[RFC PATCH 2/2] selinux: fall back to ref-walk upon LSM_AUDIT_DATA_DENTRY too

[Stephen Smalley]

On 11/21/19 7:30 PM, Paul Moore wrote:
> On Thu, Nov 21, 2019 at 7:12 PM Paul Moore <paul at paul-moore.com> wrote:
>> On Thu, Nov 21, 2019 at 9:52 AM Stephen Smalley <sds at tycho.nsa.gov> wrote:
>>> commit bda0be7ad994 ("security: make inode_follow_link RCU-walk aware")
>>> passed down the rcu flag to the SELinux AVC, but failed to adjust the
>>> test in slow_avc_audit() to also return -ECHILD on LSM_AUDIT_DATA_DENTRY.
>>> Previously, we only returned -ECHILD if generating an audit record with
>>> LSM_AUDIT_DATA_INODE since this was only relevant from inode_permission.
>>> Return -ECHILD on either LSM_AUDIT_DATA_INODE or LSM_AUDIT_DATA_DENTRY.
>>> LSM_AUDIT_DATA_INODE only requires this handling due to the fact
>>> that dump_common_audit_data() calls d_find_alias() and collects the
>>> dname from the result if any.
>>> Other cases that might require similar treatment in the future are
>>> LSM_AUDIT_DATA_PATH and LSM_AUDIT_DATA_FILE if any hook that takes
>>> a path or file is called under RCU-walk.
>>>
>>> Fixes: bda0be7ad994 ("security: make inode_follow_link RCU-walk aware")
>>> Signed-off-by: Stephen Smalley <sds at tycho.nsa.gov>
>>> ---
>>>   security/selinux/avc.c | 3 ++-
>>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/security/selinux/avc.c b/security/selinux/avc.c
>>> index 74c43ebe34bb..f1fa1072230c 100644
>>> --- a/security/selinux/avc.c
>>> +++ b/security/selinux/avc.c
>>> @@ -779,7 +779,8 @@ noinline int slow_avc_audit(struct selinux_state *state,
>>>           * during retry. However this is logically just as if the operation
>>>           * happened a little later.
>>>           */
>>> -       if ((a->type == LSM_AUDIT_DATA_INODE) &&
>>> +       if ((a->type == LSM_AUDIT_DATA_INODE ||
>>> +            a->type == LSM_AUDIT_DATA_DENTRY) &&
>>>              (flags & MAY_NOT_BLOCK))
>>>                  return -ECHILD;
> 
> With LSM_AUDIT_DATA_INODE we eventually end up calling d_find_alias()
> in dump_common_audit_data() which could block, which is bad, that I
> understand.  However, looking at LSM_AUDIT_DATA_DENTRY I'm less clear
> on why that is bad?  It makes a few audit_log*() calls and one call to
> d_backing_inode() which is non-blocking and trivial.
> 
> What am I missing?

For those who haven't, you may wish to also read the earlier thread here 
that led to this one:
https://lore.kernel.org/selinux/20191119184057.14961-1-will@kernel.org/T/#t

AFAIK, neither the LSM_AUDIT_DATA_INODE nor the LSM_AUDIT_DATA_DENTRY 
case truly block (d_find_alias does not block AFAICT, nor should 
audit_log* as long as we use audit_log_start with GFP_ATOMIC or 
GFP_NOWAIT). My impression from the comment in slow_avc_audit() is that 
the issue is not really about blocking but rather about the inability to 
safely dereference the dentry->d_name during RCU walk, which is 
something that can occur under LSM_AUDIT_DATA_INODE or _DENTRY (or _PATH 
or _FILE, but neither of the latter two are currently used from the two 
hooks that are called during RCU walk, inode_permission and 
inode_follow_link).  Originally _PATH, _DENTRY, and _INODE were all 
under a single _FS type and the original test in slow_avc_audit() was 
against LSM_AUDIT_DATA_FS before the split.

> 
>> Added the LSM list as I'm beginning to wonder if we should push this
>> logic down into common_lsm_audit(), this problem around blocking
>> shouldn't be SELinux specific.

That would require passing down the MAY_NOT_BLOCK flag or a rcu bool to 
common_lsm_audit() just so that it could immediately return if that is 
set and a->type is _INODE or _DENTRY (or _PATH or _FILE).  And the 
individual security module still needs to have its own handling of 
MAY_NOT_BLOCK/rcu for its own processing, so it won't free the security 
module authors from thinking about it.  This is only relevant for 
modules implementing the inode_permission and/or inode_follow_link 
hooks, so it only currently affects SELinux and Smack, and Smack only 
presently implements inode_permission and always returns -ECHILD if 
MAY_NOT_BLOCK (aside from a couple trivial cases), so it will never 
reach common_lsm_audit() in that case.


>>
>> For the LSM folks just joining, the full patchset can be found here:
>> * https://lore.kernel.org/selinux/20191121145245.8637-1-sds@tycho.nsa.gov/T/#t