Re: [PATCH 0/3] umount fixes and cleanups

[Jan Kiszka <jan.kiszka@web.de>]

On 05.12.18 15:52, [ext] Jan Kiszka wrote:
> On 05.12.18 15:26, Maxim Yu. Osipov wrote:
>> On 12/5/18 3:14 PM, Jan Kiszka wrote:
>>> On 05.12.18 12:12, Maxim Yu. Osipov wrote:
>>>> On 12/5/18 12:29 PM, Jan Kiszka wrote:
>>>>> Patches related to resolving the pending CI issues as well as
>>>>> simplifying the umounts used during cleanup.
>>>>
>>>> Tried to run in patch queue:
>>>>
>>>> 027b7cf Remove redundant recursive umounts
>>>> c1bdc33 isar-events: Improve umount handler
>>>> b354273 ci: Wait for bitbake worker to finish before deleting artifacts
>>>> 9cf29e6 isar-bootstrap: Fix and cleanup bind mounting
>>>> b354026 isar-image: umount base-apt when doing offline build
>>>> e965c0d gitlab-ci: Switch to ci_build.sh
>>>> ...
>>>>
>>>> After execution of problematic test case (I rebooted PC and executed steps 
>>>> in clean tree):
>>>>
>>>> my stretch Debian system was entered into unusable state
>>>> as many important mounts were disappeared (see log of mount points
>>>> before and after execution  of last command attached).
>>>
>>> OK, so I took the time and extracted the root cause. We may see a Debian 4.9 
>>> kernel bug here:
>>>
>>> # mkdir my-mnt
>>> # mount --rbind /sys my-mnt
>>> # mount --make-rslave my-mnt
>>> # umount {-R,-l,-R -l} my-mnt
>>> # rmdir my-mnt
>>> rmdir: failed to remove 'my-mnt': Device or resource busy
>>>
>>> This works fine on my Leap 15.0 kernel (4.12) as well as the 4.4 kernel 
>>> (Ubuntu) used on our CI server.
>>
>> This also works fine on my Debian 4.9 kernel (I've ran this sequence for every 
>> combination of {-R, -l , -R -l}):
>>
>>> root@tiberius:~# uname -a
>>> Linux tiberius 4.9.0-8-amd64 #1 SMP Debian 4.9.110-3+deb9u6 (2018-10-08) 
>>> x86_64 GNU/Linux
> 
> Hmpf, rechecking... I was still on 3.16 from the jessie that I upgraded from. 
> Works fine with 4.9, indeed.
> 
>>> root@tiberius:~# mkdir my-mnt
>>> root@tiberius:~# mount --rbind /sys my-mnt
>>> root@tiberius:~# mount --make-rslave my-mnt
>>> root@tiberius:~# umount -R my-mnt/
>>> root@tiberius:~# rmdir my-mnt/
>>> root@tiberius:~# mkdir my-mnt
>>> root@tiberius:~# mount --rbind /sys my-mnt
>>> root@tiberius:~# mount --make-rslave my-mnt
>>> root@tiberius:~# umount -l my-mnt
>>> root@tiberius:~# rmdir my-mnt/
>>> root@tiberius:~# mkdir my-mnt
>>> root@tiberius:~# mount --rbind /sys my-mnt
>>> root@tiberius:~# mount --make-rslave my-mnt
>>> root@tiberius:~# umount -R -l my-mnt
>>> root@tiberius:~# rmdir my-mnt/
>>> root@tiberius:~# 
>>
> 
> OK, then bind-mounting itself likely not the root cause, and we need to dig 
> further.
> 

Here is the real solution, or at least a much better workaround, without
any reverts:

diff --git a/meta/classes/wic-img.bbclass b/meta/classes/wic-img.bbclass
index 225463e..c66d208 100644
--- a/meta/classes/wic-img.bbclass
+++ b/meta/classes/wic-img.bbclass
@@ -90,7 +90,7 @@ do_wic_image() {
     for dir in ${BBLAYERS} ${STAGING_DIR} ${ISARROOT}/scripts; do
 	sudo mkdir -p ${BUILDCHROOT_DIR}/$dir
         mountpoint ${BUILDCHROOT_DIR}/$dir >/dev/null 2>&1 \
-        || sudo mount --bind $dir ${BUILDCHROOT_DIR}/$dir
+        || sudo mount --bind --make-private $dir ${BUILDCHROOT_DIR}/$dir
     done
     export FAKEROOTCMD=${FAKEROOTCMD}
     export BUILDDIR=${BUILDDIR}


I'll explain tomorrow, also what else is broken in do_wic_image.

Jan