1 .. Copyright 2020 DisplayLink (UK) Ltd.
2
3 ===================
4 Userland interfaces
5 ===================
6
7 The DRM core exports several interfaces to app
8 intended to be used through corresponding libd
9 addition, drivers export device-specific inter
10 drivers & device-aware applications through io
11
12 External interfaces include: memory mapping, c
13 operations, AGP management, vblank control, fe
14 management, and output management.
15
16 Cover generic ioctls and sysfs layout here. We
17 info, since man pages should cover the rest.
18
19 libdrm Device Lookup
20 ====================
21
22 .. kernel-doc:: drivers/gpu/drm/drm_ioctl.c
23 :doc: getunique and setversion story
24
25
26 .. _drm_primary_node:
27
28 Primary Nodes, DRM Master and Authentication
29 ============================================
30
31 .. kernel-doc:: drivers/gpu/drm/drm_auth.c
32 :doc: master and authentication
33
34 .. kernel-doc:: drivers/gpu/drm/drm_auth.c
35 :export:
36
37 .. kernel-doc:: include/drm/drm_auth.h
38 :internal:
39
40
41 .. _drm_leasing:
42
43 DRM Display Resource Leasing
44 ============================
45
46 .. kernel-doc:: drivers/gpu/drm/drm_lease.c
47 :doc: drm leasing
48
49 Open-Source Userspace Requirements
50 ==================================
51
52 The DRM subsystem has stricter requirements th
53 what the userspace side for new uAPI needs to
54 explains what exactly those requirements are,
55
56 The short summary is that any addition of DRM
57 open-sourced userspace patches, and those patc
58 merging into a suitable and canonical upstream
59
60 GFX devices (both display and render/GPU side)
61 hardware, with userspace and kernel by necessi
62 closely. The interfaces, for rendering and mo
63 and flexible, and therefore it is almost alway
64 them for every possible corner case. This in t
65 infeasible to differentiate between behaviour
66 which must not be changed to avoid regressions
67 accidental artifact of the current implementat
68
69 Without access to the full source code of all
70 becomes impossible to change the implementatio
71 depend upon the accidental behaviour of the cu
72 details. And debugging such regressions withou
73 much impossible. As a consequence this means:
74
75 - The Linux kernel's "no regression" policy ho
76 open-source userspace of the DRM subsystem.
77 if closed-source blob drivers in userspace u
78 drivers, but they must do so in the exact sa
79 Creative (ab)use of the interfaces will, and
80 to breakage.
81
82 - Any new userspace interface must have an ope
83 demonstration vehicle.
84
85 The other reason for requiring open-source use
86 kernel and userspace parts of a GFX stack must
87 review can only assess whether a new interface
88 both sides. Making sure that the interface ind
89 leads to a few additional requirements:
90
91 - The open-source userspace must not be a toy/
92 thing. Specifically it needs to handle all t
93 These are often the places where new uAPI fa
94 assess the fitness of a proposed interface.
95
96 - The userspace side must be fully reviewed an
97 userspace project. For e.g. mesa this means
98 mailing list. This is again to ensure that t
99 job done. The userspace-side reviewer shoul
100 kernel uAPI patch indicating that they belie
101 sufficiently documented and validated for us
102
103 - The userspace patches must be against the ca
104 fork. This is to make sure that no one cheat
105 requirements by doing a quick fork.
106
107 - The kernel patch can only be merged after al
108 but it **must** be merged to either drm-next
109 userspace patches land. uAPI always flows fr
110 other way round risks divergence of the uAPI
111
112 These are fairly steep requirements, but have
113 pain and experience with uAPI added hastily, a
114 just as fast. GFX devices change really fast,
115 entire new set of uAPI interfaces every few ye
116 Linux kernel's guarantee to keep existing user
117 is already rather painful for the DRM subsyste
118 for the same thing co-existing. If we add a fe
119 mix every year it would be entirely unmanageab
120
121 .. _drm_render_node:
122
123 Render nodes
124 ============
125
126 DRM core provides multiple character-devices f
127 Depending on which device is opened, user-spac
128 set of operations (mainly ioctls). The primary
129 and called card<num>. Additionally, a currentl
130 called controlD<num> is also created. The prim
131 legacy operations and historically was the onl
132 userspace. With KMS, the control node was intr
133 planned KMS control interface has never been w
134 node stays unused to date.
135
136 With the increased use of offscreen renderers
137 clients no longer require running compositors
138 make use of a GPU. But the DRM API required un
139 authenticate to a DRM-Master prior to getting
140 step and to grant clients GPU access without a
141 nodes were introduced. Render nodes solely ser
142 is, no modesetting or privileged ioctls can be
143 Only non-global rendering commands are allowed
144 render nodes, it must advertise it via the DRI
145 capability. If not supported, the primary node
146 clients together with the legacy drmAuth authe
147
148 If a driver advertises render node support, DR
149 separate render node called renderD<num>. Ther
150 per device. No ioctls except PRIME-related ioc
151 this node. Especially GEM_OPEN will be explici
152 complete list of driver-independent ioctls tha
153 nodes, see the ioctls marked DRM_RENDER_ALLOW
154 nodes are designed to avoid the buffer-leaks,
155 guess the flink names or mmap offsets on the l
156 Additionally to this basic interface, drivers
157 driver-dependent render-only ioctls as DRM_REN
158 clients can use them. Driver authors must be c
159 privileged ioctls on render nodes.
160
161 With render nodes, user-space can now control
162 via basic file-system access-modes. A running
163 authenticates clients on the privileged primar
164 required. Instead, a client can open the rende
165 granted GPU access. Communication between clie
166 via PRIME. FLINK from render node to legacy no
167 clients must not use the insecure FLINK interf
168
169 Besides dropping all modeset/global ioctls, re
170 DRM-Master concept. There is no reason to asso
171 a DRM-Master as they are independent of any gr
172 they must work without any running master, any
173 to run without a master object if they support
174 other hand, a driver requires shared state bet
175 visible to user-space and accessible beyond op
176 cannot support render nodes.
177
178 Device Hot-Unplug
179 =================
180
181 .. note::
182 The following is the plan. Implementation i
183 (2020 May).
184
185 Graphics devices (display and/or render) may b
186 display adapters or docking stations) or Thund
187 user is able to hot-unplug this kind of device
188 used, and expects that the very least the mach
189 damage from hot-unplugging a DRM device needs
190 possible and userspace must be given the chanc
191 to. Ideally, unplugging a DRM device still let
192 run, but that is going to need explicit suppor
193 graphics stack: from kernel and userspace driv
194 servers, via window system protocols, and in a
195
196 Other scenarios that should lead to the same a
197 crash, PCI device disappearing off the bus, or
198 from the physical device.
199
200 In other words, from userspace perspective eve
201 working more or less, until userspace stops us
202 device and closes it completely. Userspace wil
203 disappearance from the device removed uevent,
204 (or driver-specific ioctls returning driver-sp
205 returning ENXIO.
206
207 Only after userspace has closed all relevant D
208 descriptors and removed all mmaps, the DRM dri
209 instance for the device that no longer exists.
210 device somehow comes back in the mean time, it
211 device.
212
213 Similar to PIDs, chardev minor numbers are not
214 new DRM device always picks the next free mino
215 previous one allocated, and wraps around when
216 exhausted.
217
218 The goal raises at least the following require
219 drivers.
220
221 Requirements for KMS UAPI
222 -------------------------
223
224 - KMS connectors must change their status to d
225
226 - Legacy modesets and pageflips, and atomic co
227 TEST_ONLY, and any other ioctls either fail
228 success.
229
230 - Pending non-blocking KMS operations deliver
231 is expecting. This applies also to ioctls th
232
233 - open() on a device node whose underlying dev
234 fail with ENXIO.
235
236 - Attempting to create a DRM lease on a disapp
237 fail with ENODEV. Existing DRM leases remain
238 above.
239
240 Requirements for Render and Cross-Device UAPI
241 ---------------------------------------------
242
243 - All GPU jobs that can no longer run must hav
244 force-signalled to avoid inflicting hangs on
245 The associated error code is ENODEV.
246
247 - Some userspace APIs already define what shou
248 disappears (OpenGL, GL ES: `GL_KHR_robustnes
249 VK_ERROR_DEVICE_LOST; etc.). DRM drivers are
250 behaviour the way they see best, e.g. return
251 driver-specific ioctls and handling those in
252 rely on uevents, and so on.
253
254 - dmabuf which point to memory that has disapp
255 import with ENODEV or continue to be success
256 have succeeded before the disappearance. See
257 below for already imported dmabufs.
258
259 - Attempting to import a dmabuf to a disappear
260 with ENODEV or succeed if it would have succ
261 disappearance.
262
263 - open() on a device node whose underlying dev
264 fail with ENXIO.
265
266 .. _GL_KHR_robustness: https://www.khronos.org
267 .. _Vulkan: https://www.khronos.org/vulkan/
268
269 Requirements for Memory Maps
270 ----------------------------
271
272 Memory maps have further requirements that app
273 and maps created after the device has disappea
274 memory disappears, the map is created or modif
275 writes will still complete successfully but th
276 This applies to both userspace mmap()'d memory
277 dmabuf which might be mapped to other devices
278 imports).
279
280 Raising SIGBUS is not an option, because users
281 handle it. Signal handlers are global, which m
282 difficult to use correctly from libraries like
283 Signal handlers are not composable, you can't
284 for GPU1 and GPU2 from different vendors, and
285 mmapped regular files. Threads cause additiona
286 handling as well.
287
288 Device reset
289 ============
290
291 The GPU stack is really complex and is prone t
292 faulty applications and everything in between
293 require resetting the device in order to make
294 section describes the expectations for DRM and
295 device resets and how to propagate the reset s
296
297 Device resets can not be disabled without tain
298 hanging the entire kernel through shrinkers/mm
299 device resets is to propagate the message to t
300 special policy for blocking guilty application
301 debugging a hung GPU context require hardware
302 a GPU context while it's stopped.
303
304 Kernel Mode Driver
305 ------------------
306
307 The KMD is responsible for checking if the dev
308 it as needed. Usually a hang is detected when
309 should keep track of resets, because userspace
310 reset status for a specific context. This is n
311 the stack that a reset has happened. Currently
312 driver separately, with no common DRM interfac
313 integrated at DRM scheduler to provide a commo
314 reset, KMD should reject new command submissio
315
316 User Mode Driver
317 ----------------
318
319 After command submission, UMD should check if
320 rejected. After a reset, KMD should reject sub
321 ioctl to the KMD to check the reset status, an
322 if the UMD requires it. After detecting a rese
323 it to the application using the appropriate AP
324 section below about robustness.
325
326 Robustness
327 ----------
328
329 The only way to try to keep a graphical API co
330 it complies with the robustness aspects of the
331
332 Graphical APIs provide ways to applications to
333 there is no guarantee that the app will use su
334 userspace that doesn't support robust interfac
335 OpenGL context or API without any robustness s
336 robustness handling entirely to the userspace
337 community consensus on what the userspace driv
338 since all reasonable approaches have some clea
339
340 OpenGL
341 ~~~~~~
342
343 Apps using OpenGL should use the available rob
344 extension ``GL_ARB_robustness`` (or ``GL_EXT_r
345 interface tells if a reset has happened, and i
346 considered lost and the app proceeds by creati
347 on what to do to if robustness is not in use.
348
349 Vulkan
350 ~~~~~~
351
352 Apps using Vulkan should check for ``VK_ERROR_
353 This error code means, among other things, tha
354 it needs to recreate the contexts to keep goin
355
356 Reporting causes of resets
357 --------------------------
358
359 Apart from propagating the reset through the s
360 really useful for driver developers to learn m
361 the first place. DRM devices should make use o
362 information about the reset, so this informati
363 reports.
364
365 .. _drm_driver_ioctl:
366
367 IOCTL Support on Device Nodes
368 =============================
369
370 .. kernel-doc:: drivers/gpu/drm/drm_ioctl.c
371 :doc: driver specific ioctls
372
373 Recommended IOCTL Return Values
374 -------------------------------
375
376 In theory a driver's IOCTL callback is only al
377 codes. In practice it's good to abuse a few mo
378 practice within the DRM subsystem:
379
380 ENOENT:
381 Strictly this should only be used when
382 calling the open() syscall. We reuse t
383 lookup failure, e.g. for unknown GEM b
384 object handles and similar cases.
385
386 ENOSPC:
387 Some drivers use this to differentiate
388 of VRAM". Sometimes also applies to ot
389 rendering (e.g. when you have a specia
390 Sometimes resource allocation/reservat
391 IOCTLs are also signalled through EDEA
392
393 Simply running out of kernel/system me
394
395 EPERM/EACCES:
396 Returned for an operation that is vali
397 E.g. root-only or much more common, DR
398 this when called by unpriviledged clie
399 difference between EACCES and EPERM.
400
401 ENODEV:
402 The device is not present anymore or i
403
404 EOPNOTSUPP:
405 Feature (like PRIME, modesetting, GEM)
406
407 ENXIO:
408 Remote failure, either a hardware tran
409 when the exporting driver of a shared
410 feature needed.
411
412 EINTR:
413 DRM drivers assume that userspace rest
414 return EINTR and in such a case should
415 parameters left unchanged.
416
417 EIO:
418 The GPU died and couldn't be resurrect
419 hardware failures are signalled throug
420 property.
421
422 EINVAL:
423 Catch-all for anything that is an inva
424 cannot work.
425
426 IOCTL also use other error codes like ETIME, E
427 usage is in line with the common meanings. The
428 DRM specific patterns. Note that ENOTTY has th
429 "this IOCTL does not exist", and is used exact
430
431 .. kernel-doc:: include/drm/drm_ioctl.h
432 :internal:
433
434 .. kernel-doc:: drivers/gpu/drm/drm_ioctl.c
435 :export:
436
437 .. kernel-doc:: drivers/gpu/drm/drm_ioc32.c
438 :export:
439
440 Testing and validation
441 ======================
442
443 Testing Requirements for userspace API
444 --------------------------------------
445
446 New cross-driver userspace interface extension
447 properties, new files in sysfs or anything els
448 should have driver-agnostic testcases in IGT f
449 can be reasonably made using IGT for the targe
450
451 Validating changes with IGT
452 ---------------------------
453
454 There's a collection of tests that aims to cov
455 DRM drivers and that can be used to check that
456 core don't regress existing functionality. Thi
457 its code and instructions to build and run can
458 https://gitlab.freedesktop.org/drm/igt-gpu-too
459
460 Using VKMS to test DRM API
461 --------------------------
462
463 VKMS is a software-only model of a KMS driver
464 and for running compositors. VKMS aims to enab
465 the need for a hardware display capability. Th
466 a perfect tool for validating the DRM core beh
467 compositor developer. VKMS makes it possible t
468 virtual machine without display, simplifying t
469 core changes.
470
471 To Validate changes in DRM API with VKMS, star
472 sure to enable VKMS module; compile the kernel
473 install it in the target machine. VKMS can be
474 (QEMU, virtme or similar). It's recommended th
475 of 1GB of RAM and four cores.
476
477 It's possible to run the IGT-tests in a VM in
478
479 1. Use IGT inside a VM
480 2. Use IGT from the host machine and w
481
482 Following is an example of using a VM with a s
483 the host machine to run igt-tests. This exampl
484
485 $ virtme-run --rwdir /path/for/shared_
486
487 Run the igt-tests in the guest machine. This e
488 tests::
489
490 $ /path/for/igt-gpu-tools/scripts/run-
491
492 In this example, instead of building the igt_r
493 (-p option). It creates an HTML summary of the
494 them in the folder "igt-gpu-tools/results". It
495 matching the -t option.
496
497 Display CRC Support
498 -------------------
499
500 .. kernel-doc:: drivers/gpu/drm/drm_debugfs_cr
501 :doc: CRC ABI
502
503 .. kernel-doc:: drivers/gpu/drm/drm_debugfs_cr
504 :export:
505
506 Debugfs Support
507 ---------------
508
509 .. kernel-doc:: include/drm/drm_debugfs.h
510 :internal:
511
512 .. kernel-doc:: drivers/gpu/drm/drm_debugfs.c
513 :export:
514
515 Sysfs Support
516 =============
517
518 .. kernel-doc:: drivers/gpu/drm/drm_sysfs.c
519 :doc: overview
520
521 .. kernel-doc:: drivers/gpu/drm/drm_sysfs.c
522 :export:
523
524
525 VBlank event handling
526 =====================
527
528 The DRM core exposes two vertical blank relate
529
530 :c:macro:`DRM_IOCTL_WAIT_VBLANK`
531 This takes a struct drm_wait_vblank struct
532 it is used to block or request a signal wh
533 event occurs.
534
535 :c:macro:`DRM_IOCTL_MODESET_CTL`
536 This was only used for user-mode-settind d
537 changes to allow the kernel to update the
538 mode setting, since on many devices the ve
539 reset to 0 at some point during modeset. M
540 call this any more since with kernel mode
541
542 Userspace API Structures
543 ========================
544
545 .. kernel-doc:: include/uapi/drm/drm_mode.h
546 :doc: overview
547
548 .. _crtc_index:
549
550 CRTC index
551 ----------
552
553 CRTC's have both an object ID and an index, an
554 The index is used in cases where a densely pac
555 needed, for instance a bitmask of CRTC's. The
556 drm_mode_get_plane is an example.
557
558 :c:macro:`DRM_IOCTL_MODE_GETRESOURCES` populat
559 CRTC ID's, and the CRTC index is its position
560
561 .. kernel-doc:: include/uapi/drm/drm.h
562 :internal:
563
564 .. kernel-doc:: include/uapi/drm/drm_mode.h
565 :internal:
566
567
568 dma-buf interoperability
569 ========================
570
571 Please see Documentation/userspace-api/dma-buf
572 information on how dma-buf is integrated and e
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.