Btw, I also still don't understand why remoteproc is using dma_declare_coherent_memory to start with. The virtio code has exactly one call to dma_alloc_coherent vring_alloc_queue, a function that already switches between two different allocators. Why can't we just add a third allocator specifically for these remoteproc memory carveouts and bypass dma_declare_coherent_memory entirely?
