Re: [PATCH 3/3] x86, FPU: Do not use static_cpu_has before alternatives

[H. Peter Anvin]

On 04/29/2013 11:51 AM, Borislav Petkov wrote:
> On Mon, Apr 29, 2013 at 08:42:30AM -0700, H. Peter Anvin wrote:
>> I *was* considering adding static_cpu_has_safe() at some point which
>> would have a three-state jump, with the default (pre-alternatives)
>> jump pointing to dynamic detection code.
> 
> Actually, if we teach __static_cpu_has to do something like
> ALTERNATIVE_JUMP arch/x86/lib/copy_user_64.S but make the second
> alternative insn alt2 be none, i.e. no replacement, we can have:
> 
> * pre-alternatives: JMP dynamic_detection
> * post-alternatives:
>   - feature present: delete JMP
>   - feature absent: s/dynamic_detection/t_no/, i.e., patch only the label.

That is what I was referring to, yes.

> And even though asm goto supports multiple labels, we need to be able
> to either patch the label only or patch out the whole instruction -
> otherwise we'll be adding additional NOP bytes.

This doesn't seem to matter.

> I wonder if it would make sense to teach the alternatives to skip the
> opcode when patching so that we can say: "we only want to patch the
> label so we're patching in the offset now but leaving the single JMP
> opcode in there."
> 
> But for that we either need flags in struct alt_instr or do something
> ad-hoc apply_alternatives already does for relative jumps (0xe8).
> 
>> This might be useful here, on the other hand, perhaps it is acceptable
>> for use_eager_fpu() to be initially false?
> 
> Hmm, I don't know, FPU code is crazy. OTOH, does CR0.TS even matter on
> non-lazy FPU restore machines?

Yes, CR0.TS should be zero, or we'll get #NM traps when user space tries
to access the FPU.  For lazy restore CR0.TS should be set so we get the
#NM trap and can restore the FPU.

	-hpa



--
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/