[Intel-gfx] [PATCH 0/5] sanitize hda/i915 interface using the component fw

[Imre Deak]

On Tue, 2014-12-09 at 11:03 +0100, Daniel Vetter wrote:
> On Tue, Dec 09, 2014 at 10:59:54AM +0200, Imre Deak wrote:
> > On Mon, 2014-12-08 at 21:14 +0100, Daniel Vetter wrote:
> > > On Mon, Dec 08, 2014 at 06:42:04PM +0200, Imre Deak wrote:
> > > > The current hda/i915 interface to enable/disable power wells and query
> > > > the CD clock rate is based on looking up the relevant i915 module
> > > > symbols from the hda driver. By using the component framework we can get
> > > > rid of some global state tracking in the i915 driver and pave the way to
> > > > fully decouple the two drivers: once support is added to enable/disable
> > > > the HDMI functionality dynamically in the hda driver, it can bind/unbind
> > > > itself from the i915 component master, without the need to keep a
> > > > reference on the i915 module.
> > > > 
> > > > This also gets rid of the problem that currently the i915 driver exposes
> > > > the interface only on HSW and BDW, while it's also needed at least on
> > > > VLV/CHV.
> > > 
> > > Awesome that you're tackling this, really happy to see these hacks go.
> > > Unfortunately I think it's upside down: hda should be the component master
> > > and i915 should only register a component.
> > > 
> > > Longer story: The main reason for the component helpers is to be able to
> > > magically delay registering the main/master driver until all the
> > > components are loaded. Otherwise -EDEFER doesn't work and also the
> > > suspend/resume ordering this should result in. Master here means whatever
> > > userspace eventually sees as a device node or similar, component is
> > > anything really that this userspace interfaces needs to function
> > > correctly.
> > 
> > EDEFER doesn't solve the suspend/resume ordering, at least not in the
> > general async s/r case. In any case I don't see a problem in making the
> > hda component the master and I think it's more logical that way as you
> > said, so I changed that.
> 
> Yeah for full async s/r we're screwed as-is. But see below for some crazy
> ideas.
> > > I think what we need here is then:
> > > - Put the current azx_probe into the master_bind callback. The new
> > >   azx_probe will do nothing else than register the master component and
> > >   the component match list. To do so it checks the chip flag and if it
> > >   needs to cooperate with i915 it registers one component match for that.
> > >   The master_bind (old probe) function calls component_bind_all with the
> > >   hda_intel pointer as void *data parameter.
> > 
> > I'm not sure this is the best way since by this the i915 module would
> > need to be pinned even when no HDMI functionality is used. I think a
> > better way would be to let the probe function run as now and
> > init/register all the non-HDMI functionality. Then only the HDMI
> > functionality would be inited/registered from the bind/unbind hooks.
> 
> Hm, I wasn't sure whether alsa allows this and thought we need i915 anyway
> to be able to load the driver. But if we can defer just the hdmi part.
> 
> > But we could go either way even as a follow-up to this patchset.
> > 
> > > - i915 registers a component for the i915 gfx device. It uses the
> > >   component device to get at i915 sturctures and fills the dev+ops into
> > >   the hda_intel pointer it gets as void *data.
> > >
> > > Stuff we then should do on top:
> > > - Add deferred probing to azx_probe: Only when all components are there
> > >   should it actually register. This will take care of all the module load
> > >   order mess.
> > 
> > I agree that we should only register user interfaces when everything is
> > in place. But I'm not sure deferred probe is the best, we could do
> > without it by registering HDMI from the component bind hook.
> 
> It's mostly a question whether alsa does support delayed addition of
> interface parts. DRM most definitely does not (except for recently added
> dp mst connector hotplug). But I guess if the current driver already
> delays registering the hdmi part then we're fine. I'm not sure about
> whether it's really safe - I spotted not a lot of locking really to make
> sure there's no races between i915 loading and userspace trying to access
> the hdmi side.

Ok, I'm not sure either. If that's not possible, I agree EDEFER would be
the best and with that we could also get rid of the request_module() we
need now.

> > >   It should also take care of system suspend/resume ordering and we
> > >   should be able to delete all the early_resume/late_suspend trickery.
> > 
> > Deferred probe doesn't solve the suspend/resume ordering, we would need
> > to have a separate HDMI device that is set as a child to the i915
> > device. Alternatively we could use device_pm_wait_for_dev().
> 
> I'm not sure whether there's the possibility of deadlocks with
> device_pm_wait_for_dev. But if that works I think a component helper to
> wait for all components to resume would be really useful. Or we implement
> the full-blown pm_ops idea laid out below for components.

Yes it could deadlock with pm_async=0 (a debug thing nowadays?) and
depending on the bus address of the waiter and wait-for device. At least
for us this can't happen (with the fixed PCI addresses for the gfx and
audio devices), but I don't know if you can consider this a clean
solution. Perhaps by adding a check for this in device_pm_wait_for_dev()
and avoiding the deadlock by returning some error would be safe enough.

> > 
> > > Imo we should have things ready up to this point to make sure this
> > > refactoring actually works.
> > 
> > I think we could go with this minimal patch with your change to have the
> > hda component be master. This only adds the support for components and
> > keeps everything else the same. We could consider the bigger changes as
> > a follow-up.
> 
> Yeah that was my plan - if EDEFER isn't enough then we keep the early/late
> resume/suspend hooks for a bit longer.
> 
> > > Then there's some cool stuff we could do on top:
> > > - Register a i915-hda platform devices as a child of the gfx pci device.
> > >   Besides shuffling around a bit with the interfaces/argument casting and
> > >   the component match function this doesn't really have a functional
> > >   impact. But it makes the relationship more clear since hda doesn't
> > >   really need the entire pci device, but only the small part that does
> > >   audio.
> > > 
> > > - Replace the hand-rolled power-well interface with runtime pm on that
> > >   device node.
> > > 
> > > - If system suspend/resume doesn't work automatically with deferred
> > >   probing (tbh I'm not sure) add pm_ops to the component master. Then add
> > >   some functions as default implementations for pm_ops for components
> > >   which simply refcount all component pm_ops calls and call the master
> > >   pm_ops suspend on the first suspend call and resume on the last resume
> > >   call. That really should take care of suspend/resume ordering for good.
> > 
> > Yep, these sound good. I think having an HDMI child device is the
> > cleanest solution for the s/r ordering issue.
> 
> Ok, sounds like we have a plan. And thanks again for tackling this, I'm
> really happy to see this go away.
> 
> Cheers, Daniel