[PATCH] nptl: Use SA_RESTART for SIGCANCEL handler

[Adhemerval Zanella]

On 18/06/2021 08:38, Florian Weimer wrote:
> * Adhemerval Zanella via Libc-alpha:
> 
>> The usage of signals to implementation pthread cancellation is an
>> implementation detail and should not be visible through cancellation
>> entrypoints.
>>
>> However now that pthread_cancel always send the SIGCANCEL, some
>> entrypoint might be interruptable and return EINTR to the caller
>> (for instance on sem_wait).
>>
>> Using SA_RESTART hides this, since the cancellation handler should
>> either act uppon cancellation (if asynchronous cancellation is enable)
>> or ignore the cancellation internal signal.
> 
> I think this still needs a NEWS entry because there have been kernel
> bugs in this area (e.g. in CIFS).

Ok, I have added the following on "Deprecated and removed features, and 
other changes affecting compatibility"

* The pthread cancellation handler is now setup with SA_RESTART.  It should
  not be visible to application since the cancellation handler should either 
  act uppon cancellation (if asynchronous cancellation is enabled) or
  ignore the cancellation internal signal.

> 
>> Checked on x86_64-linux-gnu and i686-linux-gnu.
>> ---
>>  nptl/pthread_cancel.c | 6 +++++-
>>  1 file changed, 5 insertions(+), 1 deletion(-)
>>
>> diff --git a/nptl/pthread_cancel.c b/nptl/pthread_cancel.c
>> index 0698cd2046..cc25ff21f3 100644
>> --- a/nptl/pthread_cancel.c
>> +++ b/nptl/pthread_cancel.c
>> @@ -72,7 +72,11 @@ __pthread_cancel (pthread_t th)
>>      {
>>        struct sigaction sa;
>>        sa.sa_sigaction = sigcancel_handler;
>> -      sa.sa_flags = SA_SIGINFO;
>> +      /* The signal handle should be non-interruptible to avoid the risk of
>> +	 spurious EINTR caused by SIGCANCEL sent to process or if
>> +	 pthread_cancel() is called while cancellation is disabled in the
>> +	 target thread.  */
>> +      sa.sa_flags = SA_SIGINFO | SA_RESTART;
>>        __sigemptyset (&sa.sa_mask);
>>        __libc_sigaction (SIGCANCEL, &sa, NULL);
>>        atomic_store_relaxed (&init_sigcancel, 1);
> 
> I really don't feel comfortable reviewing this.  However I think it is
> still consistent with the (buggy) SYSCALL_CANCEL implementation:
> 
>         int sc_cancel_oldtype = LIBC_CANCEL_ASYNC ();                        \
>         sc_ret = INLINE_SYSCALL_CALL (__VA_ARGS__);                          \
>         LIBC_CANCEL_RESET (sc_cancel_oldtype);                               \
> 
> We temporary enable async cancellation, in which case we unwind through
> the signal handler if canceled.  We do not rely on a EINTR error return
> from the system call and a cancellation check outside of the signal
> handler.  So adding SA_RESTART should really be okay.

Yes, we still cancel the thread even for partial results (BZ#12683).