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.
*/
