[Intel-gfx] Deploying new iterator interface for dma-buf

[Christian König]

Hi everybody,

we recently found that a good bunch of the RCU accesses to the dma_resv object are actually not correctly protected.

Those where fixed by either dropping the RCU approach and taking appropriate locks or using a central function to return the current fences as array and then work with that snapshot.

This set now tries to prevent adding any new broken code by rolling out two new interfaces to access the fences in a dma_resv object:

dma_resv_for_each_fence() - Iterator which should be used while holding the reservation lock.
dma_resv_for_each_fence_unlocked() - Iterator based on RCU which can be used without holding the reservation lock and automatic restart on concurrent modification.

While doing this we also move the decision which fences to use for write and read accesses into the dma_resv object which results in a quite nice code de-duplication and simplification.

The only two remaining users of the RCU shared fence interface are removing shared fences in amdkfd and debugfs code in qxl which will both be addresses in the next patch set.

Please review and/or comment,
Christian.