Re: [RFC PATCH] tracing: Don't call wakeup() when committing the event

[Frederic Weisbecker <fweisbec@gmail.com>]

On Tue, May 03, 2011 at 02:56:11PM -0700, Vaibhav Nagarnaik wrote:
> On Tue, May 3, 2011 at 2:41 PM, Frederic Weisbecker <fweisbec@gmail.com> wrote:
> > On Tue, May 03, 2011 at 02:03:36PM -0700, Vaibhav Nagarnaik wrote:
> >> In using syscall tracing by concurrent processes, the wakeup() that is
> >> called in the event commit function causes contention on the spin lock
> >> of the waitqueue. I enabled sys_enter_getuid and sys_exit_getuid
> >> tracepoints, and by running getuid_microbench from autotest in parallel
> >> I found that the contention causes exponential latency increase in the
> >> tracing path.
> >>
> >> The autotest binary getuid_microbench calls getuid() in a tight loop for
> >> the given number of iterations and measures the average time required to
> >> complete a single invocation of syscall.
> >>
> >> The patch here points to the problem and provides a naive solution to
> >> start the discussion. It is not intended to be a definitive solution.
> >
> > Right, so another solution could be to have per cpu waitqueues for
> > the per_cpu trace_pipe/trace_pipe_raw files, and one big for the main
> > trace_pipe file.
> 
> That could be another way. But if there is still *one* common waitqueue for
> the main trace file, we are still going to get contention on waking up that
> common waitqueue.
> 
> Unless I am missing something, can you explain why there won't be contention
> in your suggested solution?

Ah right, I missed that.

I wonder if we should have a lite version of wake_up() that checks
if the list of waiters is empty before locking the queue.
After all we don't care much about tight races for tracing.

Hm?