Re: [PATCH] image: limit search for *.core to regular files

["MOESSBAUER, Felix" <felix.moessbauer@siemens.com>]

On Tue, 2024-03-12 at 09:11 +0100, 'Jan Kiszka' via isar-users wrote:
> On 12.03.24 08:39, 'Cedric Hombourger' via isar-users wrote:
> > Code to search and delete core dumps in the build tree assumes that
> > the build host has a kernel.core_pattern setting which would result
> > in core dumps having a .core file suffix: this is not guaranteed.
> > One
> > may also argue that the build should have failed if a process
> > executed
> > under qemu-user got to crash (and we should check why qemu has 
> > crashed

Well... It's not that easy. This is the third time for me that this
coredump discussion pops up somewhere.

Many builders (like CMake) use feature probing (e.g. to check for AVX2)
which execute test examples that either succeed or crash with a
coredump [1]. While most of our builders run inside the schroot, there
might still be cases outside schroot where this is the expected
behavior.

> > and fix it). My vote would be to kill that code but for now, make
> > it
> > less wrong by restricting the search to regular files suffixed with
> > .core (this would at least stop isar from moving directories such
> > as
> > "org.eclipse.equinox.p2.core" out of the image).
> > 
> > Signed-off-by: Cedric Hombourger <cedric.hombourger@siemens.com>
> > ---
> >  meta/classes/image.bbclass | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/meta/classes/image.bbclass
> > b/meta/classes/image.bbclass
> > index 73f1d52c..793c21a2 100644
> > --- a/meta/classes/image.bbclass
> > +++ b/meta/classes/image.bbclass
> > @@ -457,7 +457,7 @@ EOSUDO
> >  
> >      # Sometimes qemu-user-static generates coredumps in chroot,
> > move them
> >      # to work temporary directory and inform user about it.
> > -    for f in $(sudo find ${ROOTFSDIR} -name *.core); do
> > +    for f in $(sudo find ${ROOTFSDIR} -type f -name *.core); do
> >          sudo mv "${f}" "${WORKDIR}/temp/"
> >          bbwarn "found core dump in rootfs, check it in
> > ${WORKDIR}/temp/${f##*/}"
> >      done
> 
> Yeah, too much heuristics in play now. We could add a list of valid
> "core" files on top, but maybe we should rather demand core file
> generation being disabled during the build and enforcing that.

We had exactly that discussion on the KAS ML as well, were I tried to
introduce a warning on default coredump configurations (which are BTW
really tricky to debug on CI systems). However that was not accepted as
the builders itself shall be responsible for a suitable coredump
configuration [2]. The truths is probably somewhere in between.

Just disabling the coredump generation is not easily possible, as there
is no "coredump" namespace in the kernel. By that, you would fiddle
around with the global system config and potentially interfere with
systemd (systemd-coredump).

The probably best thing we could do is to look for common coredump
patterns and check with file if these are actual coredumps.

[1] https://github.com/DynamoRIO/dynamorio/issues/6126
[2] https://groups.google.com/g/kas-devel/c/-sEyujhICfw/m/1UtVfsDRAAAJ

Best regards,
Felix

> 
> Jan
> 
> -- 
> Siemens AG, Technology
> Linux Expert Center
> 

-- 
Siemens AG, Technology
Linux Expert Center