]>
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/notifier.h> |
8 | #include <linux/mutex.h> | |
9 | #include <linux/init.h> | |
10 | #include <linux/cpu.h> | |
3cbc5640 | 11 | #include <linux/module.h> |
e2852ae8 | 12 | #include <linux/debugobjects.h> |
3cbc5640 | 13 | |
3a8495c7 | 14 | #ifdef CONFIG_HOTPLUG_CPU |
c67ad917 | 15 | static LIST_HEAD(percpu_counters); |
d87aae2f | 16 | static DEFINE_SPINLOCK(percpu_counters_lock); |
3a8495c7 | 17 | #endif |
c67ad917 | 18 | |
e2852ae8 TH |
19 | #ifdef CONFIG_DEBUG_OBJECTS_PERCPU_COUNTER |
20 | ||
21 | static struct debug_obj_descr percpu_counter_debug_descr; | |
22 | ||
d99b1d89 | 23 | static bool percpu_counter_fixup_free(void *addr, enum debug_obj_state state) |
e2852ae8 TH |
24 | { |
25 | struct percpu_counter *fbc = addr; | |
26 | ||
27 | switch (state) { | |
28 | case ODEBUG_STATE_ACTIVE: | |
29 | percpu_counter_destroy(fbc); | |
30 | debug_object_free(fbc, &percpu_counter_debug_descr); | |
d99b1d89 | 31 | return true; |
e2852ae8 | 32 | default: |
d99b1d89 | 33 | return false; |
e2852ae8 TH |
34 | } |
35 | } | |
36 | ||
37 | static struct debug_obj_descr percpu_counter_debug_descr = { | |
38 | .name = "percpu_counter", | |
39 | .fixup_free = percpu_counter_fixup_free, | |
40 | }; | |
41 | ||
42 | static inline void debug_percpu_counter_activate(struct percpu_counter *fbc) | |
43 | { | |
44 | debug_object_init(fbc, &percpu_counter_debug_descr); | |
45 | debug_object_activate(fbc, &percpu_counter_debug_descr); | |
46 | } | |
47 | ||
48 | static inline void debug_percpu_counter_deactivate(struct percpu_counter *fbc) | |
49 | { | |
50 | debug_object_deactivate(fbc, &percpu_counter_debug_descr); | |
51 | debug_object_free(fbc, &percpu_counter_debug_descr); | |
52 | } | |
53 | ||
54 | #else /* CONFIG_DEBUG_OBJECTS_PERCPU_COUNTER */ | |
55 | static inline void debug_percpu_counter_activate(struct percpu_counter *fbc) | |
56 | { } | |
57 | static inline void debug_percpu_counter_deactivate(struct percpu_counter *fbc) | |
58 | { } | |
59 | #endif /* CONFIG_DEBUG_OBJECTS_PERCPU_COUNTER */ | |
60 | ||
3a587f47 PZ |
61 | void percpu_counter_set(struct percpu_counter *fbc, s64 amount) |
62 | { | |
63 | int cpu; | |
098faf58 | 64 | unsigned long flags; |
3a587f47 | 65 | |
098faf58 | 66 | raw_spin_lock_irqsave(&fbc->lock, flags); |
3a587f47 PZ |
67 | for_each_possible_cpu(cpu) { |
68 | s32 *pcount = per_cpu_ptr(fbc->counters, cpu); | |
69 | *pcount = 0; | |
70 | } | |
71 | fbc->count = amount; | |
098faf58 | 72 | raw_spin_unlock_irqrestore(&fbc->lock, flags); |
3a587f47 PZ |
73 | } |
74 | EXPORT_SYMBOL(percpu_counter_set); | |
75 | ||
3e8f399d NB |
76 | /** |
77 | * This function is both preempt and irq safe. The former is due to explicit | |
78 | * preemption disable. The latter is guaranteed by the fact that the slow path | |
79 | * is explicitly protected by an irq-safe spinlock whereas the fast patch uses | |
80 | * this_cpu_add which is irq-safe by definition. Hence there is no need muck | |
81 | * with irq state before calling this one | |
82 | */ | |
104b4e51 | 83 | void percpu_counter_add_batch(struct percpu_counter *fbc, s64 amount, s32 batch) |
3cbc5640 | 84 | { |
20e89767 | 85 | s64 count; |
3cbc5640 | 86 | |
ea00c30b | 87 | preempt_disable(); |
819a72af | 88 | count = __this_cpu_read(*fbc->counters) + amount; |
252e0ba6 | 89 | if (count >= batch || count <= -batch) { |
098faf58 SL |
90 | unsigned long flags; |
91 | raw_spin_lock_irqsave(&fbc->lock, flags); | |
3cbc5640 | 92 | fbc->count += count; |
d1969a84 | 93 | __this_cpu_sub(*fbc->counters, count - amount); |
098faf58 | 94 | raw_spin_unlock_irqrestore(&fbc->lock, flags); |
3cbc5640 | 95 | } else { |
74e72f89 | 96 | this_cpu_add(*fbc->counters, amount); |
3cbc5640 | 97 | } |
ea00c30b | 98 | preempt_enable(); |
3cbc5640 | 99 | } |
104b4e51 | 100 | EXPORT_SYMBOL(percpu_counter_add_batch); |
3cbc5640 RT |
101 | |
102 | /* | |
103 | * Add up all the per-cpu counts, return the result. This is a more accurate | |
104 | * but much slower version of percpu_counter_read_positive() | |
105 | */ | |
02d21168 | 106 | s64 __percpu_counter_sum(struct percpu_counter *fbc) |
3cbc5640 | 107 | { |
0216bfcf | 108 | s64 ret; |
3cbc5640 | 109 | int cpu; |
098faf58 | 110 | unsigned long flags; |
3cbc5640 | 111 | |
098faf58 | 112 | raw_spin_lock_irqsave(&fbc->lock, flags); |
3cbc5640 | 113 | ret = fbc->count; |
b4ef0296 | 114 | for_each_online_cpu(cpu) { |
0216bfcf | 115 | s32 *pcount = per_cpu_ptr(fbc->counters, cpu); |
3cbc5640 RT |
116 | ret += *pcount; |
117 | } | |
098faf58 | 118 | raw_spin_unlock_irqrestore(&fbc->lock, flags); |
bf1d89c8 | 119 | return ret; |
3cbc5640 | 120 | } |
bf1d89c8 | 121 | EXPORT_SYMBOL(__percpu_counter_sum); |
c67ad917 | 122 | |
908c7f19 | 123 | int __percpu_counter_init(struct percpu_counter *fbc, s64 amount, gfp_t gfp, |
ea319518 | 124 | struct lock_class_key *key) |
c67ad917 | 125 | { |
ebd8fef3 TH |
126 | unsigned long flags __maybe_unused; |
127 | ||
f032a450 | 128 | raw_spin_lock_init(&fbc->lock); |
ea319518 | 129 | lockdep_set_class(&fbc->lock, key); |
c67ad917 | 130 | fbc->count = amount; |
908c7f19 | 131 | fbc->counters = alloc_percpu_gfp(s32, gfp); |
833f4077 PZ |
132 | if (!fbc->counters) |
133 | return -ENOMEM; | |
e2852ae8 TH |
134 | |
135 | debug_percpu_counter_activate(fbc); | |
136 | ||
c67ad917 | 137 | #ifdef CONFIG_HOTPLUG_CPU |
8474b591 | 138 | INIT_LIST_HEAD(&fbc->list); |
ebd8fef3 | 139 | spin_lock_irqsave(&percpu_counters_lock, flags); |
c67ad917 | 140 | list_add(&fbc->list, &percpu_counters); |
ebd8fef3 | 141 | spin_unlock_irqrestore(&percpu_counters_lock, flags); |
c67ad917 | 142 | #endif |
833f4077 | 143 | return 0; |
c67ad917 | 144 | } |
ea319518 | 145 | EXPORT_SYMBOL(__percpu_counter_init); |
c67ad917 AM |
146 | |
147 | void percpu_counter_destroy(struct percpu_counter *fbc) | |
148 | { | |
ebd8fef3 TH |
149 | unsigned long flags __maybe_unused; |
150 | ||
833f4077 PZ |
151 | if (!fbc->counters) |
152 | return; | |
153 | ||
e2852ae8 TH |
154 | debug_percpu_counter_deactivate(fbc); |
155 | ||
c67ad917 | 156 | #ifdef CONFIG_HOTPLUG_CPU |
ebd8fef3 | 157 | spin_lock_irqsave(&percpu_counters_lock, flags); |
c67ad917 | 158 | list_del(&fbc->list); |
ebd8fef3 | 159 | spin_unlock_irqrestore(&percpu_counters_lock, flags); |
c67ad917 | 160 | #endif |
fd3d664f ED |
161 | free_percpu(fbc->counters); |
162 | fbc->counters = NULL; | |
c67ad917 AM |
163 | } |
164 | EXPORT_SYMBOL(percpu_counter_destroy); | |
165 | ||
179f7ebf ED |
166 | int percpu_counter_batch __read_mostly = 32; |
167 | EXPORT_SYMBOL(percpu_counter_batch); | |
168 | ||
5588f5af | 169 | static int compute_batch_value(unsigned int cpu) |
179f7ebf ED |
170 | { |
171 | int nr = num_online_cpus(); | |
172 | ||
173 | percpu_counter_batch = max(32, nr*2); | |
5588f5af | 174 | return 0; |
179f7ebf ED |
175 | } |
176 | ||
5588f5af | 177 | static int percpu_counter_cpu_dead(unsigned int cpu) |
c67ad917 | 178 | { |
179f7ebf | 179 | #ifdef CONFIG_HOTPLUG_CPU |
c67ad917 AM |
180 | struct percpu_counter *fbc; |
181 | ||
5588f5af | 182 | compute_batch_value(cpu); |
c67ad917 | 183 | |
ebd8fef3 | 184 | spin_lock_irq(&percpu_counters_lock); |
c67ad917 AM |
185 | list_for_each_entry(fbc, &percpu_counters, list) { |
186 | s32 *pcount; | |
187 | ||
aaf0f2fa | 188 | raw_spin_lock(&fbc->lock); |
c67ad917 AM |
189 | pcount = per_cpu_ptr(fbc->counters, cpu); |
190 | fbc->count += *pcount; | |
191 | *pcount = 0; | |
aaf0f2fa | 192 | raw_spin_unlock(&fbc->lock); |
c67ad917 | 193 | } |
ebd8fef3 | 194 | spin_unlock_irq(&percpu_counters_lock); |
179f7ebf | 195 | #endif |
5588f5af | 196 | return 0; |
c67ad917 AM |
197 | } |
198 | ||
27f5e0f6 TC |
199 | /* |
200 | * Compare counter against given value. | |
201 | * Return 1 if greater, 0 if equal and -1 if less | |
202 | */ | |
80188b0d | 203 | int __percpu_counter_compare(struct percpu_counter *fbc, s64 rhs, s32 batch) |
27f5e0f6 TC |
204 | { |
205 | s64 count; | |
206 | ||
207 | count = percpu_counter_read(fbc); | |
208 | /* Check to see if rough count will be sufficient for comparison */ | |
80188b0d | 209 | if (abs(count - rhs) > (batch * num_online_cpus())) { |
27f5e0f6 TC |
210 | if (count > rhs) |
211 | return 1; | |
212 | else | |
213 | return -1; | |
214 | } | |
215 | /* Need to use precise count */ | |
216 | count = percpu_counter_sum(fbc); | |
217 | if (count > rhs) | |
218 | return 1; | |
219 | else if (count < rhs) | |
220 | return -1; | |
221 | else | |
222 | return 0; | |
223 | } | |
80188b0d | 224 | EXPORT_SYMBOL(__percpu_counter_compare); |
27f5e0f6 | 225 | |
c67ad917 AM |
226 | static int __init percpu_counter_startup(void) |
227 | { | |
5588f5af SAS |
228 | int ret; |
229 | ||
230 | ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "lib/percpu_cnt:online", | |
231 | compute_batch_value, NULL); | |
232 | WARN_ON(ret < 0); | |
233 | ret = cpuhp_setup_state_nocalls(CPUHP_PERCPU_CNT_DEAD, | |
234 | "lib/percpu_cnt:dead", NULL, | |
235 | percpu_counter_cpu_dead); | |
236 | WARN_ON(ret < 0); | |
c67ad917 AM |
237 | return 0; |
238 | } | |
239 | module_init(percpu_counter_startup); |