On 28/09/2020 13:32, Jan Beulich wrote: > There's no point having every replacement variant to also specify the > INT3 - just have it once in the base macro. When patching, NOPs will get > inserted, which are fine to speculate through (until reaching the INT3). > > Signed-off-by: Jan Beulich <jbeul...@suse.com> > --- > I also wonder whether the LFENCE in IND_THUNK_RETPOLINE couldn't be > replaced by INT3 as well. Of course the effect will be marginal, as the > size of the thunk will still be 16 bytes when including tail padding > resulting from alignment.
There are surprising performance implications from the choice of speculation blocker. RSB filling in particular had a benefit (up to 6% iirc) from unrolling the loop. Any differences here are likely to be marginal, whereas for inline retpoline, the code volume reduction might easily be the winning factor. > --- > v2: New. > > --- a/xen/arch/x86/indirect-thunk.S > +++ b/xen/arch/x86/indirect-thunk.S > @@ -11,6 +11,8 @@ > > #include <asm/asm_defns.h> > > +.purgem ret This needs a comment. ~Andrew
