Re: [PATCH 21/26] smb: avoid pointless cred reference count bump

[Christian Brauner]

On Sun, Nov 24, 2024 at 06:37:43PM +0000, Al Viro wrote:
> On Sun, Nov 24, 2024 at 02:44:07PM +0100, Christian Brauner wrote:
> > No need for the extra reference count bump.
> > 
> > Signed-off-by: Christian Brauner <brauner@xxxxxxxxxx>
> > ---
> >  fs/smb/server/smb_common.c | 4 +---
> >  1 file changed, 1 insertion(+), 3 deletions(-)
> > 
> > diff --git a/fs/smb/server/smb_common.c b/fs/smb/server/smb_common.c
> > index f1d770a214c8b2c7d7dd4083ef57c7130bbce52c..a3f96804f84f03c22376769dffdf60cd66f5e3d2 100644
> > --- a/fs/smb/server/smb_common.c
> > +++ b/fs/smb/server/smb_common.c
> > @@ -780,7 +780,7 @@ int __ksmbd_override_fsids(struct ksmbd_work *work,
> >  		cred->cap_effective = cap_drop_fs_set(cred->cap_effective);
> >  
> >  	WARN_ON(work->saved_cred);
> > -	work->saved_cred = override_creds(get_new_cred(cred));
> > +	work->saved_cred = override_creds(cred);
> >  	if (!work->saved_cred) {
> >  		abort_creds(cred);
> >  		return -EINVAL;
> 
> Won't that leave a dangling pointer?

Afaict, the whole check doesn't make sense because I don't see how
override_creds() could be called on a task with current->cred == NULL.
There's no way to opt out of having current->cred set.