Locking System Performance Improvements

A number of improvements have been made to the locking system to eliminate or minimize bottlenecks identified when running various performances tests.

Preallocated RSB/LKBs

Each resource RSB now has an embedded a lock block (LKB), removing the need for a separate, contentious LKB allocation every time a new resource is allocated.

An LLB stash of LKBs is also maintained, similar to the RSB stash.

When an RSB or LKB is freed, it is returned to the LLB's stash; when the lock list itself is freed, all stashed RSB/LKBs are returned to the free pool.

Miscellaneous Locking System Improvements

The following miscellaneous locking system improvements are included in Ingres 2.6:

• The number of RSB waiters and converters are now maintained in the RSB.
• The deadlock wait-for graph lock (lkd_dlock_lock) does not need to be held if the RSB has neither waiters nor converters.
• The LKREQ built in the stack does not need to be copied to the LKB indiscriminately.
• When a lock request is blocked, the blocker's identity is now saved in the LKREQ and formatted in SYS_ERR only when the request fails.