On Tue, Nov 18, 2014 at 04:16:04PM +0100, Thierry Reding wrote:

On Tue, Nov 18, 2014 at 01:44:44PM +0100, Hans de Goede wrote:

quoted

Hi,

On 11/18/2014 12:46 PM, Hans de Goede wrote:

quoted

Hi,

On 11/18/2014 12:21 PM, Thierry Reding wrote:

quoted

On Tue, Nov 18, 2014 at 12:01:08PM +0100, Hans de Goede wrote:

quoted

Hi,

On 11/18/2014 11:19 AM, Thierry Reding wrote:

quoted

On Thu, Nov 13, 2014 at 09:58:41AM +0100, Geert Uytterhoeven wrote:

quoted

Hi Maxime,

On Thu, Nov 13, 2014 at 9:52 AM, Maxime Ripard
[off-list ref] wrote:

quoted

quoted

-module_init(simplefb_init);
+/*
+ * While this can be a module, if builtin it's most likely the console
+ * So let's leave module_exit but move module_init to an earlier place
+ */

Not really related to this patch itself, but do we want to support
simplefb as a module? It seems like it's going to be most of the time
broken.

If it depends on clocks, it won't work as a module, as CCF will have disabled
all unused clocks at that point.

If it does depend on anything beyond clocks it won't work at all. Clocks
are special because they get set up very early at boot time. If it turns
out that a simplefb ever needs a regulator to remain on, and that's even
quite likely to happen eventually, it's going to fail miserably, because
those regulators will typically be provided by a PMIC on an I2C bus. The
regulator won't be registered until very late into the boot process and
a regulator_get() call will almost certainly cause the simplefb driver
to defer probing.

Right, this has been discussed already and the plan is to have simplefb
continue its probe function and return success from it if it encounters
any -eprobe_defer errors, while tracking which resources it misses.

And then have a late_initcall which will claim any resources which failed
with -eprobe beforehand.

How do you ensure that the late_initcall gets run before any of the
other late_initcalls that disable the resources?

quoted

Also my recollection is
that deferred probing will first be triggered the first time from a
late_initcall, so chances aren't very high that all resources have shown
up by that time.

So I just looked up the relevant code, and your right, this means that
the whole model of "disable unused resources once probing is done" which
we use for e.g. clocks, is already somewhat broken since there is
no guarantee probing is really done when the cleanup code runs.

quoted

quoted

quoted

Now deferring probing is a real showstopper for simplefb, because not
only does it make the framebuffer useless as early boot console, once
probing is attempted again the clocks that it would have needed to
acquire to keep going will already have been switched off, too.

That is not true, even with the current implementation, if all necessary
drivers are built in, then simplefb will come up later, but it will still
come up before the late_initcall which disables the clocks.

Yes, in the current implementation because clocks typically are
registered very early and thus you don't hit the deferred probe. The
same is not true for other types of resources where it's actually quite
common to hit deferred probing (regulators is a very notorious one).

It doesn't matter whether a driver is built-in or not, once you hit
deferred probing you lose.

quoted

Once we do the split probing described above (which is something which
we plan to do when it becomes necessary), then simplefb will still come
up early.

It will come up early but won't have acquired all the resources that it
needs, so unless you somehow manage to order late_initcalls in exactly
the way that you need them, the frameworks will still turn off what you
haven't managed to claim yet.

If it is a resource which only shows up as a result of deferred probing,
then it may very well not have been probed & registered yet, when the
framework cleanup functions runs, and thus will not get turned off...

So yes you're right that deferred probing may cause issues, but it seems
that this is not something simplefb specific, but rather a generic problem
with deferred-probing vs subsys cleanup functions.

My view on this is simple, lets worry about this when we actually have
a board which hits these issues, and then we'll see from there.

So thinking more about this, I think this is not that hard to fix.

First lets fix the generic conflict between eprobedefer and subsys cleanup
functions. This can be done by:

1) Having a linked list of subsys cleanup functions
to call in drivers/base/dd.c

2) Have subsystems register their cleanup function rather then using
late_initcall to get it called

3) Have deferred_probe_work_func iterate over the list and call the cleanup
functions once deferred_probe_active_list goes empty. It should also remove
them once called, so that they only get called the first time
deferred_probe_active_list goes empty (iow when the deferred probing of'
build-in drivers is done).

This way cleanup functions actually get run when all probing of (buildin)
drivers is done.

That sounds like a reasonable solution to me.

quoted

Then when we move simplefb to a 2 fase probe, we can simply register the
framebuffer at the first probe call, and claim any resources we get, but
return -eprobe_defer if some resources returned that themselves.

Then when simplefb_probe gets re-called, it can first check if it did not
already register a fb, and if it did, it can use that and see which
resources are missing, and only try to claim those. If some resources
still return probe_defer, return eprobe_defer again, otherwise success.

I'm not sure you can check that the fb was already registered. When the
driver probe fails the driver core will pretty much remove any trace of
the driver, so checking for an already-reagistered framebuffer isn't
going to be easy.

But if you can manage to make that work it should give us a pretty good
solution that should be easy to make use of by other drivers as well.

What about handling priorities in the suggested cleanup mechanism, and
just add the second phase as the highest priority cleanup callback?

That way, we would be guaranteed to be run just before any cleanup
function.

Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: Digital signature
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20141118/40726180/attachment.sig>