[BUG?] fsconfig restart_syscall failed

[BUG?] fsconfig restart_syscall failed

[Cai Xinchen]

Hello:
   I am doing some test for kernel 6.4, util-linux version:2.39.1.
Have you encountered similar problems? If there is a fix, please
let me know.
Thank you very much

--------------------------------------------------

util-linux version 2.39.1 call mount use fsopen->fsconfig->fsmount->close
instead of mount syscall.

And use this shell test:

#!/bin/bash
mkdir -p /tmp/cgroup/cgrouptest
while true
do
         mount -t cgroup -o none,name=foo cgroup /tmp/cgroup/cgrouptest
         ret=$?
         if [ $ret -ne 0 ];then
                 echo "mount failed , $ret"
         fi
         umount /tmp/cgroup/cgrouptest
         ret=$?
         if [ $ret -ne 0 ];then
                 echo "umount failed, $ret"
         fi
done

And as a result, we mount cgroup immediately after umount, it will return
failed.

in fsconfig syscall, we find this stack:

SYSCALL_DEFINE5(fsconfig, ...)
         vfs_fsconfig_locked
                 if (fc->phase != FS_CONTEXT_CREATE_PARAMS)
                         return -EBUSY;

                 vfs_get_tree
                         fc->ops->get_tree // cgroup1_get_tree
                                 if (!ret && !percpu_ref_tryget_live
(&ctx->root->cgrp.self.refcnt))
                                         ret = 1;
                                 ...
                                 if (unlikely(ret > 0)) {
                                         msleep(10);
                                         restart_syscall();
                                 }
                 ...
                 fc->phase = FS_CONTEXT_FAILED;

in mount syscall, no function will check fs->phase, and fc is recreate
in monnt syscall. However, in fdconfig syscall, fc->phase is not initial as
FS_CONTEXT_CREATE_PARAMS, restart_syscall will return -EBUSY. fc is created
in fsopen syscall.

Re: [BUG?] fsconfig restart_syscall failed

[Christian Brauner]

On Fri, Sep 22, 2023 at 10:18:24AM +0800, Cai Xinchen wrote:
> Hello:
>   I am doing some test for kernel 6.4, util-linux version:2.39.1.
> Have you encountered similar problems? If there is a fix, please
> let me know.
> Thank you very much
> 
> --------------------------------------------------
> 
> util-linux version 2.39.1 call mount use fsopen->fsconfig->fsmount->close
> instead of mount syscall.
> 
> And use this shell test:
> 
> #!/bin/bash
> mkdir -p /tmp/cgroup/cgrouptest
> while true
> do
>         mount -t cgroup -o none,name=foo cgroup /tmp/cgroup/cgrouptest


> in mount syscall, no function will check fs->phase, and fc is recreate
> in monnt syscall. However, in fdconfig syscall, fc->phase is not initial as
> FS_CONTEXT_CREATE_PARAMS, restart_syscall will return -EBUSY. fc is created
> in fsopen syscall.

Mount api system calls aren't restartable so that doesn't work. cgroup2
doesn't have this issue, only cgroup1 has. So cgroup1_get_tree() should
probably be fixed if anyone cares.

Re: [BUG?] fsconfig restart_syscall failed

[Alexander Mikhalitsyn]

On Fri, 22 Sep 2023 10:08:36 +0200
Christian Brauner <brauner@kernel.org> wrote:

> On Fri, Sep 22, 2023 at 10:18:24AM +0800, Cai Xinchen wrote:
> > Hello:
> >   I am doing some test for kernel 6.4, util-linux version:2.39.1.
> > Have you encountered similar problems? If there is a fix, please
> > let me know.
> > Thank you very much
> > 
> > --------------------------------------------------
> > 
> > util-linux version 2.39.1 call mount use fsopen->fsconfig->fsmount->close
> > instead of mount syscall.
> > 
> > And use this shell test:
> > 
> > #!/bin/bash
> > mkdir -p /tmp/cgroup/cgrouptest
> > while true
> > do
> >         mount -t cgroup -o none,name=foo cgroup /tmp/cgroup/cgrouptest
> 
> 
> > in mount syscall, no function will check fs->phase, and fc is recreate
> > in monnt syscall. However, in fdconfig syscall, fc->phase is not initial as
> > FS_CONTEXT_CREATE_PARAMS, restart_syscall will return -EBUSY. fc is created
> > in fsopen syscall.
> 
> Mount api system calls aren't restartable so that doesn't work. cgroup2
> doesn't have this issue, only cgroup1 has. So cgroup1_get_tree() should
> probably be fixed if anyone cares.
> 

Dear colleagues,

I've met the same issue a few years ago and tried to fix it:
https://lore.kernel.org/all/20200923164637.13032-1-alexander.mikhalitsyn@virtuozzo.com/

but didn't come into agreement about this.

Kind regards,
Alex