Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
3cbc5640 RT |
2 | /* |
3 | * Fast batching percpu counters. | |
4 | */ | |
5 | ||
6 | #include <linux/percpu_counter.h> | |
c67ad917 AM |
7 | #include <linux/mutex.h> |
8 | #include <linux/init.h> | |
9 | #include <linux/cpu.h> | |
3cbc5640 | 10 | #include <linux/module.h> |
e2852ae8 | 11 | #include <linux/debugobjects.h> |
3cbc5640 | 12 | |
3a8495c7 | 13 | #ifdef CONFIG_HOTPLUG_CPU |
c67ad917 | 14 | static LIST_HEAD(percpu_counters); |
d87aae2f | 15 | static DEFINE_SPINLOCK(percpu_counters_lock); |
3a8495c7 | 16 | #endif |
c67ad917 | 17 | |
e2852ae8 TH |
18 | #ifdef CONFIG_DEBUG_OBJECTS_PERCPU_COUNTER |
19 | ||
f9e62f31 | 20 | static const struct debug_obj_descr percpu_counter_debug_descr; |
e2852ae8 | 21 | |
d99b1d89 | 22 | static bool percpu_counter_fixup_free(void *addr, enum debug_obj_state state) |
e2852ae8 TH |
23 | { |
24 | struct percpu_counter *fbc = addr; | |
25 | ||
26 | switch (state) { | |
27 | case ODEBUG_STATE_ACTIVE: | |
28 | percpu_counter_destroy(fbc); | |
29 | debug_object_free(fbc, &percpu_counter_debug_descr); | |
d99b1d89 | 30 | return true; |
e2852ae8 | 31 | default: |
d99b1d89 | 32 | return false; |
e2852ae8 TH |
33 | } |
34 | } | |
35 | ||
f9e62f31 | 36 | static const struct debug_obj_descr percpu_counter_debug_descr = { |
e2852ae8 TH |
37 | .name = "percpu_counter", |
38 | .fixup_free = percpu_counter_fixup_free, | |
39 | }; | |
40 | ||
41 | static inline void debug_percpu_counter_activate(struct percpu_counter *fbc) | |
42 | { | |
43 | debug_object_init(fbc, &percpu_counter_debug_descr); | |
44 | debug_object_activate(fbc, &percpu_counter_debug_descr); | |
45 | } | |
46 | ||
47 | static inline void debug_percpu_counter_deactivate(struct percpu_counter *fbc) | |
48 | { | |
49 | debug_object_deactivate(fbc, &percpu_counter_debug_descr); | |
50 | debug_object_free(fbc, &percpu_counter_debug_descr); | |
51 | } | |
52 | ||
53 | #else /* CONFIG_DEBUG_OBJECTS_PERCPU_COUNTER */ | |
54 | static inline void debug_percpu_counter_activate(struct percpu_counter *fbc) | |
55 | { } | |
56 | static inline void debug_percpu_counter_deactivate(struct percpu_counter *fbc) | |
57 | { } | |
58 | #endif /* CONFIG_DEBUG_OBJECTS_PERCPU_COUNTER */ | |
59 | ||
3a587f47 PZ |
60 | void percpu_counter_set(struct percpu_counter *fbc, s64 amount) |
61 | { | |
62 | int cpu; | |
098faf58 | 63 | unsigned long flags; |
3a587f47 | 64 | |
098faf58 | 65 | raw_spin_lock_irqsave(&fbc->lock, flags); |
3a587f47 PZ |
66 | for_each_possible_cpu(cpu) { |
67 | s32 *pcount = per_cpu_ptr(fbc->counters, cpu); | |
68 | *pcount = 0; | |
69 | } | |
70 | fbc->count = amount; | |
098faf58 | 71 | raw_spin_unlock_irqrestore(&fbc->lock, flags); |
3a587f47 PZ |
72 | } |
73 | EXPORT_SYMBOL(percpu_counter_set); | |
74 | ||
db65a867 | 75 | /* |
3e8f399d NB |
76 | * This function is both preempt and irq safe. The former is due to explicit |
77 | * preemption disable. The latter is guaranteed by the fact that the slow path | |
78 | * is explicitly protected by an irq-safe spinlock whereas the fast patch uses | |
79 | * this_cpu_add which is irq-safe by definition. Hence there is no need muck | |
80 | * with irq state before calling this one | |
81 | */ | |
104b4e51 | 82 | void percpu_counter_add_batch(struct percpu_counter *fbc, s64 amount, s32 batch) |
3cbc5640 | 83 | { |
20e89767 | 84 | s64 count; |
3cbc5640 | 85 | |
ea00c30b | 86 | preempt_disable(); |
819a72af | 87 | count = __this_cpu_read(*fbc->counters) + amount; |
1d339638 | 88 | if (abs(count) >= batch) { |
098faf58 SL |
89 | unsigned long flags; |
90 | raw_spin_lock_irqsave(&fbc->lock, flags); | |
3cbc5640 | 91 | fbc->count += count; |
d1969a84 | 92 | __this_cpu_sub(*fbc->counters, count - amount); |
098faf58 | 93 | raw_spin_unlock_irqrestore(&fbc->lock, flags); |
3cbc5640 | 94 | } else { |
74e72f89 | 95 | this_cpu_add(*fbc->counters, amount); |
3cbc5640 | 96 | } |
ea00c30b | 97 | preempt_enable(); |
3cbc5640 | 98 | } |
104b4e51 | 99 | EXPORT_SYMBOL(percpu_counter_add_batch); |
3cbc5640 | 100 | |
0a4954a8 FT |
101 | /* |
102 | * For percpu_counter with a big batch, the devication of its count could | |
103 | * be big, and there is requirement to reduce the deviation, like when the | |
104 | * counter's batch could be runtime decreased to get a better accuracy, | |
105 | * which can be achieved by running this sync function on each CPU. | |
106 | */ | |
107 | void percpu_counter_sync(struct percpu_counter *fbc) | |
108 | { | |
109 | unsigned long flags; | |
110 | s64 count; | |
111 | ||
112 | raw_spin_lock_irqsave(&fbc->lock, flags); | |
113 | count = __this_cpu_read(*fbc->counters); | |
114 | fbc->count += count; | |
115 | __this_cpu_sub(*fbc->counters, count); | |
116 | raw_spin_unlock_irqrestore(&fbc->lock, flags); | |
117 | } | |
118 | EXPORT_SYMBOL(percpu_counter_sync); | |
119 | ||
3cbc5640 RT |
120 | /* |
121 | * Add up all the per-cpu counts, return the result. This is a more accurate | |
122 | * but much slower version of percpu_counter_read_positive() | |
123 | */ | |
02d21168 | 124 | s64 __percpu_counter_sum(struct percpu_counter *fbc) |
3cbc5640 | 125 | { |
0216bfcf | 126 | s64 ret; |
3cbc5640 | 127 | int cpu; |
098faf58 | 128 | unsigned long flags; |
3cbc5640 | 129 | |
098faf58 | 130 | raw_spin_lock_irqsave(&fbc->lock, flags); |
3cbc5640 | 131 | ret = fbc->count; |
b4ef0296 | 132 | for_each_online_cpu(cpu) { |
0216bfcf | 133 | s32 *pcount = per_cpu_ptr(fbc->counters, cpu); |
3cbc5640 RT |
134 | ret += *pcount; |
135 | } | |
098faf58 | 136 | raw_spin_unlock_irqrestore(&fbc->lock, flags); |
bf1d89c8 | 137 | return ret; |
3cbc5640 | 138 | } |
bf1d89c8 | 139 | EXPORT_SYMBOL(__percpu_counter_sum); |
c67ad917 | 140 | |
908c7f19 | 141 | int __percpu_counter_init(struct percpu_counter *fbc, s64 amount, gfp_t gfp, |
ea319518 | 142 | struct lock_class_key *key) |
c67ad917 | 143 | { |
ebd8fef3 TH |
144 | unsigned long flags __maybe_unused; |
145 | ||
f032a450 | 146 | raw_spin_lock_init(&fbc->lock); |
ea319518 | 147 | lockdep_set_class(&fbc->lock, key); |
c67ad917 | 148 | fbc->count = amount; |
908c7f19 | 149 | fbc->counters = alloc_percpu_gfp(s32, gfp); |
833f4077 PZ |
150 | if (!fbc->counters) |
151 | return -ENOMEM; | |
e2852ae8 TH |
152 | |
153 | debug_percpu_counter_activate(fbc); | |
154 | ||
c67ad917 | 155 | #ifdef CONFIG_HOTPLUG_CPU |
8474b591 | 156 | INIT_LIST_HEAD(&fbc->list); |
ebd8fef3 | 157 | spin_lock_irqsave(&percpu_counters_lock, flags); |
c67ad917 | 158 | list_add(&fbc->list, &percpu_counters); |
ebd8fef3 | 159 | spin_unlock_irqrestore(&percpu_counters_lock, flags); |
c67ad917 | 160 | #endif |
833f4077 | 161 | return 0; |
c67ad917 | 162 | } |
ea319518 | 163 | EXPORT_SYMBOL(__percpu_counter_init); |
c67ad917 AM |
164 | |
165 | void percpu_counter_destroy(struct percpu_counter *fbc) | |
166 | { | |
ebd8fef3 TH |
167 | unsigned long flags __maybe_unused; |
168 | ||
833f4077 PZ |
169 | if (!fbc->counters) |
170 | return; | |
171 | ||
e2852ae8 TH |
172 | debug_percpu_counter_deactivate(fbc); |
173 | ||
c67ad917 | 174 | #ifdef CONFIG_HOTPLUG_CPU |
ebd8fef3 | 175 | spin_lock_irqsave(&percpu_counters_lock, flags); |
c67ad917 | 176 | list_del(&fbc->list); |
ebd8fef3 | 177 | spin_unlock_irqrestore(&percpu_counters_lock, flags); |
c67ad917 | 178 | #endif |
fd3d664f ED |
179 | free_percpu(fbc->counters); |
180 | fbc->counters = NULL; | |
c67ad917 AM |
181 | } |
182 | EXPORT_SYMBOL(percpu_counter_destroy); | |
183 | ||
179f7ebf ED |
184 | int percpu_counter_batch __read_mostly = 32; |
185 | EXPORT_SYMBOL(percpu_counter_batch); | |
186 | ||
5588f5af | 187 | static int compute_batch_value(unsigned int cpu) |
179f7ebf ED |
188 | { |
189 | int nr = num_online_cpus(); | |
190 | ||
191 | percpu_counter_batch = max(32, nr*2); | |
5588f5af | 192 | return 0; |
179f7ebf ED |
193 | } |
194 | ||
5588f5af | 195 | static int percpu_counter_cpu_dead(unsigned int cpu) |
c67ad917 | 196 | { |
179f7ebf | 197 | #ifdef CONFIG_HOTPLUG_CPU |
c67ad917 AM |
198 | struct percpu_counter *fbc; |
199 | ||
5588f5af | 200 | compute_batch_value(cpu); |
c67ad917 | 201 | |
ebd8fef3 | 202 | spin_lock_irq(&percpu_counters_lock); |
c67ad917 AM |
203 | list_for_each_entry(fbc, &percpu_counters, list) { |
204 | s32 *pcount; | |
205 | ||
aaf0f2fa | 206 | raw_spin_lock(&fbc->lock); |
c67ad917 AM |
207 | pcount = per_cpu_ptr(fbc->counters, cpu); |
208 | fbc->count += *pcount; | |
209 | *pcount = 0; | |
aaf0f2fa | 210 | raw_spin_unlock(&fbc->lock); |
c67ad917 | 211 | } |
ebd8fef3 | 212 | spin_unlock_irq(&percpu_counters_lock); |
179f7ebf | 213 | #endif |
5588f5af | 214 | return 0; |
c67ad917 AM |
215 | } |
216 | ||
27f5e0f6 TC |
217 | /* |
218 | * Compare counter against given value. | |
219 | * Return 1 if greater, 0 if equal and -1 if less | |
220 | */ | |
80188b0d | 221 | int __percpu_counter_compare(struct percpu_counter *fbc, s64 rhs, s32 batch) |
27f5e0f6 TC |
222 | { |
223 | s64 count; | |
224 | ||
225 | count = percpu_counter_read(fbc); | |
226 | /* Check to see if rough count will be sufficient for comparison */ | |
80188b0d | 227 | if (abs(count - rhs) > (batch * num_online_cpus())) { |
27f5e0f6 TC |
228 | if (count > rhs) |
229 | return 1; | |
230 | else | |
231 | return -1; | |
232 | } | |
233 | /* Need to use precise count */ | |
234 | count = percpu_counter_sum(fbc); | |
235 | if (count > rhs) | |
236 | return 1; | |
237 | else if (count < rhs) | |
238 | return -1; | |
239 | else | |
240 | return 0; | |
241 | } | |
80188b0d | 242 | EXPORT_SYMBOL(__percpu_counter_compare); |
27f5e0f6 | 243 | |
c67ad917 AM |
244 | static int __init percpu_counter_startup(void) |
245 | { | |
5588f5af SAS |
246 | int ret; |
247 | ||
248 | ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "lib/percpu_cnt:online", | |
249 | compute_batch_value, NULL); | |
250 | WARN_ON(ret < 0); | |
251 | ret = cpuhp_setup_state_nocalls(CPUHP_PERCPU_CNT_DEAD, | |
252 | "lib/percpu_cnt:dead", NULL, | |
253 | percpu_counter_cpu_dead); | |
254 | WARN_ON(ret < 0); | |
c67ad917 AM |
255 | return 0; |
256 | } | |
257 | module_init(percpu_counter_startup); |