On Sun, 2005-09-18 at 16:02 +0100, Chris Wakelin wrote:
We've being running with Dovecot 1.0-stable since Tuesday, and on the whole it's gone very well (meaning nobody's noticed!). My best estimate is we're using half the disk accesses and 2/3 the CPU of UW-IMAP/Qpopper.
We've quite a mixture of error messages in the logs, the most worrying of which is that several times we've had a few seconds of "xxx-login: Can't connect to auth server at default: Connection refused" (where xxx is pop or imap). It seems to sort itself out shortly afterwards.
Probably means that dovecot-auth is getting a lot of connections at the time and kernel just returns connection refused. With 1.0alphas I've already done one change for this, committing to 1.0-stable CVS too: diff -u -r1.1.1.1 auth-server-connection.c --- src/lib-auth/auth-server-connection.c 31 Jan 2005 16:37:55 -0000 1.1.1.1 +++ src/lib-auth/auth-server-connection.c 23 Sep 2005 12:49:57 -0000 @@ -192,9 +192,17 @@ struct auth_server_connection *conn; const char *handshake; pool_t pool; - int fd; + int fd, try; - fd = net_connect_unix(path); + /* max. 1 second wait here. */ + for (try = 0; try < 10; try++) { + fd = net_connect_unix(path); + if (fd != -1 || errno != EAGAIN) + break; + + /* busy. wait for a while. */ + usleep(((rand() % 10) + 1) * 10000); + } if (fd == -1) { i_error("Can't connect to auth server at %s: %m", path); return NULL; Another thing that helps is to increase LISTEN_BACLOG in src/lib/network.c (or maybe just in net_listen_unix()). Looks like with 1.0alpha I'm using 16, but even that should probably be increased.. Changed to 64 now.
Other errors we're seeing include 1) pread_full() failed with mbox file /export/mail/c/37/sau04djp/INBOX: No such file or directory (when there is!)
The error message is wrong. Should have been "X-IMAPbase uid-last unexpectedly points outside mbox file". Some problems related this this are fixed in 1.0alphas without having been backported to 1.0stable..
2) mbox sync: UID inserted in the middle of mailbox (especially with .prayer folders - configuration folders for our Webmail)
These are always a bit strange. Some might be fixed in 1.0alpha, but I know there are still some problems left..
3) file mbox-lock.c: line 505 (mbox_lock): assertion failed: (lock_type == F_RDLCK || ibox->mbox_lock_type != F_RDLCK)
I still see this in 1.0alpha too. One way to cause it is to copy mail from a mailbox within itself, but there probably are some other ways too.