xref: /linux/lib/list_sort.c (revision e2683c8868d03382da7e1ce8453b543a043066d1)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/compiler.h>
3 #include <linux/export.h>
4 #include <linux/list_sort.h>
5 #include <linux/list.h>
6 
7 /*
8  * Returns a list organized in an intermediate format suited
9  * to chaining of merge() calls: null-terminated, no reserved or
10  * sentinel head node, "prev" links not maintained.
11  */
12 __attribute__((nonnull(2,3,4)))
13 static struct list_head *merge(void *priv, list_cmp_func_t cmp,
14 				struct list_head *a, struct list_head *b)
15 {
16 	struct list_head *head, **tail = &head;
17 
18 	for (;;) {
19 		/* if equal, take 'a' -- important for sort stability */
20 		if (cmp(priv, a, b) <= 0) {
21 			*tail = a;
22 			tail = &a->next;
23 			a = a->next;
24 			if (!a) {
25 				*tail = b;
26 				break;
27 			}
28 		} else {
29 			*tail = b;
30 			tail = &b->next;
31 			b = b->next;
32 			if (!b) {
33 				*tail = a;
34 				break;
35 			}
36 		}
37 	}
38 	return head;
39 }
40 
41 /*
42  * Combine final list merge with restoration of standard doubly-linked
43  * list structure.  This approach duplicates code from merge(), but
44  * runs faster than the tidier alternatives of either a separate final
45  * prev-link restoration pass, or maintaining the prev links
46  * throughout.
47  */
48 __attribute__((nonnull(2,3,4,5)))
49 static void merge_final(void *priv, list_cmp_func_t cmp, struct list_head *head,
50 			struct list_head *a, struct list_head *b)
51 {
52 	struct list_head *tail = head;
53 
54 	for (;;) {
55 		/* if equal, take 'a' -- important for sort stability */
56 		if (cmp(priv, a, b) <= 0) {
57 			tail->next = a;
58 			a->prev = tail;
59 			tail = a;
60 			a = a->next;
61 			if (!a)
62 				break;
63 		} else {
64 			tail->next = b;
65 			b->prev = tail;
66 			tail = b;
67 			b = b->next;
68 			if (!b) {
69 				b = a;
70 				break;
71 			}
72 		}
73 	}
74 
75 	/* Finish linking remainder of list b on to tail */
76 	tail->next = b;
77 	do {
78 		b->prev = tail;
79 		tail = b;
80 		b = b->next;
81 	} while (b);
82 
83 	/* And the final links to make a circular doubly-linked list */
84 	tail->next = head;
85 	head->prev = tail;
86 }
87 
88 /**
89  * list_sort - sort a list
90  * @priv: private data, opaque to list_sort(), passed to @cmp
91  * @head: the list to sort
92  * @cmp: the elements comparison function
93  *
94  * The comparison function @cmp must return > 0 if @a should sort after
95  * @b ("@a > @b" if you want an ascending sort), and <= 0 if @a should
96  * sort before @b *or* their original order should be preserved.  It is
97  * always called with the element that came first in the input in @a,
98  * and list_sort is a stable sort, so it is not necessary to distinguish
99  * the @a < @b and @a == @b cases.
100  *
101  * The comparison function must adhere to specific mathematical properties
102  * to ensure correct and stable sorting:
103  * - Antisymmetry: cmp(@a, @b) must return the opposite sign of
104  * cmp(@b, @a).
105  * - Transitivity: if cmp(@a, @b) <= 0 and cmp(@b, @c) <= 0, then
106  * cmp(@a, @c) <= 0.
107  *
108  * This is compatible with two styles of @cmp function:
109  * - The traditional style which returns <0 / =0 / >0, or
110  * - Returning a boolean 0/1.
111  * The latter offers a chance to save a few cycles in the comparison
112  * (which is used by e.g. plug_ctx_cmp() in block/blk-mq.c).
113  *
114  * A good way to write a multi-word comparison is::
115  *
116  *	if (a->high != b->high)
117  *		return a->high > b->high;
118  *	if (a->middle != b->middle)
119  *		return a->middle > b->middle;
120  *	return a->low > b->low;
121  *
122  *
123  * This mergesort is as eager as possible while always performing at least
124  * 2:1 balanced merges.  Given two pending sublists of size 2^k, they are
125  * merged to a size-2^(k+1) list as soon as we have 2^k following elements.
126  *
127  * Thus, it will avoid cache thrashing as long as 3*2^k elements can
128  * fit into the cache.  Not quite as good as a fully-eager bottom-up
129  * mergesort, but it does use 0.2*n fewer comparisons, so is faster in
130  * the common case that everything fits into L1.
131  *
132  *
133  * The merging is controlled by "count", the number of elements in the
134  * pending lists.  This is beautifully simple code, but rather subtle.
135  *
136  * Each time we increment "count", we set one bit (bit k) and clear
137  * bits k-1 .. 0.  Each time this happens (except the very first time
138  * for each bit, when count increments to 2^k), we merge two lists of
139  * size 2^k into one list of size 2^(k+1).
140  *
141  * This merge happens exactly when the count reaches an odd multiple of
142  * 2^k, which is when we have 2^k elements pending in smaller lists,
143  * so it's safe to merge away two lists of size 2^k.
144  *
145  * After this happens twice, we have created two lists of size 2^(k+1),
146  * which will be merged into a list of size 2^(k+2) before we create
147  * a third list of size 2^(k+1), so there are never more than two pending.
148  *
149  * The number of pending lists of size 2^k is determined by the
150  * state of bit k of "count" plus two extra pieces of information:
151  *
152  * - The state of bit k-1 (when k == 0, consider bit -1 always set), and
153  * - Whether the higher-order bits are zero or non-zero (i.e.
154  *   is count >= 2^(k+1)).
155  *
156  * There are six states we distinguish.  "x" represents some arbitrary
157  * bits, and "y" represents some arbitrary non-zero bits:
158  * 0:  00x: 0 pending of size 2^k;           x pending of sizes < 2^k
159  * 1:  01x: 0 pending of size 2^k; 2^(k-1) + x pending of sizes < 2^k
160  * 2: x10x: 0 pending of size 2^k; 2^k     + x pending of sizes < 2^k
161  * 3: x11x: 1 pending of size 2^k; 2^(k-1) + x pending of sizes < 2^k
162  * 4: y00x: 1 pending of size 2^k; 2^k     + x pending of sizes < 2^k
163  * 5: y01x: 2 pending of size 2^k; 2^(k-1) + x pending of sizes < 2^k
164  * (merge and loop back to state 2)
165  *
166  * We gain lists of size 2^k in the 2->3 and 4->5 transitions (because
167  * bit k-1 is set while the more significant bits are non-zero) and
168  * merge them away in the 5->2 transition.  Note in particular that just
169  * before the 5->2 transition, all lower-order bits are 11 (state 3),
170  * so there is one list of each smaller size.
171  *
172  * When we reach the end of the input, we merge all the pending
173  * lists, from smallest to largest.  If you work through cases 2 to
174  * 5 above, you can see that the number of elements we merge with a list
175  * of size 2^k varies from 2^(k-1) (cases 3 and 5 when x == 0) to
176  * 2^(k+1) - 1 (second merge of case 5 when x == 2^(k-1) - 1).
177  */
178 __attribute__((nonnull(2,3)))
179 void list_sort(void *priv, struct list_head *head, list_cmp_func_t cmp)
180 {
181 	struct list_head *list = head->next, *pending = NULL;
182 	size_t count = 0;	/* Count of pending */
183 
184 	if (list == head->prev)	/* Zero or one elements */
185 		return;
186 
187 	/* Convert to a null-terminated singly-linked list. */
188 	head->prev->next = NULL;
189 
190 	/*
191 	 * Data structure invariants:
192 	 * - All lists are singly linked and null-terminated; prev
193 	 *   pointers are not maintained.
194 	 * - pending is a prev-linked "list of lists" of sorted
195 	 *   sublists awaiting further merging.
196 	 * - Each of the sorted sublists is power-of-two in size.
197 	 * - Sublists are sorted by size and age, smallest & newest at front.
198 	 * - There are zero to two sublists of each size.
199 	 * - A pair of pending sublists are merged as soon as the number
200 	 *   of following pending elements equals their size (i.e.
201 	 *   each time count reaches an odd multiple of that size).
202 	 *   That ensures each later final merge will be at worst 2:1.
203 	 * - Each round consists of:
204 	 *   - Merging the two sublists selected by the highest bit
205 	 *     which flips when count is incremented, and
206 	 *   - Adding an element from the input as a size-1 sublist.
207 	 */
208 	do {
209 		size_t bits;
210 		struct list_head **tail = &pending;
211 
212 		/* Find the least-significant clear bit in count */
213 		for (bits = count; bits & 1; bits >>= 1)
214 			tail = &(*tail)->prev;
215 		/* Do the indicated merge */
216 		if (likely(bits)) {
217 			struct list_head *a = *tail, *b = a->prev;
218 
219 			a = merge(priv, cmp, b, a);
220 			/* Install the merged result in place of the inputs */
221 			a->prev = b->prev;
222 			*tail = a;
223 		}
224 
225 		/* Move one element from input list to pending */
226 		list->prev = pending;
227 		pending = list;
228 		list = list->next;
229 		pending->next = NULL;
230 		count++;
231 	} while (list);
232 
233 	/* End of input; merge together all the pending lists. */
234 	list = pending;
235 	pending = pending->prev;
236 	for (;;) {
237 		struct list_head *next = pending->prev;
238 
239 		if (!next)
240 			break;
241 		list = merge(priv, cmp, pending, list);
242 		pending = next;
243 	}
244 	/* The final merge, rebuilding prev links */
245 	merge_final(priv, cmp, head, pending, list);
246 }
247 EXPORT_SYMBOL(list_sort);
248