f7d26a17 |
1 | /* |
2 | * test-read-lock-speed.c |
3343768e |
3 | * |
6087c592 |
4 | * Compare speed of : |
1b485a14 |
5 | * - spin lock / spin unlock |
fe207500 |
6 | * - rwlock read lock |
6087c592 |
7 | * - using a sequence read lock (uncontended) |
f7d26a17 |
8 | * - preempt disable/enable (RCU) |
9 | * |
10 | * Copyright 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca> |
11 | * Distributed under GPLv2 |
3343768e |
12 | */ |
13 | |
3343768e |
14 | #include <linux/jiffies.h> |
15 | #include <linux/compiler.h> |
16 | #include <linux/init.h> |
17 | #include <linux/module.h> |
6087c592 |
18 | #include <linux/math64.h> |
8593c315 |
19 | #include <linux/spinlock.h> |
20 | #include <linux/seqlock.h> |
6087c592 |
21 | #include <linux/cpumask.h> |
3343768e |
22 | #include <asm/timex.h> |
23 | #include <asm/system.h> |
24 | |
9c1292db |
25 | #ifdef CONFIG_ARM |
26 | #include <linux/trace-clock.h> |
27 | #define get_timestamp trace_clock_read64 |
28 | #else |
29 | #define get_timestamp get_cycles |
30 | #endif |
31 | |
3343768e |
32 | #define NR_LOOPS 20000 |
33 | |
6087c592 |
34 | #ifndef CONFIG_PREEMPT |
28289207 |
35 | #error "Your kernel should be built with preemption enabled" |
6087c592 |
36 | #endif |
37 | |
38 | #ifdef CONFIG_DEBUG_PREEMPT |
39 | #error "Please disable CONFIG_DEBUG_PREEMPT" |
40 | #endif |
41 | |
42 | #ifdef CONFIG_DEBUG_SPINLOCK |
43 | #error "Please disable CONFIG_DEBUG_SPINLOCK" |
44 | #endif |
45 | |
46 | #ifdef CONFIG_LOCKDEP |
47 | #error "Please disable CONFIG_LOCKDEP" |
48 | #endif |
49 | |
3343768e |
50 | int test_val; |
51 | |
52 | static void do_testbaseline(void) |
53 | { |
6087c592 |
54 | unsigned long flags; |
3343768e |
55 | unsigned int i; |
56 | cycles_t time1, time2, time; |
6087c592 |
57 | u32 rem; |
3343768e |
58 | |
59 | local_irq_save(flags); |
60 | preempt_disable(); |
9c1292db |
61 | time1 = get_timestamp(); |
3343768e |
62 | for (i = 0; i < NR_LOOPS; i++) { |
63 | asm volatile (""); |
64 | } |
9c1292db |
65 | time2 = get_timestamp(); |
3343768e |
66 | local_irq_restore(flags); |
67 | preempt_enable(); |
68 | time = time2 - time1; |
69 | |
70 | printk(KERN_ALERT "test results: time for baseline\n"); |
71 | printk(KERN_ALERT "number of loops: %d\n", NR_LOOPS); |
bc918f0f |
72 | printk(KERN_ALERT "total time: %llu\n", (unsigned long long)time); |
6087c592 |
73 | time = div_u64_rem(time, NR_LOOPS, &rem); |
bc918f0f |
74 | printk(KERN_ALERT "-> baseline takes %llu cycles\n", (unsigned long long)time); |
3343768e |
75 | printk(KERN_ALERT "test end\n"); |
76 | } |
77 | |
78 | static void do_test_spinlock(void) |
79 | { |
8593c315 |
80 | static DEFINE_SPINLOCK(mylock); |
6087c592 |
81 | unsigned long flags; |
3343768e |
82 | unsigned int i; |
83 | cycles_t time1, time2, time; |
6087c592 |
84 | u32 rem; |
3343768e |
85 | |
86 | preempt_disable(); |
8593c315 |
87 | spin_lock_irqsave(&mylock, flags); |
9c1292db |
88 | time1 = get_timestamp(); |
3343768e |
89 | for (i = 0; i < NR_LOOPS; i++) { |
0c12e051 |
90 | spin_unlock(&mylock); |
91 | spin_lock(&mylock); |
3343768e |
92 | } |
9c1292db |
93 | time2 = get_timestamp(); |
8593c315 |
94 | spin_unlock_irqrestore(&mylock, flags); |
3343768e |
95 | preempt_enable(); |
96 | time = time2 - time1; |
97 | |
98 | printk(KERN_ALERT "test results: time for spinlock\n"); |
99 | printk(KERN_ALERT "number of loops: %d\n", NR_LOOPS); |
bc918f0f |
100 | printk(KERN_ALERT "total time: %llu\n", (unsigned long long)time); |
6087c592 |
101 | time = div_u64_rem(time, NR_LOOPS, &rem); |
bc918f0f |
102 | printk(KERN_ALERT "-> spinlock takes %llu cycles\n", (unsigned long long)time); |
3343768e |
103 | printk(KERN_ALERT "test end\n"); |
104 | } |
105 | |
6bc6dad3 |
106 | static void do_test_read_rwlock(void) |
107 | { |
108 | static DEFINE_RWLOCK(mylock); |
109 | unsigned long flags; |
110 | unsigned int i; |
111 | cycles_t time1, time2, time; |
112 | u32 rem; |
113 | |
114 | preempt_disable(); |
115 | local_irq_save(flags); |
116 | read_lock(&mylock); |
9c1292db |
117 | time1 = get_timestamp(); |
6bc6dad3 |
118 | for (i = 0; i < NR_LOOPS; i++) { |
119 | read_unlock(&mylock); |
120 | read_lock(&mylock); |
121 | } |
9c1292db |
122 | time2 = get_timestamp(); |
6bc6dad3 |
123 | read_unlock(&mylock); |
124 | local_irq_restore(flags); |
125 | preempt_enable(); |
126 | time = time2 - time1; |
127 | |
128 | printk(KERN_ALERT "test results: time for read rwlock\n"); |
129 | printk(KERN_ALERT "number of loops: %d\n", NR_LOOPS); |
bc918f0f |
130 | printk(KERN_ALERT "total time: %llu\n", (unsigned long long)time); |
6bc6dad3 |
131 | time = div_u64_rem(time, NR_LOOPS, &rem); |
bc918f0f |
132 | printk(KERN_ALERT "-> read rwlock takes %llu cycles\n", (unsigned long long)time); |
6bc6dad3 |
133 | printk(KERN_ALERT "test end\n"); |
134 | } |
135 | |
3343768e |
136 | static void do_test_seqlock(void) |
137 | { |
138 | static seqlock_t test_lock; |
8593c315 |
139 | unsigned long seq; |
6087c592 |
140 | unsigned long flags; |
3343768e |
141 | unsigned int i; |
142 | cycles_t time1, time2, time; |
6087c592 |
143 | u32 rem; |
3343768e |
144 | |
472e3e15 |
145 | local_irq_save(flags); |
9c1292db |
146 | time1 = get_timestamp(); |
3343768e |
147 | for (i = 0; i < NR_LOOPS; i++) { |
148 | do { |
472e3e15 |
149 | seq = read_seqbegin(&test_lock); |
150 | } while (read_seqretry(&test_lock, seq)); |
3343768e |
151 | } |
9c1292db |
152 | time2 = get_timestamp(); |
3343768e |
153 | time = time2 - time1; |
472e3e15 |
154 | local_irq_restore(flags); |
3343768e |
155 | |
156 | printk(KERN_ALERT "test results: time for seqlock\n"); |
157 | printk(KERN_ALERT "number of loops: %d\n", NR_LOOPS); |
bc918f0f |
158 | printk(KERN_ALERT "total time: %llu\n", (unsigned long long)time); |
6087c592 |
159 | time = div_u64_rem(time, NR_LOOPS, &rem); |
bc918f0f |
160 | printk(KERN_ALERT "-> seqlock takes %llu cycles\n", (unsigned long long)time); |
3343768e |
161 | printk(KERN_ALERT "test end\n"); |
162 | } |
163 | |
164 | /* |
6087c592 |
165 | * Note : This test _should_ trigger lockdep errors due to preemption |
166 | * disabling/enabling within irq off section. Given we are only interested in |
167 | * having the most precise measurement for preemption disable/enable, we don't |
168 | * care about this. |
3343768e |
169 | */ |
170 | static void do_test_preempt(void) |
171 | { |
6087c592 |
172 | unsigned long flags; |
3343768e |
173 | unsigned int i; |
174 | cycles_t time1, time2, time; |
6087c592 |
175 | u32 rem; |
3343768e |
176 | |
177 | local_irq_save(flags); |
178 | preempt_disable(); |
9c1292db |
179 | time1 = get_timestamp(); |
3343768e |
180 | for (i = 0; i < NR_LOOPS; i++) { |
3343768e |
181 | preempt_disable(); |
54e7c224 |
182 | preempt_enable(); |
3343768e |
183 | } |
9c1292db |
184 | time2 = get_timestamp(); |
3343768e |
185 | preempt_enable(); |
186 | time = time2 - time1; |
187 | local_irq_restore(flags); |
188 | |
f7d26a17 |
189 | printk(KERN_ALERT |
190 | "test results: time for preempt disable/enable pairs\n"); |
3343768e |
191 | printk(KERN_ALERT "number of loops: %d\n", NR_LOOPS); |
bc918f0f |
192 | printk(KERN_ALERT "total time: %llu\n", (unsigned long long)time); |
6087c592 |
193 | time = div_u64_rem(time, NR_LOOPS, &rem); |
3343768e |
194 | printk(KERN_ALERT "-> preempt disable/enable pair takes %llu cycles\n", |
bc918f0f |
195 | (unsigned long long)time); |
3343768e |
196 | printk(KERN_ALERT "test end\n"); |
197 | } |
198 | |
199 | static int ltt_test_init(void) |
200 | { |
201 | printk(KERN_ALERT "test init\n"); |
202 | |
6087c592 |
203 | printk(KERN_ALERT "Number of active CPUs : %d\n", num_online_cpus()); |
3343768e |
204 | do_testbaseline(); |
205 | do_test_spinlock(); |
6bc6dad3 |
206 | do_test_read_rwlock(); |
3343768e |
207 | do_test_seqlock(); |
208 | do_test_preempt(); |
209 | return -EAGAIN; /* Fail will directly unload the module */ |
210 | } |
211 | |
212 | static void ltt_test_exit(void) |
213 | { |
214 | printk(KERN_ALERT "test exit\n"); |
215 | } |
216 | |
217 | module_init(ltt_test_init) |
218 | module_exit(ltt_test_exit) |
219 | |
220 | MODULE_LICENSE("GPL"); |
221 | MODULE_AUTHOR("Mathieu Desnoyers"); |
f7d26a17 |
222 | MODULE_DESCRIPTION("Test read lock speed"); |