t/io_uring: fix bad if
[fio.git] / t / io_uring.c
CommitLineData
c9fb4c5b
JA
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>
c3e2fc25 14#include <sys/mman.h>
e31b8288 15#include <sys/uio.h>
c9fb4c5b
JA
16#include <linux/fs.h>
17#include <fcntl.h>
18#include <unistd.h>
c9fb4c5b
JA
19#include <string.h>
20#include <pthread.h>
21#include <sched.h>
22
74efb029 23#include "../arch/arch.h"
57fa61f0 24#include "../lib/types.h"
f3e769a4 25#include "../os/linux/io_uring.h"
ac122fea 26
e31b8288 27#define min(a, b) ((a < b) ? (a) : (b))
c3e2fc25 28
e31b8288 29struct io_sq_ring {
e2239016
JA
30 unsigned *head;
31 unsigned *tail;
32 unsigned *ring_mask;
33 unsigned *ring_entries;
ce1705de 34 unsigned *flags;
e2239016 35 unsigned *array;
c9fb4c5b
JA
36};
37
e31b8288 38struct io_cq_ring {
e2239016
JA
39 unsigned *head;
40 unsigned *tail;
41 unsigned *ring_mask;
42 unsigned *ring_entries;
f0403f94 43 struct io_uring_cqe *cqes;
c9fb4c5b
JA
44};
45
701d1277 46#define DEPTH 128
c9fb4c5b 47
2e7888ef
JA
48#define BATCH_SUBMIT 32
49#define BATCH_COMPLETE 32
c9fb4c5b
JA
50
51#define BS 4096
52
a7086591
JA
53#define MAX_FDS 16
54
c3e2fc25 55static unsigned sq_ring_mask, cq_ring_mask;
e39c34dc 56
a7086591
JA
57struct file {
58 unsigned long max_blocks;
701d1277 59 unsigned pending_ios;
48e698fa
JA
60 int real_fd;
61 int fixed_fd;
a7086591
JA
62};
63
c9fb4c5b
JA
64struct submitter {
65 pthread_t thread;
f310970e 66 int ring_fd;
c9fb4c5b 67 struct drand48_data rand;
e31b8288 68 struct io_sq_ring sq_ring;
f0403f94 69 struct io_uring_sqe *sqes;
c3e2fc25 70 struct iovec iovecs[DEPTH];
e31b8288 71 struct io_cq_ring cq_ring;
c9fb4c5b
JA
72 int inflight;
73 unsigned long reaps;
74 unsigned long done;
75 unsigned long calls;
305c06ce 76 unsigned long cachehit, cachemiss;
c9fb4c5b 77 volatile int finish;
701d1277 78
48e698fa
JA
79 __s32 *fds;
80
a7086591
JA
81 struct file files[MAX_FDS];
82 unsigned nr_files;
83 unsigned cur_file;
c9fb4c5b
JA
84};
85
86static struct submitter submitters[1];
87static volatile int finish;
88
f0403f94 89static int polled = 1; /* use IO polling */
701d1277 90static int fixedbufs = 1; /* use fixed user buffers */
8c5fa755 91static int register_files = 1; /* use fixed files */
f0403f94 92static int buffered = 0; /* use buffered IO, not O_DIRECT */
3d7d00a3
JA
93static int sq_thread_poll = 0; /* use kernel submission/poller thread */
94static int sq_thread_cpu = -1; /* pin above thread to this CPU */
8025517d 95static int do_nop = 0; /* no-op SQ ring commands */
c9fb4c5b 96
2ea53ca3 97static int io_uring_register_buffers(struct submitter *s)
c9fb4c5b 98{
8025517d
JA
99 if (do_nop)
100 return 0;
101
2ea53ca3 102 return syscall(__NR_sys_io_uring_register, s->ring_fd,
919850d2 103 IORING_REGISTER_BUFFERS, s->iovecs, DEPTH);
2ea53ca3
JA
104}
105
a7abc9fb
JA
106static int io_uring_register_files(struct submitter *s)
107{
48e698fa 108 int i;
a7abc9fb 109
8025517d
JA
110 if (do_nop)
111 return 0;
112
48e698fa
JA
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;
117 }
a7abc9fb 118
48e698fa 119 return syscall(__NR_sys_io_uring_register, s->ring_fd,
919850d2 120 IORING_REGISTER_FILES, s->fds, s->nr_files);
a7abc9fb
JA
121}
122
2ea53ca3
JA
123static int io_uring_setup(unsigned entries, struct io_uring_params *p)
124{
125 return syscall(__NR_sys_io_uring_setup, entries, p);
c9fb4c5b
JA
126}
127
c3e2fc25
JA
128static int io_uring_enter(struct submitter *s, unsigned int to_submit,
129 unsigned int min_complete, unsigned int flags)
c9fb4c5b 130{
f310970e
JA
131 return syscall(__NR_sys_io_uring_enter, s->ring_fd, to_submit,
132 min_complete, flags);
c9fb4c5b
JA
133}
134
135static int gettid(void)
136{
137 return syscall(__NR_gettid);
138}
139
701d1277
JA
140static unsigned file_depth(struct submitter *s)
141{
142 return (DEPTH + s->nr_files - 1) / s->nr_files;
143}
144
a7086591 145static void init_io(struct submitter *s, unsigned index)
c9fb4c5b 146{
f0403f94 147 struct io_uring_sqe *sqe = &s->sqes[index];
c9fb4c5b 148 unsigned long offset;
a7086591 149 struct file *f;
c9fb4c5b
JA
150 long r;
151
8025517d
JA
152 if (do_nop) {
153 sqe->opcode = IORING_OP_NOP;
154 return;
155 }
156
701d1277
JA
157 if (s->nr_files == 1) {
158 f = &s->files[0];
159 } else {
160 f = &s->files[s->cur_file];
161 if (f->pending_ios >= file_depth(s)) {
162 s->cur_file++;
163 if (s->cur_file == s->nr_files)
164 s->cur_file = 0;
93d1811c 165 f = &s->files[s->cur_file];
701d1277
JA
166 }
167 }
168 f->pending_ios++;
a7086591 169
c9fb4c5b 170 lrand48_r(&s->rand, &r);
a7086591 171 offset = (r % (f->max_blocks - 1)) * BS;
c9fb4c5b 172
8c5fa755
JA
173 if (register_files) {
174 sqe->flags = IOSQE_FIXED_FILE;
175 sqe->fd = f->fixed_fd;
176 } else {
177 sqe->flags = 0;
178 sqe->fd = f->real_fd;
179 }
f0403f94 180 if (fixedbufs) {
48e698fa 181 sqe->opcode = IORING_OP_READ_FIXED;
919850d2 182 sqe->addr = (unsigned long) s->iovecs[index].iov_base;
f0403f94 183 sqe->len = BS;
2ea53ca3 184 sqe->buf_index = index;
f0403f94 185 } else {
48e698fa 186 sqe->opcode = IORING_OP_READV;
919850d2 187 sqe->addr = (unsigned long) &s->iovecs[index];
f0403f94 188 sqe->len = 1;
2ea53ca3 189 sqe->buf_index = 0;
f0403f94 190 }
f0403f94 191 sqe->ioprio = 0;
f0403f94 192 sqe->off = offset;
48e698fa 193 sqe->user_data = (unsigned long) f;
c9fb4c5b
JA
194}
195
a7086591 196static int prep_more_ios(struct submitter *s, int max_ios)
c9fb4c5b 197{
e31b8288 198 struct io_sq_ring *ring = &s->sq_ring;
e2239016 199 unsigned index, tail, next_tail, prepped = 0;
c9fb4c5b 200
c3e2fc25 201 next_tail = tail = *ring->tail;
c9fb4c5b
JA
202 do {
203 next_tail++;
679d8352 204 read_barrier();
c3e2fc25 205 if (next_tail == *ring->head)
c9fb4c5b
JA
206 break;
207
e39c34dc 208 index = tail & sq_ring_mask;
a7086591 209 init_io(s, index);
c3e2fc25 210 ring->array[index] = index;
c9fb4c5b
JA
211 prepped++;
212 tail = next_tail;
213 } while (prepped < max_ios);
214
c3e2fc25 215 if (*ring->tail != tail) {
f0403f94 216 /* order tail store with writes to sqes above */
679d8352 217 write_barrier();
c3e2fc25 218 *ring->tail = tail;
679d8352 219 write_barrier();
c9fb4c5b
JA
220 }
221 return prepped;
222}
223
a7086591 224static int get_file_size(struct file *f)
c9fb4c5b
JA
225{
226 struct stat st;
227
48e698fa 228 if (fstat(f->real_fd, &st) < 0)
c9fb4c5b
JA
229 return -1;
230 if (S_ISBLK(st.st_mode)) {
231 unsigned long long bytes;
232
48e698fa 233 if (ioctl(f->real_fd, BLKGETSIZE64, &bytes) != 0)
c9fb4c5b
JA
234 return -1;
235
a7086591 236 f->max_blocks = bytes / BS;
c9fb4c5b
JA
237 return 0;
238 } else if (S_ISREG(st.st_mode)) {
a7086591 239 f->max_blocks = st.st_size / BS;
c9fb4c5b
JA
240 return 0;
241 }
242
243 return -1;
244}
245
246static int reap_events(struct submitter *s)
247{
e31b8288 248 struct io_cq_ring *ring = &s->cq_ring;
f0403f94 249 struct io_uring_cqe *cqe;
e2239016 250 unsigned head, reaped = 0;
c9fb4c5b 251
c3e2fc25 252 head = *ring->head;
c9fb4c5b 253 do {
701d1277
JA
254 struct file *f;
255
679d8352 256 read_barrier();
c3e2fc25 257 if (head == *ring->tail)
c9fb4c5b 258 break;
f0403f94 259 cqe = &ring->cqes[head & cq_ring_mask];
8025517d 260 if (!do_nop) {
e3466352 261 f = (struct file *) (uintptr_t) cqe->user_data;
8025517d
JA
262 f->pending_ios--;
263 if (cqe->res != BS) {
264 printf("io: unexpected ret=%d\n", cqe->res);
265 return -1;
266 }
c9fb4c5b 267 }
f0403f94 268 if (cqe->flags & IOCQE_FLAG_CACHEHIT)
305c06ce
JA
269 s->cachehit++;
270 else
271 s->cachemiss++;
c9fb4c5b
JA
272 reaped++;
273 head++;
c9fb4c5b
JA
274 } while (1);
275
276 s->inflight -= reaped;
c3e2fc25 277 *ring->head = head;
679d8352 278 write_barrier();
c9fb4c5b
JA
279 return reaped;
280}
281
282static void *submitter_fn(void *data)
283{
284 struct submitter *s = data;
ce1705de 285 struct io_sq_ring *ring = &s->sq_ring;
a7086591 286 int ret, prepped;
c9fb4c5b
JA
287
288 printf("submitter=%d\n", gettid());
289
c9fb4c5b
JA
290 srand48_r(pthread_self(), &s->rand);
291
292 prepped = 0;
293 do {
f310970e 294 int to_wait, to_submit, this_reap, to_prep;
c9fb4c5b 295
f310970e
JA
296 if (!prepped && s->inflight < DEPTH) {
297 to_prep = min(DEPTH - s->inflight, BATCH_SUBMIT);
a7086591 298 prepped = prep_more_ios(s, to_prep);
f310970e 299 }
c9fb4c5b
JA
300 s->inflight += prepped;
301submit_more:
302 to_submit = prepped;
303submit:
e4db5bd9 304 if (to_submit && (s->inflight + to_submit <= DEPTH))
c9fb4c5b
JA
305 to_wait = 0;
306 else
307 to_wait = min(s->inflight + to_submit, BATCH_COMPLETE);
308
ce1705de
JA
309 /*
310 * Only need to call io_uring_enter if we're not using SQ thread
311 * poll, or if IORING_SQ_NEED_WAKEUP is set.
312 */
313 if (!sq_thread_poll || (*ring->flags & IORING_SQ_NEED_WAKEUP)) {
e0abe388
JA
314 unsigned flags = 0;
315
316 if (to_wait)
317 flags = IORING_ENTER_GETEVENTS;
e2f309e6 318 if ((*ring->flags & IORING_SQ_NEED_WAKEUP))
b532dd6d 319 flags |= IORING_ENTER_SQ_WAKEUP;
e0abe388 320 ret = io_uring_enter(s, to_submit, to_wait, flags);
ce1705de
JA
321 s->calls++;
322 }
c9fb4c5b 323
ce1705de
JA
324 /*
325 * For non SQ thread poll, we already got the events we needed
326 * through the io_uring_enter() above. For SQ thread poll, we
327 * need to loop here until we find enough events.
328 */
329 this_reap = 0;
330 do {
331 int r;
332 r = reap_events(s);
7d1435e6
JA
333 if (r == -1) {
334 s->finish = 1;
ce1705de 335 break;
7d1435e6 336 } else if (r > 0)
ce1705de
JA
337 this_reap += r;
338 } while (sq_thread_poll && this_reap < to_wait);
c9fb4c5b
JA
339 s->reaps += this_reap;
340
341 if (ret >= 0) {
342 if (!ret) {
343 to_submit = 0;
344 if (s->inflight)
345 goto submit;
346 continue;
347 } else if (ret < to_submit) {
348 int diff = to_submit - ret;
349
350 s->done += ret;
351 prepped -= diff;
352 goto submit_more;
353 }
354 s->done += ret;
355 prepped = 0;
356 continue;
357 } else if (ret < 0) {
ac122fea 358 if (errno == EAGAIN) {
c9fb4c5b
JA
359 if (s->finish)
360 break;
361 if (this_reap)
362 goto submit;
c9fb4c5b
JA
363 to_submit = 0;
364 goto submit;
365 }
ac122fea 366 printf("io_submit: %s\n", strerror(errno));
c9fb4c5b
JA
367 break;
368 }
369 } while (!s->finish);
a7086591 370
c9fb4c5b
JA
371 finish = 1;
372 return NULL;
373}
374
375static void sig_int(int sig)
376{
377 printf("Exiting on signal %d\n", sig);
378 submitters[0].finish = 1;
379 finish = 1;
380}
381
382static void arm_sig_int(void)
383{
384 struct sigaction act;
385
386 memset(&act, 0, sizeof(act));
387 act.sa_handler = sig_int;
388 act.sa_flags = SA_RESTART;
389 sigaction(SIGINT, &act, NULL);
390}
391
c3e2fc25
JA
392static int setup_ring(struct submitter *s)
393{
e31b8288
JA
394 struct io_sq_ring *sring = &s->sq_ring;
395 struct io_cq_ring *cring = &s->cq_ring;
396 struct io_uring_params p;
2ea53ca3 397 int ret, fd;
c3e2fc25 398 void *ptr;
c3e2fc25
JA
399
400 memset(&p, 0, sizeof(p));
401
7d1435e6 402 if (polled && !do_nop)
0e47f11b 403 p.flags |= IORING_SETUP_IOPOLL;
3d7d00a3 404 if (sq_thread_poll) {
2ea53ca3 405 p.flags |= IORING_SETUP_SQPOLL;
3ade18a3 406 if (sq_thread_cpu != -1) {
3d7d00a3 407 p.flags |= IORING_SETUP_SQ_AFF;
3ade18a3
JA
408 p.sq_thread_cpu = sq_thread_cpu;
409 }
c3e2fc25
JA
410 }
411
2ea53ca3 412 fd = io_uring_setup(DEPTH, &p);
c3e2fc25
JA
413 if (fd < 0) {
414 perror("io_uring_setup");
415 return 1;
416 }
f310970e 417 s->ring_fd = fd;
2ea53ca3
JA
418
419 if (fixedbufs) {
420 ret = io_uring_register_buffers(s);
421 if (ret < 0) {
a7abc9fb 422 perror("io_uring_register_buffers");
2ea53ca3
JA
423 return 1;
424 }
425 }
426
8c5fa755
JA
427 if (register_files) {
428 ret = io_uring_register_files(s);
429 if (ret < 0) {
430 perror("io_uring_register_files");
431 return 1;
432 }
a7abc9fb
JA
433 }
434
e2239016 435 ptr = mmap(0, p.sq_off.array + p.sq_entries * sizeof(__u32),
f310970e
JA
436 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
437 IORING_OFF_SQ_RING);
c3e2fc25
JA
438 printf("sq_ring ptr = 0x%p\n", ptr);
439 sring->head = ptr + p.sq_off.head;
440 sring->tail = ptr + p.sq_off.tail;
441 sring->ring_mask = ptr + p.sq_off.ring_mask;
442 sring->ring_entries = ptr + p.sq_off.ring_entries;
ce1705de 443 sring->flags = ptr + p.sq_off.flags;
ac122fea 444 sring->array = ptr + p.sq_off.array;
c3e2fc25
JA
445 sq_ring_mask = *sring->ring_mask;
446
f0403f94 447 s->sqes = mmap(0, p.sq_entries * sizeof(struct io_uring_sqe),
f310970e 448 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
f0403f94
JA
449 IORING_OFF_SQES);
450 printf("sqes ptr = 0x%p\n", s->sqes);
c3e2fc25 451
f0403f94 452 ptr = mmap(0, p.cq_off.cqes + p.cq_entries * sizeof(struct io_uring_cqe),
f310970e
JA
453 PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd,
454 IORING_OFF_CQ_RING);
c3e2fc25
JA
455 printf("cq_ring ptr = 0x%p\n", ptr);
456 cring->head = ptr + p.cq_off.head;
457 cring->tail = ptr + p.cq_off.tail;
458 cring->ring_mask = ptr + p.cq_off.ring_mask;
459 cring->ring_entries = ptr + p.cq_off.ring_entries;
f0403f94 460 cring->cqes = ptr + p.cq_off.cqes;
c3e2fc25
JA
461 cq_ring_mask = *cring->ring_mask;
462 return 0;
463}
464
2e7888ef
JA
465static void file_depths(char *buf)
466{
467 struct submitter *s = &submitters[0];
468 char *p;
469 int i;
470
ac21bfab 471 buf[0] = '\0';
2e7888ef
JA
472 p = buf;
473 for (i = 0; i < s->nr_files; i++) {
474 struct file *f = &s->files[i];
475
476 if (i + 1 == s->nr_files)
477 p += sprintf(p, "%d", f->pending_ios);
478 else
479 p += sprintf(p, "%d, ", f->pending_ios);
480 }
481}
482
c9fb4c5b
JA
483int main(int argc, char *argv[])
484{
485 struct submitter *s = &submitters[0];
305c06ce 486 unsigned long done, calls, reap, cache_hit, cache_miss;
a7086591 487 int err, i, flags, fd;
2e7888ef 488 char *fdepths;
c3e2fc25 489 void *ret;
c9fb4c5b 490
8025517d 491 if (!do_nop && argc < 2) {
c9fb4c5b
JA
492 printf("%s: filename\n", argv[0]);
493 return 1;
494 }
495
701d1277 496 flags = O_RDONLY | O_NOATIME;
a7086591
JA
497 if (!buffered)
498 flags |= O_DIRECT;
499
500 i = 1;
8025517d 501 while (!do_nop && i < argc) {
a7086591
JA
502 struct file *f = &s->files[s->nr_files];
503
504 fd = open(argv[i], flags);
505 if (fd < 0) {
506 perror("open");
507 return 1;
508 }
48e698fa 509 f->real_fd = fd;
a7086591
JA
510 if (get_file_size(f)) {
511 printf("failed getting size of device/file\n");
512 return 1;
513 }
514 if (f->max_blocks <= 1) {
515 printf("Zero file/device size?\n");
516 return 1;
517 }
518 f->max_blocks--;
519
520 printf("Added file %s\n", argv[i]);
521 s->nr_files++;
522 i++;
523 }
524
b3ce355d
JA
525 if (fixedbufs) {
526 struct rlimit rlim;
527
528 rlim.rlim_cur = RLIM_INFINITY;
529 rlim.rlim_max = RLIM_INFINITY;
530 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) {
531 perror("setrlimit");
532 return 1;
533 }
c9fb4c5b
JA
534 }
535
536 arm_sig_int();
537
c3e2fc25
JA
538 for (i = 0; i < DEPTH; i++) {
539 void *buf;
c9fb4c5b 540
c3e2fc25 541 if (posix_memalign(&buf, BS, BS)) {
c9fb4c5b
JA
542 printf("failed alloc\n");
543 return 1;
544 }
c3e2fc25
JA
545 s->iovecs[i].iov_base = buf;
546 s->iovecs[i].iov_len = BS;
c9fb4c5b
JA
547 }
548
c3e2fc25 549 err = setup_ring(s);
c9fb4c5b 550 if (err) {
c3e2fc25 551 printf("ring setup failed: %s, %d\n", strerror(errno), err);
c9fb4c5b
JA
552 return 1;
553 }
c3e2fc25
JA
554 printf("polled=%d, fixedbufs=%d, buffered=%d", polled, fixedbufs, buffered);
555 printf(" QD=%d, sq_ring=%d, cq_ring=%d\n", DEPTH, *s->sq_ring.ring_entries, *s->cq_ring.ring_entries);
c9fb4c5b
JA
556
557 pthread_create(&s->thread, NULL, submitter_fn, s);
558
2e7888ef 559 fdepths = malloc(8 * s->nr_files);
305c06ce 560 cache_hit = cache_miss = reap = calls = done = 0;
c9fb4c5b
JA
561 do {
562 unsigned long this_done = 0;
563 unsigned long this_reap = 0;
564 unsigned long this_call = 0;
305c06ce
JA
565 unsigned long this_cache_hit = 0;
566 unsigned long this_cache_miss = 0;
c9fb4c5b 567 unsigned long rpc = 0, ipc = 0;
305c06ce 568 double hit = 0.0;
c9fb4c5b
JA
569
570 sleep(1);
571 this_done += s->done;
572 this_call += s->calls;
573 this_reap += s->reaps;
305c06ce
JA
574 this_cache_hit += s->cachehit;
575 this_cache_miss += s->cachemiss;
576 if (this_cache_hit && this_cache_miss) {
577 unsigned long hits, total;
578
579 hits = this_cache_hit - cache_hit;
580 total = hits + this_cache_miss - cache_miss;
581 hit = (double) hits / (double) total;
582 hit *= 100.0;
583 }
c9fb4c5b
JA
584 if (this_call - calls) {
585 rpc = (this_done - done) / (this_call - calls);
586 ipc = (this_reap - reap) / (this_call - calls);
191561c1
JA
587 } else
588 rpc = ipc = -1;
2e7888ef
JA
589 file_depths(fdepths);
590 printf("IOPS=%lu, IOS/call=%ld/%ld, inflight=%u (%s), Cachehit=%0.2f%%\n",
c9fb4c5b 591 this_done - done, rpc, ipc, s->inflight,
2e7888ef 592 fdepths, hit);
c9fb4c5b
JA
593 done = this_done;
594 calls = this_call;
595 reap = this_reap;
305c06ce
JA
596 cache_hit = s->cachehit;
597 cache_miss = s->cachemiss;
c9fb4c5b
JA
598 } while (!finish);
599
600 pthread_join(s->thread, &ret);
e31b8288 601 close(s->ring_fd);
2e7888ef 602 free(fdepths);
c9fb4c5b
JA
603 return 0;
604}