1.. SPDX-License-Identifier: GPL-2.0
  2
  3=================================================
  4Using RCU hlist_nulls to protect list and objects
  5=================================================
  6
  7This section describes how to use hlist_nulls to
  8protect read-mostly linked lists and
  9objects using SLAB_TYPESAFE_BY_RCU allocations.
 10
 11Please read the basics in listRCU.rst.
 12
 13Using 'nulls'
 14=============
 15
 16Using special makers (called 'nulls') is a convenient way
 17to solve following problem.
 18
 19Without 'nulls', a typical RCU linked list managing objects which are
 20allocated with SLAB_TYPESAFE_BY_RCU kmem_cache can use the following
 21algorithms.  Following examples assume 'obj' is a pointer to such
 22objects, which is having below type.
 23
 24::
 25
 26  struct object {
 27    struct hlist_node obj_node;
 28    atomic_t refcnt;
 29    unsigned int key;
 30  };
 31
 321) Lookup algorithm
 33-------------------
 34
 35::
 36
 37  begin:
 38  rcu_read_lock();
 39  obj = lockless_lookup(key);
 40  if (obj) {
 41    if (!try_get_ref(obj)) { // might fail for free objects
 42      rcu_read_unlock();
 43      goto begin;
 44    }
 45    /*
 46    * Because a writer could delete object, and a writer could
 47    * reuse these object before the RCU grace period, we
 48    * must check key after getting the reference on object
 49    */
 50    if (obj->key != key) { // not the object we expected
 51      put_ref(obj);
 52      rcu_read_unlock();
 53      goto begin;
 54    }
 55  }
 56  rcu_read_unlock();
 57
 58Beware that lockless_lookup(key) cannot use traditional hlist_for_each_entry_rcu()
 59but a version with an additional memory barrier (smp_rmb())
 60
 61::
 62
 63  lockless_lookup(key)
 64  {
 65    struct hlist_node *node, *next;
 66    for (pos = rcu_dereference((head)->first);
 67         pos && ({ next = pos->next; smp_rmb(); prefetch(next); 1; }) &&
 68         ({ obj = hlist_entry(pos, typeof(*obj), obj_node); 1; });
 69         pos = rcu_dereference(next))
 70      if (obj->key == key)
 71        return obj;
 72    return NULL;
 73  }
 74
 75And note the traditional hlist_for_each_entry_rcu() misses this smp_rmb()::
 76
 77  struct hlist_node *node;
 78  for (pos = rcu_dereference((head)->first);
 79       pos && ({ prefetch(pos->next); 1; }) &&
 80       ({ obj = hlist_entry(pos, typeof(*obj), obj_node); 1; });
 81       pos = rcu_dereference(pos->next))
 82    if (obj->key == key)
 83      return obj;
 84  return NULL;
 85
 86Quoting Corey Minyard::
 87
 88  "If the object is moved from one list to another list in-between the
 89  time the hash is calculated and the next field is accessed, and the
 90  object has moved to the end of a new list, the traversal will not
 91  complete properly on the list it should have, since the object will
 92  be on the end of the new list and there's not a way to tell it's on a
 93  new list and restart the list traversal. I think that this can be
 94  solved by pre-fetching the "next" field (with proper barriers) before
 95  checking the key."
 96
 972) Insertion algorithm
 98----------------------
 99
100We need to make sure a reader cannot read the new 'obj->obj_node.next' value
101and previous value of 'obj->key'. Otherwise, an item could be deleted
102from a chain, and inserted into another chain. If new chain was empty
103before the move, 'next' pointer is NULL, and lockless reader can not
104detect the fact that it missed following items in original chain.
105
106::
107
108  /*
109   * Please note that new inserts are done at the head of list,
110   * not in the middle or end.
111   */
112  obj = kmem_cache_alloc(...);
113  lock_chain(); // typically a spin_lock()
114  obj->key = key;
115  atomic_set_release(&obj->refcnt, 1); // key before refcnt
116  hlist_add_head_rcu(&obj->obj_node, list);
117  unlock_chain(); // typically a spin_unlock()
118
119
1203) Removal algorithm
121--------------------
122
123Nothing special here, we can use a standard RCU hlist deletion.
124But thanks to SLAB_TYPESAFE_BY_RCU, beware a deleted object can be reused
125very very fast (before the end of RCU grace period)
126
127::
128
129  if (put_last_reference_on(obj) {
130    lock_chain(); // typically a spin_lock()
131    hlist_del_init_rcu(&obj->obj_node);
132    unlock_chain(); // typically a spin_unlock()
133    kmem_cache_free(cachep, obj);
134  }
135
136
137
138--------------------------------------------------------------------------
139
140Avoiding extra smp_rmb()
141========================
142
143With hlist_nulls we can avoid extra smp_rmb() in lockless_lookup().
144
145For example, if we choose to store the slot number as the 'nulls'
146end-of-list marker for each slot of the hash table, we can detect
147a race (some writer did a delete and/or a move of an object
148to another chain) checking the final 'nulls' value if
149the lookup met the end of chain. If final 'nulls' value
150is not the slot number, then we must restart the lookup at
151the beginning. If the object was moved to the same chain,
152then the reader doesn't care: It might occasionally
153scan the list again without harm.
154
155Note that using hlist_nulls means the type of 'obj_node' field of
156'struct object' becomes 'struct hlist_nulls_node'.
157
158
1591) lookup algorithm
160-------------------
161
162::
163
164  head = &table[slot];
165  begin:
166  rcu_read_lock();
167  hlist_nulls_for_each_entry_rcu(obj, node, head, obj_node) {
168    if (obj->key == key) {
169      if (!try_get_ref(obj)) { // might fail for free objects
170	rcu_read_unlock();
171        goto begin;
172      }
173      if (obj->key != key) { // not the object we expected
174        put_ref(obj);
175	rcu_read_unlock();
176        goto begin;
177      }
178      goto out;
179    }
180  }
181
182  // If the nulls value we got at the end of this lookup is
183  // not the expected one, we must restart lookup.
184  // We probably met an item that was moved to another chain.
185  if (get_nulls_value(node) != slot) {
186    put_ref(obj);
187    rcu_read_unlock();
188    goto begin;
189  }
190  obj = NULL;
191
192  out:
193  rcu_read_unlock();
194
1952) Insert algorithm
196-------------------
197
198Same to the above one, but uses hlist_nulls_add_head_rcu() instead of
199hlist_add_head_rcu().
200
201::
202
203  /*
204   * Please note that new inserts are done at the head of list,
205   * not in the middle or end.
206   */
207  obj = kmem_cache_alloc(cachep);
208  lock_chain(); // typically a spin_lock()
209  obj->key = key;
210  atomic_set_release(&obj->refcnt, 1); // key before refcnt
211  /*
212   * insert obj in RCU way (readers might be traversing chain)
213   */
214  hlist_nulls_add_head_rcu(&obj->obj_node, list);
215  unlock_chain(); // typically a spin_unlock()