Re: [PATCH] tracepoint: Do not warn on EEXIST or ENOENT

[Tetsuo Handa]

On 2021/06/26 23:18, Steven Rostedt wrote:
> On Sat, 26 Jun 2021 22:58:45 +0900
> Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx> wrote:
> 
>> syzbot is hitting WARN_ON_ONCE() at tracepoint_add_func() [1], but
>> func_add() returning -EEXIST and func_remove() returning -ENOENT are
>> not kernel bugs that can justify crashing the system.
> 
> There should be no path that registers a tracepoint twice. That's a bug
> in the kernel. Looking at the link below, I see the backtrace:
> 
> Call Trace:
>  tracepoint_probe_register_prio kernel/tracepoint.c:369 [inline]
>  tracepoint_probe_register+0x9c/0xe0 kernel/tracepoint.c:389
>  __bpf_probe_register kernel/trace/bpf_trace.c:2154 [inline]
>  bpf_probe_register+0x15a/0x1c0 kernel/trace/bpf_trace.c:2159
>  bpf_raw_tracepoint_open+0x34a/0x720 kernel/bpf/syscall.c:2878
>  __do_sys_bpf+0x2586/0x4f40 kernel/bpf/syscall.c:4435
>  do_syscall_64+0x3a/0xb0 arch/x86/entry/common.c:47
> 
> So BPF is allowing the user to register the same tracepoint more than
> once? That looks to be a bug in the BPF code where it shouldn't be
> allowing user space to register the same tracepoint multiple times.

I didn't catch your question.

  (1) func_add() can reject an attempt to add same tracepoint multiple times
      by returning -EINVAL to the caller.
  (2) But tracepoint_add_func() (the caller of func_add()) is calling WARN_ON_ONCE()
      if func_add() returned -EINVAL.
  (3) And tracepoint_add_func() is triggerable via request from userspace.
  (4) tracepoint_probe_register_prio() serializes tracepoint_add_func() call
      triggered by concurrent request from userspace using tracepoints_mutex mutex.
  (5) But tracepoint_add_func() does not check whether same tracepoint multiple
      is already registered before calling func_add().
  (6) As a result, tracepoint_add_func() receives -EINVAL from func_add(), and
      calls WARN_ON_ONCE() and the system crashes due to panic_on_warn == 1.

Why this is a bug in the BPF code? The BPF code is not allowing userspace to
register the same tracepoint multiple times. I think that tracepoint_add_func()
is stupid enough to crash the kernel instead of rejecting when an attempt to
register the same tracepoint multiple times is made.