Bug 16028

*** Bug 16027 has been marked as a duplicate of this bug. ***

Ross will try to find an owner.

Of note, not all of coreutils is redirected to uutils:

lrwxrwxrwx   1 root root        8 Sep 26 21:16 chmod -> gnuchmod
lrwxrwxrwx   1 root root        8 Sep 26 21:16 chown -> gnuchown
lrwxrwxrwx   1 root root        5 Sep 26 21:16 cp -> gnucp
lrwxrwxrwx   1 root root        5 Sep 26 21:16 df -> gnudf
lrwxrwxrwx   1 root root        5 Sep 26 21:16 mv -> gnumv
lrwxrwxrwx   1 root root        5 Sep 26 21:16 rm -> gnurm
lrwxrwxrwx   1 root root        7 Sep 26 21:16 true -> gnutrue

Easiest reproducer is glibc-locale:

ERROR: glibc-locale-2.42+git-r0 do_install: Execution of '/home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/temp/run.do_install.9977' failed with exit code 1
ERROR: Logfile of failure stored in: /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/temp/log.do_install.9977



| couldn't allocate absolute path for 'null'.
| couldn't allocate absolute path for 'null'.
| couldn't allocate absolute path for 'null'.
| couldn't allocate absolute path for 'null'.
...


And as expected, switching our hosttools/install link to to gnuinstall fixes this.

The crash is from pseudo, so the rust implementation is still calling some glibc functions that break pseudo.

So the good news is that uutils mostly wraps the libc functions, so eg install.rs ends up calling libc:chown() which is a binding to the libc function, not syscalls directly.

Which makes why it fails even more annoying...

I put PSEUDO_DEBUG=1 into glibc-locales.inc and replaced hosttools/install with a wrapper that prints the command being invoked and this happens:

DEBUG: Executing shell function do_install
openat(path /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/sources/glibc-2.42+git), flags 2000000, stat rc 0, stat mode 40755
openat(path /home/ross/poky/build/tmp/sysroots-components/x86-64-v3/glibc-stash-locale/usr/bin), flags 2144400, stat rc 0, stat mode 40755
RUNNING /home/ross/poky/build/tmp/hosttools/install -d /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/image/usr/bin/.
couldn't allocate absolute path for 'null'.
creating directory '/home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/image/usr/bin'
statx(path /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/image/usr/bin), flags 0, stat rc 0, stat uid 1750
openat(path /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/sources/glibc-2.42+git), flags 2000000, stat rc 0, stat mode 40755
openat(path /home/ross/poky/build/tmp/sysroots-components/x86-64-v3/glibc-stash-locale/usr/bin), flags 2144400, stat rc 0, stat mode 40755
chdirat : (48534) (no request)
RUNNING /home/ross/poky/build/tmp/hosttools/install -m 0755 -t /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/image/usr/bin/. /home/ross/poky/build/tmp/sysroots-components/x86-64-v3/glibc-stash-locale/usr/bin/./localedef
couldn't allocate absolute path for 'null'.
statx(path /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/image/usr/bin), flags 0, stat rc 0, stat uid 1750
statx(path /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/image/usr/bin), flags 0, stat rc 0, stat uid 1750
openat(path /home/ross/poky/build/tmp/sysroots-components/x86-64-v3/glibc-stash-locale/usr/bin/localedef), flags 2000000, stat rc 0, stat mode 100755
openat_creat: /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/image/usr/bin/localedef -> 0100755
openat(path /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/image/usr/bin/localedef), flags 2001101, stat rc 0, stat mode 100755
statx(path /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/image/usr/bin/localedef), flags 10400, stat rc 0, stat uid 1750
statx(path /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/image/usr/bin/localedef), flags 0, stat rc 0, stat uid 1750
'/home/ross/poky/build/tmp/sysroots-components/x86-64-v3/glibc-stash-locale/usr/bin/./localedef' -> '/home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/image/usr/bin/./localedef'
openat(path /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/sources/glibc-2.42+git), flags 2000000, stat rc 0, stat mode 40755
openat(path /home/ross/poky/build/tmp/sysroots-components/x86-64-v3/glibc-stash-locale/usr/lib/locale), flags 2144400, stat rc 0, stat mode 40755
...

Copying in Mark as he knows way more about pseudo than I do.

Created attachment 5150 [details]
Verbose pseudo log

Attached is a verbose log from pseudo with PSEUDO_DEBUG=foyw and breaking the install loop in glibc-locale.inc:copy_locale_files() on the first iteration (so, two calls to install, one to make a directory and the other to copy a file).

I think those are a false lead, the actual fail is a little later:

| RUNNING /home/ross/poky/build/tmp/hosttools/install -m 0644 /home/ross/poky/build/tmp/sysroots-components/x86-64-v3/glibc-stash-locale/SUPPORTED /home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/SUPPORTED
...
| wrapper called: lchown
| lchown ignored path, calling real syscall.
| wrapper completed: lchown returns -1 (errno: 22)
| /usr/bin/install: failed to chown '/home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/SUPPORTED': changing group of '/home/ross/poky/build/tmp/work/x86-64-v3-poky-linux/glibc-locale/2.42+git/SUPPORTED': Invalid argument (os error 22)

Note that this error is expected: we don't pseudo in that directory.

A minimal reproducer is this recipe:

SUMMARY = "foo"
LICENSE = "CLOSED"

export PSEUDO_DEBUG = "foyw"

do_install() {
        touch ${WORKDIR}/foo
        install ${WORKDIR}/foo ${WORKDIR}/foo2
}

I believe the problem here is that:
- install is being asked to put a file into WORKDIR
- as it's in fakeroot context, uid is root:root
- uutils-install wants to chown to root:root once it has created the file
- however, workdir is outside of the allowed pseudo directories and so it doesn't intercept the call and forwards it, which then fails because normal users can't chown(root).

Note that there are almost definitely multiple issues: #16027 reported symlinks appearing in build/ which looked like they were part of the busybox build too.

Hains: if you're willing to test, switching your system back to having coreutils-from-uutils (so /usr/bin/install is from uutils, not gnuinstall) would be interesting. I'm curious what fails if you create a clean build tree, and then switch the tmp/hosttools/install symlink to /usr/bin/gnuinstall.

I wish you could edit previous comments in bugzilla...

The spurious links are from busybox's update-alternatives running at rootfs time:

+ update-alternatives --install /bin/ash ash /bin/busybox.nosuid 50
couldn't allocate absolute path for 'null'.
couldn't allocate absolute path for 'null'.
couldn't allocate absolute path for 'null'.
couldn't allocate absolute path for 'null'.
couldn't allocate absolute path for 'null'.
couldn't allocate absolute path for ''.
tail: cannot open 'standard input' for reading: No such file or directory
couldn't allocate absolute path for 'null'.
couldn't allocate absolute path for 'null'.
ln: Permission denied
%post(busybox-1.37.0-r0.x86_64_v3): waitpid(61372) rc 61372 status 100


At least two issues:
1) bad install/chown/pseudo behaviour that doesn't match gnu install
2) "couldn't allocate absolute path for 'null'" at least spams the logs
2.5) ...and possibly causes further problems

Notably, I reproduced the rootfs failure by doing a build entirely from public sstate, so the actual packages are good and it's just the rootfs creation that is failing.

I started from scratch. I removed the entire build directory.

Beside errors during rootfs and do_install glibc, I got an error with base-password during do_package stage when coreutils-from-gnu has been removed from system.

For test I had to comment out line 112, 'addtask do_package after do_populate_sysroot'.

Created attachment 5152 [details]
Error base-passwd

Edit comment 12. glibc must be glibc-locale.

Confirmed issue from sysvinit-inittab:

3) tail: cannot open 'standard input' for reading: No such file or directory

4) various tools call statx() with NULL instead of "", which appears to be wrong according to POSIX, and causes "couldn't allocate absolute path for 'null'".

(In reply to Ross Burton from comment #16)
> 4) various tools call statx() with NULL instead of "", which appears to be
> wrong according to POSIX, and causes "couldn't allocate absolute path for
> 'null'".

I wonder if syscall has AT_EMPTY_PATH in flags, which would let kernel process/ignore path variable even when its NULL, and use the file descriptor for operation ( provided its fstatat call )

Sent a patch to prefer gnu-prefixed binaries until the breakage is fixed.

5) perl's build does a cp in a deleted directory which upsets things. https://github.com/openembedded/openembedded-core/pull/110 has a workaround, needs a little investigation to see if this is a cp-specific issue or another instance of (2).

The workaround to prefer gnu-prefixed tools has been merged in oe-core 16f2684ebeffa72b5d90525cf9102751b68c298e.

Root-causing the other bugs are not a high priority release blocker.

Issues that have been filed with uutils, which may or may not be bugs in uutils itself but are definitely either issues in pseudo, or exposed by pseudo:

https://github.com/uutils/coreutils/issues/9288
https://github.com/uutils/coreutils/issues/9292

Pulling out "couldn't allocate absolute path for 'null'" to 16075.

Created attachment 5163 [details]
fix up the implementation of realpath

The pseudo implementation of `realpath()' may return a path which doesn't exist. This is not POSIX compliant and causes troubles with uutils (coreutils in Rust).

For example, the tail commands tries to determine the file path of its standard input file descriptor by calling `realpath("/dev/fd/0")'. When the input is a pipe, the GNU C library returns NULL but pseudo returns `/proc/<pid>/fd/pipe:[xxxxxx]'. As it got a path, the tail command tries to open it and fails.

Hi all,

I attached a patch which enables to build a full image on Ubuntu 25.10. It was tested with scarthgap only. I hope this helps.

Gauthier, could you post that patch to the mailing list for review?  The last few lines of the README has the configuration for git-send-email. Thanks!

Hi Ross,

I've just done it: https://lists.yoctoproject.org/g/yocto-patches/topic/pseudo_patch_unix/116859482. Let me know if I did it wrong or missed something.

Cheers

Looks good, thanks

I think a simple test case for the tail issue (using _GNU_ tail) would be:

echo "test" | tail /dev/fd/0

without pseudo this returns the expected "tail"


with pseudo this returns:

tail: cannot open '/dev/fd/0' for reading: No such file or directory

I've sent a patch for review.  After doing further investigation the function (and what it calls) pseudo_fix_path is what needs to do special behavior to address the issue.  It appears the problems are unique to the special /proc filesystem semantics, so the fix is likely to resolve path names until /proc and then stop doing further resolution.