Deadlocks can be detected while the program is running, by running cycle detection algorithms on the graph that de?nes the current use of resources.

De?ne this graph as follows: it has one vertex for each resource (r1; : : : ; rm) and one vertex for each thread (t1; : : : ; tn). If a resource ri is held by thread tj , then we add an edge from vertex ri to vertex tj . If a thread tk is trying to acquire resource r', then we add an edge from vertex tk to vertex r'.

Given this graph, we can run a cycle detection algorithm. If a cycle is found, there is a deadlock. Detecting a deadlock ismuch easier than recovering froma deadlock. Several possible approaches might be to kill all of the threads in the cycle, or kill the threads one at a time, forcing them to release resources, and hope that this will break the deadlock.

While this may sound easy, it is not. Killing threads generally doesn't release their resources cleanly (locks,memory, ?les, etc). This is essentially the rollback problem, to back out all the actions of a thread. Databases usually include rollback mechanisms, which can be quite complicated, and it is not always possible to roll back all the actions of a thread (consider a thread which outputs hard-copy printed pages).

As a general guideline, do not use functions like pthread cancel(), which kills a thread, unless you really know what you are doing.