[Devel] [RFC][PATCH 3/4]: Enable multiple mounts of /dev/pts

[sukadev at us.ibm.com]

Pavel Emelianov [xemul at openvz.org] wrote:
| sukadev at us.ibm.com wrote:
| > Pavel Emelianov [xemul at openvz.org] wrote:
| > | Serge E. Hallyn wrote:
| > | > Quoting Pavel Emelyanov (xemul at openvz.org):
| > | >> [snip]
| > | >>
| > | >>>> Mmm. I wanted to send one small objection to Cedric's patches with mqns,
| > | >>>> but the thread was abandoned by the time I decided to do-it-right-now.
| > | >>>>
| > | >>>> So I can put it here: forcing the CLONE_NEWNS is not very good, since
| > | >>>> this makes impossible to push a bind mount inside a new namespace, which
| > | >>>> may operate in some chroot environment. But this ability is heavily
| > | >>> Which direction do you want to go?  I'm wondering whether mounts
| > | >>> propagation can address it.
| > | >> Hardly. AFAIS there's no way to let the chroot-ed tasks see parts of
| > | >> vfs tree, that left behind them after chroot, unless they are in the 
| > | >> same mntns as you, and you bind mount this parts to their tree. No?
| > | > 
| > | > Well no, but I suspect I'm just not understanding what you want to do.
| > | > But if the chroot is under /jail1, and you've done, say,
| > | > 
| > | > 	mkdir -p /share/pts
| > | > 	mkdir -p /jail1/share
| > | > 	mount --bind /share /share
| > | > 	mount --make-shared /share
| > | > 	mount --bind /share /jail1/share
| > | > 	mount --make-slave /jail1/share
| > | > 
| > | > before the chroot-ed tasks were cloned with CLONE_NEWNS, then when you
| > | > do
| > | > 
| > | > 	mount --bind /dev/pts /share/pts
| > | > 
| > | > from the parent mntns (not that I know why you'd want to do *that* :)
| > | > then the chroot'ed tasks will see the original mntns's /dev/pts under
| > | > /jail1/share.
| > | 
| > | I haven't yet tried that, but :( this function
| > | 
| > | 	static inline int check_mnt(struct vfsmount *mnt)
| > | 	{
| > | 	        return mnt->mnt_ns == current->nsproxy->mnt_ns;
| > | 	}
| > | 
| > | and this code in do_loopback
| > | 
| > |         if (!check_mnt(nd->mnt) || !check_mnt(old_nd.mnt))
| > |                 goto out;
| > | 
| > | makes me think that trying to bind a mount from another mntns
| > | ot _to_ another is prohibited... Do I miss something?
| > | 
| > | >>> Though really, I think you're right - we shouldn't break the kernel
| > | >>> doing CLONE_NEWMQ or CLONE_NEWPTS without CLONE_NEWNS, so we shouldn't
| > | >>> force the combination.
| > | >>>
| > | >>>> exploited in OpenVZ, so if we can somehow avoid forcing the NEWNS flag
| > | >>>> that would be very very good :) See my next comment about this issue.
| > | >>>>
| > | >>>>> Pavel, not long ago you said you were starting to look at tty and pty
| > | >>>>> stuff - did you have any different ideas on devpts virtualization, or
| > | >>>>> are you ok with this minus your comments thus far?
| > | >>>> I have a similar idea of how to implement this, but I didn't thought
| > | >>>> about the details. As far as this issue is concerned, I see no reasons
| > | >>>> why we need a kern_mount-ed devtpsfs instance. If we don't make such,
| > | >>>> we may safely hold the ptsns from the superblock and be happy. The
| > | >>>> same seems applicable to the mqns, no?
| > | >>> But the current->nsproxy->devpts->mnt is used in several functions in
| > | >>> patch 3.
| > | >> Indeed. I overlooked this. Then we're in a deep ... problem here.
| > | >>
| > | >> Breaking this circle was not that easy with pid namespaces, so
| > | >> I put the strut in proc_flush_task - when the last task from the
| > | >> namespace exits the kern-mount-ed vfsmnt is dropped, but we can't
| > | >> do the same stuff with devpts.
| > | > 
| > | > But I still don't see what the problem is with my proposal?  So long as
| > | > you agree that if there are no tasks remaining in the devptsns,
| > | > then any task which has its devpts mounted should see an empty directory
| > | > (due to sb->s_info being NULL), I think it works.
| > | 
| > | Well, if we _do_ can handle the races with ns->devpts_mnt switch
| > | from not NULL to NULL, then I'm fine with this approach.
| > | 
| > | I just remember, that with pid namespaces this caused a complicated
| > | locking and performance degradation. This is the problem I couldn't
| > | remember yesterday.
| > 
| > Well, iirc, one problem with pid namespaces was that we need to keep
| > the task and pid_namespace association until the task was waited on
| > (for instance the wait() call from parent needs the pid_t of the
| > child which is tied to the pid ns in struct upid).
| > 
| > For this reason, we don't drop the mnt reference in free_pid_ns() but
| > hold the reference till proc_flush_task().
| > 
| > With devpts, can't we simply drop the reference in free_pts_ns() so
| > that when the last task using the pts_ns exits, we can unmount and
| > release the mnt ?
| 
| I hope we can. The thing I'm worried about is whether we can correctly
| handle race with this pointer switch from NULL to not-NULL.
| 
| > IOW, do you suspect that the circular reference leads to leaking vfsmnts ?
| > 
| 
| Of course! If the namespace holds the vfsmnt, vfsmnt holds the superblock
| and the superblock holds the namespace we won't drop this chain ever,
| unless some other object breaks this chain.

Of course :-) I had a bug in new_pts_ns() that was masking the problem.
I had

	ns->mnt  = kern_mount_data()...

	...
	kref_init(&ns->kref);

So the kref_init() would overwrite the reference got by devpts_set_sb()
and was preventing the leaking vfsmnt in my test.

Thanks Pavel,

Sukadev
_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers