On 18/06/10 18:08 +0200, Oleg Nesterov wrote:
> On 06/18, Oleg Nesterov wrote:
> >
> > On 06/18, Louis Rilling wrote:
> > >
> > > On 17/06/10 23:20 +0200, Oleg Nesterov wrote:
> > > > On 06/16, Louis Rilling wrote:
> > > > >
> > > > > Detached tasks are not seen by zap_pid_ns_processes()->sys_wait4(), so
> > > > > that release_task()->proc_flush_task() of container init can be called
> > > > > before it is for some detached tasks in the namespace.
> > > > >
> > > > > Pin proc_mnt's in copy_process(), so that proc_flush_task() becomes safe
> > > > > whatever the ordering of tasks.
> > > >
> > > > I must have missed something, but can't we just move mntput() ?
> > >
> > > See the log of the commit introducing pid_ns_release_proc() (6f4e6433):
> > >
> > >     Sice the namespace holds the vfsmnt, vfsmnt holds the superblock and the
> > >     superblock holds the namespace we must explicitly break this circle to destroy
> > >     all the stuff.  This is done after the init of the namespace dies.
> >
> > I see thanks.
> 
> Not sure I ever understood this code. Certainly I can't say I understand
> it now. Still, do we really need this circle? I am almost sure the patch
> below is not right (and it wasn't tested at all), but could you take a
> look?

I won't pretend understanding better than you do. Still I can try to give my 2
cents.

Overall, I don't feel comfortable at being able to have a living proc_mnt
with a dead pid_ns. On the contrary, proc_mnt is almost never used from pid_ns.
proc_flush_task() excepted, the only users I saw use
current->nsproxy->pid_ns->proc_mnt, which makes them safe.

> 
> Note: afaics we have another problem. What if copy_process(CLONE_NEWPID)
> fails after pid_ns_prepare_proc() ? Who will do mntput() ? This patch
> should fix this too (again, ___if___ it correct).

Sounds true.

> 
> Oleg.
> 
>  include/linux/pid_namespace.h |    1 +
>  kernel/pid_namespace.c        |    3 +++
>  fs/proc/base.c                |    4 ----
>  fs/proc/root.c                |   18 ++++++++++++++----
>  4 files changed, 18 insertions(+), 8 deletions(-)
> 
> --- 34-rc1/include/linux/pid_namespace.h~PID_NS	2010-06-18 17:48:56.000000000 +0200
> +++ 34-rc1/include/linux/pid_namespace.h	2010-06-18 17:53:11.000000000 +0200
> @@ -26,6 +26,7 @@ struct pid_namespace {
>  	struct pid_namespace *parent;
>  #ifdef CONFIG_PROC_FS
>  	struct vfsmount *proc_mnt;
> +	struct work_struct proc_put;
>  #endif
>  #ifdef CONFIG_BSD_PROCESS_ACCT
>  	struct bsd_acct_struct *bacct;
> --- 34-rc1/kernel/pid_namespace.c~PID_NS	2010-06-18 17:48:56.000000000 +0200
> +++ 34-rc1/kernel/pid_namespace.c	2010-06-18 17:53:44.000000000 +0200
> @@ -10,6 +10,7 @@
>  
>  #include <linux/pid.h>
>  #include <linux/pid_namespace.h>
> +#include <linux/proc_fs.h>
>  #include <linux/syscalls.h>
>  #include <linux/err.h>
>  #include <linux/acct.h>
> @@ -109,6 +110,8 @@ static void destroy_pid_namespace(struct
>  {
>  	int i;
>  
> +	pid_ns_release_proc(ns);
> +
>  	for (i = 0; i < PIDMAP_ENTRIES; i++)
>  		kfree(ns->pidmap[i].page);
>  	kmem_cache_free(pid_ns_cachep, ns);
> --- 34-rc1/fs/proc/base.c~PID_NS	2010-06-18 17:48:56.000000000 +0200
> +++ 34-rc1/fs/proc/base.c	2010-06-18 17:49:45.000000000 +0200
> @@ -2720,10 +2720,6 @@ void proc_flush_task(struct task_struct 
>  		proc_flush_task_mnt(upid->ns->proc_mnt, upid->nr,
>  					tgid->numbers[i].nr);
>  	}
> -
> -	upid = &pid->numbers[pid->level];
> -	if (upid->nr == 1)
> -		pid_ns_release_proc(upid->ns);
>  }
>  
>  static struct dentry *proc_pid_instantiate(struct inode *dir,
> --- 34-rc1/fs/proc/root.c~PID_NS	2010-06-18 17:48:56.000000000 +0200
> +++ 34-rc1/fs/proc/root.c	2010-06-18 17:56:57.000000000 +0200
> @@ -31,7 +31,7 @@ static int proc_set_super(struct super_b
>  	struct pid_namespace *ns;
>  
>  	ns = (struct pid_namespace *)data;
> -	sb->s_fs_info = get_pid_ns(ns);
> +	sb->s_fs_info = ns;
>  	return set_anon_super(sb, NULL);
>  }
>  
> @@ -74,7 +74,7 @@ static int proc_get_sb(struct file_syste
>  		ei = PROC_I(sb->s_root->d_inode);
>  		if (!ei->pid) {
>  			rcu_read_lock();
> -			ei->pid = get_pid(find_pid_ns(1, ns));
> +			ei->pid = find_pid_ns(1, ns);

I don't think that this is correct. IIUC, proc_delete_inode() calls put_pid() on
ei->pid. So either a special case is added in proc_delete_inode(), or we try to
live with get_pid() here. I'm actually not sure that we can pretend that this
pid remains valid if we don't get_pid() here.

Thanks,

Louis

>  			rcu_read_unlock();
>  		}
>  
> @@ -92,7 +92,6 @@ static void proc_kill_sb(struct super_bl
>  
>  	ns = (struct pid_namespace *)sb->s_fs_info;
>  	kill_anon_super(sb);
> -	put_pid_ns(ns);
>  }
>  
>  static struct file_system_type proc_fs_type = {
> @@ -216,7 +215,18 @@ int pid_ns_prepare_proc(struct pid_names
>  	return 0;
>  }
>  
> -void pid_ns_release_proc(struct pid_namespace *ns)
> +static void proc_mntput(struct work_struct *work)
>  {
> +	struct pid_namespace *ns = container_of(work, struct pid_namespace, proc_put);
> +
> +	PROC_I(ns->proc_mnt->mnt_sb->s_root->d_inode)->pid = NULL;
>  	mntput(ns->proc_mnt);
>  }
> +
> +void pid_ns_release_proc(struct pid_namespace *ns)
> +{
> +	if (ns->proc_mnt) {
> +		INIT_WORK(&ns->proc_put, proc_mntput);
> +		schedule_work(&ns->proc_put);
> +	}
> +}
> 
> 

-- 
Dr Louis Rilling			Kerlabs
Skype: louis.rilling			Batiment Germanium
Phone: (+33|0) 6 80 89 08 23		80 avenue des Buttes de Coesmes
http://www.kerlabs.com/			35700 Rennes