[PATCH resend] vfs: Add ->statfs callback for pipefs

[PATCH resend] vfs: Add ->statfs callback for pipefs

[Cyrill Gorcunov]

From: Pavel Emelyanov <xemul@parallels.com>

This is done to make it possible to distinguish pipes
from fifos when opening one via /proc/<pid>/fd/ link.

Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
Reviewed-by: Tejun Heo <tj@kernel.org>
Acked-by: Serge Hallyn <serge.hallyn@canonical.com>
Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
---
 fs/pipe.c |    1 +
 1 file changed, 1 insertion(+)

Any objections to this one? It was a part of a patchset
but can be treated independently so I'm re-sending it alone.

Index: linux-2.6.git/fs/pipe.c
===================================================================
--- linux-2.6.git.orig/fs/pipe.c
+++ linux-2.6.git/fs/pipe.c
@@ -1254,6 +1254,7 @@ out:
 
 static const struct super_operations pipefs_ops = {
 	.destroy_inode = free_inode_nonrcu,
+	.statfs = simple_statfs,
 };
 
 /*

Re: [PATCH resend] vfs: Add ->statfs callback for pipefs

[Andrew Morton]

On Wed, 24 Aug 2011 12:09:24 +0400 Cyrill Gorcunov <gorcunov@gmail.com> wrote:

> From: Pavel Emelyanov <xemul@parallels.com>
> 
> This is done to make it possible to distinguish pipes
> from fifos when opening one via /proc/<pid>/fd/ link.
> 
> Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
> Reviewed-by: Tejun Heo <tj@kernel.org>
> Acked-by: Serge Hallyn <serge.hallyn@canonical.com>
> Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
> ---
>  fs/pipe.c |    1 +
>  1 file changed, 1 insertion(+)
> 
> Any objections to this one? It was a part of a patchset
> but can be treated independently so I'm re-sending it alone.
> 
> Index: linux-2.6.git/fs/pipe.c
> ===================================================================
> --- linux-2.6.git.orig/fs/pipe.c
> +++ linux-2.6.git/fs/pipe.c
> @@ -1254,6 +1254,7 @@ out:
>  
>  static const struct super_operations pipefs_ops = {
>  	.destroy_inode = free_inode_nonrcu,
> +	.statfs = simple_statfs,
>  };
>  
>  /*

OK, I give up - how does it work?

<goes off and reads the kernel>

So it appears that the statfs call would previously return -ENOSYS, but
with this change the statfs will succeed and userspace can then inspect
f_type.  Yes?  I will fix your changelog by adding a description along these
lines.

And I'll ask you to fix it further by telling us why we want to do
this?  The kernel has had this issue for a long time - why does it now
matter?

Re: [PATCH resend] vfs: Add ->statfs callback for pipefs

[Cyrill Gorcunov]

On Wed, Aug 24, 2011 at 01:23:28AM -0700, Andrew Morton wrote:
> On Wed, 24 Aug 2011 12:09:24 +0400 Cyrill Gorcunov <gorcunov@gmail.com> wrote:
> 
> > From: Pavel Emelyanov <xemul@parallels.com>
> > 
> > This is done to make it possible to distinguish pipes
> > from fifos when opening one via /proc/<pid>/fd/ link.
> > 
> > Signed-off-by: Pavel Emelyanov <xemul@parallels.com>
> > Reviewed-by: Tejun Heo <tj@kernel.org>
> > Acked-by: Serge Hallyn <serge.hallyn@canonical.com>
> > Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
> > ---
> >  fs/pipe.c |    1 +
> >  1 file changed, 1 insertion(+)
> > 
> > Any objections to this one? It was a part of a patchset
> > but can be treated independently so I'm re-sending it alone.
> > 
> > Index: linux-2.6.git/fs/pipe.c
> > ===================================================================
> > --- linux-2.6.git.orig/fs/pipe.c
> > +++ linux-2.6.git/fs/pipe.c
> > @@ -1254,6 +1254,7 @@ out:
> >  
> >  static const struct super_operations pipefs_ops = {
> >  	.destroy_inode = free_inode_nonrcu,
> > +	.statfs = simple_statfs,
> >  };
> >  
> >  /*
> 
> OK, I give up - how does it work?
> 
> <goes off and reads the kernel>
> 
> So it appears that the statfs call would previously return -ENOSYS, but
> with this change the statfs will succeed and userspace can then inspect
> f_type.  Yes?  I will fix your changelog by adding a description along these
> lines.

Yes, it allows to distinguish pipes from anything else. Initially this patch was
a part of early rfc on checkpoint/restore facility (mostly implemented in
userspace). So we need to distinguish somehow pipes when we dump data from
/proc/pid/fd. That's the main reason. BUT since this patch can be used without
any other c/r bits and can be treated separately I've sent it alone without
any mention of c/r at all.

> 
> And I'll ask you to fix it further by telling us why we want to do
> this?  The kernel has had this issue for a long time - why does it now
> matter?

	Cyrill

Re: [PATCH resend] vfs: Add ->statfs callback for pipefs

[Pavel Emelyanov]

> 
> OK, I give up - how does it work?
> 
> <goes off and reads the kernel>
> 
> So it appears that the statfs call would previously return -ENOSYS, but
> with this change the statfs will succeed and userspace can then inspect
> f_type.  Yes?  I will fix your changelog by adding a description along these
> lines.

Yes, the primary intention is to get the f_type.

> And I'll ask you to fix it further by telling us why we want to do
> this?  The kernel has had this issue for a long time - why does it now
> matter?

This is required by checkpoint-restart in the userspace to make it
possible to distinguish pipes from fifos. In more details:

When we dump information about task's open files we use the /proc/pid/fd
directoy's symlinks and the fact that opening any of them gives us exactly
the same dentry->inode pair as the original process has. Now if a task 
we're dumping has opened pipe and fifo we need to detect this and act
accordingly. Knowing that an fd with type S_ISFIFO resides on a pipefs
is the most precise way.

Thanks,
Pavel