Re: [RFC][PATCH] perf: sysfs type id

[Kay Sievers]

On Tue, Nov 9, 2010 at 23:22, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> On Tue, 2010-11-09 at 23:11 +0100, Kay Sievers wrote:
>> On Tue, Nov 9, 2010 at 22:45, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>> > The below is a RFC patch adding dynamic type ids to perf.
>> >
>> > We need to represent PMUs in sysfs because we want to allow multiple
>> > (loadable) PMUs and need a way to identify them.
>> >
>> > This patch creates a new device class "pmu" and adds a single attribute
>> > "type" to it. This device attribute will expose the dynamic type id as
>> > required by perf_event_attr::type.
>> >
>> > The sysfs layout looks like:
>> >
>> > [root@westmere ~]# cd /sys/class/pmu/
>>
>> Please use a 'bus_type' instead of 'class'.
>>
>> I'm very sure, some day, you'll need global attributes for the pmu
>> stuff, and class -- unlike bus -- has its own subdir where you can go
>> wild, without mixing things with the list-of-devices. :)
>
> Having its own subdir sounds like a pro, so I'm somewhat confused. Also
> the pmu (or event_source as Ingo would like it getting called) seems
> like a class of devices not a bus.
>
> USB/PCI/I2C/ISA are all buses.. pmu/event_source not so much, they are
> very different (sometimes even pure software) things that provide a
> common interface -- they generate events which we can count and sample.
>
> /me dazed & confused, please do explain.

Don't get confused by the 'bus' name, it's used for many things which
are not hardware buses. Class is just almost the same as a bus inside
and outside the kernel. Things like libudev don't even allow you to
distinguish the both, there is only a 'subsystem'. The class export in
/sys is not extensible at all, and therefore it should be avoided for
new stuff.

We will make all classes and buses show up in /sys/subsystem/. The
current silly split between a class and a bus has absolutely no
consistent meaning, and serves no purpose. After /sys/subsystem/
exists, class and bus will only be compat links pointing to the
entries in /sys/subsystem/.

There is _one_ tree at /sys/devices/, and will be _one_ classification
at /sys/subsystem/ which lists all-devices-of-the-same-subsystem and
has an extensible subdir like bus has today.

It's partly described in some outdated version in Documentation/sysfs-rules.txt.

>> No new stuff should use 'class', it's not extensible.
>
> Hrm.. when I introduced the bdi stuff class was _the_ thing to use.

Yeah, it does really matter for very simple things, but never for
stuff that reads -- before it even exists -- as "we will add links to
... and subbdirs, and have ...' :)

Kay
--
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/