Re: [patch] perf_event_open enable sysfs exported events

[Dave Jones <davej@redhat.com>]

On Wed, Jun 19, 2013 at 12:34:43PM -0400, Vince Weaver wrote:
 > On Wed, 19 Jun 2013, Dave Jones wrote:
 > 
 > > On Thu, Jun 13, 2013 at 03:58:37PM -0400, Vince Weaver wrote:
 > >  > On Thu, 13 Jun 2013, Tommi Rantala wrote:
 > >  > 
 > >  > > Unable to opendir /sys/bus/event_source/devices : Resource temporarily
 > >  > > unavailable
 > >  >
 > >  > Though in the case where /sys/bus/event_source/devices isn't available
 > >  > trinity will try opening it again each time it tries to do a 
 > >  > PERF_TYPE_READ_FROM_SYSFS type (1 time in 7).  Not sure if it's
 > >  > worth rate-limiting that.
 > > 
 > > I'm puzzled why I'm seeing opendir() fail with -ENOMEM when there's gigabytes
 > > of free memory available.  Likewise, getting -EAGAIN seems.. weird.
 > 
 > Is this easy to reproduce?  I haven't seen it locally but then again
 > I have only been running trinity with -c perf_event_open
 > I only left the debug message in because I thought it would be unusual
 > for a quick sysfs traversal to fail.

Hmm, that's weird. When I run with -c perf_event_open it doesn't happen.
So it's an interaction with something else that's causing it. Fun.
 
 > > init_pmus might be simpler if it was converted to use nftw() too.
 > 
 > I was unaware of the existence of nftw(), I'll work on updating the code
 > to use it.

I only recently learned of it myself. the tree walking in files.c got
a lot simpler. See b0d4df18a6785b0c3347f521bb2c09e88b0c0966.

	Dave