Zero-copy display-buffer import for GPU rendering to avoid rutabaga_resource_transfer_read CPU readback on custom hypervisor backend

[zhaoenjin]

We have successfully run the gfxstream solution on our self-developed hypervisor, but we've found a performance issue. The most severe one is that after the GPU finishes rendering a frame, the rendered data is read back from the GPU into the display's sanout buffer via
"rutabaga_resource_transfer_read", and this incurs significant performance overhead. So I'd like to ask: is there a way to import the display's buffer to the GPU for rendering, achieving memory sharing and avoiding this memory readback/copy?

[jmacnak]

I believe rutabaga_resource_export_blob() is what you are looking for so that you can share a handle to the buffer with the display.

[zhaoenjin]

The display buffer is allocate from the guest memory and is not the minigbm cross-domain host blob mem; Now I create the dma-buf fd from the guest mem in host, but I find the "rutabaga_resource_import" that is not implemented.

[gurchetansingh]

ResourceCreateBlob takes in a handle:

int32_t rutabaga_resource_create_blob(struct rutabaga *ptr, uint32_t ctx_id, uint32_t resource_id,
                                      const struct rutabaga_create_blob *rutabaga_create_blob,
                                      const struct rutabaga_iovecs *iovecs,
                                      const struct rutabaga_handle *handle);

You might be interested in magma-gpu/rutabaga_gfx#66. There be attempt to upstream this soon. Nonetheless, gfxstream guest does have support for this flow:

https://android-review.googlesource.com/q/topic:%22gfxstream-guest-mem%22 (now in Mesa)

[zhaoenjin]

guest_mem_dma-buf_import.patch

The issue I'm facing is not about the guest handle.It is that our resource is regular guest memory, not a blob. On the host side, I create a dma-buf fd from this guest memory, and then use "rutabaga_resource_import" as a rendering target for gfxstream. You can take a lookat the commit in the attachment — it implements this functionality.