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  * 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.
  86  */
  87 void percpu_counter_add_batch(struct percpu_counter *fbc, s64 amount, s32 batch)
  88 {
  89         s64 count;
  90         unsigned long flags;
  91
  92         local_irq_save(flags);
  93         count = __this_cpu_read(*fbc->counters) + amount;
  94         if (abs(count) >= batch) {
  95                 raw_spin_lock(&fbc->lock);
  96                 fbc->count += count;
  97                 __this_cpu_sub(*fbc->counters, count - amount);
  98                 raw_spin_unlock(&fbc->lock);
  99         } else {
 100                 this_cpu_add(*fbc->counters, amount);
 101         }
 102         local_irq_restore(flags);
 103 }
 104 EXPORT_SYMBOL(percpu_counter_add_batch);
 105
 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
 125 static s64 __percpu_counter_sum_mask(struct percpu_counter *fbc,
 126                               const struct cpumask *cpu_mask)
 127 {
 128         s64 ret;
 129         int cpu;
 130         unsigned long flags;
 131
 132         raw_spin_lock_irqsave(&fbc->lock, flags);
 133         ret = fbc->count;
 134         for_each_cpu(cpu, cpu_mask) {
 135                 s32 *pcount = per_cpu_ptr(fbc->counters, cpu);
 136                 ret += *pcount;
 137         }
 138         raw_spin_unlock_irqrestore(&fbc->lock, flags);
 139         return ret;
 140 }
 141
 142 /*
 143  * Add up all the per-cpu counts, return the result.  This is a more accurate
 144  * but much slower version of percpu_counter_read_positive()
 145  */
 146 s64 __percpu_counter_sum(struct percpu_counter *fbc)
 147 {
 148         return __percpu_counter_sum_mask(fbc, cpu_online_mask);
 149 }
 150 EXPORT_SYMBOL(__percpu_counter_sum);
 151
 152 /*
 153  * This is slower version of percpu_counter_sum as it traverses all possible
 154  * cpus. Use this only in the cases where accurate data is needed in the
 155  * presense of CPUs getting offlined.
 156  */
 157 s64 percpu_counter_sum_all(struct percpu_counter *fbc)
 158 {
 159         return __percpu_counter_sum_mask(fbc, cpu_possible_mask);
 160 }
 161 EXPORT_SYMBOL(percpu_counter_sum_all);
 162
 163 int __percpu_counter_init(struct percpu_counter *fbc, s64 amount, gfp_t gfp,
 164                           struct lock_class_key *key)
 165 {
 166         unsigned long flags __maybe_unused;
 167
 168         raw_spin_lock_init(&fbc->lock);
 169         lockdep_set_class(&fbc->lock, key);
 170         fbc->count = amount;
 171         fbc->counters = alloc_percpu_gfp(s32, gfp);
 172         if (!fbc->counters)
 173                 return -ENOMEM;
 174
 175         debug_percpu_counter_activate(fbc);
 176
 177 #ifdef CONFIG_HOTPLUG_CPU
 178         INIT_LIST_HEAD(&fbc->list);
 179         spin_lock_irqsave(&percpu_counters_lock, flags);
 180         list_add(&fbc->list, &percpu_counters);
 181         spin_unlock_irqrestore(&percpu_counters_lock, flags);
 182 #endif
 183         return 0;
 184 }
 185 EXPORT_SYMBOL(__percpu_counter_init);
 186
 187 void percpu_counter_destroy(struct percpu_counter *fbc)
 188 {
 189         unsigned long flags __maybe_unused;
 190
 191         if (!fbc->counters)
 192                 return;
 193
 194         debug_percpu_counter_deactivate(fbc);
 195
 196 #ifdef CONFIG_HOTPLUG_CPU
 197         spin_lock_irqsave(&percpu_counters_lock, flags);
 198         list_del(&fbc->list);
 199         spin_unlock_irqrestore(&percpu_counters_lock, flags);
 200 #endif
 201         free_percpu(fbc->counters);
 202         fbc->counters = NULL;
 203 }
 204 EXPORT_SYMBOL(percpu_counter_destroy);
 205
 206 int percpu_counter_batch __read_mostly = 32;
 207 EXPORT_SYMBOL(percpu_counter_batch);
 208
 209 static int compute_batch_value(unsigned int cpu)
 210 {
 211         int nr = num_online_cpus();
 212
 213         percpu_counter_batch = max(32, nr*2);
 214         return 0;
 215 }
 216
 217 static int percpu_counter_cpu_dead(unsigned int cpu)
 218 {
 219 #ifdef CONFIG_HOTPLUG_CPU
 220         struct percpu_counter *fbc;
 221
 222         compute_batch_value(cpu);
 223
 224         spin_lock_irq(&percpu_counters_lock);
 225         list_for_each_entry(fbc, &percpu_counters, list) {
 226                 s32 *pcount;
 227
 228                 raw_spin_lock(&fbc->lock);
 229                 pcount = per_cpu_ptr(fbc->counters, cpu);
 230                 fbc->count += *pcount;
 231                 *pcount = 0;
 232                 raw_spin_unlock(&fbc->lock);
 233         }
 234         spin_unlock_irq(&percpu_counters_lock);
 235 #endif
 236         return 0;
 237 }
 238
 239 /*
 240  * Compare counter against given value.
 241  * Return 1 if greater, 0 if equal and -1 if less
 242  */
 243 int __percpu_counter_compare(struct percpu_counter *fbc, s64 rhs, s32 batch)
 244 {
 245         s64     count;
 246
 247         count = percpu_counter_read(fbc);
 248         /* Check to see if rough count will be sufficient for comparison */
 249         if (abs(count - rhs) > (batch * num_online_cpus())) {
 250                 if (count > rhs)
 251                         return 1;
 252                 else
 253                         return -1;
 254         }
 255         /* Need to use precise count */
 256         count = percpu_counter_sum(fbc);
 257         if (count > rhs)
 258                 return 1;
 259         else if (count < rhs)
 260                 return -1;
 261         else
 262                 return 0;
 263 }
 264 EXPORT_SYMBOL(__percpu_counter_compare);
 265
 266 static int __init percpu_counter_startup(void)
 267 {
 268         int ret;
 269
 270         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "lib/percpu_cnt:online",
 271                                 compute_batch_value, NULL);
 272         WARN_ON(ret < 0);
 273         ret = cpuhp_setup_state_nocalls(CPUHP_PERCPU_CNT_DEAD,
 274                                         "lib/percpu_cnt:dead", NULL,
 275                                         percpu_counter_cpu_dead);
 276         WARN_ON(ret < 0);
 277         return 0;
 278 }
 279 module_init(percpu_counter_startup);