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