Re: [PATCH] Mutilated form of Andi Kleen's AMD prefetch errata patch

[Nick Piggin]

Jamie Lokier wrote:

> This builds upon Andi Kleen's excellent patch for the AMD prefetch bug
> workaround.  It is applies to 2.6.0-test6.
>
> There are four changes on top of Andi's recent patch:
>
>   1. The workaround code is only included when compiling a kernel
>      optimised for AMD processors with the bug.  For kernels optimised
>      for a different processor, there is no code size overhead.
>
>      This will make some people happier.
>
>      It will make some people unhappier, because it means a generic
>      kernel on an AMD will run fine, but won't fixup _userspace_
>      prefetch faults.  More on this below.
>
>   2. Nevertheless, when a kernel _not_ optimised for those AMD processors
>      is run on one, the "alternative" mechanism now correctly replaces
>      prefetch instructions with nops.
>
>   3. The is_prefetch() instruction decoder now handles all cases of
>      funny GDT and LDT segments, checks limits and so on.  As far as I
>      can see, there are no further bugs or flaws in that part.
>
>   4. A consequence of the is_prefetch() change is that, despite the
>      longer C code (rarely used code, for segments), is_prefetch()
>      should run marginally faster now because the access_ok() calls are
>      not required.  They're subsumed into the segment limit
>      calculation.
>
> Like Andi's patch, this removes 10k or so from current x86 kernels by
> removing the silly conditional from the prefetch() function.
>
> Controversial point coming up!

It sounds good Jamie, but could you possibly compile it in
unconditionally? Otherwise it introduces yet another caveat to the CPU
selection / build system. This can be addressed seperately.

I don't think anyone advocates attempting to handle this nicely with the
current CPU selection system, nor were they targeting this patch in
particular. It just kicked off a discussion about the shortfalls of the
selection system.


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/