blob: 4309c161ff62c7974e693c5cdeda339c2ebe1904 [file] [log] [blame]
Paolo Bonzinif42b2202012-06-09 04:01:51 +02001/*
2 * QEMU aio implementation
3 *
4 * Copyright IBM Corp., 2008
5 * Copyright Red Hat Inc., 2012
6 *
7 * Authors:
8 * Anthony Liguori <aliguori@us.ibm.com>
9 * Paolo Bonzini <pbonzini@redhat.com>
10 *
11 * This work is licensed under the terms of the GNU GPL, version 2. See
12 * the COPYING file in the top-level directory.
13 *
14 * Contributions after 2012-01-13 are licensed under the terms of the
15 * GNU GPL, version 2 or (at your option) any later version.
16 */
17
18#include "qemu-common.h"
Paolo Bonzini737e1502012-12-17 18:19:44 +010019#include "block/block.h"
Paolo Bonzini1de7afc2012-12-17 18:20:00 +010020#include "qemu/queue.h"
21#include "qemu/sockets.h"
Paolo Bonzinif42b2202012-06-09 04:01:51 +020022
23struct AioHandler {
24 EventNotifier *e;
25 EventNotifierHandler *io_notify;
Paolo Bonzinif42b2202012-06-09 04:01:51 +020026 GPollFD pfd;
27 int deleted;
28 QLIST_ENTRY(AioHandler) node;
29};
30
31void aio_set_event_notifier(AioContext *ctx,
32 EventNotifier *e,
33 EventNotifierHandler *io_notify,
34 AioFlushEventNotifierHandler *io_flush)
35{
36 AioHandler *node;
37
38 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
39 if (node->e == e && !node->deleted) {
40 break;
41 }
42 }
43
44 /* Are we deleting the fd handler? */
45 if (!io_notify) {
46 if (node) {
Paolo Bonzinie3713e02012-09-24 14:57:41 +020047 g_source_remove_poll(&ctx->source, &node->pfd);
48
Paolo Bonzinif42b2202012-06-09 04:01:51 +020049 /* If the lock is held, just mark the node as deleted */
50 if (ctx->walking_handlers) {
51 node->deleted = 1;
52 node->pfd.revents = 0;
53 } else {
54 /* Otherwise, delete it for real. We can't just mark it as
55 * deleted because deleted nodes are only cleaned up after
56 * releasing the walking_handlers lock.
57 */
58 QLIST_REMOVE(node, node);
59 g_free(node);
60 }
61 }
62 } else {
63 if (node == NULL) {
64 /* Alloc and insert if it's not already there */
65 node = g_malloc0(sizeof(AioHandler));
66 node->e = e;
67 node->pfd.fd = (uintptr_t)event_notifier_get_handle(e);
68 node->pfd.events = G_IO_IN;
69 QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node);
Paolo Bonzinie3713e02012-09-24 14:57:41 +020070
71 g_source_add_poll(&ctx->source, &node->pfd);
Paolo Bonzinif42b2202012-06-09 04:01:51 +020072 }
73 /* Update handler with latest information */
74 node->io_notify = io_notify;
Paolo Bonzinif42b2202012-06-09 04:01:51 +020075 }
Paolo Bonzini7ed2b242012-09-25 10:22:39 +020076
77 aio_notify(ctx);
Paolo Bonzinif42b2202012-06-09 04:01:51 +020078}
79
80bool aio_pending(AioContext *ctx)
81{
82 AioHandler *node;
83
84 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
85 if (node->pfd.revents && node->io_notify) {
86 return true;
87 }
88 }
89
90 return false;
91}
92
93bool aio_poll(AioContext *ctx, bool blocking)
94{
95 AioHandler *node;
96 HANDLE events[MAXIMUM_WAIT_OBJECTS + 1];
Stefan Hajnoczi164a1012013-04-11 16:56:50 +020097 bool progress;
Paolo Bonzinif42b2202012-06-09 04:01:51 +020098 int count;
99
100 progress = false;
101
102 /*
103 * If there are callbacks left that have been queued, we need to call then.
104 * Do not call select in this case, because it is possible that the caller
105 * does not need a complete flush (as is the case for qemu_aio_wait loops).
106 */
107 if (aio_bh_poll(ctx)) {
108 blocking = false;
109 progress = true;
110 }
111
112 /*
113 * Then dispatch any pending callbacks from the GSource.
114 *
115 * We have to walk very carefully in case qemu_aio_set_fd_handler is
116 * called while we're walking.
117 */
118 node = QLIST_FIRST(&ctx->aio_handlers);
119 while (node) {
120 AioHandler *tmp;
121
122 ctx->walking_handlers++;
123
124 if (node->pfd.revents && node->io_notify) {
125 node->pfd.revents = 0;
126 node->io_notify(node->e);
Stefan Hajnoczi164a1012013-04-11 16:56:50 +0200127
128 /* aio_notify() does not count as progress */
129 if (node->opaque != &ctx->notifier) {
130 progress = true;
131 }
Paolo Bonzinif42b2202012-06-09 04:01:51 +0200132 }
133
134 tmp = node;
135 node = QLIST_NEXT(node, node);
136
137 ctx->walking_handlers--;
138
139 if (!ctx->walking_handlers && tmp->deleted) {
140 QLIST_REMOVE(tmp, node);
141 g_free(tmp);
142 }
143 }
144
145 if (progress && !blocking) {
146 return true;
147 }
148
149 ctx->walking_handlers++;
150
151 /* fill fd sets */
Paolo Bonzinif42b2202012-06-09 04:01:51 +0200152 count = 0;
153 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
Paolo Bonzinif42b2202012-06-09 04:01:51 +0200154 if (!node->deleted && node->io_notify) {
155 events[count++] = event_notifier_get_handle(node->e);
156 }
157 }
158
159 ctx->walking_handlers--;
160
Stefan Hajnoczi164a1012013-04-11 16:56:50 +0200161 /* early return if we only have the aio_notify() fd */
162 if (count == 1) {
Paolo Bonzinif42b2202012-06-09 04:01:51 +0200163 return progress;
164 }
165
166 /* wait until next event */
Paolo Bonzinib022b4a2012-11-23 15:59:43 +0100167 while (count > 0) {
Paolo Bonzinif42b2202012-06-09 04:01:51 +0200168 int timeout = blocking ? INFINITE : 0;
169 int ret = WaitForMultipleObjects(count, events, FALSE, timeout);
170
171 /* if we have any signaled events, dispatch event */
172 if ((DWORD) (ret - WAIT_OBJECT_0) >= count) {
173 break;
174 }
175
176 blocking = false;
177
178 /* we have to walk very carefully in case
179 * qemu_aio_set_fd_handler is called while we're walking */
180 node = QLIST_FIRST(&ctx->aio_handlers);
181 while (node) {
182 AioHandler *tmp;
183
184 ctx->walking_handlers++;
185
186 if (!node->deleted &&
187 event_notifier_get_handle(node->e) == events[ret - WAIT_OBJECT_0] &&
188 node->io_notify) {
189 node->io_notify(node->e);
Stefan Hajnoczi164a1012013-04-11 16:56:50 +0200190
191 /* aio_notify() does not count as progress */
192 if (node->opaque != &ctx->notifier) {
193 progress = true;
194 }
Paolo Bonzinif42b2202012-06-09 04:01:51 +0200195 }
196
197 tmp = node;
198 node = QLIST_NEXT(node, node);
199
200 ctx->walking_handlers--;
201
202 if (!ctx->walking_handlers && tmp->deleted) {
203 QLIST_REMOVE(tmp, node);
204 g_free(tmp);
205 }
206 }
Paolo Bonzinib022b4a2012-11-23 15:59:43 +0100207
208 /* Try again, but only call each handler once. */
209 events[ret - WAIT_OBJECT_0] = events[--count];
Paolo Bonzinif42b2202012-06-09 04:01:51 +0200210 }
211
Stefan Hajnoczi164a1012013-04-11 16:56:50 +0200212 return progress;
Paolo Bonzinif42b2202012-06-09 04:01:51 +0200213}