Re: [PATCH -tip 0/5] kprobes: Fix stacktrace in kretprobes

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Thu, Mar 11, 2021 at 10:54:38AM +0900, Masami Hiramatsu wrote:
> On Wed, 10 Mar 2021 19:06:15 -0600
> Josh Poimboeuf <jpoimboe@redhat.com> wrote:
> 
> > On Thu, Mar 11, 2021 at 09:20:18AM +0900, Masami Hiramatsu wrote:
> > > > >  bool unwind_next_frame(struct unwind_state *state)
> > > > >  {
> > > > >  	unsigned long ip_p, sp, tmp, orig_ip = state->ip, prev_sp = state->sp;
> > > > > @@ -536,6 +561,18 @@ bool unwind_next_frame(struct unwind_state *state)
> > > > >  
> > > > >  		state->ip = ftrace_graph_ret_addr(state->task, &state->graph_idx,
> > > > >  						  state->ip, (void *)ip_p);
> > > > > +		/*
> > > > > +		 * There are special cases when the stack unwinder is called
> > > > > +		 * from the kretprobe handler or the interrupt handler which
> > > > > +		 * occurs in the kretprobe trampoline code. In those cases,
> > > > > +		 * %sp is shown on the stack instead of the return address.
> > > > > +		 * Or, when the unwinder find the return address is replaced
> > > > > +		 * by kretprobe_trampoline.
> > > > > +		 * In those cases, correct address can be found in kretprobe.
> > > > > +		 */
> > > > > +		if (state->ip == sp ||
> > > > 
> > > > Why is the 'state->ip == sp' needed?
> > > 
> > > As I commented above, until kretprobe_trampoline writes back the real
> > > address to the stack, sp value is there (which has been pushed by the
> > > 'pushq %rsp' at the entry of kretprobe_trampoline.)
> > > 
> > >         ".type kretprobe_trampoline, @function\n"
> > >         "kretprobe_trampoline:\n"
> > >         /* We don't bother saving the ss register */
> > >         "       pushq %rsp\n"				// THIS
> > >         "       pushfq\n"
> > > 
> > > Thus, from inside the kretprobe handler, like ftrace, you'll see
> > > the sp value instead of the real return address.
> > 
> > I see.  If you change is_kretprobe_trampoline_address() to include the
> > entire function, like:
> > 
> > static bool is_kretprobe_trampoline_address(unsigned long ip)
> > {
> > 	return (void *)ip >= kretprobe_trampoline &&
> > 	       (void *)ip < kretprobe_trampoline_end;
> > }
> > 
> > then the unwinder won't ever read the bogus %rsp value into state->ip,
> > and the 'state->ip == sp' check can be removed.
> 
> Hmm, I couldn't get your point. Since sp is the address of stack,
> it always out of text address.

When unwinding from trampoline_handler(), state->ip will point to the
instruction after the call:

	call trampoline_handler
	movq %rax, 19*8(%rsp)   <-- state->ip points to this insn

But then, the above version of is_kretprobe_trampoline_address() is
true, so state->ip gets immediately replaced with the real return
address:

	if (is_kretprobe_trampoline_address(state->ip))
		state->ip = orc_kretprobe_correct_ip(state);

so the unwinder skips over the kretprobe_trampoline() frame and goes
straight to the frame of the real return address.  Thus it never reads
this bogus return value into state->ip:

	pushq %rsp

which is why the weird 'state->ip == sp' check is no longer needed.

The only "downside" is that the unwinder skips the
kretprobe_trampoline() frame.  (note that downside wouldn't exist in
the case of UNWIND_HINT_REGS + valid regs->ip).

> > > > And it would make the unwinder just work automatically when unwinding
> > > > from the handler using the regs.
> > > > 
> > > > It would also work when unwinding from the handler's stack, if we put an
> > > > UNWIND_HINT_REGS after saving the regs.
> > > 
> > > At that moment, the real return address is not identified. So we can not
> > > put it.
> > 
> > True, at the time the regs are originally saved, the real return address
> > isn't available.  But by the time the user handler is called, the return
> > address *is* available.  So if the real return address were placed in
> > regs->ip before calling the handler, the unwinder could find it there,
> > when called from the handler.
> 
> OK, but this is not arch independent specification. I can make a hack
> only for x86, but that is not clean implementation, hmm.
> 
> > 
> > Then we wouldn't need the call to orc_kretprobe_correct_ip() in
> > __unwind_start().
> 
> What about the ORC implementation in other architecture? Is that for
> x86 only?

ORC is x86 only.

> > But maybe it's not possible due to the regs->ip expectations of legacy
> > handlers?
> 
> Usually, the legacy handlers will ignore it, the official way to access
> the correct return address is kretprobe_instance.ret_addr. Because it is
> arch independent.
> 
> Nowadays there are instruction_pointer() and instruction_pointer_set() APIs
> in many (not all) architecutre, so I can try to replace to use it instead
> of the kretprobe_instance.ret_addr.
> (and it will break the out-of-tree codes)

That sounds better to me, though I don't have an understanding of what
it would break.

-- 
Josh