On Thu, 24 Mar 2011, Pekka Enberg wrote:
> > I forced the fallback to the _emu function to occur but could not trigger
> > the bug in kvm.
>
> That's not the problem. I'm sure the fallback is just fine. What I'm
> saying is that the fallback is *not patched* to kernel text on Ingo's
> machines because alternative_instructions() happens late in the boot!
> So the problem is that on Ingo's boxes (that presumably have old AMD
> CPUs) we execute cmpxchg16b, not the fallback code.
But then we would get the bug in kmem_cache_alloc() and not in the
*_emu() function. So the _emu is executing but failing on Ingo's system
but not on mine. Question is why.
For some reason the first reference to %gs:(%rsi) wont work right on his
system:
>From arch/x86/lib/cmpxchg16b_emu
#
# Emulate 'cmpxchg16b %gs:(%rsi)' except we return the result in %al not
# via the ZF. Caller will access %al to get result.
#
# Note that this is only useful for a cpuops operation. Meaning that we
# do *not* have a fully atomic operation but just an operation that is
# *atomic* on a single cpu (as provided by the this_cpu_xx class of
# macros).
#
this_cpu_cmpxchg16b_emu:
pushf
cli
cmpq %gs:(%rsi), %rax
jne not_same
cmpq %gs:8(%rsi), %rdx
jne not_same
movq %rbx, %gs:(%rsi)
movq %rcx, %gs:8(%rsi)
popf
mov $1, %al
ret
not_same:
popf
xor %al,%al
ret
CFI_ENDPROC
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>