Re: perf scripting

[Tom Zanussi]

Hi,

On Fri, 2010-09-17 at 19:32 +0900, Masami Hiramatsu wrote:
> (2010/09/16 21:08), Frederic Weisbecker wrote:
> > (Sorry to answer that so late)
> > 
> > 
> > On Sat, Aug 14, 2010 at 04:04:15PM -0400, Christoph Hellwig wrote:
> >> On Fri, Jul 30, 2010 at 04:04:42PM +0200, Frederic Weisbecker wrote:
> >>> I have the feeling you've made an ad-hoc post processing script that seems
> >>> to rewrite all the format parsing, debugfs, stream handling, etc... we
> >>> have that in perf tools already.
> >>>
> >>> May be you weren't aware of what we have in perf in terms of scripting support.
> >>
> >> Frederic, any chance you could help me getting a bit more familar with
> >> the perf perl scripting.  I currently have a hacky little sequence that
> >> I use to profile what callers generate XFS log traffic, and it like to
> >> turn it into a script so that I can do a direct perf call to use it
> >> to profile things without manual work, and generate nicer output.
> >>
> >> Currently it looks like this:
> >>
> >> perf probe --add xlog_sync
> >>
> >> perf record -g -e probe:xlog_sync -a -- <insert actualy workload here>
> >>
> >> then do
> >>
> >> perf report -n -g flat
> >>
> >> to get me the callchain in a readable format.
> >>
> >> Now what I'd really like is a perl script that can read a file like
> >> latencytop.trans (or just has the information embedded) which contains
> >> functions in the backtrace that we're interested in.
> >>
> >> E.g. one simple from the report command above may look like:
> >>
> >>                 xlog_sync
> >> 		xlog_write
> >> 		xlog_cil_push
> >> 		_xfs_log_force
> >> 		xfs_log_force
> >> 		xfs_sync_data
> >> 		xfs_quiesce_data
> >> 		xfs_fs_sync_fs
> >>
> >> In which case I'm interested in xfs_log_force and xfs_fs_sync_fs.  So
> >> the output of the perl script should looks something like:
> >>
> >>
> >>   Samples	Caller
> >> 	2	xfs_fs_sync_fs
> >> 	1	xfs_file_fsync
> >> 	1	xfs_commit_dummy_trans
> 
> BTW, if you want the caller for each call, you can do with perf probe
> 
> # perf probe --add 'xlog_sync caller=+0($stack)'
> 
> then, you can see the caller address in caller argument of
> xlog_sync event record.
> 
> 
> > Somehow, that's a kind of overview you can get with
> > perf report, using the default fractal mode or the graph mode.
> > Callers are sorted by hits in these modes (actually in raw mode too).
> > 
> > But it could be interesting to add the callchains as arguments to the
> > perl/python scripting handlers for precise usecases.
> > 
> >  
> >> Or if I have a way to parse the argument of the probe (in the worst case
> >> I can replace it with a trace event if that makes it easier):
> >>
> >>   Samples	Flags		Callers
> >> 	1	sync		xfs_fs_sync_fs
> >> 	1			xfs_fs_sync_fs
> >> 	1	sync		xfs_file_fsync
> >> 	1	sync		xfs_commit_dummy_trans
> > 
> > 
> > So for example that becomes an even more precise usecase.
> > Currently the perf scripting engine doesn't give you access
> > to the callchains of a trace sample. That would be a nice feature
> > and would solve your problem.
> 
> AFAIK, perf perl script already supports getting arguments of
> events. e.g.
> 
> sub probe::xlog_sync
> {
>     my ($event_name, $context, $common_cpu, $common_secs, $common_nsecs,
>         $common_pid, $common_comm,
>         $caller) = @_;
> 
>     if (!defined($caller_list{$caller})) {
>          	$caller_list{$caller} = 0;
>     }	
>     $caller_list{$caller}++;
> }
> 
> for count up caller address.
> (However, perf perl currently doesn't have address-symbol translation
>  function. )
> 
> If perf scripting supports calling perf internally for defining
> new events for the script, it will be useful (from the viewpoint
> of script packaging).
> 
> Thank you,
> 
> > 
> > Tom, what do you think about that? This could be a special mode
> > requested by the user, or something made automatically if callchains
> > are present in samples. We could add a specific callchain extra
> > argument to the generated scripting handlers, or this could
> > be a generic extra dict argument that can contain whatever we want
> > (perf sample headers, etc...), whatever extra data the user might
> > request.
> > 
> > What do you think?
> > 

Rather than adding extra arguments to the handlers, how about providing
functions to get these, similar to how you can already get the 'less
common' fields of the event such as common_pc(), common_lock_depth(),
etc.  You call these during the event processing loop, giving them the
$context passed into the handler, and they in turn call perf internally
to get the data for the event.  This method can be used to get pretty
much anything out of perf internally.

So initially for callchains and callers we could have a couple functions
like:

get_callchain($context)

which would return an array of strings representing the callchain, and

get_caller($context)

which would just return the caller.

It would probably also make sense to have a raw version of at least
get_callchain() that would avoid the overhead of resolving the
callchains until a later time e.g. you should be able to use the raw
callchains as hash keys and resolve them only later when they get dumped
out in the report.  So maybe additionally something like:

get_raw_callchain($context)

which would return an array of u64 and a function to resolve a raw
callchain:

resolve_callchain($context, callchain[])

Or since in this case you need to keep the callchain and the $context
together, they could both just be contained in a callchain_object that
defines its own hash value and can be directly hashed or turned into the
string version by a callchain.tostring() method or something like that.

I don't know how soon I'll be able to implement this since I'm really
busy for the next several weeks, but if it makes sense, I can at least
try to do something for the first two (get_callchain() and
get_caller())...

Tom

> > Thanks.


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/