[PATCH] pidfs: detect refcount bugs

[PATCH] pidfs: detect refcount bugs

[Christian Brauner]

Now that we have pidfs_{get,register}_pid() that needs to be paired with
pidfs_put_pid() it's possible that someone pairs them with put_pid().
Thus freeing struct pid while it's still used by pidfs. Notice when that
happens. I'll also add a scheme to detect invalid uses of
pidfs_get_pid() and pidfs_put_pid() later.

Signed-off-by: Christian Brauner <brauner@kernel.org>
---
 kernel/pid.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/kernel/pid.c b/kernel/pid.c
index 26f1e136f017..8317bcbc7cf7 100644
--- a/kernel/pid.c
+++ b/kernel/pid.c
@@ -100,6 +100,7 @@ void put_pid(struct pid *pid)
 
 	ns = pid->numbers[pid->level].ns;
 	if (refcount_dec_and_test(&pid->count)) {
+		WARN_ON_ONCE(pid->stashed);
 		kmem_cache_free(ns->pid_cachep, pid);
 		put_pid_ns(ns);
 	}
-- 
2.47.2

Re: [PATCH] pidfs: detect refcount bugs

[Oleg Nesterov]

I am traveling until May 15, can't read the code, but FWIW this change
looks good to me.

Oleg.

On 05/06, Christian Brauner wrote:
>
> Now that we have pidfs_{get,register}_pid() that needs to be paired with
> pidfs_put_pid() it's possible that someone pairs them with put_pid().
> Thus freeing struct pid while it's still used by pidfs. Notice when that
> happens. I'll also add a scheme to detect invalid uses of
> pidfs_get_pid() and pidfs_put_pid() later.
> 
> Signed-off-by: Christian Brauner <brauner@kernel.org>
> ---
>  kernel/pid.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/kernel/pid.c b/kernel/pid.c
> index 26f1e136f017..8317bcbc7cf7 100644
> --- a/kernel/pid.c
> +++ b/kernel/pid.c
> @@ -100,6 +100,7 @@ void put_pid(struct pid *pid)
>  
>  	ns = pid->numbers[pid->level].ns;
>  	if (refcount_dec_and_test(&pid->count)) {
> +		WARN_ON_ONCE(pid->stashed);
>  		kmem_cache_free(ns->pid_cachep, pid);
>  		put_pid_ns(ns);
>  	}
> -- 
> 2.47.2
> 

Re: [PATCH] pidfs: detect refcount bugs

[Mateusz Guzik]

On Tue, May 06, 2025 at 01:55:54PM +0200, Christian Brauner wrote:
> Now that we have pidfs_{get,register}_pid() that needs to be paired with
> pidfs_put_pid() it's possible that someone pairs them with put_pid().
> Thus freeing struct pid while it's still used by pidfs. Notice when that
> happens. I'll also add a scheme to detect invalid uses of
> pidfs_get_pid() and pidfs_put_pid() later.
> 
> Signed-off-by: Christian Brauner <brauner@kernel.org>
> ---
>  kernel/pid.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/kernel/pid.c b/kernel/pid.c
> index 26f1e136f017..8317bcbc7cf7 100644
> --- a/kernel/pid.c
> +++ b/kernel/pid.c
> @@ -100,6 +100,7 @@ void put_pid(struct pid *pid)
>  
>  	ns = pid->numbers[pid->level].ns;
>  	if (refcount_dec_and_test(&pid->count)) {
> +		WARN_ON_ONCE(pid->stashed);
>  		kmem_cache_free(ns->pid_cachep, pid);
>  		put_pid_ns(ns);
>  	}
> -- 
> 2.47.2
> 

With the patch as proposed you are only catching the misuse if this is
the last ref though.

iow, the check should be hoisted above unrefing?

Re: [PATCH] pidfs: detect refcount bugs

[Christian Brauner]

On Tue, May 06, 2025 at 04:43:56PM +0200, Mateusz Guzik wrote:
> On Tue, May 06, 2025 at 01:55:54PM +0200, Christian Brauner wrote:
> > Now that we have pidfs_{get,register}_pid() that needs to be paired with
> > pidfs_put_pid() it's possible that someone pairs them with put_pid().
> > Thus freeing struct pid while it's still used by pidfs. Notice when that
> > happens. I'll also add a scheme to detect invalid uses of
> > pidfs_get_pid() and pidfs_put_pid() later.
> > 
> > Signed-off-by: Christian Brauner <brauner@kernel.org>
> > ---
> >  kernel/pid.c | 1 +
> >  1 file changed, 1 insertion(+)
> > 
> > diff --git a/kernel/pid.c b/kernel/pid.c
> > index 26f1e136f017..8317bcbc7cf7 100644
> > --- a/kernel/pid.c
> > +++ b/kernel/pid.c
> > @@ -100,6 +100,7 @@ void put_pid(struct pid *pid)
> >  
> >  	ns = pid->numbers[pid->level].ns;
> >  	if (refcount_dec_and_test(&pid->count)) {
> > +		WARN_ON_ONCE(pid->stashed);
> >  		kmem_cache_free(ns->pid_cachep, pid);
> >  		put_pid_ns(ns);
> >  	}
> > -- 
> > 2.47.2
> > 
> 
> With the patch as proposed you are only catching the misuse if this is
> the last ref though.
> 
> iow, the check should be hoisted above unrefing?

No, not really. If there's more than one reference then pid->stashed can
be legimitately != NULL.