18 #include <sys/types.h>
20 #include <sys/ioctl.h>
21 #include <sys/syscall.h>
22 #include <sys/resource.h>
32 #include "../arch/arch.h"
34 #include "../lib/types.h"
35 #include "../lib/roundup.h"
36 #include "../lib/rand.h"
37 #include "../minmax.h"
38 #include "../os/linux/io_uring.h"
39 #include "../engines/nvme.h"
45 unsigned *ring_entries;
54 unsigned *ring_entries;
55 struct io_uring_cqe *cqes;
59 #define BATCH_SUBMIT 32
60 #define BATCH_COMPLETE 32
65 static unsigned sq_ring_mask, cq_ring_mask;
68 unsigned long max_blocks;
69 unsigned long max_size;
70 unsigned long cur_off;
72 unsigned int nsid; /* nsid field required for nvme-passthrough */
73 unsigned int lba_shift; /* lba_shift field required for nvme-passthrough */
80 #define PLAT_VAL (1 << PLAT_BITS)
81 #define PLAT_GROUP_NR 29
82 #define PLAT_NR (PLAT_GROUP_NR * PLAT_VAL)
89 struct io_sq_ring sq_ring;
90 struct io_uring_sqe *sqes;
91 struct io_cq_ring cq_ring;
101 struct taus258_state rand_state;
103 unsigned long *clock_batch;
108 io_context_t aio_ctx;
112 const char *filename;
114 struct file files[MAX_FDS];
117 struct iovec iovecs[];
120 static struct submitter *submitter;
121 static volatile int finish;
122 static int stats_running;
123 static unsigned long max_iops;
124 static long t_io_uring_page_size;
126 static int depth = DEPTH;
127 static int batch_submit = BATCH_SUBMIT;
128 static int batch_complete = BATCH_COMPLETE;
130 static int polled = 1; /* use IO polling */
131 static int fixedbufs = 1; /* use fixed user buffers */
132 static int dma_map; /* pre-map DMA buffers */
133 static int register_files = 1; /* use fixed files */
134 static int buffered = 0; /* use buffered IO, not O_DIRECT */
135 static int sq_thread_poll = 0; /* use kernel submission/poller thread */
136 static int sq_thread_cpu = -1; /* pin above thread to this CPU */
137 static int do_nop = 0; /* no-op SQ ring commands */
138 static int nthreads = 1;
139 static int stats = 0; /* generate IO stats */
140 static int aio = 0; /* use libaio */
141 static int runtime = 0; /* runtime */
142 static int random_io = 1; /* random or sequential IO */
143 static int register_ring = 1; /* register ring */
144 static int use_sync = 0; /* use preadv2 */
145 static int numa_placement = 0; /* set to node of device */
146 static int pt = 0; /* passthrough I/O or not */
148 static unsigned long tsc_rate;
150 #define TSC_RATE_FILE "tsc-rate"
152 static int vectored = 1;
154 static float plist[] = { 1.0, 5.0, 10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0,
155 80.0, 90.0, 95.0, 99.0, 99.5, 99.9, 99.95, 99.99 };
156 static int plist_len = 17;
158 #ifndef IORING_REGISTER_MAP_BUFFERS
159 #define IORING_REGISTER_MAP_BUFFERS 26
160 struct io_uring_map_buffers {
169 static int nvme_identify(int fd, __u32 nsid, enum nvme_identify_cns cns,
170 enum nvme_csi csi, void *data)
172 struct nvme_passthru_cmd cmd = {
173 .opcode = nvme_admin_identify,
175 .addr = (__u64)(uintptr_t)data,
176 .data_len = NVME_IDENTIFY_DATA_SIZE,
178 .cdw11 = csi << NVME_IDENTIFY_CSI_SHIFT,
179 .timeout_ms = NVME_DEFAULT_IOCTL_TIMEOUT,
182 return ioctl(fd, NVME_IOCTL_ADMIN_CMD, &cmd);
185 static int nvme_get_info(int fd, __u32 *nsid, __u32 *lba_sz, __u64 *nlba)
187 struct nvme_id_ns ns;
191 namespace_id = ioctl(fd, NVME_IOCTL_ID);
192 if (namespace_id < 0) {
193 fprintf(stderr, "error failed to fetch namespace-id\n");
199 * Identify namespace to get namespace-id, namespace size in LBA's
202 err = nvme_identify(fd, namespace_id, NVME_IDENTIFY_CNS_NS,
205 fprintf(stderr, "error failed to fetch identify namespace\n");
210 *nsid = namespace_id;
211 *lba_sz = 1 << ns.lbaf[(ns.flbas & 0x0f)].ds;
217 static unsigned long cycles_to_nsec(unsigned long cycles)
224 val = cycles * 1000000000ULL;
225 return val / tsc_rate;
228 static unsigned long plat_idx_to_val(unsigned int idx)
230 unsigned int error_bits;
231 unsigned long k, base;
233 assert(idx < PLAT_NR);
235 /* MSB <= (PLAT_BITS-1), cannot be rounded off. Use
236 * all bits of the sample as index */
237 if (idx < (PLAT_VAL << 1))
238 return cycles_to_nsec(idx);
240 /* Find the group and compute the minimum value of that group */
241 error_bits = (idx >> PLAT_BITS) - 1;
242 base = ((unsigned long) 1) << (error_bits + PLAT_BITS);
244 /* Find its bucket number of the group */
247 /* Return the mean of the range of the bucket */
248 return cycles_to_nsec(base + ((k + 0.5) * (1 << error_bits)));
251 unsigned int calculate_clat_percentiles(unsigned long *io_u_plat,
252 unsigned long nr, unsigned long **output,
253 unsigned long *maxv, unsigned long *minv)
255 unsigned long sum = 0;
256 unsigned int len = plist_len, i, j = 0;
257 unsigned long *ovals = NULL;
263 ovals = malloc(len * sizeof(*ovals));
268 * Calculate bucket values, note down max and min values
271 for (i = 0; i < PLAT_NR && !is_last; i++) {
273 while (sum >= ((long double) plist[j] / 100.0 * nr)) {
274 assert(plist[j] <= 100.0);
276 ovals[j] = plat_idx_to_val(i);
277 if (ovals[j] < *minv)
279 if (ovals[j] > *maxv)
282 is_last = (j == len - 1) != 0;
291 fprintf(stderr, "error calculating latency percentiles\n");
297 static void show_clat_percentiles(unsigned long *io_u_plat, unsigned long nr,
298 unsigned int precision)
300 unsigned int divisor, len, i, j = 0;
301 unsigned long minv, maxv;
302 unsigned long *ovals;
303 int per_line, scale_down, time_width;
307 len = calculate_clat_percentiles(io_u_plat, nr, &ovals, &maxv, &minv);
314 printf(" percentiles (tsc ticks):\n |");
315 } else if (minv > 2000 && maxv > 99999) {
318 printf(" percentiles (usec):\n |");
322 printf(" percentiles (nsec):\n |");
325 time_width = max(5, (int) (log10(maxv / divisor) + 1));
326 snprintf(fmt, sizeof(fmt), " %%%u.%ufth=[%%%dllu]%%c", precision + 3,
327 precision, time_width);
328 /* fmt will be something like " %5.2fth=[%4llu]%c" */
329 per_line = (80 - 7) / (precision + 10 + time_width);
331 for (j = 0; j < len; j++) {
333 if (j != 0 && (j % per_line) == 0)
336 /* end of the list */
337 is_last = (j == len - 1) != 0;
339 for (i = 0; i < scale_down; i++)
340 ovals[j] = (ovals[j] + 999) / 1000;
342 printf(fmt, plist[j], ovals[j], is_last ? '\n' : ',');
347 if ((j % per_line) == per_line - 1) /* for formatting */
355 #ifdef ARCH_HAVE_CPU_CLOCK
356 static unsigned int plat_val_to_idx(unsigned long val)
358 unsigned int msb, error_bits, base, offset, idx;
360 /* Find MSB starting from bit 0 */
364 msb = (sizeof(val)*8) - __builtin_clzll(val) - 1;
367 * MSB <= (PLAT_BITS-1), cannot be rounded off. Use
368 * all bits of the sample as index
370 if (msb <= PLAT_BITS)
373 /* Compute the number of error bits to discard*/
374 error_bits = msb - PLAT_BITS;
376 /* Compute the number of buckets before the group */
377 base = (error_bits + 1) << PLAT_BITS;
380 * Discard the error bits and apply the mask to find the
381 * index for the buckets in the group
383 offset = (PLAT_VAL - 1) & (val >> error_bits);
385 /* Make sure the index does not exceed (array size - 1) */
386 idx = (base + offset) < (PLAT_NR - 1) ?
387 (base + offset) : (PLAT_NR - 1);
393 static void add_stat(struct submitter *s, int clock_index, int nr)
395 #ifdef ARCH_HAVE_CPU_CLOCK
396 unsigned long cycles;
399 if (!s->finish && clock_index) {
400 cycles = get_cpu_clock();
401 cycles -= s->clock_batch[clock_index];
402 pidx = plat_val_to_idx(cycles);
408 static int io_uring_map_buffers(struct submitter *s)
410 struct io_uring_map_buffers map = {
411 .fd = s->files[0].real_fd,
418 fprintf(stdout, "Mapping buffers may not work with multiple files\n");
420 return syscall(__NR_io_uring_register, s->ring_fd,
421 IORING_REGISTER_MAP_BUFFERS, &map, 1);
424 static int io_uring_register_buffers(struct submitter *s)
429 return syscall(__NR_io_uring_register, s->ring_fd,
430 IORING_REGISTER_BUFFERS, s->iovecs, roundup_pow2(depth));
433 static int io_uring_register_files(struct submitter *s)
440 s->fds = calloc(s->nr_files, sizeof(__s32));
441 for (i = 0; i < s->nr_files; i++) {
442 s->fds[i] = s->files[i].real_fd;
443 s->files[i].fixed_fd = i;
446 return syscall(__NR_io_uring_register, s->ring_fd,
447 IORING_REGISTER_FILES, s->fds, s->nr_files);
450 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
455 * Clamp CQ ring size at our SQ ring size, we don't need more entries
458 p->flags |= IORING_SETUP_CQSIZE;
459 p->cq_entries = entries;
461 p->flags |= IORING_SETUP_COOP_TASKRUN;
462 p->flags |= IORING_SETUP_SINGLE_ISSUER;
463 p->flags |= IORING_SETUP_DEFER_TASKRUN;
465 ret = syscall(__NR_io_uring_setup, entries, p);
469 if (errno == EINVAL && p->flags & IORING_SETUP_COOP_TASKRUN) {
470 p->flags &= ~IORING_SETUP_COOP_TASKRUN;
473 if (errno == EINVAL && p->flags & IORING_SETUP_SINGLE_ISSUER) {
474 p->flags &= ~IORING_SETUP_SINGLE_ISSUER;
477 if (errno == EINVAL && p->flags & IORING_SETUP_DEFER_TASKRUN) {
478 p->flags &= ~IORING_SETUP_DEFER_TASKRUN;
485 static void io_uring_probe(int fd)
487 struct io_uring_probe *p;
490 p = calloc(1, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
494 ret = syscall(__NR_io_uring_register, fd, IORING_REGISTER_PROBE, p, 256);
498 if (IORING_OP_READ > p->ops_len)
501 if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED))
507 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
508 unsigned int min_complete, unsigned int flags)
511 flags |= IORING_ENTER_REGISTERED_RING;
512 #ifdef FIO_ARCH_HAS_SYSCALL
513 return __do_syscall6(__NR_io_uring_enter, s->enter_ring_fd, to_submit,
514 min_complete, flags, NULL, 0);
516 return syscall(__NR_io_uring_enter, s->enter_ring_fd, to_submit,
517 min_complete, flags, NULL, 0);
521 static unsigned file_depth(struct submitter *s)
523 return (depth + s->nr_files - 1) / s->nr_files;
526 static unsigned long long get_offset(struct submitter *s, struct file *f)
528 unsigned long long offset;
532 unsigned long long block;
534 r = __rand64(&s->rand_state);
535 block = r % f->max_blocks;
536 offset = block * (unsigned long long) bs;
540 if (f->cur_off + bs > f->max_size)
547 static struct file *init_new_io(struct submitter *s)
551 if (s->nr_files == 1) {
554 f = &s->files[s->cur_file];
555 if (f->pending_ios >= file_depth(s)) {
557 if (s->cur_file == s->nr_files)
559 f = &s->files[s->cur_file];
567 static void init_io(struct submitter *s, unsigned index)
569 struct io_uring_sqe *sqe = &s->sqes[index];
573 sqe->opcode = IORING_OP_NOP;
579 if (register_files) {
580 sqe->flags = IOSQE_FIXED_FILE;
581 sqe->fd = f->fixed_fd;
584 sqe->fd = f->real_fd;
587 sqe->opcode = IORING_OP_READ_FIXED;
588 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
590 sqe->buf_index = index;
591 } else if (!vectored) {
592 sqe->opcode = IORING_OP_READ;
593 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
597 sqe->opcode = IORING_OP_READV;
598 sqe->addr = (unsigned long) &s->iovecs[index];
603 sqe->off = get_offset(s, f);
604 sqe->user_data = (unsigned long) f->fileno;
605 if (stats && stats_running)
606 sqe->user_data |= ((uint64_t)s->clock_index << 32);
609 static void init_io_pt(struct submitter *s, unsigned index)
611 struct io_uring_sqe *sqe = &s->sqes[index << 1];
612 unsigned long offset;
614 struct nvme_uring_cmd *cmd;
615 unsigned long long slba;
616 unsigned long long nlb;
620 offset = get_offset(s, f);
622 if (register_files) {
623 sqe->fd = f->fixed_fd;
624 sqe->flags = IOSQE_FIXED_FILE;
626 sqe->fd = f->real_fd;
629 sqe->opcode = IORING_OP_URING_CMD;
630 sqe->user_data = (unsigned long) f->fileno;
632 sqe->user_data |= ((__u64) s->clock_index << 32ULL);
633 sqe->cmd_op = NVME_URING_CMD_IO;
634 slba = offset >> f->lba_shift;
635 nlb = (bs >> f->lba_shift) - 1;
636 cmd = (struct nvme_uring_cmd *)&sqe->cmd;
637 /* cdw10 and cdw11 represent starting slba*/
638 cmd->cdw10 = slba & 0xffffffff;
639 cmd->cdw11 = slba >> 32;
640 /* cdw12 represent number of lba to be read*/
642 cmd->addr = (unsigned long) s->iovecs[index].iov_base;
645 sqe->uring_cmd_flags = IORING_URING_CMD_FIXED;
646 sqe->buf_index = index;
652 static int prep_more_ios_uring(struct submitter *s, int max_ios)
654 struct io_sq_ring *ring = &s->sq_ring;
655 unsigned head, index, tail, next_tail, prepped = 0;
658 head = atomic_load_acquire(ring->head);
662 next_tail = tail = *ring->tail;
665 if (next_tail == head)
668 index = tail & sq_ring_mask;
670 init_io_pt(s, index);
675 } while (prepped < max_ios);
678 atomic_store_release(ring->tail, tail);
682 static int get_file_size(struct file *f)
686 if (fstat(f->real_fd, &st) < 0)
693 if (!S_ISCHR(st.st_mode)) {
694 fprintf(stderr, "passthrough works with only nvme-ns "
695 "generic devices (/dev/ngXnY)\n");
698 ret = nvme_get_info(f->real_fd, &f->nsid, &lbs, &nlba);
701 if ((bs % lbs) != 0) {
702 printf("error: bs:%d should be a multiple logical_block_size:%d\n",
706 f->max_blocks = nlba;
708 f->lba_shift = ilog2(lbs);
710 } else if (S_ISBLK(st.st_mode)) {
711 unsigned long long bytes;
713 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
716 f->max_blocks = bytes / bs;
719 } else if (S_ISREG(st.st_mode)) {
720 f->max_blocks = st.st_size / bs;
721 f->max_size = st.st_size;
728 static int reap_events_uring(struct submitter *s)
730 struct io_cq_ring *ring = &s->cq_ring;
731 struct io_uring_cqe *cqe;
732 unsigned head, reaped = 0;
733 int last_idx = -1, stat_nr = 0;
739 if (head == atomic_load_acquire(ring->tail))
741 cqe = &ring->cqes[head & cq_ring_mask];
743 int fileno = cqe->user_data & 0xffffffff;
745 f = &s->files[fileno];
747 if (cqe->res != bs) {
748 printf("io: unexpected ret=%d\n", cqe->res);
749 if (polled && cqe->res == -EOPNOTSUPP)
750 printf("Your filesystem/driver/kernel doesn't support polled IO\n");
755 int clock_index = cqe->user_data >> 32;
757 if (last_idx != clock_index) {
758 if (last_idx != -1) {
759 add_stat(s, last_idx, stat_nr);
762 last_idx = clock_index;
771 add_stat(s, last_idx, stat_nr);
774 s->inflight -= reaped;
775 atomic_store_release(ring->head, head);
780 static int reap_events_uring_pt(struct submitter *s)
782 struct io_cq_ring *ring = &s->cq_ring;
783 struct io_uring_cqe *cqe;
784 unsigned head, reaped = 0;
785 int last_idx = -1, stat_nr = 0;
793 if (head == atomic_load_acquire(ring->tail))
795 index = head & cq_ring_mask;
796 cqe = &ring->cqes[index << 1];
797 fileno = cqe->user_data & 0xffffffff;
798 f = &s->files[fileno];
802 printf("io: unexpected ret=%d\n", cqe->res);
803 if (polled && cqe->res == -EINVAL)
804 printf("passthrough doesn't support polled IO\n");
808 int clock_index = cqe->user_data >> 32;
810 if (last_idx != clock_index) {
811 if (last_idx != -1) {
812 add_stat(s, last_idx, stat_nr);
815 last_idx = clock_index;
824 add_stat(s, last_idx, stat_nr);
827 s->inflight -= reaped;
828 atomic_store_release(ring->head, head);
833 static void set_affinity(struct submitter *s)
835 #ifdef CONFIG_LIBNUMA
836 struct bitmask *mask;
838 if (s->numa_node == -1)
841 numa_set_preferred(s->numa_node);
843 mask = numa_allocate_cpumask();
844 numa_node_to_cpus(s->numa_node, mask);
845 numa_sched_setaffinity(s->tid, mask);
849 static int detect_node(struct submitter *s, const char *name)
851 #ifdef CONFIG_LIBNUMA
852 const char *base = basename(name);
857 sprintf(str, "/sys/class/nvme-generic/%s/device/numa_node", base);
859 sprintf(str, "/sys/block/%s/device/numa_node", base);
860 fd = open(str, O_RDONLY);
864 ret = read(fd, str, sizeof(str));
878 static int setup_aio(struct submitter *s)
882 fprintf(stderr, "aio does not support polled IO\n");
885 if (sq_thread_poll) {
886 fprintf(stderr, "aio does not support SQPOLL IO\n");
890 fprintf(stderr, "aio does not support polled IO\n");
893 if (fixedbufs || register_files) {
894 fprintf(stderr, "aio does not support registered files or buffers\n");
895 fixedbufs = register_files = 0;
898 return io_queue_init(roundup_pow2(depth), &s->aio_ctx);
900 fprintf(stderr, "Legacy AIO not available on this system/build\n");
906 static int setup_ring(struct submitter *s)
908 struct io_sq_ring *sring = &s->sq_ring;
909 struct io_cq_ring *cring = &s->cq_ring;
910 struct io_uring_params p;
915 memset(&p, 0, sizeof(p));
917 if (polled && !do_nop)
918 p.flags |= IORING_SETUP_IOPOLL;
919 if (sq_thread_poll) {
920 p.flags |= IORING_SETUP_SQPOLL;
921 if (sq_thread_cpu != -1) {
922 p.flags |= IORING_SETUP_SQ_AFF;
923 p.sq_thread_cpu = sq_thread_cpu;
927 p.flags |= IORING_SETUP_SQE128;
928 p.flags |= IORING_SETUP_CQE32;
931 fd = io_uring_setup(depth, &p);
933 perror("io_uring_setup");
936 s->ring_fd = s->enter_ring_fd = fd;
943 rlim.rlim_cur = RLIM_INFINITY;
944 rlim.rlim_max = RLIM_INFINITY;
945 /* ignore potential error, not needed on newer kernels */
946 setrlimit(RLIMIT_MEMLOCK, &rlim);
948 ret = io_uring_register_buffers(s);
950 perror("io_uring_register_buffers");
955 ret = io_uring_map_buffers(s);
957 perror("io_uring_map_buffers");
963 if (register_files) {
964 ret = io_uring_register_files(s);
966 perror("io_uring_register_files");
971 ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
972 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
974 sring->head = ptr + p.sq_off.head;
975 sring->tail = ptr + p.sq_off.tail;
976 sring->ring_mask = ptr + p.sq_off.ring_mask;
977 sring->ring_entries = ptr + p.sq_off.ring_entries;
978 sring->flags = ptr + p.sq_off.flags;
979 sring->array = ptr + p.sq_off.array;
980 sq_ring_mask = *sring->ring_mask;
982 if (p.flags & IORING_SETUP_SQE128)
983 len = 2 * p.sq_entries * sizeof(struct io_uring_sqe);
985 len = p.sq_entries * sizeof(struct io_uring_sqe);
986 s->sqes = mmap(0, len,
987 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
990 if (p.flags & IORING_SETUP_CQE32) {
991 len = p.cq_off.cqes +
992 2 * p.cq_entries * sizeof(struct io_uring_cqe);
994 len = p.cq_off.cqes +
995 p.cq_entries * sizeof(struct io_uring_cqe);
998 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
1000 cring->head = ptr + p.cq_off.head;
1001 cring->tail = ptr + p.cq_off.tail;
1002 cring->ring_mask = ptr + p.cq_off.ring_mask;
1003 cring->ring_entries = ptr + p.cq_off.ring_entries;
1004 cring->cqes = ptr + p.cq_off.cqes;
1005 cq_ring_mask = *cring->ring_mask;
1007 for (i = 0; i < p.sq_entries; i++)
1008 sring->array[i] = i;
1013 static void *allocate_mem(struct submitter *s, int size)
1017 #ifdef CONFIG_LIBNUMA
1018 if (s->numa_node != -1)
1019 return numa_alloc_onnode(size, s->numa_node);
1022 if (posix_memalign(&buf, t_io_uring_page_size, bs)) {
1023 printf("failed alloc\n");
1030 static int submitter_init(struct submitter *s)
1032 int i, nr_batch, err;
1033 static int init_printed;
1036 printf("submitter=%d, tid=%d, file=%s, node=%d\n", s->index, s->tid,
1037 s->filename, s->numa_node);
1041 __init_rand64(&s->rand_state, s->tid);
1044 for (i = 0; i < MAX_FDS; i++)
1045 s->files[i].fileno = i;
1047 for (i = 0; i < roundup_pow2(depth); i++) {
1050 buf = allocate_mem(s, bs);
1053 s->iovecs[i].iov_base = buf;
1054 s->iovecs[i].iov_len = bs;
1058 sprintf(buf, "Engine=preadv2\n");
1061 err = setup_ring(s);
1063 sprintf(buf, "Engine=io_uring, sq_ring=%d, cq_ring=%d\n", *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
1065 sprintf(buf, "Engine=aio\n");
1069 printf("queue setup failed: %s, %d\n", strerror(errno), err);
1073 if (!init_printed) {
1074 printf("polled=%d, fixedbufs=%d/%d, register_files=%d, buffered=%d, QD=%d\n", polled, fixedbufs, dma_map, register_files, buffered, depth);
1080 nr_batch = roundup_pow2(depth / batch_submit);
1083 s->clock_batch = calloc(nr_batch, sizeof(unsigned long));
1086 s->plat = calloc(PLAT_NR, sizeof(unsigned long));
1088 s->clock_batch = NULL;
1092 /* perform the expensive command initialization part for passthrough here
1093 * rather than in the fast path
1096 for (i = 0; i < roundup_pow2(depth); i++) {
1097 struct io_uring_sqe *sqe = &s->sqes[i << 1];
1099 memset(&sqe->cmd, 0, sizeof(struct nvme_uring_cmd));
1105 #ifdef CONFIG_LIBAIO
1106 static int prep_more_ios_aio(struct submitter *s, int max_ios, struct iocb *iocbs)
1113 while (index < max_ios) {
1114 struct iocb *iocb = &iocbs[index];
1118 io_prep_pread(iocb, f->real_fd, s->iovecs[index].iov_base,
1119 s->iovecs[index].iov_len, get_offset(s, f));
1122 if (stats && stats_running)
1123 data |= (((uint64_t) s->clock_index) << 32);
1124 iocb->data = (void *) (uintptr_t) data;
1130 static int reap_events_aio(struct submitter *s, struct io_event *events, int evs)
1132 int last_idx = -1, stat_nr = 0;
1136 uint64_t data = (uintptr_t) events[reaped].data;
1137 struct file *f = &s->files[data & 0xffffffff];
1140 if (events[reaped].res != bs) {
1141 printf("io: unexpected ret=%ld\n", events[reaped].res);
1145 int clock_index = data >> 32;
1147 if (last_idx != clock_index) {
1148 if (last_idx != -1) {
1149 add_stat(s, last_idx, stat_nr);
1152 last_idx = clock_index;
1161 add_stat(s, last_idx, stat_nr);
1163 s->inflight -= reaped;
1168 static void *submitter_aio_fn(void *data)
1170 struct submitter *s = data;
1171 int i, ret, prepped;
1172 struct iocb **iocbsptr;
1174 struct io_event *events;
1175 #ifdef ARCH_HAVE_CPU_CLOCK
1179 ret = submitter_init(s);
1183 #ifdef ARCH_HAVE_CPU_CLOCK
1187 iocbsptr = calloc(depth, sizeof(struct iocb *));
1188 iocbs = calloc(depth, sizeof(struct iocb));
1189 events = calloc(depth, sizeof(struct io_event));
1191 for (i = 0; i < depth; i++)
1192 iocbsptr[i] = &iocbs[i];
1196 int to_wait, to_submit, to_prep;
1198 if (!prepped && s->inflight < depth) {
1199 to_prep = min(depth - s->inflight, batch_submit);
1200 prepped = prep_more_ios_aio(s, to_prep, iocbs);
1201 #ifdef ARCH_HAVE_CPU_CLOCK
1202 if (prepped && stats) {
1203 s->clock_batch[s->clock_index] = get_cpu_clock();
1204 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
1208 s->inflight += prepped;
1209 to_submit = prepped;
1211 if (to_submit && (s->inflight + to_submit <= depth))
1214 to_wait = min(s->inflight + to_submit, batch_complete);
1216 ret = io_submit(s->aio_ctx, to_submit, iocbsptr);
1219 perror("io_submit");
1221 } else if (ret != to_submit) {
1222 printf("submitted %d, wanted %d\n", ret, to_submit);
1231 r = io_getevents(s->aio_ctx, to_wait, to_wait, events, NULL);
1233 perror("io_getevents");
1235 } else if (r != to_wait) {
1236 printf("r=%d, wait=%d\n", r, to_wait);
1239 r = reap_events_aio(s, events, r);
1243 } while (!s->finish);
1254 static void io_uring_unregister_ring(struct submitter *s)
1256 struct io_uring_rsrc_update up = {
1257 .offset = s->enter_ring_fd,
1260 syscall(__NR_io_uring_register, s->ring_fd, IORING_UNREGISTER_RING_FDS,
1264 static int io_uring_register_ring(struct submitter *s)
1266 struct io_uring_rsrc_update up = {
1272 ret = syscall(__NR_io_uring_register, s->ring_fd,
1273 IORING_REGISTER_RING_FDS, &up, 1);
1275 s->enter_ring_fd = up.offset;
1282 static void *submitter_uring_fn(void *data)
1284 struct submitter *s = data;
1285 struct io_sq_ring *ring = &s->sq_ring;
1287 #ifdef ARCH_HAVE_CPU_CLOCK
1291 ret = submitter_init(s);
1295 #ifdef ARCH_HAVE_CPU_CLOCK
1300 io_uring_register_ring(s);
1304 int to_wait, to_submit, this_reap, to_prep;
1305 unsigned ring_flags = 0;
1307 if (!prepped && s->inflight < depth) {
1308 to_prep = min(depth - s->inflight, batch_submit);
1309 prepped = prep_more_ios_uring(s, to_prep);
1310 #ifdef ARCH_HAVE_CPU_CLOCK
1311 if (prepped && stats) {
1312 s->clock_batch[s->clock_index] = get_cpu_clock();
1313 s->clock_index = (s->clock_index + 1) & (nr_batch - 1);
1317 s->inflight += prepped;
1319 to_submit = prepped;
1321 if (to_submit && (s->inflight + to_submit <= depth))
1324 to_wait = min(s->inflight + to_submit, batch_complete);
1327 * Only need to call io_uring_enter if we're not using SQ thread
1328 * poll, or if IORING_SQ_NEED_WAKEUP is set.
1331 ring_flags = atomic_load_acquire(ring->flags);
1332 if (!sq_thread_poll || ring_flags & IORING_SQ_NEED_WAKEUP) {
1336 flags = IORING_ENTER_GETEVENTS;
1337 if (ring_flags & IORING_SQ_NEED_WAKEUP)
1338 flags |= IORING_ENTER_SQ_WAKEUP;
1339 ret = io_uring_enter(s, to_submit, to_wait, flags);
1342 /* for SQPOLL, we submitted it all effectively */
1347 * For non SQ thread poll, we already got the events we needed
1348 * through the io_uring_enter() above. For SQ thread poll, we
1349 * need to loop here until we find enough events.
1356 r = reap_events_uring_pt(s);
1358 r = reap_events_uring(s);
1364 } while (sq_thread_poll && this_reap < to_wait);
1365 s->reaps += this_reap;
1373 } else if (ret < to_submit) {
1374 int diff = to_submit - ret;
1383 } else if (ret < 0) {
1384 if (errno == EAGAIN) {
1392 printf("io_submit: %s\n", strerror(errno));
1395 } while (!s->finish);
1398 io_uring_unregister_ring(s);
1405 #ifdef CONFIG_PWRITEV2
1406 static void *submitter_sync_fn(void *data)
1408 struct submitter *s = data;
1411 if (submitter_init(s) < 0)
1420 #ifdef ARCH_HAVE_CPU_CLOCK
1422 s->clock_batch[s->clock_index] = get_cpu_clock();
1428 offset = get_offset(s, f);
1430 ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, RWF_HIPRI);
1432 ret = preadv2(f->real_fd, &s->iovecs[0], 1, offset, 0);
1437 } else if (ret != bs) {
1445 add_stat(s, s->clock_index, 1);
1446 } while (!s->finish);
1453 static void *submitter_sync_fn(void *data)
1460 static struct submitter *get_submitter(int offset)
1466 ret += offset * (sizeof(*submitter) + depth * sizeof(struct iovec));
1470 static void do_finish(const char *reason)
1474 printf("Exiting on %s\n", reason);
1475 for (j = 0; j < nthreads; j++) {
1476 struct submitter *s = get_submitter(j);
1479 if (max_iops > 1000000) {
1480 double miops = (double) max_iops / 1000000.0;
1481 printf("Maximum IOPS=%.2fM\n", miops);
1482 } else if (max_iops > 100000) {
1483 double kiops = (double) max_iops / 1000.0;
1484 printf("Maximum IOPS=%.2fK\n", kiops);
1486 printf("Maximum IOPS=%lu\n", max_iops);
1491 static void sig_int(int sig)
1493 do_finish("signal");
1496 static void arm_sig_int(void)
1498 struct sigaction act;
1500 memset(&act, 0, sizeof(act));
1501 act.sa_handler = sig_int;
1502 act.sa_flags = SA_RESTART;
1503 sigaction(SIGINT, &act, NULL);
1505 /* Windows uses SIGBREAK as a quit signal from other applications */
1507 sigaction(SIGBREAK, &act, NULL);
1511 static void usage(char *argv, int status)
1513 char runtime_str[16];
1514 snprintf(runtime_str, sizeof(runtime_str), "%d", runtime);
1515 printf("%s [options] -- [filenames]\n"
1516 " -d <int> : IO Depth, default %d\n"
1517 " -s <int> : Batch submit, default %d\n"
1518 " -c <int> : Batch complete, default %d\n"
1519 " -b <int> : Block size, default %d\n"
1520 " -p <bool> : Polled IO, default %d\n"
1521 " -B <bool> : Fixed buffers, default %d\n"
1522 " -D <bool> : DMA map fixed buffers, default %d\n"
1523 " -F <bool> : Register files, default %d\n"
1524 " -n <int> : Number of threads, default %d\n"
1525 " -O <bool> : Use O_DIRECT, default %d\n"
1526 " -N <bool> : Perform just no-op requests, default %d\n"
1527 " -t <bool> : Track IO latencies, default %d\n"
1528 " -T <int> : TSC rate in HZ\n"
1529 " -r <int> : Runtime in seconds, default %s\n"
1530 " -R <bool> : Use random IO, default %d\n"
1531 " -a <bool> : Use legacy aio, default %d\n"
1532 " -S <bool> : Use sync IO (preadv2), default %d\n"
1533 " -X <bool> : Use registered ring %d\n"
1534 " -P <bool> : Automatically place on device home node %d\n"
1535 " -u <bool> : Use nvme-passthrough I/O, default %d\n",
1536 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled,
1537 fixedbufs, dma_map, register_files, nthreads, !buffered, do_nop,
1538 stats, runtime == 0 ? "unlimited" : runtime_str, random_io, aio,
1539 use_sync, register_ring, numa_placement, pt);
1543 static void read_tsc_rate(void)
1551 fd = open(TSC_RATE_FILE, O_RDONLY);
1555 ret = read(fd, buffer, sizeof(buffer));
1561 tsc_rate = strtoul(buffer, NULL, 10);
1562 printf("Using TSC rate %luHz\n", tsc_rate);
1566 static void write_tsc_rate(void)
1572 if (!stat(TSC_RATE_FILE, &sb))
1575 fd = open(TSC_RATE_FILE, O_WRONLY | O_CREAT, 0644);
1579 memset(buffer, 0, sizeof(buffer));
1580 sprintf(buffer, "%lu", tsc_rate);
1581 ret = write(fd, buffer, strlen(buffer));
1587 int main(int argc, char *argv[])
1589 struct submitter *s;
1590 unsigned long done, calls, reap;
1591 int i, j, flags, fd, opt, threads_per_f, threads_rem = 0, nfiles;
1595 if (!do_nop && argc < 2)
1598 while ((opt = getopt(argc, argv, "d:s:c:b:p:B:F:n:N:O:t:T:a:r:D:R:X:S:P:u:h?")) != -1) {
1601 aio = !!atoi(optarg);
1604 depth = atoi(optarg);
1607 batch_submit = atoi(optarg);
1612 batch_complete = atoi(optarg);
1613 if (!batch_complete)
1620 polled = !!atoi(optarg);
1623 fixedbufs = !!atoi(optarg);
1626 register_files = !!atoi(optarg);
1629 nthreads = atoi(optarg);
1631 printf("Threads must be non-zero\n");
1636 do_nop = !!atoi(optarg);
1639 buffered = !atoi(optarg);
1642 #ifndef ARCH_HAVE_CPU_CLOCK
1643 fprintf(stderr, "Stats not supported on this CPU\n");
1646 stats = !!atoi(optarg);
1649 #ifndef ARCH_HAVE_CPU_CLOCK
1650 fprintf(stderr, "Stats not supported on this CPU\n");
1653 tsc_rate = strtoul(optarg, NULL, 10);
1657 runtime = atoi(optarg);
1660 dma_map = !!atoi(optarg);
1663 random_io = !!atoi(optarg);
1666 register_ring = !!atoi(optarg);
1669 #ifdef CONFIG_PWRITEV2
1670 use_sync = !!atoi(optarg);
1672 fprintf(stderr, "preadv2 not supported\n");
1677 numa_placement = !!atoi(optarg);
1680 pt = !!atoi(optarg);
1693 if (batch_complete > depth)
1694 batch_complete = depth;
1695 if (batch_submit > depth)
1696 batch_submit = depth;
1697 if (!fixedbufs && dma_map)
1700 submitter = calloc(nthreads, sizeof(*submitter) +
1701 roundup_pow2(depth) * sizeof(struct iovec));
1702 for (j = 0; j < nthreads; j++) {
1703 s = get_submitter(j);
1706 s->done = s->calls = s->reaps = 0;
1709 flags = O_RDONLY | O_NOATIME;
1718 printf("No files specified\n");
1721 threads_per_f = nthreads / nfiles;
1722 /* make sure each thread gets assigned files */
1723 if (threads_per_f == 0) {
1726 threads_rem = nthreads - threads_per_f * nfiles;
1729 while (!do_nop && i < argc) {
1732 memset(&f, 0, sizeof(f));
1734 fd = open(argv[i], flags);
1740 if (get_file_size(&f)) {
1741 printf("failed getting size of device/file\n");
1744 if (f.max_blocks <= 1) {
1745 printf("Zero file/device size?\n");
1750 limit = threads_per_f;
1751 limit += threads_rem > 0 ? 1 : 0;
1752 for (k = 0; k < limit; k++) {
1753 s = get_submitter((j + k) % nthreads);
1755 if (s->nr_files == MAX_FDS) {
1756 printf("Max number of files (%d) reached\n", MAX_FDS);
1760 memcpy(&s->files[s->nr_files], &f, sizeof(f));
1763 detect_node(s, argv[i]);
1765 s->filename = argv[i];
1775 t_io_uring_page_size = sysconf(_SC_PAGESIZE);
1776 if (t_io_uring_page_size < 0)
1777 t_io_uring_page_size = 4096;
1779 for (j = 0; j < nthreads; j++) {
1780 s = get_submitter(j);
1782 pthread_create(&s->thread, NULL, submitter_sync_fn, s);
1784 pthread_create(&s->thread, NULL, submitter_uring_fn, s);
1785 #ifdef CONFIG_LIBAIO
1787 pthread_create(&s->thread, NULL, submitter_aio_fn, s);
1791 reap = calls = done = 0;
1793 unsigned long this_done = 0;
1794 unsigned long this_reap = 0;
1795 unsigned long this_call = 0;
1796 unsigned long rpc = 0, ipc = 0;
1797 unsigned long iops, bw;
1800 if (runtime && !--runtime)
1801 do_finish("timeout");
1803 /* don't print partial run, if interrupted by signal */
1807 /* one second in to the run, enable stats */
1811 for (j = 0; j < nthreads; j++) {
1812 s = get_submitter(j);
1813 this_done += s->done;
1814 this_call += s->calls;
1815 this_reap += s->reaps;
1817 if (this_call - calls) {
1818 rpc = (this_done - done) / (this_call - calls);
1819 ipc = (this_reap - reap) / (this_call - calls);
1822 iops = this_done - done;
1824 bw = iops * (bs / 1048576);
1826 bw = iops / (1048576 / bs);
1827 if (iops > 1000000) {
1828 double miops = (double) iops / 1000000.0;
1829 printf("IOPS=%.2fM, ", miops);
1830 } else if (iops > 100000) {
1831 double kiops = (double) iops / 1000.0;
1832 printf("IOPS=%.2fK, ", kiops);
1834 printf("IOPS=%lu, ", iops);
1836 max_iops = max(max_iops, iops);
1839 double bw_g = (double) bw / 1000.0;
1841 printf("BW=%.2fGiB/s, ", bw_g);
1843 printf("BW=%luMiB/s, ", bw);
1846 printf("IOS/call=%ld/%ld\n", rpc, ipc);
1852 for (j = 0; j < nthreads; j++) {
1853 s = get_submitter(j);
1854 pthread_join(s->thread, &ret);
1860 printf("%d: Latency percentiles:\n", s->tid);
1861 for (i = 0, nr = 0; i < PLAT_NR; i++)
1863 show_clat_percentiles(s->plat, nr, 4);
1864 free(s->clock_batch);