https://issues.dlang.org/show_bug.cgi?id=24741

          Issue ID: 24741
           Summary: Stop the world logic is invalid and might stop a
                    thread mid critical section.
           Product: D
           Version: D2
          Hardware: x86
                OS: Windows
            Status: NEW
          Severity: enhancement
          Priority: P1
         Component: druntime
          Assignee: nobody@puremagic.com
          Reporter: deadalnix@gmail.com

A thread enter critical section by taking the criticalRegionLock and setting its m_isInCriticalRegion flag.

A thread leaves critical section in a similar way, taking the criticalRegionLock  and unsetting its m_isInCriticalRegion flag.

When stopping the world, the criticalRegionLock is taken during the whole process. For each thread, m_isInCriticalRegion before suspending. If it is set, then the lock is temporarily released and we wait to give the thread an opportunity to exit its critical section, and then then lock is taken again, and we check again if the thread is in its critical section, etc...

To suspend a thread, we send a signal using pthread_kill , which is asynchronous. This means that by the time we try to suspend thread n + 1, thread n might not be suspended yet.

If thread n + 1 is is in its critical section, we'll release criticalRegionLock, allowing thread n to enter a critical section before it had the chance to suspend.

In practice, critical section are rare enough this doesn't happen much, but a carefully designed test case can trigger the race condition, and potentially an attacker feeding malicious input to an application.

--