Linux 6.4 Lands Concurrent I/O Performance Optimizations For Device Mapper

The Device Mapper "DM" subsystem updates have been merged for the in-development Linux 6.4 kernel and it includes some notable performance optimizations.

As I wrote about back in March, Device Mapper was set to see improved concurrent I/O performance. That DM performance work has successfully landed for Linux 6.4. That concurrent I/O work is significant in that in some cases for acquiring buffers can be as much as twenty-five times faster.

Per the DM merge to Linux Git some of the highlights for the Device Mapper code this cycle includes:

- Split dm-bufio's rw_semaphore and rbtree. Offers improvements to dm-bufio's locking to allow increased concurrent IO -- particularly for read access for buffers already in dm-bufio's cache.

- Also split dm-bio-prison-v1's spinlock and rbtree with comparable aim at improving concurrent IO (for the DM thinp target).

- Both the dm-bufio and dm-bio-prison-v1 scaling of the number of locks and rbtrees used are managed by dm_num_hash_locks(). And the hash function used by both is dm_hash_locks_index().

- Allow DM targets to require DISCARD, WRITE_ZEROES and SECURE_ERASE to be split at the target specified boundary (in terms of max_discard_sectors, max_write_zeroes_sectors and max_secure_erase_sectors respectively).

- DM core ->io_hints improvements needed in support of new discard support that is added to the DM "zero" and "error" targets.

- Various small improvements to both DM core and DM targets.

Linux 6.4 has been a very active merge window thus far, which ends this weekend, with many new features and optimizations having been introduced.