Re: Thread safety of coroutine-sigaltstack

[Max Reitz <mreitz@redhat.com>]

On 20.01.21 18:25, Laszlo Ersek wrote:
> On 01/20/21 17:26, Max Reitz wrote:
>> Hi,
>>
>> I’ve run into trouble with Vladimir’s async backup series on MacOS,
>> namely that iotest 256 fails with qemu exiting because of a SIGUSR2.
>>
>> Turns out this is because MacOS (-xcode) uses coroutine-sigaltstack,
>> when I use this on Linux, I get the same error.
>>
>> (You can find the series applied on my block branch e.g. here:
>>
>> https://github.com/XanClic/qemu.git block
>> )
>>
>> Some debugging later I found that the problem seems to be two threads
>> simultaneously creating a coroutine.  It makes sense that this case
>> would appear with Vladimir’s series and iotest 256, because 256 runs two
>> backup jobs in two different threads in a transaction, i.e. they’re
>> launched simultaneously.  The async backup series makes backup use many
>> concurrent coroutines and so by default launches 64+x coroutines when
>> the backup is started.  Thus, the case of two coroutines created
>> concurrently in two threads is very likely to occur.
>>
>> I think the problem is in coroutine-sigaltstack’s qemu_coroutine_new().
>> It sets up a SIGUSR2 handler, then changes the signal handling stack,
>> then raises SIGUSR2, then reverts the signal handling stack and the
>> SIGUSR2 handler.  As far as I’m aware, setting up signal handlers and
>> changing the signal handling stack are both process-global operations,
>> and so if two threads do so concurrently, they will interfere with each
>> other.
> 
> Signal action (disposition) is process-wide.
> 
> Signal mask and signal stack are thread-specific.

Ah, OK.  Thanks for the insight!

> A signal may be pending for the whole process, or for a specific thread.
> In the former case, the signal is delivered to one of the threads that
> are not blocking the signal.
> 
>> What usually happens is that one thread sets up everything,
>> while the other is already in the process of reverting its changes: So
>> the second thread reverts the SIGUSR2 handler to the default, and then
>> the first thread raises SIGUSR2, thus making qemu exit.
> 
> I agree. The way SIGUSR2 is blocked (for the thread), made pending (for
> the thread), and then allowed to be delivered (consequently, to the
> thread), looks OK. But by the time it is delivered, the action has been
> changed.
> 
>>
>> (Could be worse though.  Both threads could set up the sigaltstack, then
>> both raise SIGUSR2, and then we get one coroutine_trampoline()
>> invocation in each thread, but both would use the same stack.  But I
>> don’t think I’ve ever seen that happen, presumably because the race time
>> window is much shorter.)
> 
> No, the "alternate stack for signal handlers" that sigaltstack()
> configures is thread-specific. (I mean one could theoretically mess it
> up if the stack were located in the same place between different
> threads, but we call qemu_alloc_stack(), so that doesn't happen.)
> 
> https://pubs.opengroup.org/onlinepubs/9699919799/functions/sigaltstack.html

Explains why I haven’t seen it. :)

>> Now, this all seems obvious to me, but I’m wondering...  If
>> coroutine-sigaltstack really couldn’t create coroutines concurrently,
>> why wouldn’t we have noticed before?  I mean, this new backup case is
>> kind of a stress test, yes, but surely we would have seen the problem
>> already, right?  That’s why I’m not sure whether my analysis is correct.
>>
>> Anyway, I’ve attached a patch that wraps the whole SIGUSR2 handling
>> section in a mutex, and that makes 256 pass reliably with Vladimir’s
>> async backup series.  Besides being unsure whether the problem is really
>> in coroutine-sigaltstack, I also don’t know whether getting out the big
>> guns and wrapping everything in the mutex is the best solution.  So,
>> it’s an RFC, I guess.
> 
> A simple grep for SIGUSR2 seems to indicate that SIGUSR2 is not used by
> system emulation for anything else, in practice. Is it possible to
> dedicate SIGUSR2 explicitly to coroutine-sigaltstack, and set up the
> action beforehand, from some init function that executes on a "central"
> thread, before qemu_coroutine_new() is ever called?

Doesn’t sound unreasonable, but wouldn’t the signal handler then have to 
check whether the SIGUSR2 comes from coroutine-sigaltstack or from the 
outside?  Or should we then keep SIGUSR2 blocked all the time?

> ... I've tried to see if POSIX says anything on signals being delivered
> with mutexen held. I can't find anything specific (the spec seems to
> talk about delivery of a signal while the thread waits in
> pthread_mutex_lock(), but that's not what we care about, here). I'm just
> somewhat uncomfortable with bracketing this whole hackery into a mutex
> even... Keeping sigaction() out of the picture could be a small
> performance benefit, too.

Speaking of signal being delivered in the mutexed section...  What would 
happen if we get an external signal after SIGUSR2 was delivered and 
coroutine_trampoline() set up the sigsetjmp(), but before the stack is 
switched back?  Wouldn’t this new signal then trash the stack?  Should 
we block all signals while using the alternate stack?

(Looking at the x64 objdump, the stack actually seems to be used to 
store @self across sigsetjmp().)

> The logic in the patch doesn't look broken, but the comments should be
> updated minimally -- the signal stack is thread-specific (similarly to
> how a thread has its own stack anyway, regardless of signals).

Sure, I can do that.

I agree that there probably are better solutions than to wrap everything 
in a lock.  OTOH, it looks to me like this lock is the most simple 
solution.  If Daniel is right[1] and we should drop 
coroutine-sigaltstack altogether (at some point...), perhaps it is best 
to go for the most simple solution now.

[1]
https://lists.nongnu.org/archive/html/qemu-block/2021-01/msg00808.html

Max