#include <urcu/arch.h>
#include <assert.h>
#include <urcu-pointer.h>
+#include <urcu/uatomic.h>
#include <stdint.h>
#include "rcuja-internal.h"
* child type.
*/
#define JA_TYPE_BITS 3
-#define JA_TYPE_MAX_NR (1U << JA_TYPE_BITS)
+#define JA_TYPE_MAX_NR (1UL << JA_TYPE_BITS)
#define JA_TYPE_MASK (JA_TYPE_MAX_NR - 1)
#define JA_PTR_MASK (~JA_TYPE_MASK)
#define JA_ENTRY_PER_NODE 256UL
#define JA_BITS_PER_BYTE 3
-#define JA_MAX_INTERNAL_DEPTH 5 /* Maximum depth, excluding leafs */
+#define JA_MAX_DEPTH 5 /* Maximum depth, including leafs */
/*
* Entry for NULL node is at index 8 of the table. It is never encoded
*/
#define NODE_INDEX_NULL 8
+/*
+ * Number of removals needed on a fallback node before we try to shrink
+ * it.
+ */
+#define JA_FALLBACK_REMOVAL_COUNT 8
+
/*
* Iteration on the array to find the right node size for the number of
* children stops when it reaches .max_child == 256 (this is the largest
static
struct cds_ja_inode_flag *ja_node_flag(struct cds_ja_inode *node,
- unsigned int type)
+ unsigned long type)
{
- assert(type < RCU_JA_NR_TYPES);
+ assert(type < (1UL << JA_TYPE_BITS));
return (struct cds_ja_inode_flag *) (((unsigned long) node) | type);
}
static
struct cds_ja_inode *ja_node_ptr(struct cds_ja_inode_flag *node)
{
- return (struct cds_ja_inode *) (((unsigned long) node) | JA_PTR_MASK);
+ return (struct cds_ja_inode *) (((unsigned long) node) & JA_PTR_MASK);
}
static
-unsigned int ja_node_type(struct cds_ja_inode_flag *node)
+unsigned long ja_node_type(struct cds_ja_inode_flag *node)
{
- unsigned int type;
+ unsigned long type;
if (ja_node_ptr(node) == NULL) {
return NODE_INDEX_NULL;
}
type = (unsigned int) ((unsigned long) node & JA_TYPE_MASK);
- assert(type < RCU_JA_NR_TYPES);
+ assert(type < (1UL << JA_TYPE_BITS));
return type;
}
assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL);
nr_child_ptr = &node->u.data[0];
+ dbg_printf("linear set nth: nr_child_ptr %p\n", nr_child_ptr);
nr_child = *nr_child_ptr;
assert(nr_child <= type->max_linear_child);
- assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_child);
values = &node->u.data[1];
for (i = 0; i < nr_child; i++) {
cmm_smp_wmb(); /* write value and pointer before nr_child */
CMM_STORE_SHARED(*nr_child_ptr, nr_child + 1);
shadow_node->nr_child++;
+ dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n",
+ (unsigned int) CMM_LOAD_SHARED(*nr_child_ptr),
+ (unsigned int) shadow_node->nr_child,
+ node, shadow_node);
+
return 0;
}
{
unsigned int new_type_index;
struct cds_ja_inode *new_node;
+ struct cds_ja_shadow_node *new_shadow_node;
const struct cds_ja_type *new_type;
struct cds_ja_inode_flag *new_node_flag;
- int new_shadow = 0;
int ret;
+ int fallback = 0;
- if (!shadow_node) {
+ if (!shadow_node || old_type_index == NODE_INDEX_NULL) {
new_type_index = 0;
} else {
new_type_index = old_type_index + 1;
}
+
+retry: /* for fallback */
+ dbg_printf("Recompact to type %d\n", new_type_index);
new_type = &ja_types[new_type_index];
new_node = alloc_cds_ja_node(new_type);
if (!new_node)
return -ENOMEM;
new_node_flag = ja_node_flag(new_node, new_type_index);
- ret = rcuja_shadow_set(ja->ht, new_node, shadow_node);
- if (ret) {
+ dbg_printf("Recompact inherit lock from %p\n", shadow_node);
+ new_shadow_node = rcuja_shadow_set(ja->ht, new_node, shadow_node);
+ if (!new_shadow_node) {
free(new_node);
- return ret;
- }
-
- if (!shadow_node) {
- shadow_node = rcuja_shadow_lookup_lock(ja->ht, new_node);
- assert(shadow_node);
- new_shadow = 1;
+ return -ENOMEM;
}
-
- /*
- * We need to clear nr_child, because it will be re-incremented
- * by _ja_node_set_nth().
- */
- shadow_node->nr_child = 0;
+ if (fallback)
+ new_shadow_node->fallback_removal_count =
+ JA_FALLBACK_REMOVAL_COUNT;
assert(old_type->type_class != RCU_JA_PIGEON);
switch (old_type->type_class) {
ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter);
if (!iter)
continue;
- ret = _ja_node_set_nth(new_type, new_node, shadow_node,
+ ret = _ja_node_set_nth(new_type, new_node,
+ new_shadow_node,
v, iter);
+ if (new_type->type_class == RCU_JA_POOL && ret) {
+ goto fallback_toosmall;
+ }
assert(!ret);
}
break;
j, &v, &iter);
if (!iter)
continue;
- ret = _ja_node_set_nth(new_type, new_node, shadow_node,
+ ret = _ja_node_set_nth(new_type, new_node,
+ new_shadow_node,
v, iter);
+ if (new_type->type_class == RCU_JA_POOL
+ && ret) {
+ goto fallback_toosmall;
+ }
assert(!ret);
}
}
break;
}
+ case RCU_JA_NULL:
+ /* Nothing to copy */
+ break;
case RCU_JA_PIGEON:
default:
assert(0);
ret = -EINVAL;
- goto unlock_new_shadow;
+ goto end;
}
/* add node */
- ret = _ja_node_set_nth(new_type, new_node, shadow_node,
+ ret = _ja_node_set_nth(new_type, new_node,
+ new_shadow_node,
n, child_node_flag);
assert(!ret);
/* Return pointer to new recompacted new through old_node_flag */
*old_node_flag = new_node_flag;
- ret = rcuja_shadow_clear(ja->ht, old_node,
- RCUJA_SHADOW_CLEAR_FREE_NODE);
- assert(!ret);
+ if (old_node) {
+ ret = rcuja_shadow_clear(ja->ht, old_node, shadow_node,
+ RCUJA_SHADOW_CLEAR_FREE_NODE);
+ assert(!ret);
+ }
ret = 0;
-
-unlock_new_shadow:
- if (new_shadow)
- rcuja_shadow_unlock(shadow_node);
+end:
return ret;
+
+fallback_toosmall:
+ /* fallback if next pool is too small */
+ ret = rcuja_shadow_clear(ja->ht, new_node, new_shadow_node,
+ RCUJA_SHADOW_CLEAR_FREE_NODE);
+ assert(!ret);
+
+ /* Last type: pigeon */
+ new_type_index = (1UL << JA_TYPE_BITS) - 1;
+ dbg_printf("Fallback to type %d\n", new_type_index);
+ uatomic_inc(&ja->nr_fallback);
+ fallback = 1;
+ goto retry;
}
/*
const struct cds_ja_type *type;
struct cds_ja_inode *node;
+ dbg_printf("ja_node_set_nth for n=%u, node %p, shadow %p\n",
+ (unsigned int) n, ja_node_ptr(*node_flag), shadow_node);
+
node = ja_node_ptr(*node_flag);
type_index = ja_node_type(*node_flag);
type = &ja_types[type_index];
struct cds_ja_node *child_node)
{
struct cds_ja_shadow_node *shadow_node = NULL,
- *parent_shadow_node = NULL;
+ *parent_shadow_node = NULL,
+ *iter_shadow_node;
struct cds_ja_inode *node = ja_node_ptr(node_flag);
struct cds_ja_inode *parent_node = ja_node_ptr(parent_node_flag);
struct cds_hlist_head head;
struct cds_ja_inode_flag *iter_node_flag, *iter_dest_node_flag;
int ret, i;
- struct cds_ja_inode_flag *created_nodes[JA_MAX_INTERNAL_DEPTH];
+ struct cds_ja_inode_flag *created_nodes[JA_MAX_DEPTH];
int nr_created_nodes = 0;
+ dbg_printf("Attach node at depth %u\n", depth);
+
assert(node);
shadow_node = rcuja_shadow_lookup_lock(ja->ht, node);
if (!shadow_node) {
}
}
+ /* Create new branch, starting from bottom */
CDS_INIT_HLIST_HEAD(&head);
cds_hlist_add_head_rcu(&child_node->list, &head);
+ iter_node_flag = (struct cds_ja_inode_flag *) head.next;
+
+ /* Create shadow node for the leaf node */
+ dbg_printf("leaf shadow node creation\n");
+ iter_shadow_node = rcuja_shadow_set(ja->ht,
+ ja_node_ptr(iter_node_flag), NULL);
+ if (!iter_shadow_node) {
+ ret = -ENOMEM;
+ goto check_error;
+ }
+ created_nodes[nr_created_nodes++] = iter_node_flag;
- iter_dest_node_flag = NULL;
- ret = ja_node_set_nth(ja, &iter_dest_node_flag,
- key >> (JA_BITS_PER_BYTE * (ja->tree_depth - 2)),
- (struct cds_ja_inode_flag *) head.next,
- NULL);
- if (ret)
- goto unlock_parent;
- created_nodes[nr_created_nodes++] = iter_dest_node_flag;
- iter_node_flag = iter_dest_node_flag;
-
- /* Create new branch, starting from bottom */
- for (i = ja->tree_depth - 2; i >= (int) depth; i--) {
+ for (i = ja->tree_depth - 1; i >= (int) depth; i--) {
+ dbg_printf("branch creation level %d, key %" PRIu64 "\n",
+ i, key >> (JA_BITS_PER_BYTE * (i - 2)));
iter_dest_node_flag = NULL;
ret = ja_node_set_nth(ja, &iter_dest_node_flag,
- key >> (JA_BITS_PER_BYTE * (i - 1)),
+ key >> (JA_BITS_PER_BYTE * (i - 2)),
iter_node_flag,
NULL);
if (ret)
iter_node_flag = iter_dest_node_flag;
}
+ if (depth > 1) {
+ /* We need to use set_nth on the previous level. */
+ iter_dest_node_flag = node_flag;
+ ret = ja_node_set_nth(ja, &iter_dest_node_flag,
+ key >> (JA_BITS_PER_BYTE * (depth - 2)),
+ iter_node_flag,
+ shadow_node);
+ if (ret)
+ goto check_error;
+ created_nodes[nr_created_nodes++] = iter_dest_node_flag;
+ iter_node_flag = iter_dest_node_flag;
+ }
+
/* Publish new branch */
+ dbg_printf("Publish branch %p, replacing %p\n",
+ iter_node_flag, *node_flag_ptr);
rcu_assign_pointer(*node_flag_ptr, iter_node_flag);
/* Success */
if (ret) {
for (i = 0; i < nr_created_nodes; i++) {
int tmpret;
+ int flags;
+
+ flags = RCUJA_SHADOW_CLEAR_FREE_LOCK;
+ if (i)
+ flags |= RCUJA_SHADOW_CLEAR_FREE_NODE;
tmpret = rcuja_shadow_clear(ja->ht,
ja_node_ptr(created_nodes[i]),
- RCUJA_SHADOW_CLEAR_FREE_NODE
- | RCUJA_SHADOW_CLEAR_FREE_LOCK);
+ NULL,
+ flags);
assert(!tmpret);
}
}
-unlock_parent:
if (parent_shadow_node)
rcuja_shadow_unlock(parent_shadow_node);
unlock_shadow:
tree_depth = ja->tree_depth;
retry:
+ dbg_printf("cds_ja_add attempt: key %" PRIu64 ", node %p\n",
+ key, new_node);
iter_key = key;
parent2_node_flag = NULL;
- parent_node_flag = NULL;
+ parent_node_flag =
+ (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */
node_flag_ptr = &ja->root;
node_flag = rcu_dereference(*node_flag_ptr);
/* Iterate on all internal levels */
- for (i = 0; i < tree_depth - 1; i++) {
+ for (i = 1; i < tree_depth; i++) {
if (!ja_node_ptr(node_flag)) {
ret = ja_attach_node(ja, node_flag_ptr,
parent_node_flag, parent2_node_flag,
const struct rcu_flavor_struct *flavor)
{
struct cds_ja *ja;
+ int ret;
+ struct cds_ja_shadow_node *root_shadow_node;
ja = calloc(sizeof(*ja), 1);
if (!ja)
/* ja->root is NULL */
/* tree_depth 0 is for pointer to root node */
ja->tree_depth = (key_bits >> JA_BITS_PER_BYTE) + 1;
- assert(ja->tree_depth <= JA_MAX_INTERNAL_DEPTH + 1);
+ assert(ja->tree_depth <= JA_MAX_DEPTH);
ja->ht = rcuja_create_ht(flavor);
if (!ja->ht)
goto ht_error;
+
+ /*
+ * Note: we should not free this node until judy array destroy.
+ */
+ root_shadow_node = rcuja_shadow_set(ja->ht,
+ ja_node_ptr((struct cds_ja_inode_flag *) &ja->root),
+ NULL);
+ if (!root_shadow_node) {
+ ret = -ENOMEM;
+ goto ht_node_error;
+ }
+
return ja;
+ht_node_error:
+ ret = rcuja_delete_ht(ja->ht);
+ assert(!ret);
ht_error:
check_error:
free(ja);
ret = rcuja_delete_ht(ja->ht);
if (ret)
return ret;
+ if (uatomic_read(&ja->nr_fallback))
+ fprintf(stderr,
+ "[warning] RCU Judy Array used %lu fallback node(s)\n",
+ uatomic_read(&ja->nr_fallback));
free(ja);
return 0;
}
projects / userspace-rcu.git / blobdiff