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

[Steven Rostedt <rostedt@goodmis.org>]

[ Added some of the affected maintainers, left off David Howells and
David Miller due to LKML Cc limit ]

On Fri, 2011-07-01 at 10:22 +0900, Masami Hiramatsu wrote:
> (2011/07/01 6:56), Peter Zijlstra wrote:
> > On Thu, 2011-06-30 at 11:51 -0400, Steven Rostedt wrote:
> >>
> >> To solve this, I've added a per_cpu variable called
> >> kprobe_preempt_disabled, that is set by the kprobe code. If it is set,
> >> the preempt_schedule() will not preempt the code.
> 
> Sorry for replying so late :(

Heh, who can blame you? Timezones make open source development a
wait-and-see affair.

> 
> > Damn this is ugly. Can we step back and see if we can make the
> > requirement for kprobe to disable preemption go away?
> 
> As I replied right now, I think we can just eliminate that
> disabling preemption code. At least we'd better try it.
> I agree with you, introducing this kind of complexity
> just for kprobes is not what I want. :(

Note, I did clean up this patch, so it is not as fugly.

> 
> > Why does it have to do that anyway? Isn't it keeping enough per-task
> > state to allow preemption over the single step?
> 
> preemption itself must not happen on single stepping, but it seems
> impossible to do heavy context switching with setting TF bit...

Yeah, if all archs single step with interrupts disabled, then we should
be fine with removing preemption.

-- Steve