Relationship between local and global deadlock detection
The four XCF messages represent one global detection cycle,
which usually takes two to four x-second intervals
to complete (where x is the number of local cycles).
Four XCF messages are required to gather and communicate the latest
information from the local deadlock detectors:
The local deadlock detector sends its information about lock waiters
to the global deadlock manager.
The global deadlock manager takes that information from all local
deadlock detectors and sends messages to each of the IRLMs in the
group. (Because the global deadlock manager is also a local deadlock
detector, it receives the same information, although somewhat quicker
than the rest of the IRLMs.)
Each local deadlock detector checks the global view of resources
and determines if it has blockers for other waiters. It passes that
information along to the global deadlock manager with its list of
waiters.
The global deadlock manager, from the information it receives
from the local deadlock detectors, determines if a global deadlock
or timeout situation exists. If a global deadlock situation exists, Db2 chooses a candidate for the
deadlock. The global deadlock manager also determines if any timeout
candidate is blocked by an incompatible waiter or holder and, if so,
presents that candidate to the owning IRLM, along with any deadlock
candidates belonging to that IRLM. When Db2 receives
this information, it determines if it should request that IRLM reject
any given timeout candidate waiter.
The following figure illustrates an example in which the deadlock
time value is set to 5 seconds. Figure 1. Global deadlock detection cycle
Deadlock detection might be delayed if any of the IRLMs in the
group encounter any of the following conditions:
XCF signaling delays
IRLM latch contention (can be encountered in systems with extremely
high IRLM locking activity)
