cp -R reads device nodes as streams instead of recreating them

[sylvestre]

Component

cp

Description

When copying directories recursively, uutils treats character and block device nodes as stream sources and reads bytes from them into regular files at the destination. This breaks device semantics: /dev/null becomes an empty regular file.

GNU cp preserves device nodes as device nodes during recursive copy by default. The --copy-contents flag explicitly enables stream-reading behavior. uutils currently lacks device node recreation (via mknod) and always treats devices as streams because is_stream() returns true for char/block devices.

This breaks recursive copies of system trees like chroot environments, initramfs staging, or container filesystems where device nodes must retain their type and major/minor numbers.

Test / Reproduction Steps

src="$(mktemp -d)"
dst="$(mktemp -d)"

sudo mknod "${src}/null" c 1 3

cp -R "${src}" "${dst}"

ls -l "${src}/null" "${dst}/null"

Impact

Recursive copies of filesystem trees containing device nodes produce incorrect results, breaking consumers that expect device semantics. Unbounded devices cause disk consumption and runtime denial-of-service.

[Thanhphan1147]

Hi, I would like to work on this, though it might take me a bit of time to understand how to recreate device nodes

[cerdelen]

I have a question regarding the intended behaviour.

I wrote code using mknod to properly create device nodes when encountering them in cp. There are some edge cases which im not sure how to handle.

when using cp to copy a device node to a destination of another device node gnu cp doesn't fail but also it doesnt overwrite the device node. Should this be recreated as a hardcoded workaround in code to silently omit these cases and return a success err code. Or should it actually contrary to GNU coreutils return an error?

[julian-klode]

Practically the same as #8679?