On Sat, Oct 17, 2015 at 03:39:20PM -0400, Rob Clark wrote:

On Sat, Oct 17, 2015 at 2:59 PM, Greg Kroah-Hartman
[off-list ref] wrote:

quoted

On Sat, Oct 17, 2015 at 02:45:34PM -0400, Rob Clark wrote:

quoted

On Sat, Oct 17, 2015 at 2:27 PM, Greg Kroah-Hartman
[off-list ref] wrote:

quoted

On Sat, Oct 17, 2015 at 01:54:43PM -0400, Rob Clark wrote:

quoted

On Sat, Oct 17, 2015 at 12:56 PM, Greg Kroah-Hartman
[off-list ref] wrote:

quoted

quoted

I'm guessing the time is a matter of probing and undoing the probes
rather than slow h/w. We could maybe improve things by making sure
drivers move what they defer on to the beginning of probe, but that
seems like a horrible, fragile hack.

How can calling probe and failing cause 2 seconds?  How many different
probe calls are failing here?  Again, a boot log graph would be great to
see as it will show the root cause, not just guessing at this.

just fwiw, but when you have a driver that depends on several other
drivers (which in turn depend on other drivers and so on), the amount
of probe-defer we end up seeing is pretty comical.  Yeah, there
probably is some room to optimize by juggling around order drivers do
things in probe.  But that doesn't solve the fundamental problem with
the current state, about probe order having no clue about
dependencies..

I can imagine it is a lot of iterations, but how long does it really
take?  How many different devices are involved that it takes multiple
loops in order to finally work out the correct order?  Where is the time
delays here, just calling probe() and having it instantly return
shouldn't take all that long.

offhand, I think the dependencies go at *least* three levels deep..
I'd say, from memory, I see drm/msm taking at least 5 or 6 tries to
get all the way through requesting it's various different
regulators/clks/gpios.

And how long does that really take?  Numbers please :)

quoted

I hadn't really paid attention to how many
tries the drivers I depend on go through.  (Of those, I take clks from
two different clk drivers (which have dependency on a 3rd clk driver),
and regulators and gpio's come from at least two places, which in turn
have dependencies on clks, etc.)  I don't have really good hard
numbers handy (since my observations of this are w/ console over uart
which effects timings, and so I see it taking much longer than 2sec)..
but the 2sec figure that Tomeu mentioned seemed pretty plausible to
me.

I can try to get better #'s... I should have my kernel hat on at least
some of the time next week.. but the 2sec figure didn't seem
unrealistic to me.

Based on the time it takes a modern laptop to boot, 2 seconds is
forever, there has to be something else going on here other than just
calling probe() a bunch of times.  Please use the tools we have to
determine this before trying to change the driver core.

yes, I am aware of the tools.. although so far I spend most of my time
just trying to get things working in the first place ;-)

And that's where most people stop, if you want to make it fast, you have
to put in more effort, sorry.  Don't expect the driver core to work
around driver bugs for you.

All I was trying to point out was that Tomeu's figures didn't really
seem unrealistic.  I mean, given that the average SoC driver probably
depends on at least one clock and at least one regulator, having to
probe each driver at least twice seems plausible.  And that having a
noticeable effect on boot time doesn't seem surprising.  I'm not sure
that saying 'modern laptop can boot in 2sec' adds much to the
discussion since I don't think you have quite so much interdependency
between devices vs random probe order.  I have seen arm devices boot
to UI in similar times, but that was pre-devicetree days.

2 extra probes add a second to the boot time?  Those sound like really
broken drivers to me :)

thanks,

greg k-h