| 1 | #include <stdio.h> |
| 2 | #include <errno.h> |
| 3 | #include <assert.h> |
| 4 | #include <stdlib.h> |
| 5 | #include <stddef.h> |
| 6 | #include <signal.h> |
| 7 | #include <inttypes.h> |
| 8 | |
| 9 | #include <sys/types.h> |
| 10 | #include <sys/stat.h> |
| 11 | #include <sys/ioctl.h> |
| 12 | #include <sys/syscall.h> |
| 13 | #include <sys/resource.h> |
| 14 | #include <sys/mman.h> |
| 15 | #include <sys/uio.h> |
| 16 | #include <linux/fs.h> |
| 17 | #include <fcntl.h> |
| 18 | #include <unistd.h> |
| 19 | #include <string.h> |
| 20 | #include <pthread.h> |
| 21 | #include <sched.h> |
| 22 | |
| 23 | #include "../arch/arch.h" |
| 24 | #include "../lib/types.h" |
| 25 | #include "../os/linux/io_uring.h" |
| 26 | |
| 27 | #define barrier() __asm__ __volatile__("": : :"memory") |
| 28 | |
| 29 | #define min(a, b) ((a < b) ? (a) : (b)) |
| 30 | |
| 31 | struct io_sq_ring { |
| 32 | unsigned *head; |
| 33 | unsigned *tail; |
| 34 | unsigned *ring_mask; |
| 35 | unsigned *ring_entries; |
| 36 | unsigned *flags; |
| 37 | unsigned *array; |
| 38 | }; |
| 39 | |
| 40 | struct io_cq_ring { |
| 41 | unsigned *head; |
| 42 | unsigned *tail; |
| 43 | unsigned *ring_mask; |
| 44 | unsigned *ring_entries; |
| 45 | struct io_uring_cqe *cqes; |
| 46 | }; |
| 47 | |
| 48 | #define DEPTH 128 |
| 49 | |
| 50 | #define BATCH_SUBMIT 64 |
| 51 | #define BATCH_COMPLETE 64 |
| 52 | |
| 53 | #define BS 4096 |
| 54 | |
| 55 | #define MAX_FDS 16 |
| 56 | |
| 57 | static unsigned sq_ring_mask, cq_ring_mask; |
| 58 | |
| 59 | struct file { |
| 60 | unsigned long max_blocks; |
| 61 | unsigned pending_ios; |
| 62 | int real_fd; |
| 63 | int fixed_fd; |
| 64 | }; |
| 65 | |
| 66 | struct submitter { |
| 67 | pthread_t thread; |
| 68 | int ring_fd; |
| 69 | struct drand48_data rand; |
| 70 | struct io_sq_ring sq_ring; |
| 71 | struct io_uring_sqe *sqes; |
| 72 | struct iovec iovecs[DEPTH]; |
| 73 | struct io_cq_ring cq_ring; |
| 74 | int inflight; |
| 75 | unsigned long reaps; |
| 76 | unsigned long done; |
| 77 | unsigned long calls; |
| 78 | unsigned long cachehit, cachemiss; |
| 79 | volatile int finish; |
| 80 | |
| 81 | __s32 *fds; |
| 82 | |
| 83 | struct file files[MAX_FDS]; |
| 84 | unsigned nr_files; |
| 85 | unsigned cur_file; |
| 86 | }; |
| 87 | |
| 88 | static struct submitter submitters[1]; |
| 89 | static volatile int finish; |
| 90 | |
| 91 | static int polled = 1; /* use IO polling */ |
| 92 | static int fixedbufs = 1; /* use fixed user buffers */ |
| 93 | static int register_files = 1; /* use fixed files */ |
| 94 | static int buffered = 0; /* use buffered IO, not O_DIRECT */ |
| 95 | static int sq_thread_poll = 0; /* use kernel submission/poller thread */ |
| 96 | static int sq_thread_cpu = -1; /* pin above thread to this CPU */ |
| 97 | static int do_nop = 0; /* no-op SQ ring commands */ |
| 98 | |
| 99 | static int io_uring_register_buffers(struct submitter *s) |
| 100 | { |
| 101 | struct io_uring_register_buffers reg = { |
| 102 | .iovecs = s->iovecs, |
| 103 | .nr_iovecs = DEPTH |
| 104 | }; |
| 105 | |
| 106 | if (do_nop) |
| 107 | return 0; |
| 108 | |
| 109 | return syscall(__NR_sys_io_uring_register, s->ring_fd, |
| 110 | IORING_REGISTER_BUFFERS, ®); |
| 111 | } |
| 112 | |
| 113 | static int io_uring_register_files(struct submitter *s) |
| 114 | { |
| 115 | struct io_uring_register_files reg; |
| 116 | int i; |
| 117 | |
| 118 | if (do_nop) |
| 119 | return 0; |
| 120 | |
| 121 | s->fds = calloc(s->nr_files, sizeof(__s32)); |
| 122 | for (i = 0; i < s->nr_files; i++) { |
| 123 | s->fds[i] = s->files[i].real_fd; |
| 124 | s->files[i].fixed_fd = i; |
| 125 | } |
| 126 | reg.fds = s->fds; |
| 127 | reg.nr_fds = s->nr_files; |
| 128 | |
| 129 | return syscall(__NR_sys_io_uring_register, s->ring_fd, |
| 130 | IORING_REGISTER_FILES, ®); |
| 131 | } |
| 132 | |
| 133 | static int io_uring_setup(unsigned entries, struct io_uring_params *p) |
| 134 | { |
| 135 | return syscall(__NR_sys_io_uring_setup, entries, p); |
| 136 | } |
| 137 | |
| 138 | static int io_uring_enter(struct submitter *s, unsigned int to_submit, |
| 139 | unsigned int min_complete, unsigned int flags) |
| 140 | { |
| 141 | return syscall(__NR_sys_io_uring_enter, s->ring_fd, to_submit, |
| 142 | min_complete, flags); |
| 143 | } |
| 144 | |
| 145 | static int gettid(void) |
| 146 | { |
| 147 | return syscall(__NR_gettid); |
| 148 | } |
| 149 | |
| 150 | static unsigned file_depth(struct submitter *s) |
| 151 | { |
| 152 | return (DEPTH + s->nr_files - 1) / s->nr_files; |
| 153 | } |
| 154 | |
| 155 | static void init_io(struct submitter *s, unsigned index) |
| 156 | { |
| 157 | struct io_uring_sqe *sqe = &s->sqes[index]; |
| 158 | unsigned long offset; |
| 159 | struct file *f; |
| 160 | long r; |
| 161 | |
| 162 | if (do_nop) { |
| 163 | sqe->opcode = IORING_OP_NOP; |
| 164 | return; |
| 165 | } |
| 166 | |
| 167 | if (s->nr_files == 1) { |
| 168 | f = &s->files[0]; |
| 169 | } else { |
| 170 | f = &s->files[s->cur_file]; |
| 171 | if (f->pending_ios >= file_depth(s)) { |
| 172 | s->cur_file++; |
| 173 | if (s->cur_file == s->nr_files) |
| 174 | s->cur_file = 0; |
| 175 | } |
| 176 | } |
| 177 | f->pending_ios++; |
| 178 | |
| 179 | lrand48_r(&s->rand, &r); |
| 180 | offset = (r % (f->max_blocks - 1)) * BS; |
| 181 | |
| 182 | if (register_files) { |
| 183 | sqe->flags = IOSQE_FIXED_FILE; |
| 184 | sqe->fd = f->fixed_fd; |
| 185 | } else { |
| 186 | sqe->flags = 0; |
| 187 | sqe->fd = f->real_fd; |
| 188 | } |
| 189 | if (fixedbufs) { |
| 190 | sqe->opcode = IORING_OP_READ_FIXED; |
| 191 | sqe->addr = s->iovecs[index].iov_base; |
| 192 | sqe->len = BS; |
| 193 | sqe->buf_index = index; |
| 194 | } else { |
| 195 | sqe->opcode = IORING_OP_READV; |
| 196 | sqe->addr = &s->iovecs[index]; |
| 197 | sqe->len = 1; |
| 198 | sqe->buf_index = 0; |
| 199 | } |
| 200 | sqe->ioprio = 0; |
| 201 | sqe->off = offset; |
| 202 | sqe->user_data = (unsigned long) f; |
| 203 | } |
| 204 | |
| 205 | static int prep_more_ios(struct submitter *s, int max_ios) |
| 206 | { |
| 207 | struct io_sq_ring *ring = &s->sq_ring; |
| 208 | unsigned index, tail, next_tail, prepped = 0; |
| 209 | |
| 210 | next_tail = tail = *ring->tail; |
| 211 | do { |
| 212 | next_tail++; |
| 213 | barrier(); |
| 214 | if (next_tail == *ring->head) |
| 215 | break; |
| 216 | |
| 217 | index = tail & sq_ring_mask; |
| 218 | init_io(s, index); |
| 219 | ring->array[index] = index; |
| 220 | prepped++; |
| 221 | tail = next_tail; |
| 222 | } while (prepped < max_ios); |
| 223 | |
| 224 | if (*ring->tail != tail) { |
| 225 | /* order tail store with writes to sqes above */ |
| 226 | barrier(); |
| 227 | *ring->tail = tail; |
| 228 | barrier(); |
| 229 | } |
| 230 | return prepped; |
| 231 | } |
| 232 | |
| 233 | static int get_file_size(struct file *f) |
| 234 | { |
| 235 | struct stat st; |
| 236 | |
| 237 | if (fstat(f->real_fd, &st) < 0) |
| 238 | return -1; |
| 239 | if (S_ISBLK(st.st_mode)) { |
| 240 | unsigned long long bytes; |
| 241 | |
| 242 | if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0) |
| 243 | return -1; |
| 244 | |
| 245 | f->max_blocks = bytes / BS; |
| 246 | return 0; |
| 247 | } else if (S_ISREG(st.st_mode)) { |
| 248 | f->max_blocks = st.st_size / BS; |
| 249 | return 0; |
| 250 | } |
| 251 | |
| 252 | return -1; |
| 253 | } |
| 254 | |
| 255 | static int reap_events(struct submitter *s) |
| 256 | { |
| 257 | struct io_cq_ring *ring = &s->cq_ring; |
| 258 | struct io_uring_cqe *cqe; |
| 259 | unsigned head, reaped = 0; |
| 260 | |
| 261 | head = *ring->head; |
| 262 | do { |
| 263 | struct file *f; |
| 264 | |
| 265 | barrier(); |
| 266 | if (head == *ring->tail) |
| 267 | break; |
| 268 | cqe = &ring->cqes[head & cq_ring_mask]; |
| 269 | if (!do_nop) { |
| 270 | f = (struct file *) (uintptr_t) cqe->user_data; |
| 271 | f->pending_ios--; |
| 272 | if (cqe->res != BS) { |
| 273 | printf("io: unexpected ret=%d\n", cqe->res); |
| 274 | return -1; |
| 275 | } |
| 276 | } |
| 277 | if (cqe->flags & IOCQE_FLAG_CACHEHIT) |
| 278 | s->cachehit++; |
| 279 | else |
| 280 | s->cachemiss++; |
| 281 | reaped++; |
| 282 | head++; |
| 283 | } while (1); |
| 284 | |
| 285 | s->inflight -= reaped; |
| 286 | *ring->head = head; |
| 287 | barrier(); |
| 288 | return reaped; |
| 289 | } |
| 290 | |
| 291 | static void *submitter_fn(void *data) |
| 292 | { |
| 293 | struct submitter *s = data; |
| 294 | struct io_sq_ring *ring = &s->sq_ring; |
| 295 | int ret, prepped; |
| 296 | |
| 297 | printf("submitter=%d\n", gettid()); |
| 298 | |
| 299 | srand48_r(pthread_self(), &s->rand); |
| 300 | |
| 301 | prepped = 0; |
| 302 | do { |
| 303 | int to_wait, to_submit, this_reap, to_prep; |
| 304 | |
| 305 | if (!prepped && s->inflight < DEPTH) { |
| 306 | to_prep = min(DEPTH - s->inflight, BATCH_SUBMIT); |
| 307 | prepped = prep_more_ios(s, to_prep); |
| 308 | } |
| 309 | s->inflight += prepped; |
| 310 | submit_more: |
| 311 | to_submit = prepped; |
| 312 | submit: |
| 313 | if (to_submit && (s->inflight + to_submit < DEPTH)) |
| 314 | to_wait = 0; |
| 315 | else |
| 316 | to_wait = min(s->inflight + to_submit, BATCH_COMPLETE); |
| 317 | |
| 318 | /* |
| 319 | * Only need to call io_uring_enter if we're not using SQ thread |
| 320 | * poll, or if IORING_SQ_NEED_WAKEUP is set. |
| 321 | */ |
| 322 | if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) { |
| 323 | unsigned flags = 0; |
| 324 | |
| 325 | if (to_wait) |
| 326 | flags = IORING_ENTER_GETEVENTS; |
| 327 | ret = io_uring_enter(s, to_submit, to_wait, flags); |
| 328 | s->calls++; |
| 329 | } |
| 330 | |
| 331 | /* |
| 332 | * For non SQ thread poll, we already got the events we needed |
| 333 | * through the io_uring_enter() above. For SQ thread poll, we |
| 334 | * need to loop here until we find enough events. |
| 335 | */ |
| 336 | this_reap = 0; |
| 337 | do { |
| 338 | int r; |
| 339 | r = reap_events(s); |
| 340 | if (r == -1) |
| 341 | break; |
| 342 | else if (r > 0) |
| 343 | this_reap += r; |
| 344 | } while (sq_thread_poll && this_reap < to_wait); |
| 345 | s->reaps += this_reap; |
| 346 | |
| 347 | if (ret >= 0) { |
| 348 | if (!ret) { |
| 349 | to_submit = 0; |
| 350 | if (s->inflight) |
| 351 | goto submit; |
| 352 | continue; |
| 353 | } else if (ret < to_submit) { |
| 354 | int diff = to_submit - ret; |
| 355 | |
| 356 | s->done += ret; |
| 357 | prepped -= diff; |
| 358 | goto submit_more; |
| 359 | } |
| 360 | s->done += ret; |
| 361 | prepped = 0; |
| 362 | continue; |
| 363 | } else if (ret < 0) { |
| 364 | if (errno == EAGAIN) { |
| 365 | if (s->finish) |
| 366 | break; |
| 367 | if (this_reap) |
| 368 | goto submit; |
| 369 | to_submit = 0; |
| 370 | goto submit; |
| 371 | } |
| 372 | printf("io_submit: %s\n", strerror(errno)); |
| 373 | break; |
| 374 | } |
| 375 | } while (!s->finish); |
| 376 | |
| 377 | finish = 1; |
| 378 | return NULL; |
| 379 | } |
| 380 | |
| 381 | static void sig_int(int sig) |
| 382 | { |
| 383 | printf("Exiting on signal %d\n", sig); |
| 384 | submitters[0].finish = 1; |
| 385 | finish = 1; |
| 386 | } |
| 387 | |
| 388 | static void arm_sig_int(void) |
| 389 | { |
| 390 | struct sigaction act; |
| 391 | |
| 392 | memset(&act, 0, sizeof(act)); |
| 393 | act.sa_handler = sig_int; |
| 394 | act.sa_flags = SA_RESTART; |
| 395 | sigaction(SIGINT, &act, NULL); |
| 396 | } |
| 397 | |
| 398 | static int setup_ring(struct submitter *s) |
| 399 | { |
| 400 | struct io_sq_ring *sring = &s->sq_ring; |
| 401 | struct io_cq_ring *cring = &s->cq_ring; |
| 402 | struct io_uring_params p; |
| 403 | int ret, fd; |
| 404 | void *ptr; |
| 405 | |
| 406 | memset(&p, 0, sizeof(p)); |
| 407 | |
| 408 | if (polled) |
| 409 | p.flags |= IORING_SETUP_IOPOLL; |
| 410 | if (sq_thread_poll) { |
| 411 | p.flags |= IORING_SETUP_SQPOLL; |
| 412 | if (sq_thread_cpu != -1) { |
| 413 | p.flags |= IORING_SETUP_SQ_AFF; |
| 414 | p.sq_thread_cpu = sq_thread_cpu; |
| 415 | } |
| 416 | } |
| 417 | |
| 418 | fd = io_uring_setup(DEPTH, &p); |
| 419 | if (fd < 0) { |
| 420 | perror("io_uring_setup"); |
| 421 | return 1; |
| 422 | } |
| 423 | s->ring_fd = fd; |
| 424 | |
| 425 | if (fixedbufs) { |
| 426 | ret = io_uring_register_buffers(s); |
| 427 | if (ret < 0) { |
| 428 | perror("io_uring_register_buffers"); |
| 429 | return 1; |
| 430 | } |
| 431 | } |
| 432 | |
| 433 | if (register_files) { |
| 434 | ret = io_uring_register_files(s); |
| 435 | if (ret < 0) { |
| 436 | perror("io_uring_register_files"); |
| 437 | return 1; |
| 438 | } |
| 439 | } |
| 440 | |
| 441 | ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32), |
| 442 | PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, |
| 443 | IORING_OFF_SQ_RING); |
| 444 | printf("sq_ring ptr = 0x%p\n", ptr); |
| 445 | sring->head = ptr + p.sq_off.head; |
| 446 | sring->tail = ptr + p.sq_off.tail; |
| 447 | sring->ring_mask = ptr + p.sq_off.ring_mask; |
| 448 | sring->ring_entries = ptr + p.sq_off.ring_entries; |
| 449 | sring->flags = ptr + p.sq_off.flags; |
| 450 | sring->array = ptr + p.sq_off.array; |
| 451 | sq_ring_mask = *sring->ring_mask; |
| 452 | |
| 453 | s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe), |
| 454 | PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, |
| 455 | IORING_OFF_SQES); |
| 456 | printf("sqes ptr = 0x%p\n", s->sqes); |
| 457 | |
| 458 | ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe), |
| 459 | PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, |
| 460 | IORING_OFF_CQ_RING); |
| 461 | printf("cq_ring ptr = 0x%p\n", ptr); |
| 462 | cring->head = ptr + p.cq_off.head; |
| 463 | cring->tail = ptr + p.cq_off.tail; |
| 464 | cring->ring_mask = ptr + p.cq_off.ring_mask; |
| 465 | cring->ring_entries = ptr + p.cq_off.ring_entries; |
| 466 | cring->cqes = ptr + p.cq_off.cqes; |
| 467 | cq_ring_mask = *cring->ring_mask; |
| 468 | return 0; |
| 469 | } |
| 470 | |
| 471 | int main(int argc, char *argv[]) |
| 472 | { |
| 473 | struct submitter *s = &submitters[0]; |
| 474 | unsigned long done, calls, reap, cache_hit, cache_miss; |
| 475 | int err, i, flags, fd; |
| 476 | void *ret; |
| 477 | |
| 478 | if (!do_nop && argc < 2) { |
| 479 | printf("%s: filename\n", argv[0]); |
| 480 | return 1; |
| 481 | } |
| 482 | |
| 483 | flags = O_RDONLY | O_NOATIME; |
| 484 | if (!buffered) |
| 485 | flags |= O_DIRECT; |
| 486 | |
| 487 | i = 1; |
| 488 | while (!do_nop && i < argc) { |
| 489 | struct file *f = &s->files[s->nr_files]; |
| 490 | |
| 491 | fd = open(argv[i], flags); |
| 492 | if (fd < 0) { |
| 493 | perror("open"); |
| 494 | return 1; |
| 495 | } |
| 496 | f->real_fd = fd; |
| 497 | if (get_file_size(f)) { |
| 498 | printf("failed getting size of device/file\n"); |
| 499 | return 1; |
| 500 | } |
| 501 | if (f->max_blocks <= 1) { |
| 502 | printf("Zero file/device size?\n"); |
| 503 | return 1; |
| 504 | } |
| 505 | f->max_blocks--; |
| 506 | |
| 507 | printf("Added file %s\n", argv[i]); |
| 508 | s->nr_files++; |
| 509 | i++; |
| 510 | } |
| 511 | |
| 512 | if (fixedbufs) { |
| 513 | struct rlimit rlim; |
| 514 | |
| 515 | rlim.rlim_cur = RLIM_INFINITY; |
| 516 | rlim.rlim_max = RLIM_INFINITY; |
| 517 | if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) { |
| 518 | perror("setrlimit"); |
| 519 | return 1; |
| 520 | } |
| 521 | } |
| 522 | |
| 523 | arm_sig_int(); |
| 524 | |
| 525 | for (i = 0; i < DEPTH; i++) { |
| 526 | void *buf; |
| 527 | |
| 528 | if (posix_memalign(&buf, BS, BS)) { |
| 529 | printf("failed alloc\n"); |
| 530 | return 1; |
| 531 | } |
| 532 | s->iovecs[i].iov_base = buf; |
| 533 | s->iovecs[i].iov_len = BS; |
| 534 | } |
| 535 | |
| 536 | err = setup_ring(s); |
| 537 | if (err) { |
| 538 | printf("ring setup failed: %s, %d\n", strerror(errno), err); |
| 539 | return 1; |
| 540 | } |
| 541 | printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered); |
| 542 | printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", DEPTH, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries); |
| 543 | |
| 544 | pthread_create(&s->thread, NULL, submitter_fn, s); |
| 545 | |
| 546 | cache_hit = cache_miss = reap = calls = done = 0; |
| 547 | do { |
| 548 | unsigned long this_done = 0; |
| 549 | unsigned long this_reap = 0; |
| 550 | unsigned long this_call = 0; |
| 551 | unsigned long this_cache_hit = 0; |
| 552 | unsigned long this_cache_miss = 0; |
| 553 | unsigned long rpc = 0, ipc = 0; |
| 554 | double hit = 0.0; |
| 555 | |
| 556 | sleep(1); |
| 557 | this_done += s->done; |
| 558 | this_call += s->calls; |
| 559 | this_reap += s->reaps; |
| 560 | this_cache_hit += s->cachehit; |
| 561 | this_cache_miss += s->cachemiss; |
| 562 | if (this_cache_hit && this_cache_miss) { |
| 563 | unsigned long hits, total; |
| 564 | |
| 565 | hits = this_cache_hit - cache_hit; |
| 566 | total = hits + this_cache_miss - cache_miss; |
| 567 | hit = (double) hits / (double) total; |
| 568 | hit *= 100.0; |
| 569 | } |
| 570 | if (this_call - calls) { |
| 571 | rpc = (this_done - done) / (this_call - calls); |
| 572 | ipc = (this_reap - reap) / (this_call - calls); |
| 573 | } else |
| 574 | rpc = ipc = -1; |
| 575 | printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (head=%u tail=%u), Cachehit=%0.2f%%\n", |
| 576 | this_done - done, rpc, ipc, s->inflight, |
| 577 | *s->cq_ring.head, *s->cq_ring.tail, hit); |
| 578 | done = this_done; |
| 579 | calls = this_call; |
| 580 | reap = this_reap; |
| 581 | cache_hit = s->cachehit; |
| 582 | cache_miss = s->cachemiss; |
| 583 | } while (!finish); |
| 584 | |
| 585 | pthread_join(s->thread, &ret); |
| 586 | close(s->ring_fd); |
| 587 | return 0; |
| 588 | } |