On 26/10/2017 18:38, Alexander Dalloz wrote:

Am 26.10.2017 um 12:20 schrieb David Zambonini:

...

There seems to be a bug with RFC822 processing in ltmp proxying that doesn't quote local parts that, for example, contain spaces.

Newer related RFCs are RFC 5321 and 5322.

Typo, meant to say RFC2822, which they still supercede, not that the local-part spec has changed. :)

[ ... ]

...

MAIL FROM:test@testdomain.com\r\n RCPT TO:\r\n

501 5.5.4 Invalid.parameters\r\n

That recipient address is totally invalid. It is neither just a local part without a domain, nor a plussed address destination.

Check your setup with i.e.

RCPT TO:<"Junk E-mail"@deemzed.uk>

or

RCPT TO:<"test+Junk E-mail"@deemzed.uk>

Apologies, I was attempting to cut the config down at the time the dump was taken. Correcting (I can provide config privately, but not share to list), I still get:

MAIL FROM:test@testdomain.com\r\n RCPT TO:<"deemzed.uk+Junk E-mail"@mailbox.localhost>\r\n DATA\r\n (etc) .\r\n

501 5.5.4 Invalid parameters\r\n

QUIT\r\n

from director -> dovecot LMTP network dump:

MAIL FROM:test@testdomain.com\r\n RCPT TO:\r\n

501 5.5.4 Invalid.parameters\r\n

The problem is that dovecot is interpreting/unquoting the local part of the address to insert into the username, but the client code in client_proxy_rcpt()/address_add_detail() [lmtp/commands.c] then inserts the username and detail directly into lmtp_rcpt.address with no attempt whatsoever to requote that string regardless of what characters it contains, leading to the situation where a straight-through proxy fails as director is generating addresses that dovecot doesn't like. It can be corrected manually using:

override_fields = destuser="%{orig_username}"@%{orig_domain}

which kind of "fixes" the issue, which I had thought sufficient last year for the limited range of inputs I have, but it turns out to break director hashing as the username is then hashed containing quotes (not to mention fun with recipient_delimiter).

Looking through RFC2822 any non-atext character in username, detail or delimiter is enough for the local part to be converted to quoted-string. Now I've traced it through in the source, I could have a look at starting to get a fix together tomorrow with an aim to providing a pull request, if it turns out there are no side-effects to treating lmtp_rcpt.address like this and you'd like an example of what I mean.

-- David Zamboninini

Hi again,

I've not heard anything further regarding this bug, so I've had a look at the code.

To restate the bug in a more precise way: LMTP in dovecot treats external RFC822 email addresses in the envelope recipient and internal usernames as almost identical/interchangeable. This is incorrect and leads to issues when attempting to use director as an LMTP proxy to proxy to recipients with quoted-local parts, as it is issuing invalid email addresses at the LMTP protocol level (it strips quotes from the local part and then does not add them again when proxying). It's also causing issues with director username hashing.

I've created a "bugfix" patch to indicate what I mean, it appears to solve the issue, although I do not think it is anywhere near a production ready change.

1. The first problem is that dovecot is not performing a full decode on the envelope recipient address when creating the username, leading to escaped characters left in escaped form, and is not treating it consistently, choosing to either strip the surrounding quotes or not depending on whether or not it contains an @. I've fixed this by changing the code in lmtp_unescape_address() [src/lmtp/commands.c] to use rfc822_parse_quoted_string().

2. This leads to the second problem where the username becomes ambiguous if the local-part contains an @ (regardless of whether or not the first fix is applied or not). I've worked around this by using strrchr() instead of strchr() on the username string to split the domain out in mail_user_hash() [src/lib-mail/mail-user-hash.c] and message_detail_address_parse() [src/lib-mail/message-address.c], although likely I've missed some place this change should be made.

3. The third problem is then re-encoding the username in the envelope recipient when proxying, which was not done at all. I've added a function lmtp_client_rfc822_escape_address(), which is similar to str_append_maybe_escape() to escape the address at protocol time in lmtp_client_send_rcpts() [src/lib-smtp/lmtp-client.c], although I suspect it should be done earlier, this is just a working proof.

The other reason I don't believe this patch is production quality is that I have not examined any interaction between these changes and sieve's use of the envelope recipient. I'm hoping that a developer can chip in here? (hint!)

(Apologies for top posting)

On 30/10/2017 13:18, David Zambonini wrote:

On 26/10/2017 19:33, David Zambonini wrote:

...

On 26/10/2017 18:38, Alexander Dalloz wrote:

...

Am 26.10.2017 um 12:20 schrieb David Zambonini:

...

There seems to be a bug with RFC822 processing in ltmp proxying that doesn't quote local parts that, for example, contain spaces.

Newer related RFCs are RFC 5321 and 5322.

Typo, meant to say RFC2822, which they still supercede, not that the local-part spec has changed. :)

...

[ ... ]

...

MAIL FROM:test@testdomain.com\r\n RCPT TO:\r\n

501 5.5.4 Invalid.parameters\r\n

That recipient address is totally invalid. It is neither just a local part without a domain, nor a plussed address destination.

Check your setup with i.e.

RCPT TO:<"Junk E-mail"@deemzed.uk>

or

RCPT TO:<"test+Junk E-mail"@deemzed.uk>

Apologies, I was attempting to cut the config down at the time the dump was taken. Correcting (I can provide config privately, but not share to list), I still get:

MAIL FROM:test@testdomain.com\r\n RCPT TO:<"deemzed.uk+Junk E-mail"@mailbox.localhost>\r\n DATA\r\n (etc) .\r\n

501 5.5.4 Invalid parameters\r\n

QUIT\r\n

from director -> dovecot LMTP network dump:

I could have a look at starting to get a fix together tomorrow with an aim to providing a pull request, if it turns out there are no side-effects to treating lmtp_rcpt.address like this and you'd like an example of what I mean.

My apologies for not adding your address on my initial response, Alexander - not sure if you noticed what I replied with or not.

Nope, this isn't going to happen. I'm not familiar with the dovecot internals but lmtp uses just the address string in the form of "full address with quotes stripped from local part but otherwise not decoded" and nothing else throughout, which touches on quite a bit of code. It makes it indeterminate and not always possible to reassemble the original, it's a bit of a trainwreck.

The sanest option to me seems to me to be to store a decoded local part and domain in addition to the detail in mail_recipient, and keeping a now properly rfc822 encoded address in sync with it. However, this would cause a deviation from existing behaviour for the full original user (the quotes would be seen).

I'm between a rock and a hard place here - at the very least I'd like this bug to be officially recognised.

-- David Zambonini

On 03/11/2017 11:48, Stephan Bosch wrote:

Hi,

Sorry, we're in a bit of a v2.3 merge frenzy. Much of the LMTP code will be replaced in v2.3, but I'll give the  older code a look as well.

This can take a while though.

Thank you very much for getting back to me, I can appreciate it can get hectic, and I don't wish to appear ungrateful, I wholeheartedly endorse/recommend dovecot and the company I work for does use paid for OX elsewhere. For my own part, the platform I manage is > 300,000 mailboxes and dovecot performs incredibly well.

I came up with some much smaller patches that accomplish the same thing in v2.2 using built-in functions and pushing the re-encoding slightly further up the call stack - address/username being interchangeable over most of the lmtp code makes significant changes problematic, so I thought it best not to try a rework.

Looking at gitub, though, I don't see any significant changes in behaviour as far as the problem I'm seeing goes, which is worrying.

What I'll do is leave the patches here for reference, and pick this up again after the v2.3 release. If you do have time for a further response, I could also provide them as pull requests against current on github if you'd like to request that.

1. Cut on the final instead of initial @ when splitting user/domain parts in LMTP, this can fix some issues where localpart contains a quoted @:

dovecot-2.2.33.2-reverse-domaincut.patch

2. Fully decode local part on receipt in LMTP, and re-encode when proxying. This fixes the issue where quoted local quotes are stripped on proxy, preventing successful proxying, and some director hashing problems (exposes str_append_maybe_escape in message-address.h, some logging is still inconsistent, though, but would require a major rework):

dovecot-2.2.33.2-quoted-local-proxy.patch

-- David Zambonini