From 701e10b12105a4f4c632915698d3ede6ad591802 Mon Sep 17 00:00:00 2001
From: Colin <colin@uninsane.org>
Date: Thu, 28 Mar 2024 18:47:22 +0000
Subject: [PATCH] hosts/common/fs: optimize NFS options

---
 hosts/common/fs.nix | 24 +++++++++++++++++++-----
 1 file changed, 19 insertions(+), 5 deletions(-)

diff --git a/hosts/common/fs.nix b/hosts/common/fs.nix
index 159fa9d1..2a7ecb9c 100644
--- a/hosts/common/fs.nix
+++ b/hosts/common/fs.nix
@@ -23,7 +23,8 @@ let
     # N.B.: `remote-fs.target` is a dependency of multi-user.target, itself of graphical.target.
     # hence, omitting `noauto` can slow down boots.
     noauto = [ "noauto" ];
-    # lazyMount: defer mounting until first access from userspace
+    # lazyMount: defer mounting until first access from userspace.
+    # see: `man systemd.automount`, `man automount`, `man autofs`
     lazyMount = noauto ++ automount;
     wg = [
       "x-systemd.requires=wireguard-wg-home.service"
@@ -67,18 +68,31 @@ let
     # actimeo=n  = how long (in seconds) to cache file/dir attributes (default: 3-60s)
     # bg         = retry failed mounts in the background
     # retry=n    = for how many minutes `mount` will retry NFS mount operation
+    # intr       = allow Ctrl+C to abort I/O (it will error with `EINTR`)
     # soft       = on "major timeout", report I/O error to userspace
+    # softreval  = on "major timeout", service the request using known-stale cache results instead of erroring -- if such cache data exists
     # retrans=n  = how many times to retry a NFS request before giving userspace a "server not responding" error (default: 3)
     # timeo=n    = number of *deciseconds* to wait for a response before retrying it (default: 600)
     #              note: client uses a linear backup, so the second request will have double this timeout, then triple, etc.
+    # proto=udp  = encapsulate protocol ops inside UDP packets instead of a TCP session.
+    #              requires `nfsvers=3` and a kernel compiled with `NFS_DISABLE_UDP_SUPPORT=n`.
+    #              UDP might be preferable to TCP because the latter is liable to hang for ~100s (kernel TCP timeout) after a link drop.
+    #              however, even UDP has issues with `umount` hanging.
+    #
+    # N.B.: don't change these without first testing the behavior of sandboxed apps on a flaky network.
     nfs = common ++ [
-      # "actimeo=10"
-      "bg"
-      "retrans=4"
+      # "actimeo=5"
+      # "bg"
+      "retrans=1"
       "retry=0"
+      # "intr"
       "soft"
-      "timeo=15"
+      "softreval"
+      "timeo=30"
       "nofail"  # don't fail remote-fs.target when this mount fails  (not an option for sshfs else would be common)
+      # "proto=udp"  # default kernel config doesn't support NFS over UDP: <https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1964093> (see comment 11).
+      # "nfsvers=3"  # NFSv4+ doesn't support UDP at *all*. it's ok to omit nfsvers -- server + client will negotiate v3 based on udp requirement. but omitting causes confusing mount errors when the server is *offline*, because the client defaults to v4 and thinks the udp option is a config error.
+      # "x-systemd.idle-timeout=10"  # auto-unmount after this much inactivity
     ];
   };
   remoteHome = host: {