- Informal test results showing the changes from 5.4 to 5.6 are available. - Reduce stalls in the kernel vm_page_alloc() code (vm_page_list_find()). - Improve page allocation algorithm to avoid re-iterating the same queues as the search is widened. - Add a vm_page_hash*() API that allows the kernel to do heuristical lockless lookups of VM pages. - Change vm_hold() and vm_unhold() semantics to not require any spin-locks. - Change vm_page_wakeup() to not require any spin-locks. - Change wiring vm_page's no longer manipulates the queue the page is on, saving a lot of overhead. Instead, the page will be removed from its queue only if the pageout demon encounters it. This allows pages to enter and leave the buffer cache quickly. - Refactor the handling of fictitious pages. - Remove m->md.pv_list entirely. VM pages in mappings no longer allocate pv_entry's, saving an enormous amount of memory when multiple processes utilize large shared memory maps (e.g. postgres database cache). - Refactor vm_object shadowing, disconnecting the backing linkages from the vm_object itself and instead organizing the linkages in a new structure called vm_map_backing which hangs off the vm_map_entry. - pmap operations now iterate vm_map_backing structures (rather than spin-locked page lists based on the vm_page and pv_entry's), and will test/match operations against the PTE found in the pmap at the requisite location. This doubles VM fault performance on shared pages and reduces the locking overhead for fault and pmap operations. - Simplify the collapse code, removing most of the original code and replacing it with simpler per-vm_map_entry optimizations to limit the shadow depth.
- Major updates to the radeon and ttm (amd support code) drivers. We have not quite gotten the AMD support up to the more modern cards or Ryzen APUs yet, however. - Improve UEFI framebuffer support. - A major deadlock has been fixed in the radeon/ttm code. - Refactor the startup delay designed to avoid conflicts between the i915 driver initialization and X startup. - Add DRM_IOCTL_GET_PCIINFO to improve mesa/libdrm support. - Fix excessive wired memory build-ups. - Fix Linux/DragonFly PAGE_MASK confusion in the DRM code. - Fix idr_*() API bugs.
- The filesystem sync code has been rewritten to significantly improve performance. - Sequential write performance also improved. - Add simple dependency tracking to prevent directory/file splits during create/rename/remove operations, for better consistency after a crash. - Refactor the snapshot code to reduce flush latency and to ensure a consistent snapshot. - Attempt to pipeline the flush code against the frontend, improving flush vs frontend write concurrency. - Improve umount operation. - Fix an allocator race that could lead to corruption. - Numerous other bugs fixed. - Improve verbosity of CHECK (CRC error) console messages.