Re: [PATCH 1/7] drm/bridge: Support hotplugging panel-bridge.

[Boris Brezillon <boris.brezillon@free-electrons.com>]

On Fri, 23 Jun 2017 09:22:15 +0200
Andrzej Hajda <a.hajda@samsung.com> wrote:

> On 22.06.2017 15:34, Boris Brezillon wrote:
> > On Thu, 22 Jun 2017 15:16:47 +0200
> > Andrzej Hajda <a.hajda@samsung.com> wrote:
> >  
> >> On 22.06.2017 14:41, Boris Brezillon wrote:  
> >>> On Thu, 22 Jun 2017 14:29:07 +0200
> >>> Andrzej Hajda <a.hajda@samsung.com> wrote:
> >>>    
> >>>> On 22.06.2017 11:23, Boris Brezillon wrote:    
> >>>>> On Thu, 22 Jun 2017 13:47:43 +0530
> >>>>> Archit Taneja <architt@codeaurora.org> wrote:
> >>>>>      
> >>>>>> On 06/22/2017 01:20 PM, Benjamin Gaignard wrote:      
> >>>>>>> 2017-06-20 19:31 GMT+02:00 Eric Anholt <eric@anholt.net>:        
> >>>>>>>> Archit Taneja <architt@codeaurora.org> writes:
> >>>>>>>>        
> >>>>>>>>> On 06/16/2017 08:13 PM, Eric Anholt wrote:        
> >>>>>>>>>> Archit Taneja <architt@codeaurora.org> writes:
> >>>>>>>>>>        
> >>>>>>>>>>> On 06/16/2017 02:11 AM, Eric Anholt wrote:        
> >>>>>>>>>>>> If the panel-bridge is being set up after the drm_mode_config_reset(),
> >>>>>>>>>>>> then the connector's state would never get initialized, and we'd
> >>>>>>>>>>>> dereference the NULL in the hotplug path.  We also need to register
> >>>>>>>>>>>> the connector, so that userspace can get at it.
> >>>>>>>>>>>>        
> >>>>>>>>>>> Shouldn't the KMS driver make sure the panel-bridge is set up before
> >>>>>>>>>>> drm_mode_config_reset? Is it the case when we're inserting the
> >>>>>>>>>>> panel-bridge driver as a module?
> >>>>>>>>>>>
> >>>>>>>>>>>
> >>>>>>>>>>> All the connectors that have been added are registered automatically
> >>>>>>>>>>> when drm_dev_register() is called by the KMS driver. Registering a
> >>>>>>>>>>> connector in the middle of setting up our driver is prone to race
> >>>>>>>>>>> conditions if the userspace decides to use them immediately.        
> >>>>>>>>>> Yeah, this is fixing initializing panel_bridge at DSI host_attach time,
> >>>>>>>>>> which in the case of a panel module that creates the DSI device
> >>>>>>>>>> (adv7533-style, like you said I should use as a reference) will be after
> >>>>>>>>>> drm_mode_config_reset() and drm_dev_register().        
> >>>>>>>>> Okay. In the case of the msm kms driver, we defer probe until the
> >>>>>>>>> adv7533 module is inserted, only then we proceed to drm_mode_config_reset()
> >>>>>>>>> and drm_dev_register(). I assumed this was the general practice followed by
> >>>>>>>>> most kms drivers. I.,e the kms driver defers probe until all connector
> >>>>>>>>> related modules are inserted, and only then proceed to create a drm device.        
> >>>>>>>> The problem, though, is the panel driver needs the MIPI DSI host to
> >>>>>>>> exist to call mipi_dsi_device_register_full() during the probe process.
> >>>>>>>> The adv7533 driver gets around this by registering the DSI device in the
> >>>>>>>> bridge attach step, but drm_panel doesn't have an attach step.        
> >>>>>> I'm not sure how we can get around this. We had discussion about this on irc
> >>>>>> recently, but couldn't come up with a good conclusion. We could come up with a
> >>>>>> panel_attach() callback to make it similar to bridges, but that's just us avoiding
> >>>>>> the real issue.      
> >>>>> How about making DSI dev registration fully asynchronous, that is, DSI
> >>>>> devs declared in the DT under the DSI host node will be
> >>>>> registered/attached at probe time, and devs using another control bus
> >>>>> (like the adv7533 controller over i2c) will be registered afterwards.
> >>>>>
> >>>>> That implies moving the drm_brige registration logic outside of the DSI
> >>>>> host ->probe() path. The idea would be to check if all devs connected
> >>>>> to the DSI bus are ready at dsi_host->attach() time. If they are, we
> >>>>> can finally register the XXX -> DSI bridge. If they're not (because
> >>>>> some devs connected to the DSI bus have not been probed yet), then we
> >>>>> do not register the drm_bridge and wait for the next dsi_host->attach()
> >>>>> event.      
> >>>> I guess you assumes that dsi-host knows all devs connected to it, thanks to:
> >>>> - subnodes of the host - ie. devices controlled via dsi bus,
> >>>> - graph links from host ports/endpoints - ie. devices controlled by
> >>>> other buses, for example adv7533.    
> >>> Yep, but I think that's already a requirement when populating devices
> >>> with the OF graph method (if one of the DSI output endpoint does not
> >>> have a drm_bridge/panel attached to it, the DSI host driver returns
> >>> -EPROBE_DEFER).
> >>>    
> >>>> I would separate both abstractions to make it more clear:
> >>>> 1. MIPI bus should be registered early - to allow create/bind devices on it,    
> >>> Exactly.
> >>>    
> >>>> 2. drm_bridge should be registered only if all required sinks
> >>>> (bridges/panels) are registered.    
> >>> That's true, until we find a solution to support add DRM bridge hotplug.
> >>>    
> >>>> First point seems OK, I am not sure about the 2nd one - if used
> >>>> consistently, it would require building pipeline from sink to source.    
> >>> Yes.    
> >> Since drm_bridge_attach requires encoder to be not null pipeline
> >> creation would be painful:
> >> 1. Every driver must call drm_of_find_panel_or_bridge on sink(s) before
> >> registering bridge and cache the result for later use.  
> > Shouldn't be hard to do since dsi_host->attach() is called each time a
> > sink is added (and ready to use). All you need to do is retrieve the
> > bridge pointer and put it in a list embedded in the DSI host priv
> > struct. Once you have all sinks added to this list (can be checked by
> > counting the number of endpoints and DSI devs at probe time), you can
> > register the DSI bridge and wait for someone to call ->attach() on it.
> >
> > In the ->attach() hook of the DSI bridge, you'll have to attach all
> > sinks stored in the list to the DSI bridge. Note that right now you have
> > a 1:1 relationship, which prevents you from having one DSI bridge that
> > can attach to different bridges available on the DSI bus (e.g. DSI ->
> > HDMI bridge on channel 0 and DSI -> LVDS bridge on channel 1).
> >  
> >> 2. After encoder finds directly connected bridge, it can attach it.  
> > I don't get that one.  
> 
> If you have pipeline:
> 
> crtc -> encoder -> bridge1 -> bridge2 -> panel
> 
> encoder knows only about bridge1, and must wait till it is registered,
> before attaching it,
> and assuming bridge must wait for its sinks before registration the
> whole pipeline construction will look like:
> 
> 0. encoder waits for bridge1, bridge1 waits for bridge2, bridge2 waits
> for panel, usually by deferring.
> 1. panel is registered.
> 2. bridge2 finds panel and is registered.
> 3. bridge1 finds bridge2 and is registered.
> 4. encoder finds bridge1 and attach it to encoder,
> 4a. bridge1->attach() attach bridge2 to encoder after bridge1
> 4b. bridge2->attach() attach panel to bridge2
> 
> This is why it seems for me quite complicated.

But that's already what happens in most drivers today (probably because
things were designed before connector hotplug was supported). I agree,
it's far from ideal, but until we get full-hotplug support, we'll have
to rely on such hacks.