lib/percpu_counter.c

   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 const 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 const 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 (abs(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  * 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
 120 static s64 __percpu_counter_sum_mask(struct percpu_counter *fbc,
 121                               const struct cpumask *cpu_mask)
 122 {
 123         s64 ret;
 124         int cpu;
 125         unsigned long flags;
 126
 127         raw_spin_lock_irqsave(&fbc->lock, flags);
 128         ret = fbc->count;
 129         for_each_cpu(cpu, cpu_mask) {
 130                 s32 *pcount = per_cpu_ptr(fbc->counters, cpu);
 131                 ret += *pcount;
 132         }
 133         raw_spin_unlock_irqrestore(&fbc->lock, flags);
 134         return ret;
 135 }
 136
 137 /*
 138  * Add up all the per-cpu counts, return the result.  This is a more accurate
 139  * but much slower version of percpu_counter_read_positive()
 140  */
 141 s64 __percpu_counter_sum(struct percpu_counter *fbc)
 142 {
 143         return __percpu_counter_sum_mask(fbc, cpu_online_mask);
 144 }
 145 EXPORT_SYMBOL(__percpu_counter_sum);
 146
 147 /*
 148  * This is slower version of percpu_counter_sum as it traverses all possible
 149  * cpus. Use this only in the cases where accurate data is needed in the
 150  * presense of CPUs getting offlined.
 151  */
 152 s64 percpu_counter_sum_all(struct percpu_counter *fbc)
 153 {
 154         return __percpu_counter_sum_mask(fbc, cpu_possible_mask);
 155 }
 156 EXPORT_SYMBOL(percpu_counter_sum_all);
 157
 158 int __percpu_counter_init(struct percpu_counter *fbc, s64 amount, gfp_t gfp,
 159                           struct lock_class_key *key)
 160 {
 161         unsigned long flags __maybe_unused;
 162
 163         raw_spin_lock_init(&fbc->lock);
 164         lockdep_set_class(&fbc->lock, key);
 165         fbc->count = amount;
 166         fbc->counters = alloc_percpu_gfp(s32, gfp);
 167         if (!fbc->counters)
 168                 return -ENOMEM;
 169
 170         debug_percpu_counter_activate(fbc);
 171
 172 #ifdef CONFIG_HOTPLUG_CPU
 173         INIT_LIST_HEAD(&fbc->list);
 174         spin_lock_irqsave(&percpu_counters_lock, flags);
 175         list_add(&fbc->list, &percpu_counters);
 176         spin_unlock_irqrestore(&percpu_counters_lock, flags);
 177 #endif
 178         return 0;
 179 }
 180 EXPORT_SYMBOL(__percpu_counter_init);
 181
 182 void percpu_counter_destroy(struct percpu_counter *fbc)
 183 {
 184         unsigned long flags __maybe_unused;
 185
 186         if (!fbc->counters)
 187                 return;
 188
 189         debug_percpu_counter_deactivate(fbc);
 190
 191 #ifdef CONFIG_HOTPLUG_CPU
 192         spin_lock_irqsave(&percpu_counters_lock, flags);
 193         list_del(&fbc->list);
 194         spin_unlock_irqrestore(&percpu_counters_lock, flags);
 195 #endif
 196         free_percpu(fbc->counters);
 197         fbc->counters = NULL;
 198 }
 199 EXPORT_SYMBOL(percpu_counter_destroy);
 200
 201 int percpu_counter_batch __read_mostly = 32;
 202 EXPORT_SYMBOL(percpu_counter_batch);
 203
 204 static int compute_batch_value(unsigned int cpu)
 205 {
 206         int nr = num_online_cpus();
 207
 208         percpu_counter_batch = max(32, nr*2);
 209         return 0;
 210 }
 211
 212 static int percpu_counter_cpu_dead(unsigned int cpu)
 213 {
 214 #ifdef CONFIG_HOTPLUG_CPU
 215         struct percpu_counter *fbc;
 216
 217         compute_batch_value(cpu);
 218
 219         spin_lock_irq(&percpu_counters_lock);
 220         list_for_each_entry(fbc, &percpu_counters, list) {
 221                 s32 *pcount;
 222
 223                 raw_spin_lock(&fbc->lock);
 224                 pcount = per_cpu_ptr(fbc->counters, cpu);
 225                 fbc->count += *pcount;
 226                 *pcount = 0;
 227                 raw_spin_unlock(&fbc->lock);
 228         }
 229         spin_unlock_irq(&percpu_counters_lock);
 230 #endif
 231         return 0;
 232 }
 233
 234 /*
 235  * Compare counter against given value.
 236  * Return 1 if greater, 0 if equal and -1 if less
 237  */
 238 int __percpu_counter_compare(struct percpu_counter *fbc, s64 rhs, s32 batch)
 239 {
 240         s64     count;
 241
 242         count = percpu_counter_read(fbc);
 243         /* Check to see if rough count will be sufficient for comparison */
 244         if (abs(count - rhs) > (batch * num_online_cpus())) {
 245                 if (count > rhs)
 246                         return 1;
 247                 else
 248                         return -1;
 249         }
 250         /* Need to use precise count */
 251         count = percpu_counter_sum(fbc);
 252         if (count > rhs)
 253                 return 1;
 254         else if (count < rhs)
 255                 return -1;
 256         else
 257                 return 0;
 258 }
 259 EXPORT_SYMBOL(__percpu_counter_compare);
 260
 261 static int __init percpu_counter_startup(void)
 262 {
 263         int ret;
 264
 265         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "lib/percpu_cnt:online",
 266                                 compute_batch_value, NULL);
 267         WARN_ON(ret < 0);
 268         ret = cpuhp_setup_state_nocalls(CPUHP_PERCPU_CNT_DEAD,
 269                                         "lib/percpu_cnt:dead", NULL,
 270                                         percpu_counter_cpu_dead);
 271         WARN_ON(ret < 0);
 272         return 0;
 273 }
 274 module_init(percpu_counter_startup);