VM_BIND design document with description of intended use cases. Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathapura@xxxxxxxxx> --- Documentation/gpu/rfc/i915_vm_bind.rst | 126 +++++++++++++++++++++++++ Documentation/gpu/rfc/index.rst | 4 + 2 files changed, 130 insertions(+) create mode 100644 Documentation/gpu/rfc/i915_vm_bind.rst diff --git a/Documentation/gpu/rfc/i915_vm_bind.rst b/Documentation/gpu/rfc/i915_vm_bind.rst new file mode 100644 index 000000000000..dbc35262a554 --- /dev/null +++ b/Documentation/gpu/rfc/i915_vm_bind.rst @@ -0,0 +1,126 @@ +========================================== +I915 VM_BIND feature design and use cases +========================================== + +VM_BIND feature +================ +DRM_I915_GEM_VM_BIND/UNBIND ioctls allows UMD to bind/unbind GEM buffer +objects (BOs) or sections of a BOs at specified GPU virtual addresses on +a specified address space (VM). + +These mappings (also referred to as persistent mappings) will be persistent +across multiple GPU submissions (execbuff) issued by the UMD, making execbuff +path leaner with fast path submission latency of O(1) w.r.t the number of +objects required for that submission. + +UMDs can still send BOs of these persistent mappings in execlist of execbuff +for specifying BO dependencies (implicit fencing) and to use BO as a batch. + +The persistent mappings are not individually tracked, instead the address +space (VM) they are mapped in is tracked to determine if the mappings are +being referred by GPU job (active) or not. + +VM_BIND features include: +- Different VA mappings can map to the same physical pages of an object + (aliasing). +- VA mapping can map to a partial section of the BO (partial binding). +- Support capture of mapping in the dump upon GPU error. +- TLB is flushed upon unbind completion. +- Asynchronous vm_bind and vm_unbind support. +- VM_BIND uses user/memory fence mechanism (explained below) for signaling + bind completion. + + +User/Memory Fence +================== +The idea is to take a user process virtual address and install an interrupt +handler to wake up the current task when the memory location passes the user +supplied filter. + +It also allows the user to emit their own MI_FLUSH/PIPE_CONTROL notify +interrupt within their batches after updating the value on the GPU to +have sub-batch precision on the wakeup. + +User/Memory fence <user address and value pair> can also be supplied to the +kernel driver to signal/wake up the user process after completion of an +asynchronous operation. + +This feature will be derived from the below original work: +https://patchwork.freedesktop.org/patch/349417/ + +When VM_BIND ioctl was provided with a user/memory fence via SYNC_FENCE +extension, it will be signaled upon the completion of binding of that +mapping. All async binds/unbinds are serialized, hence signaling of +user/memory fence also indicate the completion of all previous binds/unbinds. + + +TODOs +====== +- Rebase VM_BIND on top of ongoing i915 TTM adoption changes including + eviction support. +- Various optimizations like around LRU ordering of persistent mappings, + batching of TLB flushes etc. + + +Intended use cases +=================== + +Debugger +--------- +With debug event interface user space process (debugger) is able to keep track +of and act upon resources created by another process (debuggee) and attached +to GPU via vm_bind interface. + +Mesa/Valkun +------------ +VM_BIND can potentially reduce the CPU-overhead in Mesa thus improving +performance. For Vulkan it should be straightforward to use VM_BIND. +For Iris implicit buffer tracking must be implemented before we can harness +VM_BIND benefits. With increasing GPU hardware performance reducing CPU +overhead becomes more important. + +Page level hints settings +-------------------------- +VM_BIND allows any hints setting per mapping instead of per BO. +Possible hints include read-only, placement and atomicity. +Sub-BO level placement hint will be even more relevant with +upcoming GPU on-demand page fault support. + +Page level Cache/CLOS settings +------------------------------- +VM_BIND allows cache/CLOS settings per mapping instead of per BO. + +Compute +-------- +Usage of dma-fence expects that they complete in reasonable amount of time. +Compute on the other hand can be long running. Hence it is appropriate for +compute to use user/memory fence (explained above) and dma-fence usage will +be limited to in kernel consumption only. Compute must opt-in for this +mechanism during context creation time with a 'compute_ctx' flag. + +Where GPU page faults are not available, kernel driver upon buffer invalidation +must initiate a compute context suspend with a dma-fence attached to it. +And upon completion of that suspend fence, finish the invalidation and then +resume the compute context. + +This is much easier to support with VM_BIND instead of the current heavier +execbuff path resource attachment. + +Low Latency Submission +----------------------- +Allow compute UMDs to directly submit GPU jobs instead of through execbuff +ioctl. VM_BIND allows map/unmap of BOs required for directly submitted jobs. + +Shared Virtual Memory (SVM) support +------------------------------------ +VM_BIND interface can be used to map system memory directly (without gem BO +abstraction) using the HMM interface. + + +Links: +====== +- Reference WIP VM_BIND implementation can be found here. + https://gitlab.freedesktop.org/nvishwa1/nvishwa1-drm-tip + + NOTE: It is WIP and not fully functional. There are known issues which + are being worked upon. diff --git a/Documentation/gpu/rfc/index.rst b/Documentation/gpu/rfc/index.rst index 91e93a705230..7d10c36b268d 100644 --- a/Documentation/gpu/rfc/index.rst +++ b/Documentation/gpu/rfc/index.rst @@ -23,3 +23,7 @@ host such documentation: .. toctree:: i915_scheduler.rst + +.. toctree:: + + i915_vm_bind.rst -- 2.21.0.rc0.32.g243a4c7e27
