Fixing problems around shared fences and RCU in DMA-buf

Hi everyone and especially Daniel,

this is the revised patch set to fix and rework dma_buf_poll(). The new code should avoid problems with RCU and also now correctly waits for all fences in the resv object.

The rest of the series is then the well known change to dma_resv_test_signaled(), nouveau and now new also msm.

Then last are two patches which drop the workarounds from amdgpu, but those can wait till the next cycle.

I think it would be rather good if the have at least to change to dma_buf_poll() pushed in this merge window and maybe even CC stable since this looks really broken to me.

Please review, test and/or comment.

Thanks, Christian.