xref: /openbmc/linux/Documentation/gpu/rfc/xe.rst (revision 40aeab04)
1==========================
2Xe – Merge Acceptance Plan
3==========================
4Xe is a new driver for Intel GPUs that supports both integrated and
5discrete platforms starting with Tiger Lake (first Intel Xe Architecture).
6
7This document aims to establish a merge plan for the Xe, by writing down clear
8pre-merge goals, in order to avoid unnecessary delays.
9
10Xe – Overview
11=============
12The main motivation of Xe is to have a fresh base to work from that is
13unencumbered by older platforms, whilst also taking the opportunity to
14rearchitect our driver to increase sharing across the drm subsystem, both
15leveraging and allowing us to contribute more towards other shared components
16like TTM and drm/scheduler.
17
18This is also an opportunity to start from the beginning with a clean uAPI that is
19extensible by design and already aligned with the modern userspace needs. For
20this reason, the memory model is solely based on GPU Virtual Address space
21bind/unbind (‘VM_BIND’) of GEM buffer objects (BOs) and execution only supporting
22explicit synchronization. With persistent mapping across the execution, the
23userspace does not need to provide a list of all required mappings during each
24submission.
25
26The new driver leverages a lot from i915. As for display, the intent is to share
27the display code with the i915 driver so that there is maximum reuse there.
28
29As for the power management area, the goal is to have a much-simplified support
30for the system suspend states (S-states), PCI device suspend states (D-states),
31GPU/Render suspend states (R-states) and frequency management. It should leverage
32as much as possible all the existent PCI-subsystem infrastructure (pm and
33runtime_pm) and underlying firmware components such PCODE and GuC for the power
34states and frequency decisions.
35
36Repository:
37
38https://gitlab.freedesktop.org/drm/xe/kernel (branch drm-xe-next)
39
40Xe – Platforms
41==============
42Currently, Xe is already functional and has experimental support for multiple
43platforms starting from Tiger Lake, with initial support in userspace implemented
44in Mesa (for Iris and Anv, our OpenGL and Vulkan drivers), as well as in NEO
45(for OpenCL and Level0).
46
47During a transition period, platforms will be supported by both Xe and i915.
48However, the force_probe mechanism existent in both drivers will allow only one
49official and by-default probe at a given time.
50
51For instance, in order to probe a DG2 which PCI ID is 0x5690 by Xe instead of
52i915, the following set of parameters need to be used:
53
54```
55i915.force_probe=!5690 xe.force_probe=5690
56```
57
58In both drivers, the ‘.require_force_probe’ protection forces the user to use the
59force_probe parameter while the driver is under development. This protection is
60only removed when the support for the platform and the uAPI are stable. Stability
61which needs to be demonstrated by CI results.
62
63In order to avoid user space regressions, i915 will continue to support all the
64current platforms that are already out of this protection. Xe support will be
65forever experimental and dependent on the usage of force_probe for these
66platforms.
67
68When the time comes for Xe, the protection will be lifted on Xe and kept in i915.
69
70Xe driver will be protected with both STAGING Kconfig and force_probe. Changes in
71the uAPI are expected while the driver is behind these protections. STAGING will
72be removed when the driver uAPI gets to a mature state where we can guarantee the
73‘no regression’ rule. Then force_probe will be lifted only for future platforms
74that will be productized with Xe driver, but not with i915.
75
76Xe – Pre-Merge Goals
77====================
78
79Drm_scheduler
80-------------
81Xe primarily uses Firmware based scheduling (GuC FW). However, it will use
82drm_scheduler as the scheduler ‘frontend’ for userspace submission in order to
83resolve syncobj and dma-buf implicit sync dependencies. However, drm_scheduler is
84not yet prepared to handle the 1-to-1 relationship between drm_gpu_scheduler and
85drm_sched_entity.
86
87Deeper changes to drm_scheduler should *not* be required to get Xe accepted, but
88some consensus needs to be reached between Xe and other community drivers that
89could also benefit from this work, for coupling FW based/assisted submission such
90as the ARM’s new Mali GPU driver, and others.
91
92As a key measurable result, the patch series introducing Xe itself shall not
93depend on any other patch touching drm_scheduler itself that was not yet merged
94through drm-misc. This, by itself, already includes the reach of an agreement for
95uniform 1 to 1 relationship implementation / usage across drivers.
96
97GPU VA
98------
99Two main goals of Xe are meeting together here:
100
1011) Have an uAPI that aligns with modern UMD needs.
102
1032) Early upstream engagement.
104
105RedHat engineers working on Nouveau proposed a new DRM feature to handle keeping
106track of GPU virtual address mappings. This is still not merged upstream, but
107this aligns very well with our goals and with our VM_BIND. The engagement with
108upstream and the port of Xe towards GPUVA is already ongoing.
109
110As a key measurable result, Xe needs to be aligned with the GPU VA and working in
111our tree. Missing Nouveau patches should *not* block Xe and any needed GPUVA
112related patch should be independent and present on dri-devel or acked by
113maintainers to go along with the first Xe pull request towards drm-next.
114
115DRM_VM_BIND
116-----------
117Nouveau, and Xe are all implementing ‘VM_BIND’ and new ‘Exec’ uAPIs in order to
118fulfill the needs of the modern uAPI. Xe merge should *not* be blocked on the
119development of a common new drm_infrastructure. However, the Xe team needs to
120engage with the community to explore the options of a common API.
121
122As a key measurable result, the DRM_VM_BIND needs to be documented in this file
123below, or this entire block deleted if the consensus is for independent drivers
124vm_bind ioctls.
125
126Although having a common DRM level IOCTL for VM_BIND is not a requirement to get
127Xe merged, it is mandatory to enforce the overall locking scheme for all major
128structs and list (so vm and vma). So, a consensus is needed, and possibly some
129common helpers. If helpers are needed, they should be also documented in this
130document.
131
132ASYNC VM_BIND
133-------------
134Although having a common DRM level IOCTL for VM_BIND is not a requirement to get
135Xe merged, it is mandatory to have a consensus with other drivers and Mesa.
136It needs to be clear how to handle async VM_BIND and interactions with userspace
137memory fences. Ideally with helper support so people don't get it wrong in all
138possible ways.
139
140As a key measurable result, the benefits of ASYNC VM_BIND and a discussion of
141various flavors, error handling and a sample API should be documented here or in
142a separate document pointed to by this document.
143
144Userptr integration and vm_bind
145-------------------------------
146Different drivers implement different ways of dealing with execution of userptr.
147With multiple drivers currently introducing support to VM_BIND, the goal is to
148aim for a DRM consensus on what’s the best way to have that support. To some
149extent this is already getting addressed itself with the GPUVA where likely the
150userptr will be a GPUVA with a NULL GEM call VM bind directly on the userptr.
151However, there are more aspects around the rules for that and the usage of
152mmu_notifiers, locking and other aspects.
153
154This task here has the goal of introducing a documentation of the basic rules.
155
156The documentation *needs* to first live in this document (API session below) and
157then moved to another more specific document or at Xe level or at DRM level.
158
159Documentation should include:
160
161 * The userptr part of the VM_BIND api.
162
163 * Locking, including the page-faulting case.
164
165 * O(1) complexity under VM_BIND.
166
167Some parts of userptr like mmu_notifiers should become GPUVA or DRM helpers when
168the second driver supporting VM_BIND+userptr appears. Details to be defined when
169the time comes.
170
171Long running compute: minimal data structure/scaffolding
172--------------------------------------------------------
173The generic scheduler code needs to include the handling of endless compute
174contexts, with the minimal scaffolding for preempt-ctx fences (probably on the
175drm_sched_entity) and making sure drm_scheduler can cope with the lack of job
176completion fence.
177
178The goal is to achieve a consensus ahead of Xe initial pull-request, ideally with
179this minimal drm/scheduler work, if needed, merged to drm-misc in a way that any
180drm driver, including Xe, could re-use and add their own individual needs on top
181in a next stage. However, this should not block the initial merge.
182
183This is a non-blocker item since the driver without the support for the long
184running compute enabled is not a showstopper.
185
186Display integration with i915
187-----------------------------
188In order to share the display code with the i915 driver so that there is maximum
189reuse, the i915/display/ code is built twice, once for i915.ko and then for
190xe.ko. Currently, the i915/display code in Xe tree is polluted with many 'ifdefs'
191depending on the build target. The goal is to refactor both Xe and i915/display
192code simultaneously in order to get a clean result before they land upstream, so
193that display can already be part of the initial pull request towards drm-next.
194
195However, display code should not gate the acceptance of Xe in upstream. Xe
196patches will be refactored in a way that display code can be removed, if needed,
197from the first pull request of Xe towards drm-next. The expectation is that when
198both drivers are part of the drm-tip, the introduction of cleaner patches will be
199easier and speed up.
200
201Drm_exec
202--------
203Helper to make dma_resv locking for a big number of buffers is getting removed in
204the drm_exec series proposed in https://patchwork.freedesktop.org/patch/524376/
205If that happens, Xe needs to change and incorporate the changes in the driver.
206The goal is to engage with the Community to understand if the best approach is to
207move that to the drivers that are using it or if we should keep the helpers in
208place waiting for Xe to get merged.
209
210This item ties into the GPUVA, VM_BIND, and even long-running compute support.
211
212As a key measurable result, we need to have a community consensus documented in
213this document and the Xe driver prepared for the changes, if necessary.
214
215Dev_coredump
216------------
217
218Xe needs to align with other drivers on the way that the error states are
219dumped, avoiding a Xe only error_state solution. The goal is to use devcoredump
220infrastructure to report error states, since it produces a standardized way
221by exposing a virtual and temporary /sys/class/devcoredump device.
222
223As the key measurable result, Xe driver needs to provide GPU snapshots captured
224at hang time through devcoredump, but without depending on any core modification
225of devcoredump infrastructure itself.
226
227Later, when we are in-tree, the goal is to collaborate with devcoredump
228infrastructure with overall possible improvements, like multiple file support
229for better organization of the dumps, snapshot support, dmesg extra print,
230and whatever may make sense and help the overall infrastructure.
231
232Xe – uAPI high level overview
233=============================
234
235...Warning: To be done in follow up patches after/when/where the main consensus in various items are individually reached.
236