threadsafe_queue: Fix deadlock

This fixes a lost wakeup in SPSCQueue. If the reader is in just the right position, the writer's notification will be lost and this will be a problem if the writer then does something to wait on the reader.

This was discovered to affect my upcoming stacktrace PR. I don't think any performance decrease will be noticeable because an uncontended mutex is smart enough to skip the syscall. This PR might also resolve some rare deadlocks but I don't know of any examples.
This commit is contained in:
yzct12345 2021-08-13 19:22:51 +00:00 committed by GPUCode
parent a1443356f1
commit 637ade3b25
1 changed files with 4 additions and 6 deletions

View File

@ -48,15 +48,13 @@ public:
ElementPtr* new_ptr = new ElementPtr(); ElementPtr* new_ptr = new ElementPtr();
write_ptr->next.store(new_ptr, std::memory_order_release); write_ptr->next.store(new_ptr, std::memory_order_release);
write_ptr = new_ptr; write_ptr = new_ptr;
++size;
const size_t previous_size{size++}; // cv_mutex must be held or else there will be a missed wakeup if the other thread is in the
// line before cv.wait
// Acquire the mutex and then immediately release it as a fence.
// TODO(bunnei): This can be replaced with C++20 waitable atomics when properly supported. // TODO(bunnei): This can be replaced with C++20 waitable atomics when properly supported.
// See discussion on https://github.com/yuzu-emu/yuzu/pull/3173 for details. // See discussion on https://github.com/yuzu-emu/yuzu/pull/3173 for details.
if (previous_size == 0) { std::lock_guard lock{cv_mutex};
std::lock_guard lock{cv_mutex};
}
cv.notify_one(); cv.notify_one();
} }