Re: [PATCH 3/4] x86: open-code register save/restore in trace_hardirqs thunks

[Denys Vlasenko <vda.linux@googlemail.com>]

On Sat, Jan 10, 2015 at 10:27 PM, Linus Torvalds
<torvalds@linux-foundation.org> wrote:
> On Sat, Jan 10, 2015 at 1:09 PM, Denys Vlasenko
> <vda.linux@googlemail.com> wrote:
>>
>> I think using push/pop is okay. In the very hottest code paths
>> you may want to prefer mov's.
>
> For kernel entrypoints in particular, the code sequence is quite
> possibly constrained by the decoder and instruction fetch rather than
> the execution engine. Even if the entrypoint were to be in the L1 I$
> (which is not generally the case except in microbenchmarks), I am
> pretty sure that even Intel doesn't actually speculatively decode
> across system call boundaries, so unlike normal nice code, you don't
> have the front end running ahead of the execution engine.
>
> Looking at the system call hotpath, for example, it looks like we
> save/restore 8 registers. So 16 instructions or about 80 bytes of
> code. I could easily imagine us avoiding one cacheline access by using
> shorter 1- and 2-byte push/pop instructions (depending a bit on how
> the cacheline alignment works out, of course).
>
> Depending on how well it prefetches from L2 and/or exact decoder
> details, that kind of issue *can* overshadow the actual execution
> costs. Of course, on microbenchmarks (eg some system call benchmark
> that does "getppid()" in a loop), even the kernel side stays in the
> L1, so those might show possible execution issues more. And
> macrobenchmarks probably won't show a cycle or two in the system call
> or fault path anyway.
>
>                     Linus

Looks like consensus to me. I'm resending patches with patch #3 reworked.