On 08/08/2017 09:49, Peter Xu wrote:
> On Tue, Aug 08, 2017 at 09:26:43AM +0200, Paolo Bonzini wrote:
>> On 08/08/2017 09:00, Peter Xu wrote:
>>> We were calling rcu_init_complete() twice in the child processes when
>>> fork happened. However the pthread library does not really suggest to do
>>> it that way:
>>> "Attempting to initialise an already initialised mutex results in
>>> undefined behaviour."
>>> Actually, IMHO we can do it in a more natural way: Firstly, we only init
>>> the RCU globals once in rcu_init(). Then, in rcu_init_child(), we unlock
>>> all the locks held in rcu_init_lock() just like what we do in the parent
>>> process, then do the rest of RCU re-init (e.g., create the RCU thread).
>> This doesn't work for error-checking mutexes: rcu_init_child has a
>> different PID than the parent, so the mutexes aren't unlocked. It's
>> also true that right now we don't use error-checking mutexes (commit
>> 24fa90499f, "qemu-thread: do not use PTHREAD_MUTEX_ERRORCHECK",
>> 2015-03-10); however, that's also a bit sad.
>> The reason for the undefined behavior is probably that some operating
>> systems allocate memory in pthread_mutex_init, and initializing twice
>> causes a memory leak. One such operating system is OpenBSD. :(
> Good to know. :)
> I thought pthread_atfork() was designed to solve such a locking
> problem (in child hanlder, we unlock all the held locks). If
> PTHREAD_MUTEX_ERRORCHECK cannot coop well with it, not sure whether
> that means we should just avoid using PTHREAD_MUTEX_ERRORCHECK in such
> a use case (but we should be able to use the error checks in other
> mutexes that do not need extra fork handling)?
> Another idea is: can we just destroy the mutex first then re-init it
> in subprocess? A quick glance in libpthread code shows that at least
> pthread_mutex_destroy() won't check PTHREAD_MUTEX_ERRORCHECK.
Destroying a mutex that is locked is also undefined behavior, even if it
has no waiters. So it's a nice catch-22: we cannot destroy before
unlocking, and we cannot unlock an error-checking mutex, so we cannot
destroy before init.
Actually, unlocking in atfork's child callback is interesting even for
non-error-checking mutexes. If there are waiters, waking up the first
one in pthread_mutex_unlock fails. This is okay if (as on Linux)
pthread_mutex_lock does none of the woken-up process's work. However,
on OpenBSD pthread_mutex_lock sets mutex->owner... but in the child
there's no other thread to wake up and thus no one will set mutex->owner
to NULL. This should lead to deadlock scenarios with pthread_atfork.
So pthread_atfork is impossible to use correctly with mutexes, and
reinitializing is the best we can do (the memory leak being a lesser evil).
The workaround would be to replace mutexes with semaphores, and to
implement QemuEvent natively for OpenBSD using its __thrsleep and
__thrwakeup system calls (no, I am not going to do that :)).
Let's wait and hear what Eric has to say.
>> Eric, you chimed in on the patch that became commit 24fa90499f, what do
>> you suggest?