On Dienstag, 23. Oktober 2007, Howard Chu wrote:
On Dienstag, 23. Oktober 2007, email@example.com wrote:
For the records: I seem to be ble to consistently reproduce the issue when running slapd under valgrind;
Thanks, that seems to do the trick for me as well.
Funny, even running either slapd.1 or slapd.2 under valgrind, it still succeeds for me.
I haven't seen it yet when running the test regularly (single CPU i386, recently).
Looking at Gavin's testrun files, diffing slapd.2.log against a successful run, the main difference I see is that on the failed run some queries get added into the cache later than in the successful run. As such, the difference comes down to whether the pcache thread gets to run before the next query arrives or not. So, still non-deterministic and not really a big deal.
Ok, so it's most probably really the same issue as in ITS#5114 Followup 20. The non-deterministic behaviour got introduced by moving the actual caching of the entries to the cleanup handler. As a result of this, the ldapsearch commands in the test suite finishes (and the next operations start) before the data has been completely cached. Which I think is not a bug.
Any ideas how we could fix the test without introducing any sleep calls to it?