Hi everyone,
attached are six TTM patches and one Amdgpu patch I came up with while working on a new TTM feature.
The first patch is a bug fix. TTM was always waiting for on a BO to be idle when initially creating it. Previously that didn't matter much because a newly created BO should always be idle. But with the possibility of importing DMA-buf reservation objects into a TTM BO that sometimes isn't the case any more.
Patches 2, 3 and 4 are just cleanups, e.g. removal of unused parameters and variables.
Patches 5 and 6 now finally implement the new LRU features in TTM. Basically we want to give the driver more fine grained control about the order in which the BOs are added to the LRU.
For this patch number 5 implements an optional LRU removal callback, giving the driver the chance to updates it's internal LRU handling as well.
Patch number 6 add two callback functions which return the list head after which a BO should be inserted into the LRU. Default implementations for those are provided and wired up in all drivers with this patch.
The Amdgpu patch now uses the new callbacks to implement a LRU order so that VM page table BOs always come behind everything else. I actually don't intent to commit this patch, but it was good for testing.
In the end we want to use this LRU order for allowing HSA applications to have their BOs after the graphics allocations because preempting HSA processes under resource shortage has a huge bunch of overhead.
Another use Alex suggested is to sort the BOs in the LRU by size. So that we don't swap out a 200MB BO to make room for a 1MB allocation.
Please review and comment, Christian.