2 * This file contains job initialization and setup functions.
13 #include <sys/types.h>
23 #include "lib/getopt.h"
25 static char fio_version_string[] = "fio 1.44-rc1";
27 #define FIO_RANDSEED (0xb1899bedUL)
29 static char **ini_file;
30 static int max_jobs = MAX_JOBS;
31 static int dump_cmdline;
33 static struct thread_data def_thread;
34 struct thread_data *threads = NULL;
36 int exitall_on_terminate = 0;
39 unsigned long long mlock_size = 0;
42 char *job_section = NULL;
43 char *exec_profile = NULL;
48 static int def_timeout;
49 static int write_lat_log;
51 static int prev_group_jobs;
53 unsigned long fio_debug = 0;
54 unsigned int fio_debug_jobno = -1;
55 unsigned int *fio_debug_jobp = NULL;
58 * Command line options. These will contain the above, plus a few
59 * extra that only pertain to fio itself and not jobs.
61 static struct option l_opts[FIO_NR_OPTIONS] = {
64 .has_arg = required_argument,
69 .has_arg = required_argument,
73 .name = "latency-log",
74 .has_arg = required_argument,
78 .name = "bandwidth-log",
79 .has_arg = required_argument,
84 .has_arg = optional_argument,
89 .has_arg = no_argument,
94 .has_arg = no_argument,
99 .has_arg = optional_argument,
104 .has_arg = no_argument,
109 .has_arg = no_argument,
114 .has_arg = required_argument,
119 .has_arg = required_argument,
124 .has_arg = required_argument,
128 .name = "alloc-size",
129 .has_arg = required_argument,
134 .has_arg = required_argument,
153 * Return a free job structure.
155 static struct thread_data *get_new_job(int global, struct thread_data *parent)
157 struct thread_data *td;
161 if (thread_number >= max_jobs)
164 td = &threads[thread_number++];
167 td->o.uid = td->o.gid = -1U;
169 dup_files(td, parent);
170 options_mem_dupe(td);
172 profile_add_hooks(td);
174 td->thread_number = thread_number;
178 static void put_job(struct thread_data *td)
180 if (td == &def_thread)
186 log_info("fio: %s\n", td->verror);
188 memset(&threads[td->thread_number - 1], 0, sizeof(*td));
192 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
194 unsigned int bs = td->o.min_bs[ddir];
195 unsigned long long bytes_per_sec;
197 assert(ddir_rw(ddir));
199 if (td->o.rate[ddir])
200 bytes_per_sec = td->o.rate[ddir];
202 bytes_per_sec = td->o.rate_iops[ddir] * bs;
204 if (!bytes_per_sec) {
205 log_err("rate lower than supported\n");
209 td->rate_nsec_cycle[ddir] = 1000000000ULL / bytes_per_sec;
210 td->rate_pending_usleep[ddir] = 0;
214 static int setup_rate(struct thread_data *td)
218 if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
219 ret = __setup_rate(td, DDIR_READ);
220 if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
221 ret |= __setup_rate(td, DDIR_WRITE);
226 static int fixed_block_size(struct thread_options *o)
228 return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
229 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
230 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
234 * Lazy way of fixing up options that depend on each other. We could also
235 * define option callback handlers, but this is easier.
237 static int fixup_options(struct thread_data *td)
239 struct thread_options *o = &td->o;
241 #ifndef FIO_HAVE_PSHARED_MUTEX
242 if (!o->use_thread) {
243 log_info("fio: this platform does not support process shared"
244 " mutexes, forcing use of threads. Use the 'thread'"
245 " option to get rid of this warning.\n");
250 if (o->write_iolog_file && o->read_iolog_file) {
251 log_err("fio: read iolog overrides write_iolog\n");
252 free(o->write_iolog_file);
253 o->write_iolog_file = NULL;
257 * only really works for sequential io for now, and with 1 file
259 if (o->zone_size && td_random(td) && o->open_files == 1)
263 * Reads can do overwrites, we always need to pre-create the file
265 if (td_read(td) || td_rw(td))
268 if (!o->min_bs[DDIR_READ])
269 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
270 if (!o->max_bs[DDIR_READ])
271 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
272 if (!o->min_bs[DDIR_WRITE])
273 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
274 if (!o->max_bs[DDIR_WRITE])
275 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
277 o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
280 * For random IO, allow blockalign offset other than min_bs.
282 if (!o->ba[DDIR_READ] || !td_random(td))
283 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
284 if (!o->ba[DDIR_WRITE] || !td_random(td))
285 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
287 if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
288 o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
290 log_err("fio: Any use of blockalign= turns off randommap\n");
294 if (!o->file_size_high)
295 o->file_size_high = o->file_size_low;
297 if (o->norandommap && o->verify != VERIFY_NONE
298 && !fixed_block_size(o)) {
299 log_err("fio: norandommap given for variable block sizes, "
300 "verify disabled\n");
301 o->verify = VERIFY_NONE;
303 if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
304 log_err("fio: bs_unaligned may not work with raw io\n");
307 * thinktime_spin must be less than thinktime
309 if (o->thinktime_spin > o->thinktime)
310 o->thinktime_spin = o->thinktime;
313 * The low water mark cannot be bigger than the iodepth
315 if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
317 * syslet work around - if the workload is sequential,
318 * we want to let the queue drain all the way down to
319 * avoid seeking between async threads
321 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
324 o->iodepth_low = o->iodepth;
328 * If batch number isn't set, default to the same as iodepth
330 if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
331 o->iodepth_batch = o->iodepth;
333 if (o->nr_files > td->files_index)
334 o->nr_files = td->files_index;
336 if (o->open_files > o->nr_files || !o->open_files)
337 o->open_files = o->nr_files;
339 if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
340 ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
341 o->rate_iops_min[1]))) {
342 log_err("fio: rate and rate_iops are mutually exclusive\n");
345 if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
346 (o->rate_iops[0] < o->rate_iops_min[0]) ||
347 (o->rate_iops[1] < o->rate_iops_min[1])) {
348 log_err("fio: minimum rate exceeds rate\n");
352 if (!o->timeout && o->time_based) {
353 log_err("fio: time_based requires a runtime/timeout setting\n");
357 if (o->fill_device && !o->size)
360 if (td_rw(td) && o->verify != VERIFY_NONE)
361 log_info("fio: mixed read/write workload with verify. May not "
362 "work as expected, unless you pre-populated the file\n");
364 if (o->verify != VERIFY_NONE) {
365 o->refill_buffers = 1;
366 if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
368 o->verify_interval = o->min_bs[DDIR_WRITE];
372 o->invalidate_cache = 0;
373 if (td->io_ops->flags & FIO_PIPEIO)
374 log_info("fio: cannot pre-read files with an IO engine"
375 " that isn't seekable. Pre-read disabled.\n");
378 #ifndef FIO_HAVE_FDATASYNC
379 if (o->fdatasync_blocks) {
380 log_info("fio: this platform does not support fdatasync()"
381 " falling back to using fsync(). Use the 'fsync'"
382 " option instead of 'fdatasync' to get rid of"
384 o->fsync_blocks = o->fdatasync_blocks;
385 o->fdatasync_blocks = 0;
393 * This function leaks the buffer
395 static char *to_kmg(unsigned int val)
397 char *buf = malloc(32);
398 char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
409 snprintf(buf, 31, "%u%c", val, *p);
413 /* External engines are specified by "external:name.o") */
414 static const char *get_engine_name(const char *str)
416 char *p = strstr(str, ":");
422 strip_blank_front(&p);
427 static int exists_and_not_file(const char *filename)
431 if (lstat(filename, &sb) == -1)
434 if (S_ISREG(sb.st_mode))
440 void td_fill_rand_seeds(struct thread_data *td)
442 os_random_seed(td->rand_seeds[0], &td->bsrange_state);
443 os_random_seed(td->rand_seeds[1], &td->verify_state);
444 os_random_seed(td->rand_seeds[2], &td->rwmix_state);
446 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
447 os_random_seed(td->rand_seeds[3], &td->next_file_state);
449 os_random_seed(td->rand_seeds[5], &td->file_size_state);
450 os_random_seed(td->rand_seeds[6], &td->trim_state);
455 if (td->o.rand_repeatable)
456 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
458 os_random_seed(td->rand_seeds[4], &td->random_state);
462 * Initialize the various random states we need (random io, block size ranges,
463 * read/write mix, etc).
465 static int init_random_state(struct thread_data *td)
469 fd = open("/dev/urandom", O_RDONLY);
471 td_verror(td, errno, "open");
475 if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
476 (int) sizeof(td->rand_seeds)) {
477 td_verror(td, EIO, "read");
483 td_fill_rand_seeds(td);
488 * Adds a job to the list of things todo. Sanitizes the various options
489 * to make sure we don't have conflicts, and initializes various
492 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
494 const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
495 "randread", "randwrite", "randrw" };
498 char fname[PATH_MAX];
499 int numjobs, file_alloced;
502 * the def_thread is just for options, it's not a real job
504 if (td == &def_thread)
508 * if we are just dumping the output command line, don't add the job
515 if (profile_td_init(td))
518 engine = get_engine_name(td->o.ioengine);
519 td->io_ops = load_ioengine(td, engine);
521 log_err("fio: failed to load engine %s\n", engine);
525 if (td->o.use_thread)
531 td->io_ops->flags |= FIO_RAWIO;
534 if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
537 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
538 add_file(td, jobname);
540 for (i = 0; i < td->o.nr_files; i++) {
541 sprintf(fname, "%s.%d.%d", jobname,
542 td->thread_number, i);
548 if (fixup_options(td))
551 if (td->io_ops->flags & FIO_DISKLESSIO) {
554 for_each_file(td, f, i)
555 f->real_file_size = -1ULL;
558 td->mutex = fio_mutex_init(0);
560 td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
561 td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
562 td->ts.lat_stat[0].min_val = td->ts.lat_stat[1].min_val = ULONG_MAX;
563 td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
564 td->ddir_seq_nr = td->o.ddir_seq_nr;
566 if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
571 td->groupid = groupid;
574 if (init_random_state(td))
580 if (td->o.write_lat_log) {
581 setup_log(&td->ts.lat_log);
582 setup_log(&td->ts.slat_log);
583 setup_log(&td->ts.clat_log);
585 if (td->o.write_bw_log)
586 setup_log(&td->ts.bw_log);
589 td->o.name = strdup(jobname);
593 if (!strcmp(td->io_ops->name, "cpuio")) {
594 log_info("%s: ioengine=cpu, cpuload=%u,"
595 " cpucycle=%u\n", td->o.name,
599 char *c1, *c2, *c3, *c4;
601 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
602 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
603 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
604 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
606 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
607 " ioengine=%s, iodepth=%u\n",
608 td->o.name, td->groupid,
609 ddir_str[td->o.td_ddir],
619 } else if (job_add_num == 1)
624 * recurse add identical jobs, clear numjobs and stonewall options
625 * as they don't apply to sub-jobs
627 numjobs = td->o.numjobs;
629 struct thread_data *td_new = get_new_job(0, td);
634 td_new->o.numjobs = 1;
635 td_new->o.stonewall = 0;
636 td_new->o.new_group = 0;
639 td_new->o.filename = NULL;
640 td_new->files_index = 0;
641 td_new->files_size = 0;
642 td_new->files = NULL;
645 job_add_num = numjobs - 1;
647 if (add_job(td_new, jobname, job_add_num))
658 * Parse as if 'o' was a command line
660 void add_job_opts(const char **o)
662 struct thread_data *td, *td_parent;
663 int i, in_global = 1;
667 td_parent = td = NULL;
669 if (!strncmp(o[i], "name", 4)) {
672 add_job(td, jobname, 0);
674 sprintf(jobname, "%s", o[i] + 5);
676 if (in_global && !td_parent)
677 td_parent = get_new_job(1, &def_thread);
678 else if (!in_global && !td) {
680 td_parent = &def_thread;
681 td = get_new_job(0, td_parent);
684 fio_options_parse(td_parent, (char **) &o[i], 1);
686 fio_options_parse(td, (char **) &o[i], 1);
691 add_job(td, jobname, 0);
694 static int skip_this_section(const char *name)
698 if (!strncmp(name, "global", 6))
701 return strcmp(job_section, name);
704 static int is_empty_or_comment(char *line)
708 for (i = 0; i < strlen(line); i++) {
713 if (!isspace(line[i]) && !iscntrl(line[i]))
721 * This is our [ini] type file parser.
723 static int parse_jobs_ini(char *file, int stonewall_flag)
726 struct thread_data *td;
730 int ret = 0, stonewall;
735 int i, alloc_opts, num_opts;
737 if (!strcmp(file, "-"))
740 f = fopen(file, "r");
743 perror("fopen job file");
747 string = malloc(4096);
750 * it's really 256 + small bit, 280 should suffice
753 memset(name, 0, 280);
756 opts = malloc(sizeof(char *) * alloc_opts);
759 stonewall = stonewall_flag;
762 * if skip_fgets is set, we already have loaded a line we
766 p = fgets(string, 4095, f);
772 strip_blank_front(&p);
775 if (is_empty_or_comment(p))
777 if (sscanf(p, "[%255s]", name) != 1) {
780 log_err("fio: option <%s> outside of [] job section\n",
785 name[strlen(name) - 1] = '\0';
787 if (skip_this_section(name)) {
793 global = !strncmp(name, "global", 6);
799 log_info("--name=%s ", name);
803 td = get_new_job(global, &def_thread);
810 * Seperate multiple job files by a stonewall
812 if (!global && stonewall) {
813 td->o.stonewall = stonewall;
818 memset(opts, 0, alloc_opts * sizeof(char *));
820 while ((p = fgets(string, 4096, f)) != NULL) {
821 if (is_empty_or_comment(p))
824 strip_blank_front(&p);
827 * new section, break out and make sure we don't
828 * fgets() a new line at the top.
837 if (num_opts == alloc_opts) {
840 alloc_opts * sizeof(char *));
843 opts[num_opts] = strdup(p);
847 ret = fio_options_parse(td, opts, num_opts);
850 for (i = 0; i < num_opts; i++)
851 log_info("--%s ", opts[i]);
853 ret = add_job(td, name, 0);
855 log_err("fio: job %s dropped\n", name);
859 for (i = 0; i < num_opts; i++)
867 for (i = 0; i < num_opts; i++)
878 static int fill_def_thread(void)
880 memset(&def_thread, 0, sizeof(def_thread));
882 fio_getaffinity(getpid(), &def_thread.o.cpumask);
885 * fill default options
887 fio_fill_default_options(&def_thread);
889 def_thread.o.timeout = def_timeout;
893 static void free_shm(void)
895 struct shmid_ds sbuf;
902 fio_debug_jobp = NULL;
904 shmctl(shm_id, IPC_RMID, &sbuf);
911 * The thread area is shared between the main process and the job
912 * threads/processes. So setup a shared memory segment that will hold
913 * all the job info. We use the end of the region for keeping track of
914 * open files across jobs, for file sharing.
916 static int setup_thread_area(void)
921 * 1024 is too much on some machines, scale max_jobs if
922 * we get a failure that looks like too large a shm segment
925 size_t size = max_jobs * sizeof(struct thread_data);
927 size += file_hash_size;
928 size += sizeof(unsigned int);
930 shm_id = shmget(0, size, IPC_CREAT | 0600);
933 if (errno != EINVAL) {
944 threads = shmat(shm_id, NULL, 0);
945 if (threads == (void *) -1) {
950 memset(threads, 0, max_jobs * sizeof(struct thread_data));
951 hash = (void *) threads + max_jobs * sizeof(struct thread_data);
952 fio_debug_jobp = (void *) hash + file_hash_size;
953 *fio_debug_jobp = -1;
954 file_hash_init(hash);
959 static void usage(const char *name)
961 printf("%s\n", fio_version_string);
962 printf("%s [options] [job options] <job file(s)>\n", name);
963 printf("\t--debug=options\tEnable debug logging\n");
964 printf("\t--output\tWrite output to file\n");
965 printf("\t--timeout\tRuntime in seconds\n");
966 printf("\t--latency-log\tGenerate per-job latency logs\n");
967 printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
968 printf("\t--minimal\tMinimal (terse) output\n");
969 printf("\t--version\tPrint version info and exit\n");
970 printf("\t--help\t\tPrint this page\n");
971 printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
973 printf("\t--showcmd\tTurn a job file into command line options\n");
974 printf("\t--eta=when\tWhen ETA estimate should be printed\n");
975 printf("\t \tMay be \"always\", \"never\" or \"auto\"\n");
976 printf("\t--readonly\tTurn on safety read-only checks, preventing"
978 printf("\t--section=name\tOnly run specified section in job file\n");
979 printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
981 printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
982 printf("\n Jens Axboe <jaxboe@fusionio.com>\n");
986 struct debug_level debug_levels[] = {
987 { .name = "process", .shift = FD_PROCESS, },
988 { .name = "file", .shift = FD_FILE, },
989 { .name = "io", .shift = FD_IO, },
990 { .name = "mem", .shift = FD_MEM, },
991 { .name = "blktrace", .shift = FD_BLKTRACE },
992 { .name = "verify", .shift = FD_VERIFY },
993 { .name = "random", .shift = FD_RANDOM },
994 { .name = "parse", .shift = FD_PARSE },
995 { .name = "diskutil", .shift = FD_DISKUTIL },
996 { .name = "job", .shift = FD_JOB },
997 { .name = "mutex", .shift = FD_MUTEX },
998 { .name = "profile", .shift = FD_PROFILE },
999 { .name = "time", .shift = FD_TIME },
1003 static int set_debug(const char *string)
1005 struct debug_level *dl;
1006 char *p = (char *) string;
1010 if (!strcmp(string, "?") || !strcmp(string, "help")) {
1013 log_info("fio: dumping debug options:");
1014 for (i = 0; debug_levels[i].name; i++) {
1015 dl = &debug_levels[i];
1016 log_info("%s,", dl->name);
1022 while ((opt = strsep(&p, ",")) != NULL) {
1025 if (!strncmp(opt, "all", 3)) {
1026 log_info("fio: set all debug options\n");
1031 for (i = 0; debug_levels[i].name; i++) {
1032 dl = &debug_levels[i];
1033 found = !strncmp(opt, dl->name, strlen(dl->name));
1037 if (dl->shift == FD_JOB) {
1038 opt = strchr(opt, ':');
1040 log_err("fio: missing job number\n");
1044 fio_debug_jobno = atoi(opt);
1045 log_info("fio: set debug jobno %d\n",
1048 log_info("fio: set debug option %s\n", opt);
1049 fio_debug |= (1UL << dl->shift);
1055 log_err("fio: debug mask %s not found\n", opt);
1060 static int set_debug(const char *string)
1062 log_err("fio: debug tracing not included in build\n");
1067 static int parse_cmd_line(int argc, char *argv[])
1069 struct thread_data *td = NULL;
1070 int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1072 while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1075 smalloc_pool_size = atoi(optarg);
1078 def_timeout = atoi(optarg);
1087 f_out = fopen(optarg, "w+");
1089 perror("fopen output");
1101 exit(fio_show_option_help(optarg));
1109 printf("%s\n", fio_version_string);
1112 if (!strcmp("always", optarg))
1113 eta_print = FIO_ETA_ALWAYS;
1114 else if (!strcmp("never", optarg))
1115 eta_print = FIO_ETA_NEVER;
1118 if (set_debug(optarg))
1122 if (!strcmp(optarg, "global")) {
1123 log_err("fio: can't use global as only "
1131 job_section = strdup(optarg);
1134 exec_profile = strdup(optarg);
1136 case FIO_GETOPT_JOB: {
1137 const char *opt = l_opts[lidx].name;
1140 if (!strncmp(opt, "name", 4) && td) {
1141 ret = add_job(td, td->o.name ?: "fio", 0);
1149 int is_section = !strncmp(opt, "name", 4);
1152 if (!is_section || !strncmp(val, "global", 6))
1155 if (is_section && skip_this_section(val))
1158 td = get_new_job(global, &def_thread);
1163 ret = fio_cmd_option_parse(td, opt, val);
1178 ret = add_job(td, td->o.name ?: "fio", 0);
1183 while (optind < argc) {
1185 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1186 ini_file[ini_idx - 1] = strdup(argv[optind]);
1193 int parse_options(int argc, char *argv[])
1200 fio_options_dup_and_init(l_opts);
1202 if (setup_thread_area())
1204 if (fill_def_thread())
1207 job_files = parse_cmd_line(argc, argv);
1209 for (i = 0; i < job_files; i++) {
1210 if (fill_def_thread())
1212 if (parse_jobs_ini(ini_file[i], i))
1218 options_mem_free(&def_thread);
1220 if (!thread_number) {
1226 log_err("No jobs defined(s)\n\n");
1231 if (def_thread.o.gtod_offload) {
1233 fio_gtod_offload = 1;
1234 fio_gtod_cpu = def_thread.o.gtod_cpu;