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 io_sq_ring sq_ring;
67 struct io_uring_sqe *sqes;
68 struct io_cq_ring cq_ring;
77 struct file files[MAX_FDS];
80 struct iovec iovecs[];
83 static struct submitter *submitter;
84 static volatile int finish;
86 static int depth = DEPTH;
87 static int batch_submit = BATCH_SUBMIT;
88 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 io_uring_register_buffers(struct submitter *s)
102 return syscall(__NR_io_uring_register, s->ring_fd,
103 IORING_REGISTER_BUFFERS, s->iovecs, depth);
106 static int io_uring_register_files(struct submitter *s)
113 s->fds = calloc(s->nr_files, sizeof(__s32));
114 for (i = 0; i < s->nr_files; i++) {
115 s->fds[i] = s->files[i].real_fd;
116 s->files[i].fixed_fd = i;
119 return syscall(__NR_io_uring_register, s->ring_fd,
120 IORING_REGISTER_FILES, s->fds, s->nr_files);
123 static int io_uring_setup(unsigned entries, struct io_uring_params *p)
125 return syscall(__NR_io_uring_setup, entries, p);
128 static int io_uring_enter(struct submitter *s, unsigned int to_submit,
129 unsigned int min_complete, unsigned int flags)
131 return syscall(__NR_io_uring_enter, s->ring_fd, to_submit, min_complete,
135 #ifndef CONFIG_HAVE_GETTID
136 static int gettid(void)
138 return syscall(__NR_gettid);
142 static unsigned file_depth(struct submitter *s)
144 return (depth + s->nr_files - 1) / s->nr_files;
147 static void init_io(struct submitter *s, unsigned index)
149 struct io_uring_sqe *sqe = &s->sqes[index];
150 unsigned long offset;
155 sqe->opcode = IORING_OP_NOP;
159 if (s->nr_files == 1) {
162 f = &s->files[s->cur_file];
163 if (f->pending_ios >= file_depth(s)) {
165 if (s->cur_file == s->nr_files)
167 f = &s->files[s->cur_file];
173 offset = (r % (f->max_blocks - 1)) * BS;
175 if (register_files) {
176 sqe->flags = IOSQE_FIXED_FILE;
177 sqe->fd = f->fixed_fd;
180 sqe->fd = f->real_fd;
183 sqe->opcode = IORING_OP_READ_FIXED;
184 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
186 sqe->buf_index = index;
188 sqe->opcode = IORING_OP_READV;
189 sqe->addr = (unsigned long) &s->iovecs[index];
195 sqe->user_data = (unsigned long) f;
198 static int prep_more_ios(struct submitter *s, int max_ios)
200 struct io_sq_ring *ring = &s->sq_ring;
201 unsigned index, tail, next_tail, prepped = 0;
203 next_tail = tail = *ring->tail;
207 if (next_tail == *ring->head)
210 index = tail & sq_ring_mask;
212 ring->array[index] = index;
215 } while (prepped < max_ios);
217 if (*ring->tail != tail) {
224 static int get_file_size(struct file *f)
228 if (fstat(f->real_fd, &st) < 0)
230 if (S_ISBLK(st.st_mode)) {
231 unsigned long long bytes;
233 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
236 f->max_blocks = bytes / BS;
238 } else if (S_ISREG(st.st_mode)) {
239 f->max_blocks = st.st_size / BS;
246 static int reap_events(struct submitter *s)
248 struct io_cq_ring *ring = &s->cq_ring;
249 struct io_uring_cqe *cqe;
250 unsigned head, reaped = 0;
257 if (head == *ring->tail)
259 cqe = &ring->cqes[head & cq_ring_mask];
261 f = (struct file *) (uintptr_t) cqe->user_data;
263 if (cqe->res != BS) {
264 printf("io: unexpected ret=%d\n", cqe->res);
265 if (polled && cqe->res == -EOPNOTSUPP)
266 printf("Your filesystem/driver/kernel doesn't support polled IO\n");
274 s->inflight -= reaped;
280 static void *submitter_fn(void *data)
282 struct submitter *s = data;
283 struct io_sq_ring *ring = &s->sq_ring;
286 printf("submitter=%d\n", gettid());
288 srand48(pthread_self());
292 int to_wait, to_submit, this_reap, to_prep;
294 if (!prepped && s->inflight < depth) {
295 to_prep = min(depth - s->inflight, batch_submit);
296 prepped = prep_more_ios(s, to_prep);
298 s->inflight += prepped;
302 if (to_submit && (s->inflight + to_submit <= depth))
305 to_wait = min(s->inflight + to_submit, batch_complete);
308 * Only need to call io_uring_enter if we're not using SQ thread
309 * poll, or if IORING_SQ_NEED_WAKEUP is set.
311 if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) {
315 flags = IORING_ENTER_GETEVENTS;
316 if ((*ring->flags & IORING_SQ_NEED_WAKEUP))
317 flags |= IORING_ENTER_SQ_WAKEUP;
318 ret = io_uring_enter(s, to_submit, to_wait, flags);
323 * For non SQ thread poll, we already got the events we needed
324 * through the io_uring_enter() above. For SQ thread poll, we
325 * need to loop here until we find enough events.
336 } while (sq_thread_poll && this_reap < to_wait);
337 s->reaps += this_reap;
345 } else if (ret < to_submit) {
346 int diff = to_submit - ret;
355 } else if (ret < 0) {
356 if (errno == EAGAIN) {
364 printf("io_submit: %s\n", strerror(errno));
367 } while (!s->finish);
373 static void sig_int(int sig)
375 printf("Exiting on signal %d\n", sig);
376 submitter->finish = 1;
380 static void arm_sig_int(void)
382 struct sigaction act;
384 memset(&act, 0, sizeof(act));
385 act.sa_handler = sig_int;
386 act.sa_flags = SA_RESTART;
387 sigaction(SIGINT, &act, NULL);
390 static int setup_ring(struct submitter *s)
392 struct io_sq_ring *sring = &s->sq_ring;
393 struct io_cq_ring *cring = &s->cq_ring;
394 struct io_uring_params p;
398 memset(&p, 0, sizeof(p));
400 if (polled && !do_nop)
401 p.flags |= IORING_SETUP_IOPOLL;
402 if (sq_thread_poll) {
403 p.flags |= IORING_SETUP_SQPOLL;
404 if (sq_thread_cpu != -1) {
405 p.flags |= IORING_SETUP_SQ_AFF;
406 p.sq_thread_cpu = sq_thread_cpu;
410 fd = io_uring_setup(depth, &p);
412 perror("io_uring_setup");
418 ret = io_uring_register_buffers(s);
420 perror("io_uring_register_buffers");
425 if (register_files) {
426 ret = io_uring_register_files(s);
428 perror("io_uring_register_files");
433 ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
434 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
436 printf("sq_ring ptr = 0x%p\n", ptr);
437 sring->head = ptr + p.sq_off.head;
438 sring->tail = ptr + p.sq_off.tail;
439 sring->ring_mask = ptr + p.sq_off.ring_mask;
440 sring->ring_entries = ptr + p.sq_off.ring_entries;
441 sring->flags = ptr + p.sq_off.flags;
442 sring->array = ptr + p.sq_off.array;
443 sq_ring_mask = *sring->ring_mask;
445 s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
446 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
448 printf("sqes ptr = 0x%p\n", s->sqes);
450 ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
451 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
453 printf("cq_ring ptr = 0x%p\n", ptr);
454 cring->head = ptr + p.cq_off.head;
455 cring->tail = ptr + p.cq_off.tail;
456 cring->ring_mask = ptr + p.cq_off.ring_mask;
457 cring->ring_entries = ptr + p.cq_off.ring_entries;
458 cring->cqes = ptr + p.cq_off.cqes;
459 cq_ring_mask = *cring->ring_mask;
463 static void file_depths(char *buf)
465 struct submitter *s = submitter;
471 for (i = 0; i < s->nr_files; i++) {
472 struct file *f = &s->files[i];
474 if (i + 1 == s->nr_files)
475 p += sprintf(p, "%d", f->pending_ios);
477 p += sprintf(p, "%d, ", f->pending_ios);
481 static void usage(char *argv)
483 printf("%s [options] -- [filenames]\n"
484 " -d <int> : IO Depth, default %d\n"
485 " -s <int> : Batch submit, default %d\n"
486 " -c <int> : Batch complete, default %d\n",
487 argv, DEPTH, BATCH_SUBMIT, BATCH_COMPLETE);
491 int main(int argc, char *argv[])
494 unsigned long done, calls, reap;
495 int err, i, flags, fd, opt;
499 if (!do_nop && argc < 2) {
500 printf("%s: filename [options]\n", argv[0]);
504 while ((opt = getopt(argc, argv, "d:s:c:h?")) != -1) {
507 depth = atoi(optarg);
510 batch_submit = atoi(optarg);
513 batch_complete = atoi(optarg);
523 submitter = malloc(sizeof(*submitter) + depth * sizeof(struct iovec));
524 memset(submitter, 0, sizeof(*submitter) + depth * sizeof(struct iovec));
527 flags = O_RDONLY | O_NOATIME;
532 while (!do_nop && i < argc) {
535 if (s->nr_files == MAX_FDS) {
536 printf("Max number of files (%d) reached\n", MAX_FDS);
539 fd = open(argv[i], flags);
545 f = &s->files[s->nr_files];
547 if (get_file_size(f)) {
548 printf("failed getting size of device/file\n");
551 if (f->max_blocks <= 1) {
552 printf("Zero file/device size?\n");
557 printf("Added file %s\n", argv[i]);
565 rlim.rlim_cur = RLIM_INFINITY;
566 rlim.rlim_max = RLIM_INFINITY;
567 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
575 for (i = 0; i < depth; i++) {
578 if (posix_memalign(&buf, BS, BS)) {
579 printf("failed alloc\n");
582 s->iovecs[i].iov_base = buf;
583 s->iovecs[i].iov_len = BS;
588 printf("ring setup failed: %s, %d\n", strerror(errno), err);
591 printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
592 printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", depth, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
594 pthread_create(&s->thread, NULL, submitter_fn, s);
596 fdepths = malloc(8 * s->nr_files);
597 reap = calls = done = 0;
599 unsigned long this_done = 0;
600 unsigned long this_reap = 0;
601 unsigned long this_call = 0;
602 unsigned long rpc = 0, ipc = 0;
605 this_done += s->done;
606 this_call += s->calls;
607 this_reap += s->reaps;
608 if (this_call - calls) {
609 rpc = (this_done - done) / (this_call - calls);
610 ipc = (this_reap - reap) / (this_call - calls);
613 file_depths(fdepths);
614 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s)\n",
615 this_done - done, rpc, ipc, s->inflight,
622 pthread_join(s->thread, &ret);