]>
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> | |
13 | #include "qemu-common.h" | |
14 | #include "block/aio.h" | |
15 | #include "qemu/queue.h" | |
16 | #include "block/block.h" | |
17 | #include "block/raw-aio.h" | |
18 | #include "qemu/coroutine.h" | |
19 | #include "qapi/error.h" | |
20 | ||
21 | /* io_uring ring size */ | |
22 | #define MAX_ENTRIES 128 | |
23 | ||
24 | typedef struct LuringAIOCB { | |
25 | Coroutine *co; | |
26 | struct io_uring_sqe sqeq; | |
27 | ssize_t ret; | |
28 | QEMUIOVector *qiov; | |
29 | bool is_read; | |
30 | QSIMPLEQ_ENTRY(LuringAIOCB) next; | |
31 | ||
32 | /* | |
33 | * Buffered reads may require resubmission, see | |
34 | * luring_resubmit_short_read(). | |
35 | */ | |
36 | int total_read; | |
37 | QEMUIOVector resubmit_qiov; | |
38 | } LuringAIOCB; | |
39 | ||
40 | typedef struct LuringQueue { | |
41 | int plugged; | |
42 | unsigned int in_queue; | |
43 | unsigned int in_flight; | |
44 | bool blocked; | |
45 | QSIMPLEQ_HEAD(, LuringAIOCB) submit_queue; | |
46 | } LuringQueue; | |
47 | ||
48 | typedef struct LuringState { | |
49 | AioContext *aio_context; | |
50 | ||
51 | struct io_uring ring; | |
52 | ||
53 | /* io queue for submit at batch. Protected by AioContext lock. */ | |
54 | LuringQueue io_q; | |
55 | ||
56 | /* I/O completion processing. Only runs in I/O thread. */ | |
57 | QEMUBH *completion_bh; | |
58 | } LuringState; | |
59 | ||
60 | /** | |
61 | * luring_resubmit: | |
62 | * | |
63 | * Resubmit a request by appending it to submit_queue. The caller must ensure | |
64 | * that ioq_submit() is called later so that submit_queue requests are started. | |
65 | */ | |
66 | static void luring_resubmit(LuringState *s, LuringAIOCB *luringcb) | |
67 | { | |
68 | QSIMPLEQ_INSERT_TAIL(&s->io_q.submit_queue, luringcb, next); | |
69 | s->io_q.in_queue++; | |
70 | } | |
71 | ||
72 | /** | |
73 | * luring_resubmit_short_read: | |
74 | * | |
75 | * Before Linux commit 9d93a3f5a0c ("io_uring: punt short reads to async | |
76 | * context") a buffered I/O request with the start of the file range in the | |
77 | * page cache could result in a short read. Applications need to resubmit the | |
78 | * remaining read request. | |
79 | * | |
80 | * This is a slow path but recent kernels never take it. | |
81 | */ | |
82 | static void luring_resubmit_short_read(LuringState *s, LuringAIOCB *luringcb, | |
83 | int nread) | |
84 | { | |
85 | QEMUIOVector *resubmit_qiov; | |
86 | size_t remaining; | |
87 | ||
88 | /* Update read position */ | |
89 | luringcb->total_read = nread; | |
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 */ | |
103 | luringcb->sqeq.off = nread; | |
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--; | |
159 | ||
160 | /* total_read is non-zero only for resubmitted read requests */ | |
161 | total_bytes = ret + luringcb->total_read; | |
162 | ||
163 | if (ret < 0) { | |
164 | if (ret == -EINTR) { | |
165 | luring_resubmit(s, luringcb); | |
166 | continue; | |
167 | } | |
168 | } else if (!luringcb->qiov) { | |
169 | goto end; | |
170 | } else if (total_bytes == luringcb->qiov->size) { | |
171 | ret = 0; | |
172 | /* Only read/write */ | |
173 | } else { | |
174 | /* Short Read/Write */ | |
175 | if (luringcb->is_read) { | |
176 | if (ret > 0) { | |
177 | luring_resubmit_short_read(s, luringcb, ret); | |
178 | continue; | |
179 | } else { | |
180 | /* Pad with zeroes */ | |
181 | qemu_iovec_memset(luringcb->qiov, total_bytes, 0, | |
182 | luringcb->qiov->size - total_bytes); | |
183 | ret = 0; | |
184 | } | |
185 | } else { | |
186 | ret = -ENOSPC;; | |
187 | } | |
188 | } | |
189 | end: | |
190 | luringcb->ret = ret; | |
191 | qemu_iovec_destroy(&luringcb->resubmit_qiov); | |
192 | ||
193 | /* | |
194 | * If the coroutine is already entered it must be in ioq_submit() | |
195 | * and will notice luringcb->ret has been filled in when it | |
196 | * eventually runs later. Coroutines cannot be entered recursively | |
197 | * so avoid doing that! | |
198 | */ | |
199 | if (!qemu_coroutine_entered(luringcb->co)) { | |
200 | aio_co_wake(luringcb->co); | |
201 | } | |
202 | } | |
203 | qemu_bh_cancel(s->completion_bh); | |
204 | } | |
205 | ||
206 | static int ioq_submit(LuringState *s) | |
207 | { | |
208 | int ret = 0; | |
209 | LuringAIOCB *luringcb, *luringcb_next; | |
210 | ||
211 | while (s->io_q.in_queue > 0) { | |
212 | /* | |
213 | * Try to fetch sqes from the ring for requests waiting in | |
214 | * the overflow queue | |
215 | */ | |
216 | QSIMPLEQ_FOREACH_SAFE(luringcb, &s->io_q.submit_queue, next, | |
217 | luringcb_next) { | |
218 | struct io_uring_sqe *sqes = io_uring_get_sqe(&s->ring); | |
219 | if (!sqes) { | |
220 | break; | |
221 | } | |
222 | /* Prep sqe for submission */ | |
223 | *sqes = luringcb->sqeq; | |
224 | QSIMPLEQ_REMOVE_HEAD(&s->io_q.submit_queue, next); | |
225 | } | |
226 | ret = io_uring_submit(&s->ring); | |
227 | /* Prevent infinite loop if submission is refused */ | |
228 | if (ret <= 0) { | |
229 | if (ret == -EAGAIN) { | |
230 | continue; | |
231 | } | |
232 | break; | |
233 | } | |
234 | s->io_q.in_flight += ret; | |
235 | s->io_q.in_queue -= ret; | |
236 | } | |
237 | s->io_q.blocked = (s->io_q.in_queue > 0); | |
238 | ||
239 | if (s->io_q.in_flight) { | |
240 | /* | |
241 | * We can try to complete something just right away if there are | |
242 | * still requests in-flight. | |
243 | */ | |
244 | luring_process_completions(s); | |
245 | } | |
246 | return ret; | |
247 | } | |
248 | ||
249 | static void luring_process_completions_and_submit(LuringState *s) | |
250 | { | |
251 | aio_context_acquire(s->aio_context); | |
252 | luring_process_completions(s); | |
253 | ||
254 | if (!s->io_q.plugged && s->io_q.in_queue > 0) { | |
255 | ioq_submit(s); | |
256 | } | |
257 | aio_context_release(s->aio_context); | |
258 | } | |
259 | ||
260 | static void qemu_luring_completion_bh(void *opaque) | |
261 | { | |
262 | LuringState *s = opaque; | |
263 | luring_process_completions_and_submit(s); | |
264 | } | |
265 | ||
266 | static void qemu_luring_completion_cb(void *opaque) | |
267 | { | |
268 | LuringState *s = opaque; | |
269 | luring_process_completions_and_submit(s); | |
270 | } | |
271 | ||
272 | static void ioq_init(LuringQueue *io_q) | |
273 | { | |
274 | QSIMPLEQ_INIT(&io_q->submit_queue); | |
275 | io_q->plugged = 0; | |
276 | io_q->in_queue = 0; | |
277 | io_q->in_flight = 0; | |
278 | io_q->blocked = false; | |
279 | } | |
280 | ||
281 | void luring_io_plug(BlockDriverState *bs, LuringState *s) | |
282 | { | |
283 | s->io_q.plugged++; | |
284 | } | |
285 | ||
286 | void luring_io_unplug(BlockDriverState *bs, LuringState *s) | |
287 | { | |
288 | assert(s->io_q.plugged); | |
289 | if (--s->io_q.plugged == 0 && | |
290 | !s->io_q.blocked && s->io_q.in_queue > 0) { | |
291 | ioq_submit(s); | |
292 | } | |
293 | } | |
294 | ||
295 | /** | |
296 | * luring_do_submit: | |
297 | * @fd: file descriptor for I/O | |
298 | * @luringcb: AIO control block | |
299 | * @s: AIO state | |
300 | * @offset: offset for request | |
301 | * @type: type of request | |
302 | * | |
303 | * Fetches sqes from ring, adds to pending queue and preps them | |
304 | * | |
305 | */ | |
306 | static int luring_do_submit(int fd, LuringAIOCB *luringcb, LuringState *s, | |
307 | uint64_t offset, int type) | |
308 | { | |
309 | struct io_uring_sqe *sqes = &luringcb->sqeq; | |
310 | ||
311 | switch (type) { | |
312 | case QEMU_AIO_WRITE: | |
313 | io_uring_prep_writev(sqes, fd, luringcb->qiov->iov, | |
314 | luringcb->qiov->niov, offset); | |
315 | break; | |
316 | case QEMU_AIO_READ: | |
317 | io_uring_prep_readv(sqes, fd, luringcb->qiov->iov, | |
318 | luringcb->qiov->niov, offset); | |
319 | break; | |
320 | case QEMU_AIO_FLUSH: | |
321 | io_uring_prep_fsync(sqes, fd, IORING_FSYNC_DATASYNC); | |
322 | break; | |
323 | default: | |
324 | fprintf(stderr, "%s: invalid AIO request type, aborting 0x%x.\n", | |
325 | __func__, type); | |
326 | abort(); | |
327 | } | |
328 | io_uring_sqe_set_data(sqes, luringcb); | |
329 | ||
330 | QSIMPLEQ_INSERT_TAIL(&s->io_q.submit_queue, luringcb, next); | |
331 | s->io_q.in_queue++; | |
332 | ||
333 | if (!s->io_q.blocked && | |
334 | (!s->io_q.plugged || | |
335 | s->io_q.in_flight + s->io_q.in_queue >= MAX_ENTRIES)) { | |
336 | return ioq_submit(s); | |
337 | } | |
338 | return 0; | |
339 | } | |
340 | ||
341 | int coroutine_fn luring_co_submit(BlockDriverState *bs, LuringState *s, int fd, | |
342 | uint64_t offset, QEMUIOVector *qiov, int type) | |
343 | { | |
344 | int ret; | |
345 | LuringAIOCB luringcb = { | |
346 | .co = qemu_coroutine_self(), | |
347 | .ret = -EINPROGRESS, | |
348 | .qiov = qiov, | |
349 | .is_read = (type == QEMU_AIO_READ), | |
350 | }; | |
351 | ||
352 | ret = luring_do_submit(fd, &luringcb, s, offset, type); | |
353 | if (ret < 0) { | |
354 | return ret; | |
355 | } | |
356 | ||
357 | if (luringcb.ret == -EINPROGRESS) { | |
358 | qemu_coroutine_yield(); | |
359 | } | |
360 | return luringcb.ret; | |
361 | } | |
362 | ||
363 | void luring_detach_aio_context(LuringState *s, AioContext *old_context) | |
364 | { | |
365 | aio_set_fd_handler(old_context, s->ring.ring_fd, false, NULL, NULL, NULL, | |
366 | s); | |
367 | qemu_bh_delete(s->completion_bh); | |
368 | s->aio_context = NULL; | |
369 | } | |
370 | ||
371 | void luring_attach_aio_context(LuringState *s, AioContext *new_context) | |
372 | { | |
373 | s->aio_context = new_context; | |
374 | s->completion_bh = aio_bh_new(new_context, qemu_luring_completion_bh, s); | |
375 | aio_set_fd_handler(s->aio_context, s->ring.ring_fd, false, | |
376 | qemu_luring_completion_cb, NULL, NULL, s); | |
377 | } | |
378 | ||
379 | LuringState *luring_init(Error **errp) | |
380 | { | |
381 | int rc; | |
382 | LuringState *s = g_new0(LuringState, 1); | |
383 | struct io_uring *ring = &s->ring; | |
384 | ||
385 | rc = io_uring_queue_init(MAX_ENTRIES, ring, 0); | |
386 | if (rc < 0) { | |
387 | error_setg_errno(errp, errno, "failed to init linux io_uring ring"); | |
388 | g_free(s); | |
389 | return NULL; | |
390 | } | |
391 | ||
392 | ioq_init(&s->io_q); | |
393 | return s; | |
394 | ||
395 | } | |
396 | ||
397 | void luring_cleanup(LuringState *s) | |
398 | { | |
399 | io_uring_queue_exit(&s->ring); | |
400 | g_free(s); | |
401 | } |