drm/v3d Broadcom V3D Graphics Driver¶
This driver supports the Broadcom V3D 3.3 and 4.1 OpenGL ES GPUs. For V3D 2.x support, see the VC4 driver.
The V3D GPU includes a tiled render (composed of a bin and render pipelines), the TFU (texture formatting unit), and the CSD (compute shader dispatch).
GPU buffer object (BO) management¶
Compared to VC4 (V3D 2.x), V3D 3.3 introduces an MMU between the GPU and the bus, allowing us to use shmem objects for our storage instead of CMA.
Physically contiguous objects may still be imported to V3D, but the driver doesn't allocate physically contiguous objects on its own. Display engines requiring physically contiguous allocations should look into Mesa's "renderonly" support (as used by the Mesa pl111 driver) for an example of how to integrate with V3D.
Long term, we should support evicting pages from the MMU when under memory pressure (thus the v3d_bo_get_pages() refcounting), but that's not a high priority since our systems tend to not have swap.
Address space management¶
The V3D 3.x hardware (compared to VC4) now includes an MMU. It has a single level of page tables for the V3D's 4GB address space to map to AXI bus addresses, thus it could need up to 4MB of physically contiguous memory to store the PTEs.
Because the 4MB of contiguous memory for page tables is precious, and switching between them is expensive, we load all BOs into the same 4GB address space.
To protect clients from each other, we should use the GMP to quickly mask out (at 128kb granularity) what pages are available to each client. This is not yet implemented.
The shared DRM GPU scheduler is used to coordinate submitting jobs to the hardware. Each DRM fd (roughly a client process) gets its own scheduler entity, which will process jobs in order. The GPU scheduler will round-robin between clients to submit the next job.
For simplicity, and in order to keep latency low for interactive
jobs when bulk background jobs are queued up, we submit a new job
to the HW only when it has completed the last one, instead of
filling up the CTQ FIFOs with jobs. Similarly, we use
drm_sched_job_add_dependency() to manage the dependency between bin and
render, instead of having the clients submit jobs using the HW's
semaphores to interlock between them.
When we take a bin, render, TFU done, or CSD done interrupt, we need to signal the fence for that job so that the scheduler can queue up the next one and unblock any waiters.
When we take the binner out of memory interrupt, we need to allocate some new memory and pass it to the binner so that the current job can make progress.