1*4882a593Smuzhiyun // SPDX-License-Identifier: GPL-2.0-only
2*4882a593Smuzhiyun /*
3*4882a593Smuzhiyun * benchmark.c:
4*4882a593Smuzhiyun * Author: Konstantin Khlebnikov <koct9i@gmail.com>
5*4882a593Smuzhiyun */
6*4882a593Smuzhiyun #include <linux/radix-tree.h>
7*4882a593Smuzhiyun #include <linux/slab.h>
8*4882a593Smuzhiyun #include <linux/errno.h>
9*4882a593Smuzhiyun #include <time.h>
10*4882a593Smuzhiyun #include "test.h"
11*4882a593Smuzhiyun
12*4882a593Smuzhiyun #define NSEC_PER_SEC 1000000000L
13*4882a593Smuzhiyun
benchmark_iter(struct radix_tree_root * root,bool tagged)14*4882a593Smuzhiyun static long long benchmark_iter(struct radix_tree_root *root, bool tagged)
15*4882a593Smuzhiyun {
16*4882a593Smuzhiyun volatile unsigned long sink = 0;
17*4882a593Smuzhiyun struct radix_tree_iter iter;
18*4882a593Smuzhiyun struct timespec start, finish;
19*4882a593Smuzhiyun long long nsec;
20*4882a593Smuzhiyun int l, loops = 1;
21*4882a593Smuzhiyun void **slot;
22*4882a593Smuzhiyun
23*4882a593Smuzhiyun #ifdef BENCHMARK
24*4882a593Smuzhiyun again:
25*4882a593Smuzhiyun #endif
26*4882a593Smuzhiyun clock_gettime(CLOCK_MONOTONIC, &start);
27*4882a593Smuzhiyun for (l = 0; l < loops; l++) {
28*4882a593Smuzhiyun if (tagged) {
29*4882a593Smuzhiyun radix_tree_for_each_tagged(slot, root, &iter, 0, 0)
30*4882a593Smuzhiyun sink ^= (unsigned long)slot;
31*4882a593Smuzhiyun } else {
32*4882a593Smuzhiyun radix_tree_for_each_slot(slot, root, &iter, 0)
33*4882a593Smuzhiyun sink ^= (unsigned long)slot;
34*4882a593Smuzhiyun }
35*4882a593Smuzhiyun }
36*4882a593Smuzhiyun clock_gettime(CLOCK_MONOTONIC, &finish);
37*4882a593Smuzhiyun
38*4882a593Smuzhiyun nsec = (finish.tv_sec - start.tv_sec) * NSEC_PER_SEC +
39*4882a593Smuzhiyun (finish.tv_nsec - start.tv_nsec);
40*4882a593Smuzhiyun
41*4882a593Smuzhiyun #ifdef BENCHMARK
42*4882a593Smuzhiyun if (loops == 1 && nsec * 5 < NSEC_PER_SEC) {
43*4882a593Smuzhiyun loops = NSEC_PER_SEC / nsec / 4 + 1;
44*4882a593Smuzhiyun goto again;
45*4882a593Smuzhiyun }
46*4882a593Smuzhiyun #endif
47*4882a593Smuzhiyun
48*4882a593Smuzhiyun nsec /= loops;
49*4882a593Smuzhiyun return nsec;
50*4882a593Smuzhiyun }
51*4882a593Smuzhiyun
benchmark_insert(struct radix_tree_root * root,unsigned long size,unsigned long step)52*4882a593Smuzhiyun static void benchmark_insert(struct radix_tree_root *root,
53*4882a593Smuzhiyun unsigned long size, unsigned long step)
54*4882a593Smuzhiyun {
55*4882a593Smuzhiyun struct timespec start, finish;
56*4882a593Smuzhiyun unsigned long index;
57*4882a593Smuzhiyun long long nsec;
58*4882a593Smuzhiyun
59*4882a593Smuzhiyun clock_gettime(CLOCK_MONOTONIC, &start);
60*4882a593Smuzhiyun
61*4882a593Smuzhiyun for (index = 0 ; index < size ; index += step)
62*4882a593Smuzhiyun item_insert(root, index);
63*4882a593Smuzhiyun
64*4882a593Smuzhiyun clock_gettime(CLOCK_MONOTONIC, &finish);
65*4882a593Smuzhiyun
66*4882a593Smuzhiyun nsec = (finish.tv_sec - start.tv_sec) * NSEC_PER_SEC +
67*4882a593Smuzhiyun (finish.tv_nsec - start.tv_nsec);
68*4882a593Smuzhiyun
69*4882a593Smuzhiyun printv(2, "Size: %8ld, step: %8ld, insertion: %15lld ns\n",
70*4882a593Smuzhiyun size, step, nsec);
71*4882a593Smuzhiyun }
72*4882a593Smuzhiyun
benchmark_tagging(struct radix_tree_root * root,unsigned long size,unsigned long step)73*4882a593Smuzhiyun static void benchmark_tagging(struct radix_tree_root *root,
74*4882a593Smuzhiyun unsigned long size, unsigned long step)
75*4882a593Smuzhiyun {
76*4882a593Smuzhiyun struct timespec start, finish;
77*4882a593Smuzhiyun unsigned long index;
78*4882a593Smuzhiyun long long nsec;
79*4882a593Smuzhiyun
80*4882a593Smuzhiyun clock_gettime(CLOCK_MONOTONIC, &start);
81*4882a593Smuzhiyun
82*4882a593Smuzhiyun for (index = 0 ; index < size ; index += step)
83*4882a593Smuzhiyun radix_tree_tag_set(root, index, 0);
84*4882a593Smuzhiyun
85*4882a593Smuzhiyun clock_gettime(CLOCK_MONOTONIC, &finish);
86*4882a593Smuzhiyun
87*4882a593Smuzhiyun nsec = (finish.tv_sec - start.tv_sec) * NSEC_PER_SEC +
88*4882a593Smuzhiyun (finish.tv_nsec - start.tv_nsec);
89*4882a593Smuzhiyun
90*4882a593Smuzhiyun printv(2, "Size: %8ld, step: %8ld, tagging: %17lld ns\n",
91*4882a593Smuzhiyun size, step, nsec);
92*4882a593Smuzhiyun }
93*4882a593Smuzhiyun
benchmark_delete(struct radix_tree_root * root,unsigned long size,unsigned long step)94*4882a593Smuzhiyun static void benchmark_delete(struct radix_tree_root *root,
95*4882a593Smuzhiyun unsigned long size, unsigned long step)
96*4882a593Smuzhiyun {
97*4882a593Smuzhiyun struct timespec start, finish;
98*4882a593Smuzhiyun unsigned long index;
99*4882a593Smuzhiyun long long nsec;
100*4882a593Smuzhiyun
101*4882a593Smuzhiyun clock_gettime(CLOCK_MONOTONIC, &start);
102*4882a593Smuzhiyun
103*4882a593Smuzhiyun for (index = 0 ; index < size ; index += step)
104*4882a593Smuzhiyun item_delete(root, index);
105*4882a593Smuzhiyun
106*4882a593Smuzhiyun clock_gettime(CLOCK_MONOTONIC, &finish);
107*4882a593Smuzhiyun
108*4882a593Smuzhiyun nsec = (finish.tv_sec - start.tv_sec) * NSEC_PER_SEC +
109*4882a593Smuzhiyun (finish.tv_nsec - start.tv_nsec);
110*4882a593Smuzhiyun
111*4882a593Smuzhiyun printv(2, "Size: %8ld, step: %8ld, deletion: %16lld ns\n",
112*4882a593Smuzhiyun size, step, nsec);
113*4882a593Smuzhiyun }
114*4882a593Smuzhiyun
benchmark_size(unsigned long size,unsigned long step)115*4882a593Smuzhiyun static void benchmark_size(unsigned long size, unsigned long step)
116*4882a593Smuzhiyun {
117*4882a593Smuzhiyun RADIX_TREE(tree, GFP_KERNEL);
118*4882a593Smuzhiyun long long normal, tagged;
119*4882a593Smuzhiyun
120*4882a593Smuzhiyun benchmark_insert(&tree, size, step);
121*4882a593Smuzhiyun benchmark_tagging(&tree, size, step);
122*4882a593Smuzhiyun
123*4882a593Smuzhiyun tagged = benchmark_iter(&tree, true);
124*4882a593Smuzhiyun normal = benchmark_iter(&tree, false);
125*4882a593Smuzhiyun
126*4882a593Smuzhiyun printv(2, "Size: %8ld, step: %8ld, tagged iteration: %8lld ns\n",
127*4882a593Smuzhiyun size, step, tagged);
128*4882a593Smuzhiyun printv(2, "Size: %8ld, step: %8ld, normal iteration: %8lld ns\n",
129*4882a593Smuzhiyun size, step, normal);
130*4882a593Smuzhiyun
131*4882a593Smuzhiyun benchmark_delete(&tree, size, step);
132*4882a593Smuzhiyun
133*4882a593Smuzhiyun item_kill_tree(&tree);
134*4882a593Smuzhiyun rcu_barrier();
135*4882a593Smuzhiyun }
136*4882a593Smuzhiyun
benchmark(void)137*4882a593Smuzhiyun void benchmark(void)
138*4882a593Smuzhiyun {
139*4882a593Smuzhiyun unsigned long size[] = {1 << 10, 1 << 20, 0};
140*4882a593Smuzhiyun unsigned long step[] = {1, 2, 7, 15, 63, 64, 65,
141*4882a593Smuzhiyun 128, 256, 512, 12345, 0};
142*4882a593Smuzhiyun int c, s;
143*4882a593Smuzhiyun
144*4882a593Smuzhiyun printv(1, "starting benchmarks\n");
145*4882a593Smuzhiyun printv(1, "RADIX_TREE_MAP_SHIFT = %d\n", RADIX_TREE_MAP_SHIFT);
146*4882a593Smuzhiyun
147*4882a593Smuzhiyun for (c = 0; size[c]; c++)
148*4882a593Smuzhiyun for (s = 0; step[s]; s++)
149*4882a593Smuzhiyun benchmark_size(size[c], step[s]);
150*4882a593Smuzhiyun }
151