2 * QEMU aio implementation
4 * Copyright IBM Corp., 2008
5 * Copyright Red Hat Inc., 2012
11 * This work is licensed under the terms of the GNU GPL, version 2. See
12 * the COPYING file in the top-level directory.
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.
18 #include "qemu/osdep.h"
19 #include "qemu-common.h"
20 #include "block/block.h"
21 #include "qemu/queue.h"
22 #include "qemu/sockets.h"
23 #include "qapi/error.h"
24 #include "qemu/rcu_queue.h"
30 EventNotifierHandler *io_notify;
35 QLIST_ENTRY(AioHandler) node;
38 void aio_set_fd_handler(AioContext *ctx,
46 /* fd is a SOCKET in our case */
49 qemu_lockcnt_lock(&ctx->list_lock);
50 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
51 if (node->pfd.fd == fd && !node->deleted) {
56 /* Are we deleting the fd handler? */
57 if (!io_read && !io_write) {
59 /* If aio_poll is in progress, just mark the node as deleted */
60 if (qemu_lockcnt_count(&ctx->list_lock)) {
62 node->pfd.revents = 0;
64 /* Otherwise, delete it for real. We can't just mark it as
65 * deleted because deleted nodes are only cleaned up after
66 * releasing the list_lock.
68 QLIST_REMOVE(node, node);
77 /* Alloc and insert if it's not already there */
78 node = g_new0(AioHandler, 1);
80 QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, node, node);
85 node->pfd.events |= G_IO_IN;
88 node->pfd.events |= G_IO_OUT;
91 node->e = &ctx->notifier;
93 /* Update handler with latest information */
94 node->opaque = opaque;
95 node->io_read = io_read;
96 node->io_write = io_write;
97 node->is_external = is_external;
100 bitmask |= FD_READ | FD_ACCEPT | FD_CLOSE;
104 bitmask |= FD_WRITE | FD_CONNECT;
107 event = event_notifier_get_handle(&ctx->notifier);
108 WSAEventSelect(node->pfd.fd, event, bitmask);
111 qemu_lockcnt_unlock(&ctx->list_lock);
115 void aio_set_fd_poll(AioContext *ctx, int fd,
116 IOHandler *io_poll_begin,
117 IOHandler *io_poll_end)
119 /* Not implemented */
122 void aio_set_event_notifier(AioContext *ctx,
125 EventNotifierHandler *io_notify,
130 qemu_lockcnt_lock(&ctx->list_lock);
131 QLIST_FOREACH(node, &ctx->aio_handlers, node) {
132 if (node->e == e && !node->deleted) {
137 /* Are we deleting the fd handler? */
140 g_source_remove_poll(&ctx->source, &node->pfd);
142 /* aio_poll is in progress, just mark the node as deleted */
143 if (qemu_lockcnt_count(&ctx->list_lock)) {
145 node->pfd.revents = 0;
147 /* Otherwise, delete it for real. We can't just mark it as
148 * deleted because deleted nodes are only cleaned up after
149 * releasing the list_lock.
151 QLIST_REMOVE(node, node);
157 /* Alloc and insert if it's not already there */
158 node = g_new0(AioHandler, 1);
160 node->pfd.fd = (uintptr_t)event_notifier_get_handle(e);
161 node->pfd.events = G_IO_IN;
162 node->is_external = is_external;
163 QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, node, node);
165 g_source_add_poll(&ctx->source, &node->pfd);
167 /* Update handler with latest information */
168 node->io_notify = io_notify;
171 qemu_lockcnt_unlock(&ctx->list_lock);
175 void aio_set_event_notifier_poll(AioContext *ctx,
176 EventNotifier *notifier,
177 EventNotifierHandler *io_poll_begin,
178 EventNotifierHandler *io_poll_end)
180 /* Not implemented */
183 bool aio_prepare(AioContext *ctx)
185 static struct timeval tv0;
187 bool have_select_revents = false;
191 * We have to walk very carefully in case aio_set_fd_handler is
192 * called while we're walking.
194 qemu_lockcnt_inc(&ctx->list_lock);
199 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
201 FD_SET ((SOCKET)node->pfd.fd, &rfds);
203 if (node->io_write) {
204 FD_SET ((SOCKET)node->pfd.fd, &wfds);
208 if (select(0, &rfds, &wfds, NULL, &tv0) > 0) {
209 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
210 node->pfd.revents = 0;
211 if (FD_ISSET(node->pfd.fd, &rfds)) {
212 node->pfd.revents |= G_IO_IN;
213 have_select_revents = true;
216 if (FD_ISSET(node->pfd.fd, &wfds)) {
217 node->pfd.revents |= G_IO_OUT;
218 have_select_revents = true;
223 qemu_lockcnt_dec(&ctx->list_lock);
224 return have_select_revents;
227 bool aio_pending(AioContext *ctx)
233 * We have to walk very carefully in case aio_set_fd_handler is
234 * called while we're walking.
236 qemu_lockcnt_inc(&ctx->list_lock);
237 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
238 if (node->pfd.revents && node->io_notify) {
243 if ((node->pfd.revents & G_IO_IN) && node->io_read) {
247 if ((node->pfd.revents & G_IO_OUT) && node->io_write) {
253 qemu_lockcnt_dec(&ctx->list_lock);
257 static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event)
260 bool progress = false;
264 * We have to walk very carefully in case aio_set_fd_handler is
265 * called while we're walking.
267 QLIST_FOREACH_SAFE_RCU(node, &ctx->aio_handlers, node, tmp) {
268 int revents = node->pfd.revents;
270 if (!node->deleted &&
271 (revents || event_notifier_get_handle(node->e) == event) &&
273 node->pfd.revents = 0;
274 node->io_notify(node->e);
276 /* aio_notify() does not count as progress */
277 if (node->e != &ctx->notifier) {
282 if (!node->deleted &&
283 (node->io_read || node->io_write)) {
284 node->pfd.revents = 0;
285 if ((revents & G_IO_IN) && node->io_read) {
286 node->io_read(node->opaque);
289 if ((revents & G_IO_OUT) && node->io_write) {
290 node->io_write(node->opaque);
294 /* if the next select() will return an event, we have progressed */
295 if (event == event_notifier_get_handle(&ctx->notifier)) {
297 WSAEnumNetworkEvents(node->pfd.fd, event, &ev);
298 if (ev.lNetworkEvents) {
305 if (qemu_lockcnt_dec_if_lock(&ctx->list_lock)) {
306 QLIST_REMOVE(node, node);
308 qemu_lockcnt_inc_and_unlock(&ctx->list_lock);
316 void aio_dispatch(AioContext *ctx)
318 qemu_lockcnt_inc(&ctx->list_lock);
320 aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE);
321 qemu_lockcnt_dec(&ctx->list_lock);
322 timerlistgroup_run_timers(&ctx->tlg);
325 bool aio_poll(AioContext *ctx, bool blocking)
328 HANDLE events[MAXIMUM_WAIT_OBJECTS + 1];
329 bool progress, have_select_revents, first;
335 /* aio_notify can avoid the expensive event_notifier_set if
336 * everything (file descriptors, bottom halves, timers) will
337 * be re-evaluated before the next blocking poll(). This is
338 * already true when aio_poll is called with blocking == false;
339 * if blocking == true, it is only true after poll() returns,
340 * so disable the optimization now.
343 atomic_add(&ctx->notify_me, 2);
346 qemu_lockcnt_inc(&ctx->list_lock);
347 have_select_revents = aio_prepare(ctx);
351 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
352 if (!node->deleted && node->io_notify
353 && aio_node_check(ctx, node->is_external)) {
354 events[count++] = event_notifier_get_handle(node->e);
360 /* ctx->notifier is always registered. */
363 /* Multiple iterations, all of them non-blocking except the first,
364 * may be necessary to process all pending events. After the first
365 * WaitForMultipleObjects call ctx->notify_me will be decremented.
371 timeout = blocking && !have_select_revents
372 ? qemu_timeout_ns_to_ms(aio_compute_timeout(ctx)) : 0;
373 ret = WaitForMultipleObjects(count, events, FALSE, timeout);
376 assert(in_aio_context_home_thread(ctx));
377 atomic_sub(&ctx->notify_me, 2);
378 aio_notify_accept(ctx);
382 progress |= aio_bh_poll(ctx);
386 /* if we have any signaled events, dispatch event */
388 if ((DWORD) (ret - WAIT_OBJECT_0) < count) {
389 event = events[ret - WAIT_OBJECT_0];
390 events[ret - WAIT_OBJECT_0] = events[--count];
391 } else if (!have_select_revents) {
395 have_select_revents = false;
398 progress |= aio_dispatch_handlers(ctx, event);
401 qemu_lockcnt_dec(&ctx->list_lock);
403 progress |= timerlistgroup_run_timers(&ctx->tlg);
407 void aio_context_setup(AioContext *ctx)
411 void aio_context_destroy(AioContext *ctx)
415 void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns,
416 int64_t grow, int64_t shrink, Error **errp)
419 error_setg(errp, "AioContext polling is not implemented on Windows");