Re: [RFC] kretprobe: Prevent triggering kretprobe from within kprobe_flush_task

[Masami Hiramatsu]

On Thu, 9 Apr 2020 14:52:13 +0200
Jiri Olsa <jolsa@xxxxxxxxxx> wrote:

> On Thu, Apr 09, 2020 at 09:38:06PM +0900, Masami Hiramatsu wrote:
> > Hi Jiri,
> > 
> > On Wed,  8 Apr 2020 18:46:41 +0200
> > Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
> > 
> > > hi,
> > > Ziqian reported lockup when adding retprobe on _raw_spin_lock_irqsave.
> > 
> > Hmm, kprobe is lockless, but kretprobe involves spinlock.
> > Thus, eventually, I will blacklist the _raw_spin_lock_irqsave()
> > for kretprobe.
> 
> I thought of blacklisting, but we were using that kretprobe
> in a bcc script.. it's not overloaded by using bpf trampolines,
> but still might be useful
> 
> SNIP
> 
> > > The code within the kretprobe handler checks for probe reentrancy,
> > > so we won't trigger any _raw_spin_lock_irqsave probe in there.
> > > 
> > > The problem is in outside kprobe_flush_task, where we call:
> > > 
> > >   kprobe_flush_task
> > >     kretprobe_table_lock
> > >       raw_spin_lock_irqsave
> > >         _raw_spin_lock_irqsave
> > > 
> > > where _raw_spin_lock_irqsave triggers the kretprobe and installs
> > > kretprobe_trampoline handler on _raw_spin_lock_irqsave return.
> > 
> > Hmm, OK. In this case, I think we should mark this process is
> > going to die and never try to kretprobe on it.
> > 
> > > 
> > > The kretprobe_trampoline handler is then executed with already
> > > locked kretprobe_table_locks, and first thing it does is to
> > > lock kretprobe_table_locks ;-) the whole lockup path like:
> > > 
> > >   kprobe_flush_task
> > >     kretprobe_table_lock
> > >       raw_spin_lock_irqsave
> > >         _raw_spin_lock_irqsave ---> probe triggered, kretprobe_trampoline installed
> > > 
> > >         ---> kretprobe_table_locks locked
> > > 
> > >         kretprobe_trampoline
> > >           trampoline_handler
> > >             kretprobe_hash_lock(current, &head, &flags);  <--- deadlock
> > > 
> > > The change below sets current_kprobe in kprobe_flush_task, so the probe
> > > recursion protection check is hit and the probe is never set. It seems
> > > to fix the deadlock.
> > > 
> > > I'm not sure this is the best fix, any ideas are welcome ;-)
> > 
> > Hmm, this is a bit tricky to fix this issue. Of course, temporary disable
> > kprobes (and kretprobe) on an area by filling current_kprobe might
> > be a good idea, but it also involves other kprobes.
> > 
> > How about let kretprobe skip the task which state == TASK_DEAD ?
> 
> hum, isn't that considerable amount of paths (with state == TASK_DEAD)
> that we would kill kprobes for? someone might want to trace it

OK, and I found that even after calling kprobe_flush_task(), it may be
work because the task will not be switched. kretprobe instance will be
reclaimed.

Thank you,


-- 
Masami Hiramatsu <mhiramat@xxxxxxxxxx>