feat(busybox): run before 80base so applets replace host binaries

[nadzyah]

Distros that ship large coreutils binaries (e.g. Ubuntu 25.10 with rust-coreutils) inflate the initrd because 80base copies the host's binaries like cp, ls, mv in full.

When the busybox module is included, 80base now skips any name busybox provides as an applet. 81busybox then symlinks those names to /usr/bin/busybox, restoring the smaller initrd footprint.

Bug-Ubuntu: https://bugs.launchpad.net/bugs/2150657

Checklist

• I have tested it locally
• I have reviewed and updated any documentation if relevant
• I am providing new code and test(s) for it

[bdrung]

I would prefer a solution that would work without adding the prefer_busybox setting. So use busybox applets in case busybox is included.

I looked at git log --follow --stat modules.d/81busybox/module-setup.sh to see why busybox is included so late and found only:

Commit 4e78a87 ("fix(busybox): install busybox symlinks later in the generation process"). After a search I found the corresponding #536 PR which links to #534. This bug is fixed by commit 95ba032 but the renumbering is unrelated to it.

Commit 3975e26 ("feat(busybox): use busybox --install to install itself") says:

Previous commits in this PR exposed a bug whereby some utilities (e.g. dd) had both a busybox applet version and a coreutils version.

The busybox dracut module should not enforce taking always the busybox version. The busybox --install command is safer option that does not overwrite an existing binary of already exists.

So how about moving busybox from 81busybox to 10busybox and address #2437 to ensure that the modules work with the busybox applets. If modules do not work with the busybox applets, they can do:

[ -L "$binary" ] && rm "$binary"
inst "$binary"

What do you think? @devkontrol @Conan-Kudo

[devkontrol]

I would prefer a solution that would work without adding the prefer_busybox setting.

YES. THIS.

inflate the initrd because 80base copies the host's binaries like cp, ls, mv in full.

On Alpine when both coreutils and busybox are installed (as Alpine packages), I think this is not always the case today. Let's understand when exactly the "expected" initrd generation happening already. So there might be a way to trigger similar kind of code path with Ubuntu as well.

I am not against making additional fixes in dracut to make this scenario more accessible, but adding a new global flag is probably not the best option here.

[bdrung]

The change looks good to me. I only have some minor remarks.

How long will this test take? Maybe it could be folded into an existing test in case the setup takes some time.

We should prefer busybox over dash in shell-interpreter (separate PR).

What kind of test have you done? Did you compare the initrd size on Ubuntu with this change?

[devkontrol]

Fedora failure for test 83 might be related to #2454, please double-check (see also #2455 (comment)).

[challvy]

Fedora failure for test 83 might be related to #2454, please double-check (see also #2455 (comment)).

Yes, I think so. @devkontrol Looking at TEST-83 in PR #2451, it checks cp/ls/mv/rm/mkdir/sleep/tr under usr/bin/. My current fix(#2455 #2454 ) only handles sh (the immediate kdump blocker) on Fedora 43.

Is it necessary to normalize all applet paths to /usr/bin/? Patch like this:

diff --git a/modules.d/81busybox/module-setup.sh b/modules.d/81busybox/module-setup.sh
index 402e1fc2..b5767ea6 100755
--- a/modules.d/81busybox/module-setup.sh
+++ b/modules.d/81busybox/module-setup.sh
@@ -30,9 +30,19 @@ install() {
         _path=$(find_binary "$_i")
         [ -z "$_path" ] && continue

+        # 1. Install at host's path (preserves topology, e.g. /sbin/ifconfig)
         # do not remove existing destination files
-        [ -e "${_dstdir}/$_path" ] && continue
-
-        ln_r /usr/bin/busybox "$_path"
+        [ -e "${_dstdir}/$_path" ] || ln_r /usr/bin/busybox "$_path"
+
+        # 2. Ensure canonical /usr/bin/<applet> always exists (POSIX layout).
+        # find_binary's literal path depends on $DRACUT_PATH ordering and may
+        # resolve to e.g. /usr/sbin/sh on Fedora 42+ where the busybox package
+        # places applet symlinks under /usr/sbin/. Standard scripts expect
+        # /usr/bin/<applet> (and #!/bin/sh -> /usr/bin/sh on merged-usr), so
+        # create the canonical symlink unconditionally if absent.
+        local _canonical="/usr/bin/${_path##*/}"
+        [ "$_path" = "$_canonical" ] && continue
+        [ -e "${_dstdir}${_canonical}" ] && continue
+        ln_r /usr/bin/busybox "$_canonical"
     done
 }

[devkontrol]

Fedora failure for test 83 might be related to #2454, please double-check (see also #2455 (comment)).

Yes, I think so.

One way to proceed is to exclude test 83 for Fedora in the GitHub Action and deal with it as a follow-up PR.

[Conan-Kudo]

"host bloat" is not accurate for the issue. The issue is that you have two separate implementations of UNIX core utilities installed and you want to use different ones for initramfs and the real system. Whether your choices are correct or not are separate issues, but regardless, it isn't "bloat". Please don't misuse words here, because it will confuse people reviewing it later.

[nadzyah]

"host bloat" is not accurate for the issue. The issue is that you have two separate implementations of UNIX core utilities installed and you want to use different ones for initramfs and the real system. Whether your choices are correct or not are separate issues, but regardless, it isn't "bloat". Please don't misuse words here, because it will confuse people reviewing it later.

The wording in question was not introduced by me. The title and commit message were changed during review, and I kept that wording when amending the commit to address subsequent review comments.

I’m happy to change the terminology if “host bloat” is not considered accurate. However, the comment assumes I chose that wording intentionally, which wasn’t the case. A simple suggestion for alternative wording would have been sufficient.

[devkontrol]

The term "bloat" carries a negative connotation. Let's try to avoid it.

When prefer_busybox=yes and the busybox module is included

The PR no longer uses this flag, lets remove it from the PR description.

[bdrung]

We should install the additional items from install_optional_items after we ran the modules. Then we would not need to do this workaround here. But let's do this change separately.

[bdrung]

Done in #2511

[bdrung]

#2511 is controversial. So replacing existing binaries by symlinks would be controversial as well. So can we not do it in this PR?

[bdrung]

This means that the test needs to ignore cat, rm, etc that dracut.conf.d/fedora/01-dist.conf includes via install_optional_items on Fedora.

[bdrung]

This exclusion can be dropped. We have busybox in azurelinux now: 874c545

[collinfunk]

Note that GNU coreutils has supported building as a single binary since 8.23, released 2014-07-18. If your only concern is size, I am not sure why you would need busybox. See:

$ { ./configure --enable-single-binary CFLAGS="$(rpm --eval '%{build_cflags}') -fno-strict-aliasing -fpic"; make -j 16; strip src/coreutils; } > /dev/null 2>&1
$ du --summarize --apparent-size src/coreutils /bin/busybox
1450	src/coreutils
1408	/bin/busybox

[bdrung]

Discussion from #2512: we currently cannot use blkid from busybox. The output is different and we use several arguments that are not supported by busybox.

[devkontrol]

Discussion from #2512: we currently cannot use blkid from busybox. The output is different and we use several arguments that are not supported by busybox.

It would be beneficial to have a busybox test case, independently from resolving https://bugs.launchpad.net/bugs/2150657.

Perhaps test case should be factored out into its own PR and for now the test case would check for blkid (or some other applet) coming from busybox (and not for the problematic cp, ls, mv binaried).

One of my concern is that distro's package busybox quite differently (different path, different applets, etc). Thsi test case would allow us to have some experience with busybox in our CI.