WISHLIST

Wish list:

	Things to do before the stable release:

	Remove this file from the stable release.

	Things to do after the stable release:

	Spellcheck and double-word check.

	Don't forget Apple's code donation for fetching mail from
	IMAP server.

	Make "rename" the default when postmapping a DB file
	(later: use copy+rename for postmap -i, postmap -d).

	Service-name parameters aren't documented in daemon manpages.

	When faking up the DSN ORCPT, don't send bare usernames
	from local command-line submission.

	lmtp_assume_final is broken. A 2XX response does not imply
	final delivery. The Sieve language implements accept-then-bounce.

	postscreen event-driven plug-in interface to send out a
	query in parallel with the Pregreet and DNSBL tests, using
	a simplified version of the policy delegation protocol.

	Parallelized queue preprocessing: rip out the queue manager
	code to read queue files and resolve recipients, and run
	it in parallel processes. The queue manager then processes
	their results as they become available.  This would eliminate
	the qmgr<->trivial-rewrite bottleneck. This can also eliminate
	much of the scheduling disadvantage of a single queue manager
	compared to hundreds of mail receiving or sending processes
	(especially if there is a way to scan the queue in parallel).

	Memory pools for same-type memory objects.  This can be
	used to either increase memory locality for frequently-allocated
	objects (MRU allocation) or to make use-after-free bugs
	more detectable (use LRU allocation and wipe the object
	immediately after free().  Finally, same-type memory pools
	prevent object type errors with use-after-free bugs.

	"no-cache" option for selected postscreen tests?

	Need a new DICT flag to indicate that a map handle supports
	locking.  If it doesn't (as with memcache or proxymap
	handles), then postscreen etc. don't need to close a cache
	file after "postfix reload".  After a fork() it is OK to
	keep using a memcache or proxymap handle, because the parent
	exits immediately.  For this to work, the memcache client
	needs to propagate the flag from a persistent backup map,
	but the proxymap protocol should not propagate this to the
	client.

	Different TTL values for different DNSBL sources?

	Replace master(8) SIGHUP by very simple socket protocol to
	allow reload of a specific service.

	postscreen: in the dummy SMTP engine, log the protocol state
	at time of violation (like smtpd, set state->where initially
	to CONNECT, then update it with the name of the last "known"
	command, or set it to "unimplemented").

	The discussion of postscreen cache configuration is in the
	wrong place (how whitelisting works). Move it to the section
	about configuring postscreen.

	Before proxymap can be exposed to the network (primarily
	to share postscreen or verify caches), need to enforce
	limits on attribute string name and value length in IPC
	protocols.  10-20KB seems OK. We need to enforce content
	sanity checks (for example, no control characters; Postfix
	does not pass around multi-line data in table lookups). The
	VSTREAM library already supports read/write deadlines.  We
	need to use attack-resistant code for numeric conversion.

	move flush_init() etc. from defer service clients to the
	bounce daemon? Postfix works best when work can be spread
	out over many clients, instead of over a few servers.

	multi_connect() function that takes a list of inet:host:port
	and/or unix:pathname specs, with an explicit "inet" prefix
	argument to handle applications that use host:port only.
	This will simplify multi-host implementation for memcache
	client, dovecot client, and other.

	dict_memcache: treat "bad" key as cache miss, i.e.  read/write
	the backup database as if the cache did not exist. This
	does not help because most Postfix maps (virtual, canonical,
	access, transport, ...) also don't support spaces in keys.

	postscreen: keep the cache open after "postfix reload" when
	it is remote (type memcache: or proxy:). This does not work
	because memcache can use a non-proxied file as backup).

	What is the feasibility of adding an mta_name (personality)
	attribute that is propagated via queue files and delivery
	agent requests? It would default to myhostname.

	Major performance improvement opportunity (that is until
	everyone runs Postfix queues on SSDs).  Investigate the
	viability of a daemon that produces incoming and postdrop
	queue files on request (in reality it would maintain a
	limited queue of "spare" files). Central queue file allocation
	reduces the I/O performance disadvantage that qmgr has when
	100 smtpd processes are receiving mail, or when lots of
	mail is submitted with the sendmail command line.  When an
	smtpd process accepts MAIL FROM, a cleanup daemon requests
	a queue file and receives a queue ID + file handle from the
	queue file daemon.  If the queue file daemon is down, the
	cleanup daemon creates the file itself like it does now;
	this can be hidden in the mail_stream library module.  If
	the mail transaction is aborted, then the cleanup daemon
	gives the queue file back to the queue file daemon's "spare"
	file pool, saving most of the overhead of creating and
	deleting a queue file (the file would still need to be
	renamed at the start of the next mail transaction).  If the
	cleanup daemon is unable to give a file back, then it can
	delete the file like it does now; this can be hidden in the
	mail_stream library module.  The whole thing can be
	transparently added to Postfix by adding calls to a
	queue-file-service client to the mail_queue_enter() and
	mail_queue_remove() library routines.  Other advantages:
	1) negligible performance hit when queue file allocation
	happens earlier, so that logging and milters have a queue
	ID for the whole transaction not just the first valid
	recipient; 2) by not removing every queue files we get most
	of the performance gain of a queue based on append/truncate
	instead of the much more expensive create/delete.

	Investigate viability of Sendmail dns maps.

	Check if FILTER_README has the "postsuper -r" workaround

	Bounces without <> in the plaintext section. Apparently,
	some software renders the text as HTML (and therefore 
	does not render addresses and other text inside <> ).

	Make the rules for how to use close-on-exec more explicit.

	Provide separate timeout control for dict_proxy client,
	rewrite client, resolve client, cleanup client, and so on.
	Perhaps a timeout argument to the mail_connect() routines.

	Trick from amavisd: save listen socket/fifo/etc state, clear
	their close-on-exec flags, exec the same program file to
	re-initialize (with saved socket state on command line or
	in environment), then restore the listen socket/fifo/etc
	close-on-exec flags.  This could be a way to mitigate the
	impact of memory/file leaks, and to implement "postfix
	reload" support for master(8) features that currently don't
	support this. 

	Sub-second time resolution. The first benefit is to make
	per-destination rate delays more usable. Other applications
	will come up once the support exists. The straightforward
	approach is to represent all time intervals in milliseconds,
	and to update all code that makes system calls with a time
	argument (as well as the compiled-in upper and lower time
	parameter bounds, which are currently in seconds).
	Unfortunately, that limits he maximum time interval to less
	than 25 days on 32-bit systems, and is likely to break
	compatibility (for starters, it cannot even deal with the
	compiled-in 100d upper bound on the queue file lifetime).
	A second option is to have a "compatibility" time base
	switch between milliseconds and seconds; this means extra
	changes to all code that makes system calls with a time
	argument, and the way that the compiled-in upper and lower
	bounds are specified.  Some of this can be encapsulated in
	macros like time_to_sec(t), time_to_msec(t) and sec_to_time(t).
	Finally, it is relatively easy to replace the events(3)
	interface to use "double" for the time delay arguments, but
	it is a major pain to convert all main.cf time parameters
	into doubles (converting only some leads to a documentation
	nightmare).

	Address verify cache: allow a negative cache "refresh"
	result to purge a "positive" cache entry in some safe manner.
	Currently, the negative cache "refresh" result is discarded,
	address verify cache lookup returns OK, and each lookup
	forces a "refresh" probe until the entry expires.

	Some Sendmail configurations trigger sub-optimal behavior
	when the postscreen_whitelist_interfaces parameter lists
	primary MX addresses only.  When postscreen's "deep protocol
	tests" are successful on the primary MX address (i.e. they
	result in 4XX responses to RCPT TO), some Sendmail
	configurations keep the primary MX connection open until
	AFTER they finish talking to the backup MX address.  The
	problem is that the backup connection runs into a WHITELIST
	VETO condition because the whitelisting database has not
	yet been updated with the PASS NEW result for the primary
	MX connection.  Unfortunately postscreen can't update the
	whitelisting database before the primary MX connection is
	closed, because a client may still make a mistake.

	Simplify postscreen logic. Individual "fail" flags help to
	avoid repeated testing/logging the same mistake.  Individual
	"pass" flags provide evidence that the client didn't skip
	tests by hanging up early. The current global "noforward"
	flag implements the wrong model: instead we need an indicator
	that a client has passed all tests or that all mistakes
	were forgiven.

	In the SMTP server, check if the connection is closed before
	replying to ".", and discard the message if the reply can't
	be sent. This reduces the time window for RFC 1047 message
	duplication, and may even prevent the delivery of some spam.
	http://www.exim.org/lurker/message/20070416.103159.9d5ff0ce.en.html
	This requires splitting the SMTP server's commit operation
	into two operations: first, a tentative commit operation
	that performs most of the I/O and processing in milters and
	in the cleanup server; second, a final commit operation
	that is executed only if the remote SMTP client hasn't hung
	up in the mean time. Unfortunately, SMTP-based before-queue
	content filters don't support a tentative commit operation.

	Find out how to reproduce Berkeley DB bogus ENOENT errors.
	postscreen does not log this with Berkeley DB 1 (FreeBSD
	4..8), 4.7.25 (Ubuntu 9.04) and 4.8.24 (Ubuntu 10.04).

	postconf command-line option to show the compile-time
	settings (CCARGS, AUXLIBS) in case binary packages
	don't install the makedefs.out file.

	events.c: cache the side effects of file descriptor event
	enable/disable operations in user space, and do bulk kernel
	updates at event_loop() time.  This can eliminate costly
	system calls with successive event disable/enable operations
	on the same file descriptor. This can also eliminate the
	need for tricky code that tries to avoid the expense of
	successive disable/enable operations. Such code is likely
	to introduce bugs.

	When does it pay off to send domains in the active queue
	to a DNS prefetch daemon? Could this generalize to a dynamic
	transport map that piggy-backs domains with the same MX
	host into the same mail delivery transaction?

	inline table where the "whitespace replacement" character
	is specified in-line. Ex: inline:XYname1Xvalue1Yname2Xvalue2
	would instantiate a table with (name1, value1) and (name2,
	value2).  I'm afraid this is just too ugly.

	tlsproxy(8) should receive TLS preferences from postscreen(8)
	and smtpd(8), instead of reading them from main.cf. This
	means that many tlsproxy_ parameters become postscreen_
	parameters, and that tls_server_init() parameters move to
	to tls_server_start(). That is a significant API change.
	It also means tlsproxy can't open all files before chroot().

	anvil rate limit for sasl_username.

	Encapsulate nbbio buffer access and update by tlsproxy.

	Full-duplex support for tlsproxy(8). This requires updating
	events(3) and nbbio(3).

	Register automagic destructor for object attached to VSTREAM.

	Use different ipc time limits for email message transactions
	(smtpd, pickup)->cleanup and for quick query/reply transactions
	such as address rewriting/resolution. Beware of large time
	limits for local or virtual alias expansion.

	permit_tempfail_action (default: defer_if_reject) to be
	used as the default value for dnswl_tempfail_action and
	rhswl_tempfail_action. Steal liberally from the code that
	implements unverified_recipient_tempfail_action etc.

	Support filtering of messages that are generated by Postfix:
	This would apply to postmaster notices and bounce messages
	(DKIM), and address verification (BATV).

	Consistency: in postconf.proto make <dt>..</dt> tags bold.

	postscreen(8): listen on multiple IP addresses and enforce
	that the client contacts the primary MX address first (i.e.
	punish hosts that contact the secondary before the primary).
	The downside with any approach that relies on temporary
	punishment is that it does not scale to configurations
	with multiple equal-preference MX hosts. Such hosts would
	have to share the postscreen cache, causing an unacceptable
	performance bottleneck and a single point of failure.

	According to a paper by Ted Unangst at BSDCON09, kqueue
	reports state changes, i.e. kqueue indicates when the socket
	becomes readable. Specifically, he writes when kqueue reports
	a socket becomes readable but no data is read from that
	socket, later kqueue calls won't report the socket as
	readable. That's not what happens on FreeBSD 8.0, where
	kqueue will keep reporting the socket as readable when
	nothing is read.  Also, FreeBSD 8.0 kqueue still reports
	the socket as readable after a read operation does not empty
	the kernel buffer.  We need a test program for this that 
	repeats these tests with OpenBSD and NetBSD (and MacOS X
	once they fix their kqueue implementation).

	Would it help if there were different cleanup_service
	parameter names for different message paths? smtpd(8) uses
	the same cleanup_service value for receiving remote mail
	and for submitting postmaster problem reports. Do we need
	separate mumble_cleanup_service_name parameters for "inject",
	"notify" and "forward" (with backwards compatible defaults)?

	IF/ENDIF support for CIDR tables.

	Need a regular expression table to translate address
	verification responses into hard/soft/accept reply codes.

	Is there a way to make sendmail -V work after local alias
	expansion? Majordomo-like mailing lists would benefit from
	this; the example in VERP_README does not work in the general
	case.

	When an alias is a member of an :include: list with owner-
	alias, local(8) needs an option to deliver alias or alias->user
	indirectly. What happens when an :include: list with owner-
	alias includes another list?

	Don't allow empty result values in pcre and regexp maps.
	Postfix doesn't allow them anywhere else (check this).

	Make PCRE_MAX_CAPTURE configurable.

	Add some checks for tokens starting with #. A challenge
	is to report sensible context from the guts of some low-level
	parser, without introducing a great deal of clumsiness.

	Add sendmail macros for {verify} and maybe other TLS info.

	Find out if we are doing the correct thing by looking at
	state->milter_reject_text when expanding {rcpt_addr} or
	{rcpt_host}.

	Find out why post_mail() etc. block when the qmgr fifo is
	full (answer: trigger_timeout). How can this cause delays
	in the queue manager?  When a recipient bounces during
	(transport, nexthop, address) resolution, it is redirected
	to the error or retry mailer; and bounce-after-delivery is
	asynchrounous so it can't block the queue manager, either.

	How to ensure that proxy_read_maps is processed after all
	its dependencies are initialized, or just bite the bullet
	and rewrite the parameter initialization code.

	The cleanup virtual alias expansion limit does not really
	deliver on its promises. 1) It promises to truncate the
	result without aborting delivery, which would be undesirable
	anyway, but that is not what it does, so that is good.  2)
	It keeps all the recipients from multi-recipient database
	lookup, then terminates further recursion when the result
	exceeds the expansion limit. This behavior achieves the
	original goal that all things shall have a finite size (even
	though but we don'really care how large they are) but may
	result in surprises when recipients are listed in virtual
	alias domains or need expansion for other reasons.  In a
	phone call with Victor, a reasonable way out is to set the
	limit to some large number (100000) and abort delivery when
	the result exceeds the limit.

	Should the postscreen save permanent white/black list lookup
	results to the temporary cache, and query the temporary
	cache first? Skipping white/black list lookups will speed
	up the handling of "good" clients without a permanent
	whitelist entry.  Of course, this means that updates to the
	white/black lists do not immediately take effect. Workarounds:
	1) use a shorter temporary cache TTL for clients on the
	permanent black/white lists; 2) ignore cached white/black
	list lookup results after "postfix reload"; 2) adjust the
	logging, for example "WHITELISTED address (cached)" and
	"BLACKLISTED address (cached)" to eliminate surprises.
	Comparing the cache entry time with the white/blacklist
	file modification time is not foolproof: for example, pcre
	or CIDR tables are read only once.

	It would be nice if the generic dict_cache(3) cache manager
	could postpone process suicide until cache cleanup is
	completed (but that is not possible when postscreen forks
	into the background to finish already-accepted connections,
	and it is not desirable when a host is being shut down).

	When postscreen drops a connection, a 521 "greeting" should
	be of the form "521 servername..." and not have an enhanced
	status code. The "521 5.7.1" form can be used after EHLO.
	Of course no spammer is going to complain about Postfix
	SMTP compliance.

	Find a place to document all the mail routing mechanisms
	in one place so people can figure out how Postfix works.

	The access map BCC action is marked "not stable", perhaps
	because people would also expect BCC actions in header/body_checks.
	How much would it take to make the queue file editing code
	generally usable?

	Move smtpd_command_filter into smtpd_chat_query() and update
	the session transcript (see smtp_chat_reply() for an example).

	SMTP connection caching without storing connections, to
	improve TLS mail delivery performance.

	Should not milter8_mail_event() unset the "hold" default
	reply? Better, the default reply should not be used for
	this purpose.

	Don't send MASTER_STAT_TAKEN/MASTER_STAT_AVAIL when a server
	runs with process limit of 1. But this means the master
	never learns that the process is successful and will always
	pause $service_throttle_time before restarting a failed service.

	Don't bother maintaining a per-service lockfile when a
	server runs with process limit of 1. The purpose of the
	lockfile is to avoid thundering herd problems when the kernel
	wakes up multiple processes for each new client connection.

	Implement PREPEND action for milter_header_checks. Save the
	to-be-prepended text to buffer, then emit it along with the
	new header.

	Fix the header_body_checks API, so that the name of the map
	class (e.g. milter_header_checks) is available for logging.

	Fix the mime_state and header_body_checks APIs, so that
	they use VSTRINGs. This simplifies REPLACE actions.

	Update FILTER_README for multi-instance support, and rename
	the old document to FILTER_LEGACY_README.

	Need to sign delivery status notifications, to avoid surprises
	when eventually people start enforcing DKIM etc. signatures.

	Either document or remove the internal_mail_filter_classes
	feature (it's disabled by default).

	Make the "unknown recipient" test configurable as
	first|last|never, with "yes"=="last" for backwards
	compatibility. The "first" setting is good for performance
	(stress=yes) when all users are defined in local files; but
	it may perform worse when users are in networked tables.

	Cleanup: make DNSBL query format configurable beyond the
	client's reversed IP address.

	With 'final delivery' in the LMTP client, need an option
	to also add delivered-to and other pipe(8) features.  This
	requires making mail_copy() functionality available in
	non-mailbox context.

	Cleanup: modernize the "add missing From: header" code, to
	``phrase <addr>'' form. Most likely, quote the entire phrase
	if it contains any text that is special, then rfc822_externalize
	the whole thing.

	SMTP server: make the server_addr and server_port available
	to policy server, Dovecot, and perhaps Milters.

	Med: local and remote source port and IP address for smtpd
	policy hook.

	Maybe change maps_rbl_reject_code default to 521, and
	update wording in STRESS_README.

	Encapsulate time_t comparisons so that they can be made
	system dependent (use difftime() where available).

	Encapsulate time_t conversions (e.g. REC_TYPE_TIME) so that
	they can be made system dependent.

	Plan for time_t larger than long, or wait for LP64 to
	dominate the world?

	Make "AUTH=<>" appendage to MAIL FROM configurable, enabled
	by default.

	To support ternary operator without a huge parsing effort,
	consider ${value?{xxx}:{yyy}} where ${name} is existing
	syntax, and where ?{text} and :{text} are new syntax that
	is unlikely to break existing configurations. Or perhaps
	it's just too ugly.

	Write delivery rate delay example (which _README?) and auth
	failure cache example (SASL_README). Then include them in
	SOHO_README.

	Look for alternatives for the use of non_smtpd_milters.
	This involves some way to force local submissions to go
	through a local SMTP client and server, without triggering
	"mail loops back to myself" false alarms. The advantage is
	that it makes smtpd_mumble_restrictions available for local
	and remote mail; the disadvantage is that it makes local
	submissions more dependent on networking.  One possibility
	is to use "pickup -o content_filter=smtp:127.0.0.1:10025",
	or a dedicated SMTP client/server on UNIX-domain sockets;
	we could also decide to always suppress "mail loop" detection
	for loopback connections.  Another option is to have the
	pickup or cleanup server drive an SMTP client directly;
	this would require extension of the mail_stream() interface,
	plus a way to handle bounced/deferred recipients intelligently,
	but it would be at odds with Postfix design where delivery
	agents access queue files directly; exposing delivery agents
	to raw queue files violates another Postfix design principle.

	Consolidate duplicated code in *_server_accept_{pass,inet}().

	Consolidate duplicated code in {inet,unix,upass}_trigger.c.

	In the SMTP client, handle 421 replies in smtp_loop() by
	having the input function raise a flag after detecting 421
	(kill connection caching and be sure to do the right thing
	with RSET probes), leave the smtp_loop() per-command reply
	handlers unchanged, and have the smtp_loop() reader loop
	bail out with smtp_site_fail("server disconnected after
	%s", where), but only in the case that it isn't already in
	the final state. But first we need to clean up the handling
	of do/don't cache, expired, bad and dead sessions.

	Combine smtpd_peer.c and qmqpd_peer.c into a single function
	that produces a client context object, and provide attribute
	print/scan routines that pass these client context objects
	around. With this, we no longer have to update multiple
	pieces of code when a client attribute is added. Ditto for
	SASL and TLS context.

	Don't log "warning: XXXXX: undeliverable postmaster
	notification discarded" for spam from outside.

	Really need a cleanup driver that allows testing against
	Milter applications instead of synthetic events. This would
	have to provide stubs for clients that talk to Postfix
	daemon processes. See if this approach can also be used for
	other daemons.

	smtpd(8) exempts $address_verify_sender from access controls,
	but it doesn't know whether cleanup(8) or delivery agents
	modify the sender. Would it be possible to "calibrate" this
	exemption, perhaps by having delivery agents pass the probe
	sender to the verify server, keeping in mind that the probe
	sender may differ per delivery agent due to output rewriting.

	Update attr_print/scan() so they can send/receive file
	descriptors. This simplifies kludgy code in many daemons.

	Would there be a problem adding $smtpd_mumble_restrictions
	and $smtpd_sender_login_maps to the default proxy_read_maps
	settings?

	Remove defer(8) and trace(8) references and man pages. These
	are services not program names. On the other hand we have
	man pages for lmtp(8) and smtp(8), but not for relay(8).
	Likewise, retry(8) does not have a man page.

	Bind all deliveries to the same local delivery process,
	making Postfix perform as poorly as monolithic mailers, but
	giving a possibility to eliminate duplicate deliveries.

	Maybe declare loop when resolve_local(mxhost) is true?

	Update message content length when adding/removing headers.

	Need scache size limit.

	REDIRECT should override original recipient info, and
	probably override DSN as well.

	Update FILTER_README with mailing list suggestions to tag
	with a badness indicator and then filter down-stream.

	Make null local-part handling configurable: either expand
	into mailer-daemon (current bahavior) or disallow (strict
	behavior, currently implemented only in the SMTP server).

	The type of var_message_limit (and other file size/offset
	configuration parameters or internal protocol attributes)
	should be changed from int to off_t.  This also requires
	checking all expressions in which var_message_limit etc.
	appears: qmqpd, netstring, deliver_request, ...

	Add M flag (enable multi-recipient delivery) to pipe daemon.

	The usage of TLScontext->cache_type is unclear. It specifies
	a TLS session cache type (smtpd, smtp, or lmtp), but it is
	sometimes used as an indicator that TLS session caching is
	unavailable.  In reality, that decision is made by not
	registering call-back functions for cache maintenance.

	Postfix TLS library code should copy any strings that it
	receives from the application, instead of passing them
	around as pointers. TLScontext->cache_type is a case in
	point.

	Are transport:nexthop null fields the same as in the case
	of default_transport etc. parameters?

	Don't lose bits when converting st_dev into maildir file
	name. It's 64 bits on Linux. Found with the BEAM source
	code analyzer. Is this really a problem, or are they just
	using 64 bits for upwards compatibility with LP64 systems?

	Do or don't introduce unknown_reverse_client_reject_code.

	Check that "UINT32 == unsigned int" choice is ok (i.e. LP64
	UNIX).

	Tempfail when a Milter application tries to negotiate content
	access, while it is configured in an SMTP server that runs
	before the smtpd_proxy filter.

	Log DSN original recipient when rejecting mail.

	Keep whitespace between label and ":"?

	Make the map case folding/locking options configurable, if
	not at run-time then at least at compile time so we get
	consistent behavior across applications.

	Investigate what it would take to eliminate oqmgr, and to
	make the old behavior configurable in a unified queue
	manager.  This would shave another 2.7 KLOC from the source
	footprint.

	Document the case folding strategy for match_list like
	features.

	Eliminate the (incoming,deferred)->active rename operation.
	This requires an in-memory hash of queue file names to avoid 
	duplicate open() operations.

	Softbounce fallback-to-ISP for SOHO users. This heuristic
	assumes that when direct-to-MX delivery fails with 5XX,
	delivery via the ISP may still succeed.  This could be
	implemented by enabling soft bounces for destinations other
	than the smtp_fallback_relay. So the only benefit of this
	over the existing soft_bounce feature is that it has no
	effect on smtp_fallback_relay deliveries.

	Centralize main.cf parameter input so that defaults work
	consistently. What about parameter names that are prefixed
	with mail delivery transport names?

	Fix default time unit handling so that we can have a default
	bounce lifetime of $maximal_queue_lifetime, without causing
	panics when a non-default maximal_queue_lifetime setting
	includes no time unit.

	After the 20051222 ISASCII paranoia, lowercase() lowercases
	ASCII text only.

	Privacy: remove local command/pathname details from remote
	delivery status reports, and log them via local msg_warn().

	Is it safe to cache a connection after it has been used for
	more than some number of address verification probes?

	Try to recognize that Resent- headers appear in blocks,
	newest block first. But don't break on incorrect header
	block organization.

	Hard limits on cache sizes (anvil, specifically).

	Laptop friendliness: make the qmgr remember when the next
	deferred queue scan needs to be done, and have the pickup
	server stat() the maildrop directory before searching it.

	Low: replace_sender/replace_recipient actions in access
	maps, so they can be used in policy servers?

	Low: configurable order of local(8) delivery methods.

	Med: smtp_connect_timeout_budget (default: 3x smtp_connect_timeout)
	to limit the total time spent trying to connect.

	Med: transform IPv4-in-IPv6 address literals to IPv4 form
	when comparing against local IP addresses?

	Med: transform IPv4-in-IPv6 address literals to IPv4 form
	when eliminating MX mailer loops?

	Med: Postfix requires [] around IPv6 address information
	in match lists such as mynetworks, debug_peer_list etc.,
	but the [] must not be specified in access(5) maps. Other
	places don't care.  For now, this gotcha is documented in
	IPV6_README and in postconf(5) with each feature that may
	use IPv6 address information. The general recommendation
	is not to use [] unless absolutely necessary.

	Med: the partial address matching of IPv6 addresses in
	access(5) maps is a bit lame: it repeatedly truncates the
	last ":octetpair" from the printable address representation
	until a match is found or until truncation is no longer
	possible.  Since one or more ":" are usually omitted from
	the printable IPv6 address representation, this does not
	really try all the possibilities that one might expect to
	be tried. For now, this gotcha is documented in access(5).

	Low: reject HELO with any domain name or IP address that
	this MTA is the final destination for.

	Low: should the Delivered-To: test in local(8) be configurable?

	Low: make mail_addr_find() lookup configurable.

	Low: update events.c so that 1-second timer requests do not
	suffer from rounding errors. This is needed for 1-second
	SMTP session caching time limits. A 1-second interval would
	become arbitrarily short when an event is scheduled just
	before the current second rolls over.

	Low: configurable internal/system locking method.

	Low: add INSTALL section for pre-existing Postfix systems.

	Low: add INSTALL section for pre-existing RPM Postfixes.

	Low: disallow smtpd_recipient_limit < 100 (the RFC minimum).

	Low: noise filter: allow smtp(8) to retry immediately if
	all MXes return a quick ECONNRESET or 4xx reply during the
	initial handshake. Retry once? How many times?

	Low: make post-install a "postfix-only script" so it can
	take data from the environment instead of main.cf.

	Low: randomize deferred mail backoff.

	Med: separate ulimit for delivery to command?

	Med: postsuper -r should do something with recipients in
	bounce logfiles, to make sure the sender will be notified.
	To be perfectly safe, no process other than the queue manager
	should move a queue file away from the active queue.

	This could involve tagging a queue file, and use up another
	permission bit (postsuper tags a "hot" file, qmgr requeues it).

	Low: postsuper re-run after renaming files, but only a
	limited number of times.

	Low: smtp-source may block when sending large test messages.

	Med: find a way to log the sender address when MAIL FROM
	is rejected due to lack of disk space.

	Low: revise other local delivery agent duplicate filters.

	Low: all table lookups should consistently use internalized
	(unquoted) or externalized (quoted) forms as lookup keys.
	smtpd, qmgr, local, etc. use unquoted address forms as keys.
	cleanup uses quoted forms.

	Low: have a configurable list of errno values for mailbox
	or maildir delivery that result in deferral rather than
	bouncing mail. What about "killed by signal" exits?

	Low: after reorganizing configuration parameters, add flags
	to all parameters whose value can be read from file.

	Medium: need in-process caching for map lookups. LDAP servers
	seem to need this in particular. Need a way to expire cached
	results that are too old.

	Low: generic showq protocol, to allow for more intelligent
	processing than just mailq. Maybe marry this with postsuper.

	Low: default domain for appending to unqualified recipients,
	so that unqualified names can be delivered locally.

	Low: The $process_id_directory setting is not used anywhere
	in Postfix. Problem reported by Michael Smith, texas.net.
	This should be documented, or better, the code should warn
	about attempts to set read-only parameters.

	Low: while converting 8bit text to quoted-printable, perhaps
	use =46rom to avoid having to produce >From when delivering
	to mailbox.

	virtual_mailbox_path expression like forward_path, so that
	people can specify prefix and suffix.