X-Git-Url: http://git.liburcu.org/?a=blobdiff_plain;f=rcuja%2Frcuja.c;h=84c1fb73c26afdc3bc09b0bd037edf6a728d4027;hb=75d573aa7d4609a0845220527665aba86d9c32dc;hp=92e35ea91d4fcb9ada8f185971514d30495aaf56;hpb=2e3136706921bc05b9bbc260d3155f2c49b2ea1a;p=userspace-rcu.git diff --git a/rcuja/rcuja.c b/rcuja/rcuja.c index 92e35ea..84c1fb7 100644 --- a/rcuja/rcuja.c +++ b/rcuja/rcuja.c @@ -24,6 +24,7 @@ #include #include #include +#include #include #include #include @@ -35,6 +36,10 @@ #include "rcuja-internal.h" #include "bitfield.h" +#ifndef abs +#define abs_int(a) ((int) (a) > 0 ? (int) (a) : -((int) (a))) +#endif + enum cds_ja_type_class { RCU_JA_LINEAR = 0, /* Type A */ /* 32-bit: 1 to 25 children, 8 to 128 bytes */ @@ -61,33 +66,6 @@ struct cds_ja_type { uint16_t pool_size_order; /* pool size */ }; -/* - * Number of least significant pointer bits reserved to represent the - * child type. - */ -#define JA_TYPE_BITS 3 -#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_LOG2_BITS_PER_BYTE 3U -#define JA_BITS_PER_BYTE (1U << JA_LOG2_BITS_PER_BYTE) - -#define JA_MAX_DEPTH 9 /* Maximum depth, including leafs */ - -/* - * Entry for NULL node is at index 8 of the table. It is never encoded - * in flags. - */ -#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 @@ -147,8 +125,8 @@ const struct cds_ja_type ja_types[] = { { .type_class = RCU_JA_POOL, .min_child = 45, .max_child = ja_type_6_max_child, .max_linear_child = ja_type_6_max_linear_child, .order = 9, .nr_pool_order = ja_type_6_nr_pool_order, .pool_size_order = 7, }, /* - * TODO: Upon node removal below min_child, if child pool is - * filled beyond capacity, we need to roll back to pigeon. + * Upon node removal below min_child, if child pool is filled + * beyond capacity, we roll back to pigeon. */ { .type_class = RCU_JA_PIGEON, .min_child = 89, .max_child = ja_type_7_max_child, .order = 10, }, @@ -195,8 +173,8 @@ const struct cds_ja_type ja_types[] = { { .type_class = RCU_JA_POOL, .min_child = 51, .max_child = ja_type_6_max_child, .max_linear_child = ja_type_6_max_linear_child, .order = 10, .nr_pool_order = ja_type_6_nr_pool_order, .pool_size_order = 8, }, /* - * TODO: Upon node removal below min_child, if child pool is - * filled beyond capacity, we need to roll back to pigeon. + * Upon node removal below min_child, if child pool is filled + * beyond capacity, we roll back to pigeon. */ { .type_class = RCU_JA_PIGEON, .min_child = 101, .max_child = ja_type_7_max_child, .order = 11, }, @@ -261,31 +239,27 @@ struct cds_ja_inode { }; enum ja_recompact { - JA_RECOMPACT, - JA_RECOMPACT_ADD, + JA_RECOMPACT_ADD_SAME, + JA_RECOMPACT_ADD_NEXT, JA_RECOMPACT_DEL, }; static -struct cds_ja_inode_flag *ja_node_flag(struct cds_ja_inode *node, - unsigned long type) -{ - assert(type < (1UL << JA_TYPE_BITS)); - return (struct cds_ja_inode_flag *) (((unsigned long) node) | type); -} +unsigned long node_fallback_count_distribution[JA_ENTRY_PER_NODE]; +static +unsigned long nr_nodes_allocated, nr_nodes_freed; static -struct cds_ja_inode *ja_node_ptr(struct cds_ja_inode_flag *node) +struct cds_ja_inode *_ja_node_mask_ptr(struct cds_ja_inode_flag *node) { return (struct cds_ja_inode *) (((unsigned long) node) & JA_PTR_MASK); } -static unsigned long ja_node_type(struct cds_ja_inode_flag *node) { unsigned long type; - if (ja_node_ptr(node) == NULL) { + if (_ja_node_mask_ptr(node) == NULL) { return NODE_INDEX_NULL; } type = (unsigned int) ((unsigned long) node & JA_TYPE_MASK); @@ -293,14 +267,50 @@ unsigned long ja_node_type(struct cds_ja_inode_flag *node) return type; } +struct cds_ja_inode *ja_node_ptr(struct cds_ja_inode_flag *node) +{ + unsigned long type_index = ja_node_type(node); + const struct cds_ja_type *type; + + type = &ja_types[type_index]; + switch (type->type_class) { + case RCU_JA_LINEAR: + case RCU_JA_PIGEON: /* fall-through */ + case RCU_JA_NULL: /* fall-through */ + default: /* fall-through */ + return _ja_node_mask_ptr(node); + case RCU_JA_POOL: + switch (type->nr_pool_order) { + case 1: + return (struct cds_ja_inode *) (((unsigned long) node) & ~(JA_POOL_1D_MASK | JA_TYPE_MASK)); + case 2: + return (struct cds_ja_inode *) (((unsigned long) node) & ~(JA_POOL_2D_MASK | JA_POOL_1D_MASK | JA_TYPE_MASK)); + default: + assert(0); + } + } +} + struct cds_ja_inode *alloc_cds_ja_node(const struct cds_ja_type *ja_type) { - return calloc(1U << ja_type->order, sizeof(char)); + size_t len = 1U << ja_type->order; + void *p; + int ret; + + ret = posix_memalign(&p, len, len); + if (ret || !p) { + return NULL; + } + memset(p, 0, len); + uatomic_inc(&nr_nodes_allocated); + return p; } void free_cds_ja_node(struct cds_ja_inode *node) { free(node); + if (node) + uatomic_inc(&nr_nodes_freed); } #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask)) @@ -330,7 +340,7 @@ uint8_t ja_linear_node_get_nr_child(const struct cds_ja_type *type, static struct cds_ja_inode_flag *ja_linear_node_get_nth(const struct cds_ja_type *type, struct cds_ja_inode *node, - struct cds_ja_inode_flag ***child_node_flag_ptr, + struct cds_ja_inode_flag ***node_flag_ptr, uint8_t n) { uint8_t nr_child; @@ -351,12 +361,15 @@ struct cds_ja_inode_flag *ja_linear_node_get_nth(const struct cds_ja_type *type, if (CMM_LOAD_SHARED(values[i]) == n) break; } - if (i >= nr_child) + if (i >= nr_child) { + if (caa_unlikely(node_flag_ptr)) + *node_flag_ptr = NULL; return NULL; + } pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); ptr = rcu_dereference(pointers[i]); - if (caa_unlikely(child_node_flag_ptr) && ptr) - *child_node_flag_ptr = &pointers[i]; + if (caa_unlikely(node_flag_ptr)) + *node_flag_ptr = &pointers[i]; return ptr; } @@ -382,19 +395,44 @@ void ja_linear_node_get_ith_pos(const struct cds_ja_type *type, static struct cds_ja_inode_flag *ja_pool_node_get_nth(const struct cds_ja_type *type, struct cds_ja_inode *node, - struct cds_ja_inode_flag ***child_node_flag_ptr, + struct cds_ja_inode_flag *node_flag, + struct cds_ja_inode_flag ***node_flag_ptr, uint8_t n) { struct cds_ja_inode *linear; assert(type->type_class == RCU_JA_POOL); - /* - * TODO: currently, we select the pool by highest bits. We - * should support various encodings. - */ - linear = (struct cds_ja_inode *) - &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; - return ja_linear_node_get_nth(type, linear, child_node_flag_ptr, n); + + switch (type->nr_pool_order) { + case 1: + { + unsigned long bitsel, index; + + bitsel = ja_node_pool_1d_bitsel(node_flag); + assert(bitsel < CHAR_BIT); + index = ((unsigned long) n >> bitsel) & 0x1; + linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; + break; + } + case 2: + { + unsigned long bitsel[2], index[2], rindex; + + ja_node_pool_2d_bitsel(node_flag, bitsel); + assert(bitsel[0] < CHAR_BIT); + assert(bitsel[1] < CHAR_BIT); + index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; + index[0] <<= 1; + index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; + rindex = index[0] | index[1]; + linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; + break; + } + default: + linear = NULL; + assert(0); + } + return ja_linear_node_get_nth(type, linear, node_flag_ptr, n); } static @@ -410,18 +448,20 @@ struct cds_ja_inode *ja_pool_node_get_ith_pool(const struct cds_ja_type *type, static struct cds_ja_inode_flag *ja_pigeon_node_get_nth(const struct cds_ja_type *type, struct cds_ja_inode *node, - struct cds_ja_inode_flag ***child_node_flag_ptr, + struct cds_ja_inode_flag ***node_flag_ptr, uint8_t n) { - struct cds_ja_inode_flag **child_node_flag; + struct cds_ja_inode_flag **child_node_flag_ptr; + struct cds_ja_inode_flag *child_node_flag; assert(type->type_class == RCU_JA_PIGEON); - child_node_flag = &((struct cds_ja_inode_flag **) node->u.data)[n]; + child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; + child_node_flag = rcu_dereference(*child_node_flag_ptr); dbg_printf("ja_pigeon_node_get_nth child_node_flag_ptr %p\n", - child_node_flag); - if (caa_unlikely(child_node_flag_ptr) && *child_node_flag) - *child_node_flag_ptr = child_node_flag; - return rcu_dereference(*child_node_flag); + child_node_flag_ptr); + if (caa_unlikely(node_flag_ptr)) + *node_flag_ptr = child_node_flag_ptr; + return child_node_flag; } static @@ -437,8 +477,8 @@ struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *t * node_flag is already rcu_dereference'd. */ static -struct cds_ja_inode_flag * ja_node_get_nth(struct cds_ja_inode_flag *node_flag, - struct cds_ja_inode_flag ***child_node_flag_ptr, +struct cds_ja_inode_flag *ja_node_get_nth(struct cds_ja_inode_flag *node_flag, + struct cds_ja_inode_flag ***node_flag_ptr, uint8_t n) { unsigned int type_index; @@ -453,13 +493,13 @@ struct cds_ja_inode_flag * ja_node_get_nth(struct cds_ja_inode_flag *node_flag, switch (type->type_class) { case RCU_JA_LINEAR: return ja_linear_node_get_nth(type, node, - child_node_flag_ptr, n); + node_flag_ptr, n); case RCU_JA_POOL: - return ja_pool_node_get_nth(type, node, - child_node_flag_ptr, n); + return ja_pool_node_get_nth(type, node, node_flag, + node_flag_ptr, n); case RCU_JA_PIGEON: return ja_pigeon_node_get_nth(type, node, - child_node_flag_ptr, n); + node_flag_ptr, n); default: assert(0); return (void *) -1UL; @@ -481,7 +521,8 @@ int ja_linear_node_set_nth(const struct cds_ja_type *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); + dbg_printf("linear set nth: n %u, nr_child_ptr %p\n", + (unsigned int) n, nr_child_ptr); nr_child = *nr_child_ptr; assert(nr_child <= type->max_linear_child); @@ -527,6 +568,7 @@ int ja_linear_node_set_nth(const struct cds_ja_type *type, static int ja_pool_node_set_nth(const struct cds_ja_type *type, struct cds_ja_inode *node, + struct cds_ja_inode_flag *node_flag, struct cds_ja_shadow_node *shadow_node, uint8_t n, struct cds_ja_inode_flag *child_node_flag) @@ -534,8 +576,37 @@ int ja_pool_node_set_nth(const struct cds_ja_type *type, struct cds_ja_inode *linear; assert(type->type_class == RCU_JA_POOL); - linear = (struct cds_ja_inode *) - &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; + + switch (type->nr_pool_order) { + case 1: + { + unsigned long bitsel, index; + + bitsel = ja_node_pool_1d_bitsel(node_flag); + assert(bitsel < CHAR_BIT); + index = ((unsigned long) n >> bitsel) & 0x1; + linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; + break; + } + case 2: + { + unsigned long bitsel[2], index[2], rindex; + + ja_node_pool_2d_bitsel(node_flag, bitsel); + assert(bitsel[0] < CHAR_BIT); + assert(bitsel[1] < CHAR_BIT); + index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; + index[0] <<= 1; + index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; + rindex = index[0] | index[1]; + linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; + break; + } + default: + linear = NULL; + assert(0); + } + return ja_linear_node_set_nth(type, linear, shadow_node, n, child_node_flag); } @@ -565,6 +636,7 @@ int ja_pigeon_node_set_nth(const struct cds_ja_type *type, static int _ja_node_set_nth(const struct cds_ja_type *type, struct cds_ja_inode *node, + struct cds_ja_inode_flag *node_flag, struct cds_ja_shadow_node *shadow_node, uint8_t n, struct cds_ja_inode_flag *child_node_flag) @@ -574,7 +646,7 @@ int _ja_node_set_nth(const struct cds_ja_type *type, return ja_linear_node_set_nth(type, node, shadow_node, n, child_node_flag); case RCU_JA_POOL: - return ja_pool_node_set_nth(type, node, shadow_node, n, + return ja_pool_node_set_nth(type, node, node_flag, shadow_node, n, child_node_flag); case RCU_JA_PIGEON: return ja_pigeon_node_set_nth(type, node, shadow_node, n, @@ -590,45 +662,30 @@ int _ja_node_set_nth(const struct cds_ja_type *type, } static -int ja_linear_node_clear_nth(const struct cds_ja_type *type, +int ja_linear_node_clear_ptr(const struct cds_ja_type *type, struct cds_ja_inode *node, struct cds_ja_shadow_node *shadow_node, - uint8_t n) + struct cds_ja_inode_flag **node_flag_ptr) { uint8_t nr_child; - uint8_t *values, *nr_child_ptr; - struct cds_ja_inode_flag **pointers; - unsigned int i; + uint8_t *nr_child_ptr; assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); nr_child_ptr = &node->u.data[0]; - dbg_printf("linear clear nth: nr_child_ptr %p\n", nr_child_ptr); nr_child = *nr_child_ptr; assert(nr_child <= type->max_linear_child); - values = &node->u.data[1]; - pointers = (struct cds_ja_inode_flag **) align_ptr_size(&values[type->max_linear_child]); - for (i = 0; i < nr_child; i++) { - if (values[i] == n) { - if (pointers[i]) - break; - else - return -ENOENT; - } - } - if (i >= nr_child) - return -ENOENT; - if (shadow_node->fallback_removal_count) { - shadow_node->fallback_removal_count--; - } else { + if (type->type_class == RCU_JA_LINEAR) { + assert(!shadow_node->fallback_removal_count); if (shadow_node->nr_child <= type->min_child) { /* We need to try recompacting the node */ return -EFBIG; } } - assert(pointers[i] != NULL); - rcu_assign_pointer(pointers[i], NULL); + dbg_printf("linear clear ptr: nr_child_ptr %p\n", nr_child_ptr); + assert(*node_flag_ptr != NULL); + rcu_assign_pointer(*node_flag_ptr, NULL); /* * Value and nr_child are never changed (would cause ABA issue). * Instead, we leave the pointer to NULL and recompact the node @@ -637,62 +694,106 @@ int ja_linear_node_clear_nth(const struct cds_ja_type *type, * Only update the shadow node accounting. */ shadow_node->nr_child--; - dbg_printf("linear set nth: %u child, shadow: %u child, for node %p shadow %p\n", + dbg_printf("linear clear ptr: %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; } static -int ja_pool_node_clear_nth(const struct cds_ja_type *type, +int ja_pool_node_clear_ptr(const struct cds_ja_type *type, struct cds_ja_inode *node, + struct cds_ja_inode_flag *node_flag, struct cds_ja_shadow_node *shadow_node, + struct cds_ja_inode_flag **node_flag_ptr, uint8_t n) { struct cds_ja_inode *linear; assert(type->type_class == RCU_JA_POOL); - linear = (struct cds_ja_inode *) - &node->u.data[((unsigned long) n >> (CHAR_BIT - type->nr_pool_order)) << type->pool_size_order]; - return ja_linear_node_clear_nth(type, linear, shadow_node, n); + + if (shadow_node->fallback_removal_count) { + shadow_node->fallback_removal_count--; + } else { + /* We should try recompacting the node */ + if (shadow_node->nr_child <= type->min_child) + return -EFBIG; + } + + switch (type->nr_pool_order) { + case 1: + { + unsigned long bitsel, index; + + bitsel = ja_node_pool_1d_bitsel(node_flag); + assert(bitsel < CHAR_BIT); + index = ((unsigned long) n >> bitsel) & type->nr_pool_order; + linear = (struct cds_ja_inode *) &node->u.data[index << type->pool_size_order]; + break; + } + case 2: + { + unsigned long bitsel[2], index[2], rindex; + + ja_node_pool_2d_bitsel(node_flag, bitsel); + assert(bitsel[0] < CHAR_BIT); + assert(bitsel[1] < CHAR_BIT); + index[0] = ((unsigned long) n >> bitsel[0]) & 0x1; + index[0] <<= 1; + index[1] = ((unsigned long) n >> bitsel[1]) & 0x1; + rindex = index[0] | index[1]; + linear = (struct cds_ja_inode *) &node->u.data[rindex << type->pool_size_order]; + break; + } + default: + linear = NULL; + assert(0); + } + + return ja_linear_node_clear_ptr(type, linear, shadow_node, node_flag_ptr); } static -int ja_pigeon_node_clear_nth(const struct cds_ja_type *type, +int ja_pigeon_node_clear_ptr(const struct cds_ja_type *type, struct cds_ja_inode *node, struct cds_ja_shadow_node *shadow_node, - uint8_t n) + struct cds_ja_inode_flag **node_flag_ptr) { - struct cds_ja_inode_flag **ptr; - assert(type->type_class == RCU_JA_PIGEON); - ptr = &((struct cds_ja_inode_flag **) node->u.data)[n]; - if (!*ptr) - return -ENOENT; - rcu_assign_pointer(*ptr, NULL); + + if (shadow_node->fallback_removal_count) { + shadow_node->fallback_removal_count--; + } else { + /* We should try recompacting the node */ + if (shadow_node->nr_child <= type->min_child) + return -EFBIG; + } + dbg_printf("ja_pigeon_node_clear_ptr: clearing ptr: %p\n", *node_flag_ptr); + rcu_assign_pointer(*node_flag_ptr, NULL); shadow_node->nr_child--; return 0; } /* - * _ja_node_clear_nth: clear nth item within a node. Return an error + * _ja_node_clear_ptr: clear ptr item within a node. Return an error * (negative error value) if it is not found (-ENOENT). */ static -int _ja_node_clear_nth(const struct cds_ja_type *type, +int _ja_node_clear_ptr(const struct cds_ja_type *type, struct cds_ja_inode *node, + struct cds_ja_inode_flag *node_flag, struct cds_ja_shadow_node *shadow_node, + struct cds_ja_inode_flag **node_flag_ptr, uint8_t n) { switch (type->type_class) { case RCU_JA_LINEAR: - return ja_linear_node_clear_nth(type, node, shadow_node, n); + return ja_linear_node_clear_ptr(type, node, shadow_node, node_flag_ptr); case RCU_JA_POOL: - return ja_pool_node_clear_nth(type, node, shadow_node, n); + return ja_pool_node_clear_ptr(type, node, node_flag, shadow_node, node_flag_ptr, n); case RCU_JA_PIGEON: - return ja_pigeon_node_clear_nth(type, node, shadow_node, n); + return ja_pigeon_node_clear_ptr(type, node, shadow_node, node_flag_ptr); case RCU_JA_NULL: return -ENOENT; default: @@ -704,102 +805,48 @@ int _ja_node_clear_nth(const struct cds_ja_type *type, } /* - * ja_node_recompact_add: recompact a node, adding a new child. - * TODO: for pool type, take selection bit(s) into account. - * Return 0 on success, -EAGAIN if need to retry, or other negative - * error value otherwise. + * Calculate bit distribution. Returns the bit (0 to 7) that splits the + * distribution in two sub-distributions containing as much elements one + * compared to the other. */ static -int ja_node_recompact(enum ja_recompact mode, +unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode, struct cds_ja *ja, - unsigned int old_type_index, - const struct cds_ja_type *old_type, - struct cds_ja_inode *old_node, + unsigned int type_index, + const struct cds_ja_type *type, + struct cds_ja_inode *node, struct cds_ja_shadow_node *shadow_node, - struct cds_ja_inode_flag **old_node_flag, uint8_t n, - struct cds_ja_inode_flag *child_node_flag) + uint8_t n, + struct cds_ja_inode_flag *child_node_flag, + struct cds_ja_inode_flag **nullify_node_flag_ptr) { - 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 ret; - int fallback = 0; - - switch (mode) { - case JA_RECOMPACT: - new_type_index = old_type_index; - break; - case JA_RECOMPACT_ADD: - if (!shadow_node || old_type_index == NODE_INDEX_NULL) { - new_type_index = 0; - } else { - new_type_index = old_type_index + 1; - } - break; - case JA_RECOMPACT_DEL: - if (old_type_index == 0) { - new_type_index = NODE_INDEX_NULL; - } else { - new_type_index = old_type_index - 1; - } - break; - default: - assert(0); - } - -retry: /* for fallback */ - dbg_printf("Recompact from type %d to type %d\n", - old_type_index, new_type_index); - new_type = &ja_types[new_type_index]; - if (new_type_index != NODE_INDEX_NULL) { - new_node = alloc_cds_ja_node(new_type); - if (!new_node) - return -ENOMEM; - new_node_flag = ja_node_flag(new_node, new_type_index); - 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 -ENOMEM; - } - if (fallback) - new_shadow_node->fallback_removal_count = - JA_FALLBACK_REMOVAL_COUNT; - } else { - new_node = NULL; - new_node_flag = NULL; - } - - assert(mode != JA_RECOMPACT_ADD || old_type->type_class != RCU_JA_PIGEON); + uint8_t nr_one[JA_BITS_PER_BYTE]; + unsigned int bitsel = 0, bit_i, overall_best_distance = UINT_MAX; + unsigned int distrib_nr_child = 0; - if (new_type_index == NODE_INDEX_NULL) - goto skip_copy; + memset(nr_one, 0, sizeof(nr_one)); - switch (old_type->type_class) { + switch (type->type_class) { case RCU_JA_LINEAR: { uint8_t nr_child = - ja_linear_node_get_nr_child(old_type, old_node); + ja_linear_node_get_nr_child(type, node); unsigned int i; for (i = 0; i < nr_child; i++) { struct cds_ja_inode_flag *iter; uint8_t v; - ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter); + ja_linear_node_get_ith_pos(type, node, i, &v, &iter); if (!iter) continue; - if (mode == JA_RECOMPACT_DEL && v == n) + if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) continue; - 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; + for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { + if (v & (1U << bit_i)) + nr_one[bit_i]++; } - assert(!ret); + distrib_nr_child++; } break; } @@ -807,147 +854,697 @@ retry: /* for fallback */ { unsigned int pool_nr; - for (pool_nr = 0; pool_nr < (1U << old_type->nr_pool_order); pool_nr++) { + for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { struct cds_ja_inode *pool = - ja_pool_node_get_ith_pool(old_type, - old_node, pool_nr); + ja_pool_node_get_ith_pool(type, + node, pool_nr); uint8_t nr_child = - ja_linear_node_get_nr_child(old_type, pool); + ja_linear_node_get_nr_child(type, pool); unsigned int j; for (j = 0; j < nr_child; j++) { struct cds_ja_inode_flag *iter; uint8_t v; - ja_linear_node_get_ith_pos(old_type, pool, + ja_linear_node_get_ith_pos(type, pool, j, &v, &iter); if (!iter) continue; - if (mode == JA_RECOMPACT_DEL && v == n) + if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) continue; - 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; + for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { + if (v & (1U << bit_i)) + nr_one[bit_i]++; } - assert(!ret); + distrib_nr_child++; } } break; } - case RCU_JA_NULL: - assert(mode == JA_RECOMPACT_ADD); - break; case RCU_JA_PIGEON: { - uint8_t nr_child; unsigned int i; assert(mode == JA_RECOMPACT_DEL); - nr_child = shadow_node->nr_child; - for (i = 0; i < nr_child; i++) { + for (i = 0; i < JA_ENTRY_PER_NODE; i++) { struct cds_ja_inode_flag *iter; - iter = ja_pigeon_node_get_ith_pos(old_type, old_node, i); + iter = ja_pigeon_node_get_ith_pos(type, node, i); if (!iter) continue; - if (mode == JA_RECOMPACT_DEL && i == n) + if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) continue; - ret = _ja_node_set_nth(new_type, new_node, - new_shadow_node, - i, iter); - if (new_type->type_class == RCU_JA_POOL && ret) { - goto fallback_toosmall; + for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { + if (i & (1U << bit_i)) + nr_one[bit_i]++; } - assert(!ret); + distrib_nr_child++; } break; } + case RCU_JA_NULL: + assert(mode == JA_RECOMPACT_ADD_NEXT); + break; default: assert(0); - ret = -EINVAL; - goto end; - } -skip_copy: - - if (JA_RECOMPACT_ADD) { - /* add node */ - ret = _ja_node_set_nth(new_type, new_node, - new_shadow_node, - n, child_node_flag); - assert(!ret); + break; } - /* Return pointer to new recompacted node through old_node_flag */ - *old_node_flag = new_node_flag; - if (old_node) { - int flags; - flags = RCUJA_SHADOW_CLEAR_FREE_NODE; - /* - * It is OK to free the lock associated with a node - * going to NULL, since we are holding the parent lock. - * This synchronizes removal with re-add of that node. - */ - if (new_type_index == NODE_INDEX_NULL) - flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; - ret = rcuja_shadow_clear(ja->ht, old_node, shadow_node, - flags); - assert(!ret); + if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { + for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { + if (n & (1U << bit_i)) + nr_one[bit_i]++; + } + distrib_nr_child++; } - ret = 0; -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); + /* + * The best bit selector is that for which the number of ones is + * closest to half of the number of children in the + * distribution. We calculate the distance using the double of + * the sub-distribution sizes to eliminate truncation error. + */ + for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { + unsigned int distance_to_best; - /* Choose fallback 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; + distance_to_best = abs_int((nr_one[bit_i] << 1U) - distrib_nr_child); + if (distance_to_best < overall_best_distance) { + overall_best_distance = distance_to_best; + bitsel = bit_i; + } + } + dbg_printf("1 dimension pool bit selection: (%u)\n", bitsel); + return bitsel; } /* - * Return 0 on success, -EAGAIN if need to retry, or other negative - * error value otherwise. + * Calculate bit distribution in two dimensions. Returns the two bits + * (each 0 to 7) that splits the distribution in four sub-distributions + * containing as much elements one compared to the other. */ static -int ja_node_set_nth(struct cds_ja *ja, - struct cds_ja_inode_flag **node_flag, uint8_t n, +void ja_node_sum_distribution_2d(enum ja_recompact mode, + struct cds_ja *ja, + unsigned int type_index, + const struct cds_ja_type *type, + struct cds_ja_inode *node, + struct cds_ja_shadow_node *shadow_node, + uint8_t n, struct cds_ja_inode_flag *child_node_flag, - struct cds_ja_shadow_node *shadow_node) + struct cds_ja_inode_flag **nullify_node_flag_ptr, + unsigned int *_bitsel) { - int ret; - unsigned int type_index; - 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); + uint8_t nr_2d_11[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], + nr_2d_10[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], + nr_2d_01[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE], + nr_2d_00[JA_BITS_PER_BYTE][JA_BITS_PER_BYTE]; + unsigned int bitsel[2] = { 0, 1 }; + unsigned int bit_i, bit_j; + int overall_best_distance = INT_MAX; + unsigned int distrib_nr_child = 0; + + memset(nr_2d_11, 0, sizeof(nr_2d_11)); + memset(nr_2d_10, 0, sizeof(nr_2d_10)); + memset(nr_2d_01, 0, sizeof(nr_2d_01)); + memset(nr_2d_00, 0, sizeof(nr_2d_00)); + + switch (type->type_class) { + case RCU_JA_LINEAR: + { + uint8_t nr_child = + ja_linear_node_get_nr_child(type, node); + unsigned int i; + + for (i = 0; i < nr_child; i++) { + struct cds_ja_inode_flag *iter; + uint8_t v; + + ja_linear_node_get_ith_pos(type, node, i, &v, &iter); + if (!iter) + continue; + if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) + continue; + for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { + for (bit_j = 0; bit_j < bit_i; bit_j++) { + if ((v & (1U << bit_i)) && (v & (1U << bit_j))) { + nr_2d_11[bit_i][bit_j]++; + } + if ((v & (1U << bit_i)) && !(v & (1U << bit_j))) { + nr_2d_10[bit_i][bit_j]++; + } + if (!(v & (1U << bit_i)) && (v & (1U << bit_j))) { + nr_2d_01[bit_i][bit_j]++; + } + if (!(v & (1U << bit_i)) && !(v & (1U << bit_j))) { + nr_2d_00[bit_i][bit_j]++; + } + } + } + distrib_nr_child++; + } + break; + } + case RCU_JA_POOL: + { + unsigned int pool_nr; + + for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { + struct cds_ja_inode *pool = + ja_pool_node_get_ith_pool(type, + node, pool_nr); + uint8_t nr_child = + ja_linear_node_get_nr_child(type, pool); + unsigned int j; + + for (j = 0; j < nr_child; j++) { + struct cds_ja_inode_flag *iter; + uint8_t v; + + ja_linear_node_get_ith_pos(type, pool, + j, &v, &iter); + if (!iter) + continue; + if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) + continue; + for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { + for (bit_j = 0; bit_j < bit_i; bit_j++) { + if ((v & (1U << bit_i)) && (v & (1U << bit_j))) { + nr_2d_11[bit_i][bit_j]++; + } + if ((v & (1U << bit_i)) && !(v & (1U << bit_j))) { + nr_2d_10[bit_i][bit_j]++; + } + if (!(v & (1U << bit_i)) && (v & (1U << bit_j))) { + nr_2d_01[bit_i][bit_j]++; + } + if (!(v & (1U << bit_i)) && !(v & (1U << bit_j))) { + nr_2d_00[bit_i][bit_j]++; + } + } + } + distrib_nr_child++; + } + } + break; + } + case RCU_JA_PIGEON: + { + unsigned int i; + + assert(mode == JA_RECOMPACT_DEL); + for (i = 0; i < JA_ENTRY_PER_NODE; i++) { + struct cds_ja_inode_flag *iter; + + iter = ja_pigeon_node_get_ith_pos(type, node, i); + if (!iter) + continue; + if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) + continue; + for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { + for (bit_j = 0; bit_j < bit_i; bit_j++) { + if ((i & (1U << bit_i)) && (i & (1U << bit_j))) { + nr_2d_11[bit_i][bit_j]++; + } + if ((i & (1U << bit_i)) && !(i & (1U << bit_j))) { + nr_2d_10[bit_i][bit_j]++; + } + if (!(i & (1U << bit_i)) && (i & (1U << bit_j))) { + nr_2d_01[bit_i][bit_j]++; + } + if (!(i & (1U << bit_i)) && !(i & (1U << bit_j))) { + nr_2d_00[bit_i][bit_j]++; + } + } + } + distrib_nr_child++; + } + break; + } + case RCU_JA_NULL: + assert(mode == JA_RECOMPACT_ADD_NEXT); + break; + default: + assert(0); + break; + } + + if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { + for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { + for (bit_j = 0; bit_j < bit_i; bit_j++) { + if ((n & (1U << bit_i)) && (n & (1U << bit_j))) { + nr_2d_11[bit_i][bit_j]++; + } + if ((n & (1U << bit_i)) && !(n & (1U << bit_j))) { + nr_2d_10[bit_i][bit_j]++; + } + if (!(n & (1U << bit_i)) && (n & (1U << bit_j))) { + nr_2d_01[bit_i][bit_j]++; + } + if (!(n & (1U << bit_i)) && !(n & (1U << bit_j))) { + nr_2d_00[bit_i][bit_j]++; + } + } + } + distrib_nr_child++; + } + + /* + * The best bit selector is that for which the number of nodes + * in each sub-class is closest to one-fourth of the number of + * children in the distribution. We calculate the distance using + * 4 times the size of the sub-distribution to eliminate + * truncation error. + */ + for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { + for (bit_j = 0; bit_j < bit_i; bit_j++) { + int distance_to_best[4]; + + distance_to_best[0] = (nr_2d_11[bit_i][bit_j] << 2U) - distrib_nr_child; + distance_to_best[1] = (nr_2d_10[bit_i][bit_j] << 2U) - distrib_nr_child; + distance_to_best[2] = (nr_2d_01[bit_i][bit_j] << 2U) - distrib_nr_child; + distance_to_best[3] = (nr_2d_00[bit_i][bit_j] << 2U) - distrib_nr_child; + + /* Consider worse distance above best */ + if (distance_to_best[1] > 0 && distance_to_best[1] > distance_to_best[0]) + distance_to_best[0] = distance_to_best[1]; + if (distance_to_best[2] > 0 && distance_to_best[2] > distance_to_best[0]) + distance_to_best[0] = distance_to_best[2]; + if (distance_to_best[3] > 0 && distance_to_best[3] > distance_to_best[0]) + distance_to_best[0] = distance_to_best[3]; + + /* + * If our worse distance is better than overall, + * we become new best candidate. + */ + if (distance_to_best[0] < overall_best_distance) { + overall_best_distance = distance_to_best[0]; + bitsel[0] = bit_i; + bitsel[1] = bit_j; + } + } + } + + dbg_printf("2 dimensions pool bit selection: (%u,%u)\n", bitsel[0], bitsel[1]); + + /* Return our bit selection */ + _bitsel[0] = bitsel[0]; + _bitsel[1] = bitsel[1]; +} + +static +unsigned int find_nearest_type_index(unsigned int type_index, + unsigned int nr_nodes) +{ + const struct cds_ja_type *type; + + assert(type_index != NODE_INDEX_NULL); + if (nr_nodes == 0) + return NODE_INDEX_NULL; + for (;;) { + type = &ja_types[type_index]; + if (nr_nodes < type->min_child) + type_index--; + else if (nr_nodes > type->max_child) + type_index++; + else + break; + } + return type_index; +} + +/* + * ja_node_recompact_add: recompact a node, adding a new child. + * Return 0 on success, -EAGAIN if need to retry, or other negative + * error value otherwise. + */ +static +int ja_node_recompact(enum ja_recompact mode, + struct cds_ja *ja, + unsigned int old_type_index, + const struct cds_ja_type *old_type, + struct cds_ja_inode *old_node, + struct cds_ja_shadow_node *shadow_node, + struct cds_ja_inode_flag **old_node_flag_ptr, uint8_t n, + struct cds_ja_inode_flag *child_node_flag, + struct cds_ja_inode_flag **nullify_node_flag_ptr, + int level) +{ + unsigned int new_type_index; + struct cds_ja_inode *new_node; + struct cds_ja_shadow_node *new_shadow_node = NULL; + const struct cds_ja_type *new_type; + struct cds_ja_inode_flag *new_node_flag, *old_node_flag; + int ret; + int fallback = 0; + + old_node_flag = *old_node_flag_ptr; + + /* + * Need to find nearest type index even for ADD_SAME, because + * this recompaction, when applied to linear nodes, will garbage + * collect dummy (NULL) entries, and can therefore cause a few + * linear representations to be skipped. + */ + switch (mode) { + case JA_RECOMPACT_ADD_SAME: + new_type_index = find_nearest_type_index(old_type_index, + shadow_node->nr_child + 1); + dbg_printf("Recompact for node with %u children\n", + shadow_node->nr_child + 1); + break; + case JA_RECOMPACT_ADD_NEXT: + if (!shadow_node || old_type_index == NODE_INDEX_NULL) { + new_type_index = 0; + dbg_printf("Recompact for NULL\n"); + } else { + new_type_index = find_nearest_type_index(old_type_index, + shadow_node->nr_child + 1); + dbg_printf("Recompact for node with %u children\n", + shadow_node->nr_child + 1); + } + break; + case JA_RECOMPACT_DEL: + new_type_index = find_nearest_type_index(old_type_index, + shadow_node->nr_child - 1); + dbg_printf("Recompact for node with %u children\n", + shadow_node->nr_child - 1); + break; + default: + assert(0); + } + +retry: /* for fallback */ + dbg_printf("Recompact from type %d to type %d\n", + old_type_index, new_type_index); + new_type = &ja_types[new_type_index]; + if (new_type_index != NODE_INDEX_NULL) { + new_node = alloc_cds_ja_node(new_type); + if (!new_node) + return -ENOMEM; + + if (new_type->type_class == RCU_JA_POOL) { + switch (new_type->nr_pool_order) { + case 1: + { + unsigned int node_distrib_bitsel; + + node_distrib_bitsel = + ja_node_sum_distribution_1d(mode, ja, + old_type_index, old_type, + old_node, shadow_node, + n, child_node_flag, + nullify_node_flag_ptr); + assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); + new_node_flag = ja_node_flag_pool_1d(new_node, + new_type_index, node_distrib_bitsel); + break; + } + case 2: + { + unsigned int node_distrib_bitsel[2]; + + ja_node_sum_distribution_2d(mode, ja, + old_type_index, old_type, + old_node, shadow_node, + n, child_node_flag, + nullify_node_flag_ptr, + node_distrib_bitsel); + assert(!((unsigned long) new_node & JA_POOL_1D_MASK)); + assert(!((unsigned long) new_node & JA_POOL_2D_MASK)); + new_node_flag = ja_node_flag_pool_2d(new_node, + new_type_index, node_distrib_bitsel); + break; + } + default: + assert(0); + } + } else { + new_node_flag = ja_node_flag(new_node, new_type_index); + } + + dbg_printf("Recompact inherit lock from %p\n", shadow_node); + new_shadow_node = rcuja_shadow_set(ja->ht, new_node_flag, shadow_node, ja, level); + if (!new_shadow_node) { + free_cds_ja_node(new_node); + return -ENOMEM; + } + if (fallback) + new_shadow_node->fallback_removal_count = + JA_FALLBACK_REMOVAL_COUNT; + } else { + new_node = NULL; + new_node_flag = NULL; + } + + assert(mode != JA_RECOMPACT_ADD_NEXT || old_type->type_class != RCU_JA_PIGEON); + + if (new_type_index == NODE_INDEX_NULL) + goto skip_copy; + + switch (old_type->type_class) { + case RCU_JA_LINEAR: + { + uint8_t nr_child = + ja_linear_node_get_nr_child(old_type, old_node); + unsigned int i; + + for (i = 0; i < nr_child; i++) { + struct cds_ja_inode_flag *iter; + uint8_t v; + + ja_linear_node_get_ith_pos(old_type, old_node, i, &v, &iter); + if (!iter) + continue; + if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) + continue; + ret = _ja_node_set_nth(new_type, new_node, new_node_flag, + new_shadow_node, + v, iter); + if (new_type->type_class == RCU_JA_POOL && ret) { + goto fallback_toosmall; + } + assert(!ret); + } + break; + } + case RCU_JA_POOL: + { + unsigned int pool_nr; + + for (pool_nr = 0; pool_nr < (1U << old_type->nr_pool_order); pool_nr++) { + struct cds_ja_inode *pool = + ja_pool_node_get_ith_pool(old_type, + old_node, pool_nr); + uint8_t nr_child = + ja_linear_node_get_nr_child(old_type, pool); + unsigned int j; + + for (j = 0; j < nr_child; j++) { + struct cds_ja_inode_flag *iter; + uint8_t v; + + ja_linear_node_get_ith_pos(old_type, pool, + j, &v, &iter); + if (!iter) + continue; + if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) + continue; + ret = _ja_node_set_nth(new_type, new_node, new_node_flag, + new_shadow_node, + v, iter); + if (new_type->type_class == RCU_JA_POOL + && ret) { + goto fallback_toosmall; + } + assert(!ret); + } + } + break; + } + case RCU_JA_NULL: + assert(mode == JA_RECOMPACT_ADD_NEXT); + break; + case RCU_JA_PIGEON: + { + unsigned int i; + + assert(mode == JA_RECOMPACT_DEL); + for (i = 0; i < JA_ENTRY_PER_NODE; i++) { + struct cds_ja_inode_flag *iter; + + iter = ja_pigeon_node_get_ith_pos(old_type, old_node, i); + if (!iter) + continue; + if (mode == JA_RECOMPACT_DEL && *nullify_node_flag_ptr == iter) + continue; + ret = _ja_node_set_nth(new_type, new_node, new_node_flag, + new_shadow_node, + i, iter); + if (new_type->type_class == RCU_JA_POOL && ret) { + goto fallback_toosmall; + } + assert(!ret); + } + break; + } + default: + assert(0); + ret = -EINVAL; + goto end; + } +skip_copy: + + if (mode == JA_RECOMPACT_ADD_NEXT || mode == JA_RECOMPACT_ADD_SAME) { + /* add node */ + ret = _ja_node_set_nth(new_type, new_node, new_node_flag, + new_shadow_node, + n, child_node_flag); + if (new_type->type_class == RCU_JA_POOL && ret) { + goto fallback_toosmall; + } + assert(!ret); + } + + if (fallback) { + dbg_printf("Using fallback for %u children, node type index: %u, mode %s\n", + new_shadow_node->nr_child, old_type_index, mode == JA_RECOMPACT_ADD_NEXT ? "add_next" : + (mode == JA_RECOMPACT_DEL ? "del" : "add_same")); + uatomic_inc(&node_fallback_count_distribution[new_shadow_node->nr_child]); + } + + /* Return pointer to new recompacted node through old_node_flag_ptr */ + *old_node_flag_ptr = new_node_flag; + if (old_node) { + int flags; + + flags = RCUJA_SHADOW_CLEAR_FREE_NODE; + /* + * It is OK to free the lock associated with a node + * going to NULL, since we are holding the parent lock. + * This synchronizes removal with re-add of that node. + */ + if (new_type_index == NODE_INDEX_NULL) + flags |= RCUJA_SHADOW_CLEAR_FREE_LOCK; + ret = rcuja_shadow_clear(ja->ht, old_node_flag, shadow_node, + flags); + assert(!ret); + } + + ret = 0; +end: + return ret; + +fallback_toosmall: + /* fallback if next pool is too small */ + assert(new_shadow_node); + ret = rcuja_shadow_clear(ja->ht, new_node_flag, new_shadow_node, + RCUJA_SHADOW_CLEAR_FREE_NODE); + assert(!ret); + + switch (mode) { + case JA_RECOMPACT_ADD_SAME: + /* + * JA_RECOMPACT_ADD_SAME is only triggered if a linear + * node within a pool has unused entries. It should + * therefore _never_ be too small. + */ + assert(0); + + /* Fall-through */ + case JA_RECOMPACT_ADD_NEXT: + { + const struct cds_ja_type *next_type; + + /* + * Recompaction attempt on add failed. Should only + * happen if target node type is pool. Caused by + * hard-to-split distribution. Recompact using the next + * distribution size. + */ + assert(new_type->type_class == RCU_JA_POOL); + next_type = &ja_types[new_type_index + 1]; + /* + * Try going to the next pool size if our population + * fits within its range. This is not flagged as a + * fallback. + */ + if (shadow_node->nr_child + 1 >= next_type->min_child + && shadow_node->nr_child + 1 <= next_type->max_child) { + new_type_index++; + goto retry; + } else { + new_type_index++; + dbg_printf("Add fallback to type %d\n", new_type_index); + uatomic_inc(&ja->nr_fallback); + fallback = 1; + goto retry; + } + break; + } + case JA_RECOMPACT_DEL: + /* + * Recompaction attempt on delete failed. Should only + * happen if target node type is pool. This is caused by + * a hard-to-split distribution. Recompact on same node + * size, but flag current node as "fallback" to ensure + * we don't attempt recompaction before some activity + * has reshuffled our node. + */ + assert(new_type->type_class == RCU_JA_POOL); + new_type_index = old_type_index; + dbg_printf("Delete fallback keeping type %d\n", new_type_index); + uatomic_inc(&ja->nr_fallback); + fallback = 1; + goto retry; + default: + assert(0); + return -EINVAL; + } + + /* + * Last resort fallback: 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; +} + +/* + * Return 0 on success, -EAGAIN if need to retry, or other negative + * error value otherwise. + */ +static +int ja_node_set_nth(struct cds_ja *ja, + struct cds_ja_inode_flag **node_flag, uint8_t n, + struct cds_ja_inode_flag *child_node_flag, + struct cds_ja_shadow_node *shadow_node, + int level) +{ + int ret; + unsigned int type_index; + 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]; - ret = _ja_node_set_nth(type, node, shadow_node, + ret = _ja_node_set_nth(type, node, *node_flag, shadow_node, n, child_node_flag); switch (ret) { case -ENOSPC: - /* Not enough space in node, need to recompact. */ - ret = ja_node_recompact(JA_RECOMPACT_ADD, ja, type_index, type, node, - shadow_node, node_flag, n, child_node_flag); + /* Not enough space in node, need to recompact to next type. */ + ret = ja_node_recompact(JA_RECOMPACT_ADD_NEXT, ja, type_index, type, node, + shadow_node, node_flag, n, child_node_flag, NULL, level); break; case -ERANGE: /* Node needs to be recompacted. */ - ret = ja_node_recompact(JA_RECOMPACT, ja, type_index, type, node, - shadow_node, node_flag, n, child_node_flag); + ret = ja_node_recompact(JA_RECOMPACT_ADD_SAME, ja, type_index, type, node, + shadow_node, node_flag, n, child_node_flag, NULL, level); break; } return ret; @@ -958,58 +1555,62 @@ int ja_node_set_nth(struct cds_ja *ja, * error value otherwise. */ static -int ja_node_clear_nth(struct cds_ja *ja, - struct cds_ja_inode_flag **node_flag, uint8_t n, - struct cds_ja_shadow_node *shadow_node) +int ja_node_clear_ptr(struct cds_ja *ja, + struct cds_ja_inode_flag **node_flag_ptr, /* Pointer to location to nullify */ + struct cds_ja_inode_flag **parent_node_flag_ptr, /* Address of parent ptr in its parent */ + struct cds_ja_shadow_node *shadow_node, /* of parent */ + uint8_t n, int level) { int ret; unsigned int type_index; const struct cds_ja_type *type; struct cds_ja_inode *node; - dbg_printf("ja_node_clear_nth for n=%u, node %p, shadow %p\n", - (unsigned int) n, ja_node_ptr(*node_flag), shadow_node); + dbg_printf("ja_node_clear_ptr for node %p, shadow %p, target ptr %p\n", + ja_node_ptr(*parent_node_flag_ptr), shadow_node, node_flag_ptr); - node = ja_node_ptr(*node_flag); - type_index = ja_node_type(*node_flag); + node = ja_node_ptr(*parent_node_flag_ptr); + type_index = ja_node_type(*parent_node_flag_ptr); type = &ja_types[type_index]; - ret = _ja_node_clear_nth(type, node, shadow_node, n); + ret = _ja_node_clear_ptr(type, node, *parent_node_flag_ptr, shadow_node, node_flag_ptr, n); if (ret == -EFBIG) { - /* Should to try recompaction. */ + /* Should try recompaction. */ ret = ja_node_recompact(JA_RECOMPACT_DEL, ja, type_index, type, node, - shadow_node, node_flag, n, NULL); + shadow_node, parent_node_flag_ptr, n, NULL, + node_flag_ptr, level); } return ret; } -struct cds_hlist_head *cds_ja_lookup(struct cds_ja *ja, uint64_t key) +struct cds_hlist_head cds_ja_lookup(struct cds_ja *ja, uint64_t key) { unsigned int tree_depth, i; struct cds_ja_inode_flag *node_flag; + struct cds_hlist_head head = { NULL }; if (caa_unlikely(key > ja->key_max)) - return NULL; + return head; tree_depth = ja->tree_depth; node_flag = rcu_dereference(ja->root); /* level 0: root node */ if (!ja_node_ptr(node_flag)) - return NULL; + return head; for (i = 1; i < tree_depth; i++) { uint8_t iter_key; iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); - node_flag = ja_node_get_nth(node_flag, NULL, - iter_key); + node_flag = ja_node_get_nth(node_flag, NULL, iter_key); dbg_printf("cds_ja_lookup iter key lookup %u finds node_flag %p\n", (unsigned int) iter_key, node_flag); if (!ja_node_ptr(node_flag)) - return NULL; + return head; } /* Last level lookup succeded. We got an actual match. */ - return (struct cds_hlist_head *) node_flag; + head.next = (struct cds_hlist_node *) node_flag; + return head; } /* @@ -1025,96 +1626,114 @@ struct cds_hlist_head *cds_ja_lookup(struct cds_ja *ja, uint64_t key) * TODO: we currently always take the parent lock even when not needed. */ static -int ja_attach_node(struct cds_ja *ja, - struct cds_ja_inode_flag **node_flag_ptr, - struct cds_ja_inode_flag *node_flag, - struct cds_ja_inode_flag *parent_node_flag, +int ja_attach_node(struct cds_ja *ja, + struct cds_ja_inode_flag **attach_node_flag_ptr, + struct cds_ja_inode_flag *attach_node_flag, + struct cds_ja_inode_flag *parent_attach_node_flag, + struct cds_ja_inode_flag **old_node_flag_ptr, + struct cds_ja_inode_flag *old_node_flag, uint64_t key, unsigned int level, struct cds_ja_node *child_node) { struct cds_ja_shadow_node *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); + *parent_shadow_node = NULL; 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_DEPTH]; int nr_created_nodes = 0; - dbg_printf("Attach node at level %u (node %p, node_flag %p)\n", - level, node, node_flag); + dbg_printf("Attach node at level %u (old_node_flag %p, attach_node_flag_ptr %p attach_node_flag %p, parent_attach_node_flag %p)\n", + level, old_node_flag, attach_node_flag_ptr, attach_node_flag, parent_attach_node_flag); - assert(node); - shadow_node = rcuja_shadow_lookup_lock(ja->ht, node); - if (!shadow_node) { - ret = -EAGAIN; - goto end; + assert(!old_node_flag); + if (attach_node_flag) { + shadow_node = rcuja_shadow_lookup_lock(ja->ht, attach_node_flag); + if (!shadow_node) { + ret = -EAGAIN; + goto end; + } } - if (parent_node) { + if (parent_attach_node_flag) { parent_shadow_node = rcuja_shadow_lookup_lock(ja->ht, - parent_node); + parent_attach_node_flag); if (!parent_shadow_node) { ret = -EAGAIN; goto unlock_shadow; } } + if (old_node_flag_ptr && ja_node_ptr(*old_node_flag_ptr)) { + /* + * Target node has been updated between RCU lookup and + * lock acquisition. We need to re-try lookup and + * attach. + */ + ret = -EAGAIN; + goto unlock_parent; + } + + if (attach_node_flag_ptr && ja_node_ptr(*attach_node_flag_ptr) != + ja_node_ptr(attach_node_flag)) { + /* + * Target node has been updated between RCU lookup and + * lock acquisition. We need to re-try lookup and + * attach. + */ + ret = -EAGAIN; + goto unlock_parent; + } + /* 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; - - for (i = ja->tree_depth; i > (int) level; i--) { + for (i = ja->tree_depth - 1; i >= (int) level; i--) { uint8_t iter_key; - iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i))); + iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - i - 1))); dbg_printf("branch creation level %d, key %u\n", - i - 1, (unsigned int) iter_key); + i, (unsigned int) iter_key); iter_dest_node_flag = NULL; ret = ja_node_set_nth(ja, &iter_dest_node_flag, iter_key, iter_node_flag, - NULL); + NULL, i); if (ret) goto check_error; created_nodes[nr_created_nodes++] = iter_dest_node_flag; iter_node_flag = iter_dest_node_flag; } + assert(level > 0); - if (level > 1) { + /* Publish branch */ + if (level == 1) { + /* + * Attaching to root node. + */ + rcu_assign_pointer(ja->root, iter_node_flag); + } else { uint8_t iter_key; iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); + dbg_printf("publish branch at level %d, key %u\n", + level - 1, (unsigned int) iter_key); /* We need to use set_nth on the previous level. */ - iter_dest_node_flag = node_flag; + iter_dest_node_flag = attach_node_flag; ret = ja_node_set_nth(ja, &iter_dest_node_flag, iter_key, iter_node_flag, - shadow_node); + shadow_node, level - 1); if (ret) goto check_error; - created_nodes[nr_created_nodes++] = iter_dest_node_flag; - iter_node_flag = iter_dest_node_flag; + /* + * Attach branch + */ + rcu_assign_pointer(*attach_node_flag_ptr, 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 */ ret = 0; @@ -1128,12 +1747,13 @@ check_error: if (i) flags |= RCUJA_SHADOW_CLEAR_FREE_NODE; tmpret = rcuja_shadow_clear(ja->ht, - ja_node_ptr(created_nodes[i]), + created_nodes[i], NULL, flags); assert(!tmpret); } } +unlock_parent: if (parent_shadow_node) rcuja_shadow_unlock(parent_shadow_node); unlock_shadow: @@ -1144,39 +1764,54 @@ end: } /* - * Lock the hlist head shadow node mutex, and add node to list of - * duplicates. Failure can happen if concurrent removal removes the last - * node with same key before we get the lock. + * Lock the parent containing the hlist head pointer, and add node to list of + * duplicates. Failure can happen if concurrent update changes the + * parent before we get the lock. We return -EAGAIN in that case. * Return 0 on success, negative error value on failure. */ static int ja_chain_node(struct cds_ja *ja, - struct cds_hlist_head *head, + struct cds_ja_inode_flag *parent_node_flag, + struct cds_ja_inode_flag **node_flag_ptr, + struct cds_ja_inode_flag *node_flag, struct cds_ja_node *node) { struct cds_ja_shadow_node *shadow_node; + int ret = 0; - shadow_node = rcuja_shadow_lookup_lock(ja->ht, - (struct cds_ja_inode *) head); - if (!shadow_node) + shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); + if (!shadow_node) { return -EAGAIN; - cds_hlist_add_head_rcu(&node->list, head); + } + if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { + ret = -EAGAIN; + goto end; + } + cds_hlist_add_head_rcu(&node->list, (struct cds_hlist_head *) node_flag_ptr); +end: rcuja_shadow_unlock(shadow_node); - return 0; + return ret; } -int cds_ja_add(struct cds_ja *ja, uint64_t key, - struct cds_ja_node *new_node) +static +int _cds_ja_add(struct cds_ja *ja, uint64_t key, + struct cds_ja_node *new_node, + struct cds_ja_node **unique_node_ret) { unsigned int tree_depth, i; - struct cds_ja_inode_flag **node_flag_ptr; /* in parent */ - struct cds_ja_inode_flag *node_flag, + struct cds_ja_inode_flag *attach_node_flag, *parent_node_flag, - *parent2_node_flag; + *parent2_node_flag, + *node_flag, + *parent_attach_node_flag; + struct cds_ja_inode_flag **attach_node_flag_ptr, + **parent_node_flag_ptr, + **node_flag_ptr; int ret; - if (caa_unlikely(key > ja->key_max)) + if (caa_unlikely(key > ja->key_max)) { return -EINVAL; + } tree_depth = ja->tree_depth; retry: @@ -1185,51 +1820,403 @@ retry: parent2_node_flag = NULL; parent_node_flag = (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */ + parent_node_flag_ptr = NULL; + node_flag = rcu_dereference(ja->root); node_flag_ptr = &ja->root; - node_flag = rcu_dereference(*node_flag_ptr); /* Iterate on all internal levels */ for (i = 1; i < tree_depth; i++) { uint8_t iter_key; - dbg_printf("cds_ja_add iter node_flag_ptr %p node_flag %p\n", - *node_flag_ptr, node_flag); - if (!ja_node_ptr(node_flag)) { - ret = ja_attach_node(ja, node_flag_ptr, - parent_node_flag, parent2_node_flag, - key, i, new_node); - if (ret == -EAGAIN || ret == -EEXIST) - goto retry; - else - goto end; - } + if (!ja_node_ptr(node_flag)) + break; + dbg_printf("cds_ja_add iter parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", + parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); parent2_node_flag = parent_node_flag; parent_node_flag = node_flag; + parent_node_flag_ptr = node_flag_ptr; node_flag = ja_node_get_nth(node_flag, &node_flag_ptr, iter_key); - dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p node_flag_ptr %p\n", - (unsigned int) iter_key, node_flag, *node_flag_ptr); } /* - * We reached bottom of tree, simply add node to last internal - * level, or chain it if key is already present. + * We reached either bottom of tree or internal NULL node, + * simply add node to last internal level, or chain it if key is + * already present. */ if (!ja_node_ptr(node_flag)) { - dbg_printf("cds_ja_add last node_flag_ptr %p node_flag %p\n", - *node_flag_ptr, node_flag); - ret = ja_attach_node(ja, node_flag_ptr, parent_node_flag, - parent2_node_flag, key, i, new_node); + dbg_printf("cds_ja_add NULL parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", + parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); + + attach_node_flag = parent_node_flag; + attach_node_flag_ptr = parent_node_flag_ptr; + parent_attach_node_flag = parent2_node_flag; + + ret = ja_attach_node(ja, attach_node_flag_ptr, + attach_node_flag, + parent_attach_node_flag, + node_flag_ptr, + node_flag, + key, i, new_node); } else { + if (unique_node_ret) { + *unique_node_ret = (struct cds_ja_node *) ja_node_ptr(node_flag); + return -EEXIST; + } + + dbg_printf("cds_ja_add duplicate parent2_node_flag %p parent_node_flag %p node_flag_ptr %p node_flag %p\n", + parent2_node_flag, parent_node_flag, node_flag_ptr, node_flag); + + attach_node_flag = node_flag; + attach_node_flag_ptr = node_flag_ptr; + parent_attach_node_flag = parent_node_flag; + ret = ja_chain_node(ja, - (struct cds_hlist_head *) ja_node_ptr(node_flag), + parent_attach_node_flag, + attach_node_flag_ptr, + attach_node_flag, new_node); } - if (ret == -EAGAIN) + if (ret == -EAGAIN || ret == -EEXIST) goto retry; + + return ret; +} + +int cds_ja_add(struct cds_ja *ja, uint64_t key, + struct cds_ja_node *new_node) +{ + return _cds_ja_add(ja, key, new_node, NULL); +} + +struct cds_ja_node *cds_ja_add_unique(struct cds_ja *ja, uint64_t key, + struct cds_ja_node *new_node) +{ + int ret; + struct cds_ja_node *ret_node; + + ret = _cds_ja_add(ja, key, new_node, &ret_node); + if (ret == -EEXIST) + return ret_node; + else + return new_node; +} + +/* + * Note: there is no need to lookup the pointer address associated with + * each node's nth item after taking the lock: it's already been done by + * cds_ja_del while holding the rcu read-side lock, and our node rules + * ensure that when a match value -> pointer is found in a node, it is + * _NEVER_ changed for that node without recompaction, and recompaction + * reallocates the node. + * However, when a child is removed from "linear" nodes, its pointer + * is set to NULL. We therefore check, while holding the locks, if this + * pointer is NULL, and return -ENOENT to the caller if it is the case. + */ +static +int ja_detach_node(struct cds_ja *ja, + struct cds_ja_inode_flag **snapshot, + struct cds_ja_inode_flag ***snapshot_ptr, + uint8_t *snapshot_n, + int nr_snapshot, + uint64_t key, + struct cds_ja_node *node) +{ + struct cds_ja_shadow_node *shadow_nodes[JA_MAX_DEPTH]; + struct cds_ja_inode_flag **node_flag_ptr = NULL, + *parent_node_flag = NULL, + **parent_node_flag_ptr = NULL; + struct cds_ja_inode_flag *iter_node_flag; + int ret, i, nr_shadow = 0, nr_clear = 0, nr_branch = 0; + uint8_t n = 0; + + assert(nr_snapshot == ja->tree_depth + 1); + + /* + * From the last internal level node going up, get the node + * lock, check if the node has only one child left. If it is the + * case, we continue iterating upward. When we reach a node + * which has more that one child left, we lock the parent, and + * proceed to the node deletion (removing its children too). + */ + for (i = nr_snapshot - 2; i >= 1; i--) { + struct cds_ja_shadow_node *shadow_node; + + shadow_node = rcuja_shadow_lookup_lock(ja->ht, + snapshot[i]); + if (!shadow_node) { + ret = -EAGAIN; + goto end; + } + shadow_nodes[nr_shadow++] = shadow_node; + + /* + * Check if node has been removed between RCU + * lookup and lock acquisition. + */ + assert(snapshot_ptr[i + 1]); + if (ja_node_ptr(*snapshot_ptr[i + 1]) + != ja_node_ptr(snapshot[i + 1])) { + ret = -ENOENT; + goto end; + } + + assert(shadow_node->nr_child > 0); + if (shadow_node->nr_child == 1 && i > 1) + nr_clear++; + nr_branch++; + if (shadow_node->nr_child > 1 || i == 1) { + /* Lock parent and break */ + shadow_node = rcuja_shadow_lookup_lock(ja->ht, + snapshot[i - 1]); + if (!shadow_node) { + ret = -EAGAIN; + goto end; + } + shadow_nodes[nr_shadow++] = shadow_node; + + /* + * Check if node has been removed between RCU + * lookup and lock acquisition. + */ + assert(snapshot_ptr[i]); + if (ja_node_ptr(*snapshot_ptr[i]) + != ja_node_ptr(snapshot[i])) { + ret = -ENOENT; + goto end; + } + + node_flag_ptr = snapshot_ptr[i + 1]; + n = snapshot_n[i + 1]; + parent_node_flag_ptr = snapshot_ptr[i]; + parent_node_flag = snapshot[i]; + + if (i > 1) { + /* + * Lock parent's parent, in case we need + * to recompact parent. + */ + shadow_node = rcuja_shadow_lookup_lock(ja->ht, + snapshot[i - 2]); + if (!shadow_node) { + ret = -EAGAIN; + goto end; + } + shadow_nodes[nr_shadow++] = shadow_node; + + /* + * Check if node has been removed between RCU + * lookup and lock acquisition. + */ + assert(snapshot_ptr[i - 1]); + if (ja_node_ptr(*snapshot_ptr[i - 1]) + != ja_node_ptr(snapshot[i - 1])) { + ret = -ENOENT; + goto end; + } + } + + break; + } + } + + /* + * At this point, we want to delete all nodes that are about to + * be removed from shadow_nodes (except the last one, which is + * either the root or the parent of the upmost node with 1 + * child). OK to free lock here, because RCU read lock is held, + * and free only performed in call_rcu. + */ + + for (i = 0; i < nr_clear; i++) { + ret = rcuja_shadow_clear(ja->ht, + shadow_nodes[i]->node_flag, + shadow_nodes[i], + RCUJA_SHADOW_CLEAR_FREE_NODE + | RCUJA_SHADOW_CLEAR_FREE_LOCK); + assert(!ret); + } + + iter_node_flag = parent_node_flag; + /* Remove from parent */ + ret = ja_node_clear_ptr(ja, + node_flag_ptr, /* Pointer to location to nullify */ + &iter_node_flag, /* Old new parent ptr in its parent */ + shadow_nodes[nr_branch - 1], /* of parent */ + n, nr_branch - 1); + if (ret) + goto end; + + dbg_printf("ja_detach_node: publish %p instead of %p\n", + iter_node_flag, *parent_node_flag_ptr); + /* Update address of parent ptr in its parent */ + rcu_assign_pointer(*parent_node_flag_ptr, iter_node_flag); + +end: + for (i = 0; i < nr_shadow; i++) + rcuja_shadow_unlock(shadow_nodes[i]); + return ret; +} + +static +int ja_unchain_node(struct cds_ja *ja, + struct cds_ja_inode_flag *parent_node_flag, + struct cds_ja_inode_flag **node_flag_ptr, + struct cds_ja_inode_flag *node_flag, + struct cds_ja_node *node) +{ + struct cds_ja_shadow_node *shadow_node; + struct cds_hlist_node *hlist_node; + struct cds_hlist_head hlist_head; + int ret = 0, count = 0, found = 0; + + shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); + if (!shadow_node) + return -EAGAIN; + if (ja_node_ptr(*node_flag_ptr) != ja_node_ptr(node_flag)) { + ret = -EAGAIN; + goto end; + } + hlist_head.next = (struct cds_hlist_node *) ja_node_ptr(node_flag); + /* + * Retry if another thread removed all but one of duplicates + * since check (this check was performed without lock). + * Ensure that the node we are about to remove is still in the + * list (while holding lock). + */ + cds_hlist_for_each_rcu(hlist_node, &hlist_head) { + if (count == 0) { + /* FIXME: currently a work-around */ + hlist_node->prev = (struct cds_hlist_node *) node_flag_ptr; + } + count++; + if (hlist_node == &node->list) + found++; + } + assert(found <= 1); + if (!found || count == 1) { + ret = -EAGAIN; + goto end; + } + cds_hlist_del_rcu(&node->list); + /* + * Validate that we indeed removed the node from linked list. + */ + assert(ja_node_ptr(*node_flag_ptr) != (struct cds_ja_inode *) node); end: + rcuja_shadow_unlock(shadow_node); + return ret; +} + +/* + * Called with RCU read lock held. + */ +int cds_ja_del(struct cds_ja *ja, uint64_t key, + struct cds_ja_node *node) +{ + unsigned int tree_depth, i; + struct cds_ja_inode_flag *snapshot[JA_MAX_DEPTH]; + struct cds_ja_inode_flag **snapshot_ptr[JA_MAX_DEPTH]; + uint8_t snapshot_n[JA_MAX_DEPTH]; + struct cds_ja_inode_flag *node_flag; + struct cds_ja_inode_flag **prev_node_flag_ptr, + **node_flag_ptr; + int nr_snapshot; + int ret; + + if (caa_unlikely(key > ja->key_max)) + return -EINVAL; + tree_depth = ja->tree_depth; + +retry: + nr_snapshot = 0; + dbg_printf("cds_ja_del attempt: key %" PRIu64 ", node %p\n", + key, node); + + /* snapshot for level 0 is only for shadow node lookup */ + snapshot_n[0] = 0; + snapshot_n[1] = 0; + snapshot_ptr[nr_snapshot] = NULL; + snapshot[nr_snapshot++] = (struct cds_ja_inode_flag *) &ja->root; + node_flag = rcu_dereference(ja->root); + prev_node_flag_ptr = &ja->root; + node_flag_ptr = &ja->root; + + /* Iterate on all internal levels */ + for (i = 1; i < tree_depth; i++) { + uint8_t iter_key; + + dbg_printf("cds_ja_del iter node_flag %p\n", + node_flag); + if (!ja_node_ptr(node_flag)) { + return -ENOENT; + } + iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - i - 1))); + snapshot_n[nr_snapshot + 1] = iter_key; + snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; + snapshot[nr_snapshot++] = node_flag; + node_flag = ja_node_get_nth(node_flag, + &node_flag_ptr, + iter_key); + if (node_flag) + prev_node_flag_ptr = node_flag_ptr; + dbg_printf("cds_ja_del iter key lookup %u finds node_flag %p, prev_node_flag_ptr %p\n", + (unsigned int) iter_key, node_flag, + prev_node_flag_ptr); + } + /* + * We reached bottom of tree, try to find the node we are trying + * to remove. Fail if we cannot find it. + */ + if (!ja_node_ptr(node_flag)) { + dbg_printf("cds_ja_del: no node found for key %" PRIu64 "\n", + key); + return -ENOENT; + } else { + struct cds_hlist_head hlist_head; + struct cds_hlist_node *hlist_node; + struct cds_ja_node *entry, *match = NULL; + int count = 0; + + hlist_head.next = + (struct cds_hlist_node *) ja_node_ptr(node_flag); + cds_hlist_for_each_entry_rcu(entry, + hlist_node, + &hlist_head, + list) { + dbg_printf("cds_ja_del: compare %p with entry %p\n", node, entry); + if (entry == node) + match = entry; + count++; + } + if (!match) { + dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64 "\n", node, key); + return -ENOENT; + } + assert(count > 0); + if (count == 1) { + /* + * Removing last of duplicates. Last snapshot + * does not have a shadow node (external leafs). + */ + snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; + snapshot[nr_snapshot++] = node_flag; + ret = ja_detach_node(ja, snapshot, snapshot_ptr, + snapshot_n, nr_snapshot, key, node); + } else { + ret = ja_unchain_node(ja, snapshot[nr_snapshot - 1], + node_flag_ptr, node_flag, match); + } + } + /* + * Explanation of -ENOENT handling: caused by concurrent delete + * between RCU lookup and actual removal. Need to re-do the + * lookup and removal attempt. + */ + if (ret == -EAGAIN || ret == -ENOENT) + goto retry; return ret; } @@ -1246,13 +2233,13 @@ struct cds_ja *_cds_ja_new(unsigned int key_bits, switch (key_bits) { case 8: - ja->key_max = UINT8_MAX; - break; case 16: - ja->key_max = UINT16_MAX; - break; + case 24: case 32: - ja->key_max = UINT32_MAX; + case 40: + case 48: + case 56: + ja->key_max = (1ULL << key_bits) - 1; break; case 64: ja->key_max = UINT64_MAX; @@ -1273,13 +2260,12 @@ struct cds_ja *_cds_ja_new(unsigned int key_bits, * 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); + (struct cds_ja_inode_flag *) &ja->root, + NULL, ja, 0); if (!root_shadow_node) { ret = -ENOMEM; goto ht_node_error; } - root_shadow_node->is_root = 1; return ja; @@ -1293,23 +2279,147 @@ ja_error: return NULL; } +/* + * Called from RCU read-side CS. + */ +__attribute__((visibility("protected"))) +void rcuja_free_all_children(struct cds_ja_shadow_node *shadow_node, + struct cds_ja_inode_flag *node_flag, + void (*free_node_cb)(struct rcu_head *head)) +{ + const struct rcu_flavor_struct *flavor; + unsigned int type_index; + struct cds_ja_inode *node; + const struct cds_ja_type *type; + + flavor = cds_lfht_rcu_flavor(shadow_node->ja->ht); + node = ja_node_ptr(node_flag); + assert(node != NULL); + type_index = ja_node_type(node_flag); + type = &ja_types[type_index]; + + switch (type->type_class) { + case RCU_JA_LINEAR: + { + uint8_t nr_child = + ja_linear_node_get_nr_child(type, node); + unsigned int i; + + for (i = 0; i < nr_child; i++) { + struct cds_ja_inode_flag *iter; + struct cds_hlist_head head; + struct cds_ja_node *entry; + struct cds_hlist_node *pos, *tmp; + uint8_t v; + + ja_linear_node_get_ith_pos(type, node, i, &v, &iter); + if (!iter) + continue; + head.next = (struct cds_hlist_node *) iter; + cds_hlist_for_each_entry_safe(entry, pos, tmp, &head, list) { + flavor->update_call_rcu(&entry->head, free_node_cb); + } + } + break; + } + case RCU_JA_POOL: + { + unsigned int pool_nr; + + for (pool_nr = 0; pool_nr < (1U << type->nr_pool_order); pool_nr++) { + struct cds_ja_inode *pool = + ja_pool_node_get_ith_pool(type, node, pool_nr); + uint8_t nr_child = + ja_linear_node_get_nr_child(type, pool); + unsigned int j; + + for (j = 0; j < nr_child; j++) { + struct cds_ja_inode_flag *iter; + struct cds_hlist_head head; + struct cds_ja_node *entry; + struct cds_hlist_node *pos, *tmp; + uint8_t v; + + ja_linear_node_get_ith_pos(type, pool, j, &v, &iter); + if (!iter) + continue; + head.next = (struct cds_hlist_node *) iter; + cds_hlist_for_each_entry_safe(entry, pos, tmp, &head, list) { + flavor->update_call_rcu(&entry->head, free_node_cb); + } + } + } + break; + } + case RCU_JA_NULL: + break; + case RCU_JA_PIGEON: + { + unsigned int i; + + for (i = 0; i < JA_ENTRY_PER_NODE; i++) { + struct cds_ja_inode_flag *iter; + struct cds_hlist_head head; + struct cds_ja_node *entry; + struct cds_hlist_node *pos, *tmp; + + iter = ja_pigeon_node_get_ith_pos(type, node, i); + if (!iter) + continue; + head.next = (struct cds_hlist_node *) iter; + cds_hlist_for_each_entry_safe(entry, pos, tmp, &head, list) { + flavor->update_call_rcu(&entry->head, free_node_cb); + } + } + break; + } + default: + assert(0); + } +} + +static +void print_debug_fallback_distribution(void) +{ + int i; + + fprintf(stderr, "Fallback node distribution:\n"); + for (i = 0; i < JA_ENTRY_PER_NODE; i++) { + if (!node_fallback_count_distribution[i]) + continue; + fprintf(stderr, " %3u: %4lu\n", + i, node_fallback_count_distribution[i]); + } +} + /* * There should be no more concurrent add to the judy array while it is * being destroyed (ensured by the caller). */ -int cds_ja_destroy(struct cds_ja *ja) +int cds_ja_destroy(struct cds_ja *ja, + void (*free_node_cb)(struct rcu_head *head)) { + const struct rcu_flavor_struct *flavor; int ret; + flavor = cds_lfht_rcu_flavor(ja->ht); rcuja_shadow_prune(ja->ht, - RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK); + RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK, + free_node_cb); + flavor->thread_offline(); ret = rcuja_delete_ht(ja->ht); if (ret) return ret; + flavor->thread_online(); if (uatomic_read(&ja->nr_fallback)) fprintf(stderr, "[warning] RCU Judy Array used %lu fallback node(s)\n", uatomic_read(&ja->nr_fallback)); + fprintf(stderr, "Nodes allocated: %lu, Nodes freed: %lu. Fallback ratio: %g\n", + uatomic_read(&nr_nodes_allocated), + uatomic_read(&nr_nodes_freed), + (double) uatomic_read(&ja->nr_fallback) / (double) uatomic_read(&nr_nodes_allocated)); + print_debug_fallback_distribution(); free(ja); return 0; }