Re: [RFC][PATCH 2/6] perf, arch: Rework perf_event_index()

[Will Deacon]

On Tue, Nov 22, 2011 at 11:26:20AM +0000, Peter Zijlstra wrote:
> On Mon, 2011-11-21 at 22:43 +0000, Will Deacon wrote:
> > Perhaps we could disable it while per-cpu events are running, although I
> > think this will probably just lead to SIGILL central for anybody trying to
> > use the counters in userspace.
> 
> One possibility would be to do as I did in patch 4, except ARM has it
> disabled by default and the folks who think they know WTF they're doing
> can enable it or so.

The problem is that everybody thinks they know WTF they're doing!

> Ostrikov mentioned on #kernelnewbies he wanted to have this enabled
> because apparently games use it. He mentioned toggling the user access
> on/off depending on if the kernel was using perf or not, but that would
> create very unreliable service.

Well we already have a reserve/release PMU thing which perf honours so we
could conceivably do this. I still reckon this will just lead to SIGILLs in
userspace though because we can't sanely notify tasks that they should leave
the PMU alone for a bit.

> Best would be to use perf to program the thing and use the userspace
> read and simply agree not to write to these registers (and pray people
> don't)..

But you know that the first thing people will do is zero the registers.

> Also, for those ARMs that do have a user readable clock, you could
> support the new time_{mult,shift,offset} from patch 5.

The user-readable clock will first appear in Cortex-A15, so the code for
that still needs to hit mainline before I can look at doing this in perf.

Will
--
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/