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" |
855dc4d4 AG |
27 | #include "nvme.h" |
28 | ||
29 | #include <sys/stat.h> | |
30 | ||
31 | enum uring_cmd_type { | |
32 | FIO_URING_CMD_NVME = 1, | |
33 | }; | |
9a2d78b3 | 34 | |
bffad86f | 35 | struct io_sq_ring { |
e2239016 JA |
36 | unsigned *head; |
37 | unsigned *tail; | |
38 | unsigned *ring_mask; | |
39 | unsigned *ring_entries; | |
40 | unsigned *flags; | |
41 | unsigned *array; | |
52885fa2 JA |
42 | }; |
43 | ||
bffad86f | 44 | struct io_cq_ring { |
e2239016 JA |
45 | unsigned *head; |
46 | unsigned *tail; | |
47 | unsigned *ring_mask; | |
48 | unsigned *ring_entries; | |
f0403f94 | 49 | struct io_uring_cqe *cqes; |
9a2d78b3 JA |
50 | }; |
51 | ||
bffad86f | 52 | struct ioring_mmap { |
9a2d78b3 JA |
53 | void *ptr; |
54 | size_t len; | |
52885fa2 JA |
55 | }; |
56 | ||
bffad86f | 57 | struct ioring_data { |
9a2d78b3 JA |
58 | int ring_fd; |
59 | ||
52885fa2 JA |
60 | struct io_u **io_u_index; |
61 | ||
5ffd5626 JA |
62 | int *fds; |
63 | ||
bffad86f | 64 | struct io_sq_ring sq_ring; |
f0403f94 | 65 | struct io_uring_sqe *sqes; |
9a2d78b3 | 66 | struct iovec *iovecs; |
b87aa01a | 67 | unsigned sq_ring_mask; |
52885fa2 | 68 | |
bffad86f | 69 | struct io_cq_ring cq_ring; |
b87aa01a | 70 | unsigned cq_ring_mask; |
52885fa2 JA |
71 | |
72 | int queued; | |
73 | int cq_ring_off; | |
b87aa01a | 74 | unsigned iodepth; |
5a59a81d | 75 | int prepped; |
96563db9 | 76 | |
bffad86f | 77 | struct ioring_mmap mmap[3]; |
d6cbeab4 NC |
78 | |
79 | struct cmdprio cmdprio; | |
52885fa2 JA |
80 | }; |
81 | ||
bffad86f | 82 | struct ioring_options { |
a48f0cc7 | 83 | struct thread_data *td; |
52885fa2 | 84 | unsigned int hipri; |
d6cbeab4 | 85 | struct cmdprio_options cmdprio_options; |
52885fa2 | 86 | unsigned int fixedbufs; |
5ffd5626 | 87 | unsigned int registerfiles; |
3d7d00a3 | 88 | unsigned int sqpoll_thread; |
2ea53ca3 JA |
89 | unsigned int sqpoll_set; |
90 | unsigned int sqpoll_cpu; | |
b10b1e70 | 91 | unsigned int nonvectored; |
4a87b584 | 92 | unsigned int uncached; |
7d42e66e | 93 | unsigned int nowait; |
5a59a81d | 94 | unsigned int force_async; |
855dc4d4 | 95 | enum uring_cmd_type cmd_type; |
52885fa2 JA |
96 | }; |
97 | ||
b10b1e70 JA |
98 | static const int ddir_to_op[2][2] = { |
99 | { IORING_OP_READV, IORING_OP_READ }, | |
100 | { IORING_OP_WRITEV, IORING_OP_WRITE } | |
101 | }; | |
102 | ||
3f1e3af7 KB |
103 | static const int fixed_ddir_to_op[2] = { |
104 | IORING_OP_READ_FIXED, | |
105 | IORING_OP_WRITE_FIXED | |
106 | }; | |
107 | ||
2ea53ca3 | 108 | static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val) |
a90cd050 | 109 | { |
bffad86f | 110 | struct ioring_options *o = data; |
a90cd050 | 111 | |
2ea53ca3 JA |
112 | o->sqpoll_cpu = *val; |
113 | o->sqpoll_set = 1; | |
a90cd050 JA |
114 | return 0; |
115 | } | |
116 | ||
52885fa2 JA |
117 | static struct fio_option options[] = { |
118 | { | |
119 | .name = "hipri", | |
120 | .lname = "High Priority", | |
121 | .type = FIO_OPT_STR_SET, | |
bffad86f | 122 | .off1 = offsetof(struct ioring_options, hipri), |
52885fa2 JA |
123 | .help = "Use polled IO completions", |
124 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 125 | .group = FIO_OPT_G_IOURING, |
52885fa2 | 126 | }, |
b2a432bf PC |
127 | #ifdef FIO_HAVE_IOPRIO_CLASS |
128 | { | |
129 | .name = "cmdprio_percentage", | |
130 | .lname = "high priority percentage", | |
131 | .type = FIO_OPT_INT, | |
e9f6567a | 132 | .off1 = offsetof(struct ioring_options, |
d6cbeab4 | 133 | cmdprio_options.percentage[DDIR_READ]), |
e9f6567a | 134 | .off2 = offsetof(struct ioring_options, |
d6cbeab4 | 135 | cmdprio_options.percentage[DDIR_WRITE]), |
e9f6567a | 136 | .minval = 0, |
b2a432bf PC |
137 | .maxval = 100, |
138 | .help = "Send high priority I/O this percentage of the time", | |
139 | .category = FIO_OPT_C_ENGINE, | |
140 | .group = FIO_OPT_G_IOURING, | |
141 | }, | |
12f9d54a DLM |
142 | { |
143 | .name = "cmdprio_class", | |
144 | .lname = "Asynchronous I/O priority class", | |
145 | .type = FIO_OPT_INT, | |
146 | .off1 = offsetof(struct ioring_options, | |
d6cbeab4 | 147 | cmdprio_options.class[DDIR_READ]), |
12f9d54a | 148 | .off2 = offsetof(struct ioring_options, |
d6cbeab4 | 149 | cmdprio_options.class[DDIR_WRITE]), |
12f9d54a DLM |
150 | .help = "Set asynchronous IO priority class", |
151 | .minval = IOPRIO_MIN_PRIO_CLASS + 1, | |
152 | .maxval = IOPRIO_MAX_PRIO_CLASS, | |
153 | .interval = 1, | |
154 | .category = FIO_OPT_C_ENGINE, | |
155 | .group = FIO_OPT_G_IOURING, | |
156 | }, | |
157 | { | |
158 | .name = "cmdprio", | |
159 | .lname = "Asynchronous I/O priority level", | |
160 | .type = FIO_OPT_INT, | |
161 | .off1 = offsetof(struct ioring_options, | |
d6cbeab4 | 162 | cmdprio_options.level[DDIR_READ]), |
12f9d54a | 163 | .off2 = offsetof(struct ioring_options, |
d6cbeab4 | 164 | cmdprio_options.level[DDIR_WRITE]), |
12f9d54a DLM |
165 | .help = "Set asynchronous IO priority level", |
166 | .minval = IOPRIO_MIN_PRIO, | |
167 | .maxval = IOPRIO_MAX_PRIO, | |
168 | .interval = 1, | |
169 | .category = FIO_OPT_C_ENGINE, | |
170 | .group = FIO_OPT_G_IOURING, | |
171 | }, | |
a48f0cc7 DLM |
172 | { |
173 | .name = "cmdprio_bssplit", | |
174 | .lname = "Priority percentage block size split", | |
d6cbeab4 NC |
175 | .type = FIO_OPT_STR_STORE, |
176 | .off1 = offsetof(struct ioring_options, | |
177 | cmdprio_options.bssplit_str), | |
a48f0cc7 DLM |
178 | .help = "Set priority percentages for different block sizes", |
179 | .category = FIO_OPT_C_ENGINE, | |
180 | .group = FIO_OPT_G_IOURING, | |
181 | }, | |
b2a432bf PC |
182 | #else |
183 | { | |
184 | .name = "cmdprio_percentage", | |
185 | .lname = "high priority percentage", | |
186 | .type = FIO_OPT_UNSUPPORTED, | |
187 | .help = "Your platform does not support I/O priority classes", | |
188 | }, | |
12f9d54a DLM |
189 | { |
190 | .name = "cmdprio_class", | |
191 | .lname = "Asynchronous I/O priority class", | |
192 | .type = FIO_OPT_UNSUPPORTED, | |
193 | .help = "Your platform does not support I/O priority classes", | |
194 | }, | |
195 | { | |
196 | .name = "cmdprio", | |
197 | .lname = "Asynchronous I/O priority level", | |
198 | .type = FIO_OPT_UNSUPPORTED, | |
199 | .help = "Your platform does not support I/O priority classes", | |
200 | }, | |
a48f0cc7 DLM |
201 | { |
202 | .name = "cmdprio_bssplit", | |
203 | .lname = "Priority percentage block size split", | |
204 | .type = FIO_OPT_UNSUPPORTED, | |
205 | .help = "Your platform does not support I/O priority classes", | |
206 | }, | |
b2a432bf | 207 | #endif |
52885fa2 JA |
208 | { |
209 | .name = "fixedbufs", | |
210 | .lname = "Fixed (pre-mapped) IO buffers", | |
211 | .type = FIO_OPT_STR_SET, | |
bffad86f | 212 | .off1 = offsetof(struct ioring_options, fixedbufs), |
52885fa2 JA |
213 | .help = "Pre map IO buffers", |
214 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 215 | .group = FIO_OPT_G_IOURING, |
52885fa2 | 216 | }, |
5ffd5626 JA |
217 | { |
218 | .name = "registerfiles", | |
219 | .lname = "Register file set", | |
220 | .type = FIO_OPT_STR_SET, | |
221 | .off1 = offsetof(struct ioring_options, registerfiles), | |
222 | .help = "Pre-open/register files", | |
223 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 224 | .group = FIO_OPT_G_IOURING, |
5ffd5626 | 225 | }, |
771c9901 JA |
226 | { |
227 | .name = "sqthread_poll", | |
3d7d00a3 JA |
228 | .lname = "Kernel SQ thread polling", |
229 | .type = FIO_OPT_INT, | |
230 | .off1 = offsetof(struct ioring_options, sqpoll_thread), | |
231 | .help = "Offload submission/completion to kernel thread", | |
232 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 233 | .group = FIO_OPT_G_IOURING, |
3d7d00a3 JA |
234 | }, |
235 | { | |
236 | .name = "sqthread_poll_cpu", | |
237 | .lname = "SQ Thread Poll CPU", | |
2ea53ca3 JA |
238 | .type = FIO_OPT_INT, |
239 | .cb = fio_ioring_sqpoll_cb, | |
3d7d00a3 | 240 | .help = "What CPU to run SQ thread polling on", |
a90cd050 | 241 | .category = FIO_OPT_C_ENGINE, |
27f436d9 | 242 | .group = FIO_OPT_G_IOURING, |
a90cd050 | 243 | }, |
b10b1e70 JA |
244 | { |
245 | .name = "nonvectored", | |
246 | .lname = "Non-vectored", | |
247 | .type = FIO_OPT_INT, | |
248 | .off1 = offsetof(struct ioring_options, nonvectored), | |
556d8415 | 249 | .def = "-1", |
b10b1e70 JA |
250 | .help = "Use non-vectored read/write commands", |
251 | .category = FIO_OPT_C_ENGINE, | |
252 | .group = FIO_OPT_G_IOURING, | |
253 | }, | |
4a87b584 JA |
254 | { |
255 | .name = "uncached", | |
256 | .lname = "Uncached", | |
257 | .type = FIO_OPT_INT, | |
258 | .off1 = offsetof(struct ioring_options, uncached), | |
259 | .help = "Use RWF_UNCACHED for buffered read/writes", | |
260 | .category = FIO_OPT_C_ENGINE, | |
261 | .group = FIO_OPT_G_IOURING, | |
262 | }, | |
7d42e66e KK |
263 | { |
264 | .name = "nowait", | |
265 | .lname = "RWF_NOWAIT", | |
266 | .type = FIO_OPT_BOOL, | |
267 | .off1 = offsetof(struct ioring_options, nowait), | |
268 | .help = "Use RWF_NOWAIT for reads/writes", | |
269 | .category = FIO_OPT_C_ENGINE, | |
270 | .group = FIO_OPT_G_IOURING, | |
271 | }, | |
5a59a81d JA |
272 | { |
273 | .name = "force_async", | |
274 | .lname = "Force async", | |
275 | .type = FIO_OPT_INT, | |
276 | .off1 = offsetof(struct ioring_options, force_async), | |
277 | .help = "Set IOSQE_ASYNC every N requests", | |
278 | .category = FIO_OPT_C_ENGINE, | |
279 | .group = FIO_OPT_G_IOURING, | |
280 | }, | |
855dc4d4 AG |
281 | { |
282 | .name = "cmd_type", | |
283 | .lname = "Uring cmd type", | |
284 | .type = FIO_OPT_STR, | |
285 | .off1 = offsetof(struct ioring_options, cmd_type), | |
286 | .help = "Specify uring-cmd type", | |
287 | .def = "nvme", | |
288 | .posval = { | |
289 | { .ival = "nvme", | |
290 | .oval = FIO_URING_CMD_NVME, | |
291 | .help = "Issue nvme-uring-cmd", | |
292 | }, | |
293 | }, | |
294 | .category = FIO_OPT_C_ENGINE, | |
295 | .group = FIO_OPT_G_IOURING, | |
296 | }, | |
52885fa2 JA |
297 | { |
298 | .name = NULL, | |
299 | }, | |
300 | }; | |
301 | ||
bffad86f | 302 | static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit, |
52885fa2 JA |
303 | unsigned int min_complete, unsigned int flags) |
304 | { | |
c377f4f8 JA |
305 | #ifdef FIO_ARCH_HAS_SYSCALL |
306 | return __do_syscall6(__NR_io_uring_enter, ld->ring_fd, to_submit, | |
307 | min_complete, flags, NULL, 0); | |
308 | #else | |
bfed648c | 309 | return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit, |
521164fa | 310 | min_complete, flags, NULL, 0); |
c377f4f8 | 311 | #endif |
52885fa2 JA |
312 | } |
313 | ||
bffad86f | 314 | static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 315 | { |
bffad86f | 316 | struct ioring_data *ld = td->io_ops_data; |
cfcc8564 | 317 | struct ioring_options *o = td->eo; |
52885fa2 | 318 | struct fio_file *f = io_u->file; |
f0403f94 | 319 | struct io_uring_sqe *sqe; |
52885fa2 | 320 | |
f0403f94 | 321 | sqe = &ld->sqes[io_u->index]; |
34d6090e | 322 | |
5ffd5626 JA |
323 | if (o->registerfiles) { |
324 | sqe->fd = f->engine_pos; | |
325 | sqe->flags = IOSQE_FIXED_FILE; | |
326 | } else { | |
327 | sqe->fd = f->fd; | |
87b69ef2 | 328 | sqe->flags = 0; |
5ffd5626 | 329 | } |
52885fa2 | 330 | |
e3970057 | 331 | if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) { |
f0403f94 | 332 | if (o->fixedbufs) { |
3f1e3af7 | 333 | sqe->opcode = fixed_ddir_to_op[io_u->ddir]; |
919850d2 | 334 | sqe->addr = (unsigned long) io_u->xfer_buf; |
f0403f94 | 335 | sqe->len = io_u->xfer_buflen; |
2ea53ca3 | 336 | sqe->buf_index = io_u->index; |
cfcc8564 | 337 | } else { |
832faaaf JA |
338 | struct iovec *iov = &ld->iovecs[io_u->index]; |
339 | ||
340 | /* | |
341 | * Update based on actual io_u, requeue could have | |
342 | * adjusted these | |
343 | */ | |
344 | iov->iov_base = io_u->xfer_buf; | |
345 | iov->iov_len = io_u->xfer_buflen; | |
346 | ||
3f1e3af7 | 347 | sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored]; |
b10b1e70 | 348 | if (o->nonvectored) { |
832faaaf JA |
349 | sqe->addr = (unsigned long) iov->iov_base; |
350 | sqe->len = iov->iov_len; | |
b10b1e70 | 351 | } else { |
832faaaf | 352 | sqe->addr = (unsigned long) iov; |
b10b1e70 JA |
353 | sqe->len = 1; |
354 | } | |
cfcc8564 | 355 | } |
fd70e361 | 356 | sqe->rw_flags = 0; |
4a87b584 | 357 | if (!td->o.odirect && o->uncached) |
fd70e361 | 358 | sqe->rw_flags |= RWF_UNCACHED; |
7d42e66e KK |
359 | if (o->nowait) |
360 | sqe->rw_flags |= RWF_NOWAIT; | |
8ff6b289 NC |
361 | |
362 | /* | |
363 | * Since io_uring can have a submission context (sqthread_poll) | |
364 | * that is different from the process context, we cannot rely on | |
365 | * the IO priority set by ioprio_set() (option prio/prioclass) | |
366 | * to be inherited. | |
367 | * td->ioprio will have the value of the "default prio", so set | |
368 | * this unconditionally. This value might get overridden by | |
ff00f247 | 369 | * fio_ioring_cmdprio_prep() if the option cmdprio_percentage or |
8ff6b289 NC |
370 | * cmdprio_bssplit is used. |
371 | */ | |
372 | sqe->ioprio = td->ioprio; | |
f0403f94 | 373 | sqe->off = io_u->offset; |
48e698fa | 374 | } else if (ddir_sync(io_u->ddir)) { |
7c70f506 | 375 | sqe->ioprio = 0; |
01387bfe AF |
376 | if (io_u->ddir == DDIR_SYNC_FILE_RANGE) { |
377 | sqe->off = f->first_write; | |
378 | sqe->len = f->last_write - f->first_write; | |
379 | sqe->sync_range_flags = td->o.sync_file_range; | |
380 | sqe->opcode = IORING_OP_SYNC_FILE_RANGE; | |
381 | } else { | |
7c70f506 JA |
382 | sqe->off = 0; |
383 | sqe->addr = 0; | |
384 | sqe->len = 0; | |
01387bfe AF |
385 | if (io_u->ddir == DDIR_DATASYNC) |
386 | sqe->fsync_flags |= IORING_FSYNC_DATASYNC; | |
387 | sqe->opcode = IORING_OP_FSYNC; | |
388 | } | |
48e698fa | 389 | } |
52885fa2 | 390 | |
5a59a81d JA |
391 | if (o->force_async && ++ld->prepped == o->force_async) { |
392 | ld->prepped = 0; | |
393 | sqe->flags |= IOSQE_ASYNC; | |
394 | } | |
395 | ||
48e698fa | 396 | sqe->user_data = (unsigned long) io_u; |
52885fa2 JA |
397 | return 0; |
398 | } | |
399 | ||
855dc4d4 AG |
400 | static int fio_ioring_cmd_prep(struct thread_data *td, struct io_u *io_u) |
401 | { | |
402 | struct ioring_data *ld = td->io_ops_data; | |
403 | struct ioring_options *o = td->eo; | |
404 | struct fio_file *f = io_u->file; | |
3ce6a3de | 405 | struct nvme_uring_cmd *cmd; |
855dc4d4 | 406 | struct io_uring_sqe *sqe; |
855dc4d4 | 407 | |
3ce6a3de JA |
408 | /* only supports nvme_uring_cmd */ |
409 | if (o->cmd_type != FIO_URING_CMD_NVME) | |
410 | return -EINVAL; | |
855dc4d4 | 411 | |
3ce6a3de | 412 | sqe = &ld->sqes[(io_u->index) << 1]; |
855dc4d4 | 413 | |
3ce6a3de JA |
414 | if (o->registerfiles) { |
415 | sqe->fd = f->engine_pos; | |
416 | sqe->flags = IOSQE_FIXED_FILE; | |
417 | } else { | |
418 | sqe->fd = f->fd; | |
419 | } | |
420 | sqe->rw_flags = 0; | |
421 | if (!td->o.odirect && o->uncached) | |
422 | sqe->rw_flags |= RWF_UNCACHED; | |
423 | if (o->nowait) | |
424 | sqe->rw_flags |= RWF_NOWAIT; | |
855dc4d4 | 425 | |
3ce6a3de JA |
426 | sqe->opcode = IORING_OP_URING_CMD; |
427 | sqe->user_data = (unsigned long) io_u; | |
428 | if (o->nonvectored) | |
429 | sqe->cmd_op = NVME_URING_CMD_IO; | |
430 | else | |
431 | sqe->cmd_op = NVME_URING_CMD_IO_VEC; | |
432 | if (o->force_async && ++ld->prepped == o->force_async) { | |
433 | ld->prepped = 0; | |
434 | sqe->flags |= IOSQE_ASYNC; | |
855dc4d4 | 435 | } |
3ce6a3de JA |
436 | |
437 | cmd = (struct nvme_uring_cmd *)sqe->cmd; | |
438 | return fio_nvme_uring_cmd_prep(cmd, io_u, | |
439 | o->nonvectored ? NULL : &ld->iovecs[io_u->index]); | |
855dc4d4 AG |
440 | } |
441 | ||
bffad86f | 442 | static struct io_u *fio_ioring_event(struct thread_data *td, int event) |
52885fa2 | 443 | { |
bffad86f | 444 | struct ioring_data *ld = td->io_ops_data; |
f0403f94 | 445 | struct io_uring_cqe *cqe; |
52885fa2 | 446 | struct io_u *io_u; |
b87aa01a | 447 | unsigned index; |
52885fa2 | 448 | |
b87aa01a | 449 | index = (event + ld->cq_ring_off) & ld->cq_ring_mask; |
52885fa2 | 450 | |
f0403f94 | 451 | cqe = &ld->cq_ring.cqes[index]; |
e3466352 | 452 | io_u = (struct io_u *) (uintptr_t) cqe->user_data; |
52885fa2 | 453 | |
f0403f94 JA |
454 | if (cqe->res != io_u->xfer_buflen) { |
455 | if (cqe->res > io_u->xfer_buflen) | |
456 | io_u->error = -cqe->res; | |
52885fa2 | 457 | else |
f0403f94 | 458 | io_u->resid = io_u->xfer_buflen - cqe->res; |
52885fa2 JA |
459 | } else |
460 | io_u->error = 0; | |
461 | ||
462 | return io_u; | |
463 | } | |
464 | ||
855dc4d4 AG |
465 | static struct io_u *fio_ioring_cmd_event(struct thread_data *td, int event) |
466 | { | |
467 | struct ioring_data *ld = td->io_ops_data; | |
468 | struct ioring_options *o = td->eo; | |
469 | struct io_uring_cqe *cqe; | |
470 | struct io_u *io_u; | |
471 | unsigned index; | |
472 | ||
473 | index = (event + ld->cq_ring_off) & ld->cq_ring_mask; | |
474 | if (o->cmd_type == FIO_URING_CMD_NVME) | |
475 | index <<= 1; | |
476 | ||
477 | cqe = &ld->cq_ring.cqes[index]; | |
478 | io_u = (struct io_u *) (uintptr_t) cqe->user_data; | |
479 | ||
480 | if (cqe->res != 0) | |
481 | io_u->error = -cqe->res; | |
482 | else | |
483 | io_u->error = 0; | |
484 | ||
485 | return io_u; | |
486 | } | |
487 | ||
bffad86f | 488 | static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events, |
52885fa2 JA |
489 | unsigned int max) |
490 | { | |
bffad86f JA |
491 | struct ioring_data *ld = td->io_ops_data; |
492 | struct io_cq_ring *ring = &ld->cq_ring; | |
e2239016 | 493 | unsigned head, reaped = 0; |
52885fa2 | 494 | |
9a2d78b3 | 495 | head = *ring->head; |
52885fa2 | 496 | do { |
9e26aff9 | 497 | if (head == atomic_load_acquire(ring->tail)) |
52885fa2 JA |
498 | break; |
499 | reaped++; | |
500 | head++; | |
52885fa2 JA |
501 | } while (reaped + events < max); |
502 | ||
76ce63dd AB |
503 | if (reaped) |
504 | atomic_store_release(ring->head, head); | |
505 | ||
52885fa2 JA |
506 | return reaped; |
507 | } | |
508 | ||
bffad86f JA |
509 | static int fio_ioring_getevents(struct thread_data *td, unsigned int min, |
510 | unsigned int max, const struct timespec *t) | |
52885fa2 | 511 | { |
bffad86f | 512 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 | 513 | unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min; |
bffad86f JA |
514 | struct ioring_options *o = td->eo; |
515 | struct io_cq_ring *ring = &ld->cq_ring; | |
b87aa01a JA |
516 | unsigned events = 0; |
517 | int r; | |
52885fa2 | 518 | |
9a2d78b3 | 519 | ld->cq_ring_off = *ring->head; |
52885fa2 | 520 | do { |
bffad86f | 521 | r = fio_ioring_cqring_reap(td, events, max); |
52885fa2 JA |
522 | if (r) { |
523 | events += r; | |
f7cbbbf8 ST |
524 | if (actual_min != 0) |
525 | actual_min -= r; | |
52885fa2 JA |
526 | continue; |
527 | } | |
528 | ||
3d7d00a3 | 529 | if (!o->sqpoll_thread) { |
9a2d78b3 JA |
530 | r = io_uring_enter(ld, 0, actual_min, |
531 | IORING_ENTER_GETEVENTS); | |
771c9901 | 532 | if (r < 0) { |
f6abd731 | 533 | if (errno == EAGAIN || errno == EINTR) |
771c9901 | 534 | continue; |
1816895b | 535 | r = -errno; |
9a2d78b3 | 536 | td_verror(td, errno, "io_uring_enter"); |
771c9901 JA |
537 | break; |
538 | } | |
52885fa2 JA |
539 | } |
540 | } while (events < min); | |
541 | ||
542 | return r < 0 ? r : events; | |
543 | } | |
544 | ||
127715b6 NC |
545 | static inline void fio_ioring_cmdprio_prep(struct thread_data *td, |
546 | struct io_u *io_u) | |
b2a432bf | 547 | { |
b2a432bf | 548 | struct ioring_data *ld = td->io_ops_data; |
d6cbeab4 | 549 | struct cmdprio *cmdprio = &ld->cmdprio; |
127715b6 NC |
550 | |
551 | if (fio_cmdprio_set_ioprio(td, cmdprio, io_u)) | |
552 | ld->sqes[io_u->index].ioprio = io_u->ioprio; | |
b2a432bf PC |
553 | } |
554 | ||
bffad86f JA |
555 | static enum fio_q_status fio_ioring_queue(struct thread_data *td, |
556 | struct io_u *io_u) | |
52885fa2 | 557 | { |
bffad86f JA |
558 | struct ioring_data *ld = td->io_ops_data; |
559 | struct io_sq_ring *ring = &ld->sq_ring; | |
52885fa2 JA |
560 | unsigned tail, next_tail; |
561 | ||
562 | fio_ro_check(td, io_u); | |
563 | ||
b87aa01a | 564 | if (ld->queued == ld->iodepth) |
52885fa2 JA |
565 | return FIO_Q_BUSY; |
566 | ||
52885fa2 JA |
567 | if (io_u->ddir == DDIR_TRIM) { |
568 | if (ld->queued) | |
569 | return FIO_Q_BUSY; | |
570 | ||
571 | do_io_u_trim(td, io_u); | |
572 | io_u_mark_submit(td, 1); | |
573 | io_u_mark_complete(td, 1); | |
574 | return FIO_Q_COMPLETED; | |
575 | } | |
576 | ||
9a2d78b3 | 577 | tail = *ring->tail; |
52885fa2 | 578 | next_tail = tail + 1; |
9e26aff9 | 579 | if (next_tail == atomic_load_acquire(ring->head)) |
52885fa2 JA |
580 | return FIO_Q_BUSY; |
581 | ||
d6cbeab4 | 582 | if (ld->cmdprio.mode != CMDPRIO_MODE_NONE) |
ff00f247 NC |
583 | fio_ioring_cmdprio_prep(td, io_u); |
584 | ||
b87aa01a | 585 | ring->array[tail & ld->sq_ring_mask] = io_u->index; |
9e26aff9 | 586 | atomic_store_release(ring->tail, next_tail); |
52885fa2 JA |
587 | |
588 | ld->queued++; | |
589 | return FIO_Q_QUEUED; | |
590 | } | |
591 | ||
bffad86f | 592 | static void fio_ioring_queued(struct thread_data *td, int start, int nr) |
52885fa2 | 593 | { |
bffad86f | 594 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
595 | struct timespec now; |
596 | ||
597 | if (!fio_fill_issue_time(td)) | |
598 | return; | |
599 | ||
600 | fio_gettime(&now, NULL); | |
601 | ||
602 | while (nr--) { | |
bffad86f | 603 | struct io_sq_ring *ring = &ld->sq_ring; |
9a2d78b3 | 604 | int index = ring->array[start & ld->sq_ring_mask]; |
f8289afc | 605 | struct io_u *io_u = ld->io_u_index[index]; |
52885fa2 JA |
606 | |
607 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
608 | io_u_queued(td, io_u); | |
609 | ||
610 | start++; | |
52885fa2 | 611 | } |
39f56400 VF |
612 | |
613 | /* | |
614 | * only used for iolog | |
615 | */ | |
616 | if (td->o.read_iolog_file) | |
617 | memcpy(&td->last_issue, &now, sizeof(now)); | |
52885fa2 JA |
618 | } |
619 | ||
bffad86f | 620 | static int fio_ioring_commit(struct thread_data *td) |
52885fa2 | 621 | { |
bffad86f JA |
622 | struct ioring_data *ld = td->io_ops_data; |
623 | struct ioring_options *o = td->eo; | |
52885fa2 JA |
624 | int ret; |
625 | ||
626 | if (!ld->queued) | |
627 | return 0; | |
628 | ||
3d7d00a3 JA |
629 | /* |
630 | * Kernel side does submission. just need to check if the ring is | |
631 | * flagged as needing a kick, if so, call io_uring_enter(). This | |
632 | * only happens if we've been idle too long. | |
633 | */ | |
634 | if (o->sqpoll_thread) { | |
bffad86f | 635 | struct io_sq_ring *ring = &ld->sq_ring; |
2dd96cc4 | 636 | unsigned flags; |
4cdbc048 | 637 | |
2dd96cc4 JA |
638 | flags = atomic_load_acquire(ring->flags); |
639 | if (flags & IORING_SQ_NEED_WAKEUP) | |
b532dd6d JA |
640 | io_uring_enter(ld, ld->queued, 0, |
641 | IORING_ENTER_SQ_WAKEUP); | |
771c9901 JA |
642 | ld->queued = 0; |
643 | return 0; | |
644 | } | |
645 | ||
52885fa2 | 646 | do { |
9a2d78b3 | 647 | unsigned start = *ld->sq_ring.head; |
52885fa2 JA |
648 | long nr = ld->queued; |
649 | ||
9a2d78b3 | 650 | ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS); |
52885fa2 | 651 | if (ret > 0) { |
bffad86f | 652 | fio_ioring_queued(td, start, ret); |
52885fa2 JA |
653 | io_u_mark_submit(td, ret); |
654 | ||
655 | ld->queued -= ret; | |
656 | ret = 0; | |
a90cd050 JA |
657 | } else if (!ret) { |
658 | io_u_mark_submit(td, ret); | |
52885fa2 | 659 | continue; |
a90cd050 | 660 | } else { |
f6abd731 | 661 | if (errno == EAGAIN || errno == EINTR) { |
bffad86f | 662 | ret = fio_ioring_cqring_reap(td, 0, ld->queued); |
a90cd050 JA |
663 | if (ret) |
664 | continue; | |
665 | /* Shouldn't happen */ | |
666 | usleep(1); | |
667 | continue; | |
52885fa2 | 668 | } |
1816895b | 669 | ret = -errno; |
9a2d78b3 | 670 | td_verror(td, errno, "io_uring_enter submit"); |
52885fa2 | 671 | break; |
a90cd050 | 672 | } |
52885fa2 JA |
673 | } while (ld->queued); |
674 | ||
675 | return ret; | |
676 | } | |
677 | ||
bffad86f | 678 | static void fio_ioring_unmap(struct ioring_data *ld) |
52885fa2 | 679 | { |
9a2d78b3 | 680 | int i; |
52885fa2 | 681 | |
59f94d26 | 682 | for (i = 0; i < FIO_ARRAY_SIZE(ld->mmap); i++) |
9a2d78b3 JA |
683 | munmap(ld->mmap[i].ptr, ld->mmap[i].len); |
684 | close(ld->ring_fd); | |
b87aa01a JA |
685 | } |
686 | ||
bffad86f | 687 | static void fio_ioring_cleanup(struct thread_data *td) |
52885fa2 | 688 | { |
bffad86f | 689 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
690 | |
691 | if (ld) { | |
52885fa2 | 692 | if (!(td->flags & TD_F_CHILD)) |
bffad86f | 693 | fio_ioring_unmap(ld); |
9a2d78b3 | 694 | |
d6cbeab4 | 695 | fio_cmdprio_cleanup(&ld->cmdprio); |
52885fa2 | 696 | free(ld->io_u_index); |
9a2d78b3 | 697 | free(ld->iovecs); |
5ffd5626 | 698 | free(ld->fds); |
52885fa2 JA |
699 | free(ld); |
700 | } | |
701 | } | |
702 | ||
bffad86f | 703 | static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p) |
9a2d78b3 | 704 | { |
bffad86f JA |
705 | struct io_sq_ring *sring = &ld->sq_ring; |
706 | struct io_cq_ring *cring = &ld->cq_ring; | |
9a2d78b3 JA |
707 | void *ptr; |
708 | ||
e2239016 | 709 | ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32); |
9a2d78b3 JA |
710 | ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE, |
711 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
712 | IORING_OFF_SQ_RING); | |
713 | ld->mmap[0].ptr = ptr; | |
714 | sring->head = ptr + p->sq_off.head; | |
715 | sring->tail = ptr + p->sq_off.tail; | |
716 | sring->ring_mask = ptr + p->sq_off.ring_mask; | |
717 | sring->ring_entries = ptr + p->sq_off.ring_entries; | |
718 | sring->flags = ptr + p->sq_off.flags; | |
ac122fea | 719 | sring->array = ptr + p->sq_off.array; |
9a2d78b3 JA |
720 | ld->sq_ring_mask = *sring->ring_mask; |
721 | ||
855dc4d4 AG |
722 | if (p->flags & IORING_SETUP_SQE128) |
723 | ld->mmap[1].len = 2 * p->sq_entries * sizeof(struct io_uring_sqe); | |
724 | else | |
725 | ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe); | |
f0403f94 | 726 | ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE, |
9a2d78b3 | 727 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, |
f0403f94 JA |
728 | IORING_OFF_SQES); |
729 | ld->mmap[1].ptr = ld->sqes; | |
9a2d78b3 | 730 | |
855dc4d4 AG |
731 | if (p->flags & IORING_SETUP_CQE32) { |
732 | ld->mmap[2].len = p->cq_off.cqes + | |
733 | 2 * p->cq_entries * sizeof(struct io_uring_cqe); | |
734 | } else { | |
735 | ld->mmap[2].len = p->cq_off.cqes + | |
736 | p->cq_entries * sizeof(struct io_uring_cqe); | |
737 | } | |
9a2d78b3 JA |
738 | ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE, |
739 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
740 | IORING_OFF_CQ_RING); | |
741 | ld->mmap[2].ptr = ptr; | |
742 | cring->head = ptr + p->cq_off.head; | |
743 | cring->tail = ptr + p->cq_off.tail; | |
744 | cring->ring_mask = ptr + p->cq_off.ring_mask; | |
745 | cring->ring_entries = ptr + p->cq_off.ring_entries; | |
f0403f94 | 746 | cring->cqes = ptr + p->cq_off.cqes; |
9a2d78b3 JA |
747 | ld->cq_ring_mask = *cring->ring_mask; |
748 | return 0; | |
749 | } | |
750 | ||
556d8415 JA |
751 | static void fio_ioring_probe(struct thread_data *td) |
752 | { | |
753 | struct ioring_data *ld = td->io_ops_data; | |
754 | struct ioring_options *o = td->eo; | |
755 | struct io_uring_probe *p; | |
756 | int ret; | |
757 | ||
758 | /* already set by user, don't touch */ | |
759 | if (o->nonvectored != -1) | |
760 | return; | |
761 | ||
762 | /* default to off, as that's always safe */ | |
763 | o->nonvectored = 0; | |
764 | ||
765 | p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op)); | |
766 | if (!p) | |
767 | return; | |
768 | ||
769 | memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op)); | |
770 | ret = syscall(__NR_io_uring_register, ld->ring_fd, | |
771 | IORING_REGISTER_PROBE, p, 256); | |
772 | if (ret < 0) | |
773 | goto out; | |
774 | ||
775 | if (IORING_OP_WRITE > p->ops_len) | |
776 | goto out; | |
777 | ||
778 | if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED) && | |
779 | (p->ops[IORING_OP_WRITE].flags & IO_URING_OP_SUPPORTED)) | |
780 | o->nonvectored = 1; | |
781 | out: | |
782 | free(p); | |
783 | } | |
784 | ||
bffad86f | 785 | static int fio_ioring_queue_init(struct thread_data *td) |
52885fa2 | 786 | { |
bffad86f JA |
787 | struct ioring_data *ld = td->io_ops_data; |
788 | struct ioring_options *o = td->eo; | |
52885fa2 | 789 | int depth = td->o.iodepth; |
bffad86f | 790 | struct io_uring_params p; |
9a2d78b3 JA |
791 | int ret; |
792 | ||
793 | memset(&p, 0, sizeof(p)); | |
52885fa2 JA |
794 | |
795 | if (o->hipri) | |
bffad86f | 796 | p.flags |= IORING_SETUP_IOPOLL; |
3d7d00a3 JA |
797 | if (o->sqpoll_thread) { |
798 | p.flags |= IORING_SETUP_SQPOLL; | |
799 | if (o->sqpoll_set) { | |
800 | p.flags |= IORING_SETUP_SQ_AFF; | |
801 | p.sq_thread_cpu = o->sqpoll_cpu; | |
802 | } | |
f635f1fb | 803 | } |
a90cd050 | 804 | |
1db268db JA |
805 | /* |
806 | * Clamp CQ ring size at our SQ ring size, we don't need more entries | |
807 | * than that. | |
808 | */ | |
809 | p.flags |= IORING_SETUP_CQSIZE; | |
810 | p.cq_entries = depth; | |
811 | ||
4d22c103 JA |
812 | /* |
813 | * Setup COOP_TASKRUN as we don't need to get IPI interrupted for | |
814 | * completing IO operations. | |
815 | */ | |
816 | p.flags |= IORING_SETUP_COOP_TASKRUN; | |
817 | ||
e453f369 JA |
818 | /* |
819 | * io_uring is always a single issuer, and we can defer task_work | |
820 | * runs until we reap events. | |
821 | */ | |
822 | p.flags |= IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN; | |
823 | ||
b5e99df6 | 824 | retry: |
bfed648c | 825 | ret = syscall(__NR_io_uring_setup, depth, &p); |
b5e99df6 | 826 | if (ret < 0) { |
e453f369 JA |
827 | if (errno == EINVAL && p.flags & IORING_SETUP_DEFER_TASKRUN) { |
828 | p.flags &= ~IORING_SETUP_DEFER_TASKRUN; | |
829 | p.flags &= ~IORING_SETUP_SINGLE_ISSUER; | |
830 | goto retry; | |
831 | } | |
4d22c103 JA |
832 | if (errno == EINVAL && p.flags & IORING_SETUP_COOP_TASKRUN) { |
833 | p.flags &= ~IORING_SETUP_COOP_TASKRUN; | |
834 | goto retry; | |
835 | } | |
b5e99df6 JA |
836 | if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) { |
837 | p.flags &= ~IORING_SETUP_CQSIZE; | |
838 | goto retry; | |
839 | } | |
9a2d78b3 | 840 | return ret; |
b5e99df6 | 841 | } |
9a2d78b3 JA |
842 | |
843 | ld->ring_fd = ret; | |
2ea53ca3 | 844 | |
556d8415 JA |
845 | fio_ioring_probe(td); |
846 | ||
2ea53ca3 | 847 | if (o->fixedbufs) { |
bfed648c | 848 | ret = syscall(__NR_io_uring_register, ld->ring_fd, |
919850d2 | 849 | IORING_REGISTER_BUFFERS, ld->iovecs, depth); |
2ea53ca3 JA |
850 | if (ret < 0) |
851 | return ret; | |
852 | } | |
853 | ||
bffad86f | 854 | return fio_ioring_mmap(ld, &p); |
52885fa2 JA |
855 | } |
856 | ||
855dc4d4 AG |
857 | static int fio_ioring_cmd_queue_init(struct thread_data *td) |
858 | { | |
859 | struct ioring_data *ld = td->io_ops_data; | |
860 | struct ioring_options *o = td->eo; | |
861 | int depth = td->o.iodepth; | |
862 | struct io_uring_params p; | |
863 | int ret; | |
864 | ||
865 | memset(&p, 0, sizeof(p)); | |
866 | ||
867 | if (o->hipri) | |
868 | p.flags |= IORING_SETUP_IOPOLL; | |
869 | if (o->sqpoll_thread) { | |
870 | p.flags |= IORING_SETUP_SQPOLL; | |
871 | if (o->sqpoll_set) { | |
872 | p.flags |= IORING_SETUP_SQ_AFF; | |
873 | p.sq_thread_cpu = o->sqpoll_cpu; | |
874 | } | |
875 | } | |
876 | if (o->cmd_type == FIO_URING_CMD_NVME) { | |
877 | p.flags |= IORING_SETUP_SQE128; | |
878 | p.flags |= IORING_SETUP_CQE32; | |
879 | } | |
880 | ||
881 | /* | |
882 | * Clamp CQ ring size at our SQ ring size, we don't need more entries | |
883 | * than that. | |
884 | */ | |
885 | p.flags |= IORING_SETUP_CQSIZE; | |
886 | p.cq_entries = depth; | |
887 | ||
888 | retry: | |
889 | ret = syscall(__NR_io_uring_setup, depth, &p); | |
890 | if (ret < 0) { | |
891 | if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) { | |
892 | p.flags &= ~IORING_SETUP_CQSIZE; | |
893 | goto retry; | |
894 | } | |
895 | return ret; | |
896 | } | |
897 | ||
898 | ld->ring_fd = ret; | |
899 | ||
900 | fio_ioring_probe(td); | |
901 | ||
902 | if (o->fixedbufs) { | |
903 | ret = syscall(__NR_io_uring_register, ld->ring_fd, | |
904 | IORING_REGISTER_BUFFERS, ld->iovecs, depth); | |
905 | if (ret < 0) | |
906 | return ret; | |
907 | } | |
908 | ||
909 | return fio_ioring_mmap(ld, &p); | |
910 | } | |
911 | ||
5ffd5626 JA |
912 | static int fio_ioring_register_files(struct thread_data *td) |
913 | { | |
914 | struct ioring_data *ld = td->io_ops_data; | |
915 | struct fio_file *f; | |
916 | unsigned int i; | |
917 | int ret; | |
918 | ||
919 | ld->fds = calloc(td->o.nr_files, sizeof(int)); | |
920 | ||
921 | for_each_file(td, f, i) { | |
922 | ret = generic_open_file(td, f); | |
923 | if (ret) | |
924 | goto err; | |
925 | ld->fds[i] = f->fd; | |
926 | f->engine_pos = i; | |
927 | } | |
928 | ||
bfed648c | 929 | ret = syscall(__NR_io_uring_register, ld->ring_fd, |
5ffd5626 JA |
930 | IORING_REGISTER_FILES, ld->fds, td->o.nr_files); |
931 | if (ret) { | |
932 | err: | |
933 | free(ld->fds); | |
934 | ld->fds = NULL; | |
935 | } | |
936 | ||
937 | /* | |
938 | * Pretend the file is closed again, and really close it if we hit | |
939 | * an error. | |
940 | */ | |
941 | for_each_file(td, f, i) { | |
942 | if (ret) { | |
943 | int fio_unused ret2; | |
944 | ret2 = generic_close_file(td, f); | |
945 | } else | |
946 | f->fd = -1; | |
947 | } | |
948 | ||
949 | return ret; | |
950 | } | |
951 | ||
bffad86f | 952 | static int fio_ioring_post_init(struct thread_data *td) |
52885fa2 | 953 | { |
bffad86f | 954 | struct ioring_data *ld = td->io_ops_data; |
5ffd5626 | 955 | struct ioring_options *o = td->eo; |
52885fa2 | 956 | struct io_u *io_u; |
650346e1 | 957 | int err, i; |
52885fa2 | 958 | |
650346e1 JA |
959 | for (i = 0; i < td->o.iodepth; i++) { |
960 | struct iovec *iov = &ld->iovecs[i]; | |
9a2d78b3 | 961 | |
650346e1 JA |
962 | io_u = ld->io_u_index[i]; |
963 | iov->iov_base = io_u->buf; | |
964 | iov->iov_len = td_max_bs(td); | |
52885fa2 JA |
965 | } |
966 | ||
bffad86f | 967 | err = fio_ioring_queue_init(td); |
52885fa2 | 968 | if (err) { |
0442b53f | 969 | int init_err = errno; |
c4f5c92f | 970 | |
0442b53f | 971 | if (init_err == ENOSYS) |
c4f5c92f | 972 | log_err("fio: your kernel doesn't support io_uring\n"); |
0442b53f | 973 | td_verror(td, init_err, "io_queue_init"); |
52885fa2 JA |
974 | return 1; |
975 | } | |
976 | ||
7c70f506 JA |
977 | for (i = 0; i < td->o.iodepth; i++) { |
978 | struct io_uring_sqe *sqe; | |
979 | ||
980 | sqe = &ld->sqes[i]; | |
981 | memset(sqe, 0, sizeof(*sqe)); | |
982 | } | |
983 | ||
5ffd5626 JA |
984 | if (o->registerfiles) { |
985 | err = fio_ioring_register_files(td); | |
986 | if (err) { | |
987 | td_verror(td, errno, "ioring_register_files"); | |
988 | return 1; | |
989 | } | |
990 | } | |
991 | ||
52885fa2 JA |
992 | return 0; |
993 | } | |
994 | ||
855dc4d4 AG |
995 | static int fio_ioring_cmd_post_init(struct thread_data *td) |
996 | { | |
997 | struct ioring_data *ld = td->io_ops_data; | |
998 | struct ioring_options *o = td->eo; | |
999 | struct io_u *io_u; | |
1000 | int err, i; | |
1001 | ||
1002 | for (i = 0; i < td->o.iodepth; i++) { | |
1003 | struct iovec *iov = &ld->iovecs[i]; | |
1004 | ||
1005 | io_u = ld->io_u_index[i]; | |
1006 | iov->iov_base = io_u->buf; | |
1007 | iov->iov_len = td_max_bs(td); | |
1008 | } | |
1009 | ||
1010 | err = fio_ioring_cmd_queue_init(td); | |
1011 | if (err) { | |
1012 | int init_err = errno; | |
1013 | ||
1014 | td_verror(td, init_err, "io_queue_init"); | |
1015 | return 1; | |
1016 | } | |
1017 | ||
1018 | for (i = 0; i < td->o.iodepth; i++) { | |
1019 | struct io_uring_sqe *sqe; | |
1020 | ||
1021 | if (o->cmd_type == FIO_URING_CMD_NVME) { | |
1022 | sqe = &ld->sqes[i << 1]; | |
1023 | memset(sqe, 0, 2 * sizeof(*sqe)); | |
1024 | } else { | |
1025 | sqe = &ld->sqes[i]; | |
1026 | memset(sqe, 0, sizeof(*sqe)); | |
1027 | } | |
1028 | } | |
1029 | ||
1030 | if (o->registerfiles) { | |
1031 | err = fio_ioring_register_files(td); | |
1032 | if (err) { | |
1033 | td_verror(td, errno, "ioring_register_files"); | |
1034 | return 1; | |
1035 | } | |
1036 | } | |
1037 | ||
1038 | return 0; | |
1039 | } | |
1040 | ||
bffad86f | 1041 | static int fio_ioring_init(struct thread_data *td) |
52885fa2 | 1042 | { |
5ffd5626 | 1043 | struct ioring_options *o = td->eo; |
bffad86f | 1044 | struct ioring_data *ld; |
e9f6567a | 1045 | int ret; |
52885fa2 | 1046 | |
5ffd5626 JA |
1047 | /* sqthread submission requires registered files */ |
1048 | if (o->sqpoll_thread) | |
1049 | o->registerfiles = 1; | |
1050 | ||
1051 | if (o->registerfiles && td->o.nr_files != td->o.open_files) { | |
1052 | log_err("fio: io_uring registered files require nr_files to " | |
1053 | "be identical to open_files\n"); | |
1054 | return 1; | |
1055 | } | |
1056 | ||
52885fa2 JA |
1057 | ld = calloc(1, sizeof(*ld)); |
1058 | ||
b87aa01a JA |
1059 | /* ring depth must be a power-of-2 */ |
1060 | ld->iodepth = td->o.iodepth; | |
1061 | td->o.iodepth = roundup_pow2(td->o.iodepth); | |
1062 | ||
52885fa2 JA |
1063 | /* io_u index */ |
1064 | ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *)); | |
650346e1 | 1065 | ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec)); |
52885fa2 JA |
1066 | |
1067 | td->io_ops_data = ld; | |
b2a432bf | 1068 | |
d6cbeab4 | 1069 | ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options); |
e9f6567a DLM |
1070 | if (ret) { |
1071 | td_verror(td, EINVAL, "fio_ioring_init"); | |
b2a432bf PC |
1072 | return 1; |
1073 | } | |
1af44196 | 1074 | |
52885fa2 JA |
1075 | return 0; |
1076 | } | |
1077 | ||
bffad86f | 1078 | static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 1079 | { |
bffad86f | 1080 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
1081 | |
1082 | ld->io_u_index[io_u->index] = io_u; | |
1083 | return 0; | |
1084 | } | |
1085 | ||
5ffd5626 JA |
1086 | static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f) |
1087 | { | |
1088 | struct ioring_data *ld = td->io_ops_data; | |
1089 | struct ioring_options *o = td->eo; | |
1090 | ||
17318cf6 | 1091 | if (!ld || !o->registerfiles) |
5ffd5626 JA |
1092 | return generic_open_file(td, f); |
1093 | ||
1094 | f->fd = ld->fds[f->engine_pos]; | |
1095 | return 0; | |
1096 | } | |
1097 | ||
855dc4d4 AG |
1098 | static int fio_ioring_cmd_open_file(struct thread_data *td, struct fio_file *f) |
1099 | { | |
1100 | struct ioring_data *ld = td->io_ops_data; | |
1101 | struct ioring_options *o = td->eo; | |
1102 | ||
1103 | if (o->cmd_type == FIO_URING_CMD_NVME) { | |
1104 | struct nvme_data *data = NULL; | |
1105 | unsigned int nsid, lba_size = 0; | |
1106 | unsigned long long nlba = 0; | |
1107 | int ret; | |
1108 | ||
1109 | /* Store the namespace-id and lba size. */ | |
1110 | data = FILE_ENG_DATA(f); | |
1111 | if (data == NULL) { | |
1112 | ret = fio_nvme_get_info(f, &nsid, &lba_size, &nlba); | |
1113 | if (ret) | |
1114 | return ret; | |
1115 | ||
1116 | data = calloc(1, sizeof(struct nvme_data)); | |
1117 | data->nsid = nsid; | |
1118 | data->lba_shift = ilog2(lba_size); | |
1119 | ||
1120 | FILE_SET_ENG_DATA(f, data); | |
1121 | } | |
1122 | } | |
1123 | if (!ld || !o->registerfiles) | |
1124 | return generic_open_file(td, f); | |
1125 | ||
1126 | f->fd = ld->fds[f->engine_pos]; | |
1127 | return 0; | |
1128 | } | |
1129 | ||
5ffd5626 JA |
1130 | static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f) |
1131 | { | |
17318cf6 | 1132 | struct ioring_data *ld = td->io_ops_data; |
5ffd5626 JA |
1133 | struct ioring_options *o = td->eo; |
1134 | ||
17318cf6 | 1135 | if (!ld || !o->registerfiles) |
5ffd5626 JA |
1136 | return generic_close_file(td, f); |
1137 | ||
1138 | f->fd = -1; | |
1139 | return 0; | |
1140 | } | |
1141 | ||
855dc4d4 AG |
1142 | static int fio_ioring_cmd_close_file(struct thread_data *td, |
1143 | struct fio_file *f) | |
1144 | { | |
1145 | struct ioring_data *ld = td->io_ops_data; | |
1146 | struct ioring_options *o = td->eo; | |
1147 | ||
1148 | if (o->cmd_type == FIO_URING_CMD_NVME) { | |
1149 | struct nvme_data *data = FILE_ENG_DATA(f); | |
1150 | ||
1151 | FILE_SET_ENG_DATA(f, NULL); | |
1152 | free(data); | |
1153 | } | |
1154 | if (!ld || !o->registerfiles) | |
1155 | return generic_close_file(td, f); | |
1156 | ||
1157 | f->fd = -1; | |
1158 | return 0; | |
1159 | } | |
1160 | ||
1161 | static int fio_ioring_cmd_get_file_size(struct thread_data *td, | |
1162 | struct fio_file *f) | |
1163 | { | |
1164 | struct ioring_options *o = td->eo; | |
1165 | ||
1166 | if (fio_file_size_known(f)) | |
1167 | return 0; | |
1168 | ||
1169 | if (o->cmd_type == FIO_URING_CMD_NVME) { | |
1170 | struct nvme_data *data = NULL; | |
1171 | unsigned int nsid, lba_size = 0; | |
1172 | unsigned long long nlba = 0; | |
1173 | int ret; | |
1174 | ||
1175 | ret = fio_nvme_get_info(f, &nsid, &lba_size, &nlba); | |
1176 | if (ret) | |
1177 | return ret; | |
1178 | ||
1179 | data = calloc(1, sizeof(struct nvme_data)); | |
1180 | data->nsid = nsid; | |
1181 | data->lba_shift = ilog2(lba_size); | |
1182 | ||
1183 | f->real_file_size = lba_size * nlba; | |
1184 | fio_file_set_size_known(f); | |
1185 | ||
1186 | FILE_SET_ENG_DATA(f, data); | |
1187 | return 0; | |
1188 | } | |
1189 | return generic_get_file_size(td, f); | |
1190 | } | |
1191 | ||
3d05e0ff AK |
1192 | static int fio_ioring_cmd_get_zoned_model(struct thread_data *td, |
1193 | struct fio_file *f, | |
1194 | enum zbd_zoned_model *model) | |
1195 | { | |
1196 | return fio_nvme_get_zoned_model(td, f, model); | |
1197 | } | |
1198 | ||
1199 | static int fio_ioring_cmd_report_zones(struct thread_data *td, | |
1200 | struct fio_file *f, uint64_t offset, | |
1201 | struct zbd_zone *zbdz, | |
1202 | unsigned int nr_zones) | |
1203 | { | |
1204 | return fio_nvme_report_zones(td, f, offset, zbdz, nr_zones); | |
1205 | } | |
1206 | ||
1207 | static int fio_ioring_cmd_reset_wp(struct thread_data *td, struct fio_file *f, | |
1208 | uint64_t offset, uint64_t length) | |
1209 | { | |
1210 | return fio_nvme_reset_wp(td, f, offset, length); | |
1211 | } | |
1212 | ||
1213 | static int fio_ioring_cmd_get_max_open_zones(struct thread_data *td, | |
1214 | struct fio_file *f, | |
1215 | unsigned int *max_open_zones) | |
1216 | { | |
1217 | return fio_nvme_get_max_open_zones(td, f, max_open_zones); | |
1218 | } | |
1219 | ||
855dc4d4 | 1220 | static struct ioengine_ops ioengine_uring = { |
bffad86f | 1221 | .name = "io_uring", |
52885fa2 | 1222 | .version = FIO_IOOPS_VERSION, |
4e7e7898 VF |
1223 | .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD | |
1224 | FIO_ASYNCIO_SETS_ISSUE_TIME, | |
bffad86f JA |
1225 | .init = fio_ioring_init, |
1226 | .post_init = fio_ioring_post_init, | |
1227 | .io_u_init = fio_ioring_io_u_init, | |
1228 | .prep = fio_ioring_prep, | |
1229 | .queue = fio_ioring_queue, | |
1230 | .commit = fio_ioring_commit, | |
1231 | .getevents = fio_ioring_getevents, | |
1232 | .event = fio_ioring_event, | |
1233 | .cleanup = fio_ioring_cleanup, | |
5ffd5626 JA |
1234 | .open_file = fio_ioring_open_file, |
1235 | .close_file = fio_ioring_close_file, | |
52885fa2 JA |
1236 | .get_file_size = generic_get_file_size, |
1237 | .options = options, | |
bffad86f | 1238 | .option_struct_size = sizeof(struct ioring_options), |
52885fa2 JA |
1239 | }; |
1240 | ||
855dc4d4 AG |
1241 | static struct ioengine_ops ioengine_uring_cmd = { |
1242 | .name = "io_uring_cmd", | |
1243 | .version = FIO_IOOPS_VERSION, | |
4e7e7898 VF |
1244 | .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD | |
1245 | FIO_MEMALIGN | FIO_RAWIO | | |
1246 | FIO_ASYNCIO_SETS_ISSUE_TIME, | |
855dc4d4 AG |
1247 | .init = fio_ioring_init, |
1248 | .post_init = fio_ioring_cmd_post_init, | |
1249 | .io_u_init = fio_ioring_io_u_init, | |
1250 | .prep = fio_ioring_cmd_prep, | |
1251 | .queue = fio_ioring_queue, | |
1252 | .commit = fio_ioring_commit, | |
1253 | .getevents = fio_ioring_getevents, | |
1254 | .event = fio_ioring_cmd_event, | |
1255 | .cleanup = fio_ioring_cleanup, | |
1256 | .open_file = fio_ioring_cmd_open_file, | |
1257 | .close_file = fio_ioring_cmd_close_file, | |
1258 | .get_file_size = fio_ioring_cmd_get_file_size, | |
3d05e0ff AK |
1259 | .get_zoned_model = fio_ioring_cmd_get_zoned_model, |
1260 | .report_zones = fio_ioring_cmd_report_zones, | |
1261 | .reset_wp = fio_ioring_cmd_reset_wp, | |
1262 | .get_max_open_zones = fio_ioring_cmd_get_max_open_zones, | |
855dc4d4 AG |
1263 | .options = options, |
1264 | .option_struct_size = sizeof(struct ioring_options), | |
1265 | }; | |
1266 | ||
bffad86f | 1267 | static void fio_init fio_ioring_register(void) |
52885fa2 | 1268 | { |
855dc4d4 AG |
1269 | register_ioengine(&ioengine_uring); |
1270 | register_ioengine(&ioengine_uring_cmd); | |
52885fa2 JA |
1271 | } |
1272 | ||
bffad86f | 1273 | static void fio_exit fio_ioring_unregister(void) |
52885fa2 | 1274 | { |
855dc4d4 AG |
1275 | unregister_ioengine(&ioengine_uring); |
1276 | unregister_ioengine(&ioengine_uring_cmd); | |
52885fa2 | 1277 | } |
1f90e9bb | 1278 | #endif |