Re: [PATCH] x86/selftests: add clobbers for int80 on x86_64

[Dmitry Safonov]

On 02/10/2017 07:45 PM, Andy Lutomirski wrote:
> On Fri, Feb 10, 2017 at 8:28 AM, Dmitry Safonov <dsafonov@xxxxxxxxxxxxx> wrote:
> > On 02/10/2017 07:13 PM, Andy Lutomirski wrote:
> > > On Fri, Feb 10, 2017 at 3:52 AM, Dmitry Safonov <dsafonov@xxxxxxxxxxxxx>
> > > wrote:
> > > > Kernel erases R8..R11 registers prior returning to userspace
> > > > from int80: https://lkml.org/lkml/2009/10/1/164
> > > >
> > > > GCC can reuse this registers and doesn't expect them to change
> > > > during syscall invocation. I met this kind of bug in CRIU once
> > > > gcc 6.1 and clang stored local variables in those registers
> > > > and the kernel zerofied them during syscall:
> > > >
> > > > https://github.com/xemul/criu/commit/990d33f1a1cdd17bca6c2eb059ab3be2564f7fa2
> > > >
> > > > By that reason I suggest to add those registers to clobbers
> > > > in selftests.
> > >
> > >
> > > Seems reasonable, but presumably INT80_CLOBBERS should be defined the
> > > same way in all the tests.  IOW, if the "flags" clobber is actually
> > > needed, it should be "flags", INT80_CLOBBERS (possibly without the
> > > comma if it's problematic).
> >
> > Well, that was my initial attempt: I've defined it as:
> > +# define INT80_CLOBBERS , "r8", "r9", "r10", "r11"
> >
> > But that hanging comma looks awful, so I added "flags" there.
> > And if I do define it without coma and leave it in asm statement,
> > 32-bit version would be unhappy.
> > So, I found that it's easier to define it with flags included.
>
> Woudl the right answer be to get rid of "flags" in the test where it
> appears?  I'm not sure it's needed in the first place.

I think it can live without it.
But I didn't want to change it in the same patch and wasn't sure if I
fail to see the reason for it.
So, I'll resend with flags removing, thanks.

--
             Dmitry