Re: [PATCH 0/7] de-stage SW_SYNC validation frawework

[Pavel Machek]

On Tue 2016-08-09 08:04:54, Daniel Vetter wrote:
> On Sun, Jul 24, 2016 at 05:00:31PM +0200, Pavel Machek wrote:
> > On Mon 2016-08-08 16:08:12, Gustavo Padovan wrote:
> > > 2016-08-07 Pavel Machek <pavel@xxxxxx>:
> > > 
> > > > On Sun 2016-07-24 15:21:11, Greg Kroah-Hartman wrote:
> > > > > On Mon, Jul 18, 2016 at 04:12:45PM -0300, Gustavo Padovan wrote:
> > > > > > Hi,
> > > > > > 
> > > > > > Do you think there is time to get this in for 4.8?
> > > > > 
> > > > > No, it was too late on my end, due to travel and vacation, sorry.  I'll
> > > > > queue it up for 4.9-rc1.
> > > > 
> > > > Could we get some documentation what this does? Is it visilble to
> > > > userspace?
> > > 
> > > This interface is only intended for testing and validation, there are
> > > ioctls on the debugfs file that can be accessed by userspace but there
> > > isn't any exported kernel header with this info. The tester should know
> > > and add a internal header to be able to access it. We want to prevent
> > > people from misusing this feature by not advertising it nor providing
> > > documentation.
> > 
> > You are playing dangerous game here. debugfs is not normally considered stable,
> > but otoh... ioctls on debugfs?
> 
> It's not considered stable. The idea is that we also add the existing
> testcases to kselftest. It's purely a bit of interface to be able to drive
> run the test logic for real fences. What it really tests is the fence
> interface (which is public in the uapi headers and all that), but to be
> able to do that we need some (hw-independent way) to expose fences, which
> this provides.
> 
> Long term we might even do this as a proper interface (with some
> restrictions to make it safe and avoid userspace pulling the kernel over
> the table). And then rip out sw_sync entirely.
> 
> Imo there's no need at all for docs for this.

There's full directory of files, with absolutely zero comments/documentation. They
are quite hard to understand. Plus it has userland interface.

IMO comment should be added, explaining what it is testing, that interface is not considered
stable, and where the test lives.

I know what "fence" is in the cpu sense (mfence and friends), but I'm not sure 
what "real fence" is in this context.

Best regards,
										Pavel

-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html