Re: Do x86 NX and AMD prefetch check cause page fault infinite loop?

[Jamie Lokier <jamie@shareable.org>]

Ingo, I think I now know what must be added to your 32-bit NX patch to
prevent the "infinite loop without a signal" problem.

It appears the correct way to prevent that one possibility I thought
of, with no side effects, is to add this test in
i386/mm/fault.c:is_prefetch():

        /* Catch an obscure case of prefetch inside an NX page. */
        if (error_code & 16)
                return 0;

That means that it doesn't count as a prefetch fault if it's an
_instruction_ fault.  I.e. an instruction fault will always raise a
signal.  Bit 4 of error_code was kindly added alongside the NX feature
by AMD.

(Tweak: Because early Intel 64-bit chips don't have NX, perhaps it
should say "if ((error_code & 16) && boot_cpu_has(X86_FEATURE_NX))"
instead -- if we find the bit isn't architecturally set to 0 for those
chips).

This test isn't needed in the plain, non-NX i386 kernel, because the
condition can never occur.  (Actually it can once, a really obscure
condition due to separate ITLB and DTLB loading and page table races
with other CPUs, but it's transient so won't loop infinitely).

Enjoy,
-- Jamie