[Dovecot] Maildir synchronization

I thought I'd send this here too, since I spent so much time writing it as comment to maildir-sync.c :) Now only thing left to do is to write the code to match it.

/* Here's a description of how we handle Maildir synchronization and it's problems:

We want to be as efficient as we can. The most efficient way to check if changes have occured is to stat() the new/ and cur/ directories and uidlist file - if their mtimes hasn't changed, there's no changes and we don't need to do anything.

Problem 1: Multiple changes can happen within a single second - nothing guarantees that once we synced it, someone else didn't just then make a modification. Such modifications wouldn't get noticed until a new modification occured later.

Problem 2: Syncing cur/ directory is much more costly than syncing new/. Moving mails from new/ to cur/ will always change mtime of cur/ causing us to sync it as well.

Problem 3: We may not be able to move mail from new/ to cur/ because we're out of quota, or simply because we're accessing a read-only mailbox.

MAILDIR_SYNC_SECS

Several checks below use MAILDIR_SYNC_SECS, which should be maximum clock drift between all computers accessing the maildir (eg. via NFS), rounded up to next second. Our default is 1 second, since everyone should be using NTP.

Note that setting it to 0 works only if there's only one computer accessing the maildir. It's practically impossible to make two clocks _exactly_ synchronized.

cur directory

We have maildir_cur_dirty variable which is set to cur/ directory's mtime when it's >= time() - MAILDIR_SYNC_SECS and we _think_ we have synchronized the directory.

When maildir_cur_dirty is non-zero, we don't synchronize the cur/ directory until

  a) cur/'s mtime changes
  b) opening a mail fails with ENOENT
  c) time() > maildir_cur_dirty + MAILDIR_SYNC_SECS

This allows us to modify the maildir multiple times without having to sync it at every change. The sync will eventually be done to make sure we didn't miss any external changes.

The maildir_cur_dirty is set when:

  - we change message flags
  - we expunge messages
  - we move mail from new/ to cur/
  - we sync cur/ directory and it's mtime is
    >= time() - MAILDIR_SYNC_SECS

It's unset when we do the final syncing, ie. when mtime is older than time() - MAILDIR_SYNC_SECS.

new directory

If new/'s mtime is >= time() - MAILDIR_SYNC_SECS, always synchronize it. maildir_cur_dirty-like feature might save us a few syncs, but that might break a client which saves a mail in one connection and tries to fetch it in another one. new/ directory is almost always empty, so syncing it should be very fast anyway.

Normally we move all mails from new/ to cur/ whenever we sync it. If it's not possible for some reason, we set maildir_have_new flag on which instructs synchronization to check files in new/ directory as well. maildir_keep_new flag is also set which instructs syncing to not even try to move mails to cur/ anymore.

If client tries to change a flag for message in new/, we try to rename() it into cur/. If it's successful, we clear the maildir_keep_new flag so at next sync we'll try to move all of them to cur/. When all of them have been moved, maildir_have_new flag is cleared as well. Expunges will also clear maildir_keep_new flag.

If rename() still fails because of ENOSPC or EDQUOT, we still save the flag changes in index with dirty-flag on. When moving the mail to cur/ directory, or when we notice it's already moved there, we apply the flag changes to the filename, rename it and remove the dirty flag. If there's dirty flags, this should be tried every time after expunge or when closing the mailbox.

broken clients

Originally the middle identifier in Maildir filename was specified only as <process id>_<delivery counter>. That however created a problem with randomized PIDs which made it possible that the same PID was reused within one second.

So if within one second a mail was delivered, MUA moved it to cur/ and another mail was delivered by a new process using same PID as the first one, we likely ended up overwriting the first mail when the second mail was moved over it.

Nowadays everyone should be giving a bit more specific identifier, for example include microseconds in it which Dovecot does.

There's a simple way to prevent this from happening in some cases: Don't move the mail from new/ to cur/ if it's mtime is >= time() - MAILDIR_SYNC_SECS. The second delivery's link() call then fails because the file is already in new/, and it will then use a different filename. There's a few problems with this however:

  - it requires extra stat() call which is unneeded extra I/O
  - another MUA might still move the mail to cur/
  - if first file's flags are modified by either Dovecot or another
    MUA, it's moved to cur/ (you _could_ just do the dirty-flagging
but that'd be ugly)

Because this is useful only for very few people and it requires extra I/O, I decided not to implement this. It should be however quite easy to do since we need to be able to deal with files in new/ in any case.

It's also possible to never accidentally overwrite a mail by using link() + unlink() rather than rename(). This however isn't very good idea as it introduces potential race conditions when multiple clients are accessing the mailbox:

Trying to move the same mail from new/ to cur/ at the same time:

  a) Client 1 uses slightly different filename than client 2,
     for example one sets read-flag on but the other doesn't.
 You have the same mail duplicated now.

  b) Client 3 sees the mail between Client 1's and 2's link() calls
     and changes it's flag. You have the same mail duplicated now.

And it gets worse when they're unlink()ing in cur/ directory:

  c) Client 1 changes mails's flag and client 2 changes it back
     between 1's link() and unlink(). The mail is now expunged.

  d) If you try to deal with the duplicates by unlink()ing another
     one of them, you might end up unlinking both of them.

So, what should we do then if we notice a duplicate? First of all, it might not be a duplicate at all, readdir() might have just returned it twice because it was just renamed. What we should do is create a completely new base name for it and rename() it to that. If the call fails with ENOENT, it only means that it wasn't a duplicate after all. */