On Tue, Jul 23, 2024 at 3:04 AM Xu Kuohai <xukuohai@xxxxxxxxxxxxxxx> wrote:
> On 7/23/2024 5:35 AM, Paul Moore wrote:
> > On Sat, Jul 20, 2024 at 5:31 AM Xu Kuohai <xukuohai@xxxxxxxxxxxxxxx> wrote:
> >> On 7/19/2024 10:08 AM, Paul Moore wrote:
> >>> On Jul 11, 2024 Xu Kuohai <xukuohai@xxxxxxxxxxxxxxx> wrote:
> >>>>
> >>>> To be consistent with most LSM hooks, convert the return value of
> >>>> hook key_getsecurity to 0 or a negative error code.
> >>>>
> >>>> Before:
> >>>> - Hook key_getsecurity returns length of value on success or a
> >>>> negative error code on failure.
> >>>>
> >>>> After:
> >>>> - Hook key_getsecurity returns 0 on success or a negative error
> >>>> code on failure. An output parameter @len is introduced to hold
> >>>> the length of value on success.
> >>>>
> >>>> Signed-off-by: Xu Kuohai <xukuohai@xxxxxxxxxx>
> >>>> ---
> >>>> include/linux/lsm_hook_defs.h | 3 ++-
> >>>> include/linux/security.h | 6 ++++--
> >>>> security/keys/keyctl.c | 11 ++++++++---
> >>>> security/security.c | 26 +++++++++++++++++++++-----
> >>>> security/selinux/hooks.c | 11 +++++------
> >>>> security/smack/smack_lsm.c | 21 +++++++++++----------
> >>>> 6 files changed, 51 insertions(+), 27 deletions(-)
> >
> > ...
> >
> >>>> diff --git a/security/security.c b/security/security.c
> >>>> index 9dd2ae6cf763..2c161101074d 100644
> >>>> --- a/security/security.c
> >>>> +++ b/security/security.c
> >>>> @@ -5338,19 +5338,35 @@ int security_key_permission(key_ref_t key_ref, const struct cred *cred,
> >>>> * security_key_getsecurity() - Get the key's security label
> >>>> * @key: key
> >>>> * @buffer: security label buffer
> >>>> + * @len: the length of @buffer (including terminating NULL) on success
> >>>> *
> >>>> * Get a textual representation of the security context attached to a key for
> >>>> * the purposes of honouring KEYCTL_GETSECURITY. This function allocates the
> >>>> * storage for the NUL-terminated string and the caller should free it.
> >>>> *
> >>>> - * Return: Returns the length of @buffer (including terminating NUL) or -ve if
> >>>> - * an error occurs. May also return 0 (and a NULL buffer pointer) if
> >>>> - * there is no security label assigned to the key.
> >>>> + * Return: Returns 0 on success or -ve if an error occurs. May also return 0
> >>>> + * (and a NULL buffer pointer) if there is no security label assigned
> >>>> + * to the key.
> >>>> */
> >>>> -int security_key_getsecurity(struct key *key, char **buffer)
> >>>> +int security_key_getsecurity(struct key *key, char **buffer, size_t *len)
> >>>> {
> >>>> + int rc;
> >>>> + size_t n = 0;
> >>>> + struct security_hook_list *hp;
> >>>> +
> >>>> *buffer = NULL;
> >>>> - return call_int_hook(key_getsecurity, key, buffer);
> >>>> +
> >>>> + hlist_for_each_entry(hp, &security_hook_heads.key_getsecurity, list) {
> >>>> + rc = hp->hook.key_getsecurity(key, buffer, &n);
> >>>> + if (rc < 0)
> >>>> + return rc;
> >>>> + if (n)
> >>>> + break;
> >>>> + }
> >>>> +
> >>>> + *len = n;
> >>>> +
> >>>> + return 0;
> >>>> }
> >>>
> >>> Help me understand why we can't continue to use the call_int_hook()
> >>> macro here?
> >>>
> >>
> >> Before this patch, the hook may return +ve, 0, or -ve, and call_int_hook
> >> breaks the loop when the hook return value is not 0.
> >>
> >> After this patch, the +ve is stored in @n, so @n and return value should
> >> both be checked to determine whether to break the loop. This is not
> >> feasible with call_int_hook.
> >
> > Yes, gotcha. I was focused on the error condition and wasn't thinking
> > about the length getting zero'd out by a trailing callback.
> > Unfortunately, we *really* want to stick with the
> > call_{int,void}_hook() macros so I think we either need to find a way
> > to work within that constraint for existing macro callers, or we have
> > to leave this hook as-is for the moment.
> >
>
> Let's leave it as is. So we ultimately have four hooks that can be
> converted, two of which require adding additional output parameter to
> hold odd return values. These output parameters require extra work
> on the BPF verifier, and it doesn't seem worthwhile just for two
> hooks. So I prefer to keep only the two patches that handle
> conversion without adding output parameters (patch 1 and 5).
Fair enough. Thanks for working on this, between the changes to the
LSM framework and the BPF verifier, I think this is still a nice
improvement.
--
paul-moore.com
