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

From: Masami Hiramatsu <mhiramat@kernel.org>
To: Steven Rostedt <rostedt@goodmis.org>
Cc: Peter Zijlstra <peterz@infradead.org>,
	Alexei Starovoitov <alexei.starovoitov@gmail.com>,
	Masami Hiramatsu <mhiramat@kernel.org>,
	Nikolay Borisov <nborisov@suse.com>,
	LKML <linux-kernel@vger.kernel.org>,
	Alexei Starovoitov <ast@kernel.org>, bpf <bpf@vger.kernel.org>,
	Josh Poimboeuf <jpoimboe@redhat.com>
Subject: Re: kprobes broken since 0d00449c7a28 ("x86: Replace ist_enter() with nmi_enter()")
Date: Wed, 3 Feb 2021 22:33:28 +0900	[thread overview]
Message-ID: <20210203223328.9e99548d19d482ac2e2cda81@kernel.org> (raw)
In-Reply-To: <20210202160513.38ada3a7@gandalf.local.home>

On Tue, 2 Feb 2021 16:05:13 -0500
Steven Rostedt <rostedt@goodmis.org> wrote:

> On Tue, 2 Feb 2021 19:30:34 +0100
> Peter Zijlstra <peterz@infradead.org> wrote:
> 
> > That does mean that kprobes are then fundamentally running from
> > in_nmi(), which is what started all this.
> 
> I just thought about the fact that tracing records the context of the
> function it is called in. If you set "in_nmi()" for all ftrace handlers,
> then all functions will look like they are in an NMI context during tracing.
> 
> That is, the preempt count is checked to fill in the flags in the ring
> buffer that denotes what context the event (in this case the function) was
> called in.

Ah, that is what I worried about. ftrace and kprobes handler usually want to
know "what is the actual status of the system where the probe hits".

If the new kernel_exception_enter() for ftrace/kprobes or any other kernel
instrumention does

  __preempt_count_add(KEX_OFFSET + NMI_OFFSET + HARDIRQ_OFFSET);

And we can distinguish the KEX from NMI, and get the original status of the context.
What would you think about?

Thank you,

-- 
Masami Hiramatsu <mhiramat@kernel.org>