File descriptor leak in Nix master

[thufschmitt]

Describe the bug

Building a derivation with Nix master leaves a number of file descriptors that aren't closed until Nix exits

Steps To Reproduce

1. In the background, monitor the number of FDs used by a nix-build process (replace by nix-store if you're using the daemon)

   $ while true; do lsof -c nix-build | wc -l; done

2. Build a big-enough number of independant derivations:

   $ nix-build many-drvs.nix

   where many-drvs.nix is as follows:

   let pkgs = import <nixpkgs> {}; in
   
   builtins.map
     (i: pkgs.writeText "foo" (toString builtins.currentTime + toString i))
     (pkgs.lib.genList (i: i) 1000)

   watch the output of lsof grow (up to ~5000 in my tests).

Expected behavior

The number of opened file descriptors should be bounded (it stays under 100 for nix 2.3.7)

nix-env --version output

$ nix-env --version
nix-env (Nix) 3.0pre20200829_f156513

Additional context

Looking at the lsof output, I see a lot of .lock (~half of the output) and .drv.bz2 (~1/4th of the output) files. I guess we forget to release these at some point

[stale]

I marked this as stale due to inactivity. → More info

[cole-h]

I just tried this on nix-env (Nix) 2.4pre20210308_1c0e3e4 and it stayed at about 58 the entire time. It seems like this was resolved in some intermediate commit?

[edolstra]

Yeah, I can't reproduce this. Top-level substitution goals had a similar FD leak (8a29052).

[veprbl]

Still reproducible on 2.7.0

[thufschmitt]

@veprbl how did you manage to reproduce it? Just tried on 2.8.0pre20220317_a0b517d, but the FD count stayed nicely bounded

[veprbl]

Just evaluated texlive from master on macOS and observed "Too many open files" on Nix 2.7.0. Trying on 2.3.16 fixed the issue.

[thufschmitt]

Oh it’s happening at eval-time? Interesting…

Still can’t reproduce though. Maybe it’s a darwin-specific thing ? 🤔

[veprbl]

I meant build. I remember originally I also saw a similar behaviour on Linux, it was opening some extra files (not the .lock's) in the store.

[tobiasBora]

I can experience this again, on NixOs when installing texlive with the documentation NixOS/nixpkgs#171218

I can fix it by building first my system with

  security.pam.loginLimits = [{
    domain = "*";
    type = "soft";
    item = "nofile";
    value = "65535";
  }];

[aameen-tulip]

This was causing issues for me building ( an admittingly large ) Yarn project on Debian 11 ( barebones cloud-init image with nothing other than nix installed ), Nix 2.7.0, multi-user.

Nix was trying to build ~6,000 derivations at once, so in fairness, it was a heavy load.

I tried several of the most common ways to raise the max file limit with little success; but what finally worked was overriding the limit in the service:

sudo systemctl edit nix-daemon.service;
# In your editor set these fields
[Service]
LimitNOFILE=65535
LimitNOFILESoft=65535