view TODO @ 956:26cafa3dc09c HEAD

minor optimization
author Timo Sirainen <>
date Sun, 12 Jan 2003 01:49:45 +0200
parents 80b847900dfd
children 6f005d5d9931
line wrap: on
line source

 - bugs
    - maildir: if mail file isn't found, it may be because it was renamed
      (flag changed). we must then sync the directory and see again if the mail
      is found
    - we may override mbox dotlock file, but then get stuck at fcntl/flock.
      we should check for those before overriding the mbox lock..
    - mail-lockdir.c isn't 100% safe.. stale locks are detected by checking
      that hard link count is 1, then it's unlink()ed. but what if another
      process did the same unlink() + creat() in the middle of our
      stat()..unlink()? no easy way to fix this really, just replace it with a
      fcntl() lock.
    - SEARCH FROM/TO/CC/BCC now generates the field from ENVELOPE which it
      uses for matching. This however may give different results than when
      matching from headers.
    - SIGHUP didn't update imap_listen. this is a bit annoying to fix though,
      since new listen() may fail for a few times because auth processes may
      not die immediately..
    - SIGHUP doesn't update log file location.
    - We can use Linux sendfile() up to 2GB, after that we get EOVERFLOW and
      fail. We should rather fallback to mmap+write at that point.
    - unlink_directory() is racy with symlink handling, see if that could be

 - reliability fixes:
    - if we deleted mail from index but didn't write modify log, other
      dovecots don't handle it properly. they either assert at index-sync.c:42
      or if new mails have also been added since, they don't notice it at all
      actually, that breaks reads as well since we get expunges only from
      the old file.. and check that deleting file does "inconsistency error"
    - if imap process notices that both modify logs are getting full because
      it's client isn't syncing, the client should be disconnected
    - what happens if .customflags can't be locked while opening index?
    - we don't handle out of memory conditions too well, malloc failing kills
      the process which is good enough (and likely never happens), but mmap()
      failures aren't handled too well. Rather should be handled in similiar
      way to locking failures, so that at least we don't try to rebuild the
      index because of it.
    - limit folder hierarchy levels? user can now create eg. a/a/a/a/...
      and then start renaming them from end to beginning, which probably will
      at some point start causing syscall failures which will fill up logs.
    - fsck should check binary tree

 - checks:
   - if we have entries in modifylog with UID 10..11, 9..12, 8..13 etc.
     do they work correctly?
   - check that search message-id worked properly always
   - check that search's OR and () work properly
   - Should SEARCH SENT* apply timezone?
   - make sure SELECT rebuilds index properly when next_uid is near 32bit value
   - make sure connection limits work

 - enhancements:
    - option to disable SORT, SEARCH and other memory/cpu-intensive features.
      defaults and per-user by imap-auth.
    - optionally don't fail if index is locked, but build it in memory
    - when fetching body/envelope/etc we could try to cache it immediately if
      we can get lock with try_lock.
    - optionally use only in-memory indexes
    - maildir could support also the dirty-flag in messages. files would be
      renamed "whenever there's time" (that'd require the indexer program, or
      forking and doing it in background)
    - optionally keep the message file name as it's UID. Then we don't have to
      save the filename anywhere.
    - send EXISTS immediately after new mail arrives.
        - linux: we can use dnotify for maildir (but not mbox I think, we'd
	  get interrupted all the time if we checked eg. large /var/spool/mail)
	- *bsd: kqueue() can notify changes in mbox and maildir
    - .subscriptions currently uses fcntl() locking - maybe we should instead
      just write to temp file and rename()? optionally at least, so it
      works with NFS.
    - OpenSSL: support generated DH parameters
    - multiline headers can cause our memory usage to go up. that should be
      fixed somehow. try to change things to be able to handle one line at a
      time? Well, other IMAP servers have same problem - post 1.0 problem.
    - check with strace what dovecot does when evolution checks new mail,
      it's quite a lot. some things probably wouldn't need to be done
      (mkdirs/symlinking inbox) and other things could be cached in memory.
    - sort: we could create alternative binary tree file(s) for different sort
      conditions, ".tree-sort" or something. sort code itself already supports
      this optimization.
    - sort: we could use some more memory so we don't need to parse the
      envelopes over and over again. for example with subject sorting we could
      store the 4 first bytes of message into integer and compare it. only when
      they compare equal we'd fetch the full subject and compare it.
    - tree file: should we instead use b+-tree or something similiar? or at
      least try to do some defragmentation with it, so that the root nodes
      would be kept at the beginning of the file.
    - use vsftpd-like safebufs, ie. keep non-rwx page before and after the
      memory we want to use.
        - mmap_anon()
        - mmap()ing files would probably need to first go through anon_mmap()
	  and then use MAP_FIXED. annoying that it slows the mmaping..
	- data stack should use mmap_anon()
    - option: copy /var/mail/$user to INBOX when logged in. nice for not missing
      any mails with quota enabled

 - lib-charset:
     - utf8_toupper() is a must. and a bit difficult if we want to do it right.
     - add support for other things than iconv() as well? we could reuse
       the code from cyrus or courier
     - cache iconvs? they'd probably be faster if we just reset the
       conversion instead of opening new one every time. and there will likely
       be only one or two charsets which are used for nearly all conversions.

 - should we allow following symlinks in mbox/maildirs? they are now.
     - if we implement shared mailboxes with shared indexes, never do that or
       others could symlink your personal mailboxes and see the indexes
       created for it which may contain envelope etc. data
     - this allows circular mailbox hierarchies which should be prevented by
       eg. allowing max. 20 hierarchies.

 - allow index files to be in completely separate location than mail data.
   mails could be read through slow NFS access but indexes from fast local
   disk. with this thinking it makes more sense to create larger index files
   to save for example mail headers. also index rebuilding should be very
   light operation, the indexes would be filled while the data is being
   accessed by the imap client. of course all this should be optional so
   we don't slow down when mails and indexes are stored in same disk.

 - we need permanent storage for UIDs. with mbox use X-UID like UW-IMAP,
   with maildir a) file:2,flags,Uuid b) file,U=uid:2,flags. uid validity
   would be in .uidvalidity file. the b-case would require that to be done
   by the client moving it from new/ to cur/

 - mbox:
     - if a file isn't valid mbox and it's tried to be opened, say it in one
       line in error log, not 6..
     - empty lines at beginning of file still aren't ignored
     - UW-IMAPd writes empty spaces after X-Keywords which it uses so that
       it doesn't have to rewrite the whole file if status flags changed
       in the beginning of it. We could do that too.
     - When expunging the first message we could move the X-IMAPbase header
       to next message to avoid full rewriting later.
     - We shouldn't send X-IMAPbase, Status, X-Status, X-Keywords, X-UID, etc.
       headers to client - they may change and clients must see messages as
     - COPY 1 copies X-IMAPbase header too which isn't good idea. save() could
       actually strip this (and X-UID) while also fixing From-lines etc.
     - we need either From-line escaping or writing Content-Length when saving
     - two adjacent From-lines breaks us. not too easy to fix though.
 - we could try compressing same from/to/subject fields into a single
   location in data file. requires larger changes..
 - read-only support for mailboxes where we don't have write-access
 - we should try to avoid completely rebuilding indexes unless they're
   corrupted. especially if we later want to support some read-only boxes
   and keep the mail flags only in index file. fsck() could verify that
   records are ok, and that if data file isn't ok the record is deleted.
 - if .customflags is removed and Maildir files have custom flags, add
   "unknown1" "unknown2" etc. flags to .customflags file for each found flag
 - when index is being rebuilt, it always complains about tree/modifylog
   having wrong indexid..
 - log transferred amount of bytes. just a bit problematic who logs it, since
   imap-login does SSL transfers but not unencrypted.. could also log SSL
   settings (especially compression). login process doesn't currently know
   the userid.. maybe it should get it from auth as a reply. also log amount
   of invalid password tries if all failed.
 - if we wanted to support huge mailboxes with small memory usage, it'd now
   be possible if we just instead of mmap()ing the whole index files would
   have maybe 3-4 256k mmap()ed areas which we move based on the need.
     - should work fine with .imap.index and
     - log files aren't affected by mailbox size
     - if the tree file also kept constantly moving the nodes so that
       tree's root was at the beginning of the file, we could use this mmap
       caching with it too
     - but, is it worth the trouble really? the OS can do all this itself,
       only thing we're doing is keeping the processes virtual memory usage

 - support multiple mailbox formats and locations for one user. that would
   require support for multiple MailStorages, and since we're chroot()ed,
   usually the only way to communicate with others would be to create
   RemoteMailStorage which would use TCP/UNIX sockets to connect to another 
   imap session.
     - message_body_search() could accept multiple search keywords so we
       wouldn't need to call it separately for each one (so we wouldn't need
       to parse the message multiple times).
     - message_body_search() could support NULL MessagePart and the searching
       could be done while parsing the message. this would need changes to
       message_parse() as well.
     - could optionally support scanning inside file attachments and use
       plugins to extract text out of them (word, excel, pdf, etc. etc.)
     - use a trie index for fast text searching, like cyrus squat?
     - Create our own extension: When searching with TEXT/BODY, return
       the message text surrounding the keywords just like web search engines
       do. like: SEARCH X-PRINT-MATCHES TEXT "hello" -> * SEARCH 1 "He said:
       Hello world!" 2 "Hello, I'm ...". This would be especially useful with
       the above attachment scanning.
 - DELETE/RENAME: when someone else had the mailbox open, we should
   disconnect it (when stat() fails with ENOENT while syncing). Also deleting
   selected mailbox begins giving internal error messages.
 - RENAME INBOX isn't atomic with Maildir. And in general, RENAME can't
   move mails between different storages. Maybe support doing also using
   COPY + delete once COPY is atomic?
 - maildir: atomic COPY could be done by having transaction directories.
   Make a "tra" directory at the same level as cur/new/tmp, and make it
   have subdirectories in the same way as tmp has temp files. Directory
   begins with a "." as long as transaction isn't finished, rename()ing
   it away finishes it. All mails under finished dirs must be moved into
   new/ directory and the directory removed by any process who notices them.
 - we should probably do some light checking that appended mails actually
   look like valid rfc822 mails..
 - maybe limit the length of custom flags? we don't really have a problem
   with them, but with mbox a long X-IMAPbase could break something.. Maybe
   configurable, default to 50 chars?
 - we could send flag changes after all commands by making expunge/flags sync
   counters separate for modify log. flags would need to update the seq
   though, too slow?
 - things calling message_send() could verify that it wrote enough data.
   if not, fill the rest with spaces and return failure. -1 = error,
   0 = filled, 1 = ok.

 - sieve (rfc3028)
 - rfc2231 continuation support
 - rfc2557 support for BODYSTRUCTURE, as specified by latest IMAP4rev1 draft

 - create indexer binary
 - should we bother checking if there's invalid 8bit headers in
   BODY/BODYSTRUCTURE output and converting them to quoted printable? well,
   several of them are now but not all..
 - update docs/index.txt
 - support Maildir++ quota
 - maybe give more untagged NO/ALERT replies? like when mailbox is in
   inconsistent state. and when UIDs are reordered because they're too large.
 - imap/ and lib-imap/ should allow infinite number of custom flags, it's
   storage's problem if it can't handle too many of them.

auth / login:
 - kchuid, SRP, anonymous SASL
 - PAM: support some options so /etc/passwd-lookup isn't needed. uid=x, gid=y,
   mailroot=/var/mail. maildirs should be then created when needed
 - Digest-MD5: support integrity protection, and maybe crypting. Do it
   through imap-login like SSL is done?
 - for invalid user/pass, wait for a while before giving a reply to user
 - imap-auth should limit how fast authentication requests are allowed from
   login processes. especially if there's one login/connection the speed
   should be something like once/sec. also limit how fast to accept new

cleanups / checks:
 - grep for FIXME
 - check if t_push()/t_pop() should be added somewhere
 - ..wonder what it would look like if I did s/FooBarBaz/struct foo_bar_baz/..
 - rfc822-tokenize: make it work with callbacks instead?
 - try to fix @UNSAFE code to use buffer API instead
     - subscription-file.c, custom_flags, ioloop
     - [io]stream-file.c?

optional optimizations:
 - provide some helper binary to save new mail into mailboxes with CR+LF
   line breaks?
 - disk I/O is the biggest problem, so split the mail into multiple computers
   based on user and have a proxy in the front redirecting the connection.
   cyrus had something like this except a lot more complicated - it tried
   to fix the problem of having shared mailboxes. we have the same problem
   with local shared mailboxes as we don't use same UID for everyone's mail
   and we may be chrooted, so locally we could communicate with UNIX sockets,
   remotely that could be done with TCP sockets.

 - preferrably all should be possible to #ifdef away by a configure
   option (--without-capabilities=acl,namespace,...)
 - possibility to disable them from config file
 - acl (rfc2086, draft-ietf-imapext-acl), namespace (rfc2342)
     - probably do it like cyrus. "user.<username>" to access other
       users, with "" defaulting to "user.<myself>". these should be
       configurable however.
     - shared namespaces? maybe configurable in config file
     - easiest way to do ACL would be to use unix modes, but is that
       useful at all? Well, ACL2 has a bit better support for that, so
       maybe we could support it.
     - otherwise gets a bit trickly, we could keep all mail in "imapmail"
       group and 0600/0700 mode by default, but when mail is shared to others,
       the group read/write access bits would be set. or alternatively we
       could launch another imap process to handle it, which we should support
       anyway. ACLs could be stored into ".acl" ascii file in each folder.
     - support for private and shared flags, configurable by mailbox admin.
       this isn't in any draft yet, but ACL2 author was going to create one.
       [SHAREDFLAGS (...)] would specify which ones are shared, don't know yet
       how they would be configured.
 - quota (rfc2087, draft-cridland-imap-quota)
     - give filesystem values only to admins
     - support for Maildir++, probably no need to support more.
       quota capability supports complex quota configuration, but if
       no mailer supports them we probably shouldn't bother either
 - id (rfc2971)
     - must be configurable what gets sent, default to only name=Dovecot
     - separate pre/post-login settings
     - optionally log configured parts of the client information, but only
       once, probably at the same time as logging "Logged in",
       "Disconnected", etc.
     - remember to force truncating values longer than 30 chars,
       especially before logging
 - mailbox-referrals (rfc2193)
     - this is useful whenever we would otherwise need to make the
       connection ourself. for example load balancing and shared mailboxes
       requiring another UID to run.
     - this rfc defines no exact way for server to detect if client
       supports referrals or not. I don't think there's much point in
       supporting only referrals, as most clients don't support them.
       Instead we should return referrals when we know that client
       supports them, otherwise do the connecting ourself. If client
       issues RLIST or RLSUB command, it's safe to assume it supports
     - for load balancing this works just fine, but what about shared
       mailboxes which require different UID? If we login with our own
       username, we end up with our own UID instead of what we wanted.
       IMAP URLs don't support separated authorization id which would
       have made this very easy.. We could give the "userid@group" as
       userid, but clients probably treat it as different userid and
       ask the password again.
     - problems, problems, .. maybe not worth the trouble.
 - literal+ (rfc2088)
     - simple. in case of invalid data, just disconnect client.
 - idle (rfc2177)
     - just call the syncing every few seconds (configurable)
     - with Linux we can use fcntl() and F_SETSIG to provide fast checks.
       just make sure sync() still won't be called more than once in a
       few seconds
 - uidplus (rfc2359)
     - uid expunge: no problem
     - append, copy: oh no. these would slow down things and make
       handling them much more difficult. currently we just store the
       mails to destination mailbox without touching the indexes. since
       we'd need to know their final UID, we'd have to lock the indexes
       and mbox) fsck() first and append() next to find out the uid,
       maildir) move the mail directly into cur/ and index it.
 - unselect (no draft or anything AFAIK)
     - like CLOSE, but doesn't expunge mails. easy.
 - drafts:
     - multiappend (draft-crispin-imap-multiappend)
	 - shouldn't have any problems
     - listext (draft-ietf-imapext-list-extensions)
	 - well, it expired January 2002.. I like it though.
     - children (draft-gahrns-imap-child-mailbox)
	 - I like listext more.. They have the same functionality though,
	   so pretty easy to support both if needed
     - annotate (draft-ietf-imapext-annotate)
	 - per-message annotations. this will be major change. especially
	   because currently there's no suitable storage for them, and
	   they'll probably change all the time.. maybe if we moved into
	   berkeley db to store the .data file and these annotations.
     - annotatemore (draft-daboo-imap-annotatemore)
	 - server and per-mailbox annotations. much easier than
	   per-message annotations, but they'd be easier to place into
	   db as well.
     - binary (draft-nerenberg-imap-binary)
	 - perhaps not too useful. I'd like to make Dovecot fully
	   binary-safe though.
     - thread (draft-ietf-imapext-thread)
         - basically SORT but reply with thread lists
	 - possibly use a binary tree too .. or maybe it's enough to use the
	   sort-tree and then just pick up the references separately? have to
	   check more carefully later.
     - view (draft-ietf-imapext-view)
         - slow, complex, luckily draft expired almost two years ago.
	   i hope i don't have to implement this :)
	 - can be done client-side just fine (evolution's virtual folders)