]>
Commit | Line | Data |
---|---|---|
6663a0a3 AM |
1 | /* |
2 | * Linux io_uring support. | |
3 | * | |
4 | * Copyright (C) 2009 IBM, Corp. | |
5 | * Copyright (C) 2009 Red Hat, Inc. | |
6 | * Copyright (C) 2019 Aarushi Mehta | |
7 | * | |
8 | * This work is licensed under the terms of the GNU GPL, version 2 or later. | |
9 | * See the COPYING file in the top-level directory. | |
10 | */ | |
11 | #include "qemu/osdep.h" | |
12 | #include <liburing.h> | |
6663a0a3 AM |
13 | #include "block/aio.h" |
14 | #include "qemu/queue.h" | |
15 | #include "block/block.h" | |
16 | #include "block/raw-aio.h" | |
17 | #include "qemu/coroutine.h" | |
18 | #include "qapi/error.h" | |
6a6da231 | 19 | #include "sysemu/block-backend.h" |
d803f590 | 20 | #include "trace.h" |
6663a0a3 | 21 | |
a75e4e43 EGE |
22 | /* Only used for assertions. */ |
23 | #include "qemu/coroutine_int.h" | |
24 | ||
6663a0a3 AM |
25 | /* io_uring ring size */ |
26 | #define MAX_ENTRIES 128 | |
27 | ||
28 | typedef struct LuringAIOCB { | |
29 | Coroutine *co; | |
30 | struct io_uring_sqe sqeq; | |
31 | ssize_t ret; | |
32 | QEMUIOVector *qiov; | |
33 | bool is_read; | |
34 | QSIMPLEQ_ENTRY(LuringAIOCB) next; | |
35 | ||
36 | /* | |
37 | * Buffered reads may require resubmission, see | |
38 | * luring_resubmit_short_read(). | |
39 | */ | |
40 | int total_read; | |
41 | QEMUIOVector resubmit_qiov; | |
42 | } LuringAIOCB; | |
43 | ||
44 | typedef struct LuringQueue { | |
6663a0a3 AM |
45 | unsigned int in_queue; |
46 | unsigned int in_flight; | |
47 | bool blocked; | |
48 | QSIMPLEQ_HEAD(, LuringAIOCB) submit_queue; | |
49 | } LuringQueue; | |
50 | ||
51 | typedef struct LuringState { | |
52 | AioContext *aio_context; | |
53 | ||
54 | struct io_uring ring; | |
55 | ||
a75e4e43 | 56 | /* No locking required, only accessed from AioContext home thread */ |
6663a0a3 AM |
57 | LuringQueue io_q; |
58 | ||
6663a0a3 AM |
59 | QEMUBH *completion_bh; |
60 | } LuringState; | |
61 | ||
62 | /** | |
63 | * luring_resubmit: | |
64 | * | |
65 | * Resubmit a request by appending it to submit_queue. The caller must ensure | |
66 | * that ioq_submit() is called later so that submit_queue requests are started. | |
67 | */ | |
68 | static void luring_resubmit(LuringState *s, LuringAIOCB *luringcb) | |
69 | { | |
70 | QSIMPLEQ_INSERT_TAIL(&s->io_q.submit_queue, luringcb, next); | |
71 | s->io_q.in_queue++; | |
72 | } | |
73 | ||
74 | /** | |
75 | * luring_resubmit_short_read: | |
76 | * | |
be6a166f SH |
77 | * Short reads are rare but may occur. The remaining read request needs to be |
78 | * resubmitted. | |
6663a0a3 AM |
79 | */ |
80 | static void luring_resubmit_short_read(LuringState *s, LuringAIOCB *luringcb, | |
81 | int nread) | |
82 | { | |
83 | QEMUIOVector *resubmit_qiov; | |
84 | size_t remaining; | |
85 | ||
d803f590 AM |
86 | trace_luring_resubmit_short_read(s, luringcb, nread); |
87 | ||
6663a0a3 | 88 | /* Update read position */ |
c06fc7ce | 89 | luringcb->total_read += nread; |
6663a0a3 AM |
90 | remaining = luringcb->qiov->size - luringcb->total_read; |
91 | ||
92 | /* Shorten qiov */ | |
93 | resubmit_qiov = &luringcb->resubmit_qiov; | |
94 | if (resubmit_qiov->iov == NULL) { | |
95 | qemu_iovec_init(resubmit_qiov, luringcb->qiov->niov); | |
96 | } else { | |
97 | qemu_iovec_reset(resubmit_qiov); | |
98 | } | |
99 | qemu_iovec_concat(resubmit_qiov, luringcb->qiov, luringcb->total_read, | |
100 | remaining); | |
101 | ||
102 | /* Update sqe */ | |
c06fc7ce | 103 | luringcb->sqeq.off += nread; |
6663a0a3 AM |
104 | luringcb->sqeq.addr = (__u64)(uintptr_t)luringcb->resubmit_qiov.iov; |
105 | luringcb->sqeq.len = luringcb->resubmit_qiov.niov; | |
106 | ||
107 | luring_resubmit(s, luringcb); | |
108 | } | |
109 | ||
110 | /** | |
111 | * luring_process_completions: | |
112 | * @s: AIO state | |
113 | * | |
114 | * Fetches completed I/O requests, consumes cqes and invokes their callbacks | |
115 | * The function is somewhat tricky because it supports nested event loops, for | |
116 | * example when a request callback invokes aio_poll(). | |
117 | * | |
118 | * Function schedules BH completion so it can be called again in a nested | |
119 | * event loop. When there are no events left to complete the BH is being | |
120 | * canceled. | |
121 | * | |
122 | */ | |
123 | static void luring_process_completions(LuringState *s) | |
124 | { | |
125 | struct io_uring_cqe *cqes; | |
126 | int total_bytes; | |
127 | /* | |
128 | * Request completion callbacks can run the nested event loop. | |
129 | * Schedule ourselves so the nested event loop will "see" remaining | |
130 | * completed requests and process them. Without this, completion | |
131 | * callbacks that wait for other requests using a nested event loop | |
132 | * would hang forever. | |
133 | * | |
134 | * This workaround is needed because io_uring uses poll_wait, which | |
135 | * is woken up when new events are added to the uring, thus polling on | |
136 | * the same uring fd will block unless more events are received. | |
137 | * | |
138 | * Other leaf block drivers (drivers that access the data themselves) | |
139 | * are networking based, so they poll sockets for data and run the | |
140 | * correct coroutine. | |
141 | */ | |
142 | qemu_bh_schedule(s->completion_bh); | |
143 | ||
144 | while (io_uring_peek_cqe(&s->ring, &cqes) == 0) { | |
145 | LuringAIOCB *luringcb; | |
146 | int ret; | |
147 | ||
148 | if (!cqes) { | |
149 | break; | |
150 | } | |
151 | ||
152 | luringcb = io_uring_cqe_get_data(cqes); | |
153 | ret = cqes->res; | |
154 | io_uring_cqe_seen(&s->ring, cqes); | |
155 | cqes = NULL; | |
156 | ||
157 | /* Change counters one-by-one because we can be nested. */ | |
158 | s->io_q.in_flight--; | |
d803f590 | 159 | trace_luring_process_completion(s, luringcb, ret); |
6663a0a3 AM |
160 | |
161 | /* total_read is non-zero only for resubmitted read requests */ | |
162 | total_bytes = ret + luringcb->total_read; | |
163 | ||
164 | if (ret < 0) { | |
54caccb3 FE |
165 | /* |
166 | * Only writev/readv/fsync requests on regular files or host block | |
167 | * devices are submitted. Therefore -EAGAIN is not expected but it's | |
168 | * known to happen sometimes with Linux SCSI. Submit again and hope | |
169 | * the request completes successfully. | |
170 | * | |
171 | * For more information, see: | |
172 | * https://lore.kernel.org/io-uring/20210727165811.284510-3-axboe@kernel.dk/T/#u | |
173 | * | |
174 | * If the code is changed to submit other types of requests in the | |
175 | * future, then this workaround may need to be extended to deal with | |
176 | * genuine -EAGAIN results that should not be resubmitted | |
177 | * immediately. | |
178 | */ | |
179 | if (ret == -EINTR || ret == -EAGAIN) { | |
6663a0a3 AM |
180 | luring_resubmit(s, luringcb); |
181 | continue; | |
182 | } | |
183 | } else if (!luringcb->qiov) { | |
184 | goto end; | |
185 | } else if (total_bytes == luringcb->qiov->size) { | |
186 | ret = 0; | |
187 | /* Only read/write */ | |
188 | } else { | |
189 | /* Short Read/Write */ | |
190 | if (luringcb->is_read) { | |
191 | if (ret > 0) { | |
192 | luring_resubmit_short_read(s, luringcb, ret); | |
193 | continue; | |
194 | } else { | |
195 | /* Pad with zeroes */ | |
196 | qemu_iovec_memset(luringcb->qiov, total_bytes, 0, | |
197 | luringcb->qiov->size - total_bytes); | |
198 | ret = 0; | |
199 | } | |
200 | } else { | |
74e4a8a9 | 201 | ret = -ENOSPC; |
6663a0a3 AM |
202 | } |
203 | } | |
204 | end: | |
205 | luringcb->ret = ret; | |
206 | qemu_iovec_destroy(&luringcb->resubmit_qiov); | |
207 | ||
208 | /* | |
209 | * If the coroutine is already entered it must be in ioq_submit() | |
210 | * and will notice luringcb->ret has been filled in when it | |
211 | * eventually runs later. Coroutines cannot be entered recursively | |
212 | * so avoid doing that! | |
213 | */ | |
a75e4e43 | 214 | assert(luringcb->co->ctx == s->aio_context); |
6663a0a3 AM |
215 | if (!qemu_coroutine_entered(luringcb->co)) { |
216 | aio_co_wake(luringcb->co); | |
217 | } | |
218 | } | |
219 | qemu_bh_cancel(s->completion_bh); | |
220 | } | |
221 | ||
222 | static int ioq_submit(LuringState *s) | |
223 | { | |
224 | int ret = 0; | |
225 | LuringAIOCB *luringcb, *luringcb_next; | |
226 | ||
227 | while (s->io_q.in_queue > 0) { | |
228 | /* | |
229 | * Try to fetch sqes from the ring for requests waiting in | |
230 | * the overflow queue | |
231 | */ | |
232 | QSIMPLEQ_FOREACH_SAFE(luringcb, &s->io_q.submit_queue, next, | |
233 | luringcb_next) { | |
234 | struct io_uring_sqe *sqes = io_uring_get_sqe(&s->ring); | |
235 | if (!sqes) { | |
236 | break; | |
237 | } | |
238 | /* Prep sqe for submission */ | |
239 | *sqes = luringcb->sqeq; | |
240 | QSIMPLEQ_REMOVE_HEAD(&s->io_q.submit_queue, next); | |
241 | } | |
242 | ret = io_uring_submit(&s->ring); | |
d803f590 | 243 | trace_luring_io_uring_submit(s, ret); |
6663a0a3 AM |
244 | /* Prevent infinite loop if submission is refused */ |
245 | if (ret <= 0) { | |
b4e44c99 | 246 | if (ret == -EAGAIN || ret == -EINTR) { |
6663a0a3 AM |
247 | continue; |
248 | } | |
249 | break; | |
250 | } | |
251 | s->io_q.in_flight += ret; | |
252 | s->io_q.in_queue -= ret; | |
253 | } | |
254 | s->io_q.blocked = (s->io_q.in_queue > 0); | |
255 | ||
256 | if (s->io_q.in_flight) { | |
257 | /* | |
258 | * We can try to complete something just right away if there are | |
259 | * still requests in-flight. | |
260 | */ | |
261 | luring_process_completions(s); | |
262 | } | |
263 | return ret; | |
264 | } | |
265 | ||
266 | static void luring_process_completions_and_submit(LuringState *s) | |
267 | { | |
6663a0a3 AM |
268 | luring_process_completions(s); |
269 | ||
6a6da231 | 270 | if (s->io_q.in_queue > 0) { |
6663a0a3 AM |
271 | ioq_submit(s); |
272 | } | |
6663a0a3 AM |
273 | } |
274 | ||
275 | static void qemu_luring_completion_bh(void *opaque) | |
276 | { | |
277 | LuringState *s = opaque; | |
278 | luring_process_completions_and_submit(s); | |
279 | } | |
280 | ||
281 | static void qemu_luring_completion_cb(void *opaque) | |
282 | { | |
283 | LuringState *s = opaque; | |
284 | luring_process_completions_and_submit(s); | |
285 | } | |
286 | ||
daffeb02 AM |
287 | static bool qemu_luring_poll_cb(void *opaque) |
288 | { | |
289 | LuringState *s = opaque; | |
daffeb02 | 290 | |
826cc324 SH |
291 | return io_uring_cq_ready(&s->ring); |
292 | } | |
293 | ||
294 | static void qemu_luring_poll_ready(void *opaque) | |
295 | { | |
296 | LuringState *s = opaque; | |
daffeb02 | 297 | |
826cc324 | 298 | luring_process_completions_and_submit(s); |
daffeb02 AM |
299 | } |
300 | ||
6663a0a3 AM |
301 | static void ioq_init(LuringQueue *io_q) |
302 | { | |
303 | QSIMPLEQ_INIT(&io_q->submit_queue); | |
6663a0a3 AM |
304 | io_q->in_queue = 0; |
305 | io_q->in_flight = 0; | |
306 | io_q->blocked = false; | |
307 | } | |
308 | ||
6a6da231 | 309 | static void luring_unplug_fn(void *opaque) |
6663a0a3 | 310 | { |
6a6da231 SH |
311 | LuringState *s = opaque; |
312 | trace_luring_unplug_fn(s, s->io_q.blocked, s->io_q.in_queue, | |
313 | s->io_q.in_flight); | |
314 | if (!s->io_q.blocked && s->io_q.in_queue > 0) { | |
6663a0a3 AM |
315 | ioq_submit(s); |
316 | } | |
317 | } | |
318 | ||
319 | /** | |
320 | * luring_do_submit: | |
321 | * @fd: file descriptor for I/O | |
322 | * @luringcb: AIO control block | |
323 | * @s: AIO state | |
324 | * @offset: offset for request | |
325 | * @type: type of request | |
326 | * | |
327 | * Fetches sqes from ring, adds to pending queue and preps them | |
328 | * | |
329 | */ | |
330 | static int luring_do_submit(int fd, LuringAIOCB *luringcb, LuringState *s, | |
331 | uint64_t offset, int type) | |
332 | { | |
d803f590 | 333 | int ret; |
6663a0a3 AM |
334 | struct io_uring_sqe *sqes = &luringcb->sqeq; |
335 | ||
336 | switch (type) { | |
337 | case QEMU_AIO_WRITE: | |
338 | io_uring_prep_writev(sqes, fd, luringcb->qiov->iov, | |
339 | luringcb->qiov->niov, offset); | |
340 | break; | |
4751d09a SL |
341 | case QEMU_AIO_ZONE_APPEND: |
342 | io_uring_prep_writev(sqes, fd, luringcb->qiov->iov, | |
343 | luringcb->qiov->niov, offset); | |
344 | break; | |
6663a0a3 AM |
345 | case QEMU_AIO_READ: |
346 | io_uring_prep_readv(sqes, fd, luringcb->qiov->iov, | |
347 | luringcb->qiov->niov, offset); | |
348 | break; | |
349 | case QEMU_AIO_FLUSH: | |
350 | io_uring_prep_fsync(sqes, fd, IORING_FSYNC_DATASYNC); | |
351 | break; | |
352 | default: | |
353 | fprintf(stderr, "%s: invalid AIO request type, aborting 0x%x.\n", | |
354 | __func__, type); | |
355 | abort(); | |
356 | } | |
357 | io_uring_sqe_set_data(sqes, luringcb); | |
358 | ||
359 | QSIMPLEQ_INSERT_TAIL(&s->io_q.submit_queue, luringcb, next); | |
360 | s->io_q.in_queue++; | |
6a6da231 SH |
361 | trace_luring_do_submit(s, s->io_q.blocked, s->io_q.in_queue, |
362 | s->io_q.in_flight); | |
363 | if (!s->io_q.blocked) { | |
364 | if (s->io_q.in_flight + s->io_q.in_queue >= MAX_ENTRIES) { | |
365 | ret = ioq_submit(s); | |
366 | trace_luring_do_submit_done(s, ret); | |
367 | return ret; | |
368 | } | |
369 | ||
370 | blk_io_plug_call(luring_unplug_fn, s); | |
6663a0a3 AM |
371 | } |
372 | return 0; | |
373 | } | |
374 | ||
a75e4e43 EGE |
375 | int coroutine_fn luring_co_submit(BlockDriverState *bs, int fd, uint64_t offset, |
376 | QEMUIOVector *qiov, int type) | |
6663a0a3 AM |
377 | { |
378 | int ret; | |
a75e4e43 EGE |
379 | AioContext *ctx = qemu_get_current_aio_context(); |
380 | LuringState *s = aio_get_linux_io_uring(ctx); | |
6663a0a3 AM |
381 | LuringAIOCB luringcb = { |
382 | .co = qemu_coroutine_self(), | |
383 | .ret = -EINPROGRESS, | |
384 | .qiov = qiov, | |
385 | .is_read = (type == QEMU_AIO_READ), | |
386 | }; | |
d803f590 AM |
387 | trace_luring_co_submit(bs, s, &luringcb, fd, offset, qiov ? qiov->size : 0, |
388 | type); | |
6663a0a3 | 389 | ret = luring_do_submit(fd, &luringcb, s, offset, type); |
d803f590 | 390 | |
6663a0a3 AM |
391 | if (ret < 0) { |
392 | return ret; | |
393 | } | |
394 | ||
395 | if (luringcb.ret == -EINPROGRESS) { | |
396 | qemu_coroutine_yield(); | |
397 | } | |
398 | return luringcb.ret; | |
399 | } | |
400 | ||
401 | void luring_detach_aio_context(LuringState *s, AioContext *old_context) | |
402 | { | |
60f782b6 | 403 | aio_set_fd_handler(old_context, s->ring.ring_fd, |
826cc324 | 404 | NULL, NULL, NULL, NULL, s); |
6663a0a3 AM |
405 | qemu_bh_delete(s->completion_bh); |
406 | s->aio_context = NULL; | |
407 | } | |
408 | ||
409 | void luring_attach_aio_context(LuringState *s, AioContext *new_context) | |
410 | { | |
411 | s->aio_context = new_context; | |
412 | s->completion_bh = aio_bh_new(new_context, qemu_luring_completion_bh, s); | |
60f782b6 | 413 | aio_set_fd_handler(s->aio_context, s->ring.ring_fd, |
826cc324 SH |
414 | qemu_luring_completion_cb, NULL, |
415 | qemu_luring_poll_cb, qemu_luring_poll_ready, s); | |
6663a0a3 AM |
416 | } |
417 | ||
418 | LuringState *luring_init(Error **errp) | |
419 | { | |
420 | int rc; | |
421 | LuringState *s = g_new0(LuringState, 1); | |
422 | struct io_uring *ring = &s->ring; | |
423 | ||
d803f590 AM |
424 | trace_luring_init_state(s, sizeof(*s)); |
425 | ||
6663a0a3 AM |
426 | rc = io_uring_queue_init(MAX_ENTRIES, ring, 0); |
427 | if (rc < 0) { | |
428 | error_setg_errno(errp, errno, "failed to init linux io_uring ring"); | |
429 | g_free(s); | |
430 | return NULL; | |
431 | } | |
432 | ||
433 | ioq_init(&s->io_q); | |
e2848bc5 | 434 | return s; |
7845e731 | 435 | |
6663a0a3 AM |
436 | } |
437 | ||
438 | void luring_cleanup(LuringState *s) | |
439 | { | |
440 | io_uring_queue_exit(&s->ring); | |
d803f590 | 441 | trace_luring_cleanup_state(s); |
bd89f936 | 442 | g_free(s); |
6663a0a3 | 443 | } |