xref: /netbsd-src/external/ibm-public/postfix/dist/WISHLIST (revision 059c16a85b0b39d60ad6d18f53c09510815afa2b)

Wish list:

	Things to do before the stable release:

	make pre-release-check, HTML validator check.

	Disable -DSNAPSHOT and -DNONPROD in makedefs.

	Scan Postfix code with github.com/googleprojectzero/weggli
	(depends on "rust").

	Investigate clang-format compatibility as a possible migration
	away from indent. This requires that the output is stable.

	Check out https://github.com/milter-manager/milter-manager/

	Check out https://cutter.osdn.jp/ (C/C++ unit tests).

	Follow https://github.com/vdukhovni/postfix/commits/rpk

	Multi-recipient support in sender/recipient_bcc_maps and
	always_bcc.

	Test for dns_rr_copy() + dns_rr_free().


	mail_conf_xxx supprt for non-negative numbers (i.e.
	numbers with a lower bound of zero).

	Log anvil transgressions with their address range (in
	addition to the offending IP address. We should not disclose
	to random clients how we aggregate anvil event counters.

	Should "postconf -f" pretty-print text inside {}?

	Is there any code that calls attr_scan*() and that works
	when the number of attributes received < the expected number?
	If there is no such code, then we can simplify a few things.

	Update TLS_README diagram, tlsmgr no longer manages cert
	info.

	Consider renaming local_header_rewrite_clients to
	local_header_canonicalize_clients, as a more accurate name.
	Optionally support "local_header_canonicalize_classes =
	rewrite_addresses, add_missing_headers" (default setting).

	And ditto for remote_header_rewrite_domain, whether it
	should rewrite address, add missing headers, or both.

	Add weight factors to randmap, for example randmap:{{result1}*99,
	{result2}*1}. To parse out weights, see postscreen.

	randmap already allows randmap:{{result}, ...}, to support
	whitespace and comma in result values, but it should also
	extract the value from {}.

	Migrate masquerade_domains from ARGV to STRING_LIST, or
	deprecate this feature because it breaks table-driven address
	validation.

	Enforce var_line_limit in util/attr_scan*c. This is needed if
	we want to expose Postfix RPC protocols externally.

	Can tests use LD_PRELOAD to inject fake modules such as
	fake_dns(3), fake_msg(3), fake_myaddrinfo(3) and so on?
	One limitation is that functions etc. in a preloaded object
	always take precedence, even in code that is not being
	tested.

	'%l' support, similar to %D in the Dovecot LDAP driver.
	Subject: Feature request: '%l' expansion for ldap_table,
	Date: Tue, 5 Apr 2022.  Message-ID:
	<ef7c661c-d86a-2366-6a73-ec8d51d75012@dev.snart.me>

	WARN_IF_REJECT like prefix that disables the error counter increment.

	Consider migrating Postfix server sockets from directory
	$queue_directory/public to $queue_directory/protected. The
	directory $queue_directory/public can then be used for
	non-Postfix listeners (one subdirectory per application).

	FILTER_README needs some text on multi-instance implementations,
	and existing multi-instance references need to be updated.

	Fix code that still uses "long" for data_size and data_offset,
	and sscanf("%ld or strtou?l()). This seems relevant for 32-bit
	systems. This would use a new REC_TYPE_OFFS with a corresponding
	data type of off_t, using off_cvt() for conversion from string,
	and new code to convert off_t to string.

	A smart query service for live Postfix tables that outputs JSON?
	If the idea is to introspect on a running Postfix system, this
	involves adding an RPC endpoint to specific Postfix services.
	That could work for single-instance services like qmgr, verify,
	postscreen.

	JSON logging?

	default_transport_maps? This would simplify configuration.

	Add a pointer to
	https://fabianlee.org/2019/10/23/docker-running-a-postfix-container-for-testing-mail-during-development/
	and https://github.com/docker-mailserver/docker-mailserver

	Add a pointer to
	https://github.com/tarickb/sasl-xoauth2 and/or
	http://mmogilvi.users.sourceforge.net/software/oauthbearer.html
	in documentation or on-line howtos.

	Read the above links and see how we can improve usability on
	the Postfix side.

	Add verp=+= to the qmgr "from=" logging. This is already
	implemented but not yet integrated.

	Need canonical Dovecot example that has virtual_mailbox_domains,
	(virtual_mailbox_maps or reject unverified_recipient), and
	virtual_transport.

	Make smtpd_relay_before_recipient_restrictions settable
	in smtpd_checks tests.

	Make the DNS resolver library pluggable, so that we can a)
	plug in a fake resolver library for DNS-related regression
	tests and make DNS tests hermetic (no external dependency;
	b) add support for non-libbind resolvers. Gracefully handle
	requests for unsupported functionality; return an error status,
	instead of terminating.

	Add a robust dnssec_probe regression test (success and fail)
	that does not break existing regression tests.

	smtp_sasl_tls_security_options = noanonymous, and make
	smtp_sasl_security_options the default dependent on the
	smtp_sasl_tls_security_options default (i.e. reverse the
	dependency). Or make them independent.

	Try to make the master throttle more distrusting. Currently,
	the master throttles a service after a child process cannot be
	created (fork() fails), or if a child process fails upon its
	first use. The master always unthrottles the service if a process
	handles a client successfully. This is sufficient to mitigate
	local errors that break all attempts to use a service. It also
	slows down stupid remote attacks as long as malicious traffic
	dominates benign traffic. Perhaps monitor a crashing percentage?
	If 50% of all connections to a service result in abnormal
	termination, that would be bad even under a non-attack scenario.

	More accurate address verification: do a quota check before
	reporting that a local(8) or virtual(8) recipient is deliverable.

	Eliminate duplicate mail submission permission checks from
	sendmail, so that they happen in postdrop only. Then, pass the
	result through the postdrop-to-sendmail protocol. This requires
	that postdrop reads all inputs before responding (the
	local_login_sender_maps check depends on the envelope
	sender). Then sendmail can save input to dead.letter (no setgid
	privilege, but it would still have to use safe_open() to avoid
	clobbering files).

	Consider removing compat_level_from_numbers() and aliases,
	because they are no longer used anywhere.

	Allow '}' at the beginning of a line. This would make multi-line
	configuration settings easier to enter. This may be true
	for main.cf, master.cf and similar files (such as database
	configuration files, but not necessarily elsewhere). So it
	may have to be a readlline flag.

	Understand what happens with DNSSEC related status fields
	in posttls-finger when resolv.conf points to a host that
	runs no DNS server.

	Hardening the half-dane behavior: some sites may rely on
	current behavior which allows original MX domain name for
	certificate matches. Requires a new (compatibility) parameter
	setting?

	Code deduplication: migrate multi_server applications to
	event_server, because the multi_server and event_server
	skeletons are much more similar than other skeletons. In
	addition to the default event_server accept() handler, also
	register a read event callback for handling post_accept
	events. But the currrent multi_server API fits typical usage
	better.

	When a secondary instance has no multi_instance_name set,
	postmulti -i won't be able to find it.

	nbbio: exercise the sanity checks with fake msg(3) functions.

	optreset (bsd-ism) how badly do we need it?

	transport policy protocol (clone of check_policy).

	See also postscreen event-driven client for policy delegation
	below.

	smtp_line_length_limit can insert a line break in the middle
	of a multi-byte character (which is not necessarily UTF-8,
	so we can't simply look at the 8th bit). Also, note that a
	multi-byte character may span queue file record boundaries,
	for example if line_length_limit == smtp_line_length_limit.
	The only way to fix this is to make the smtp_text_out()
	routine aware of every possible multi-byte encoding.

	Replace ad-hoc code for pipe(8) flags handling, with
	infrastructure that was built for smtp(8).

	Move map descriptions from postconf(1) to DATABASE_README
	and point there. The text in DATABASE_README is less complete
	than that in postconf(1).

	make tls_pre_jail_init() safe by design for use in programs
	that implement both clients and servers.

	In smtpd(8) and postscreen(8), set the ehlo_discard_mask
	to ~0 so that STARTTLS, BDAT, DSN, etc. work only for clients
	that send EHLO.

	Wordsmithing: "replace by X" -> "replace with X" unless X
	is "responsible" for making the substitution.

	In postscreen, don't fork after 'postfix reload' when
	psc_check_queue_length (and psc_post_queue_length?) is zero.

	After I/O error, store errno in VSTREAM object before errno
	may be overwritten.

	Add some tips for logging from container:
	https://www.projectatomic.io/blog/2016/10/playing-with-docker-logging/;
	syslog_name = $myhostname/postfix; mkdir queue and data
	dir; postfix check to create queue subdirectories.

	Add postwhite as a postscreen-related project.
	https://github.com/stevejenkins/postwhite/blob/master/README.md

	XFORWARD attributes in policy protocol?

	Document postsrsd and postforward for srs-ifying. Would
	more fine-grained smtp_generic_maps support help?

	Decide whether to deprecate database configuration pathnames
	that start with ".", for example, ldap:./file/name. These forms
	are documented for ldap:, memcache:, mysql:, pgsql:, and sqlite:
	maps. Postfix daemon processes will look up files relative to the
	queue directory, but with postmap command-line processes it would
	be more natural to interpret relative pathnames relative to the
	current directory of the calling process (it would be a surprise
	if "postmap hash:./foo" would access "/var/spool/postfix/foo",
	or if "postmap hash:foo" and or "postmap hash:./foo" would access
	different files).

	Convert postalias(1) to store external-form keys, and convert
	aliases(5) to perform external-first lookup with fallback to
	internal form, to make it consistent with the rest of Postfix.
	In several years we may remove the internal-form fallbacks
	with a compatibility_level safety net.

	In the bounce daemon, set util_utf8_enable if returning an
	SMTPUTF8 message. This is wrong; if SMTPUTF8 is disabled,
	then Postfix must not turn it on.

	Add a header_body_checks extension callback in smtp_proto.c
	that implements the PASS action.

	Propagate SMTPD_PEER_CODE_XXX from smtpd(8) to cleanup(8),
	so that {client_resolve} and {_} produce consistent results.

	NO_IP_CYRUS_SASL_AUTH should be a main.cf parameter.

	Modeline support in config files to enable/disable trailing
	#comment, and to give hints about how to handle an LHS or
	RHS. This will not preserve trailing comments in lines that
	are modified with "postconf -e" and the like.

	Maintainability: replace lengthy libmilter-API argument lists
	with named parameters, as with the libtls API.

	Fix buflen integer overflow detection in dict*sql.c.

	Fix "make test" bitrot.

	Move DNS-based tests from porcupine.org to postfix.org, or use
	a mock DNS library (a library that presents the same API as the
	real library, but that produces canned responses).

	Document dns_ncache_ttl_fix_enable use case in POSTSCREEN_README
	and RELEASE_NOTES.

	Remove this file from the stable release.

	Things to do after the stable release:

	Specify WARN_UNUSED_RESULT for all library functions that
	pass, deliver, bounce or defer a delivery request.

	Invent some kind of type-checking wrappers for htable(3),
	ctable(3) and other modules that take and return a void*
	pointer. We already did that for variadic functions.

	TLS certificate provenance: indicate whether a subject
	name/issuer are verified or not (for example, change the
	attribute name to unverified_ccert_subject etc.).  This is
	relevant only for fingerprint-based authentication including
	DANE, and affects logging, SMTPD policy, and Milters.

	Generalize the daemon '-S' stand-alone mode, so that it can
	be used with custom configuration settings for request/reply
	regression testing. This would use the existing "-o name=value"
	support to override parameters. For example, queue_directory
	would point to a directory with sockets for fake versions of
	Postfix-internal services.

	Update the list of Sendmail macros that Postfix can send
	to Milters (auth_ssf and TLS-related).

	Update smtpd command count when rejecting or skipping input
	before command-table lookup. But then we need to count
	commands that are rejected (malformed UTF-8, tokenizer
	error, forbidden command), or skipped (noop).

	What is the best place to detect spaces in pathnames during
	installation/upgrade/packaging? postfix-install for early
	warning, and post-install as a safety net?

	When the service basename differs from the program file
	basename, either prepend the service name to the syslogname (as
	if syslog_name=postfix/service/program), or prepend the service
	name to the process name (perhaps too confusing).  The service
	indication is desirable for mail delivery transports (smtp
	versus relay) as it identifies what scheduler parameters are
	in effect, but it is also desirable for mail receiving services
	(smtp versus submission verus smtps as configured in the stock
	master.cf file). This requires exceptions for some program names
	(exclude smtpd to avoid logging postfix/smtp/smtpd which could
	result in more confusion, and maybe other program names).

	UTF8 DNS[BW]L domain name.

	Consolidate maps flags in mail_params.h instead of having
	multiple copies scattered across programs.

	Try to allow UTF-8 myhostname/mydomain, at least in bounce
	template expansion.

	In the SMTP server, do not issue an enhanced status code when
	rejecting a connection before the HELO handshake is completed.

	Maybe don't whitelist a client that has maxed out its
	per-MTA connection count limit.

	Log command=good/bad statistics in postscreen?

	smtpd_checks tests either must use a DNS dummy resolver
	(override the res_search API) or all names must be under
	test.postfix.org (but that does not work for address->name
	lookups, and cannot simulate some errors).

	Reporting the original Message-ID in a bounce message
	In-Reply-To: or References: header.  In the cleanup daemon,
	grab a copy of the Message-ID and export it along with other
	header-extracted information at the top of the "extracted"
	queue file segment.  In the queue manager, extract this
	along with other header-extracted information, and forward
	the Message-ID in the bounce server notification request.

	Clobber ORCPT when sender is owner-mumble?

	Add milter_mumble_macros to the list of per-macro features.

	The pickup daemon logs warnings only when the cleanup daemon
	dit not provide a "reason" attribute. Is this logic right?

	up-convert myhostname to UTF-8 in MIME boundary strings?

	Eliminate code duplication between pcf_print_master_field()
	and pcf_print_master_entry().

	Error reporting: see if pcf_check_master_entry() and children
	can return error descriptions instead of terminating with
	a fatal error.

	Add a switch to consider postscreen deep protocol tests as
	"completed" when receiving "RSET" after "RCPT TO" and the
	session has passed all tests up to that point. RSET becomes
	like QUIT except perhaps that it does not hang up.

	apipe: map, splits results into address lists and performs
	lookups for the invidual addresses, converting back and
	forth between external and internal forms.

	Clarify that receive_override_options have no effect with
	smtpd_proxy_filter.

	Document the relative order of header_checks, address
	rewriting, milters.

	NOT: Table-driven case folding and case-insensitive string
	comparison specifically for UTF-8. Use libicu functions
	instead.

	When downgrading message/global to 7bit, is quoted-printable
	the appropriate encoding? Should it be base64?

	Should we encode headers with RFC 2047, when that is the
	only reason that Postfix cannot deliver to a non-UTF8SMTP
	server? Probably not in the general case. What about
	Postfix as a gateway server that converts UTF8SMTP
	for delivery to non-UTF8SMTP environments?

	Document and test restriction_classes example for
	smtpd_policy_service_default_action.

	Don't accept AUTH or other features that are not announced
	in the EHLO response.

	Suggested at Mailserver conference: Postscreen RDNS-based
	reputation (but this makes postscreen performance highly
	unpredicable because it introduces a dependency on random
	DNS servers).

	Suggested at Mailserver conference: a way to select a
	specific field in a table, presumably as the result value.
	This may be done with a filtermap{i,j,...}: table that propagates
	only the specified field(s).

	Discourage the use of "after 220" tests in POSTSCREEN_README
	and the documentation of individual parameter settings.

	To un-break "make tests" under src/smtpd, make tests
	independent from the DNS and native routines for host
	name/address lookup.

	Make been_here flag BH_FLAG_FOLD configurable for masochists.

	Replace some redundant TLS_README sections with pointers
	to FORWARD_SECRECY_README.

	Move html/index.html source to proto/.

	How hard is it to follow canonical or virtual mapping
	for the purpose of address validation? We must never
	reject a valid address.

	Preserve case in smtpd_resolve_addr() and add a structure
	member for the case-folded address. IIRC some Milter macro
	needs to show the unfolded address.

	Per SASL account rate limits. This requires new infrastructure
	that maintains stats by SASL account instead of client IP
	address.

	Watchdog timer in postmap/postalias.

	Begin code revision, after DANE support stabilizes.  This
	should be one pass that changes only names and no code.

	recipient_delimiters = $recipient_delimiter for BC

	All source code must specify its original author and
	license statement. Some code modules specify Lutz Jaenicke
	as the original author and fall under his liberal license.
	Code that is added to such a module has the same license
	(or at least something that is not more restrictive). Code
	modules without input from Lutz Jaenicke must state its
	original author and license (preferably no more restrictive
	than Postfix's own license). Currently, too many files list
	Wietse as the original author, and Lutz Jaenicke's license,
	which is wrong.

	We have smtp_host_lookup, smtp_dns_resolver_options, and
	now smtp_dns_support_level.  Of these, smtp_dns_resolver_options
	is orthogonal but the rest has overlap.

	There needs to be support for automatic migration from the
	deprecated disable_dns_lookups feature to the preferred
	smtp_dns_support_level feature. This support needs to exist
	for several releases before the deprecated feature can be
	removed.

	End code revision, after DANE support stabilizes.

	It would be nice if "bare username" lookup is not hard-coded
	for domains in the local address class.

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

	Should postconf -o refuse to work without the -x option?

	Make 30s caching (feature 20070414) configurable, such that
	0 means no caching.

	Make errno white/blacklist for getpwnam_r etc. and mailbox
	write errors.

	smtpd_muble_restrictions rule names are case-insensitive.
	restriction_classes values are case-sensitive but should
	be case-insensitive for consistency with smtpd_muble_restrictions.

	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.

	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.

	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?

	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.

	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.

	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?

	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 behavior) or disallow (strict
	behavior, currently implemented only in the SMTP server).

	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.