* (not visible to lookups anymore) before the RCU read-side critical
* section held across removal ends. Furthermore, this ensures that
* the node with "removed" flag set is removed from the linked-list
- * before its memory is reclaimed. Only the thread which removal
- * successfully set the "removed" flag (with a cmpxchg) into a node's
- * next pointer is considered to have succeeded its removal (and thus
- * owns the node to reclaim). Because we garbage-collect starting from
- * an invariant node (the start-of-bucket bucket node) up to the
- * "removed" node (or find a reverse-hash that is higher), we are sure
- * that a successful traversal of the chain leads to a chain that is
- * present in the linked-list (the start node is never removed) and
- * that is does not contain the "removed" node anymore, even if
- * concurrent delete/add operations are changing the structure of the
- * list concurrently.
- * - The add operation performs gargage collection of buckets if it
+ * before its memory is reclaimed. After setting the "removal" flag,
+ * only the thread which removal is the first to set the "removal
+ * owner" flag (with an xchg) into a node's next pointer is considered
+ * to have succeeded its removal (and thus owns the node to reclaim).
+ * Because we garbage-collect starting from an invariant node (the
+ * start-of-bucket bucket node) up to the "removed" node (or find a
+ * reverse-hash that is higher), we are sure that a successful
+ * traversal of the chain leads to a chain that is present in the
+ * linked-list (the start node is never removed) and that is does not
+ * contain the "removed" node anymore, even if concurrent delete/add
+ * operations are changing the structure of the list concurrently.
+ * - The add operation performs garbage collection of buckets if it
* encounters nodes with removed flag set in the bucket where it wants
* to add its new node. This ensures lock-freedom of add operation by
* helping the remover unlink nodes from the list rather than to wait
* hash table nodes. These tables are invariant after they are
* populated into the hash table.
*
+ * Ordering Guarantees:
+ *
+ * To discuss these guarantees, we first define "read" operation as any
+ * of the the basic cds_lfht_lookup, cds_lfht_next_duplicate,
+ * cds_lfht_first, cds_lfht_next operation, as well as
+ * cds_lfht_add_unique (failure).
+ *
+ * We define "read traversal" operation as any of the following
+ * group of operations
+ * - cds_lfht_lookup followed by iteration with cds_lfht_next_duplicate
+ * (and/or cds_lfht_next, although less common).
+ * - cds_lfht_add_unique (failure) followed by iteration with
+ * cds_lfht_next_duplicate (and/or cds_lfht_next, although less
+ * common).
+ * - cds_lfht_first followed iteration with cds_lfht_next (and/or
+ * cds_lfht_next_duplicate, although less common).
+ *
+ * We define "write" operations as any of cds_lfht_add,
+ * cds_lfht_add_unique (success), cds_lfht_add_replace, cds_lfht_del.
+ *
+ * When cds_lfht_add_unique succeeds (returns the node passed as
+ * parameter), it acts as a "write" operation. When cds_lfht_add_unique
+ * fails (returns a node different from the one passed as parameter), it
+ * acts as a "read" operation. A cds_lfht_add_unique failure is a
+ * cds_lfht_lookup "read" operation, therefore, any ordering guarantee
+ * referring to "lookup" imply any of "lookup" or cds_lfht_add_unique
+ * (failure).
+ *
+ * We define "prior" and "later" node as nodes observable by reads and
+ * read traversals respectively before and after a write or sequence of
+ * write operations.
+ *
+ * Hash-table operations are often cascaded, for example, the pointer
+ * returned by a cds_lfht_lookup() might be passed to a cds_lfht_next(),
+ * whose return value might in turn be passed to another hash-table
+ * operation. This entire cascaded series of operations must be enclosed
+ * by a pair of matching rcu_read_lock() and rcu_read_unlock()
+ * operations.
+ *
+ * The following ordering guarantees are offered by this hash table:
+ *
+ * A.1) "read" after "write": if there is ordering between a write and a
+ * later read, then the read is guaranteed to see the write or some
+ * later write.
+ * A.2) "read traversal" after "write": given that there is dependency
+ * ordering between reads in a "read traversal", if there is
+ * ordering between a write and the first read of the traversal,
+ * then the "read traversal" is guaranteed to see the write or
+ * some later write.
+ * B.1) "write" after "read": if there is ordering between a read and a
+ * later write, then the read will never see the write.
+ * B.2) "write" after "read traversal": given that there is dependency
+ * ordering between reads in a "read traversal", if there is
+ * ordering between the last read of the traversal and a later
+ * write, then the "read traversal" will never see the write.
+ * C) "write" while "read traversal": if a write occurs during a "read
+ * traversal", the traversal may, or may not, see the write.
+ * D.1) "write" after "write": if there is ordering between a write and
+ * a later write, then the later write is guaranteed to see the
+ * effects of the first write.
+ * D.2) Concurrent "write" pairs: The system will assign an arbitrary
+ * order to any pair of concurrent conflicting writes.
+ * Non-conflicting writes (for example, to different keys) are
+ * unordered.
+ * E) If a grace period separates a "del" or "replace" operation
+ * and a subsequent operation, then that subsequent operation is
+ * guaranteed not to see the removed item.
+ * F) Uniqueness guarantee: given a hash table that does not contain
+ * duplicate items for a given key, there will only be one item in
+ * the hash table after an arbitrary sequence of add_unique and/or
+ * add_replace operations. Note, however, that a pair of
+ * concurrent read operations might well access two different items
+ * with that key.
+ * G.1) If a pair of lookups for a given key are ordered (e.g. by a
+ * memory barrier), then the second lookup will return the same
+ * node as the previous lookup, or some later node.
+ * G.2) A "read traversal" that starts after the end of a prior "read
+ * traversal" (ordered by memory barriers) is guaranteed to see the
+ * same nodes as the previous traversal, or some later nodes.
+ * G.3) Concurrent "read" pairs: concurrent reads are unordered. For
+ * example, if a pair of reads to the same key run concurrently
+ * with an insertion of that same key, the reads remain unordered
+ * regardless of their return values. In other words, you cannot
+ * rely on the values returned by the reads to deduce ordering.
+ *
+ * Progress guarantees:
+ *
+ * * Reads are wait-free. These operations always move forward in the
+ * hash table linked list, and this list has no loop.
+ * * Writes are lock-free. Any retry loop performed by a write operation
+ * is triggered by progress made within another update operation.
+ *
* Bucket node tables:
*
* hash table hash table the last all bucket node tables
*/
#define _LGPL_SOURCE
+#define _GNU_SOURCE
#include <stdlib.h>
#include <errno.h>
#include <assert.h>
#include <stdio.h>
#include <stdint.h>
#include <string.h>
+#include <sched.h>
#include "config.h"
#include <urcu.h>
*/
#define REMOVED_FLAG (1UL << 0)
#define BUCKET_FLAG (1UL << 1)
-#define FLAGS_MASK ((1UL << 2) - 1)
+#define REMOVAL_OWNER_FLAG (1UL << 2)
+#define FLAGS_MASK ((1UL << 3) - 1)
/* Value of the end pointer. Should not interact with flags. */
#define END_VALUE NULL
}
#endif
-unsigned int fls_ulong(unsigned long x)
+unsigned int cds_lfht_fls_ulong(unsigned long x)
{
#if (CAA_BITS_PER_LONG == 32)
return fls_u32(x);
* Return the minimum order for which x <= (1UL << order).
* Return -1 if x is 0.
*/
-int get_count_order_u32(uint32_t x)
+int cds_lfht_get_count_order_u32(uint32_t x)
{
if (!x)
return -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 cds_lfht_get_count_order_ulong(unsigned long x)
{
if (!x)
return -1;
- return fls_ulong(x - 1);
+ return cds_lfht_fls_ulong(x - 1);
}
static
* round up number of CPUs to next power of two, so we
* can use & for modulo.
*/
- maxcpus = 1UL << get_count_order_ulong(maxcpus);
+ maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus);
nr_cpus_mask = maxcpus - 1;
}
#else /* #if defined(HAVE_SYSCONF) */
chain_len);
if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD)
cds_lfht_resize_lazy_grow(ht, size,
- get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
+ cds_lfht_get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1)));
}
static
return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG);
}
+static
+int is_removal_owner(struct cds_lfht_node *node)
+{
+ return ((unsigned long) node) & REMOVAL_OWNER_FLAG;
+}
+
+static
+struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node)
+{
+ return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG);
+}
+
static
struct cds_lfht_node *get_end(void)
{
new_next = clear_flag(next);
(void) uatomic_cmpxchg(&iter_prev->next, iter, new_next);
}
- return;
}
static
*/
return -ENOENT;
}
- assert(!is_bucket(old_next));
- assert(new_node != clear_flag(old_next));
- new_node->next = clear_flag(old_next);
+ assert(old_next == clear_flag(old_next));
+ assert(new_node != old_next);
+ new_node->next = old_next;
/*
* Here is the whole trick for lock-free replace: we add
* the replacement node _after_ the node we want to
* 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.
+ * This is a replacement of a node with another node
+ * that has the same value: we are therefore not
+ * removing a value from the hash table.
*/
ret_next = uatomic_cmpxchg(&old_node->next,
old_next, flag_removed(new_node));
*/
static
void _cds_lfht_add(struct cds_lfht *ht,
+ unsigned long hash,
cds_lfht_match_fct match,
const void *key,
unsigned long size,
assert(!is_bucket(node));
assert(!is_removed(node));
- bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash));
+ bucket = lookup_bucket(ht, size, hash);
for (;;) {
uint32_t chain_len = 0;
static
int _cds_lfht_del(struct cds_lfht *ht, unsigned long size,
- struct cds_lfht_node *node,
- int bucket_removal)
+ struct cds_lfht_node *node)
{
- struct cds_lfht_node *bucket, *next, *old;
+ struct cds_lfht_node *bucket, *next;
if (!node) /* Return -ENOENT if asked to delete NULL node */
return -ENOENT;
/* logically delete the node */
assert(!is_bucket(node));
assert(!is_removed(node));
- old = rcu_dereference(node->next);
- do {
- struct cds_lfht_node *new_next;
+ assert(!is_removal_owner(node));
- next = old;
- if (caa_unlikely(is_removed(next)))
- return -ENOENT;
- if (bucket_removal)
- assert(is_bucket(next));
- else
- assert(!is_bucket(next));
- new_next = flag_removed(next);
- old = uatomic_cmpxchg(&node->next, next, new_next);
- } while (old != next);
+ /*
+ * We are first checking if the node had previously been
+ * logically removed (this check is not atomic with setting the
+ * logical removal flag). Return -ENOENT if the node had
+ * previously been removed.
+ */
+ next = rcu_dereference(node->next);
+ if (caa_unlikely(is_removed(next)))
+ return -ENOENT;
+ assert(!is_bucket(next));
+ /*
+ * We set the REMOVED_FLAG unconditionally. Note that there may
+ * be more than one concurrent thread setting this flag.
+ * Knowing which wins the race will be known after the garbage
+ * collection phase, stay tuned!
+ */
+ uatomic_or(&node->next, REMOVED_FLAG);
/* We performed the (logical) deletion. */
/*
_cds_lfht_gc_bucket(bucket, node);
assert(is_removed(rcu_dereference(node->next)));
- return 0;
+ /*
+ * Last phase: atomically exchange node->next with a version
+ * having "REMOVAL_OWNER_FLAG" set. If the returned node->next
+ * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own
+ * the node and win the removal race.
+ * It is interesting to note that all "add" paths are forbidden
+ * to change the next pointer starting from the point where the
+ * REMOVED_FLAG is set, so here using a read, followed by a
+ * xchg() suffice to guarantee that the xchg() will ever only
+ * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag
+ * was already set).
+ */
+ if (!is_removal_owner(uatomic_xchg(&node->next,
+ flag_removal_owner(node->next))))
+ return 0;
+ else
+ return -ENOENT;
}
static
} else {
nr_threads = 1;
}
- partition_len = len >> get_count_order_ulong(nr_threads);
+ partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads);
work = calloc(nr_threads, sizeof(*work));
assert(work);
for (thread = 0; thread < nr_threads; thread++) {
dbg_printf("init populate: order %lu index %lu hash %lu\n",
i, j, j);
new_node->reverse_hash = bit_reverse_ulong(j);
- _cds_lfht_add(ht, NULL, NULL, size, new_node, NULL, 1);
+ _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1);
}
ht->flavor->read_unlock();
}
assert(i > MIN_TABLE_ORDER);
ht->flavor->read_lock();
for (j = size + start; j < size + start + len; j++) {
- struct cds_lfht_node *fini_node = bucket_at(ht, j);
+ struct cds_lfht_node *fini_bucket = bucket_at(ht, j);
+ struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size);
assert(j >= size && j < (size << 1));
dbg_printf("remove entry: order %lu index %lu hash %lu\n",
i, j, j);
- fini_node->reverse_hash = bit_reverse_ulong(j);
- (void) _cds_lfht_del(ht, size, fini_node, 1);
+ /* Set the REMOVED_FLAG to freeze the ->next for gc */
+ uatomic_or(&fini_bucket->next, REMOVED_FLAG);
+ _cds_lfht_gc_bucket(parent_bucket, fini_bucket);
}
ht->flavor->read_unlock();
}
node->next = flag_bucket(get_end());
node->reverse_hash = 0;
- for (order = 1; order < get_count_order_ulong(size) + 1; order++) {
+ for (order = 1; order < cds_lfht_get_count_order_ulong(size) + 1; order++) {
len = 1UL << (order - 1);
cds_lfht_alloc_bucket_table(ht, order);
* Memory management plugin default.
*/
if (!mm) {
- if (!max_nr_buckets) {
- /*
- * If the maximum number of buckets is not
- * specified, we cannot use the mmap allocator,
- * so fallback on order allocator.
- */
- mm = &cds_lfht_mm_order;
- } else if (CAA_BITS_PER_LONG > 32
+ if (CAA_BITS_PER_LONG > 32
+ && max_nr_buckets
&& max_nr_buckets <= (1ULL << 32)) {
/*
* For 64-bit architectures, with max number of
alloc_split_items_count(ht);
/* this mutex should not nest in read-side C.S. */
pthread_mutex_init(&ht->resize_mutex, NULL);
- order = get_count_order_ulong(init_size);
+ order = cds_lfht_get_count_order_ulong(init_size);
ht->resize_target = 1UL << order;
cds_lfht_create_bucket(ht, 1UL << order);
ht->size = 1UL << order;
{
unsigned long size;
- node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
+ node->reverse_hash = bit_reverse_ulong(hash);
size = rcu_dereference(ht->size);
- _cds_lfht_add(ht, NULL, NULL, size, node, NULL, 0);
+ _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0);
ht_count_add(ht, size, hash);
}
unsigned long size;
struct cds_lfht_iter iter;
- node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
+ node->reverse_hash = bit_reverse_ulong(hash);
size = rcu_dereference(ht->size);
- _cds_lfht_add(ht, match, key, size, node, &iter, 0);
+ _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
if (iter.node == node)
ht_count_add(ht, size, hash);
return iter.node;
unsigned long size;
struct cds_lfht_iter iter;
- node->reverse_hash = bit_reverse_ulong((unsigned long) hash);
+ node->reverse_hash = bit_reverse_ulong(hash);
size = rcu_dereference(ht->size);
for (;;) {
- _cds_lfht_add(ht, match, key, size, node, &iter, 0);
+ _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0);
if (iter.node == node) {
ht_count_add(ht, size, hash);
return NULL;
}
}
-int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter,
+int cds_lfht_replace(struct cds_lfht *ht,
+ struct cds_lfht_iter *old_iter,
+ unsigned long hash,
+ cds_lfht_match_fct match,
+ const void *key,
struct cds_lfht_node *new_node)
{
unsigned long size;
+ new_node->reverse_hash = bit_reverse_ulong(hash);
+ if (!old_iter->node)
+ return -ENOENT;
+ if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash))
+ return -EINVAL;
+ if (caa_unlikely(!match(old_iter->node, key)))
+ return -EINVAL;
size = rcu_dereference(ht->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)
+int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node)
{
unsigned long size, hash;
int ret;
size = rcu_dereference(ht->size);
- ret = _cds_lfht_del(ht, size, iter->node, 0);
+ ret = _cds_lfht_del(ht, size, node);
if (!ret) {
- hash = bit_reverse_ulong(iter->node->reverse_hash);
+ hash = bit_reverse_ulong(node->reverse_hash);
ht_count_del(ht, size, hash);
}
return ret;
}
+int cds_lfht_is_node_deleted(struct cds_lfht_node *node)
+{
+ return is_removed(rcu_dereference(node->next));
+}
+
static
int cds_lfht_delete_bucket(struct cds_lfht *ht)
{
assert(is_bucket(node->next));
}
- for (order = get_count_order_ulong(size); (long)order >= 0; order--)
+ for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--)
cds_lfht_free_bucket_table(ht, order);
return 0;
void cds_lfht_count_nodes(struct cds_lfht *ht,
long *approx_before,
unsigned long *count,
- unsigned long *removed,
long *approx_after)
{
struct cds_lfht_node *node, *next;
- unsigned long nr_bucket = 0;
+ unsigned long nr_bucket = 0, nr_removed = 0;
*approx_before = 0;
if (ht->split_count) {
}
*count = 0;
- *removed = 0;
/* Count non-bucket nodes in the table */
node = bucket_at(ht, 0);
next = rcu_dereference(node->next);
if (is_removed(next)) {
if (!is_bucket(next))
- (*removed)++;
+ (nr_removed)++;
else
(nr_bucket)++;
} else if (!is_bucket(next))
(nr_bucket)++;
node = clear_flag(next);
} while (!is_end(node));
+ dbg_printf("number of logically removed nodes: %lu\n", nr_removed);
dbg_printf("number of bucket nodes: %lu\n", nr_bucket);
*approx_after = 0;
if (ht->split_count) {
{
unsigned long old_order, new_order;
- old_order = get_count_order_ulong(old_size);
- new_order = get_count_order_ulong(new_size);
+ old_order = cds_lfht_get_count_order_ulong(old_size);
+ new_order = cds_lfht_get_count_order_ulong(new_size);
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);
unsigned long old_order, new_order;
new_size = max(new_size, MIN_TABLE_SIZE);
- old_order = get_count_order_ulong(old_size);
- new_order = get_count_order_ulong(new_size);
+ old_order = cds_lfht_get_count_order_ulong(old_size);
+ new_order = cds_lfht_get_count_order_ulong(new_size);
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);
return;
}
work = malloc(sizeof(*work));
+ if (work == NULL) {
+ dbg_printf("error allocating resize work, bailing out\n");
+ uatomic_dec(&ht->in_progress_resize);
+ return;
+ }
work->ht = ht;
ht->flavor->update_call_rcu(&work->head, do_resize_cb);
CMM_STORE_SHARED(ht->resize_initiated, 1);
projects / urcu.git / blobdiff