| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | /* |
| 3 | * Fast batching percpu counters. |
| 4 | */ |
| 5 | |
| 6 | #include <linux/percpu_counter.h> |
| 7 | #include <linux/mutex.h> |
| 8 | #include <linux/init.h> |
| 9 | #include <linux/cpu.h> |
| 10 | #include <linux/module.h> |
| 11 | #include <linux/debugobjects.h> |
| 12 | |
| 13 | #ifdef CONFIG_HOTPLUG_CPU |
| 14 | static LIST_HEAD(percpu_counters); |
| 15 | static DEFINE_SPINLOCK(percpu_counters_lock); |
| 16 | #endif |
| 17 | |
| 18 | #ifdef CONFIG_DEBUG_OBJECTS_PERCPU_COUNTER |
| 19 | |
| 20 | static struct debug_obj_descr percpu_counter_debug_descr; |
| 21 | |
| 22 | static bool percpu_counter_fixup_free(void *addr, enum debug_obj_state state) |
| 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); |
| 30 | return true; |
| 31 | default: |
| 32 | return false; |
| 33 | } |
| 34 | } |
| 35 | |
| 36 | static struct debug_obj_descr percpu_counter_debug_descr = { |
| 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 | |
| 60 | void percpu_counter_set(struct percpu_counter *fbc, s64 amount) |
| 61 | { |
| 62 | int cpu; |
| 63 | unsigned long flags; |
| 64 | |
| 65 | raw_spin_lock_irqsave(&fbc->lock, flags); |
| 66 | for_each_possible_cpu(cpu) { |
| 67 | s32 *pcount = per_cpu_ptr(fbc->counters, cpu); |
| 68 | *pcount = 0; |
| 69 | } |
| 70 | fbc->count = amount; |
| 71 | raw_spin_unlock_irqrestore(&fbc->lock, flags); |
| 72 | } |
| 73 | EXPORT_SYMBOL(percpu_counter_set); |
| 74 | |
| 75 | /** |
| 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 | */ |
| 82 | void percpu_counter_add_batch(struct percpu_counter *fbc, s64 amount, s32 batch) |
| 83 | { |
| 84 | s64 count; |
| 85 | |
| 86 | preempt_disable(); |
| 87 | count = __this_cpu_read(*fbc->counters) + amount; |
| 88 | if (count >= batch || count <= -batch) { |
| 89 | unsigned long flags; |
| 90 | raw_spin_lock_irqsave(&fbc->lock, flags); |
| 91 | fbc->count += count; |
| 92 | __this_cpu_sub(*fbc->counters, count - amount); |
| 93 | raw_spin_unlock_irqrestore(&fbc->lock, flags); |
| 94 | } else { |
| 95 | this_cpu_add(*fbc->counters, amount); |
| 96 | } |
| 97 | preempt_enable(); |
| 98 | } |
| 99 | EXPORT_SYMBOL(percpu_counter_add_batch); |
| 100 | |
| 101 | /* |
| 102 | * Add up all the per-cpu counts, return the result. This is a more accurate |
| 103 | * but much slower version of percpu_counter_read_positive() |
| 104 | */ |
| 105 | s64 __percpu_counter_sum(struct percpu_counter *fbc) |
| 106 | { |
| 107 | s64 ret; |
| 108 | int cpu; |
| 109 | unsigned long flags; |
| 110 | |
| 111 | raw_spin_lock_irqsave(&fbc->lock, flags); |
| 112 | ret = fbc->count; |
| 113 | for_each_online_cpu(cpu) { |
| 114 | s32 *pcount = per_cpu_ptr(fbc->counters, cpu); |
| 115 | ret += *pcount; |
| 116 | } |
| 117 | raw_spin_unlock_irqrestore(&fbc->lock, flags); |
| 118 | return ret; |
| 119 | } |
| 120 | EXPORT_SYMBOL(__percpu_counter_sum); |
| 121 | |
| 122 | int __percpu_counter_init(struct percpu_counter *fbc, s64 amount, gfp_t gfp, |
| 123 | struct lock_class_key *key) |
| 124 | { |
| 125 | unsigned long flags __maybe_unused; |
| 126 | |
| 127 | raw_spin_lock_init(&fbc->lock); |
| 128 | lockdep_set_class(&fbc->lock, key); |
| 129 | fbc->count = amount; |
| 130 | fbc->counters = alloc_percpu_gfp(s32, gfp); |
| 131 | if (!fbc->counters) |
| 132 | return -ENOMEM; |
| 133 | |
| 134 | debug_percpu_counter_activate(fbc); |
| 135 | |
| 136 | #ifdef CONFIG_HOTPLUG_CPU |
| 137 | INIT_LIST_HEAD(&fbc->list); |
| 138 | spin_lock_irqsave(&percpu_counters_lock, flags); |
| 139 | list_add(&fbc->list, &percpu_counters); |
| 140 | spin_unlock_irqrestore(&percpu_counters_lock, flags); |
| 141 | #endif |
| 142 | return 0; |
| 143 | } |
| 144 | EXPORT_SYMBOL(__percpu_counter_init); |
| 145 | |
| 146 | void percpu_counter_destroy(struct percpu_counter *fbc) |
| 147 | { |
| 148 | unsigned long flags __maybe_unused; |
| 149 | |
| 150 | if (!fbc->counters) |
| 151 | return; |
| 152 | |
| 153 | debug_percpu_counter_deactivate(fbc); |
| 154 | |
| 155 | #ifdef CONFIG_HOTPLUG_CPU |
| 156 | spin_lock_irqsave(&percpu_counters_lock, flags); |
| 157 | list_del(&fbc->list); |
| 158 | spin_unlock_irqrestore(&percpu_counters_lock, flags); |
| 159 | #endif |
| 160 | free_percpu(fbc->counters); |
| 161 | fbc->counters = NULL; |
| 162 | } |
| 163 | EXPORT_SYMBOL(percpu_counter_destroy); |
| 164 | |
| 165 | int percpu_counter_batch __read_mostly = 32; |
| 166 | EXPORT_SYMBOL(percpu_counter_batch); |
| 167 | |
| 168 | static int compute_batch_value(unsigned int cpu) |
| 169 | { |
| 170 | int nr = num_online_cpus(); |
| 171 | |
| 172 | percpu_counter_batch = max(32, nr*2); |
| 173 | return 0; |
| 174 | } |
| 175 | |
| 176 | static int percpu_counter_cpu_dead(unsigned int cpu) |
| 177 | { |
| 178 | #ifdef CONFIG_HOTPLUG_CPU |
| 179 | struct percpu_counter *fbc; |
| 180 | |
| 181 | compute_batch_value(cpu); |
| 182 | |
| 183 | spin_lock_irq(&percpu_counters_lock); |
| 184 | list_for_each_entry(fbc, &percpu_counters, list) { |
| 185 | s32 *pcount; |
| 186 | |
| 187 | raw_spin_lock(&fbc->lock); |
| 188 | pcount = per_cpu_ptr(fbc->counters, cpu); |
| 189 | fbc->count += *pcount; |
| 190 | *pcount = 0; |
| 191 | raw_spin_unlock(&fbc->lock); |
| 192 | } |
| 193 | spin_unlock_irq(&percpu_counters_lock); |
| 194 | #endif |
| 195 | return 0; |
| 196 | } |
| 197 | |
| 198 | /* |
| 199 | * Compare counter against given value. |
| 200 | * Return 1 if greater, 0 if equal and -1 if less |
| 201 | */ |
| 202 | int __percpu_counter_compare(struct percpu_counter *fbc, s64 rhs, s32 batch) |
| 203 | { |
| 204 | s64 count; |
| 205 | |
| 206 | count = percpu_counter_read(fbc); |
| 207 | /* Check to see if rough count will be sufficient for comparison */ |
| 208 | if (abs(count - rhs) > (batch * num_online_cpus())) { |
| 209 | if (count > rhs) |
| 210 | return 1; |
| 211 | else |
| 212 | return -1; |
| 213 | } |
| 214 | /* Need to use precise count */ |
| 215 | count = percpu_counter_sum(fbc); |
| 216 | if (count > rhs) |
| 217 | return 1; |
| 218 | else if (count < rhs) |
| 219 | return -1; |
| 220 | else |
| 221 | return 0; |
| 222 | } |
| 223 | EXPORT_SYMBOL(__percpu_counter_compare); |
| 224 | |
| 225 | static int __init percpu_counter_startup(void) |
| 226 | { |
| 227 | int ret; |
| 228 | |
| 229 | ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "lib/percpu_cnt:online", |
| 230 | compute_batch_value, NULL); |
| 231 | WARN_ON(ret < 0); |
| 232 | ret = cpuhp_setup_state_nocalls(CPUHP_PERCPU_CNT_DEAD, |
| 233 | "lib/percpu_cnt:dead", NULL, |
| 234 | percpu_counter_cpu_dead); |
| 235 | WARN_ON(ret < 0); |
| 236 | return 0; |
| 237 | } |
| 238 | module_init(percpu_counter_startup); |