Loading...
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
33static 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 still 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 */
58static 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 */
83static 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 */
111static 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/*
121 * The values are arbitrary as long as they can't be confused with
122 * a pointer, but small integers make for the smallest compare
123 * instructions.
124 */
125#define SWAP_WORDS_64 (swap_func_t)0
126#define SWAP_WORDS_32 (swap_func_t)1
127#define SWAP_BYTES (swap_func_t)2
128
129/*
130 * The function pointer is last to make tail calls most efficient if the
131 * compiler decides not to inline this function.
132 */
133static void do_swap(void *a, void *b, size_t size, swap_func_t swap_func)
134{
135 if (swap_func == SWAP_WORDS_64)
136 swap_words_64(a, b, size);
137 else if (swap_func == SWAP_WORDS_32)
138 swap_words_32(a, b, size);
139 else if (swap_func == SWAP_BYTES)
140 swap_bytes(a, b, size);
141 else
142 swap_func(a, b, (int)size);
143}
144
145#define _CMP_WRAPPER ((cmp_r_func_t)0L)
146
147static int do_cmp(const void *a, const void *b, cmp_r_func_t cmp, const void *priv)
148{
149 if (cmp == _CMP_WRAPPER)
150 return ((cmp_func_t)(priv))(a, b);
151 return cmp(a, b, priv);
152}
153
154/**
155 * parent - given the offset of the child, find the offset of the parent.
156 * @i: the offset of the heap element whose parent is sought. Non-zero.
157 * @lsbit: a precomputed 1-bit mask, equal to "size & -size"
158 * @size: size of each element
159 *
160 * In terms of array indexes, the parent of element j = @i/@size is simply
161 * (j-1)/2. But when working in byte offsets, we can't use implicit
162 * truncation of integer divides.
163 *
164 * Fortunately, we only need one bit of the quotient, not the full divide.
165 * @size has a least significant bit. That bit will be clear if @i is
166 * an even multiple of @size, and set if it's an odd multiple.
167 *
168 * Logically, we're doing "if (i & lsbit) i -= size;", but since the
169 * branch is unpredictable, it's done with a bit of clever branch-free
170 * code instead.
171 */
172__attribute_const__ __always_inline
173static size_t parent(size_t i, unsigned int lsbit, size_t size)
174{
175 i -= size;
176 i -= size & -(i & lsbit);
177 return i / 2;
178}
179
180/**
181 * sort_r - sort an array of elements
182 * @base: pointer to data to sort
183 * @num: number of elements
184 * @size: size of each element
185 * @cmp_func: pointer to comparison function
186 * @swap_func: pointer to swap function or NULL
187 * @priv: third argument passed to comparison function
188 *
189 * This function does a heapsort on the given array. You may provide
190 * a swap_func function if you need to do something more than a memory
191 * copy (e.g. fix up pointers or auxiliary data), but the built-in swap
192 * avoids a slow retpoline and so is significantly faster.
193 *
194 * Sorting time is O(n log n) both on average and worst-case. While
195 * quicksort is slightly faster on average, it suffers from exploitable
196 * O(n*n) worst-case behavior and extra memory requirements that make
197 * it less suitable for kernel use.
198 */
199void sort_r(void *base, size_t num, size_t size,
200 cmp_r_func_t cmp_func,
201 swap_func_t swap_func,
202 const void *priv)
203{
204 /* pre-scale counters for performance */
205 size_t n = num * size, a = (num/2) * size;
206 const unsigned int lsbit = size & -size; /* Used to find parent */
207
208 if (!a) /* num < 2 || size == 0 */
209 return;
210
211 if (!swap_func) {
212 if (is_aligned(base, size, 8))
213 swap_func = SWAP_WORDS_64;
214 else if (is_aligned(base, size, 4))
215 swap_func = SWAP_WORDS_32;
216 else
217 swap_func = SWAP_BYTES;
218 }
219
220 /*
221 * Loop invariants:
222 * 1. elements [a,n) satisfy the heap property (compare greater than
223 * all of their children),
224 * 2. elements [n,num*size) are sorted, and
225 * 3. a <= b <= c <= d <= n (whenever they are valid).
226 */
227 for (;;) {
228 size_t b, c, d;
229
230 if (a) /* Building heap: sift down --a */
231 a -= size;
232 else if (n -= size) /* Sorting: Extract root to --n */
233 do_swap(base, base + n, size, swap_func);
234 else /* Sort complete */
235 break;
236
237 /*
238 * Sift element at "a" down into heap. This is the
239 * "bottom-up" variant, which significantly reduces
240 * calls to cmp_func(): we find the sift-down path all
241 * the way to the leaves (one compare per level), then
242 * backtrack to find where to insert the target element.
243 *
244 * Because elements tend to sift down close to the leaves,
245 * this uses fewer compares than doing two per level
246 * on the way down. (A bit more than half as many on
247 * average, 3/4 worst-case.)
248 */
249 for (b = a; c = 2*b + size, (d = c + size) < n;)
250 b = do_cmp(base + c, base + d, cmp_func, priv) >= 0 ? c : d;
251 if (d == n) /* Special case last leaf with no sibling */
252 b = c;
253
254 /* Now backtrack from "b" to the correct location for "a" */
255 while (b != a && do_cmp(base + a, base + b, cmp_func, priv) >= 0)
256 b = parent(b, lsbit, size);
257 c = b; /* Where "a" belongs */
258 while (b != a) { /* Shift it into place */
259 b = parent(b, lsbit, size);
260 do_swap(base + b, base + c, size, swap_func);
261 }
262 }
263}
264EXPORT_SYMBOL(sort_r);
265
266void sort(void *base, size_t num, size_t size,
267 cmp_func_t cmp_func,
268 swap_func_t swap_func)
269{
270 return sort_r(base, num, size, _CMP_WRAPPER, swap_func, cmp_func);
271}
272EXPORT_SYMBOL(sort);
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 * Quicksort 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#include <linux/types.h>
14#include <linux/export.h>
15#include <linux/sort.h>
16
17/**
18 * is_aligned - is this pointer & size okay for word-wide copying?
19 * @base: pointer to data
20 * @size: size of each element
21 * @align: required alignment (typically 4 or 8)
22 *
23 * Returns true if elements can be copied using word loads and stores.
24 * The size must be a multiple of the alignment, and the base address must
25 * be if we do not have CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS.
26 *
27 * For some reason, gcc doesn't know to optimize "if (a & mask || b & mask)"
28 * to "if ((a | b) & mask)", so we do that by hand.
29 */
30__attribute_const__ __always_inline
31static bool is_aligned(const void *base, size_t size, unsigned char align)
32{
33 unsigned char lsbits = (unsigned char)size;
34
35 (void)base;
36#ifndef CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS
37 lsbits |= (unsigned char)(uintptr_t)base;
38#endif
39 return (lsbits & (align - 1)) == 0;
40}
41
42/**
43 * swap_words_32 - swap two elements in 32-bit chunks
44 * @a: pointer to the first element to swap
45 * @b: pointer to the second element to swap
46 * @n: element size (must be a multiple of 4)
47 *
48 * Exchange the two objects in memory. This exploits base+index addressing,
49 * which basically all CPUs have, to minimize loop overhead computations.
50 *
51 * For some reason, on x86 gcc 7.3.0 adds a redundant test of n at the
52 * bottom of the loop, even though the zero flag is still valid from the
53 * subtract (since the intervening mov instructions don't alter the flags).
54 * Gcc 8.1.0 doesn't have that problem.
55 */
56static void swap_words_32(void *a, void *b, size_t n)
57{
58 do {
59 u32 t = *(u32 *)(a + (n -= 4));
60 *(u32 *)(a + n) = *(u32 *)(b + n);
61 *(u32 *)(b + n) = t;
62 } while (n);
63}
64
65/**
66 * swap_words_64 - swap two elements in 64-bit chunks
67 * @a: pointer to the first element to swap
68 * @b: pointer to the second element to swap
69 * @n: element size (must be a multiple of 8)
70 *
71 * Exchange the two objects in memory. This exploits base+index
72 * addressing, which basically all CPUs have, to minimize loop overhead
73 * computations.
74 *
75 * We'd like to use 64-bit loads if possible. If they're not, emulating
76 * one requires base+index+4 addressing which x86 has but most other
77 * processors do not. If CONFIG_64BIT, we definitely have 64-bit loads,
78 * but it's possible to have 64-bit loads without 64-bit pointers (e.g.
79 * x32 ABI). Are there any cases the kernel needs to worry about?
80 */
81static void swap_words_64(void *a, void *b, size_t n)
82{
83 do {
84#ifdef CONFIG_64BIT
85 u64 t = *(u64 *)(a + (n -= 8));
86 *(u64 *)(a + n) = *(u64 *)(b + n);
87 *(u64 *)(b + n) = t;
88#else
89 /* Use two 32-bit transfers to avoid base+index+4 addressing */
90 u32 t = *(u32 *)(a + (n -= 4));
91 *(u32 *)(a + n) = *(u32 *)(b + n);
92 *(u32 *)(b + n) = t;
93
94 t = *(u32 *)(a + (n -= 4));
95 *(u32 *)(a + n) = *(u32 *)(b + n);
96 *(u32 *)(b + n) = t;
97#endif
98 } while (n);
99}
100
101/**
102 * swap_bytes - swap two elements a byte at a time
103 * @a: pointer to the first element to swap
104 * @b: pointer to the second element to swap
105 * @n: element size
106 *
107 * This is the fallback if alignment doesn't allow using larger chunks.
108 */
109static void swap_bytes(void *a, void *b, size_t n)
110{
111 do {
112 char t = ((char *)a)[--n];
113 ((char *)a)[n] = ((char *)b)[n];
114 ((char *)b)[n] = t;
115 } while (n);
116}
117
118/*
119 * The values are arbitrary as long as they can't be confused with
120 * a pointer, but small integers make for the smallest compare
121 * instructions.
122 */
123#define SWAP_WORDS_64 (swap_r_func_t)0
124#define SWAP_WORDS_32 (swap_r_func_t)1
125#define SWAP_BYTES (swap_r_func_t)2
126#define SWAP_WRAPPER (swap_r_func_t)3
127
128struct wrapper {
129 cmp_func_t cmp;
130 swap_func_t swap;
131};
132
133/*
134 * The function pointer is last to make tail calls most efficient if the
135 * compiler decides not to inline this function.
136 */
137static void do_swap(void *a, void *b, size_t size, swap_r_func_t swap_func, const void *priv)
138{
139 if (swap_func == SWAP_WRAPPER) {
140 ((const struct wrapper *)priv)->swap(a, b, (int)size);
141 return;
142 }
143
144 if (swap_func == SWAP_WORDS_64)
145 swap_words_64(a, b, size);
146 else if (swap_func == SWAP_WORDS_32)
147 swap_words_32(a, b, size);
148 else if (swap_func == SWAP_BYTES)
149 swap_bytes(a, b, size);
150 else
151 swap_func(a, b, (int)size, priv);
152}
153
154#define _CMP_WRAPPER ((cmp_r_func_t)0L)
155
156static int do_cmp(const void *a, const void *b, cmp_r_func_t cmp, const void *priv)
157{
158 if (cmp == _CMP_WRAPPER)
159 return ((const struct wrapper *)priv)->cmp(a, b);
160 return cmp(a, b, priv);
161}
162
163/**
164 * parent - given the offset of the child, find the offset of the parent.
165 * @i: the offset of the heap element whose parent is sought. Non-zero.
166 * @lsbit: a precomputed 1-bit mask, equal to "size & -size"
167 * @size: size of each element
168 *
169 * In terms of array indexes, the parent of element j = @i/@size is simply
170 * (j-1)/2. But when working in byte offsets, we can't use implicit
171 * truncation of integer divides.
172 *
173 * Fortunately, we only need one bit of the quotient, not the full divide.
174 * @size has a least significant bit. That bit will be clear if @i is
175 * an even multiple of @size, and set if it's an odd multiple.
176 *
177 * Logically, we're doing "if (i & lsbit) i -= size;", but since the
178 * branch is unpredictable, it's done with a bit of clever branch-free
179 * code instead.
180 */
181__attribute_const__ __always_inline
182static size_t parent(size_t i, unsigned int lsbit, size_t size)
183{
184 i -= size;
185 i -= size & -(i & lsbit);
186 return i / 2;
187}
188
189/**
190 * sort_r - sort an array of elements
191 * @base: pointer to data to sort
192 * @num: number of elements
193 * @size: size of each element
194 * @cmp_func: pointer to comparison function
195 * @swap_func: pointer to swap function or NULL
196 * @priv: third argument passed to comparison function
197 *
198 * This function does a heapsort on the given array. You may provide
199 * a swap_func function if you need to do something more than a memory
200 * copy (e.g. fix up pointers or auxiliary data), but the built-in swap
201 * avoids a slow retpoline and so is significantly faster.
202 *
203 * Sorting time is O(n log n) both on average and worst-case. While
204 * quicksort is slightly faster on average, it suffers from exploitable
205 * O(n*n) worst-case behavior and extra memory requirements that make
206 * it less suitable for kernel use.
207 */
208void sort_r(void *base, size_t num, size_t size,
209 cmp_r_func_t cmp_func,
210 swap_r_func_t swap_func,
211 const void *priv)
212{
213 /* pre-scale counters for performance */
214 size_t n = num * size, a = (num/2) * size;
215 const unsigned int lsbit = size & -size; /* Used to find parent */
216 size_t shift = 0;
217
218 if (!a) /* num < 2 || size == 0 */
219 return;
220
221 /* called from 'sort' without swap function, let's pick the default */
222 if (swap_func == SWAP_WRAPPER && !((struct wrapper *)priv)->swap)
223 swap_func = NULL;
224
225 if (!swap_func) {
226 if (is_aligned(base, size, 8))
227 swap_func = SWAP_WORDS_64;
228 else if (is_aligned(base, size, 4))
229 swap_func = SWAP_WORDS_32;
230 else
231 swap_func = SWAP_BYTES;
232 }
233
234 /*
235 * Loop invariants:
236 * 1. elements [a,n) satisfy the heap property (compare greater than
237 * all of their children),
238 * 2. elements [n,num*size) are sorted, and
239 * 3. a <= b <= c <= d <= n (whenever they are valid).
240 */
241 for (;;) {
242 size_t b, c, d;
243
244 if (a) /* Building heap: sift down a */
245 a -= size << shift;
246 else if (n > 3 * size) { /* Sorting: Extract two largest elements */
247 n -= size;
248 do_swap(base, base + n, size, swap_func, priv);
249 shift = do_cmp(base + size, base + 2 * size, cmp_func, priv) <= 0;
250 a = size << shift;
251 n -= size;
252 do_swap(base + a, base + n, size, swap_func, priv);
253 } else { /* Sort complete */
254 break;
255 }
256
257 /*
258 * Sift element at "a" down into heap. This is the
259 * "bottom-up" variant, which significantly reduces
260 * calls to cmp_func(): we find the sift-down path all
261 * the way to the leaves (one compare per level), then
262 * backtrack to find where to insert the target element.
263 *
264 * Because elements tend to sift down close to the leaves,
265 * this uses fewer compares than doing two per level
266 * on the way down. (A bit more than half as many on
267 * average, 3/4 worst-case.)
268 */
269 for (b = a; c = 2*b + size, (d = c + size) < n;)
270 b = do_cmp(base + c, base + d, cmp_func, priv) > 0 ? c : d;
271 if (d == n) /* Special case last leaf with no sibling */
272 b = c;
273
274 /* Now backtrack from "b" to the correct location for "a" */
275 while (b != a && do_cmp(base + a, base + b, cmp_func, priv) >= 0)
276 b = parent(b, lsbit, size);
277 c = b; /* Where "a" belongs */
278 while (b != a) { /* Shift it into place */
279 b = parent(b, lsbit, size);
280 do_swap(base + b, base + c, size, swap_func, priv);
281 }
282 }
283
284 n -= size;
285 do_swap(base, base + n, size, swap_func, priv);
286 if (n == size * 2 && do_cmp(base, base + size, cmp_func, priv) > 0)
287 do_swap(base, base + size, size, swap_func, priv);
288}
289EXPORT_SYMBOL(sort_r);
290
291void sort(void *base, size_t num, size_t size,
292 cmp_func_t cmp_func,
293 swap_func_t swap_func)
294{
295 struct wrapper w = {
296 .cmp = cmp_func,
297 .swap = swap_func,
298 };
299
300 return sort_r(base, num, size, _CMP_WRAPPER, SWAP_WRAPPER, &w);
301}
302EXPORT_SYMBOL(sort);