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 | ||
bffad86f | 53 | struct io_sq_ring sq_ring; |
f0403f94 | 54 | struct io_uring_sqe *sqes; |
9a2d78b3 | 55 | struct iovec *iovecs; |
b87aa01a | 56 | unsigned sq_ring_mask; |
52885fa2 | 57 | |
bffad86f | 58 | struct io_cq_ring cq_ring; |
b87aa01a | 59 | unsigned cq_ring_mask; |
52885fa2 JA |
60 | |
61 | int queued; | |
62 | int cq_ring_off; | |
b87aa01a | 63 | unsigned iodepth; |
96563db9 | 64 | |
bffad86f | 65 | struct ioring_mmap mmap[3]; |
52885fa2 JA |
66 | }; |
67 | ||
bffad86f | 68 | struct ioring_options { |
52885fa2 JA |
69 | void *pad; |
70 | unsigned int hipri; | |
71 | unsigned int fixedbufs; | |
3d7d00a3 | 72 | unsigned int sqpoll_thread; |
2ea53ca3 JA |
73 | unsigned int sqpoll_set; |
74 | unsigned int sqpoll_cpu; | |
52885fa2 JA |
75 | }; |
76 | ||
2ea53ca3 | 77 | static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val) |
a90cd050 | 78 | { |
bffad86f | 79 | struct ioring_options *o = data; |
a90cd050 | 80 | |
2ea53ca3 JA |
81 | o->sqpoll_cpu = *val; |
82 | o->sqpoll_set = 1; | |
a90cd050 JA |
83 | return 0; |
84 | } | |
85 | ||
52885fa2 JA |
86 | static struct fio_option options[] = { |
87 | { | |
88 | .name = "hipri", | |
89 | .lname = "High Priority", | |
90 | .type = FIO_OPT_STR_SET, | |
bffad86f | 91 | .off1 = offsetof(struct ioring_options, hipri), |
52885fa2 JA |
92 | .help = "Use polled IO completions", |
93 | .category = FIO_OPT_C_ENGINE, | |
94 | .group = FIO_OPT_G_LIBAIO, | |
95 | }, | |
96 | { | |
97 | .name = "fixedbufs", | |
98 | .lname = "Fixed (pre-mapped) IO buffers", | |
99 | .type = FIO_OPT_STR_SET, | |
bffad86f | 100 | .off1 = offsetof(struct ioring_options, fixedbufs), |
52885fa2 JA |
101 | .help = "Pre map IO buffers", |
102 | .category = FIO_OPT_C_ENGINE, | |
103 | .group = FIO_OPT_G_LIBAIO, | |
104 | }, | |
771c9901 JA |
105 | { |
106 | .name = "sqthread_poll", | |
3d7d00a3 JA |
107 | .lname = "Kernel SQ thread polling", |
108 | .type = FIO_OPT_INT, | |
109 | .off1 = offsetof(struct ioring_options, sqpoll_thread), | |
110 | .help = "Offload submission/completion to kernel thread", | |
111 | .category = FIO_OPT_C_ENGINE, | |
112 | .group = FIO_OPT_G_LIBAIO, | |
113 | }, | |
114 | { | |
115 | .name = "sqthread_poll_cpu", | |
116 | .lname = "SQ Thread Poll CPU", | |
2ea53ca3 JA |
117 | .type = FIO_OPT_INT, |
118 | .cb = fio_ioring_sqpoll_cb, | |
3d7d00a3 | 119 | .help = "What CPU to run SQ thread polling on", |
a90cd050 JA |
120 | .category = FIO_OPT_C_ENGINE, |
121 | .group = FIO_OPT_G_LIBAIO, | |
122 | }, | |
52885fa2 JA |
123 | { |
124 | .name = NULL, | |
125 | }, | |
126 | }; | |
127 | ||
bffad86f | 128 | static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit, |
52885fa2 JA |
129 | unsigned int min_complete, unsigned int flags) |
130 | { | |
9a2d78b3 | 131 | return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit, |
521164fa | 132 | min_complete, flags, NULL, 0); |
52885fa2 JA |
133 | } |
134 | ||
bffad86f | 135 | static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 136 | { |
bffad86f | 137 | struct ioring_data *ld = td->io_ops_data; |
cfcc8564 | 138 | struct ioring_options *o = td->eo; |
52885fa2 | 139 | struct fio_file *f = io_u->file; |
f0403f94 | 140 | struct io_uring_sqe *sqe; |
52885fa2 | 141 | |
f0403f94 JA |
142 | sqe = &ld->sqes[io_u->index]; |
143 | sqe->fd = f->fd; | |
144 | sqe->flags = 0; | |
145 | sqe->ioprio = 0; | |
2ea53ca3 | 146 | sqe->buf_index = 0; |
52885fa2 | 147 | |
e3970057 | 148 | if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) { |
f0403f94 | 149 | if (o->fixedbufs) { |
48e698fa JA |
150 | if (io_u->ddir == DDIR_READ) |
151 | sqe->opcode = IORING_OP_READ_FIXED; | |
152 | else | |
153 | sqe->opcode = IORING_OP_WRITE_FIXED; | |
919850d2 | 154 | sqe->addr = (unsigned long) io_u->xfer_buf; |
f0403f94 | 155 | sqe->len = io_u->xfer_buflen; |
2ea53ca3 | 156 | sqe->buf_index = io_u->index; |
cfcc8564 | 157 | } else { |
48e698fa JA |
158 | if (io_u->ddir == DDIR_READ) |
159 | sqe->opcode = IORING_OP_READV; | |
160 | else | |
161 | sqe->opcode = IORING_OP_WRITEV; | |
919850d2 | 162 | sqe->addr = (unsigned long) &ld->iovecs[io_u->index]; |
f0403f94 | 163 | sqe->len = 1; |
cfcc8564 | 164 | } |
f0403f94 | 165 | sqe->off = io_u->offset; |
48e698fa JA |
166 | } else if (ddir_sync(io_u->ddir)) { |
167 | sqe->fsync_flags = 0; | |
168 | if (io_u->ddir == DDIR_DATASYNC) | |
169 | sqe->fsync_flags |= IORING_FSYNC_DATASYNC; | |
f0403f94 | 170 | sqe->opcode = IORING_OP_FSYNC; |
48e698fa | 171 | } |
52885fa2 | 172 | |
48e698fa | 173 | sqe->user_data = (unsigned long) io_u; |
52885fa2 JA |
174 | return 0; |
175 | } | |
176 | ||
bffad86f | 177 | static struct io_u *fio_ioring_event(struct thread_data *td, int event) |
52885fa2 | 178 | { |
bffad86f | 179 | struct ioring_data *ld = td->io_ops_data; |
f0403f94 | 180 | struct io_uring_cqe *cqe; |
52885fa2 | 181 | struct io_u *io_u; |
b87aa01a | 182 | unsigned index; |
52885fa2 | 183 | |
b87aa01a | 184 | index = (event + ld->cq_ring_off) & ld->cq_ring_mask; |
52885fa2 | 185 | |
f0403f94 | 186 | cqe = &ld->cq_ring.cqes[index]; |
e3466352 | 187 | io_u = (struct io_u *) (uintptr_t) cqe->user_data; |
52885fa2 | 188 | |
f0403f94 JA |
189 | if (cqe->res != io_u->xfer_buflen) { |
190 | if (cqe->res > io_u->xfer_buflen) | |
191 | io_u->error = -cqe->res; | |
52885fa2 | 192 | else |
f0403f94 | 193 | io_u->resid = io_u->xfer_buflen - cqe->res; |
52885fa2 JA |
194 | } else |
195 | io_u->error = 0; | |
196 | ||
197 | return io_u; | |
198 | } | |
199 | ||
bffad86f | 200 | static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events, |
52885fa2 JA |
201 | unsigned int max) |
202 | { | |
bffad86f JA |
203 | struct ioring_data *ld = td->io_ops_data; |
204 | struct io_cq_ring *ring = &ld->cq_ring; | |
e2239016 | 205 | unsigned head, reaped = 0; |
52885fa2 | 206 | |
9a2d78b3 | 207 | head = *ring->head; |
52885fa2 JA |
208 | do { |
209 | read_barrier(); | |
9a2d78b3 | 210 | if (head == *ring->tail) |
52885fa2 JA |
211 | break; |
212 | reaped++; | |
213 | head++; | |
52885fa2 JA |
214 | } while (reaped + events < max); |
215 | ||
9a2d78b3 | 216 | *ring->head = head; |
52885fa2 JA |
217 | write_barrier(); |
218 | return reaped; | |
219 | } | |
220 | ||
bffad86f JA |
221 | static int fio_ioring_getevents(struct thread_data *td, unsigned int min, |
222 | unsigned int max, const struct timespec *t) | |
52885fa2 | 223 | { |
bffad86f | 224 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 | 225 | unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min; |
bffad86f JA |
226 | struct ioring_options *o = td->eo; |
227 | struct io_cq_ring *ring = &ld->cq_ring; | |
b87aa01a JA |
228 | unsigned events = 0; |
229 | int r; | |
52885fa2 | 230 | |
9a2d78b3 | 231 | ld->cq_ring_off = *ring->head; |
52885fa2 | 232 | do { |
bffad86f | 233 | r = fio_ioring_cqring_reap(td, events, max); |
52885fa2 JA |
234 | if (r) { |
235 | events += r; | |
f7cbbbf8 ST |
236 | if (actual_min != 0) |
237 | actual_min -= r; | |
52885fa2 JA |
238 | continue; |
239 | } | |
240 | ||
3d7d00a3 | 241 | if (!o->sqpoll_thread) { |
9a2d78b3 JA |
242 | r = io_uring_enter(ld, 0, actual_min, |
243 | IORING_ENTER_GETEVENTS); | |
771c9901 JA |
244 | if (r < 0) { |
245 | if (errno == EAGAIN) | |
246 | continue; | |
9a2d78b3 | 247 | td_verror(td, errno, "io_uring_enter"); |
771c9901 JA |
248 | break; |
249 | } | |
52885fa2 JA |
250 | } |
251 | } while (events < min); | |
252 | ||
253 | return r < 0 ? r : events; | |
254 | } | |
255 | ||
bffad86f JA |
256 | static enum fio_q_status fio_ioring_queue(struct thread_data *td, |
257 | struct io_u *io_u) | |
52885fa2 | 258 | { |
bffad86f JA |
259 | struct ioring_data *ld = td->io_ops_data; |
260 | struct io_sq_ring *ring = &ld->sq_ring; | |
52885fa2 JA |
261 | unsigned tail, next_tail; |
262 | ||
263 | fio_ro_check(td, io_u); | |
264 | ||
b87aa01a | 265 | if (ld->queued == ld->iodepth) |
52885fa2 JA |
266 | return FIO_Q_BUSY; |
267 | ||
52885fa2 JA |
268 | if (io_u->ddir == DDIR_TRIM) { |
269 | if (ld->queued) | |
270 | return FIO_Q_BUSY; | |
271 | ||
272 | do_io_u_trim(td, io_u); | |
273 | io_u_mark_submit(td, 1); | |
274 | io_u_mark_complete(td, 1); | |
275 | return FIO_Q_COMPLETED; | |
276 | } | |
277 | ||
9a2d78b3 | 278 | tail = *ring->tail; |
52885fa2 | 279 | next_tail = tail + 1; |
52885fa2 | 280 | read_barrier(); |
9a2d78b3 | 281 | if (next_tail == *ring->head) |
52885fa2 JA |
282 | return FIO_Q_BUSY; |
283 | ||
2b9415dd JA |
284 | /* ensure sqe stores are ordered with tail update */ |
285 | write_barrier(); | |
b87aa01a | 286 | ring->array[tail & ld->sq_ring_mask] = io_u->index; |
9a2d78b3 | 287 | *ring->tail = next_tail; |
52885fa2 JA |
288 | write_barrier(); |
289 | ||
290 | ld->queued++; | |
291 | return FIO_Q_QUEUED; | |
292 | } | |
293 | ||
bffad86f | 294 | static void fio_ioring_queued(struct thread_data *td, int start, int nr) |
52885fa2 | 295 | { |
bffad86f | 296 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
297 | struct timespec now; |
298 | ||
299 | if (!fio_fill_issue_time(td)) | |
300 | return; | |
301 | ||
302 | fio_gettime(&now, NULL); | |
303 | ||
304 | while (nr--) { | |
bffad86f | 305 | struct io_sq_ring *ring = &ld->sq_ring; |
9a2d78b3 | 306 | int index = ring->array[start & ld->sq_ring_mask]; |
f8289afc | 307 | struct io_u *io_u = ld->io_u_index[index]; |
52885fa2 JA |
308 | |
309 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
310 | io_u_queued(td, io_u); | |
311 | ||
312 | start++; | |
52885fa2 JA |
313 | } |
314 | } | |
315 | ||
bffad86f | 316 | static int fio_ioring_commit(struct thread_data *td) |
52885fa2 | 317 | { |
bffad86f JA |
318 | struct ioring_data *ld = td->io_ops_data; |
319 | struct ioring_options *o = td->eo; | |
52885fa2 JA |
320 | int ret; |
321 | ||
322 | if (!ld->queued) | |
323 | return 0; | |
324 | ||
3d7d00a3 JA |
325 | /* |
326 | * Kernel side does submission. just need to check if the ring is | |
327 | * flagged as needing a kick, if so, call io_uring_enter(). This | |
328 | * only happens if we've been idle too long. | |
329 | */ | |
330 | if (o->sqpoll_thread) { | |
bffad86f | 331 | struct io_sq_ring *ring = &ld->sq_ring; |
4cdbc048 | 332 | |
2ea53ca3 | 333 | read_barrier(); |
9a2d78b3 | 334 | if (*ring->flags & IORING_SQ_NEED_WAKEUP) |
b532dd6d JA |
335 | io_uring_enter(ld, ld->queued, 0, |
336 | IORING_ENTER_SQ_WAKEUP); | |
771c9901 JA |
337 | ld->queued = 0; |
338 | return 0; | |
339 | } | |
340 | ||
52885fa2 | 341 | do { |
9a2d78b3 | 342 | unsigned start = *ld->sq_ring.head; |
52885fa2 JA |
343 | long nr = ld->queued; |
344 | ||
9a2d78b3 | 345 | ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS); |
52885fa2 | 346 | if (ret > 0) { |
bffad86f | 347 | fio_ioring_queued(td, start, ret); |
52885fa2 JA |
348 | io_u_mark_submit(td, ret); |
349 | ||
350 | ld->queued -= ret; | |
351 | ret = 0; | |
a90cd050 JA |
352 | } else if (!ret) { |
353 | io_u_mark_submit(td, ret); | |
52885fa2 | 354 | continue; |
a90cd050 JA |
355 | } else { |
356 | if (errno == EAGAIN) { | |
bffad86f | 357 | ret = fio_ioring_cqring_reap(td, 0, ld->queued); |
a90cd050 JA |
358 | if (ret) |
359 | continue; | |
360 | /* Shouldn't happen */ | |
361 | usleep(1); | |
362 | continue; | |
52885fa2 | 363 | } |
9a2d78b3 | 364 | td_verror(td, errno, "io_uring_enter submit"); |
52885fa2 | 365 | break; |
a90cd050 | 366 | } |
52885fa2 JA |
367 | } while (ld->queued); |
368 | ||
369 | return ret; | |
370 | } | |
371 | ||
bffad86f | 372 | static void fio_ioring_unmap(struct ioring_data *ld) |
52885fa2 | 373 | { |
9a2d78b3 | 374 | int i; |
52885fa2 | 375 | |
9a2d78b3 JA |
376 | for (i = 0; i < ARRAY_SIZE(ld->mmap); i++) |
377 | munmap(ld->mmap[i].ptr, ld->mmap[i].len); | |
378 | close(ld->ring_fd); | |
b87aa01a JA |
379 | } |
380 | ||
bffad86f | 381 | static void fio_ioring_cleanup(struct thread_data *td) |
52885fa2 | 382 | { |
bffad86f | 383 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
384 | |
385 | if (ld) { | |
52885fa2 | 386 | if (!(td->flags & TD_F_CHILD)) |
bffad86f | 387 | fio_ioring_unmap(ld); |
9a2d78b3 | 388 | |
52885fa2 | 389 | free(ld->io_u_index); |
9a2d78b3 | 390 | free(ld->iovecs); |
52885fa2 JA |
391 | free(ld); |
392 | } | |
393 | } | |
394 | ||
bffad86f | 395 | static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p) |
9a2d78b3 | 396 | { |
bffad86f JA |
397 | struct io_sq_ring *sring = &ld->sq_ring; |
398 | struct io_cq_ring *cring = &ld->cq_ring; | |
9a2d78b3 JA |
399 | void *ptr; |
400 | ||
e2239016 | 401 | ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32); |
9a2d78b3 JA |
402 | ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE, |
403 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
404 | IORING_OFF_SQ_RING); | |
405 | ld->mmap[0].ptr = ptr; | |
406 | sring->head = ptr + p->sq_off.head; | |
407 | sring->tail = ptr + p->sq_off.tail; | |
408 | sring->ring_mask = ptr + p->sq_off.ring_mask; | |
409 | sring->ring_entries = ptr + p->sq_off.ring_entries; | |
410 | sring->flags = ptr + p->sq_off.flags; | |
ac122fea | 411 | sring->array = ptr + p->sq_off.array; |
9a2d78b3 JA |
412 | ld->sq_ring_mask = *sring->ring_mask; |
413 | ||
f0403f94 JA |
414 | ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe); |
415 | ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE, | |
9a2d78b3 | 416 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, |
f0403f94 JA |
417 | IORING_OFF_SQES); |
418 | ld->mmap[1].ptr = ld->sqes; | |
9a2d78b3 | 419 | |
f0403f94 JA |
420 | ld->mmap[2].len = p->cq_off.cqes + |
421 | p->cq_entries * sizeof(struct io_uring_cqe); | |
9a2d78b3 JA |
422 | ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE, |
423 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
424 | IORING_OFF_CQ_RING); | |
425 | ld->mmap[2].ptr = ptr; | |
426 | cring->head = ptr + p->cq_off.head; | |
427 | cring->tail = ptr + p->cq_off.tail; | |
428 | cring->ring_mask = ptr + p->cq_off.ring_mask; | |
429 | cring->ring_entries = ptr + p->cq_off.ring_entries; | |
f0403f94 | 430 | cring->cqes = ptr + p->cq_off.cqes; |
9a2d78b3 JA |
431 | ld->cq_ring_mask = *cring->ring_mask; |
432 | return 0; | |
433 | } | |
434 | ||
bffad86f | 435 | static int fio_ioring_queue_init(struct thread_data *td) |
52885fa2 | 436 | { |
bffad86f JA |
437 | struct ioring_data *ld = td->io_ops_data; |
438 | struct ioring_options *o = td->eo; | |
52885fa2 | 439 | int depth = td->o.iodepth; |
bffad86f | 440 | struct io_uring_params p; |
9a2d78b3 JA |
441 | int ret; |
442 | ||
443 | memset(&p, 0, sizeof(p)); | |
52885fa2 JA |
444 | |
445 | if (o->hipri) | |
bffad86f | 446 | p.flags |= IORING_SETUP_IOPOLL; |
3d7d00a3 JA |
447 | if (o->sqpoll_thread) { |
448 | p.flags |= IORING_SETUP_SQPOLL; | |
449 | if (o->sqpoll_set) { | |
450 | p.flags |= IORING_SETUP_SQ_AFF; | |
451 | p.sq_thread_cpu = o->sqpoll_cpu; | |
452 | } | |
f635f1fb | 453 | } |
a90cd050 | 454 | |
2ea53ca3 | 455 | ret = syscall(__NR_sys_io_uring_setup, depth, &p); |
9a2d78b3 JA |
456 | if (ret < 0) |
457 | return ret; | |
458 | ||
459 | ld->ring_fd = ret; | |
2ea53ca3 JA |
460 | |
461 | if (o->fixedbufs) { | |
2d644205 JA |
462 | struct rlimit rlim = { |
463 | .rlim_cur = RLIM_INFINITY, | |
464 | .rlim_max = RLIM_INFINITY, | |
465 | }; | |
466 | ||
467 | if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) | |
468 | return -1; | |
469 | ||
2ea53ca3 | 470 | ret = syscall(__NR_sys_io_uring_register, ld->ring_fd, |
919850d2 | 471 | IORING_REGISTER_BUFFERS, ld->iovecs, depth); |
2ea53ca3 JA |
472 | if (ret < 0) |
473 | return ret; | |
474 | } | |
475 | ||
bffad86f | 476 | return fio_ioring_mmap(ld, &p); |
52885fa2 JA |
477 | } |
478 | ||
bffad86f | 479 | static int fio_ioring_post_init(struct thread_data *td) |
52885fa2 | 480 | { |
bffad86f | 481 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 | 482 | struct io_u *io_u; |
650346e1 | 483 | int err, i; |
52885fa2 | 484 | |
650346e1 JA |
485 | for (i = 0; i < td->o.iodepth; i++) { |
486 | struct iovec *iov = &ld->iovecs[i]; | |
9a2d78b3 | 487 | |
650346e1 JA |
488 | io_u = ld->io_u_index[i]; |
489 | iov->iov_base = io_u->buf; | |
490 | iov->iov_len = td_max_bs(td); | |
52885fa2 JA |
491 | } |
492 | ||
bffad86f | 493 | err = fio_ioring_queue_init(td); |
52885fa2 | 494 | if (err) { |
d63a472d | 495 | td_verror(td, errno, "io_queue_init"); |
52885fa2 JA |
496 | return 1; |
497 | } | |
498 | ||
499 | return 0; | |
500 | } | |
501 | ||
9a2d78b3 JA |
502 | static unsigned roundup_pow2(unsigned depth) |
503 | { | |
504 | return 1UL << __fls(depth - 1); | |
505 | } | |
506 | ||
bffad86f | 507 | static int fio_ioring_init(struct thread_data *td) |
52885fa2 | 508 | { |
bffad86f | 509 | struct ioring_data *ld; |
52885fa2 | 510 | |
52885fa2 JA |
511 | ld = calloc(1, sizeof(*ld)); |
512 | ||
b87aa01a JA |
513 | /* ring depth must be a power-of-2 */ |
514 | ld->iodepth = td->o.iodepth; | |
515 | td->o.iodepth = roundup_pow2(td->o.iodepth); | |
516 | ||
52885fa2 JA |
517 | /* io_u index */ |
518 | ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *)); | |
650346e1 | 519 | ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec)); |
52885fa2 JA |
520 | |
521 | td->io_ops_data = ld; | |
522 | return 0; | |
523 | } | |
524 | ||
bffad86f | 525 | static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 526 | { |
bffad86f | 527 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
528 | |
529 | ld->io_u_index[io_u->index] = io_u; | |
530 | return 0; | |
531 | } | |
532 | ||
533 | static struct ioengine_ops ioengine = { | |
bffad86f | 534 | .name = "io_uring", |
52885fa2 | 535 | .version = FIO_IOOPS_VERSION, |
04ba61df | 536 | .flags = FIO_ASYNCIO_SYNC_TRIM, |
bffad86f JA |
537 | .init = fio_ioring_init, |
538 | .post_init = fio_ioring_post_init, | |
539 | .io_u_init = fio_ioring_io_u_init, | |
540 | .prep = fio_ioring_prep, | |
541 | .queue = fio_ioring_queue, | |
542 | .commit = fio_ioring_commit, | |
543 | .getevents = fio_ioring_getevents, | |
544 | .event = fio_ioring_event, | |
545 | .cleanup = fio_ioring_cleanup, | |
52885fa2 JA |
546 | .open_file = generic_open_file, |
547 | .close_file = generic_close_file, | |
548 | .get_file_size = generic_get_file_size, | |
549 | .options = options, | |
bffad86f | 550 | .option_struct_size = sizeof(struct ioring_options), |
52885fa2 JA |
551 | }; |
552 | ||
bffad86f | 553 | static void fio_init fio_ioring_register(void) |
52885fa2 | 554 | { |
52885fa2 | 555 | register_ioengine(&ioengine); |
52885fa2 JA |
556 | } |
557 | ||
bffad86f | 558 | static void fio_exit fio_ioring_unregister(void) |
52885fa2 | 559 | { |
52885fa2 | 560 | unregister_ioengine(&ioengine); |
52885fa2 | 561 | } |
1f90e9bb | 562 | #endif |