Re: [PATCH v2 00/14] perf, persistent: Kernel updates for perf toolintegration

From: Robert Richter
Date: Tue Jun 25 2013 - 12:56:49 EST


On 24.06.13 12:22:00, Peter Zijlstra wrote:
> On Tue, Jun 11, 2013 at 06:42:26PM +0200, Robert Richter wrote:
> > Note that perf tools need to support the 'attr<num>' syntax that is
> > added in a separate patch set. With it we are able to run perf tool
> > commands to read persistent events, e.g.:
>
> where is this patch? I can't find it.

I posted it with that subject:

[PATCH 0/4] perf tools: Persistent events, changes for perf tool integration

> I also find attr<num>:<bit> a bit weird. So far we've used
> <perf_event_attr::fieldname>:<bits>, so while initializing anonymous
> unions is a bit difficult with 'older' GCCs and we cannot actually do:
>
> struct perf_event_attr {
> ...
> union {
> u64 flags;
> u64 disabled : 1,
> ...
> __reserved_1 : X;
> }
> ...
> };
>
> We could fake it in userspace by allowing things like: flags:23. It
> would not be a much worse hack than attr<num>:<bit> I suppose.

A goal is to be able to entirely describe the complete attr structure
with sysfs. perf tools should work out-of-the-box for every kind of
event. For this particular case a 'flags' syntax would be ok, but not
if you like to fill in other attr members.

With attr<num>:<bit> you are flexible to put in *everything* into
attr. And for better readability there are abstractions as formats
like:

/sys/bus/event_source/devices/persistent/format/persistent:attr5:23

Not sure if the event parser handles recursion already, but we could
also support the 'flag' format you describe that way:

/sys/bus/event_source/devices/persistent/format/flags:attr5:0-63
/sys/bus/event_source/devices/persistent/format/persistent:flags:23

We could even list events for other pmus with small changes in the
kernel, e.g:

/sys/bus/event_source/devices/persistent/format/type:attr0:0-31
/sys/bus/event_source/devices/persistent/events/some_tracepoint:type=2,config=<tp_id>

The example is a bit weird. But in general we can translate struct
perf_event_attr to a string and vice versa. And perf tools understand
how to set up every bit of the event attr structure.

-Robert
--
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/