Willy, On 4/29/22 14:39, Willy Tarreau wrote:
However once we manage to get rid of all of them, it would be interesting to enable them in the CI so that new regressions can be caught. But until this happens, it would only be reports for known failures.
I agree and I planned to propose that once I've worked through the backlog for my production config.
Anyway your test is useful in that it reported quite a significant number of entries at once, we rarely see so many, so it will be a good starting point about new locations to look for.
Those in Ilya's test are "false positives" in so far, as `-cc` currently does not yet use deinit_and_exit, but only exit. So there's a huge number of live allocations we can already clean.
Currently a deinit only happens for: - haproxy -vv - haproxy -c (if the check is successful, i.e. exit 0). - SIGUSR1 Best regards Tim Düsterhus
