Berkeley DB infinite sleep on Solaris [#3245]
davids at webmaster.com
Mon Jan 8 22:35:41 UTC 2001
> The Solaris man page (_lwp_cond_signal) is explicit:
> "Both functions should be called under the protection of the
> same LWP mutex lock that is used with the LWP condition
> variable being signaled. Otherwise, the condition variable
> may be signalled between the test of the associated condi-
> tion and blocking in _lwp_cond_wait(). This can cause an
> infinite wait."
This is flat-out incorrect. What it's suggesting is that the following can
<-- signal can occur here and be lost
True, that wakeup would be lost. But since the thread holds a lock on the
predicate, the value of the predicate is the same, hence the value of
'Test_Predicate' will be the same. Hence the thread, if it didn't lose the
wakeup, would just go back to sleep.
Since the running thread holds the lock, no other thread can change the
predicate inbetween the 'Test' and the 'Block'. Thus the predicate cannot
have changed state inbetween the 'Test' and the 'Block'. So even if there
was a lost wakeup, the thread would still want to block because the
predicate is still false.
The decision to block or not is based upon the predicate. So long as the
predicate does not change (and it can't because the thread holds a lock on
it), if it decided to sleep before the lost signal, it would just decide to
sleep again after it. So the lost signal simply saves a loop.
More information about the inn-workers