From 7c9b2bf67775ecc1359ce973580807d173e7f710 Mon Sep 17 00:00:00 2001 From: Paolo Bonzini Date: Wed, 12 Aug 2015 15:38:18 +0200 Subject: qemu-thread: add a fast path to the Win32 QemuEvent QemuEvents are used heavily by call_rcu. We do not want them to be slow, but the current implementation does a kernel call on every invocation of qemu_event_* and won't cut it. So, wrap a Win32 manual-reset event with a fast userspace path. The states and transitions are the same as for the futex and mutex/condvar implementations, but the slow path is different of course. The idea is to reset the Win32 event lazily, as part of a test-reset-test-wait sequence. Such a sequence is, indeed, how QemuEvents are used by RCU and other subsystems! The patch includes a formal model of the algorithm. Tested-by: Stefan Weil Signed-off-by: Paolo Bonzini Signed-off-by: Stefan Weil --- util/qemu-thread-win32.c | 66 +++++++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 62 insertions(+), 4 deletions(-) (limited to 'util') diff --git a/util/qemu-thread-win32.c b/util/qemu-thread-win32.c index 406b52f..6cdd553 100644 --- a/util/qemu-thread-win32.c +++ b/util/qemu-thread-win32.c @@ -238,10 +238,34 @@ void qemu_sem_wait(QemuSemaphore *sem) } } +/* Wrap a Win32 manual-reset event with a fast userspace path. The idea + * is to reset the Win32 event lazily, as part of a test-reset-test-wait + * sequence. Such a sequence is, indeed, how QemuEvents are used by + * RCU and other subsystems! + * + * Valid transitions: + * - free->set, when setting the event + * - busy->set, when setting the event, followed by futex_wake + * - set->free, when resetting the event + * - free->busy, when waiting + * + * set->busy does not happen (it can be observed from the outside but + * it really is set->free->busy). + * + * busy->free provably cannot happen; to enforce it, the set->free transition + * is done with an OR, which becomes a no-op if the event has concurrently + * transitioned to free or busy (and is faster than cmpxchg). + */ + +#define EV_SET 0 +#define EV_FREE 1 +#define EV_BUSY -1 + void qemu_event_init(QemuEvent *ev, bool init) { /* Manual reset. */ - ev->event = CreateEvent(NULL, TRUE, init, NULL); + ev->event = CreateEvent(NULL, TRUE, TRUE, NULL); + ev->value = (init ? EV_SET : EV_FREE); } void qemu_event_destroy(QemuEvent *ev) @@ -251,17 +275,51 @@ void qemu_event_destroy(QemuEvent *ev) void qemu_event_set(QemuEvent *ev) { - SetEvent(ev->event); + if (atomic_mb_read(&ev->value) != EV_SET) { + if (atomic_xchg(&ev->value, EV_SET) == EV_BUSY) { + /* There were waiters, wake them up. */ + SetEvent(ev->event); + } + } } void qemu_event_reset(QemuEvent *ev) { - ResetEvent(ev->event); + if (atomic_mb_read(&ev->value) == EV_SET) { + /* If there was a concurrent reset (or even reset+wait), + * do nothing. Otherwise change EV_SET->EV_FREE. + */ + atomic_or(&ev->value, EV_FREE); + } } void qemu_event_wait(QemuEvent *ev) { - WaitForSingleObject(ev->event, INFINITE); + unsigned value; + + value = atomic_mb_read(&ev->value); + if (value != EV_SET) { + if (value == EV_FREE) { + /* qemu_event_set is not yet going to call SetEvent, but we are + * going to do another check for EV_SET below when setting EV_BUSY. + * At that point it is safe to call WaitForSingleObject. + */ + ResetEvent(ev->event); + + /* Tell qemu_event_set that there are waiters. No need to retry + * because there cannot be a concurent busy->free transition. + * After the CAS, the event will be either set or busy. + */ + if (atomic_cmpxchg(&ev->value, EV_FREE, EV_BUSY) == EV_SET) { + value = EV_SET; + } else { + value = EV_BUSY; + } + } + if (value == EV_BUSY) { + WaitForSingleObject(ev->event, INFINITE); + } + } } struct QemuThreadData { -- cgit v1.1