Re: [musl] Mutexes are not unlocking

[Jonathan Wakely <jwakely@redhat.com>]

On 23/11/20 09:56 -0500, Rich Felker wrote:
>On Mon, Nov 23, 2020 at 12:24:28PM +0000, Jonathan Wakely wrote:
>> On 22/11/20 23:05 +0300, Арсений wrote:
>> >
>> >Hello,
>> > 
>> >I fixed the problem by making a workaround. Specifying -D_GLIBCXX_GTHREAD_USE_WEAK=0 forces libstdc++ headers do not use weak symbols. Mutexes are correctly locked and unlocked now.
>>
>> That might "work" but is unsupported, because that macro is for
>> libstdc++'s internal use, not for users to define/undefine.
>>
>> But then libstdc++ doesn't support being compiled/linked against a
>> glibc libstdc++ and then running against musl at runtime (I didn't
>> even know that was an option until today) so one more unsupported
>> thing probably won't hurt :-)
>>
>> Maybe we should just bless the use of that macro as supported, which
>> would solve https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89714
>
>As you noted in comment 2, that won't fix the uses internal to
>libstdc++.{so,a}, only the inlined ones. I think this could give the
>wrong behavior in the opposite direction -- calling unlock without
>lock, thereby causing an error (for error-checking mutexes) or trap
>(if UB catching traps are in place for other types).

I've just had a quick look, and (for linux targets) the only case I
found where there's a lock outside the library (in user code) and a
corresponding unlock inside the library, or vice versa, is in
std::notify_all_at_thread_exit. (I only looked very quickly, so there
might be other cases I missed).

It does look like it would be possible for user code to lock the mutex
(via a direct all to the non-weak pthread_mutex_lock) and then call
std::notify_all_at_thread_exit() which unlock the mutex, which would
test __gthread_active_p, which would incorrectly think the program is
single-threaded, and not call pthread_mutex_unlock.

So recommending -D_GLIBCXX_GTHREAD_USE_WEAK=0 is a bad idea. We would
need a better way for users to override the checks.