Re: [musl] Difference in pthread behavior between glibc and musl

[Gabriel Ravier <gabravier@gmail.com>]

On 7/12/23 04:48, Rich Felker wrote:
> On Tue, Jul 11, 2023 at 07:19:50PM +0000, Appelmans, Madeleine wrote:
>> Hello,
>>
>> There seems to be a difference in pthread behavior when compiling
>> with glibc and using the musl-gcc wrapper. The attached snippet of
>> code creates a destructor attribute which deletes a pthread key. The
>> code never actually creates the pthread key. This code segfaults
>> when compiled using musl-gcc, and does not segfault when compiled
>> with gcc.
>>
>> Best guess at what is going on: When creating a pthread key, musl
>> initializes a field called
>> tsd<https://git.musl-libc.org/cgit/musl/tree/src/thread/pthread_key_create.c#n37>.
>> When deleting a key, musl assumes that initialization has been done,
>> and dereferences tsd without checking that it exists: see
>> here<https://git.musl-libc.org/cgit/musl/tree/src/thread/pthread_key_create.c#n65>.
>> This dereference may be the source of the segfault.
> This is completely expected; the behavior is undefined because you
> passed to pthread_key_delete a value which was not acquired via
> pthread_key_create.
>
> If it happens not to crash on glibc, that doesn't mean it's okay to do
> it. It will end up deleting whatever key happens to correspond to the
> zero-initialized pthread_key_t object, which may be a key that was
> allocated for use by some other part of the program when it called
> pthread_key_create. (In other words, you have a type of double-free or
> use-after-free bug.) Your program logic must ensure you refrain from
> doing that.
>
> Rich

Hmmm, this does indeed seem to be the case ever since 
SUSv4/XPG7/POSIX.1-2008 removed the EINVAL error from the specification 
of pthread_key_delete, but the requirement to detect this error is 
present in SUSv3/XPG6/POSIX.1-2001 (up to and including the 2004 
edition). so if you want musl to be able to say it conforms to that 
standard, it would have to implement this, which after a quick look at 
musl's source code w.r.t. pthread keys doesn't seem particularly burdensome.

Personally I'd be in favor of having this detection occur regardless of 
whether musl aims for conformance to older standards given that POSIX 
still now explicitly recommends it even in the standards where it is not 
mandatory anymore, which is something I've usually seen done in cases 
where they wanted to specify a behavior but had to back down on making 
it mandatory on the basis of certain implementations complaining that it 
was too complicated to implement and/or wanting to retain another 
behavior for backwards compatibility or other stuff like that - i.e. 
they would very much prefer implementations where it isn't particularly 
burdensome to implement it to do so.