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 JA |
64 | |
65 | uint64_t cachehit; | |
66 | uint64_t cachemiss; | |
9a2d78b3 | 67 | |
bffad86f | 68 | struct ioring_mmap mmap[3]; |
52885fa2 JA |
69 | }; |
70 | ||
bffad86f | 71 | struct ioring_options { |
52885fa2 JA |
72 | void *pad; |
73 | unsigned int hipri; | |
74 | unsigned int fixedbufs; | |
3d7d00a3 | 75 | unsigned int sqpoll_thread; |
2ea53ca3 JA |
76 | unsigned int sqpoll_set; |
77 | unsigned int sqpoll_cpu; | |
52885fa2 JA |
78 | }; |
79 | ||
2ea53ca3 | 80 | static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val) |
a90cd050 | 81 | { |
bffad86f | 82 | struct ioring_options *o = data; |
a90cd050 | 83 | |
2ea53ca3 JA |
84 | o->sqpoll_cpu = *val; |
85 | o->sqpoll_set = 1; | |
a90cd050 JA |
86 | return 0; |
87 | } | |
88 | ||
52885fa2 JA |
89 | static struct fio_option options[] = { |
90 | { | |
91 | .name = "hipri", | |
92 | .lname = "High Priority", | |
93 | .type = FIO_OPT_STR_SET, | |
bffad86f | 94 | .off1 = offsetof(struct ioring_options, hipri), |
52885fa2 JA |
95 | .help = "Use polled IO completions", |
96 | .category = FIO_OPT_C_ENGINE, | |
97 | .group = FIO_OPT_G_LIBAIO, | |
98 | }, | |
99 | { | |
100 | .name = "fixedbufs", | |
101 | .lname = "Fixed (pre-mapped) IO buffers", | |
102 | .type = FIO_OPT_STR_SET, | |
bffad86f | 103 | .off1 = offsetof(struct ioring_options, fixedbufs), |
52885fa2 JA |
104 | .help = "Pre map IO buffers", |
105 | .category = FIO_OPT_C_ENGINE, | |
106 | .group = FIO_OPT_G_LIBAIO, | |
107 | }, | |
771c9901 JA |
108 | { |
109 | .name = "sqthread_poll", | |
3d7d00a3 JA |
110 | .lname = "Kernel SQ thread polling", |
111 | .type = FIO_OPT_INT, | |
112 | .off1 = offsetof(struct ioring_options, sqpoll_thread), | |
113 | .help = "Offload submission/completion to kernel thread", | |
114 | .category = FIO_OPT_C_ENGINE, | |
115 | .group = FIO_OPT_G_LIBAIO, | |
116 | }, | |
117 | { | |
118 | .name = "sqthread_poll_cpu", | |
119 | .lname = "SQ Thread Poll CPU", | |
2ea53ca3 JA |
120 | .type = FIO_OPT_INT, |
121 | .cb = fio_ioring_sqpoll_cb, | |
3d7d00a3 | 122 | .help = "What CPU to run SQ thread polling on", |
a90cd050 JA |
123 | .category = FIO_OPT_C_ENGINE, |
124 | .group = FIO_OPT_G_LIBAIO, | |
125 | }, | |
52885fa2 JA |
126 | { |
127 | .name = NULL, | |
128 | }, | |
129 | }; | |
130 | ||
bffad86f | 131 | static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit, |
52885fa2 JA |
132 | unsigned int min_complete, unsigned int flags) |
133 | { | |
9a2d78b3 | 134 | return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit, |
521164fa | 135 | min_complete, flags, NULL, 0); |
52885fa2 JA |
136 | } |
137 | ||
bffad86f | 138 | static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 139 | { |
bffad86f | 140 | struct ioring_data *ld = td->io_ops_data; |
cfcc8564 | 141 | struct ioring_options *o = td->eo; |
52885fa2 | 142 | struct fio_file *f = io_u->file; |
f0403f94 | 143 | struct io_uring_sqe *sqe; |
52885fa2 | 144 | |
f0403f94 JA |
145 | sqe = &ld->sqes[io_u->index]; |
146 | sqe->fd = f->fd; | |
147 | sqe->flags = 0; | |
148 | sqe->ioprio = 0; | |
2ea53ca3 | 149 | sqe->buf_index = 0; |
52885fa2 | 150 | |
e3970057 | 151 | if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) { |
f0403f94 | 152 | if (o->fixedbufs) { |
48e698fa JA |
153 | if (io_u->ddir == DDIR_READ) |
154 | sqe->opcode = IORING_OP_READ_FIXED; | |
155 | else | |
156 | sqe->opcode = IORING_OP_WRITE_FIXED; | |
919850d2 | 157 | sqe->addr = (unsigned long) io_u->xfer_buf; |
f0403f94 | 158 | sqe->len = io_u->xfer_buflen; |
2ea53ca3 | 159 | sqe->buf_index = io_u->index; |
cfcc8564 | 160 | } else { |
48e698fa JA |
161 | if (io_u->ddir == DDIR_READ) |
162 | sqe->opcode = IORING_OP_READV; | |
163 | else | |
164 | sqe->opcode = IORING_OP_WRITEV; | |
919850d2 | 165 | sqe->addr = (unsigned long) &ld->iovecs[io_u->index]; |
f0403f94 | 166 | sqe->len = 1; |
cfcc8564 | 167 | } |
f0403f94 | 168 | sqe->off = io_u->offset; |
48e698fa JA |
169 | } else if (ddir_sync(io_u->ddir)) { |
170 | sqe->fsync_flags = 0; | |
171 | if (io_u->ddir == DDIR_DATASYNC) | |
172 | sqe->fsync_flags |= IORING_FSYNC_DATASYNC; | |
f0403f94 | 173 | sqe->opcode = IORING_OP_FSYNC; |
48e698fa | 174 | } |
52885fa2 | 175 | |
48e698fa | 176 | sqe->user_data = (unsigned long) io_u; |
52885fa2 JA |
177 | return 0; |
178 | } | |
179 | ||
bffad86f | 180 | static struct io_u *fio_ioring_event(struct thread_data *td, int event) |
52885fa2 | 181 | { |
bffad86f | 182 | struct ioring_data *ld = td->io_ops_data; |
f0403f94 | 183 | struct io_uring_cqe *cqe; |
52885fa2 | 184 | struct io_u *io_u; |
b87aa01a | 185 | unsigned index; |
52885fa2 | 186 | |
b87aa01a | 187 | index = (event + ld->cq_ring_off) & ld->cq_ring_mask; |
52885fa2 | 188 | |
f0403f94 | 189 | cqe = &ld->cq_ring.cqes[index]; |
e3466352 | 190 | io_u = (struct io_u *) (uintptr_t) cqe->user_data; |
52885fa2 | 191 | |
f0403f94 JA |
192 | if (cqe->res != io_u->xfer_buflen) { |
193 | if (cqe->res > io_u->xfer_buflen) | |
194 | io_u->error = -cqe->res; | |
52885fa2 | 195 | else |
f0403f94 | 196 | io_u->resid = io_u->xfer_buflen - cqe->res; |
52885fa2 JA |
197 | } else |
198 | io_u->error = 0; | |
199 | ||
96563db9 | 200 | if (io_u->ddir == DDIR_READ) { |
f0403f94 | 201 | if (cqe->flags & IOCQE_FLAG_CACHEHIT) |
96563db9 JA |
202 | ld->cachehit++; |
203 | else | |
204 | ld->cachemiss++; | |
205 | } | |
206 | ||
52885fa2 JA |
207 | return io_u; |
208 | } | |
209 | ||
bffad86f | 210 | static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events, |
52885fa2 JA |
211 | unsigned int max) |
212 | { | |
bffad86f JA |
213 | struct ioring_data *ld = td->io_ops_data; |
214 | struct io_cq_ring *ring = &ld->cq_ring; | |
e2239016 | 215 | unsigned head, reaped = 0; |
52885fa2 | 216 | |
9a2d78b3 | 217 | head = *ring->head; |
52885fa2 JA |
218 | do { |
219 | read_barrier(); | |
9a2d78b3 | 220 | if (head == *ring->tail) |
52885fa2 JA |
221 | break; |
222 | reaped++; | |
223 | head++; | |
52885fa2 JA |
224 | } while (reaped + events < max); |
225 | ||
9a2d78b3 | 226 | *ring->head = head; |
52885fa2 JA |
227 | write_barrier(); |
228 | return reaped; | |
229 | } | |
230 | ||
bffad86f JA |
231 | static int fio_ioring_getevents(struct thread_data *td, unsigned int min, |
232 | unsigned int max, const struct timespec *t) | |
52885fa2 | 233 | { |
bffad86f | 234 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 | 235 | unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min; |
bffad86f JA |
236 | struct ioring_options *o = td->eo; |
237 | struct io_cq_ring *ring = &ld->cq_ring; | |
b87aa01a JA |
238 | unsigned events = 0; |
239 | int r; | |
52885fa2 | 240 | |
9a2d78b3 | 241 | ld->cq_ring_off = *ring->head; |
52885fa2 | 242 | do { |
bffad86f | 243 | r = fio_ioring_cqring_reap(td, events, max); |
52885fa2 JA |
244 | if (r) { |
245 | events += r; | |
246 | continue; | |
247 | } | |
248 | ||
3d7d00a3 | 249 | if (!o->sqpoll_thread) { |
9a2d78b3 JA |
250 | r = io_uring_enter(ld, 0, actual_min, |
251 | IORING_ENTER_GETEVENTS); | |
771c9901 JA |
252 | if (r < 0) { |
253 | if (errno == EAGAIN) | |
254 | continue; | |
9a2d78b3 | 255 | td_verror(td, errno, "io_uring_enter"); |
771c9901 JA |
256 | break; |
257 | } | |
52885fa2 JA |
258 | } |
259 | } while (events < min); | |
260 | ||
261 | return r < 0 ? r : events; | |
262 | } | |
263 | ||
bffad86f JA |
264 | static enum fio_q_status fio_ioring_queue(struct thread_data *td, |
265 | struct io_u *io_u) | |
52885fa2 | 266 | { |
bffad86f JA |
267 | struct ioring_data *ld = td->io_ops_data; |
268 | struct io_sq_ring *ring = &ld->sq_ring; | |
52885fa2 JA |
269 | unsigned tail, next_tail; |
270 | ||
271 | fio_ro_check(td, io_u); | |
272 | ||
b87aa01a | 273 | if (ld->queued == ld->iodepth) |
52885fa2 JA |
274 | return FIO_Q_BUSY; |
275 | ||
52885fa2 JA |
276 | if (io_u->ddir == DDIR_TRIM) { |
277 | if (ld->queued) | |
278 | return FIO_Q_BUSY; | |
279 | ||
280 | do_io_u_trim(td, io_u); | |
281 | io_u_mark_submit(td, 1); | |
282 | io_u_mark_complete(td, 1); | |
283 | return FIO_Q_COMPLETED; | |
284 | } | |
285 | ||
9a2d78b3 | 286 | tail = *ring->tail; |
52885fa2 | 287 | next_tail = tail + 1; |
52885fa2 | 288 | read_barrier(); |
9a2d78b3 | 289 | if (next_tail == *ring->head) |
52885fa2 JA |
290 | return FIO_Q_BUSY; |
291 | ||
2b9415dd JA |
292 | /* ensure sqe stores are ordered with tail update */ |
293 | write_barrier(); | |
b87aa01a | 294 | ring->array[tail & ld->sq_ring_mask] = io_u->index; |
9a2d78b3 | 295 | *ring->tail = next_tail; |
52885fa2 JA |
296 | write_barrier(); |
297 | ||
298 | ld->queued++; | |
299 | return FIO_Q_QUEUED; | |
300 | } | |
301 | ||
bffad86f | 302 | static void fio_ioring_queued(struct thread_data *td, int start, int nr) |
52885fa2 | 303 | { |
bffad86f | 304 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
305 | struct timespec now; |
306 | ||
307 | if (!fio_fill_issue_time(td)) | |
308 | return; | |
309 | ||
310 | fio_gettime(&now, NULL); | |
311 | ||
312 | while (nr--) { | |
bffad86f | 313 | struct io_sq_ring *ring = &ld->sq_ring; |
9a2d78b3 | 314 | int index = ring->array[start & ld->sq_ring_mask]; |
f8289afc | 315 | struct io_u *io_u = ld->io_u_index[index]; |
52885fa2 JA |
316 | |
317 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
318 | io_u_queued(td, io_u); | |
319 | ||
320 | start++; | |
52885fa2 JA |
321 | } |
322 | } | |
323 | ||
bffad86f | 324 | static int fio_ioring_commit(struct thread_data *td) |
52885fa2 | 325 | { |
bffad86f JA |
326 | struct ioring_data *ld = td->io_ops_data; |
327 | struct ioring_options *o = td->eo; | |
52885fa2 JA |
328 | int ret; |
329 | ||
330 | if (!ld->queued) | |
331 | return 0; | |
332 | ||
3d7d00a3 JA |
333 | /* |
334 | * Kernel side does submission. just need to check if the ring is | |
335 | * flagged as needing a kick, if so, call io_uring_enter(). This | |
336 | * only happens if we've been idle too long. | |
337 | */ | |
338 | if (o->sqpoll_thread) { | |
bffad86f | 339 | struct io_sq_ring *ring = &ld->sq_ring; |
4cdbc048 | 340 | |
2ea53ca3 | 341 | read_barrier(); |
9a2d78b3 | 342 | if (*ring->flags & IORING_SQ_NEED_WAKEUP) |
b532dd6d JA |
343 | io_uring_enter(ld, ld->queued, 0, |
344 | IORING_ENTER_SQ_WAKEUP); | |
771c9901 JA |
345 | ld->queued = 0; |
346 | return 0; | |
347 | } | |
348 | ||
52885fa2 | 349 | do { |
9a2d78b3 | 350 | unsigned start = *ld->sq_ring.head; |
52885fa2 JA |
351 | long nr = ld->queued; |
352 | ||
9a2d78b3 | 353 | ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS); |
52885fa2 | 354 | if (ret > 0) { |
bffad86f | 355 | fio_ioring_queued(td, start, ret); |
52885fa2 JA |
356 | io_u_mark_submit(td, ret); |
357 | ||
358 | ld->queued -= ret; | |
359 | ret = 0; | |
a90cd050 JA |
360 | } else if (!ret) { |
361 | io_u_mark_submit(td, ret); | |
52885fa2 | 362 | continue; |
a90cd050 JA |
363 | } else { |
364 | if (errno == EAGAIN) { | |
bffad86f | 365 | ret = fio_ioring_cqring_reap(td, 0, ld->queued); |
a90cd050 JA |
366 | if (ret) |
367 | continue; | |
368 | /* Shouldn't happen */ | |
369 | usleep(1); | |
370 | continue; | |
52885fa2 | 371 | } |
9a2d78b3 | 372 | td_verror(td, errno, "io_uring_enter submit"); |
52885fa2 | 373 | break; |
a90cd050 | 374 | } |
52885fa2 JA |
375 | } while (ld->queued); |
376 | ||
377 | return ret; | |
378 | } | |
379 | ||
bffad86f | 380 | static void fio_ioring_unmap(struct ioring_data *ld) |
52885fa2 | 381 | { |
9a2d78b3 | 382 | int i; |
52885fa2 | 383 | |
9a2d78b3 JA |
384 | for (i = 0; i < ARRAY_SIZE(ld->mmap); i++) |
385 | munmap(ld->mmap[i].ptr, ld->mmap[i].len); | |
386 | close(ld->ring_fd); | |
b87aa01a JA |
387 | } |
388 | ||
bffad86f | 389 | static void fio_ioring_cleanup(struct thread_data *td) |
52885fa2 | 390 | { |
bffad86f | 391 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
392 | |
393 | if (ld) { | |
96563db9 JA |
394 | td->ts.cachehit += ld->cachehit; |
395 | td->ts.cachemiss += ld->cachemiss; | |
396 | ||
52885fa2 | 397 | if (!(td->flags & TD_F_CHILD)) |
bffad86f | 398 | fio_ioring_unmap(ld); |
9a2d78b3 | 399 | |
52885fa2 | 400 | free(ld->io_u_index); |
9a2d78b3 | 401 | free(ld->iovecs); |
52885fa2 JA |
402 | free(ld); |
403 | } | |
404 | } | |
405 | ||
bffad86f | 406 | static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p) |
9a2d78b3 | 407 | { |
bffad86f JA |
408 | struct io_sq_ring *sring = &ld->sq_ring; |
409 | struct io_cq_ring *cring = &ld->cq_ring; | |
9a2d78b3 JA |
410 | void *ptr; |
411 | ||
e2239016 | 412 | ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32); |
9a2d78b3 JA |
413 | ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE, |
414 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
415 | IORING_OFF_SQ_RING); | |
416 | ld->mmap[0].ptr = ptr; | |
417 | sring->head = ptr + p->sq_off.head; | |
418 | sring->tail = ptr + p->sq_off.tail; | |
419 | sring->ring_mask = ptr + p->sq_off.ring_mask; | |
420 | sring->ring_entries = ptr + p->sq_off.ring_entries; | |
421 | sring->flags = ptr + p->sq_off.flags; | |
ac122fea | 422 | sring->array = ptr + p->sq_off.array; |
9a2d78b3 JA |
423 | ld->sq_ring_mask = *sring->ring_mask; |
424 | ||
f0403f94 JA |
425 | ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe); |
426 | ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE, | |
9a2d78b3 | 427 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, |
f0403f94 JA |
428 | IORING_OFF_SQES); |
429 | ld->mmap[1].ptr = ld->sqes; | |
9a2d78b3 | 430 | |
f0403f94 JA |
431 | ld->mmap[2].len = p->cq_off.cqes + |
432 | p->cq_entries * sizeof(struct io_uring_cqe); | |
9a2d78b3 JA |
433 | ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE, |
434 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
435 | IORING_OFF_CQ_RING); | |
436 | ld->mmap[2].ptr = ptr; | |
437 | cring->head = ptr + p->cq_off.head; | |
438 | cring->tail = ptr + p->cq_off.tail; | |
439 | cring->ring_mask = ptr + p->cq_off.ring_mask; | |
440 | cring->ring_entries = ptr + p->cq_off.ring_entries; | |
f0403f94 | 441 | cring->cqes = ptr + p->cq_off.cqes; |
9a2d78b3 JA |
442 | ld->cq_ring_mask = *cring->ring_mask; |
443 | return 0; | |
444 | } | |
445 | ||
bffad86f | 446 | static int fio_ioring_queue_init(struct thread_data *td) |
52885fa2 | 447 | { |
bffad86f JA |
448 | struct ioring_data *ld = td->io_ops_data; |
449 | struct ioring_options *o = td->eo; | |
52885fa2 | 450 | int depth = td->o.iodepth; |
bffad86f | 451 | struct io_uring_params p; |
9a2d78b3 JA |
452 | int ret; |
453 | ||
454 | memset(&p, 0, sizeof(p)); | |
52885fa2 JA |
455 | |
456 | if (o->hipri) | |
bffad86f | 457 | p.flags |= IORING_SETUP_IOPOLL; |
3d7d00a3 JA |
458 | if (o->sqpoll_thread) { |
459 | p.flags |= IORING_SETUP_SQPOLL; | |
460 | if (o->sqpoll_set) { | |
461 | p.flags |= IORING_SETUP_SQ_AFF; | |
462 | p.sq_thread_cpu = o->sqpoll_cpu; | |
463 | } | |
f635f1fb | 464 | } |
a90cd050 | 465 | |
2ea53ca3 | 466 | ret = syscall(__NR_sys_io_uring_setup, depth, &p); |
9a2d78b3 JA |
467 | if (ret < 0) |
468 | return ret; | |
469 | ||
470 | ld->ring_fd = ret; | |
2ea53ca3 JA |
471 | |
472 | if (o->fixedbufs) { | |
2d644205 JA |
473 | struct rlimit rlim = { |
474 | .rlim_cur = RLIM_INFINITY, | |
475 | .rlim_max = RLIM_INFINITY, | |
476 | }; | |
477 | ||
478 | if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) | |
479 | return -1; | |
480 | ||
2ea53ca3 | 481 | ret = syscall(__NR_sys_io_uring_register, ld->ring_fd, |
919850d2 | 482 | IORING_REGISTER_BUFFERS, ld->iovecs, depth); |
2ea53ca3 JA |
483 | if (ret < 0) |
484 | return ret; | |
485 | } | |
486 | ||
bffad86f | 487 | return fio_ioring_mmap(ld, &p); |
52885fa2 JA |
488 | } |
489 | ||
bffad86f | 490 | static int fio_ioring_post_init(struct thread_data *td) |
52885fa2 | 491 | { |
bffad86f | 492 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 | 493 | struct io_u *io_u; |
650346e1 | 494 | int err, i; |
52885fa2 | 495 | |
650346e1 JA |
496 | for (i = 0; i < td->o.iodepth; i++) { |
497 | struct iovec *iov = &ld->iovecs[i]; | |
9a2d78b3 | 498 | |
650346e1 JA |
499 | io_u = ld->io_u_index[i]; |
500 | iov->iov_base = io_u->buf; | |
501 | iov->iov_len = td_max_bs(td); | |
52885fa2 JA |
502 | } |
503 | ||
bffad86f | 504 | err = fio_ioring_queue_init(td); |
52885fa2 | 505 | if (err) { |
d63a472d | 506 | td_verror(td, errno, "io_queue_init"); |
52885fa2 JA |
507 | return 1; |
508 | } | |
509 | ||
510 | return 0; | |
511 | } | |
512 | ||
9a2d78b3 JA |
513 | static unsigned roundup_pow2(unsigned depth) |
514 | { | |
515 | return 1UL << __fls(depth - 1); | |
516 | } | |
517 | ||
bffad86f | 518 | static int fio_ioring_init(struct thread_data *td) |
52885fa2 | 519 | { |
bffad86f | 520 | struct ioring_data *ld; |
52885fa2 | 521 | |
52885fa2 JA |
522 | ld = calloc(1, sizeof(*ld)); |
523 | ||
b87aa01a JA |
524 | /* ring depth must be a power-of-2 */ |
525 | ld->iodepth = td->o.iodepth; | |
526 | td->o.iodepth = roundup_pow2(td->o.iodepth); | |
527 | ||
52885fa2 JA |
528 | /* io_u index */ |
529 | ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *)); | |
650346e1 | 530 | ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec)); |
52885fa2 JA |
531 | |
532 | td->io_ops_data = ld; | |
533 | return 0; | |
534 | } | |
535 | ||
bffad86f | 536 | static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 537 | { |
bffad86f | 538 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
539 | |
540 | ld->io_u_index[io_u->index] = io_u; | |
541 | return 0; | |
542 | } | |
543 | ||
544 | static struct ioengine_ops ioengine = { | |
bffad86f | 545 | .name = "io_uring", |
52885fa2 | 546 | .version = FIO_IOOPS_VERSION, |
bffad86f JA |
547 | .init = fio_ioring_init, |
548 | .post_init = fio_ioring_post_init, | |
549 | .io_u_init = fio_ioring_io_u_init, | |
550 | .prep = fio_ioring_prep, | |
551 | .queue = fio_ioring_queue, | |
552 | .commit = fio_ioring_commit, | |
553 | .getevents = fio_ioring_getevents, | |
554 | .event = fio_ioring_event, | |
555 | .cleanup = fio_ioring_cleanup, | |
52885fa2 JA |
556 | .open_file = generic_open_file, |
557 | .close_file = generic_close_file, | |
558 | .get_file_size = generic_get_file_size, | |
559 | .options = options, | |
bffad86f | 560 | .option_struct_size = sizeof(struct ioring_options), |
52885fa2 JA |
561 | }; |
562 | ||
bffad86f | 563 | static void fio_init fio_ioring_register(void) |
52885fa2 | 564 | { |
52885fa2 | 565 | register_ioengine(&ioengine); |
52885fa2 JA |
566 | } |
567 | ||
bffad86f | 568 | static void fio_exit fio_ioring_unregister(void) |
52885fa2 | 569 | { |
52885fa2 | 570 | unregister_ioengine(&ioengine); |
52885fa2 | 571 | } |
1f90e9bb | 572 | #endif |