* operation.
* - The resize operation for larger tables (and available through an
* API) allows both expanding and shrinking the hash table.
- * - Per-CPU Split-counters are used to keep track of the number of
+ * - Split-counters are used to keep track of the number of
* nodes within the hash table for automatic resize triggering.
* - Resize operation initiated by long chain detection is executed by a
* call_rcu thread, which keeps lock-freedom of add and remove.
#endif
/*
- * Per-CPU split-counters lazily update the global counter each 1024
+ * Split-counters lazily update the global counter each 1024
* addition/removal. It automatically keeps track of resize required.
* We use the bucket length as indicator for need to expand for small
* tables and machines lacking per-cpu data suppport.
*/
#define COUNT_COMMIT_ORDER 10
+#define DEFAULT_SPLIT_COUNT_MASK 0xFUL
#define CHAIN_LEN_TARGET 1
#define CHAIN_LEN_RESIZE_THRESHOLD 3
/* Value of the end pointer. Should not interact with flags. */
#define END_VALUE NULL
+/*
+ * ht_items_count: Split-counters counting the number of node addition
+ * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag
+ * is set at hash table creation.
+ *
+ * These are free-running counters, never reset to zero. They count the
+ * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER)
+ * operations to update the global counter. We choose a power-of-2 value
+ * for the trigger to deal with 32 or 64-bit overflow of the counter.
+ */
struct ht_items_count {
unsigned long add, del;
} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
+/*
+ * rcu_level: Contains the per order-index-level dummy node table. The
+ * size of each dummy node table is half the number of hashes contained
+ * in this order (except for order 0). The minimum allocation size
+ * parameter allows combining the dummy node arrays of the lowermost
+ * levels to improve cache locality for small index orders.
+ */
struct rcu_level {
/* Note: manually update allocation length when adding a field */
- struct _cds_lfht_node nodes[0];
+ struct cds_lfht_node nodes[0];
};
+/*
+ * rcu_table: Contains the size and desired new size if a resize
+ * operation is in progress, as well as the statically-sized array of
+ * rcu_level pointers.
+ */
struct rcu_table {
unsigned long size; /* always a power of 2, shared (RCU) */
unsigned long resize_target;
struct rcu_level *tbl[MAX_TABLE_ORDER];
};
+/*
+ * cds_lfht: Top-level data structure representing a lock-free hash
+ * table. Defined in the implementation file to make it be an opaque
+ * cookie to users.
+ */
struct cds_lfht {
struct rcu_table t;
- cds_lfht_hash_fct hash_fct;
- cds_lfht_compare_fct compare_fct;
- unsigned long hash_seed;
+ unsigned long min_alloc_order;
+ unsigned long min_alloc_size;
int flags;
/*
* We need to put the work threads offline (QSBR) when taking this
void (*cds_lfht_rcu_unregister_thread)(void);
pthread_attr_t *resize_attr; /* Resize threads attributes */
long count; /* global approximate item count */
- struct ht_items_count *percpu_count; /* per-cpu item count */
+ struct ht_items_count *split_count; /* split item count */
};
+/*
+ * rcu_resize_work: Contains arguments passed to RCU worker thread
+ * responsible for performing lazy resize.
+ */
struct rcu_resize_work {
struct rcu_head head;
struct cds_lfht *ht;
};
+/*
+ * partition_resize_work: Contains arguments passed to worker threads
+ * executing the hash table resize on partitions of the hash table
+ * assigned to each processor's worker thread.
+ */
struct partition_resize_work {
pthread_t thread_id;
struct cds_lfht *ht;
static
void _cds_lfht_add(struct cds_lfht *ht,
+ cds_lfht_match_fct match,
+ void *key,
unsigned long size,
struct cds_lfht_node *node,
struct cds_lfht_iter *unique_ret,
unsigned int fls_ulong(unsigned long x)
{
-#if (CAA_BITS_PER_lONG == 32)
+#if (CAA_BITS_PER_LONG == 32)
return fls_u32(x);
#else
return fls_u64(x);
}
#ifdef POISON_FREE
-#define poison_free(ptr) \
- do { \
- memset(ptr, 0x42, sizeof(*(ptr))); \
- free(ptr); \
+#define poison_free(ptr) \
+ do { \
+ if (ptr) { \
+ memset(ptr, 0x42, sizeof(*(ptr))); \
+ free(ptr); \
+ } \
} while (0)
#else
#define poison_free(ptr) free(ptr)
#endif
static
-void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, 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)
+void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth);
static
void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
unsigned long count);
static long nr_cpus_mask = -1;
+static long split_count_mask = -1;
+
+#if defined(HAVE_SYSCONF)
+static void ht_init_nr_cpus_mask(void)
+{
+ long maxcpus;
+
+ maxcpus = sysconf(_SC_NPROCESSORS_CONF);
+ if (maxcpus <= 0) {
+ nr_cpus_mask = -2;
+ return;
+ }
+ /*
+ * 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;
+}
+#else /* #if defined(HAVE_SYSCONF) */
+static void ht_init_nr_cpus_mask(void)
+{
+ nr_cpus_mask = -2;
+}
+#endif /* #else #if defined(HAVE_SYSCONF) */
static
-struct ht_items_count *alloc_per_cpu_items_count(void)
+void alloc_split_items_count(struct cds_lfht *ht)
{
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;
+ if (nr_cpus_mask == -1) {
+ ht_init_nr_cpus_mask();
+ if (nr_cpus_mask < 0)
+ split_count_mask = DEFAULT_SPLIT_COUNT_MASK;
+ else
+ split_count_mask = nr_cpus_mask;
}
- /* Fall-through */
- default:
- return calloc(nr_cpus_mask + 1, sizeof(*count));
+
+ assert(split_count_mask >= 0);
+
+ if (ht->flags & CDS_LFHT_ACCOUNTING) {
+ ht->split_count = calloc(split_count_mask + 1, sizeof(*count));
+ assert(ht->split_count);
+ } else {
+ ht->split_count = NULL;
}
}
static
-void free_per_cpu_items_count(struct ht_items_count *count)
+void free_split_items_count(struct cds_lfht *ht)
{
- poison_free(count);
+ poison_free(ht->split_count);
}
+#if defined(HAVE_SCHED_GETCPU)
static
-int ht_get_cpu(void)
+int ht_get_split_count_index(unsigned long hash)
{
int cpu;
- assert(nr_cpus_mask >= 0);
+ assert(split_count_mask >= 0);
cpu = sched_getcpu();
- if (unlikely(cpu < 0))
- return cpu;
+ if (caa_unlikely(cpu < 0))
+ return hash & split_count_mask;
else
- return cpu & nr_cpus_mask;
+ return cpu & split_count_mask;
}
+#else /* #if defined(HAVE_SCHED_GETCPU) */
+static
+int ht_get_split_count_index(unsigned long hash)
+{
+ return hash & split_count_mask;
+}
+#endif /* #else #if defined(HAVE_SCHED_GETCPU) */
static
-void ht_count_add(struct cds_lfht *ht, unsigned long size)
+void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash)
{
- unsigned long percpu_count;
- int cpu;
+ unsigned long split_count;
+ int index;
- if (unlikely(!ht->percpu_count))
- return;
- cpu = ht_get_cpu();
- if (unlikely(cpu < 0))
+ if (caa_unlikely(!ht->split_count))
return;
- percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
- if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
+ index = ht_get_split_count_index(hash);
+ split_count = uatomic_add_return(&ht->split_count[index].add, 1);
+ if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
long count;
- dbg_printf("add percpu %lu\n", percpu_count);
+ dbg_printf("add split count %lu\n", split_count);
count = uatomic_add_return(&ht->count,
1UL << COUNT_COMMIT_ORDER);
/* If power of 2 */
}
static
-void ht_count_del(struct cds_lfht *ht, unsigned long size)
+void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash)
{
- unsigned long percpu_count;
- int cpu;
+ unsigned long split_count;
+ int index;
- if (unlikely(!ht->percpu_count))
- return;
- cpu = ht_get_cpu();
- if (unlikely(cpu < 0))
+ if (caa_unlikely(!ht->split_count))
return;
- percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1);
- if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
+ index = ht_get_split_count_index(hash);
+ split_count = uatomic_add_return(&ht->split_count[index].del, 1);
+ if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
long count;
- dbg_printf("del percpu %lu\n", percpu_count);
+ dbg_printf("del split count %lu\n", split_count);
count = uatomic_add_return(&ht->count,
-(1UL << COUNT_COMMIT_ORDER));
/* If power of 2 */
* Don't shrink table if the number of nodes is below a
* certain threshold.
*/
- if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1))
+ if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1))
return;
cds_lfht_resize_lazy_count(ht, size,
count >> (CHAIN_LEN_TARGET - 1));
}
}
-#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
-
-static const long nr_cpus_mask = -2;
-
-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, unsigned long size)
-{
-}
-
-static
-void ht_count_del(struct cds_lfht *ht, unsigned long size)
-{
-}
-
-#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
-
-
static
void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len)
{
dbg_printf("WARNING: large chain length: %u.\n",
chain_len);
if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
- cds_lfht_resize_lazy(ht, size,
+ cds_lfht_resize_lazy_grow(ht, size,
get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
}
}
static
-unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
+unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr,
+ unsigned long v)
{
unsigned long old1, old2;
if (old2 >= v)
return old2;
} while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2);
- return v;
+ return old2;
}
static
-struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
+struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
unsigned long hash)
{
unsigned long index, order;
assert(size > 0);
index = hash & (size - 1);
+
+ if (index < ht->min_alloc_size) {
+ dbg_printf("lookup hash %lu index %lu order 0 aridx 0\n",
+ hash, index);
+ return &ht->t.tbl[0]->nodes[index];
+ }
/*
* equivalent to get_count_order_ulong(index + 1), but optimizes
* away the non-existing 0 special-case for
* get_count_order_ulong.
*/
order = fls_ulong(index);
-
dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
- hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
-
- return &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
+ hash, index, order, index & ((1UL << (order - 1)) - 1));
+ return &ht->t.tbl[order]->nodes[index & ((1UL << (order - 1)) - 1)];
}
/*
for (;;) {
iter_prev = dummy;
/* We can always skip the dummy node initially */
- iter = rcu_dereference(iter_prev->p.next);
+ iter = rcu_dereference(iter_prev->next);
assert(!is_removed(iter));
- assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
+ assert(iter_prev->reverse_hash <= node->reverse_hash);
/*
* We should never be called with dummy (start of chain)
* and logically removed node (end of path compression
*/
assert(dummy != node);
for (;;) {
- if (unlikely(is_end(iter)))
+ if (caa_unlikely(is_end(iter)))
return;
- if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
+ if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
return;
- next = rcu_dereference(clear_flag(iter)->p.next);
- if (likely(is_removed(next)))
+ next = rcu_dereference(clear_flag(iter)->next);
+ if (caa_likely(is_removed(next)))
break;
iter_prev = clear_flag(iter);
iter = next;
new_next = flag_dummy(clear_flag(next));
else
new_next = clear_flag(next);
- (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
+ (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
}
return;
}
struct cds_lfht_node *old_next,
struct cds_lfht_node *new_node)
{
- struct cds_lfht_node *dummy, *ret_next;
- struct _cds_lfht_node *lookup;
+ struct cds_lfht_node *bucket, *ret_next;
if (!old_node) /* Return -ENOENT if asked to replace NULL node */
return -ENOENT;
}
assert(!is_dummy(old_next));
assert(new_node != clear_flag(old_next));
- new_node->p.next = clear_flag(old_next);
+ new_node->next = clear_flag(old_next);
/*
* Here is the whole trick for lock-free replace: we add
* the replacement node _after_ the node we want to
* to the removal flag and see the new node, or use
* the old node, but will not see the new one.
*/
- ret_next = uatomic_cmpxchg(&old_node->p.next,
+ ret_next = uatomic_cmpxchg(&old_node->next,
old_next, flag_removed(new_node));
if (ret_next == old_next)
break; /* We performed the replacement. */
* lookup for the node, and remove it (along with any other
* logically removed node) if found.
*/
- lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash));
- dummy = (struct cds_lfht_node *) lookup;
- _cds_lfht_gc_bucket(dummy, new_node);
+ bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash));
+ _cds_lfht_gc_bucket(bucket, new_node);
- assert(is_removed(rcu_dereference(old_node->p.next)));
+ assert(is_removed(rcu_dereference(old_node->next)));
return 0;
}
*/
static
void _cds_lfht_add(struct cds_lfht *ht,
+ cds_lfht_match_fct match,
+ void *key,
unsigned long size,
struct cds_lfht_node *node,
struct cds_lfht_iter *unique_ret,
{
struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
*return_node;
- struct _cds_lfht_node *lookup;
+ struct cds_lfht_node *bucket;
assert(!is_dummy(node));
assert(!is_removed(node));
- lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
+ bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
for (;;) {
uint32_t chain_len = 0;
* iter_prev points to the non-removed node prior to the
* insert location.
*/
- iter_prev = (struct cds_lfht_node *) lookup;
+ iter_prev = bucket;
/* We can always skip the dummy node initially */
- iter = rcu_dereference(iter_prev->p.next);
- assert(iter_prev->p.reverse_hash <= node->p.reverse_hash);
+ iter = rcu_dereference(iter_prev->next);
+ assert(iter_prev->reverse_hash <= node->reverse_hash);
for (;;) {
- if (unlikely(is_end(iter)))
+ if (caa_unlikely(is_end(iter)))
goto insert;
- if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
+ if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash))
goto insert;
/* dummy node is the first node of the identical-hash-value chain */
- if (dummy && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash)
+ if (dummy && clear_flag(iter)->reverse_hash == node->reverse_hash)
goto insert;
- next = rcu_dereference(clear_flag(iter)->p.next);
- if (unlikely(is_removed(next)))
+ next = rcu_dereference(clear_flag(iter)->next);
+ if (caa_unlikely(is_removed(next)))
goto gc_node;
/* uniquely add */
if (unique_ret
&& !is_dummy(next)
- && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) {
+ && clear_flag(iter)->reverse_hash == node->reverse_hash) {
struct cds_lfht_iter d_iter = { .node = node, .next = iter, };
/*
* (including observe one node by one node
* by forward iterations)
*/
- cds_lfht_next_duplicate(ht, &d_iter);
+ cds_lfht_next_duplicate(ht, match, key, &d_iter);
if (!d_iter.node)
goto insert;
}
/* Only account for identical reverse hash once */
- if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
+ if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash
&& !is_dummy(next))
check_resize(ht, size, ++chain_len);
iter_prev = clear_flag(iter);
assert(!is_removed(iter));
assert(iter_prev != node);
if (!dummy)
- node->p.next = clear_flag(iter);
+ node->next = clear_flag(iter);
else
- node->p.next = flag_dummy(clear_flag(iter));
+ node->next = flag_dummy(clear_flag(iter));
if (is_dummy(iter))
new_node = flag_dummy(node);
else
new_node = node;
- if (uatomic_cmpxchg(&iter_prev->p.next, iter,
+ if (uatomic_cmpxchg(&iter_prev->next, iter,
new_node) != iter) {
continue; /* retry */
} else {
new_next = flag_dummy(clear_flag(next));
else
new_next = clear_flag(next);
- (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
+ (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
/* retry */
}
end:
struct cds_lfht_node *node,
int dummy_removal)
{
- struct cds_lfht_node *dummy, *next, *old;
- struct _cds_lfht_node *lookup;
+ struct cds_lfht_node *bucket, *next, *old;
if (!node) /* Return -ENOENT if asked to delete NULL node */
return -ENOENT;
/* logically delete the node */
assert(!is_dummy(node));
assert(!is_removed(node));
- old = rcu_dereference(node->p.next);
+ old = rcu_dereference(node->next);
do {
struct cds_lfht_node *new_next;
next = old;
- if (unlikely(is_removed(next)))
+ if (caa_unlikely(is_removed(next)))
return -ENOENT;
if (dummy_removal)
assert(is_dummy(next));
else
assert(!is_dummy(next));
new_next = flag_removed(next);
- old = uatomic_cmpxchg(&node->p.next, next, new_next);
+ old = uatomic_cmpxchg(&node->next, next, new_next);
} while (old != next);
/* We performed the (logical) deletion. */
* the node, and remove it (along with any other logically removed node)
* if found.
*/
- lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
- dummy = (struct cds_lfht_node *) lookup;
- _cds_lfht_gc_bucket(dummy, node);
+ bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
+ _cds_lfht_gc_bucket(bucket, node);
- assert(is_removed(rcu_dereference(node->p.next)));
+ assert(is_removed(rcu_dereference(node->next)));
return 0;
}
{
unsigned long j;
- assert(i > 0);
+ assert(i > ht->min_alloc_order);
ht->cds_lfht_rcu_read_lock();
for (j = start; j < start + len; j++) {
- struct cds_lfht_node *new_node =
- (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
+ struct cds_lfht_node *new_node = &ht->t.tbl[i]->nodes[j];
dbg_printf("init populate: i %lu j %lu hash %lu\n",
i, j, (1UL << (i - 1)) + j);
- new_node->p.reverse_hash =
+ new_node->reverse_hash =
bit_reverse_ulong((1UL << (i - 1)) + j);
- _cds_lfht_add(ht, 1UL << (i - 1),
+ _cds_lfht_add(ht, NULL, NULL, 1UL << (i - 1),
new_node, NULL, 1);
}
ht->cds_lfht_rcu_read_unlock();
dbg_printf("init table: first_order %lu last_order %lu\n",
first_order, last_order);
- assert(first_order > 0);
+ assert(first_order > ht->min_alloc_order);
for (i = first_order; i <= last_order; i++) {
unsigned long len;
if (CMM_LOAD_SHARED(ht->t.resize_target) < (1UL << i))
break;
- ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node));
+ ht->t.tbl[i] = calloc(1, len * sizeof(struct cds_lfht_node));
assert(ht->t.tbl[i]);
/*
{
unsigned long j;
- assert(i > 0);
+ assert(i > ht->min_alloc_order);
ht->cds_lfht_rcu_read_lock();
for (j = start; j < start + len; j++) {
- struct cds_lfht_node *fini_node =
- (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
+ struct cds_lfht_node *fini_node = &ht->t.tbl[i]->nodes[j];
dbg_printf("remove entry: i %lu j %lu hash %lu\n",
i, j, (1UL << (i - 1)) + j);
- fini_node->p.reverse_hash =
+ fini_node->reverse_hash =
bit_reverse_ulong((1UL << (i - 1)) + j);
(void) _cds_lfht_del(ht, 1UL << (i - 1), fini_node, 1);
}
dbg_printf("fini table: first_order %lu last_order %lu\n",
first_order, last_order);
- assert(first_order > 0);
+ assert(first_order > ht->min_alloc_order);
for (i = last_order; i >= first_order; i--) {
unsigned long len;
static
void cds_lfht_create_dummy(struct cds_lfht *ht, unsigned long size)
{
- struct _cds_lfht_node *prev, *node;
+ struct cds_lfht_node *prev, *node;
unsigned long order, len, i, j;
- ht->t.tbl[0] = calloc(1, sizeof(struct _cds_lfht_node));
+ ht->t.tbl[0] = calloc(1, ht->min_alloc_size * sizeof(struct cds_lfht_node));
assert(ht->t.tbl[0]);
dbg_printf("create dummy: order %lu index %lu hash %lu\n", 0, 0, 0);
for (order = 1; order < get_count_order_ulong(size) + 1; order++) {
len = 1UL << (order - 1);
- ht->t.tbl[order] = calloc(1, len * sizeof(struct _cds_lfht_node));
- assert(ht->t.tbl[order]);
+ if (order <= ht->min_alloc_order) {
+ ht->t.tbl[order] = (struct rcu_level *) (ht->t.tbl[0]->nodes + len);
+ } else {
+ ht->t.tbl[order] = calloc(1, len * sizeof(struct cds_lfht_node));
+ assert(ht->t.tbl[order]);
+ }
i = 0;
prev = ht->t.tbl[i]->nodes;
node->next = prev->next;
assert(is_dummy(node->next));
node->reverse_hash = bit_reverse_ulong(j + len);
- prev->next = flag_dummy((struct cds_lfht_node *)node);
+ prev->next = flag_dummy(node);
}
}
}
-struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct,
- cds_lfht_compare_fct compare_fct,
- unsigned long hash_seed,
- unsigned long init_size,
+struct cds_lfht *_cds_lfht_new(unsigned long init_size,
+ unsigned long min_alloc_size,
int flags,
void (*cds_lfht_call_rcu)(struct rcu_head *head,
void (*func)(struct rcu_head *head)),
struct cds_lfht *ht;
unsigned long order;
+ /* min_alloc_size must be power of two */
+ if (!min_alloc_size || (min_alloc_size & (min_alloc_size - 1)))
+ return NULL;
/* init_size must be power of two */
- if (init_size && (init_size & (init_size - 1)))
+ if (!init_size || (init_size & (init_size - 1)))
return NULL;
+ min_alloc_size = max(min_alloc_size, MIN_TABLE_SIZE);
+ init_size = max(init_size, min_alloc_size);
ht = calloc(1, sizeof(struct cds_lfht));
assert(ht);
- ht->hash_fct = hash_fct;
- ht->compare_fct = compare_fct;
- ht->hash_seed = hash_seed;
+ ht->flags = flags;
ht->cds_lfht_call_rcu = cds_lfht_call_rcu;
ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu;
ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock;
ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread;
ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread;
ht->resize_attr = attr;
- ht->percpu_count = alloc_per_cpu_items_count();
+ alloc_split_items_count(ht);
/* this mutex should not nest in read-side C.S. */
pthread_mutex_init(&ht->resize_mutex, NULL);
- ht->flags = flags;
- order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE));
+ order = get_count_order_ulong(init_size);
ht->t.resize_target = 1UL << order;
+ ht->min_alloc_size = min_alloc_size;
+ ht->min_alloc_order = get_count_order_ulong(min_alloc_size);
cds_lfht_create_dummy(ht, 1UL << order);
ht->t.size = 1UL << order;
return ht;
}
-void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
- struct cds_lfht_iter *iter)
+void cds_lfht_lookup(struct cds_lfht *ht, cds_lfht_match_fct match,
+ unsigned long hash, void *key, struct cds_lfht_iter *iter)
{
- struct cds_lfht_node *node, *next, *dummy_node;
- struct _cds_lfht_node *lookup;
- unsigned long hash, reverse_hash, size;
+ struct cds_lfht_node *node, *next, *bucket;
+ unsigned long reverse_hash, size;
- hash = ht->hash_fct(key, key_len, ht->hash_seed);
reverse_hash = bit_reverse_ulong(hash);
size = rcu_dereference(ht->t.size);
- lookup = lookup_bucket(ht, size, hash);
- dummy_node = (struct cds_lfht_node *) lookup;
+ bucket = lookup_bucket(ht, size, hash);
/* We can always skip the dummy node initially */
- node = rcu_dereference(dummy_node->p.next);
+ node = rcu_dereference(bucket->next);
node = clear_flag(node);
for (;;) {
- if (unlikely(is_end(node))) {
+ if (caa_unlikely(is_end(node))) {
node = next = NULL;
break;
}
- if (unlikely(node->p.reverse_hash > reverse_hash)) {
+ if (caa_unlikely(node->reverse_hash > reverse_hash)) {
node = next = NULL;
break;
}
- next = rcu_dereference(node->p.next);
- if (likely(!is_removed(next))
+ next = rcu_dereference(node->next);
+ assert(node == clear_flag(node));
+ if (caa_likely(!is_removed(next))
&& !is_dummy(next)
- && clear_flag(node)->p.reverse_hash == reverse_hash
- && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
+ && node->reverse_hash == reverse_hash
+ && caa_likely(match(node, key))) {
break;
}
node = clear_flag(next);
}
- assert(!node || !is_dummy(rcu_dereference(node->p.next)));
+ assert(!node || !is_dummy(rcu_dereference(node->next)));
iter->node = node;
iter->next = next;
}
-void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
+void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match,
+ void *key, struct cds_lfht_iter *iter)
{
struct cds_lfht_node *node, *next;
unsigned long reverse_hash;
- void *key;
- size_t key_len;
node = iter->node;
- reverse_hash = node->p.reverse_hash;
- key = node->key;
- key_len = node->key_len;
+ reverse_hash = node->reverse_hash;
next = iter->next;
node = clear_flag(next);
for (;;) {
- if (unlikely(is_end(node))) {
+ if (caa_unlikely(is_end(node))) {
node = next = NULL;
break;
}
- if (unlikely(node->p.reverse_hash > reverse_hash)) {
+ if (caa_unlikely(node->reverse_hash > reverse_hash)) {
node = next = NULL;
break;
}
- next = rcu_dereference(node->p.next);
- if (likely(!is_removed(next))
+ next = rcu_dereference(node->next);
+ if (caa_likely(!is_removed(next))
&& !is_dummy(next)
- && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) {
+ && caa_likely(match(node, key))) {
break;
}
node = clear_flag(next);
}
- assert(!node || !is_dummy(rcu_dereference(node->p.next)));
+ assert(!node || !is_dummy(rcu_dereference(node->next)));
iter->node = node;
iter->next = next;
}
node = clear_flag(iter->next);
for (;;) {
- if (unlikely(is_end(node))) {
+ if (caa_unlikely(is_end(node))) {
node = next = NULL;
break;
}
- next = rcu_dereference(node->p.next);
- if (likely(!is_removed(next))
+ next = rcu_dereference(node->next);
+ if (caa_likely(!is_removed(next))
&& !is_dummy(next)) {
break;
}
node = clear_flag(next);
}
- assert(!node || !is_dummy(rcu_dereference(node->p.next)));
+ assert(!node || !is_dummy(rcu_dereference(node->next)));
iter->node = node;
iter->next = next;
}
void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter)
{
- struct _cds_lfht_node *lookup;
+ struct cds_lfht_node *lookup;
/*
* Get next after first dummy node. The first dummy node is the
cds_lfht_next(ht, iter);
}
-void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
+void cds_lfht_add(struct cds_lfht *ht, unsigned long hash,
+ struct cds_lfht_node *node)
{
- unsigned long hash, size;
-
- hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
- node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
+ unsigned long size;
+ node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
size = rcu_dereference(ht->t.size);
- _cds_lfht_add(ht, size, node, NULL, 0);
- ht_count_add(ht, size);
+ _cds_lfht_add(ht, NULL, NULL, size, node, NULL, 0);
+ ht_count_add(ht, size, hash);
}
struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
+ cds_lfht_match_fct match,
+ void *key,
+ unsigned long hash,
struct cds_lfht_node *node)
{
- unsigned long hash, size;
+ unsigned long size;
struct cds_lfht_iter iter;
- hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
- node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
-
+ node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
size = rcu_dereference(ht->t.size);
- _cds_lfht_add(ht, size, node, &iter, 0);
+ _cds_lfht_add(ht, match, key, size, node, &iter, 0);
if (iter.node == node)
- ht_count_add(ht, size);
+ ht_count_add(ht, size, hash);
return iter.node;
}
struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
+ cds_lfht_match_fct match,
+ void *key,
+ unsigned long hash,
struct cds_lfht_node *node)
{
- unsigned long hash, size;
+ unsigned long size;
struct cds_lfht_iter iter;
- hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed);
- node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
-
+ node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
size = rcu_dereference(ht->t.size);
for (;;) {
- _cds_lfht_add(ht, size, node, &iter, 0);
+ _cds_lfht_add(ht, match, key, size, node, &iter, 0);
if (iter.node == node) {
- ht_count_add(ht, size);
+ ht_count_add(ht, size, hash);
return NULL;
}
int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
{
- unsigned long size;
+ unsigned long size, hash;
int ret;
size = rcu_dereference(ht->t.size);
ret = _cds_lfht_del(ht, size, iter->node, 0);
- if (!ret)
- ht_count_del(ht, size);
+ if (!ret) {
+ hash = bit_reverse_ulong(iter->node->reverse_hash);
+ ht_count_del(ht, size, hash);
+ }
return ret;
}
int cds_lfht_delete_dummy(struct cds_lfht *ht)
{
struct cds_lfht_node *node;
- struct _cds_lfht_node *lookup;
unsigned long order, i, size;
/* Check that the table is empty */
- lookup = &ht->t.tbl[0]->nodes[0];
- node = (struct cds_lfht_node *) lookup;
+ node = &ht->t.tbl[0]->nodes[0];
do {
- node = clear_flag(node)->p.next;
+ node = clear_flag(node)->next;
if (!is_dummy(node))
return -EPERM;
assert(!is_removed(node));
bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash));
assert(is_dummy(ht->t.tbl[order]->nodes[i].next));
}
- poison_free(ht->t.tbl[order]);
+
+ if (order == ht->min_alloc_order)
+ poison_free(ht->t.tbl[0]);
+ else if (order > ht->min_alloc_order)
+ poison_free(ht->t.tbl[order]);
+ /* Nothing to delete for order < ht->min_alloc_order */
}
return 0;
}
ret = cds_lfht_delete_dummy(ht);
if (ret)
return ret;
- free_per_cpu_items_count(ht->percpu_count);
+ free_split_items_count(ht);
if (attr)
*attr = ht->resize_attr;
poison_free(ht);
long *approx_after)
{
struct cds_lfht_node *node, *next;
- struct _cds_lfht_node *lookup;
unsigned long nr_dummy = 0;
*approx_before = 0;
- if (nr_cpus_mask >= 0) {
+ if (ht->split_count) {
int i;
- for (i = 0; i < nr_cpus_mask + 1; i++) {
- *approx_before += uatomic_read(&ht->percpu_count[i].add);
- *approx_before -= uatomic_read(&ht->percpu_count[i].del);
+ for (i = 0; i < split_count_mask + 1; i++) {
+ *approx_before += uatomic_read(&ht->split_count[i].add);
+ *approx_before -= uatomic_read(&ht->split_count[i].del);
}
}
*removed = 0;
/* Count non-dummy nodes in the table */
- lookup = &ht->t.tbl[0]->nodes[0];
- node = (struct cds_lfht_node *) lookup;
+ node = &ht->t.tbl[0]->nodes[0];
do {
- next = rcu_dereference(node->p.next);
+ next = rcu_dereference(node->next);
if (is_removed(next)) {
if (!is_dummy(next))
(*removed)++;
} while (!is_end(node));
dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
*approx_after = 0;
- if (nr_cpus_mask >= 0) {
+ if (ht->split_count) {
int i;
- for (i = 0; i < nr_cpus_mask + 1; i++) {
- *approx_after += uatomic_read(&ht->percpu_count[i].add);
- *approx_after -= uatomic_read(&ht->percpu_count[i].del);
+ for (i = 0; i < split_count_mask + 1; i++) {
+ *approx_after += uatomic_read(&ht->split_count[i].add);
+ *approx_after -= uatomic_read(&ht->split_count[i].del);
}
}
}
{
unsigned long old_order, new_order;
- new_size = max(new_size, MIN_TABLE_SIZE);
+ new_size = max(new_size, ht->min_alloc_size);
old_order = get_count_order_ulong(old_size);
new_order = get_count_order_ulong(new_size);
dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
}
static
-unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size,
- int growth_order)
+unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size)
{
- return _uatomic_max(&ht->t.resize_target,
- size << growth_order);
+ return _uatomic_xchg_monotonic_increase(&ht->t.resize_target, new_size);
}
static
void resize_target_update_count(struct cds_lfht *ht,
unsigned long count)
{
- count = max(count, MIN_TABLE_SIZE);
+ count = max(count, ht->min_alloc_size);
uatomic_set(&ht->t.resize_target, count);
}
}
static
-void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth)
+void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht)
{
struct rcu_resize_work *work;
- unsigned long target_size;
- target_size = resize_target_update(ht, size, growth);
/* Store resize_target before read resize_initiated */
cmm_smp_mb();
- if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
+ if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
uatomic_inc(&ht->in_progress_resize);
cmm_smp_mb(); /* increment resize count before load destroy */
if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
}
}
-#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
+static
+void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth)
+{
+ unsigned long target_size = size << growth;
+
+ if (resize_target_grow(ht, target_size) >= target_size)
+ return;
+
+ __cds_lfht_resize_lazy_launch(ht);
+}
+/*
+ * We favor grow operations over shrink. A shrink operation never occurs
+ * if a grow operation is queued for lazy execution. A grow operation
+ * cancels any pending shrink lazy execution.
+ */
static
void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size,
unsigned long count)
{
- struct rcu_resize_work *work;
-
if (!(ht->flags & CDS_LFHT_AUTO_RESIZE))
return;
- resize_target_update_count(ht, count);
- /* Store resize_target before read resize_initiated */
- cmm_smp_mb();
- if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
- uatomic_inc(&ht->in_progress_resize);
- cmm_smp_mb(); /* increment resize count before load destroy */
- if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
- uatomic_dec(&ht->in_progress_resize);
+ count = max(count, ht->min_alloc_size);
+ if (count == size)
+ return; /* Already the right size, no resize needed */
+ if (count > size) { /* lazy grow */
+ if (resize_target_grow(ht, count) >= count)
return;
+ } else { /* lazy shrink */
+ for (;;) {
+ unsigned long s;
+
+ s = uatomic_cmpxchg(&ht->t.resize_target, size, count);
+ if (s == size)
+ break; /* no resize needed */
+ if (s > size)
+ return; /* growing is/(was just) in progress */
+ if (s <= count)
+ return; /* some other thread do shrink */
+ size = s;
}
- work = malloc(sizeof(*work));
- work->ht = ht;
- ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
- CMM_STORE_SHARED(ht->t.resize_initiated, 1);
}
+ __cds_lfht_resize_lazy_launch(ht);
}
-
-#endif
projects / urcu.git / blobdiff