File access time is not normalized when importing to store

[infinisil]

Describe the bug

The file access time is not normalized when importing that file to the store, leading to a potential build-time impurity.

Steps To Reproduce

1. Create a test file:

   touch test
   

2. Create a default.nix file with these contents:

   (import <nixpkgs> {}).runCommand "test" {} ''
     ls --full-time --time=atime ${./test} > $out
   ''

3. Build it and show the result, notice how the resulting access time is not normalized to unix epoch 1:

   $ nix-build
   this derivation will be built:
     /nix/store/zc8kj3fgb5rxij4kbx6p9ynlldig9kla-test.drv
   building '/nix/store/zc8kj3fgb5rxij4kbx6p9ynlldig9kla-test.drv'...
   /nix/store/gdzg99df8f8a6zpxxrzj2m7p439b09ln-test
   
   $ cat result
   -r--r--r-- 2 nobody nogroup 0 2023-05-31 19:01:19.000000000 +0000 /nix/store/vd14zxkay784callbkhnmd88qv3p2rip-test
   
   $ cp --copy-contents result before
   

4. Remove the imported file from the Nix store (requires a workaround, see Cannot delete store path with no roots #6141)

   $ rm result
   
   $ nix-store --query --referrers-closure /nix/store/vd14zxkay784callbkhnmd88qv3p2rip-test | xargs nix-store --delete
   finding garbage collector roots...
   removing stale link from '/nix/var/nix/gcroots/auto/7rfrjpxmdf6qk5q4h20s8ikk9lnw3347' to '/home/tweagysil/test/time3/result'
   deleting '/nix/store/zc8kj3fgb5rxij4kbx6p9ynlldig9kla-test.drv'
   deleting '/nix/store/gdzg99df8f8a6zpxxrzj2m7p439b09ln-test'
   deleting '/nix/store/vd14zxkay784callbkhnmd88qv3p2rip-test'
   deleting unused links...
   note: currently hard linking saves 15891.46 MiB
   3 store paths deleted, 0.00 MiB freed
   

5. Update the access time of the file:

   $ touch test
   

6. Build it again and diff the result, noticing that the result is different:

   $ nix-build
   this derivation will be built:
     /nix/store/zc8kj3fgb5rxij4kbx6p9ynlldig9kla-test.drv
   building '/nix/store/zc8kj3fgb5rxij4kbx6p9ynlldig9kla-test.drv'...
   /nix/store/gdzg99df8f8a6zpxxrzj2m7p439b09ln-test
   
   $ diff before result
   1c1
   < -r--r--r-- 2 nobody nogroup 0 2023-05-31 19:04:54.000000000 +0000 /nix/store/vd14zxkay784callbkhnmd88qv3p2rip-test
   ---
   > -r--r--r-- 2 nobody nogroup 0 2023-05-31 19:06:44.000000000 +0000 /nix/store/vd14zxkay784callbkhnmd88qv3p2rip-test
   

   Notice also how the hash of the file (/nix/store/vd14zxkay784callbkhnmd88qv3p2rip-test) did not change between the two builds, but the access time did change.

Expected behavior

The access time should get normalized to unix epoch 1, the same as is done to the modification time (mtime) already.

Additional context

The change time (ctime) and birth time (on some filesystems) is also not normalized, but this can't reasonable be fixed.

This work is sponsored by Antithesis ✨

Priorities

Add 👍 to issues you find important.

[thufschmitt]

I confirm it isn't. I suspect it's indeed a bug, but I'd like to get @edolstra's opinion as it could have been a deliberate tradeoff from the origins of Nix (also note that build outputs aren't normalized either, except for their mtime)

[roberth]

atime and GC

atime could be useful for informing what garbage to collect first, as a store's garbage is also a cache, and atime is a good indicator of relevance. Performance may or may not be a challenge for the atime for GC idea.

Related:

• idea: nix-collect-garbage option to also respect creation date of a derivation #2793 (also atime is discussed)