On Sun, Jun 04, 2017 at 01:24:13AM -0400, Matt Brown wrote:
> On 06/03/2017 02:33 AM, Al Viro wrote:
> > On Sat, Jun 03, 2017 at 01:53:51AM -0400, Matt Brown wrote:
> >
> > > +static int tpe_bprm_set_creds(struct linux_binprm *bprm)
> > > +{
> > > + struct file *file = bprm->file;
> > > + struct inode *inode = d_backing_inode(file->f_path.dentry->d_parent);
> > > + struct inode *file_inode = d_backing_inode(file->f_path.dentry);
> >
> > Bloody wonderful. Do tell, what *does* prevent a race with rename(2) here,
> > somehow making sure that your 'inode' won't get freed right under you?
> >
>
> Good catch. How does this look:
>
> spin_lock(&inode->i_lock);
> spin_lock(&file_inode->i_lock);
> if (global_nonroot(inode->i_uid) && !uid_eq(inode->i_uid, cred->uid))
> reason1 = "directory not owned by user";
> else if (inode->i_mode & 0002)
> reason1 = "file in world-writable directory";
> else if ((inode->i_mode & 0020) && global_nonroot_gid(inode->i_gid))
> reason1 = "file in group-writable directory";
> else if (file_inode->i_mode & 0002)
> reason1 = "file is world-writable";
> spin_unlock(&inode->i_lock);
> spin_unlock(&file_inode->i_lock);
>
> and likewise for other places in the code?
Er... You have a pointer to object that might get freed by a thread
running on another CPU. So you attempt to take a spinlock sitting
inside that object. How exactly is that supposed to help anything?
