On Sat, 07/18 22:21, Paolo Bonzini wrote:
> It is pretty rare for aio_notify to actually set the EventNotifier. It
> can happen with worker threads such as thread-pool.c's, but otherwise it
> should never be set thanks to the ctx->notify_me optimization. The
> previous patch, unfortunately, added an unconditional call to
> event_notifier_test_and_clear; now add a userspace fast path that
> avoids the call.
>
> Note that it is not possible to do the same with event_notifier_set;
> it would break, as proved (again) by the included formal model.
>
> This patch survived over 800 reboots on aarch64 KVM.
For aio-posix, how about keeping the optimization local which doesn't need
atomic operation? (no idea for win32 :)
diff --git a/aio-posix.c b/aio-posix.c
index 5c8b266..7e98123 100644
--- a/aio-posix.c
+++ b/aio-posix.c
@@ -236,6 +236,7 @@ bool aio_poll(AioContext *ctx, bool blocking)
int i, ret;
bool progress;
int64_t timeout;
+ int aio_notifier_idx = -1;
aio_context_acquire(ctx);
progress = false;
@@ -256,11 +257,18 @@ bool aio_poll(AioContext *ctx, bool blocking)
assert(npfd == 0);
/* fill pollfds */
+ i = 0;
QLIST_FOREACH(node, &ctx->aio_handlers, node) {
if (!node->deleted && node->pfd.events) {
add_pollfd(node);
+ if (node->pfd.fd == event_notifier_get_fd(&ctx->notifier)) {
+ assert(aio_notifier_idx == -1);
+ aio_notifier_idx = i;
+ }
+ i++;
}
}
+ assert(aio_notifier_idx != -1);
timeout = blocking ? aio_compute_timeout(ctx) : 0;
@@ -276,7 +284,9 @@ bool aio_poll(AioContext *ctx, bool blocking)
aio_context_acquire(ctx);
}
- event_notifier_test_and_clear(&ctx->notifier);
+ if (pollfds[aio_notifier_idx].revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) {
+ event_notifier_test_and_clear(&ctx->notifier);
+ }
/* if we have any readable fds, dispatch event */
if (ret > 0) {
