On Thu, 2017-04-06 at 16:03 +1000, Michael Ellerman wrote:

Alistair Popple [off-list ref] writes:

quoted

On Wed, 23 Nov 2016 10:40:36 PM Balbir Singh wrote:

quoted

quoted

But overall I guess it's OK. We'd want to do a quick benchmark to make
sure it's not adding any overhead.

OK.. I'll try and find a benchmark and run it with traces disabled.

For what's it's worth I didn't notice any slow down running a NAMD
test with tracing disabled and ~400000 tlbie's in 34s. Turning tracing
on could have slowed things down ever so slightly, but I didn't notice
as it may have just been in the noise of the benchmark I was running
(I wasn't specifically looking at timing, hence the vagueness of the
remarks).

Otherwise the patch worked and would be useful - I have already had to
setup tlbie counting/tracing several times in the last 12 months.

Tested-by: Alistair Popple <redacted>

OK thanks.

To get it merged I'd like:
 - always called after the barriers.
 - no calls in loops, instead add a "count" field to the trace point and
   when we call it multiple times in a loop we pass the count.

So you'd see entries something like:

<...>-5141  [062]  1354.486693: tlbie: lpid=0, local=1, rb=0x7b5d0ff874f11f1, rs=0, ric=0, prs=0, r=0, count=128

I'll double check the patches and repost if required.

Balbir Singh.