2 * Linux native AIO support.
4 * Copyright (C) 2009 IBM, Corp.
5 * Copyright (C) 2009 Red Hat, Inc.
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.
10 #include "qemu-common.h"
11 #include "block/aio.h"
12 #include "qemu/queue.h"
13 #include "block/raw-aio.h"
14 #include "qemu/event_notifier.h"
19 * Queue size (per-device).
21 * XXX: eventually we need to communicate this to the guest and/or make it
22 * tunable by the guest. If we get more outstanding requests at a time
23 * than this we will get EAGAIN from io_submit which is communicated to
24 * the guest as an I/O error.
26 #define MAX_EVENTS 128
28 #define MAX_QUEUED_IO 128
32 struct qemu_laio_state *ctx;
38 QSIMPLEQ_ENTRY(qemu_laiocb) next;
45 QSIMPLEQ_HEAD(, qemu_laiocb) pending;
48 struct qemu_laio_state {
52 /* io queue for submit at batch */
55 /* I/O completion processing */
56 QEMUBH *completion_bh;
57 struct io_event events[MAX_EVENTS];
62 static int ioq_submit(struct qemu_laio_state *s);
64 static inline ssize_t io_event_ret(struct io_event *ev)
66 return (ssize_t)(((uint64_t)ev->res2 << 32) | ev->res);
70 * Completes an AIO request (calls the callback and frees the ACB).
72 static void qemu_laio_process_completion(struct qemu_laio_state *s,
73 struct qemu_laiocb *laiocb)
78 if (ret != -ECANCELED) {
79 if (ret == laiocb->nbytes) {
81 } else if (ret >= 0) {
82 /* Short reads mean EOF, pad with zeros. */
83 if (laiocb->is_read) {
84 qemu_iovec_memset(laiocb->qiov, ret, 0,
85 laiocb->qiov->size - ret);
91 laiocb->common.cb(laiocb->common.opaque, ret);
93 qemu_aio_unref(laiocb);
96 /* The completion BH fetches completed I/O requests and invokes their
99 * The function is somewhat tricky because it supports nested event loops, for
100 * example when a request callback invokes aio_poll(). In order to do this,
101 * the completion events array and index are kept in qemu_laio_state. The BH
102 * reschedules itself as long as there are completions pending so it will
103 * either be called again in a nested event loop or will be called after all
104 * events have been completed. When there are no events left to complete, the
105 * BH returns without rescheduling.
107 static void qemu_laio_completion_bh(void *opaque)
109 struct qemu_laio_state *s = opaque;
111 /* Fetch more completion events when empty */
112 if (s->event_idx == s->event_max) {
114 struct timespec ts = { 0 };
115 s->event_max = io_getevents(s->ctx, MAX_EVENTS, MAX_EVENTS,
117 } while (s->event_max == -EINTR);
120 if (s->event_max <= 0) {
122 return; /* no more events */
126 /* Reschedule so nested event loops see currently pending completions */
127 qemu_bh_schedule(s->completion_bh);
129 /* Process completion events */
130 while (s->event_idx < s->event_max) {
131 struct iocb *iocb = s->events[s->event_idx].obj;
132 struct qemu_laiocb *laiocb =
133 container_of(iocb, struct qemu_laiocb, iocb);
135 laiocb->ret = io_event_ret(&s->events[s->event_idx]);
138 qemu_laio_process_completion(s, laiocb);
141 if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
146 static void qemu_laio_completion_cb(EventNotifier *e)
148 struct qemu_laio_state *s = container_of(e, struct qemu_laio_state, e);
150 if (event_notifier_test_and_clear(&s->e)) {
151 qemu_bh_schedule(s->completion_bh);
155 static void laio_cancel(BlockAIOCB *blockacb)
157 struct qemu_laiocb *laiocb = (struct qemu_laiocb *)blockacb;
158 struct io_event event;
161 if (laiocb->ret != -EINPROGRESS) {
164 ret = io_cancel(laiocb->ctx->ctx, &laiocb->iocb, &event);
165 laiocb->ret = -ECANCELED;
167 /* iocb is not cancelled, cb will be called by the event loop later */
171 laiocb->common.cb(laiocb->common.opaque, laiocb->ret);
174 static const AIOCBInfo laio_aiocb_info = {
175 .aiocb_size = sizeof(struct qemu_laiocb),
176 .cancel_async = laio_cancel,
179 static void ioq_init(LaioQueue *io_q)
181 QSIMPLEQ_INIT(&io_q->pending);
184 io_q->blocked = false;
187 static int ioq_submit(struct qemu_laio_state *s)
190 struct qemu_laiocb *aiocb;
191 struct iocb *iocbs[MAX_QUEUED_IO];
195 QSIMPLEQ_FOREACH(aiocb, &s->io_q.pending, next) {
196 iocbs[len++] = &aiocb->iocb;
197 if (len == MAX_QUEUED_IO) {
202 ret = io_submit(s->ctx, len, iocbs);
203 if (ret == -EAGAIN) {
210 for (i = 0; i < ret; i++) {
212 QSIMPLEQ_REMOVE_HEAD(&s->io_q.pending, next);
214 } while (ret == len && !QSIMPLEQ_EMPTY(&s->io_q.pending));
215 s->io_q.blocked = (s->io_q.idx > 0);
220 void laio_io_plug(BlockDriverState *bs, void *aio_ctx)
222 struct qemu_laio_state *s = aio_ctx;
227 int laio_io_unplug(BlockDriverState *bs, void *aio_ctx, bool unplug)
229 struct qemu_laio_state *s = aio_ctx;
232 assert(s->io_q.plugged > 0 || !unplug);
234 if (unplug && --s->io_q.plugged > 0) {
238 if (!s->io_q.blocked && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
245 BlockAIOCB *laio_submit(BlockDriverState *bs, void *aio_ctx, int fd,
246 int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
247 BlockCompletionFunc *cb, void *opaque, int type)
249 struct qemu_laio_state *s = aio_ctx;
250 struct qemu_laiocb *laiocb;
252 off_t offset = sector_num * 512;
254 laiocb = qemu_aio_get(&laio_aiocb_info, bs, cb, opaque);
255 laiocb->nbytes = nb_sectors * 512;
257 laiocb->ret = -EINPROGRESS;
258 laiocb->is_read = (type == QEMU_AIO_READ);
261 iocbs = &laiocb->iocb;
265 io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
268 io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
270 /* Currently Linux kernel does not support other operations */
272 fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
276 io_set_eventfd(&laiocb->iocb, event_notifier_get_fd(&s->e));
278 QSIMPLEQ_INSERT_TAIL(&s->io_q.pending, laiocb, next);
280 if (!s->io_q.blocked &&
281 (!s->io_q.plugged || s->io_q.idx >= MAX_QUEUED_IO)) {
284 return &laiocb->common;
287 qemu_aio_unref(laiocb);
291 void laio_detach_aio_context(void *s_, AioContext *old_context)
293 struct qemu_laio_state *s = s_;
295 aio_set_event_notifier(old_context, &s->e, NULL);
296 qemu_bh_delete(s->completion_bh);
299 void laio_attach_aio_context(void *s_, AioContext *new_context)
301 struct qemu_laio_state *s = s_;
303 s->completion_bh = aio_bh_new(new_context, qemu_laio_completion_bh, s);
304 aio_set_event_notifier(new_context, &s->e, qemu_laio_completion_cb);
307 void *laio_init(void)
309 struct qemu_laio_state *s;
311 s = g_malloc0(sizeof(*s));
312 if (event_notifier_init(&s->e, false) < 0) {
316 if (io_setup(MAX_EVENTS, &s->ctx) != 0) {
325 event_notifier_cleanup(&s->e);
331 void laio_cleanup(void *s_)
333 struct qemu_laio_state *s = s_;
335 event_notifier_cleanup(&s->e);
337 if (io_destroy(s->ctx) != 0) {
338 fprintf(stderr, "%s: destroy AIO context %p failed\n",