Re: [PATCH] ring-buffer: add paranoid checks for loops

[Steven Rostedt]

On Thu, 30 Oct 2008, Ingo Molnar wrote:
> 
> hm, all those magic constants look a bit like voodoo and make the 
> patch ugly, and people who read this will be confused about the 
> purpose for sure.

Point taken.

> 
> But the checks are still worth having in practice. So could you please 
> improve the comments, to come up with some tangible calculation that 
> leads to these constants?
> 
> For example the '1000' constant, how did you come to that? Could you 
> estimate what type of interrupt storm is needed to trigger it falsely? 
> So instead of this comment:

My original number was 100,000, but I thought that a bit high ;-)
Since it is OK for an interrupt to preempt this code and perform a trace, 
which would make the condition fail by the one being preempted. The 
likelyhood of an interrupt coming in at that location 1000 times in a row 
seems to be awefully low. It's not enough that a 1000 interrupts come in, 
the task being preempted must loop 1000 times and have a trace interrupt 
cause the condition to fail each time. I'll explain it this way in the 
comments.

I picked a big number because I can see a traced interrupt that is very 
active causing several interruptions in this code.

> 
> > +	 * If we loop here 1,000 times, that means we are either
> > +	 * in an interrupt storm, or we have something buggy.
> > +	 * Bail!
> 
> something like this might look more acceptable:
> 
> > +	 * If we loop here 1,000 times, that means we are either
> > +	 * in an interrupt storm that preempted the same trace-entry
> > +	 * attempt 1000 times in a row, or we have a bug in the tracer.
> > +	 * Bail!
> 
> i.e. please exaplain every single magic number there so that it can be 
> followed how you got to that number, and what precise effects that 
> number has.
> 
> In the cases where you just guessed a number based on experiments, 
> please think it through and insert an analysis about the effects of 
> that number.
> 
> Would this be doable?

Again, there are small "allowable" races that causes the code to loop a 
few times.  I'll try to explain them a bit better in the comments.
There's small races between the reader and writer that can hit just right 
to cause a "loop again". But these chances are much smaller than the 
interrupt tracing situation.

I'll look deeper at the reasons for the races and explain them a bit 
better.

Thanks,

-- Steve

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/