Re: [RFCv3 0/9] perf tool: parser generator for events parsing

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Tue, 2012-01-24 at 16:53 -0800, Greg KH wrote:

> Well, what's to keep someone from exploding one of those files to go
> over the buffer size without knowing it?

The content is known at compile time, we could put checks in. But yeah
4k isn't that much text..

> Even after reading the above link, I can't really understand what this
> is being used for.  As it's sysfs files, why aren't Documentation/ABI/
> files also being created with the patch explaining it all?

Ah, yeah, documentation, wasn't that written in C :-)

But fair enough... Anyway, the purpose is to describe the magic bits
that go into perf_event_attr::config[012] for a particular pmu.

Basically they're the 'hardware' bitmasks we want to export, so that
userspace can deal with sane name/value pairs.

> Again, if at all possible, sysfs should be one value per file.  Please
> NEVER create a sysfs file that requires a parser to determine what is
> going on in it.  It should be a simple 'read the value' type thing.

The whole purpose was to drive a parser :-) Anyway, is:
"config:0-7,32-35" acceptable for a single file?

This means, the 'config' member of struct perf_event_attr bits 0-7,32-35
form a bitfield whose name is then given by the filename that has this
content.

> So yes, multiple sysfs files do make sense, the resource load should be
> almost non-existant for new ones.

Surely all these attribute objects take more space than a few lines of
text, but ok, I guess multiple files it is.