linux/Documentation/gpu/todo.rst

728 lines
27 KiB
ReStructuredText

.. _todo:
=========
TODO list
=========
This section contains a list of smaller janitorial tasks in the kernel DRM
graphics subsystem useful as newbie projects. Or for slow rainy days.
Difficulty
----------
To make it easier task are categorized into different levels:
Starter: Good tasks to get started with the DRM subsystem.
Intermediate: Tasks which need some experience with working in the DRM
subsystem, or some specific GPU/display graphics knowledge. For debugging issue
it's good to have the relevant hardware (or a virtual driver set up) available
for testing.
Advanced: Tricky tasks that need fairly good understanding of the DRM subsystem
and graphics topics. Generally need the relevant hardware for development and
testing.
Expert: Only attempt these if you've successfully completed some tricky
refactorings already and are an expert in the specific area
Subsystem-wide refactorings
===========================
Remove custom dumb_map_offset implementations
---------------------------------------------
All GEM based drivers should be using drm_gem_create_mmap_offset() instead.
Audit each individual driver, make sure it'll work with the generic
implementation (there's lots of outdated locking leftovers in various
implementations), and then remove it.
Contact: Daniel Vetter, respective driver maintainers
Level: Intermediate
Convert existing KMS drivers to atomic modesetting
--------------------------------------------------
3.19 has the atomic modeset interfaces and helpers, so drivers can now be
converted over. Modern compositors like Wayland or Surfaceflinger on Android
really want an atomic modeset interface, so this is all about the bright
future.
There is a conversion guide for atomic and all you need is a GPU for a
non-converted driver (again virtual HW drivers for KVM are still all
suitable).
As part of this drivers also need to convert to universal plane (which means
exposing primary & cursor as proper plane objects). But that's much easier to
do by directly using the new atomic helper driver callbacks.
Contact: Daniel Vetter, respective driver maintainers
Level: Advanced
Clean up the clipped coordination confusion around planes
---------------------------------------------------------
We have a helper to get this right with drm_plane_helper_check_update(), but
it's not consistently used. This should be fixed, preferrably in the atomic
helpers (and drivers then moved over to clipped coordinates). Probably the
helper should also be moved from drm_plane_helper.c to the atomic helpers, to
avoid confusion - the other helpers in that file are all deprecated legacy
helpers.
Contact: Ville Syrjälä, Daniel Vetter, driver maintainers
Level: Advanced
Improve plane atomic_check helpers
----------------------------------
Aside from the clipped coordinates right above there's a few suboptimal things
with the current helpers:
- drm_plane_helper_funcs->atomic_check gets called for enabled or disabled
planes. At best this seems to confuse drivers, worst it means they blow up
when the plane is disabled without the CRTC. The only special handling is
resetting values in the plane state structures, which instead should be moved
into the drm_plane_funcs->atomic_duplicate_state functions.
- Once that's done, helpers could stop calling ->atomic_check for disabled
planes.
- Then we could go through all the drivers and remove the more-or-less confused
checks for plane_state->fb and plane_state->crtc.
Contact: Daniel Vetter
Level: Advanced
Convert early atomic drivers to async commit helpers
----------------------------------------------------
For the first year the atomic modeset helpers didn't support asynchronous /
nonblocking commits, and every driver had to hand-roll them. This is fixed
now, but there's still a pile of existing drivers that easily could be
converted over to the new infrastructure.
One issue with the helpers is that they require that drivers handle completion
events for atomic commits correctly. But fixing these bugs is good anyway.
Somewhat related is the legacy_cursor_update hack, which should be replaced with
the new atomic_async_check/commit functionality in the helpers in drivers that
still look at that flag.
Contact: Daniel Vetter, respective driver maintainers
Level: Advanced
Fallout from atomic KMS
-----------------------
``drm_atomic_helper.c`` provides a batch of functions which implement legacy
IOCTLs on top of the new atomic driver interface. Which is really nice for
gradual conversion of drivers, but unfortunately the semantic mismatches are
a bit too severe. So there's some follow-up work to adjust the function
interfaces to fix these issues:
* atomic needs the lock acquire context. At the moment that's passed around
implicitly with some horrible hacks, and it's also allocate with
``GFP_NOFAIL`` behind the scenes. All legacy paths need to start allocating
the acquire context explicitly on stack and then also pass it down into
drivers explicitly so that the legacy-on-atomic functions can use them.
Except for some driver code this is done. This task should be finished by
adding WARN_ON(!drm_drv_uses_atomic_modeset) in drm_modeset_lock_all().
* A bunch of the vtable hooks are now in the wrong place: DRM has a split
between core vfunc tables (named ``drm_foo_funcs``), which are used to
implement the userspace ABI. And then there's the optional hooks for the
helper libraries (name ``drm_foo_helper_funcs``), which are purely for
internal use. Some of these hooks should be move from ``_funcs`` to
``_helper_funcs`` since they are not part of the core ABI. There's a
``FIXME`` comment in the kerneldoc for each such case in ``drm_crtc.h``.
Contact: Daniel Vetter
Level: Intermediate
Get rid of dev->struct_mutex from GEM drivers
---------------------------------------------
``dev->struct_mutex`` is the Big DRM Lock from legacy days and infested
everything. Nowadays in modern drivers the only bit where it's mandatory is
serializing GEM buffer object destruction. Which unfortunately means drivers
have to keep track of that lock and either call ``unreference`` or
``unreference_locked`` depending upon context.
Core GEM doesn't have a need for ``struct_mutex`` any more since kernel 4.8,
and there's a GEM object ``free`` callback for any drivers which are
entirely ``struct_mutex`` free.
For drivers that need ``struct_mutex`` it should be replaced with a driver-
private lock. The tricky part is the BO free functions, since those can't
reliably take that lock any more. Instead state needs to be protected with
suitable subordinate locks or some cleanup work pushed to a worker thread. For
performance-critical drivers it might also be better to go with a more
fine-grained per-buffer object and per-context lockings scheme. Currently only
the ``msm`` and `i915` drivers use ``struct_mutex``.
Contact: Daniel Vetter, respective driver maintainers
Level: Advanced
Move Buffer Object Locking to dma_resv_lock()
---------------------------------------------
Many drivers have their own per-object locking scheme, usually using
mutex_lock(). This causes all kinds of trouble for buffer sharing, since
depending which driver is the exporter and importer, the locking hierarchy is
reversed.
To solve this we need one standard per-object locking mechanism, which is
dma_resv_lock(). This lock needs to be called as the outermost lock, with all
other driver specific per-object locks removed. The problem is tha rolling out
the actual change to the locking contract is a flag day, due to struct dma_buf
buffer sharing.
Level: Expert
Convert logging to drm_* functions with drm_device paramater
------------------------------------------------------------
For drivers which could have multiple instances, it is necessary to
differentiate between which is which in the logs. Since DRM_INFO/WARN/ERROR
don't do this, drivers used dev_info/warn/err to make this differentiation. We
now have drm_* variants of the drm print functions, so we can start to convert
those drivers back to using drm-formatted specific log messages.
Before you start this conversion please contact the relevant maintainers to make
sure your work will be merged - not everyone agrees that the DRM dmesg macros
are better.
Contact: Sean Paul, Maintainer of the driver you plan to convert
Level: Starter
Convert drivers to use simple modeset suspend/resume
----------------------------------------------------
Most drivers (except i915 and nouveau) that use
drm_atomic_helper_suspend/resume() can probably be converted to use
drm_mode_config_helper_suspend/resume(). Also there's still open-coded version
of the atomic suspend/resume code in older atomic modeset drivers.
Contact: Maintainer of the driver you plan to convert
Level: Intermediate
Convert drivers to use drm_fbdev_generic_setup()
------------------------------------------------
Most drivers can use drm_fbdev_generic_setup(). Driver have to implement
atomic modesetting and GEM vmap support. Historically, generic fbdev emulation
expected the framebuffer in system memory or system-like memory. By employing
struct dma_buf_map, drivers with frambuffers in I/O memory can be supported
as well.
Contact: Maintainer of the driver you plan to convert
Level: Intermediate
Reimplement functions in drm_fbdev_fb_ops without fbdev
-------------------------------------------------------
A number of callback functions in drm_fbdev_fb_ops could benefit from
being rewritten without dependencies on the fbdev module. Some of the
helpers could further benefit from using struct dma_buf_map instead of
raw pointers.
Contact: Thomas Zimmermann <tzimmermann@suse.de>, Daniel Vetter
Level: Advanced
drm_framebuffer_funcs and drm_mode_config_funcs.fb_create cleanup
-----------------------------------------------------------------
A lot more drivers could be switched over to the drm_gem_framebuffer helpers.
Various hold-ups:
- Need to switch over to the generic dirty tracking code using
drm_atomic_helper_dirtyfb first (e.g. qxl).
- Need to switch to drm_fbdev_generic_setup(), otherwise a lot of the custom fb
setup code can't be deleted.
- Many drivers wrap drm_gem_fb_create() only to check for valid formats. For
atomic drivers we could check for valid formats by calling
drm_plane_check_pixel_format() against all planes, and pass if any plane
supports the format. For non-atomic that's not possible since like the format
list for the primary plane is fake and we'd therefor reject valid formats.
- Many drivers subclass drm_framebuffer, we'd need a embedding compatible
version of the varios drm_gem_fb_create functions. Maybe called
drm_gem_fb_create/_with_dirty/_with_funcs as needed.
Contact: Daniel Vetter
Level: Intermediate
Clean up mmap forwarding
------------------------
A lot of drivers forward gem mmap calls to dma-buf mmap for imported buffers.
And also a lot of them forward dma-buf mmap to the gem mmap implementations.
There's drm_gem_prime_mmap() for this now, but still needs to be rolled out.
Contact: Daniel Vetter
Level: Intermediate
Generic fbdev defio support
---------------------------
The defio support code in the fbdev core has some very specific requirements,
which means drivers need to have a special framebuffer for fbdev. The main
issue is that it uses some fields in struct page itself, which breaks shmem
gem objects (and other things). To support defio, affected drivers require
the use of a shadow buffer, which may add CPU and memory overhead.
Possible solution would be to write our own defio mmap code in the drm fbdev
emulation. It would need to fully wrap the existing mmap ops, forwarding
everything after it has done the write-protect/mkwrite trickery:
- In the drm_fbdev_fb_mmap helper, if we need defio, change the
default page prots to write-protected with something like this::
vma->vm_page_prot = pgprot_wrprotect(vma->vm_page_prot);
- Set the mkwrite and fsync callbacks with similar implementions to the core
fbdev defio stuff. These should all work on plain ptes, they don't actually
require a struct page. uff. These should all work on plain ptes, they don't
actually require a struct page.
- Track the dirty pages in a separate structure (bitfield with one bit per page
should work) to avoid clobbering struct page.
Might be good to also have some igt testcases for this.
Contact: Daniel Vetter, Noralf Tronnes
Level: Advanced
Garbage collect fbdev scrolling acceleration
--------------------------------------------
Scroll acceleration is disabled in fbcon by hard-wiring p->scrollmode =
SCROLL_REDRAW. There's a ton of code this will allow us to remove:
- lots of code in fbcon.c
- a bunch of the hooks in fbcon_ops, maybe the remaining hooks could be called
directly instead of the function table (with a switch on p->rotate)
- fb_copyarea is unused after this, and can be deleted from all drivers
Note that not all acceleration code can be deleted, since clearing and cursor
support is still accelerated, which might be good candidates for further
deletion projects.
Contact: Daniel Vetter
Level: Intermediate
idr_init_base()
---------------
DRM core&drivers uses a lot of idr (integer lookup directories) for mapping
userspace IDs to internal objects, and in most places ID=0 means NULL and hence
is never used. Switching to idr_init_base() for these would make the idr more
efficient.
Contact: Daniel Vetter
Level: Starter
struct drm_gem_object_funcs
---------------------------
GEM objects can now have a function table instead of having the callbacks on the
DRM driver struct. This is now the preferred way. Callbacks in drivers have been
converted, except for struct drm_driver.gem_prime_mmap.
Level: Intermediate
Use DRM_MODESET_LOCK_ALL_* helpers instead of boilerplate
---------------------------------------------------------
For cases where drivers are attempting to grab the modeset locks with a local
acquire context. Replace the boilerplate code surrounding
drm_modeset_lock_all_ctx() with DRM_MODESET_LOCK_ALL_BEGIN() and
DRM_MODESET_LOCK_ALL_END() instead.
This should also be done for all places where drm_modeset_lock_all() is still
used.
As a reference, take a look at the conversions already completed in drm core.
Contact: Sean Paul, respective driver maintainers
Level: Starter
Rename CMA helpers to DMA helpers
---------------------------------
CMA (standing for contiguous memory allocator) is really a bit an accident of
what these were used for first, a much better name would be DMA helpers. In the
text these should even be called coherent DMA memory helpers (so maybe CDM, but
no one knows what that means) since underneath they just use dma_alloc_coherent.
Contact: Laurent Pinchart, Daniel Vetter
Level: Intermediate (mostly because it is a huge tasks without good partial
milestones, not technically itself that challenging)
connector register/unregister fixes
-----------------------------------
- For most connectors it's a no-op to call drm_connector_register/unregister
directly from driver code, drm_dev_register/unregister take care of this
already. We can remove all of them.
- For dp drivers it's a bit more a mess, since we need the connector to be
registered when calling drm_dp_aux_register. Fix this by instead calling
drm_dp_aux_init, and moving the actual registering into a late_register
callback as recommended in the kerneldoc.
Level: Intermediate
Remove load/unload callbacks from all non-DRIVER_LEGACY drivers
---------------------------------------------------------------
The load/unload callbacks in struct &drm_driver are very much midlayers, plus
for historical reasons they get the ordering wrong (and we can't fix that)
between setting up the &drm_driver structure and calling drm_dev_register().
- Rework drivers to no longer use the load/unload callbacks, directly coding the
load/unload sequence into the driver's probe function.
- Once all non-DRIVER_LEGACY drivers are converted, disallow the load/unload
callbacks for all modern drivers.
Contact: Daniel Vetter
Level: Intermediate
Replace drm_detect_hdmi_monitor() with drm_display_info.is_hdmi
---------------------------------------------------------------
Once EDID is parsed, the monitor HDMI support information is available through
drm_display_info.is_hdmi. Many drivers still call drm_detect_hdmi_monitor() to
retrieve the same information, which is less efficient.
Audit each individual driver calling drm_detect_hdmi_monitor() and switch to
drm_display_info.is_hdmi if applicable.
Contact: Laurent Pinchart, respective driver maintainers
Level: Intermediate
Consolidate custom driver modeset properties
--------------------------------------------
Before atomic modeset took place, many drivers where creating their own
properties. Among other things, atomic brought the requirement that custom,
driver specific properties should not be used.
For this task, we aim to introduce core helpers or reuse the existing ones
if available:
A quick, unconfirmed, examples list.
Introduce core helpers:
- audio (amdgpu, intel, gma500, radeon)
- brightness, contrast, etc (armada, nouveau) - overlay only (?)
- broadcast rgb (gma500, intel)
- colorkey (armada, nouveau, rcar) - overlay only (?)
- dither (amdgpu, nouveau, radeon) - varies across drivers
- underscan family (amdgpu, radeon, nouveau)
Already in core:
- colorspace (sti)
- tv format names, enhancements (gma500, intel)
- tv overscan, margins, etc. (gma500, intel)
- zorder (omapdrm) - same as zpos (?)
Contact: Emil Velikov, respective driver maintainers
Level: Intermediate
Plumb drm_atomic_state all over
-------------------------------
Currently various atomic functions take just a single or a handful of
object states (eg. plane state). While that single object state can
suffice for some simple cases, we often have to dig out additional
object states for dealing with various dependencies between the individual
objects or the hardware they represent. The process of digging out the
additional states is rather non-intuitive and error prone.
To fix that most functions should rather take the overall
drm_atomic_state as one of their parameters. The other parameters
would generally be the object(s) we mainly want to interact with.
For example, instead of
.. code-block:: c
int (*atomic_check)(struct drm_plane *plane, struct drm_plane_state *state);
we would have something like
.. code-block:: c
int (*atomic_check)(struct drm_plane *plane, struct drm_atomic_state *state);
The implementation can then trivially gain access to any required object
state(s) via drm_atomic_get_plane_state(), drm_atomic_get_new_plane_state(),
drm_atomic_get_old_plane_state(), and their equivalents for
other object types.
Additionally many drivers currently access the object->state pointer
directly in their commit functions. That is not going to work if we
eg. want to allow deeper commit pipelines as those pointers could
then point to the states corresponding to a future commit instead of
the current commit we're trying to process. Also non-blocking commits
execute locklessly so there are serious concerns with dereferencing
the object->state pointers without holding the locks that protect them.
Use of drm_atomic_get_new_plane_state(), drm_atomic_get_old_plane_state(),
etc. avoids these problems as well since they relate to a specific
commit via the passed in drm_atomic_state.
Contact: Ville Syrjälä, Daniel Vetter
Level: Intermediate
Use struct dma_buf_map throughout codebase
------------------------------------------
Pointers to shared device memory are stored in struct dma_buf_map. Each
instance knows whether it refers to system or I/O memory. Most of the DRM-wide
interface have been converted to use struct dma_buf_map, but implementations
often still use raw pointers.
The task is to use struct dma_buf_map where it makes sense.
* Memory managers should use struct dma_buf_map for dma-buf-imported buffers.
* TTM might benefit from using struct dma_buf_map internally.
* Framebuffer copying and blitting helpers should operate on struct dma_buf_map.
Contact: Thomas Zimmermann <tzimmermann@suse.de>, Christian König, Daniel Vetter
Level: Intermediate
Core refactorings
=================
Make panic handling work
------------------------
This is a really varied tasks with lots of little bits and pieces:
* The panic path can't be tested currently, leading to constant breaking. The
main issue here is that panics can be triggered from hardirq contexts and
hence all panic related callback can run in hardirq context. It would be
awesome if we could test at least the fbdev helper code and driver code by
e.g. trigger calls through drm debugfs files. hardirq context could be
achieved by using an IPI to the local processor.
* There's a massive confusion of different panic handlers. DRM fbdev emulation
helpers have one, but on top of that the fbcon code itself also has one. We
need to make sure that they stop fighting over each another.
* ``drm_can_sleep()`` is a mess. It hides real bugs in normal operations and
isn't a full solution for panic paths. We need to make sure that it only
returns true if there's a panic going on for real, and fix up all the
fallout.
* The panic handler must never sleep, which also means it can't ever
``mutex_lock()``. Also it can't grab any other lock unconditionally, not
even spinlocks (because NMI and hardirq can panic too). We need to either
make sure to not call such paths, or trylock everything. Really tricky.
* For the above locking troubles reasons it's pretty much impossible to
attempt a synchronous modeset from panic handlers. The only thing we could
try to achive is an atomic ``set_base`` of the primary plane, and hope that
it shows up. Everything else probably needs to be delayed to some worker or
something else which happens later on. Otherwise it just kills the box
harder, prevent the panic from going out on e.g. netconsole.
* There's also proposal for a simplied DRM console instead of the full-blown
fbcon and DRM fbdev emulation. Any kind of panic handling tricks should
obviously work for both console, in case we ever get kmslog merged.
Contact: Daniel Vetter
Level: Advanced
Clean up the debugfs support
----------------------------
There's a bunch of issues with it:
- The drm_info_list ->show() function doesn't even bother to cast to the drm
structure for you. This is lazy.
- We probably want to have some support for debugfs files on crtc/connectors and
maybe other kms objects directly in core. There's even drm_print support in
the funcs for these objects to dump kms state, so it's all there. And then the
->show() functions should obviously give you a pointer to the right object.
- The drm_info_list stuff is centered on drm_minor instead of drm_device. For
anything we want to print drm_device (or maybe drm_file) is the right thing.
- The drm_driver->debugfs_init hooks we have is just an artifact of the old
midlayered load sequence. DRM debugfs should work more like sysfs, where you
can create properties/files for an object anytime you want, and the core
takes care of publishing/unpuplishing all the files at register/unregister
time. Drivers shouldn't need to worry about these technicalities, and fixing
this (together with the drm_minor->drm_device move) would allow us to remove
debugfs_init.
Contact: Daniel Vetter
Level: Intermediate
KMS cleanups
------------
Some of these date from the very introduction of KMS in 2008 ...
- Make ->funcs and ->helper_private vtables optional. There's a bunch of empty
function tables in drivers, but before we can remove them we need to make sure
that all the users in helpers and drivers do correctly check for a NULL
vtable.
- Cleanup up the various ->destroy callbacks. A lot of them just wrapt the
drm_*_cleanup implementations and can be removed. Some tack a kfree() at the
end, for which we could add drm_*_cleanup_kfree(). And then there's the (for
historical reasons) misnamed drm_primary_helper_destroy() function.
Level: Intermediate
Better Testing
==============
Enable trinity for DRM
----------------------
And fix up the fallout. Should be really interesting ...
Level: Advanced
Make KMS tests in i-g-t generic
-------------------------------
The i915 driver team maintains an extensive testsuite for the i915 DRM driver,
including tons of testcases for corner-cases in the modesetting API. It would
be awesome if those tests (at least the ones not relying on Intel-specific GEM
features) could be made to run on any KMS driver.
Basic work to run i-g-t tests on non-i915 is done, what's now missing is mass-
converting things over. For modeset tests we also first need a bit of
infrastructure to use dumb buffers for untiled buffers, to be able to run all
the non-i915 specific modeset tests.
Level: Advanced
Extend virtual test driver (VKMS)
---------------------------------
See the documentation of :ref:`VKMS <vkms>` for more details. This is an ideal
internship task, since it only requires a virtual machine and can be sized to
fit the available time.
Contact: Daniel Vetter
Level: See details
Backlight Refactoring
---------------------
Backlight drivers have a triple enable/disable state, which is a bit overkill.
Plan to fix this:
1. Roll out backlight_enable() and backlight_disable() helpers everywhere. This
has started already.
2. In all, only look at one of the three status bits set by the above helpers.
3. Remove the other two status bits.
Contact: Daniel Vetter
Level: Intermediate
Driver Specific
===============
AMD DC Display Driver
---------------------
AMD DC is the display driver for AMD devices starting with Vega. There has been
a bunch of progress cleaning it up but there's still plenty of work to be done.
See drivers/gpu/drm/amd/display/TODO for tasks.
Contact: Harry Wentland, Alex Deucher
Bootsplash
==========
There is support in place now for writing internal DRM clients making it
possible to pick up the bootsplash work that was rejected because it was written
for fbdev.
- [v6,8/8] drm/client: Hack: Add bootsplash example
https://patchwork.freedesktop.org/patch/306579/
- [RFC PATCH v2 00/13] Kernel based bootsplash
https://lore.kernel.org/r/20171213194755.3409-1-mstaudt@suse.de
Contact: Sam Ravnborg
Level: Advanced
Outside DRM
===========
Convert fbdev drivers to DRM
----------------------------
There are plenty of fbdev drivers for older hardware. Some hwardware has
become obsolete, but some still provides good(-enough) framebuffers. The
drivers that are still useful should be converted to DRM and afterwards
removed from fbdev.
Very simple fbdev drivers can best be converted by starting with a new
DRM driver. Simple KMS helpers and SHMEM should be able to handle any
existing hardware. The new driver's call-back functions are filled from
existing fbdev code.
More complex fbdev drivers can be refactored step-by-step into a DRM
driver with the help of the DRM fbconv helpers. [1] These helpers provide
the transition layer between the DRM core infrastructure and the fbdev
driver interface. Create a new DRM driver on top of the fbconv helpers,
copy over the fbdev driver, and hook it up to the DRM code. Examples for
several fbdev drivers are available at [1] and a tutorial of this process
available at [2]. The result is a primitive DRM driver that can run X11
and Weston.
- [1] https://gitlab.freedesktop.org/tzimmermann/linux/tree/fbconv
- [2] https://gitlab.freedesktop.org/tzimmermann/linux/blob/fbconv/drivers/gpu/drm/drm_fbconv_helper.c
Contact: Thomas Zimmermann <tzimmermann@suse.de>
Level: Advanced