On Fri, Mar 19, 2021 at 5:14 PM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Fri, Mar 19, 2021 at 5:00 PM Stanislav Fomichev <sdf@xxxxxxxxxx> wrote:
> >
> > __bpf_arch_text_poke does rewrite only for atomic nop5, emit_nops(xxx, 5)
> > emits non-atomic one which breaks fentry/fexit with k8 atomics:
> >
> > P6_NOP5 == P6_NOP5_ATOMIC (0f1f440000 == 0f1f440000)
> > K8_NOP5 != K8_NOP5_ATOMIC (6666906690 != 6666666690)
> >
> > Can be reproduced by doing "ideal_nops = k8_nops" in "arch_init_ideal_nops()
> > and running fexit_bpf2bpf selftest.
> >
> > Fixes: e21aa341785c ("bpf: Fix fexit trampoline.")
> > Signed-off-by: Stanislav Fomichev <sdf@xxxxxxxxxx>
> > ---
> > arch/x86/net/bpf_jit_comp.c | 3 ++-
> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
> > index 72b5a57e9e31..b35fc8023884 100644
> > --- a/arch/x86/net/bpf_jit_comp.c
> > +++ b/arch/x86/net/bpf_jit_comp.c
> > @@ -2012,7 +2012,8 @@ int arch_prepare_bpf_trampoline(struct bpf_tramp_image *im, void *image, void *i
> > /* remember return value in a stack for bpf prog to access */
> > emit_stx(&prog, BPF_DW, BPF_REG_FP, BPF_REG_0, -8);
> > im->ip_after_call = prog;
> > - emit_nops(&prog, 5);
> > + memcpy(prog, ideal_nops[NOP_ATOMIC5], X86_PATCH_SIZE);
> > + prog += X86_PATCH_SIZE;
>
> I'm well aware, but ideal_nops are pretty much gone already.
> The changes are already in the -tip tree.
> So I decided to reduce the conflicts for the merge window.
>
> Do you actually see the breakage or it's purely theoretical?
We do see it, but it's on our tree that pulls from bpf.
And it obviously doesn't have that "x86: Remove dynamic NOP selection" yet.
Thanks for the pointer, I guess I can just wait for the real merge then.
