1=================================
2Red-black Trees (rbtree) in Linux
3=================================
4
5
6:Date: January 18, 2007
7:Author: Rob Landley <rob@landley.net>
8
9What are red-black trees, and what are they for?
10------------------------------------------------
11
12Red-black trees are a type of self-balancing binary search tree, used for
13storing sortable key/value data pairs.  This differs from radix trees (which
14are used to efficiently store sparse arrays and thus use long integer indexes
15to insert/access/delete nodes) and hash tables (which are not kept sorted to
16be easily traversed in order, and must be tuned for a specific size and
17hash function where rbtrees scale gracefully storing arbitrary keys).
18
19Red-black trees are similar to AVL trees, but provide faster real-time bounded
20worst case performance for insertion and deletion (at most two rotations and
21three rotations, respectively, to balance the tree), with slightly slower
22(but still O(log n)) lookup time.
23
24To quote Linux Weekly News:
25
26    There are a number of red-black trees in use in the kernel.
27    The deadline and CFQ I/O schedulers employ rbtrees to
28    track requests; the packet CD/DVD driver does the same.
29    The high-resolution timer code uses an rbtree to organize outstanding
30    timer requests.  The ext3 filesystem tracks directory entries in a
31    red-black tree.  Virtual memory areas (VMAs) are tracked with red-black
32    trees, as are epoll file descriptors, cryptographic keys, and network
33    packets in the "hierarchical token bucket" scheduler.
34
35This document covers use of the Linux rbtree implementation.  For more
36information on the nature and implementation of Red Black Trees,  see:
37
38  Linux Weekly News article on red-black trees
39    https://lwn.net/Articles/184495/
40
41  Wikipedia entry on red-black trees
42    https://en.wikipedia.org/wiki/Red-black_tree
43
44Linux implementation of red-black trees
45---------------------------------------
46
47Linux's rbtree implementation lives in the file "lib/rbtree.c".  To use it,
48"#include <linux/rbtree.h>".
49
50The Linux rbtree implementation is optimized for speed, and thus has one
51less layer of indirection (and better cache locality) than more traditional
52tree implementations.  Instead of using pointers to separate rb_node and data
53structures, each instance of struct rb_node is embedded in the data structure
54it organizes.  And instead of using a comparison callback function pointer,
55users are expected to write their own tree search and insert functions
56which call the provided rbtree functions.  Locking is also left up to the
57user of the rbtree code.
58
59Creating a new rbtree
60---------------------
61
62Data nodes in an rbtree tree are structures containing a struct rb_node member::
63
64  struct mytype {
65  	struct rb_node node;
66  	char *keystring;
67  };
68
69When dealing with a pointer to the embedded struct rb_node, the containing data
70structure may be accessed with the standard container_of() macro.  In addition,
71individual members may be accessed directly via rb_entry(node, type, member).
72
73At the root of each rbtree is an rb_root structure, which is initialized to be
74empty via:
75
76  struct rb_root mytree = RB_ROOT;
77
78Searching for a value in an rbtree
79----------------------------------
80
81Writing a search function for your tree is fairly straightforward: start at the
82root, compare each value, and follow the left or right branch as necessary.
83
84Example::
85
86  struct mytype *my_search(struct rb_root *root, char *string)
87  {
88  	struct rb_node *node = root->rb_node;
89
90  	while (node) {
91  		struct mytype *data = container_of(node, struct mytype, node);
92		int result;
93
94		result = strcmp(string, data->keystring);
95
96		if (result < 0)
97  			node = node->rb_left;
98		else if (result > 0)
99  			node = node->rb_right;
100		else
101  			return data;
102	}
103	return NULL;
104  }
105
106Inserting data into an rbtree
107-----------------------------
108
109Inserting data in the tree involves first searching for the place to insert the
110new node, then inserting the node and rebalancing ("recoloring") the tree.
111
112The search for insertion differs from the previous search by finding the
113location of the pointer on which to graft the new node.  The new node also
114needs a link to its parent node for rebalancing purposes.
115
116Example::
117
118  int my_insert(struct rb_root *root, struct mytype *data)
119  {
120  	struct rb_node **new = &(root->rb_node), *parent = NULL;
121
122  	/* Figure out where to put new node */
123  	while (*new) {
124  		struct mytype *this = container_of(*new, struct mytype, node);
125  		int result = strcmp(data->keystring, this->keystring);
126
127		parent = *new;
128  		if (result < 0)
129  			new = &((*new)->rb_left);
130  		else if (result > 0)
131  			new = &((*new)->rb_right);
132  		else
133  			return FALSE;
134  	}
135
136  	/* Add new node and rebalance tree. */
137  	rb_link_node(&data->node, parent, new);
138  	rb_insert_color(&data->node, root);
139
140	return TRUE;
141  }
142
143Removing or replacing existing data in an rbtree
144------------------------------------------------
145
146To remove an existing node from a tree, call::
147
148  void rb_erase(struct rb_node *victim, struct rb_root *tree);
149
150Example::
151
152  struct mytype *data = mysearch(&mytree, "walrus");
153
154  if (data) {
155  	rb_erase(&data->node, &mytree);
156  	myfree(data);
157  }
158
159To replace an existing node in a tree with a new one with the same key, call::
160
161  void rb_replace_node(struct rb_node *old, struct rb_node *new,
162  			struct rb_root *tree);
163
164Replacing a node this way does not re-sort the tree: If the new node doesn't
165have the same key as the old node, the rbtree will probably become corrupted.
166
167Iterating through the elements stored in an rbtree (in sort order)
168------------------------------------------------------------------
169
170Four functions are provided for iterating through an rbtree's contents in
171sorted order.  These work on arbitrary trees, and should not need to be
172modified or wrapped (except for locking purposes)::
173
174  struct rb_node *rb_first(struct rb_root *tree);
175  struct rb_node *rb_last(struct rb_root *tree);
176  struct rb_node *rb_next(struct rb_node *node);
177  struct rb_node *rb_prev(struct rb_node *node);
178
179To start iterating, call rb_first() or rb_last() with a pointer to the root
180of the tree, which will return a pointer to the node structure contained in
181the first or last element in the tree.  To continue, fetch the next or previous
182node by calling rb_next() or rb_prev() on the current node.  This will return
183NULL when there are no more nodes left.
184
185The iterator functions return a pointer to the embedded struct rb_node, from
186which the containing data structure may be accessed with the container_of()
187macro, and individual members may be accessed directly via
188rb_entry(node, type, member).
189
190Example::
191
192  struct rb_node *node;
193  for (node = rb_first(&mytree); node; node = rb_next(node))
194	printk("key=%s\n", rb_entry(node, struct mytype, node)->keystring);
195
196Cached rbtrees
197--------------
198
199Computing the leftmost (smallest) node is quite a common task for binary
200search trees, such as for traversals or users relying on a the particular
201order for their own logic. To this end, users can use 'struct rb_root_cached'
202to optimize O(logN) rb_first() calls to a simple pointer fetch avoiding
203potentially expensive tree iterations. This is done at negligible runtime
204overhead for maintanence; albeit larger memory footprint.
205
206Similar to the rb_root structure, cached rbtrees are initialized to be
207empty via::
208
209  struct rb_root_cached mytree = RB_ROOT_CACHED;
210
211Cached rbtree is simply a regular rb_root with an extra pointer to cache the
212leftmost node. This allows rb_root_cached to exist wherever rb_root does,
213which permits augmented trees to be supported as well as only a few extra
214interfaces::
215
216  struct rb_node *rb_first_cached(struct rb_root_cached *tree);
217  void rb_insert_color_cached(struct rb_node *, struct rb_root_cached *, bool);
218  void rb_erase_cached(struct rb_node *node, struct rb_root_cached *);
219
220Both insert and erase calls have their respective counterpart of augmented
221trees::
222
223  void rb_insert_augmented_cached(struct rb_node *node, struct rb_root_cached *,
224				  bool, struct rb_augment_callbacks *);
225  void rb_erase_augmented_cached(struct rb_node *, struct rb_root_cached *,
226				 struct rb_augment_callbacks *);
227
228
229Support for Augmented rbtrees
230-----------------------------
231
232Augmented rbtree is an rbtree with "some" additional data stored in
233each node, where the additional data for node N must be a function of
234the contents of all nodes in the subtree rooted at N. This data can
235be used to augment some new functionality to rbtree. Augmented rbtree
236is an optional feature built on top of basic rbtree infrastructure.
237An rbtree user who wants this feature will have to call the augmentation
238functions with the user provided augmentation callback when inserting
239and erasing nodes.
240
241C files implementing augmented rbtree manipulation must include
242<linux/rbtree_augmented.h> instead of <linux/rbtree.h>. Note that
243linux/rbtree_augmented.h exposes some rbtree implementations details
244you are not expected to rely on; please stick to the documented APIs
245there and do not include <linux/rbtree_augmented.h> from header files
246either so as to minimize chances of your users accidentally relying on
247such implementation details.
248
249On insertion, the user must update the augmented information on the path
250leading to the inserted node, then call rb_link_node() as usual and
251rb_augment_inserted() instead of the usual rb_insert_color() call.
252If rb_augment_inserted() rebalances the rbtree, it will callback into
253a user provided function to update the augmented information on the
254affected subtrees.
255
256When erasing a node, the user must call rb_erase_augmented() instead of
257rb_erase(). rb_erase_augmented() calls back into user provided functions
258to updated the augmented information on affected subtrees.
259
260In both cases, the callbacks are provided through struct rb_augment_callbacks.
2613 callbacks must be defined:
262
263- A propagation callback, which updates the augmented value for a given
264  node and its ancestors, up to a given stop point (or NULL to update
265  all the way to the root).
266
267- A copy callback, which copies the augmented value for a given subtree
268  to a newly assigned subtree root.
269
270- A tree rotation callback, which copies the augmented value for a given
271  subtree to a newly assigned subtree root AND recomputes the augmented
272  information for the former subtree root.
273
274The compiled code for rb_erase_augmented() may inline the propagation and
275copy callbacks, which results in a large function, so each augmented rbtree
276user should have a single rb_erase_augmented() call site in order to limit
277compiled code size.
278
279
280Sample usage
281^^^^^^^^^^^^
282
283Interval tree is an example of augmented rb tree. Reference -
284"Introduction to Algorithms" by Cormen, Leiserson, Rivest and Stein.
285More details about interval trees:
286
287Classical rbtree has a single key and it cannot be directly used to store
288interval ranges like [lo:hi] and do a quick lookup for any overlap with a new
289lo:hi or to find whether there is an exact match for a new lo:hi.
290
291However, rbtree can be augmented to store such interval ranges in a structured
292way making it possible to do efficient lookup and exact match.
293
294This "extra information" stored in each node is the maximum hi
295(max_hi) value among all the nodes that are its descendants. This
296information can be maintained at each node just be looking at the node
297and its immediate children. And this will be used in O(log n) lookup
298for lowest match (lowest start address among all possible matches)
299with something like::
300
301  struct interval_tree_node *
302  interval_tree_first_match(struct rb_root *root,
303			    unsigned long start, unsigned long last)
304  {
305	struct interval_tree_node *node;
306
307	if (!root->rb_node)
308		return NULL;
309	node = rb_entry(root->rb_node, struct interval_tree_node, rb);
310
311	while (true) {
312		if (node->rb.rb_left) {
313			struct interval_tree_node *left =
314				rb_entry(node->rb.rb_left,
315					 struct interval_tree_node, rb);
316			if (left->__subtree_last >= start) {
317				/*
318				 * Some nodes in left subtree satisfy Cond2.
319				 * Iterate to find the leftmost such node N.
320				 * If it also satisfies Cond1, that's the match
321				 * we are looking for. Otherwise, there is no
322				 * matching interval as nodes to the right of N
323				 * can't satisfy Cond1 either.
324				 */
325				node = left;
326				continue;
327			}
328		}
329		if (node->start <= last) {		/* Cond1 */
330			if (node->last >= start)	/* Cond2 */
331				return node;	/* node is leftmost match */
332			if (node->rb.rb_right) {
333				node = rb_entry(node->rb.rb_right,
334					struct interval_tree_node, rb);
335				if (node->__subtree_last >= start)
336					continue;
337			}
338		}
339		return NULL;	/* No match */
340	}
341  }
342
343Insertion/removal are defined using the following augmented callbacks::
344
345  static inline unsigned long
346  compute_subtree_last(struct interval_tree_node *node)
347  {
348	unsigned long max = node->last, subtree_last;
349	if (node->rb.rb_left) {
350		subtree_last = rb_entry(node->rb.rb_left,
351			struct interval_tree_node, rb)->__subtree_last;
352		if (max < subtree_last)
353			max = subtree_last;
354	}
355	if (node->rb.rb_right) {
356		subtree_last = rb_entry(node->rb.rb_right,
357			struct interval_tree_node, rb)->__subtree_last;
358		if (max < subtree_last)
359			max = subtree_last;
360	}
361	return max;
362  }
363
364  static void augment_propagate(struct rb_node *rb, struct rb_node *stop)
365  {
366	while (rb != stop) {
367		struct interval_tree_node *node =
368			rb_entry(rb, struct interval_tree_node, rb);
369		unsigned long subtree_last = compute_subtree_last(node);
370		if (node->__subtree_last == subtree_last)
371			break;
372		node->__subtree_last = subtree_last;
373		rb = rb_parent(&node->rb);
374	}
375  }
376
377  static void augment_copy(struct rb_node *rb_old, struct rb_node *rb_new)
378  {
379	struct interval_tree_node *old =
380		rb_entry(rb_old, struct interval_tree_node, rb);
381	struct interval_tree_node *new =
382		rb_entry(rb_new, struct interval_tree_node, rb);
383
384	new->__subtree_last = old->__subtree_last;
385  }
386
387  static void augment_rotate(struct rb_node *rb_old, struct rb_node *rb_new)
388  {
389	struct interval_tree_node *old =
390		rb_entry(rb_old, struct interval_tree_node, rb);
391	struct interval_tree_node *new =
392		rb_entry(rb_new, struct interval_tree_node, rb);
393
394	new->__subtree_last = old->__subtree_last;
395	old->__subtree_last = compute_subtree_last(old);
396  }
397
398  static const struct rb_augment_callbacks augment_callbacks = {
399	augment_propagate, augment_copy, augment_rotate
400  };
401
402  void interval_tree_insert(struct interval_tree_node *node,
403			    struct rb_root *root)
404  {
405	struct rb_node **link = &root->rb_node, *rb_parent = NULL;
406	unsigned long start = node->start, last = node->last;
407	struct interval_tree_node *parent;
408
409	while (*link) {
410		rb_parent = *link;
411		parent = rb_entry(rb_parent, struct interval_tree_node, rb);
412		if (parent->__subtree_last < last)
413			parent->__subtree_last = last;
414		if (start < parent->start)
415			link = &parent->rb.rb_left;
416		else
417			link = &parent->rb.rb_right;
418	}
419
420	node->__subtree_last = last;
421	rb_link_node(&node->rb, rb_parent, link);
422	rb_insert_augmented(&node->rb, root, &augment_callbacks);
423  }
424
425  void interval_tree_remove(struct interval_tree_node *node,
426			    struct rb_root *root)
427  {
428	rb_erase_augmented(&node->rb, root, &augment_callbacks);
429  }
430