Re: [PATCH 1/6] drm: Add writeback connector type

From: Brian Starkey
Date: Wed Apr 19 2017 - 05:51:39 EST

On Tue, Apr 18, 2017 at 09:57:17PM +0200, Boris Brezillon wrote:
Hi Brian,

On Tue, 18 Apr 2017 18:34:43 +0100
Brian Starkey <brian.starkey@xxxxxxx> wrote:

>> @@ -214,6 +214,19 @@ struct drm_connector_state {
>> struct drm_encoder *best_encoder;
>>
>> struct drm_atomic_state *state;
>> +
>> + /**
>> + * @writeback_job: Writeback job for writeback connectors
>> + *
>> + * Holds the framebuffer for a writeback connector. As the writeback
>> + * completion may be asynchronous to the normal commit cycle, the
>> + * writeback job lifetime is managed separately from the normal atomic
>> + * state by this object.
>> + *
>> + * See also: drm_writeback_queue_job() and
>> + * drm_writeback_signal_completion()
>> + */
>> + struct drm_writeback_job *writeback_job;
>
>Maybe I'm wrong, but is feels weird to have the writeback_job field
>directly embedded in drm_connector_state, while drm_writeback_connector
>inherits from drm_connector.
>
>IMO, either you decide to directly put the drm_writeback_connector's
>job_xxx fields in drm_connector and keep the drm_connector_state as is,
>or you create a drm_writeback_connector_state which inherits from
>drm_connector_state and embeds the writeback_job field.

I did spend a decent amount of time looking at tracking the writeback
state along with the normal connector state. I couldn't come up with
anything I liked.

As the comment mentions, one of the problems is that you have to make
sure the relevant parts of the connector_state stay around until the
writeback is finished. That means you've got to block before
"swap_state()" until the previous writeback is done, and that
effectively limits your frame rate to refresh/2.

The Mali-DP HW doesn't have that limitation - we can queue up a new
commit while the current writeback is ongoing. For that reason I
didn't want to impose such a limitation in the framework.

In v1 I allowed that by making the Mali-DP driver hold its own
references to the relevant bits of the state for as long as it needed
them. In v3 I moved most of that code back to the core (in part
because Gustavo didn't like me signalling the DRM-"owned" fence from
my driver code directly). I think the new approach of "queue_job()"
and "signal_job()" reduces the amount of tricky code in drivers, and
is generally more clear (also familiar, when compared to vsync
events).

I'm certain there's other ways to do it (refcount atomic states?), but
it seemed like a biggish overhaul to achieve what would basically be
the same thing.

I was expecting each driver supporting writeback to have its own
different requirements around writeback lifetime/duration. For example
I think VC4 specifically came up, in that its writeback could take
several frames, whereas on Mali-DP we either finish within the frame
or we fail.

Letting the driver manage its writeback_job lifetime seemed like a
reasonable way to handle all that, with the documentation stating the
only behaviour which is guaranteed to work on all drivers:

* Userspace should wait for this fence to signal before making another
* commit affecting any of the same CRTCs, Planes or Connectors.
* **Failure to do so will result in undefined behaviour.**
* For this reason it is strongly recommended that all userspace
* applications making use of writeback connectors *always* retrieve an
* out-fence for the commit and use it appropriately.



... so all of that is why the _job fields don't live in a *_state
structure directly, and instead have to live in the separately-managed
structure pointed to by ->writeback_job.

Now, I did look at creating drm_writeback_connector_state, but as it
would only be holding the job pointer (see above) it didn't seem worth
scattering around the

if (conn_state->connector->connector_type ==
DRM_MODE_CONNECTOR_WRITEBACK)

checks everywhere before up-casting - {clear,reset,duplicate}_state(),
prepare_signalling(), complete_signalling(), etc. It just touched a
lot of code for the sake of an extra pointer field in each connector
state.

I can easily revisit that part if you like.

I think there's a misunderstanding. I was just suggesting to be
consistent in the inheritance vs 'one object to handle everything'
approach.

doh.. right yeah I misread. Sorry for the tangent. :-)


I'm perfectly fine with embedding the writeback_job pointer directly in
drm_connector_state, but then it would IMO make more sense to do the
same for the drm_connector object (embed drm_writeback_connector fields
into drm_connector instead of making drm_writeback_connector inherit
from drm_connector).


I agree that it's inconsistent. I guess I did it out of pragmatism -
there's quite a lot of new fields in drm_writeback_connector, and the
code needed to support it was comparatively small. On the other hand
there's only one additional field in drm_connector_state and the code
required to subclass it looked comparatively large.

Anyway, that's just a detail.


>
>Anyway, wait for Daniel's feedback before doing this change.
>

Am I expecting some more feedback from Daniel?

No, I was just saying that before doing the changes I was suggesting,
you should wait for Daniel's opinion, because I might be wrong.


>> };
>>
>> /**
>> diff --git a/include/drm/drm_mode_config.h b/include/drm/drm_mode_config.h
>> index bf9991b2..3d3d07f 100644
>> --- a/include/drm/drm_mode_config.h
>> +++ b/include/drm/drm_mode_config.h
>> @@ -634,6 +634,20 @@ struct drm_mode_config {
>> */
>> struct drm_property *suggested_y_property;
>>
>> + /**
>> + * @writeback_fb_id_property: Property for writeback connectors, storing
>> + * the ID of the output framebuffer.
>> + * See also: drm_writeback_connector_init()
>> + */
>> + struct drm_property *writeback_fb_id_property;
>> + /**
>> + * @writeback_pixel_formats_property: Property for writeback connectors,
>> + * storing an array of the supported pixel formats for the writeback
>> + * engine (read-only).
>> + * See also: drm_writeback_connector_init()
>> + */
>> + struct drm_property *writeback_pixel_formats_property;
>> +
>> /* dumb ioctl parameters */
>> uint32_t preferred_depth, prefer_shadow;
>>
>> diff --git a/include/drm/drm_writeback.h b/include/drm/drm_writeback.h
>> new file mode 100644
>> index 0000000..6b2ac45
>> --- /dev/null
>> +++ b/include/drm/drm_writeback.h
>> @@ -0,0 +1,78 @@
>> +/*
>> + * (C) COPYRIGHT 2016 ARM Limited. All rights reserved.
>> + * Author: Brian Starkey <brian.starkey@xxxxxxx>
>> + *
>> + * This program is free software and is provided to you under the terms of the
>> + * GNU General Public License version 2 as published by the Free Software
>> + * Foundation, and any use by you of this program is subject to the terms
>> + * of such GNU licence.
>> + */
>> +
>> +#ifndef __DRM_WRITEBACK_H__
>> +#define __DRM_WRITEBACK_H__
>> +#include <drm/drm_connector.h>
>> +#include <linux/workqueue.h>
>> +
>> +struct drm_writeback_connector {
>> + struct drm_connector base;
>
>AFAIU, a writeback connector will always require an 'dummy' encoder to
>make the DRM framework happy (AFAIK, a connector is always connected to
>a CRTC through an encoder).
>
>Wouldn't it make more sense to have a drm_encoder object embedded in
>drm_writeback_connector so that people don't have to declare an extra
>structure containing both the drm_writeback_connector connector and a
>drm_encoder? Is there a good reason to keep them separate?
>

Yeah that's not a bad idea. The encoder funcs could be passed in to
drm_writeback_connector_init() (in which case adding a writeback
encoder type would also make sense).

Well, the more I look at it the more I find it weird to represent this
writeback feature as a connector. To me, it seems to be closer to a
plane object (DRM_PLANE_TYPE_WRITEBACK?) than a real connector.
Actually, the Atmel HLCDC IP use the same register interface to expose
the overlay planes and writeback features.
Of course, representing it as a plane requires patching the core to
allow enabling a CRTC that has no active connectors connected to it if
at least one writeback plane is enabled and connected to the CRTC, but
it should be doable.

Could you expand a bit on how you think planes fit better? It is
something we've previously talked about internally, but so far I'm not
convinced :-)

By doing that, we would also get rid of these fake connector/encoder
objects as well as the fake modes we are returning in
connector->get_modes().

What makes the connector/encoder fake? They represent a real piece of
hardware just the same as a drm_plane would.

I don't mind dropping the mode list and letting userspace just make
up whatever timing it wants - it'd need to do the same if writeback
was a plane - but in some respects giving it a list of modes the same
way we normally do seems nicer for userspace.


As far as I can tell, the VC4 and Atmel HLCDC IP should fit pretty well
in this model, not sure about the mali-dp though.

Brian, did you consider this approach, and if you did, can you detail
why you decided to expose this feature as a connector?

Daniel (or anyone else), please step in if you think this is a stupid
idea :-).

Ville originally suggested using a connector, which Eric followed up
by saying that's what he was thinking of for VC4 writeback too[1].
That was my initial reason for focussing on a connector-based
approach.

I prefer connector over plane conceptually because it keeps with the
established data flow: planes are sources, connectors are sinks.

In some respects the plane _object_ looks like it would fit - it has a
pixel format list and an FB_ID. But everything else would be acting
the exact opposite to a normal plane, and I think there's a bunch of
baked-in assumptions in the kernel and userspace around CRTCs always
having at least one connector.

On the other hand, a writeback connector gains a few extra properties
over a normal connector, but most stuff stays the same. The pipeline
setup looks the same as normal to userspace, you don't need a CRTC to
be active with no connectors, output cloning is the same etc.

Thanks,
-Brian

[1] https://lists.freedesktop.org/archives/dri-devel/2016-July/113329.html