Hi Maxim,

On 4/25/25 3:27 PM, Maxim Cournoyer wrote:
Hi,

Giacomo Leidi <goodoldpaul@autistici.org> writes:

This commit adds a password-file to the postgresql-role field.  It
allows users to provision Postgres roles with a set password.

* gnu/services/databases.scm (postgresql-role): Add password-file field;
(postgresql-role-configuration): add requirement field;
(postgresql-create-roles): add support for setting passwords from a
file without leaking passwords to the command line;
(postgresql-role-shepherd-service): add support for customizable
requirements.
* gnu/tests/databases.scm: Test it.
* doc/guix.texi: Document it.
Looks useful!

Nitpick: Each change can be a sentence, starting with a capitalized
letter and ending with a period.
Should be fixed now.

Change-Id: I3aabaa10b0c5e826c5aa874e5649e25a3508a585
---
 doc/guix.texi              | 15 +++++++++---
 gnu/services/databases.scm | 48 +++++++++++++++++++++++++++++++++-----
 gnu/tests/databases.scm    | 38 +++++++++++++++++++++++++++---
 3 files changed, 89 insertions(+), 12 deletions(-)

diff --git a/doc/guix.texi b/doc/guix.texi
index bee80cd4e2b..8b32557f76a 100644
--- a/doc/guix.texi
+++ b/doc/guix.texi
@@ -27588,9 +27588,10 @@ Database Services
 
 @lisp
 (service-extension postgresql-role-service-type
-                   (const (postgresql-role
-                           (name "alice")
-                           (create-database? #t))))
+                   (const (list
+                           (postgresql-role
+                            (name "alice")
+                            (create-database? #t)))))
Are you sure this is correct?  If it's a bug fix, it's not listed in the
changelog.

Yes I am sure, the service type has


(compose concatenate)


this is now mentioned in the changelog.

             postgresql-role?
             postgresql-role-name
+            postgresql-role-password-file
             postgresql-role-permissions
             postgresql-role-create-database?
             postgresql-role-configuration
             postgresql-role-configuration?
             postgresql-role-configuration-host
+            postgresql-role-configuration-requirement
             postgresql-role-configuration-roles
 
             postgresql-role-service-type
@@ -374,6 +378,8 @@ (define-record-type* <postgresql-role>
   postgresql-role make-postgresql-role
   postgresql-role?
   (name             postgresql-role-name) ;string
+  (password-file    postgresql-role-password-file  ;string
+                    (default #f))
   (permissions      postgresql-role-permissions
                     (default '(createdb login))) ;list
   (create-database? postgresql-role-create-database?  ;boolean
@@ -392,6 +398,8 @@ (define-record-type* <postgresql-role-configuration>
   postgresql-role-configuration?
   (host             postgresql-role-configuration-host ;string
                     (default "/var/run/postgresql"))
+  (requirement      postgresql-role-configuration-requirement ;list-of-symbols
+                    (default '()))
as mentioned above, this should be 'shepherd-requirement'.
I know at least two examples that are already in master which use 'requirement' (I may be biased as I authored both of them): the restic-backup-service-type and the oci-container-service-type. If you don't find this to be a blocker I would defer this change to a specific commit setting a convention for the whole codebase after an analysis of which services use simply 'requirement' and which use 'shepherd-requirement'. Is it ok for you?

      
   (log              postgresql-role-configuration-log ;string
                     (default "/var/log/postgresql_roles.log"))
   (roles            postgresql-role-configuration-roles
@@ -409,19 +417,36 @@ (define (postgresql-create-roles config)
                                permissions)
                    " ")))
 
+  (define (password-value role)
+    (string-append "password_" (postgresql-role-name role)))
+
+  (define (role->password-variable role)
+    (define file-name
+      (postgresql-role-password-file role))
I'd maybe use a let here, just for stylistic preferences.
I changed the define to let

+    (if (string? file-name)
+        ;; This way passwords do not leak to the command line
Please use complete sentences, including the trailing period, for
stand-alone comments.
I added the period

+        #~(string-append "-v \"" #$(password-value role)
+                         "=$(" #+coreutils "/bin/cat " #$file-name ")\"")
Are you sure this should use ungexp-native (#+) on coreutils?  If this
runs in a snippet, it should use the regular ungexp (#$). It could also
use file-append.  Only if it runs as part of a derivation build should
it use #+.
this runs at system activation so in my understanding it should be #$ . Thank you for explaining the difference, I was not aware :)

+        ""))
+
   (define (roles->queries roles)
     (apply mixed-text-file "queries"
            (append-map
             (lambda (role)
               (match-record role <postgresql-role>
                 (name permissions create-database? encoding collation ctype
-                      template)
+                      template password-file)
                 `("SELECT NOT(EXISTS(SELECT 1 FROM pg_catalog.pg_roles WHERE \
 rolname = '" ,name "')) as not_exists;\n"
 "\\gset\n"
 "\\if :not_exists\n"
 "CREATE ROLE \"" ,name "\""
 " WITH " ,(format-permissions permissions)
+,(if (and (string? password-file)
+          (not (string-null? password-file)))
Since you already check for the string-null? case, perhaps the default
password value should be "" instead of #f?  And then you could do away
with the string? check (though ideally there could be a sanitizer to
verify that a string is passed -- perhaps best leaving it for the day
this gets ported to define-configuration, which auto-generates sanitizers
based on the type specified for the field in its form.
in its ideal form (using define-configuration) this would be a 'maybe-string' type, as it represents an optional path. if you don't find this a blocker I would leave this for a later change where we could use 'maybe-value-set?' and leave the type check to a sanitizer. What do you think?

      
+     (string-append
+      "\nPASSWORD :'" (password-value role) "'")
+     "")
 ";\n"
 ,@(if create-database?
       `("CREATE DATABASE \"" ,name "\""
@@ -436,20 +461,30 @@ (define (postgresql-create-roles config)
 
   (let ((host (postgresql-role-configuration-host config))
         (roles (postgresql-role-configuration-roles config)))
-    #~(let ((psql #$(file-append postgresql "/bin/psql")))
-        (list psql "-a" "-h" #$host "-f" #$(roles->queries roles)))))
+    (program-file "run-queries"
+      #~(let ((bash #$(file-append bash-minimal "/bin/bash"))
+              (psql #$(file-append postgresql "/bin/psql")))
+          (define command
+            (string-append
+             "set -e; exec " psql " -a -h " #$host " -f "
+             #$(roles->queries roles) " "
+             (string-join
+              (list
+               #$@(map role->password-variable roles))
+              " ")))
+          (execlp bash bash "-c" command)))))
 
 (define (postgresql-role-shepherd-service config)
   (match-record config <postgresql-role-configuration>
-    (log)
+    (log requirement)
     (list (shepherd-service
-           (requirement '(user-processes postgres))
+           (requirement `(user-processes postgres ,@requirement))
I see two styles being used in Guix services: 'shepherd-requirement'
extend the requirements, or overrides it, with the 'user-processes at
least normally found in the default values.  I think at this point in
time the later fits better in Guix (think of arguments such as #:modules
and #:imported-modules, etc.  the user has full control on them, for the
better and worst).
Thank you for pointing this out, I created a %postgresql-role-shepherd-requirement value and exported it to allow users which want to override the Guix defaults, while retaining them can simply append to this list.
            
            (provision '(postgres-roles))
            (one-shot? #t)
            (start
             #~(lambda args
                 (zero? (spawn-command
-                        #$(postgresql-create-roles config)
+                        (list #$(postgresql-create-roles config))
Why is this change now necessary?  I didn't follow.

the return value of postgresql-create-roles changed, before this change it returned a list containing a psql command line, after this change it returns a gexp which will evaluate to the entrypoint responsible for reading password files and then calling psql.


                         #:user "postgres"
                         #:group "postgres"
                         ;; XXX: As of Shepherd 1.0.2, #:log-file is not
@@ -468,6 +503,7 @@ (define postgresql-role-service-type
                           (match-record config <postgresql-role-configuration>
                             (host roles)
                             (postgresql-role-configuration
+                             (inherit config)
and this one?  Was it a preexisting bug?
Yes, for example without this change we would lose the 'log' field's value upon extension.

                              (host host)
                              (roles (append roles extended-roles))))))
                 (default-value (postgresql-role-configuration))
diff --git a/gnu/tests/databases.scm b/gnu/tests/databases.scm
index fd5041344b6..c5da603565d 100644
--- a/gnu/tests/databases.scm
+++ b/gnu/tests/databases.scm
@@ -142,6 +142,8 @@ (define %role-log-file
 
 (define %postgresql-os
   (simple-operating-system
+   (extra-special-file "/password"
+                       (plain-file "password" "hello"))
    (service postgresql-service-type
             (postgresql-configuration
              (postgresql postgresql)
@@ -158,6 +160,10 @@ (define %postgresql-os
              (roles
               (list (postgresql-role
                      (name "root")
+                     (create-database? #t))
+                    (postgresql-role
+                     (name "alice")
+                     (password-file "/password")
                      (create-database? #t))))))))
 
 (define (run-postgresql-test)
@@ -230,14 +236,40 @@ (define (run-postgresql-test)

      
             (marionette-eval
              '(begin
                 (use-modules (gnu services herd)
+                             (srfi srfi-1)
                              (ice-9 popen))
                 (current-output-port
                  (open-file "/dev/console" "w0"))
+                (every
+                 (lambda (role)
+                   (let* ((port (open-pipe*
+                                 OPEN_READ
+                                 #$(file-append postgresql "/bin/psql")
+                                 "-tA" "-c"
+                                 (string-append
+                                  "SELECT 1 FROM pg_database WHERE"
+                                  " datname='" role "'")))
+                          (output (get-string-all port)))
+                     (close-pipe port)
+                     (string-contains output "1")))
+                 '("root" "alice")))
+             marionette))
+
+          (test-assert "database passwords are set"
+            (marionette-eval
+             '(begin
+                (use-modules (gnu services herd)
+                             (ice-9 match)
+                             (ice-9 popen))
+                (current-output-port
+                 (open-file "/dev/console" "w0"))
+                (setgid (passwd:gid (getpwnam "alice")))
+                (setuid (passwd:uid (getpw "alice")))
+                (setenv "PGPASSWORD" "hello")
                 (let* ((port (open-pipe*
                               OPEN_READ
-                              #$(file-append postgresql "/bin/psql")
-                              "-tA" "-c" "SELECT 1 FROM pg_database WHERE
- datname='root'"))
+                              #$(file-append postgresql "/bin/psql") "-tA" "-c"
+                              "SELECT 1 FROM pg_database WHERE datname='alice'"))
Nice to have tests!  Could you please send a v7 taking care of my
comments?  Then I guess it'd be good to go.


I should have addressed all your comments, I'm about to send a v7. Thank you for your help!


cheers

giacomo