rcu: add rcu_access_pointer and rcu_dereference_protected

This patch adds variants of rcu_dereference() that handle situations where the RCU-protected data structure cannot change, perhaps due to our holding the update-side lock, or where the RCU-protected pointer is only to be fetched, not dereferenced. These are needed due to some performance concerns with using rcu_dereference() where it is not required, aside from the need for lockdep/sparse checking.

The new rcu_access_pointer() primitive is for the case where the pointer is be fetch and not dereferenced. This primitive may be used without protection, RCU or otherwise, due to the fact that it uses ACCESS_ONCE().

The new rcu_dereference_protected() primitive is for the case where updates are prevented, for example, due to holding the update-side lock. This primitive does neither ACCESS_ONCE() nor smp_read_barrier_depends(), so can only be used when updates are somehow prevented.

Suggested-by David Howells <[email protected]>

Signed-off-by Paul E. McKenney <[email protected]>

This patch adds variants of rcu_dereference() that handle situations where the RCU-protected data structure cannot change, perhaps due to our holding the update-side lock, or where the RCU-protected pointer is only to be fetched, not dereferenced. These are needed due to some performance concerns with using rcu_dereference() where it is not required, aside from the need for lockdep/sparse checking.

The new rcu_access_pointer() primitive is for the case where the pointer is be fetch and not dereferenced. This primitive may be used without protection, RCU or otherwise, due to the fact that it uses ACCESS_ONCE().

The new rcu_dereference_protected() primitive is for the case where updates are prevented, for example, due to holding the update-side lock. This primitive does neither ACCESS_ONCE() nor smp_read_barrier_depends(), so can only be used when updates are somehow prevented.

Suggested-by David Howells <[email protected]>

Signed-off-by Paul E. McKenney <[email protected]>

 rcupdate.h |   38 ++++++++++++++++++++++++++++++++++++++
 1 file changed, 38 insertions(+)

include/linux/rcupdate.h +38 -0

index 872a98e..3f06b3d 100644
@@ -209,9 +209,47 @@ static inline int rcu_read_lock_sched_held(void)
 		rcu_dereference_raw(p); \
 	})
 
+/**
+ * rcu_access_pointer - fetch RCU pointer with no dereferencing
+ *
+ * Return the value of the specified RCU-protected pointer, but omit the
+ * smp_read_barrier_depends() and keep the ACCESS_ONCE().  This is useful
+ * when the value of this pointer is accessed, but the pointer is not
+ * dereferenced, for example, when testing an RCU-protected pointer against
+ * NULL.  This may also be used in cases where update-side locks prevent
+ * the value of the pointer from changing, but rcu_dereference_protected()
+ * is a lighter-weight primitive for this use case.
+ */
+#define rcu_access_pointer(p, c) \
+	({ \
+		if (debug_lockdep_rcu_enabled() && !(c)) \
+			lockdep_rcu_dereference(__FILE__, __LINE__); \
+		ACCESS_ONCE(p); \
+	})
+
+/**
+ * rcu_dereference_protected - fetch RCU pointer when updates prevented
+ *
+ * Return the value of the specified RCU-protected pointer, but omit
+ * both the smp_read_barrier_depends() and the ACCESS_ONCE().  This
+ * is useful in cases where update-side locks prevent the value of the
+ * pointer from changing.  Please note that this primitive does -not-
+ * prevent the compiler from repeating this reference or combining it
+ * with other references, so it should not be used without protection
+ * of appropriate locks.
+ */
+#define rcu_dereference_protected(p, c) \
+	({ \
+		if (debug_lockdep_rcu_enabled() && !(c)) \
+			lockdep_rcu_dereference(__FILE__, __LINE__); \
+		(p); \
+	})
+
 #else /* #ifdef CONFIG_PROVE_RCU */
 
 #define rcu_dereference_check(p, c)	rcu_dereference_raw(p)
+#define rcu_access_pointer(p, c)	ACCESS_ONCE(p)
+#define rcu_dereference_protected(p, c) (p)
 
 #endif /* #else #ifdef CONFIG_PROVE_RCU */
 

The new rcu_access_pointer() primitive is for the case where the pointer is be fetch and not dereferenced. This primitive may be used without protection, RCU or otherwise, due to the fact that it uses ACCESS_ONCE().
...
+#define rcu_access_pointer(p, c) \

Quoted 5 lines · Paul E. McKenney

The new rcu_access_pointer() primitive is for the case where the pointer is be fetch and not dereferenced. This primitive may be used without protection, RCU or otherwise, due to the fact that it uses ACCESS_ONCE().
...
+#define rcu_access_pointer(p, c) \

NAK. This shouldn't have the conditional parameter 'c'. Given that 'c' (by analogy to rcu_dereference_check()) is there to describe the conditions under which it's permitted to dereference the pointer, why is that relevant here? What is it you're proving?

In other cases, there will be a reference counter or a "not yet fully initialized" flag that can (and should) be tested.

Quoted 2 lines · Paul E. McKenney

In other cases, there will be a reference counter or a "not yet fully initialized" flag that can (and should) be tested.

Why would you be using rcu_access_pointer() there? Why wouldn't you be using rcu_dereference_protected()?

Also, one other thing: Should the default versions of these functions make some reference to 'c' to prevent compiler warnings? Should:

	#define rcu_dereference_check(p, c)	rcu_dereference_raw(p)

for example, be:

	#define rcu_dereference_check(p, c)	\
	({					\
		if (1 || !(c))			\
			rcu_dereference_raw(p);	\
	})

I'm not sure it's necessary, but it's possible to envisage a situation where someone calculates something specifically for use in 'c', which will cause an warning from the compiler if c isn't then checked.

Quoted 7 lines · David Howells

Quoted 2 lines · Paul E. McKenney

In other cases, there will be a reference counter or a "not yet fully initialized" flag that can (and should) be tested.

Why would you be using rcu_access_pointer() there? Why wouldn't you be using rcu_dereference_protected()?

Excellent question. I am writing up the documentation now, and will either (1) have a good use case or (2) remove the condition.

And all of the examples I could come up with that had c!=1 were contorted, even by my standards. So you were right, and I will drop the "c" on my next set of patches.

Quoted 3 lines · Paul E. McKenney

And all of the examples I could come up with that had c!=1 were contorted, even by my standards. So you were right, and I will drop the "c" on my next set of patches.

:-)

When it's done, I'll rebuild my keys and NFS patches on top of it.