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; |
7d42e66e | 83 | unsigned int nowait; |
52885fa2 JA |
84 | }; |
85 | ||
b10b1e70 JA |
86 | static const int ddir_to_op[2][2] = { |
87 | { IORING_OP_READV, IORING_OP_READ }, | |
88 | { IORING_OP_WRITEV, IORING_OP_WRITE } | |
89 | }; | |
90 | ||
3f1e3af7 KB |
91 | static const int fixed_ddir_to_op[2] = { |
92 | IORING_OP_READ_FIXED, | |
93 | IORING_OP_WRITE_FIXED | |
94 | }; | |
95 | ||
2ea53ca3 | 96 | static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val) |
a90cd050 | 97 | { |
bffad86f | 98 | struct ioring_options *o = data; |
a90cd050 | 99 | |
2ea53ca3 JA |
100 | o->sqpoll_cpu = *val; |
101 | o->sqpoll_set = 1; | |
a90cd050 JA |
102 | return 0; |
103 | } | |
104 | ||
52885fa2 JA |
105 | static struct fio_option options[] = { |
106 | { | |
107 | .name = "hipri", | |
108 | .lname = "High Priority", | |
109 | .type = FIO_OPT_STR_SET, | |
bffad86f | 110 | .off1 = offsetof(struct ioring_options, hipri), |
52885fa2 JA |
111 | .help = "Use polled IO completions", |
112 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 113 | .group = FIO_OPT_G_IOURING, |
52885fa2 | 114 | }, |
b2a432bf PC |
115 | #ifdef FIO_HAVE_IOPRIO_CLASS |
116 | { | |
117 | .name = "cmdprio_percentage", | |
118 | .lname = "high priority percentage", | |
119 | .type = FIO_OPT_INT, | |
120 | .off1 = offsetof(struct ioring_options, cmdprio_percentage), | |
121 | .minval = 1, | |
122 | .maxval = 100, | |
123 | .help = "Send high priority I/O this percentage of the time", | |
124 | .category = FIO_OPT_C_ENGINE, | |
125 | .group = FIO_OPT_G_IOURING, | |
126 | }, | |
127 | #else | |
128 | { | |
129 | .name = "cmdprio_percentage", | |
130 | .lname = "high priority percentage", | |
131 | .type = FIO_OPT_UNSUPPORTED, | |
132 | .help = "Your platform does not support I/O priority classes", | |
133 | }, | |
134 | #endif | |
52885fa2 JA |
135 | { |
136 | .name = "fixedbufs", | |
137 | .lname = "Fixed (pre-mapped) IO buffers", | |
138 | .type = FIO_OPT_STR_SET, | |
bffad86f | 139 | .off1 = offsetof(struct ioring_options, fixedbufs), |
52885fa2 JA |
140 | .help = "Pre map IO buffers", |
141 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 142 | .group = FIO_OPT_G_IOURING, |
52885fa2 | 143 | }, |
5ffd5626 JA |
144 | { |
145 | .name = "registerfiles", | |
146 | .lname = "Register file set", | |
147 | .type = FIO_OPT_STR_SET, | |
148 | .off1 = offsetof(struct ioring_options, registerfiles), | |
149 | .help = "Pre-open/register files", | |
150 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 151 | .group = FIO_OPT_G_IOURING, |
5ffd5626 | 152 | }, |
771c9901 JA |
153 | { |
154 | .name = "sqthread_poll", | |
3d7d00a3 JA |
155 | .lname = "Kernel SQ thread polling", |
156 | .type = FIO_OPT_INT, | |
157 | .off1 = offsetof(struct ioring_options, sqpoll_thread), | |
158 | .help = "Offload submission/completion to kernel thread", | |
159 | .category = FIO_OPT_C_ENGINE, | |
27f436d9 | 160 | .group = FIO_OPT_G_IOURING, |
3d7d00a3 JA |
161 | }, |
162 | { | |
163 | .name = "sqthread_poll_cpu", | |
164 | .lname = "SQ Thread Poll CPU", | |
2ea53ca3 JA |
165 | .type = FIO_OPT_INT, |
166 | .cb = fio_ioring_sqpoll_cb, | |
3d7d00a3 | 167 | .help = "What CPU to run SQ thread polling on", |
a90cd050 | 168 | .category = FIO_OPT_C_ENGINE, |
27f436d9 | 169 | .group = FIO_OPT_G_IOURING, |
a90cd050 | 170 | }, |
b10b1e70 JA |
171 | { |
172 | .name = "nonvectored", | |
173 | .lname = "Non-vectored", | |
174 | .type = FIO_OPT_INT, | |
175 | .off1 = offsetof(struct ioring_options, nonvectored), | |
176 | .help = "Use non-vectored read/write commands", | |
177 | .category = FIO_OPT_C_ENGINE, | |
178 | .group = FIO_OPT_G_IOURING, | |
179 | }, | |
4a87b584 JA |
180 | { |
181 | .name = "uncached", | |
182 | .lname = "Uncached", | |
183 | .type = FIO_OPT_INT, | |
184 | .off1 = offsetof(struct ioring_options, uncached), | |
185 | .help = "Use RWF_UNCACHED for buffered read/writes", | |
186 | .category = FIO_OPT_C_ENGINE, | |
187 | .group = FIO_OPT_G_IOURING, | |
188 | }, | |
7d42e66e KK |
189 | { |
190 | .name = "nowait", | |
191 | .lname = "RWF_NOWAIT", | |
192 | .type = FIO_OPT_BOOL, | |
193 | .off1 = offsetof(struct ioring_options, nowait), | |
194 | .help = "Use RWF_NOWAIT for reads/writes", | |
195 | .category = FIO_OPT_C_ENGINE, | |
196 | .group = FIO_OPT_G_IOURING, | |
197 | }, | |
52885fa2 JA |
198 | { |
199 | .name = NULL, | |
200 | }, | |
201 | }; | |
202 | ||
bffad86f | 203 | static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit, |
52885fa2 JA |
204 | unsigned int min_complete, unsigned int flags) |
205 | { | |
bfed648c | 206 | return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit, |
521164fa | 207 | min_complete, flags, NULL, 0); |
52885fa2 JA |
208 | } |
209 | ||
bffad86f | 210 | static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 211 | { |
bffad86f | 212 | struct ioring_data *ld = td->io_ops_data; |
cfcc8564 | 213 | struct ioring_options *o = td->eo; |
52885fa2 | 214 | struct fio_file *f = io_u->file; |
f0403f94 | 215 | struct io_uring_sqe *sqe; |
52885fa2 | 216 | |
f0403f94 | 217 | sqe = &ld->sqes[io_u->index]; |
34d6090e AF |
218 | |
219 | /* zero out fields not used in this submission */ | |
220 | memset(sqe, 0, sizeof(*sqe)); | |
221 | ||
5ffd5626 JA |
222 | if (o->registerfiles) { |
223 | sqe->fd = f->engine_pos; | |
224 | sqe->flags = IOSQE_FIXED_FILE; | |
225 | } else { | |
226 | sqe->fd = f->fd; | |
5ffd5626 | 227 | } |
52885fa2 | 228 | |
e3970057 | 229 | if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) { |
f0403f94 | 230 | if (o->fixedbufs) { |
3f1e3af7 | 231 | sqe->opcode = fixed_ddir_to_op[io_u->ddir]; |
919850d2 | 232 | sqe->addr = (unsigned long) io_u->xfer_buf; |
f0403f94 | 233 | sqe->len = io_u->xfer_buflen; |
2ea53ca3 | 234 | sqe->buf_index = io_u->index; |
cfcc8564 | 235 | } else { |
3f1e3af7 | 236 | sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored]; |
b10b1e70 JA |
237 | if (o->nonvectored) { |
238 | sqe->addr = (unsigned long) | |
239 | ld->iovecs[io_u->index].iov_base; | |
240 | sqe->len = ld->iovecs[io_u->index].iov_len; | |
241 | } else { | |
242 | sqe->addr = (unsigned long) &ld->iovecs[io_u->index]; | |
243 | sqe->len = 1; | |
244 | } | |
cfcc8564 | 245 | } |
4a87b584 JA |
246 | if (!td->o.odirect && o->uncached) |
247 | sqe->rw_flags = RWF_UNCACHED; | |
7d42e66e KK |
248 | if (o->nowait) |
249 | sqe->rw_flags |= RWF_NOWAIT; | |
1af44196 | 250 | if (ld->ioprio_class_set) |
b7ed2a86 | 251 | sqe->ioprio = td->o.ioprio_class << 13; |
1af44196 | 252 | if (ld->ioprio_set) |
b7ed2a86 | 253 | sqe->ioprio |= td->o.ioprio; |
f0403f94 | 254 | sqe->off = io_u->offset; |
48e698fa | 255 | } else if (ddir_sync(io_u->ddir)) { |
01387bfe AF |
256 | if (io_u->ddir == DDIR_SYNC_FILE_RANGE) { |
257 | sqe->off = f->first_write; | |
258 | sqe->len = f->last_write - f->first_write; | |
259 | sqe->sync_range_flags = td->o.sync_file_range; | |
260 | sqe->opcode = IORING_OP_SYNC_FILE_RANGE; | |
261 | } else { | |
01387bfe AF |
262 | if (io_u->ddir == DDIR_DATASYNC) |
263 | sqe->fsync_flags |= IORING_FSYNC_DATASYNC; | |
264 | sqe->opcode = IORING_OP_FSYNC; | |
265 | } | |
48e698fa | 266 | } |
52885fa2 | 267 | |
48e698fa | 268 | sqe->user_data = (unsigned long) io_u; |
52885fa2 JA |
269 | return 0; |
270 | } | |
271 | ||
bffad86f | 272 | static struct io_u *fio_ioring_event(struct thread_data *td, int event) |
52885fa2 | 273 | { |
bffad86f | 274 | struct ioring_data *ld = td->io_ops_data; |
f0403f94 | 275 | struct io_uring_cqe *cqe; |
52885fa2 | 276 | struct io_u *io_u; |
b87aa01a | 277 | unsigned index; |
52885fa2 | 278 | |
b87aa01a | 279 | index = (event + ld->cq_ring_off) & ld->cq_ring_mask; |
52885fa2 | 280 | |
f0403f94 | 281 | cqe = &ld->cq_ring.cqes[index]; |
e3466352 | 282 | io_u = (struct io_u *) (uintptr_t) cqe->user_data; |
52885fa2 | 283 | |
f0403f94 JA |
284 | if (cqe->res != io_u->xfer_buflen) { |
285 | if (cqe->res > io_u->xfer_buflen) | |
286 | io_u->error = -cqe->res; | |
52885fa2 | 287 | else |
f0403f94 | 288 | io_u->resid = io_u->xfer_buflen - cqe->res; |
52885fa2 JA |
289 | } else |
290 | io_u->error = 0; | |
291 | ||
292 | return io_u; | |
293 | } | |
294 | ||
bffad86f | 295 | static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events, |
52885fa2 JA |
296 | unsigned int max) |
297 | { | |
bffad86f JA |
298 | struct ioring_data *ld = td->io_ops_data; |
299 | struct io_cq_ring *ring = &ld->cq_ring; | |
e2239016 | 300 | unsigned head, reaped = 0; |
52885fa2 | 301 | |
9a2d78b3 | 302 | head = *ring->head; |
52885fa2 | 303 | do { |
9e26aff9 | 304 | if (head == atomic_load_acquire(ring->tail)) |
52885fa2 JA |
305 | break; |
306 | reaped++; | |
307 | head++; | |
52885fa2 JA |
308 | } while (reaped + events < max); |
309 | ||
9e26aff9 | 310 | atomic_store_release(ring->head, head); |
52885fa2 JA |
311 | return reaped; |
312 | } | |
313 | ||
bffad86f JA |
314 | static int fio_ioring_getevents(struct thread_data *td, unsigned int min, |
315 | unsigned int max, const struct timespec *t) | |
52885fa2 | 316 | { |
bffad86f | 317 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 | 318 | unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min; |
bffad86f JA |
319 | struct ioring_options *o = td->eo; |
320 | struct io_cq_ring *ring = &ld->cq_ring; | |
b87aa01a JA |
321 | unsigned events = 0; |
322 | int r; | |
52885fa2 | 323 | |
9a2d78b3 | 324 | ld->cq_ring_off = *ring->head; |
52885fa2 | 325 | do { |
bffad86f | 326 | r = fio_ioring_cqring_reap(td, events, max); |
52885fa2 JA |
327 | if (r) { |
328 | events += r; | |
f7cbbbf8 ST |
329 | if (actual_min != 0) |
330 | actual_min -= r; | |
52885fa2 JA |
331 | continue; |
332 | } | |
333 | ||
3d7d00a3 | 334 | if (!o->sqpoll_thread) { |
9a2d78b3 JA |
335 | r = io_uring_enter(ld, 0, actual_min, |
336 | IORING_ENTER_GETEVENTS); | |
771c9901 | 337 | if (r < 0) { |
f6abd731 | 338 | if (errno == EAGAIN || errno == EINTR) |
771c9901 | 339 | continue; |
9a2d78b3 | 340 | td_verror(td, errno, "io_uring_enter"); |
771c9901 JA |
341 | break; |
342 | } | |
52885fa2 JA |
343 | } |
344 | } while (events < min); | |
345 | ||
346 | return r < 0 ? r : events; | |
347 | } | |
348 | ||
b2a432bf PC |
349 | static void fio_ioring_prio_prep(struct thread_data *td, struct io_u *io_u) |
350 | { | |
351 | struct ioring_options *o = td->eo; | |
352 | struct ioring_data *ld = td->io_ops_data; | |
353 | if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) { | |
354 | ld->sqes[io_u->index].ioprio = IOPRIO_CLASS_RT << IOPRIO_CLASS_SHIFT; | |
355 | io_u->flags |= IO_U_F_PRIORITY; | |
356 | } | |
357 | return; | |
358 | } | |
359 | ||
bffad86f JA |
360 | static enum fio_q_status fio_ioring_queue(struct thread_data *td, |
361 | struct io_u *io_u) | |
52885fa2 | 362 | { |
bffad86f JA |
363 | struct ioring_data *ld = td->io_ops_data; |
364 | struct io_sq_ring *ring = &ld->sq_ring; | |
b2a432bf | 365 | struct ioring_options *o = td->eo; |
52885fa2 JA |
366 | unsigned tail, next_tail; |
367 | ||
368 | fio_ro_check(td, io_u); | |
369 | ||
b87aa01a | 370 | if (ld->queued == ld->iodepth) |
52885fa2 JA |
371 | return FIO_Q_BUSY; |
372 | ||
52885fa2 JA |
373 | if (io_u->ddir == DDIR_TRIM) { |
374 | if (ld->queued) | |
375 | return FIO_Q_BUSY; | |
376 | ||
377 | do_io_u_trim(td, io_u); | |
378 | io_u_mark_submit(td, 1); | |
379 | io_u_mark_complete(td, 1); | |
380 | return FIO_Q_COMPLETED; | |
381 | } | |
382 | ||
9a2d78b3 | 383 | tail = *ring->tail; |
52885fa2 | 384 | next_tail = tail + 1; |
9e26aff9 | 385 | if (next_tail == atomic_load_acquire(ring->head)) |
52885fa2 JA |
386 | return FIO_Q_BUSY; |
387 | ||
b2a432bf PC |
388 | if (o->cmdprio_percentage) |
389 | fio_ioring_prio_prep(td, io_u); | |
b87aa01a | 390 | ring->array[tail & ld->sq_ring_mask] = io_u->index; |
9e26aff9 | 391 | atomic_store_release(ring->tail, next_tail); |
52885fa2 JA |
392 | |
393 | ld->queued++; | |
394 | return FIO_Q_QUEUED; | |
395 | } | |
396 | ||
bffad86f | 397 | static void fio_ioring_queued(struct thread_data *td, int start, int nr) |
52885fa2 | 398 | { |
bffad86f | 399 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
400 | struct timespec now; |
401 | ||
402 | if (!fio_fill_issue_time(td)) | |
403 | return; | |
404 | ||
405 | fio_gettime(&now, NULL); | |
406 | ||
407 | while (nr--) { | |
bffad86f | 408 | struct io_sq_ring *ring = &ld->sq_ring; |
9a2d78b3 | 409 | int index = ring->array[start & ld->sq_ring_mask]; |
f8289afc | 410 | struct io_u *io_u = ld->io_u_index[index]; |
52885fa2 JA |
411 | |
412 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
413 | io_u_queued(td, io_u); | |
414 | ||
415 | start++; | |
52885fa2 JA |
416 | } |
417 | } | |
418 | ||
bffad86f | 419 | static int fio_ioring_commit(struct thread_data *td) |
52885fa2 | 420 | { |
bffad86f JA |
421 | struct ioring_data *ld = td->io_ops_data; |
422 | struct ioring_options *o = td->eo; | |
52885fa2 JA |
423 | int ret; |
424 | ||
425 | if (!ld->queued) | |
426 | return 0; | |
427 | ||
3d7d00a3 JA |
428 | /* |
429 | * Kernel side does submission. just need to check if the ring is | |
430 | * flagged as needing a kick, if so, call io_uring_enter(). This | |
431 | * only happens if we've been idle too long. | |
432 | */ | |
433 | if (o->sqpoll_thread) { | |
bffad86f | 434 | struct io_sq_ring *ring = &ld->sq_ring; |
4cdbc048 | 435 | |
2ea53ca3 | 436 | read_barrier(); |
9a2d78b3 | 437 | if (*ring->flags & IORING_SQ_NEED_WAKEUP) |
b532dd6d JA |
438 | io_uring_enter(ld, ld->queued, 0, |
439 | IORING_ENTER_SQ_WAKEUP); | |
771c9901 JA |
440 | ld->queued = 0; |
441 | return 0; | |
442 | } | |
443 | ||
52885fa2 | 444 | do { |
9a2d78b3 | 445 | unsigned start = *ld->sq_ring.head; |
52885fa2 JA |
446 | long nr = ld->queued; |
447 | ||
9a2d78b3 | 448 | ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS); |
52885fa2 | 449 | if (ret > 0) { |
bffad86f | 450 | fio_ioring_queued(td, start, ret); |
52885fa2 JA |
451 | io_u_mark_submit(td, ret); |
452 | ||
453 | ld->queued -= ret; | |
454 | ret = 0; | |
a90cd050 JA |
455 | } else if (!ret) { |
456 | io_u_mark_submit(td, ret); | |
52885fa2 | 457 | continue; |
a90cd050 | 458 | } else { |
f6abd731 | 459 | if (errno == EAGAIN || errno == EINTR) { |
bffad86f | 460 | ret = fio_ioring_cqring_reap(td, 0, ld->queued); |
a90cd050 JA |
461 | if (ret) |
462 | continue; | |
463 | /* Shouldn't happen */ | |
464 | usleep(1); | |
465 | continue; | |
52885fa2 | 466 | } |
9a2d78b3 | 467 | td_verror(td, errno, "io_uring_enter submit"); |
52885fa2 | 468 | break; |
a90cd050 | 469 | } |
52885fa2 JA |
470 | } while (ld->queued); |
471 | ||
472 | return ret; | |
473 | } | |
474 | ||
bffad86f | 475 | static void fio_ioring_unmap(struct ioring_data *ld) |
52885fa2 | 476 | { |
9a2d78b3 | 477 | int i; |
52885fa2 | 478 | |
9a2d78b3 JA |
479 | for (i = 0; i < ARRAY_SIZE(ld->mmap); i++) |
480 | munmap(ld->mmap[i].ptr, ld->mmap[i].len); | |
481 | close(ld->ring_fd); | |
b87aa01a JA |
482 | } |
483 | ||
bffad86f | 484 | static void fio_ioring_cleanup(struct thread_data *td) |
52885fa2 | 485 | { |
bffad86f | 486 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
487 | |
488 | if (ld) { | |
52885fa2 | 489 | if (!(td->flags & TD_F_CHILD)) |
bffad86f | 490 | fio_ioring_unmap(ld); |
9a2d78b3 | 491 | |
52885fa2 | 492 | free(ld->io_u_index); |
9a2d78b3 | 493 | free(ld->iovecs); |
5ffd5626 | 494 | free(ld->fds); |
52885fa2 JA |
495 | free(ld); |
496 | } | |
497 | } | |
498 | ||
bffad86f | 499 | static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p) |
9a2d78b3 | 500 | { |
bffad86f JA |
501 | struct io_sq_ring *sring = &ld->sq_ring; |
502 | struct io_cq_ring *cring = &ld->cq_ring; | |
9a2d78b3 JA |
503 | void *ptr; |
504 | ||
e2239016 | 505 | ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32); |
9a2d78b3 JA |
506 | ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE, |
507 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
508 | IORING_OFF_SQ_RING); | |
509 | ld->mmap[0].ptr = ptr; | |
510 | sring->head = ptr + p->sq_off.head; | |
511 | sring->tail = ptr + p->sq_off.tail; | |
512 | sring->ring_mask = ptr + p->sq_off.ring_mask; | |
513 | sring->ring_entries = ptr + p->sq_off.ring_entries; | |
514 | sring->flags = ptr + p->sq_off.flags; | |
ac122fea | 515 | sring->array = ptr + p->sq_off.array; |
9a2d78b3 JA |
516 | ld->sq_ring_mask = *sring->ring_mask; |
517 | ||
f0403f94 JA |
518 | ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe); |
519 | ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE, | |
9a2d78b3 | 520 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, |
f0403f94 JA |
521 | IORING_OFF_SQES); |
522 | ld->mmap[1].ptr = ld->sqes; | |
9a2d78b3 | 523 | |
f0403f94 JA |
524 | ld->mmap[2].len = p->cq_off.cqes + |
525 | p->cq_entries * sizeof(struct io_uring_cqe); | |
9a2d78b3 JA |
526 | ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE, |
527 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, | |
528 | IORING_OFF_CQ_RING); | |
529 | ld->mmap[2].ptr = ptr; | |
530 | cring->head = ptr + p->cq_off.head; | |
531 | cring->tail = ptr + p->cq_off.tail; | |
532 | cring->ring_mask = ptr + p->cq_off.ring_mask; | |
533 | cring->ring_entries = ptr + p->cq_off.ring_entries; | |
f0403f94 | 534 | cring->cqes = ptr + p->cq_off.cqes; |
9a2d78b3 JA |
535 | ld->cq_ring_mask = *cring->ring_mask; |
536 | return 0; | |
537 | } | |
538 | ||
bffad86f | 539 | static int fio_ioring_queue_init(struct thread_data *td) |
52885fa2 | 540 | { |
bffad86f JA |
541 | struct ioring_data *ld = td->io_ops_data; |
542 | struct ioring_options *o = td->eo; | |
52885fa2 | 543 | int depth = td->o.iodepth; |
bffad86f | 544 | struct io_uring_params p; |
9a2d78b3 JA |
545 | int ret; |
546 | ||
547 | memset(&p, 0, sizeof(p)); | |
52885fa2 JA |
548 | |
549 | if (o->hipri) | |
bffad86f | 550 | p.flags |= IORING_SETUP_IOPOLL; |
3d7d00a3 JA |
551 | if (o->sqpoll_thread) { |
552 | p.flags |= IORING_SETUP_SQPOLL; | |
553 | if (o->sqpoll_set) { | |
554 | p.flags |= IORING_SETUP_SQ_AFF; | |
555 | p.sq_thread_cpu = o->sqpoll_cpu; | |
556 | } | |
f635f1fb | 557 | } |
a90cd050 | 558 | |
bfed648c | 559 | ret = syscall(__NR_io_uring_setup, depth, &p); |
9a2d78b3 JA |
560 | if (ret < 0) |
561 | return ret; | |
562 | ||
563 | ld->ring_fd = ret; | |
2ea53ca3 JA |
564 | |
565 | if (o->fixedbufs) { | |
2d644205 JA |
566 | struct rlimit rlim = { |
567 | .rlim_cur = RLIM_INFINITY, | |
568 | .rlim_max = RLIM_INFINITY, | |
569 | }; | |
570 | ||
571 | if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) | |
572 | return -1; | |
573 | ||
bfed648c | 574 | ret = syscall(__NR_io_uring_register, ld->ring_fd, |
919850d2 | 575 | IORING_REGISTER_BUFFERS, ld->iovecs, depth); |
2ea53ca3 JA |
576 | if (ret < 0) |
577 | return ret; | |
578 | } | |
579 | ||
bffad86f | 580 | return fio_ioring_mmap(ld, &p); |
52885fa2 JA |
581 | } |
582 | ||
5ffd5626 JA |
583 | static int fio_ioring_register_files(struct thread_data *td) |
584 | { | |
585 | struct ioring_data *ld = td->io_ops_data; | |
586 | struct fio_file *f; | |
587 | unsigned int i; | |
588 | int ret; | |
589 | ||
590 | ld->fds = calloc(td->o.nr_files, sizeof(int)); | |
591 | ||
592 | for_each_file(td, f, i) { | |
593 | ret = generic_open_file(td, f); | |
594 | if (ret) | |
595 | goto err; | |
596 | ld->fds[i] = f->fd; | |
597 | f->engine_pos = i; | |
598 | } | |
599 | ||
bfed648c | 600 | ret = syscall(__NR_io_uring_register, ld->ring_fd, |
5ffd5626 JA |
601 | IORING_REGISTER_FILES, ld->fds, td->o.nr_files); |
602 | if (ret) { | |
603 | err: | |
604 | free(ld->fds); | |
605 | ld->fds = NULL; | |
606 | } | |
607 | ||
608 | /* | |
609 | * Pretend the file is closed again, and really close it if we hit | |
610 | * an error. | |
611 | */ | |
612 | for_each_file(td, f, i) { | |
613 | if (ret) { | |
614 | int fio_unused ret2; | |
615 | ret2 = generic_close_file(td, f); | |
616 | } else | |
617 | f->fd = -1; | |
618 | } | |
619 | ||
620 | return ret; | |
621 | } | |
622 | ||
bffad86f | 623 | static int fio_ioring_post_init(struct thread_data *td) |
52885fa2 | 624 | { |
bffad86f | 625 | struct ioring_data *ld = td->io_ops_data; |
5ffd5626 | 626 | struct ioring_options *o = td->eo; |
52885fa2 | 627 | struct io_u *io_u; |
650346e1 | 628 | int err, i; |
52885fa2 | 629 | |
650346e1 JA |
630 | for (i = 0; i < td->o.iodepth; i++) { |
631 | struct iovec *iov = &ld->iovecs[i]; | |
9a2d78b3 | 632 | |
650346e1 JA |
633 | io_u = ld->io_u_index[i]; |
634 | iov->iov_base = io_u->buf; | |
635 | iov->iov_len = td_max_bs(td); | |
52885fa2 JA |
636 | } |
637 | ||
bffad86f | 638 | err = fio_ioring_queue_init(td); |
52885fa2 | 639 | if (err) { |
d63a472d | 640 | td_verror(td, errno, "io_queue_init"); |
52885fa2 JA |
641 | return 1; |
642 | } | |
643 | ||
5ffd5626 JA |
644 | if (o->registerfiles) { |
645 | err = fio_ioring_register_files(td); | |
646 | if (err) { | |
647 | td_verror(td, errno, "ioring_register_files"); | |
648 | return 1; | |
649 | } | |
650 | } | |
651 | ||
52885fa2 JA |
652 | return 0; |
653 | } | |
654 | ||
9a2d78b3 JA |
655 | static unsigned roundup_pow2(unsigned depth) |
656 | { | |
657 | return 1UL << __fls(depth - 1); | |
658 | } | |
659 | ||
bffad86f | 660 | static int fio_ioring_init(struct thread_data *td) |
52885fa2 | 661 | { |
5ffd5626 | 662 | struct ioring_options *o = td->eo; |
bffad86f | 663 | struct ioring_data *ld; |
b2a432bf | 664 | struct thread_options *to = &td->o; |
52885fa2 | 665 | |
5ffd5626 JA |
666 | /* sqthread submission requires registered files */ |
667 | if (o->sqpoll_thread) | |
668 | o->registerfiles = 1; | |
669 | ||
670 | if (o->registerfiles && td->o.nr_files != td->o.open_files) { | |
671 | log_err("fio: io_uring registered files require nr_files to " | |
672 | "be identical to open_files\n"); | |
673 | return 1; | |
674 | } | |
675 | ||
52885fa2 JA |
676 | ld = calloc(1, sizeof(*ld)); |
677 | ||
b87aa01a JA |
678 | /* ring depth must be a power-of-2 */ |
679 | ld->iodepth = td->o.iodepth; | |
680 | td->o.iodepth = roundup_pow2(td->o.iodepth); | |
681 | ||
52885fa2 JA |
682 | /* io_u index */ |
683 | ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *)); | |
650346e1 | 684 | ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec)); |
52885fa2 JA |
685 | |
686 | td->io_ops_data = ld; | |
b2a432bf PC |
687 | |
688 | /* | |
689 | * Check for option conflicts | |
690 | */ | |
691 | if ((fio_option_is_set(to, ioprio) || fio_option_is_set(to, ioprio_class)) && | |
692 | o->cmdprio_percentage != 0) { | |
693 | log_err("%s: cmdprio_percentage option and mutually exclusive " | |
694 | "prio or prioclass option is set, exiting\n", to->name); | |
695 | td_verror(td, EINVAL, "fio_io_uring_init"); | |
696 | return 1; | |
697 | } | |
1af44196 XW |
698 | |
699 | if (fio_option_is_set(&td->o, ioprio_class)) | |
700 | ld->ioprio_class_set = true; | |
701 | if (fio_option_is_set(&td->o, ioprio)) | |
702 | ld->ioprio_set = true; | |
703 | ||
52885fa2 JA |
704 | return 0; |
705 | } | |
706 | ||
bffad86f | 707 | static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u) |
52885fa2 | 708 | { |
bffad86f | 709 | struct ioring_data *ld = td->io_ops_data; |
52885fa2 JA |
710 | |
711 | ld->io_u_index[io_u->index] = io_u; | |
712 | return 0; | |
713 | } | |
714 | ||
5ffd5626 JA |
715 | static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f) |
716 | { | |
717 | struct ioring_data *ld = td->io_ops_data; | |
718 | struct ioring_options *o = td->eo; | |
719 | ||
17318cf6 | 720 | if (!ld || !o->registerfiles) |
5ffd5626 JA |
721 | return generic_open_file(td, f); |
722 | ||
723 | f->fd = ld->fds[f->engine_pos]; | |
724 | return 0; | |
725 | } | |
726 | ||
727 | static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f) | |
728 | { | |
17318cf6 | 729 | struct ioring_data *ld = td->io_ops_data; |
5ffd5626 JA |
730 | struct ioring_options *o = td->eo; |
731 | ||
17318cf6 | 732 | if (!ld || !o->registerfiles) |
5ffd5626 JA |
733 | return generic_close_file(td, f); |
734 | ||
735 | f->fd = -1; | |
736 | return 0; | |
737 | } | |
738 | ||
52885fa2 | 739 | static struct ioengine_ops ioengine = { |
bffad86f | 740 | .name = "io_uring", |
52885fa2 | 741 | .version = FIO_IOOPS_VERSION, |
04ba61df | 742 | .flags = FIO_ASYNCIO_SYNC_TRIM, |
bffad86f JA |
743 | .init = fio_ioring_init, |
744 | .post_init = fio_ioring_post_init, | |
745 | .io_u_init = fio_ioring_io_u_init, | |
746 | .prep = fio_ioring_prep, | |
747 | .queue = fio_ioring_queue, | |
748 | .commit = fio_ioring_commit, | |
749 | .getevents = fio_ioring_getevents, | |
750 | .event = fio_ioring_event, | |
751 | .cleanup = fio_ioring_cleanup, | |
5ffd5626 JA |
752 | .open_file = fio_ioring_open_file, |
753 | .close_file = fio_ioring_close_file, | |
52885fa2 JA |
754 | .get_file_size = generic_get_file_size, |
755 | .options = options, | |
bffad86f | 756 | .option_struct_size = sizeof(struct ioring_options), |
52885fa2 JA |
757 | }; |
758 | ||
bffad86f | 759 | static void fio_init fio_ioring_register(void) |
52885fa2 | 760 | { |
52885fa2 | 761 | register_ioengine(&ioengine); |
52885fa2 JA |
762 | } |
763 | ||
bffad86f | 764 | static void fio_exit fio_ioring_unregister(void) |
52885fa2 | 765 | { |
52885fa2 | 766 | unregister_ioengine(&ioengine); |
52885fa2 | 767 | } |
1f90e9bb | 768 | #endif |