(ITS#4254) back-hdb deadlock under load

[gael.roualland@oleane.net]

Full_Name: Gaël Roualland
Version: 2.3.13
OS: Linux
URL: ftp://ftp.openldap.org/incoming/gael-roualland-051208.log
Submission from: (NULL) (213.56.0.199)


Hello,

While benchmarking back-hdb on a large test directory running OpenLDAP 2.3.13,
after a while all threads stop responding and the slapd process comes to a
deadlock.
 
For the purpose of this bug report, I ran the test with slapd configured to run
with only 4 threads, and I have attached to the locked threads with gdb.
Relevant parts of the gdb session are in the uploaded file.

As you can see, the 4 worker threads are competing for the lru mutex and cache
rwlock. It looks like :
- thread 6 wants to lock the lru mutex at cache.c:442, while it has the rwlock
locked from cache.c:419
- thread 7 wants to lock the rwlock at cache.c:602, while it has the lru mutex
locked from cache.c:823

Thanks,

Gaël.