xref: /openbmc/linux/Documentation/gpu/todo.rst (revision 4da722ca)
1.. _todo:
2
3=========
4TODO list
5=========
6
7This section contains a list of smaller janitorial tasks in the kernel DRM
8graphics subsystem useful as newbie projects. Or for slow rainy days.
9
10Subsystem-wide refactorings
11===========================
12
13De-midlayer drivers
14-------------------
15
16With the recent ``drm_bus`` cleanup patches for 3.17 it is no longer required
17to have a ``drm_bus`` structure set up. Drivers can directly set up the
18``drm_device`` structure instead of relying on bus methods in ``drm_usb.c``
19and ``drm_pci.c``. The goal is to get rid of the driver's ``->load`` /
20``->unload`` callbacks and open-code the load/unload sequence properly, using
21the new two-stage ``drm_device`` setup/teardown.
22
23Once all existing drivers are converted we can also remove those bus support
24files for USB and platform devices.
25
26All you need is a GPU for a non-converted driver (currently almost all of
27them, but also all the virtual ones used by KVM, so everyone qualifies).
28
29Contact: Daniel Vetter, Thierry Reding, respective driver maintainers
30
31Switch from reference/unreference to get/put
32--------------------------------------------
33
34For some reason DRM core uses ``reference``/``unreference`` suffixes for
35refcounting functions, but kernel uses ``get``/``put`` (e.g.
36``kref_get``/``put()``). It would be good to switch over for consistency, and
37it's shorter. Needs to be done in 3 steps for each pair of functions:
38
39* Create new ``get``/``put`` functions, define the old names as compatibility
40  wrappers
41* Switch over each file/driver using a cocci-generated spatch.
42* Once all users of the old names are gone, remove them.
43
44This way drivers/patches in the progress of getting merged won't break.
45
46Contact: Daniel Vetter
47
48Convert existing KMS drivers to atomic modesetting
49--------------------------------------------------
50
513.19 has the atomic modeset interfaces and helpers, so drivers can now be
52converted over. Modern compositors like Wayland or Surfaceflinger on Android
53really want an atomic modeset interface, so this is all about the bright
54future.
55
56There is a conversion guide for atomic and all you need is a GPU for a
57non-converted driver (again virtual HW drivers for KVM are still all
58suitable).
59
60As part of this drivers also need to convert to universal plane (which means
61exposing primary & cursor as proper plane objects). But that's much easier to
62do by directly using the new atomic helper driver callbacks.
63
64Contact: Daniel Vetter, respective driver maintainers
65
66Clean up the clipped coordination confusion around planes
67---------------------------------------------------------
68
69We have a helper to get this right with drm_plane_helper_check_update(), but
70it's not consistently used. This should be fixed, preferrably in the atomic
71helpers (and drivers then moved over to clipped coordinates). Probably the
72helper should also be moved from drm_plane_helper.c to the atomic helpers, to
73avoid confusion - the other helpers in that file are all deprecated legacy
74helpers.
75
76Contact: Ville Syrjälä, Daniel Vetter, driver maintainers
77
78Implement deferred fbdev setup in the helper
79--------------------------------------------
80
81Many (especially embedded drivers) want to delay fbdev setup until there's a
82real screen plugged in. This is to avoid the dreaded fallback to the low-res
83fbdev default. Many drivers have a hacked-up (and often broken) version of this,
84better to do it once in the shared helpers. Thierry has a patch series, but that
85one needs to be rebased and final polish applied.
86
87Contact: Thierry Reding, Daniel Vetter, driver maintainers
88
89Convert early atomic drivers to async commit helpers
90----------------------------------------------------
91
92For the first year the atomic modeset helpers didn't support asynchronous /
93nonblocking commits, and every driver had to hand-roll them. This is fixed
94now, but there's still a pile of existing drivers that easily could be
95converted over to the new infrastructure.
96
97One issue with the helpers is that they require that drivers handle completion
98events for atomic commits correctly. But fixing these bugs is good anyway.
99
100Contact: Daniel Vetter, respective driver maintainers
101
102Better manual-upload support for atomic
103---------------------------------------
104
105This would be especially useful for tinydrm:
106
107- Add a struct drm_rect dirty_clip to drm_crtc_state. When duplicating the
108  crtc state, clear that to the max values, x/y = 0 and w/h = MAX_INT, in
109  __drm_atomic_helper_crtc_duplicate_state().
110
111- Move tinydrm_merge_clips into drm_framebuffer.c, dropping the tinydrm_
112  prefix ofc and using drm_fb_. drm_framebuffer.c makes sense since this
113  is a function useful to implement the fb->dirty function.
114
115- Create a new drm_fb_dirty function which does essentially what e.g.
116  mipi_dbi_fb_dirty does. You can use e.g. drm_atomic_helper_update_plane as the
117  template. But instead of doing a simple full-screen plane update, this new
118  helper also sets crtc_state->dirty_clip to the right coordinates. And of
119  course it needs to check whether the fb is actually active (and maybe where),
120  so there's some book-keeping involved. There's also some good fun involved in
121  scaling things appropriately. For that case we might simply give up and
122  declare the entire area covered by the plane as dirty.
123
124Contact: Noralf Trønnes, Daniel Vetter
125
126Fallout from atomic KMS
127-----------------------
128
129``drm_atomic_helper.c`` provides a batch of functions which implement legacy
130IOCTLs on top of the new atomic driver interface. Which is really nice for
131gradual conversion of drivers, but unfortunately the semantic mismatches are
132a bit too severe. So there's some follow-up work to adjust the function
133interfaces to fix these issues:
134
135* atomic needs the lock acquire context. At the moment that's passed around
136  implicitly with some horrible hacks, and it's also allocate with
137  ``GFP_NOFAIL`` behind the scenes. All legacy paths need to start allocating
138  the acquire context explicitly on stack and then also pass it down into
139  drivers explicitly so that the legacy-on-atomic functions can use them.
140
141* A bunch of the vtable hooks are now in the wrong place: DRM has a split
142  between core vfunc tables (named ``drm_foo_funcs``), which are used to
143  implement the userspace ABI. And then there's the optional hooks for the
144  helper libraries (name ``drm_foo_helper_funcs``), which are purely for
145  internal use. Some of these hooks should be move from ``_funcs`` to
146  ``_helper_funcs`` since they are not part of the core ABI. There's a
147  ``FIXME`` comment in the kerneldoc for each such case in ``drm_crtc.h``.
148
149* There's a new helper ``drm_atomic_helper_best_encoder()`` which could be
150  used by all atomic drivers which don't select the encoder for a given
151  connector at runtime. That's almost all of them, and would allow us to get
152  rid of a lot of ``best_encoder`` boilerplate in drivers.
153
154Contact: Daniel Vetter
155
156Get rid of dev->struct_mutex from GEM drivers
157---------------------------------------------
158
159``dev->struct_mutex`` is the Big DRM Lock from legacy days and infested
160everything. Nowadays in modern drivers the only bit where it's mandatory is
161serializing GEM buffer object destruction. Which unfortunately means drivers
162have to keep track of that lock and either call ``unreference`` or
163``unreference_locked`` depending upon context.
164
165Core GEM doesn't have a need for ``struct_mutex`` any more since kernel 4.8,
166and there's a ``gem_free_object_unlocked`` callback for any drivers which are
167entirely ``struct_mutex`` free.
168
169For drivers that need ``struct_mutex`` it should be replaced with a driver-
170private lock. The tricky part is the BO free functions, since those can't
171reliably take that lock any more. Instead state needs to be protected with
172suitable subordinate locks or some cleanup work pushed to a worker thread. For
173performance-critical drivers it might also be better to go with a more
174fine-grained per-buffer object and per-context lockings scheme. Currently the
175following drivers still use ``struct_mutex``: ``msm``, ``omapdrm`` and
176``udl``.
177
178Contact: Daniel Vetter, respective driver maintainers
179
180Core refactorings
181=================
182
183Use new IDR deletion interface to clean up drm_gem_handle_delete()
184------------------------------------------------------------------
185
186See the "This is gross" comment -- apparently the IDR system now can return an
187error code instead of oopsing.
188
189Clean up the DRM header mess
190----------------------------
191
192Currently the DRM subsystem has only one global header, ``drmP.h``. This is
193used both for functions exported to helper libraries and drivers and functions
194only used internally in the ``drm.ko`` module. The goal would be to move all
195header declarations not needed outside of ``drm.ko`` into
196``drivers/gpu/drm/drm_*_internal.h`` header files. ``EXPORT_SYMBOL`` also
197needs to be dropped for these functions.
198
199This would nicely tie in with the below task to create kerneldoc after the API
200is cleaned up. Or with the "hide legacy cruft better" task.
201
202Note that this is well in progress, but ``drmP.h`` is still huge. The updated
203plan is to switch to per-file driver API headers, which will also structure
204the kerneldoc better. This should also allow more fine-grained ``#include``
205directives.
206
207In the end no .c file should need to include ``drmP.h`` anymore.
208
209Contact: Daniel Vetter
210
211Add missing kerneldoc for exported functions
212--------------------------------------------
213
214The DRM reference documentation is still lacking kerneldoc in a few areas. The
215task would be to clean up interfaces like moving functions around between
216files to better group them and improving the interfaces like dropping return
217values for functions that never fail. Then write kerneldoc for all exported
218functions and an overview section and integrate it all into the drm book.
219
220See https://dri.freedesktop.org/docs/drm/ for what's there already.
221
222Contact: Daniel Vetter
223
224Hide legacy cruft better
225------------------------
226
227Way back DRM supported only drivers which shadow-attached to PCI devices with
228userspace or fbdev drivers setting up outputs. Modern DRM drivers take charge
229of the entire device, you can spot them with the DRIVER_MODESET flag.
230
231Unfortunately there's still large piles of legacy code around which needs to
232be hidden so that driver writers don't accidentally end up using it. And to
233prevent security issues in those legacy IOCTLs from being exploited on modern
234drivers. This has multiple possible subtasks:
235
236* Extract support code for legacy features into a ``drm-legacy.ko`` kernel
237  module and compile it only when one of the legacy drivers is enabled.
238
239This is mostly done, the only thing left is to split up ``drm_irq.c`` into
240legacy cruft and the parts needed by modern KMS drivers.
241
242Contact: Daniel Vetter
243
244Make panic handling work
245------------------------
246
247This is a really varied tasks with lots of little bits and pieces:
248
249* The panic path can't be tested currently, leading to constant breaking. The
250  main issue here is that panics can be triggered from hardirq contexts and
251  hence all panic related callback can run in hardirq context. It would be
252  awesome if we could test at least the fbdev helper code and driver code by
253  e.g. trigger calls through drm debugfs files. hardirq context could be
254  achieved by using an IPI to the local processor.
255
256* There's a massive confusion of different panic handlers. DRM fbdev emulation
257  helpers have one, but on top of that the fbcon code itself also has one. We
258  need to make sure that they stop fighting over each another.
259
260* ``drm_can_sleep()`` is a mess. It hides real bugs in normal operations and
261  isn't a full solution for panic paths. We need to make sure that it only
262  returns true if there's a panic going on for real, and fix up all the
263  fallout.
264
265* The panic handler must never sleep, which also means it can't ever
266  ``mutex_lock()``. Also it can't grab any other lock unconditionally, not
267  even spinlocks (because NMI and hardirq can panic too). We need to either
268  make sure to not call such paths, or trylock everything. Really tricky.
269
270* For the above locking troubles reasons it's pretty much impossible to
271  attempt a synchronous modeset from panic handlers. The only thing we could
272  try to achive is an atomic ``set_base`` of the primary plane, and hope that
273  it shows up. Everything else probably needs to be delayed to some worker or
274  something else which happens later on. Otherwise it just kills the box
275  harder, prevent the panic from going out on e.g. netconsole.
276
277* There's also proposal for a simplied DRM console instead of the full-blown
278  fbcon and DRM fbdev emulation. Any kind of panic handling tricks should
279  obviously work for both console, in case we ever get kmslog merged.
280
281Contact: Daniel Vetter
282
283Clean up the debugfs support
284----------------------------
285
286There's a bunch of issues with it:
287
288- The drm_info_list ->show() function doesn't even bother to cast to the drm
289  structure for you. This is lazy.
290
291- We probably want to have some support for debugfs files on crtc/connectors and
292  maybe other kms objects directly in core. There's even drm_print support in
293  the funcs for these objects to dump kms state, so it's all there. And then the
294  ->show() functions should obviously give you a pointer to the right object.
295
296- The drm_info_list stuff is centered on drm_minor instead of drm_device. For
297  anything we want to print drm_device (or maybe drm_file) is the right thing.
298
299- The drm_driver->debugfs_init hooks we have is just an artifact of the old
300  midlayered load sequence. DRM debugfs should work more like sysfs, where you
301  can create properties/files for an object anytime you want, and the core
302  takes care of publishing/unpuplishing all the files at register/unregister
303  time. Drivers shouldn't need to worry about these technicalities, and fixing
304  this (together with the drm_minor->drm_device move) would allow us to remove
305  debugfs_init.
306
307Contact: Daniel Vetter
308
309Better Testing
310==============
311
312Enable trinity for DRM
313----------------------
314
315And fix up the fallout. Should be really interesting ...
316
317Make KMS tests in i-g-t generic
318-------------------------------
319
320The i915 driver team maintains an extensive testsuite for the i915 DRM driver,
321including tons of testcases for corner-cases in the modesetting API. It would
322be awesome if those tests (at least the ones not relying on Intel-specific GEM
323features) could be made to run on any KMS driver.
324
325Basic work to run i-g-t tests on non-i915 is done, what's now missing is mass-
326converting things over. For modeset tests we also first need a bit of
327infrastructure to use dumb buffers for untiled buffers, to be able to run all
328the non-i915 specific modeset tests.
329
330Contact: Daniel Vetter
331
332Create a virtual KMS driver for testing (vkms)
333----------------------------------------------
334
335With all the latest helpers it should be fairly simple to create a virtual KMS
336driver useful for testing, or for running X or similar on headless machines
337(to be able to still use the GPU). This would be similar to vgem, but aimed at
338the modeset side.
339
340Once the basics are there there's tons of possibilities to extend it.
341
342Contact: Daniel Vetter
343
344Driver Specific
345===============
346
347tinydrm
348-------
349
350Tinydrm is the helper driver for really simple fb drivers. The goal is to make
351those drivers as simple as possible, so lots of room for refactoring:
352
353- backlight helpers, probably best to put them into a new drm_backlight.c.
354  This is because drivers/video is de-facto unmaintained. We could also
355  move drivers/video/backlight to drivers/gpu/backlight and take it all
356  over within drm-misc, but that's more work.
357
358- spi helpers, probably best put into spi core/helper code. Thierry said
359  the spi maintainer is fast&reactive, so shouldn't be a big issue.
360
361- extract the mipi-dbi helper (well, the non-tinydrm specific parts at
362  least) into a separate helper, like we have for mipi-dsi already. Or follow
363  one of the ideas for having a shared dsi/dbi helper, abstracting away the
364  transport details more.
365
366- tinydrm_lastclose could be drm_fb_helper_lastclose. Only thing we need
367  for that is to store the drm_fb_helper pointer somewhere in
368  drm_device->mode_config. And then we could roll that out to all the
369  drivers.
370
371- tinydrm_gem_cma_prime_import_sg_table should probably go into the cma
372  helpers, as a _vmapped variant (since not every driver needs the vmap).
373  And tinydrm_gem_cma_free_object could the be merged into
374  drm_gem_cma_free_object().
375
376- tinydrm_fb_create we could move into drm_simple_pipe, only need to add
377  the fb_create hook to drm_simple_pipe_funcs, which would again simplify a
378  bunch of things (since it gives you a one-stop vfunc for simple drivers).
379
380- Quick aside: The unregister devm stuff is kinda getting the lifetimes of
381  a drm_device wrong. Doesn't matter, since everyone else gets it wrong
382  too :-)
383
384- With the fbdev pointer in dev->mode_config we could also make
385  suspend/resume helpers entirely generic, at least if we add a
386  dev->mode_config.suspend_state. We could even provide a generic pm_ops
387  structure with those.
388
389- also rework the drm_framebuffer_funcs->dirty hook wire-up, see above.
390
391Contact: Noralf Trønnes, Daniel Vetter
392
393Outside DRM
394===========
395