kernel crash in mknod

[Christian Brauner]

On Thu, Mar 28, 2024 at 01:24:25PM +0200, Roberto Sassu wrote:
> On 3/28/2024 12:08 PM, Christian Brauner wrote:
> > On Thu, Mar 28, 2024 at 12:53:40PM +0200, Roberto Sassu wrote:
> > > On 3/26/2024 12:40 PM, Christian Brauner wrote:
> > > > > we can change the parameter of security_path_post_mknod() from
> > > > > dentry to inode?
> > > > 
> > > > If all current callers only operate on the inode then it seems the best
> > > > to only pass the inode. If there's some reason someone later needs a
> > > > dentry the hook can always be changed.
> > > 
> > > Ok, so the crash is likely caused by:
> > > 
> > > void security_path_post_mknod(struct mnt_idmap *idmap, struct dentry
> > > *dentry)
> > > {
> > >          if (unlikely(IS_PRIVATE(d_backing_inode(dentry))))
> > > 
> > > I guess we can also simply check if there is an inode attached to the
> > > dentry, to minimize the changes. I can do both.
> > > 
> > > More technical question, do I need to do extra checks on the dentry before
> > > calling security_path_post_mknod()?
> > 
> > Why do you need the dentry? The two users I see are ima in [1] and evm in [2].
> > Both of them don't care about the dentry. They only care about the
> > inode. So why is that hook not just:
> 
> Sure, I can definitely do that. Seems an easier fix to do an extra check in
> security_path_post_mknod(), rather than changing the parameter everywhere.

You only have two callers and the generic implementation.

> 
> Next time, when we introduce new LSM hooks we can try to introduce more
> specific parameters.
> 
> Also, consider that the pre hook security_path_mknod() has the dentry as
> parameter. For symmetry, we could keep it in the post hook.

I think that's not that important.

> 
> What I was also asking is if I can still call d_backing_inode() on the
> dentry without extra checks, and avoiding the IS_PRIVATE() check if the
> former returns NULL.