Re: [PATCH 00/20] drm: Split out the formats API and move it to a common place

From: Laurent Pinchart
Date: Sat May 11 2019 - 15:28:22 EST


Hi Daniel,

On Tue, Apr 23, 2019 at 09:18:52PM +0200, Daniel Vetter wrote:
> On Tue, Apr 23, 2019 at 5:45 PM Laurent Pinchart wrote:
> > On Tue, Apr 23, 2019 at 09:25:54AM +0200, Daniel Vetter wrote:
> >> On Sun, Apr 21, 2019 at 01:59:04AM +0300, Laurent Pinchart wrote:
> >>> On Thu, Apr 18, 2019 at 12:07:44PM +0200, Daniel Vetter wrote:
> >>>> On Thu, Apr 18, 2019 at 11:02 AM Maxime Ripard wrote:
> >>>>> On Thu, Apr 18, 2019 at 09:52:10AM +0200, Daniel Vetter wrote:
> >>>>>> On Thu, Apr 18, 2019 at 8:22 AM Maxime Ripard wrote:
> >>>>>>> On Wed, Apr 17, 2019 at 05:41:21PM +0200, Daniel Vetter wrote:
> >>>>>>>> On Wed, Apr 17, 2019 at 09:54:26AM +0200, Maxime Ripard wrote:
> >>>>>>>>> Hi,
> >>>>>>>>>
> >>>>>>>>> DRM comes with an extensive format support to retrieve the various
> >>>>>>>>> parameters associated with a given format (such as the subsampling, or the
> >>>>>>>>> bits per pixel), as well as some helpers and utilities to ease the driver
> >>>>>>>>> development.
> >>>>>>>>>
> >>>>>>>>> v4l2, on the other side, doesn't provide such facilities, leaving each
> >>>>>>>>> driver reimplement a subset of the formats parameters for the one supported
> >>>>>>>>> by that particular driver. This leads to a lot of duplication and
> >>>>>>>>> boilerplate code in the v4l2 drivers.
> >>>>>>>>>
> >>>>>>>>> This series tries to address this by moving the DRM format API into lib and
> >>>>>>>>> turning it into a more generic API. In order to do this, we've needed to do
> >>>>>>>>> some preliminary changes on the DRM drivers, then moved the API and finally
> >>>>>>>>> converted a v4l2 driver to give an example of how such library could be
> >>>>>>>>> used.
> >>>>>>>>>
> >>>>>>>>> Let me know what you think,
> >>>>>>>>> Maxime
> >>>>>>>>>
> >>>>>>>>> Changes from RFC:
> >>>>>>>>> - Rebased on next
> >>>>>>>>> - Fixed the various formats mapping
> >>>>>>>>> - Added tags
> >>>>>>>>> - Did most of the formats functions as inline functions
> >>>>>>>>> - Used a consistent prefix for all the utilities functions
> >>>>>>>>> - Fixed the compilation breakages, and did a run of allmodconfig for arm,
> >>>>>>>>> arm64 and x86_64
> >>>>>>>>> - Fixed out of array bounds warnings in the image_format_info_block_*
> >>>>>>>>> functions
> >>>>>>>>> - Added License and copyright headers on missing files
> >>>>>>>>>
> >>>>>>>>> Maxime Ripard (20):
> >>>>>>>>> drm: Remove users of drm_format_num_planes
> >>>>>>>>> drm: Remove users of drm_format_(horz|vert)_chroma_subsampling
> >>>>>>>>> drm/fourcc: Pass the format_info pointer to drm_format_plane_cpp
> >>>>>>>>> drm/fourcc: Pass the format_info pointer to drm_format_plane_width/height
> >>>>>>>>> drm: Replace instances of drm_format_info by drm_get_format_info
> >>>>>>>>> lib: Add video format information library
> >>>>>>>>> drm/fb: Move from drm_format_info to image_format_info
> >>>>>>>>> drm/malidp: Convert to generic image format library
> >>>>>>>>> drm/client: Convert to generic image format library
> >>>>>>>>> drm/exynos: Convert to generic image format library
> >>>>>>>>> drm/i915: Convert to generic image format library
> >>>>>>>>> drm/ipuv3: Convert to generic image format library
> >>>>>>>>> drm/msm: Convert to generic image format library
> >>>>>>>>> drm/omap: Convert to generic image format library
> >>>>>>>>> drm/rockchip: Convert to generic image format library
> >>>>>>>>> drm/tegra: Convert to generic image format library
> >>>>>>>>> drm/fourcc: Remove old DRM format API
> >>>>>>>>> lib: image-formats: Add v4l2 formats support
> >>>>>>>>> lib: image-formats: Add more functions
> >>>>>>>>> media: sun6i: Convert to the image format API
> >>>>>>>>
> >>>>>>>> In the interest of making myself unpopular: Why move this out of drm?
> >>>>>>>>
> >>>>>>>> We do have a bunch of other such shared helpers already (mostly in
> >>>>>>>> drivers/video) for dt videomode and hdmi infoframes, and I'm not super
> >>>>>>>> sure that's going better than keeping it maintained in drm.
> >>>
> >>> That's a bit of a different situation as both DRM and FBDEV address the
> >>> same features (display output), and FBDEV is deprecared and replaced by
> >>> DRM.
> >>>
> >>> I'm not against maintaining a 4CC library in DRM (adding a third-party
> >>> maintainer would likely create more problems than it would solve), but
> >>> that doesn't mean the library has to live in drivers/gpu/, nor that it
> >>> needs to have the drm_ prefix. I would actually advocate to make it live
> >>> in a neutral directory, with a neutral prefix (kudos to anyone who can
> >>> propose a nice naming convention that would establish a new shared
> >>> ground for image/video-related Linux APIs), and maintained through the
> >>> DRM tree (possibly with extra entries in MAINTAINERS to ensure it
> >>> reaches all the related folks).
> >>>
> >>>>>>>> Plus the uapi is already that you include drm_fourcc.h to get at these,
> >>>>>>>> dropping the drm prefix isn't going to help I think.
> >>>>>>>>
> >>>>>>>> Of course we'd need to make it a separate drm_formats.ko (so that v4l can
> >>>>>>>> use it without dragging in all of drm), and we need to add some fields for
> >>>>>>>> converting to v4l fourcc codes (which are different). But that should be
> >>>>>>>> all possible. And I don't think the drm_ prefix in v4l code is a problem,
> >>>>>>>> it's actually a feature: It makes it really clear that these are the drm
> >>>>>>>> fourcc codes, as allocated in drm_fourcc.h, plus their modifiers, and all
> >>>>>>>> that. That allocation authority is also already baked into various khr/ext
> >>>>>>>> standards, too.
> >>>
> >>> There's one thing that V4L2 has and DRM hasn't for 4CCs: good
> >>> documentation. Look at
> >>> https://linuxtv.org/downloads/v4l-dvb-apis-new/uapi/v4l/pixfmt-packed-rgb.html
> >>> or
> >>> https://linuxtv.org/downloads/v4l-dvb-apis-new/uapi/v4l/yuv-formats.html
> >>> for instance. It's painful to write, painful to read, but defines the
> >>> 4CCs very clearly without ambiguity. I wouldn't be surprised if
> >>> different drivers used the same DRM 4CC for different formats due to the
> >>> lack of detailed documentation. Moving to a shared library for 4CCs
> >>> should also address the documentation side, and any new format added to
> >>> the kernel, whether from the V4L2 side or the DRM side, would be
> >>> required to provide detailed documentation. That would be a great
> >>> improvement for DRM 4CC handling.
> >>>
> >>>>>>> The way I see it, there's a fundamental difference between the UAPI
> >>>>>>> and the kernel. I don't suggest we change anything about the UAPI: the
> >>>>>>> drm formats should keep their prefix, drm_fourcc.h can remain that
> >>>>>>> authority, it's all fine.
> >>>>>>>
> >>>>>>> Internally however, the long term goal is to share the fourcc's
> >>>>>>> between DRM and V4L2 for the same formats. It basically means that of
> >>>>>>> course v4l2 should be using the DRM fourcc when a format exists in DRM
> >>>>>>> and not v4l2, but also that DRM should use v4l2 fourcc when the format
> >>>>>>> exists in v4l2 but not DRM, and that is far more likely, given the
> >>>>>>> already extensive v4l2 formats support.
> >>>>>>
> >>>>>> Uh no. We did look at v4l fourcc extensively when deciding upon a drm
> >>>>>> format identifier space.
> >>>>>
> >>>>> No to what exactly?
> >>>>>
> >>>>>> And a lot of people pushed for the "fourcc is a standard", when
> >>>>>> really it's totally not.
> >>>>>
> >>>>> Even if it's not a standard, having consistency would be a good thing.
> >>>>>
> >>>>> And you said yourself that DRM fourcc is now pretty much an authority
> >>>>> for the fourcc, so it definitely looks like a standard to me.
> >>>>
> >>>> drm fourcc is the authority for drm fourcc codes. Not for any of the
> >>>> others (and there's lots of them). Now it's used in a bunch of other
> >>>> places (khr standards, dri protocols in wayland/X11), but they're
> >>>> still only drm fourcc. There is no overall fourcc standard.
> >>>>
> >>>>>> v4l tends to conflate pixel format with stuff that we tend to encode
> >>>>>> in modifiers a lot more.
> >>>>>
> >>>>> Boris is working on adding the modifiers concept to v4l2, so we're
> >>>>> converging here, and we can totally have a layer in v4l2 to convert
> >>>>> between old v4l2 "format+modifiers" formats, and DRM style formats.
> >>>>>
> >>>>>> There's a bunch of reasons we can't just use v4l, and they're as
> >>>>>> valid as ever:
> >>>>>>
> >>>>>> - We overlap badly in some areas, so even if fourcc codes match, we
> >>>>>> can't use them and need a duplicated DRM_FOURCC code.
> >>>>>
> >>>>> Do yo have an example of one of those areas?
> >>>>
> >>>> I think the rgb stuff was one of the big reasons to not reuse any
> >>>> existing fourcc standard (whether v4l, or another one from e.g. video
> >>>> container formats). We had initial patches to reuse the fourcc that
> >>>> existed, but the end result was really inconsistent, so we ditch that
> >>>> and rolled out a new set of entirely drm specific fourcc codes for
> >>>> rgba.
> >>>
> >>> Could you give one or a couple of examples of V4L2 4CCs that are not
> >>> OCD-compatible ? :-)
> >>>
> >>>>>> - v4l encodes some metadata into the fourcc that we encode elsewhere,
> >>>>>> e.g. offset for planar yuv formats, or tiling mode
> >>>>>
> >>>>> As I was saying, this changes on the v4l2 side, and converging to
> >>>>> what DRM is doing.
> >>>>>
> >>>>>> - drm fourcc code doesn't actually define the drm_format_info
> >>>>>> uniquely, drivers can override that (that's an explicit design
> >>>>>> intent of modifiers, to allow drivers to add another plane for
> >>>>>> e.g. compression information). You'd need to pull that driver
> >>>>>> knowledge into your format library.
> >>>
> >>> That's a mistake in my opinion. We tried that in V4L2 to store metadata
> >>> in a separate plane, and had to go another route eventually as it
> >>> created a very bad mess.
> >>
> >> Just quick clarification in the middle here: This is how the hw works.
> >
> > The hardware takes parameters from a buffer, but it doesn't mandate how
> > that buffer is exposed to userspace :-) Using an extra plane is one
> > option, but other APIs are possible.
>
> I think Daniel Stone explains fairly well why some of our additional
> metadata is included as a plane, while a lot of the other metadata
> involved in rendering/compute the framebuffer isn't. Not really
> anything to add here.
>
> >> It's not metadata that sw ever touches (in general, testcases to make sure
> >> we display these correctly excepted).
> >>
> >> There has been some talking to add maybe a bit more mixed metadata, for
> >> fast-clear colors (which isn't used by any display engine afaik yet). That
> >
> > What are fast-clear colors ?
>
> hw offers enormous amounts of tricks to make glClear O(1), or at least
> close enough. glClear is usually what's done at the start of every
> frame, and clears the entire framebuffer to a uniform color. This is
> achieved usually through 3 pieces:
> - actual framebuffer plane with the pixel data
> - a 2nd plane that (usually, but there's lots of tricks here) contains
> a bit of metadata per cacheline/block/whatever in the framebuffer to
> indicate how/if those pixels are compressed, or whether they are still
> the uniform color supplied through glClear. For actual O(1) this
> metadata is hierarchical, so that a glClear really only sets the
> top-level metadata to "all subordinate blocks still have are the clear
> color". hw tends to have some pretty strong opinions on where it's
> going to look for that 2nd plane.
> - curiously on most hw the actual clear color is supplied separately
> (and our plan is to just stuff it into a 3rd plane)

Just to clarify, is this needed for display engines ? Does the GPU
render to a 3 planes buffer with glClear()-related data in planes 2 and
3, with the buffer then being passed to the display engine that knows
how to interpret this, or are those extra planes only needed for GPU
rendering ?

> >> would generally be written by the cpu (in the gl stack), but again read by
> >> the hw (loaded as indirect state packet most likely, or something like
> >> that). So again hw specific layout, because the hw needs to read it.
> >>
> >> Pure metadata only of interest for the cpu/sw stack has been shot down
> >> completely on the drm side too.
> >>
> >>> There's a tendency in both subsystems to look at the other as a bit of a
> >>> retarded relative, and that's a shame, we have lots to learn from each
> >>> other's mistakes. That wouldn't be too difficult if people started
> >>> talking to each other.
> >>>
> >>> A semi-related comment: DRM also carries other mistakes of its own, I'm
> >>> thinking about DRM_FORMAT_BIG_ENDIAN in particular
> >>
> >> Yeah that one is hilarios, but in practice big endian is dead, except for
> >> a very few server chips, and there I think Gerd's work mostly fixed up
> >> that mess.
> >>
> >>>>> I'm not sure how my patches are changing anything here. This is
> >>>>> litterally the same code, with the functions renamed.
> >>>>>
> >>>>> If drivers want to override that, then yeah, fine, we can let them do
> >>>>> that. Just like any helper this just provides a default that covers
> >>>>> most of the cases.
> >>>>>
> >>>>>> Iow there's no way we can easily adopt v4l fourcc, except if we do
> >>>>>> something like a new addfb flag.
> >>>>>
> >>>>> For the formats that would be described as a modifier, sure. For all
> >>>>> the others (that are not yet supported by DRM), then I don't really
> >>>>> see why not.
> >>>>
> >>>> See above, we tried that initially, didn't pass the ocd tests when
> >>>> reviewing. Maybe the situation is better outside of rbgx/a formats,
> >>>> and I think we do at least try to use the same fourcc codes there when
> >>>> there already is one. But then the details occasionally look
> >>>> different.
> >>>>
> >>>>>>> And given how the current state is a mess in this regard, I'm not too
> >>>>>>> optimistic about keeping the formats in their relevant frameworks.
> >>>>>>>
> >>>>>>> Having a shared library, governed by both, will make this far easier,
> >>>>>>> since it will be easy to discover the formats that are already
> >>>>>>> supported by the other subsystem.
> >>>>>>
> >>>>>> I think a compat library that (tries to, best effort) convert between
> >>>>>> v4l and drm fourcc would make sense. Somewhere in drivers/video, next
> >>>>>> to the conversion functions for videomode <-> drm_display_mode
> >>>>>> perhaps. That should be useful for drivers.
> >>>>>
> >>>>> That's not really what this series is about though. That series is
> >>>>> about sharing the (image|pixels) formats database across everyone so
> >>>>> that everyone can benefit from it.
> >>>>
> >>>> Yeah I know. I still think leaving the drm fourcc with the drm prefix
> >>>> would be good, since there's really no standard here.
> >>>>
> >>>>>> Unifying the formats themselves, and all the associated metadata is
> >>>>>> imo a no-go, and was a pretty conscious decision when we implemented
> >>>>>> drm_fourcc a few years back.
> >>>>>>
> >>>>>>> If we want to keep the current library in DRM, we have two options
> >>>>>>> then:
> >>>>>>>
> >>>>>>> - Support all the v4l2 formats in the DRM library, which is
> >>>>>>> essentially what I'm doing in the last patches. However, that
> >>>>>>> would require to have the v4l2 developpers also reviewing stuff
> >>>>>>> there. And given how busy they are, I cannot really see how that
> >>>>>>> would work.
> >>>>>>
> >>>>>> Well, if we end up with a common library then yes we need cross
> >>>>>> review. There's no way around that. Doesn't matter where exactly that
> >>>>>> library is in the filesystem tree, and adding a special MAINTAINERS
> >>>>>> entry for anything related to fourcc (both drm and v4l) to make sure
> >>>>>> they get cross-posted is easy. No file renaming needed.
> >>>>>
> >>>>> That would create some governing issues as well. For example, if you
> >>>>> ever have a patch from one fourcc addition (that might or might not be
> >>>>> covered by v4l2), will you wait for any v4l2 developper to review it?
> >>>>
> >>>> None of this is fixed by code renaming or locations. Either way we
> >>>> need to figure that out.
> >>>>
> >>>> And yes part of the reasons for not moving this out of drm is that I'm
> >>>> not a fan of boutique trees for small stuff. If sharing means we need
> >>>> to split the drm_fourcc code and library out of drm trees, I'm not
> >>>> sure that's a good idea. We're already super liberal with merging
> >>>> anything through driver trees with acks, and integrating them quickly
> >>>> into drm-next. This would still be workable if v4l sends regular pull
> >>>> requests to drm-next (every 1-2 weeks, like the other big gpu trees
> >>>> do). If we can only sync up once per merge window with a shared
> >>>> boutique tree for formats only, life is going to be painful.
> >>>
> >>> That should be solved by the proposal above, maintaining the shared
> >>> library in the DRM tree. We would need to be careful there, and ideally
> >>> maintain that in a separate branch that could be merged in both DRM and
> >>> V4L2 without having to merge most of the other subsystem's pending
> >>> changes at the same time, but I think it's doable without any big issue.
> >>>
> >>>>> If it's shared code, then it should be shared, and every client
> >>>>> framework put on an equal footing.
> >>>>>
> >>>>>>> - Develop the same library from within v4l2. That is really a poor
> >>>>>>> solution, since the information would be greatly duplicated
> >>>>>>> between the two, and in terms of maintainance, code, and binary
> >>>>>>> size that would be duplicated too.
> >>>>>>
> >>>>>> It's essentially what we decided to do for drm years back.
> >>>>>
> >>>>> And it was probably the right solution back then, but I'm really not
> >>>>> convinced it's still the right thing to do today.
> >>>>>
> >>>>>>> Having it shared allows to easily share, and discover formats from the
> >>>>>>> other subsystem, and to have a single, unique place where this is
> >>>>>>> centralized.
> >>>>>>
> >>>>>> What I think could work as middle ground:
> >>>>>> - Put drm_format stuff into a separate .ko
> >>>>>> - Add a MAINTAINERS entry to make sure all things fourcc are cross
> >>>>>> posted to both drm and v4l lists. Easy on the drm side, since it's all
> >>>>>> separate files. Not sure it's so convenient for v4l uapi.
> >>>>>> - Add a conversion library that tries to best-effort map between drm
> >>>>>> and v4l formats. On the drm side that most likely means you need
> >>>>>> offsets for planes, and modifiers too (since those are implied in some
> >>>>>> v4l fourcc). Emphasis on "best effort" i.e. only support as much as
> >>>>>> the drivers that use this library need.
> >>>>>> - Add drm_fourcc as-needed by these drivers that want to use a unified
> >>>>>> format space.
> >>>>>>
> >>>>>> Forcing this unification on everyone otoh is imo way too much.
> >>>>>
> >>>>> v4l2 is starting to converge with DRM, and we're using the DRM API
> >>>>> pretty much untouched for that library, so I'm not really sure how
> >>>>> anyone is hurt by that unification.
> >>>>
> >>>> It might make sense to regularly pull v4l updates into drm-next then
> >>>> anyway. That would also remove the need to have the format library
> >>>> somewhere else.
> >>>
> >>> Are you saying it should the live in V4L2 ? ;-)
> >>
> >> Maybe a few clarifications on how the drm shared core thing usually works,
> >> and why I'm a sticker here. Bottom reply since I'm not sure where to put
> >> it:
> >>
> >> - Refactorings usually go in through drm-misc (at least since a few
> >> years).
> >>
> >> - Small patches go in through the relevant driver tree (which is often
> >> drm-misc, but not always), with an Ack from drm maintainers.
> >>
> >> - No topic branches, everyone just pushes patches where it's most
> >> convenient.
> >>
> >> We get away with this mess because everyone sends regular pull requests to
> >> drm, where the entire history is baked in and others can backmerge/fast
> >> forward/rebase. Worst case you wait one month (around the merge window,
> >> when drm-next is closed for features), but usually it's just 1-2 weeks.
> >> Plus we tend to have fairly big trees, with good chances that the next
> >> patch series lands in the same tree again and no work at all is needed.
> >>
> >> If we start regularly sharing lots of code with v4l (which seems to be the
> >> long term goal here), then I think we need something equally convenient
> >> for all that.
> >>
> >> We're not going to be able to teach some complicated topic branch scheme
> >> to the 50+ submaintainers/committers we have in drm - a lot much more
> >> basic stuff already takes lots of work to get it to stick. If the proposal
> >> is "to be careful" and "maintain it in a separate branch", I'm not in
> >> favour because I think that just wouldn't work.
> >
> > Why not ? It can be a fast-moving branch that gets merged in drm-misc
> > as often as you want (even at every commit if that's desired). We're
> > dealing with a limited amount of code here, and there's no more reason
> > that V4L2 should pull in drm-misc to get 4CC updates than DRM should
> > pull V4L2 for the same. I have no objection against a 4CC branch being
> > officially maintained under the DRM umbrella, but I think the code
> > should live elsewhere than drivers/gpu/drm/, have a neutral prefix, and
> > not require pulling an entire subsystem in.
>
> I think small boutique trees are a problem themselves, not a solution.
> So if you're creating a new small boutique tree to fix a problem, you
> then have 2. Yes, assuming sufficient expenditure of energy it can be
> made to work, but I'd prefer to make my own life as easy and lazy as
> possible :-) And I think I've been fairly successful at that within
> drivers/gpu at least.
>
> Imo the proper fix is to merge v4l and drm, at a process/maintainer
> level. That would solve both the original issue and the 2ndary one of
> the permanent topic branch.

Proposals are welcome ;-)

--
Regards,

Laurent Pinchart