Re: [PATCH] tracing: Refuse fprobe if RCU is not watching

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Mar 21, 2023 at 10:17:11AM -0400, Steven Rostedt wrote:
> On Tue, 21 Mar 2023 02:01:03 +0000
> Yafang Shao <laoar.shao@gmail.com> wrote:
> 
> > It hits below warning on my test machine when running
> > selftests/bpf/test_progs,
> > 
> > [  702.223611] ------------[ cut here ]------------
> > [  702.224168] RCU not on for: preempt_count_sub+0x0/0xa0
> > [  702.224770] WARNING: CPU: 14 PID: 5267 at include/linux/trace_recursion.h:162 fprobe_handler.part.0+0x1b8/0x1c0
> > [  702.231740] CPU: 14 PID: 5267 Comm: main_amd64 Kdump: loaded Tainted: G           O       6.2.0+ #584
> > [  702.233169] RIP: 0010:fprobe_handler.part.0+0x1b8/0x1c0
> > [  702.241388] Call Trace:
> > [  702.241615]  <TASK>
> > [  702.241811]  fprobe_handler+0x22/0x30
> > [  702.242129]  0xffffffffc04710f7
> > [  702.242417] RIP: 0010:preempt_count_sub+0x5/0xa0
> > [  702.242809] Code: c8 50 68 94 42 0e b5 48 cf e9 f9 fd ff ff 0f 1f 80 00 00 00 00 90 90 90 90 90 90 90 90 90 90 90 90 90 90 90 90 e8 4b cd 38 0b <55> 8b 0d 9c d0 cf 02 48 89 e5 85 c9 75 1b 65 8b 05 be 78 f4 4a 89
> > [  702.244752] RSP: 0018:ffffaf6187d27f10 EFLAGS: 00000082 ORIG_RAX: 0000000000000000
> > [  702.245801] RAX: 000000000000000e RBX: 0000000001b6ab72 RCX: 0000000000000000
> > [  702.246804] RDX: 0000000000000000 RSI: ffffffffb627967d RDI: 0000000000000001
> > [  702.247801] RBP: ffffaf6187d27f30 R08: 0000000000000000 R09: 0000000000000000
> > [  702.248786] R10: 0000000000000000 R11: 0000000000000000 R12: 00000000000000ca
> > [  702.249782] R13: ffffaf6187d27f58 R14: 0000000000000000 R15: 0000000000000000
> > [  702.250785]  ? preempt_count_sub+0x5/0xa0
> > [  702.251540]  ? syscall_enter_from_user_mode+0x96/0xc0
> > [  702.252368]  ? preempt_count_sub+0x5/0xa0
> > [  702.253104]  ? syscall_enter_from_user_mode+0x96/0xc0
> > [  702.253918]  do_syscall_64+0x16/0x90
> > [  702.254613]  entry_SYSCALL_64_after_hwframe+0x72/0xdc
> > [  702.255422] RIP: 0033:0x46b793
> > 
> > This issue happens under CONFIG_CONTEXT_TRACKING_USER=y. When a task
> > enters from user mode to kernel mode, or enters from user mode to irq,
> > it excutes preempt_count_sub before RCU begins watching, and thus this
> > warning is triggered.
> > 
> > We should not handle fprobe if RCU is not watching.
> > 
> > Suggested-by: Alexei Starovoitov <alexei.starovoitov@gmail.com>
> > Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
> > Cc: Alexei Starovoitov <alexei.starovoitov@gmail.com>
> > Cc: Andrii Nakryiko <andrii.nakryiko@gmail.com>
> > Cc: Jiri Olsa <olsajiri@gmail.com>
> > ---
> >  kernel/trace/fprobe.c | 3 +++
> >  1 file changed, 3 insertions(+)
> > 
> > diff --git a/kernel/trace/fprobe.c b/kernel/trace/fprobe.c
> > index e8143e3..fe4b248 100644
> > --- a/kernel/trace/fprobe.c
> > +++ b/kernel/trace/fprobe.c
> > @@ -27,6 +27,9 @@ static void fprobe_handler(unsigned long ip, unsigned long parent_ip,
> >  	struct fprobe *fp;
> >  	int bit;
> >  
> > +	if (!rcu_is_watching())
> > +		return;
> 
> Hmm, at least on 6.3, this should not be an issue anymore. I believe that
> all locations that have ftrace callbacks should now have rcu watching?
> 
> I think we *want* a warn on when this happens.
> 
> Peter?

You always want a wanring, because silently dropping stuff is very poor
form. But yes, we must not enter tracing then RCU isn't watching, that's
a fundamental fail and should be fixed.