lib/percpu-refcount.c

   1 #define pr_fmt(fmt) "%s: " fmt "\n", __func__
   2
   3 #include <linux/kernel.h>
   4 #include <linux/percpu-refcount.h>
   5
   6 /*
   7  * Initially, a percpu refcount is just a set of percpu counters. Initially, we
   8  * don't try to detect the ref hitting 0 - which means that get/put can just
   9  * increment or decrement the local counter. Note that the counter on a
  10  * particular cpu can (and will) wrap - this is fine, when we go to shutdown the
  11  * percpu counters will all sum to the correct value
  12  *
  13  * (More precisely: because moduler arithmatic is commutative the sum of all the
  14  * pcpu_count vars will be equal to what it would have been if all the gets and
  15  * puts were done to a single integer, even if some of the percpu integers
  16  * overflow or underflow).
  17  *
  18  * The real trick to implementing percpu refcounts is shutdown. We can't detect
  19  * the ref hitting 0 on every put - this would require global synchronization
  20  * and defeat the whole purpose of using percpu refs.
  21  *
  22  * What we do is require the user to keep track of the initial refcount; we know
  23  * the ref can't hit 0 before the user drops the initial ref, so as long as we
  24  * convert to non percpu mode before the initial ref is dropped everything
  25  * works.
  26  *
  27  * Converting to non percpu mode is done with some RCUish stuff in
  28  * percpu_ref_kill. Additionally, we need a bias value so that the atomic_t
  29  * can't hit 0 before we've added up all the percpu refs.
  30  */
  31
  32 #define PCPU_COUNT_BIAS         (1U << 31)
  33
  34 static unsigned __percpu *pcpu_count_ptr(struct percpu_ref *ref)
  35 {
  36         return (unsigned __percpu *)(ref->pcpu_count_ptr & ~PCPU_REF_DEAD);
  37 }
  38
  39 /**
  40  * percpu_ref_init - initialize a percpu refcount
  41  * @ref: percpu_ref to initialize
  42  * @release: function which will be called when refcount hits 0
  43  * @gfp: allocation mask to use
  44  *
  45  * Initializes the refcount in single atomic counter mode with a refcount of 1;
  46  * analagous to atomic_set(ref, 1).
  47  *
  48  * Note that @release must not sleep - it may potentially be called from RCU
  49  * callback context by percpu_ref_kill().
  50  */
  51 int percpu_ref_init(struct percpu_ref *ref, percpu_ref_func_t *release,
  52                     gfp_t gfp)
  53 {
  54         atomic_set(&ref->count, 1 + PCPU_COUNT_BIAS);
  55
  56         ref->pcpu_count_ptr = (unsigned long)alloc_percpu_gfp(unsigned, gfp);
  57         if (!ref->pcpu_count_ptr)
  58                 return -ENOMEM;
  59
  60         ref->release = release;
  61         return 0;
  62 }
  63 EXPORT_SYMBOL_GPL(percpu_ref_init);
  64
  65 /**
  66  * percpu_ref_reinit - re-initialize a percpu refcount
  67  * @ref: perpcu_ref to re-initialize
  68  *
  69  * Re-initialize @ref so that it's in the same state as when it finished
  70  * percpu_ref_init().  @ref must have been initialized successfully, killed
  71  * and reached 0 but not exited.
  72  *
  73  * Note that percpu_ref_tryget[_live]() are safe to perform on @ref while
  74  * this function is in progress.
  75  */
  76 void percpu_ref_reinit(struct percpu_ref *ref)
  77 {
  78         unsigned __percpu *pcpu_count = pcpu_count_ptr(ref);
  79         int cpu;
  80
  81         BUG_ON(!pcpu_count);
  82         WARN_ON(!percpu_ref_is_zero(ref));
  83
  84         atomic_set(&ref->count, 1 + PCPU_COUNT_BIAS);
  85
  86         /*
  87          * Restore per-cpu operation.  smp_store_release() is paired with
  88          * smp_read_barrier_depends() in __pcpu_ref_alive() and guarantees
  89          * that the zeroing is visible to all percpu accesses which can see
  90          * the following PCPU_REF_DEAD clearing.
  91          */
  92         for_each_possible_cpu(cpu)
  93                 *per_cpu_ptr(pcpu_count, cpu) = 0;
  94
  95         smp_store_release(&ref->pcpu_count_ptr,
  96                           ref->pcpu_count_ptr & ~PCPU_REF_DEAD);
  97 }
  98 EXPORT_SYMBOL_GPL(percpu_ref_reinit);
  99
 100 /**
 101  * percpu_ref_exit - undo percpu_ref_init()
 102  * @ref: percpu_ref to exit
 103  *
 104  * This function exits @ref.  The caller is responsible for ensuring that
 105  * @ref is no longer in active use.  The usual places to invoke this
 106  * function from are the @ref->release() callback or in init failure path
 107  * where percpu_ref_init() succeeded but other parts of the initialization
 108  * of the embedding object failed.
 109  */
 110 void percpu_ref_exit(struct percpu_ref *ref)
 111 {
 112         unsigned __percpu *pcpu_count = pcpu_count_ptr(ref);
 113
 114         if (pcpu_count) {
 115                 free_percpu(pcpu_count);
 116                 ref->pcpu_count_ptr = PCPU_REF_DEAD;
 117         }
 118 }
 119 EXPORT_SYMBOL_GPL(percpu_ref_exit);
 120
 121 static void percpu_ref_kill_rcu(struct rcu_head *rcu)
 122 {
 123         struct percpu_ref *ref = container_of(rcu, struct percpu_ref, rcu);
 124         unsigned __percpu *pcpu_count = pcpu_count_ptr(ref);
 125         unsigned count = 0;
 126         int cpu;
 127
 128         for_each_possible_cpu(cpu)
 129                 count += *per_cpu_ptr(pcpu_count, cpu);
 130
 131         pr_debug("global %i pcpu %i", atomic_read(&ref->count), (int) count);
 132
 133         /*
 134          * It's crucial that we sum the percpu counters _before_ adding the sum
 135          * to &ref->count; since gets could be happening on one cpu while puts
 136          * happen on another, adding a single cpu's count could cause
 137          * @ref->count to hit 0 before we've got a consistent value - but the
 138          * sum of all the counts will be consistent and correct.
 139          *
 140          * Subtracting the bias value then has to happen _after_ adding count to
 141          * &ref->count; we need the bias value to prevent &ref->count from
 142          * reaching 0 before we add the percpu counts. But doing it at the same
 143          * time is equivalent and saves us atomic operations:
 144          */
 145
 146         atomic_add((int) count - PCPU_COUNT_BIAS, &ref->count);
 147
 148         WARN_ONCE(atomic_read(&ref->count) <= 0,
 149                   "percpu ref (%pf) <= 0 (%i) after killed",
 150                   ref->release, atomic_read(&ref->count));
 151
 152         /* @ref is viewed as dead on all CPUs, send out kill confirmation */
 153         if (ref->confirm_kill)
 154                 ref->confirm_kill(ref);
 155
 156         /*
 157          * Now we're in single atomic_t mode with a consistent refcount, so it's
 158          * safe to drop our initial ref:
 159          */
 160         percpu_ref_put(ref);
 161 }
 162
 163 /**
 164  * percpu_ref_kill_and_confirm - drop the initial ref and schedule confirmation
 165  * @ref: percpu_ref to kill
 166  * @confirm_kill: optional confirmation callback
 167  *
 168  * Equivalent to percpu_ref_kill() but also schedules kill confirmation if
 169  * @confirm_kill is not NULL.  @confirm_kill, which may not block, will be
 170  * called after @ref is seen as dead from all CPUs - all further
 171  * invocations of percpu_ref_tryget() will fail.  See percpu_ref_tryget()
 172  * for more details.
 173  *
 174  * Due to the way percpu_ref is implemented, @confirm_kill will be called
 175  * after at least one full RCU grace period has passed but this is an
 176  * implementation detail and callers must not depend on it.
 177  */
 178 void percpu_ref_kill_and_confirm(struct percpu_ref *ref,
 179                                  percpu_ref_func_t *confirm_kill)
 180 {
 181         WARN_ONCE(ref->pcpu_count_ptr & PCPU_REF_DEAD,
 182                   "percpu_ref_kill() called more than once on %pf!",
 183                   ref->release);
 184
 185         ref->pcpu_count_ptr |= PCPU_REF_DEAD;
 186         ref->confirm_kill = confirm_kill;
 187
 188         call_rcu_sched(&ref->rcu, percpu_ref_kill_rcu);
 189 }
 190 EXPORT_SYMBOL_GPL(percpu_ref_kill_and_confirm);