--- Begin Message ---
Subject: |
Invalid 'location' field generated in dovecot configuration |
Date: |
Sun, 20 Nov 2022 21:53:01 +0000 |
User-agent: |
mu4e 1.8.11; emacs 28.2 |
Hi Guix!
After updating the system, the dovecot service got confused and started
moving around all mailboxes. I looked up the configuration and noticed
strange invalid syntax for the location field:
--8<---------------cut here---------------start------------->8---
location=#<<location> file: "path/to/config.scm" line: 297 column: 20>
--8<---------------cut here---------------end--------------->8---
Because the # character is interpreted as a comment, dovecot doesn't
crash and instead moves mailboxes around in weird ways I don't quite
understand :-/.
This can actually be reproduced locally with the dovecot system test if
one dumps the following expression to check the configuration:
--8<---------------cut here---------------start------------->8---
(format #t "~a\n" (wait-for-file "/etc/dovecot/dovecot.conf"
marionette
#:read 'get-string-all))
--8<---------------cut here---------------end--------------->8---
Giving us the snippets like this in the config:
--8<---------------cut here---------------start------------->8---
$ make check-system TESTS="dovecot" VERBOSE=1
...
namespace inbox {
type=private
separator=
prefix=
location=#<<location> file: "gnu/tests/mail.scm" line: 297 column: 20>
inbox=yes
hidden=no
list=yes
subscriptions=yes
mailbox "Drafts" {
auto=no
special_use=\Drafts
}
mailbox "Junk" {
auto=no
special_use=\Junk
}
mailbox "Trash" {
auto=no
special_use=\Trash
}
mailbox "Sent" {
auto=no
special_use=\Sent
}
mailbox "Sent Messages" {
auto=no
special_use=\Sent
}
mailbox "Drafts" {
auto=no
special_use=\Drafts
}
}
...
--8<---------------cut here---------------end--------------->8---
I did a `git bisect` with `guix time-machine` (this tool is invaluable)
and found the issue started with this commit:
--8<---------------cut here---------------start------------->8---
commit 543d971ed2a1d9eb934af1f51930741d7cc4e7ef
Author: Maxim Cournoyer <maxim.cournoyer@gmail.com>
Date: Fri Oct 28 17:06:16 2022 -0400
services: configuration: Re-order generated record fields.
This is so that the first field of the generated record matches the first
one
declared, which makes 'define-configuration' record API compatible with
define-record-type* ones.
* gnu/services/configuration.scm (define-configuration-helper): Move the
%location field below the ones declared by the user.
* gnu/services/monitoring.scm (zabbix-front-end-config): Adjust match
pattern
accordingly.
--8<---------------cut here---------------end--------------->8---
Sooo, I'm guessing this is something to do with the configuration field
being named "location", and /maybe/ we're patching it with the origin
location of the configuration, something like that? I don't understand
how this works well enough to be able to thing of any fixes.
Thanks,
Pierre
signature.asc
Description: PGP signature
--- End Message ---
--- Begin Message ---
Subject: |
Re: bug#59423: Invalid 'location' field generated in dovecot configuration |
Date: |
Fri, 02 Dec 2022 22:05:12 -0500 |
User-agent: |
Gnus/5.13 (Gnus v5.13) Emacs/28.2 (gnu/linux) |
Hi Ludovic,
Ludovic Courtès <ludo@gnu.org> writes:
> Ludovic Courtès <ludo@gnu.org> skribis:
>
>> Maxim Cournoyer <maxim.cournoyer@gmail.com> skribis:
>
> [...]
>
>>> No :-). I meant why do we even set a default accessor for the *source
>>> location* information (in the (gnu service configuration) macros); it's
>>> that one that doesn't seem to get used (or I'm blind to it!), at least
>>> via this accessor. If it's not strictly necessary, we can stop
>>> producing it, and that would solve the problem.
>>
>> Like I wrote, I think it’s necessary, even if not used now.
>
> To complement this answer: key high-level record types usually have a
> ‘location’ field: <package>, <channel>, <mapped-device>, <file-system>,
> <service-type>, etc. The rationale is that it allows us to report
> accurate location info for errors and warnings, to jump to their
> definition, and so on.
>
> For configuration records this is not a common pattern, but the
> rationale holds. ‘zabbix-front-end-config’ uses the ‘%location’ field,
> but it seems to be the only one so far.
Thanks for this extra bit of information and for spotting this usage. I
think "location" is likely to conflict for the general purpose
'define-configuration' generated records, so I've renamed the "location"
*accessor* to "source-location".
In the near future I want to migrate more service configurations to the
'define-configuration' machinery, to benefit from its useful
self-validating property. For now I wouldn't feel at ease doing so
unless raw record matching (not yet using 'match-record') works the same
way, since we still have many occurrences making use of that (often via
match-lambda). For that reason, I prefer to not revert the record
layout until we've gotten rid of all the match-lambda matching record
fields directly (which will take some time).
I've applied the rename fix to master.
--
Thanks,
Maxim
--- End Message ---