A bug about system call on ARM

[Will Deacon]

On Mon, Jun 03, 2013 at 11:45:34AM +0100, Russell King - ARM Linux wrote:
> On Mon, Jun 03, 2013 at 11:27:23AM +0100, Will Deacon wrote:
> > On Mon, Jun 03, 2013 at 11:18:09AM +0100, Russell King - ARM Linux wrote:
> > > On Thu, May 30, 2013 at 12:41:12PM +0100, Will Deacon wrote:
> > > > +#if defined(CONFIG_OABI_COMPAT) || !defined(CONFIG_AEABI)
> > > > +	/*
> > > > +	 * We may have faulted trying to load the SWI instruction due to
> > > > +	 * concurrent page aging on another CPU. In this case, return
> > > > +	 * back to the swi instruction and fault the page back.
> > > > +	 */
> > > > +9001:
> > > > +	sub	lr, lr, #4
> > > > +	str	lr, [sp, #S_PC]
> > > > +	b	ret_fast_syscall
> > > > +#endif
> > > 
> > > The comment is wrong.  If we get here, it means that the fault from
> > > trying to loading the instruction can't be fixed up.  Arguably, that
> > > should result in a SIGSEGV being sent immediately, but we'll get to
> > > that when we then try to re-load the instruction.
> > 
> > Why would we kill the application in this case? The reported problem is
> > where one CPU ages the page containing the swi instruction (mkold =>
> > clears L_PTE_YOUNG => write 0 to the pte) in between the other CPU executing
> > the swi and the kernel trying to read the immediate. The VMA is fine.
> 
> If you mark the instruction was a user-accessing instruction, the kernel
> will handle the resulting exception, trying to make the page accessible.
> If it is successful, then execution resumes as normal at the faulting
> instruction and continues as if nothing happened.
> 
> If it can't make the page accessible (eg, out of memory) the exception
> handler path (your code above) will be called instead.  Normal action in
> that case would be for a system call to return -EFAULT, but in this case
> we can't know what the syscall was, so we don't know if userspace will
> even pay attention to the returned error code.  In any case, if the page
> is no longer accessible, it's going to end up being killed by a SEGV
> when we eventually return to userspace anyway.

Yes, of course, the fault handling will sort out non-fatal faults for us, so
I'll update the comment.

Thanks,

Will