Re: [PATCH] x86/ibt: make is_endbr() notrace

[Menglong Dong]

On Fri, Sep 19, 2025 at 1:54 AM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Thu, Sep 18, 2025 at 9:59 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > On Thu, Sep 18, 2025 at 09:02:31AM -0700, Alexei Starovoitov wrote:
> > > On Thu, Sep 18, 2025 at 6:32???AM Menglong Dong <menglong8.dong@xxxxxxxxx> wrote:
> > > >
> > > > On Thu, Sep 18, 2025 at 9:05???PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > > > >
> > > > > On Thu, Sep 18, 2025 at 08:09:39PM +0800, Menglong Dong wrote:
> > > > > > is_endbr() is called in __ftrace_return_to_handler -> fprobe_return ->
> > > > > > kprobe_multi_link_exit_handler -> is_endbr.
> > > > > >
> > > > > > It is not protected by the "bpf_prog_active", so it can't be traced by
> > > > > > kprobe-multi, which can cause recurring and panic the kernel. Fix it by
> > > > > > make it notrace.
> > > > >
> > > > > This is very much a riddle wrapped in an enigma. Notably
> > > > > kprobe_multi_link_exit_handler() does not call is_endbr(). Nor is that
> > > > > cryptic next line sufficient to explain why its a problem.
> > > > >
> > > > > I suspect the is_endbr() you did mean is the one in
> > > > > arch_ftrace_get_symaddr(), but who knows.
> > > >
> > > > Yeah, I mean
> > > > kprobe_multi_link_exit_handler -> ftrace_get_entry_ip ->
> > > > arch_ftrace_get_symaddr -> is_endbr
> > > > actually. And CONFIG_X86_KERNEL_IBT is enabled of course.
> > >
> > > All this makes sense to me.
> >
> > As written down, I'm still clueless.

Ok, let me describe the problem in deetail.

First of all, it has nothing to do with kprobe. The bpf program of type
kprobe-multi based on fprobe, and fprobe base on fgraph. So it's all
about the ftrace, which means __fentry__.

Second, let me explain the recur detection of the kprobe-multi. Let's
take the is_endbr() for example. When it is hooked by the bpf program
of type kretprobe-multi, following calling chain will happen:

  is_endbr -> __ftrace_return_to_handler -> fprobe_return ->
  kprobe_multi_link_exit_handler -> ftrace_get_entry_ip ->
  arch_ftrace_get_symaddr -> is_endbr

Look, is_endbr() is called again during the ftrace handler, so it will
trigger the ftrace handler(__ftrace_return_to_handler) again, which
causes recurrence.

Such recurrence can be detected. In kprobe_multi_link_prog_run(),
the percpu various "bpf_prog_active" will be increased by 1 before we
run the bpf progs, and decrease by 1 after the bpf progs finish. If the
kprobe_multi_link_prog_run() is triggered again during bpf progs run,
it will check if bpf_prog_active is zero, and return directly if it is not.
Therefore, recurrence can't happen within the "bpf_prog_active" protection.

However, the calling to is_endbr() is not within that scope, which makes
the recurrence happen.

Hope I described it clearly 😉

Thanks!
Menglong Dong

> >
> > > __noendbr bool is_endbr(u32 *val) needs "notrace",
> > > since it's in alternative.c and won't get inlined (unless LTO+luck).
> >
> > notrace don't help with kprobes in general, only with __fentry__ sites.
>
> Are you sure ? Last time I checked "notrace" prevents kprobing
> anywhere in that function.