Re: pt_regs leak into userspace (was Re: [PATCH v3 20/71] ARC: Signal handling)

[James Hogan <james.hogan@imgtec.com>]

On 11/02/13 10:53, Jonas Bonn wrote:
> On 11 February 2013 11:28, James Hogan <james.hogan@imgtec.com> wrote:
>> On 11/02/13 10:13, Vineet Gupta wrote:
>>> On Monday 11 February 2013 03:06 PM, Jonas Bonn wrote:
>>>> On 11 February 2013 08:26, Vineet Gupta <Vineet.Gupta1@synopsys.com> wrote:
>>>>
>>>>> The only downside of this patch is that userspace signal stack grows in size,
>>>>> since signal frame only cares about scratch regs (pt_regs), but has to accommodate
>>>>> unused placeholder for callee regs too by virtue of using user_regs_struct.
>>>> Is this really true?  Don't setcontext and friends require that _all_
>>>> the registers be part of sigcontext?
>>>
>>> But for an ABI - callee saved regs will anyhow be saved/restored even in
>>> setcontext case ! So collecting it for that purpose seems useless, or am I missing
>>> something here.
>>
>> I think Jonas' point was that signals are asynchronous, i.e. you could
>> get interrupted by a signal at virtually any time during the program's
>> execution.
> 
> No, I agree that the callee-saved regs don't need to be saved across a
> signal handler invocation.  It's really just the setcontext case that
> wants to be able to swap out the callee-saved regs.

Yes, I was getting muddled, sorry!

> And now that I think about it some more, I think this is done
> incorrectly in the openrisc arch, too, as the fast-path for
> rt_sigreturn probably only restores the call-clobbered regs.
> sigreturn probably needs to be special-cased to _always_ restore all
> the regs on its way back to userspace.  Not for the "signal" case, but
> for the "setcontext" case; but these two are pretty-much
> indistinguishable.

Cheers
James