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 1.50-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 /* \\.\ is the device namespace in Windows, where every file
435 * is a device node */
436 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
442 void td_fill_rand_seeds(struct thread_data *td)
444 os_random_seed(td->rand_seeds[0], &td->bsrange_state);
445 os_random_seed(td->rand_seeds[1], &td->verify_state);
446 os_random_seed(td->rand_seeds[2], &td->rwmix_state);
448 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
449 os_random_seed(td->rand_seeds[3], &td->next_file_state);
451 os_random_seed(td->rand_seeds[5], &td->file_size_state);
452 os_random_seed(td->rand_seeds[6], &td->trim_state);
457 if (td->o.rand_repeatable)
458 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
460 os_random_seed(td->rand_seeds[4], &td->random_state);
464 * Initialize the various random states we need (random io, block size ranges,
465 * read/write mix, etc).
467 static int init_random_state(struct thread_data *td)
471 fd = open("/dev/urandom", O_RDONLY);
473 td_verror(td, errno, "open");
477 if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
478 (int) sizeof(td->rand_seeds)) {
479 td_verror(td, EIO, "read");
485 td_fill_rand_seeds(td);
490 * Adds a job to the list of things todo. Sanitizes the various options
491 * to make sure we don't have conflicts, and initializes various
494 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
496 const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
497 "randread", "randwrite", "randrw" };
500 char fname[PATH_MAX];
501 int numjobs, file_alloced;
504 * the def_thread is just for options, it's not a real job
506 if (td == &def_thread)
510 * if we are just dumping the output command line, don't add the job
517 if (profile_td_init(td))
520 engine = get_engine_name(td->o.ioengine);
521 td->io_ops = load_ioengine(td, engine);
523 log_err("fio: failed to load engine %s\n", engine);
527 if (td->o.use_thread)
533 td->io_ops->flags |= FIO_RAWIO;
536 if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
539 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
540 add_file(td, jobname);
542 for (i = 0; i < td->o.nr_files; i++) {
543 sprintf(fname, "%s.%d.%d", jobname,
544 td->thread_number, i);
550 if (fixup_options(td))
553 if (td->io_ops->flags & FIO_DISKLESSIO) {
556 for_each_file(td, f, i)
557 f->real_file_size = -1ULL;
560 td->mutex = fio_mutex_init(0);
562 td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
563 td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
564 td->ts.lat_stat[0].min_val = td->ts.lat_stat[1].min_val = ULONG_MAX;
565 td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
566 td->ddir_seq_nr = td->o.ddir_seq_nr;
568 if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
573 td->groupid = groupid;
576 if (init_random_state(td))
582 if (td->o.write_lat_log) {
583 setup_log(&td->ts.lat_log);
584 setup_log(&td->ts.slat_log);
585 setup_log(&td->ts.clat_log);
587 if (td->o.write_bw_log)
588 setup_log(&td->ts.bw_log);
591 td->o.name = strdup(jobname);
595 if (!strcmp(td->io_ops->name, "cpuio")) {
596 log_info("%s: ioengine=cpu, cpuload=%u,"
597 " cpucycle=%u\n", td->o.name,
601 char *c1, *c2, *c3, *c4;
603 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
604 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
605 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
606 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
608 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
609 " ioengine=%s, iodepth=%u\n",
610 td->o.name, td->groupid,
611 ddir_str[td->o.td_ddir],
621 } else if (job_add_num == 1)
626 * recurse add identical jobs, clear numjobs and stonewall options
627 * as they don't apply to sub-jobs
629 numjobs = td->o.numjobs;
631 struct thread_data *td_new = get_new_job(0, td);
636 td_new->o.numjobs = 1;
637 td_new->o.stonewall = 0;
638 td_new->o.new_group = 0;
641 td_new->o.filename = NULL;
642 td_new->files_index = 0;
643 td_new->files_size = 0;
644 td_new->files = NULL;
647 job_add_num = numjobs - 1;
649 if (add_job(td_new, jobname, job_add_num))
660 * Parse as if 'o' was a command line
662 void add_job_opts(const char **o)
664 struct thread_data *td, *td_parent;
665 int i, in_global = 1;
669 td_parent = td = NULL;
671 if (!strncmp(o[i], "name", 4)) {
674 add_job(td, jobname, 0);
676 sprintf(jobname, "%s", o[i] + 5);
678 if (in_global && !td_parent)
679 td_parent = get_new_job(1, &def_thread);
680 else if (!in_global && !td) {
682 td_parent = &def_thread;
683 td = get_new_job(0, td_parent);
686 fio_options_parse(td_parent, (char **) &o[i], 1);
688 fio_options_parse(td, (char **) &o[i], 1);
693 add_job(td, jobname, 0);
696 static int skip_this_section(const char *name)
700 if (!strncmp(name, "global", 6))
703 return strcmp(job_section, name);
706 static int is_empty_or_comment(char *line)
710 for (i = 0; i < strlen(line); i++) {
715 if (!isspace(line[i]) && !iscntrl(line[i]))
723 * This is our [ini] type file parser.
725 static int parse_jobs_ini(char *file, int stonewall_flag)
728 struct thread_data *td;
732 int ret = 0, stonewall;
737 int i, alloc_opts, num_opts;
739 if (!strcmp(file, "-"))
742 f = fopen(file, "r");
745 perror("fopen job file");
749 string = malloc(4096);
752 * it's really 256 + small bit, 280 should suffice
755 memset(name, 0, 280);
758 opts = malloc(sizeof(char *) * alloc_opts);
761 stonewall = stonewall_flag;
764 * if skip_fgets is set, we already have loaded a line we
768 p = fgets(string, 4095, f);
774 strip_blank_front(&p);
777 if (is_empty_or_comment(p))
779 if (sscanf(p, "[%255s]", name) != 1) {
782 log_err("fio: option <%s> outside of [] job section\n",
787 name[strlen(name) - 1] = '\0';
789 if (skip_this_section(name)) {
795 global = !strncmp(name, "global", 6);
801 log_info("--name=%s ", name);
805 td = get_new_job(global, &def_thread);
812 * Seperate multiple job files by a stonewall
814 if (!global && stonewall) {
815 td->o.stonewall = stonewall;
820 memset(opts, 0, alloc_opts * sizeof(char *));
822 while ((p = fgets(string, 4096, f)) != NULL) {
823 if (is_empty_or_comment(p))
826 strip_blank_front(&p);
829 * new section, break out and make sure we don't
830 * fgets() a new line at the top.
839 if (num_opts == alloc_opts) {
842 alloc_opts * sizeof(char *));
845 opts[num_opts] = strdup(p);
849 ret = fio_options_parse(td, opts, num_opts);
852 for (i = 0; i < num_opts; i++)
853 log_info("--%s ", opts[i]);
855 ret = add_job(td, name, 0);
857 log_err("fio: job %s dropped\n", name);
861 for (i = 0; i < num_opts; i++)
869 for (i = 0; i < num_opts; i++)
880 static int fill_def_thread(void)
882 memset(&def_thread, 0, sizeof(def_thread));
884 fio_getaffinity(getpid(), &def_thread.o.cpumask);
887 * fill default options
889 fio_fill_default_options(&def_thread);
891 def_thread.o.timeout = def_timeout;
895 static void free_shm(void)
897 struct shmid_ds sbuf;
904 fio_debug_jobp = NULL;
906 shmctl(shm_id, IPC_RMID, &sbuf);
913 * The thread area is shared between the main process and the job
914 * threads/processes. So setup a shared memory segment that will hold
915 * all the job info. We use the end of the region for keeping track of
916 * open files across jobs, for file sharing.
918 static int setup_thread_area(void)
923 * 1024 is too much on some machines, scale max_jobs if
924 * we get a failure that looks like too large a shm segment
927 size_t size = max_jobs * sizeof(struct thread_data);
929 size += file_hash_size;
930 size += sizeof(unsigned int);
932 shm_id = shmget(0, size, IPC_CREAT | 0600);
935 if (errno != EINVAL) {
946 threads = shmat(shm_id, NULL, 0);
947 if (threads == (void *) -1) {
952 memset(threads, 0, max_jobs * sizeof(struct thread_data));
953 hash = (void *) threads + max_jobs * sizeof(struct thread_data);
954 fio_debug_jobp = (void *) hash + file_hash_size;
955 *fio_debug_jobp = -1;
956 file_hash_init(hash);
961 static void usage(const char *name)
963 printf("%s [options] [job options] <job file(s)>\n", name);
964 printf("\t--debug=options\tEnable debug logging\n");
965 printf("\t--output\tWrite output to file\n");
966 printf("\t--timeout\tRuntime in seconds\n");
967 printf("\t--latency-log\tGenerate per-job latency logs\n");
968 printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
969 printf("\t--minimal\tMinimal (terse) output\n");
970 printf("\t--version\tPrint version info and exit\n");
971 printf("\t--help\t\tPrint this page\n");
972 printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
974 printf("\t--showcmd\tTurn a job file into command line options\n");
975 printf("\t--eta=when\tWhen ETA estimate should be printed\n");
976 printf("\t \tMay be \"always\", \"never\" or \"auto\"\n");
977 printf("\t--readonly\tTurn on safety read-only checks, preventing"
979 printf("\t--section=name\tOnly run specified section in job file\n");
980 printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
982 printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
983 printf("\n Jens Axboe <jaxboe@fusionio.com>\n");
987 struct debug_level debug_levels[] = {
988 { .name = "process", .shift = FD_PROCESS, },
989 { .name = "file", .shift = FD_FILE, },
990 { .name = "io", .shift = FD_IO, },
991 { .name = "mem", .shift = FD_MEM, },
992 { .name = "blktrace", .shift = FD_BLKTRACE },
993 { .name = "verify", .shift = FD_VERIFY },
994 { .name = "random", .shift = FD_RANDOM },
995 { .name = "parse", .shift = FD_PARSE },
996 { .name = "diskutil", .shift = FD_DISKUTIL },
997 { .name = "job", .shift = FD_JOB },
998 { .name = "mutex", .shift = FD_MUTEX },
999 { .name = "profile", .shift = FD_PROFILE },
1000 { .name = "time", .shift = FD_TIME },
1004 static int set_debug(const char *string)
1006 struct debug_level *dl;
1007 char *p = (char *) string;
1011 if (!strcmp(string, "?") || !strcmp(string, "help")) {
1014 log_info("fio: dumping debug options:");
1015 for (i = 0; debug_levels[i].name; i++) {
1016 dl = &debug_levels[i];
1017 log_info("%s,", dl->name);
1023 while ((opt = strsep(&p, ",")) != NULL) {
1026 if (!strncmp(opt, "all", 3)) {
1027 log_info("fio: set all debug options\n");
1032 for (i = 0; debug_levels[i].name; i++) {
1033 dl = &debug_levels[i];
1034 found = !strncmp(opt, dl->name, strlen(dl->name));
1038 if (dl->shift == FD_JOB) {
1039 opt = strchr(opt, ':');
1041 log_err("fio: missing job number\n");
1045 fio_debug_jobno = atoi(opt);
1046 log_info("fio: set debug jobno %d\n",
1049 log_info("fio: set debug option %s\n", opt);
1050 fio_debug |= (1UL << dl->shift);
1056 log_err("fio: debug mask %s not found\n", opt);
1061 static int set_debug(const char *string)
1063 log_err("fio: debug tracing not included in build\n");
1068 static int parse_cmd_line(int argc, char *argv[])
1070 struct thread_data *td = NULL;
1071 int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1073 while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1076 smalloc_pool_size = atoi(optarg);
1079 def_timeout = atoi(optarg);
1088 f_out = fopen(optarg, "w+");
1090 perror("fopen output");
1102 exit(fio_show_option_help(optarg));
1110 /* already being printed, just quit */
1113 if (!strcmp("always", optarg))
1114 eta_print = FIO_ETA_ALWAYS;
1115 else if (!strcmp("never", optarg))
1116 eta_print = FIO_ETA_NEVER;
1119 if (set_debug(optarg))
1123 if (!strcmp(optarg, "global")) {
1124 log_err("fio: can't use global as only "
1132 job_section = strdup(optarg);
1135 exec_profile = strdup(optarg);
1137 case FIO_GETOPT_JOB: {
1138 const char *opt = l_opts[lidx].name;
1141 if (!strncmp(opt, "name", 4) && td) {
1142 ret = add_job(td, td->o.name ?: "fio", 0);
1150 int is_section = !strncmp(opt, "name", 4);
1153 if (!is_section || !strncmp(val, "global", 6))
1156 if (is_section && skip_this_section(val))
1159 td = get_new_job(global, &def_thread);
1164 ret = fio_cmd_option_parse(td, opt, val);
1179 ret = add_job(td, td->o.name ?: "fio", 0);
1184 while (optind < argc) {
1186 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1187 ini_file[ini_idx - 1] = strdup(argv[optind]);
1194 int parse_options(int argc, char *argv[])
1201 log_info("%s\n", fio_version_string);
1203 fio_options_dup_and_init(l_opts);
1205 if (setup_thread_area())
1207 if (fill_def_thread())
1210 job_files = parse_cmd_line(argc, argv);
1212 for (i = 0; i < job_files; i++) {
1213 if (fill_def_thread())
1215 if (parse_jobs_ini(ini_file[i], i))
1221 options_mem_free(&def_thread);
1223 if (!thread_number) {
1229 log_err("No jobs(s) defined\n\n");
1234 if (def_thread.o.gtod_offload) {
1236 fio_gtod_offload = 1;
1237 fio_gtod_cpu = def_thread.o.gtod_cpu;