On Mon, Nov 29, 2021 at 11:57:30AM +0100, Marco Elver wrote:

On Mon, Nov 29, 2021 at 04:47PM +0800, Boqun Feng wrote:

quoted

Hi Marco,

On Thu, Nov 18, 2021 at 09:10:07AM +0100, Marco Elver wrote:

quoted

Avoid checking scoped accesses from nested contexts (such as nested
interrupts or in scheduler code) which share the same kcsan_ctx.

This is to avoid detecting false positive races of accesses in the same

Could you provide an example for a false positive?

I think we do want to detect the following race:

	static int v = SOME_VALUE; // a percpu variable.
	static int other_v = ... ;

	void foo(..)
	{
		int tmp;
		int other_tmp;

		preempt_disable();
		{
			ASSERT_EXCLUSIVE_ACCESSS_SCOPED(v);
			tmp = v;
			
			other_tmp = other_v; // int_handler() may run here
			
			v = tmp + 2;
		}
		preempt_enabled();
	}

	void int_handler() // an interrupt handler
	{
		v++;
	}

, if I understand correctly, we can detect this currently, but with this
patch, we cannot detect this if the interrupt happens while we're doing
the check for "other_tmp = other_v;", right? Of course, running tests
multiple times may eventually catch this, but I just want to understand
what's this patch for, thanks!

The above will still be detected. Task and interrupt contexts in this
case are distinct, i.e. kcsan_ctx differ (see get_ctx()).

Ok, I was missing that.

But there are rare cases where kcsan_ctx is shared, such as nested
interrupts (NMI?), or when entering scheduler code -- which currently
has a KCSAN_SANITIZE := n, but I occasionally test it, which is how I
found this problem. The problem occurs frequently when enabling KCSAN in
kernel/sched and placing a random ASSERT_EXCLUSIVE_ACCESS_SCOPED() in
task context, or just enable "weak memory modeling" without this fix.
You also need CONFIG_PREEMPT=y + CONFIG_KCSAN_INTERRUPT_WATCHER=y.

Thanks for the background, it's now more clear that the problem is
triggered ;-)

The emphasis here really is on _shared kcsan_ctx_, which is not too
common. As noted in the commit description, we need to "[...] setting up
a watchpoint for a non-scoped (normal) access that also "conflicts" with
a current scoped access."

Consider this:

	static int v;
	int foo(..)
	{
		ASSERT_EXCLUSIVE_ACCESS_SCOPED(v);
		v++; // preempted during watchpoint for 'v++'
	}

Here we set up a scoped_access to be checked for v. Then on v++, a
watchpoint is set up for the normal access. While the watchpoint is set
up, the task is preempted and upon entering scheduler code, we're still
in_task() and 'current' is still the same, thus get_ctx() returns a
kcsan_ctx where the scoped_accesses list is non-empty containing the
scoped access for foo()'s ASSERT_EXCLUSIVE.

That means, when instrumenting scheduler code or any other code called
by scheduler code or nested interrupts (anything where get_ctx() still
returns the same as parent context), it'd now perform checks based on
the parent context's scoped access, and because the parent context also
has a watchpoint set up on the variable that conflicts with the scoped
access we'd report a nonsensical race.

Agreed.

This case is also possible:

	static int v;
	static int x;
	int foo(..)
	{
		ASSERT_EXCLUSIVE_ACCESS_SCOPED(v);
		x++; // preempted during watchpoint for 'v' after checking x++
	}

Here, all we need is for the scoped access to be checked after x++, end
up with a watchpoint for it, then enter scheduler code, which then
checked 'v', sees the conflicting watchpoint, and reports a nonsensical
race again.

Just to be clear, in both examples, the assumption is that 'v' is a
variable that scheduler code doesn't access, right? Because if scheduler
code does access 'v', then it's a problem that KCSAN should report. Yes,
I don't know any variable that scheduler exports, just to make sure
here.

By disallowing scoped access checking for a kcsan_ctx, we simply make
sure that in such nested contexts where kcsan_ctx is shared, none of
these nonsensical races would be detected nor reported.

Hopefully that clarifies what this is about.

Make sense to me, thanks.

Regards,
Boqun

Thanks,
-- Marco