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"
12 #include "block_int.h"
13 #include "block/raw-posix-aio.h"
15 #include <sys/eventfd.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
29 BlockDriverAIOCB common;
30 struct qemu_laio_state *ctx;
35 QLIST_ENTRY(qemu_laiocb) node;
38 struct qemu_laio_state {
42 QLIST_HEAD(, qemu_laiocb) completed_reqs;
45 static inline ssize_t io_event_ret(struct io_event *ev)
47 return (ssize_t)(((uint64_t)ev->res2 << 32) | ev->res);
51 * Completes an AIO request (calls the callback and frees the ACB).
52 * Be sure to be in the right AsyncContext before calling this function.
54 static void qemu_laio_process_completion(struct qemu_laio_state *s,
55 struct qemu_laiocb *laiocb)
62 if (ret != -ECANCELED) {
63 if (ret == laiocb->nbytes)
68 laiocb->common.cb(laiocb->common.opaque, ret);
71 qemu_aio_release(laiocb);
75 * Processes all queued AIO requests, i.e. requests that have return from OS
76 * but their callback was not called yet. Requests that cannot have their
77 * callback called in the current AsyncContext, remain in the queue.
79 * Returns 1 if at least one request could be completed, 0 otherwise.
81 static int qemu_laio_process_requests(void *opaque)
83 struct qemu_laio_state *s = opaque;
84 struct qemu_laiocb *laiocb, *next;
87 QLIST_FOREACH_SAFE (laiocb, &s->completed_reqs, node, next) {
88 if (laiocb->async_context_id == get_async_context_id()) {
89 qemu_laio_process_completion(s, laiocb);
90 QLIST_REMOVE(laiocb, node);
99 * Puts a request in the completion queue so that its callback is called the
100 * next time when it's possible. If we already are in the right AsyncContext,
101 * the request is completed immediately instead.
103 static void qemu_laio_enqueue_completed(struct qemu_laio_state *s,
104 struct qemu_laiocb* laiocb)
106 if (laiocb->async_context_id == get_async_context_id()) {
107 qemu_laio_process_completion(s, laiocb);
109 QLIST_INSERT_HEAD(&s->completed_reqs, laiocb, node);
113 static void qemu_laio_completion_cb(void *opaque)
115 struct qemu_laio_state *s = opaque;
118 struct io_event events[MAX_EVENTS];
121 struct timespec ts = { 0 };
125 ret = read(s->efd, &val, sizeof(val));
126 } while (ret == -1 && errno == EINTR);
128 if (ret == -1 && errno == EAGAIN)
135 nevents = io_getevents(s->ctx, val, MAX_EVENTS, events, &ts);
136 } while (nevents == -EINTR);
138 for (i = 0; i < nevents; i++) {
139 struct iocb *iocb = events[i].obj;
140 struct qemu_laiocb *laiocb =
141 container_of(iocb, struct qemu_laiocb, iocb);
143 laiocb->ret = io_event_ret(&events[i]);
144 qemu_laio_enqueue_completed(s, laiocb);
149 static int qemu_laio_flush_cb(void *opaque)
151 struct qemu_laio_state *s = opaque;
153 return (s->count > 0) ? 1 : 0;
156 static void laio_cancel(BlockDriverAIOCB *blockacb)
158 struct qemu_laiocb *laiocb = (struct qemu_laiocb *)blockacb;
159 struct io_event event;
162 if (laiocb->ret != -EINPROGRESS)
166 * Note that as of Linux 2.6.31 neither the block device code nor any
167 * filesystem implements cancellation of AIO request.
168 * Thus the polling loop below is the normal code path.
170 ret = io_cancel(laiocb->ctx->ctx, &laiocb->iocb, &event);
172 laiocb->ret = -ECANCELED;
177 * We have to wait for the iocb to finish.
179 * The only way to get the iocb status update is by polling the io context.
180 * We might be able to do this slightly more optimal by removing the
183 while (laiocb->ret == -EINPROGRESS)
184 qemu_laio_completion_cb(laiocb->ctx);
187 static AIOPool laio_pool = {
188 .aiocb_size = sizeof(struct qemu_laiocb),
189 .cancel = laio_cancel,
192 BlockDriverAIOCB *laio_submit(BlockDriverState *bs, void *aio_ctx, int fd,
193 int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
194 BlockDriverCompletionFunc *cb, void *opaque, int type)
196 struct qemu_laio_state *s = aio_ctx;
197 struct qemu_laiocb *laiocb;
199 off_t offset = sector_num * 512;
201 laiocb = qemu_aio_get(&laio_pool, bs, cb, opaque);
204 laiocb->nbytes = nb_sectors * 512;
206 laiocb->ret = -EINPROGRESS;
207 laiocb->async_context_id = get_async_context_id();
209 iocbs = &laiocb->iocb;
213 io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
216 io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
219 fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
223 io_set_eventfd(&laiocb->iocb, s->efd);
226 if (io_submit(s->ctx, 1, &iocbs) < 0)
228 return &laiocb->common;
231 qemu_aio_release(laiocb);
237 void *laio_init(void)
239 struct qemu_laio_state *s;
241 s = qemu_mallocz(sizeof(*s));
242 QLIST_INIT(&s->completed_reqs);
243 s->efd = eventfd(0, 0);
246 fcntl(s->efd, F_SETFL, O_NONBLOCK);
248 if (io_setup(MAX_EVENTS, &s->ctx) != 0)
251 qemu_aio_set_fd_handler(s->efd, qemu_laio_completion_cb, NULL,
252 qemu_laio_flush_cb, qemu_laio_process_requests, s);