Re: perf on 2.6.38-rc4 wedges my box

[Arnaldo Carvalho de Melo <acme@ghostprotocols.net>]

Em Wed, Feb 09, 2011 at 01:12:22PM -0700, David Ahern escreveu:
> 
> 
> On 02/09/11 11:22, Jeff Moyer wrote:
> > David Ahern <dsahern@gmail.com> writes:
> > 
> >> Have you tried '-e cpu-clock' for S/W based profiling vs the default H/W
> >> profiling? Add -v to see if the fallback to S/W is happening now.
> > 
> > Thanks for the suggestion, David.  I tried:
> > 
> > # perf record -v ls
> >   Warning:  ... trying to fall back to cpu-clock-ticks
> > 
> > couldn't open /proc/-1/status
> > couldn't open /proc/-1/maps
> > [ls output]
> > [ perf record: Woken up 1 times to write data ]
> > [ perf record: Captured and wrote 0.008 MB perf.data (~363 samples) ]
> > 
> > If I explicitly set '-e cpu-clock', then the output is the same,
> > except that the warning is gone.  What's up with the /proc/-1/*?
> 
> target_{pid,tid} are initialized to -1 in builtin-record.c I believe the
> tid version is making its way through the event__synthesize_xxx code
> (event__synthesize_thread -> __event__synthesize_thread ->
> event__synthesize_comm and event__synthesize_mmap_events).

Yes, I'm working on a patch, probably just this will fix it:

diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index 07f8d6d..dd27b9f 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -680,7 +680,7 @@ static int __cmd_record(int argc, const char **argv)
 					       perf_event__synthesize_guest_os);
 
 	if (!system_wide)
-		perf_event__synthesize_thread(target_tid,
+		perf_event__synthesize_thread(evsel_list->threads->map[0],
 					      process_synthesized_event,
 					      session);
 	else
 
---

So that it gets the child_pid or the tid passed via --tid.

But thinking more I think that the correct is to pass the thread_map
evsel_list->threads, that will cover the case of just one thread
(--tid), or all a group of threads (--pid, when the thread_map will have
more than one tid).

I'll get a patch done later.

> > 
> > Now, when running perf record -e cpu-clock on the aio-stress run,
> > unsurprisingly, I get the same result:
> > 
> > # perf record -e cpu-clock -v -- ./aio-stress -O -o 0 -r 4 -d 32 -b 16 /dev/sds
> > couldn't open /proc/-1/status
> > couldn't open /proc/-1/maps
> > adding stage write
> > starting with write
> > file size 1024MB, record size 4KB, depth 32, ios per iteration 8
> > max io_submit 16, buffer alignment set to 4KB
> > threads 1 files 1 contexts 1 context offset 2MB verification off
> > adding file /dev/sds thread 0
> > 
> > and there it sits.  In this case, however, I did not see the NOHZ
> > warnings on the console, and this time the machine is still responding
> > to ping, but nothing else.
> 
> cpu-clock is handled through hrtimers if that helps understand the lockup.
> 
> David
> 
> > 
> > Cheers,
> > Jeff