>>> 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 'll take a look at it for the mq namespace.

we will need to flush the dcache in some way nop ? to make sure the lookup 
in the directory fails to return anything after the ns has become NULL. 
I'm not an fs expert so I might be completely wrong there but I'll study 
in this direction to see if we can drop the CLONE_NEWNS.
 
> I just remember, that with pid namespaces this caused a complicated
> locking and performance degradation. This is the problem I couldn't
> remember yesterday.

That might have been bc you had to invalidate the /proc dentries ? 

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