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;
70 #define PLAT_VAL (1 << PLAT_BITS)
71 #define PLAT_GROUP_NR 29
72 #define PLAT_NR (PLAT_GROUP_NR * PLAT_VAL)
78 struct io_sq_ring sq_ring;
79 struct io_uring_sqe *sqes;
80 struct io_cq_ring cq_ring;
90 struct taus258_state rand_state;
92 unsigned long *clock_batch;
100 struct file files[MAX_FDS];
103 struct iovec iovecs[];
106 static struct submitter *submitter;
107 static volatile int finish;
108 static int stats_running;
109 static unsigned long max_iops;
111 static int depth = DEPTH;
112 static int batch_submit = BATCH_SUBMIT;
113 static int batch_complete = BATCH_COMPLETE;
115 static int polled = 1; /* use IO polling */
116 static int fixedbufs = 1; /* use fixed user buffers */
117 static int dma_map; /* pre-map DMA buffers */
118 static int register_files = 1; /* use fixed files */
119 static int buffered = 0; /* use buffered IO, not O_DIRECT */
120 static int sq_thread_poll = 0; /* use kernel submission/poller thread */
121 static int sq_thread_cpu = -1; /* pin above thread to this CPU */
122 static int do_nop = 0; /* no-op SQ ring commands */
123 static int nthreads = 1;
124 static int stats = 0; /* generate IO stats */
125 static int aio = 0; /* use libaio */
126 static int runtime = 0; /* runtime */
128 static unsigned long tsc_rate;
130 #define TSC_RATE_FILE "tsc-rate"
132 static int vectored = 1;
134 static float plist[] = { 1.0, 5.0, 10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0,
135 80.0, 90.0, 95.0, 99.0, 99.5, 99.9, 99.95, 99.99 };
136 static int plist_len = 17;
138 #ifndef IORING_REGISTER_MAP_BUFFERS
139 #define IORING_REGISTER_MAP_BUFFERS 20
140 struct io_uring_map_buffers {
149 static unsigned long cycles_to_nsec(unsigned long cycles)
156 val = cycles * 1000000000ULL;
157 return val / tsc_rate;
160 static unsigned long plat_idx_to_val(unsigned int idx)
162 unsigned int error_bits;
163 unsigned long k, base;
165 assert(idx < PLAT_NR);
167 /* MSB <= (PLAT_BITS-1), cannot be rounded off. Use
168 * all bits of the sample as index */
169 if (idx < (PLAT_VAL << 1))
170 return cycles_to_nsec(idx);
172 /* Find the group and compute the minimum value of that group */
173 error_bits = (idx >> PLAT_BITS) - 1;
174 base = ((unsigned long) 1) << (error_bits + PLAT_BITS);
176 /* Find its bucket number of the group */
179 /* Return the mean of the range of the bucket */
180 return cycles_to_nsec(base + ((k + 0.5) * (1 << error_bits)));
183 unsigned int calc_clat_percentiles(unsigned long *io_u_plat, unsigned long nr,
184 unsigned long **output,
185 unsigned long *maxv, unsigned long *minv)
187 unsigned long sum = 0;
188 unsigned int len = plist_len, i, j = 0;
189 unsigned long *ovals = NULL;
195 ovals = malloc(len * sizeof(*ovals));
200 * Calculate bucket values, note down max and min values
203 for (i = 0; i < PLAT_NR && !is_last; i++) {
205 while (sum >= ((long double) plist[j] / 100.0 * nr)) {
206 assert(plist[j] <= 100.0);
208 ovals[j] = plat_idx_to_val(i);
209 if (ovals[j] < *minv)
211 if (ovals[j] > *maxv)
214 is_last = (j == len - 1) != 0;
223 fprintf(stderr, "error calculating latency percentiles\n");
229 static void show_clat_percentiles(unsigned long *io_u_plat, unsigned long nr,
230 unsigned int precision)
232 unsigned int divisor, len, i, j = 0;
233 unsigned long minv, maxv;
234 unsigned long *ovals;
235 int per_line, scale_down, time_width;
239 len = calc_clat_percentiles(io_u_plat, nr, &ovals, &maxv, &minv);
246 printf(" percentiles (tsc ticks):\n |");
247 } else if (minv > 2000 && maxv > 99999) {
250 printf(" percentiles (usec):\n |");
254 printf(" percentiles (nsec):\n |");
257 time_width = max(5, (int) (log10(maxv / divisor) + 1));
258 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
259 precision, time_width);
260 /* fmt will be something like " %5.2fth=[%4llu]%c" */
261 per_line = (80 - 7) / (precision + 10 + time_width);
263 for (j = 0; j < len; j++) {
265 if (j != 0 && (j % per_line) == 0)
268 /* end of the list */
269 is_last = (j == len - 1) != 0;
271 for (i = 0; i < scale_down; i++)
272 ovals[j] = (ovals[j] + 999) / 1000;
274 printf(fmt, plist[j], ovals[j], is_last ? '\n' : ',');
279 if ((j % per_line) == per_line - 1) /* for formatting */
287 static unsigned int plat_val_to_idx(unsigned long val)
289 unsigned int msb, error_bits, base, offset, idx;
291 /* Find MSB starting from bit 0 */
295 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
298 * MSB <= (PLAT_BITS-1), cannot be rounded off. Use
299 * all bits of the sample as index
301 if (msb <= PLAT_BITS)
304 /* Compute the number of error bits to discard*/
305 error_bits = msb - PLAT_BITS;
307 /* Compute the number of buckets before the group */
308 base = (error_bits + 1) << PLAT_BITS;
311 * Discard the error bits and apply the mask to find the
312 * index for the buckets in the group
314 offset = (PLAT_VAL - 1) & (val >> error_bits);
316 /* Make sure the index does not exceed (array size - 1) */
317 idx = (base + offset) < (PLAT_NR - 1) ?
318 (base + offset) : (PLAT_NR - 1);
323 static void add_stat(struct submitter *s, int clock_index, int nr)
325 #ifdef ARCH_HAVE_CPU_CLOCK
326 unsigned long cycles;
329 if (!s->finish && clock_index) {
330 cycles = get_cpu_clock();
331 cycles -= s->clock_batch[clock_index];
332 pidx = plat_val_to_idx(cycles);
338 static int io_uring_map_buffers(struct submitter *s)
340 struct io_uring_map_buffers map = {
341 .fd = s->files[0].real_fd,
347 if (s->nr_files > 1) {
348 fprintf(stderr, "Can't map buffers with multiple files\n");
352 return syscall(__NR_io_uring_register, s->ring_fd,
353 IORING_REGISTER_MAP_BUFFERS, &map, 1);
356 static int io_uring_register_buffers(struct submitter *s)
361 return syscall(__NR_io_uring_register, s->ring_fd,
362 IORING_REGISTER_BUFFERS, s->iovecs, depth);
365 static int io_uring_register_files(struct submitter *s)
372 s->fds = calloc(s->nr_files, sizeof(__s32));
373 for (i = 0; i < s->nr_files; i++) {
374 s->fds[i] = s->files[i].real_fd;
375 s->files[i].fixed_fd = i;
378 return syscall(__NR_io_uring_register, s->ring_fd,
379 IORING_REGISTER_FILES, s->fds, s->nr_files);
382 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
384 return syscall(__NR_io_uring_setup, entries, p);
387 static void io_uring_probe(int fd)
389 struct io_uring_probe *p;
392 p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
396 memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
397 ret = syscall(__NR_io_uring_register, fd, IORING_REGISTER_PROBE, p, 256);
401 if (IORING_OP_READ > p->ops_len)
404 if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED))
410 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
411 unsigned int min_complete, unsigned int flags)
413 return syscall(__NR_io_uring_enter, s->ring_fd, to_submit, min_complete,
417 #ifndef CONFIG_HAVE_GETTID
418 static int gettid(void)
420 return syscall(__NR_gettid);
424 static unsigned file_depth(struct submitter *s)
426 return (depth + s->nr_files - 1) / s->nr_files;
429 static void init_io(struct submitter *s, unsigned index)
431 struct io_uring_sqe *sqe = &s->sqes[index];
432 unsigned long offset;
437 sqe->opcode = IORING_OP_NOP;
441 if (s->nr_files == 1) {
444 f = &s->files[s->cur_file];
445 if (f->pending_ios >= file_depth(s)) {
447 if (s->cur_file == s->nr_files)
449 f = &s->files[s->cur_file];
454 r = __rand64(&s->rand_state);
455 offset = (r % (f->max_blocks - 1)) * bs;
457 if (register_files) {
458 sqe->flags = IOSQE_FIXED_FILE;
459 sqe->fd = f->fixed_fd;
462 sqe->fd = f->real_fd;
465 sqe->opcode = IORING_OP_READ_FIXED;
466 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
468 sqe->buf_index = index;
469 } else if (!vectored) {
470 sqe->opcode = IORING_OP_READ;
471 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
475 sqe->opcode = IORING_OP_READV;
476 sqe->addr = (unsigned long) &s->iovecs[index];
482 sqe->user_data = (unsigned long) f->fileno;
483 if (stats && stats_running)
484 sqe->user_data |= ((unsigned long)s->clock_index << 32);
487 static int prep_more_ios_uring(struct submitter *s, int max_ios)
489 struct io_sq_ring *ring = &s->sq_ring;
490 unsigned index, tail, next_tail, prepped = 0;
492 next_tail = tail = *ring->tail;
495 if (next_tail == atomic_load_acquire(ring->head))
498 index = tail & sq_ring_mask;
500 ring->array[index] = index;
503 } while (prepped < max_ios);
506 atomic_store_release(ring->tail, tail);
510 static int get_file_size(struct file *f)
514 if (fstat(f->real_fd, &st) < 0)
516 if (S_ISBLK(st.st_mode)) {
517 unsigned long long bytes;
519 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
522 f->max_blocks = bytes / bs;
524 } else if (S_ISREG(st.st_mode)) {
525 f->max_blocks = st.st_size / bs;
532 static int reap_events_uring(struct submitter *s)
534 struct io_cq_ring *ring = &s->cq_ring;
535 struct io_uring_cqe *cqe;
536 unsigned head, reaped = 0;
537 int last_idx = -1, stat_nr = 0;
544 if (head == atomic_load_acquire(ring->tail))
546 cqe = &ring->cqes[head & cq_ring_mask];
548 int fileno = cqe->user_data & 0xffffffff;
550 f = &s->files[fileno];
552 if (cqe->res != bs) {
553 printf("io: unexpected ret=%d\n", cqe->res);
554 if (polled && cqe->res == -EOPNOTSUPP)
555 printf("Your filesystem/driver/kernel doesn't support polled IO\n");
560 int clock_index = cqe->user_data >> 32;
562 if (last_idx != clock_index) {
563 if (last_idx != -1) {
564 add_stat(s, last_idx, stat_nr);
567 last_idx = clock_index;
576 add_stat(s, last_idx, stat_nr);
579 s->inflight -= reaped;
580 atomic_store_release(ring->head, head);
585 static int submitter_init(struct submitter *s)
590 printf("submitter=%d, tid=%d\n", s->index, s->tid);
592 __init_rand64(&s->rand_state, pthread_self());
593 srand48(pthread_self());
595 for (i = 0; i < MAX_FDS; i++)
596 s->files[i].fileno = i;
599 nr_batch = roundup_pow2(depth / batch_submit);
602 s->clock_batch = calloc(nr_batch, sizeof(unsigned long));
605 s->plat = calloc(PLAT_NR, sizeof(unsigned long));
607 s->clock_batch = NULL;
616 static int prep_more_ios_aio(struct submitter *s, int max_ios, struct iocb *iocbs)
618 unsigned long offset, data;
624 while (index < max_ios) {
625 struct iocb *iocb = &iocbs[index];
627 if (s->nr_files == 1) {
630 f = &s->files[s->cur_file];
631 if (f->pending_ios >= file_depth(s)) {
633 if (s->cur_file == s->nr_files)
635 f = &s->files[s->cur_file];
641 offset = (r % (f->max_blocks - 1)) * bs;
642 io_prep_pread(iocb, f->real_fd, s->iovecs[index].iov_base,
643 s->iovecs[index].iov_len, offset);
646 if (stats && stats_running)
647 data |= ((unsigned long) s->clock_index << 32);
648 iocb->data = (void *) (uintptr_t) data;
654 static int reap_events_aio(struct submitter *s, struct io_event *events, int evs)
656 int last_idx = -1, stat_nr = 0;
660 unsigned long data = (uintptr_t) events[reaped].data;
661 struct file *f = &s->files[data & 0xffffffff];
664 if (events[reaped].res != bs) {
665 printf("io: unexpected ret=%ld\n", events[reaped].res);
669 int clock_index = data >> 32;
671 if (last_idx != clock_index) {
672 if (last_idx != -1) {
673 add_stat(s, last_idx, stat_nr);
676 last_idx = clock_index;
685 add_stat(s, last_idx, stat_nr);
687 s->inflight -= reaped;
692 static void *submitter_aio_fn(void *data)
694 struct submitter *s = data;
695 int i, ret, prepped, nr_batch;
696 struct iocb **iocbsptr;
698 struct io_event *events;
700 nr_batch = submitter_init(s);
702 iocbsptr = calloc(depth, sizeof(struct iocb *));
703 iocbs = calloc(depth, sizeof(struct iocb));
704 events = calloc(depth, sizeof(struct io_event));
706 for (i = 0; i < depth; i++)
707 iocbsptr[i] = &iocbs[i];
711 int to_wait, to_submit, to_prep;
713 if (!prepped && s->inflight < depth) {
714 to_prep = min(depth - s->inflight, batch_submit);
715 prepped = prep_more_ios_aio(s, to_prep, iocbs);
716 #ifdef ARCH_HAVE_CPU_CLOCK
717 if (prepped && stats) {
718 s->clock_batch[s->clock_index] = get_cpu_clock();
719 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
723 s->inflight += prepped;
726 if (to_submit && (s->inflight + to_submit <= depth))
729 to_wait = min(s->inflight + to_submit, batch_complete);
731 ret = io_submit(s->aio_ctx, to_submit, iocbsptr);
736 } else if (ret != to_submit) {
737 printf("submitted %d, wanted %d\n", ret, to_submit);
746 r = io_getevents(s->aio_ctx, to_wait, to_wait, events, NULL);
748 perror("io_getevents");
750 } else if (r != to_wait) {
751 printf("r=%d, wait=%d\n", r, to_wait);
754 r = reap_events_aio(s, events, r);
758 } while (!s->finish);
768 static void *submitter_uring_fn(void *data)
770 struct submitter *s = data;
771 struct io_sq_ring *ring = &s->sq_ring;
772 int ret, prepped, nr_batch;
774 nr_batch = submitter_init(s);
778 int to_wait, to_submit, this_reap, to_prep;
779 unsigned ring_flags = 0;
781 if (!prepped && s->inflight < depth) {
782 to_prep = min(depth - s->inflight, batch_submit);
783 prepped = prep_more_ios_uring(s, to_prep);
784 #ifdef ARCH_HAVE_CPU_CLOCK
785 if (prepped && stats) {
786 s->clock_batch[s->clock_index] = get_cpu_clock();
787 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
791 s->inflight += prepped;
795 if (to_submit && (s->inflight + to_submit <= depth))
798 to_wait = min(s->inflight + to_submit, batch_complete);
801 * Only need to call io_uring_enter if we're not using SQ thread
802 * poll, or if IORING_SQ_NEED_WAKEUP is set.
805 ring_flags = atomic_load_acquire(ring->flags);
806 if (!sq_thread_poll || ring_flags & IORING_SQ_NEED_WAKEUP) {
810 flags = IORING_ENTER_GETEVENTS;
811 if (ring_flags & IORING_SQ_NEED_WAKEUP)
812 flags |= IORING_ENTER_SQ_WAKEUP;
813 ret = io_uring_enter(s, to_submit, to_wait, flags);
816 /* for SQPOLL, we submitted it all effectively */
821 * For non SQ thread poll, we already got the events we needed
822 * through the io_uring_enter() above. For SQ thread poll, we
823 * need to loop here until we find enough events.
829 r = reap_events_uring(s);
835 } while (sq_thread_poll && this_reap < to_wait);
836 s->reaps += this_reap;
844 } else if (ret < to_submit) {
845 int diff = to_submit - ret;
854 } else if (ret < 0) {
855 if (errno == EAGAIN) {
863 printf("io_submit: %s\n", strerror(errno));
866 } while (!s->finish);
872 static struct submitter *get_submitter(int offset)
878 ret += offset * (sizeof(*submitter) + depth * sizeof(struct iovec));
882 static void do_finish(const char *reason)
885 printf("Exiting on %s\n", reason);
886 for (j = 0; j < nthreads; j++) {
887 struct submitter *s = get_submitter(j);
890 if (max_iops > 100000)
891 printf("Maximum IOPS=%luK\n", max_iops / 1000);
893 printf("Maximum IOPS=%lu\n", max_iops);
897 static void sig_int(int sig)
902 static void arm_sig_int(void)
904 struct sigaction act;
906 memset(&act, 0, sizeof(act));
907 act.sa_handler = sig_int;
908 act.sa_flags = SA_RESTART;
909 sigaction(SIGINT, &act, NULL);
911 /* Windows uses SIGBREAK as a quit signal from other applications */
913 sigaction(SIGBREAK, &act, NULL);
917 static int setup_aio(struct submitter *s)
921 fprintf(stderr, "aio does not support polled IO\n");
924 if (sq_thread_poll) {
925 fprintf(stderr, "aio does not support SQPOLL IO\n");
929 fprintf(stderr, "aio does not support polled IO\n");
932 if (fixedbufs || register_files) {
933 fprintf(stderr, "aio does not support registered files or buffers\n");
934 fixedbufs = register_files = 0;
937 return io_queue_init(depth, &s->aio_ctx);
939 fprintf(stderr, "Legacy AIO not available on this system/build\n");
945 static int setup_ring(struct submitter *s)
947 struct io_sq_ring *sring = &s->sq_ring;
948 struct io_cq_ring *cring = &s->cq_ring;
949 struct io_uring_params p;
953 memset(&p, 0, sizeof(p));
955 if (polled && !do_nop)
956 p.flags |= IORING_SETUP_IOPOLL;
957 if (sq_thread_poll) {
958 p.flags |= IORING_SETUP_SQPOLL;
959 if (sq_thread_cpu != -1) {
960 p.flags |= IORING_SETUP_SQ_AFF;
961 p.sq_thread_cpu = sq_thread_cpu;
965 fd = io_uring_setup(depth, &p);
967 perror("io_uring_setup");
977 rlim.rlim_cur = RLIM_INFINITY;
978 rlim.rlim_max = RLIM_INFINITY;
979 /* ignore potential error, not needed on newer kernels */
980 setrlimit(RLIMIT_MEMLOCK, &rlim);
982 ret = io_uring_register_buffers(s);
984 perror("io_uring_register_buffers");
989 ret = io_uring_map_buffers(s);
991 perror("io_uring_map_buffers");
997 if (register_files) {
998 ret = io_uring_register_files(s);
1000 perror("io_uring_register_files");
1005 ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
1006 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1007 IORING_OFF_SQ_RING);
1008 sring->head = ptr + p.sq_off.head;
1009 sring->tail = ptr + p.sq_off.tail;
1010 sring->ring_mask = ptr + p.sq_off.ring_mask;
1011 sring->ring_entries = ptr + p.sq_off.ring_entries;
1012 sring->flags = ptr + p.sq_off.flags;
1013 sring->array = ptr + p.sq_off.array;
1014 sq_ring_mask = *sring->ring_mask;
1016 s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
1017 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1020 ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
1021 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1022 IORING_OFF_CQ_RING);
1023 cring->head = ptr + p.cq_off.head;
1024 cring->tail = ptr + p.cq_off.tail;
1025 cring->ring_mask = ptr + p.cq_off.ring_mask;
1026 cring->ring_entries = ptr + p.cq_off.ring_entries;
1027 cring->cqes = ptr + p.cq_off.cqes;
1028 cq_ring_mask = *cring->ring_mask;
1032 static void file_depths(char *buf)
1040 for (j = 0; j < nthreads; j++) {
1041 struct submitter *s = get_submitter(j);
1043 for (i = 0; i < s->nr_files; i++) {
1044 struct file *f = &s->files[i];
1047 p += sprintf(p, " %d", f->pending_ios);
1049 p += sprintf(p, "%d", f->pending_ios);
1055 static void usage(char *argv, int status)
1057 char runtime_str[16];
1058 snprintf(runtime_str, sizeof(runtime_str), "%d", runtime);
1059 printf("%s [options] -- [filenames]\n"
1060 " -d <int> : IO Depth, default %d\n"
1061 " -s <int> : Batch submit, default %d\n"
1062 " -c <int> : Batch complete, default %d\n"
1063 " -b <int> : Block size, default %d\n"
1064 " -p <bool> : Polled IO, default %d\n"
1065 " -B <bool> : Fixed buffers, default %d\n"
1066 " -D <bool> : DMA map fixed buffers, default %d\n"
1067 " -F <bool> : Register files, default %d\n"
1068 " -n <int> : Number of threads, default %d\n"
1069 " -O <bool> : Use O_DIRECT, default %d\n"
1070 " -N <bool> : Perform just no-op requests, default %d\n"
1071 " -t <bool> : Track IO latencies, default %d\n"
1072 " -T <int> : TSC rate in HZ\n"
1073 " -a <bool> : Use legacy aio, default %d\n"
1074 " -r <int> : Runtime in seconds, default %s\n",
1075 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled,
1076 fixedbufs, dma_map, register_files, nthreads, !buffered, do_nop,
1077 stats, aio, runtime == 0 ? "unlimited" : runtime_str);
1081 static void read_tsc_rate(void)
1089 fd = open(TSC_RATE_FILE, O_RDONLY);
1093 ret = read(fd, buffer, sizeof(buffer));
1099 tsc_rate = strtoul(buffer, NULL, 10);
1100 printf("Using TSC rate %luHz\n", tsc_rate);
1104 static void write_tsc_rate(void)
1110 if (!stat(TSC_RATE_FILE, &sb))
1113 fd = open(TSC_RATE_FILE, O_WRONLY | O_CREAT, 0644);
1117 memset(buffer, 0, sizeof(buffer));
1118 sprintf(buffer, "%lu", tsc_rate);
1119 ret = write(fd, buffer, strlen(buffer));
1125 int main(int argc, char *argv[])
1127 struct submitter *s;
1128 unsigned long done, calls, reap;
1129 int err, i, j, flags, fd, opt, threads_per_f, threads_rem = 0, nfiles;
1134 if (!do_nop && argc < 2)
1137 while ((opt = getopt(argc, argv, "d:s:c:b:p:B:F:n:N:O:t:T:a:r:D:h?")) != -1) {
1140 aio = !!atoi(optarg);
1143 depth = atoi(optarg);
1146 batch_submit = atoi(optarg);
1151 batch_complete = atoi(optarg);
1152 if (!batch_complete)
1159 polled = !!atoi(optarg);
1162 fixedbufs = !!atoi(optarg);
1165 register_files = !!atoi(optarg);
1168 nthreads = atoi(optarg);
1170 printf("Threads must be non-zero\n");
1175 do_nop = !!atoi(optarg);
1178 buffered = !atoi(optarg);
1181 #ifndef ARCH_HAVE_CPU_CLOCK
1182 fprintf(stderr, "Stats not supported on this CPU\n");
1185 stats = !!atoi(optarg);
1188 #ifndef ARCH_HAVE_CPU_CLOCK
1189 fprintf(stderr, "Stats not supported on this CPU\n");
1192 tsc_rate = strtoul(optarg, NULL, 10);
1196 runtime = atoi(optarg);
1199 dma_map = !!atoi(optarg);
1212 if (batch_complete > depth)
1213 batch_complete = depth;
1214 if (batch_submit > depth)
1215 batch_submit = depth;
1216 if (!fixedbufs && dma_map)
1219 submitter = calloc(nthreads, sizeof(*submitter) +
1220 depth * sizeof(struct iovec));
1221 for (j = 0; j < nthreads; j++) {
1222 s = get_submitter(j);
1224 s->done = s->calls = s->reaps = 0;
1227 flags = O_RDONLY | O_NOATIME;
1236 printf("No files specified\n");
1239 threads_per_f = nthreads / nfiles;
1240 /* make sure each thread gets assigned files */
1241 if (threads_per_f == 0) {
1244 threads_rem = nthreads - threads_per_f * nfiles;
1247 while (!do_nop && i < argc) {
1250 memset(&f, 0, sizeof(f));
1252 fd = open(argv[i], flags);
1258 if (get_file_size(&f)) {
1259 printf("failed getting size of device/file\n");
1262 if (f.max_blocks <= 1) {
1263 printf("Zero file/device size?\n");
1268 limit = threads_per_f;
1269 limit += threads_rem > 0 ? 1 : 0;
1270 for (k = 0; k < limit; k++) {
1271 s = get_submitter((j + k) % nthreads);
1273 if (s->nr_files == MAX_FDS) {
1274 printf("Max number of files (%d) reached\n", MAX_FDS);
1278 memcpy(&s->files[s->nr_files], &f, sizeof(f));
1280 printf("Added file %s (submitter %d)\n", argv[i], s->index);
1290 for (j = 0; j < nthreads; j++) {
1291 s = get_submitter(j);
1292 for (i = 0; i < depth; i++) {
1295 if (posix_memalign(&buf, bs, bs)) {
1296 printf("failed alloc\n");
1299 s->iovecs[i].iov_base = buf;
1300 s->iovecs[i].iov_len = bs;
1304 for (j = 0; j < nthreads; j++) {
1305 s = get_submitter(j);
1308 err = setup_ring(s);
1312 printf("ring setup failed: %s, %d\n", strerror(errno), err);
1316 s = get_submitter(0);
1317 printf("polled=%d, fixedbufs=%d/%d, register_files=%d, buffered=%d, QD=%d\n", polled, fixedbufs, dma_map, register_files, buffered, depth);
1319 printf("Engine=io_uring, sq_ring=%d, cq_ring=%d\n", *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
1321 printf("Engine=aio\n");
1323 for (j = 0; j < nthreads; j++) {
1324 s = get_submitter(j);
1326 pthread_create(&s->thread, NULL, submitter_uring_fn, s);
1327 #ifdef CONFIG_LIBAIO
1329 pthread_create(&s->thread, NULL, submitter_aio_fn, s);
1333 fdepths = malloc(8 * s->nr_files * nthreads);
1334 reap = calls = done = 0;
1336 unsigned long this_done = 0;
1337 unsigned long this_reap = 0;
1338 unsigned long this_call = 0;
1339 unsigned long rpc = 0, ipc = 0;
1340 unsigned long iops, bw;
1343 if (runtime && !--runtime)
1344 do_finish("timeout");
1346 /* don't print partial run, if interrupted by signal */
1350 /* one second in to the run, enable stats */
1354 for (j = 0; j < nthreads; j++) {
1355 s = get_submitter(j);
1356 this_done += s->done;
1357 this_call += s->calls;
1358 this_reap += s->reaps;
1360 if (this_call - calls) {
1361 rpc = (this_done - done) / (this_call - calls);
1362 ipc = (this_reap - reap) / (this_call - calls);
1365 file_depths(fdepths);
1366 iops = this_done - done;
1368 bw = iops * (bs / 1048576);
1370 bw = iops / (1048576 / bs);
1372 printf("IOPS=%luK, ", iops / 1000);
1374 printf("IOPS=%lu, ", iops);
1375 max_iops = max(max_iops, iops);
1377 printf("BW=%luMiB/s, ", bw);
1378 printf("IOS/call=%ld/%ld, inflight=(%s)\n", rpc, ipc, fdepths);
1384 for (j = 0; j < nthreads; j++) {
1385 s = get_submitter(j);
1386 pthread_join(s->thread, &ret);
1392 printf("%d: Latency percentiles:\n", s->tid);
1393 for (i = 0, nr = 0; i < PLAT_NR; i++)
1395 show_clat_percentiles(s->plat, nr, 4);
1396 free(s->clock_batch);