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 /* max size we will create a log file */
  40 size_t binlog_size_limit = 10 << 20;
  41
  42 char *binlog_dir = NULL;
  43 static int binlog_index = 0;
  44 static int binlog_fd = -1, next_binlog_fd = -1;
  45 static int binlog_version = 2;
  46 static size_t binlog_space = 0, next_binlog_space = 0;
  47 static size_t binlog_reserved = 0, next_binlog_reserved = 0;
  48 static size_t bytes_written;
  49 static int lock_fd;
  50
  51 static binlog first_binlog = NULL, last_binlog = NULL, next_binlog = NULL;
  52
  53 static int
  54 binlog_scan_dir()
  55 {
  56     DIR *dirp;
  57     struct dirent *dp;
  58     long min = 0;
  59     long max = 0;
  60     long val;
  61     char *endptr;
  62     size_t name_len;
  63
  64     dirp = opendir(binlog_dir);
  65     if (!dirp) return 0;
  66
  67     while ((dp = readdir(dirp)) != NULL) {
  68         name_len = strlen(dp->d_name);
  69         if (name_len > 7 && !strncmp("binlog.", dp->d_name, 7)) {
  70             val = strtol(dp->d_name + 7, &endptr, 10);
  71             if (endptr && *endptr == 0) {
  72                 if (max == 0 || val > max) max = val;
  73                 if (min == 0 || val < min) min = val;
  74             }
  75         }
  76     }
  77
  78     closedir(dirp);
  79     binlog_index = (int) max;
  80     return (int) min;
  81 }
  82
  83 static void
  84 binlog_remove_first()
  85 {
  86     binlog b = first_binlog;
  87
  88     if (!b) return;
  89
  90     first_binlog = b->next;
  91     if (!first_binlog) last_binlog = NULL;
  92
  93     unlink(b->path);
  94     free(b);
  95 }
  96
  97 static binlog
  98 binlog_iref(binlog b)
  99 {
 100     if (b) b->refs++;
 101     return b;
 102 }
 103
 104 static void
 105 binlog_dref(binlog b)
 106 {
 107     if (!b) return;
 108     if (b->refs < 1) return twarnx("refs is zero for binlog: %s", b->path);
 109
 110     --b->refs;
 111     if (b->refs < 1) {
 112         while (first_binlog && first_binlog->refs == 0) binlog_remove_first();
 113     }
 114 }
 115
 116 /*
 117 static void
 118 binlog_warn(int fd, const char* path, const char *msg)
 119 {
 120     warnx("WARNING, %s at %s:%u.\n%s", msg, path, lseek(fd, 0, SEEK_CUR),
 121           "  Continuing. You may be missing data.");
 122 }
 123 */
 124
 125 #define binlog_warn(fd, path, fmt, args...) \
 126     warnx("WARNING, " fmt " at %s:%u. %s: ", \
 127           ##args, path, lseek(fd, 0, SEEK_CUR), \
 128           "Continuing. You may be missing data.")
 129
 130 static void
 131 binlog_read_one(int fd, job binlog_jobs, const char *path)
 132 {
 133     struct job js;
 134     tube t;
 135     job j;
 136     char tubename[MAX_TUBE_NAME_LEN];
 137     size_t namelen;
 138     ssize_t r;
 139     int version;
 140
 141     r = read(fd, &version, sizeof(version));
 142     if (r == -1) return twarn("read()");
 143     if (r < sizeof(version)) {
 144         return binlog_warn(fd, path, "EOF while reading version record");
 145     }
 146
 147     if (version != binlog_version) {
 148         return warnx("%s: binlog version mismatch %d %d", path, version,
 149                      binlog_version);
 150     }
 151
 152     while (read(fd, &namelen, sizeof(size_t)) == sizeof(size_t)) {
 153         if (namelen > 0) {
 154             r = read(fd, tubename, namelen);
 155             if (r == -1) return twarn("read()");
 156             if (r < namelen) {
 157                 lseek(fd, SEEK_CUR, 0);
 158                 return binlog_warn(fd, path, "EOF while reading tube name");
 159             }
 160         }
 161
 162         tubename[namelen] = '\0';
 163         r = read(fd, &js, sizeof(struct job));
 164         if (r == -1) return twarn("read()");
 165         if (r < sizeof(struct job)) {
 166           return binlog_warn(fd, path, "EOF while reading job record");
 167         }
 168
 169         if (!js.id) break;
 170
 171         j = job_find(js.id);
 172         switch (js.state) {
 173         case JOB_STATE_INVALID:
 174             if (j) {
 175                 job_remove(j);
 176                 binlog_dref(j->binlog);
 177                 job_free(j);
 178                 j = NULL;
 179             }
 180             break;
 181         case JOB_STATE_READY:
 182         case JOB_STATE_DELAYED:
 183             if (!j) {
 184                 t = tube_find_or_make(tubename);
 185                 j = make_job_with_id(js.pri, js.delay, js.ttr, js.body_size,
 186                                      t, js.id);
 187                 j->next = j->prev = j;
 188                 j->creation = js.creation;
 189                 job_insert(binlog_jobs, j);
 190             }
 191             if (js.body_size) {
 192                 if (js.body_size > j->body_size) {
 193                     warnx("job size increased from %zu to %zu", j->body_size,
 194                           js.body_size);
 195                     job_remove(j);
 196                     binlog_dref(j->binlog);
 197                     job_free(j);
 198                     return binlog_warn(fd, path, "EOF while reading job body");
 199                 }
 200                 r = read(fd, j->body, js.body_size);
 201                 if (r == -1) return twarn("read()");
 202                 if (r < js.body_size) {
 203                     warnx("dropping incomplete job %llu", j->id);
 204                     job_remove(j);
 205                     binlog_dref(j->binlog);
 206                     job_free(j);
 207                     return binlog_warn(fd, path, "EOF while reading job body");
 208                 }
 209             }
 210             break;
 211         }
 212         if (j) {
 213             j->state = js.state;
 214             j->deadline = js.deadline;
 215             j->pri = js.pri;
 216             j->delay = js.delay;
 217             j->ttr = js.ttr;
 218             j->timeout_ct = js.timeout_ct;
 219             j->release_ct = js.release_ct;
 220             j->bury_ct = js.bury_ct;
 221             j->kick_ct = js.kick_ct;
 222
 223             /* this is a complete record, so we can move the binlog ref */
 224             if (namelen && js.body_size) {
 225                 binlog_dref(j->binlog);
 226                 j->binlog = binlog_iref(last_binlog);
 227             }
 228         }
 229     }
 230 }
 231
 232 static void
 233 binlog_close_last()
 234 {
 235     if (binlog_fd < 0) return;
 236     close(binlog_fd);
 237     binlog_dref(last_binlog);
 238     binlog_fd = -1;
 239 }
 240
 241 static binlog
 242 make_binlog(char *path)
 243 {
 244     binlog b;
 245
 246     b = (binlog) malloc(sizeof(struct binlog) + strlen(path) + 1);
 247     if (!b) return twarnx("OOM"), NULL;
 248     strcpy(b->path, path);
 249     b->refs = 0;
 250     b->next = NULL;
 251     return b;
 252 }
 253
 254 static binlog
 255 make_next_binlog()
 256 {
 257     int r;
 258     char path[PATH_MAX];
 259
 260     if (!binlog_dir) return NULL;
 261
 262     r = snprintf(path, PATH_MAX, "%s/binlog.%d", binlog_dir, ++binlog_index);
 263     if (r > PATH_MAX) return twarnx("path too long: %s", binlog_dir), NULL;
 264
 265     return make_binlog(path);
 266 }
 267
 268 static binlog
 269 add_binlog(binlog b)
 270 {
 271     if (last_binlog) last_binlog->next = b;
 272     last_binlog = b;
 273     if (!first_binlog) first_binlog = b;
 274
 275     return b;
 276 }
 277
 278 static int
 279 binlog_open(binlog log)
 280 {
 281     int fd, r;
 282
 283     if (!binlog_iref(log)) return -1;
 284
 285     fd = open(log->path, O_WRONLY | O_CREAT, 0400);
 286
 287     if (fd < 0) return twarn("Cannot open binlog %s", log->path), -1;
 288
 289 #ifdef HAVE_POSIX_FALLOCATE
 290     r = posix_fallocate(fd, 0, binlog_size_limit);
 291 #else
 292     #warning no known method to preallocate files on this platform
 293     r = 0;
 294 #endif
 295     if (r) {
 296         close(fd);
 297         binlog_dref(log);
 298         errno = r;
 299         return twarn("Cannot allocate space for binlog %s", log->path), -1;
 300     }
 301
 302     bytes_written = write(fd, &binlog_version, sizeof(int));
 303
 304     if (bytes_written < sizeof(int)) {
 305         twarn("Cannot write to binlog");
 306         close(fd);
 307         binlog_dref(log);
 308         return -1;
 309     }
 310
 311     return fd;
 312 }
 313
 314 /* returns 1 on success, 0 on error. */
 315 static int
 316 binlog_use_next()
 317 {
 318     if (binlog_fd < 0) return 0;
 319     if (next_binlog_fd < 0) return 0;
 320     if (binlog_reserved > next_binlog_space) return twarnx("overextended"), 0;
 321
 322     binlog_close_last();
 323
 324     binlog_fd = next_binlog_fd;
 325     add_binlog(next_binlog);
 326
 327     next_binlog = NULL;
 328     next_binlog_fd = -1;
 329
 330     binlog_space = next_binlog_space - binlog_reserved;
 331     binlog_reserved = next_binlog_reserved + binlog_reserved;
 332
 333     next_binlog_reserved = next_binlog_space = 0;
 334     return 1;
 335 }
 336
 337 void
 338 binlog_close()
 339 {
 340     binlog_use_next();
 341     binlog_close_last();
 342 }
 343
 344 /* Returns the number of jobs successfully written (either 0 or 1). */
 345 /* If we are not using the binlog at all (binlog_fd < 0), then we pretend to
 346    have made a successful write and return 1. */
 347 int
 348 binlog_write_job(job j)
 349 {
 350     ssize_t written;
 351     size_t tube_namelen, to_write = 0;
 352     struct iovec vec[4], *vptr;
 353     int vcnt = 3, r;
 354
 355     if (binlog_fd < 0) return 1;
 356     tube_namelen = 0;
 357
 358     vec[0].iov_base = (char *) &tube_namelen;
 359     to_write += vec[0].iov_len = sizeof(size_t);
 360
 361     vec[1].iov_base = j->tube->name;
 362     vec[1].iov_len = 0;
 363
 364     /* we could save some bytes in the binlog file by only saving some parts of
 365      * the job struct */
 366     vec[2].iov_base = (char *) j;
 367     to_write += vec[2].iov_len = sizeof(struct job);
 368
 369     printf("writing job %lld state %d\n", j->id, j->state);
 370
 371     if (j->state == JOB_STATE_READY || j->state == JOB_STATE_DELAYED) {
 372         if (!j->binlog) {
 373             tube_namelen = strlen(j->tube->name);
 374             to_write += vec[1].iov_len = tube_namelen;
 375             vcnt = 4;
 376             vec[3].iov_base = j->body;
 377             to_write += vec[3].iov_len = j->body_size;
 378         }
 379     } else if (j->state == JOB_STATE_INVALID) {
 380         if (j->binlog) binlog_dref(j->binlog);
 381         j->binlog = NULL;
 382     } else {
 383         return twarnx("unserializable job state: %d", j->state), 0;
 384     }
 385
 386     if (to_write > binlog_reserved) {
 387         r = binlog_use_next();
 388         if (!r) return twarnx("failed to use next binlog"), 0;
 389     }
 390
 391     if (j->state && !j->binlog) j->binlog = binlog_iref(last_binlog);
 392
 393     while (to_write > 0) {
 394         written = writev(binlog_fd, vec, vcnt);
 395
 396         if (written < 0) {
 397             if (errno == EAGAIN) continue;
 398             if (errno == EINTR) continue;
 399
 400             twarn("writev");
 401             binlog_close_last();
 402             return 0;
 403         }
 404
 405         bytes_written += written;
 406         to_write -= written;
 407         if (to_write > 0 && written > 0) {
 408             for (vptr = vec; written >= vptr->iov_len; vptr++) {
 409                 written -= vptr->iov_len;
 410                 vptr->iov_len = 0;
 411             }
 412             vptr->iov_base = (char *) vptr->iov_base + written;
 413             vptr->iov_len -= written;
 414         }
 415     }
 416
 417     return 1;
 418 }
 419
 420 /* Returns the number of bytes successfully reserved: either 0 or n. */
 421 static size_t
 422 binlog_reserve_space(size_t n)
 423 {
 424     /* This value must be nonzero but is otherwise ignored. */
 425     if (binlog_fd < 0) return 1;
 426
 427     if (n <= binlog_space) {
 428         binlog_space -= n;
 429         binlog_reserved += n;
 430         return n;
 431     }
 432
 433     if (n <= next_binlog_space) {
 434         next_binlog_space -= n;
 435         next_binlog_reserved += n;
 436         return n;
 437     }
 438
 439     /* The next binlog is already allocated and it is full. */
 440     if (next_binlog_fd >= 0) return 0;
 441
 442     /* open a new binlog with more space to reserve */
 443     next_binlog = make_next_binlog();
 444     if (!next_binlog) return twarnx("error making next binlog"), 0;
 445     next_binlog_fd = binlog_open(next_binlog);
 446
 447     /* open failed, so we can't reserve any space */
 448     if (next_binlog_fd < 0) return 0;
 449
 450     next_binlog_space = binlog_size_limit - bytes_written - n;
 451     next_binlog_reserved = n;
 452
 453     return n;
 454 }
 455
 456 /* Returns the number of bytes reserved. */
 457 size_t
 458 binlog_reserve_space_put(job j)
 459 {
 460     size_t z = 0;
 461
 462     /* reserve space for the initial job record */
 463     z += sizeof(size_t);
 464     z += strlen(j->tube->name);
 465     z += sizeof(struct job);
 466     z += j->body_size;
 467
 468     /* plus space for a delete to come later */
 469     z += sizeof(size_t);
 470     z += sizeof(struct job);
 471
 472     return binlog_reserve_space(z);
 473 }
 474
 475 size_t
 476 binlog_reserve_space_update(job j)
 477 {
 478     size_t z = 0;
 479
 480     z += sizeof(size_t);
 481     z += sizeof(struct job);
 482     return binlog_reserve_space(z);
 483 }
 484
 485 void
 486 binlog_read(job binlog_jobs)
 487 {
 488     int binlog_index_min;
 489     struct stat sbuf;
 490     int fd, idx, r;
 491     char path[PATH_MAX];
 492     binlog b;
 493
 494     if (!binlog_dir) return;
 495
 496     if (stat(binlog_dir, &sbuf) < 0) {
 497         if (mkdir(binlog_dir, 0700) < 0) return twarn("%s", binlog_dir);
 498     } else if (!(sbuf.st_mode & S_IFDIR)) {
 499         twarnx("%s", binlog_dir);
 500         return;
 501     }
 502
 503     binlog_index_min = binlog_scan_dir();
 504
 505     if (binlog_index_min) {
 506         for (idx = binlog_index_min; idx <= binlog_index; idx++) {
 507             r = snprintf(path, PATH_MAX, "%s/binlog.%d", binlog_dir, idx);
 508             if (r > PATH_MAX) return twarnx("path too long: %s", binlog_dir);
 509
 510             fd = open(path, O_RDONLY);
 511
 512             if (fd < 0) {
 513                 twarn("%s", path);
 514             } else {
 515                 b = binlog_iref(add_binlog(make_binlog(path)));
 516                 binlog_read_one(fd, binlog_jobs, path);
 517                 close(fd);
 518                 binlog_dref(b);
 519             }
 520         }
 521
 522     }
 523 }
 524
 525 int
 526 binlog_lock()
 527 {
 528     int r;
 529     struct flock lock;
 530     char path[PATH_MAX];
 531
 532     r = snprintf(path, PATH_MAX, "%s/lock", binlog_dir);
 533     if (r > PATH_MAX) return twarnx("path too long: %s", binlog_dir), 0;
 534
 535     lock_fd = open(path, O_WRONLY|O_CREAT, 0600);
 536     if (lock_fd == -1) return twarn("open"), 0;
 537
 538     lock.l_type = F_WRLCK;
 539     lock.l_whence = SEEK_SET;
 540     lock.l_start = 0;
 541     lock.l_len = 0;
 542     r = fcntl(lock_fd, F_SETLK, &lock);
 543     if (r) return twarn("fcntl"), 0;
 544
 545     return 1;
 546 }
 547
 548 void
 549 binlog_init()
 550 {
 551     binlog log;
 552
 553     if (!binlog_dir) return;
 554
 555     log = make_next_binlog();
 556     if (!log) return twarnx("error making first binlog");
 557     binlog_fd = binlog_open(log);
 558     if (binlog_fd >= 0) add_binlog(log);
 559 }
 560
 561 const char *
 562 binlog_oldest_index()
 563 {
 564     if (!first_binlog) return "0";
 565
 566     return strrchr(first_binlog->path, '.') + 1;
 567 }
 568
 569 const char *
 570 binlog_current_index()
 571 {
 572     if (!last_binlog) return "0";
 573
 574     return strrchr(last_binlog->path, '.') + 1;
 575 }