kern/179827: [hwpmc] process-mode counters aren't correctly read on multi-core machines

[hiren panchasara]

The following reply was made to PR kern/179827; it has been noted by GNATS.

From: hiren panchasara <hiren.panchasara at gmail.com>
To: Adrian Chadd <adrian at freebsd.org>
Cc: bug-followup at freebsd.org
Subject: Re: kern/179827: [hwpmc] process-mode counters aren't correctly
 read on multi-core machines
Date: Sat, 22 Jun 2013 18:05:13 -0700

 On Sat, Jun 22, 2013 at 1:23 PM, Adrian Chadd <adrian at freebsd.org> wrote:
 > Right. Do that but with the test running in another window, so you
 > don't get the message overlap.
 
 Here is how it looks:
 
 #  p/instructions
         293898819
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
                 0
 #  p/instructions
      201677536862
 
 >
 > THen look at how the results show lots of '0' from pmcstat, and then
 > correlate that with the one-second "PMC,OPS" sample lines. See how
 > whenever they're run whilst the process is running (ie, the last event
 > was SWI) and it hasn't yet been de-scheduled. because it's still
 > running, the saved counter is never updated with the PMC counter and
 > subsequent reads (until it does get de-scheduled!) return the same
 > cached value. Hence, lots of '0's, followed by a big, big counter
 > value.
 
 Precisely how it looks.
 
 Thanks for the explanation.
 
 cheers,
 Hiren