2005-01-02 15:45:57

by Andrea Arcangeli

[permalink] [raw]
Subject: Re: VM fixes [->used_math to PF_USED_MATH] [6/4]

On Sun, Dec 26, 2004 at 11:03:09PM -0800, Andy Isaacson wrote:
> I can't find any authoritative source for that assertion, but google
> supports it:
> http://sources.redhat.com/ml/libc-alpha/2002-09/msg00328.html

This is a nice reference, thanks for the info (I personally only worked
with ev6 so I probably never run in the exact features and history of
the older chips).

All issues with ev4+smp/preempt and ev5+smp/preempt should be fixed with
the two incremental patches I posted last week that make PF_MEMDIE a
TIF_MEMDIE and used_math a PF_USED_MATH and that fixes the PF_MEMDIE
race that existed on all archs in mainline 2.6 and that AFIK Wli even
managed to reproduce once.


2005-01-08 17:18:16

by William Lee Irwin III

[permalink] [raw]
Subject: Re: VM fixes [->used_math to PF_USED_MATH] [6/4]

On Sun, Dec 26, 2004 at 11:03:09PM -0800, Andy Isaacson wrote:
>> I can't find any authoritative source for that assertion, but google
>> supports it:
>> http://sources.redhat.com/ml/libc-alpha/2002-09/msg00328.html

On Sun, Jan 02, 2005 at 04:41:48PM +0100, Andrea Arcangeli wrote:
> This is a nice reference, thanks for the info (I personally only worked
> with ev6 so I probably never run in the exact features and history of
> the older chips).
> All issues with ev4+smp/preempt and ev5+smp/preempt should be fixed with
> the two incremental patches I posted last week that make PF_MEMDIE a
> TIF_MEMDIE and used_math a PF_USED_MATH and that fixes the PF_MEMDIE
> race that existed on all archs in mainline 2.6 and that AFIK Wli even
> managed to reproduce once.

Reproduced only once ever on x86-64, but reliably reproducible on all
other 64-bit architectures. 32-bit architectures have resource
scalability issues that make the particular exploit I wrote ineffective.


-- wli