! predicate(entry))
 385          continue;
 386
 387       count++;
 388    }
 389
 390    if (count == 0) {
 391       closedir(dir);
 392       return NULL;
 393    }
 394
 395    victim = rand() % count;
 396
 397    rewinddir(dir);
 398    count = 0;
 399
 400    while (1) {
 401       entry = readdir(dir);
 402       if (entry == NULL)
 403          break;
 404       if (! predicate(entry))
 405          continue;
 406       if (count == victim)
 407          break;
 408
 409       count++;
 410    }
 411
 412    if (entry == NULL) {
 413       closedir(dir);
 414       return NULL;
 415    }
 416
 417    if (asprintf(&filename, "%s/%s", dir_path, entry->d_name) < 0)
 418       filename = NULL;
 419
 420    closedir(dir);
 421
 422    return filename;
 423 }
 424
 425 /* Is entry a regular file, and not having a name with a trailing
 426  * ".tmp"
 427  */
 428 static bool
 429 is_regular_non_tmp_file(struct dirent *entry)
 430 {
 431    size_t len;
 432
 433    if (entry->d_type != DT_REG)
 434       return false;
 435
 436    len = strlen (entry->d_name);
 437    if (len >= 4 && strcmp(&entry->d_name[len-4], ".tmp") == 0)
 438       return false;
 439
 440    return true;
 441 }
 442
 443 /* Returns the size of the deleted file, (or 0 on any error). */
 444 static size_t
 445 unlink_random_file_from_directory(const char *path)
 446 {
 447    struct stat sb;
 448    char *filename;
 449
 450    filename = choose_random_file_matching(path, is_regular_non_tmp_file);
 451    if (filename == NULL)
 452       return 0;
 453
 454    if (stat(filename, &sb) == -1) {
 455       free (filename);
 456       return 0;
 457    }
 458
 459    unlink(filename);
 460
 461    free (filename);
 462
 463    return sb.st_size;
 464 }
 465
 466 /* Is entry a directory with a two-character name, (and not the
 467  * special name of "..")
 468  */
 469 static bool
 470 is_two_character_sub_directory(struct dirent *entry)
 471 {
 472    if (entry->d_type != DT_DIR)
 473       return false;
 474
 475    if (strlen(entry->d_name) != 2)
 476       return false;
 477
 478    if (strcmp(entry->d_name, "..") == 0)
 479       return false;
 480
 481    return true;
 482 }
 483
 484 static void
 485 evict_random_item(struct disk_cache *cache)
 486 {
 487    const char hex[] = "0123456789abcde";
 488    char *dir_path;
 489    int a, b;
 490    size_t size;
 491
 492    /* With a reasonably-sized, full cache, (and with keys generated
 493     * from a cryptographic hash), we can choose two random hex digits
 494     * and reasonably expect the directory to exist with a file in it.
 495     */
 496    a = rand() % 16;
 497    b = rand() % 16;
 498
 499    if (asprintf(&dir_path, "%s/%c%c", cache->path, hex[a], hex[b]) < 0)
 500       return;
 501
 502    size = unlink_random_file_from_directory(dir_path);
 503
 504    free(dir_path);
 505
 506    if (size) {
 507       p_atomic_add(cache->size, - size);
 508       return;
 509    }
 510
 511    /* In the case where the random choice of directory didn't find
 512     * something, we choose randomly from the existing directories.
 513     *
 514     * Really, the only reason this code exists is to allow the unit
 515     * tests to work, (which use an artificially-small cache to be able
 516     * to force a single cached item to be evicted).
 517     */
 518    dir_path = choose_random_file_matching(cache->path,
 519                                           is_two_character_sub_directory);
 520    if (dir_path == NULL)
 521       return;
 522
 523    size = unlink_random_file_from_directory(dir_path);
 524
 525    free(dir_path);
 526
 527    if (size)
 528       p_atomic_add(cache->size, - size);
 529 }
 530
 531 void
 532 disk_cache_put(struct disk_cache *cache,
 533           cache_key key,
 534           const void *data,
 535           size_t size)
 536 {
 537    int fd = -1, fd_final = -1, err, ret;
 538    size_t len;
 539    char *filename = NULL, *filename_tmp = NULL;
 540    const char *p = data;
 541
 542    filename = get_cache_file(cache, key);
 543    if (filename == NULL)
 544       goto done;
 545
 546    /* Write to a temporary file to allow for an atomic rename to the
 547     * final destination filename, (to prevent any readers from seeing
 548     * a partially written file).
 549     */
 550    filename_tmp = ralloc_asprintf(cache, "%s.tmp", filename);
 551    if (filename_tmp == NULL)
 552       goto done;
 553
 554    fd = open(filename_tmp, O_WRONLY | O_CLOEXEC | O_CREAT, 0644);
 555
 556    /* Make the two-character subdirectory within the cache as needed. */
 557    if (fd == -1) {
 558       if (errno != ENOENT)
 559          goto done;
 560
 561       make_cache_file_directory(cache, key);
 562
 563       fd = open(filename_tmp, O_WRONLY | O_CLOEXEC | O_CREAT, 0644);
 564       if (fd == -1)
 565          goto done;
 566    }
 567
 568    /* With the temporary file open, we take an exclusive flock on
 569     * it. If the flock fails, then another process still has the file
 570     * open with the flock held. So just let that file be responsible
 571     * for writing the file.
 572     */
 573    err = flock(fd, LOCK_EX | LOCK_NB);
 574    if (err == -1)
 575       goto done;
 576
 577    /* Now that we have the lock on the open temporary file, we can
 578     * check to see if the destination file already exists. If so,
 579     * another process won the race between when we saw that the file
 580     * didn't exist and now. In this case, we don't do anything more,
 581     * (to ensure the size accounting of the cache doesn't get off).
 582     */
 583    fd_final = open(filename, O_RDONLY | O_CLOEXEC);
 584    if (fd_final != -1)
 585       goto done;
 586
 587    /* OK, we're now on the hook to write out a file that we know is
 588     * not in the cache, and is also not being written out to the cache
 589     * by some other process.
 590     *
 591     * Before we do that, if the cache is too large, evict something
 592     * else first.
 593     */
 594    if (*cache->size + size > cache->max_size)
 595       evict_random_item(cache);
 596
 597    /* Now, finally, write out the contents to the temporary file, then
 598     * rename them atomically to the destination filename, and also
 599     * perform an atomic increment of the total cache size.
 600     */
 601    for (len = 0; len < size; len += ret) {
 602       ret = write(fd, p + len, size - len);
 603       if (ret == -1) {
 604          unlink(filename_tmp);
 605          goto done;
 606       }
 607    }
 608
 609    rename(filename_tmp, filename);
 610
 611    p_atomic_add(cache->size, size);
 612
 613  done:
 614    if (fd_final != -1)
 615       close(fd_final);
 616    /* This close finally releases the flock, (now that the final dile
 617     * has been renamed into place and the size has been added).
 618     */
 619    if (fd != -1)
 620       close(fd);
 621    if (filename_tmp)
 622       ralloc_free(filename_tmp);
 623    if (filename)
 624       ralloc_free(filename);
 625 }
 626
 627 void *
 628 disk_cache_get(struct disk_cache *cache, cache_key key, size_t *size)
 629 {
 630    int fd = -1, ret, len;
 631    struct stat sb;
 632    char *filename = NULL;
 633    uint8_t *data = NULL;
 634
 635    if (size)
 636       *size = 0;
 637
 638    filename = get_cache_file(cache, key);
 639    if (filename == NULL)
 640       goto fail;
 641
 642    fd = open(filename, O_RDONLY | O_CLOEXEC);
 643    if (fd == -1)
 644       goto fail;
 645
 646    if (fstat(fd, &sb) == -1)
 647       goto fail;
 648
 649    data = malloc(sb.st_size);
 650    if (data == NULL)
 651       goto fail;
 652
 653    for (len = 0; len < sb.st_size; len += ret) {
 654       ret = read(fd, data + len, sb.st_size - len);
 655       if (ret == -1)
 656          goto fail;
 657    }
 658
 659    ralloc_free(filename);
 660    close(fd);
 661
 662    if (size)
 663       *size = sb.st_size;
 664
 665    return data;
 666
 667  fail:
 668    if (data)
 669       free(data);
 670    if (filename)
 671       ralloc_free(filename);
 672    if (fd != -1)
 673       close(fd);
 674
 675    return NULL;
 676 }
 677
 678 void
 679 disk_cache_put_key(struct disk_cache *cache, cache_key key)
 680 {
 681    uint32_t *key_chunk = (uint32_t *) key;
 682    int i = *key_chunk & CACHE_INDEX_KEY_MASK;
 683    unsigned char *entry;
 684
 685    entry = &cache->stored_keys[i + CACHE_KEY_SIZE];
 686
 687    memcpy(entry, key, CACHE_KEY_SIZE);
 688 }
 689
 690 /* This function lets us test whether a given key was previously
 691  * stored in the cache with disk_cache_put_key(). The implement is
 692  * efficient by not using syscalls or hitting the disk. It's not
 693  * race-free, but the races are benign. If we race with someone else
 694  * calling disk_cache_put_key, then that's just an extra cache miss and an
 695  * extra recompile.
 696  */
 697 bool
 698 disk_cache_has_key(struct disk_cache *cache, cache_key key)
 699 {
 700    uint32_t *key_chunk = (uint32_t *) key;
 701    int i = *key_chunk & CACHE_INDEX_KEY_MASK;
 702    unsigned char *entry;
 703
 704    entry = &cache->stored_keys[i + CACHE_KEY_SIZE];
 705
 706    return memcmp(entry, key, CACHE_KEY_SIZE) == 0;
 707 }