* Add a migration to delete any existing DKIM key so that existing machines get a fresh 2048-bit key. (Sadly we don't support key rotation so the change is immediate.)
* Because the DNS record for a 2048-bit key is so much longer, the way we read OpenDKIM's DNS record text file had to be modified to combine an arbitrary number of TXT record quoted ("...") strings.
* When writing out the TXT record value, the string must be split into quoted ("...") strings with a maximum length of 255 bytes each, per the DNS spec.
* Added a changelog entry.
* Split the nginx templates again so we have just the part needed to make a domain do a redirect separate from the rest.
* Add server blocks to the nginx config for these domains.
* List these domains in the SSL certificate install admin panel.
* Generate default 'www' records just for domains we provide default redirects for.
Fixes#321.
I changed my mind. In 1bf8f1991f I allowed Unicode domain names to go into the database. I thought that was nice because it's what the user *means*. But it's not how the web works. Web and DNS were working, but mail wasn't. Postfix (as shipped with Ubuntu 14.04 without support for SMTPUTF8) exists in an ASCII-only world. When it goes to the users/aliases table, it queries in ASCII (IDNA) only and had no hope of delivering mail if the domain was in full Unicode in the database. I was thinking ahead to SMTPUTF8, where we *could* put Unicode in the database (though that would prevent IDNA-encoded addressing from being deliverable) not realizing it isn't well supported yet anyway.
It's IDNA that goes on the wire in most places anyway (SMTP without SMTPUTF8 (and therefore how Postfix queries our users/aliases tables), DNS zone files, nginx config, CSR 'CN' field, X509 Common Name and Subject Alternative Names fields), so we should really be talking in terms of IDNA (i.e. ASCII).
This partially reverts commit 1bf8f1991f, where I added a lot of Unicode=>IDNA conversions when writing configuration files. Instead I'm doing Unicode=>IDNA before email addresses get into the users/aliases table. Now we assume the database uses IDNA-encoded ASCII domain names. When adding/removing aliases, addresses are converted to ASCII (w/ IDNA). User accounts must be ASCII-only anyway because of Dovecot's auth limitations, so we don't do any IDNA conversion (don't want to change the user's login info behind their back!). The aliases control panel page converts domains back to Unicode for display to be nice. The status checks converts the domains to Unicode just for the output headings.
A migration is added to convert existing aliases with Unicode domains into IDNA. Any custom DNS or web settings with Unicode may need to be changed.
Future support for SMTPUTF8 will probably need to add columns in the users/aliases table so that it lists both IDNA and Unicode forms.
* For non-ASCII domain names, we will keep the Unicode encoding in our users/aliases table. This is nice for the user and also simplifies things like sorting domain names (using Unicode lexicographic order is good, using ASCII lexicogrpahic order on IDNA is confusing).
* Write nsd config, nsd zone files, nginx config, and SSL CSRs with domains in IDNA-encoded ASCII.
* When checking SSL certificates, treat the CN and SANs as IDNA.
* Since Chrome has an interesting feature of converting Unicode to IDNA in <input type="email"> form fields, we'll also forcibly convert IDNA to Unicode in the domain part of email addresses before saving email addresses in the users/aliases tables so that the table is normalized to Unicode.
* Don't allow non-ASCII characters in user account email addresses. Dovecot gets confused when querying the Sqlite database (which we observed even for non-word ASCII characters too, so it may not be related to the character encoding).
Relative hostnames have a fall-back lookup with the machine's hostname appended, which makes no sense. Add a period, e.g. "my.hostname.com" => "my.hostname.com.", to prevent that.
This caused false positive Spamhaus checks. Fixes#185.
The problem was that custom records defined for a subdomain where implicit
records are otherwise defined (e.g. A/AAAA records for the root) were ignored.
Though additional_records for a subdomain are processed in the base call to
build_zone (the call for the parent domain), and so custom records that don't
override implicits were working fine, those that overrode implicits were
ignored.
This was because the recursive call to build_zone for the subdomain creates the
implicit records (including A/AAAA records for the root), and so by relying on
the base call to add the additional_records fails because has_rec returned
true.
Adding a subdomain's additional_records in the child call works because has_rec
returns false when testing whether to add an e.g. A/AAAA override for the root,
as the defaults have not yet been added.
Brian Bustin
Joshua Tauberer
David Piggott
Ben Schumacher
sfPlayer1