xref: /openbmc/linux/lib/sort.c (revision adb80490)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * A fast, small, non-recursive O(n log n) sort for the Linux kernel
4  *
5  * This performs n*log2(n) + 0.37*n + o(n) comparisons on average,
6  * and 1.5*n*log2(n) + O(n) in the (very contrived) worst case.
7  *
8  * Glibc qsort() manages n*log2(n) - 1.26*n for random inputs (1.63*n
9  * better) at the expense of stack usage and much larger code to avoid
10  * quicksort's O(n^2) worst case.
11  */
12 
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14 
15 #include <linux/types.h>
16 #include <linux/export.h>
17 #include <linux/sort.h>
18 
19 /**
20  * is_aligned - is this pointer & size okay for word-wide copying?
21  * @base: pointer to data
22  * @size: size of each element
23  * @align: required alignment (typically 4 or 8)
24  *
25  * Returns true if elements can be copied using word loads and stores.
26  * The size must be a multiple of the alignment, and the base address must
27  * be if we do not have CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS.
28  *
29  * For some reason, gcc doesn't know to optimize "if (a & mask || b & mask)"
30  * to "if ((a | b) & mask)", so we do that by hand.
31  */
32 __attribute_const__ __always_inline
33 static bool is_aligned(const void *base, size_t size, unsigned char align)
34 {
35 	unsigned char lsbits = (unsigned char)size;
36 
37 	(void)base;
38 #ifndef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
39 	lsbits |= (unsigned char)(uintptr_t)base;
40 #endif
41 	return (lsbits & (align - 1)) == 0;
42 }
43 
44 /**
45  * swap_words_32 - swap two elements in 32-bit chunks
46  * @a: pointer to the first element to swap
47  * @b: pointer to the second element to swap
48  * @n: element size (must be a multiple of 4)
49  *
50  * Exchange the two objects in memory.  This exploits base+index addressing,
51  * which basically all CPUs have, to minimize loop overhead computations.
52  *
53  * For some reason, on x86 gcc 7.3.0 adds a redundant test of n at the
54  * bottom of the loop, even though the zero flag is stil valid from the
55  * subtract (since the intervening mov instructions don't alter the flags).
56  * Gcc 8.1.0 doesn't have that problem.
57  */
58 static void swap_words_32(void *a, void *b, size_t n)
59 {
60 	do {
61 		u32 t = *(u32 *)(a + (n -= 4));
62 		*(u32 *)(a + n) = *(u32 *)(b + n);
63 		*(u32 *)(b + n) = t;
64 	} while (n);
65 }
66 
67 /**
68  * swap_words_64 - swap two elements in 64-bit chunks
69  * @a: pointer to the first element to swap
70  * @b: pointer to the second element to swap
71  * @n: element size (must be a multiple of 8)
72  *
73  * Exchange the two objects in memory.  This exploits base+index
74  * addressing, which basically all CPUs have, to minimize loop overhead
75  * computations.
76  *
77  * We'd like to use 64-bit loads if possible.  If they're not, emulating
78  * one requires base+index+4 addressing which x86 has but most other
79  * processors do not.  If CONFIG_64BIT, we definitely have 64-bit loads,
80  * but it's possible to have 64-bit loads without 64-bit pointers (e.g.
81  * x32 ABI).  Are there any cases the kernel needs to worry about?
82  */
83 static void swap_words_64(void *a, void *b, size_t n)
84 {
85 	do {
86 #ifdef CONFIG_64BIT
87 		u64 t = *(u64 *)(a + (n -= 8));
88 		*(u64 *)(a + n) = *(u64 *)(b + n);
89 		*(u64 *)(b + n) = t;
90 #else
91 		/* Use two 32-bit transfers to avoid base+index+4 addressing */
92 		u32 t = *(u32 *)(a + (n -= 4));
93 		*(u32 *)(a + n) = *(u32 *)(b + n);
94 		*(u32 *)(b + n) = t;
95 
96 		t = *(u32 *)(a + (n -= 4));
97 		*(u32 *)(a + n) = *(u32 *)(b + n);
98 		*(u32 *)(b + n) = t;
99 #endif
100 	} while (n);
101 }
102 
103 /**
104  * swap_bytes - swap two elements a byte at a time
105  * @a: pointer to the first element to swap
106  * @b: pointer to the second element to swap
107  * @n: element size
108  *
109  * This is the fallback if alignment doesn't allow using larger chunks.
110  */
111 static void swap_bytes(void *a, void *b, size_t n)
112 {
113 	do {
114 		char t = ((char *)a)[--n];
115 		((char *)a)[n] = ((char *)b)[n];
116 		((char *)b)[n] = t;
117 	} while (n);
118 }
119 
120 typedef void (*swap_func_t)(void *a, void *b, int size);
121 
122 /*
123  * The values are arbitrary as long as they can't be confused with
124  * a pointer, but small integers make for the smallest compare
125  * instructions.
126  */
127 #define SWAP_WORDS_64 (swap_func_t)0
128 #define SWAP_WORDS_32 (swap_func_t)1
129 #define SWAP_BYTES    (swap_func_t)2
130 
131 /*
132  * The function pointer is last to make tail calls most efficient if the
133  * compiler decides not to inline this function.
134  */
135 static void do_swap(void *a, void *b, size_t size, swap_func_t swap_func)
136 {
137 	if (swap_func == SWAP_WORDS_64)
138 		swap_words_64(a, b, size);
139 	else if (swap_func == SWAP_WORDS_32)
140 		swap_words_32(a, b, size);
141 	else if (swap_func == SWAP_BYTES)
142 		swap_bytes(a, b, size);
143 	else
144 		swap_func(a, b, (int)size);
145 }
146 
147 /**
148  * parent - given the offset of the child, find the offset of the parent.
149  * @i: the offset of the heap element whose parent is sought.  Non-zero.
150  * @lsbit: a precomputed 1-bit mask, equal to "size & -size"
151  * @size: size of each element
152  *
153  * In terms of array indexes, the parent of element j = @i/@size is simply
154  * (j-1)/2.  But when working in byte offsets, we can't use implicit
155  * truncation of integer divides.
156  *
157  * Fortunately, we only need one bit of the quotient, not the full divide.
158  * @size has a least significant bit.  That bit will be clear if @i is
159  * an even multiple of @size, and set if it's an odd multiple.
160  *
161  * Logically, we're doing "if (i & lsbit) i -= size;", but since the
162  * branch is unpredictable, it's done with a bit of clever branch-free
163  * code instead.
164  */
165 __attribute_const__ __always_inline
166 static size_t parent(size_t i, unsigned int lsbit, size_t size)
167 {
168 	i -= size;
169 	i -= size & -(i & lsbit);
170 	return i / 2;
171 }
172 
173 /**
174  * sort - sort an array of elements
175  * @base: pointer to data to sort
176  * @num: number of elements
177  * @size: size of each element
178  * @cmp_func: pointer to comparison function
179  * @swap_func: pointer to swap function or NULL
180  *
181  * This function does a heapsort on the given array.  You may provide
182  * a swap_func function if you need to do something more than a memory
183  * copy (e.g. fix up pointers or auxiliary data), but the built-in swap
184  * avoids a slow retpoline and so is significantly faster.
185  *
186  * Sorting time is O(n log n) both on average and worst-case. While
187  * quicksort is slightly faster on average, it suffers from exploitable
188  * O(n*n) worst-case behavior and extra memory requirements that make
189  * it less suitable for kernel use.
190  */
191 void sort(void *base, size_t num, size_t size,
192 	  int (*cmp_func)(const void *, const void *),
193 	  void (*swap_func)(void *, void *, int size))
194 {
195 	/* pre-scale counters for performance */
196 	size_t n = num * size, a = (num/2) * size;
197 	const unsigned int lsbit = size & -size;  /* Used to find parent */
198 
199 	if (!a)		/* num < 2 || size == 0 */
200 		return;
201 
202 	if (!swap_func) {
203 		if (is_aligned(base, size, 8))
204 			swap_func = SWAP_WORDS_64;
205 		else if (is_aligned(base, size, 4))
206 			swap_func = SWAP_WORDS_32;
207 		else
208 			swap_func = SWAP_BYTES;
209 	}
210 
211 	/*
212 	 * Loop invariants:
213 	 * 1. elements [a,n) satisfy the heap property (compare greater than
214 	 *    all of their children),
215 	 * 2. elements [n,num*size) are sorted, and
216 	 * 3. a <= b <= c <= d <= n (whenever they are valid).
217 	 */
218 	for (;;) {
219 		size_t b, c, d;
220 
221 		if (a)			/* Building heap: sift down --a */
222 			a -= size;
223 		else if (n -= size)	/* Sorting: Extract root to --n */
224 			do_swap(base, base + n, size, swap_func);
225 		else			/* Sort complete */
226 			break;
227 
228 		/*
229 		 * Sift element at "a" down into heap.  This is the
230 		 * "bottom-up" variant, which significantly reduces
231 		 * calls to cmp_func(): we find the sift-down path all
232 		 * the way to the leaves (one compare per level), then
233 		 * backtrack to find where to insert the target element.
234 		 *
235 		 * Because elements tend to sift down close to the leaves,
236 		 * this uses fewer compares than doing two per level
237 		 * on the way down.  (A bit more than half as many on
238 		 * average, 3/4 worst-case.)
239 		 */
240 		for (b = a; c = 2*b + size, (d = c + size) < n;)
241 			b = cmp_func(base + c, base + d) >= 0 ? c : d;
242 		if (d == n)	/* Special case last leaf with no sibling */
243 			b = c;
244 
245 		/* Now backtrack from "b" to the correct location for "a" */
246 		while (b != a && cmp_func(base + a, base + b) >= 0)
247 			b = parent(b, lsbit, size);
248 		c = b;			/* Where "a" belongs */
249 		while (b != a) {	/* Shift it into place */
250 			b = parent(b, lsbit, size);
251 			do_swap(base + b, base + c, size, swap_func);
252 		}
253 	}
254 }
255 EXPORT_SYMBOL(sort);
256