linux-aio: poll ring for completions
[qemu.git] / block / linux-aio.c
1 /*
2 * Linux native AIO support.
3 *
4 * Copyright (C) 2009 IBM, Corp.
5 * Copyright (C) 2009 Red Hat, Inc.
6 *
7 * This work is licensed under the terms of the GNU GPL, version 2 or later.
8 * See the COPYING file in the top-level directory.
9 */
10 #include "qemu/osdep.h"
11 #include "qemu-common.h"
12 #include "block/aio.h"
13 #include "qemu/queue.h"
14 #include "block/block.h"
15 #include "block/raw-aio.h"
16 #include "qemu/event_notifier.h"
17 #include "qemu/coroutine.h"
18
19 #include <libaio.h>
20
21 /*
22 * Queue size (per-device).
23 *
24 * XXX: eventually we need to communicate this to the guest and/or make it
25 * tunable by the guest. If we get more outstanding requests at a time
26 * than this we will get EAGAIN from io_submit which is communicated to
27 * the guest as an I/O error.
28 */
29 #define MAX_EVENTS 128
30
31 struct qemu_laiocb {
32 BlockAIOCB common;
33 Coroutine *co;
34 LinuxAioState *ctx;
35 struct iocb iocb;
36 ssize_t ret;
37 size_t nbytes;
38 QEMUIOVector *qiov;
39 bool is_read;
40 QSIMPLEQ_ENTRY(qemu_laiocb) next;
41 };
42
43 typedef struct {
44 int plugged;
45 unsigned int in_queue;
46 unsigned int in_flight;
47 bool blocked;
48 QSIMPLEQ_HEAD(, qemu_laiocb) pending;
49 } LaioQueue;
50
51 struct LinuxAioState {
52 AioContext *aio_context;
53
54 io_context_t ctx;
55 EventNotifier e;
56
57 /* io queue for submit at batch */
58 LaioQueue io_q;
59
60 /* I/O completion processing */
61 QEMUBH *completion_bh;
62 int event_idx;
63 int event_max;
64 };
65
66 static void ioq_submit(LinuxAioState *s);
67
68 static inline ssize_t io_event_ret(struct io_event *ev)
69 {
70 return (ssize_t)(((uint64_t)ev->res2 << 32) | ev->res);
71 }
72
73 /*
74 * Completes an AIO request (calls the callback and frees the ACB).
75 */
76 static void qemu_laio_process_completion(struct qemu_laiocb *laiocb)
77 {
78 int ret;
79
80 ret = laiocb->ret;
81 if (ret != -ECANCELED) {
82 if (ret == laiocb->nbytes) {
83 ret = 0;
84 } else if (ret >= 0) {
85 /* Short reads mean EOF, pad with zeros. */
86 if (laiocb->is_read) {
87 qemu_iovec_memset(laiocb->qiov, ret, 0,
88 laiocb->qiov->size - ret);
89 } else {
90 ret = -ENOSPC;
91 }
92 }
93 }
94
95 laiocb->ret = ret;
96 if (laiocb->co) {
97 /* If the coroutine is already entered it must be in ioq_submit() and
98 * will notice laio->ret has been filled in when it eventually runs
99 * later. Coroutines cannot be entered recursively so avoid doing
100 * that!
101 */
102 if (!qemu_coroutine_entered(laiocb->co)) {
103 qemu_coroutine_enter(laiocb->co);
104 }
105 } else {
106 laiocb->common.cb(laiocb->common.opaque, ret);
107 qemu_aio_unref(laiocb);
108 }
109 }
110
111 /**
112 * aio_ring buffer which is shared between userspace and kernel.
113 *
114 * This copied from linux/fs/aio.c, common header does not exist
115 * but AIO exists for ages so we assume ABI is stable.
116 */
117 struct aio_ring {
118 unsigned id; /* kernel internal index number */
119 unsigned nr; /* number of io_events */
120 unsigned head; /* Written to by userland or by kernel. */
121 unsigned tail;
122
123 unsigned magic;
124 unsigned compat_features;
125 unsigned incompat_features;
126 unsigned header_length; /* size of aio_ring */
127
128 struct io_event io_events[0];
129 };
130
131 /**
132 * io_getevents_peek:
133 * @ctx: AIO context
134 * @events: pointer on events array, output value
135
136 * Returns the number of completed events and sets a pointer
137 * on events array. This function does not update the internal
138 * ring buffer, only reads head and tail. When @events has been
139 * processed io_getevents_commit() must be called.
140 */
141 static inline unsigned int io_getevents_peek(io_context_t ctx,
142 struct io_event **events)
143 {
144 struct aio_ring *ring = (struct aio_ring *)ctx;
145 unsigned int head = ring->head, tail = ring->tail;
146 unsigned int nr;
147
148 nr = tail >= head ? tail - head : ring->nr - head;
149 *events = ring->io_events + head;
150 /* To avoid speculative loads of s->events[i] before observing tail.
151 Paired with smp_wmb() inside linux/fs/aio.c: aio_complete(). */
152 smp_rmb();
153
154 return nr;
155 }
156
157 /**
158 * io_getevents_commit:
159 * @ctx: AIO context
160 * @nr: the number of events on which head should be advanced
161 *
162 * Advances head of a ring buffer.
163 */
164 static inline void io_getevents_commit(io_context_t ctx, unsigned int nr)
165 {
166 struct aio_ring *ring = (struct aio_ring *)ctx;
167
168 if (nr) {
169 ring->head = (ring->head + nr) % ring->nr;
170 }
171 }
172
173 /**
174 * io_getevents_advance_and_peek:
175 * @ctx: AIO context
176 * @events: pointer on events array, output value
177 * @nr: the number of events on which head should be advanced
178 *
179 * Advances head of a ring buffer and returns number of elements left.
180 */
181 static inline unsigned int
182 io_getevents_advance_and_peek(io_context_t ctx,
183 struct io_event **events,
184 unsigned int nr)
185 {
186 io_getevents_commit(ctx, nr);
187 return io_getevents_peek(ctx, events);
188 }
189
190 /**
191 * qemu_laio_process_completions:
192 * @s: AIO state
193 *
194 * Fetches completed I/O requests and invokes their callbacks.
195 *
196 * The function is somewhat tricky because it supports nested event loops, for
197 * example when a request callback invokes aio_poll(). In order to do this,
198 * indices are kept in LinuxAioState. Function schedules BH completion so it
199 * can be called again in a nested event loop. When there are no events left
200 * to complete the BH is being canceled.
201 */
202 static void qemu_laio_process_completions(LinuxAioState *s)
203 {
204 struct io_event *events;
205
206 /* Reschedule so nested event loops see currently pending completions */
207 qemu_bh_schedule(s->completion_bh);
208
209 while ((s->event_max = io_getevents_advance_and_peek(s->ctx, &events,
210 s->event_idx))) {
211 for (s->event_idx = 0; s->event_idx < s->event_max; ) {
212 struct iocb *iocb = events[s->event_idx].obj;
213 struct qemu_laiocb *laiocb =
214 container_of(iocb, struct qemu_laiocb, iocb);
215
216 laiocb->ret = io_event_ret(&events[s->event_idx]);
217
218 /* Change counters one-by-one because we can be nested. */
219 s->io_q.in_flight--;
220 s->event_idx++;
221 qemu_laio_process_completion(laiocb);
222 }
223 }
224
225 qemu_bh_cancel(s->completion_bh);
226
227 /* If we are nested we have to notify the level above that we are done
228 * by setting event_max to zero, upper level will then jump out of it's
229 * own `for` loop. If we are the last all counters droped to zero. */
230 s->event_max = 0;
231 s->event_idx = 0;
232 }
233
234 static void qemu_laio_process_completions_and_submit(LinuxAioState *s)
235 {
236 qemu_laio_process_completions(s);
237 if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
238 ioq_submit(s);
239 }
240 }
241
242 static void qemu_laio_completion_bh(void *opaque)
243 {
244 LinuxAioState *s = opaque;
245
246 qemu_laio_process_completions_and_submit(s);
247 }
248
249 static void qemu_laio_completion_cb(EventNotifier *e)
250 {
251 LinuxAioState *s = container_of(e, LinuxAioState, e);
252
253 if (event_notifier_test_and_clear(&s->e)) {
254 qemu_laio_process_completions_and_submit(s);
255 }
256 }
257
258 static bool qemu_laio_poll_cb(void *opaque)
259 {
260 EventNotifier *e = opaque;
261 LinuxAioState *s = container_of(e, LinuxAioState, e);
262 struct io_event *events;
263
264 if (!io_getevents_peek(s->ctx, &events)) {
265 return false;
266 }
267
268 qemu_laio_process_completions_and_submit(s);
269 return true;
270 }
271
272 static void laio_cancel(BlockAIOCB *blockacb)
273 {
274 struct qemu_laiocb *laiocb = (struct qemu_laiocb *)blockacb;
275 struct io_event event;
276 int ret;
277
278 if (laiocb->ret != -EINPROGRESS) {
279 return;
280 }
281 ret = io_cancel(laiocb->ctx->ctx, &laiocb->iocb, &event);
282 laiocb->ret = -ECANCELED;
283 if (ret != 0) {
284 /* iocb is not cancelled, cb will be called by the event loop later */
285 return;
286 }
287
288 laiocb->common.cb(laiocb->common.opaque, laiocb->ret);
289 }
290
291 static const AIOCBInfo laio_aiocb_info = {
292 .aiocb_size = sizeof(struct qemu_laiocb),
293 .cancel_async = laio_cancel,
294 };
295
296 static void ioq_init(LaioQueue *io_q)
297 {
298 QSIMPLEQ_INIT(&io_q->pending);
299 io_q->plugged = 0;
300 io_q->in_queue = 0;
301 io_q->in_flight = 0;
302 io_q->blocked = false;
303 }
304
305 static void ioq_submit(LinuxAioState *s)
306 {
307 int ret, len;
308 struct qemu_laiocb *aiocb;
309 struct iocb *iocbs[MAX_EVENTS];
310 QSIMPLEQ_HEAD(, qemu_laiocb) completed;
311
312 do {
313 if (s->io_q.in_flight >= MAX_EVENTS) {
314 break;
315 }
316 len = 0;
317 QSIMPLEQ_FOREACH(aiocb, &s->io_q.pending, next) {
318 iocbs[len++] = &aiocb->iocb;
319 if (s->io_q.in_flight + len >= MAX_EVENTS) {
320 break;
321 }
322 }
323
324 ret = io_submit(s->ctx, len, iocbs);
325 if (ret == -EAGAIN) {
326 break;
327 }
328 if (ret < 0) {
329 /* Fail the first request, retry the rest */
330 aiocb = QSIMPLEQ_FIRST(&s->io_q.pending);
331 QSIMPLEQ_REMOVE_HEAD(&s->io_q.pending, next);
332 s->io_q.in_queue--;
333 aiocb->ret = ret;
334 qemu_laio_process_completion(aiocb);
335 continue;
336 }
337
338 s->io_q.in_flight += ret;
339 s->io_q.in_queue -= ret;
340 aiocb = container_of(iocbs[ret - 1], struct qemu_laiocb, iocb);
341 QSIMPLEQ_SPLIT_AFTER(&s->io_q.pending, aiocb, next, &completed);
342 } while (ret == len && !QSIMPLEQ_EMPTY(&s->io_q.pending));
343 s->io_q.blocked = (s->io_q.in_queue > 0);
344
345 if (s->io_q.in_flight) {
346 /* We can try to complete something just right away if there are
347 * still requests in-flight. */
348 qemu_laio_process_completions(s);
349 /*
350 * Even we have completed everything (in_flight == 0), the queue can
351 * have still pended requests (in_queue > 0). We do not attempt to
352 * repeat submission to avoid IO hang. The reason is simple: s->e is
353 * still set and completion callback will be called shortly and all
354 * pended requests will be submitted from there.
355 */
356 }
357 }
358
359 void laio_io_plug(BlockDriverState *bs, LinuxAioState *s)
360 {
361 s->io_q.plugged++;
362 }
363
364 void laio_io_unplug(BlockDriverState *bs, LinuxAioState *s)
365 {
366 assert(s->io_q.plugged);
367 if (--s->io_q.plugged == 0 &&
368 !s->io_q.blocked && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
369 ioq_submit(s);
370 }
371 }
372
373 static int laio_do_submit(int fd, struct qemu_laiocb *laiocb, off_t offset,
374 int type)
375 {
376 LinuxAioState *s = laiocb->ctx;
377 struct iocb *iocbs = &laiocb->iocb;
378 QEMUIOVector *qiov = laiocb->qiov;
379
380 switch (type) {
381 case QEMU_AIO_WRITE:
382 io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
383 break;
384 case QEMU_AIO_READ:
385 io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
386 break;
387 /* Currently Linux kernel does not support other operations */
388 default:
389 fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
390 __func__, type);
391 return -EIO;
392 }
393 io_set_eventfd(&laiocb->iocb, event_notifier_get_fd(&s->e));
394
395 QSIMPLEQ_INSERT_TAIL(&s->io_q.pending, laiocb, next);
396 s->io_q.in_queue++;
397 if (!s->io_q.blocked &&
398 (!s->io_q.plugged ||
399 s->io_q.in_flight + s->io_q.in_queue >= MAX_EVENTS)) {
400 ioq_submit(s);
401 }
402
403 return 0;
404 }
405
406 int coroutine_fn laio_co_submit(BlockDriverState *bs, LinuxAioState *s, int fd,
407 uint64_t offset, QEMUIOVector *qiov, int type)
408 {
409 int ret;
410 struct qemu_laiocb laiocb = {
411 .co = qemu_coroutine_self(),
412 .nbytes = qiov->size,
413 .ctx = s,
414 .ret = -EINPROGRESS,
415 .is_read = (type == QEMU_AIO_READ),
416 .qiov = qiov,
417 };
418
419 ret = laio_do_submit(fd, &laiocb, offset, type);
420 if (ret < 0) {
421 return ret;
422 }
423
424 if (laiocb.ret == -EINPROGRESS) {
425 qemu_coroutine_yield();
426 }
427 return laiocb.ret;
428 }
429
430 BlockAIOCB *laio_submit(BlockDriverState *bs, LinuxAioState *s, int fd,
431 int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
432 BlockCompletionFunc *cb, void *opaque, int type)
433 {
434 struct qemu_laiocb *laiocb;
435 off_t offset = sector_num * BDRV_SECTOR_SIZE;
436 int ret;
437
438 laiocb = qemu_aio_get(&laio_aiocb_info, bs, cb, opaque);
439 laiocb->nbytes = nb_sectors * BDRV_SECTOR_SIZE;
440 laiocb->ctx = s;
441 laiocb->ret = -EINPROGRESS;
442 laiocb->is_read = (type == QEMU_AIO_READ);
443 laiocb->qiov = qiov;
444
445 ret = laio_do_submit(fd, laiocb, offset, type);
446 if (ret < 0) {
447 qemu_aio_unref(laiocb);
448 return NULL;
449 }
450
451 return &laiocb->common;
452 }
453
454 void laio_detach_aio_context(LinuxAioState *s, AioContext *old_context)
455 {
456 aio_set_event_notifier(old_context, &s->e, false, NULL, NULL);
457 qemu_bh_delete(s->completion_bh);
458 }
459
460 void laio_attach_aio_context(LinuxAioState *s, AioContext *new_context)
461 {
462 s->aio_context = new_context;
463 s->completion_bh = aio_bh_new(new_context, qemu_laio_completion_bh, s);
464 aio_set_event_notifier(new_context, &s->e, false,
465 qemu_laio_completion_cb,
466 qemu_laio_poll_cb);
467 }
468
469 LinuxAioState *laio_init(void)
470 {
471 LinuxAioState *s;
472
473 s = g_malloc0(sizeof(*s));
474 if (event_notifier_init(&s->e, false) < 0) {
475 goto out_free_state;
476 }
477
478 if (io_setup(MAX_EVENTS, &s->ctx) != 0) {
479 goto out_close_efd;
480 }
481
482 ioq_init(&s->io_q);
483
484 return s;
485
486 out_close_efd:
487 event_notifier_cleanup(&s->e);
488 out_free_state:
489 g_free(s);
490 return NULL;
491 }
492
493 void laio_cleanup(LinuxAioState *s)
494 {
495 event_notifier_cleanup(&s->e);
496
497 if (io_destroy(s->ctx) != 0) {
498 fprintf(stderr, "%s: destroy AIO context %p failed\n",
499 __func__, &s->ctx);
500 }
501 g_free(s);
502 }