NFS: Difference between revisions

From NixOS Wiki
imported>Allopsychic21
adds systemd.mounts and systemd.automounts method of configuration
Nh2 (talk | contribs)
add TODO regarding local_lock=all on Nix store
 
(7 intermediate revisions by 6 users not shown)
Line 9: Line 9:
$ mkdir /export
$ mkdir /export
</syntaxhighlight>
</syntaxhighlight>
You may need to change ownership of the <code>/export</code> directory to <code>nobody:nogroup</code>


Then we have to either move our already-existing directories inside <code>/export</code> (using <code>mv</code> from the command line) or bind-mount them there:
Then we have to either move our already-existing directories inside <code>/export</code> (using <code>mv</code> from the command line) or bind-mount them there:
Line 88: Line 90:
}
}
</syntaxhighlight>
</syntaxhighlight>
Note that clients see exposed shares as if they were exposed at the root level - i.e. <code>/export/foo</code> becomes <code>/foo</code> (in the <code>device</code> option). Other, regular '''fileSystems''' options apply.
Replace "server" in the above device attribute with the IP address or DNS entry of the NFS server. Note that clients see exposed shares as if they were exposed at the root level - i.e. <code>/export/foo</code> becomes <code>/foo</code> (in the <code>device</code> option). Other, regular '''fileSystems''' options apply.


=== Specifying NFS version ===
=== Specifying NFS version ===
Line 130: Line 132:
=== Using systemd.mounts and systemd.automounts ===
=== Using systemd.mounts and systemd.automounts ===


Using <code>systemd.mounts</code> and <code>systemd.automounts</code> was the only way to solve NFS client hangs experienced with the above configurations. Here is an example with auto-disconnecting and lazy-mounting implemented, and the <code>noatime</code> mount option added.
Here is an example with auto-disconnecting and lazy-mounting implemented, and the <code>noatime</code> mount option added.


Note that <code>wantedBy = [ "multi-user.target" ];</code> is required for the automount unit to start at boot.  
Note that <code>wantedBy = [ "multi-user.target" ];</code> is required for the automount unit to start at boot.  
Line 138: Line 140:
<syntaxhighlight lang="nix">
<syntaxhighlight lang="nix">
{
{
   systemd.mounts = {
  services.rpcbind.enable = true; # needed for NFS
   systemd.mounts = [{
     type = "nfs";
     type = "nfs";
     mountConfig = {
     mountConfig = {
Line 145: Line 148:
     what = "server:/tomoyo";
     what = "server:/tomoyo";
     where = "/mnt/tomoyo";
     where = "/mnt/tomoyo";
   };
   }];


   systemd.automounts = {
   systemd.automounts = [{
     wantedBy = [ "multi-user.target" ];
     wantedBy = [ "multi-user.target" ];
     automountConfig = {
     automountConfig = {
Line 153: Line 156:
     };
     };
     where = "/mnt/tomoyo";
     where = "/mnt/tomoyo";
   };
   }];
}
}
</syntaxhighlight>
</syntaxhighlight>
Line 161: Line 164:
<syntaxhighlight lang="nix">
<syntaxhighlight lang="nix">
{
{
  services.rpcbind.enable = true; # needed for NFS
   systemd.mounts = let commonMountOptions = {
   systemd.mounts = let commonMountOptions = {
     type = "nfs";
     type = "nfs";
Line 202: Line 206:
In a single-user setup ('''not on Nixos''') the Nix store can be also exported over NFS (common in HPC clusters) to share package over the networks. The only requirement is to also pass <code>local_lock=flock</code> or <code>local_lock=all</code> as mount option to allow the nix packages to take locks on modifications. Example entry in <code>fstab</code>:
In a single-user setup ('''not on Nixos''') the Nix store can be also exported over NFS (common in HPC clusters) to share package over the networks. The only requirement is to also pass <code>local_lock=flock</code> or <code>local_lock=all</code> as mount option to allow the nix packages to take locks on modifications. Example entry in <code>fstab</code>:


<syntaxhighlight lang="console"><host_or_ip>/nix /nix nfs nofail,x-systemd.device-timeout=4,local_lock=all 0 0</syntaxhighlight>
<syntaxhighlight lang="console"><host_or_ip>/nix /nix nfs nofail,x-systemd.device-timeout=4,local_lock=all 0 0</syntaxhighlight>'''TODO:''' Why this? That seems extremely unsafe. This disables NFS locks (which apply to all NFS clients), and makes locks ''local'', meaning a lock taken by one NFS client isn't seen by another, and both can take their locks. So this removes protection against concurrent writes, which Nix assumes.
[[Category:Filesystem]]

Latest revision as of 01:41, 22 July 2024

Server

Let's say that we've got one server-machine with 4 directories that we want to share: /mnt/kotomi, /mnt/mafuyu, /mnt/sen and /mnt/tomoyo.

First, we have to create a dedicated directory from which our NFS server will access the data:

$ mkdir /export

You may need to change ownership of the /export directory to nobody:nogroup

Then we have to either move our already-existing directories inside /export (using mv from the command line) or bind-mount them there:

{
  fileSystems."/export/mafuyu" = {
    device = "/mnt/mafuyu";
    options = [ "bind" ];
  };

  fileSystems."/export/sen" = {
    device = "/mnt/sen";
    options = [ "bind" ];
  };

  fileSystems."/export/tomoyo" = {
    device = "/mnt/tomoyo";
    options = [ "bind" ];
  };

  fileSystems."/export/kotomi" = {
    device = "/mnt/kotomi";
    options = [ "bind" ];
  };
}

Having the filesystem ready, we can proceed to configure the NFS server itself:

{
  services.nfs.server.enable = true;
  services.nfs.server.exports = ''
    /export         192.168.1.10(rw,fsid=0,no_subtree_check) 192.168.1.15(rw,fsid=0,no_subtree_check)
    /export/kotomi  192.168.1.10(rw,nohide,insecure,no_subtree_check) 192.168.1.15(rw,nohide,insecure,no_subtree_check)
    /export/mafuyu  192.168.1.10(rw,nohide,insecure,no_subtree_check) 192.168.1.15(rw,nohide,insecure,no_subtree_check)
    /export/sen     192.168.1.10(rw,nohide,insecure,no_subtree_check) 192.168.1.15(rw,nohide,insecure,no_subtree_check)
    /export/tomoyo  192.168.1.10(rw,nohide,insecure,no_subtree_check) 192.168.1.15(rw,nohide,insecure,no_subtree_check)
  '';
}

This configuration exposes all our shares to 2 local IPs; you can find more examples at Gentoo's wiki [1].

Other options are available on the NixOS option page or via the nixos-option command.

Firewall

If your server-machine has a firewall turned on (as NixOS does by default, for instance), don't forget to open appropriate ports; e.g. for NFSv4:

networking.firewall.allowedTCPPorts = [ 2049 ];

Many clients only support NFSv3, which requires the server to have fixed ports:

  services.nfs.server = {
    enable = true;
    # fixed rpc.statd port; for firewall
    lockdPort = 4001;
    mountdPort = 4002;
    statdPort = 4000;
    extraNfsdConfig = '''';
  };
  networking.firewall = {
    enable = true;
      # for NFSv3; view with `rpcinfo -p`
    allowedTCPPorts = [ 111  2049 4000 4001 4002 20048 ];
    allowedUDPPorts = [ 111 2049 4000 4001  4002 20048 ];
  };

Client

Continuing the server example, mounting the now-exposed tomoyo share on another box (on a client) is as simple as:

{
  fileSystems."/mnt/tomoyo" = {
    device = "server:/tomoyo";
    fsType = "nfs";
  };
}

Replace "server" in the above device attribute with the IP address or DNS entry of the NFS server. Note that clients see exposed shares as if they were exposed at the root level - i.e. /export/foo becomes /foo (in the device option). Other, regular fileSystems options apply.

Specifying NFS version

You can specify NFS version by adding the "nfsvers=" option:

{
  fileSystems."/mnt/tomoyo" = {
    # ...
    options = [ "nfsvers=4.2" ];
  };
}

Lazy-mounting

By default, all shares will be mounted right when your machine starts - apart from being simply unwanted sometimes, this may also cause issues when your computer doesn't have a stable network connection or uses WiFi; you can fix this by telling systemd to mount your shares the first time they are accessed (instead of keeping them mounted at all times):

{
  fileSystems."/mnt/tomoyo" = {
    # ...
    options = [ "x-systemd.automount" "noauto" ];
  };
}

Auto-disconnecting

You can tell systemd to disconnect your NFS-client from the NFS-server when the directory has not been accessed for some time:

{
  fileSystems."/mnt/tomoyo" = {
    # ...
    options = [ "x-systemd.idle-timeout=600" ]; # disconnects after 10 minutes (i.e. 600 seconds)
  };
}

Using systemd.mounts and systemd.automounts

Here is an example with auto-disconnecting and lazy-mounting implemented, and the noatime mount option added.

Note that wantedBy = [ "multi-user.target" ]; is required for the automount unit to start at boot.

Also note that x-systemd mount options are unneeded, as they are a representation of systemd options in fstab(5) format. They get parsed and converted to unit files by systemd-fstab-generator(8) as mentioned in systemd.mount(5).

{
  services.rpcbind.enable = true; # needed for NFS
  systemd.mounts = [{
    type = "nfs";
    mountConfig = {
      Options = "noatime";
    };
    what = "server:/tomoyo";
    where = "/mnt/tomoyo";
  }];

  systemd.automounts = [{
    wantedBy = [ "multi-user.target" ];
    automountConfig = {
      TimeoutIdleSec = "600";
    };
    where = "/mnt/tomoyo";
  }];
}

Multiple mounts with the exact same options can benefit from abstraction.

{
  services.rpcbind.enable = true; # needed for NFS
  systemd.mounts = let commonMountOptions = {
    type = "nfs";
    mountConfig = {
      Options = "noatime";
    };
  };

  in

  [
    (commonMountOptions // {
      what = "server:/tomoyo";
      where = "/mnt/tomoyo";
    })

    (commonMountOptions // {
      what = "server:/oyomot";
      where = "/mnt/oyomot";
    })
  ];

  systemd.automounts = let commonAutoMountOptions = {
    wantedBy = [ "multi-user.target" ];
    automountConfig = {
      TimeoutIdleSec = "600";
    };
  };

  in

  [
    (commonAutoMountOptions // { where = "/mnt/tomoyo"; })
    (commonAutoMountOptions // { where = "/mnt/oyomot"; })
  ];
}

Nix store on NFS

In a single-user setup (not on Nixos) the Nix store can be also exported over NFS (common in HPC clusters) to share package over the networks. The only requirement is to also pass local_lock=flock or local_lock=all as mount option to allow the nix packages to take locks on modifications. Example entry in fstab:

<host_or_ip>/nix /nix nfs nofail,x-systemd.device-timeout=4,local_lock=all 0 0

TODO: Why this? That seems extremely unsafe. This disables NFS locks (which apply to all NFS clients), and makes locks local, meaning a lock taken by one NFS client isn't seen by another, and both can take their locks. So this removes protection against concurrent writes, which Nix assumes.