Re: [PATCH v4 2/3] binder: use cred instead of task for getsecid

[Dan Carpenter]

On Mon, Oct 11, 2021 at 02:59:13PM -0700, Casey Schaufler wrote:
> On 10/11/2021 2:33 PM, Paul Moore wrote:
> > On Wed, Oct 6, 2021 at 8:46 PM Todd Kjos <tkjos@xxxxxxxxxx> wrote:
> >> Use the 'struct cred' saved at binder_open() to lookup
> >> the security ID via security_cred_getsecid(). This
> >> ensures that the security context that opened binder
> >> is the one used to generate the secctx.
> >>
> >> Fixes: ec74136ded79 ("binder: create node flag to request sender's
> >> security context")
> >> Signed-off-by: Todd Kjos <tkjos@xxxxxxxxxx>
> >> Suggested-by: Stephen Smalley <stephen.smalley.work@xxxxxxxxx>
> >> Reported-by: kernel test robot <lkp@xxxxxxxxx>
> >> Cc: stable@xxxxxxxxxxxxxxx # 5.4+
> >> ---
> >> v3: added this patch to series
> >> v4: fix build-break for !CONFIG_SECURITY
> >>
> >>  drivers/android/binder.c | 11 +----------
> >>  include/linux/security.h |  4 ++++
> >>  2 files changed, 5 insertions(+), 10 deletions(-)
> >>
> >> diff --git a/drivers/android/binder.c b/drivers/android/binder.c
> >> index ca599ebdea4a..989afd0804ca 100644
> >> --- a/drivers/android/binder.c
> >> +++ b/drivers/android/binder.c
> >> @@ -2722,16 +2722,7 @@ static void binder_transaction(struct binder_proc *proc,
> >>                 u32 secid;
> >>                 size_t added_size;
> >>
> >> -               /*
> >> -                * Arguably this should be the task's subjective LSM secid but
> >> -                * we can't reliably access the subjective creds of a task
> >> -                * other than our own so we must use the objective creds, which
> >> -                * are safe to access.  The downside is that if a task is
> >> -                * temporarily overriding it's creds it will not be reflected
> >> -                * here; however, it isn't clear that binder would handle that
> >> -                * case well anyway.
> >> -                */
> >> -               security_task_getsecid_obj(proc->tsk, &secid);
> >> +               security_cred_getsecid(proc->cred, &secid);
> >>                 ret = security_secid_to_secctx(secid, &secctx, &secctx_sz);
> >>                 if (ret) {
> >>                         return_error = BR_FAILED_REPLY;
> >> diff --git a/include/linux/security.h b/include/linux/security.h
> >> index 6344d3362df7..f02cc0211b10 100644
> >> --- a/include/linux/security.h
> >> +++ b/include/linux/security.h
> >> @@ -1041,6 +1041,10 @@ static inline void security_transfer_creds(struct cred *new,
> >>  {
> >>  }
> >>
> >> +static inline void security_cred_getsecid(const struct cred *c, u32 *secid)
> >> +{
> >> +}
> > Since security_cred_getsecid() doesn't return an error code we should
> > probably set the secid to 0 in this case, for example:
> >
> >   static inline void security_cred_getsecid(...)
> >   {
> >     *secid = 0;
> >   }
> 
> If CONFIG_SECURITY is unset there shouldn't be any case where
> the secid value is ever used for anything. Are you suggesting that
> it be set out of an abundance of caution?

The security_secid_to_secctx() function is probably inlined so probably
KMSan will not warn about this.  But Smatch will warn about passing
unitialized variables.  You probably wouldn't recieve and email about
it, and I would just add an exception that security_cred_getsecid()
should be ignored.

regards,
dan carpenter