[AGENT++] [PATCH] agent++: locking fixes and improvements

[dominik.vogt at external.thalesgroup.com]

On Tue, Aug 17, 2010 at 10:24:47PM +0200, Jochen Katz wrote:
>    >   This fixes a race condition or crash caused by unlocking the mib
>    >   in the wrong place (there are three nested loops; the lock is
>    >   managed in the second loop, but the particular unlock takes
>    >   place right before the inner loop is restarted).
> 
>    maybe it is already too late, but I cannot find a way through the code
>    where the lock is not locked when this unlock/continue is called.

The code is structured like this:

  for (/*outer loop*/)
  {
    for (/*middle loop*/)
    {
A     lock_mib()
      do
      {
B       ... <------------------
        if (/*something*/)     |
        {                      | inner loop condition ture
C         unlock_mib();        |
D         continue;  ---------------
        }                           |
        ...                         | inner loop condition false
      } while (/*inner loop*/);     |
E     ... <-------------------------
F     unlock_mib();
      ...
    }
  }

1. The program first enters the middle loop, it locks the Mib at
   A and then enters the inner loop.
2. At the beginning of the inner loop (B), the Mib is locked.
3. The program reaches C inside the inner loop and unlocks the
   Mib.
4. At D, the program jumps to the end of the inner loop and either
   breaks the inner loop and proceeds with E (if the condition is
   false) or restarts the *inner* loop and jumps to B.
5. If it jumps to B, the inner loop is now running *without* the
   lock.  If it proceeds with E, it eventually reaches F and tries
   to unlock the Mib although it is not locked.

With posix mutexes it would not do any harm to unlock a mutex
twice that was locked exactly once.  But it *does* harm to run the
inner loop without the lock.

Ciao

Dominik ^_^  ^_^

-- 

Dominik Vogt