2 * This file contains job initialization and setup functions.
14 #include <sys/types.h>
23 static char fio_version_string[] = "fio 1.28";
25 #define FIO_RANDSEED (0xb1899bedUL)
27 static char **ini_file;
28 static int max_jobs = MAX_JOBS;
29 static int dump_cmdline;
31 static struct thread_data def_thread;
32 struct thread_data *threads = NULL;
34 int exitall_on_terminate = 0;
37 unsigned long long mlock_size = 0;
40 char *job_section = NULL;
45 static int def_timeout;
46 static int write_lat_log;
48 static int prev_group_jobs;
50 unsigned long fio_debug = 0;
51 unsigned int fio_debug_jobno = -1;
52 unsigned int *fio_debug_jobp = NULL;
55 * Command line options. These will contain the above, plus a few
56 * extra that only pertain to fio itself and not jobs.
58 static struct option l_opts[FIO_NR_OPTIONS] = {
61 .has_arg = required_argument,
66 .has_arg = required_argument,
70 .name = "latency-log",
71 .has_arg = required_argument,
75 .name = "bandwidth-log",
76 .has_arg = required_argument,
81 .has_arg = optional_argument,
86 .has_arg = no_argument,
91 .has_arg = no_argument,
96 .has_arg = optional_argument,
101 .has_arg = no_argument,
106 .has_arg = no_argument,
111 .has_arg = required_argument,
116 .has_arg = required_argument,
121 .has_arg = required_argument,
125 .name = "alloc-size",
126 .has_arg = required_argument,
145 * Return a free job structure.
147 static struct thread_data *get_new_job(int global, struct thread_data *parent)
149 struct thread_data *td;
153 if (thread_number >= max_jobs)
156 td = &threads[thread_number++];
159 dup_files(td, parent);
160 options_mem_dupe(td);
162 td->thread_number = thread_number;
166 static void put_job(struct thread_data *td)
168 if (td == &def_thread)
172 log_info("fio: %s\n", td->verror);
174 memset(&threads[td->thread_number - 1], 0, sizeof(*td));
178 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
180 unsigned int bs = td->o.min_bs[ddir];
181 unsigned long long rate;
182 unsigned long ios_per_msec;
184 if (td->o.rate[ddir]) {
185 rate = td->o.rate[ddir];
186 ios_per_msec = (rate * 1000LL) / bs;
188 ios_per_msec = td->o.rate_iops[ddir] * 1000UL;
191 log_err("rate lower than supported\n");
195 td->rate_usec_cycle[ddir] = 1000000000ULL / ios_per_msec;
196 td->rate_pending_usleep[ddir] = 0;
200 static int setup_rate(struct thread_data *td)
204 if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
205 ret = __setup_rate(td, DDIR_READ);
206 if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
207 ret |= __setup_rate(td, DDIR_WRITE);
212 static int fixed_block_size(struct thread_options *o)
214 return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
215 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
216 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
220 * Lazy way of fixing up options that depend on each other. We could also
221 * define option callback handlers, but this is easier.
223 static int fixup_options(struct thread_data *td)
225 struct thread_options *o = &td->o;
227 #ifndef FIO_HAVE_PSHARED_MUTEX
228 if (!td->o.use_thread) {
229 log_info("fio: this platform does not support process shared"
230 " mutexes, forcing use of threads. Use the 'thread'"
231 " option to get rid of this warning.\n");
232 td->o.use_thread = 1;
236 #ifndef FIO_HAVE_CPU_AFFINITY
237 if (td->o.gtod_cpu) {
238 log_err("fio: platform must support CPU affinity for"
239 "gettimeofday() offloading\n");
245 if (read_only && td_write(td)) {
246 log_err("fio: job <%s> has write bit set, but fio is in"
247 " read-only mode\n", td->o.name);
252 if (o->write_iolog_file && o->read_iolog_file) {
253 log_err("fio: read iolog overrides write_iolog\n");
254 free(o->write_iolog_file);
255 o->write_iolog_file = NULL;
259 * only really works for sequential io for now, and with 1 file
261 if (o->zone_size && td_random(td) && o->open_files == 1)
265 * Reads can do overwrites, we always need to pre-create the file
267 if (td_read(td) || td_rw(td))
270 if (!o->min_bs[DDIR_READ])
271 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
272 if (!o->max_bs[DDIR_READ])
273 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
274 if (!o->min_bs[DDIR_WRITE])
275 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
276 if (!o->max_bs[DDIR_WRITE])
277 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
279 o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
282 * For random IO, allow blockalign offset other than min_bs.
284 if (!o->ba[DDIR_READ] || !td_random(td))
285 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
286 if (!o->ba[DDIR_WRITE] || !td_random(td))
287 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
289 if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
290 o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
291 !td->o.norandommap) {
292 log_err("fio: Any use of blockalign= turns off randommap\n");
293 td->o.norandommap = 1;
296 if (!o->file_size_high)
297 o->file_size_high = o->file_size_low;
299 if (o->norandommap && o->verify != VERIFY_NONE
300 && !fixed_block_size(o)) {
301 log_err("fio: norandommap given for variable block sizes, "
302 "verify disabled\n");
303 o->verify = VERIFY_NONE;
305 if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
306 log_err("fio: bs_unaligned may not work with raw io\n");
309 * thinktime_spin must be less than thinktime
311 if (o->thinktime_spin > o->thinktime)
312 o->thinktime_spin = o->thinktime;
315 * The low water mark cannot be bigger than the iodepth
317 if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
319 * syslet work around - if the workload is sequential,
320 * we want to let the queue drain all the way down to
321 * avoid seeking between async threads
323 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
326 o->iodepth_low = o->iodepth;
330 * If batch number isn't set, default to the same as iodepth
332 if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
333 o->iodepth_batch = o->iodepth;
335 if (o->nr_files > td->files_index)
336 o->nr_files = td->files_index;
338 if (o->open_files > o->nr_files || !o->open_files)
339 o->open_files = o->nr_files;
341 if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
342 ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
343 o->rate_iops_min[1]))) {
344 log_err("fio: rate and rate_iops are mutually exclusive\n");
347 if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
348 (o->rate_iops[0] < o->rate_iops_min[0]) ||
349 (o->rate_iops[1] < o->rate_iops_min[1])) {
350 log_err("fio: minimum rate exceeds rate\n");
354 if (!o->timeout && o->time_based) {
355 log_err("fio: time_based requires a runtime/timeout setting\n");
359 if (o->fill_device && !o->size)
362 if (td_rw(td) && td->o.verify != VERIFY_NONE)
363 log_info("fio: mixed read/write workload with verify. May not "
364 "work as expected, unless you pre-populated the file\n");
366 if (td->o.verify != VERIFY_NONE)
367 td->o.refill_buffers = 1;
370 td->o.invalidate_cache = 0;
372 if (td->o.mem_align) {
373 if (td->o.odirect && !is_power_of_2(td->o.mem_align)) {
374 log_err("fio: given IO mem alignment conflicts with"
375 " direct=1. Resetting.\n");
376 td->o.mem_align = page_mask;
384 * This function leaks the buffer
386 static char *to_kmg(unsigned int val)
388 char *buf = malloc(32);
389 char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
400 snprintf(buf, 31, "%u%c", val, *p);
404 /* External engines are specified by "external:name.o") */
405 static const char *get_engine_name(const char *str)
407 char *p = strstr(str, ":");
413 strip_blank_front(&p);
418 static int exists_and_not_file(const char *filename)
422 if (lstat(filename, &sb) == -1)
425 if (S_ISREG(sb.st_mode))
431 void td_fill_rand_seeds(struct thread_data *td)
433 os_random_seed(td->rand_seeds[0], &td->bsrange_state);
434 os_random_seed(td->rand_seeds[1], &td->verify_state);
435 os_random_seed(td->rand_seeds[2], &td->rwmix_state);
437 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
438 os_random_seed(td->rand_seeds[3], &td->next_file_state);
440 os_random_seed(td->rand_seeds[5], &td->file_size_state);
445 if (td->o.rand_repeatable)
446 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
448 os_random_seed(td->rand_seeds[4], &td->random_state);
452 * Initialize the various random states we need (random io, block size ranges,
453 * read/write mix, etc).
455 static int init_random_state(struct thread_data *td)
459 fd = open("/dev/urandom", O_RDONLY);
461 td_verror(td, errno, "open");
465 if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
466 (int) sizeof(td->rand_seeds)) {
467 td_verror(td, EIO, "read");
473 td_fill_rand_seeds(td);
478 * Adds a job to the list of things todo. Sanitizes the various options
479 * to make sure we don't have conflicts, and initializes various
482 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
484 const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
485 "randread", "randwrite", "randrw" };
488 char fname[PATH_MAX];
489 int numjobs, file_alloced;
492 * the def_thread is just for options, it's not a real job
494 if (td == &def_thread)
498 * if we are just dumping the output command line, don't add the job
505 engine = get_engine_name(td->o.ioengine);
506 td->io_ops = load_ioengine(td, engine);
508 log_err("fio: failed to load engine %s\n", engine);
512 if (td->o.use_thread)
518 td->io_ops->flags |= FIO_RAWIO;
521 if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
524 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
525 add_file(td, jobname);
527 for (i = 0; i < td->o.nr_files; i++) {
528 sprintf(fname, "%s.%d.%d", jobname,
529 td->thread_number, i);
535 if (fixup_options(td))
538 if (td->io_ops->flags & FIO_DISKLESSIO) {
541 for_each_file(td, f, i)
542 f->real_file_size = -1ULL;
545 td->mutex = fio_mutex_init(0);
547 td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
548 td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
549 td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
550 td->ddir_nr = td->o.ddir_nr;
552 if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
553 && prev_group_jobs) {
558 td->groupid = groupid;
561 if (init_random_state(td))
567 if (td->o.write_lat_log) {
568 setup_log(&td->ts.slat_log);
569 setup_log(&td->ts.clat_log);
571 if (td->o.write_bw_log)
572 setup_log(&td->ts.bw_log);
575 td->o.name = strdup(jobname);
579 if (!strcmp(td->io_ops->name, "cpuio")) {
580 log_info("%s: ioengine=cpu, cpuload=%u,"
581 " cpucycle=%u\n", td->o.name,
585 char *c1, *c2, *c3, *c4;
587 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
588 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
589 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
590 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
592 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
593 " ioengine=%s, iodepth=%u\n",
594 td->o.name, td->groupid,
595 ddir_str[td->o.td_ddir],
605 } else if (job_add_num == 1)
610 * recurse add identical jobs, clear numjobs and stonewall options
611 * as they don't apply to sub-jobs
613 numjobs = td->o.numjobs;
615 struct thread_data *td_new = get_new_job(0, td);
620 td_new->o.numjobs = 1;
621 td_new->o.stonewall = 0;
622 td_new->o.new_group = 0;
625 td_new->o.filename = NULL;
626 td_new->files_index = 0;
627 td_new->files_size = 0;
628 td_new->files = NULL;
631 job_add_num = numjobs - 1;
633 if (add_job(td_new, jobname, job_add_num))
643 static int skip_this_section(const char *name)
647 if (!strncmp(name, "global", 6))
650 return strcmp(job_section, name);
653 static int is_empty_or_comment(char *line)
657 for (i = 0; i < strlen(line); i++) {
662 if (!isspace(line[i]) && !iscntrl(line[i]))
670 * This is our [ini] type file parser.
672 static int parse_jobs_ini(char *file, int stonewall_flag)
675 struct thread_data *td;
679 int ret = 0, stonewall;
684 int i, alloc_opts, num_opts;
686 if (!strcmp(file, "-"))
689 f = fopen(file, "r");
692 perror("fopen job file");
696 string = malloc(4096);
699 * it's really 256 + small bit, 280 should suffice
702 memset(name, 0, 280);
705 opts = malloc(sizeof(char *) * alloc_opts);
708 stonewall = stonewall_flag;
711 * if skip_fgets is set, we already have loaded a line we
715 p = fgets(string, 4095, f);
721 strip_blank_front(&p);
724 if (is_empty_or_comment(p))
726 if (sscanf(p, "[%255s]", name) != 1) {
729 log_err("fio: option <%s> outside of [] job section\n",
734 if (skip_this_section(name)) {
740 global = !strncmp(name, "global", 6);
742 name[strlen(name) - 1] = '\0';
748 log_info("--name=%s ", name);
752 td = get_new_job(global, &def_thread);
759 * Seperate multiple job files by a stonewall
761 if (!global && stonewall) {
762 td->o.stonewall = stonewall;
767 memset(opts, 0, alloc_opts * sizeof(char *));
769 while ((p = fgets(string, 4096, f)) != NULL) {
770 if (is_empty_or_comment(p))
773 strip_blank_front(&p);
776 * new section, break out and make sure we don't
777 * fgets() a new line at the top.
786 if (num_opts == alloc_opts) {
789 alloc_opts * sizeof(char *));
792 opts[num_opts] = strdup(p);
796 ret = fio_options_parse(td, opts, num_opts);
799 for (i = 0; i < num_opts; i++)
800 log_info("--%s ", opts[i]);
802 ret = add_job(td, name, 0);
804 log_err("fio: job %s dropped\n", name);
808 for (i = 0; i < num_opts; i++)
816 for (i = 0; i < num_opts; i++)
827 static int fill_def_thread(void)
829 memset(&def_thread, 0, sizeof(def_thread));
831 fio_getaffinity(getpid(), &def_thread.o.cpumask);
834 * fill default options
836 fio_fill_default_options(&def_thread);
838 def_thread.o.timeout = def_timeout;
839 def_thread.o.write_bw_log = write_bw_log;
840 def_thread.o.write_lat_log = write_lat_log;
845 static void free_shm(void)
847 struct shmid_ds sbuf;
854 fio_debug_jobp = NULL;
856 shmctl(shm_id, IPC_RMID, &sbuf);
863 * The thread area is shared between the main process and the job
864 * threads/processes. So setup a shared memory segment that will hold
865 * all the job info. We use the end of the region for keeping track of
866 * open files across jobs, for file sharing.
868 static int setup_thread_area(void)
873 * 1024 is too much on some machines, scale max_jobs if
874 * we get a failure that looks like too large a shm segment
877 size_t size = max_jobs * sizeof(struct thread_data);
879 size += file_hash_size;
880 size += sizeof(unsigned int);
882 shm_id = shmget(0, size, IPC_CREAT | 0600);
885 if (errno != EINVAL) {
896 threads = shmat(shm_id, NULL, 0);
897 if (threads == (void *) -1) {
902 memset(threads, 0, max_jobs * sizeof(struct thread_data));
903 hash = (void *) threads + max_jobs * sizeof(struct thread_data);
904 fio_debug_jobp = (void *) hash + file_hash_size;
905 *fio_debug_jobp = -1;
906 file_hash_init(hash);
911 static void usage(const char *name)
913 printf("%s\n", fio_version_string);
914 printf("%s [options] [job options] <job file(s)>\n", name);
915 printf("\t--debug=options\tEnable debug logging\n");
916 printf("\t--output\tWrite output to file\n");
917 printf("\t--timeout\tRuntime in seconds\n");
918 printf("\t--latency-log\tGenerate per-job latency logs\n");
919 printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
920 printf("\t--minimal\tMinimal (terse) output\n");
921 printf("\t--version\tPrint version info and exit\n");
922 printf("\t--help\t\tPrint this page\n");
923 printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
925 printf("\t--showcmd\tTurn a job file into command line options\n");
926 printf("\t--eta=when\tWhen ETA estimate should be printed\n");
927 printf("\t \tMay be \"always\", \"never\" or \"auto\"\n");
928 printf("\t--readonly\tTurn on safety read-only checks, preventing"
930 printf("\t--section=name\tOnly run specified section in job file\n");
931 printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
933 printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>\n");
937 struct debug_level debug_levels[] = {
938 { .name = "process", .shift = FD_PROCESS, },
939 { .name = "file", .shift = FD_FILE, },
940 { .name = "io", .shift = FD_IO, },
941 { .name = "mem", .shift = FD_MEM, },
942 { .name = "blktrace", .shift = FD_BLKTRACE },
943 { .name = "verify", .shift = FD_VERIFY },
944 { .name = "random", .shift = FD_RANDOM },
945 { .name = "parse", .shift = FD_PARSE },
946 { .name = "diskutil", .shift = FD_DISKUTIL },
947 { .name = "job", .shift = FD_JOB },
948 { .name = "mutex", .shift = FD_MUTEX },
952 static int set_debug(const char *string)
954 struct debug_level *dl;
955 char *p = (char *) string;
959 if (!strcmp(string, "?") || !strcmp(string, "help")) {
962 log_info("fio: dumping debug options:");
963 for (i = 0; debug_levels[i].name; i++) {
964 dl = &debug_levels[i];
965 log_info("%s,", dl->name);
971 while ((opt = strsep(&p, ",")) != NULL) {
974 if (!strncmp(opt, "all", 3)) {
975 log_info("fio: set all debug options\n");
980 for (i = 0; debug_levels[i].name; i++) {
981 dl = &debug_levels[i];
982 found = !strncmp(opt, dl->name, strlen(dl->name));
986 if (dl->shift == FD_JOB) {
987 opt = strchr(opt, ':');
989 log_err("fio: missing job number\n");
993 fio_debug_jobno = atoi(opt);
994 log_info("fio: set debug jobno %d\n",
997 log_info("fio: set debug option %s\n", opt);
998 fio_debug |= (1UL << dl->shift);
1004 log_err("fio: debug mask %s not found\n", opt);
1009 static int set_debug(const char *string)
1011 log_err("fio: debug tracing not included in build\n");
1016 static int parse_cmd_line(int argc, char *argv[])
1018 struct thread_data *td = NULL;
1019 int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1021 while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1024 smalloc_pool_size = atoi(optarg);
1027 def_timeout = atoi(optarg);
1036 f_out = fopen(optarg, "w+");
1038 perror("fopen output");
1050 exit(fio_show_option_help(optarg));
1058 printf("%s\n", fio_version_string);
1061 if (!strcmp("always", optarg))
1062 eta_print = FIO_ETA_ALWAYS;
1063 else if (!strcmp("never", optarg))
1064 eta_print = FIO_ETA_NEVER;
1067 if (set_debug(optarg))
1071 if (!strcmp(optarg, "global")) {
1072 log_err("fio: can't use global as only "
1080 job_section = strdup(optarg);
1082 case FIO_GETOPT_JOB: {
1083 const char *opt = l_opts[lidx].name;
1086 if (!strncmp(opt, "name", 4) && td) {
1087 ret = add_job(td, td->o.name ?: "fio", 0);
1095 int is_section = !strncmp(opt, "name", 4);
1098 if (!is_section || !strncmp(val, "global", 6))
1101 if (is_section && skip_this_section(val))
1104 td = get_new_job(global, &def_thread);
1109 ret = fio_cmd_option_parse(td, opt, val);
1124 ret = add_job(td, td->o.name ?: "fio", 0);
1129 while (optind < argc) {
1131 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1132 ini_file[ini_idx - 1] = strdup(argv[optind]);
1139 int parse_options(int argc, char *argv[])
1146 fio_options_dup_and_init(l_opts);
1148 if (setup_thread_area())
1150 if (fill_def_thread())
1153 job_files = parse_cmd_line(argc, argv);
1155 for (i = 0; i < job_files; i++) {
1156 if (fill_def_thread())
1158 if (parse_jobs_ini(ini_file[i], i))
1164 options_mem_free(&def_thread);
1166 if (!thread_number) {
1170 log_err("No jobs defined(s)\n\n");
1175 if (def_thread.o.gtod_offload) {
1177 fio_gtod_offload = 1;
1178 fio_gtod_cpu = def_thread.o.gtod_cpu;