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 vectored = 1;
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))
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 (!vectored) {
213 sqe->opcode = IORING_OP_READ;
214 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
218 sqe->opcode = IORING_OP_READV;
219 sqe->addr = (unsigned long) &s->iovecs[index];
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;
236 if (next_tail == atomic_load_acquire(ring->head))
239 index = tail & sq_ring_mask;
241 ring->array[index] = index;
244 } while (prepped < max_ios);
247 atomic_store_release(ring->tail, tail);
251 static int get_file_size(struct file *f)
255 if (fstat(f->real_fd, &st) < 0)
257 if (S_ISBLK(st.st_mode)) {
258 unsigned long long bytes;
260 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
263 f->max_blocks = bytes / bs;
265 } else if (S_ISREG(st.st_mode)) {
266 f->max_blocks = st.st_size / bs;
273 static int reap_events(struct submitter *s)
275 struct io_cq_ring *ring = &s->cq_ring;
276 struct io_uring_cqe *cqe;
277 unsigned head, reaped = 0;
284 if (head == atomic_load_acquire(ring->tail))
286 cqe = &ring->cqes[head & cq_ring_mask];
288 f = (struct file *) (uintptr_t) cqe->user_data;
290 if (cqe->res != bs) {
291 printf("io: unexpected ret=%d\n", cqe->res);
292 if (polled && cqe->res == -EOPNOTSUPP)
293 printf("Your filesystem/driver/kernel doesn't support polled IO\n");
302 s->inflight -= reaped;
303 atomic_store_release(ring->head, head);
308 static void *submitter_fn(void *data)
310 struct submitter *s = data;
311 struct io_sq_ring *ring = &s->sq_ring;
314 printf("submitter=%d\n", gettid());
316 srand48(pthread_self());
320 int to_wait, to_submit, this_reap, to_prep;
321 unsigned ring_flags = 0;
323 if (!prepped && s->inflight < depth) {
324 to_prep = min(depth - s->inflight, batch_submit);
325 prepped = prep_more_ios(s, to_prep);
327 s->inflight += prepped;
331 if (to_submit && (s->inflight + to_submit <= depth))
334 to_wait = min(s->inflight + to_submit, batch_complete);
337 * Only need to call io_uring_enter if we're not using SQ thread
338 * poll, or if IORING_SQ_NEED_WAKEUP is set.
341 ring_flags = atomic_load_acquire(ring->flags);
342 if (!sq_thread_poll || ring_flags & IORING_SQ_NEED_WAKEUP) {
346 flags = IORING_ENTER_GETEVENTS;
347 if (ring_flags & IORING_SQ_NEED_WAKEUP)
348 flags |= IORING_ENTER_SQ_WAKEUP;
349 ret = io_uring_enter(s, to_submit, to_wait, flags);
352 /* for SQPOLL, we submitted it all effectively */
357 * For non SQ thread poll, we already got the events we needed
358 * through the io_uring_enter() above. For SQ thread poll, we
359 * need to loop here until we find enough events.
370 } while (sq_thread_poll && this_reap < to_wait);
371 s->reaps += this_reap;
379 } else if (ret < to_submit) {
380 int diff = to_submit - ret;
389 } else if (ret < 0) {
390 if (errno == EAGAIN) {
398 printf("io_submit: %s\n", strerror(errno));
401 } while (!s->finish);
407 static void sig_int(int sig)
409 printf("Exiting on signal %d\n", sig);
410 submitter->finish = 1;
414 static void arm_sig_int(void)
416 struct sigaction act;
418 memset(&act, 0, sizeof(act));
419 act.sa_handler = sig_int;
420 act.sa_flags = SA_RESTART;
421 sigaction(SIGINT, &act, NULL);
424 static int setup_ring(struct submitter *s)
426 struct io_sq_ring *sring = &s->sq_ring;
427 struct io_cq_ring *cring = &s->cq_ring;
428 struct io_uring_params p;
432 memset(&p, 0, sizeof(p));
434 if (polled && !do_nop)
435 p.flags |= IORING_SETUP_IOPOLL;
436 if (sq_thread_poll) {
437 p.flags |= IORING_SETUP_SQPOLL;
438 if (sq_thread_cpu != -1) {
439 p.flags |= IORING_SETUP_SQ_AFF;
440 p.sq_thread_cpu = sq_thread_cpu;
444 fd = io_uring_setup(depth, &p);
446 perror("io_uring_setup");
454 ret = io_uring_register_buffers(s);
456 perror("io_uring_register_buffers");
461 if (register_files) {
462 ret = io_uring_register_files(s);
464 perror("io_uring_register_files");
469 ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
470 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
472 printf("sq_ring ptr = 0x%p\n", ptr);
473 sring->head = ptr + p.sq_off.head;
474 sring->tail = ptr + p.sq_off.tail;
475 sring->ring_mask = ptr + p.sq_off.ring_mask;
476 sring->ring_entries = ptr + p.sq_off.ring_entries;
477 sring->flags = ptr + p.sq_off.flags;
478 sring->array = ptr + p.sq_off.array;
479 sq_ring_mask = *sring->ring_mask;
481 s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
482 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
484 printf("sqes ptr = 0x%p\n", s->sqes);
486 ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
487 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
489 printf("cq_ring ptr = 0x%p\n", ptr);
490 cring->head = ptr + p.cq_off.head;
491 cring->tail = ptr + p.cq_off.tail;
492 cring->ring_mask = ptr + p.cq_off.ring_mask;
493 cring->ring_entries = ptr + p.cq_off.ring_entries;
494 cring->cqes = ptr + p.cq_off.cqes;
495 cq_ring_mask = *cring->ring_mask;
499 static void file_depths(char *buf)
501 struct submitter *s = submitter;
507 for (i = 0; i < s->nr_files; i++) {
508 struct file *f = &s->files[i];
510 if (i + 1 == s->nr_files)
511 p += sprintf(p, "%d", f->pending_ios);
513 p += sprintf(p, "%d, ", f->pending_ios);
517 static void usage(char *argv)
519 printf("%s [options] -- [filenames]\n"
520 " -d <int> : IO Depth, default %d\n"
521 " -s <int> : Batch submit, default %d\n"
522 " -c <int> : Batch complete, default %d\n"
523 " -b <int> : Block size, default %d\n"
524 " -p <bool> : Polled IO, default %d\n",
525 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE, BS, polled);
529 int main(int argc, char *argv[])
532 unsigned long done, calls, reap;
533 int err, i, flags, fd, opt;
537 if (!do_nop && argc < 2) {
538 printf("%s: filename [options]\n", argv[0]);
542 while ((opt = getopt(argc, argv, "d:s:c:b:p:B:F:h?")) != -1) {
545 depth = atoi(optarg);
548 batch_submit = atoi(optarg);
551 batch_complete = atoi(optarg);
557 polled = !!atoi(optarg);
560 fixedbufs = !!atoi(optarg);
563 register_files = !!atoi(optarg);
573 submitter = malloc(sizeof(*submitter) + depth * sizeof(struct iovec));
574 memset(submitter, 0, sizeof(*submitter) + depth * sizeof(struct iovec));
577 flags = O_RDONLY | O_NOATIME;
582 while (!do_nop && i < argc) {
585 if (s->nr_files == MAX_FDS) {
586 printf("Max number of files (%d) reached\n", MAX_FDS);
589 fd = open(argv[i], flags);
595 f = &s->files[s->nr_files];
597 if (get_file_size(f)) {
598 printf("failed getting size of device/file\n");
601 if (f->max_blocks <= 1) {
602 printf("Zero file/device size?\n");
607 printf("Added file %s\n", argv[i]);
615 rlim.rlim_cur = RLIM_INFINITY;
616 rlim.rlim_max = RLIM_INFINITY;
617 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
625 for (i = 0; i < depth; i++) {
628 if (posix_memalign(&buf, bs, bs)) {
629 printf("failed alloc\n");
632 s->iovecs[i].iov_base = buf;
633 s->iovecs[i].iov_len = bs;
638 printf("ring setup failed: %s, %d\n", strerror(errno), err);
641 printf("polled=%d, fixedbufs=%d, register_files=%d, buffered=%d", polled, fixedbufs, register_files, buffered);
642 printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", depth, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
644 pthread_create(&s->thread, NULL, submitter_fn, s);
646 fdepths = malloc(8 * s->nr_files);
647 reap = calls = done = 0;
649 unsigned long this_done = 0;
650 unsigned long this_reap = 0;
651 unsigned long this_call = 0;
652 unsigned long rpc = 0, ipc = 0;
655 this_done += s->done;
656 this_call += s->calls;
657 this_reap += s->reaps;
658 if (this_call - calls) {
659 rpc = (this_done - done) / (this_call - calls);
660 ipc = (this_reap - reap) / (this_call - calls);
663 file_depths(fdepths);
664 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s)\n",
665 this_done - done, rpc, ipc, s->inflight,
672 pthread_join(s->thread, &ret);