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
53 static unsigned sq_ring_mask, cq_ring_mask;
56 unsigned long max_blocks;
65 struct io_sq_ring sq_ring;
66 struct io_uring_sqe *sqes;
67 struct io_cq_ring cq_ring;
76 struct file files[MAX_FDS];
79 struct iovec iovecs[];
82 static struct submitter *submitter;
83 static volatile int finish;
85 static int depth = DEPTH;
86 static int batch_submit = BATCH_SUBMIT;
87 static int batch_complete = BATCH_COMPLETE;
89 static int polled = 1; /* use IO polling */
90 static int fixedbufs = 1; /* use fixed user buffers */
91 static int register_files = 1; /* use fixed files */
92 static int buffered = 0; /* use buffered IO, not O_DIRECT */
93 static int sq_thread_poll = 0; /* use kernel submission/poller thread */
94 static int sq_thread_cpu = -1; /* pin above thread to this CPU */
95 static int do_nop = 0; /* no-op SQ ring commands */
97 static int read_op = IORING_OP_READV;
99 static int io_uring_register_buffers(struct submitter *s)
104 return syscall(__NR_io_uring_register, s->ring_fd,
105 IORING_REGISTER_BUFFERS, s->iovecs, depth);
108 static int io_uring_register_files(struct submitter *s)
115 s->fds = calloc(s->nr_files, sizeof(__s32));
116 for (i = 0; i < s->nr_files; i++) {
117 s->fds[i] = s->files[i].real_fd;
118 s->files[i].fixed_fd = i;
121 return syscall(__NR_io_uring_register, s->ring_fd,
122 IORING_REGISTER_FILES, s->fds, s->nr_files);
125 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
127 return syscall(__NR_io_uring_setup, entries, p);
130 static void io_uring_probe(int fd)
132 struct io_uring_probe *p;
135 p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
139 memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
140 ret = syscall(__NR_io_uring_register, fd, IORING_REGISTER_PROBE, p, 256);
144 if (IORING_OP_READ > p->ops_len)
147 if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED))
148 read_op = IORING_OP_READ;
153 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
154 unsigned int min_complete, unsigned int flags)
156 return syscall(__NR_io_uring_enter, s->ring_fd, to_submit, min_complete,
160 #ifndef CONFIG_HAVE_GETTID
161 static int gettid(void)
163 return syscall(__NR_gettid);
167 static unsigned file_depth(struct submitter *s)
169 return (depth + s->nr_files - 1) / s->nr_files;
172 static void init_io(struct submitter *s, unsigned index)
174 struct io_uring_sqe *sqe = &s->sqes[index];
175 unsigned long offset;
180 sqe->opcode = IORING_OP_NOP;
184 if (s->nr_files == 1) {
187 f = &s->files[s->cur_file];
188 if (f->pending_ios >= file_depth(s)) {
190 if (s->cur_file == s->nr_files)
192 f = &s->files[s->cur_file];
198 offset = (r % (f->max_blocks - 1)) * bs;
200 if (register_files) {
201 sqe->flags = IOSQE_FIXED_FILE;
202 sqe->fd = f->fixed_fd;
205 sqe->fd = f->real_fd;
208 sqe->opcode = IORING_OP_READ_FIXED;
209 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
211 sqe->buf_index = index;
212 } else if (read_op == IORING_OP_READV) {
213 sqe->opcode = IORING_OP_READV;
214 sqe->addr = (unsigned long) &s->iovecs[index];
218 sqe->opcode = IORING_OP_READ;
219 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
225 sqe->user_data = (unsigned long) f;
228 static int prep_more_ios(struct submitter *s, int max_ios)
230 struct io_sq_ring *ring = &s->sq_ring;
231 unsigned index, tail, next_tail, prepped = 0;
233 next_tail = tail = *ring->tail;
237 if (next_tail == *ring->head)
240 index = tail & sq_ring_mask;
242 ring->array[index] = index;
245 } while (prepped < max_ios);
247 if (*ring->tail != tail) {
254 static int get_file_size(struct file *f)
258 if (fstat(f->real_fd, &st) < 0)
260 if (S_ISBLK(st.st_mode)) {
261 unsigned long long bytes;
263 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
266 f->max_blocks = bytes / bs;
268 } else if (S_ISREG(st.st_mode)) {
269 f->max_blocks = st.st_size / bs;
276 static int reap_events(struct submitter *s)
278 struct io_cq_ring *ring = &s->cq_ring;
279 struct io_uring_cqe *cqe;
280 unsigned head, reaped = 0;
287 if (head == *ring->tail)
289 cqe = &ring->cqes[head & cq_ring_mask];
291 f = (struct file *) (uintptr_t) cqe->user_data;
293 if (cqe->res != bs) {
294 printf("io: unexpected ret=%d\n", cqe->res);
295 if (polled && cqe->res == -EOPNOTSUPP)
296 printf("Your filesystem/driver/kernel doesn't support polled IO\n");
304 s->inflight -= reaped;
310 static void *submitter_fn(void *data)
312 struct submitter *s = data;
313 struct io_sq_ring *ring = &s->sq_ring;
316 printf("submitter=%d\n", gettid());
318 srand48(pthread_self());
322 int to_wait, to_submit, this_reap, to_prep;
324 if (!prepped && s->inflight < depth) {
325 to_prep = min(depth - s->inflight, batch_submit);
326 prepped = prep_more_ios(s, to_prep);
328 s->inflight += prepped;
332 if (to_submit && (s->inflight + to_submit <= depth))
335 to_wait = min(s->inflight + to_submit, batch_complete);
338 * Only need to call io_uring_enter if we're not using SQ thread
339 * poll, or if IORING_SQ_NEED_WAKEUP is set.
341 if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) {
345 flags = IORING_ENTER_GETEVENTS;
346 if ((*ring->flags & IORING_SQ_NEED_WAKEUP))
347 flags |= IORING_ENTER_SQ_WAKEUP;
348 ret = io_uring_enter(s, to_submit, to_wait, flags);
353 * For non SQ thread poll, we already got the events we needed
354 * through the io_uring_enter() above. For SQ thread poll, we
355 * need to loop here until we find enough events.
366 } while (sq_thread_poll && this_reap < to_wait);
367 s->reaps += this_reap;
375 } else if (ret < to_submit) {
376 int diff = to_submit - ret;
385 } else if (ret < 0) {
386 if (errno == EAGAIN) {
394 printf("io_submit: %s\n", strerror(errno));
397 } while (!s->finish);
403 static void sig_int(int sig)
405 printf("Exiting on signal %d\n", sig);
406 submitter->finish = 1;
410 static void arm_sig_int(void)
412 struct sigaction act;
414 memset(&act, 0, sizeof(act));
415 act.sa_handler = sig_int;
416 act.sa_flags = SA_RESTART;
417 sigaction(SIGINT, &act, NULL);
420 static int setup_ring(struct submitter *s)
422 struct io_sq_ring *sring = &s->sq_ring;
423 struct io_cq_ring *cring = &s->cq_ring;
424 struct io_uring_params p;
428 memset(&p, 0, sizeof(p));
430 if (polled && !do_nop)
431 p.flags |= IORING_SETUP_IOPOLL;
432 if (sq_thread_poll) {
433 p.flags |= IORING_SETUP_SQPOLL;
434 if (sq_thread_cpu != -1) {
435 p.flags |= IORING_SETUP_SQ_AFF;
436 p.sq_thread_cpu = sq_thread_cpu;
440 fd = io_uring_setup(depth, &p);
442 perror("io_uring_setup");
450 ret = io_uring_register_buffers(s);
452 perror("io_uring_register_buffers");
457 if (register_files) {
458 ret = io_uring_register_files(s);
460 perror("io_uring_register_files");
465 ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
466 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
468 printf("sq_ring ptr = 0x%p\n", ptr);
469 sring->head = ptr + p.sq_off.head;
470 sring->tail = ptr + p.sq_off.tail;
471 sring->ring_mask = ptr + p.sq_off.ring_mask;
472 sring->ring_entries = ptr + p.sq_off.ring_entries;
473 sring->flags = ptr + p.sq_off.flags;
474 sring->array = ptr + p.sq_off.array;
475 sq_ring_mask = *sring->ring_mask;
477 s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
478 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
480 printf("sqes ptr = 0x%p\n", s->sqes);
482 ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
483 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
485 printf("cq_ring ptr = 0x%p\n", ptr);
486 cring->head = ptr + p.cq_off.head;
487 cring->tail = ptr + p.cq_off.tail;
488 cring->ring_mask = ptr + p.cq_off.ring_mask;
489 cring->ring_entries = ptr + p.cq_off.ring_entries;
490 cring->cqes = ptr + p.cq_off.cqes;
491 cq_ring_mask = *cring->ring_mask;
495 static void file_depths(char *buf)
497 struct submitter *s = submitter;
503 for (i = 0; i < s->nr_files; i++) {
504 struct file *f = &s->files[i];
506 if (i + 1 == s->nr_files)
507 p += sprintf(p, "%d", f->pending_ios);
509 p += sprintf(p, "%d, ", f->pending_ios);
513 static void usage(char *argv)
515 printf("%s [options] -- [filenames]\n"
516 " -d <int> : IO Depth, default %d\n"
517 " -s <int> : Batch submit, default %d\n"
518 " -c <int> : Batch complete, default %d\n"
519 " -b <int> : Block size, default %d\n"
520 " -p <bool> : Polled IO, default %d\n",
521 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled);
525 int main(int argc, char *argv[])
528 unsigned long done, calls, reap;
529 int err, i, flags, fd, opt;
533 if (!do_nop && argc < 2) {
534 printf("%s: filename [options]\n", argv[0]);
538 while ((opt = getopt(argc, argv, "d:s:c:b:p:h?")) != -1) {
541 depth = atoi(optarg);
544 batch_submit = atoi(optarg);
547 batch_complete = atoi(optarg);
553 polled = !!atoi(optarg);
563 submitter = malloc(sizeof(*submitter) + depth * sizeof(struct iovec));
564 memset(submitter, 0, sizeof(*submitter) + depth * sizeof(struct iovec));
567 flags = O_RDONLY | O_NOATIME;
572 while (!do_nop && i < argc) {
575 if (s->nr_files == MAX_FDS) {
576 printf("Max number of files (%d) reached\n", MAX_FDS);
579 fd = open(argv[i], flags);
585 f = &s->files[s->nr_files];
587 if (get_file_size(f)) {
588 printf("failed getting size of device/file\n");
591 if (f->max_blocks <= 1) {
592 printf("Zero file/device size?\n");
597 printf("Added file %s\n", argv[i]);
605 rlim.rlim_cur = RLIM_INFINITY;
606 rlim.rlim_max = RLIM_INFINITY;
607 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
615 for (i = 0; i < depth; i++) {
618 if (posix_memalign(&buf, bs, bs)) {
619 printf("failed alloc\n");
622 s->iovecs[i].iov_base = buf;
623 s->iovecs[i].iov_len = bs;
628 printf("ring setup failed: %s, %d\n", strerror(errno), err);
631 printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
632 printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", depth, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
634 pthread_create(&s->thread, NULL, submitter_fn, s);
636 fdepths = malloc(8 * s->nr_files);
637 reap = calls = done = 0;
639 unsigned long this_done = 0;
640 unsigned long this_reap = 0;
641 unsigned long this_call = 0;
642 unsigned long rpc = 0, ipc = 0;
645 this_done += s->done;
646 this_call += s->calls;
647 this_reap += s->reaps;
648 if (this_call - calls) {
649 rpc = (this_done - done) / (this_call - calls);
650 ipc = (this_reap - reap) / (this_call - calls);
653 file_depths(fdepths);
654 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s)\n",
655 this_done - done, rpc, ipc, s->inflight,
662 pthread_join(s->thread, &ret);