Re: kprobes broken since 0d00449c7a28 ("x86: Replace ist_enter() with nmi_enter()")

[Alexei Starovoitov]

On Sat, Jan 30, 2021 at 11:02:49AM +0900, Masami Hiramatsu wrote:
> On Fri, 29 Jan 2021 18:59:43 +0100
> Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> > On Fri, Jan 29, 2021 at 09:45:48AM -0800, Alexei Starovoitov wrote:
> > > Same things apply to bpf side. We can statically prove safety for
> > > ftrace and kprobe attaching whereas to deal with NMI situation we
> > > have to use run-time checks for recursion prevention, etc.
> > 
> > I have no idea what you're saying. You can attach to functions that are
> > called with random locks held, you can create kprobes in some very
> > sensitive places.
> > 
> > What can you staticlly prove about that?
> 
> For the bpf and the kprobe tracer, if a probe hits in the NMI context,
> it can call the handler with another handler processing events.
> 
> kprobes is carefully avoiding the deadlock by checking recursion
> with per-cpu variable. But if the handler is shared with the other events
> like tracepoints, it needs to its own recursion cheker too.
> 
> So, Alexei, maybe you need something like this instead of in_nmi() check.
> 
> DEFINE_PER_CPU(bool, under_running_bpf);
> 
> common_handler()
> {
> 	if (__this_cpu_read(under_running_bpf))
> 		return;
> 	__this_cpu_write(under_running_bpf, true);
> 	/* execute bpf prog */
> 	__this_cpu_write(under_running_bpf, false);	
> }
> 
> Does this work for you?

This exactly check is already in trace_call_bpf.
Right after if (in_nmi()).
See bpf_prog_active. It serves different purpose though.
Simply removing if (in_nmi()) from trace_call_bpf is a bit scary.
I need to analyze all code paths first.