Re: [RFC][PATCH] kprobes: Add separate preempt_disabling for kprobes

[Steven Rostedt <rostedt@goodmis.org>]

On Thu, 2011-06-30 at 18:14 +0200, Frederic Weisbecker wrote:


> That's a bit sad we need to bloat preempt_schedule() with a new test, especially
> as kprobes should not add overhead in the off case and then I guess many distros
> enable kprobes by default, so it's probably not just enabled on some debug kernel.

A simple per_cpu var test is not that bad, and that's also why I put it
where I did. It only gets checked after all the other locations fail. I
doubt this really adds any measurable overhead. Note, most distro's
don't even enable CONFIG_PREEMPT so this isn't even touched by them.

> 
> Another idea could be to turn current_thread_info()->preempt_count into a local_t
> var which ensures a single incrementation is performed in a single instruction.

Not all archs support such a thing.

> 
> Well, in the hope that every arch can do something like:
> 
> inc (mem_addr)
> 
> IIRC, I believe arm can't... And the default local_inc is probably not helpful
> in our case...

Some archs (I believe) perform local_inc() slower than i++. In which
case, this solution will more likely slow things down even more than
what I proposed. As the impact will be on every preempt_disable and
enable.

And this would not even help for the place that I actually hit the crash
on. Which was in the scheduler. It wasn't the addition of the
preempt_count that caused my crash, it was just reading it.

	if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
		__schedule_bug(prev);

The in_atomic_preempt_off() is what blew up on me.

#define in_atomic_preempt_off() \
		((preempt_count() & ~PREEMPT_ACTIVE) != PREEMPT_CHECK_OFFSET)

What happened was that the reading of preempt_count was single stepped.
And that had preempt_count set to one more due to the preempt_disable()
in kprobes. Even if preempt_count was a local_t, this bug would have
still triggered, and my system would have crashed anyway. (every
schedule caused a printk to happen).


-- Steve