with the patch above the kms value calculated more quickly because of next improvements:
1) interval tree was used for the going through the list of locks instead the direct linked list
search.
2) skip the subtree when maximum value in this subtree is less then new kms
3) kms search stopped when lock with size equal to the old kms value found.

this optimization shows speed increase from 20 to 40% according to the tests results.

https://bugzilla.lustre.org/show_bug.cgi?id=22060 - contains the whole story