bash_5 fails on hydra but succeeds locally

[meithecatte]

Describe the bug
The most recent Hydra build of bash_5 failed with

configure: WARNING: unrecognized options: --disable-static
building
build flags: -j2 -l2 SHELL=/nix/store/f7jzmxq9bpbxsg69cszx56mw14n115n5-bash-4.4-p23/bin/bash
cd . && autoconf
/nix/store/f7jzmxq9bpbxsg69cszx56mw14n115n5-bash-4.4-p23/bin/bash: autoconf: command not found
make: *** [Makefile:785: configure] Error 127

However, running nix-build -A bash_5 succeeds and produces the store path /nix/store/yqk7pzd398gxnd7ph6vbka7cm1hvb6ai-bash-5.1-p4 - exactly as listed on the Hydra page.

Expected behavior
The results should agree.

Notify maintainers
The package is maintained by @peti @dtzWill but I suspect there's something deeper at play here.

Metadata

• system: "x86_64-linux"
• host os: Linux 5.4.99, NixOS, 21.05pre272380.f6b5bfdb470 (Okapi)
• multi-user?: yes
• sandbox: yes
• version: nix-env (Nix) 2.3.10
• channels(kuba): "nixpkgs-21.05pre272380.f6b5bfdb470"
• channels(root): "nixos-21.05pre272380.f6b5bfdb470"
• nixpkgs: /home/kuba/.nix-defexpr/channels/nixpkgs

Maintainer information:

# a list of nixpkgs attributes affected by the problem
attribute: bash_5 bashInteractive_5
# a list of nixos modules affected by the problem
module:

[andersk]

This is a nondeterministic failure. I just saw it happen locally, but I reran the same build and it succeeded. I think I figured out why—note the following difference between the bash log and the bash_5 log.

bash:

applying patch /nix/store/d6a4ngpmijv1sk5q7al7ka2lhy8m5ryc-pgrp-pipe-4.4.patch
patching file ./configure
patching file ./configure.ac

bash_5:

applying patch /nix/store/hr5f8216rr5dzbl4ib4gzl76yp4bbrip-pgrp-pipe-5.1.patch
patching file ./configure
Hunk #1 succeeded at 20407 with fuzz 2 (offset 4095 lines).
Not setting time of file ./configure (time mismatch)
patching file ./configure.ac
Hunk #1 succeeded at 1159 with fuzz 2 (offset 51 lines).
Not setting time of file ./configure.ac (time mismatch)

The 4.4 patch applies cleanly with matched timestamps, so configure and configure.ac get their timestamps set to the ones in the patch file, with configure getting a significantly newer timestamp than configure.ac. But the 5.1 patch applies with mismatched timestamps, so configure and configure.ac both get their timestamps set to the current time, in that order. It’s then up to rounding whether this rule in Makefile.in gets triggered:

# comment out for distribution
$(srcdir)/configure:	$(srcdir)/configure.ac $(srcdir)/aclocal.m4 $(srcdir)/config.h.in
	cd $(srcdir) && autoconf

This patch was added in #77196 (cc @raboof @FRidh).

[meithecatte]

If we're patching configure.ac, perhaps configure should get generated by autoconf and not patched in?

[vcunat]

Well, I suspect it might complicate bootstrapping if we add those dependencies here. Note that the patch fixes both of the files equally, so it should be just a timestamp issue. 🤔 perhaps just swap the two parts inside each patch, so that configure is patched after configure.ac?

[raboof]

the 5.1 patch applies with mismatched timestamps, so configure and configure.ac both get their timestamps set to the current time, in that order. It’s then up to rounding whether this rule in Makefile.in gets triggered

Ouch, great find!

[raboof]

It would be cool if we could tell patch to fail rather than warn when the timestamps don't match, but it looks like that isn't possible right now (https://savannah.gnu.org/bugs/index.php?60266)

[vcunat]

Fixes merged.