security: smack: add a hash table to quicken smk_find_entry()

Accepted for the smack-next tree after changing the number of
slots from 128 to 16.

This patch adds a hash table to quicken searching of a smack label by its name.

Basically, the patch improves performance of SMACK initialization.  Parsing of
rules involves translation from a string to a smack_known (aka label) entity
which is done in smk_find_entry().

The current implementation of the function iterates over a global list of
smack_known resulting in O(N) complexity for smk_find_entry().  The total
complexity of SMACK initialization becomes O(rules * labels).  Therefore it
scales quadratically with a complexity of a system.

Applying the patch reduced the complexity of smk_find_entry() to O(1) as long
as number of label is in hundreds. If the number of labels is increased please
update SMACK_HASH_SLOTS constant defined in security/smack/smack.h. Introducing
the configuration of this constant with Kconfig or cmdline might be a good
idea.

The size of the hash table was adjusted experimentally.  The rule set used by
TIZEN contains circa 17K rules for 500 labels.  The table above contains
results of SMACK initialization using 'time smackctl apply' bash command.
The 'Ref' is a kernel without this patch applied. The consecutive values
refers to value of SMACK_HASH_SLOTS.  Every measurement was repeated three
times to reduce noise.

     |  Ref  |   1   |   2   |   4   |   8   |   16  |   32  |   64  |  128  |  256  |  512
--------------------------------------------------------------------------------------------
Run1 | 1.156 | 1.096 | 0.883 | 0.764 | 0.692 | 0.667 | 0.649 | 0.633 | 0.634 | 0.629 | 0.620
Run2 | 1.156 | 1.111 | 0.885 | 0.764 | 0.694 | 0.661 | 0.649 | 0.651 | 0.634 | 0.638 | 0.623
Run3 | 1.160 | 1.107 | 0.886 | 0.764 | 0.694 | 0.671 | 0.661 | 0.638 | 0.631 | 0.624 | 0.638
AVG  | 1.157 | 1.105 | 0.885 | 0.764 | 0.693 | 0.666 | 0.653 | 0.641 | 0.633 | 0.630 | 0.627

Surprisingly, a single hlist is slightly faster than a double-linked list.
The speed-up saturates near 64 slots.  Therefore I chose value 128 to provide
some margin if more labels were used.
It looks that IO becomes a new bottleneck.

Signed-off-by: Tomasz Stanislawski <t.stanislaws@samsung.com>

1 files changed, 26 insertions, 3 deletions

diff --git a/security/smack/smack_access.c b/security/smack/smack_access.c
index 6a0377f38620..b3b59b1e93d6 100644
--- a/security/smack/smack_access.c
+++ b/security/smack/smack_access.c
@@ -325,6 +325,25 @@ void smack_log(char *subject_label, char *object_label, int request,
 
 DEFINE_MUTEX(smack_known_lock);
 
+struct hlist_head smack_known_hash[SMACK_HASH_SLOTS];
+
+/**
+ * smk_insert_entry - insert a smack label into a hash map,
+ *
+ * this function must be called under smack_known_lock
+ */
+void smk_insert_entry(struct smack_known *skp)
+{
+	unsigned int hash;
+	struct hlist_head *head;
+
+	hash = full_name_hash(skp->smk_known, strlen(skp->smk_known));
+	head = &smack_known_hash[hash & (SMACK_HASH_SLOTS - 1)];
+
+	hlist_add_head_rcu(&skp->smk_hashed, head);
+	list_add_rcu(&skp->list, &smack_known_list);
+}
+
 /**
  * smk_find_entry - find a label on the list, return the list entry
  * @string: a text string that might be a Smack label
@@ -334,12 +353,16 @@ DEFINE_MUTEX(smack_known_lock);
  */
 struct smack_known *smk_find_entry(const char *string)
 {
+	unsigned int hash;
+	struct hlist_head *head;
 	struct smack_known *skp;
 
-	list_for_each_entry_rcu(skp, &smack_known_list, list) {
+	hash = full_name_hash(string, strlen(string));
+	head = &smack_known_hash[hash & (SMACK_HASH_SLOTS - 1)];
+
+	hlist_for_each_entry_rcu(skp, head, smk_hashed)
 		if (strcmp(skp->smk_known, string) == 0)
 			return skp;
-	}
 
 	return NULL;
 }
@@ -475,7 +498,7 @@ struct smack_known *smk_import_entry(const char *string, int len)
 		 * Make sure that the entry is actually
 		 * filled before putting it on the list.
 		 */
-		list_add_rcu(&skp->list, &smack_known_list);
+		smk_insert_entry(skp);
 		goto unlockout;
 	}
 	/*