Subject: [patch] lockdep core: improve the lock-chain-hash

Subject: [patch] lockdep core: improve the lock-chain-hash

[Ingo Molnar]

From: Ingo Molnar <mingo@elte.hu>

With CONFIG_DEBUG_LOCK_ALLOC turned off i was getting sporadic
failures in the locking self-test:

------------>
| Locking API testsuite:
----------------------------------------------------------------------------
                                 | spin |wlock |rlock |mutex | wsem | rsem |
  --------------------------------------------------------------------------
                     A-A deadlock:  ok  |  ok  |  ok  |  ok  |  ok  |  ok  |
                 A-B-B-A deadlock:  ok  |  ok  |  ok  |  ok  |  ok  |  ok  |
             A-B-B-C-C-A deadlock:  ok  |  ok  |  ok  |  ok  |  ok  |  ok  |
             A-B-C-A-B-C deadlock:  ok  |  ok  |  ok  |  ok  |  ok  |  ok  |
         A-B-B-C-C-D-D-A deadlock:  ok  |FAILED|  ok  |  ok  |  ok  |  ok  |
         A-B-C-D-B-D-D-A deadlock:  ok  |  ok  |  ok  |  ok  |  ok  |  ok  |
         A-B-C-D-B-C-D-A deadlock:  ok  |  ok  |  ok  |  ok  |  ok  |FAILED|

after much debugging it turned out to be caused by accidental chain-hash
key collisions. The current hash is:

 #define iterate_chain_key(key1, key2) \
	(((key1) << MAX_LOCKDEP_KEYS_BITS/2) ^ \
	((key1) >> (64-MAX_LOCKDEP_KEYS_BITS/2)) ^ \
 	(key2))

where MAX_LOCKDEP_KEYS_BITS is 11. This hash is pretty good as it will
shift by 5 bits in every iteration, where every new ID 'mixed' into the
hash would have up to 11 bits. But because there was a 6 bits overlap
between subsequent IDs and their high bits tended to be similar, there
was a chance for accidental chain-hash collision for a low number of
locks held.

the solution is to shift by 11 bits:

 #define iterate_chain_key(key1, key2) \
	(((key1) << MAX_LOCKDEP_KEYS_BITS) ^ \
	((key1) >> (64-MAX_LOCKDEP_KEYS_BITS)) ^ \
 	(key2))

This keeps the hash perfect up to 5 locks held, but even above that the
hash is still good because 11 bits is a relative prime to the total 64
bits, so a complete match will only occur after 64 held locks (which
doesnt happen in Linux). Even after 5 locks held, entropy of the 5 IDs
mixed into the hash is already good enough so that overlap doesnt
generate a colliding hash ID.

which this change the false positives went away.

Signed-off-by: Ingo Molnar <mingo@elte.hu>

---
 kernel/lockdep.c |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

Index: linux/kernel/lockdep.c
===================================================================
--- linux.orig/kernel/lockdep.c
+++ linux/kernel/lockdep.c
@@ -121,8 +121,8 @@ static struct list_head chainhash_table[
  * unique.
  */
 #define iterate_chain_key(key1, key2) \
-	(((key1) << MAX_LOCKDEP_KEYS_BITS/2) ^ \
-	((key1) >> (64-MAX_LOCKDEP_KEYS_BITS/2)) ^ \
+	(((key1) << MAX_LOCKDEP_KEYS_BITS) ^ \
+	((key1) >> (64-MAX_LOCKDEP_KEYS_BITS)) ^ \
 	(key2))
 
 void lockdep_off(void)