linux-kernel - Re: perf_events: zero time running and enabled, but non-zero count

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <4AE50B55.4070706@linux.vnet.ibm.com>
Date:	Sun, 25 Oct 2009 19:37:09 -0700
From:	Corey Ashford <cjashfor@...ux.vnet.ibm.com>
To:	Ingo Molnar <mingo@...e.hu>, paulus@....ibm.com
CC:	Peter Zijlstra <a.p.zijlstra@...llo.nl>,
	LKML <linux-kernel@...r.kernel.org>
Subject: Re: perf_events: zero time running and enabled, but non-zero count

On 10/23/2009 01:57 PM, Ingo Molnar wrote:
>
> * Corey Ashford<cjashfor@...ux.vnet.ibm.com>  wrote:
>
>>> Btw., you might want to consider putting such testcases into a new
>>> 'perf test' kind of utility (under tools/perf/builtin-test.c) that
>>> does various tests such as reading a count and validating the
>>> time-running/time-enabled values - so that we can see any bugs
>>> directly.
>>
>> I'll see if I can put the test case in some sort of acceptable format
>> like that.  Maybe we can use this one test case as the starting point
>> for builtin-test.c.
>
> Yeah. Feel free to add just something minimal that solves your problem.
> Others will add their testcases too i'm sure, once the basic command is
> there.
>
> 	Ingo

Hi Ingo and Paul,

I followed the directions Ingo provided for getting the -tip tree and 
then attempted to build the perf tool that's in it and it doesn't build. 
  I'm getting this error on a Power5 box:

../../lib/find_next_bit.c: In function ‘ext2_swabp’:
../../lib/find_next_bit.c:168: warning: passing argument 1 of 
‘__swab64p’ from incompatible pointer type
make: *** [util/find_next_bit.o] Error 1

I looked at the code and it looks ok to me.  There's a cast from an 
"unsigned long *" to a "unsigned long long *", but those types should be 
of the same size (in -m64 mode), so I don't know why this is failing.

Some good news is that I have created a simple test case which 
demonstrates the bug, but I haven't formulated it in the builtin-test.c 
format yet, since I have been able to build perf.

If you toss the attached file into the perf source directory and compile 
it as follows:

gcc -m64 -o petb perf_events_time_bug.c -lrt

it will show the error:
% ./petb
[0] = 1
[1] = 0 <- time enabled
[2] = 0 <- time running
[3] = 4559d1f8
[4] = 1
test failed!

perf_events_time_bug.c -lrt

If you compile it with:

gcc -DDO_DISABLE -m66 -o petb perf_events_time_bug.c -lrt

This will disable the event before reading it and then you will get the 
correct times.

% ./petb
[0] = 1
[1] = 2a50c1c8 <- time enabled
[2] = 2a50c1c8 <- time running
[3] = 460dea97
[4] = 2
test passed!

-- 
Regards,

- Corey

Corey Ashford
Software Engineer
IBM Linux Technology Center, Linux Toolchain
Beaverton, OR
cjashfor@...ibm.com

View attachment "perf_events_time_bug.c" of type "text/plain" (2755 bytes)