[08/08] nptl: arm: Fix Race conditions in pthread cancellation (BZ#12683)

  On 07-09-2015 14:17, Phil Blundell wrote:
> On Mon, 2015-08-31 at 18:11 -0300, Adhemerval Zanella wrote:
>> This patch adds the ARM modifications required for the BZ#12683 fix.
>> It basically removes the enable_asynccancel/disable_asynccancel function
>> usage on code, provide a arch-specific symbol that contains global
>> markers to be used in SIGCANCEL handler.
> 
> I looked at this in a bit more detail.  Here are some further comments:
> 
>>  	.fnstart;	/* matched by the .fnend in UNDOARGS below.  */	\
>> -	DOCARGS_##args;	/* save syscall args etc. around CENABLE.  */	\
>> -	CENABLE;							\
>> -	mov ip, r0;		/* put mask in safe place.  */		\
>> -	UNDOCARGS_##args;	/* restore syscall args.  */		\
>> -	ldr	r7, =SYS_ify (syscall_name);				\
>> -	swi	0x0;		/* do the call.  */			\
>> -	mov	r7, r0;		/* save syscall return value.  */	\
>> -	mov	r0, ip;		/* get mask back.  */			\
>> -	CDISABLE;							\
>> -	mov	r0, r7;		/* retrieve return value.  */		\
>> -	RESTORE_LR_##args;						\
> 
> After this change, DOCARGS(), UNDOCARGS() and RESTORE_LR() are all dead
> code.  Please delete them.

I noted it on your first reply and I already removed them in my branch.

> 
>> -	UNDOARGS_##args;						\
>> +	push	{r4, r5, lr};						\
>> +	.save   {r4, r5, lr};						\
>> +	PSEUDO_CANCEL_BEFORE;						\
>> +	movw	r0, SYS_ify (syscall_name);				\
> 
> This fails when compiling for non-thumb2.  Please use "ldr r0, =..."
> instead.
> 

I will change it

>> +	PSEUDO_CANCEL_AFTER;						\
>> +	pop	{r4, r5, pc};						\
> 

Indeed 'pc' seems wrong, I think it should be 'lr' instead.  I will change it.

> Popping {pc} here causes an immediate return, which means that the errno
> handling code which follows is never executed.  Somewhat embarrassingly
> it seems that there is no existing glibc test which catches this
> failure.  I've attached a proof of concept which demonstrates it, but I
> rather wonder whether we should extend the test harness so that
> some/most of the existing glibc tests are run both single-threaded (as
> now) and with an additional dummy thread created at startup in order to
> force this code down the multi-threaded path.
> 
> Also note that the two testcases in the attached patch give slightly
> different results and I think they would continue to do so (in a
> different way) if the bug above was fixed.  It's not entirely clear to
> me that this part of __syscall_cancel() from your other patch is
> correct:
> 
> +  /* If cancellation is not enabled, call the syscall directly.  */
> +  if (pd->cancelhandling & CANCELSTATE_BITMASK)
> +    {
> +      INTERNAL_SYSCALL_DECL (err);
> +      result = INTERNAL_SYSCALL_NCS (nr, err, 6, a1, a2, a3, a4, a5, a6);
> +      return INTERNAL_SYSCALL_ERROR_P (result, err) ? -result : result;
> +    }
> 
> p.
> 

I will add your testcases on my patch to fix the NPLT tests for this mechanism
[1].  Also the inline cancellation call is fact wrong: it should only negate
the value depending if the architecture set the syscall return to be negate
in case of a error (powerpc for instance).

With these changes:

I saw no regression on arm with both NPTL current tests and the ones 
you have added.  I also checked on other ports (x86, i386, powerpc64, and aarch64).

Thanks for the review.

[1] https://sourceware.org/ml/libc-alpha/2015-08/msg01287.html