xref: /openbmc/linux/Documentation/gpu/todo.rst (revision 50df3be7)
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
13Remove custom dumb_map_offset implementations
14---------------------------------------------
15
16All GEM based drivers should be using drm_gem_create_mmap_offset() instead.
17Audit each individual driver, make sure it'll work with the generic
18implementation (there's lots of outdated locking leftovers in various
19implementations), and then remove it.
20
21Contact: Daniel Vetter, respective driver maintainers
22
23Convert existing KMS drivers to atomic modesetting
24--------------------------------------------------
25
263.19 has the atomic modeset interfaces and helpers, so drivers can now be
27converted over. Modern compositors like Wayland or Surfaceflinger on Android
28really want an atomic modeset interface, so this is all about the bright
29future.
30
31There is a conversion guide for atomic and all you need is a GPU for a
32non-converted driver (again virtual HW drivers for KVM are still all
33suitable).
34
35As part of this drivers also need to convert to universal plane (which means
36exposing primary & cursor as proper plane objects). But that's much easier to
37do by directly using the new atomic helper driver callbacks.
38
39Contact: Daniel Vetter, respective driver maintainers
40
41Clean up the clipped coordination confusion around planes
42---------------------------------------------------------
43
44We have a helper to get this right with drm_plane_helper_check_update(), but
45it's not consistently used. This should be fixed, preferrably in the atomic
46helpers (and drivers then moved over to clipped coordinates). Probably the
47helper should also be moved from drm_plane_helper.c to the atomic helpers, to
48avoid confusion - the other helpers in that file are all deprecated legacy
49helpers.
50
51Contact: Ville Syrjälä, Daniel Vetter, driver maintainers
52
53Convert early atomic drivers to async commit helpers
54----------------------------------------------------
55
56For the first year the atomic modeset helpers didn't support asynchronous /
57nonblocking commits, and every driver had to hand-roll them. This is fixed
58now, but there's still a pile of existing drivers that easily could be
59converted over to the new infrastructure.
60
61One issue with the helpers is that they require that drivers handle completion
62events for atomic commits correctly. But fixing these bugs is good anyway.
63
64Contact: Daniel Vetter, respective driver maintainers
65
66Fallout from atomic KMS
67-----------------------
68
69``drm_atomic_helper.c`` provides a batch of functions which implement legacy
70IOCTLs on top of the new atomic driver interface. Which is really nice for
71gradual conversion of drivers, but unfortunately the semantic mismatches are
72a bit too severe. So there's some follow-up work to adjust the function
73interfaces to fix these issues:
74
75* atomic needs the lock acquire context. At the moment that's passed around
76  implicitly with some horrible hacks, and it's also allocate with
77  ``GFP_NOFAIL`` behind the scenes. All legacy paths need to start allocating
78  the acquire context explicitly on stack and then also pass it down into
79  drivers explicitly so that the legacy-on-atomic functions can use them.
80
81  Except for some driver code this is done. This task should be finished by
82  adding WARN_ON(!drm_drv_uses_atomic_modeset) in drm_modeset_lock_all().
83
84* A bunch of the vtable hooks are now in the wrong place: DRM has a split
85  between core vfunc tables (named ``drm_foo_funcs``), which are used to
86  implement the userspace ABI. And then there's the optional hooks for the
87  helper libraries (name ``drm_foo_helper_funcs``), which are purely for
88  internal use. Some of these hooks should be move from ``_funcs`` to
89  ``_helper_funcs`` since they are not part of the core ABI. There's a
90  ``FIXME`` comment in the kerneldoc for each such case in ``drm_crtc.h``.
91
92Contact: Daniel Vetter
93
94Get rid of dev->struct_mutex from GEM drivers
95---------------------------------------------
96
97``dev->struct_mutex`` is the Big DRM Lock from legacy days and infested
98everything. Nowadays in modern drivers the only bit where it's mandatory is
99serializing GEM buffer object destruction. Which unfortunately means drivers
100have to keep track of that lock and either call ``unreference`` or
101``unreference_locked`` depending upon context.
102
103Core GEM doesn't have a need for ``struct_mutex`` any more since kernel 4.8,
104and there's a ``gem_free_object_unlocked`` callback for any drivers which are
105entirely ``struct_mutex`` free.
106
107For drivers that need ``struct_mutex`` it should be replaced with a driver-
108private lock. The tricky part is the BO free functions, since those can't
109reliably take that lock any more. Instead state needs to be protected with
110suitable subordinate locks or some cleanup work pushed to a worker thread. For
111performance-critical drivers it might also be better to go with a more
112fine-grained per-buffer object and per-context lockings scheme. Currently only the
113``msm`` driver still use ``struct_mutex``.
114
115Contact: Daniel Vetter, respective driver maintainers
116
117Convert instances of dev_info/dev_err/dev_warn to their DRM_DEV_* equivalent
118----------------------------------------------------------------------------
119
120For drivers which could have multiple instances, it is necessary to
121differentiate between which is which in the logs. Since DRM_INFO/WARN/ERROR
122don't do this, drivers used dev_info/warn/err to make this differentiation. We
123now have DRM_DEV_* variants of the drm print macros, so we can start to convert
124those drivers back to using drm-formwatted specific log messages.
125
126Before you start this conversion please contact the relevant maintainers to make
127sure your work will be merged - not everyone agrees that the DRM dmesg macros
128are better.
129
130Contact: Sean Paul, Maintainer of the driver you plan to convert
131
132Convert drivers to use simple modeset suspend/resume
133----------------------------------------------------
134
135Most drivers (except i915 and nouveau) that use
136drm_atomic_helper_suspend/resume() can probably be converted to use
137drm_mode_config_helper_suspend/resume(). Also there's still open-coded version
138of the atomic suspend/resume code in older atomic modeset drivers.
139
140Contact: Maintainer of the driver you plan to convert
141
142Convert drivers to use drm_fb_helper_fbdev_setup/teardown()
143-----------------------------------------------------------
144
145Most drivers can use drm_fb_helper_fbdev_setup() except maybe:
146
147- amdgpu which has special logic to decide whether to call
148  drm_helper_disable_unused_functions()
149
150- armada which isn't atomic and doesn't call
151  drm_helper_disable_unused_functions()
152
153- i915 which calls drm_fb_helper_initial_config() in a worker
154
155Drivers that use drm_framebuffer_remove() to clean up the fbdev framebuffer can
156probably use drm_fb_helper_fbdev_teardown().
157
158Contact: Maintainer of the driver you plan to convert
159
160Clean up mmap forwarding
161------------------------
162
163A lot of drivers forward gem mmap calls to dma-buf mmap for imported buffers.
164And also a lot of them forward dma-buf mmap to the gem mmap implementations.
165There's drm_gem_prime_mmap() for this now, but still needs to be rolled out.
166
167Contact: Daniel Vetter
168
169Generic fbdev defio support
170---------------------------
171
172The defio support code in the fbdev core has some very specific requirements,
173which means drivers need to have a special framebuffer for fbdev. Which prevents
174us from using the generic fbdev emulation code everywhere. The main issue is
175that it uses some fields in struct page itself, which breaks shmem gem objects
176(and other things).
177
178Possible solution would be to write our own defio mmap code in the drm fbdev
179emulation. It would need to fully wrap the existing mmap ops, forwarding
180everything after it has done the write-protect/mkwrite trickery:
181
182- In the drm_fbdev_fb_mmap helper, if we need defio, change the
183  default page prots to write-protected with something like this::
184
185      vma->vm_page_prot = pgprot_wrprotect(vma->vm_page_prot);
186
187- Set the mkwrite and fsync callbacks with similar implementions to the core
188  fbdev defio stuff. These should all work on plain ptes, they don't actually
189  require a struct page.  uff. These should all work on plain ptes, they don't
190  actually require a struct page.
191
192- Track the dirty pages in a separate structure (bitfield with one bit per page
193  should work) to avoid clobbering struct page.
194
195Might be good to also have some igt testcases for this.
196
197Contact: Daniel Vetter, Noralf Tronnes
198
199idr_init_base()
200---------------
201
202DRM core&drivers uses a lot of idr (integer lookup directories) for mapping
203userspace IDs to internal objects, and in most places ID=0 means NULL and hence
204is never used. Switching to idr_init_base() for these would make the idr more
205efficient.
206
207Contact: Daniel Vetter
208
209struct drm_gem_object_funcs
210---------------------------
211
212GEM objects can now have a function table instead of having the callbacks on the
213DRM driver struct. This is now the preferred way and drivers can be moved over.
214
215We also need a 2nd version of the CMA define that doesn't require the
216vmapping to be present (different hook for prime importing). Plus this needs to
217be rolled out to all drivers using their own implementations, too.
218
219Use DRM_MODESET_LOCK_ALL_* helpers instead of boilerplate
220---------------------------------------------------------
221
222For cases where drivers are attempting to grab the modeset locks with a local
223acquire context. Replace the boilerplate code surrounding
224drm_modeset_lock_all_ctx() with DRM_MODESET_LOCK_ALL_BEGIN() and
225DRM_MODESET_LOCK_ALL_END() instead.
226
227This should also be done for all places where drm_modest_lock_all() is still
228used.
229
230As a reference, take a look at the conversions already completed in drm core.
231
232Contact: Sean Paul, respective driver maintainers
233
234Rename CMA helpers to DMA helpers
235---------------------------------
236
237CMA (standing for contiguous memory allocator) is really a bit an accident of
238what these were used for first, a much better name would be DMA helpers. In the
239text these should even be called coherent DMA memory helpers (so maybe CDM, but
240no one knows what that means) since underneath they just use dma_alloc_coherent.
241
242Contact: Laurent Pinchart, Daniel Vetter
243
244Convert direct mode.vrefresh accesses to use drm_mode_vrefresh()
245----------------------------------------------------------------
246
247drm_display_mode.vrefresh isn't guaranteed to be populated. As such, using it
248is risky and has been known to cause div-by-zero bugs. Fortunately, drm core
249has helper which will use mode.vrefresh if it's !0 and will calculate it from
250the timings when it's 0.
251
252Use simple search/replace, or (more fun) cocci to replace instances of direct
253vrefresh access with a call to the helper. Check out
254https://lists.freedesktop.org/archives/dri-devel/2019-January/205186.html for
255inspiration.
256
257Once all instances of vrefresh have been converted, remove vrefresh from
258drm_display_mode to avoid future use.
259
260Contact: Sean Paul
261
262Remove drm_display_mode.hsync
263-----------------------------
264
265We have drm_mode_hsync() to calculate this from hsync_start/end, since drivers
266shouldn't/don't use this, remove this member to avoid any temptations to use it
267in the future. If there is any debug code using drm_display_mode.hsync, convert
268it to use drm_mode_hsync() instead.
269
270Contact: Sean Paul
271
272drm_fb_helper tasks
273-------------------
274
275- drm_fb_helper_restore_fbdev_mode_unlocked() should call restore_fbdev_mode()
276  not the _force variant so it can bail out if there is a master. But first
277  these igt tests need to be fixed: kms_fbcon_fbt@psr and
278  kms_fbcon_fbt@psr-suspend.
279
280- The max connector argument for drm_fb_helper_init() and
281  drm_fb_helper_fbdev_setup() isn't used anymore and can be removed.
282
283- The helper doesn't keep an array of connectors anymore so these can be
284  removed: drm_fb_helper_single_add_all_connectors(),
285  drm_fb_helper_add_one_connector() and drm_fb_helper_remove_one_connector().
286
287Core refactorings
288=================
289
290Clean up the DRM header mess
291----------------------------
292
293The DRM subsystem originally had only one huge global header, ``drmP.h``. This
294is now split up, but many source files still include it. The remaining part of
295the cleanup work here is to replace any ``#include <drm/drmP.h>`` by only the
296headers needed (and fixing up any missing pre-declarations in the headers).
297
298In the end no .c file should need to include ``drmP.h`` anymore.
299
300Contact: Daniel Vetter
301
302Make panic handling work
303------------------------
304
305This is a really varied tasks with lots of little bits and pieces:
306
307* The panic path can't be tested currently, leading to constant breaking. The
308  main issue here is that panics can be triggered from hardirq contexts and
309  hence all panic related callback can run in hardirq context. It would be
310  awesome if we could test at least the fbdev helper code and driver code by
311  e.g. trigger calls through drm debugfs files. hardirq context could be
312  achieved by using an IPI to the local processor.
313
314* There's a massive confusion of different panic handlers. DRM fbdev emulation
315  helpers have one, but on top of that the fbcon code itself also has one. We
316  need to make sure that they stop fighting over each another.
317
318* ``drm_can_sleep()`` is a mess. It hides real bugs in normal operations and
319  isn't a full solution for panic paths. We need to make sure that it only
320  returns true if there's a panic going on for real, and fix up all the
321  fallout.
322
323* The panic handler must never sleep, which also means it can't ever
324  ``mutex_lock()``. Also it can't grab any other lock unconditionally, not
325  even spinlocks (because NMI and hardirq can panic too). We need to either
326  make sure to not call such paths, or trylock everything. Really tricky.
327
328* For the above locking troubles reasons it's pretty much impossible to
329  attempt a synchronous modeset from panic handlers. The only thing we could
330  try to achive is an atomic ``set_base`` of the primary plane, and hope that
331  it shows up. Everything else probably needs to be delayed to some worker or
332  something else which happens later on. Otherwise it just kills the box
333  harder, prevent the panic from going out on e.g. netconsole.
334
335* There's also proposal for a simplied DRM console instead of the full-blown
336  fbcon and DRM fbdev emulation. Any kind of panic handling tricks should
337  obviously work for both console, in case we ever get kmslog merged.
338
339Contact: Daniel Vetter
340
341Clean up the debugfs support
342----------------------------
343
344There's a bunch of issues with it:
345
346- The drm_info_list ->show() function doesn't even bother to cast to the drm
347  structure for you. This is lazy.
348
349- We probably want to have some support for debugfs files on crtc/connectors and
350  maybe other kms objects directly in core. There's even drm_print support in
351  the funcs for these objects to dump kms state, so it's all there. And then the
352  ->show() functions should obviously give you a pointer to the right object.
353
354- The drm_info_list stuff is centered on drm_minor instead of drm_device. For
355  anything we want to print drm_device (or maybe drm_file) is the right thing.
356
357- The drm_driver->debugfs_init hooks we have is just an artifact of the old
358  midlayered load sequence. DRM debugfs should work more like sysfs, where you
359  can create properties/files for an object anytime you want, and the core
360  takes care of publishing/unpuplishing all the files at register/unregister
361  time. Drivers shouldn't need to worry about these technicalities, and fixing
362  this (together with the drm_minor->drm_device move) would allow us to remove
363  debugfs_init.
364
365- Drop the return code and error checking from all debugfs functions. Greg KH is
366  working on this already.
367
368Contact: Daniel Vetter
369
370KMS cleanups
371------------
372
373Some of these date from the very introduction of KMS in 2008 ...
374
375- Make ->funcs and ->helper_private vtables optional. There's a bunch of empty
376  function tables in drivers, but before we can remove them we need to make sure
377  that all the users in helpers and drivers do correctly check for a NULL
378  vtable.
379
380- Cleanup up the various ->destroy callbacks. A lot of them just wrapt the
381  drm_*_cleanup implementations and can be removed. Some tack a kfree() at the
382  end, for which we could add drm_*_cleanup_kfree(). And then there's the (for
383  historical reasons) misnamed drm_primary_helper_destroy() function.
384
385Better Testing
386==============
387
388Enable trinity for DRM
389----------------------
390
391And fix up the fallout. Should be really interesting ...
392
393Make KMS tests in i-g-t generic
394-------------------------------
395
396The i915 driver team maintains an extensive testsuite for the i915 DRM driver,
397including tons of testcases for corner-cases in the modesetting API. It would
398be awesome if those tests (at least the ones not relying on Intel-specific GEM
399features) could be made to run on any KMS driver.
400
401Basic work to run i-g-t tests on non-i915 is done, what's now missing is mass-
402converting things over. For modeset tests we also first need a bit of
403infrastructure to use dumb buffers for untiled buffers, to be able to run all
404the non-i915 specific modeset tests.
405
406Extend virtual test driver (VKMS)
407---------------------------------
408
409See the documentation of :ref:`VKMS <vkms>` for more details. This is an ideal
410internship task, since it only requires a virtual machine and can be sized to
411fit the available time.
412
413Contact: Daniel Vetter
414
415Backlight Refactoring
416---------------------
417
418Backlight drivers have a triple enable/disable state, which is a bit overkill.
419Plan to fix this:
420
4211. Roll out backlight_enable() and backlight_disable() helpers everywhere. This
422   has started already.
4232. In all, only look at one of the three status bits set by the above helpers.
4243. Remove the other two status bits.
425
426Contact: Daniel Vetter
427
428Driver Specific
429===============
430
431AMD DC Display Driver
432---------------------
433
434AMD DC is the display driver for AMD devices starting with Vega. There has been
435a bunch of progress cleaning it up but there's still plenty of work to be done.
436
437See drivers/gpu/drm/amd/display/TODO for tasks.
438
439Contact: Harry Wentland, Alex Deucher
440
441i915
442----
443
444- Our early/late pm callbacks could be removed in favour of using
445  device_link_add to model the dependency between i915 and snd_had. See
446  https://dri.freedesktop.org/docs/drm/driver-api/device_link.html
447
448Bootsplash
449==========
450
451There is support in place now for writing internal DRM clients making it
452possible to pick up the bootsplash work that was rejected because it was written
453for fbdev.
454
455- [v6,8/8] drm/client: Hack: Add bootsplash example
456  https://patchwork.freedesktop.org/patch/306579/
457
458- [RFC PATCH v2 00/13] Kernel based bootsplash
459  https://lkml.org/lkml/2017/12/13/764
460
461Contact: Sam Ravnborg
462
463Outside DRM
464===========
465