This series makes sure we never reallocated a busy extent. This is absolutely
needed for a working online discard implementation, and should also speed
up large concurrent workloads that would trip over a busy extents a lot.
The two most significant patches are from a series Dave posted mored than
two years ago, although the main patch to avoid busy ranges in the normal
allocator need a lot of bug fixes to survived QA. In addition to that we
now also completely avoid resues for AGFL based allocations. While this
might not be important from a performance perspectice it is critical for
the discard implementation I'm currently working on.
|