binlog.c

   1 /* binlog.c - binary log implementation */
   2
   3 /* Copyright (C) 2008 Graham Barr
   4
   5  * This program is free software: you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License as published by
   7  * the Free Software Foundation, either version 3 of the License, or
   8  * (at your option) any later version.
   9
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
  17  */
  18
  19 #include <stdlib.h>
  20 #include <stdio.h>
  21 #include <fcntl.h>
  22 #include <unistd.h>
  23 #include <string.h>
  24 #include <errno.h>
  25 #include <dirent.h>
  26 #include <sys/resource.h>
  27 #include <sys/param.h>
  28 #include <sys/uio.h>
  29 #include <sys/stat.h>
  30 #include <stdarg.h>
  31 #include <limits.h>
  32
  33 #include "tube.h"
  34 #include "job.h"
  35 #include "binlog.h"
  36 #include "util.h"
  37 #include "config.h"
  38
  39 typedef struct binlog *binlog;
  40
  41 struct binlog {
  42   binlog next;
  43   unsigned int refs;
  44   int fd;
  45   size_t free;
  46   size_t reserved;
  47   char path[];
  48 };
  49
  50 /* max size we will create a log file */
  51 size_t binlog_size_limit = 10 << 20;
  52
  53 char *binlog_dir = NULL;
  54 static int binlog_index = 0;
  55 static int binlog_version = 2;
  56 static int lock_fd;
  57
  58 static binlog oldest_binlog = 0,
  59               current_binlog = 0,
  60               newest_binlog = 0;
  61
  62 static int
  63 binlog_scan_dir()
  64 {
  65     DIR *dirp;
  66     struct dirent *dp;
  67     long min = 0;
  68     long max = 0;
  69     long val;
  70     char *endptr;
  71     size_t name_len;
  72
  73     dirp = opendir(binlog_dir);
  74     if (!dirp) return 0;
  75
  76     while ((dp = readdir(dirp)) != NULL) {
  77         name_len = strlen(dp->d_name);
  78         if (name_len > 7 && !strncmp("binlog.", dp->d_name, 7)) {
  79             val = strtol(dp->d_name + 7, &endptr, 10);
  80             if (endptr && *endptr == 0) {
  81                 if (max == 0 || val > max) max = val;
  82                 if (min == 0 || val < min) min = val;
  83             }
  84         }
  85     }
  86
  87     closedir(dirp);
  88     binlog_index = (int) max;
  89     return (int) min;
  90 }
  91
  92 static void
  93 binlog_remove_oldest()
  94 {
  95     binlog b = oldest_binlog;
  96
  97     if (!b) return;
  98
  99     oldest_binlog = b->next;
 100
 101     unlink(b->path);
 102     free(b);
 103 }
 104
 105 static binlog
 106 binlog_iref(binlog b)
 107 {
 108     if (b) b->refs++;
 109     return b;
 110 }
 111
 112 static void
 113 binlog_dref(binlog b)
 114 {
 115     if (!b) return;
 116     if (b->refs < 1) return twarnx("refs is zero for binlog: %s", b->path);
 117
 118     --b->refs;
 119     if (b->refs < 1) {
 120         while (oldest_binlog && oldest_binlog->refs == 0) {
 121             binlog_remove_oldest();
 122         }
 123     }
 124 }
 125
 126 /*
 127 static void
 128 binlog_warn(int fd, const char* path, const char *msg)
 129 {
 130     warnx("WARNING, %s at %s:%u.\n%s", msg, path, lseek(fd, 0, SEEK_CUR),
 131           "  Continuing. You may be missing data.");
 132 }
 133 */
 134
 135 #define binlog_warn(b, fmt, args...) \
 136     warnx("WARNING, " fmt " at %s:%u. %s: ", \
 137           ##args, b->path, lseek(b->fd, 0, SEEK_CUR), \
 138           "Continuing. You may be missing data.")
 139
 140 static void
 141 binlog_read_log_file(binlog b, job binlog_jobs)
 142 {
 143     struct job js;
 144     tube t;
 145     job j;
 146     char tubename[MAX_TUBE_NAME_LEN];
 147     size_t namelen;
 148     ssize_t r;
 149     int version;
 150
 151     r = read(b->fd, &version, sizeof(version));
 152     if (r == -1) return twarn("read()");
 153     if (r < sizeof(version)) {
 154         return binlog_warn(b, "EOF while reading version record");
 155     }
 156
 157     if (version != binlog_version) {
 158         return warnx("%s: binlog version mismatch %d %d", b->path, version,
 159                      binlog_version);
 160     }
 161
 162     while (read(b->fd, &namelen, sizeof(size_t)) == sizeof(size_t)) {
 163         if (namelen > 0) {
 164             r = read(b->fd, tubename, namelen);
 165             if (r == -1) return twarn("read()");
 166             if (r < namelen) {
 167                 lseek(b->fd, SEEK_CUR, 0);
 168                 return binlog_warn(b, "EOF while reading tube name");
 169             }
 170         }
 171
 172         tubename[namelen] = '\0';
 173         r = read(b->fd, &js, sizeof(struct job));
 174         if (r == -1) return twarn("read()");
 175         if (r < sizeof(struct job)) {
 176           return binlog_warn(b, "EOF while reading job record");
 177         }
 178
 179         if (!js.id) break;
 180
 181         j = job_find(js.id);
 182         switch (js.state) {
 183         case JOB_STATE_INVALID:
 184             if (j) {
 185                 job_remove(j);
 186                 binlog_dref(j->binlog);
 187                 job_free(j);
 188                 j = NULL;
 189             }
 190             break;
 191         case JOB_STATE_READY:
 192         case JOB_STATE_DELAYED:
 193             if (!j) {
 194                 t = tube_find_or_make(tubename);
 195                 j = make_job_with_id(js.pri, js.delay, js.ttr, js.body_size,
 196                                      t, js.id);
 197                 j->next = j->prev = j;
 198                 j->creation = js.creation;
 199                 job_insert(binlog_jobs, j);
 200             }
 201             if (js.body_size) {
 202                 if (js.body_size > j->body_size) {
 203                     warnx("job size increased from %zu to %zu", j->body_size,
 204                           js.body_size);
 205                     job_remove(j);
 206                     binlog_dref(j->binlog);
 207                     job_free(j);
 208                     return binlog_warn(b, "EOF while reading job body");
 209                 }
 210                 r = read(b->fd, j->body, js.body_size);
 211                 if (r == -1) return twarn("read()");
 212                 if (r < js.body_size) {
 213                     warnx("dropping incomplete job %llu", j->id);
 214                     job_remove(j);
 215                     binlog_dref(j->binlog);
 216                     job_free(j);
 217                     return binlog_warn(b, "EOF while reading job body");
 218                 }
 219             }
 220             break;
 221         }
 222         if (j) {
 223             j->state = js.state;
 224             j->deadline = js.deadline;
 225             j->pri = js.pri;
 226             j->delay = js.delay;
 227             j->ttr = js.ttr;
 228             j->timeout_ct = js.timeout_ct;
 229             j->release_ct = js.release_ct;
 230             j->bury_ct = js.bury_ct;
 231             j->kick_ct = js.kick_ct;
 232
 233             /* this is a complete record, so we can move the binlog ref */
 234             if (namelen && js.body_size) {
 235                 binlog_dref(j->binlog);
 236                 j->binlog = binlog_iref(b);
 237             }
 238         }
 239     }
 240 }
 241
 242 static void
 243 binlog_close(binlog b)
 244 {
 245     if (!b) return;
 246     if (b->fd < 0) return;
 247     close(b->fd);
 248     binlog_dref(b);
 249     b->fd = -1;
 250 }
 251
 252 static binlog
 253 make_binlog(char *path)
 254 {
 255     binlog b;
 256
 257     b = (binlog) malloc(sizeof(struct binlog) + strlen(path) + 1);
 258     if (!b) return twarnx("OOM"), (binlog) 0;
 259     strcpy(b->path, path);
 260     b->refs = 0;
 261     b->next = NULL;
 262     b->fd = -1;
 263     b->free = 0;
 264     b->reserved = 0;
 265     return b;
 266 }
 267
 268 static binlog
 269 make_next_binlog()
 270 {
 271     int r;
 272     char path[PATH_MAX];
 273
 274     if (!binlog_dir) return NULL;
 275
 276     r = snprintf(path, PATH_MAX, "%s/binlog.%d", binlog_dir, ++binlog_index);
 277     if (r > PATH_MAX) return twarnx("path too long: %s", binlog_dir), (binlog)0;
 278
 279     return make_binlog(path);
 280 }
 281
 282 static binlog
 283 add_binlog(binlog b)
 284 {
 285     if (newest_binlog) newest_binlog->next = b;
 286     newest_binlog = b;
 287     if (!oldest_binlog) oldest_binlog = b;
 288
 289     return b;
 290 }
 291
 292 static void
 293 binlog_open(binlog log, size_t *written)
 294 {
 295     int fd;
 296     size_t bytes_written;
 297
 298     if (!binlog_iref(log)) return;
 299
 300     fd = open(log->path, O_WRONLY | O_CREAT, 0400);
 301
 302     if (fd < 0) return twarn("Cannot open binlog %s", log->path);
 303
 304 #ifdef HAVE_POSIX_FALLOCATE
 305     {
 306         int r;
 307         r = posix_fallocate(fd, 0, binlog_size_limit);
 308         if (r) {
 309             close(fd);
 310             binlog_dref(log);
 311             errno = r;
 312             return twarn("Cannot allocate space for binlog %s", log->path);
 313         }
 314     }
 315 #else
 316     /* Allocate space in a slow but portable way. */
 317     {
 318         size_t i;
 319         ssize_t w;
 320         off_t p;
 321         #define ZERO_BUF_SIZE 512
 322         char buf[ZERO_BUF_SIZE] = {}; /* initialize to zero */
 323
 324         for (i = 0; i < binlog_size_limit; i += w) {
 325             w = write(fd, &buf, ZERO_BUF_SIZE);
 326             if (w == -1) {
 327                 twarn("Cannot allocate space for binlog %s", log->path);
 328                 close(fd);
 329                 binlog_dref(log);
 330                 return;
 331             }
 332         }
 333
 334         p = lseek(fd, 0, SEEK_SET);
 335         if (p == -1) {
 336             twarn("lseek");
 337             close(fd);
 338             binlog_dref(log);
 339             return;
 340         }
 341     }
 342 #endif
 343
 344     bytes_written = write(fd, &binlog_version, sizeof(int));
 345
 346     if (bytes_written < sizeof(int)) {
 347         twarn("Cannot write to binlog");
 348         close(fd);
 349         binlog_dref(log);
 350         return;
 351     }
 352
 353     if (written) *written = bytes_written;
 354     log->fd = fd;
 355 }
 356
 357 /* returns 1 on success, 0 on error. */
 358 static int
 359 binlog_use_next()
 360 {
 361     binlog next;
 362
 363     if (!current_binlog) return 0;
 364
 365     next = current_binlog->next;
 366
 367     if (!next) return 0;
 368
 369     /* assert(current_binlog->reserved == 0); */
 370
 371     binlog_close(current_binlog);
 372     current_binlog = next;
 373
 374     return 1;
 375 }
 376
 377 void
 378 binlog_shutdown()
 379 {
 380     binlog_use_next();
 381     binlog_close(current_binlog);
 382 }
 383
 384 /* Returns the number of jobs successfully written (either 0 or 1). */
 385 /* If we are not using the binlog at all (!current_binlog), then we pretend to
 386    have made a successful write and return 1. */
 387 int
 388 binlog_write_job(job j)
 389 {
 390     ssize_t written;
 391     size_t tube_namelen, to_write = 0;
 392     struct iovec vec[4], *vptr;
 393     int vcnt = 3, r;
 394
 395     if (!current_binlog) return 1;
 396     tube_namelen = 0;
 397
 398     vec[0].iov_base = (char *) &tube_namelen;
 399     to_write += vec[0].iov_len = sizeof(size_t);
 400
 401     vec[1].iov_base = j->tube->name;
 402     vec[1].iov_len = 0;
 403
 404     /* we could save some bytes in the binlog file by only saving some parts of
 405      * the job struct */
 406     vec[2].iov_base = (char *) j;
 407     to_write += vec[2].iov_len = sizeof(struct job);
 408
 409     printf("writing job %lld state %d\n", j->id, j->state);
 410
 411     if (j->state == JOB_STATE_READY || j->state == JOB_STATE_DELAYED) {
 412         if (!j->binlog) {
 413             tube_namelen = strlen(j->tube->name);
 414             to_write += vec[1].iov_len = tube_namelen;
 415             vcnt = 4;
 416             vec[3].iov_base = j->body;
 417             to_write += vec[3].iov_len = j->body_size;
 418         }
 419     } else if (j->state == JOB_STATE_INVALID) {
 420         if (j->binlog) binlog_dref(j->binlog);
 421         j->binlog = NULL;
 422     } else {
 423         return twarnx("unserializable job state: %d", j->state), 0;
 424     }
 425
 426     if (to_write > current_binlog->reserved) {
 427         r = binlog_use_next();
 428         if (!r) return twarnx("failed to use next binlog"), 0;
 429     }
 430
 431     if (j->state && !j->binlog) j->binlog = binlog_iref(current_binlog);
 432
 433     while (to_write > 0) {
 434         written = writev(current_binlog->fd, vec, vcnt);
 435
 436         if (written < 0) {
 437             if (errno == EAGAIN) continue;
 438             if (errno == EINTR) continue;
 439
 440             twarn("writev");
 441             binlog_close(current_binlog);
 442             current_binlog = 0;
 443             return 0;
 444         }
 445
 446         to_write -= written;
 447         if (to_write > 0 && written > 0) {
 448             for (vptr = vec; written >= vptr->iov_len; vptr++) {
 449                 written -= vptr->iov_len;
 450                 vptr->iov_len = 0;
 451             }
 452             vptr->iov_base = (char *) vptr->iov_base + written;
 453             vptr->iov_len -= written;
 454         }
 455     }
 456
 457     return 1;
 458 }
 459
 460 static binlog
 461 make_future_binlog()
 462 {
 463     binlog b;
 464     size_t header;
 465
 466     /* open a new binlog with more space to reserve */
 467     b = make_next_binlog();
 468     if (!b) return twarnx("error making next binlog"), (binlog) 0;
 469     binlog_open(b, &header);
 470
 471     /* open failed, so we can't reserve any space */
 472     if (b->fd < 0) {
 473         free(b);
 474         return 0;
 475     }
 476
 477     b->free = binlog_size_limit - header;
 478     b->reserved = 0;
 479     return b;
 480 }
 481
 482 static int
 483 can_move_reserved(size_t n, binlog from, binlog to)
 484 {
 485     return from->reserved >= n && to->free >= n;
 486 }
 487
 488 static void
 489 move_reserved(size_t n, binlog from, binlog to)
 490 {
 491     from->reserved -= n;
 492     from->free += n;
 493     to->reserved += n;
 494     to->free -= n;
 495 }
 496
 497 static size_t
 498 ensure_free_space(size_t n)
 499 {
 500     binlog fb;
 501
 502     if (newest_binlog->free >= n) return n;
 503
 504     /* open a new binlog */
 505     fb = make_future_binlog();
 506     if (!fb) return twarnx("make_future_binlog"), 0;
 507
 508     add_binlog(fb);
 509     return n;
 510 }
 511
 512 /* Preserve some invariants immediately after any space reservation.
 513  * Invariant 1: current_binlog->reserved >= n.
 514  * Invariant 2: current_binlog->reserved is congruent to n (mod z), where z
 515  * is the size of a delete record in the binlog. */
 516 static size_t
 517 maintain_invariant(size_t n)
 518 {
 519     size_t reserved_later, remainder, complement, z, r;
 520
 521     /* In this function, reserved bytes are conserved (they are neither created
 522      * nor destroyed). We just move them around to preserve the invariant. We
 523      * might have to create new free space (i.e. allocate a new binlog file),
 524      * though. */
 525
 526     /* Invariant 1. */
 527     /* This is a loop, but it's guaranteed to run at most once. The proof is
 528      * left as an exercise for the reader. */
 529     while (current_binlog->reserved < n) {
 530         size_t to_move = current_binlog->reserved;
 531
 532         r = ensure_free_space(to_move);
 533         if (r != to_move) {
 534             /* TODO unreserve n bytes */
 535             return twarnx("ensure_free_space"), 0;
 536         }
 537
 538         move_reserved(to_move, current_binlog, newest_binlog);
 539         binlog_use_next();
 540     }
 541
 542     /* Invariant 2. */
 543
 544
 545     z = sizeof(size_t) + sizeof(struct job);
 546     reserved_later = current_binlog->reserved - n;
 547     remainder = reserved_later % z;
 548     if (remainder == 0) return n;
 549     complement = z - remainder;
 550     if (can_move_reserved(complement, newest_binlog, current_binlog)) {
 551         move_reserved(complement, newest_binlog, current_binlog);
 552         return n;
 553     }
 554
 555     r = ensure_free_space(remainder);
 556     if (r != remainder) {
 557         /* TODO unreserve n bytes */
 558         return twarnx("ensure_free_space"), 0;
 559     }
 560     move_reserved(remainder, current_binlog, newest_binlog);
 561
 562     return n;
 563 }
 564
 565 /* Returns the number of bytes successfully reserved: either 0 or n. */
 566 static size_t
 567 binlog_reserve_space(size_t n)
 568 {
 569     /* This return value must be nonzero but is otherwise ignored. */
 570     if (!current_binlog) return 1;
 571
 572     if (current_binlog->free >= n) {
 573         current_binlog->free -= n;
 574         current_binlog->reserved += n;
 575         return maintain_invariant(n);
 576     }
 577
 578     ensure_free_space(n);
 579
 580     newest_binlog->free -= n;
 581     newest_binlog->reserved += n;
 582     return maintain_invariant(n);
 583 }
 584
 585 /* Returns the number of bytes reserved. */
 586 size_t
 587 binlog_reserve_space_put(job j)
 588 {
 589     size_t z = 0;
 590
 591     /* reserve space for the initial job record */
 592     z += sizeof(size_t);
 593     z += strlen(j->tube->name);
 594     z += sizeof(struct job);
 595     z += j->body_size;
 596
 597     /* plus space for a delete to come later */
 598     z += sizeof(size_t);
 599     z += sizeof(struct job);
 600
 601     return binlog_reserve_space(z);
 602 }
 603
 604 size_t
 605 binlog_reserve_space_update(job j)
 606 {
 607     size_t z = 0;
 608
 609     z += sizeof(size_t);
 610     z += sizeof(struct job);
 611     return binlog_reserve_space(z);
 612 }
 613
 614 int
 615 binlog_lock()
 616 {
 617     int r;
 618     struct flock lock;
 619     char path[PATH_MAX];
 620
 621     r = snprintf(path, PATH_MAX, "%s/lock", binlog_dir);
 622     if (r > PATH_MAX) return twarnx("path too long: %s", binlog_dir), 0;
 623
 624     lock_fd = open(path, O_WRONLY|O_CREAT, 0600);
 625     if (lock_fd == -1) return twarn("open"), 0;
 626
 627     lock.l_type = F_WRLCK;
 628     lock.l_whence = SEEK_SET;
 629     lock.l_start = 0;
 630     lock.l_len = 0;
 631     r = fcntl(lock_fd, F_SETLK, &lock);
 632     if (r) return twarn("fcntl"), 0;
 633
 634     return 1;
 635 }
 636
 637 void
 638 binlog_init(job binlog_jobs)
 639 {
 640     int binlog_index_min;
 641     struct stat sbuf;
 642     int fd, idx, r;
 643     char path[PATH_MAX];
 644     binlog b;
 645
 646     if (!binlog_dir) return;
 647
 648     /* Recover any jobs in old binlogs */
 649
 650     if (stat(binlog_dir, &sbuf) < 0) {
 651         if (mkdir(binlog_dir, 0700) < 0) return twarn("%s", binlog_dir);
 652     } else if (!(sbuf.st_mode & S_IFDIR)) {
 653         twarnx("%s", binlog_dir);
 654         return;
 655     }
 656
 657     binlog_index_min = binlog_scan_dir();
 658
 659     if (binlog_index_min) {
 660         for (idx = binlog_index_min; idx <= binlog_index; idx++) {
 661             r = snprintf(path, PATH_MAX, "%s/binlog.%d", binlog_dir, idx);
 662             if (r > PATH_MAX) return twarnx("path too long: %s", binlog_dir);
 663
 664             fd = open(path, O_RDONLY);
 665
 666             if (fd < 0) {
 667                 twarn("%s", path);
 668             } else {
 669                 b = binlog_iref(add_binlog(make_binlog(path)));
 670                 b->fd = fd;
 671                 binlog_read_log_file(b, binlog_jobs);
 672                 close(fd);
 673                 b->fd = -1;
 674                 binlog_dref(b);
 675             }
 676         }
 677
 678     }
 679
 680
 681     /* Set up for writing out new jobs */
 682
 683     b = make_next_binlog();
 684     if (!b) return twarnx("error making first binlog");
 685     binlog_open(b, 0);
 686     if (b->fd >= 0) add_binlog(b);
 687     current_binlog = b;
 688 }
 689
 690 const char *
 691 binlog_oldest_index()
 692 {
 693     if (!oldest_binlog) return "0";
 694
 695     return strrchr(oldest_binlog->path, '.') + 1;
 696 }
 697
 698 const char *
 699 binlog_current_index()
 700 {
 701     if (!newest_binlog) return "0";
 702
 703     return strrchr(newest_binlog->path, '.') + 1;
 704 }