Message-ID: <20100831122343.GB27532@console-pimps.org>
Date:	Tue, 31 Aug 2010 13:23:43 +0100
From:	Matt Fleming <matt@...sole-pimps.org>
To:	Robert Richter <robert.richter@....com>
Cc:	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	Will Deacon <will.deacon@....com>,
	Paul Mundt <lethal@...ux-sh.org>,
	Russell King <linux@....linux.org.uk>,
	"linux-arm-kernel@...ts.infradead.org" 
	<linux-arm-kernel@...ts.infradead.org>,
	"linux-sh@...r.kernel.org" <linux-sh@...r.kernel.org>,
	Peter Zijlstra <peterz@...radead.org>,
	Ingo Molnar <mingo@...e.hu>,
	Frederic Weisbecker <fweisbec@...il.com>,
	Arnaldo Carvalho de Melo <acme@...hat.com>,
	"linux-arch@...r.kernel.org" <linux-arch@...r.kernel.org>
Subject: Re: [PATCH V2 4/4] sh: Use the perf-events backend for oprofile

On Tue, Aug 31, 2010 at 01:28:41PM +0200, Robert Richter wrote:
> On 27.08.10 16:19:46, Matt Fleming wrote:
> > On Fri, Aug 27, 2010 at 04:59:01PM +0200, Robert Richter wrote:
> > > On 26.08.10 15:09:19, Matt Fleming wrote:
> > > > Use the perf-events based wrapper for oprofile available in
> > > > drivers/oprofile. This allows us to centralise the code to control
> > > > performance counters.
> > > > 
> > > > Signed-off-by: Matt Fleming <matt@...sole-pimps.org>
> > > > ---
> > > > 
> > > > Paul,
> > > > 
> > > > I dropped the CONFIG_PERF_EVENTS dependency from the Makefile in this
> > > > version because to do anything useful we need perf events anyway.
> > > 
> > > Initialization should simply fail with a printk message for this case,
> > > implement function stubs for the !CONFIG_PERF_EVENTS case instead in
> > > the oprofile.h header file.
> > 
> > I didn't do this because I was hoping that eventually we'd make
> > CONFIG_OPROFILE select PERF_EVENTS. Would you be OK making that change
> > instead? Runtime failure is best avoided where possible, especially when
> > we can sort this out at compile time.
> 
> Ok, we don't need it if we add architectural dependencies to Kconfig
> for those architectures requiring perf.

OK cool.

> > > > -static int op_sh_start(void)
> > > > +static char *op_name_from_perf_name(const char *name)
> > > >  {
> > > > -	/* Enable performance monitoring for all counters.  */
> > > > -	on_each_cpu(model->cpu_start, NULL, 1);
> > > > +	if (!strcmp(name, "SH-4A"))
> > > > +		return "sh/sh4a";
> > > > +	if (!strcmp(name, "SH7750"))
> > > > +		return "sh/sh7750";
> > > 
> > > With that implementation we always have to touch the code for new
> > > cpus. Maybe we derive it from the perf name, e.g. making all lowercase
> > > and removing dashes?
> > 
> > Is this code really that bad that we need to start playing string
> > manipulation games?
> 
> No, but with that implementation we always have to update the cpu
> string with each new cpu though nothing else changes. We may keep this
> code. But, shouldn't we return a default string "sh/<name>" for all
> other cases? We will then need to update only the oprofile userland
> with new cpus.

These names are actually the names of types of performance counters,
not a specific cpu. All SH-4 cpus that have performance counters have
7750-style performance counters and all SH-4A cpus have SH-4A-style
counters.

It's unlikely we'd have to update this code in the near future. Paul,
correct me if I'm wrong here.

> > > > +	ops->setup		= oprofile_perf_setup;
> > > > +	ops->create_files	= oprofile_perf_create_files;
> > > > +	ops->start		= oprofile_perf_start;
> > > > +	ops->stop		= oprofile_perf_stop;
> > > > +	ops->cpu_type		= op_name_from_perf_name(sh_pmu_name());
> > > >  
> > > > -	model = lmodel;
> > > > +	oprofile_perf_set_num_counters(sh_pmu_num_events());
> > > >  
> > > > -	ops->setup		= op_sh_setup;
> > > > -	ops->create_files	= op_sh_create_files;
> > > > -	ops->start		= op_sh_start;
> > > > -	ops->stop		= op_sh_stop;
> > > > -	ops->cpu_type		= lmodel->cpu_type;
> > > > +	ret = oprofile_perf_init();
> > > 
> > > Instead of exporting all the functions above implement something like:
> > > 
> > > 	name = op_name_from_perf_name(sh_pmu_name());
> > > 	num_events = sh_pmu_num_events();
> > > 	ret = oprofile_perf_init(ops, name, num_events);
> > > 
> > > We will then have only oprofile_perf_init() and oprofile_perf_exit()
> > > as interface which is much cleaner.
> > 
> > Well, the reason that I left it this way is so that architectures can
> > choose to implement wrappers around the oprofile_perf_* functions. I
> > don't think ARM or SH actually need wrappers (the only extra thing that
> > ARM does is locking which SH should probably do too) but I assumed there
> > was a reason that these functions pointers were exposed originally. I
> > haven't look at what other architectures would do. I'll take a look at
> > that.
> 
> I am not sure if we need such wrappers, and if so we could implement
> it anyway, e.g.:
> 
>  oprofile_perf_init(perf_ops, name, num_events);
> 
>  op_sh_setup():
> 
> 	/* setup something */
> 	...
> 
> 	perf_ops->setup();
> 
> 	/* setup more */
> 	...
> 
> But I don't think we need this. And the above makes the interface much
> cleaner.

OK, seeing as the two architectures that will use this initially don't
require wrappers I've no problem doing it your way. It can always be
extended later if necessary. And more importantly, with a proper
usecase we'll be able to see exactly _how_ it needs to be extended.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists