On Thu, Aug 18, 2022 at 12:45 AM Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Wed, Aug 17, 2022 at 03:08:01PM -0700, Nick Desaulniers wrote:
> > On Wed, Aug 17, 2022 at 11:56 AM Nathan Chancellor <nathan@kernel.org> wrote:
> > >
> > > LLVM 16 will have support for this flag so move it out of the GCC-only
> > > block to allow LLVM builds to take advantage of it.
> > >
> > > Link: https://github.com/ClangBuiltLinux/linux/issues/1665
> > > Link: https://github.com/llvm/llvm-project/commit/6f867f9102838ebe314c1f3661fdf95700386e5a
> > > Signed-off-by: Nathan Chancellor <nathan@kernel.org>
> >
> > Thanks for the patch, I also boot tested this (in virtual guests for
> > both 64b and 32b).
> >
> > Reviewed-by: Nick Desaulniers <ndesaulniers@google.com>
> > Tested-by: Nick Desaulniers <ndesaulniers@google.com>
> >
> > Peter, is there anything else special about these prefixes needed to
> > make use of them?
>
> The thing to do is boot with (warning, *very* verbose):
>
>         "spectre_v2=retpoline,lfence debug-alternative"

ah, debug-alternative also needs loglevel=8. Was wondering why I
wasn't getting anything new in the output...

No kidding on the verbosity. For a defconfig, booting in QEMU and
redirecting stdout to a file produced a 95257 line file.

$ grep "SMP alternatives: retpoline" /tmp/log.txt | wc -l
18623
$ du -h /tmp/log.txt
11M /tmp/log.txt
$ grep "SMP alternatives: retpoline" /tmp/log.txt | tr -s ' '
...
[ 56.736261] SMP alternatives: retpoline at:
pm_check_save_msr+0x30/0x80 (ffffffffaa34b320) len: 6 to:
__x86_indirect_thunk_r11+0x0/0x20
$ grep "SMP alternatives: retpoline" /tmp/log.txt | tr -s ' ' | cut -d
' ' -f 10 | sort -u
5
6
# Curious about the 5, looks like there's only 2 cases:
$ grep "SMP alternatives: retpoline" /tmp/log.txt | grep 'len: 5'
[    0.147261] SMP alternatives: retpoline at: ret_from_fork+0x1d/0x30
(ffffffffa9601dbd) len: 5 to: __x86_indirect_thunk_rbx+0x0/0x20
[    2.274261] SMP alternatives: retpoline at: __efi_call+0x23/0x30
(ffffffffa9671663) len: 5 to: __x86_indirect_thunk_rdi+0x0/0x20


Is there anything else I can do to help verify Nathan and Joao's
patches?  If not, would you mind Ack'ing this so the x86 maintainers
can pick it up?


>
> and observe that the retpoline sites all replace:
>
>         "cs call __x86_indirect_thunk_r11" (6 bytes)
>
> with:
>
>         "lfence; jmp *%r11" (6 bytes)
>
>
> This being clang, you'll ofcourse still have a few weird:
>
>         "Jcc.d32 __x86_indirect_thunk_r11"
>
> sites that will not be able to be replaced, but there's nothing I can do
> about that other than to continue to encourage y'all to stop emitting
> them ;-)

Noting that Joao's patch was applied in my tests.
https://reviews.llvm.org/D134915
I plan to land Joao's patch imminently.
-- 
Thanks,
~Nick Desaulniers