11 #include <sys/ioctl.h>
12 #include <sys/syscall.h>
13 #include <sys/resource.h>
23 #include "../arch/arch.h"
24 #include "../lib/types.h"
25 #include "../os/linux/io_uring.h"
27 #define min(a, b) ((a < b) ? (a) : (b))
33 unsigned *ring_entries;
42 unsigned *ring_entries;
43 struct io_uring_cqe *cqes;
47 #define BATCH_SUBMIT 32
48 #define BATCH_COMPLETE 32
54 static unsigned sq_ring_mask, cq_ring_mask;
57 unsigned long max_blocks;
66 struct drand48_data rand;
67 struct io_sq_ring sq_ring;
68 struct io_uring_sqe *sqes;
69 struct io_cq_ring cq_ring;
78 struct file files[MAX_FDS];
81 struct iovec iovecs[];
84 static struct submitter *submitter;
85 static volatile int finish;
87 static int depth = DEPTH;
88 static int batch_submit = BATCH_SUBMIT;
89 static int batch_complete = BATCH_COMPLETE;
90 static int polled = 1; /* use IO polling */
91 static int fixedbufs = 1; /* use fixed user buffers */
92 static int register_files = 1; /* use fixed files */
93 static int buffered = 0; /* use buffered IO, not O_DIRECT */
94 static int sq_thread_poll = 0; /* use kernel submission/poller thread */
95 static int sq_thread_cpu = -1; /* pin above thread to this CPU */
96 static int do_nop = 0; /* no-op SQ ring commands */
98 static int io_uring_register_buffers(struct submitter *s)
103 return syscall(__NR_sys_io_uring_register, s->ring_fd,
104 IORING_REGISTER_BUFFERS, s->iovecs, depth);
107 static int io_uring_register_files(struct submitter *s)
114 s->fds = calloc(s->nr_files, sizeof(__s32));
115 for (i = 0; i < s->nr_files; i++) {
116 s->fds[i] = s->files[i].real_fd;
117 s->files[i].fixed_fd = i;
120 return syscall(__NR_sys_io_uring_register, s->ring_fd,
121 IORING_REGISTER_FILES, s->fds, s->nr_files);
124 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
126 return syscall(__NR_sys_io_uring_setup, entries, p);
129 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
130 unsigned int min_complete, unsigned int flags)
132 return syscall(__NR_sys_io_uring_enter, s->ring_fd, to_submit,
133 min_complete, flags, NULL, 0);
136 static int gettid(void)
138 return syscall(__NR_gettid);
141 static unsigned file_depth(struct submitter *s)
143 return (depth + s->nr_files - 1) / s->nr_files;
146 static void init_io(struct submitter *s, unsigned index)
148 struct io_uring_sqe *sqe = &s->sqes[index];
149 unsigned long offset;
154 sqe->opcode = IORING_OP_NOP;
158 if (s->nr_files == 1) {
161 f = &s->files[s->cur_file];
162 if (f->pending_ios >= file_depth(s)) {
164 if (s->cur_file == s->nr_files)
166 f = &s->files[s->cur_file];
171 lrand48_r(&s->rand, &r);
172 offset = (r % (f->max_blocks - 1)) * BS;
174 if (register_files) {
175 sqe->flags = IOSQE_FIXED_FILE;
176 sqe->fd = f->fixed_fd;
179 sqe->fd = f->real_fd;
182 sqe->opcode = IORING_OP_READ_FIXED;
183 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
185 sqe->buf_index = index;
187 sqe->opcode = IORING_OP_READV;
188 sqe->addr = (unsigned long) &s->iovecs[index];
194 sqe->user_data = (unsigned long) f;
197 static int prep_more_ios(struct submitter *s, int max_ios)
199 struct io_sq_ring *ring = &s->sq_ring;
200 unsigned index, tail, next_tail, prepped = 0;
202 next_tail = tail = *ring->tail;
206 if (next_tail == *ring->head)
209 index = tail & sq_ring_mask;
211 ring->array[index] = index;
214 } while (prepped < max_ios);
216 if (*ring->tail != tail) {
217 /* order tail store with writes to sqes above */
225 static int get_file_size(struct file *f)
229 if (fstat(f->real_fd, &st) < 0)
231 if (S_ISBLK(st.st_mode)) {
232 unsigned long long bytes;
234 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
237 f->max_blocks = bytes / BS;
239 } else if (S_ISREG(st.st_mode)) {
240 f->max_blocks = st.st_size / BS;
247 static int reap_events(struct submitter *s)
249 struct io_cq_ring *ring = &s->cq_ring;
250 struct io_uring_cqe *cqe;
251 unsigned head, reaped = 0;
258 if (head == *ring->tail)
260 cqe = &ring->cqes[head & cq_ring_mask];
262 f = (struct file *) (uintptr_t) cqe->user_data;
264 if (cqe->res != BS) {
265 printf("io: unexpected ret=%d\n", cqe->res);
266 if (polled && cqe->res == -EOPNOTSUPP)
267 printf("Your filesystem/driver/kernel doesn't support polled IO\n");
275 s->inflight -= reaped;
281 static void *submitter_fn(void *data)
283 struct submitter *s = data;
284 struct io_sq_ring *ring = &s->sq_ring;
287 printf("submitter=%d\n", gettid());
289 srand48_r(pthread_self(), &s->rand);
293 int to_wait, to_submit, this_reap, to_prep;
295 if (!prepped && s->inflight < depth) {
296 to_prep = min(depth - s->inflight, batch_submit);
297 prepped = prep_more_ios(s, to_prep);
299 s->inflight += prepped;
303 if (to_submit && (s->inflight + to_submit <= depth))
306 to_wait = min(s->inflight + to_submit, batch_complete);
309 * Only need to call io_uring_enter if we're not using SQ thread
310 * poll, or if IORING_SQ_NEED_WAKEUP is set.
312 if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) {
316 flags = IORING_ENTER_GETEVENTS;
317 if ((*ring->flags & IORING_SQ_NEED_WAKEUP))
318 flags |= IORING_ENTER_SQ_WAKEUP;
319 ret = io_uring_enter(s, to_submit, to_wait, flags);
324 * For non SQ thread poll, we already got the events we needed
325 * through the io_uring_enter() above. For SQ thread poll, we
326 * need to loop here until we find enough events.
337 } while (sq_thread_poll && this_reap < to_wait);
338 s->reaps += this_reap;
346 } else if (ret < to_submit) {
347 int diff = to_submit - ret;
356 } else if (ret < 0) {
357 if (errno == EAGAIN) {
365 printf("io_submit: %s\n", strerror(errno));
368 } while (!s->finish);
374 static void sig_int(int sig)
376 printf("Exiting on signal %d\n", sig);
377 submitter->finish = 1;
381 static void arm_sig_int(void)
383 struct sigaction act;
385 memset(&act, 0, sizeof(act));
386 act.sa_handler = sig_int;
387 act.sa_flags = SA_RESTART;
388 sigaction(SIGINT, &act, NULL);
391 static int setup_ring(struct submitter *s)
393 struct io_sq_ring *sring = &s->sq_ring;
394 struct io_cq_ring *cring = &s->cq_ring;
395 struct io_uring_params p;
399 memset(&p, 0, sizeof(p));
401 if (polled && !do_nop)
402 p.flags |= IORING_SETUP_IOPOLL;
403 if (sq_thread_poll) {
404 p.flags |= IORING_SETUP_SQPOLL;
405 if (sq_thread_cpu != -1) {
406 p.flags |= IORING_SETUP_SQ_AFF;
407 p.sq_thread_cpu = sq_thread_cpu;
411 fd = io_uring_setup(depth, &p);
413 perror("io_uring_setup");
419 ret = io_uring_register_buffers(s);
421 perror("io_uring_register_buffers");
426 if (register_files) {
427 ret = io_uring_register_files(s);
429 perror("io_uring_register_files");
434 ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
435 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
437 printf("sq_ring ptr = 0x%p\n", ptr);
438 sring->head = ptr + p.sq_off.head;
439 sring->tail = ptr + p.sq_off.tail;
440 sring->ring_mask = ptr + p.sq_off.ring_mask;
441 sring->ring_entries = ptr + p.sq_off.ring_entries;
442 sring->flags = ptr + p.sq_off.flags;
443 sring->array = ptr + p.sq_off.array;
444 sq_ring_mask = *sring->ring_mask;
446 s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
447 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
449 printf("sqes ptr = 0x%p\n", s->sqes);
451 ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
452 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
454 printf("cq_ring ptr = 0x%p\n", ptr);
455 cring->head = ptr + p.cq_off.head;
456 cring->tail = ptr + p.cq_off.tail;
457 cring->ring_mask = ptr + p.cq_off.ring_mask;
458 cring->ring_entries = ptr + p.cq_off.ring_entries;
459 cring->cqes = ptr + p.cq_off.cqes;
460 cq_ring_mask = *cring->ring_mask;
464 static void file_depths(char *buf)
466 struct submitter *s = submitter;
472 for (i = 0; i < s->nr_files; i++) {
473 struct file *f = &s->files[i];
475 if (i + 1 == s->nr_files)
476 p += sprintf(p, "%d", f->pending_ios);
478 p += sprintf(p, "%d, ", f->pending_ios);
482 static void usage(char *argv)
484 printf("%s [options] -- [filenames]\n"
485 " -d <int> : IO Depth, default %d\n"
486 " -s <int> : Batch submit, default %d\n"
487 " -c <int> : Batch complete, default %d\n",
488 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE);
492 int main(int argc, char *argv[])
495 unsigned long done, calls, reap;
496 int err, i, flags, fd, opt;
500 if (!do_nop && argc < 2) {
501 printf("%s: filename [options]\n", argv[0]);
505 while ((opt = getopt(argc, argv, "d:s:c:h?")) != -1) {
508 depth = atoi(optarg);
511 batch_submit = atoi(optarg);
514 batch_complete = atoi(optarg);
524 submitter = malloc(sizeof(*submitter) + depth * sizeof(struct iovec));
525 memset(submitter, 0, sizeof(*submitter) + depth * sizeof(struct iovec));
528 flags = O_RDONLY | O_NOATIME;
533 while (!do_nop && i < argc) {
536 if (s->nr_files == MAX_FDS) {
537 printf("Max number of files (%d) reached\n", MAX_FDS);
540 fd = open(argv[i], flags);
546 f = &s->files[s->nr_files];
548 if (get_file_size(f)) {
549 printf("failed getting size of device/file\n");
552 if (f->max_blocks <= 1) {
553 printf("Zero file/device size?\n");
558 printf("Added file %s\n", argv[i]);
566 rlim.rlim_cur = RLIM_INFINITY;
567 rlim.rlim_max = RLIM_INFINITY;
568 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
576 for (i = 0; i < depth; i++) {
579 if (posix_memalign(&buf, BS, BS)) {
580 printf("failed alloc\n");
583 s->iovecs[i].iov_base = buf;
584 s->iovecs[i].iov_len = BS;
589 printf("ring setup failed: %s, %d\n", strerror(errno), err);
592 printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
593 printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", depth, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
595 pthread_create(&s->thread, NULL, submitter_fn, s);
597 fdepths = malloc(8 * s->nr_files);
598 reap = calls = done = 0;
600 unsigned long this_done = 0;
601 unsigned long this_reap = 0;
602 unsigned long this_call = 0;
603 unsigned long rpc = 0, ipc = 0;
606 this_done += s->done;
607 this_call += s->calls;
608 this_reap += s->reaps;
609 if (this_call - calls) {
610 rpc = (this_done - done) / (this_call - calls);
611 ipc = (this_reap - reap) / (this_call - calls);
614 file_depths(fdepths);
615 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s)\n",
616 this_done - done, rpc, ipc, s->inflight,
623 pthread_join(s->thread, &ret);