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 | /* |
805afd83 MS |
76 | * local_irq_save() is needed to make the function irq safe: |
77 | * - The slow path would be ok as protected by an irq-safe spinlock. | |
78 | * - this_cpu_add would be ok as it is irq-safe by definition. | |
79 | * But: | |
80 | * The decision slow path/fast path and the actual update must be atomic, too. | |
81 | * Otherwise a call in process context could check the current values and | |
82 | * decide that the fast path can be used. If now an interrupt occurs before | |
83 | * the this_cpu_add(), and the interrupt updates this_cpu(*fbc->counters), | |
84 | * then the this_cpu_add() that is executed after the interrupt has completed | |
85 | * can produce values larger than "batch" or even overflows. | |
3e8f399d | 86 | */ |
104b4e51 | 87 | void percpu_counter_add_batch(struct percpu_counter *fbc, s64 amount, s32 batch) |
3cbc5640 | 88 | { |
20e89767 | 89 | s64 count; |
805afd83 | 90 | unsigned long flags; |
3cbc5640 | 91 | |
805afd83 | 92 | local_irq_save(flags); |
819a72af | 93 | count = __this_cpu_read(*fbc->counters) + amount; |
1d339638 | 94 | if (abs(count) >= batch) { |
805afd83 | 95 | raw_spin_lock(&fbc->lock); |
3cbc5640 | 96 | fbc->count += count; |
d1969a84 | 97 | __this_cpu_sub(*fbc->counters, count - amount); |
805afd83 | 98 | raw_spin_unlock(&fbc->lock); |
3cbc5640 | 99 | } else { |
74e72f89 | 100 | this_cpu_add(*fbc->counters, amount); |
3cbc5640 | 101 | } |
805afd83 | 102 | local_irq_restore(flags); |
3cbc5640 | 103 | } |
104b4e51 | 104 | EXPORT_SYMBOL(percpu_counter_add_batch); |
3cbc5640 | 105 | |
0a4954a8 FT |
106 | /* |
107 | * For percpu_counter with a big batch, the devication of its count could | |
108 | * be big, and there is requirement to reduce the deviation, like when the | |
109 | * counter's batch could be runtime decreased to get a better accuracy, | |
110 | * which can be achieved by running this sync function on each CPU. | |
111 | */ | |
112 | void percpu_counter_sync(struct percpu_counter *fbc) | |
113 | { | |
114 | unsigned long flags; | |
115 | s64 count; | |
116 | ||
117 | raw_spin_lock_irqsave(&fbc->lock, flags); | |
118 | count = __this_cpu_read(*fbc->counters); | |
119 | fbc->count += count; | |
120 | __this_cpu_sub(*fbc->counters, count); | |
121 | raw_spin_unlock_irqrestore(&fbc->lock, flags); | |
122 | } | |
123 | EXPORT_SYMBOL(percpu_counter_sync); | |
124 | ||
f689054a SB |
125 | /* |
126 | * Add up all the per-cpu counts, return the result. This is a more accurate | |
8b57b11c DC |
127 | * but much slower version of percpu_counter_read_positive(). |
128 | * | |
129 | * We use the cpu mask of (cpu_online_mask | cpu_dying_mask) to capture sums | |
130 | * from CPUs that are in the process of being taken offline. Dying cpus have | |
131 | * been removed from the online mask, but may not have had the hotplug dead | |
132 | * notifier called to fold the percpu count back into the global counter sum. | |
133 | * By including dying CPUs in the iteration mask, we avoid this race condition | |
134 | * so __percpu_counter_sum() just does the right thing when CPUs are being taken | |
135 | * offline. | |
f689054a SB |
136 | */ |
137 | s64 __percpu_counter_sum(struct percpu_counter *fbc) | |
138 | { | |
e9b60c7f DC |
139 | s64 ret; |
140 | int cpu; | |
141 | unsigned long flags; | |
8b57b11c | 142 | |
e9b60c7f DC |
143 | raw_spin_lock_irqsave(&fbc->lock, flags); |
144 | ret = fbc->count; | |
145 | for_each_cpu_or(cpu, cpu_online_mask, cpu_dying_mask) { | |
146 | s32 *pcount = per_cpu_ptr(fbc->counters, cpu); | |
147 | ret += *pcount; | |
148 | } | |
149 | raw_spin_unlock_irqrestore(&fbc->lock, flags); | |
150 | return ret; | |
f689054a | 151 | } |
bf1d89c8 | 152 | EXPORT_SYMBOL(__percpu_counter_sum); |
c67ad917 | 153 | |
908c7f19 | 154 | int __percpu_counter_init(struct percpu_counter *fbc, s64 amount, gfp_t gfp, |
ea319518 | 155 | struct lock_class_key *key) |
c67ad917 | 156 | { |
ebd8fef3 TH |
157 | unsigned long flags __maybe_unused; |
158 | ||
f032a450 | 159 | raw_spin_lock_init(&fbc->lock); |
ea319518 | 160 | lockdep_set_class(&fbc->lock, key); |
c67ad917 | 161 | fbc->count = amount; |
908c7f19 | 162 | fbc->counters = alloc_percpu_gfp(s32, gfp); |
833f4077 PZ |
163 | if (!fbc->counters) |
164 | return -ENOMEM; | |
e2852ae8 TH |
165 | |
166 | debug_percpu_counter_activate(fbc); | |
167 | ||
c67ad917 | 168 | #ifdef CONFIG_HOTPLUG_CPU |
8474b591 | 169 | INIT_LIST_HEAD(&fbc->list); |
ebd8fef3 | 170 | spin_lock_irqsave(&percpu_counters_lock, flags); |
c67ad917 | 171 | list_add(&fbc->list, &percpu_counters); |
ebd8fef3 | 172 | spin_unlock_irqrestore(&percpu_counters_lock, flags); |
c67ad917 | 173 | #endif |
833f4077 | 174 | return 0; |
c67ad917 | 175 | } |
ea319518 | 176 | EXPORT_SYMBOL(__percpu_counter_init); |
c67ad917 AM |
177 | |
178 | void percpu_counter_destroy(struct percpu_counter *fbc) | |
179 | { | |
ebd8fef3 TH |
180 | unsigned long flags __maybe_unused; |
181 | ||
833f4077 PZ |
182 | if (!fbc->counters) |
183 | return; | |
184 | ||
e2852ae8 TH |
185 | debug_percpu_counter_deactivate(fbc); |
186 | ||
c67ad917 | 187 | #ifdef CONFIG_HOTPLUG_CPU |
ebd8fef3 | 188 | spin_lock_irqsave(&percpu_counters_lock, flags); |
c67ad917 | 189 | list_del(&fbc->list); |
ebd8fef3 | 190 | spin_unlock_irqrestore(&percpu_counters_lock, flags); |
c67ad917 | 191 | #endif |
fd3d664f ED |
192 | free_percpu(fbc->counters); |
193 | fbc->counters = NULL; | |
c67ad917 AM |
194 | } |
195 | EXPORT_SYMBOL(percpu_counter_destroy); | |
196 | ||
179f7ebf ED |
197 | int percpu_counter_batch __read_mostly = 32; |
198 | EXPORT_SYMBOL(percpu_counter_batch); | |
199 | ||
5588f5af | 200 | static int compute_batch_value(unsigned int cpu) |
179f7ebf ED |
201 | { |
202 | int nr = num_online_cpus(); | |
203 | ||
204 | percpu_counter_batch = max(32, nr*2); | |
5588f5af | 205 | return 0; |
179f7ebf ED |
206 | } |
207 | ||
5588f5af | 208 | static int percpu_counter_cpu_dead(unsigned int cpu) |
c67ad917 | 209 | { |
179f7ebf | 210 | #ifdef CONFIG_HOTPLUG_CPU |
c67ad917 AM |
211 | struct percpu_counter *fbc; |
212 | ||
5588f5af | 213 | compute_batch_value(cpu); |
c67ad917 | 214 | |
ebd8fef3 | 215 | spin_lock_irq(&percpu_counters_lock); |
c67ad917 AM |
216 | list_for_each_entry(fbc, &percpu_counters, list) { |
217 | s32 *pcount; | |
218 | ||
aaf0f2fa | 219 | raw_spin_lock(&fbc->lock); |
c67ad917 AM |
220 | pcount = per_cpu_ptr(fbc->counters, cpu); |
221 | fbc->count += *pcount; | |
222 | *pcount = 0; | |
aaf0f2fa | 223 | raw_spin_unlock(&fbc->lock); |
c67ad917 | 224 | } |
ebd8fef3 | 225 | spin_unlock_irq(&percpu_counters_lock); |
179f7ebf | 226 | #endif |
5588f5af | 227 | return 0; |
c67ad917 AM |
228 | } |
229 | ||
27f5e0f6 TC |
230 | /* |
231 | * Compare counter against given value. | |
232 | * Return 1 if greater, 0 if equal and -1 if less | |
233 | */ | |
80188b0d | 234 | int __percpu_counter_compare(struct percpu_counter *fbc, s64 rhs, s32 batch) |
27f5e0f6 TC |
235 | { |
236 | s64 count; | |
237 | ||
238 | count = percpu_counter_read(fbc); | |
239 | /* Check to see if rough count will be sufficient for comparison */ | |
80188b0d | 240 | if (abs(count - rhs) > (batch * num_online_cpus())) { |
27f5e0f6 TC |
241 | if (count > rhs) |
242 | return 1; | |
243 | else | |
244 | return -1; | |
245 | } | |
246 | /* Need to use precise count */ | |
247 | count = percpu_counter_sum(fbc); | |
248 | if (count > rhs) | |
249 | return 1; | |
250 | else if (count < rhs) | |
251 | return -1; | |
252 | else | |
253 | return 0; | |
254 | } | |
80188b0d | 255 | EXPORT_SYMBOL(__percpu_counter_compare); |
27f5e0f6 | 256 | |
c67ad917 AM |
257 | static int __init percpu_counter_startup(void) |
258 | { | |
5588f5af SAS |
259 | int ret; |
260 | ||
261 | ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "lib/percpu_cnt:online", | |
262 | compute_batch_value, NULL); | |
263 | WARN_ON(ret < 0); | |
264 | ret = cpuhp_setup_state_nocalls(CPUHP_PERCPU_CNT_DEAD, | |
265 | "lib/percpu_cnt:dead", NULL, | |
266 | percpu_counter_cpu_dead); | |
267 | WARN_ON(ret < 0); | |
c67ad917 AM |
268 | return 0; |
269 | } | |
270 | module_init(percpu_counter_startup); |