14 #include <sys/types.h>
16 #include <sys/ioctl.h>
17 #include <sys/syscall.h>
18 #include <sys/resource.h>
28 #include "../arch/arch.h"
29 #include "../lib/types.h"
30 #include "../lib/roundup.h"
31 #include "../lib/rand.h"
32 #include "../minmax.h"
33 #include "../os/linux/io_uring.h"
39 unsigned *ring_entries;
48 unsigned *ring_entries;
49 struct io_uring_cqe *cqes;
53 #define BATCH_SUBMIT 32
54 #define BATCH_COMPLETE 32
59 static unsigned sq_ring_mask, cq_ring_mask;
62 unsigned long max_blocks;
63 unsigned long max_size;
64 unsigned long cur_off;
72 #define PLAT_VAL (1 << PLAT_BITS)
73 #define PLAT_GROUP_NR 29
74 #define PLAT_NR (PLAT_GROUP_NR * PLAT_VAL)
80 struct io_sq_ring sq_ring;
81 struct io_uring_sqe *sqes;
82 struct io_cq_ring cq_ring;
92 struct taus258_state rand_state;
94 unsigned long *clock_batch;
102 struct file files[MAX_FDS];
105 struct iovec iovecs[];
108 static struct submitter *submitter;
109 static volatile int finish;
110 static int stats_running;
111 static unsigned long max_iops;
113 static int depth = DEPTH;
114 static int batch_submit = BATCH_SUBMIT;
115 static int batch_complete = BATCH_COMPLETE;
117 static int polled = 1; /* use IO polling */
118 static int fixedbufs = 1; /* use fixed user buffers */
119 static int dma_map; /* pre-map DMA buffers */
120 static int register_files = 1; /* use fixed files */
121 static int buffered = 0; /* use buffered IO, not O_DIRECT */
122 static int sq_thread_poll = 0; /* use kernel submission/poller thread */
123 static int sq_thread_cpu = -1; /* pin above thread to this CPU */
124 static int do_nop = 0; /* no-op SQ ring commands */
125 static int nthreads = 1;
126 static int stats = 0; /* generate IO stats */
127 static int aio = 0; /* use libaio */
128 static int runtime = 0; /* runtime */
129 static int random_io = 1; /* random or sequential IO */
131 static unsigned long tsc_rate;
133 #define TSC_RATE_FILE "tsc-rate"
135 static int vectored = 1;
137 static float plist[] = { 1.0, 5.0, 10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0,
138 80.0, 90.0, 95.0, 99.0, 99.5, 99.9, 99.95, 99.99 };
139 static int plist_len = 17;
141 #ifndef IORING_REGISTER_MAP_BUFFERS
142 #define IORING_REGISTER_MAP_BUFFERS 20
143 struct io_uring_map_buffers {
152 static unsigned long cycles_to_nsec(unsigned long cycles)
159 val = cycles * 1000000000ULL;
160 return val / tsc_rate;
163 static unsigned long plat_idx_to_val(unsigned int idx)
165 unsigned int error_bits;
166 unsigned long k, base;
168 assert(idx < PLAT_NR);
170 /* MSB <= (PLAT_BITS-1), cannot be rounded off. Use
171 * all bits of the sample as index */
172 if (idx < (PLAT_VAL << 1))
173 return cycles_to_nsec(idx);
175 /* Find the group and compute the minimum value of that group */
176 error_bits = (idx >> PLAT_BITS) - 1;
177 base = ((unsigned long) 1) << (error_bits + PLAT_BITS);
179 /* Find its bucket number of the group */
182 /* Return the mean of the range of the bucket */
183 return cycles_to_nsec(base + ((k + 0.5) * (1 << error_bits)));
186 unsigned int calc_clat_percentiles(unsigned long *io_u_plat, unsigned long nr,
187 unsigned long **output,
188 unsigned long *maxv, unsigned long *minv)
190 unsigned long sum = 0;
191 unsigned int len = plist_len, i, j = 0;
192 unsigned long *ovals = NULL;
198 ovals = malloc(len * sizeof(*ovals));
203 * Calculate bucket values, note down max and min values
206 for (i = 0; i < PLAT_NR && !is_last; i++) {
208 while (sum >= ((long double) plist[j] / 100.0 * nr)) {
209 assert(plist[j] <= 100.0);
211 ovals[j] = plat_idx_to_val(i);
212 if (ovals[j] < *minv)
214 if (ovals[j] > *maxv)
217 is_last = (j == len - 1) != 0;
226 fprintf(stderr, "error calculating latency percentiles\n");
232 static void show_clat_percentiles(unsigned long *io_u_plat, unsigned long nr,
233 unsigned int precision)
235 unsigned int divisor, len, i, j = 0;
236 unsigned long minv, maxv;
237 unsigned long *ovals;
238 int per_line, scale_down, time_width;
242 len = calc_clat_percentiles(io_u_plat, nr, &ovals, &maxv, &minv);
249 printf(" percentiles (tsc ticks):\n |");
250 } else if (minv > 2000 && maxv > 99999) {
253 printf(" percentiles (usec):\n |");
257 printf(" percentiles (nsec):\n |");
260 time_width = max(5, (int) (log10(maxv / divisor) + 1));
261 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
262 precision, time_width);
263 /* fmt will be something like " %5.2fth=[%4llu]%c" */
264 per_line = (80 - 7) / (precision + 10 + time_width);
266 for (j = 0; j < len; j++) {
268 if (j != 0 && (j % per_line) == 0)
271 /* end of the list */
272 is_last = (j == len - 1) != 0;
274 for (i = 0; i < scale_down; i++)
275 ovals[j] = (ovals[j] + 999) / 1000;
277 printf(fmt, plist[j], ovals[j], is_last ? '\n' : ',');
282 if ((j % per_line) == per_line - 1) /* for formatting */
290 static unsigned int plat_val_to_idx(unsigned long val)
292 unsigned int msb, error_bits, base, offset, idx;
294 /* Find MSB starting from bit 0 */
298 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
301 * MSB <= (PLAT_BITS-1), cannot be rounded off. Use
302 * all bits of the sample as index
304 if (msb <= PLAT_BITS)
307 /* Compute the number of error bits to discard*/
308 error_bits = msb - PLAT_BITS;
310 /* Compute the number of buckets before the group */
311 base = (error_bits + 1) << PLAT_BITS;
314 * Discard the error bits and apply the mask to find the
315 * index for the buckets in the group
317 offset = (PLAT_VAL - 1) & (val >> error_bits);
319 /* Make sure the index does not exceed (array size - 1) */
320 idx = (base + offset) < (PLAT_NR - 1) ?
321 (base + offset) : (PLAT_NR - 1);
326 static void add_stat(struct submitter *s, int clock_index, int nr)
328 #ifdef ARCH_HAVE_CPU_CLOCK
329 unsigned long cycles;
332 if (!s->finish && clock_index) {
333 cycles = get_cpu_clock();
334 cycles -= s->clock_batch[clock_index];
335 pidx = plat_val_to_idx(cycles);
341 static int io_uring_map_buffers(struct submitter *s)
343 struct io_uring_map_buffers map = {
344 .fd = s->files[0].real_fd,
350 if (s->nr_files > 1) {
351 fprintf(stderr, "Can't map buffers with multiple files\n");
355 return syscall(__NR_io_uring_register, s->ring_fd,
356 IORING_REGISTER_MAP_BUFFERS, &map, 1);
359 static int io_uring_register_buffers(struct submitter *s)
364 return syscall(__NR_io_uring_register, s->ring_fd,
365 IORING_REGISTER_BUFFERS, s->iovecs, depth);
368 static int io_uring_register_files(struct submitter *s)
375 s->fds = calloc(s->nr_files, sizeof(__s32));
376 for (i = 0; i < s->nr_files; i++) {
377 s->fds[i] = s->files[i].real_fd;
378 s->files[i].fixed_fd = i;
381 return syscall(__NR_io_uring_register, s->ring_fd,
382 IORING_REGISTER_FILES, s->fds, s->nr_files);
385 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
387 return syscall(__NR_io_uring_setup, entries, p);
390 static void io_uring_probe(int fd)
392 struct io_uring_probe *p;
395 p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
399 memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
400 ret = syscall(__NR_io_uring_register, fd, IORING_REGISTER_PROBE, p, 256);
404 if (IORING_OP_READ > p->ops_len)
407 if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED))
413 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
414 unsigned int min_complete, unsigned int flags)
416 return syscall(__NR_io_uring_enter, s->ring_fd, to_submit, min_complete,
420 #ifndef CONFIG_HAVE_GETTID
421 static int gettid(void)
423 return syscall(__NR_gettid);
427 static unsigned file_depth(struct submitter *s)
429 return (depth + s->nr_files - 1) / s->nr_files;
432 static void init_io(struct submitter *s, unsigned index)
434 struct io_uring_sqe *sqe = &s->sqes[index];
435 unsigned long offset;
440 sqe->opcode = IORING_OP_NOP;
444 if (s->nr_files == 1) {
447 f = &s->files[s->cur_file];
448 if (f->pending_ios >= file_depth(s)) {
450 if (s->cur_file == s->nr_files)
452 f = &s->files[s->cur_file];
458 r = __rand64(&s->rand_state);
459 offset = (r % (f->max_blocks - 1)) * bs;
463 if (f->cur_off + bs > f->max_size)
467 if (register_files) {
468 sqe->flags = IOSQE_FIXED_FILE;
469 sqe->fd = f->fixed_fd;
472 sqe->fd = f->real_fd;
475 sqe->opcode = IORING_OP_READ_FIXED;
476 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
478 sqe->buf_index = index;
479 } else if (!vectored) {
480 sqe->opcode = IORING_OP_READ;
481 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
485 sqe->opcode = IORING_OP_READV;
486 sqe->addr = (unsigned long) &s->iovecs[index];
492 sqe->user_data = (unsigned long) f->fileno;
493 if (stats && stats_running)
494 sqe->user_data |= ((unsigned long)s->clock_index << 32);
497 static int prep_more_ios_uring(struct submitter *s, int max_ios)
499 struct io_sq_ring *ring = &s->sq_ring;
500 unsigned index, tail, next_tail, prepped = 0;
502 next_tail = tail = *ring->tail;
505 if (next_tail == atomic_load_acquire(ring->head))
508 index = tail & sq_ring_mask;
510 ring->array[index] = index;
513 } while (prepped < max_ios);
516 atomic_store_release(ring->tail, tail);
520 static int get_file_size(struct file *f)
524 if (fstat(f->real_fd, &st) < 0)
526 if (S_ISBLK(st.st_mode)) {
527 unsigned long long bytes;
529 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
532 f->max_blocks = bytes / bs;
535 } else if (S_ISREG(st.st_mode)) {
536 f->max_blocks = st.st_size / bs;
537 f->max_size = st.st_size;
544 static int reap_events_uring(struct submitter *s)
546 struct io_cq_ring *ring = &s->cq_ring;
547 struct io_uring_cqe *cqe;
548 unsigned head, reaped = 0;
549 int last_idx = -1, stat_nr = 0;
556 if (head == atomic_load_acquire(ring->tail))
558 cqe = &ring->cqes[head & cq_ring_mask];
560 int fileno = cqe->user_data & 0xffffffff;
562 f = &s->files[fileno];
564 if (cqe->res != bs) {
565 printf("io: unexpected ret=%d\n", cqe->res);
566 if (polled && cqe->res == -EOPNOTSUPP)
567 printf("Your filesystem/driver/kernel doesn't support polled IO\n");
572 int clock_index = cqe->user_data >> 32;
574 if (last_idx != clock_index) {
575 if (last_idx != -1) {
576 add_stat(s, last_idx, stat_nr);
579 last_idx = clock_index;
588 add_stat(s, last_idx, stat_nr);
591 s->inflight -= reaped;
592 atomic_store_release(ring->head, head);
597 static int submitter_init(struct submitter *s)
602 printf("submitter=%d, tid=%d\n", s->index, s->tid);
604 __init_rand64(&s->rand_state, pthread_self());
605 srand48(pthread_self());
607 for (i = 0; i < MAX_FDS; i++)
608 s->files[i].fileno = i;
611 nr_batch = roundup_pow2(depth / batch_submit);
614 s->clock_batch = calloc(nr_batch, sizeof(unsigned long));
617 s->plat = calloc(PLAT_NR, sizeof(unsigned long));
619 s->clock_batch = NULL;
628 static int prep_more_ios_aio(struct submitter *s, int max_ios, struct iocb *iocbs)
630 unsigned long offset, data;
636 while (index < max_ios) {
637 struct iocb *iocb = &iocbs[index];
639 if (s->nr_files == 1) {
642 f = &s->files[s->cur_file];
643 if (f->pending_ios >= file_depth(s)) {
645 if (s->cur_file == s->nr_files)
647 f = &s->files[s->cur_file];
653 offset = (r % (f->max_blocks - 1)) * bs;
654 io_prep_pread(iocb, f->real_fd, s->iovecs[index].iov_base,
655 s->iovecs[index].iov_len, offset);
658 if (stats && stats_running)
659 data |= ((unsigned long) s->clock_index << 32);
660 iocb->data = (void *) (uintptr_t) data;
666 static int reap_events_aio(struct submitter *s, struct io_event *events, int evs)
668 int last_idx = -1, stat_nr = 0;
672 unsigned long data = (uintptr_t) events[reaped].data;
673 struct file *f = &s->files[data & 0xffffffff];
676 if (events[reaped].res != bs) {
677 printf("io: unexpected ret=%ld\n", events[reaped].res);
681 int clock_index = data >> 32;
683 if (last_idx != clock_index) {
684 if (last_idx != -1) {
685 add_stat(s, last_idx, stat_nr);
688 last_idx = clock_index;
697 add_stat(s, last_idx, stat_nr);
699 s->inflight -= reaped;
704 static void *submitter_aio_fn(void *data)
706 struct submitter *s = data;
707 int i, ret, prepped, nr_batch;
708 struct iocb **iocbsptr;
710 struct io_event *events;
712 nr_batch = submitter_init(s);
714 iocbsptr = calloc(depth, sizeof(struct iocb *));
715 iocbs = calloc(depth, sizeof(struct iocb));
716 events = calloc(depth, sizeof(struct io_event));
718 for (i = 0; i < depth; i++)
719 iocbsptr[i] = &iocbs[i];
723 int to_wait, to_submit, to_prep;
725 if (!prepped && s->inflight < depth) {
726 to_prep = min(depth - s->inflight, batch_submit);
727 prepped = prep_more_ios_aio(s, to_prep, iocbs);
728 #ifdef ARCH_HAVE_CPU_CLOCK
729 if (prepped && stats) {
730 s->clock_batch[s->clock_index] = get_cpu_clock();
731 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
735 s->inflight += prepped;
738 if (to_submit && (s->inflight + to_submit <= depth))
741 to_wait = min(s->inflight + to_submit, batch_complete);
743 ret = io_submit(s->aio_ctx, to_submit, iocbsptr);
748 } else if (ret != to_submit) {
749 printf("submitted %d, wanted %d\n", ret, to_submit);
758 r = io_getevents(s->aio_ctx, to_wait, to_wait, events, NULL);
760 perror("io_getevents");
762 } else if (r != to_wait) {
763 printf("r=%d, wait=%d\n", r, to_wait);
766 r = reap_events_aio(s, events, r);
770 } while (!s->finish);
780 static void *submitter_uring_fn(void *data)
782 struct submitter *s = data;
783 struct io_sq_ring *ring = &s->sq_ring;
784 int ret, prepped, nr_batch;
786 nr_batch = submitter_init(s);
790 int to_wait, to_submit, this_reap, to_prep;
791 unsigned ring_flags = 0;
793 if (!prepped && s->inflight < depth) {
794 to_prep = min(depth - s->inflight, batch_submit);
795 prepped = prep_more_ios_uring(s, to_prep);
796 #ifdef ARCH_HAVE_CPU_CLOCK
797 if (prepped && stats) {
798 s->clock_batch[s->clock_index] = get_cpu_clock();
799 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
803 s->inflight += prepped;
807 if (to_submit && (s->inflight + to_submit <= depth))
810 to_wait = min(s->inflight + to_submit, batch_complete);
813 * Only need to call io_uring_enter if we're not using SQ thread
814 * poll, or if IORING_SQ_NEED_WAKEUP is set.
817 ring_flags = atomic_load_acquire(ring->flags);
818 if (!sq_thread_poll || ring_flags & IORING_SQ_NEED_WAKEUP) {
822 flags = IORING_ENTER_GETEVENTS;
823 if (ring_flags & IORING_SQ_NEED_WAKEUP)
824 flags |= IORING_ENTER_SQ_WAKEUP;
825 ret = io_uring_enter(s, to_submit, to_wait, flags);
828 /* for SQPOLL, we submitted it all effectively */
833 * For non SQ thread poll, we already got the events we needed
834 * through the io_uring_enter() above. For SQ thread poll, we
835 * need to loop here until we find enough events.
841 r = reap_events_uring(s);
847 } while (sq_thread_poll && this_reap < to_wait);
848 s->reaps += this_reap;
856 } else if (ret < to_submit) {
857 int diff = to_submit - ret;
866 } else if (ret < 0) {
867 if (errno == EAGAIN) {
875 printf("io_submit: %s\n", strerror(errno));
878 } while (!s->finish);
884 static struct submitter *get_submitter(int offset)
890 ret += offset * (sizeof(*submitter) + depth * sizeof(struct iovec));
894 static void do_finish(const char *reason)
897 printf("Exiting on %s\n", reason);
898 for (j = 0; j < nthreads; j++) {
899 struct submitter *s = get_submitter(j);
902 if (max_iops > 100000)
903 printf("Maximum IOPS=%luK\n", max_iops / 1000);
905 printf("Maximum IOPS=%lu\n", max_iops);
909 static void sig_int(int sig)
914 static void arm_sig_int(void)
916 struct sigaction act;
918 memset(&act, 0, sizeof(act));
919 act.sa_handler = sig_int;
920 act.sa_flags = SA_RESTART;
921 sigaction(SIGINT, &act, NULL);
923 /* Windows uses SIGBREAK as a quit signal from other applications */
925 sigaction(SIGBREAK, &act, NULL);
929 static int setup_aio(struct submitter *s)
933 fprintf(stderr, "aio does not support polled IO\n");
936 if (sq_thread_poll) {
937 fprintf(stderr, "aio does not support SQPOLL IO\n");
941 fprintf(stderr, "aio does not support polled IO\n");
944 if (fixedbufs || register_files) {
945 fprintf(stderr, "aio does not support registered files or buffers\n");
946 fixedbufs = register_files = 0;
949 return io_queue_init(depth, &s->aio_ctx);
951 fprintf(stderr, "Legacy AIO not available on this system/build\n");
957 static int setup_ring(struct submitter *s)
959 struct io_sq_ring *sring = &s->sq_ring;
960 struct io_cq_ring *cring = &s->cq_ring;
961 struct io_uring_params p;
965 memset(&p, 0, sizeof(p));
967 if (polled && !do_nop)
968 p.flags |= IORING_SETUP_IOPOLL;
969 if (sq_thread_poll) {
970 p.flags |= IORING_SETUP_SQPOLL;
971 if (sq_thread_cpu != -1) {
972 p.flags |= IORING_SETUP_SQ_AFF;
973 p.sq_thread_cpu = sq_thread_cpu;
977 fd = io_uring_setup(depth, &p);
979 perror("io_uring_setup");
989 rlim.rlim_cur = RLIM_INFINITY;
990 rlim.rlim_max = RLIM_INFINITY;
991 /* ignore potential error, not needed on newer kernels */
992 setrlimit(RLIMIT_MEMLOCK, &rlim);
994 ret = io_uring_register_buffers(s);
996 perror("io_uring_register_buffers");
1001 ret = io_uring_map_buffers(s);
1003 perror("io_uring_map_buffers");
1009 if (register_files) {
1010 ret = io_uring_register_files(s);
1012 perror("io_uring_register_files");
1017 ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
1018 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1019 IORING_OFF_SQ_RING);
1020 sring->head = ptr + p.sq_off.head;
1021 sring->tail = ptr + p.sq_off.tail;
1022 sring->ring_mask = ptr + p.sq_off.ring_mask;
1023 sring->ring_entries = ptr + p.sq_off.ring_entries;
1024 sring->flags = ptr + p.sq_off.flags;
1025 sring->array = ptr + p.sq_off.array;
1026 sq_ring_mask = *sring->ring_mask;
1028 s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
1029 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1032 ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
1033 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1034 IORING_OFF_CQ_RING);
1035 cring->head = ptr + p.cq_off.head;
1036 cring->tail = ptr + p.cq_off.tail;
1037 cring->ring_mask = ptr + p.cq_off.ring_mask;
1038 cring->ring_entries = ptr + p.cq_off.ring_entries;
1039 cring->cqes = ptr + p.cq_off.cqes;
1040 cq_ring_mask = *cring->ring_mask;
1044 static void file_depths(char *buf)
1052 for (j = 0; j < nthreads; j++) {
1053 struct submitter *s = get_submitter(j);
1055 for (i = 0; i < s->nr_files; i++) {
1056 struct file *f = &s->files[i];
1059 p += sprintf(p, " %d", f->pending_ios);
1061 p += sprintf(p, "%d", f->pending_ios);
1067 static void usage(char *argv, int status)
1069 char runtime_str[16];
1070 snprintf(runtime_str, sizeof(runtime_str), "%d", runtime);
1071 printf("%s [options] -- [filenames]\n"
1072 " -d <int> : IO Depth, default %d\n"
1073 " -s <int> : Batch submit, default %d\n"
1074 " -c <int> : Batch complete, default %d\n"
1075 " -b <int> : Block size, default %d\n"
1076 " -p <bool> : Polled IO, default %d\n"
1077 " -B <bool> : Fixed buffers, default %d\n"
1078 " -D <bool> : DMA map fixed buffers, default %d\n"
1079 " -F <bool> : Register files, default %d\n"
1080 " -n <int> : Number of threads, default %d\n"
1081 " -O <bool> : Use O_DIRECT, default %d\n"
1082 " -N <bool> : Perform just no-op requests, default %d\n"
1083 " -t <bool> : Track IO latencies, default %d\n"
1084 " -T <int> : TSC rate in HZ\n"
1085 " -r <int> : Runtime in seconds, default %s\n"
1086 " -R <bool> : Use random IO, default %d\n"
1087 " -a <bool> : Use legacy aio, default %d\n",
1088 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled,
1089 fixedbufs, dma_map, register_files, nthreads, !buffered, do_nop,
1090 stats, runtime == 0 ? "unlimited" : runtime_str, aio, random_io);
1094 static void read_tsc_rate(void)
1102 fd = open(TSC_RATE_FILE, O_RDONLY);
1106 ret = read(fd, buffer, sizeof(buffer));
1112 tsc_rate = strtoul(buffer, NULL, 10);
1113 printf("Using TSC rate %luHz\n", tsc_rate);
1117 static void write_tsc_rate(void)
1123 if (!stat(TSC_RATE_FILE, &sb))
1126 fd = open(TSC_RATE_FILE, O_WRONLY | O_CREAT, 0644);
1130 memset(buffer, 0, sizeof(buffer));
1131 sprintf(buffer, "%lu", tsc_rate);
1132 ret = write(fd, buffer, strlen(buffer));
1138 int main(int argc, char *argv[])
1140 struct submitter *s;
1141 unsigned long done, calls, reap;
1142 int err, i, j, flags, fd, opt, threads_per_f, threads_rem = 0, nfiles;
1147 if (!do_nop && argc < 2)
1150 while ((opt = getopt(argc, argv, "d:s:c:b:p:B:F:n:N:O:t:T:a:r:D:R:h?")) != -1) {
1153 aio = !!atoi(optarg);
1156 depth = atoi(optarg);
1159 batch_submit = atoi(optarg);
1164 batch_complete = atoi(optarg);
1165 if (!batch_complete)
1172 polled = !!atoi(optarg);
1175 fixedbufs = !!atoi(optarg);
1178 register_files = !!atoi(optarg);
1181 nthreads = atoi(optarg);
1183 printf("Threads must be non-zero\n");
1188 do_nop = !!atoi(optarg);
1191 buffered = !atoi(optarg);
1194 #ifndef ARCH_HAVE_CPU_CLOCK
1195 fprintf(stderr, "Stats not supported on this CPU\n");
1198 stats = !!atoi(optarg);
1201 #ifndef ARCH_HAVE_CPU_CLOCK
1202 fprintf(stderr, "Stats not supported on this CPU\n");
1205 tsc_rate = strtoul(optarg, NULL, 10);
1209 runtime = atoi(optarg);
1212 dma_map = !!atoi(optarg);
1215 random_io = !!atoi(optarg);
1228 if (batch_complete > depth)
1229 batch_complete = depth;
1230 if (batch_submit > depth)
1231 batch_submit = depth;
1232 if (!fixedbufs && dma_map)
1235 submitter = calloc(nthreads, sizeof(*submitter) +
1236 depth * sizeof(struct iovec));
1237 for (j = 0; j < nthreads; j++) {
1238 s = get_submitter(j);
1240 s->done = s->calls = s->reaps = 0;
1243 flags = O_RDONLY | O_NOATIME;
1252 printf("No files specified\n");
1255 threads_per_f = nthreads / nfiles;
1256 /* make sure each thread gets assigned files */
1257 if (threads_per_f == 0) {
1260 threads_rem = nthreads - threads_per_f * nfiles;
1263 while (!do_nop && i < argc) {
1266 memset(&f, 0, sizeof(f));
1268 fd = open(argv[i], flags);
1274 if (get_file_size(&f)) {
1275 printf("failed getting size of device/file\n");
1278 if (f.max_blocks <= 1) {
1279 printf("Zero file/device size?\n");
1284 limit = threads_per_f;
1285 limit += threads_rem > 0 ? 1 : 0;
1286 for (k = 0; k < limit; k++) {
1287 s = get_submitter((j + k) % nthreads);
1289 if (s->nr_files == MAX_FDS) {
1290 printf("Max number of files (%d) reached\n", MAX_FDS);
1294 memcpy(&s->files[s->nr_files], &f, sizeof(f));
1296 printf("Added file %s (submitter %d)\n", argv[i], s->index);
1306 for (j = 0; j < nthreads; j++) {
1307 s = get_submitter(j);
1308 for (i = 0; i < depth; i++) {
1311 if (posix_memalign(&buf, bs, bs)) {
1312 printf("failed alloc\n");
1315 s->iovecs[i].iov_base = buf;
1316 s->iovecs[i].iov_len = bs;
1320 for (j = 0; j < nthreads; j++) {
1321 s = get_submitter(j);
1324 err = setup_ring(s);
1328 printf("ring setup failed: %s, %d\n", strerror(errno), err);
1332 s = get_submitter(0);
1333 printf("polled=%d, fixedbufs=%d/%d, register_files=%d, buffered=%d, QD=%d\n", polled, fixedbufs, dma_map, register_files, buffered, depth);
1335 printf("Engine=io_uring, sq_ring=%d, cq_ring=%d\n", *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
1337 printf("Engine=aio\n");
1339 for (j = 0; j < nthreads; j++) {
1340 s = get_submitter(j);
1342 pthread_create(&s->thread, NULL, submitter_uring_fn, s);
1343 #ifdef CONFIG_LIBAIO
1345 pthread_create(&s->thread, NULL, submitter_aio_fn, s);
1349 fdepths = malloc(8 * s->nr_files * nthreads);
1350 reap = calls = done = 0;
1352 unsigned long this_done = 0;
1353 unsigned long this_reap = 0;
1354 unsigned long this_call = 0;
1355 unsigned long rpc = 0, ipc = 0;
1356 unsigned long iops, bw;
1359 if (runtime && !--runtime)
1360 do_finish("timeout");
1362 /* don't print partial run, if interrupted by signal */
1366 /* one second in to the run, enable stats */
1370 for (j = 0; j < nthreads; j++) {
1371 s = get_submitter(j);
1372 this_done += s->done;
1373 this_call += s->calls;
1374 this_reap += s->reaps;
1376 if (this_call - calls) {
1377 rpc = (this_done - done) / (this_call - calls);
1378 ipc = (this_reap - reap) / (this_call - calls);
1381 file_depths(fdepths);
1382 iops = this_done - done;
1384 bw = iops * (bs / 1048576);
1386 bw = iops / (1048576 / bs);
1388 printf("IOPS=%luK, ", iops / 1000);
1390 printf("IOPS=%lu, ", iops);
1391 max_iops = max(max_iops, iops);
1393 printf("BW=%luMiB/s, ", bw);
1394 printf("IOS/call=%ld/%ld, inflight=(%s)\n", rpc, ipc, fdepths);
1400 for (j = 0; j < nthreads; j++) {
1401 s = get_submitter(j);
1402 pthread_join(s->thread, &ret);
1408 printf("%d: Latency percentiles:\n", s->tid);
1409 for (i = 0, nr = 0; i < PLAT_NR; i++)
1411 show_clat_percentiles(s->plat, nr, 4);
1412 free(s->clock_batch);