Re: [Dovecot] Handling ENFILE

On Mon, 2009-04-13 at 14:29 -0500, Mike Abbott wrote:

My mail server consumed all of its configured file table slots and
started returning ENFILE ("Too many open files in system") for many
operations. This wreaked havoc with dovecot-1.1.13. Here are some
areas where dovecot should detect and more gracefully handle ENFILE
error returns:

I also just tried making nfs_safe_open() fail randomly and reproducing these.

1. Deliver should report an error more helpful than "Unknown internal
   error" when fdatasync_path fails: deliver(user): Apr 10 14:53:20 Error: fdatasync_path(/Volumes/MailData/ user/new) failed: Too many open files in system deliver(user): Apr 10 14:53:20 Info: msgid=<...>: save failed to
   INBOX: BUG: Unknown internal error

http://hg.dovecot.org/dovecot-1.1/rev/576e77047959

2. Index files should not be fscked when they could not be opened due
   to ENFILE: ..
3. Dovecot should not crash when mail transaction logs become
   corrupt, or appear to have become corrupt due to ENFILE.

I made nfs_safe_open() return failures randomly and found all kinds of bugs with it. I committed several fixes to hg, but I still get some weird errors once in a while, I'll try to debug more later..