On Wed, 2018-02-21 at 19:43 -0800, Alexei Starovoitov wrote:
> On Wed, Feb 21, 2018 at 07:04:02PM -0800, Eric Dumazet wrote:
> > On Thu, 2018-02-22 at 01:05 +0100, Daniel Borkmann wrote:
> >
> > ...
> >
> > > +/* Instead of plain jmp %rax, we emit a retpoline to control
> > > + * speculative execution for the indirect branch.
> > > + */
> > > +static void emit_retpoline_rax_trampoline(u8 **pprog)
> > > +{
> > > + u8 *prog = *pprog;
> > > + int cnt = 0;
> > > +
> > > + EMIT1_off32(0xE8, 7); /* callq <set_up_target> */
> > > + /* capture_spec: */
> > > + EMIT2(0xF3, 0x90); /* pause */
> > > + EMIT3(0x0F, 0xAE, 0xE8); /* lfence */
> > > + EMIT2(0xEB, 0xF9); /* jmp <capture_spec> */
> > > + /* set_up_target: */
> > > + EMIT4(0x48, 0x89, 0x04, 0x24); /* mov %rax,(%rsp) */
> > > + EMIT1(0xC3); /* retq */
> > > +
> > > + BUILD_BUG_ON(cnt != RETPOLINE_SIZE);
> > > + *pprog = prog;
> >
> > You might define the actual code sequence (and length) in
> > arch/x86/include/asm/nospec-branch.h
> >
> > If we need to adjust code sequences for RETPOLINE, then we wont
> > forget/miss that arch/x86/net/bpf_jit_comp.c had it hard-coded.
>
> like adding a comment to asm/nospec-branch.h that says
> "dont forget to adjust bpf_jit_comp.c" ?
> but clang/gcc generate slightly different sequences for
> retpoline anyway, so even if '.macro RETPOLINE_JMP' in
> nospec-branch.h changes it doesn't mean that x64 jit has to change.
> So what kinda comment there would make sense?
I was thinking of something very explicit :
/* byte sequence for following assembly code used by eBPF
call ...
...
retq
*/
#define RETPOLINE_RAX_DIRECT_FOR_EBPF \
EMIT1_off32(0xE8, 7); /* callq <set_up_target> */ \
/* capture_spec: */ \
EMIT2(0xF3, 0x90); /* pause */ \
EMIT3(0x0F, 0xAE, 0xE8); /* lfence */ \
EMIT2(0xEB, 0xF9); /* jmp <capture_spec> */ \
/* set_up_target: */ \
EMIT4(0x48, 0x89, 0x04, 0x24); /* mov %rax,(%rsp) */ \
EMIT1(0xC3); /* retq */ \
Might be simply byte encoded, (array of 17 bytes)
Well, something like that anyway...
