[PATCH] Add a refcount check in dput()

[PATCH] Add a refcount check in dput()

[Paul Menage]

Add a BUG_ON() to check for passing an unreferenced dentry to dput().

This is analogous to the similar check in dget(), and will make
reference-counting bugs in filesystems more immediately obvious. (I
just spent a while debugging an oops that turned out to be due to
broken fs reference counting.)

Signed-off-by: Paul Menage <menage@google.com>

---
 fs/dcache.c |    1 +
 1 file changed, 1 insertion(+)

Index: container-2.6.23-rc3-mm1/fs/dcache.c
===================================================================
--- container-2.6.23-rc3-mm1.orig/fs/dcache.c
+++ container-2.6.23-rc3-mm1/fs/dcache.c
@@ -176,6 +176,7 @@ void dput(struct dentry *dentry)
 	if (!dentry)
 		return;
 
+	BUG_ON(!atomic_read(&dentry->d_count));
 repeat:
 	if (atomic_read(&dentry->d_count) == 1)
 		might_sleep();

Re: [PATCH] Add a refcount check in dput()

[Andrew Morton]

On Mon, 10 Sep 2007 15:13:19 -0700 (PDT) menage@google.com (Paul Menage) wrote:

> Add a BUG_ON() to check for passing an unreferenced dentry to dput().
> 
> This is analogous to the similar check in dget(), and will make
> reference-counting bugs in filesystems more immediately obvious. (I
> just spent a while debugging an oops that turned out to be due to
> broken fs reference counting.)
> 
> Signed-off-by: Paul Menage <menage@google.com>
> 
> ---
>  fs/dcache.c |    1 +
>  1 file changed, 1 insertion(+)
> 
> Index: container-2.6.23-rc3-mm1/fs/dcache.c
> ===================================================================
> --- container-2.6.23-rc3-mm1.orig/fs/dcache.c
> +++ container-2.6.23-rc3-mm1/fs/dcache.c
> @@ -176,6 +176,7 @@ void dput(struct dentry *dentry)
>  	if (!dentry)
>  		return;
>  
> +	BUG_ON(!atomic_read(&dentry->d_count));
>  repeat:
>  	if (atomic_read(&dentry->d_count) == 1)
>  		might_sleep();

eek, much too aggressive.

I added this:

--- a/fs/dcache.c~add-a-refcount-check-in-dput
+++ a/fs/dcache.c
@@ -176,6 +176,7 @@ void dput(struct dentry *dentry)
 	if (!dentry)
 		return;
 
+	WARN_ON_ONCE(!atomic_read(&dentry->d_count));
 repeat:
 	if (atomic_read(&dentry->d_count) == 1)
 		might_sleep();
_

to the -mm-only-debugging-patches section of the -mm tree.

Re: [PATCH] Add a refcount check in dput()

[Paul Menage]

On 9/15/07, Andrew Morton <akpm@linux-foundation.org> wrote:
> >
> > +     BUG_ON(!atomic_read(&dentry->d_count));
> >  repeat:
> >       if (atomic_read(&dentry->d_count) == 1)
> >               might_sleep();
>
> eek, much too aggressive.

How about the equivalent BUG_ON() in dget()? I figure that they ought
to both be of the same strictness.

Paul

Re: [PATCH] Add a refcount check in dput()

[Andrew Morton]

On Mon, 17 Sep 2007 11:21:35 -0700
"Paul Menage" <menage@google.com> wrote:

> On 9/15/07, Andrew Morton <akpm@linux-foundation.org> wrote:
> > >
> > > +     BUG_ON(!atomic_read(&dentry->d_count));
> > >  repeat:
> > >       if (atomic_read(&dentry->d_count) == 1)
> > >               might_sleep();
> >
> > eek, much too aggressive.
> 
> How about the equivalent BUG_ON() in dget()? I figure that they ought
> to both be of the same strictness.

The one in dget() is known not to trigger - it's been there for a long time.

My problem with new BUG_ON's is that they get added thinking "no way will
this trigger" and lo, they do trigger and lots of people get their testing
disrupted for a whole release.



Long-standing checks like the one in dget() should perhaps be removed now
that we know they don't trigger - move them under some CONFIG_DEBUG_foo
option.