[sane-devel] Fwd: Re: pthread_detach after pthread_join generates segfault

[Olaf Meeuwissen]

Hi,

# I had meant to send this to the list as well.  It's a week later now
# and I have already requested merged the pu/ branch mentioned below.

I'm no pthread buff either, but

Jens-U. Mozdzen writes:

> Hi *,
>
> quoting Allan:
>>
>> I'm no expert either, but your basic analysis seems correct- once you
>> have returned from pthread_join(), the thread is gone, and its return
>> value has been popped.
>
> that's indeed the way it works. Upon return of pthread_join(), the  
> thread is *gone*... unless it returns with a non-zero rc.
>
>> I suppose it is possible that the pthread_join
>> could fail in a way that might leave the thread running, perhaps
>> because of a deadlock? In that case I suppose we could call
>> pthread_detatch only if there is an error?
>
> Other reasons to for pthread_join() to fail would be an already  
> pending pthread_join() in a different thread.
>
> It's not necessary/valid to call pthread_detach() on all error cases -  
> i.e. in case of ESRCH, there is no thread with that id. Calling  
> pthread_detach() then will only return ESRCH itself, the same holds  
> true for EINVAL.

The only valid, documented return value of pthread_join() where calling
pthread_detach() might make sense is EDEADLK.  However, I am not sure
what the effect is of detaching the calling thread.  Perhaps adding a
check for that makes sense.

I've pushed a proposed upgrade.  If that looks okay, please merge.  The
changes can be found at

  https://gitlab.com/sane-project/backends/commits/pu/dont-detach-after-join

> What doesn't fit the problem description is that calling  
> pthread_detach() for an already terminated thread results in a segv.  
> I'd rather have expected a return of ESRCH in such case. The only  
> unspecified result (per man page) is for calling pthread_detach()  
> multiple times on the same thread. Might this be the case here?
>
> quoting Gerhard
>> I wonder why it took 11 years to spot that ;)
>
> If it is a case of calling pthread_detach() multiple times, it may be  
> because the code got compiled for a new platform. "undefined results"  
> include causes a segv :D. Or maybe the pthread lib contains a bug,  
> i.e. segfaulting on detaching a no longer existing thread, rather than  
> returning ESRCH, or some race condition.
>
> Regards,
> Jens

Hope this helps,
-- 
Olaf Meeuwissen, LPIC-2            FSF Associate Member since 2004-01-27
Support Free Software               Support the Free Software Foundation
https://my.fsf.org/donate                        https://my.fsf.org/join