V1->V2: - Works now on top of 4.20-rc7 - A couple of bug fixes This is a patchset on top of Matthew Wilcox Xarray code and implements slab object migration for xarray nodes. The migration is integrated into the defragmetation and shrinking logic of the slab allocator. Defragmentation will ensure that all xarray slab pages have less objects available than specified by the slab defrag ratio. Slab shrinking will create a slab cache with optimal object density. Only one slab page will have available objects per node. To test apply this patchset and run a workload that uses lots of radix tree objects Then go to /sys/kernel/slab/radix_tree_node Inspect the number of total objects that the slab can handle cat total_objects qmdr:/sys/kernel/slab/radix_tree_node# cat objects 868 N0=448 N1=168 N2=56 N3=196 And the number of slab pages used for those cat slabs qmdr:/sys/kernel/slab/radix_tree_node# cat slabs 31 N0=16 N1=6 N2=2 N3=7 Perform a cache shrink operation echo 1 >shrink Now see how the slab has changed: qmdr:/sys/kernel/slab/radix_tree_node# cat slabs 30 N0=15 N1=6 N2=2 N3=7 qmdr:/sys/kernel/slab/radix_tree_node# cat objects 713 N0=349 N1=141 N2=52 N3=171 This is just a barebones approach using a special mode of the slab migration patchset that does not require refcounts. If this is acceptable then additional functionality can be added: 1. Migration of objects to a specific node. Not sure how to implement that. Using another sysfs field? 2. Dispersion of objects across all nodes (MPOL_INTERLEAVE) 3. Subsystems can request to move an object to a specific node. How to design such functionality best? 4. Tying into the page migration and page defragmentation logic so that so far unmovable pages that are in the way of creating a contiguous block of memory will become movable. This would mean checking for slab pages in the migration logic and calling slab to see if it can move the page by migrating all objects. This is only possible for xarray for now but it would be worthwhile to extend this to dentries and inodes.
