lkml.org 
[lkml]   [2017]   [Sep]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 2/2] tracing: Add support for critical section events
On Mon, Sep 04, 2017 at 12:04:44PM -0700, Joel Fernandes wrote:
> Hi Peter,
>
> On Mon, Sep 4, 2017 at 12:56 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> > On Sun, Sep 03, 2017 at 01:50:51AM -0700, Joel Fernandes wrote:
> >> Critical section trace events can be used for tracing the start and
> >> end of a critical section which can be used by a trace viewer like
> >> systrace to graphically view the start and end of a critical section and
> >> correlate them with latencies and scheduling issues.
> >>
> >> Reason for starting critical section:
> >> Either preemption or interrupts were turned off.
> >> Reason for ending critical section:
> >> Both preemption and interrupts are now turned back on.
> >
> > Please don't use the name critical section for this.
>
> I can change the name to something else, but at the moment I can't
> think of anything better. Could you suggest a better name? Also btw,
> 'critical timings' is the terminology used within the irqsoff tracer
> so this is in line with that.

So 'critical section' is what some mis-guided people call the locked
region of a lock :-) Using it for something else is prone to cause more
confusion...

I would simply call them what they are: irq_disable,irq_enable
preempt_disable,preempt_enable.

> > Also IRQ and preempt already have a gazillion trace hooks, why do we need more?
>
> The goal of the patch is not to add more hooks, but to make it
> possible for userspace to use the trace-events mechanism to see these
> events, so that these events can be seen along with other trace events
> when using event tracing, using a convenient trace event interface
> which can be enabled by userspace.

But only when you're already building a debug kernel and we already have
these irq/preempt trace things in, right?

> This makes it possible to visually
> see these events as a timeline along with other kernel and userspace
> events. See more description in my coverletter [1] and a screenshot of
> how this is used [2] by the systrace trace viewer. These patches make
> it possible for this.

Yeah, for some reason these graphics things don't work for me... And
regular traces already have the irq-off and preempt-depth column, which
typically is enough. But I suppose I can see the value of allowing
explicit events for them.

> Also, this work is along the same lines of what we discussed in a
> recent conference about adding irqsoff and preemptoff as real trace
> events, and then using synthetic events (which can combine start and
> end event into a single event) to redo the irqsoff tracer, so in that
> sense it is in the right direction. Synthetic events is some time away
> though from being merged AFAIK.

I have no memories of that, but that sounds OK :-)

\
 
 \ /
  Last update: 2017-09-04 21:45    [W:0.079 / U:1.324 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site