rculfhash: count nodes (per-cpu) as expand criterion
authorMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Tue, 6 Sep 2011 17:44:30 +0000 (13:44 -0400)
committerMathieu Desnoyers <mathieu.desnoyers@efficios.com>
Tue, 6 Sep 2011 17:44:30 +0000 (13:44 -0400)
Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
rculfhash.c

index 174c11cf7babb4a35a7d533b3dd557ab221ae3dc..6bc25aed32e09817bcf74280c19bc0022789551c 100644 (file)
 #include <stdint.h>
 #include <string.h>
 
+#include "config.h"
 #include <urcu.h>
 #include <urcu-call-rcu.h>
 #include <urcu/arch.h>
 #define CHAIN_LEN_TARGET               4
 #define CHAIN_LEN_RESIZE_THRESHOLD     8
 
+/* Commit counter changes to global counter each 1024 steps */
+#define COUNT_COMMIT_ORDER             10
+
 #ifndef max
 #define max(a, b)      ((a) > (b) ? (a) : (b))
 #endif
 #define DUMMY_FLAG             (1UL << 1)
 #define FLAGS_MASK             ((1UL << 2) - 1)
 
+struct ht_items_count {
+       unsigned long v;
+} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
+
 struct rcu_table {
        unsigned long size;     /* always a power of 2 */
        unsigned long resize_target;
@@ -151,6 +159,8 @@ struct cds_lfht {
        unsigned int in_progress_resize, in_progress_destroy;
        void (*cds_lfht_call_rcu)(struct rcu_head *head,
                      void (*func)(struct rcu_head *head));
+       unsigned long count;            /* global approximate item count */
+       struct ht_items_count *percpu_count;    /* per-cpu item count */
 };
 
 struct rcu_resize_work {
@@ -353,10 +363,152 @@ int get_count_order_ulong(unsigned long x)
 static
 void cds_lfht_resize_lazy(struct cds_lfht *ht, struct rcu_table *t, int growth);
 
+/*
+ * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
+ * available, then we support hash table item accounting.
+ * In the unfortunate event the number of CPUs reported would be
+ * inaccurate, we use modulo arithmetic on the number of CPUs we got.
+ */
+
+#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
+
+static long nr_cpus_mask = -1;
+
+static
+struct ht_items_count *alloc_per_cpu_items_count(void)
+{
+       struct ht_items_count *count;
+
+       switch (nr_cpus_mask) {
+       case -2:
+               return NULL;
+       case -1:
+       {
+               long maxcpus;
+
+               maxcpus = sysconf(_SC_NPROCESSORS_CONF);
+               if (maxcpus <= 0) {
+                       nr_cpus_mask = -2;
+                       return NULL;
+               }
+               /*
+                * round up number of CPUs to next power of two, so we
+                * can use & for modulo.
+                */
+               maxcpus = 1UL << get_count_order_ulong(maxcpus);
+               nr_cpus_mask = maxcpus - 1;
+       }
+               /* Fall-through */
+       default:
+               return calloc(nr_cpus_mask + 1, sizeof(*count));
+       }
+}
+
+static
+void free_per_cpu_items_count(struct ht_items_count *count)
+{
+       free(count);
+}
+
+static
+int ht_get_cpu(void)
+{
+       int cpu;
+
+       assert(nr_cpus_mask >= 0);
+       cpu = sched_getcpu();
+       if (unlikely(cpu < 0))
+               return cpu;
+       else
+               return cpu & nr_cpus_mask;
+}
+
+static
+unsigned long ht_count_update(struct cds_lfht *ht, unsigned int value)
+{
+       int cpu;
+
+       if (unlikely(!ht->percpu_count))
+               return 0;
+       cpu = ht_get_cpu();
+       if (unlikely(cpu < 0))
+               return 0;
+       return uatomic_add_return(&ht->percpu_count[cpu].v, 1);
+}
+
+static
+void ht_count_add(struct cds_lfht *ht, struct rcu_table *t)
+{
+       unsigned long percpu_count;
+
+       percpu_count = ht_count_update(ht, 1);
+       if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
+               unsigned long count;
+
+               dbg_printf("add percpu %lu\n", percpu_count);
+               count = uatomic_add_return(&ht->count,
+                                          1UL << COUNT_COMMIT_ORDER);
+               /* If power of 2 */
+               if (!(count & (count - 1))) {
+                       dbg_printf("add global %lu\n", count);
+                       cds_lfht_resize_lazy(ht, t, 1);
+               }
+       }
+}
+
+static
+void ht_count_remove(struct cds_lfht *ht, struct rcu_table *t)
+{
+       unsigned long percpu_count;
+
+       percpu_count = ht_count_update(ht, -1);
+       if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
+               unsigned long count;
+
+               dbg_printf("remove percpu %lu\n", percpu_count);
+               count = uatomic_add_return(&ht->count,
+                                          1UL << COUNT_COMMIT_ORDER);
+               /* If power of 2 */
+               if (!(count & (count - 1))) {
+                       dbg_printf("remove global %lu\n", count);
+                       cds_lfht_resize_lazy(ht, t, -1);
+               }
+       }
+}
+
+#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
+
+static const long nr_cpus_mask = -1;
+
+static
+struct ht_items_count *alloc_per_cpu_items_count(void)
+{
+       return NULL;
+}
+
+static
+void free_per_cpu_items_count(struct ht_items_count *count)
+{
+}
+
+static
+void ht_count_add(struct cds_lfht *ht)
+{
+}
+
+static
+void ht_count_remove(struct cds_lfht *ht)
+{
+}
+
+#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
+
+
 static
 void check_resize(struct cds_lfht *ht, struct rcu_table *t,
                  uint32_t chain_len)
 {
+       return;         //TEST
        if (chain_len > 100)
                dbg_printf("WARNING: large chain length: %u.\n",
                           chain_len);
@@ -630,6 +782,7 @@ struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
        ht->hash_seed = hash_seed;
        ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
        ht->in_progress_resize = 0;
+       ht->percpu_count = alloc_per_cpu_items_count();
        /* this mutex should not nest in read-side C.S. */
        pthread_mutex_init(&ht->resize_mutex, NULL);
        order = get_count_order_ulong(max(init_size, 1)) + 1;
@@ -721,6 +874,7 @@ void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
 
        t = rcu_dereference(ht->t);
        (void) _cds_lfht_add(ht, t, node, 0, 0);
+       ht_count_add(ht, t);
 }
 
 struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
@@ -728,20 +882,28 @@ struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
 {
        struct rcu_table *t;
        unsigned long hash;
+       struct cds_lfht_node *ret;
 
        hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
        node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
 
        t = rcu_dereference(ht->t);
-       return _cds_lfht_add(ht, t, node, 1, 0);
+       ret = _cds_lfht_add(ht, t, node, 1, 0);
+       if (ret != node)
+               ht_count_add(ht, t);
+       return ret;
 }
 
 int cds_lfht_remove(struct cds_lfht *ht, struct cds_lfht_node *node)
 {
        struct rcu_table *t;
+       int ret;
 
        t = rcu_dereference(ht->t);
-       return _cds_lfht_remove(ht, t, node);
+       ret = _cds_lfht_remove(ht, t, node);
+       if (!ret)
+               ht_count_remove(ht, t);
+       return ret;
 }
 
 static
@@ -794,6 +956,7 @@ int cds_lfht_destroy(struct cds_lfht *ht)
        if (ret)
                return ret;
        free(ht->t);
+       free_per_cpu_items_count(ht->percpu_count);
        free(ht);
        return ret;
 }
@@ -851,7 +1014,7 @@ void _do_cds_lfht_resize(struct cds_lfht *ht)
        if (old_size == new_size)
                return;
        new_order = get_count_order_ulong(new_size) + 1;
-       dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
+       printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
               old_size, old_order, new_size, new_order);
        new_t = malloc(sizeof(struct cds_lfht)
                        + (new_order * sizeof(struct _cds_lfht_node *)));
This page took 0.029482 seconds and 4 git commands to generate.