#include <urcu-call-rcu.h>
#include <urcu/arch.h>
#include <urcu/uatomic.h>
-#include <urcu/jhash.h>
#include <urcu/compiler.h>
#include <urcu/rculfhash.h>
#include <stdio.h>
#define CHAIN_LEN_RESIZE_THRESHOLD 3
/*
- * Define the minimum table size. Protects against hash table resize overload
- * when too many entries are added quickly before the resize can complete.
- * This is especially the case if the table could be shrinked to a size of 1.
- * TODO: we might want to make the add/remove operations help the resize to
- * add or remove dummy nodes when a resize is ongoing to ensure upper-bound on
- * chain length.
+ * Define the minimum table size.
*/
-#define MIN_TABLE_SIZE 128
+#define MIN_TABLE_SIZE 1
#if (CAA_BITS_PER_LONG == 32)
#define MAX_TABLE_ORDER 32
#define MAX_TABLE_ORDER 64
#endif
+/*
+ * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink.
+ */
+#define MIN_PARTITION_PER_THREAD_ORDER 12
+#define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER)
+
#ifndef min
#define min(a, b) ((a) < (b) ? (a) : (b))
#endif
/*
* The removed flag needs to be updated atomically with the pointer.
+ * It indicates that no node must attach to the node scheduled for
+ * removal, and that node garbage collection must be performed.
* The dummy flag does not require to be updated atomically with the
* pointer, but it is added as a pointer low bit flag to save space.
*/
#define DUMMY_FLAG (1UL << 1)
#define FLAGS_MASK ((1UL << 2) - 1)
+/* Value of the end pointer. Should not interact with flags. */
+#define END_VALUE NULL
+
struct ht_items_count {
- unsigned long add, remove;
+ unsigned long add, del;
} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
struct rcu_level {
- struct rcu_head head;
+ /* Note: manually update allocation length when adding a field */
struct _cds_lfht_node nodes[0];
};
void (*cds_lfht_rcu_read_unlock)(void);
void (*cds_lfht_rcu_thread_offline)(void);
void (*cds_lfht_rcu_thread_online)(void);
- unsigned long count; /* global approximate item count */
+ void (*cds_lfht_rcu_register_thread)(void);
+ 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 cds_lfht *ht;
};
+struct partition_resize_work {
+ pthread_t thread_id;
+ struct cds_lfht *ht;
+ unsigned long i, start, len;
+ void (*fct)(struct cds_lfht *ht, unsigned long i,
+ unsigned long start, unsigned long len);
+};
+
+enum add_mode {
+ ADD_DEFAULT = 0,
+ ADD_UNIQUE = 1,
+ ADD_REPLACE = 2,
+};
+
+static
+struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
+ unsigned long size,
+ struct cds_lfht_node *node,
+ enum add_mode mode, int dummy);
+
/*
* Algorithm to reverse bits in a word by lookup table, extended to
* 64-bit words.
#endif
}
+/*
+ * Return the minimum order for which x <= (1UL << order).
+ * Return -1 if x is 0.
+ */
int get_count_order_u32(uint32_t x)
{
- int order;
+ if (!x)
+ return -1;
- order = fls_u32(x) - 1;
- if (x & (x - 1))
- order++;
- return order;
+ return fls_u32(x - 1);
}
+/*
+ * Return the minimum order for which x <= (1UL << order).
+ * Return -1 if x is 0.
+ */
int get_count_order_ulong(unsigned long x)
{
- int order;
+ if (!x)
+ return -1;
- order = fls_ulong(x) - 1;
- if (x & (x - 1))
- order++;
- return order;
+ return fls_ulong(x - 1);
}
#ifdef POISON_FREE
return;
percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1);
if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
- unsigned long count;
+ long count;
dbg_printf("add percpu %lu\n", percpu_count);
count = uatomic_add_return(&ht->count,
if (!(count & (count - 1))) {
if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size)
return;
- dbg_printf("add set global %lu\n", count);
+ dbg_printf("add set global %ld\n", count);
cds_lfht_resize_lazy_count(ht, size,
count >> (CHAIN_LEN_TARGET - 1));
}
}
static
-void ht_count_remove(struct cds_lfht *ht, unsigned long size)
+void ht_count_del(struct cds_lfht *ht, unsigned long size)
{
unsigned long percpu_count;
int cpu;
cpu = ht_get_cpu();
if (unlikely(cpu < 0))
return;
- percpu_count = uatomic_add_return(&ht->percpu_count[cpu].remove, -1);
+ percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1);
if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) {
- unsigned long count;
+ long count;
- dbg_printf("remove percpu %lu\n", percpu_count);
+ dbg_printf("del percpu %lu\n", percpu_count);
count = uatomic_add_return(&ht->count,
-(1UL << COUNT_COMMIT_ORDER));
/* If power of 2 */
if (!(count & (count - 1))) {
if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size)
return;
- dbg_printf("remove set global %lu\n", count);
+ dbg_printf("del set global %ld\n", count);
+ /*
+ * Don't shrink table if the number of nodes is below a
+ * certain threshold.
+ */
+ if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_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 = -1;
+static const long nr_cpus_mask = -2;
static
struct ht_items_count *alloc_per_cpu_items_count(void)
}
static
-void ht_count_remove(struct cds_lfht *ht, unsigned long size)
+void ht_count_del(struct cds_lfht *ht, unsigned long size)
{
}
{
return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG);
}
-
+
+static
+struct cds_lfht_node *get_end(void)
+{
+ return (struct cds_lfht_node *) END_VALUE;
+}
+
+static
+int is_end(struct cds_lfht_node *node)
+{
+ return clear_flag(node) == (struct cds_lfht_node *) END_VALUE;
+}
+
static
unsigned long _uatomic_max(unsigned long *ptr, unsigned long v)
{
}
static
-void cds_lfht_free_level(struct rcu_head *head)
+struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size,
+ unsigned long hash)
{
- struct rcu_level *l =
- caa_container_of(head, struct rcu_level, head);
- poison_free(l);
+ unsigned long index, order;
+
+ assert(size > 0);
+ index = hash & (size - 1);
+ /*
+ * 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))];
}
/*
*/
assert(dummy != node);
for (;;) {
- if (unlikely(!clear_flag(iter)))
+ if (unlikely(is_end(iter)))
return;
if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
return;
new_next = flag_dummy(clear_flag(next));
else
new_next = clear_flag(next);
+ if (is_removed(iter))
+ new_next = flag_removed(new_next);
(void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
}
+ return;
+}
+
+static
+int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size,
+ struct cds_lfht_node *old_node,
+ struct cds_lfht_node *ret_next,
+ struct cds_lfht_node *new_node)
+{
+ struct cds_lfht_node *dummy, *old_next;
+ struct _cds_lfht_node *lookup;
+ int flagged = 0;
+
+ if (!old_node) /* Return -ENOENT if asked to replace NULL node */
+ goto end;
+
+ assert(!is_removed(old_node));
+ assert(!is_dummy(old_node));
+ assert(!is_removed(new_node));
+ assert(!is_dummy(new_node));
+ assert(new_node != old_node);
+ do {
+ /* Insert after node to be replaced */
+ old_next = ret_next;
+ if (is_removed(old_next)) {
+ /*
+ * Too late, the old node has been removed under us
+ * between lookup and replace. Fail.
+ */
+ goto end;
+ }
+ assert(!is_dummy(old_next));
+ assert(new_node != clear_flag(old_next));
+ new_node->p.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
+ * replace by atomically setting its next pointer at the
+ * same time we set its removal flag. Given that
+ * the lookups/get next use an iterator aware of the
+ * next pointer, they will either skip the old node due
+ * 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,
+ old_next, flag_removed(new_node));
+ } while (ret_next != old_next);
+
+ /* We performed the replacement. */
+ flagged = 1;
+
+ /*
+ * Ensure that the old node is not visible to readers anymore:
+ * 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);
+end:
+ /*
+ * Only the flagging action indicated that we (and no other)
+ * replaced the node from the hash table.
+ */
+ if (flagged) {
+ assert(is_removed(rcu_dereference(old_node->p.next)));
+ return 0;
+ } else {
+ return -ENOENT;
+ }
}
static
struct cds_lfht_node *_cds_lfht_add(struct cds_lfht *ht,
unsigned long size,
struct cds_lfht_node *node,
- int unique, int dummy)
+ enum add_mode mode, int dummy)
{
struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next,
- *dummy_node;
+ *return_node;
struct _cds_lfht_node *lookup;
- unsigned long hash, index, order;
assert(!is_dummy(node));
assert(!is_removed(node));
if (!size) {
assert(dummy);
- node->p.next = flag_dummy(NULL);
+ node->p.next = flag_dummy(get_end());
return node; /* Initial first add (head) */
}
- hash = bit_reverse_ulong(node->p.reverse_hash);
+ lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
for (;;) {
uint32_t chain_len = 0;
* iter_prev points to the non-removed node prior to the
* insert location.
*/
- index = hash & (size - 1);
- order = get_count_order_ulong(index + 1);
- lookup = &ht->t.tbl[order]->nodes[index & ((!order ? 0 : (1UL << (order - 1))) - 1)];
iter_prev = (struct cds_lfht_node *) lookup;
/* 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);
for (;;) {
- /* TODO: check if removed */
- if (unlikely(!clear_flag(iter)))
+ if (unlikely(is_end(iter)))
goto insert;
- /* TODO: check if removed */
if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash))
goto insert;
next = rcu_dereference(clear_flag(iter)->p.next);
if (unlikely(is_removed(next)))
goto gc_node;
- if (unique
+ if ((mode == ADD_UNIQUE || mode == ADD_REPLACE)
&& !is_dummy(next)
&& !ht->compare_fct(node->key, node->key_len,
clear_flag(iter)->key,
- clear_flag(iter)->key_len))
- return clear_flag(iter);
+ clear_flag(iter)->key_len)) {
+ if (mode == ADD_UNIQUE)
+ return clear_flag(iter);
+ else /* mode == ADD_REPLACE */
+ goto replace;
+ }
/* Only account for identical reverse hash once */
if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash
&& !is_dummy(next))
iter_prev = clear_flag(iter);
iter = next;
}
+
insert:
assert(node != clear_flag(iter));
assert(!is_removed(iter_prev));
else
new_node = node;
if (uatomic_cmpxchg(&iter_prev->p.next, iter,
- new_node) != iter)
+ new_node) != iter) {
continue; /* retry */
- else
- goto gc_end;
+ } else {
+ if (mode == ADD_REPLACE)
+ return_node = NULL;
+ else /* ADD_DEFAULT and ADD_UNIQUE */
+ return_node = node;
+ goto end;
+ }
+
+ replace:
+
+ if (!_cds_lfht_replace(ht, size, clear_flag(iter), next,
+ node)) {
+ return_node = clear_flag(iter);
+ goto end; /* gc already done */
+ } else {
+ continue; /* retry */
+ }
+
gc_node:
assert(!is_removed(iter));
if (is_dummy(iter))
(void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next);
/* retry */
}
-gc_end:
- /* Garbage collect logically removed nodes in the bucket */
- index = hash & (size - 1);
- order = get_count_order_ulong(index + 1);
- lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
- dummy_node = (struct cds_lfht_node *) lookup;
- _cds_lfht_gc_bucket(dummy_node, node);
- return node;
+end:
+ return return_node;
}
static
-int _cds_lfht_remove(struct cds_lfht *ht, unsigned long size,
+int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
struct cds_lfht_node *node,
int dummy_removal)
{
struct cds_lfht_node *dummy, *next, *old;
struct _cds_lfht_node *lookup;
int flagged = 0;
- unsigned long hash, index, order;
+
+ if (!node) /* Return -ENOENT if asked to delete NULL node */
+ goto end;
/* logically delete the node */
assert(!is_dummy(node));
assert(!is_removed(node));
old = rcu_dereference(node->p.next);
do {
+ struct cds_lfht_node *new_next;
+
next = old;
if (unlikely(is_removed(next)))
goto end;
assert(is_dummy(next));
else
assert(!is_dummy(next));
- old = uatomic_cmpxchg(&node->p.next, next,
- flag_removed(next));
+ new_next = flag_removed(next);
+ old = uatomic_cmpxchg(&node->p.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.
*/
- hash = bit_reverse_ulong(node->p.reverse_hash);
- assert(size > 0);
- index = hash & (size - 1);
- order = get_count_order_ulong(index + 1);
- lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))];
+ lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash));
dummy = (struct cds_lfht_node *) lookup;
_cds_lfht_gc_bucket(dummy, node);
end:
if (flagged) {
assert(is_removed(rcu_dereference(node->p.next)));
return 0;
- } else
+ } else {
return -ENOENT;
+ }
}
static
-void init_table_hash(struct cds_lfht *ht, unsigned long i,
- unsigned long len)
+void *partition_resize_thread(void *arg)
{
- unsigned long j;
+ struct partition_resize_work *work = arg;
- for (j = 0; j < len; j++) {
- struct cds_lfht_node *new_node =
- (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
+ work->ht->cds_lfht_rcu_register_thread();
+ work->fct(work->ht, work->i, work->start, work->len);
+ work->ht->cds_lfht_rcu_unregister_thread();
+ return NULL;
+}
- dbg_printf("init hash entry: i %lu j %lu hash %lu\n",
- i, j, !i ? 0 : (1UL << (i - 1)) + j);
- new_node->p.reverse_hash =
- bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
- if (CMM_LOAD_SHARED(ht->in_progress_destroy))
- break;
+static
+void partition_resize_helper(struct cds_lfht *ht, unsigned long i,
+ unsigned long len,
+ void (*fct)(struct cds_lfht *ht, unsigned long i,
+ unsigned long start, unsigned long len))
+{
+ unsigned long partition_len;
+ struct partition_resize_work *work;
+ int thread, ret;
+ unsigned long nr_threads;
+
+ /*
+ * Note: nr_cpus_mask + 1 is always power of 2.
+ * We spawn just the number of threads we need to satisfy the minimum
+ * partition size, up to the number of CPUs in the system.
+ */
+ if (nr_cpus_mask > 0) {
+ nr_threads = min(nr_cpus_mask + 1,
+ len >> MIN_PARTITION_PER_THREAD_ORDER);
+ } else {
+ nr_threads = 1;
}
+ partition_len = len >> get_count_order_ulong(nr_threads);
+ work = calloc(nr_threads, sizeof(*work));
+ assert(work);
+ for (thread = 0; thread < nr_threads; thread++) {
+ work[thread].ht = ht;
+ work[thread].i = i;
+ work[thread].len = partition_len;
+ work[thread].start = thread * partition_len;
+ work[thread].fct = fct;
+ ret = pthread_create(&(work[thread].thread_id), ht->resize_attr,
+ partition_resize_thread, &work[thread]);
+ assert(!ret);
+ }
+ for (thread = 0; thread < nr_threads; thread++) {
+ ret = pthread_join(work[thread].thread_id, NULL);
+ assert(!ret);
+ }
+ free(work);
}
+/*
+ * Holding RCU read lock to protect _cds_lfht_add against memory
+ * reclaim that could be performed by other call_rcu worker threads (ABA
+ * problem).
+ *
+ * When we reach a certain length, we can split this population phase over
+ * many worker threads, based on the number of CPUs available in the system.
+ * This should therefore take care of not having the expand lagging behind too
+ * many concurrent insertion threads by using the scheduler's ability to
+ * schedule dummy node population fairly with insertions.
+ */
static
-void init_table_link(struct cds_lfht *ht, unsigned long i, unsigned long len)
+void init_table_populate_partition(struct cds_lfht *ht, unsigned long i,
+ unsigned long start, unsigned long len)
{
unsigned long j;
- ht->cds_lfht_rcu_thread_online();
ht->cds_lfht_rcu_read_lock();
- for (j = 0; j < len; j++) {
+ for (j = start; j < start + len; j++) {
struct cds_lfht_node *new_node =
(struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
- dbg_printf("init link: i %lu j %lu hash %lu\n",
+ dbg_printf("init populate: i %lu j %lu hash %lu\n",
i, j, !i ? 0 : (1UL << (i - 1)) + j);
+ new_node->p.reverse_hash =
+ bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
(void) _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)),
- new_node, 0, 1);
- if (CMM_LOAD_SHARED(ht->in_progress_destroy))
- break;
+ new_node, ADD_DEFAULT, 1);
}
ht->cds_lfht_rcu_read_unlock();
- ht->cds_lfht_rcu_thread_offline();
}
-/*
- * Holding RCU read lock to protect _cds_lfht_add against memory
- * reclaim that could be performed by other call_rcu worker threads (ABA
- * problem).
- */
+static
+void init_table_populate(struct cds_lfht *ht, unsigned long i,
+ unsigned long len)
+{
+ assert(nr_cpus_mask != -1);
+ if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
+ ht->cds_lfht_rcu_thread_online();
+ init_table_populate_partition(ht, i, 0, len);
+ ht->cds_lfht_rcu_thread_offline();
+ return;
+ }
+ partition_resize_helper(ht, i, len, init_table_populate_partition);
+}
+
static
void init_table(struct cds_lfht *ht,
unsigned long first_order, unsigned long len_order)
len = !i ? 1 : 1UL << (i - 1);
dbg_printf("init order %lu len: %lu\n", i, len);
- ht->t.tbl[i] = calloc(1, sizeof(struct rcu_level)
- + (len * sizeof(struct _cds_lfht_node)));
- /* Set all dummy nodes reverse hash values for a level */
- init_table_hash(ht, i, len);
+ /* Stop expand if the resize target changes under us */
+ if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i)))
+ break;
+
+ ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node));
+ assert(ht->t.tbl[i]);
/*
- * Link all dummy nodes into the table. Concurrent
- * add/remove are helping us.
+ * Set all dummy nodes reverse hash values for a level and
+ * link all dummy nodes into the table.
*/
- init_table_link(ht, i, len);
+ init_table_populate(ht, i, len);
/*
- * Update table size (after init for now, because no
- * concurrent updater help (TODO)).
+ * Update table size.
*/
cmm_smp_wmb(); /* populate data before RCU size */
CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i));
+
dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i));
if (CMM_LOAD_SHARED(ht->in_progress_destroy))
break;
}
}
+/*
+ * Holding RCU read lock to protect _cds_lfht_remove against memory
+ * reclaim that could be performed by other call_rcu worker threads (ABA
+ * problem).
+ * For a single level, we logically remove and garbage collect each node.
+ *
+ * As a design choice, we perform logical removal and garbage collection on a
+ * node-per-node basis to simplify this algorithm. We also assume keeping good
+ * cache locality of the operation would overweight possible performance gain
+ * that could be achieved by batching garbage collection for multiple levels.
+ * However, this would have to be justified by benchmarks.
+ *
+ * Concurrent removal and add operations are helping us perform garbage
+ * collection of logically removed nodes. We guarantee that all logically
+ * removed nodes have been garbage-collected (unlinked) before call_rcu is
+ * invoked to free a hole level of dummy nodes (after a grace period).
+ *
+ * Logical removal and garbage collection can therefore be done in batch or on a
+ * node-per-node basis, as long as the guarantee above holds.
+ *
+ * When we reach a certain length, we can split this removal over many worker
+ * threads, based on the number of CPUs available in the system. This should
+ * take care of not letting resize process lag behind too many concurrent
+ * updater threads actively inserting into the hash table.
+ */
static
-void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
+void remove_table_partition(struct cds_lfht *ht, unsigned long i,
+ unsigned long start, unsigned long len)
{
unsigned long j;
- ht->cds_lfht_rcu_thread_online();
ht->cds_lfht_rcu_read_lock();
- for (j = 0; j < len; j++) {
+ for (j = start; j < start + len; j++) {
struct cds_lfht_node *fini_node =
(struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j];
i, j, !i ? 0 : (1UL << (i - 1)) + j);
fini_node->p.reverse_hash =
bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j);
- (void) _cds_lfht_remove(ht, !i ? 0 : (1UL << (i - 1)),
+ (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)),
fini_node, 1);
- if (CMM_LOAD_SHARED(ht->in_progress_destroy))
- break;
}
ht->cds_lfht_rcu_read_unlock();
- ht->cds_lfht_rcu_thread_offline();
}
-/*
- * Holding RCU read lock to protect _cds_lfht_remove against memory
- * reclaim that could be performed by other call_rcu worker threads (ABA
- * problem).
- */
+static
+void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len)
+{
+
+ assert(nr_cpus_mask != -1);
+ if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) {
+ ht->cds_lfht_rcu_thread_online();
+ remove_table_partition(ht, i, 0, len);
+ ht->cds_lfht_rcu_thread_offline();
+ return;
+ }
+ partition_resize_helper(ht, i, len, remove_table_partition);
+}
+
static
void fini_table(struct cds_lfht *ht,
unsigned long first_order, unsigned long len_order)
{
long i, end_order;
+ void *free_by_rcu = NULL;
dbg_printf("fini table: first_order %lu end_order %lu\n",
first_order, first_order + len_order);
end_order = first_order + len_order;
assert(first_order > 0);
- assert(ht->t.size == (1UL << (first_order - 1)));
for (i = end_order - 1; i >= first_order; i--) {
unsigned long len;
len = !i ? 1 : 1UL << (i - 1);
dbg_printf("fini order %lu len: %lu\n", i, len);
+ /* Stop shrink if the resize target changes under us */
+ if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1)))
+ break;
+
+ cmm_smp_wmb(); /* populate data before RCU size */
+ CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1));
+
+ /*
+ * We need to wait for all add operations to reach Q.S. (and
+ * thus use the new table for lookups) before we can start
+ * releasing the old dummy nodes. Otherwise their lookup will
+ * return a logically removed node as insert position.
+ */
+ ht->cds_lfht_synchronize_rcu();
+ if (free_by_rcu)
+ free(free_by_rcu);
+
/*
* Set "removed" flag in dummy nodes about to be removed.
* Unlink all now-logically-removed dummy node pointers.
*/
remove_table(ht, i, len);
- ht->cds_lfht_call_rcu(&ht->t.tbl[i]->head, cds_lfht_free_level);
+ free_by_rcu = ht->t.tbl[i];
dbg_printf("fini new size: %lu\n", 1UL << i);
if (CMM_LOAD_SHARED(ht->in_progress_destroy))
break;
}
+
+ if (free_by_rcu) {
+ ht->cds_lfht_synchronize_rcu();
+ free(free_by_rcu);
+ }
}
-struct cds_lfht *cds_lfht_new(cds_lfht_hash_fct hash_fct,
+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,
void (*cds_lfht_rcu_read_lock)(void),
void (*cds_lfht_rcu_read_unlock)(void),
void (*cds_lfht_rcu_thread_offline)(void),
- void (*cds_lfht_rcu_thread_online)(void))
+ void (*cds_lfht_rcu_thread_online)(void),
+ void (*cds_lfht_rcu_register_thread)(void),
+ void (*cds_lfht_rcu_unregister_thread)(void),
+ pthread_attr_t *attr)
{
struct cds_lfht *ht;
unsigned long order;
if (init_size && (init_size & (init_size - 1)))
return NULL;
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->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock;
ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline;
ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online;
+ 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();
/* this mutex should not nest in read-side C.S. */
pthread_mutex_init(&ht->resize_mutex, NULL);
ht->flags = flags;
ht->cds_lfht_rcu_thread_offline();
pthread_mutex_lock(&ht->resize_mutex);
+ ht->t.resize_target = 1UL << (order - 1);
init_table(ht, 0, order);
pthread_mutex_unlock(&ht->resize_mutex);
ht->cds_lfht_rcu_thread_online();
return ht;
}
-struct cds_lfht_node *cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len)
+void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len,
+ struct cds_lfht_iter *iter)
{
- struct cds_lfht_node *node, *next;
+ struct cds_lfht_node *node, *next, *dummy_node;
struct _cds_lfht_node *lookup;
- unsigned long hash, reverse_hash, index, order, size;
+ unsigned long hash, 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);
- index = hash & (size - 1);
- order = get_count_order_ulong(index + 1);
- lookup = &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1))) - 1)];
- dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n",
- hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1)));
- node = (struct cds_lfht_node *) lookup;
+ lookup = lookup_bucket(ht, size, hash);
+ dummy_node = (struct cds_lfht_node *) lookup;
+ /* We can always skip the dummy node initially */
+ node = rcu_dereference(dummy_node->p.next);
+ node = clear_flag(node);
for (;;) {
- if (unlikely(!node))
+ if (unlikely(is_end(node))) {
+ node = next = NULL;
break;
+ }
if (unlikely(node->p.reverse_hash > reverse_hash)) {
- node = NULL;
+ node = next = NULL;
break;
}
next = rcu_dereference(node->p.next);
node = clear_flag(next);
}
assert(!node || !is_dummy(rcu_dereference(node->p.next)));
- return node;
+ iter->node = node;
+ iter->next = next;
}
-struct cds_lfht_node *cds_lfht_next(struct cds_lfht *ht,
- struct cds_lfht_node *node)
+void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter)
{
- struct cds_lfht_node *next;
+ 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;
- next = rcu_dereference(node->p.next);
+ next = iter->next;
node = clear_flag(next);
for (;;) {
- if (unlikely(!node))
+ if (unlikely(is_end(node))) {
+ node = next = NULL;
break;
+ }
if (unlikely(node->p.reverse_hash > reverse_hash)) {
- node = NULL;
+ node = next = NULL;
break;
}
next = rcu_dereference(node->p.next);
node = clear_flag(next);
}
assert(!node || !is_dummy(rcu_dereference(node->p.next)));
- return node;
+ iter->node = node;
+ iter->next = next;
+}
+
+void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter)
+{
+ struct cds_lfht_node *node, *next;
+
+ node = clear_flag(iter->next);
+ for (;;) {
+ if (unlikely(is_end(node))) {
+ node = next = NULL;
+ break;
+ }
+ next = rcu_dereference(node->p.next);
+ if (likely(!is_removed(next))
+ && !is_dummy(next)) {
+ break;
+ }
+ node = clear_flag(next);
+ }
+ assert(!node || !is_dummy(rcu_dereference(node->p.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;
+
+ /*
+ * Get next after first dummy node. The first dummy node is the
+ * first node of the linked list.
+ */
+ lookup = &ht->t.tbl[0]->nodes[0];
+ iter->next = lookup->next;
+ cds_lfht_next(ht, iter);
}
void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node)
node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
size = rcu_dereference(ht->t.size);
- (void) _cds_lfht_add(ht, size, node, 0, 0);
+ (void) _cds_lfht_add(ht, size, node, ADD_DEFAULT, 0);
ht_count_add(ht, size);
}
struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht,
- struct cds_lfht_node *node)
+ struct cds_lfht_node *node)
+{
+ unsigned long hash, size;
+ 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);
+
+ size = rcu_dereference(ht->t.size);
+ ret = _cds_lfht_add(ht, size, node, ADD_UNIQUE, 0);
+ if (ret == node)
+ ht_count_add(ht, size);
+ return ret;
+}
+
+struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht,
+ struct cds_lfht_node *node)
{
unsigned long hash, size;
struct cds_lfht_node *ret;
node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash);
size = rcu_dereference(ht->t.size);
- ret = _cds_lfht_add(ht, size, node, 1, 0);
- if (ret != node)
+ ret = _cds_lfht_add(ht, size, node, ADD_REPLACE, 0);
+ if (ret == NULL)
ht_count_add(ht, size);
return ret;
}
-int cds_lfht_remove(struct cds_lfht *ht, struct cds_lfht_node *node)
+int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
+ struct cds_lfht_node *new_node)
+{
+ unsigned long size;
+
+ size = rcu_dereference(ht->t.size);
+ return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next,
+ new_node);
+}
+
+int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter)
{
unsigned long size;
int ret;
size = rcu_dereference(ht->t.size);
- ret = _cds_lfht_remove(ht, size, node, 0);
+ ret = _cds_lfht_del(ht, size, iter->node, 0);
if (!ret)
- ht_count_remove(ht, size);
+ ht_count_del(ht, size);
return ret;
}
if (!is_dummy(node))
return -EPERM;
assert(!is_removed(node));
- } while (clear_flag(node));
+ } while (!is_end(node));
/*
* size accessed without rcu_dereference because hash table is
* being destroyed.
* Should only be called when no more concurrent readers nor writers can
* possibly access the table.
*/
-int cds_lfht_destroy(struct cds_lfht *ht)
+int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr)
{
int ret;
/* Wait for in-flight resize operations to complete */
- CMM_STORE_SHARED(ht->in_progress_destroy, 1);
+ _CMM_STORE_SHARED(ht->in_progress_destroy, 1);
+ cmm_smp_mb(); /* Store destroy before load resize */
while (uatomic_read(&ht->in_progress_resize))
poll(NULL, 0, 100); /* wait for 100ms */
ret = cds_lfht_delete_dummy(ht);
if (ret)
return ret;
free_per_cpu_items_count(ht->percpu_count);
+ if (attr)
+ *attr = ht->resize_attr;
poison_free(ht);
return ret;
}
void cds_lfht_count_nodes(struct cds_lfht *ht,
+ long *approx_before,
unsigned long *count,
- unsigned long *removed)
+ unsigned long *removed,
+ 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) {
+ 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);
+ }
+ }
+
*count = 0;
*removed = 0;
do {
next = rcu_dereference(node->p.next);
if (is_removed(next)) {
- assert(!is_dummy(next));
- (*removed)++;
+ if (!is_dummy(next))
+ (*removed)++;
+ else
+ (nr_dummy)++;
} else if (!is_dummy(next))
(*count)++;
else
(nr_dummy)++;
node = clear_flag(next);
- } while (node);
+ } while (!is_end(node));
dbg_printf("number of dummy nodes: %lu\n", nr_dummy);
+ *approx_after = 0;
+ if (nr_cpus_mask >= 0) {
+ 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);
+ }
+ }
}
/* called with resize mutex held */
old_order = get_count_order_ulong(old_size) + 1;
new_order = get_count_order_ulong(new_size) + 1;
- printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
- old_size, old_order, new_size, new_order);
+ dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
+ old_size, old_order, new_size, new_order);
assert(new_size > old_size);
init_table(ht, old_order, new_order - old_order);
}
new_size = max(new_size, MIN_TABLE_SIZE);
old_order = get_count_order_ulong(old_size) + 1;
new_order = get_count_order_ulong(new_size) + 1;
- printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
- old_size, old_order, new_size, new_order);
+ dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n",
+ old_size, old_order, new_size, new_order);
assert(new_size < old_size);
- cmm_smp_wmb(); /* populate data before RCU size */
- CMM_STORE_SHARED(ht->t.size, new_size);
-
- /*
- * We need to wait for all add operations to reach Q.S. (and
- * thus use the new table for lookups) before we can start
- * releasing the old dummy nodes. Otherwise their lookup will
- * return a logically removed node as insert position.
- */
- ht->cds_lfht_synchronize_rcu();
-
/* Remove and unlink all dummy nodes to remove. */
fini_table(ht, new_order, old_order - new_order);
}
* Resize table, re-do if the target size has changed under us.
*/
do {
+ assert(uatomic_read(&ht->in_progress_resize));
+ if (CMM_LOAD_SHARED(ht->in_progress_destroy))
+ break;
ht->t.resize_initiated = 1;
old_size = ht->t.size;
new_size = CMM_LOAD_SHARED(ht->t.resize_target);
ht->t.resize_initiated = 0;
/* write resize_initiated before read resize_target */
cmm_smp_mb();
- } while (new_size != CMM_LOAD_SHARED(ht->t.resize_target));
+ } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target));
}
static
cmm_smp_mb();
if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) {
uatomic_inc(&ht->in_progress_resize);
- cmm_smp_mb(); /* increment resize count before calling it */
+ cmm_smp_mb(); /* increment resize count before load destroy */
+ if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
+ uatomic_dec(&ht->in_progress_resize);
+ return;
+ }
work = malloc(sizeof(*work));
work->ht = ht;
ht->cds_lfht_call_rcu(&work->head, do_resize_cb);
cmm_smp_mb();
if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) {
uatomic_inc(&ht->in_progress_resize);
- cmm_smp_mb(); /* increment resize count before calling it */
+ cmm_smp_mb(); /* increment resize count before load destroy */
+ if (CMM_LOAD_SHARED(ht->in_progress_destroy)) {
+ uatomic_dec(&ht->in_progress_resize);
+ return;
+ }
work = malloc(sizeof(*work));
work->ht = ht;
ht->cds_lfht_call_rcu(&work->head, do_resize_cb);