Re: [PATCH 2/2 -tip] perf_counter: parse-events.c introduce alias member in event_symbol

[Ingo Molnar <mingo@elte.hu>]

* Jaswinder Singh Rajput <jaswinder@kernel.org> wrote:

> On Tue, 2009-06-23 at 10:24 +0200, Ingo Molnar wrote:
> > * Jaswinder Singh Rajput <jaswinder@kernel.org> wrote:
> > 
> > > 
> > > perf test also need some command to execute otherwise it will also 
> > > show long list of <inactive>
> > 
> > I think what it should do is to execute test-cases _internally_. Not 
> > just execute some random command on the system and hope for events.
> 
> Can you suggest some good test cases where we can get numbers for 
> almost all the events on each and every run.

the testcases should be specific to the counter type. Some counters 
are easy (such as cycles or instructions - you in fact cannot even 
get zero out of them even if you tried) - some are harder - such as 
migrations, or cache-misses or page-faults.

Small functions that trigger them for sure would do the trick. For 
example, to provoke a minor page fault:

static void trigger_minor_page_fault(void)
{
	void *page;

	page = mmap(0, PAGE_SIZE, PROT_READ|PROT_WRITE, MAP_ANONYMOUS, -1, 0);

	/* Clearing the page will trigger a minor fault: */
	memset(page, 0, PAGE_SIZE);

	munmap(page, PAGE_SIZE);
}

These small test-functions should be attached to the event array as 
function pointers - hence the generic testing code would just do a 
->event.trigger_event() callback, it wouldnt need to know much about 
the event itself.

> > >  context-switches                                         7956
> > >  CPU-migrations                                              7
> > 
> > this needs to be provoked intentionally via sched_setaffinity(): 
> > first migrate to cpu0, then to cpu1.
> 
> There should be some option from user or we test it each time.

If this is builtin-test.c - then yes, testing it all the time would 
be a good default. (there might be other aspects of testing that 
would require options to this command in the future.)

	Ingo