selinux: reduce path walk overhead

selinux: reduce path walk overhead

Reduce the SELinux performance overhead during path walks through the
use of a per-task directory access cache and some minor code
optimizations.  The directory access cache is per-task because it allows
for a lockless cache while also fitting well with a common application
pattern of heavily accessing a relatively small number of SELinux
directory labels.  The cache is inherited by child processes when the
child runs with the same SELinux domain as the parent, and invalidated
on changes to the task's SELinux domain or the loaded SELinux policy.
A cache of four entries was chosen based on testing with the Fedora
"targeted" policy, a SELinux Reference Policy variant, and
'make allmodconfig' on Linux v6.14.

Code optimizations include better use of inline functions to reduce
function calls in the common case, especially in the inode revalidation
code paths, and elimination of redundant checks between the LSM and
SELinux layers.

As mentioned briefly above, aside from general use and regression
testing with the selinux-testsuite, performance was measured using
'make allmodconfig' with Linux v6.14 as a base reference.  As expected,
there were variations from one test run to another, but the measurements
below are a good representation of the test results seen on my test
system.

 * Linux v6.14
   REF
     1.26%  [k] __d_lookup_rcu
   SELINUX (1.31%)
     0.58%  [k] selinux_inode_permission
     0.29%  [k] avc_lookup
     0.25%  [k] avc_has_perm_noaudit
     0.19%  [k] __inode_security_revalidate

 * Linux v6.14 + patch
   REF
     1.41%  [k] __d_lookup_rcu
   SELINUX (0.89%)
     0.65%  [k] selinux_inode_permission
     0.15%  [k] avc_lookup
     0.05%  [k] avc_has_perm_noaudit
     0.04%  [k] avc_policy_seqno
     X.XX%  [k] __inode_security_revalidate (now inline)

In both cases the __d_lookup_rcu() function was used as a reference
point to establish a context for the SELinux related functions.  On a
unpatched Linux v6.14 system we see the time spent in the combined
SELinux functions exceeded that of __d_lookup_rcu(), 1.31% compared to
1.26%.  However, with this patch applied the time spent in the combined
SELinux functions dropped to roughly 65% of the time spent in
__d_lookup_rcu(), 0.89% compared to 1.41%.  Aside from the significant
decrease in time spent in the SELinux AVC, it appears that any additional
time spent searching and updating the cache is offset by other code
improvements, e.g. time spent in selinux_inode_permission() +
__inode_security_revalidate() + avc_policy_seqno() is less on the
patched kernel than the unpatched kernel.

It is worth noting that in this patch the use of the per-task cache is
limited to the security_inode_permission() LSM callback,
selinux_inode_permission(), but future work could expand the cache into
inode_has_perm(), likely through consolidation of the two functions.
While this would likely have little to no impact on path walks, it
may benefit other operations.

Reviewed-by: Stephen Smalley <stephen.smalley.work@gmail.com>
Tested-by: Stephen Smalley <stephen.smalley.work@gmail.com>
Signed-off-by: Paul Moore <paul@paul-moore.com>