Bug 146549

Description of problem:
Ever so often with scriptlets that return rather quickly (i.e. path 
of execution in script is small) rpm will deadlock waiting for the 
script to return.  I have been seeing this with my rpm test hanrness 
for about a year now, but recently have reproduced this on RHEL 3.

Version-Release number of selected component (if applicable):
4.2.3-10

How reproducible:
Very Rare, but it is reproducible if you wait long enough (no pun 
intended)

Steps to Reproduce:
1. Create a small two small spec files that contains %pre, %post, %
preun and %postun scriptlets that create a semaphore file and then 
exit 0.  The first should something like x-1-1 and the second x-1-2.
2.  Bulid the packages from the spec file.
3.  Now install the first, upgrade to the second, erase them and 
start all over in a loop:

    while :
    do
      rpm -Uvh x-1-1.noarch.rpm
      rpm -Uvh x-1-2.noarch.rpm
      rpm -e x
    done

  
Actual results:
Eventually, after perhaps thousands of passes it will lockup.  When 
it does, strace will reveal it is haning on a futex, and gdb will 
show it is running pthread_cond_wait().

Expected results:
That it would not hang ever.

Additional info:
I spent a lot of time trying to figure out what was going on 
(probably more than I should have).  The first thing I found was that
rpm was using pthread_cond_signal() in its signal handler and 
pthread_cond_wait() in rpmsqWaitUnregister() without proper locking 
of the signaler and the waiter (it was using the conditions mutex 
with the wait, but not with the signaler).  I initially created a 
patch to fix this, only to find that the signal handler (and thus 
the caller of pthread_cond_signal()) where in the same thread of 
execution so my initial patch actually create a far easier to step 
on deadlock.

Once I realized that pthread_cond_signal() and pthread_cond_wait() 
were in the same thread of execution, I arrived at what I think is 
a better solution.  Ultimately, you can't use pthread_cond stuff in 
the same thread of execution (quick read of the man page will show 
this), so instead I thought through a way of using a single mutex
to accomplish the synchronization that the pthread_cond's were used
to provide.  Its pretty simple:

    Before the fork crate and lock the rpmsq objects mutex.
    In the signal handler:
        unlock the mutex.
    In rpmsqWaitUnregister():
        try to lock the mutex()

since it starts off locked rpmsqWaitUnregister() will not be able to 
move forward (i.e. will block) until the signal handler unlocks the 
mutex.

Also, I would understand if this was not fixed in RHEL 3, I merely 
logging this as a bug with a possible patch.  What you guys do with 
it is up to you (I would do nothing until it has been tested out in 
Fedora land if I where you).