i915_scheduler.rst (7208B)
1========================================= 2I915 GuC Submission/DRM Scheduler Section 3========================================= 4 5Upstream plan 6============= 7For upstream the overall plan for landing GuC submission and integrating the 8i915 with the DRM scheduler is: 9 10* Merge basic GuC submission 11 * Basic submission support for all gen11+ platforms 12 * Not enabled by default on any current platforms but can be enabled via 13 modparam enable_guc 14 * Lots of rework will need to be done to integrate with DRM scheduler so 15 no need to nit pick everything in the code, it just should be 16 functional, no major coding style / layering errors, and not regress 17 execlists 18 * Update IGTs / selftests as needed to work with GuC submission 19 * Enable CI on supported platforms for a baseline 20 * Rework / get CI heathly for GuC submission in place as needed 21* Merge new parallel submission uAPI 22 * Bonding uAPI completely incompatible with GuC submission, plus it has 23 severe design issues in general, which is why we want to retire it no 24 matter what 25 * New uAPI adds I915_CONTEXT_ENGINES_EXT_PARALLEL context setup step 26 which configures a slot with N contexts 27 * After I915_CONTEXT_ENGINES_EXT_PARALLEL a user can submit N batches to 28 a slot in a single execbuf IOCTL and the batches run on the GPU in 29 paralllel 30 * Initially only for GuC submission but execlists can be supported if 31 needed 32* Convert the i915 to use the DRM scheduler 33 * GuC submission backend fully integrated with DRM scheduler 34 * All request queues removed from backend (e.g. all backpressure 35 handled in DRM scheduler) 36 * Resets / cancels hook in DRM scheduler 37 * Watchdog hooks into DRM scheduler 38 * Lots of complexity of the GuC backend can be pulled out once 39 integrated with DRM scheduler (e.g. state machine gets 40 simplier, locking gets simplier, etc...) 41 * Execlists backend will minimum required to hook in the DRM scheduler 42 * Legacy interface 43 * Features like timeslicing / preemption / virtual engines would 44 be difficult to integrate with the DRM scheduler and these 45 features are not required for GuC submission as the GuC does 46 these things for us 47 * ROI low on fully integrating into DRM scheduler 48 * Fully integrating would add lots of complexity to DRM 49 scheduler 50 * Port i915 priority inheritance / boosting feature in DRM scheduler 51 * Used for i915 page flip, may be useful to other DRM drivers as 52 well 53 * Will be an optional feature in the DRM scheduler 54 * Remove in-order completion assumptions from DRM scheduler 55 * Even when using the DRM scheduler the backends will handle 56 preemption, timeslicing, etc... so it is possible for jobs to 57 finish out of order 58 * Pull out i915 priority levels and use DRM priority levels 59 * Optimize DRM scheduler as needed 60 61TODOs for GuC submission upstream 62================================= 63 64* Need an update to GuC firmware / i915 to enable error state capture 65* Open source tool to decode GuC logs 66* Public GuC spec 67 68New uAPI for basic GuC submission 69================================= 70No major changes are required to the uAPI for basic GuC submission. The only 71change is a new scheduler attribute: I915_SCHEDULER_CAP_STATIC_PRIORITY_MAP. 72This attribute indicates the 2k i915 user priority levels are statically mapped 73into 3 levels as follows: 74 75* -1k to -1 Low priority 76* 0 Medium priority 77* 1 to 1k High priority 78 79This is needed because the GuC only has 4 priority bands. The highest priority 80band is reserved with the kernel. This aligns with the DRM scheduler priority 81levels too. 82 83Spec references: 84---------------- 85* https://www.khronos.org/registry/EGL/extensions/IMG/EGL_IMG_context_priority.txt 86* https://www.khronos.org/registry/vulkan/specs/1.2-extensions/html/chap5.html#devsandqueues-priority 87* https://spec.oneapi.com/level-zero/latest/core/api.html#ze-command-queue-priority-t 88 89New parallel submission uAPI 90============================ 91The existing bonding uAPI is completely broken with GuC submission because 92whether a submission is a single context submit or parallel submit isn't known 93until execbuf time activated via the I915_SUBMIT_FENCE. To submit multiple 94contexts in parallel with the GuC the context must be explicitly registered with 95N contexts and all N contexts must be submitted in a single command to the GuC. 96The GuC interfaces do not support dynamically changing between N contexts as the 97bonding uAPI does. Hence the need for a new parallel submission interface. Also 98the legacy bonding uAPI is quite confusing and not intuitive at all. Furthermore 99I915_SUBMIT_FENCE is by design a future fence, so not really something we should 100continue to support. 101 102The new parallel submission uAPI consists of 3 parts: 103 104* Export engines logical mapping 105* A 'set_parallel' extension to configure contexts for parallel 106 submission 107* Extend execbuf2 IOCTL to support submitting N BBs in a single IOCTL 108 109Export engines logical mapping 110------------------------------ 111Certain use cases require BBs to be placed on engine instances in logical order 112(e.g. split-frame on gen11+). The logical mapping of engine instances can change 113based on fusing. Rather than making UMDs be aware of fusing, simply expose the 114logical mapping with the existing query engine info IOCTL. Also the GuC 115submission interface currently only supports submitting multiple contexts to 116engines in logical order which is a new requirement compared to execlists. 117Lastly, all current platforms have at most 2 engine instances and the logical 118order is the same as uAPI order. This will change on platforms with more than 2 119engine instances. 120 121A single bit will be added to drm_i915_engine_info.flags indicating that the 122logical instance has been returned and a new field, 123drm_i915_engine_info.logical_instance, returns the logical instance. 124 125A 'set_parallel' extension to configure contexts for parallel submission 126------------------------------------------------------------------------ 127The 'set_parallel' extension configures a slot for parallel submission of N BBs. 128It is a setup step that must be called before using any of the contexts. See 129I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE or I915_CONTEXT_ENGINES_EXT_BOND for 130similar existing examples. Once a slot is configured for parallel submission the 131execbuf2 IOCTL can be called submitting N BBs in a single IOCTL. Initially only 132supports GuC submission. Execlists supports can be added later if needed. 133 134Add I915_CONTEXT_ENGINES_EXT_PARALLEL_SUBMIT and 135drm_i915_context_engines_parallel_submit to the uAPI to implement this 136extension. 137 138.. kernel-doc:: include/uapi/drm/i915_drm.h 139 :functions: i915_context_engines_parallel_submit 140 141Extend execbuf2 IOCTL to support submitting N BBs in a single IOCTL 142------------------------------------------------------------------- 143Contexts that have been configured with the 'set_parallel' extension can only 144submit N BBs in a single execbuf2 IOCTL. The BBs are either the last N objects 145in the drm_i915_gem_exec_object2 list or the first N if I915_EXEC_BATCH_FIRST is 146set. The number of BBs is implicit based on the slot submitted and how it has 147been configured by 'set_parallel' or other extensions. No uAPI changes are 148required to the execbuf2 IOCTL.