xref: /qemu/block/linux-aio.c (revision a942d8fa)
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 #define MAX_QUEUED_IO  128
32 
33 struct qemu_laiocb {
34     BlockAIOCB common;
35     Coroutine *co;
36     LinuxAioState *ctx;
37     struct iocb iocb;
38     ssize_t ret;
39     size_t nbytes;
40     QEMUIOVector *qiov;
41     bool is_read;
42     QSIMPLEQ_ENTRY(qemu_laiocb) next;
43 };
44 
45 typedef struct {
46     int plugged;
47     unsigned int n;
48     bool blocked;
49     QSIMPLEQ_HEAD(, qemu_laiocb) pending;
50 } LaioQueue;
51 
52 struct LinuxAioState {
53     io_context_t ctx;
54     EventNotifier e;
55 
56     /* io queue for submit at batch */
57     LaioQueue io_q;
58 
59     /* I/O completion processing */
60     QEMUBH *completion_bh;
61     struct io_event events[MAX_EVENTS];
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         qemu_coroutine_enter(laiocb->co, NULL);
98     } else {
99         laiocb->common.cb(laiocb->common.opaque, ret);
100         qemu_aio_unref(laiocb);
101     }
102 }
103 
104 /* The completion BH fetches completed I/O requests and invokes their
105  * callbacks.
106  *
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.
114  */
115 static void qemu_laio_completion_bh(void *opaque)
116 {
117     LinuxAioState *s = opaque;
118 
119     /* Fetch more completion events when empty */
120     if (s->event_idx == s->event_max) {
121         do {
122             struct timespec ts = { 0 };
123             s->event_max = io_getevents(s->ctx, MAX_EVENTS, MAX_EVENTS,
124                                         s->events, &ts);
125         } while (s->event_max == -EINTR);
126 
127         s->event_idx = 0;
128         if (s->event_max <= 0) {
129             s->event_max = 0;
130             return; /* no more events */
131         }
132     }
133 
134     /* Reschedule so nested event loops see currently pending completions */
135     qemu_bh_schedule(s->completion_bh);
136 
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);
142 
143         laiocb->ret = io_event_ret(&s->events[s->event_idx]);
144         s->event_idx++;
145 
146         qemu_laio_process_completion(laiocb);
147     }
148 
149     if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
150         ioq_submit(s);
151     }
152 
153     qemu_bh_cancel(s->completion_bh);
154 }
155 
156 static void qemu_laio_completion_cb(EventNotifier *e)
157 {
158     LinuxAioState *s = container_of(e, LinuxAioState, e);
159 
160     if (event_notifier_test_and_clear(&s->e)) {
161         qemu_laio_completion_bh(s);
162     }
163 }
164 
165 static void laio_cancel(BlockAIOCB *blockacb)
166 {
167     struct qemu_laiocb *laiocb = (struct qemu_laiocb *)blockacb;
168     struct io_event event;
169     int ret;
170 
171     if (laiocb->ret != -EINPROGRESS) {
172         return;
173     }
174     ret = io_cancel(laiocb->ctx->ctx, &laiocb->iocb, &event);
175     laiocb->ret = -ECANCELED;
176     if (ret != 0) {
177         /* iocb is not cancelled, cb will be called by the event loop later */
178         return;
179     }
180 
181     laiocb->common.cb(laiocb->common.opaque, laiocb->ret);
182 }
183 
184 static const AIOCBInfo laio_aiocb_info = {
185     .aiocb_size         = sizeof(struct qemu_laiocb),
186     .cancel_async       = laio_cancel,
187 };
188 
189 static void ioq_init(LaioQueue *io_q)
190 {
191     QSIMPLEQ_INIT(&io_q->pending);
192     io_q->plugged = 0;
193     io_q->n = 0;
194     io_q->blocked = false;
195 }
196 
197 static void ioq_submit(LinuxAioState *s)
198 {
199     int ret, len;
200     struct qemu_laiocb *aiocb;
201     struct iocb *iocbs[MAX_QUEUED_IO];
202     QSIMPLEQ_HEAD(, qemu_laiocb) completed;
203 
204     do {
205         len = 0;
206         QSIMPLEQ_FOREACH(aiocb, &s->io_q.pending, next) {
207             iocbs[len++] = &aiocb->iocb;
208             if (len == MAX_QUEUED_IO) {
209                 break;
210             }
211         }
212 
213         ret = io_submit(s->ctx, len, iocbs);
214         if (ret == -EAGAIN) {
215             break;
216         }
217         if (ret < 0) {
218             abort();
219         }
220 
221         s->io_q.n -= ret;
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);
226 }
227 
228 void laio_io_plug(BlockDriverState *bs, LinuxAioState *s)
229 {
230     assert(!s->io_q.plugged);
231     s->io_q.plugged = 1;
232 }
233 
234 void laio_io_unplug(BlockDriverState *bs, LinuxAioState *s)
235 {
236     assert(s->io_q.plugged);
237     s->io_q.plugged = 0;
238     if (!s->io_q.blocked && !QSIMPLEQ_EMPTY(&s->io_q.pending)) {
239         ioq_submit(s);
240     }
241 }
242 
243 static int laio_do_submit(int fd, struct qemu_laiocb *laiocb, off_t offset,
244                           int type)
245 {
246     LinuxAioState *s = laiocb->ctx;
247     struct iocb *iocbs = &laiocb->iocb;
248     QEMUIOVector *qiov = laiocb->qiov;
249 
250     switch (type) {
251     case QEMU_AIO_WRITE:
252         io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset);
253 	break;
254     case QEMU_AIO_READ:
255         io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset);
256 	break;
257     /* Currently Linux kernel does not support other operations */
258     default:
259         fprintf(stderr, "%s: invalid AIO request type 0x%x.\n",
260                         __func__, type);
261         return -EIO;
262     }
263     io_set_eventfd(&laiocb->iocb, event_notifier_get_fd(&s->e));
264 
265     QSIMPLEQ_INSERT_TAIL(&s->io_q.pending, laiocb, next);
266     s->io_q.n++;
267     if (!s->io_q.blocked &&
268         (!s->io_q.plugged || s->io_q.n >= MAX_QUEUED_IO)) {
269         ioq_submit(s);
270     }
271 
272     return 0;
273 }
274 
275 int coroutine_fn laio_co_submit(BlockDriverState *bs, LinuxAioState *s, int fd,
276                                 uint64_t offset, QEMUIOVector *qiov, int type)
277 {
278     int ret;
279     struct qemu_laiocb laiocb = {
280         .co         = qemu_coroutine_self(),
281         .nbytes     = qiov->size,
282         .ctx        = s,
283         .is_read    = (type == QEMU_AIO_READ),
284         .qiov       = qiov,
285     };
286 
287     ret = laio_do_submit(fd, &laiocb, offset, type);
288     if (ret < 0) {
289         return ret;
290     }
291 
292     qemu_coroutine_yield();
293     return laiocb.ret;
294 }
295 
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)
299 {
300     struct qemu_laiocb *laiocb;
301     off_t offset = sector_num * BDRV_SECTOR_SIZE;
302     int ret;
303 
304     laiocb = qemu_aio_get(&laio_aiocb_info, bs, cb, opaque);
305     laiocb->nbytes = nb_sectors * BDRV_SECTOR_SIZE;
306     laiocb->ctx = s;
307     laiocb->ret = -EINPROGRESS;
308     laiocb->is_read = (type == QEMU_AIO_READ);
309     laiocb->qiov = qiov;
310 
311     ret = laio_do_submit(fd, laiocb, offset, type);
312     if (ret < 0) {
313         qemu_aio_unref(laiocb);
314         return NULL;
315     }
316 
317     return &laiocb->common;
318 }
319 
320 void laio_detach_aio_context(LinuxAioState *s, AioContext *old_context)
321 {
322     aio_set_event_notifier(old_context, &s->e, false, NULL);
323     qemu_bh_delete(s->completion_bh);
324 }
325 
326 void laio_attach_aio_context(LinuxAioState *s, AioContext *new_context)
327 {
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);
331 }
332 
333 LinuxAioState *laio_init(void)
334 {
335     LinuxAioState *s;
336 
337     s = g_malloc0(sizeof(*s));
338     if (event_notifier_init(&s->e, false) < 0) {
339         goto out_free_state;
340     }
341 
342     if (io_setup(MAX_EVENTS, &s->ctx) != 0) {
343         goto out_close_efd;
344     }
345 
346     ioq_init(&s->io_q);
347 
348     return s;
349 
350 out_close_efd:
351     event_notifier_cleanup(&s->e);
352 out_free_state:
353     g_free(s);
354     return NULL;
355 }
356 
357 void laio_cleanup(LinuxAioState *s)
358 {
359     event_notifier_cleanup(&s->e);
360 
361     if (io_destroy(s->ctx) != 0) {
362         fprintf(stderr, "%s: destroy AIO context %p failed\n",
363                         __func__, &s->ctx);
364     }
365     g_free(s);
366 }
367