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 /*
 126  * Add up all the per-cpu counts, return the result.  This is a more accurate
 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.
 136  */
 137 s64 __percpu_counter_sum(struct percpu_counter *fbc)
 138 {
 139         s64 ret;
 140         int cpu;
 141         unsigned long flags;
 142
 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;
 151 }
 152 EXPORT_SYMBOL(__percpu_counter_sum);
 153
 154 int __percpu_counter_init_many(struct percpu_counter *fbc, s64 amount,
 155                                gfp_t gfp, u32 nr_counters,
 156                                struct lock_class_key *key)
 157 {
 158         unsigned long flags __maybe_unused;
 159         size_t counter_size;
 160         s32 __percpu *counters;
 161         u32 i;
 162
 163         counter_size = ALIGN(sizeof(*counters), __alignof__(*counters));
 164         counters = __alloc_percpu_gfp(nr_counters * counter_size,
 165                                       __alignof__(*counters), gfp);
 166         if (!counters) {
 167                 fbc[0].counters = NULL;
 168                 return -ENOMEM;
 169         }
 170
 171         for (i = 0; i < nr_counters; i++) {
 172                 raw_spin_lock_init(&fbc[i].lock);
 173                 lockdep_set_class(&fbc[i].lock, key);
 174 #ifdef CONFIG_HOTPLUG_CPU
 175                 INIT_LIST_HEAD(&fbc[i].list);
 176 #endif
 177                 fbc[i].count = amount;
 178                 fbc[i].counters = (void *)counters + (i * counter_size);
 179
 180                 debug_percpu_counter_activate(&fbc[i]);
 181         }
 182
 183 #ifdef CONFIG_HOTPLUG_CPU
 184         spin_lock_irqsave(&percpu_counters_lock, flags);
 185         for (i = 0; i < nr_counters; i++)
 186                 list_add(&fbc[i].list, &percpu_counters);
 187         spin_unlock_irqrestore(&percpu_counters_lock, flags);
 188 #endif
 189         return 0;
 190 }
 191 EXPORT_SYMBOL(__percpu_counter_init_many);
 192
 193 void percpu_counter_destroy_many(struct percpu_counter *fbc, u32 nr_counters)
 194 {
 195         unsigned long flags __maybe_unused;
 196         u32 i;
 197
 198         if (WARN_ON_ONCE(!fbc))
 199                 return;
 200
 201         if (!fbc[0].counters)
 202                 return;
 203
 204         for (i = 0; i < nr_counters; i++)
 205                 debug_percpu_counter_deactivate(&fbc[i]);
 206
 207 #ifdef CONFIG_HOTPLUG_CPU
 208         spin_lock_irqsave(&percpu_counters_lock, flags);
 209         for (i = 0; i < nr_counters; i++)
 210                 list_del(&fbc[i].list);
 211         spin_unlock_irqrestore(&percpu_counters_lock, flags);
 212 #endif
 213
 214         free_percpu(fbc[0].counters);
 215
 216         for (i = 0; i < nr_counters; i++)
 217                 fbc[i].counters = NULL;
 218 }
 219 EXPORT_SYMBOL(percpu_counter_destroy_many);
 220
 221 int percpu_counter_batch __read_mostly = 32;
 222 EXPORT_SYMBOL(percpu_counter_batch);
 223
 224 static int compute_batch_value(unsigned int cpu)
 225 {
 226         int nr = num_online_cpus();
 227
 228         percpu_counter_batch = max(32, nr*2);
 229         return 0;
 230 }
 231
 232 static int percpu_counter_cpu_dead(unsigned int cpu)
 233 {
 234 #ifdef CONFIG_HOTPLUG_CPU
 235         struct percpu_counter *fbc;
 236
 237         compute_batch_value(cpu);
 238
 239         spin_lock_irq(&percpu_counters_lock);
 240         list_for_each_entry(fbc, &percpu_counters, list) {
 241                 s32 *pcount;
 242
 243                 raw_spin_lock(&fbc->lock);
 244                 pcount = per_cpu_ptr(fbc->counters, cpu);
 245                 fbc->count += *pcount;
 246                 *pcount = 0;
 247                 raw_spin_unlock(&fbc->lock);
 248         }
 249         spin_unlock_irq(&percpu_counters_lock);
 250 #endif
 251         return 0;
 252 }
 253
 254 /*
 255  * Compare counter against given value.
 256  * Return 1 if greater, 0 if equal and -1 if less
 257  */
 258 int __percpu_counter_compare(struct percpu_counter *fbc, s64 rhs, s32 batch)
 259 {
 260         s64     count;
 261
 262         count = percpu_counter_read(fbc);
 263         /* Check to see if rough count will be sufficient for comparison */
 264         if (abs(count - rhs) > (batch * num_online_cpus())) {
 265                 if (count > rhs)
 266                         return 1;
 267                 else
 268                         return -1;
 269         }
 270         /* Need to use precise count */
 271         count = percpu_counter_sum(fbc);
 272         if (count > rhs)
 273                 return 1;
 274         else if (count < rhs)
 275                 return -1;
 276         else
 277                 return 0;
 278 }
 279 EXPORT_SYMBOL(__percpu_counter_compare);
 280
 281 /*
 282  * Compare counter, and add amount if total is: less than or equal to limit if
 283  * amount is positive, or greater than or equal to limit if amount is negative.
 284  * Return true if amount is added, or false if total would be beyond the limit.
 285  *
 286  * Negative limit is allowed, but unusual.
 287  * When negative amounts (subs) are given to percpu_counter_limited_add(),
 288  * the limit would most naturally be 0 - but other limits are also allowed.
 289  *
 290  * Overflow beyond S64_MAX is not allowed for: counter, limit and amount
 291  * are all assumed to be sane (far from S64_MIN and S64_MAX).
 292  */
 293 bool __percpu_counter_limited_add(struct percpu_counter *fbc,
 294                                   s64 limit, s64 amount, s32 batch)
 295 {
 296         s64 count;
 297         s64 unknown;
 298         unsigned long flags;
 299         bool good = false;
 300
 301         if (amount == 0)
 302                 return true;
 303
 304         local_irq_save(flags);
 305         unknown = batch * num_online_cpus();
 306         count = __this_cpu_read(*fbc->counters);
 307
 308         /* Skip taking the lock when safe */
 309         if (abs(count + amount) <= batch &&
 310             ((amount > 0 && fbc->count + unknown <= limit) ||
 311              (amount < 0 && fbc->count - unknown >= limit))) {
 312                 this_cpu_add(*fbc->counters, amount);
 313                 local_irq_restore(flags);
 314                 return true;
 315         }
 316
 317         raw_spin_lock(&fbc->lock);
 318         count = fbc->count + amount;
 319
 320         /* Skip percpu_counter_sum() when safe */
 321         if (amount > 0) {
 322                 if (count - unknown > limit)
 323                         goto out;
 324                 if (count + unknown <= limit)
 325                         good = true;
 326         } else {
 327                 if (count + unknown < limit)
 328                         goto out;
 329                 if (count - unknown >= limit)
 330                         good = true;
 331         }
 332
 333         if (!good) {
 334                 s32 *pcount;
 335                 int cpu;
 336
 337                 for_each_cpu_or(cpu, cpu_online_mask, cpu_dying_mask) {
 338                         pcount = per_cpu_ptr(fbc->counters, cpu);
 339                         count += *pcount;
 340                 }
 341                 if (amount > 0) {
 342                         if (count > limit)
 343                                 goto out;
 344                 } else {
 345                         if (count < limit)
 346                                 goto out;
 347                 }
 348                 good = true;
 349         }
 350
 351         count = __this_cpu_read(*fbc->counters);
 352         fbc->count += count + amount;
 353         __this_cpu_sub(*fbc->counters, count);
 354 out:
 355         raw_spin_unlock(&fbc->lock);
 356         local_irq_restore(flags);
 357         return good;
 358 }
 359
 360 static int __init percpu_counter_startup(void)
 361 {
 362         int ret;
 363
 364         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "lib/percpu_cnt:online",
 365                                 compute_batch_value, NULL);
 366         WARN_ON(ret < 0);
 367         ret = cpuhp_setup_state_nocalls(CPUHP_PERCPU_CNT_DEAD,
 368                                         "lib/percpu_cnt:dead", NULL,
 369                                         percpu_counter_cpu_dead);
 370         WARN_ON(ret < 0);
 371         return 0;
 372 }
 373 module_init(percpu_counter_startup);