| 1 | |
| 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 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 |
|
| 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 |
|
| 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 | |
| 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_r_func_t)0
|
| 126 | #define SWAP_WORDS_32 (swap_r_func_t)1
|
| 127 | #define SWAP_BYTES (swap_r_func_t)2
|
| 128 | #define SWAP_WRAPPER (swap_r_func_t)3
|
| 129 |
|
| 130 | struct wrapper {
|
| 131 | cmp_func_t cmp;
|
| 132 | swap_func_t swap;
|
| 133 | };
|
| 134 |
|
| 135 | |
| 136 | * The function pointer is last to make tail calls most efficient if the
|
| 137 | * compiler decides not to inline this function.
|
| 138 |
|
| 139 | static void do_swap(void *a, void *b, size_t size, swap_r_func_t swap_func, const void *priv)
|
| 140 | {
|
| 141 | if (swap_func == SWAP_WRAPPER) {
|
| 142 | ((const struct wrapper *)priv)->swap(a, b, (int)size);
|
| 143 | return;
|
| 144 | }
|
| 145 |
|
| 146 | if (swap_func == SWAP_WORDS_64)
|
| 147 | swap_words_64(a, b, size);
|
| 148 | else if (swap_func == SWAP_WORDS_32)
|
| 149 | swap_words_32(a, b, size);
|
| 150 | else if (swap_func == SWAP_BYTES)
|
| 151 | swap_bytes(a, b, size);
|
| 152 | else
|
| 153 | swap_func(a, b, (int)size, priv);
|
| 154 | }
|
| 155 |
|
| 156 | #define _CMP_WRAPPER ((cmp_r_func_t)0L)
|
| 157 |
|
| 158 | static int do_cmp(const void *a, const void *b, cmp_r_func_t cmp, const void *priv)
|
| 159 | {
|
| 160 | if (cmp == _CMP_WRAPPER)
|
| 161 | return ((const struct wrapper *)priv)->cmp(a, b);
|
| 162 | return cmp(a, b, priv);
|
| 163 | }
|
| 164 |
|
| 165 | |
| 166 | * parent - given the offset of the child, find the offset of the parent.
|
| 167 | * @i: the offset of the heap element whose parent is sought. Non-zero.
|
| 168 | * @lsbit: a precomputed 1-bit mask, equal to "size & -size"
|
| 169 | * @size: size of each element
|
| 170 | *
|
| 171 | * In terms of array indexes, the parent of element j = @i/@size is simply
|
| 172 | * (j-1)/2. But when working in byte offsets, we can't use implicit
|
| 173 | * truncation of integer divides.
|
| 174 | *
|
| 175 | * Fortunately, we only need one bit of the quotient, not the full divide.
|
| 176 | * @size has a least significant bit. That bit will be clear if @i is
|
| 177 | * an even multiple of @size, and set if it's an odd multiple.
|
| 178 | *
|
| 179 | * Logically, we're doing "if (i & lsbit) i -= size;", but since the
|
| 180 | * branch is unpredictable, it's done with a bit of clever branch-free
|
| 181 | * code instead.
|
| 182 |
|
| 183 | __attribute_const__ __always_inline
|
| 184 | static size_t parent(size_t i, unsigned int lsbit, size_t size)
|
| 185 | {
|
| 186 | i -= size;
|
| 187 | i -= size & -(i & lsbit);
|
| 188 | return i / 2;
|
| 189 | }
|
| 190 |
|
| 191 | |
| 192 | * sort_r - sort an array of elements
|
| 193 | * @base: pointer to data to sort
|
| 194 | * @num: number of elements
|
| 195 | * @size: size of each element
|
| 196 | * @cmp_func: pointer to comparison function
|
| 197 | * @swap_func: pointer to swap function or NULL
|
| 198 | * @priv: third argument passed to comparison function
|
| 199 | *
|
| 200 | * This function does a heapsort on the given array. You may provide
|
| 201 | * a swap_func function if you need to do something more than a memory
|
| 202 | * copy (e.g. fix up pointers or auxiliary data), but the built-in swap
|
| 203 | * avoids a slow retpoline and so is significantly faster.
|
| 204 | *
|
| 205 | * Sorting time is O(n log n) both on average and worst-case. While
|
| 206 | * quicksort is slightly faster on average, it suffers from exploitable
|
| 207 | * O(n*n) worst-case behavior and extra memory requirements that make
|
| 208 | * it less suitable for kernel use.
|
| 209 |
|
| 210 | void sort_r(void *base, size_t num, size_t size,
|
| 211 | cmp_r_func_t cmp_func,
|
| 212 | swap_r_func_t swap_func,
|
| 213 | const void *priv)
|
| 214 | {
|
| 215 |
|
| 216 | size_t n = num * size, a = (num/2) * size;
|
| 217 | const unsigned int lsbit = size & -size;
|
| 218 |
|
| 219 | if (!a)
|
| 220 | return;
|
| 221 |
|
| 222 |
|
| 223 | if (swap_func == SWAP_WRAPPER && !((struct wrapper *)priv)->swap)
|
| 224 | swap_func = NULL;
|
| 225 |
|
| 226 | if (!swap_func) {
|
| 227 | if (is_aligned(base, size, 8))
|
| 228 | swap_func = SWAP_WORDS_64;
|
| 229 | else if (is_aligned(base, size, 4))
|
| 230 | swap_func = SWAP_WORDS_32;
|
| 231 | else
|
| 232 | swap_func = SWAP_BYTES;
|
| 233 | }
|
| 234 |
|
| 235 | |
| 236 | * Loop invariants:
|
| 237 | * 1. elements [a,n) satisfy the heap property (compare greater than
|
| 238 | * all of their children),
|
| 239 | * 2. elements [n,num*size) are sorted, and
|
| 240 | * 3. a <= b <= c <= d <= n (whenever they are valid).
|
| 241 |
|
| 242 | for (;;) {
|
| 243 | size_t b, c, d;
|
| 244 |
|
| 245 | if (a)
|
| 246 | a -= size;
|
| 247 | else if (n -= size)
|
| 248 | do_swap(base, base + n, size, swap_func, priv);
|
| 249 | else
|
| 250 | break;
|
| 251 |
|
| 252 | |
| 253 | * Sift element at "a" down into heap. This is the
|
| 254 | * "bottom-up" variant, which significantly reduces
|
| 255 | * calls to cmp_func(): we find the sift-down path all
|
| 256 | * the way to the leaves (one compare per level), then
|
| 257 | * backtrack to find where to insert the target element.
|
| 258 | *
|
| 259 | * Because elements tend to sift down close to the leaves,
|
| 260 | * this uses fewer compares than doing two per level
|
| 261 | * on the way down. (A bit more than half as many on
|
| 262 | * average, 3/4 worst-case.)
|
| 263 |
|
| 264 | for (b = a; c = 2*b + size, (d = c + size) < n;)
|
| 265 | b = do_cmp(base + c, base + d, cmp_func, priv) >= 0 ? c : d;
|
| 266 | if (d == n)
|
| 267 | b = c;
|
| 268 |
|
| 269 |
|
| 270 | while (b != a && do_cmp(base + a, base + b, cmp_func, priv) >= 0)
|
| 271 | b = parent(b, lsbit, size);
|
| 272 | c = b;
|
| 273 | while (b != a) {
|
| 274 | b = parent(b, lsbit, size);
|
| 275 | do_swap(base + b, base + c, size, swap_func, priv);
|
| 276 | }
|
| 277 | }
|
| 278 | }
|
| 279 | EXPORT_SYMBOL(sort_r);
|
| 280 |
|
| 281 | void sort(void *base, size_t num, size_t size,
|
| 282 | cmp_func_t cmp_func,
|
| 283 | swap_func_t swap_func)
|
| 284 | {
|
| 285 | struct wrapper w = {
|
| 286 | .cmp = cmp_func,
|
| 287 | .swap = swap_func,
|
| 288 | };
|
| 289 |
|
| 290 | return sort_r(base, num, size, _CMP_WRAPPER, SWAP_WRAPPER, &w);
|
| 291 | }
|
| 292 | EXPORT_SYMBOL(sort);
|