From e24182764def764b84146c62a69059ef60f3af86 Mon Sep 17 00:00:00 2001
From: Jean-Paul Calderone <exarkun@twistedmatrix.com>
Date: Mon, 26 Sep 2022 12:22:39 -0400
Subject: [PATCH] Declare that the Tahoe unit *Requires* its corresponding
Socket unit
Otherwise Tahoe might start without the socket unit and not have access to the
socket that's bound to the port and so just fail to start.
---
nixos/modules/private-storage.nix | 18 ++++++++----------
nixos/modules/tahoe.nix | 21 ++++++++++++++++++---
2 files changed, 26 insertions(+), 13 deletions(-)
diff --git a/nixos/modules/private-storage.nix b/nixos/modules/private-storage.nix
index 7b7892cd..3a716cf0 100644
--- a/nixos/modules/private-storage.nix
+++ b/nixos/modules/private-storage.nix
@@ -111,21 +111,19 @@ in
node =
# XXX Should try to name that is unique across the grid.
{ nickname = "${storage-node-name}";
- # We have the web port active because the CLI uses it. We may
- # eventually turn this off, or at least have it off by default (with
- # an option to turn it on). I don't know how much we'll use the CLI
- # on the nodes. Maybe very little? Or maybe it will be part of a
- # health check for the node... In any case, we tell it to bind to
- # localhost so no one *else* can use it. And the principle of the
- # web interface is that merely having access to it doesn't grant
- # access to any data. It does grant access to storage capabilities
- # but with our plugin configuration you still need ZKAPs to use
- # those...
+
+ # We have the web port active because the CLI uses it and because it
+ # exposes a metrics endpoint for our monitoring system. The actual
+ # port configuration lives in systemd so that it can order binding
+ # the socket correctly with other dependencies (which we can't
+ # reliably do with Tahoe without a bunch of other work).
"web.port" = "systemd:domain=INET:index=0";
+
# We have to tell Tahoe-LAFS where to listen for Foolscap
# connections for the storage protocol. We have to tell it twice.
# First, in the syntax which it uses to listen.
"tub.port" = "tcp:${toString cfg.publicStoragePort}";
+
# Second, in the syntax it advertises to in the fURL.
"tub.location" = "tcp:${cfg.publicAddress}:${toString cfg.publicStoragePort}";
};
diff --git a/nixos/modules/tahoe.nix b/nixos/modules/tahoe.nix
index 3c0691f7..51c86954 100644
--- a/nixos/modules/tahoe.nix
+++ b/nixos/modules/tahoe.nix
@@ -188,7 +188,6 @@ in
systemd.sockets."tahoe.storage" = {
description = "Tahoe Web Server Socket";
wantedBy = [ "sockets.target" ];
- before = [ "multi-user.target" ];
socketConfig = {
ListenStream = "127.0.0.1:3456";
};
@@ -203,10 +202,18 @@ in
eliotLog = "file:${nodedir}/logs/eliot.json,rotate_length=${toString (1024 * 1024 * 32)},max_rotated_files=32";
in nameValuePair "tahoe.${node}" {
description = "Tahoe LAFS node ${node}";
+
+ # We are partially socket activated but only for the web API port.
+ # For the actual storage service port, we bind ourselves. So make
+ # sure we actually do start up early in case storage requests come
wantedBy = [ "multi-user.target" ];
+
path = [ settings.package ];
- restartTriggers = [
- config.environment.etc."tahoe-lafs/${node}.cfg".source ];
+
+ # We don't know how to re-read our configuration file at runtime
+ # so restart if it ever changes.
+ restartTriggers = [ config.environment.etc."tahoe-lafs/${node}.cfg".source ];
+
serviceConfig = {
Type = "simple";
PIDFile = pidfile;
@@ -254,6 +261,14 @@ in
# now. So it makes sense to have the limit be 2^15 right now.
LimitNOFILE = 32768;
};
+
+ unitConfig = {
+ # Our config doesn't know how to bind all of its sockets on its
+ # own so don't start without the systemd units that *do* know
+ # how to bind them.
+ Requires = [ "tahoe.${node}.socket" ];
+ };
+
preStart =
let
created = "${nodedir}.created";
--
GitLab