On 24/11/2015 16:43, David Engraf wrote: > Commit afbe70535ff1a8a7a32910cc15ebecc0ba92e7da introduced the function > qemu_mutex_iothread_locked to avoid recursive locking of the iothread > lock. The iothread_locked variable uses the __thread attribute which > results in a per thread storage location whereas the qemu_global_mutex > is not thread specific. This leads to race conditions because the > variable is not in sync between threads.
Frankly, this makes no sense. You're modifying the qemu_mutex_iothread_locked function to return whether _some_ thread has taken the mutex, but the function tells you whether _the calling_ thread has taken the mutex (that's the point about recursive locking). This breaks the callers of qemu_mutex_iothread_locked completely. The variable need not be in sync between the different threads; each thread only needs to know about itself. The current code works because: 1) the iothread mutex is not locked before qemu_mutex_lock_iothread 2) the iothread mutex is not locked after qemu_mutex_lock_iothread 3) qemu_cond_wait doesn't matter because the thread does not run during a qemu_cond_wait. > I triggered this problem reproducible on a Windows machine whereas Linux > works fine. What are the symptoms, and what is your Windows compiler? Paolo
