Re: autofs: add autofs_parse_fd()

[Ian Kent <raven@themaw.net>]

On 20/10/23 17:02, Arnd Bergmann wrote:
> On Fri, Oct 20, 2023, at 09:48, Naresh Kamboju wrote:
>> On Fri, 20 Oct 2023 at 12:07, Arnd Bergmann <arnd@arndb.de> wrote:
>>> On Thu, Oct 19, 2023, at 17:27, Naresh Kamboju wrote:
>>>> The qemu-x86_64 and x86_64 booting with 64bit kernel and 32bit rootfs we call
>>>> it as compat mode boot testing. Recently it started to failed to get login
>>>> prompt.
>>>>
>>>> We have not seen any kernel crash logs.
>>>>
>>>> Anders, bisection is pointing to first bad commit,
>>>> 546694b8f658 autofs: add autofs_parse_fd()
>>>>
>>>> Reported-by: Linux Kernel Functional Testing <lkft@linaro.org>
>>>> Reported-by: Anders Roxell <anders.roxell@linaro.org>
>>> I tried to find something in that commit that would be different
>>> in compat mode, but don't see anything at all -- this appears
>>> to be just a simple refactoring of the code, unlike the commits
>>> that immediately follow it and that do change the mount
>>> interface.
>>>
>>> Unfortunately this makes it impossible to just revert the commit
>>> on top of linux-next. Can you double-check your bisection by
>>> testing 546694b8f658 and the commit before it again?
>> I will try your suggested ways.
>>
>> Is this information helpful ?
>> Linux-next the regression started happening from next-20230925.
>>
>> GOOD: next-20230925
>> BAD: next-20230926
>>
>> $ git log --oneline next-20230925..next-20230926 -- fs/autofs/
>> dede367149c4 autofs: fix protocol sub version setting
>> e6ec453bd0f0 autofs: convert autofs to use the new mount api
>> 1f50012d9c63 autofs: validate protocol version
>> 9b2731666d1d autofs: refactor parse_options()
>> 7efd93ea790e autofs: reformat 0pt enum declaration
>> a7467430b4de autofs: refactor super block info init
>> 546694b8f658 autofs: add autofs_parse_fd()
>> bc69fdde0ae1 autofs: refactor autofs_prepare_pipe()
> Right, and it looks like the bottom five patches of this
> should be fairly harmless as they only try to move code
> around in preparation of the later changes, and even the
> other ones should not cause any difference between a 32-bit
> or a 64-bit /sbin/mount binary.
>
> If the native (full 64-bit or full 32-bit) test run still
> works with the same version, there may be some other difference
> here.
>
>>> What are the exact mount options you pass to autofs in your fstab?
>> mount output shows like this,
>> systemd-1 on /proc/sys/fs/binfmt_misc type autofs
>> (rw,relatime,fd=30,pgrp=1,timeout=0,minproto=5,maxproto=5,direct,pipe_ino=1421)
> This is only the binfmt-misc mount, which should not
> prevent your rootfs from getting mounted, but it's possible
> that failure to mount this prevents you from running
> 32-bit binaries.
>
> I see this comes from the "proc-sys-fs-binfmt_misc.automount"
> service in systemd.  I see this is defined in
> https://github.com/systemd/systemd/blob/main/units/proc-sys-fs-binfmt_misc.automount
> but I don't know exactly what its purpose is here. On a
> 64-bit system, you normally use compat_binfmt_elf.ko to run
> 32-bit binaries, and this does not require any specific mount
> points. Alternatively, you could use binfmt_misc.ko with
> the procfs mount to configure running arbitrary binary
> formats such as arm32 on x86_64 with qemu-user emulation.
>
> I double-checked your rootfs image from
> https://storage.tuxboot.com/debian/bookworm/i386/rootfs.ext4.xz
> to ensure that this indeed contains i386 executables rather than
> arm32 ones, and that is all fine.
>
> I also see in your log file at
> https://qa-reports.linaro.org/lkft/linux-next-master/build/next-20230926/testrun/20125035/suite/boot/test/gcc-13-lkftconfig-compat/log
> that it is running the i386 binaries from the rootfs, but
> it does get stuck soon after trying to set up the binfmt-misc
> mount at the end of the log:
>
> [[0;32m  OK  [0m] Reached target [0;1;39mlocal-fs.target[0m - Local File Systems.
>           Starting [0;1;39msystemd-binfmt.se…et Up Additional Binary Formats...
>           Starting [0;1;39msystemd-tmpfiles-… Volatile Files and Directories...
>           Starting [0;1;39msystemd-udevd.ser…ger for Device Events and Files...
> [   15.869404] igb 0000:01:00.0 eno1: renamed from eth0 (while UP)
> [   15.883753] igb 0000:02:00.0 eno2: renamed from eth1
> [   20.053885] (udev-worker) (175) used greatest stack depth: 12416 bytes left
> \x1dquit

Were there any console log messages at the time the problem occurred?


>
> I'm a bit out of ideas at that point, my best guess now is
> that your bisection points to something in autofs that makes
> it hang while setting up autofs, but that neither autofs
> nor binfmt-misc are actually being used otherwise.
>
> Maybe you can try to modify your rootfs to disable or remove
> the systemd-binfmt.service, to confirm that autofs is not
> actually needed here but does cause the crash?
>
>       Arnd