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/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"
22 * Queue size (per-device).
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.
29 #define MAX_EVENTS 128
31 #define MAX_QUEUED_IO 128
42 QSIMPLEQ_ENTRY(qemu_laiocb) next;
49 QSIMPLEQ_HEAD(, qemu_laiocb) pending;
52 struct LinuxAioState {
56 /* io queue for submit at batch */
59 /* I/O completion processing */
60 QEMUBH *completion_bh;
61 struct io_event events[MAX_EVENTS];
66 static void ioq_submit(LinuxAioState *s);
68 static inline ssize_t io_event_ret(struct io_event *ev)
70 return (ssize_t)(((uint64_t)ev->res2 << 32) | ev->res);
74 * Completes an AIO request (calls the callback and frees the ACB).
76 static void qemu_laio_process_completion(struct qemu_laiocb *laiocb)
81 if (ret != -ECANCELED) {
82 if (ret == laiocb->nbytes) {
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);
97 qemu_coroutine_enter(laiocb->co, NULL);
99 laiocb->common.cb(laiocb->common.opaque, ret);
100 qemu_aio_unref(laiocb);
104 /* The completion BH fetches completed I/O requests and invokes their
107 * The function is somewhat tricky because it supports nested event loops, for
108 * example when a request callback invokes aio_poll(). In order to do this,
109 * the completion events array and index are kept in LinuxAioState. The BH
110 * reschedules itself as long as there are completions pending so it will
111 * either be called again in a nested event loop or will be called after all
112 * events have been completed. When there are no events left to complete, the
113 * BH returns without rescheduling.
115 static void qemu_laio_completion_bh(void *opaque)
117 LinuxAioState *s = opaque;
119 /* Fetch more completion events when empty */
120 if (s->event_idx == s->event_max) {
122 struct timespec ts = { 0 };
123 s->event_max = io_getevents(s->ctx, MAX_EVENTS, MAX_EVENTS,
125 } while (s->event_max == -EINTR);
128 if (s->event_max <= 0) {
130 return; /* no more events */
134 /* Reschedule so nested event loops see currently pending completions */
135 qemu_bh_schedule(s->completion_bh);
137 /* Process completion events */
138 while (s->event_idx < s->event_max) {
139 struct iocb *iocb = s->events[s->event_idx].obj;
140 struct qemu_laiocb *laiocb =
141 container_of(iocb, struct qemu_laiocb, iocb);
143 laiocb->ret = io_event_ret(&s->events[s->event_idx]);
146 qemu_laio_process_completion(laiocb);
149 if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
153 qemu_bh_cancel(s->completion_bh);
156 static void qemu_laio_completion_cb(EventNotifier *e)
158 LinuxAioState *s = container_of(e, LinuxAioState, e);
160 if (event_notifier_test_and_clear(&s->e)) {
161 qemu_laio_completion_bh(s);
165 static void laio_cancel(BlockAIOCB *blockacb)
167 struct qemu_laiocb *laiocb = (struct qemu_laiocb *)blockacb;
168 struct io_event event;
171 if (laiocb->ret != -EINPROGRESS) {
174 ret = io_cancel(laiocb->ctx->ctx, &laiocb->iocb, &event);
175 laiocb->ret = -ECANCELED;
177 /* iocb is not cancelled, cb will be called by the event loop later */
181 laiocb->common.cb(laiocb->common.opaque, laiocb->ret);
184 static const AIOCBInfo laio_aiocb_info = {
185 .aiocb_size = sizeof(struct qemu_laiocb),
186 .cancel_async = laio_cancel,
189 static void ioq_init(LaioQueue *io_q)
191 QSIMPLEQ_INIT(&io_q->pending);
194 io_q->blocked = false;
197 static void ioq_submit(LinuxAioState *s)
200 struct qemu_laiocb *aiocb;
201 struct iocb *iocbs[MAX_QUEUED_IO];
202 QSIMPLEQ_HEAD(, qemu_laiocb) completed;
206 QSIMPLEQ_FOREACH(aiocb, &s->io_q.pending, next) {
207 iocbs[len++] = &aiocb->iocb;
208 if (len == MAX_QUEUED_IO) {
213 ret = io_submit(s->ctx, len, iocbs);
214 if (ret == -EAGAIN) {
222 aiocb = container_of(iocbs[ret - 1], struct qemu_laiocb, iocb);
223 QSIMPLEQ_SPLIT_AFTER(&s->io_q.pending, aiocb, next, &completed);
224 } while (ret == len && !QSIMPLEQ_EMPTY(&s->io_q.pending));
225 s->io_q.blocked = (s->io_q.n > 0);
228 void laio_io_plug(BlockDriverState *bs, LinuxAioState *s)
230 assert(!s->io_q.plugged);
234 void laio_io_unplug(BlockDriverState *bs, LinuxAioState *s)
236 assert(s->io_q.plugged);
238 if (!s->io_q.blocked && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
243 static int laio_do_submit(int fd, struct qemu_laiocb *laiocb, off_t offset,
246 LinuxAioState *s = laiocb->ctx;
247 struct iocb *iocbs = &laiocb->iocb;
248 QEMUIOVector *qiov = laiocb->qiov;
252 io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
255 io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
257 /* Currently Linux kernel does not support other operations */
259 fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
263 io_set_eventfd(&laiocb->iocb, event_notifier_get_fd(&s->e));
265 QSIMPLEQ_INSERT_TAIL(&s->io_q.pending, laiocb, next);
267 if (!s->io_q.blocked &&
268 (!s->io_q.plugged || s->io_q.n >= MAX_QUEUED_IO)) {
275 int coroutine_fn laio_co_submit(BlockDriverState *bs, LinuxAioState *s, int fd,
276 uint64_t offset, QEMUIOVector *qiov, int type)
279 struct qemu_laiocb laiocb = {
280 .co = qemu_coroutine_self(),
281 .nbytes = qiov->size,
283 .is_read = (type == QEMU_AIO_READ),
287 ret = laio_do_submit(fd, &laiocb, offset, type);
292 qemu_coroutine_yield();
296 BlockAIOCB *laio_submit(BlockDriverState *bs, LinuxAioState *s, int fd,
297 int64_t sector_num, QEMUIOVector *qiov, int nb_sectors,
298 BlockCompletionFunc *cb, void *opaque, int type)
300 struct qemu_laiocb *laiocb;
301 off_t offset = sector_num * BDRV_SECTOR_SIZE;
304 laiocb = qemu_aio_get(&laio_aiocb_info, bs, cb, opaque);
305 laiocb->nbytes = nb_sectors * BDRV_SECTOR_SIZE;
307 laiocb->ret = -EINPROGRESS;
308 laiocb->is_read = (type == QEMU_AIO_READ);
311 ret = laio_do_submit(fd, laiocb, offset, type);
313 qemu_aio_unref(laiocb);
317 return &laiocb->common;
320 void laio_detach_aio_context(LinuxAioState *s, AioContext *old_context)
322 aio_set_event_notifier(old_context, &s->e, false, NULL);
323 qemu_bh_delete(s->completion_bh);
326 void laio_attach_aio_context(LinuxAioState *s, AioContext *new_context)
328 s->completion_bh = aio_bh_new(new_context, qemu_laio_completion_bh, s);
329 aio_set_event_notifier(new_context, &s->e, false,
330 qemu_laio_completion_cb);
333 LinuxAioState *laio_init(void)
337 s = g_malloc0(sizeof(*s));
338 if (event_notifier_init(&s->e, false) < 0) {
342 if (io_setup(MAX_EVENTS, &s->ctx) != 0) {
351 event_notifier_cleanup(&s->e);
357 void laio_cleanup(LinuxAioState *s)
359 event_notifier_cleanup(&s->e);
361 if (io_destroy(s->ctx) != 0) {
362 fprintf(stderr, "%s: destroy AIO context %p failed\n",