lockdep: Add lock_set_cmp_fn() annotation

This implements a new interface to lockdep, lock_set_cmp_fn(), for
defining a custom ordering when taking multiple locks of the same
class.

This is an alternative to subclasses, but can not fully replace them
since subclasses allow lock hierarchies with other clasees
inter-twined, while this relies on pure class nesting.

Specifically, if A is our nesting class then:

  A/0 <- B <- A/1

Would be a valid lock order with subclasses (each subclass really is a
full class from the validation PoV) but not with this annotation,
which requires all nesting to be consecutive.

Example output:

| ============================================
| WARNING: possible recursive locking detected
| 6.2.0-rc8-00003-g7d81e591ca6a-dirty #15 Not tainted
| --------------------------------------------
| kworker/14:3/938 is trying to acquire lock:
| ffff8880143218c8 (&b->lock l=0 0:2803368){++++}-{3:3}, at: bch_btree_node_get.part.0+0x81/0x2b0
|
| but task is already holding lock:
| ffff8880143de8c8 (&b->lock l=1 1048575:9223372036854775807){++++}-{3:3}, at: __bch_btree_map_nodes+0xea/0x1e0
| and the lock comparison function returns 1:
|
| other info that might help us debug this:
|  Possible unsafe locking scenario:
|
|        CPU0
|        ----
|   lock(&b->lock l=1 1048575:9223372036854775807);
|   lock(&b->lock l=0 0:2803368);
|
|  *** DEADLOCK ***
|
|  May be due to missing lock nesting notation
|
| 3 locks held by kworker/14:3/938:
|  #0: ffff888005ea9d38 ((wq_completion)bcache){+.+.}-{0:0}, at: process_one_work+0x1ec/0x530
|  #1: ffff8880098c3e70 ((work_completion)(&cl->work)#3){+.+.}-{0:0}, at: process_one_work+0x1ec/0x530
|  #2: ffff8880143de8c8 (&b->lock l=1 1048575:9223372036854775807){++++}-{3:3}, at: __bch_btree_map_nodes+0xea/0x1e0

[peterz: extended changelog]
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20230509195847.1745548-1-kent.overstreet@linux.dev

include/linux/lockdep.h

@@ -434,6 +434,14 @@ extern int lockdep_is_held(const void *);
 
 #endif /* !LOCKDEP */
 
+#ifdef CONFIG_PROVE_LOCKING
+void lockdep_set_lock_cmp_fn(struct lockdep_map *, lock_cmp_fn, lock_print_fn);
+
+#define lock_set_cmp_fn(lock, ...)	lockdep_set_lock_cmp_fn(&(lock)->dep_map, __VA_ARGS__)
+#else
+#define lock_set_cmp_fn(lock, ...)	do { } while (0)
+#endif
+
 enum xhlock_context_t {
 	XHLOCK_HARD,
 	XHLOCK_SOFT,

include/linux/lockdep_types.h

@@ -84,6 +84,11 @@ struct lock_trace;
 
 #define LOCKSTAT_POINTS		4
 
+struct lockdep_map;
+typedef int (*lock_cmp_fn)(const struct lockdep_map *a,
+			   const struct lockdep_map *b);
+typedef void (*lock_print_fn)(const struct lockdep_map *map);
+
 /*
  * The lock-class itself. The order of the structure members matters.
  * reinit_class() zeroes the key member and all subsequent members.
@@ -109,6 +114,9 @@ struct lock_class {
 	struct list_head		locks_after, locks_before;
 
 	const struct lockdep_subclass_key *key;
+	lock_cmp_fn			cmp_fn;
+	lock_print_fn			print_fn;
+
 	unsigned int			subclass;
 	unsigned int			dep_gen_id;