10 #include <sys/types.h>
16 #define DEF_TIMEOUT (0)
17 #define DEF_RATE_CYCLE (1000)
18 #define DEF_ODIRECT (1)
19 #define DEF_IO_ENGINE (FIO_SYNCIO)
20 #define DEF_IO_ENGINE_NAME "sync"
21 #define DEF_SEQUENTIAL (1)
22 #define DEF_RAND_REPEAT (1)
23 #define DEF_OVERWRITE (1)
24 #define DEF_CREATE (1)
25 #define DEF_INVALIDATE (1)
26 #define DEF_SYNCIO (0)
27 #define DEF_RANDSEED (0xb1899bedUL)
28 #define DEF_BWAVGTIME (500)
29 #define DEF_CREATE_SER (1)
30 #define DEF_CREATE_FSYNC (1)
32 #define DEF_VERIFY (0)
33 #define DEF_STONEWALL (0)
34 #define DEF_NUMJOBS (1)
35 #define DEF_USE_THREAD (0)
36 #define DEF_FILE_SIZE (1024 * 1024 * 1024UL)
37 #define DEF_ZONE_SIZE (0)
38 #define DEF_ZONE_SKIP (0)
39 #define DEF_RWMIX_CYCLE (500)
40 #define DEF_RWMIX_READ (50)
42 static char fio_version_string[] = "fio 1.3";
44 static int repeatable = DEF_RAND_REPEAT;
45 static char *ini_file;
46 static int max_jobs = MAX_JOBS;
48 struct thread_data def_thread;
49 struct thread_data *threads = NULL;
52 int write_lat_log = 0;
54 int exitall_on_terminate = 0;
55 unsigned long long mlock_size = 0;
57 static int setup_rate(struct thread_data *td)
64 if (td->rate < td->ratemin) {
65 fprintf(stderr, "min rate larger than nominal rate\n");
69 nr_reads_per_sec = (td->rate * 1024) / td->min_bs;
70 td->rate_usec_cycle = 1000000 / nr_reads_per_sec;
71 td->rate_pending_usleep = 0;
75 static void setup_log(struct io_log **log)
77 struct io_log *l = malloc(sizeof(*l));
80 l->max_samples = 1024;
81 l->log = malloc(l->max_samples * sizeof(struct io_sample));
85 void finish_log(struct thread_data *td, struct io_log *log, const char *name)
91 snprintf(file_name, 200, "client%d_%s.log", td->thread_number, name);
92 f = fopen(file_name, "w");
98 for (i = 0; i < log->nr_samples; i++)
99 fprintf(f, "%lu, %lu, %u\n", log->log[i].time, log->log[i].val, log->log[i].ddir);
106 static struct thread_data *get_new_job(int global, struct thread_data *parent)
108 struct thread_data *td;
112 if (thread_number >= max_jobs)
115 td = &threads[thread_number++];
119 memset(td, 0, sizeof(*td));
122 td->thread_number = thread_number;
124 td->ddir = parent->ddir;
125 td->ioprio = parent->ioprio;
126 td->sequential = parent->sequential;
128 td->min_bs = parent->min_bs;
129 td->max_bs = parent->max_bs;
130 td->odirect = parent->odirect;
131 td->thinktime = parent->thinktime;
132 td->fsync_blocks = parent->fsync_blocks;
133 td->start_delay = parent->start_delay;
134 td->timeout = parent->timeout;
135 td->io_engine = parent->io_engine;
136 td->create_file = parent->create_file;
137 td->overwrite = parent->overwrite;
138 td->invalidate_cache = parent->invalidate_cache;
139 td->file_size = parent->file_size;
140 td->file_offset = parent->file_offset;
141 td->zone_size = parent->zone_size;
142 td->zone_skip = parent->zone_skip;
143 td->rate = parent->rate;
144 td->ratemin = parent->ratemin;
145 td->ratecycle = parent->ratecycle;
146 td->iodepth = parent->iodepth;
147 td->sync_io = parent->sync_io;
148 td->mem_type = parent->mem_type;
149 td->bw_avg_time = parent->bw_avg_time;
150 td->create_serialize = parent->create_serialize;
151 td->create_fsync = parent->create_fsync;
152 td->loops = parent->loops;
153 td->verify = parent->verify;
154 td->stonewall = parent->stonewall;
155 td->numjobs = parent->numjobs;
156 td->use_thread = parent->use_thread;
157 td->do_disk_util = parent->do_disk_util;
158 memcpy(&td->cpumask, &parent->cpumask, sizeof(td->cpumask));
159 strcpy(td->io_engine_name, parent->io_engine_name);
164 static void put_job(struct thread_data *td)
166 memset(&threads[td->thread_number - 1], 0, sizeof(*td));
170 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
172 char *ddir_str[] = { "read", "write", "randread", "randwrite",
173 "rw", NULL, "randrw" };
177 #ifndef FIO_HAVE_LIBAIO
178 if (td->io_engine == FIO_LIBAIO) {
179 fprintf(stderr, "Linux libaio not available\n");
183 #ifndef FIO_HAVE_POSIXAIO
184 if (td->io_engine == FIO_POSIXAIO) {
185 fprintf(stderr, "posix aio not available\n");
191 * the def_thread is just for options, it's not a real job
193 if (td == &def_thread)
196 if (td->io_engine & FIO_SYNCIO)
204 * only really works for sequential io for now
206 if (td->zone_size && !td->sequential)
209 td->filetype = FIO_TYPE_FILE;
210 if (!stat(jobname, &sb)) {
211 if (S_ISBLK(sb.st_mode))
212 td->filetype = FIO_TYPE_BD;
213 else if (S_ISCHR(sb.st_mode))
214 td->filetype = FIO_TYPE_CHAR;
217 if (td->filetype == FIO_TYPE_FILE) {
218 if (td->directory[0] != '\0')
219 sprintf(td->file_name, "%s/%s.%d", td->directory, jobname, td->jobnum);
221 sprintf(td->file_name, "%s.%d", jobname, td->jobnum);
223 strncpy(td->file_name, jobname, sizeof(td->file_name) - 1);
225 sem_init(&td->mutex, 0, 0);
227 td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
228 td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
229 td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
231 if (td->min_bs == -1U)
233 if (td->max_bs == -1U)
235 if (td_read(td) && !td_rw(td))
238 if (td->stonewall && td->thread_number > 1)
241 td->groupid = groupid;
247 setup_log(&td->slat_log);
248 setup_log(&td->clat_log);
251 setup_log(&td->bw_log);
253 ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
256 printf("Client%d (g=%d): rw=%s, odir=%d, bs=%d-%d, rate=%d, ioengine=%s, iodepth=%d\n", td->thread_number, td->groupid, ddir_str[ddir], td->odirect, td->min_bs, td->max_bs, td->rate, td->io_engine_name, td->iodepth);
257 else if (job_add_num == 1)
261 * recurse add identical jobs, clear numjobs and stonewall options
262 * as they don't apply to sub-jobs
264 numjobs = td->numjobs;
266 struct thread_data *td_new = get_new_job(0, td);
272 td_new->stonewall = 0;
273 td_new->jobnum = numjobs;
274 job_add_num = numjobs - 1;
276 if (add_job(td_new, jobname, job_add_num))
285 int init_random_state(struct thread_data *td)
287 unsigned long seeds[4];
288 int fd, num_maps, blocks;
290 fd = open("/dev/urandom", O_RDONLY);
292 td_verror(td, errno);
296 if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
304 srand48_r(seeds[0], &td->bsrange_state);
305 srand48_r(seeds[1], &td->verify_state);
306 srand48_r(seeds[2], &td->rwmix_state);
312 seeds[3] = DEF_RANDSEED;
314 blocks = (td->io_size + td->min_bs - 1) / td->min_bs;
315 num_maps = blocks / BLOCKS_PER_MAP;
316 td->file_map = malloc(num_maps * sizeof(long));
317 td->num_maps = num_maps;
318 memset(td->file_map, 0, num_maps * sizeof(long));
320 srand48_r(seeds[3], &td->random_state);
324 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
326 #ifdef FIO_HAVE_CPU_AFFINITY
331 for (i = 0; i < sizeof(int) * 8; i++) {
333 CPU_SET(i, &cpumask);
338 static unsigned long get_mult(char c)
349 return 1024 * 1024 * 1024;
356 * convert string after '=' into decimal value, noting any size suffix
358 static int str_cnv(char *p, unsigned long long *val)
363 str = strchr(p, '=');
370 *val = strtoul(str, NULL, 10);
371 if (*val == ULONG_MAX && errno == ERANGE)
374 *val *= get_mult(str[len - 1]);
378 static int check_strcnv(char *p, char *name, unsigned long long *val)
380 if (strncmp(p, name, strlen(name) - 1))
383 return str_cnv(p, val);
386 static void strip_blank_front(char **p)
394 static void strip_blank_end(char *p)
396 char *s = p + strlen(p) - 1;
398 while (isspace(*s) || iscntrl(*s))
404 typedef int (str_cb_fn)(struct thread_data *, char *);
406 static int check_str(char *p, char *name, str_cb_fn *cb, struct thread_data *td)
408 char *s = strstr(p, name);
418 strip_blank_front(&s);
422 static int check_strstore(char *p, char *name, char *dest)
424 char *s = strstr(p, name);
434 strip_blank_front(&s);
440 static int __check_range(char *str, unsigned long *val)
444 if (sscanf(str, "%lu%c", val, &suffix) == 2) {
445 *val *= get_mult(suffix);
449 if (sscanf(str, "%lu", val) == 1)
455 static int check_range(char *p, char *name, unsigned long *s, unsigned long *e)
463 str = strstr(p, name);
469 str = strchr(p, '=');
474 * 'p' now holds whatever is after the '=' sign
479 * terminate p1 at the '-' sign
488 if (!__check_range(p1, s) && !__check_range(p2, e))
494 static int check_int(char *p, char *name, unsigned int *val)
498 if (strncmp(p, name, strlen(name)))
501 str = strstr(p, name);
505 str = strchr(p, '=');
511 if (sscanf(str, "%u", val) == 1)
517 static int check_strset(char *p, char *name)
519 return strncmp(p, name, strlen(name));
522 static int is_empty_or_comment(char *line)
526 for (i = 0; i < strlen(line); i++) {
529 if (!isspace(line[i]) && !iscntrl(line[i]))
536 static int str_rw_cb(struct thread_data *td, char *mem)
538 if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
539 td->ddir = DDIR_READ;
542 } else if (!strncmp(mem, "randread", 8)) {
543 td->ddir = DDIR_READ;
546 } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
547 td->ddir = DDIR_WRITE;
550 } else if (!strncmp(mem, "randwrite", 9)) {
551 td->ddir = DDIR_WRITE;
554 } else if (!strncmp(mem, "rw", 2)) {
559 } else if (!strncmp(mem, "randrw", 6)) {
566 fprintf(stderr, "bad data direction: %s\n", mem);
570 static int str_verify_cb(struct thread_data *td, char *mem)
572 if (!strncmp(mem, "0", 1)) {
573 td->verify = VERIFY_NONE;
575 } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
576 td->verify = VERIFY_MD5;
578 } else if (!strncmp(mem, "crc32", 5)) {
579 td->verify = VERIFY_CRC32;
583 fprintf(stderr, "bad verify type: %s\n", mem);
587 static int str_mem_cb(struct thread_data *td, char *mem)
589 if (!strncmp(mem, "malloc", 6)) {
590 td->mem_type = MEM_MALLOC;
592 } else if (!strncmp(mem, "shm", 3)) {
593 td->mem_type = MEM_SHM;
595 } else if (!strncmp(mem, "mmap", 4)) {
596 td->mem_type = MEM_MMAP;
600 fprintf(stderr, "bad mem type: %s\n", mem);
604 static int str_ioengine_cb(struct thread_data *td, char *str)
606 if (!strncmp(str, "linuxaio", 8) || !strncmp(str, "aio", 3) ||
607 !strncmp(str, "libaio", 6)) {
608 strcpy(td->io_engine_name, "libaio");
609 td->io_engine = FIO_LIBAIO;
611 } else if (!strncmp(str, "posixaio", 8)) {
612 strcpy(td->io_engine_name, "posixaio");
613 td->io_engine = FIO_POSIXAIO;
615 } else if (!strncmp(str, "sync", 4)) {
616 strcpy(td->io_engine_name, "sync");
617 td->io_engine = FIO_SYNCIO;
619 } else if (!strncmp(str, "mmap", 4)) {
620 strcpy(td->io_engine_name, "mmap");
621 td->io_engine = FIO_MMAPIO;
623 } else if (!strncmp(str, "sgio", 4)) {
624 strcpy(td->io_engine_name, "sgio");
625 td->io_engine = FIO_SGIO;
627 } else if (!strncmp(str, "splice", 6)) {
628 strcpy(td->io_engine_name, "splice");
629 td->io_engine = FIO_SPLICEIO;
633 fprintf(stderr, "bad ioengine type: %s\n", str);
637 static int str_iolog_cb(struct thread_data *td, char *file)
639 strncpy(td->iolog_file, file, sizeof(td->iolog_file) - 1);
644 int parse_jobs_ini(char *file)
646 unsigned int prioclass, prio, cpu, global, il;
647 unsigned long long ull;
648 unsigned long ul1, ul2;
649 struct thread_data *td;
655 f = fopen(file, "r");
657 perror("fopen job file");
661 string = malloc(4096);
664 while ((p = fgets(string, 4096, f)) != NULL) {
665 if (is_empty_or_comment(p))
667 if (sscanf(p, "[%s]", name) != 1)
670 global = !strncmp(name, "global", 6);
672 name[strlen(name) - 1] = '\0';
674 td = get_new_job(global, &def_thread);
679 while ((p = fgets(string, 4096, f)) != NULL) {
680 if (is_empty_or_comment(p))
684 strip_blank_front(&p);
687 if (!check_int(p, "prio", &prio)) {
688 #ifndef FIO_HAVE_IOPRIO
689 fprintf(stderr, "io priorities not available\n");
696 if (!check_int(p, "prioclass", &prioclass)) {
697 #ifndef FIO_HAVE_IOPRIO
698 fprintf(stderr, "io priorities not available\n");
701 td->ioprio |= prioclass << IOPRIO_CLASS_SHIFT;
705 if (!check_int(p, "direct", &td->odirect)) {
709 if (!check_int(p, "rate", &td->rate)) {
713 if (!check_int(p, "ratemin", &td->ratemin)) {
717 if (!check_int(p, "ratecycle", &td->ratecycle)) {
721 if (!check_int(p, "thinktime", &td->thinktime)) {
725 if (!check_int(p, "cpumask", &cpu)) {
726 #ifndef FIO_HAVE_CPU_AFFINITY
727 fprintf(stderr, "cpu affinity not available\n");
730 fill_cpu_mask(td->cpumask, cpu);
734 if (!check_int(p, "fsync", &td->fsync_blocks)) {
739 if (!check_int(p, "startdelay", &td->start_delay)) {
743 if (!check_int(p, "timeout", &td->timeout)) {
747 if (!check_int(p, "invalidate",&td->invalidate_cache)) {
751 if (!check_int(p, "iodepth", &td->iodepth)) {
755 if (!check_int(p, "sync", &td->sync_io)) {
759 if (!check_int(p, "bwavgtime", &td->bw_avg_time)) {
763 if (!check_int(p, "create_serialize", &td->create_serialize)) {
767 if (!check_int(p, "create_fsync", &td->create_fsync)) {
771 if (!check_int(p, "end_fsync", &td->end_fsync)) {
775 if (!check_int(p, "loops", &td->loops)) {
779 if (!check_int(p, "numjobs", &td->numjobs)) {
783 if (!check_int(p, "overwrite", &td->overwrite)) {
787 if (!check_int(p, "rwmixcycle", &td->rwmixcycle)) {
791 if (!check_int(p, "rwmixread", &il)) {
798 if (!check_int(p, "rwmixwrite", &il)) {
801 td->rwmixread = 100 - il;
805 if (!check_range(p, "bsrange", &ul1, &ul2)) {
816 if (!check_strcnv(p, "bs", &ull)) {
821 if (!check_strcnv(p, "size", &td->file_size)) {
825 if (!check_strcnv(p, "offset", &td->file_offset)) {
829 if (!check_strcnv(p, "zonesize", &td->zone_size)) {
833 if (!check_strcnv(p, "zoneskip", &td->zone_skip)) {
837 if (!check_strcnv(p, "lockmem", &mlock_size)) {
841 if (!check_strstore(p, "directory", td->directory)) {
845 if (!check_str(p, "mem", str_mem_cb, td)) {
849 if (!check_str(p, "verify", str_verify_cb, td)) {
853 if (!check_str(p, "rw", str_rw_cb, td)) {
857 if (!check_str(p, "ioengine", str_ioengine_cb, td)) {
861 if (!check_strset(p, "create")) {
866 if (!check_strset(p, "exitall")) {
867 exitall_on_terminate = 1;
871 if (!check_strset(p, "stonewall")) {
876 if (!check_strset(p, "thread")) {
881 if (!check_str(p, "iolog", str_iolog_cb, td)) {
887 printf("Client%d: bad option %s\n",td->thread_number,p);
892 if (add_job(td, name, 0))
902 static int fill_def_thread(void)
904 memset(&def_thread, 0, sizeof(def_thread));
906 if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
907 perror("sched_getaffinity");
914 def_thread.ddir = DDIR_READ;
915 def_thread.iomix = 0;
916 def_thread.bs = DEF_BS;
917 def_thread.min_bs = -1;
918 def_thread.max_bs = -1;
919 def_thread.io_engine = DEF_IO_ENGINE;
920 strcpy(def_thread.io_engine_name, DEF_IO_ENGINE_NAME);
921 def_thread.odirect = DEF_ODIRECT;
922 def_thread.ratecycle = DEF_RATE_CYCLE;
923 def_thread.sequential = DEF_SEQUENTIAL;
924 def_thread.timeout = DEF_TIMEOUT;
925 def_thread.create_file = DEF_CREATE;
926 def_thread.overwrite = DEF_OVERWRITE;
927 def_thread.invalidate_cache = DEF_INVALIDATE;
928 def_thread.sync_io = DEF_SYNCIO;
929 def_thread.mem_type = MEM_MALLOC;
930 def_thread.bw_avg_time = DEF_BWAVGTIME;
931 def_thread.create_serialize = DEF_CREATE_SER;
932 def_thread.create_fsync = DEF_CREATE_FSYNC;
933 def_thread.loops = DEF_LOOPS;
934 def_thread.verify = DEF_VERIFY;
935 def_thread.stonewall = DEF_STONEWALL;
936 def_thread.numjobs = DEF_NUMJOBS;
937 def_thread.use_thread = DEF_USE_THREAD;
938 def_thread.rwmixcycle = DEF_RWMIX_CYCLE;
939 def_thread.rwmixread = DEF_RWMIX_READ;
940 #ifdef FIO_HAVE_DISK_UTIL
941 def_thread.do_disk_util = 1;
947 static void usage(char *name)
949 printf("%s\n", fio_version_string);
950 printf("\t-s IO is sequential\n");
951 printf("\t-b Block size in KiB for each IO\n");
952 printf("\t-t Runtime in seconds\n");
953 printf("\t-R Exit all threads on failure to meet rate goal\n");
954 printf("\t-o Use O_DIRECT\n");
955 printf("\t-l Generate per-job latency logs\n");
956 printf("\t-w Generate per-job bandwidth logs\n");
957 printf("\t-f Job file (Required)\n");
958 printf("\t-v Print version info and exit\n");
961 static void parse_cmd_line(int argc, char *argv[])
965 while ((c = getopt(argc, argv, "s:b:t:r:R:o:f:lwvh")) != EOF) {
968 def_thread.sequential = !!atoi(optarg);
971 def_thread.bs = atoi(optarg);
972 def_thread.bs <<= 10;
973 if (!def_thread.bs) {
974 printf("bad block size\n");
975 def_thread.bs = DEF_BS;
979 def_thread.timeout = atoi(optarg);
982 repeatable = !!atoi(optarg);
985 rate_quit = !!atoi(optarg);
988 def_thread.odirect = !!atoi(optarg);
991 ini_file = strdup(optarg);
1003 printf("%s\n", fio_version_string);
1008 if (!ini_file && argc > 1 && argv[argc - 1][0] != '-')
1009 ini_file = strdup(argv[argc - 1]);
1012 static void free_shm(void)
1014 struct shmid_ds sbuf;
1019 shmctl(shm_id, IPC_RMID, &sbuf);
1023 static int setup_thread_area(void)
1026 * 1024 is too much on some machines, scale max_jobs if
1027 * we get a failure that looks like too large a shm segment
1030 int s = max_jobs * sizeof(struct thread_data);
1032 shm_id = shmget(0, s, IPC_CREAT | 0600);
1035 if (errno != EINVAL) {
1046 threads = shmat(shm_id, NULL, 0);
1047 if (threads == (void *) -1) {
1056 int parse_options(int argc, char *argv[])
1058 if (setup_thread_area())
1060 if (fill_def_thread())
1063 parse_cmd_line(argc, argv);
1066 printf("Need job file\n");
1071 if (parse_jobs_ini(ini_file)) {