This patch implements my attempt at dmabuf synchronization. The core idea is that a lot of devices will have their own methods of synchronization, but more complicated devices allow some way of fencing, so why not export those as dma-buf?
This patchset implements dmabufmgr, which is based on ttm's code. The ttm code deals with a lot more than just reservation however, I took out almost all the code not dealing with reservations.
I used the drm-intel-next-queued tree as base. It contains some i915 flushing changes. I would rather use linux-next, but the deferred fput code makes my system unbootable. That is unfortunate since it would reduce the deadlocks happening in dma_buf_put when 2 devices release each other's dmabuf.
The i915 changes implement a simple cpu wait only, the nouveau code imports the sync dmabuf read-only and maps it to affected channels, then performs a wait on it in hardware. Since the hardware may still be processing other commands, it could be the case that no hardware wait would have to be performed at all.
Only the nouveau nv84 code is tested, but the nvc0 code should work as well.