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" |
6d975f2c | 20 | #include "../lib/roundup.h" |
52885fa2 | 21 | |
bffad86f | 22 | #ifdef ARCH_HAVE_IOURING |
52885fa2 | 23 | |
57fa61f0 | 24 | #include "../lib/types.h" |
f3e769a4 | 25 | #include "../os/linux/io_uring.h" |
e9f6567a | 26 | #include "cmdprio.h" |
9a2d78b3 | 27 | |
bffad86f | 28 | struct io_sq_ring { |
e2239016 JA |
29 | unsigned *head; |
30 | unsigned *tail; | |
31 | unsigned *ring_mask; | |
32 | unsigned *ring_entries; | |
33 | unsigned *flags; | |
34 | unsigned *array; | |
52885fa2 JA |
35 | }; |
36 | ||
bffad86f | 37 | struct io_cq_ring { |
e2239016 JA |
38 | unsigned *head; |
39 | unsigned *tail; | |
40 | unsigned *ring_mask; | |
41 | unsigned *ring_entries; | |
f0403f94 | 42 | struct io_uring_cqe *cqes; |
9a2d78b3 JA |
43 | }; |
44 | ||
bffad86f | 45 | struct ioring_mmap { |
9a2d78b3 JA |
46 | void *ptr; |
47 | size_t len; | |
52885fa2 JA |
48 | }; |
49 | ||
bffad86f | 50 | struct ioring_data { |
9a2d78b3 JA |
51 | int ring_fd; |
52 | ||
52885fa2 JA |
53 | struct io_u **io_u_index; |
54 | ||
5ffd5626 JA |
55 | int *fds; |
56 | ||
bffad86f | 57 | struct io_sq_ring sq_ring; |
f0403f94 | 58 | struct io_uring_sqe *sqes; |
9a2d78b3 | 59 | struct iovec *iovecs; |
b87aa01a | 60 | unsigned sq_ring_mask; |
52885fa2 | 61 | |
bffad86f | 62 | struct io_cq_ring cq_ring; |
b87aa01a | 63 | unsigned cq_ring_mask; |
52885fa2 JA |
64 | |
65 | int queued; | |
66 | int cq_ring_off; | |
b87aa01a | 67 | unsigned iodepth; |
5a59a81d | 68 | int prepped; |
96563db9 | 69 | |
bffad86f | 70 | struct ioring_mmap mmap[3]; |
d6cbeab4 NC |
71 | |
72 | struct cmdprio cmdprio; | |
52885fa2 JA |
73 | }; |
74 | ||
bffad86f | 75 | struct ioring_options { |
a48f0cc7 | 76 | struct thread_data *td; |
52885fa2 | 77 | unsigned int hipri; |
d6cbeab4 | 78 | struct cmdprio_options cmdprio_options; |
52885fa2 | 79 | unsigned int fixedbufs; |
5ffd5626 | 80 | unsigned int registerfiles; |
3d7d00a3 | 81 | unsigned int sqpoll_thread; |
2ea53ca3 JA |
82 | unsigned int sqpoll_set; |
83 | unsigned int sqpoll_cpu; | |
b10b1e70 | 84 | unsigned int nonvectored; |
4a87b584 | 85 | unsigned int uncached; |
7d42e66e | 86 | unsigned int nowait; |
5a59a81d | 87 | unsigned int force_async; |
52885fa2 JA |
88 | }; |
89 | ||
b10b1e70 JA |
90 | static const int ddir_to_op[2][2] = { |
91 | { IORING_OP_READV, IORING_OP_READ }, | |
92 | { IORING_OP_WRITEV, IORING_OP_WRITE } | |
93 | }; | |
94 | ||
3f1e3af7 KB |
95 | static const int fixed_ddir_to_op[2] = { |
96 | IORING_OP_READ_FIXED, | |
97 | IORING_OP_WRITE_FIXED | |
98 | }; | |
99 | ||
2ea53ca3 | 100 | static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val) |
a90cd050 | 101 | { |
bffad86f | 102 | struct ioring_options *o = data; |
a90cd050 | 103 | |
2ea53ca3 JA |
104 | o->sqpoll_cpu = *val; |
105 | o->sqpoll_set = 1; | |
a90cd050 JA |
106 | return 0; |
107 | } | |
108 | ||
52885fa2 JA |
109 | static struct fio_option options[] = { |
110 | { | |
111 | .name = "hipri", | |
112 | .lname = "High Priority", | |
113 | .type = FIO_OPT_STR_SET, | |
bffad86f | 114 | .off1 = offsetof(struct ioring_options, hipri), |
52885fa2 JA |
115 | .help = "Use polled IO completions", |
116 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 117 | .group = FIO_OPT_G_IOURING, |
52885fa2 | 118 | }, |
b2a432bf PC |
119 | #ifdef FIO_HAVE_IOPRIO_CLASS |
120 | { | |
121 | .name = "cmdprio_percentage", | |
122 | .lname = "high priority percentage", | |
123 | .type = FIO_OPT_INT, | |
e9f6567a | 124 | .off1 = offsetof(struct ioring_options, |
d6cbeab4 | 125 | cmdprio_options.percentage[DDIR_READ]), |
e9f6567a | 126 | .off2 = offsetof(struct ioring_options, |
d6cbeab4 | 127 | cmdprio_options.percentage[DDIR_WRITE]), |
e9f6567a | 128 | .minval = 0, |
b2a432bf PC |
129 | .maxval = 100, |
130 | .help = "Send high priority I/O this percentage of the time", | |
131 | .category = FIO_OPT_C_ENGINE, | |
132 | .group = FIO_OPT_G_IOURING, | |
133 | }, | |
12f9d54a DLM |
134 | { |
135 | .name = "cmdprio_class", | |
136 | .lname = "Asynchronous I/O priority class", | |
137 | .type = FIO_OPT_INT, | |
138 | .off1 = offsetof(struct ioring_options, | |
d6cbeab4 | 139 | cmdprio_options.class[DDIR_READ]), |
12f9d54a | 140 | .off2 = offsetof(struct ioring_options, |
d6cbeab4 | 141 | cmdprio_options.class[DDIR_WRITE]), |
12f9d54a DLM |
142 | .help = "Set asynchronous IO priority class", |
143 | .minval = IOPRIO_MIN_PRIO_CLASS + 1, | |
144 | .maxval = IOPRIO_MAX_PRIO_CLASS, | |
145 | .interval = 1, | |
146 | .category = FIO_OPT_C_ENGINE, | |
147 | .group = FIO_OPT_G_IOURING, | |
148 | }, | |
149 | { | |
150 | .name = "cmdprio", | |
151 | .lname = "Asynchronous I/O priority level", | |
152 | .type = FIO_OPT_INT, | |
153 | .off1 = offsetof(struct ioring_options, | |
d6cbeab4 | 154 | cmdprio_options.level[DDIR_READ]), |
12f9d54a | 155 | .off2 = offsetof(struct ioring_options, |
d6cbeab4 | 156 | cmdprio_options.level[DDIR_WRITE]), |
12f9d54a DLM |
157 | .help = "Set asynchronous IO priority level", |
158 | .minval = IOPRIO_MIN_PRIO, | |
159 | .maxval = IOPRIO_MAX_PRIO, | |
160 | .interval = 1, | |
161 | .category = FIO_OPT_C_ENGINE, | |
162 | .group = FIO_OPT_G_IOURING, | |
163 | }, | |
a48f0cc7 DLM |
164 | { |
165 | .name = "cmdprio_bssplit", | |
166 | .lname = "Priority percentage block size split", | |
d6cbeab4 NC |
167 | .type = FIO_OPT_STR_STORE, |
168 | .off1 = offsetof(struct ioring_options, | |
169 | cmdprio_options.bssplit_str), | |
a48f0cc7 DLM |
170 | .help = "Set priority percentages for different block sizes", |
171 | .category = FIO_OPT_C_ENGINE, | |
172 | .group = FIO_OPT_G_IOURING, | |
173 | }, | |
b2a432bf PC |
174 | #else |
175 | { | |
176 | .name = "cmdprio_percentage", | |
177 | .lname = "high priority percentage", | |
178 | .type = FIO_OPT_UNSUPPORTED, | |
179 | .help = "Your platform does not support I/O priority classes", | |
180 | }, | |
12f9d54a DLM |
181 | { |
182 | .name = "cmdprio_class", | |
183 | .lname = "Asynchronous I/O priority class", | |
184 | .type = FIO_OPT_UNSUPPORTED, | |
185 | .help = "Your platform does not support I/O priority classes", | |
186 | }, | |
187 | { | |
188 | .name = "cmdprio", | |
189 | .lname = "Asynchronous I/O priority level", | |
190 | .type = FIO_OPT_UNSUPPORTED, | |
191 | .help = "Your platform does not support I/O priority classes", | |
192 | }, | |
a48f0cc7 DLM |
193 | { |
194 | .name = "cmdprio_bssplit", | |
195 | .lname = "Priority percentage block size split", | |
196 | .type = FIO_OPT_UNSUPPORTED, | |
197 | .help = "Your platform does not support I/O priority classes", | |
198 | }, | |
b2a432bf | 199 | #endif |
52885fa2 JA |
200 | { |
201 | .name = "fixedbufs", | |
202 | .lname = "Fixed (pre-mapped) IO buffers", | |
203 | .type = FIO_OPT_STR_SET, | |
bffad86f | 204 | .off1 = offsetof(struct ioring_options, fixedbufs), |
52885fa2 JA |
205 | .help = "Pre map IO buffers", |
206 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 207 | .group = FIO_OPT_G_IOURING, |
52885fa2 | 208 | }, |
5ffd5626 JA |
209 | { |
210 | .name = "registerfiles", | |
211 | .lname = "Register file set", | |
212 | .type = FIO_OPT_STR_SET, | |
213 | .off1 = offsetof(struct ioring_options, registerfiles), | |
214 | .help = "Pre-open/register files", | |
215 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 216 | .group = FIO_OPT_G_IOURING, |
5ffd5626 | 217 | }, |
771c9901 JA |
218 | { |
219 | .name = "sqthread_poll", | |
3d7d00a3 JA |
220 | .lname = "Kernel SQ thread polling", |
221 | .type = FIO_OPT_INT, | |
222 | .off1 = offsetof(struct ioring_options, sqpoll_thread), | |
223 | .help = "Offload submission/completion to kernel thread", | |
224 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 225 | .group = FIO_OPT_G_IOURING, |
3d7d00a3 JA |
226 | }, |
227 | { | |
228 | .name = "sqthread_poll_cpu", | |
229 | .lname = "SQ Thread Poll CPU", | |
2ea53ca3 JA |
230 | .type = FIO_OPT_INT, |
231 | .cb = fio_ioring_sqpoll_cb, | |
3d7d00a3 | 232 | .help = "What CPU to run SQ thread polling on", |
a90cd050 | 233 | .category = FIO_OPT_C_ENGINE, |
27f436d9 | 234 | .group = FIO_OPT_G_IOURING, |
a90cd050 | 235 | }, |
b10b1e70 JA |
236 | { |
237 | .name = "nonvectored", | |
238 | .lname = "Non-vectored", | |
239 | .type = FIO_OPT_INT, | |
240 | .off1 = offsetof(struct ioring_options, nonvectored), | |
556d8415 | 241 | .def = "-1", |
b10b1e70 JA |
242 | .help = "Use non-vectored read/write commands", |
243 | .category = FIO_OPT_C_ENGINE, | |
244 | .group = FIO_OPT_G_IOURING, | |
245 | }, | |
4a87b584 JA |
246 | { |
247 | .name = "uncached", | |
248 | .lname = "Uncached", | |
249 | .type = FIO_OPT_INT, | |
250 | .off1 = offsetof(struct ioring_options, uncached), | |
251 | .help = "Use RWF_UNCACHED for buffered read/writes", | |
252 | .category = FIO_OPT_C_ENGINE, | |
253 | .group = FIO_OPT_G_IOURING, | |
254 | }, | |
7d42e66e KK |
255 | { |
256 | .name = "nowait", | |
257 | .lname = "RWF_NOWAIT", | |
258 | .type = FIO_OPT_BOOL, | |
259 | .off1 = offsetof(struct ioring_options, nowait), | |
260 | .help = "Use RWF_NOWAIT for reads/writes", | |
261 | .category = FIO_OPT_C_ENGINE, | |
262 | .group = FIO_OPT_G_IOURING, | |
263 | }, | |
5a59a81d JA |
264 | { |
265 | .name = "force_async", | |
266 | .lname = "Force async", | |
267 | .type = FIO_OPT_INT, | |
268 | .off1 = offsetof(struct ioring_options, force_async), | |
269 | .help = "Set IOSQE_ASYNC every N requests", | |
270 | .category = FIO_OPT_C_ENGINE, | |
271 | .group = FIO_OPT_G_IOURING, | |
272 | }, | |
52885fa2 JA |
273 | { |
274 | .name = NULL, | |
275 | }, | |
276 | }; | |
277 | ||
bffad86f | 278 | static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit, |
52885fa2 JA |
279 | unsigned int min_complete, unsigned int flags) |
280 | { | |
bfed648c | 281 | return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit, |
521164fa | 282 | min_complete, flags, NULL, 0); |
52885fa2 JA |
283 | } |
284 | ||
bffad86f | 285 | static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 286 | { |
bffad86f | 287 | struct ioring_data *ld = td->io_ops_data; |
cfcc8564 | 288 | struct ioring_options *o = td->eo; |
52885fa2 | 289 | struct fio_file *f = io_u->file; |
f0403f94 | 290 | struct io_uring_sqe *sqe; |
52885fa2 | 291 | |
f0403f94 | 292 | sqe = &ld->sqes[io_u->index]; |
34d6090e | 293 | |
5ffd5626 JA |
294 | if (o->registerfiles) { |
295 | sqe->fd = f->engine_pos; | |
296 | sqe->flags = IOSQE_FIXED_FILE; | |
297 | } else { | |
298 | sqe->fd = f->fd; | |
87b69ef2 | 299 | sqe->flags = 0; |
5ffd5626 | 300 | } |
52885fa2 | 301 | |
e3970057 | 302 | if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) { |
f0403f94 | 303 | if (o->fixedbufs) { |
3f1e3af7 | 304 | sqe->opcode = fixed_ddir_to_op[io_u->ddir]; |
919850d2 | 305 | sqe->addr = (unsigned long) io_u->xfer_buf; |
f0403f94 | 306 | sqe->len = io_u->xfer_buflen; |
2ea53ca3 | 307 | sqe->buf_index = io_u->index; |
cfcc8564 | 308 | } else { |
832faaaf JA |
309 | struct iovec *iov = &ld->iovecs[io_u->index]; |
310 | ||
311 | /* | |
312 | * Update based on actual io_u, requeue could have | |
313 | * adjusted these | |
314 | */ | |
315 | iov->iov_base = io_u->xfer_buf; | |
316 | iov->iov_len = io_u->xfer_buflen; | |
317 | ||
3f1e3af7 | 318 | sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored]; |
b10b1e70 | 319 | if (o->nonvectored) { |
832faaaf JA |
320 | sqe->addr = (unsigned long) iov->iov_base; |
321 | sqe->len = iov->iov_len; | |
b10b1e70 | 322 | } else { |
832faaaf | 323 | sqe->addr = (unsigned long) iov; |
b10b1e70 JA |
324 | sqe->len = 1; |
325 | } | |
cfcc8564 | 326 | } |
fd70e361 | 327 | sqe->rw_flags = 0; |
4a87b584 | 328 | if (!td->o.odirect && o->uncached) |
fd70e361 | 329 | sqe->rw_flags |= RWF_UNCACHED; |
7d42e66e KK |
330 | if (o->nowait) |
331 | sqe->rw_flags |= RWF_NOWAIT; | |
8ff6b289 NC |
332 | |
333 | /* | |
334 | * Since io_uring can have a submission context (sqthread_poll) | |
335 | * that is different from the process context, we cannot rely on | |
336 | * the IO priority set by ioprio_set() (option prio/prioclass) | |
337 | * to be inherited. | |
338 | * td->ioprio will have the value of the "default prio", so set | |
339 | * this unconditionally. This value might get overridden by | |
ff00f247 | 340 | * fio_ioring_cmdprio_prep() if the option cmdprio_percentage or |
8ff6b289 NC |
341 | * cmdprio_bssplit is used. |
342 | */ | |
343 | sqe->ioprio = td->ioprio; | |
f0403f94 | 344 | sqe->off = io_u->offset; |
48e698fa | 345 | } else if (ddir_sync(io_u->ddir)) { |
7c70f506 | 346 | sqe->ioprio = 0; |
01387bfe AF |
347 | if (io_u->ddir == DDIR_SYNC_FILE_RANGE) { |
348 | sqe->off = f->first_write; | |
349 | sqe->len = f->last_write - f->first_write; | |
350 | sqe->sync_range_flags = td->o.sync_file_range; | |
351 | sqe->opcode = IORING_OP_SYNC_FILE_RANGE; | |
352 | } else { | |
7c70f506 JA |
353 | sqe->off = 0; |
354 | sqe->addr = 0; | |
355 | sqe->len = 0; | |
01387bfe AF |
356 | if (io_u->ddir == DDIR_DATASYNC) |
357 | sqe->fsync_flags |= IORING_FSYNC_DATASYNC; | |
358 | sqe->opcode = IORING_OP_FSYNC; | |
359 | } | |
48e698fa | 360 | } |
52885fa2 | 361 | |
5a59a81d JA |
362 | if (o->force_async && ++ld->prepped == o->force_async) { |
363 | ld->prepped = 0; | |
364 | sqe->flags |= IOSQE_ASYNC; | |
365 | } | |
366 | ||
48e698fa | 367 | sqe->user_data = (unsigned long) io_u; |
52885fa2 JA |
368 | return 0; |
369 | } | |
370 | ||
bffad86f | 371 | static struct io_u *fio_ioring_event(struct thread_data *td, int event) |
52885fa2 | 372 | { |
bffad86f | 373 | struct ioring_data *ld = td->io_ops_data; |
f0403f94 | 374 | struct io_uring_cqe *cqe; |
52885fa2 | 375 | struct io_u *io_u; |
b87aa01a | 376 | unsigned index; |
52885fa2 | 377 | |
b87aa01a | 378 | index = (event + ld->cq_ring_off) & ld->cq_ring_mask; |
52885fa2 | 379 | |
f0403f94 | 380 | cqe = &ld->cq_ring.cqes[index]; |
e3466352 | 381 | io_u = (struct io_u *) (uintptr_t) cqe->user_data; |
52885fa2 | 382 | |
f0403f94 JA |
383 | if (cqe->res != io_u->xfer_buflen) { |
384 | if (cqe->res > io_u->xfer_buflen) | |
385 | io_u->error = -cqe->res; | |
52885fa2 | 386 | else |
f0403f94 | 387 | io_u->resid = io_u->xfer_buflen - cqe->res; |
52885fa2 JA |
388 | } else |
389 | io_u->error = 0; | |
390 | ||
391 | return io_u; | |
392 | } | |
393 | ||
bffad86f | 394 | static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events, |
52885fa2 JA |
395 | unsigned int max) |
396 | { | |
bffad86f JA |
397 | struct ioring_data *ld = td->io_ops_data; |
398 | struct io_cq_ring *ring = &ld->cq_ring; | |
e2239016 | 399 | unsigned head, reaped = 0; |
52885fa2 | 400 | |
9a2d78b3 | 401 | head = *ring->head; |
52885fa2 | 402 | do { |
9e26aff9 | 403 | if (head == atomic_load_acquire(ring->tail)) |
52885fa2 JA |
404 | break; |
405 | reaped++; | |
406 | head++; | |
52885fa2 JA |
407 | } while (reaped + events < max); |
408 | ||
76ce63dd AB |
409 | if (reaped) |
410 | atomic_store_release(ring->head, head); | |
411 | ||
52885fa2 JA |
412 | return reaped; |
413 | } | |
414 | ||
bffad86f JA |
415 | static int fio_ioring_getevents(struct thread_data *td, unsigned int min, |
416 | unsigned int max, const struct timespec *t) | |
52885fa2 | 417 | { |
bffad86f | 418 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 | 419 | unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min; |
bffad86f JA |
420 | struct ioring_options *o = td->eo; |
421 | struct io_cq_ring *ring = &ld->cq_ring; | |
b87aa01a JA |
422 | unsigned events = 0; |
423 | int r; | |
52885fa2 | 424 | |
9a2d78b3 | 425 | ld->cq_ring_off = *ring->head; |
52885fa2 | 426 | do { |
bffad86f | 427 | r = fio_ioring_cqring_reap(td, events, max); |
52885fa2 JA |
428 | if (r) { |
429 | events += r; | |
f7cbbbf8 ST |
430 | if (actual_min != 0) |
431 | actual_min -= r; | |
52885fa2 JA |
432 | continue; |
433 | } | |
434 | ||
3d7d00a3 | 435 | if (!o->sqpoll_thread) { |
9a2d78b3 JA |
436 | r = io_uring_enter(ld, 0, actual_min, |
437 | IORING_ENTER_GETEVENTS); | |
771c9901 | 438 | if (r < 0) { |
f6abd731 | 439 | if (errno == EAGAIN || errno == EINTR) |
771c9901 | 440 | continue; |
9a2d78b3 | 441 | td_verror(td, errno, "io_uring_enter"); |
771c9901 JA |
442 | break; |
443 | } | |
52885fa2 JA |
444 | } |
445 | } while (events < min); | |
446 | ||
447 | return r < 0 ? r : events; | |
448 | } | |
449 | ||
127715b6 NC |
450 | static inline void fio_ioring_cmdprio_prep(struct thread_data *td, |
451 | struct io_u *io_u) | |
b2a432bf | 452 | { |
b2a432bf | 453 | struct ioring_data *ld = td->io_ops_data; |
d6cbeab4 | 454 | struct cmdprio *cmdprio = &ld->cmdprio; |
127715b6 NC |
455 | |
456 | if (fio_cmdprio_set_ioprio(td, cmdprio, io_u)) | |
457 | ld->sqes[io_u->index].ioprio = io_u->ioprio; | |
b2a432bf PC |
458 | } |
459 | ||
bffad86f JA |
460 | static enum fio_q_status fio_ioring_queue(struct thread_data *td, |
461 | struct io_u *io_u) | |
52885fa2 | 462 | { |
bffad86f JA |
463 | struct ioring_data *ld = td->io_ops_data; |
464 | struct io_sq_ring *ring = &ld->sq_ring; | |
52885fa2 JA |
465 | unsigned tail, next_tail; |
466 | ||
467 | fio_ro_check(td, io_u); | |
468 | ||
b87aa01a | 469 | if (ld->queued == ld->iodepth) |
52885fa2 JA |
470 | return FIO_Q_BUSY; |
471 | ||
52885fa2 JA |
472 | if (io_u->ddir == DDIR_TRIM) { |
473 | if (ld->queued) | |
474 | return FIO_Q_BUSY; | |
475 | ||
476 | do_io_u_trim(td, io_u); | |
477 | io_u_mark_submit(td, 1); | |
478 | io_u_mark_complete(td, 1); | |
479 | return FIO_Q_COMPLETED; | |
480 | } | |
481 | ||
9a2d78b3 | 482 | tail = *ring->tail; |
52885fa2 | 483 | next_tail = tail + 1; |
9e26aff9 | 484 | if (next_tail == atomic_load_acquire(ring->head)) |
52885fa2 JA |
485 | return FIO_Q_BUSY; |
486 | ||
d6cbeab4 | 487 | if (ld->cmdprio.mode != CMDPRIO_MODE_NONE) |
ff00f247 NC |
488 | fio_ioring_cmdprio_prep(td, io_u); |
489 | ||
b87aa01a | 490 | ring->array[tail & ld->sq_ring_mask] = io_u->index; |
9e26aff9 | 491 | atomic_store_release(ring->tail, next_tail); |
52885fa2 JA |
492 | |
493 | ld->queued++; | |
494 | return FIO_Q_QUEUED; | |
495 | } | |
496 | ||
bffad86f | 497 | static void fio_ioring_queued(struct thread_data *td, int start, int nr) |
52885fa2 | 498 | { |
bffad86f | 499 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
500 | struct timespec now; |
501 | ||
502 | if (!fio_fill_issue_time(td)) | |
503 | return; | |
504 | ||
505 | fio_gettime(&now, NULL); | |
506 | ||
507 | while (nr--) { | |
bffad86f | 508 | struct io_sq_ring *ring = &ld->sq_ring; |
9a2d78b3 | 509 | int index = ring->array[start & ld->sq_ring_mask]; |
f8289afc | 510 | struct io_u *io_u = ld->io_u_index[index]; |
52885fa2 JA |
511 | |
512 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
513 | io_u_queued(td, io_u); | |
514 | ||
515 | start++; | |
52885fa2 JA |
516 | } |
517 | } | |
518 | ||
bffad86f | 519 | static int fio_ioring_commit(struct thread_data *td) |
52885fa2 | 520 | { |
bffad86f JA |
521 | struct ioring_data *ld = td->io_ops_data; |
522 | struct ioring_options *o = td->eo; | |
52885fa2 JA |
523 | int ret; |
524 | ||
525 | if (!ld->queued) | |
526 | return 0; | |
527 | ||
3d7d00a3 JA |
528 | /* |
529 | * Kernel side does submission. just need to check if the ring is | |
530 | * flagged as needing a kick, if so, call io_uring_enter(). This | |
531 | * only happens if we've been idle too long. | |
532 | */ | |
533 | if (o->sqpoll_thread) { | |
bffad86f | 534 | struct io_sq_ring *ring = &ld->sq_ring; |
2dd96cc4 | 535 | unsigned flags; |
4cdbc048 | 536 | |
2dd96cc4 JA |
537 | flags = atomic_load_acquire(ring->flags); |
538 | if (flags & IORING_SQ_NEED_WAKEUP) | |
b532dd6d JA |
539 | io_uring_enter(ld, ld->queued, 0, |
540 | IORING_ENTER_SQ_WAKEUP); | |
771c9901 JA |
541 | ld->queued = 0; |
542 | return 0; | |
543 | } | |
544 | ||
52885fa2 | 545 | do { |
9a2d78b3 | 546 | unsigned start = *ld->sq_ring.head; |
52885fa2 JA |
547 | long nr = ld->queued; |
548 | ||
9a2d78b3 | 549 | ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS); |
52885fa2 | 550 | if (ret > 0) { |
bffad86f | 551 | fio_ioring_queued(td, start, ret); |
52885fa2 JA |
552 | io_u_mark_submit(td, ret); |
553 | ||
554 | ld->queued -= ret; | |
555 | ret = 0; | |
a90cd050 JA |
556 | } else if (!ret) { |
557 | io_u_mark_submit(td, ret); | |
52885fa2 | 558 | continue; |
a90cd050 | 559 | } else { |
f6abd731 | 560 | if (errno == EAGAIN || errno == EINTR) { |
bffad86f | 561 | ret = fio_ioring_cqring_reap(td, 0, ld->queued); |
a90cd050 JA |
562 | if (ret) |
563 | continue; | |
564 | /* Shouldn't happen */ | |
565 | usleep(1); | |
566 | continue; | |
52885fa2 | 567 | } |
9a2d78b3 | 568 | td_verror(td, errno, "io_uring_enter submit"); |
52885fa2 | 569 | break; |
a90cd050 | 570 | } |
52885fa2 JA |
571 | } while (ld->queued); |
572 | ||
573 | return ret; | |
574 | } | |
575 | ||
bffad86f | 576 | static void fio_ioring_unmap(struct ioring_data *ld) |
52885fa2 | 577 | { |
9a2d78b3 | 578 | int i; |
52885fa2 | 579 | |
59f94d26 | 580 | for (i = 0; i < FIO_ARRAY_SIZE(ld->mmap); i++) |
9a2d78b3 JA |
581 | munmap(ld->mmap[i].ptr, ld->mmap[i].len); |
582 | close(ld->ring_fd); | |
b87aa01a JA |
583 | } |
584 | ||
bffad86f | 585 | static void fio_ioring_cleanup(struct thread_data *td) |
52885fa2 | 586 | { |
bffad86f | 587 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
588 | |
589 | if (ld) { | |
52885fa2 | 590 | if (!(td->flags & TD_F_CHILD)) |
bffad86f | 591 | fio_ioring_unmap(ld); |
9a2d78b3 | 592 | |
d6cbeab4 | 593 | fio_cmdprio_cleanup(&ld->cmdprio); |
52885fa2 | 594 | free(ld->io_u_index); |
9a2d78b3 | 595 | free(ld->iovecs); |
5ffd5626 | 596 | free(ld->fds); |
52885fa2 JA |
597 | free(ld); |
598 | } | |
599 | } | |
600 | ||
bffad86f | 601 | static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p) |
9a2d78b3 | 602 | { |
bffad86f JA |
603 | struct io_sq_ring *sring = &ld->sq_ring; |
604 | struct io_cq_ring *cring = &ld->cq_ring; | |
9a2d78b3 JA |
605 | void *ptr; |
606 | ||
e2239016 | 607 | ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32); |
9a2d78b3 JA |
608 | ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE, |
609 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
610 | IORING_OFF_SQ_RING); | |
611 | ld->mmap[0].ptr = ptr; | |
612 | sring->head = ptr + p->sq_off.head; | |
613 | sring->tail = ptr + p->sq_off.tail; | |
614 | sring->ring_mask = ptr + p->sq_off.ring_mask; | |
615 | sring->ring_entries = ptr + p->sq_off.ring_entries; | |
616 | sring->flags = ptr + p->sq_off.flags; | |
ac122fea | 617 | sring->array = ptr + p->sq_off.array; |
9a2d78b3 JA |
618 | ld->sq_ring_mask = *sring->ring_mask; |
619 | ||
f0403f94 JA |
620 | ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe); |
621 | ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE, | |
9a2d78b3 | 622 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, |
f0403f94 JA |
623 | IORING_OFF_SQES); |
624 | ld->mmap[1].ptr = ld->sqes; | |
9a2d78b3 | 625 | |
f0403f94 JA |
626 | ld->mmap[2].len = p->cq_off.cqes + |
627 | p->cq_entries * sizeof(struct io_uring_cqe); | |
9a2d78b3 JA |
628 | ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE, |
629 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
630 | IORING_OFF_CQ_RING); | |
631 | ld->mmap[2].ptr = ptr; | |
632 | cring->head = ptr + p->cq_off.head; | |
633 | cring->tail = ptr + p->cq_off.tail; | |
634 | cring->ring_mask = ptr + p->cq_off.ring_mask; | |
635 | cring->ring_entries = ptr + p->cq_off.ring_entries; | |
f0403f94 | 636 | cring->cqes = ptr + p->cq_off.cqes; |
9a2d78b3 JA |
637 | ld->cq_ring_mask = *cring->ring_mask; |
638 | return 0; | |
639 | } | |
640 | ||
556d8415 JA |
641 | static void fio_ioring_probe(struct thread_data *td) |
642 | { | |
643 | struct ioring_data *ld = td->io_ops_data; | |
644 | struct ioring_options *o = td->eo; | |
645 | struct io_uring_probe *p; | |
646 | int ret; | |
647 | ||
648 | /* already set by user, don't touch */ | |
649 | if (o->nonvectored != -1) | |
650 | return; | |
651 | ||
652 | /* default to off, as that's always safe */ | |
653 | o->nonvectored = 0; | |
654 | ||
655 | p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op)); | |
656 | if (!p) | |
657 | return; | |
658 | ||
659 | memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op)); | |
660 | ret = syscall(__NR_io_uring_register, ld->ring_fd, | |
661 | IORING_REGISTER_PROBE, p, 256); | |
662 | if (ret < 0) | |
663 | goto out; | |
664 | ||
665 | if (IORING_OP_WRITE > p->ops_len) | |
666 | goto out; | |
667 | ||
668 | if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED) && | |
669 | (p->ops[IORING_OP_WRITE].flags & IO_URING_OP_SUPPORTED)) | |
670 | o->nonvectored = 1; | |
671 | out: | |
672 | free(p); | |
673 | } | |
674 | ||
bffad86f | 675 | static int fio_ioring_queue_init(struct thread_data *td) |
52885fa2 | 676 | { |
bffad86f JA |
677 | struct ioring_data *ld = td->io_ops_data; |
678 | struct ioring_options *o = td->eo; | |
52885fa2 | 679 | int depth = td->o.iodepth; |
bffad86f | 680 | struct io_uring_params p; |
9a2d78b3 JA |
681 | int ret; |
682 | ||
683 | memset(&p, 0, sizeof(p)); | |
52885fa2 JA |
684 | |
685 | if (o->hipri) | |
bffad86f | 686 | p.flags |= IORING_SETUP_IOPOLL; |
3d7d00a3 JA |
687 | if (o->sqpoll_thread) { |
688 | p.flags |= IORING_SETUP_SQPOLL; | |
689 | if (o->sqpoll_set) { | |
690 | p.flags |= IORING_SETUP_SQ_AFF; | |
691 | p.sq_thread_cpu = o->sqpoll_cpu; | |
692 | } | |
f635f1fb | 693 | } |
a90cd050 | 694 | |
1db268db JA |
695 | /* |
696 | * Clamp CQ ring size at our SQ ring size, we don't need more entries | |
697 | * than that. | |
698 | */ | |
699 | p.flags |= IORING_SETUP_CQSIZE; | |
700 | p.cq_entries = depth; | |
701 | ||
b5e99df6 | 702 | retry: |
bfed648c | 703 | ret = syscall(__NR_io_uring_setup, depth, &p); |
b5e99df6 JA |
704 | if (ret < 0) { |
705 | if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) { | |
706 | p.flags &= ~IORING_SETUP_CQSIZE; | |
707 | goto retry; | |
708 | } | |
9a2d78b3 | 709 | return ret; |
b5e99df6 | 710 | } |
9a2d78b3 JA |
711 | |
712 | ld->ring_fd = ret; | |
2ea53ca3 | 713 | |
556d8415 JA |
714 | fio_ioring_probe(td); |
715 | ||
2ea53ca3 | 716 | if (o->fixedbufs) { |
bfed648c | 717 | ret = syscall(__NR_io_uring_register, ld->ring_fd, |
919850d2 | 718 | IORING_REGISTER_BUFFERS, ld->iovecs, depth); |
2ea53ca3 JA |
719 | if (ret < 0) |
720 | return ret; | |
721 | } | |
722 | ||
bffad86f | 723 | return fio_ioring_mmap(ld, &p); |
52885fa2 JA |
724 | } |
725 | ||
5ffd5626 JA |
726 | static int fio_ioring_register_files(struct thread_data *td) |
727 | { | |
728 | struct ioring_data *ld = td->io_ops_data; | |
729 | struct fio_file *f; | |
730 | unsigned int i; | |
731 | int ret; | |
732 | ||
733 | ld->fds = calloc(td->o.nr_files, sizeof(int)); | |
734 | ||
735 | for_each_file(td, f, i) { | |
736 | ret = generic_open_file(td, f); | |
737 | if (ret) | |
738 | goto err; | |
739 | ld->fds[i] = f->fd; | |
740 | f->engine_pos = i; | |
741 | } | |
742 | ||
bfed648c | 743 | ret = syscall(__NR_io_uring_register, ld->ring_fd, |
5ffd5626 JA |
744 | IORING_REGISTER_FILES, ld->fds, td->o.nr_files); |
745 | if (ret) { | |
746 | err: | |
747 | free(ld->fds); | |
748 | ld->fds = NULL; | |
749 | } | |
750 | ||
751 | /* | |
752 | * Pretend the file is closed again, and really close it if we hit | |
753 | * an error. | |
754 | */ | |
755 | for_each_file(td, f, i) { | |
756 | if (ret) { | |
757 | int fio_unused ret2; | |
758 | ret2 = generic_close_file(td, f); | |
759 | } else | |
760 | f->fd = -1; | |
761 | } | |
762 | ||
763 | return ret; | |
764 | } | |
765 | ||
bffad86f | 766 | static int fio_ioring_post_init(struct thread_data *td) |
52885fa2 | 767 | { |
bffad86f | 768 | struct ioring_data *ld = td->io_ops_data; |
5ffd5626 | 769 | struct ioring_options *o = td->eo; |
52885fa2 | 770 | struct io_u *io_u; |
650346e1 | 771 | int err, i; |
52885fa2 | 772 | |
650346e1 JA |
773 | for (i = 0; i < td->o.iodepth; i++) { |
774 | struct iovec *iov = &ld->iovecs[i]; | |
9a2d78b3 | 775 | |
650346e1 JA |
776 | io_u = ld->io_u_index[i]; |
777 | iov->iov_base = io_u->buf; | |
778 | iov->iov_len = td_max_bs(td); | |
52885fa2 JA |
779 | } |
780 | ||
bffad86f | 781 | err = fio_ioring_queue_init(td); |
52885fa2 | 782 | if (err) { |
0442b53f | 783 | int init_err = errno; |
c4f5c92f | 784 | |
0442b53f | 785 | if (init_err == ENOSYS) |
c4f5c92f | 786 | log_err("fio: your kernel doesn't support io_uring\n"); |
0442b53f | 787 | td_verror(td, init_err, "io_queue_init"); |
52885fa2 JA |
788 | return 1; |
789 | } | |
790 | ||
7c70f506 JA |
791 | for (i = 0; i < td->o.iodepth; i++) { |
792 | struct io_uring_sqe *sqe; | |
793 | ||
794 | sqe = &ld->sqes[i]; | |
795 | memset(sqe, 0, sizeof(*sqe)); | |
796 | } | |
797 | ||
5ffd5626 JA |
798 | if (o->registerfiles) { |
799 | err = fio_ioring_register_files(td); | |
800 | if (err) { | |
801 | td_verror(td, errno, "ioring_register_files"); | |
802 | return 1; | |
803 | } | |
804 | } | |
805 | ||
52885fa2 JA |
806 | return 0; |
807 | } | |
808 | ||
bffad86f | 809 | static int fio_ioring_init(struct thread_data *td) |
52885fa2 | 810 | { |
5ffd5626 | 811 | struct ioring_options *o = td->eo; |
bffad86f | 812 | struct ioring_data *ld; |
e9f6567a | 813 | int ret; |
52885fa2 | 814 | |
5ffd5626 JA |
815 | /* sqthread submission requires registered files */ |
816 | if (o->sqpoll_thread) | |
817 | o->registerfiles = 1; | |
818 | ||
819 | if (o->registerfiles && td->o.nr_files != td->o.open_files) { | |
820 | log_err("fio: io_uring registered files require nr_files to " | |
821 | "be identical to open_files\n"); | |
822 | return 1; | |
823 | } | |
824 | ||
52885fa2 JA |
825 | ld = calloc(1, sizeof(*ld)); |
826 | ||
b87aa01a JA |
827 | /* ring depth must be a power-of-2 */ |
828 | ld->iodepth = td->o.iodepth; | |
829 | td->o.iodepth = roundup_pow2(td->o.iodepth); | |
830 | ||
52885fa2 JA |
831 | /* io_u index */ |
832 | ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *)); | |
650346e1 | 833 | ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec)); |
52885fa2 JA |
834 | |
835 | td->io_ops_data = ld; | |
b2a432bf | 836 | |
d6cbeab4 | 837 | ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options); |
e9f6567a DLM |
838 | if (ret) { |
839 | td_verror(td, EINVAL, "fio_ioring_init"); | |
b2a432bf PC |
840 | return 1; |
841 | } | |
1af44196 | 842 | |
52885fa2 JA |
843 | return 0; |
844 | } | |
845 | ||
bffad86f | 846 | static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 847 | { |
bffad86f | 848 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
849 | |
850 | ld->io_u_index[io_u->index] = io_u; | |
851 | return 0; | |
852 | } | |
853 | ||
5ffd5626 JA |
854 | static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f) |
855 | { | |
856 | struct ioring_data *ld = td->io_ops_data; | |
857 | struct ioring_options *o = td->eo; | |
858 | ||
17318cf6 | 859 | if (!ld || !o->registerfiles) |
5ffd5626 JA |
860 | return generic_open_file(td, f); |
861 | ||
862 | f->fd = ld->fds[f->engine_pos]; | |
863 | return 0; | |
864 | } | |
865 | ||
866 | static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f) | |
867 | { | |
17318cf6 | 868 | struct ioring_data *ld = td->io_ops_data; |
5ffd5626 JA |
869 | struct ioring_options *o = td->eo; |
870 | ||
17318cf6 | 871 | if (!ld || !o->registerfiles) |
5ffd5626 JA |
872 | return generic_close_file(td, f); |
873 | ||
874 | f->fd = -1; | |
875 | return 0; | |
876 | } | |
877 | ||
52885fa2 | 878 | static struct ioengine_ops ioengine = { |
bffad86f | 879 | .name = "io_uring", |
52885fa2 | 880 | .version = FIO_IOOPS_VERSION, |
8bfe330e | 881 | .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD, |
bffad86f JA |
882 | .init = fio_ioring_init, |
883 | .post_init = fio_ioring_post_init, | |
884 | .io_u_init = fio_ioring_io_u_init, | |
885 | .prep = fio_ioring_prep, | |
886 | .queue = fio_ioring_queue, | |
887 | .commit = fio_ioring_commit, | |
888 | .getevents = fio_ioring_getevents, | |
889 | .event = fio_ioring_event, | |
890 | .cleanup = fio_ioring_cleanup, | |
5ffd5626 JA |
891 | .open_file = fio_ioring_open_file, |
892 | .close_file = fio_ioring_close_file, | |
52885fa2 JA |
893 | .get_file_size = generic_get_file_size, |
894 | .options = options, | |
bffad86f | 895 | .option_struct_size = sizeof(struct ioring_options), |
52885fa2 JA |
896 | }; |
897 | ||
bffad86f | 898 | static void fio_init fio_ioring_register(void) |
52885fa2 | 899 | { |
52885fa2 | 900 | register_ioengine(&ioengine); |
52885fa2 JA |
901 | } |
902 | ||
bffad86f | 903 | static void fio_exit fio_ioring_unregister(void) |
52885fa2 | 904 | { |
52885fa2 | 905 | unregister_ioengine(&ioengine); |
52885fa2 | 906 | } |
1f90e9bb | 907 | #endif |