Re: [PATCH v4 1/5] x86/bugs: Only harden syscalls when needed

[Josh Poimboeuf <jpoimboe@kernel.org>]

On Mon, Apr 22, 2024 at 04:09:33PM +0800, Yujie Liu wrote:
> On Fri, Apr 19, 2024 at 02:09:47PM -0700, Josh Poimboeuf wrote:
> > Syscall hardening (converting the syscall indirect branch to a series of
> > direct branches) has shown some performance regressions:
> >
> > - Red Hat internal testing showed up to 12% slowdowns in database
> >   benchmark testing on Sapphire Rapids when the DB was stressed with 80+
> >   users to cause contention.
> >
> > - The kernel test robot's will-it-scale benchmarks showed significant
> >   regressions on Skylake with IBRS:
> >   https://lkml.kernel.org/lkml/202404191333.178a0eed-yujie.liu@intel.com
> 
> To clarify, we reported a +1.4% improvement (not regression) of
> will-it-scale futex4 benchmark on Skylake. Meanwhile we did observe some
> regressions by running other benchmarks on Ice Lake, such as:
> 
>     stress-ng.null.ops_per_sec -4.0% regression on Intel Xeon Gold 6346 (Ice Lake)
>     unixbench.fsbuffer.throughput -1.4% regression on Intel Xeon Gold 6346 (Ice Lake)

Thanks for clarifying that.  I'm not sure what I was looking at.

I also saw your email where Ice Lake showed a ~10% regression for
1e3ad78334a6.  Unfortunately my patch wouldn't help with that, as it's
designed to help with older systems (e.g., Skylake) and newer (e.g.,
Sapphire Rapids) but not Ice/Cascade Lake.

Whether 1e3ad78334a6 helps or hurts seems very workload-dependent.

It would be especially interesting to see if my patch helps on the newer
systems which have the HW mitigation: Raptor Lake, Meteor Lake, Sapphire
Rapids, Emerald Rapids.

For now, maybe I'll just table this patch until we have more data.

-- 
Josh