Commit | Line | Data |
---|---|---|
52885fa2 | 1 | /* |
bffad86f | 2 | * io_uring engine |
52885fa2 | 3 | * |
bffad86f | 4 | * IO engine using the new native Linux aio io_uring interface. See: |
a90cd050 | 5 | * |
bffad86f | 6 | * http://git.kernel.dk/cgit/linux-block/log/?h=io_uring |
52885fa2 JA |
7 | * |
8 | */ | |
9 | #include <stdlib.h> | |
10 | #include <unistd.h> | |
11 | #include <errno.h> | |
52885fa2 JA |
12 | #include <sys/time.h> |
13 | #include <sys/resource.h> | |
14 | ||
15 | #include "../fio.h" | |
16 | #include "../lib/pow2.h" | |
17 | #include "../optgroup.h" | |
18 | #include "../lib/memalign.h" | |
b87aa01a | 19 | #include "../lib/fls.h" |
52885fa2 | 20 | |
bffad86f | 21 | #ifdef ARCH_HAVE_IOURING |
52885fa2 | 22 | |
57fa61f0 | 23 | #include "../lib/types.h" |
f3e769a4 | 24 | #include "../os/linux/io_uring.h" |
9a2d78b3 | 25 | |
bffad86f | 26 | struct io_sq_ring { |
e2239016 JA |
27 | unsigned *head; |
28 | unsigned *tail; | |
29 | unsigned *ring_mask; | |
30 | unsigned *ring_entries; | |
31 | unsigned *flags; | |
32 | unsigned *array; | |
52885fa2 JA |
33 | }; |
34 | ||
bffad86f | 35 | struct io_cq_ring { |
e2239016 JA |
36 | unsigned *head; |
37 | unsigned *tail; | |
38 | unsigned *ring_mask; | |
39 | unsigned *ring_entries; | |
f0403f94 | 40 | struct io_uring_cqe *cqes; |
9a2d78b3 JA |
41 | }; |
42 | ||
bffad86f | 43 | struct ioring_mmap { |
9a2d78b3 JA |
44 | void *ptr; |
45 | size_t len; | |
52885fa2 JA |
46 | }; |
47 | ||
bffad86f | 48 | struct ioring_data { |
9a2d78b3 JA |
49 | int ring_fd; |
50 | ||
52885fa2 JA |
51 | struct io_u **io_u_index; |
52 | ||
5ffd5626 JA |
53 | int *fds; |
54 | ||
bffad86f | 55 | struct io_sq_ring sq_ring; |
f0403f94 | 56 | struct io_uring_sqe *sqes; |
9a2d78b3 | 57 | struct iovec *iovecs; |
b87aa01a | 58 | unsigned sq_ring_mask; |
52885fa2 | 59 | |
bffad86f | 60 | struct io_cq_ring cq_ring; |
b87aa01a | 61 | unsigned cq_ring_mask; |
52885fa2 JA |
62 | |
63 | int queued; | |
64 | int cq_ring_off; | |
b87aa01a | 65 | unsigned iodepth; |
1af44196 XW |
66 | bool ioprio_class_set; |
67 | bool ioprio_set; | |
96563db9 | 68 | |
bffad86f | 69 | struct ioring_mmap mmap[3]; |
52885fa2 JA |
70 | }; |
71 | ||
bffad86f | 72 | struct ioring_options { |
52885fa2 JA |
73 | void *pad; |
74 | unsigned int hipri; | |
b2a432bf | 75 | unsigned int cmdprio_percentage; |
52885fa2 | 76 | unsigned int fixedbufs; |
5ffd5626 | 77 | unsigned int registerfiles; |
3d7d00a3 | 78 | unsigned int sqpoll_thread; |
2ea53ca3 JA |
79 | unsigned int sqpoll_set; |
80 | unsigned int sqpoll_cpu; | |
b10b1e70 | 81 | unsigned int nonvectored; |
4a87b584 | 82 | unsigned int uncached; |
52885fa2 JA |
83 | }; |
84 | ||
b10b1e70 JA |
85 | static const int ddir_to_op[2][2] = { |
86 | { IORING_OP_READV, IORING_OP_READ }, | |
87 | { IORING_OP_WRITEV, IORING_OP_WRITE } | |
88 | }; | |
89 | ||
3f1e3af7 KB |
90 | static const int fixed_ddir_to_op[2] = { |
91 | IORING_OP_READ_FIXED, | |
92 | IORING_OP_WRITE_FIXED | |
93 | }; | |
94 | ||
2ea53ca3 | 95 | static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val) |
a90cd050 | 96 | { |
bffad86f | 97 | struct ioring_options *o = data; |
a90cd050 | 98 | |
2ea53ca3 JA |
99 | o->sqpoll_cpu = *val; |
100 | o->sqpoll_set = 1; | |
a90cd050 JA |
101 | return 0; |
102 | } | |
103 | ||
52885fa2 JA |
104 | static struct fio_option options[] = { |
105 | { | |
106 | .name = "hipri", | |
107 | .lname = "High Priority", | |
108 | .type = FIO_OPT_STR_SET, | |
bffad86f | 109 | .off1 = offsetof(struct ioring_options, hipri), |
52885fa2 JA |
110 | .help = "Use polled IO completions", |
111 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 112 | .group = FIO_OPT_G_IOURING, |
52885fa2 | 113 | }, |
b2a432bf PC |
114 | #ifdef FIO_HAVE_IOPRIO_CLASS |
115 | { | |
116 | .name = "cmdprio_percentage", | |
117 | .lname = "high priority percentage", | |
118 | .type = FIO_OPT_INT, | |
119 | .off1 = offsetof(struct ioring_options, cmdprio_percentage), | |
120 | .minval = 1, | |
121 | .maxval = 100, | |
122 | .help = "Send high priority I/O this percentage of the time", | |
123 | .category = FIO_OPT_C_ENGINE, | |
124 | .group = FIO_OPT_G_IOURING, | |
125 | }, | |
126 | #else | |
127 | { | |
128 | .name = "cmdprio_percentage", | |
129 | .lname = "high priority percentage", | |
130 | .type = FIO_OPT_UNSUPPORTED, | |
131 | .help = "Your platform does not support I/O priority classes", | |
132 | }, | |
133 | #endif | |
52885fa2 JA |
134 | { |
135 | .name = "fixedbufs", | |
136 | .lname = "Fixed (pre-mapped) IO buffers", | |
137 | .type = FIO_OPT_STR_SET, | |
bffad86f | 138 | .off1 = offsetof(struct ioring_options, fixedbufs), |
52885fa2 JA |
139 | .help = "Pre map IO buffers", |
140 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 141 | .group = FIO_OPT_G_IOURING, |
52885fa2 | 142 | }, |
5ffd5626 JA |
143 | { |
144 | .name = "registerfiles", | |
145 | .lname = "Register file set", | |
146 | .type = FIO_OPT_STR_SET, | |
147 | .off1 = offsetof(struct ioring_options, registerfiles), | |
148 | .help = "Pre-open/register files", | |
149 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 150 | .group = FIO_OPT_G_IOURING, |
5ffd5626 | 151 | }, |
771c9901 JA |
152 | { |
153 | .name = "sqthread_poll", | |
3d7d00a3 JA |
154 | .lname = "Kernel SQ thread polling", |
155 | .type = FIO_OPT_INT, | |
156 | .off1 = offsetof(struct ioring_options, sqpoll_thread), | |
157 | .help = "Offload submission/completion to kernel thread", | |
158 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 159 | .group = FIO_OPT_G_IOURING, |
3d7d00a3 JA |
160 | }, |
161 | { | |
162 | .name = "sqthread_poll_cpu", | |
163 | .lname = "SQ Thread Poll CPU", | |
2ea53ca3 JA |
164 | .type = FIO_OPT_INT, |
165 | .cb = fio_ioring_sqpoll_cb, | |
3d7d00a3 | 166 | .help = "What CPU to run SQ thread polling on", |
a90cd050 | 167 | .category = FIO_OPT_C_ENGINE, |
27f436d9 | 168 | .group = FIO_OPT_G_IOURING, |
a90cd050 | 169 | }, |
b10b1e70 JA |
170 | { |
171 | .name = "nonvectored", | |
172 | .lname = "Non-vectored", | |
173 | .type = FIO_OPT_INT, | |
174 | .off1 = offsetof(struct ioring_options, nonvectored), | |
175 | .help = "Use non-vectored read/write commands", | |
176 | .category = FIO_OPT_C_ENGINE, | |
177 | .group = FIO_OPT_G_IOURING, | |
178 | }, | |
4a87b584 JA |
179 | { |
180 | .name = "uncached", | |
181 | .lname = "Uncached", | |
182 | .type = FIO_OPT_INT, | |
183 | .off1 = offsetof(struct ioring_options, uncached), | |
184 | .help = "Use RWF_UNCACHED for buffered read/writes", | |
185 | .category = FIO_OPT_C_ENGINE, | |
186 | .group = FIO_OPT_G_IOURING, | |
187 | }, | |
52885fa2 JA |
188 | { |
189 | .name = NULL, | |
190 | }, | |
191 | }; | |
192 | ||
bffad86f | 193 | static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit, |
52885fa2 JA |
194 | unsigned int min_complete, unsigned int flags) |
195 | { | |
bfed648c | 196 | return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit, |
521164fa | 197 | min_complete, flags, NULL, 0); |
52885fa2 JA |
198 | } |
199 | ||
bffad86f | 200 | static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 201 | { |
bffad86f | 202 | struct ioring_data *ld = td->io_ops_data; |
cfcc8564 | 203 | struct ioring_options *o = td->eo; |
52885fa2 | 204 | struct fio_file *f = io_u->file; |
f0403f94 | 205 | struct io_uring_sqe *sqe; |
52885fa2 | 206 | |
f0403f94 | 207 | sqe = &ld->sqes[io_u->index]; |
34d6090e AF |
208 | |
209 | /* zero out fields not used in this submission */ | |
210 | memset(sqe, 0, sizeof(*sqe)); | |
211 | ||
5ffd5626 JA |
212 | if (o->registerfiles) { |
213 | sqe->fd = f->engine_pos; | |
214 | sqe->flags = IOSQE_FIXED_FILE; | |
215 | } else { | |
216 | sqe->fd = f->fd; | |
5ffd5626 | 217 | } |
52885fa2 | 218 | |
e3970057 | 219 | if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) { |
f0403f94 | 220 | if (o->fixedbufs) { |
3f1e3af7 | 221 | sqe->opcode = fixed_ddir_to_op[io_u->ddir]; |
919850d2 | 222 | sqe->addr = (unsigned long) io_u->xfer_buf; |
f0403f94 | 223 | sqe->len = io_u->xfer_buflen; |
2ea53ca3 | 224 | sqe->buf_index = io_u->index; |
cfcc8564 | 225 | } else { |
3f1e3af7 | 226 | sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored]; |
b10b1e70 JA |
227 | if (o->nonvectored) { |
228 | sqe->addr = (unsigned long) | |
229 | ld->iovecs[io_u->index].iov_base; | |
230 | sqe->len = ld->iovecs[io_u->index].iov_len; | |
231 | } else { | |
232 | sqe->addr = (unsigned long) &ld->iovecs[io_u->index]; | |
233 | sqe->len = 1; | |
234 | } | |
cfcc8564 | 235 | } |
4a87b584 JA |
236 | if (!td->o.odirect && o->uncached) |
237 | sqe->rw_flags = RWF_UNCACHED; | |
1af44196 | 238 | if (ld->ioprio_class_set) |
b7ed2a86 | 239 | sqe->ioprio = td->o.ioprio_class << 13; |
1af44196 | 240 | if (ld->ioprio_set) |
b7ed2a86 | 241 | sqe->ioprio |= td->o.ioprio; |
f0403f94 | 242 | sqe->off = io_u->offset; |
48e698fa | 243 | } else if (ddir_sync(io_u->ddir)) { |
01387bfe AF |
244 | if (io_u->ddir == DDIR_SYNC_FILE_RANGE) { |
245 | sqe->off = f->first_write; | |
246 | sqe->len = f->last_write - f->first_write; | |
247 | sqe->sync_range_flags = td->o.sync_file_range; | |
248 | sqe->opcode = IORING_OP_SYNC_FILE_RANGE; | |
249 | } else { | |
01387bfe AF |
250 | if (io_u->ddir == DDIR_DATASYNC) |
251 | sqe->fsync_flags |= IORING_FSYNC_DATASYNC; | |
252 | sqe->opcode = IORING_OP_FSYNC; | |
253 | } | |
48e698fa | 254 | } |
52885fa2 | 255 | |
48e698fa | 256 | sqe->user_data = (unsigned long) io_u; |
52885fa2 JA |
257 | return 0; |
258 | } | |
259 | ||
bffad86f | 260 | static struct io_u *fio_ioring_event(struct thread_data *td, int event) |
52885fa2 | 261 | { |
bffad86f | 262 | struct ioring_data *ld = td->io_ops_data; |
f0403f94 | 263 | struct io_uring_cqe *cqe; |
52885fa2 | 264 | struct io_u *io_u; |
b87aa01a | 265 | unsigned index; |
52885fa2 | 266 | |
b87aa01a | 267 | index = (event + ld->cq_ring_off) & ld->cq_ring_mask; |
52885fa2 | 268 | |
f0403f94 | 269 | cqe = &ld->cq_ring.cqes[index]; |
e3466352 | 270 | io_u = (struct io_u *) (uintptr_t) cqe->user_data; |
52885fa2 | 271 | |
f0403f94 JA |
272 | if (cqe->res != io_u->xfer_buflen) { |
273 | if (cqe->res > io_u->xfer_buflen) | |
274 | io_u->error = -cqe->res; | |
52885fa2 | 275 | else |
f0403f94 | 276 | io_u->resid = io_u->xfer_buflen - cqe->res; |
52885fa2 JA |
277 | } else |
278 | io_u->error = 0; | |
279 | ||
280 | return io_u; | |
281 | } | |
282 | ||
bffad86f | 283 | static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events, |
52885fa2 JA |
284 | unsigned int max) |
285 | { | |
bffad86f JA |
286 | struct ioring_data *ld = td->io_ops_data; |
287 | struct io_cq_ring *ring = &ld->cq_ring; | |
e2239016 | 288 | unsigned head, reaped = 0; |
52885fa2 | 289 | |
9a2d78b3 | 290 | head = *ring->head; |
52885fa2 JA |
291 | do { |
292 | read_barrier(); | |
9a2d78b3 | 293 | if (head == *ring->tail) |
52885fa2 JA |
294 | break; |
295 | reaped++; | |
296 | head++; | |
52885fa2 JA |
297 | } while (reaped + events < max); |
298 | ||
9a2d78b3 | 299 | *ring->head = head; |
52885fa2 JA |
300 | write_barrier(); |
301 | return reaped; | |
302 | } | |
303 | ||
bffad86f JA |
304 | static int fio_ioring_getevents(struct thread_data *td, unsigned int min, |
305 | unsigned int max, const struct timespec *t) | |
52885fa2 | 306 | { |
bffad86f | 307 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 | 308 | unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min; |
bffad86f JA |
309 | struct ioring_options *o = td->eo; |
310 | struct io_cq_ring *ring = &ld->cq_ring; | |
b87aa01a JA |
311 | unsigned events = 0; |
312 | int r; | |
52885fa2 | 313 | |
9a2d78b3 | 314 | ld->cq_ring_off = *ring->head; |
52885fa2 | 315 | do { |
bffad86f | 316 | r = fio_ioring_cqring_reap(td, events, max); |
52885fa2 JA |
317 | if (r) { |
318 | events += r; | |
f7cbbbf8 ST |
319 | if (actual_min != 0) |
320 | actual_min -= r; | |
52885fa2 JA |
321 | continue; |
322 | } | |
323 | ||
3d7d00a3 | 324 | if (!o->sqpoll_thread) { |
9a2d78b3 JA |
325 | r = io_uring_enter(ld, 0, actual_min, |
326 | IORING_ENTER_GETEVENTS); | |
771c9901 | 327 | if (r < 0) { |
f6abd731 | 328 | if (errno == EAGAIN || errno == EINTR) |
771c9901 | 329 | continue; |
9a2d78b3 | 330 | td_verror(td, errno, "io_uring_enter"); |
771c9901 JA |
331 | break; |
332 | } | |
52885fa2 JA |
333 | } |
334 | } while (events < min); | |
335 | ||
336 | return r < 0 ? r : events; | |
337 | } | |
338 | ||
b2a432bf PC |
339 | static void fio_ioring_prio_prep(struct thread_data *td, struct io_u *io_u) |
340 | { | |
341 | struct ioring_options *o = td->eo; | |
342 | struct ioring_data *ld = td->io_ops_data; | |
343 | if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) { | |
344 | ld->sqes[io_u->index].ioprio = IOPRIO_CLASS_RT << IOPRIO_CLASS_SHIFT; | |
345 | io_u->flags |= IO_U_F_PRIORITY; | |
346 | } | |
347 | return; | |
348 | } | |
349 | ||
bffad86f JA |
350 | static enum fio_q_status fio_ioring_queue(struct thread_data *td, |
351 | struct io_u *io_u) | |
52885fa2 | 352 | { |
bffad86f JA |
353 | struct ioring_data *ld = td->io_ops_data; |
354 | struct io_sq_ring *ring = &ld->sq_ring; | |
b2a432bf | 355 | struct ioring_options *o = td->eo; |
52885fa2 JA |
356 | unsigned tail, next_tail; |
357 | ||
358 | fio_ro_check(td, io_u); | |
359 | ||
b87aa01a | 360 | if (ld->queued == ld->iodepth) |
52885fa2 JA |
361 | return FIO_Q_BUSY; |
362 | ||
52885fa2 JA |
363 | if (io_u->ddir == DDIR_TRIM) { |
364 | if (ld->queued) | |
365 | return FIO_Q_BUSY; | |
366 | ||
367 | do_io_u_trim(td, io_u); | |
368 | io_u_mark_submit(td, 1); | |
369 | io_u_mark_complete(td, 1); | |
370 | return FIO_Q_COMPLETED; | |
371 | } | |
372 | ||
9a2d78b3 | 373 | tail = *ring->tail; |
52885fa2 | 374 | next_tail = tail + 1; |
52885fa2 | 375 | read_barrier(); |
9a2d78b3 | 376 | if (next_tail == *ring->head) |
52885fa2 JA |
377 | return FIO_Q_BUSY; |
378 | ||
b2a432bf PC |
379 | if (o->cmdprio_percentage) |
380 | fio_ioring_prio_prep(td, io_u); | |
b87aa01a | 381 | ring->array[tail & ld->sq_ring_mask] = io_u->index; |
9a2d78b3 | 382 | *ring->tail = next_tail; |
52885fa2 JA |
383 | write_barrier(); |
384 | ||
385 | ld->queued++; | |
386 | return FIO_Q_QUEUED; | |
387 | } | |
388 | ||
bffad86f | 389 | static void fio_ioring_queued(struct thread_data *td, int start, int nr) |
52885fa2 | 390 | { |
bffad86f | 391 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
392 | struct timespec now; |
393 | ||
394 | if (!fio_fill_issue_time(td)) | |
395 | return; | |
396 | ||
397 | fio_gettime(&now, NULL); | |
398 | ||
399 | while (nr--) { | |
bffad86f | 400 | struct io_sq_ring *ring = &ld->sq_ring; |
9a2d78b3 | 401 | int index = ring->array[start & ld->sq_ring_mask]; |
f8289afc | 402 | struct io_u *io_u = ld->io_u_index[index]; |
52885fa2 JA |
403 | |
404 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
405 | io_u_queued(td, io_u); | |
406 | ||
407 | start++; | |
52885fa2 JA |
408 | } |
409 | } | |
410 | ||
bffad86f | 411 | static int fio_ioring_commit(struct thread_data *td) |
52885fa2 | 412 | { |
bffad86f JA |
413 | struct ioring_data *ld = td->io_ops_data; |
414 | struct ioring_options *o = td->eo; | |
52885fa2 JA |
415 | int ret; |
416 | ||
417 | if (!ld->queued) | |
418 | return 0; | |
419 | ||
3d7d00a3 JA |
420 | /* |
421 | * Kernel side does submission. just need to check if the ring is | |
422 | * flagged as needing a kick, if so, call io_uring_enter(). This | |
423 | * only happens if we've been idle too long. | |
424 | */ | |
425 | if (o->sqpoll_thread) { | |
bffad86f | 426 | struct io_sq_ring *ring = &ld->sq_ring; |
4cdbc048 | 427 | |
2ea53ca3 | 428 | read_barrier(); |
9a2d78b3 | 429 | if (*ring->flags & IORING_SQ_NEED_WAKEUP) |
b532dd6d JA |
430 | io_uring_enter(ld, ld->queued, 0, |
431 | IORING_ENTER_SQ_WAKEUP); | |
771c9901 JA |
432 | ld->queued = 0; |
433 | return 0; | |
434 | } | |
435 | ||
52885fa2 | 436 | do { |
9a2d78b3 | 437 | unsigned start = *ld->sq_ring.head; |
52885fa2 JA |
438 | long nr = ld->queued; |
439 | ||
9a2d78b3 | 440 | ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS); |
52885fa2 | 441 | if (ret > 0) { |
bffad86f | 442 | fio_ioring_queued(td, start, ret); |
52885fa2 JA |
443 | io_u_mark_submit(td, ret); |
444 | ||
445 | ld->queued -= ret; | |
446 | ret = 0; | |
a90cd050 JA |
447 | } else if (!ret) { |
448 | io_u_mark_submit(td, ret); | |
52885fa2 | 449 | continue; |
a90cd050 | 450 | } else { |
f6abd731 | 451 | if (errno == EAGAIN || errno == EINTR) { |
bffad86f | 452 | ret = fio_ioring_cqring_reap(td, 0, ld->queued); |
a90cd050 JA |
453 | if (ret) |
454 | continue; | |
455 | /* Shouldn't happen */ | |
456 | usleep(1); | |
457 | continue; | |
52885fa2 | 458 | } |
9a2d78b3 | 459 | td_verror(td, errno, "io_uring_enter submit"); |
52885fa2 | 460 | break; |
a90cd050 | 461 | } |
52885fa2 JA |
462 | } while (ld->queued); |
463 | ||
464 | return ret; | |
465 | } | |
466 | ||
bffad86f | 467 | static void fio_ioring_unmap(struct ioring_data *ld) |
52885fa2 | 468 | { |
9a2d78b3 | 469 | int i; |
52885fa2 | 470 | |
9a2d78b3 JA |
471 | for (i = 0; i < ARRAY_SIZE(ld->mmap); i++) |
472 | munmap(ld->mmap[i].ptr, ld->mmap[i].len); | |
473 | close(ld->ring_fd); | |
b87aa01a JA |
474 | } |
475 | ||
bffad86f | 476 | static void fio_ioring_cleanup(struct thread_data *td) |
52885fa2 | 477 | { |
bffad86f | 478 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
479 | |
480 | if (ld) { | |
52885fa2 | 481 | if (!(td->flags & TD_F_CHILD)) |
bffad86f | 482 | fio_ioring_unmap(ld); |
9a2d78b3 | 483 | |
52885fa2 | 484 | free(ld->io_u_index); |
9a2d78b3 | 485 | free(ld->iovecs); |
5ffd5626 | 486 | free(ld->fds); |
52885fa2 JA |
487 | free(ld); |
488 | } | |
489 | } | |
490 | ||
bffad86f | 491 | static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p) |
9a2d78b3 | 492 | { |
bffad86f JA |
493 | struct io_sq_ring *sring = &ld->sq_ring; |
494 | struct io_cq_ring *cring = &ld->cq_ring; | |
9a2d78b3 JA |
495 | void *ptr; |
496 | ||
e2239016 | 497 | ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32); |
9a2d78b3 JA |
498 | ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE, |
499 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
500 | IORING_OFF_SQ_RING); | |
501 | ld->mmap[0].ptr = ptr; | |
502 | sring->head = ptr + p->sq_off.head; | |
503 | sring->tail = ptr + p->sq_off.tail; | |
504 | sring->ring_mask = ptr + p->sq_off.ring_mask; | |
505 | sring->ring_entries = ptr + p->sq_off.ring_entries; | |
506 | sring->flags = ptr + p->sq_off.flags; | |
ac122fea | 507 | sring->array = ptr + p->sq_off.array; |
9a2d78b3 JA |
508 | ld->sq_ring_mask = *sring->ring_mask; |
509 | ||
f0403f94 JA |
510 | ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe); |
511 | ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE, | |
9a2d78b3 | 512 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, |
f0403f94 JA |
513 | IORING_OFF_SQES); |
514 | ld->mmap[1].ptr = ld->sqes; | |
9a2d78b3 | 515 | |
f0403f94 JA |
516 | ld->mmap[2].len = p->cq_off.cqes + |
517 | p->cq_entries * sizeof(struct io_uring_cqe); | |
9a2d78b3 JA |
518 | ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE, |
519 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
520 | IORING_OFF_CQ_RING); | |
521 | ld->mmap[2].ptr = ptr; | |
522 | cring->head = ptr + p->cq_off.head; | |
523 | cring->tail = ptr + p->cq_off.tail; | |
524 | cring->ring_mask = ptr + p->cq_off.ring_mask; | |
525 | cring->ring_entries = ptr + p->cq_off.ring_entries; | |
f0403f94 | 526 | cring->cqes = ptr + p->cq_off.cqes; |
9a2d78b3 JA |
527 | ld->cq_ring_mask = *cring->ring_mask; |
528 | return 0; | |
529 | } | |
530 | ||
bffad86f | 531 | static int fio_ioring_queue_init(struct thread_data *td) |
52885fa2 | 532 | { |
bffad86f JA |
533 | struct ioring_data *ld = td->io_ops_data; |
534 | struct ioring_options *o = td->eo; | |
52885fa2 | 535 | int depth = td->o.iodepth; |
bffad86f | 536 | struct io_uring_params p; |
9a2d78b3 JA |
537 | int ret; |
538 | ||
539 | memset(&p, 0, sizeof(p)); | |
52885fa2 JA |
540 | |
541 | if (o->hipri) | |
bffad86f | 542 | p.flags |= IORING_SETUP_IOPOLL; |
3d7d00a3 JA |
543 | if (o->sqpoll_thread) { |
544 | p.flags |= IORING_SETUP_SQPOLL; | |
545 | if (o->sqpoll_set) { | |
546 | p.flags |= IORING_SETUP_SQ_AFF; | |
547 | p.sq_thread_cpu = o->sqpoll_cpu; | |
548 | } | |
f635f1fb | 549 | } |
a90cd050 | 550 | |
bfed648c | 551 | ret = syscall(__NR_io_uring_setup, depth, &p); |
9a2d78b3 JA |
552 | if (ret < 0) |
553 | return ret; | |
554 | ||
555 | ld->ring_fd = ret; | |
2ea53ca3 JA |
556 | |
557 | if (o->fixedbufs) { | |
2d644205 JA |
558 | struct rlimit rlim = { |
559 | .rlim_cur = RLIM_INFINITY, | |
560 | .rlim_max = RLIM_INFINITY, | |
561 | }; | |
562 | ||
563 | if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) | |
564 | return -1; | |
565 | ||
bfed648c | 566 | ret = syscall(__NR_io_uring_register, ld->ring_fd, |
919850d2 | 567 | IORING_REGISTER_BUFFERS, ld->iovecs, depth); |
2ea53ca3 JA |
568 | if (ret < 0) |
569 | return ret; | |
570 | } | |
571 | ||
bffad86f | 572 | return fio_ioring_mmap(ld, &p); |
52885fa2 JA |
573 | } |
574 | ||
5ffd5626 JA |
575 | static int fio_ioring_register_files(struct thread_data *td) |
576 | { | |
577 | struct ioring_data *ld = td->io_ops_data; | |
578 | struct fio_file *f; | |
579 | unsigned int i; | |
580 | int ret; | |
581 | ||
582 | ld->fds = calloc(td->o.nr_files, sizeof(int)); | |
583 | ||
584 | for_each_file(td, f, i) { | |
585 | ret = generic_open_file(td, f); | |
586 | if (ret) | |
587 | goto err; | |
588 | ld->fds[i] = f->fd; | |
589 | f->engine_pos = i; | |
590 | } | |
591 | ||
bfed648c | 592 | ret = syscall(__NR_io_uring_register, ld->ring_fd, |
5ffd5626 JA |
593 | IORING_REGISTER_FILES, ld->fds, td->o.nr_files); |
594 | if (ret) { | |
595 | err: | |
596 | free(ld->fds); | |
597 | ld->fds = NULL; | |
598 | } | |
599 | ||
600 | /* | |
601 | * Pretend the file is closed again, and really close it if we hit | |
602 | * an error. | |
603 | */ | |
604 | for_each_file(td, f, i) { | |
605 | if (ret) { | |
606 | int fio_unused ret2; | |
607 | ret2 = generic_close_file(td, f); | |
608 | } else | |
609 | f->fd = -1; | |
610 | } | |
611 | ||
612 | return ret; | |
613 | } | |
614 | ||
bffad86f | 615 | static int fio_ioring_post_init(struct thread_data *td) |
52885fa2 | 616 | { |
bffad86f | 617 | struct ioring_data *ld = td->io_ops_data; |
5ffd5626 | 618 | struct ioring_options *o = td->eo; |
52885fa2 | 619 | struct io_u *io_u; |
650346e1 | 620 | int err, i; |
52885fa2 | 621 | |
650346e1 JA |
622 | for (i = 0; i < td->o.iodepth; i++) { |
623 | struct iovec *iov = &ld->iovecs[i]; | |
9a2d78b3 | 624 | |
650346e1 JA |
625 | io_u = ld->io_u_index[i]; |
626 | iov->iov_base = io_u->buf; | |
627 | iov->iov_len = td_max_bs(td); | |
52885fa2 JA |
628 | } |
629 | ||
bffad86f | 630 | err = fio_ioring_queue_init(td); |
52885fa2 | 631 | if (err) { |
d63a472d | 632 | td_verror(td, errno, "io_queue_init"); |
52885fa2 JA |
633 | return 1; |
634 | } | |
635 | ||
5ffd5626 JA |
636 | if (o->registerfiles) { |
637 | err = fio_ioring_register_files(td); | |
638 | if (err) { | |
639 | td_verror(td, errno, "ioring_register_files"); | |
640 | return 1; | |
641 | } | |
642 | } | |
643 | ||
52885fa2 JA |
644 | return 0; |
645 | } | |
646 | ||
9a2d78b3 JA |
647 | static unsigned roundup_pow2(unsigned depth) |
648 | { | |
649 | return 1UL << __fls(depth - 1); | |
650 | } | |
651 | ||
bffad86f | 652 | static int fio_ioring_init(struct thread_data *td) |
52885fa2 | 653 | { |
5ffd5626 | 654 | struct ioring_options *o = td->eo; |
bffad86f | 655 | struct ioring_data *ld; |
b2a432bf | 656 | struct thread_options *to = &td->o; |
52885fa2 | 657 | |
5ffd5626 JA |
658 | /* sqthread submission requires registered files */ |
659 | if (o->sqpoll_thread) | |
660 | o->registerfiles = 1; | |
661 | ||
662 | if (o->registerfiles && td->o.nr_files != td->o.open_files) { | |
663 | log_err("fio: io_uring registered files require nr_files to " | |
664 | "be identical to open_files\n"); | |
665 | return 1; | |
666 | } | |
667 | ||
52885fa2 JA |
668 | ld = calloc(1, sizeof(*ld)); |
669 | ||
b87aa01a JA |
670 | /* ring depth must be a power-of-2 */ |
671 | ld->iodepth = td->o.iodepth; | |
672 | td->o.iodepth = roundup_pow2(td->o.iodepth); | |
673 | ||
52885fa2 JA |
674 | /* io_u index */ |
675 | ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *)); | |
650346e1 | 676 | ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec)); |
52885fa2 JA |
677 | |
678 | td->io_ops_data = ld; | |
b2a432bf PC |
679 | |
680 | /* | |
681 | * Check for option conflicts | |
682 | */ | |
683 | if ((fio_option_is_set(to, ioprio) || fio_option_is_set(to, ioprio_class)) && | |
684 | o->cmdprio_percentage != 0) { | |
685 | log_err("%s: cmdprio_percentage option and mutually exclusive " | |
686 | "prio or prioclass option is set, exiting\n", to->name); | |
687 | td_verror(td, EINVAL, "fio_io_uring_init"); | |
688 | return 1; | |
689 | } | |
1af44196 XW |
690 | |
691 | if (fio_option_is_set(&td->o, ioprio_class)) | |
692 | ld->ioprio_class_set = true; | |
693 | if (fio_option_is_set(&td->o, ioprio)) | |
694 | ld->ioprio_set = true; | |
695 | ||
52885fa2 JA |
696 | return 0; |
697 | } | |
698 | ||
bffad86f | 699 | static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 700 | { |
bffad86f | 701 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
702 | |
703 | ld->io_u_index[io_u->index] = io_u; | |
704 | return 0; | |
705 | } | |
706 | ||
5ffd5626 JA |
707 | static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f) |
708 | { | |
709 | struct ioring_data *ld = td->io_ops_data; | |
710 | struct ioring_options *o = td->eo; | |
711 | ||
17318cf6 | 712 | if (!ld || !o->registerfiles) |
5ffd5626 JA |
713 | return generic_open_file(td, f); |
714 | ||
715 | f->fd = ld->fds[f->engine_pos]; | |
716 | return 0; | |
717 | } | |
718 | ||
719 | static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f) | |
720 | { | |
17318cf6 | 721 | struct ioring_data *ld = td->io_ops_data; |
5ffd5626 JA |
722 | struct ioring_options *o = td->eo; |
723 | ||
17318cf6 | 724 | if (!ld || !o->registerfiles) |
5ffd5626 JA |
725 | return generic_close_file(td, f); |
726 | ||
727 | f->fd = -1; | |
728 | return 0; | |
729 | } | |
730 | ||
52885fa2 | 731 | static struct ioengine_ops ioengine = { |
bffad86f | 732 | .name = "io_uring", |
52885fa2 | 733 | .version = FIO_IOOPS_VERSION, |
04ba61df | 734 | .flags = FIO_ASYNCIO_SYNC_TRIM, |
bffad86f JA |
735 | .init = fio_ioring_init, |
736 | .post_init = fio_ioring_post_init, | |
737 | .io_u_init = fio_ioring_io_u_init, | |
738 | .prep = fio_ioring_prep, | |
739 | .queue = fio_ioring_queue, | |
740 | .commit = fio_ioring_commit, | |
741 | .getevents = fio_ioring_getevents, | |
742 | .event = fio_ioring_event, | |
743 | .cleanup = fio_ioring_cleanup, | |
5ffd5626 JA |
744 | .open_file = fio_ioring_open_file, |
745 | .close_file = fio_ioring_close_file, | |
52885fa2 JA |
746 | .get_file_size = generic_get_file_size, |
747 | .options = options, | |
bffad86f | 748 | .option_struct_size = sizeof(struct ioring_options), |
52885fa2 JA |
749 | }; |
750 | ||
bffad86f | 751 | static void fio_init fio_ioring_register(void) |
52885fa2 | 752 | { |
52885fa2 | 753 | register_ioengine(&ioengine); |
52885fa2 JA |
754 | } |
755 | ||
bffad86f | 756 | static void fio_exit fio_ioring_unregister(void) |
52885fa2 | 757 | { |
52885fa2 | 758 | unregister_ioengine(&ioengine); |
52885fa2 | 759 | } |
1f90e9bb | 760 | #endif |