https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98033
Bug ID: 98033
Summary: ABA problem in atomic wait
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: libstdc++
Assignee: unassigned at gcc dot gnu.org
Reporter: redi at gcc dot gnu.org
Target Milestone: ---
In __atomic_wait in include/bits/atomic_wait.h we have:
while (!__pred())
{
if constexpr (__platform_wait_uses_type<_Tp>)
{
__platform_wait(__addr, __old);
}
If the predicate is initially false (i.e. *addr == old) we enter the loop. If
the value changes before calling __platform_wait then that will return
immediately. If it then changes back to the old value the predicate will be
false again and we will re-enter the loop. This time __platform_wait will block
waiting for the value to change.
The standard says we're allowed to miss such transient values, and so the code
above is conforming. But I think we can make it work on linux for the
std::atomic_signed_lock_free type at least (which isn't defined, but should be
atomic<int>).
If __platform_wait returned a boolean indicating whether the not-equal
condition was met (either because it was already true, the EAGAIN case, or
because the futex wait got woken up) then the caller could check that and not
re-check the predicate.
