fix(userspace/falco): fix watchdog race condition on timeout exchange

The watchdog thread and stop() consume the timeout pointer with
m_timeout.exchange(nullptr, ...). That exchange was using
memory_order_release. The load part of the RMW needs acquire
semantics so it synchronizes-with the release store in
set_timeout()/cancel_timeout(); otherwise the consumer can see
the pointer value without seeing the writes that initialized
the timeout_data and payload (data race).
Use memory_order_acq_rel on the consumer exchanges so the load
synchronizes-with the producer and the pointed-to memory is
visible before use.

Signed-off-by: irozzo-1A <iacopo@sysdig.com>

Author: irozzo-1A

userspace/falco/watchdog.h

@@ -35,7 +35,7 @@ class watchdog {
 			const auto no_deadline = time_point{};
 			timeout_data curr;
 			while(m_is_running.load(std::memory_order_acquire)) {
-				auto t = m_timeout.exchange(nullptr, std::memory_order_release);
+				auto t = m_timeout.exchange(nullptr, std::memory_order_acq_rel);
 				if(t) {
 					curr = *t;
 					delete t;
@@ -56,7 +56,7 @@ class watchdog {
 			if(m_thread.joinable()) {
 				m_thread.join();
 			}
-			delete m_timeout.exchange(nullptr, std::memory_order_release);
+			delete m_timeout.exchange(nullptr, std::memory_order_acq_rel);
 		}
 	}