rcu: Document alternative RCU/reference-count algorithms
authorPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Sat, 27 Oct 2012 23:34:51 +0000 (16:34 -0700)
committerPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Thu, 8 Nov 2012 19:44:38 +0000 (11:44 -0800)
The approach for mixing RCU and reference counting listed in the RCU
documentation only describes one possible approach.  This approach can
result in failure on the read side, which is nice if you want fresh data,
but not so good if you want simple code.  This commit therefore adds
two additional approaches that feature unconditional reference-count
acquisition by RCU readers.  These approaches are very similar to that
used in the security code.

Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Documentation/RCU/rcuref.txt

index 4202ad0..141d531 100644 (file)
@@ -20,7 +20,7 @@ release_referenced()                  delete()
 {                                      {
     ...                                            write_lock(&list_lock);
     atomic_dec(&el->rc, relfunc)           ...
-    ...                                            delete_element
+    ...                                            remove_element
 }                                          write_unlock(&list_lock);
                                            ...
                                            if (atomic_dec_and_test(&el->rc))
@@ -52,7 +52,7 @@ release_referenced()                  delete()
 {                                      {
     ...                                            spin_lock(&list_lock);
     if (atomic_dec_and_test(&el->rc))       ...
-        call_rcu(&el->head, el_free);       delete_element
+        call_rcu(&el->head, el_free);       remove_element
     ...                                     spin_unlock(&list_lock);
 }                                          ...
                                            if (atomic_dec_and_test(&el->rc))
@@ -64,3 +64,60 @@ Sometimes, a reference to the element needs to be obtained in the
 update (write) stream.  In such cases, atomic_inc_not_zero() might be
 overkill, since we hold the update-side spinlock.  One might instead
 use atomic_inc() in such cases.
+
+It is not always convenient to deal with "FAIL" in the
+search_and_reference() code path.  In such cases, the
+atomic_dec_and_test() may be moved from delete() to el_free()
+as follows:
+
+1.                                     2.
+add()                                  search_and_reference()
+{                                      {
+    alloc_object                           rcu_read_lock();
+    ...                                            search_for_element
+    atomic_set(&el->rc, 1);                atomic_inc(&el->rc);
+    spin_lock(&list_lock);                 ...
+
+    add_element                                    rcu_read_unlock();
+    ...                                        }
+    spin_unlock(&list_lock);           4.
+}                                      delete()
+3.                                     {
+release_referenced()                       spin_lock(&list_lock);
+{                                          ...
+    ...                                            remove_element
+    if (atomic_dec_and_test(&el->rc))       spin_unlock(&list_lock);
+        kfree(el);                         ...
+    ...                                     call_rcu(&el->head, el_free);
+}                                          ...
+5.                                     }
+void el_free(struct rcu_head *rhp)
+{
+    release_referenced();
+}
+
+The key point is that the initial reference added by add() is not removed
+until after a grace period has elapsed following removal.  This means that
+search_and_reference() cannot find this element, which means that the value
+of el->rc cannot increase.  Thus, once it reaches zero, there are no
+readers that can or ever will be able to reference the element.  The
+element can therefore safely be freed.  This in turn guarantees that if
+any reader finds the element, that reader may safely acquire a reference
+without checking the value of the reference counter.
+
+In cases where delete() can sleep, synchronize_rcu() can be called from
+delete(), so that el_free() can be subsumed into delete as follows:
+
+4.
+delete()
+{
+    spin_lock(&list_lock);
+    ...
+    remove_element
+    spin_unlock(&list_lock);
+    ...
+    synchronize_rcu();
+    if (atomic_dec_and_test(&el->rc))
+       kfree(el);
+    ...
+}