Hi everybody,
Here's a summary of the CDF BoF that took place at the ELC 2013.
I'd like to start by thanking all the participants who provided valuable feedback (and those who didn't, but who now know a bit more about CDF and will, I have no doubt about that, contribute in the future :-)). Thank you also to Linus Walleij and Jesse Barker for taking notes during the meeting while I was presenting. And obviously, thank you to Jesse Barker for organizing the BoF.
I've tried to be as accurate as possible in this summary, but I might have made mistakes. If you have attended the meeting, please point out any issue, inconsistency, or just points I might have forgotten.
----
As not all attendees were familiar with CDF I started by briefly introducing the problems that prompted me to start working on CDF.
CDF started as GPF, the Generic Panel Framework. While working on DT support for a display controller driver I realized that panel control code was located in board file. Moving the code somewhere in drivers/ was thus a prerequisite, but it turned out that no framework existed in the kernel to support that tasks. Several major display controller drivers (TI DSS and Samsung Exynos to name a few) had a platform-specific panel driver framework, but the resulting panel drivers wouldn't be reusable across different display controllers. A need for a new framework became pretty evident to me.
After drafting an initial proposal and discussing it with several people online and offline (in Helsinki with Tomi Valkeinen from TI, in Copenhagen at Linaro Connect with Marcus Lorentzon from ST-Ericsson, and in Brussels during a BoF at the FOSDEM) the need to support encoders in addition to panels quickly arose, and GPF turned into CDF.
I then pursued with an overview of the latest CDF code and its key concepts. While I was expecting this to be a short overview followed by more in-depth discussions, it turned out to support our discussions for the whole 2 hours meeting.
The latest available version at the time of the BoF (posted to the linaro-mm- sig mailing list in reply to the BoF's annoucement) was the "non-quite-v3" version. It incorporated feedback received on v2 but hadn't been properly tested yet.
The basic CDF building block is called a display entity, modeled as an instance of struct display_entity. They have sink ports through which they receive video data and/or source ports through which they transmit video data. Entities are chained via their ports to create a display pipeline.
From the outside world entities are interfaced through two sets of abstract
operations they must provide:
- Control operations are called from "upper layers" (usually to implement userspace requests) to get and set entity parameters (such as the physical size, video modes, operation states, bus parameters, ...). Those operations are implemented at the entity level.
Google asked how partial updates were handled, I answered they're not handled yet (this is a key concept behind the CDF RFCs: while I try to make sure all devices can be supported, I first concentrate on hardware features required for the devices I work on). Linus Walleij mentioned he thought that partial updates were becoming out of fashion, but larger display sizes might keep them useful in the future.
- Video operations control video streams. They're implemented by entities on their source ports, and are called in the upstream (from a video pipeline point of view) direction. A panel will call video operations of the entity it gets its video stream from (this could be an HDMI transmitter, the display controller directly, ...) to control the video stream it receives.
Video operations are split in a set of common operations and sets of display bus specific operations (for DPI, DBI, DSI, ...). Some discussion around ops that might be needed in some cases but not others indicate that the ops structures are not quite finished for all bus types (and/or that some ops might be considered for "promotion" to common). In particular the current DSI implementation is copied from a proposal posted by Tomasz Figa from Samsung. As I have no DSI hardware to test it on I have kept it as-is.
Jesse Barker pointed out that to make this fly we willl need to get CDF into a number of implementations, in particular the Samsung Exynos SoCs (needing DSI). Several efforts are ongoing:
- Marcus Lorentzon (ST Ericsson, Linaro) is working on porting ST Ericsson code to CDF, and in particular on the DSI interface.
- Tomasz Figa (Samsung) has worked on porting the Exynos display controller driver to CDF and provided a DSI implementation.
- Tomi Valkeinen (TI) is working on porting the TI DSS driver to CDF (or rather his own version of CDF as a first step, to avoid depending on an ever- moving target right now) independently from Linaro.
- Alison Chaiken (Mentor Embedded Software) mentioned that Pengutronix is working on panels support for the Freescale i.MX family.
- Linaro can probably also help extending the test coverage to various platforms from its member companies.
- Finally, I'm working on CDF support for two display controllers found in Renesas SoCs. One of them support DBI and DPI, the other supports DPI only. However, I can't easily test DBI support, as I don't have access to the necessary hardware.
I explained at that point that there is currently no clear agreement on a bus and operations model. The initial CDF proposal created a Linux busses for DBI and DSI (similar to I2C and SPI busses), with access to the control bus implemented through those Linux busses, and access to the video bus implemented through video operations on display entities. Tomi Valkeinen then advocated for getting rid of the DBI and DSI Linux busses and implementing access to both control and video through the display entity operations, while Marcus Lorentzon wanted to implement all those operations at the Linux bus level instead. The best way to arbitrate this will probably to work on several implementations and find out which one works better.
SONY Mobile currently supports DSI auto-probing, with plug-n-play detection of DSI panels. The panel ID is first retrieved, and the correct panel driver is then loaded. We will likely need to support a similar model. Another option would be to write a single panel-dcs driver to support all DSI panels that conform with the DSI and DCS standards (although we will very likely need panel-specific quirks in that case). The two options could also coexist.
We then moved to how display entities should be handled by KMS drivers and mapped to KMS objects. The KMS model hardcodes the following fixed pipeline
CRTC -> encoder -> connector
The CRTC is controlled by the display controller driver, and panels can be mapped to KMS connector objects. What goes in-between is more of a gray area, as hardware pipeline can have several encoders chained together.
I've presented one possible control flow that could solution the problem by grouping multiple objects into an abstract entity. The right-most entity would be a standalone entity, and every encoder but the left-most one in the chain would hide the entities connected at their output. This results in a "russian dolls" model, where encoders forward control operations to the entities they embed, and forward video operations to the entity at their sink side.
This can quickly become very complex, especially when locking and reference counting are added to the model. Furthermore, this solution could only handle linear pipelines, which will likely become a severe limitation in the future, especially on embedded devices (for instance splitting a video stream between two panels at the encoder level is a common use case, or driving a two-inputs panel from two CRTCs).
Google asked whether this model tries to address both panels and VGA(/HDMI/...) outputs. From what I've seen so far the only limits come from the hardware engineers (often^H^H^H^H^Hsometimes troubled) minds, all kinds of data streams may appear in practice. As most systems will have one CRTS, one encoder and one panel (or connector), we should probably try to keep the model simple to start with with 1:1 mappings between the KMS CRTC/encoder/connector model and the CDF model. If we try to solve every possible problem right now the complexity will explode and we won't be able to handle it. Getting a simple solution upstream now and refactoring it later (there is no userspace API involved, so no backward compatibility issue) might be the right answer. I have no strong feeling about it, but I certainly want something I can get upstream in a reasonable time frame.
Keith Packard bluntly (and totally rightfully) whether CDF is not just duplicating part of the KMS API, and whether we shouldn't instead extend the in-kernel KMS model to handle multiple encoders.
One reason that drove the creation of CDF outside of KMS was to support sharing a single driver between multiple subsystems. For instance an HDMI encoder could be connected to the output of a display controller handled by a KMS driver on one board, and to the output of a video processor handled by a V4L2 driver on another board. A panel could also be connected to a display controller handled by a KMS driver on one board, and to a display controller handled by an FBDEV driver on another board. Having a single driver for those encoders or panels is one of the goals of CDF.
After publishing the first CDF RFC I realized there was a global consensus in the kernel display community to deprecate FBDEV at some point. Sharing panel drivers between KMS and FBDEV then became a "nice to have, but not important" feature. As V4L2 doesn't handle panels (and shouldn't be extended to do so) only encoders drivers would need to be shared, between KMS and V4L2.
It's important to note here that we don't need to share a given encoder between two subsystems at runtime. On a given board the encoder will need to be controlled by KMS or V4L2, but never both at the same time. In the CDF context driver sharing refers to the ability to control a given driver from either the KMS or V4L2 subsystem.
The discussion then moved to why V4L2 drivers for devices connected to an encoder couldn't be moved to KMS. All display devices should be handled by KMS, but we still have use cases where V4L2 need to handle video outputs. For instance a system with the following pipeline
HDMI con. -> HDMI RX -> Processing Engine -> HDMI TX -> HDMI con.
doesn't involve memory buffers in the processing pipeline. This can't be handled by KMS, as KMS cannot reporesent a video pipeline without memory in- between the receiving side and the display side. Hans Verkuil also mentioned that for certain applications one prefers to center the API around frames, and that V4L2 is ideal for instance for video conferencing/telephony.
Keith Packard thought we should just extend KMS to handle the V4L2 use cases. V4L2 would then (somehow) plug its infrastructure into KMS. This topic has already been discussed in the past, and I agree that extending the KMS model to support "live sources" for CRTCs will be needed in the near future. This could be the basis of other KMS enhancements to support more complex pipelines. Making KMS and V4L2 cooperate is also desirable on the display side to write the output of the CRTC back to memory. KMS has no write-back feature in the API, V4L2 could come to the rescue there.
With this kind of extension it might be possible to handle the display part of memory-less pipelines in KMS, although that might be quite a challenge. There was no clear consensus on whether this was desirable.
Furthermore, only two HDMI encoders currently need to be shared (both are only supported out-of-tree at the moment). As we don't expect more than a handful of such use cases in the near future, it might not be worth the hasle to create a complete infrastructure to handle a use case that might disappear if we later move all the display-side drivers to KMS.
Another solution mentioned by Hans Verkuil would be to create helper functions to translate V4L2 calls to KMS calls (to be clear, this only covers in-kernel calls to encoders).
There was no clear consensus on this topic.
We then moved on to the hot-plug (and hot-unplug) issues following a question from Google. Hot-plug is currently not supported. We would need to add hot- plugging notifiers and possibly a couple of other operations. However, the video common operations structure has bind/unbind operations, that can serve as a basis.
The hard part in hot-plugging support is actually hot-unplugging, as we need to ensure that devices don't disappear all of a sudden while still in use. This was a design goal of CDF from the start, and any issue there will need to be resolved. Panels shouldn't be handled differently than HDMI connectors, CDF will provide a common hot-plugging model.
Keith Packard then explained that DRM and KMS will likely be split in the future. The main link between the DRM and KMS APIs is GEM objects. With the recent addition of dmabuf to the Linux kernel the DRM and KMS APIs could be split and use dmabuf to share buffers. DRM and KMS would then be exposed on two separate device nodes. It would be a good idea to revisit the whole KMS/V4L2 unification discussion when DRM and KMS will be split.
We briefly touched the subject of namespaces, and whether CDF should use the KMS namespace (drm_*). There is some resistance on the V4L2 side on having CDF structures be KMS objects.
It was then time to wrap up the meeting, and I asked the audience one final question: should we shoehorn complex pipelines into the KMS three-stages model, or should we extend the KMS model? That was unfortunately answered by silence, showing that more thinking is needed.
A couple more minutes of offline discussions briefly touched the topics of GPU driver reverse engineering and whether we could, after the KMS/DRM split, set a kernel-side standard for embedded GPU drivers. As interesting as this topic is, CDF will not solve that problem :-)