Documentation/RCU/rcuref.txt

   1 Reference-count design for elements of lists/arrays protected by RCU.
   2
   3 Reference counting on elements of lists which are protected by traditional
   4 reader/writer spinlocks or semaphores are straightforward:
   5
   6 1.                              2.
   7 add()                           search_and_reference()
   8 {                               {
   9     alloc_object                    read_lock(&list_lock);
  10     ...                             search_for_element
  11     atomic_set(&el->rc, 1);         atomic_inc(&el->rc);
  12     write_lock(&list_lock);          ...
  13     add_element                     read_unlock(&list_lock);
  14     ...                             ...
  15     write_unlock(&list_lock);   }
  16 }
  17
  18 3.                                      4.
  19 release_referenced()                    delete()
  20 {                                       {
  21     ...                                     write_lock(&list_lock);
  22     atomic_dec(&el->rc, relfunc)            ...
  23     ...                                     remove_element
  24 }                                           write_unlock(&list_lock);
  25                                             ...
  26                                             if (atomic_dec_and_test(&el->rc))
  27                                                 kfree(el);
  28                                             ...
  29                                         }
  30
  31 If this list/array is made lock free using RCU as in changing the
  32 write_lock() in add() and delete() to spin_lock() and changing read_lock()
  33 in search_and_reference() to rcu_read_lock(), the atomic_inc() in
  34 search_and_reference() could potentially hold reference to an element which
  35 has already been deleted from the list/array.  Use atomic_inc_not_zero()
  36 in this scenario as follows:
  37
  38 1.                                      2.
  39 add()                                   search_and_reference()
  40 {                                       {
  41     alloc_object                            rcu_read_lock();
  42     ...                                     search_for_element
  43     atomic_set(&el->rc, 1);                 if (!atomic_inc_not_zero(&el->rc)) {
  44     spin_lock(&list_lock);                      rcu_read_unlock();
  45                                                 return FAIL;
  46     add_element                             }
  47     ...                                     ...
  48     spin_unlock(&list_lock);                rcu_read_unlock();
  49 }                                       }
  50 3.                                      4.
  51 release_referenced()                    delete()
  52 {                                       {
  53     ...                                     spin_lock(&list_lock);
  54     if (atomic_dec_and_test(&el->rc))       ...
  55         call_rcu(&el->head, el_free);       remove_element
  56     ...                                     spin_unlock(&list_lock);
  57 }                                           ...
  58                                             if (atomic_dec_and_test(&el->rc))
  59                                                 call_rcu(&el->head, el_free);
  60                                             ...
  61                                         }
  62
  63 Sometimes, a reference to the element needs to be obtained in the
  64 update (write) stream.  In such cases, atomic_inc_not_zero() might be
  65 overkill, since we hold the update-side spinlock.  One might instead
  66 use atomic_inc() in such cases.
  67
  68 It is not always convenient to deal with "FAIL" in the
  69 search_and_reference() code path.  In such cases, the
  70 atomic_dec_and_test() may be moved from delete() to el_free()
  71 as follows:
  72
  73 1.                                      2.
  74 add()                                   search_and_reference()
  75 {                                       {
  76     alloc_object                            rcu_read_lock();
  77     ...                                     search_for_element
  78     atomic_set(&el->rc, 1);                 atomic_inc(&el->rc);
  79     spin_lock(&list_lock);                  ...
  80
  81     add_element                             rcu_read_unlock();
  82     ...                                 }
  83     spin_unlock(&list_lock);            4.
  84 }                                       delete()
  85 3.                                      {
  86 release_referenced()                        spin_lock(&list_lock);
  87 {                                           ...
  88     ...                                     remove_element
  89     if (atomic_dec_and_test(&el->rc))       spin_unlock(&list_lock);
  90         kfree(el);                          ...
  91     ...                                     call_rcu(&el->head, el_free);
  92 }                                           ...
  93 5.                                      }
  94 void el_free(struct rcu_head *rhp)
  95 {
  96     release_referenced();
  97 }
  98
  99 The key point is that the initial reference added by add() is not removed
 100 until after a grace period has elapsed following removal.  This means that
 101 search_and_reference() cannot find this element, which means that the value
 102 of el->rc cannot increase.  Thus, once it reaches zero, there are no
 103 readers that can or ever will be able to reference the element.  The
 104 element can therefore safely be freed.  This in turn guarantees that if
 105 any reader finds the element, that reader may safely acquire a reference
 106 without checking the value of the reference counter.
 107
 108 In cases where delete() can sleep, synchronize_rcu() can be called from
 109 delete(), so that el_free() can be subsumed into delete as follows:
 110
 111 4.
 112 delete()
 113 {
 114     spin_lock(&list_lock);
 115     ...
 116     remove_element
 117     spin_unlock(&list_lock);
 118     ...
 119     synchronize_rcu();
 120     if (atomic_dec_and_test(&el->rc))
 121         kfree(el);
 122     ...
 123 }