Re: Regression due to 0c44c2d0f459 x86: Use asm goto to implement better modify_and_test() functions

[Alex Williamson <alex.williamson@redhat.com>]

On Wed, 2013-12-04 at 08:44 -0800, H. Peter Anvin wrote:
> On 12/04/2013 12:59 AM, Ingo Molnar wrote:
> > 
> > * H. Peter Anvin <hpa@zytor.com> wrote:
> > 
> >> Hi guys,
> >>
> >> 0c44c2d0f459 x86: Use asm goto to implement better modify_and_test()
> >> functions
> >>
> >> causes a regression, because it incorrectly changed the constraints of
> >> bitops.
> >>
> >> Specifically, the GEN_BINARY_RMWcc() hardcodes a constraint as "er", but
> >> it needs to be "Ir" for the bitops themselves.  "I" is correct (as
> >> opposed to "J" even on 64 bits, because we only generate the 64-bit
> >> version when we have a register operand.
> >>
> >> Unfortunately there isn't a way we can get gcc+gas to generate a version
> >> with an offset pointer.
> > 
> > Does the regression manifest itself in any actual breakage - if yes, 
> > how does it look like? (People experiencing similar symptoms will be 
> > helped by seeing a fix matching their problems.)
> > 
> 
> It was discovered because it caused a build failure in a
> not-yet-submitted driver patch.  This happens when someone uses
> test_and_set_bit() or another similar operation on a fixed bit index
> above 255; the assembler throws an error at that point and the build fails.
> 
> *HOWEVER*, for bit indicies in the range 32-255, the current code will
> instead silently miscompile, as the CPU will truncate the argument to 5
> bits.  I don't know if there are any such instances in the current
> kernel, but it is entirely possible there is, with unknown but
> potentially disastrous results.

I'm seeing reports from folks using vfio for PCI device assignment
through QEMU that hugepages aren't being released on VM shutdown and
reverting 0c44c2d0f459 resolves the problem.  I haven't confirmed this
myself nor do I pretend to understand the changes this commit makes.
Thanks,

Alex