On Oct 6, 2008, at 2:20 AM, Steven Winikoff wrote:
- recompiled dovecot without optimization (ie, gcc -O0 instead of
-O2) (as per http://www.mail-archive.com/dovecot@dovecot.org/msg12936.html)
The problem here turned out to be a bug in Sun CC (and gcc -O2
apparently worked after all, so maybe a problem with testing).
- reverted to using flock for file locking instead of fcntl (we'd
been using flock with dovecot 1.0.10; 1.1.3 used fcntl by default in a couple of places, so I changed it back to see what would happen)
This shouldn't really make a difference.
- we still see the occasional assertion failure at line 1091 of index-mail.c; this is the one that I'd hoped to solve by recompiling without optimization, but at least it seems to be happening less frequently than before
I had hoped v1.1.3 would have fixed this. But perhaps this is also
related to the disk quota issues. Are there any other errors that are
logged before the assert?
- this doesn't happen often, but once in a while we see assertion failures at line 163 of src/pop3-login/client-authenticate.c or line 200 of src/imap-login/client-authenticate.c
I think I've managed to get this also sometimes when killing dovecot-
auth or doing something else "wrong" (I don't remember exactly), but
not during normal operation. I should have noted down how to reproduce
this when I still knew how to reproduce it..
I missed this originally, but on reflection I notice in the logs that every time the "record size too small" error appears, inevitably the user involved is over disk quota and out of grace period. For
example,
Yes, this is most likely what's the problem. I remember testing v1.0
quite a lot in "out of disk space" conditions, but I haven't really
done any such testing with v1.1. I'll try that and see if I can get
these problems reproduced and fixed.