xref: /openbmc/linux/lib/list_sort.c (revision 3dc4b6fb)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/kernel.h>
3 #include <linux/bug.h>
4 #include <linux/compiler.h>
5 #include <linux/export.h>
6 #include <linux/string.h>
7 #include <linux/list_sort.h>
8 #include <linux/list.h>
9 
10 typedef int __attribute__((nonnull(2,3))) (*cmp_func)(void *,
11 		struct list_head const *, struct list_head const *);
12 
13 /*
14  * Returns a list organized in an intermediate format suited
15  * to chaining of merge() calls: null-terminated, no reserved or
16  * sentinel head node, "prev" links not maintained.
17  */
18 __attribute__((nonnull(2,3,4)))
19 static struct list_head *merge(void *priv, cmp_func cmp,
20 				struct list_head *a, struct list_head *b)
21 {
22 	struct list_head *head, **tail = &head;
23 
24 	for (;;) {
25 		/* if equal, take 'a' -- important for sort stability */
26 		if (cmp(priv, a, b) <= 0) {
27 			*tail = a;
28 			tail = &a->next;
29 			a = a->next;
30 			if (!a) {
31 				*tail = b;
32 				break;
33 			}
34 		} else {
35 			*tail = b;
36 			tail = &b->next;
37 			b = b->next;
38 			if (!b) {
39 				*tail = a;
40 				break;
41 			}
42 		}
43 	}
44 	return head;
45 }
46 
47 /*
48  * Combine final list merge with restoration of standard doubly-linked
49  * list structure.  This approach duplicates code from merge(), but
50  * runs faster than the tidier alternatives of either a separate final
51  * prev-link restoration pass, or maintaining the prev links
52  * throughout.
53  */
54 __attribute__((nonnull(2,3,4,5)))
55 static void merge_final(void *priv, cmp_func cmp, struct list_head *head,
56 			struct list_head *a, struct list_head *b)
57 {
58 	struct list_head *tail = head;
59 	u8 count = 0;
60 
61 	for (;;) {
62 		/* if equal, take 'a' -- important for sort stability */
63 		if (cmp(priv, a, b) <= 0) {
64 			tail->next = a;
65 			a->prev = tail;
66 			tail = a;
67 			a = a->next;
68 			if (!a)
69 				break;
70 		} else {
71 			tail->next = b;
72 			b->prev = tail;
73 			tail = b;
74 			b = b->next;
75 			if (!b) {
76 				b = a;
77 				break;
78 			}
79 		}
80 	}
81 
82 	/* Finish linking remainder of list b on to tail */
83 	tail->next = b;
84 	do {
85 		/*
86 		 * If the merge is highly unbalanced (e.g. the input is
87 		 * already sorted), this loop may run many iterations.
88 		 * Continue callbacks to the client even though no
89 		 * element comparison is needed, so the client's cmp()
90 		 * routine can invoke cond_resched() periodically.
91 		 */
92 		if (unlikely(!++count))
93 			cmp(priv, b, b);
94 		b->prev = tail;
95 		tail = b;
96 		b = b->next;
97 	} while (b);
98 
99 	/* And the final links to make a circular doubly-linked list */
100 	tail->next = head;
101 	head->prev = tail;
102 }
103 
104 /**
105  * list_sort - sort a list
106  * @priv: private data, opaque to list_sort(), passed to @cmp
107  * @head: the list to sort
108  * @cmp: the elements comparison function
109  *
110  * The comparison funtion @cmp must return > 0 if @a should sort after
111  * @b ("@a > @b" if you want an ascending sort), and <= 0 if @a should
112  * sort before @b *or* their original order should be preserved.  It is
113  * always called with the element that came first in the input in @a,
114  * and list_sort is a stable sort, so it is not necessary to distinguish
115  * the @a < @b and @a == @b cases.
116  *
117  * This is compatible with two styles of @cmp function:
118  * - The traditional style which returns <0 / =0 / >0, or
119  * - Returning a boolean 0/1.
120  * The latter offers a chance to save a few cycles in the comparison
121  * (which is used by e.g. plug_ctx_cmp() in block/blk-mq.c).
122  *
123  * A good way to write a multi-word comparison is::
124  *
125  *	if (a->high != b->high)
126  *		return a->high > b->high;
127  *	if (a->middle != b->middle)
128  *		return a->middle > b->middle;
129  *	return a->low > b->low;
130  *
131  *
132  * This mergesort is as eager as possible while always performing at least
133  * 2:1 balanced merges.  Given two pending sublists of size 2^k, they are
134  * merged to a size-2^(k+1) list as soon as we have 2^k following elements.
135  *
136  * Thus, it will avoid cache thrashing as long as 3*2^k elements can
137  * fit into the cache.  Not quite as good as a fully-eager bottom-up
138  * mergesort, but it does use 0.2*n fewer comparisons, so is faster in
139  * the common case that everything fits into L1.
140  *
141  *
142  * The merging is controlled by "count", the number of elements in the
143  * pending lists.  This is beautiully simple code, but rather subtle.
144  *
145  * Each time we increment "count", we set one bit (bit k) and clear
146  * bits k-1 .. 0.  Each time this happens (except the very first time
147  * for each bit, when count increments to 2^k), we merge two lists of
148  * size 2^k into one list of size 2^(k+1).
149  *
150  * This merge happens exactly when the count reaches an odd multiple of
151  * 2^k, which is when we have 2^k elements pending in smaller lists,
152  * so it's safe to merge away two lists of size 2^k.
153  *
154  * After this happens twice, we have created two lists of size 2^(k+1),
155  * which will be merged into a list of size 2^(k+2) before we create
156  * a third list of size 2^(k+1), so there are never more than two pending.
157  *
158  * The number of pending lists of size 2^k is determined by the
159  * state of bit k of "count" plus two extra pieces of information:
160  *
161  * - The state of bit k-1 (when k == 0, consider bit -1 always set), and
162  * - Whether the higher-order bits are zero or non-zero (i.e.
163  *   is count >= 2^(k+1)).
164  *
165  * There are six states we distinguish.  "x" represents some arbitrary
166  * bits, and "y" represents some arbitrary non-zero bits:
167  * 0:  00x: 0 pending of size 2^k;           x pending of sizes < 2^k
168  * 1:  01x: 0 pending of size 2^k; 2^(k-1) + x pending of sizes < 2^k
169  * 2: x10x: 0 pending of size 2^k; 2^k     + x pending of sizes < 2^k
170  * 3: x11x: 1 pending of size 2^k; 2^(k-1) + x pending of sizes < 2^k
171  * 4: y00x: 1 pending of size 2^k; 2^k     + x pending of sizes < 2^k
172  * 5: y01x: 2 pending of size 2^k; 2^(k-1) + x pending of sizes < 2^k
173  * (merge and loop back to state 2)
174  *
175  * We gain lists of size 2^k in the 2->3 and 4->5 transitions (because
176  * bit k-1 is set while the more significant bits are non-zero) and
177  * merge them away in the 5->2 transition.  Note in particular that just
178  * before the 5->2 transition, all lower-order bits are 11 (state 3),
179  * so there is one list of each smaller size.
180  *
181  * When we reach the end of the input, we merge all the pending
182  * lists, from smallest to largest.  If you work through cases 2 to
183  * 5 above, you can see that the number of elements we merge with a list
184  * of size 2^k varies from 2^(k-1) (cases 3 and 5 when x == 0) to
185  * 2^(k+1) - 1 (second merge of case 5 when x == 2^(k-1) - 1).
186  */
187 __attribute__((nonnull(2,3)))
188 void list_sort(void *priv, struct list_head *head,
189 		int (*cmp)(void *priv, struct list_head *a,
190 			struct list_head *b))
191 {
192 	struct list_head *list = head->next, *pending = NULL;
193 	size_t count = 0;	/* Count of pending */
194 
195 	if (list == head->prev)	/* Zero or one elements */
196 		return;
197 
198 	/* Convert to a null-terminated singly-linked list. */
199 	head->prev->next = NULL;
200 
201 	/*
202 	 * Data structure invariants:
203 	 * - All lists are singly linked and null-terminated; prev
204 	 *   pointers are not maintained.
205 	 * - pending is a prev-linked "list of lists" of sorted
206 	 *   sublists awaiting further merging.
207 	 * - Each of the sorted sublists is power-of-two in size.
208 	 * - Sublists are sorted by size and age, smallest & newest at front.
209 	 * - There are zero to two sublists of each size.
210 	 * - A pair of pending sublists are merged as soon as the number
211 	 *   of following pending elements equals their size (i.e.
212 	 *   each time count reaches an odd multiple of that size).
213 	 *   That ensures each later final merge will be at worst 2:1.
214 	 * - Each round consists of:
215 	 *   - Merging the two sublists selected by the highest bit
216 	 *     which flips when count is incremented, and
217 	 *   - Adding an element from the input as a size-1 sublist.
218 	 */
219 	do {
220 		size_t bits;
221 		struct list_head **tail = &pending;
222 
223 		/* Find the least-significant clear bit in count */
224 		for (bits = count; bits & 1; bits >>= 1)
225 			tail = &(*tail)->prev;
226 		/* Do the indicated merge */
227 		if (likely(bits)) {
228 			struct list_head *a = *tail, *b = a->prev;
229 
230 			a = merge(priv, (cmp_func)cmp, b, a);
231 			/* Install the merged result in place of the inputs */
232 			a->prev = b->prev;
233 			*tail = a;
234 		}
235 
236 		/* Move one element from input list to pending */
237 		list->prev = pending;
238 		pending = list;
239 		list = list->next;
240 		pending->next = NULL;
241 		count++;
242 	} while (list);
243 
244 	/* End of input; merge together all the pending lists. */
245 	list = pending;
246 	pending = pending->prev;
247 	for (;;) {
248 		struct list_head *next = pending->prev;
249 
250 		if (!next)
251 			break;
252 		list = merge(priv, (cmp_func)cmp, pending, list);
253 		pending = next;
254 	}
255 	/* The final merge, rebuilding prev links */
256 	merge_final(priv, (cmp_func)cmp, head, pending, list);
257 }
258 EXPORT_SYMBOL(list_sort);
259