On Wednesday 23 January 2013 00:40:24 Maciej W. Rozycki wrote:
> On Wed, 23 Jan 2013, Mike Frysinger wrote:
> > >  We have an issue with the INTERNAL_SYSCALL_NCS wrapper in that it does
> > > not respect the kernel's syscall restart convention.
> > > 
> > >  That convention requires the instruction immediately preceding SYSCALL
> > > to initialize $v0 with the syscall number.  Then if a restart triggers,
> > > $v0 will have been clobbered by the syscall interrupted, and needs to
> > > be reinititalized.  The kernel will decrement the PC by 4 before
> > > switching back to the user mode so that $v0 has been reloaded before
> > > SYSCALL is executed again.  This implies the place $v0 is loaded from
> > > must be preserved across a syscall, e.g. an immediate, static
> > > register, stack slot, etc.
> > 
> > naïvely, but why can't the mips kernel paths take care of the reload
> > itself ? other arches have scratch space in their pt_regs for doing just
> > this (a bunch of arches use the orig_<reg> convention).
> 
>  I agree it would be the most reasonable approach if designing from
> scratch; unfortunately what we have is how the ABI has been set back in
> 1994.  You won't be able to patch up all the kernel binaries out there,
> sigh...

sure, you won't be able to retroactively fixing kernels.  but you'll be able to 
make future kernels more robust against shady userlands.  as you've pointed 
out, this is an extremely subtle bug that can easily go unnoticed for a long 
time which simply injects random flakiness into the runtime system.
-mike