On Thu, 29 Oct 2020 09:40:01 -0400
Steven Rostedt <rost...@goodmis.org> wrote:
> On Thu, 29 Oct 2020 16:58:03 +0900
> Masami Hiramatsu <mhira...@kernel.org> wrote:
>
> > Hi Steve,
> >
> > On Wed, 28 Oct 2020 07:52:49 -0400
> > Steven Rostedt <rost...@goodmis.org> wrote:
> >
> > > From: "Steven Rostedt (VMware)" <rost...@goodmis.org>
> > >
> > > If a ftrace callback does not supply its own recursion protection and
> > > does not set the RECURSION_SAFE flag in its ftrace_ops, then ftrace will
> > > make a helper trampoline to do so before calling the callback instead of
> > > just calling the callback directly.
> >
> > So in that case the handlers will be called without preempt disabled?
> >
> >
> > > The default for ftrace_ops is going to assume recursion protection unless
> > > otherwise specified.
> >
> > This seems to skip entier handler if ftrace finds recursion.
> > I would like to increment the missed counter even in that case.
>
> Note, this code does not change the functionality at this point, because
> without having the FL_RECURSION flag set (which kprobes does not even in
> this patch), it always gets called from the helper function that does this:
>
> bit = trace_test_and_set_recursion(TRACE_LIST_START, TRACE_LIST_MAX);
> if (bit < 0)
> return;
>
> preempt_disable_notrace();
>
> op->func(ip, parent_ip, op, regs);
>
> preempt_enable_notrace();
> trace_clear_recursion(bit);
>
> Where this function gets called by op->func().
>
> In other words, you don't get that count anyway, and I don't think you want
> it. Because it means you traced something that your callback calls.
Got it. So nmissed count increment will be an improvement.
>
> That bit check is basically a nop, because the last patch in this series
> will make the default that everything has recursion protection, but at this
> patch the test does this:
>
> /* A previous recursion check was made */
> if ((val & TRACE_CONTEXT_MASK) > max)
> return 0;
>
> Which would always return true, because this function is called via the
> helper that already did the trace_test_and_set_recursion() which, if it
> made it this far, the val would always be greater than max.
OK, let me check the last patch too.
>
> >
> > [...]
> > e.g.
> >
> > > diff --git a/arch/csky/kernel/probes/ftrace.c
> > > b/arch/csky/kernel/probes/ftrace.c
> > > index 5264763d05be..5eb2604fdf71 100644
> > > --- a/arch/csky/kernel/probes/ftrace.c
> > > +++ b/arch/csky/kernel/probes/ftrace.c
> > > @@ -13,16 +13,21 @@ int arch_check_ftrace_location(struct kprobe *p)
> > > void kprobe_ftrace_handler(unsigned long ip, unsigned long parent_ip,
> > > struct ftrace_ops *ops, struct pt_regs *regs)
> > > {
> > > + int bit;
> > > bool lr_saver = false;
> > > struct kprobe *p;
> > > struct kprobe_ctlblk *kcb;
> > >
> > > - /* Preempt is disabled by ftrace */
> > > + bit = ftrace_test_recursion_trylock();
> >
> > > +
> > > + preempt_disable_notrace();
> > > p = get_kprobe((kprobe_opcode_t *)ip);
> > > if (!p) {
> > > p = get_kprobe((kprobe_opcode_t *)(ip - MCOUNT_INSN_SIZE));
> > > if (unlikely(!p) || kprobe_disabled(p))
> > > - return;
> > > + goto out;
> > > lr_saver = true;
> > > }
> >
> > if (bit < 0) {
> > kprobes_inc_nmissed_count(p);
> > goto out;
> > }
>
> If anything called in get_kprobe() or kprobes_inc_nmissed_count() gets
> traced here, you have zero recursion protection, and this will crash the
> machine with a likely reboot (triple fault).
Oops, ok, those can be traced.
>
> Note, the recursion handles interrupts and wont stop them. bit < 0 only
> happens if you recurse because this function called something that ends up
> calling itself. Really, why would you care about missing a kprobe on the
> same kprobe?
Usually, sw-breakpoint based kprobes will count that case. Moreover, kprobes
shares one ftrace_ops among all kprobes. I guess in that case any kprobes
in kprobes (e.g. recursive call inside kprobe pre_handlers) will be skipped
by ftrace_test_recursion_trylock(), is that correct?
Thank you,
>
> -- Steve
--
Masami Hiramatsu <mhira...@kernel.org>
