Commit | Line | Data |
---|---|---|
52885fa2 | 1 | /* |
bffad86f | 2 | * io_uring engine |
52885fa2 | 3 | * |
bffad86f | 4 | * IO engine using the new native Linux aio io_uring interface. See: |
a90cd050 | 5 | * |
bffad86f | 6 | * http://git.kernel.dk/cgit/linux-block/log/?h=io_uring |
52885fa2 JA |
7 | * |
8 | */ | |
9 | #include <stdlib.h> | |
10 | #include <unistd.h> | |
11 | #include <errno.h> | |
52885fa2 JA |
12 | #include <sys/time.h> |
13 | #include <sys/resource.h> | |
14 | ||
15 | #include "../fio.h" | |
16 | #include "../lib/pow2.h" | |
17 | #include "../optgroup.h" | |
18 | #include "../lib/memalign.h" | |
b87aa01a | 19 | #include "../lib/fls.h" |
52885fa2 | 20 | |
bffad86f | 21 | #ifdef ARCH_HAVE_IOURING |
52885fa2 | 22 | |
57fa61f0 | 23 | #include "../lib/types.h" |
f3e769a4 | 24 | #include "../os/linux/io_uring.h" |
9a2d78b3 | 25 | |
bffad86f | 26 | struct io_sq_ring { |
e2239016 JA |
27 | unsigned *head; |
28 | unsigned *tail; | |
29 | unsigned *ring_mask; | |
30 | unsigned *ring_entries; | |
31 | unsigned *flags; | |
32 | unsigned *array; | |
52885fa2 JA |
33 | }; |
34 | ||
bffad86f | 35 | struct io_cq_ring { |
e2239016 JA |
36 | unsigned *head; |
37 | unsigned *tail; | |
38 | unsigned *ring_mask; | |
39 | unsigned *ring_entries; | |
f0403f94 | 40 | struct io_uring_cqe *cqes; |
9a2d78b3 JA |
41 | }; |
42 | ||
bffad86f | 43 | struct ioring_mmap { |
9a2d78b3 JA |
44 | void *ptr; |
45 | size_t len; | |
52885fa2 JA |
46 | }; |
47 | ||
bffad86f | 48 | struct ioring_data { |
9a2d78b3 JA |
49 | int ring_fd; |
50 | ||
52885fa2 JA |
51 | struct io_u **io_u_index; |
52 | ||
5ffd5626 JA |
53 | int *fds; |
54 | ||
bffad86f | 55 | struct io_sq_ring sq_ring; |
f0403f94 | 56 | struct io_uring_sqe *sqes; |
9a2d78b3 | 57 | struct iovec *iovecs; |
b87aa01a | 58 | unsigned sq_ring_mask; |
52885fa2 | 59 | |
bffad86f | 60 | struct io_cq_ring cq_ring; |
b87aa01a | 61 | unsigned cq_ring_mask; |
52885fa2 JA |
62 | |
63 | int queued; | |
64 | int cq_ring_off; | |
b87aa01a | 65 | unsigned iodepth; |
96563db9 | 66 | |
bffad86f | 67 | struct ioring_mmap mmap[3]; |
52885fa2 JA |
68 | }; |
69 | ||
bffad86f | 70 | struct ioring_options { |
52885fa2 JA |
71 | void *pad; |
72 | unsigned int hipri; | |
73 | unsigned int fixedbufs; | |
5ffd5626 | 74 | unsigned int registerfiles; |
3d7d00a3 | 75 | unsigned int sqpoll_thread; |
2ea53ca3 JA |
76 | unsigned int sqpoll_set; |
77 | unsigned int sqpoll_cpu; | |
4a87b584 | 78 | unsigned int uncached; |
52885fa2 JA |
79 | }; |
80 | ||
2ea53ca3 | 81 | static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val) |
a90cd050 | 82 | { |
bffad86f | 83 | struct ioring_options *o = data; |
a90cd050 | 84 | |
2ea53ca3 JA |
85 | o->sqpoll_cpu = *val; |
86 | o->sqpoll_set = 1; | |
a90cd050 JA |
87 | return 0; |
88 | } | |
89 | ||
52885fa2 JA |
90 | static struct fio_option options[] = { |
91 | { | |
92 | .name = "hipri", | |
93 | .lname = "High Priority", | |
94 | .type = FIO_OPT_STR_SET, | |
bffad86f | 95 | .off1 = offsetof(struct ioring_options, hipri), |
52885fa2 JA |
96 | .help = "Use polled IO completions", |
97 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 98 | .group = FIO_OPT_G_IOURING, |
52885fa2 JA |
99 | }, |
100 | { | |
101 | .name = "fixedbufs", | |
102 | .lname = "Fixed (pre-mapped) IO buffers", | |
103 | .type = FIO_OPT_STR_SET, | |
bffad86f | 104 | .off1 = offsetof(struct ioring_options, fixedbufs), |
52885fa2 JA |
105 | .help = "Pre map IO buffers", |
106 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 107 | .group = FIO_OPT_G_IOURING, |
52885fa2 | 108 | }, |
5ffd5626 JA |
109 | { |
110 | .name = "registerfiles", | |
111 | .lname = "Register file set", | |
112 | .type = FIO_OPT_STR_SET, | |
113 | .off1 = offsetof(struct ioring_options, registerfiles), | |
114 | .help = "Pre-open/register files", | |
115 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 116 | .group = FIO_OPT_G_IOURING, |
5ffd5626 | 117 | }, |
771c9901 JA |
118 | { |
119 | .name = "sqthread_poll", | |
3d7d00a3 JA |
120 | .lname = "Kernel SQ thread polling", |
121 | .type = FIO_OPT_INT, | |
122 | .off1 = offsetof(struct ioring_options, sqpoll_thread), | |
123 | .help = "Offload submission/completion to kernel thread", | |
124 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 125 | .group = FIO_OPT_G_IOURING, |
3d7d00a3 JA |
126 | }, |
127 | { | |
128 | .name = "sqthread_poll_cpu", | |
129 | .lname = "SQ Thread Poll CPU", | |
2ea53ca3 JA |
130 | .type = FIO_OPT_INT, |
131 | .cb = fio_ioring_sqpoll_cb, | |
3d7d00a3 | 132 | .help = "What CPU to run SQ thread polling on", |
a90cd050 | 133 | .category = FIO_OPT_C_ENGINE, |
27f436d9 | 134 | .group = FIO_OPT_G_IOURING, |
a90cd050 | 135 | }, |
4a87b584 JA |
136 | { |
137 | .name = "uncached", | |
138 | .lname = "Uncached", | |
139 | .type = FIO_OPT_INT, | |
140 | .off1 = offsetof(struct ioring_options, uncached), | |
141 | .help = "Use RWF_UNCACHED for buffered read/writes", | |
142 | .category = FIO_OPT_C_ENGINE, | |
143 | .group = FIO_OPT_G_IOURING, | |
144 | }, | |
52885fa2 JA |
145 | { |
146 | .name = NULL, | |
147 | }, | |
148 | }; | |
149 | ||
bffad86f | 150 | static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit, |
52885fa2 JA |
151 | unsigned int min_complete, unsigned int flags) |
152 | { | |
9a2d78b3 | 153 | return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit, |
521164fa | 154 | min_complete, flags, NULL, 0); |
52885fa2 JA |
155 | } |
156 | ||
bffad86f | 157 | static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 158 | { |
bffad86f | 159 | struct ioring_data *ld = td->io_ops_data; |
cfcc8564 | 160 | struct ioring_options *o = td->eo; |
52885fa2 | 161 | struct fio_file *f = io_u->file; |
f0403f94 | 162 | struct io_uring_sqe *sqe; |
52885fa2 | 163 | |
f0403f94 | 164 | sqe = &ld->sqes[io_u->index]; |
34d6090e AF |
165 | |
166 | /* zero out fields not used in this submission */ | |
167 | memset(sqe, 0, sizeof(*sqe)); | |
168 | ||
5ffd5626 JA |
169 | if (o->registerfiles) { |
170 | sqe->fd = f->engine_pos; | |
171 | sqe->flags = IOSQE_FIXED_FILE; | |
172 | } else { | |
173 | sqe->fd = f->fd; | |
5ffd5626 | 174 | } |
52885fa2 | 175 | |
e3970057 | 176 | if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) { |
f0403f94 | 177 | if (o->fixedbufs) { |
48e698fa JA |
178 | if (io_u->ddir == DDIR_READ) |
179 | sqe->opcode = IORING_OP_READ_FIXED; | |
180 | else | |
181 | sqe->opcode = IORING_OP_WRITE_FIXED; | |
919850d2 | 182 | sqe->addr = (unsigned long) io_u->xfer_buf; |
f0403f94 | 183 | sqe->len = io_u->xfer_buflen; |
2ea53ca3 | 184 | sqe->buf_index = io_u->index; |
cfcc8564 | 185 | } else { |
48e698fa JA |
186 | if (io_u->ddir == DDIR_READ) |
187 | sqe->opcode = IORING_OP_READV; | |
188 | else | |
189 | sqe->opcode = IORING_OP_WRITEV; | |
919850d2 | 190 | sqe->addr = (unsigned long) &ld->iovecs[io_u->index]; |
f0403f94 | 191 | sqe->len = 1; |
cfcc8564 | 192 | } |
4a87b584 JA |
193 | if (!td->o.odirect && o->uncached) |
194 | sqe->rw_flags = RWF_UNCACHED; | |
f0403f94 | 195 | sqe->off = io_u->offset; |
48e698fa | 196 | } else if (ddir_sync(io_u->ddir)) { |
01387bfe AF |
197 | if (io_u->ddir == DDIR_SYNC_FILE_RANGE) { |
198 | sqe->off = f->first_write; | |
199 | sqe->len = f->last_write - f->first_write; | |
200 | sqe->sync_range_flags = td->o.sync_file_range; | |
201 | sqe->opcode = IORING_OP_SYNC_FILE_RANGE; | |
202 | } else { | |
01387bfe AF |
203 | if (io_u->ddir == DDIR_DATASYNC) |
204 | sqe->fsync_flags |= IORING_FSYNC_DATASYNC; | |
205 | sqe->opcode = IORING_OP_FSYNC; | |
206 | } | |
48e698fa | 207 | } |
52885fa2 | 208 | |
48e698fa | 209 | sqe->user_data = (unsigned long) io_u; |
52885fa2 JA |
210 | return 0; |
211 | } | |
212 | ||
bffad86f | 213 | static struct io_u *fio_ioring_event(struct thread_data *td, int event) |
52885fa2 | 214 | { |
bffad86f | 215 | struct ioring_data *ld = td->io_ops_data; |
f0403f94 | 216 | struct io_uring_cqe *cqe; |
52885fa2 | 217 | struct io_u *io_u; |
b87aa01a | 218 | unsigned index; |
52885fa2 | 219 | |
b87aa01a | 220 | index = (event + ld->cq_ring_off) & ld->cq_ring_mask; |
52885fa2 | 221 | |
f0403f94 | 222 | cqe = &ld->cq_ring.cqes[index]; |
e3466352 | 223 | io_u = (struct io_u *) (uintptr_t) cqe->user_data; |
52885fa2 | 224 | |
f0403f94 JA |
225 | if (cqe->res != io_u->xfer_buflen) { |
226 | if (cqe->res > io_u->xfer_buflen) | |
227 | io_u->error = -cqe->res; | |
52885fa2 | 228 | else |
f0403f94 | 229 | io_u->resid = io_u->xfer_buflen - cqe->res; |
52885fa2 JA |
230 | } else |
231 | io_u->error = 0; | |
232 | ||
233 | return io_u; | |
234 | } | |
235 | ||
bffad86f | 236 | static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events, |
52885fa2 JA |
237 | unsigned int max) |
238 | { | |
bffad86f JA |
239 | struct ioring_data *ld = td->io_ops_data; |
240 | struct io_cq_ring *ring = &ld->cq_ring; | |
e2239016 | 241 | unsigned head, reaped = 0; |
52885fa2 | 242 | |
9a2d78b3 | 243 | head = *ring->head; |
52885fa2 JA |
244 | do { |
245 | read_barrier(); | |
9a2d78b3 | 246 | if (head == *ring->tail) |
52885fa2 JA |
247 | break; |
248 | reaped++; | |
249 | head++; | |
52885fa2 JA |
250 | } while (reaped + events < max); |
251 | ||
9a2d78b3 | 252 | *ring->head = head; |
52885fa2 JA |
253 | write_barrier(); |
254 | return reaped; | |
255 | } | |
256 | ||
bffad86f JA |
257 | static int fio_ioring_getevents(struct thread_data *td, unsigned int min, |
258 | unsigned int max, const struct timespec *t) | |
52885fa2 | 259 | { |
bffad86f | 260 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 | 261 | unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min; |
bffad86f JA |
262 | struct ioring_options *o = td->eo; |
263 | struct io_cq_ring *ring = &ld->cq_ring; | |
b87aa01a JA |
264 | unsigned events = 0; |
265 | int r; | |
52885fa2 | 266 | |
9a2d78b3 | 267 | ld->cq_ring_off = *ring->head; |
52885fa2 | 268 | do { |
bffad86f | 269 | r = fio_ioring_cqring_reap(td, events, max); |
52885fa2 JA |
270 | if (r) { |
271 | events += r; | |
f7cbbbf8 ST |
272 | if (actual_min != 0) |
273 | actual_min -= r; | |
52885fa2 JA |
274 | continue; |
275 | } | |
276 | ||
3d7d00a3 | 277 | if (!o->sqpoll_thread) { |
9a2d78b3 JA |
278 | r = io_uring_enter(ld, 0, actual_min, |
279 | IORING_ENTER_GETEVENTS); | |
771c9901 | 280 | if (r < 0) { |
f6abd731 | 281 | if (errno == EAGAIN || errno == EINTR) |
771c9901 | 282 | continue; |
9a2d78b3 | 283 | td_verror(td, errno, "io_uring_enter"); |
771c9901 JA |
284 | break; |
285 | } | |
52885fa2 JA |
286 | } |
287 | } while (events < min); | |
288 | ||
289 | return r < 0 ? r : events; | |
290 | } | |
291 | ||
bffad86f JA |
292 | static enum fio_q_status fio_ioring_queue(struct thread_data *td, |
293 | struct io_u *io_u) | |
52885fa2 | 294 | { |
bffad86f JA |
295 | struct ioring_data *ld = td->io_ops_data; |
296 | struct io_sq_ring *ring = &ld->sq_ring; | |
52885fa2 JA |
297 | unsigned tail, next_tail; |
298 | ||
299 | fio_ro_check(td, io_u); | |
300 | ||
b87aa01a | 301 | if (ld->queued == ld->iodepth) |
52885fa2 JA |
302 | return FIO_Q_BUSY; |
303 | ||
52885fa2 JA |
304 | if (io_u->ddir == DDIR_TRIM) { |
305 | if (ld->queued) | |
306 | return FIO_Q_BUSY; | |
307 | ||
308 | do_io_u_trim(td, io_u); | |
309 | io_u_mark_submit(td, 1); | |
310 | io_u_mark_complete(td, 1); | |
311 | return FIO_Q_COMPLETED; | |
312 | } | |
313 | ||
9a2d78b3 | 314 | tail = *ring->tail; |
52885fa2 | 315 | next_tail = tail + 1; |
52885fa2 | 316 | read_barrier(); |
9a2d78b3 | 317 | if (next_tail == *ring->head) |
52885fa2 JA |
318 | return FIO_Q_BUSY; |
319 | ||
2b9415dd JA |
320 | /* ensure sqe stores are ordered with tail update */ |
321 | write_barrier(); | |
b87aa01a | 322 | ring->array[tail & ld->sq_ring_mask] = io_u->index; |
9a2d78b3 | 323 | *ring->tail = next_tail; |
52885fa2 JA |
324 | write_barrier(); |
325 | ||
326 | ld->queued++; | |
327 | return FIO_Q_QUEUED; | |
328 | } | |
329 | ||
bffad86f | 330 | static void fio_ioring_queued(struct thread_data *td, int start, int nr) |
52885fa2 | 331 | { |
bffad86f | 332 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
333 | struct timespec now; |
334 | ||
335 | if (!fio_fill_issue_time(td)) | |
336 | return; | |
337 | ||
338 | fio_gettime(&now, NULL); | |
339 | ||
340 | while (nr--) { | |
bffad86f | 341 | struct io_sq_ring *ring = &ld->sq_ring; |
9a2d78b3 | 342 | int index = ring->array[start & ld->sq_ring_mask]; |
f8289afc | 343 | struct io_u *io_u = ld->io_u_index[index]; |
52885fa2 JA |
344 | |
345 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
346 | io_u_queued(td, io_u); | |
347 | ||
348 | start++; | |
52885fa2 JA |
349 | } |
350 | } | |
351 | ||
bffad86f | 352 | static int fio_ioring_commit(struct thread_data *td) |
52885fa2 | 353 | { |
bffad86f JA |
354 | struct ioring_data *ld = td->io_ops_data; |
355 | struct ioring_options *o = td->eo; | |
52885fa2 JA |
356 | int ret; |
357 | ||
358 | if (!ld->queued) | |
359 | return 0; | |
360 | ||
3d7d00a3 JA |
361 | /* |
362 | * Kernel side does submission. just need to check if the ring is | |
363 | * flagged as needing a kick, if so, call io_uring_enter(). This | |
364 | * only happens if we've been idle too long. | |
365 | */ | |
366 | if (o->sqpoll_thread) { | |
bffad86f | 367 | struct io_sq_ring *ring = &ld->sq_ring; |
4cdbc048 | 368 | |
2ea53ca3 | 369 | read_barrier(); |
9a2d78b3 | 370 | if (*ring->flags & IORING_SQ_NEED_WAKEUP) |
b532dd6d JA |
371 | io_uring_enter(ld, ld->queued, 0, |
372 | IORING_ENTER_SQ_WAKEUP); | |
771c9901 JA |
373 | ld->queued = 0; |
374 | return 0; | |
375 | } | |
376 | ||
52885fa2 | 377 | do { |
9a2d78b3 | 378 | unsigned start = *ld->sq_ring.head; |
52885fa2 JA |
379 | long nr = ld->queued; |
380 | ||
9a2d78b3 | 381 | ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS); |
52885fa2 | 382 | if (ret > 0) { |
bffad86f | 383 | fio_ioring_queued(td, start, ret); |
52885fa2 JA |
384 | io_u_mark_submit(td, ret); |
385 | ||
386 | ld->queued -= ret; | |
387 | ret = 0; | |
a90cd050 JA |
388 | } else if (!ret) { |
389 | io_u_mark_submit(td, ret); | |
52885fa2 | 390 | continue; |
a90cd050 | 391 | } else { |
f6abd731 | 392 | if (errno == EAGAIN || errno == EINTR) { |
bffad86f | 393 | ret = fio_ioring_cqring_reap(td, 0, ld->queued); |
a90cd050 JA |
394 | if (ret) |
395 | continue; | |
396 | /* Shouldn't happen */ | |
397 | usleep(1); | |
398 | continue; | |
52885fa2 | 399 | } |
9a2d78b3 | 400 | td_verror(td, errno, "io_uring_enter submit"); |
52885fa2 | 401 | break; |
a90cd050 | 402 | } |
52885fa2 JA |
403 | } while (ld->queued); |
404 | ||
405 | return ret; | |
406 | } | |
407 | ||
bffad86f | 408 | static void fio_ioring_unmap(struct ioring_data *ld) |
52885fa2 | 409 | { |
9a2d78b3 | 410 | int i; |
52885fa2 | 411 | |
9a2d78b3 JA |
412 | for (i = 0; i < ARRAY_SIZE(ld->mmap); i++) |
413 | munmap(ld->mmap[i].ptr, ld->mmap[i].len); | |
414 | close(ld->ring_fd); | |
b87aa01a JA |
415 | } |
416 | ||
bffad86f | 417 | static void fio_ioring_cleanup(struct thread_data *td) |
52885fa2 | 418 | { |
bffad86f | 419 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
420 | |
421 | if (ld) { | |
52885fa2 | 422 | if (!(td->flags & TD_F_CHILD)) |
bffad86f | 423 | fio_ioring_unmap(ld); |
9a2d78b3 | 424 | |
52885fa2 | 425 | free(ld->io_u_index); |
9a2d78b3 | 426 | free(ld->iovecs); |
5ffd5626 | 427 | free(ld->fds); |
52885fa2 JA |
428 | free(ld); |
429 | } | |
430 | } | |
431 | ||
bffad86f | 432 | static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p) |
9a2d78b3 | 433 | { |
bffad86f JA |
434 | struct io_sq_ring *sring = &ld->sq_ring; |
435 | struct io_cq_ring *cring = &ld->cq_ring; | |
9a2d78b3 JA |
436 | void *ptr; |
437 | ||
e2239016 | 438 | ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32); |
9a2d78b3 JA |
439 | ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE, |
440 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
441 | IORING_OFF_SQ_RING); | |
442 | ld->mmap[0].ptr = ptr; | |
443 | sring->head = ptr + p->sq_off.head; | |
444 | sring->tail = ptr + p->sq_off.tail; | |
445 | sring->ring_mask = ptr + p->sq_off.ring_mask; | |
446 | sring->ring_entries = ptr + p->sq_off.ring_entries; | |
447 | sring->flags = ptr + p->sq_off.flags; | |
ac122fea | 448 | sring->array = ptr + p->sq_off.array; |
9a2d78b3 JA |
449 | ld->sq_ring_mask = *sring->ring_mask; |
450 | ||
f0403f94 JA |
451 | ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe); |
452 | ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE, | |
9a2d78b3 | 453 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, |
f0403f94 JA |
454 | IORING_OFF_SQES); |
455 | ld->mmap[1].ptr = ld->sqes; | |
9a2d78b3 | 456 | |
f0403f94 JA |
457 | ld->mmap[2].len = p->cq_off.cqes + |
458 | p->cq_entries * sizeof(struct io_uring_cqe); | |
9a2d78b3 JA |
459 | ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE, |
460 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
461 | IORING_OFF_CQ_RING); | |
462 | ld->mmap[2].ptr = ptr; | |
463 | cring->head = ptr + p->cq_off.head; | |
464 | cring->tail = ptr + p->cq_off.tail; | |
465 | cring->ring_mask = ptr + p->cq_off.ring_mask; | |
466 | cring->ring_entries = ptr + p->cq_off.ring_entries; | |
f0403f94 | 467 | cring->cqes = ptr + p->cq_off.cqes; |
9a2d78b3 JA |
468 | ld->cq_ring_mask = *cring->ring_mask; |
469 | return 0; | |
470 | } | |
471 | ||
bffad86f | 472 | static int fio_ioring_queue_init(struct thread_data *td) |
52885fa2 | 473 | { |
bffad86f JA |
474 | struct ioring_data *ld = td->io_ops_data; |
475 | struct ioring_options *o = td->eo; | |
52885fa2 | 476 | int depth = td->o.iodepth; |
bffad86f | 477 | struct io_uring_params p; |
9a2d78b3 JA |
478 | int ret; |
479 | ||
480 | memset(&p, 0, sizeof(p)); | |
52885fa2 JA |
481 | |
482 | if (o->hipri) | |
bffad86f | 483 | p.flags |= IORING_SETUP_IOPOLL; |
3d7d00a3 JA |
484 | if (o->sqpoll_thread) { |
485 | p.flags |= IORING_SETUP_SQPOLL; | |
486 | if (o->sqpoll_set) { | |
487 | p.flags |= IORING_SETUP_SQ_AFF; | |
488 | p.sq_thread_cpu = o->sqpoll_cpu; | |
489 | } | |
f635f1fb | 490 | } |
a90cd050 | 491 | |
2ea53ca3 | 492 | ret = syscall(__NR_sys_io_uring_setup, depth, &p); |
9a2d78b3 JA |
493 | if (ret < 0) |
494 | return ret; | |
495 | ||
496 | ld->ring_fd = ret; | |
2ea53ca3 JA |
497 | |
498 | if (o->fixedbufs) { | |
2d644205 JA |
499 | struct rlimit rlim = { |
500 | .rlim_cur = RLIM_INFINITY, | |
501 | .rlim_max = RLIM_INFINITY, | |
502 | }; | |
503 | ||
504 | if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) | |
505 | return -1; | |
506 | ||
2ea53ca3 | 507 | ret = syscall(__NR_sys_io_uring_register, ld->ring_fd, |
919850d2 | 508 | IORING_REGISTER_BUFFERS, ld->iovecs, depth); |
2ea53ca3 JA |
509 | if (ret < 0) |
510 | return ret; | |
511 | } | |
512 | ||
bffad86f | 513 | return fio_ioring_mmap(ld, &p); |
52885fa2 JA |
514 | } |
515 | ||
5ffd5626 JA |
516 | static int fio_ioring_register_files(struct thread_data *td) |
517 | { | |
518 | struct ioring_data *ld = td->io_ops_data; | |
519 | struct fio_file *f; | |
520 | unsigned int i; | |
521 | int ret; | |
522 | ||
523 | ld->fds = calloc(td->o.nr_files, sizeof(int)); | |
524 | ||
525 | for_each_file(td, f, i) { | |
526 | ret = generic_open_file(td, f); | |
527 | if (ret) | |
528 | goto err; | |
529 | ld->fds[i] = f->fd; | |
530 | f->engine_pos = i; | |
531 | } | |
532 | ||
533 | ret = syscall(__NR_sys_io_uring_register, ld->ring_fd, | |
534 | IORING_REGISTER_FILES, ld->fds, td->o.nr_files); | |
535 | if (ret) { | |
536 | err: | |
537 | free(ld->fds); | |
538 | ld->fds = NULL; | |
539 | } | |
540 | ||
541 | /* | |
542 | * Pretend the file is closed again, and really close it if we hit | |
543 | * an error. | |
544 | */ | |
545 | for_each_file(td, f, i) { | |
546 | if (ret) { | |
547 | int fio_unused ret2; | |
548 | ret2 = generic_close_file(td, f); | |
549 | } else | |
550 | f->fd = -1; | |
551 | } | |
552 | ||
553 | return ret; | |
554 | } | |
555 | ||
bffad86f | 556 | static int fio_ioring_post_init(struct thread_data *td) |
52885fa2 | 557 | { |
bffad86f | 558 | struct ioring_data *ld = td->io_ops_data; |
5ffd5626 | 559 | struct ioring_options *o = td->eo; |
52885fa2 | 560 | struct io_u *io_u; |
650346e1 | 561 | int err, i; |
52885fa2 | 562 | |
650346e1 JA |
563 | for (i = 0; i < td->o.iodepth; i++) { |
564 | struct iovec *iov = &ld->iovecs[i]; | |
9a2d78b3 | 565 | |
650346e1 JA |
566 | io_u = ld->io_u_index[i]; |
567 | iov->iov_base = io_u->buf; | |
568 | iov->iov_len = td_max_bs(td); | |
52885fa2 JA |
569 | } |
570 | ||
bffad86f | 571 | err = fio_ioring_queue_init(td); |
52885fa2 | 572 | if (err) { |
d63a472d | 573 | td_verror(td, errno, "io_queue_init"); |
52885fa2 JA |
574 | return 1; |
575 | } | |
576 | ||
5ffd5626 JA |
577 | if (o->registerfiles) { |
578 | err = fio_ioring_register_files(td); | |
579 | if (err) { | |
580 | td_verror(td, errno, "ioring_register_files"); | |
581 | return 1; | |
582 | } | |
583 | } | |
584 | ||
52885fa2 JA |
585 | return 0; |
586 | } | |
587 | ||
9a2d78b3 JA |
588 | static unsigned roundup_pow2(unsigned depth) |
589 | { | |
590 | return 1UL << __fls(depth - 1); | |
591 | } | |
592 | ||
bffad86f | 593 | static int fio_ioring_init(struct thread_data *td) |
52885fa2 | 594 | { |
5ffd5626 | 595 | struct ioring_options *o = td->eo; |
bffad86f | 596 | struct ioring_data *ld; |
52885fa2 | 597 | |
5ffd5626 JA |
598 | /* sqthread submission requires registered files */ |
599 | if (o->sqpoll_thread) | |
600 | o->registerfiles = 1; | |
601 | ||
602 | if (o->registerfiles && td->o.nr_files != td->o.open_files) { | |
603 | log_err("fio: io_uring registered files require nr_files to " | |
604 | "be identical to open_files\n"); | |
605 | return 1; | |
606 | } | |
607 | ||
52885fa2 JA |
608 | ld = calloc(1, sizeof(*ld)); |
609 | ||
b87aa01a JA |
610 | /* ring depth must be a power-of-2 */ |
611 | ld->iodepth = td->o.iodepth; | |
612 | td->o.iodepth = roundup_pow2(td->o.iodepth); | |
613 | ||
52885fa2 JA |
614 | /* io_u index */ |
615 | ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *)); | |
650346e1 | 616 | ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec)); |
52885fa2 JA |
617 | |
618 | td->io_ops_data = ld; | |
619 | return 0; | |
620 | } | |
621 | ||
bffad86f | 622 | static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 623 | { |
bffad86f | 624 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
625 | |
626 | ld->io_u_index[io_u->index] = io_u; | |
627 | return 0; | |
628 | } | |
629 | ||
5ffd5626 JA |
630 | static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f) |
631 | { | |
632 | struct ioring_data *ld = td->io_ops_data; | |
633 | struct ioring_options *o = td->eo; | |
634 | ||
17318cf6 | 635 | if (!ld || !o->registerfiles) |
5ffd5626 JA |
636 | return generic_open_file(td, f); |
637 | ||
638 | f->fd = ld->fds[f->engine_pos]; | |
639 | return 0; | |
640 | } | |
641 | ||
642 | static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f) | |
643 | { | |
17318cf6 | 644 | struct ioring_data *ld = td->io_ops_data; |
5ffd5626 JA |
645 | struct ioring_options *o = td->eo; |
646 | ||
17318cf6 | 647 | if (!ld || !o->registerfiles) |
5ffd5626 JA |
648 | return generic_close_file(td, f); |
649 | ||
650 | f->fd = -1; | |
651 | return 0; | |
652 | } | |
653 | ||
52885fa2 | 654 | static struct ioengine_ops ioengine = { |
bffad86f | 655 | .name = "io_uring", |
52885fa2 | 656 | .version = FIO_IOOPS_VERSION, |
04ba61df | 657 | .flags = FIO_ASYNCIO_SYNC_TRIM, |
bffad86f JA |
658 | .init = fio_ioring_init, |
659 | .post_init = fio_ioring_post_init, | |
660 | .io_u_init = fio_ioring_io_u_init, | |
661 | .prep = fio_ioring_prep, | |
662 | .queue = fio_ioring_queue, | |
663 | .commit = fio_ioring_commit, | |
664 | .getevents = fio_ioring_getevents, | |
665 | .event = fio_ioring_event, | |
666 | .cleanup = fio_ioring_cleanup, | |
5ffd5626 JA |
667 | .open_file = fio_ioring_open_file, |
668 | .close_file = fio_ioring_close_file, | |
52885fa2 JA |
669 | .get_file_size = generic_get_file_size, |
670 | .options = options, | |
bffad86f | 671 | .option_struct_size = sizeof(struct ioring_options), |
52885fa2 JA |
672 | }; |
673 | ||
bffad86f | 674 | static void fio_init fio_ioring_register(void) |
52885fa2 | 675 | { |
52885fa2 | 676 | register_ioengine(&ioengine); |
52885fa2 JA |
677 | } |
678 | ||
bffad86f | 679 | static void fio_exit fio_ioring_unregister(void) |
52885fa2 | 680 | { |
52885fa2 | 681 | unregister_ioengine(&ioengine); |
52885fa2 | 682 | } |
1f90e9bb | 683 | #endif |