X-Git-Url: http://git.liburcu.org/?a=blobdiff_plain;f=rcuja%2Frcuja.c;h=9398a222bcd526fc4d3dcd1f86e851a02a17ced4;hb=refs%2Fheads%2Furcu%2Frcuja-range;hp=1eeb6552e8380ccbae9863a28b65b074bee21ea2;hpb=c76c77cb74e78ee5916ca7a86d0b164124d62337;p=userspace-rcu.git diff --git a/rcuja/rcuja.c b/rcuja/rcuja.c index 1eeb655..9398a22 100644 --- a/rcuja/rcuja.c +++ b/rcuja/rcuja.c @@ -3,7 +3,8 @@ * * Userspace RCU library - RCU Judy Array * - * Copyright 2012 - Mathieu Desnoyers + * Copyright (C) 2000 - 2002 Hewlett-Packard Company + * Copyright 2012-2013 - Mathieu Desnoyers * * This library is free software; you can redistribute it and/or * modify it under the terms of the GNU Lesser General Public @@ -25,16 +26,14 @@ #include #include #include +#include #include #include #include -#include #include #include -#include #include "rcuja-internal.h" -#include "bitfield.h" #ifndef abs #define abs_int(a) ((int) (a) > 0 ? (int) (a) : -((int) (a))) @@ -121,14 +120,16 @@ const struct cds_ja_type ja_types[] = { { .type_class = RCU_JA_LINEAR, .min_child = 10, .max_child = ja_type_4_max_child, .max_linear_child = ja_type_4_max_linear_child, .order = 7, }, /* Pools may fill sooner than max_child */ + /* This pool is hardcoded at index 5. See ja_node_ptr(). */ { .type_class = RCU_JA_POOL, .min_child = 20, .max_child = ja_type_5_max_child, .max_linear_child = ja_type_5_max_linear_child, .order = 8, .nr_pool_order = ja_type_5_nr_pool_order, .pool_size_order = 7, }, + /* This pool is hardcoded at index 6. See ja_node_ptr(). */ { .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, }, /* * 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, }, + { .type_class = RCU_JA_PIGEON, .min_child = 83, .max_child = ja_type_7_max_child, .order = 10, }, { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, }; @@ -169,14 +170,16 @@ const struct cds_ja_type ja_types[] = { { .type_class = RCU_JA_LINEAR, .min_child = 10, .max_child = ja_type_4_max_child, .max_linear_child = ja_type_4_max_linear_child, .order = 8, }, /* Pools may fill sooner than max_child. */ + /* This pool is hardcoded at index 5. See ja_node_ptr(). */ { .type_class = RCU_JA_POOL, .min_child = 22, .max_child = ja_type_5_max_child, .max_linear_child = ja_type_5_max_linear_child, .order = 9, .nr_pool_order = ja_type_5_nr_pool_order, .pool_size_order = 8, }, + /* This pool is hardcoded at index 6. See ja_node_ptr(). */ { .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, }, /* * 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, }, + { .type_class = RCU_JA_PIGEON, .min_child = 95, .max_child = ja_type_7_max_child, .order = 11, }, { .type_class = RCU_JA_NULL, .min_child = 0, .max_child = ja_type_8_max_child, }, }; @@ -244,8 +247,17 @@ enum ja_recompact { JA_RECOMPACT_DEL, }; -static -unsigned long node_fallback_count_distribution[JA_ENTRY_PER_NODE]; +enum ja_lookup_inequality { + JA_LOOKUP_BE, + JA_LOOKUP_AE, +}; + +enum ja_direction { + JA_LEFT, + JA_RIGHT, + JA_LEFTMOST, + JA_RIGHTMOST, +}; static struct cds_ja_inode *_ja_node_mask_ptr(struct cds_ja_inode_flag *node) @@ -265,31 +277,9 @@ 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) +static +struct cds_ja_inode *alloc_cds_ja_node(struct cds_ja *ja, + const struct cds_ja_type *ja_type) { size_t len = 1U << ja_type->order; void *p; @@ -300,12 +290,16 @@ struct cds_ja_inode *alloc_cds_ja_node(const struct cds_ja_type *ja_type) return NULL; } memset(p, 0, len); + if (ja_debug_counters()) + uatomic_inc(&ja->nr_nodes_allocated); return p; } -void free_cds_ja_node(struct cds_ja_inode *node) +void free_cds_ja_node(struct cds_ja *ja, struct cds_ja_inode *node) { free(node); + if (ja_debug_counters() && node) + uatomic_inc(&ja->nr_nodes_freed); } #define __JA_ALIGN_MASK(v, mask) (((v) + (mask)) & ~(mask)) @@ -335,8 +329,6 @@ 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 **child_node_flag_v, struct cds_ja_inode_flag ***node_flag_ptr, uint8_t n) { @@ -365,15 +357,69 @@ struct cds_ja_inode_flag *ja_linear_node_get_nth(const struct cds_ja_type *type, } 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(child_node_flag_v) && ptr) - *child_node_flag_v = ptr; if (caa_unlikely(node_flag_ptr)) *node_flag_ptr = &pointers[i]; return ptr; } +static +struct cds_ja_inode_flag *ja_linear_node_get_direction(const struct cds_ja_type *type, + struct cds_ja_inode *node, + int n, uint8_t *result_key, + enum ja_direction dir) +{ + uint8_t nr_child; + uint8_t *values; + struct cds_ja_inode_flag **pointers; + struct cds_ja_inode_flag *ptr, *match_ptr = NULL; + unsigned int i; + int match_v; + + assert(type->type_class == RCU_JA_LINEAR || type->type_class == RCU_JA_POOL); + assert(dir == JA_LEFT || dir == JA_RIGHT); + + if (dir == JA_LEFT) { + match_v = -1; + } else { + match_v = JA_ENTRY_PER_NODE; + } + + nr_child = ja_linear_node_get_nr_child(type, node); + cmm_smp_rmb(); /* read nr_child before values and pointers */ + assert(nr_child <= type->max_linear_child); + assert(type->type_class != RCU_JA_LINEAR || nr_child >= type->min_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++) { + unsigned int v; + + v = CMM_LOAD_SHARED(values[i]); + ptr = CMM_LOAD_SHARED(pointers[i]); + if (!ptr) + continue; + if (dir == JA_LEFT) { + if ((int) v < n && (int) v > match_v) { + match_v = v; + match_ptr = ptr; + } + } else { + if ((int) v > n && (int) v < match_v) { + match_v = v; + match_ptr = ptr; + } + } + } + + if (!match_ptr) { + return NULL; + } + assert(match_v >= 0 && match_v < JA_ENTRY_PER_NODE); + + *result_key = (uint8_t) match_v; + return rcu_dereference(match_ptr); +} + static void ja_linear_node_get_ith_pos(const struct cds_ja_type *type, struct cds_ja_inode *node, @@ -397,8 +443,6 @@ 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 *node_flag, - struct cds_ja_inode_flag ***child_node_flag_ptr, - struct cds_ja_inode_flag **child_node_flag_v, struct cds_ja_inode_flag ***node_flag_ptr, uint8_t n) { @@ -435,8 +479,7 @@ struct cds_ja_inode_flag *ja_pool_node_get_nth(const struct cds_ja_type *type, linear = NULL; assert(0); } - return ja_linear_node_get_nth(type, linear, child_node_flag_ptr, - child_node_flag_v, node_flag_ptr, n); + return ja_linear_node_get_nth(type, linear, node_flag_ptr, n); } static @@ -449,29 +492,117 @@ struct cds_ja_inode *ja_pool_node_get_ith_pool(const struct cds_ja_type *type, &node->u.data[(unsigned int) i << type->pool_size_order]; } +static +struct cds_ja_inode_flag *ja_pool_node_get_direction(const struct cds_ja_type *type, + struct cds_ja_inode *node, + int n, uint8_t *result_key, + enum ja_direction dir) +{ + unsigned int pool_nr; + int match_v; + struct cds_ja_inode_flag *match_node_flag = NULL; + + assert(type->type_class == RCU_JA_POOL); + assert(dir == JA_LEFT || dir == JA_RIGHT); + + if (dir == JA_LEFT) { + match_v = -1; + } else { + match_v = JA_ENTRY_PER_NODE; + } + + 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 (dir == JA_LEFT) { + if ((int) v < n && (int) v > match_v) { + match_v = v; + match_node_flag = iter; + } + } else { + if ((int) v > n && (int) v < match_v) { + match_v = v; + match_node_flag = iter; + } + } + } + } + if (match_node_flag) + *result_key = (uint8_t) match_v; + return match_node_flag; +} + 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 **child_node_flag_v, 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_read; + 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_read = rcu_dereference(*child_node_flag); + 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_read) - *child_node_flag_ptr = child_node_flag; - if (caa_unlikely(child_node_flag_v) && child_node_flag_read) - *child_node_flag_v = child_node_flag_read; + child_node_flag_ptr); if (caa_unlikely(node_flag_ptr)) - *node_flag_ptr = child_node_flag; - return child_node_flag_read; + *node_flag_ptr = child_node_flag_ptr; + return child_node_flag; +} + +static +struct cds_ja_inode_flag *ja_pigeon_node_get_direction(const struct cds_ja_type *type, + struct cds_ja_inode *node, + int n, uint8_t *result_key, + enum ja_direction dir) +{ + struct cds_ja_inode_flag **child_node_flag_ptr; + struct cds_ja_inode_flag *child_node_flag; + int i; + + assert(type->type_class == RCU_JA_PIGEON); + assert(dir == JA_LEFT || dir == JA_RIGHT); + + if (dir == JA_LEFT) { + /* n - 1 is first value left of n */ + for (i = n - 1; i >= 0; i--) { + child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[i]; + child_node_flag = rcu_dereference(*child_node_flag_ptr); + if (child_node_flag) { + dbg_printf("ja_pigeon_node_get_left child_node_flag %p\n", + child_node_flag); + *result_key = (uint8_t) i; + return child_node_flag; + } + } + } else { + /* n + 1 is first value right of n */ + for (i = n + 1; i < JA_ENTRY_PER_NODE; i++) { + child_node_flag_ptr = &((struct cds_ja_inode_flag **) node->u.data)[i]; + child_node_flag = rcu_dereference(*child_node_flag_ptr); + if (child_node_flag) { + dbg_printf("ja_pigeon_node_get_right child_node_flag %p\n", + child_node_flag); + *result_key = (uint8_t) i; + return child_node_flag; + } + } + } + return NULL; } static @@ -479,7 +610,7 @@ struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *t struct cds_ja_inode *node, uint8_t i) { - return ja_pigeon_node_get_nth(type, node, NULL, NULL, NULL, i); + return ja_pigeon_node_get_nth(type, node, NULL, i); } /* @@ -488,8 +619,6 @@ struct cds_ja_inode_flag *ja_pigeon_node_get_ith_pos(const struct cds_ja_type *t */ 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 **child_node_flag, struct cds_ja_inode_flag ***node_flag_ptr, uint8_t n) { @@ -505,15 +634,12 @@ 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, child_node_flag, node_flag_ptr, n); case RCU_JA_POOL: return ja_pool_node_get_nth(type, node, node_flag, - child_node_flag_ptr, child_node_flag, node_flag_ptr, n); case RCU_JA_PIGEON: return ja_pigeon_node_get_nth(type, node, - child_node_flag_ptr, child_node_flag, node_flag_ptr, n); default: assert(0); @@ -521,6 +647,58 @@ struct cds_ja_inode_flag *ja_node_get_nth(struct cds_ja_inode_flag *node_flag, } } +static +struct cds_ja_inode_flag *ja_node_get_direction(struct cds_ja_inode_flag *node_flag, + int n, uint8_t *result_key, + enum ja_direction dir) +{ + unsigned int type_index; + struct cds_ja_inode *node; + const struct cds_ja_type *type; + + 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: + return ja_linear_node_get_direction(type, node, n, result_key, dir); + case RCU_JA_POOL: + return ja_pool_node_get_direction(type, node, n, result_key, dir); + case RCU_JA_PIGEON: + return ja_pigeon_node_get_direction(type, node, n, result_key, dir); + default: + assert(0); + return (void *) -1UL; + } +} + +static +struct cds_ja_inode_flag *ja_node_get_leftright(struct cds_ja_inode_flag *node_flag, + unsigned int n, uint8_t *result_key, + enum ja_direction dir) +{ + return ja_node_get_direction(node_flag, n, result_key, dir); +} + +static +struct cds_ja_inode_flag *ja_node_get_minmax(struct cds_ja_inode_flag *node_flag, + uint8_t *result_key, + enum ja_direction dir) +{ + switch (dir) { + case JA_LEFTMOST: + return ja_node_get_direction(node_flag, + -1, result_key, JA_RIGHT); + case JA_RIGHTMOST: + return ja_node_get_direction(node_flag, + JA_ENTRY_PER_NODE, result_key, JA_LEFT); + default: + assert(0); + } +} + static int ja_linear_node_set_nth(const struct cds_ja_type *type, struct cds_ja_inode *node, @@ -536,7 +714,8 @@ 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); @@ -690,11 +869,9 @@ int ja_linear_node_clear_ptr(const struct cds_ja_type *type, nr_child = *nr_child_ptr; assert(nr_child <= type->max_linear_child); - if (shadow_node->fallback_removal_count) { - shadow_node->fallback_removal_count--; - } else { - if (type->type_class == RCU_JA_LINEAR - && shadow_node->nr_child <= type->min_child) { + 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; } @@ -899,12 +1076,10 @@ unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode, } 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(type, node, i); @@ -945,7 +1120,7 @@ unsigned int ja_node_sum_distribution_1d(enum ja_recompact mode, for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { unsigned int distance_to_best; - distance_to_best = abs_int((nr_one[bit_i] << 1U) - distrib_nr_child); + distance_to_best = abs_int(((unsigned 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; @@ -977,11 +1152,14 @@ void ja_node_sum_distribution_2d(enum ja_recompact mode, 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, overall_best_distance = UINT_MAX; + 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: @@ -1001,17 +1179,18 @@ void ja_node_sum_distribution_2d(enum ja_recompact mode, 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]++; + if (v & (1U << bit_i)) { + if (v & (1U << bit_j)) { + nr_2d_11[bit_i][bit_j]++; + } else { + nr_2d_10[bit_i][bit_j]++; + } + } else { + if (v & (1U << bit_j)) { + nr_2d_01[bit_i][bit_j]++; + } else { + nr_2d_00[bit_i][bit_j]++; + } } } } @@ -1043,17 +1222,18 @@ void ja_node_sum_distribution_2d(enum ja_recompact mode, 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]++; + if (v & (1U << bit_i)) { + if (v & (1U << bit_j)) { + nr_2d_11[bit_i][bit_j]++; + } else { + nr_2d_10[bit_i][bit_j]++; + } + } else { + if (v & (1U << bit_j)) { + nr_2d_01[bit_i][bit_j]++; + } else { + nr_2d_00[bit_i][bit_j]++; + } } } } @@ -1064,12 +1244,10 @@ void ja_node_sum_distribution_2d(enum ja_recompact mode, } 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(type, node, i); @@ -1079,17 +1257,18 @@ void ja_node_sum_distribution_2d(enum ja_recompact mode, 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]++; + if (i & (1U << bit_i)) { + if (i & (1U << bit_j)) { + nr_2d_11[bit_i][bit_j]++; + } else { + nr_2d_10[bit_i][bit_j]++; + } + } else { + if (i & (1U << bit_j)) { + nr_2d_01[bit_i][bit_j]++; + } else { + nr_2d_00[bit_i][bit_j]++; + } } } } @@ -1108,17 +1287,18 @@ void ja_node_sum_distribution_2d(enum ja_recompact mode, 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]++; + if (n & (1U << bit_i)) { + if (n & (1U << bit_j)) { + nr_2d_11[bit_i][bit_j]++; + } else { + nr_2d_10[bit_i][bit_j]++; + } + } else { + if (n & (1U << bit_j)) { + nr_2d_01[bit_i][bit_j]++; + } else { + nr_2d_00[bit_i][bit_j]++; + } } } } @@ -1134,20 +1314,21 @@ void ja_node_sum_distribution_2d(enum ja_recompact mode, */ for (bit_i = 0; bit_i < JA_BITS_PER_BYTE; bit_i++) { for (bit_j = 0; bit_j < bit_i; bit_j++) { - unsigned int distance_to_best[4]; + int distance_to_best[4]; - distance_to_best[0] = abs_int((nr_2d_11[bit_i][bit_j] << 2U) - distrib_nr_child); - distance_to_best[1] = abs_int((nr_2d_10[bit_i][bit_j] << 2U) - distrib_nr_child); - distance_to_best[2] = abs_int((nr_2d_01[bit_i][bit_j] << 2U) - distrib_nr_child); - distance_to_best[3] = abs_int((nr_2d_00[bit_i][bit_j] << 2U) - distrib_nr_child); + distance_to_best[0] = ((unsigned int) nr_2d_11[bit_i][bit_j] << 2U) - distrib_nr_child; + distance_to_best[1] = ((unsigned int) nr_2d_10[bit_i][bit_j] << 2U) - distrib_nr_child; + distance_to_best[2] = ((unsigned int) nr_2d_01[bit_i][bit_j] << 2U) - distrib_nr_child; + distance_to_best[3] = ((unsigned int) nr_2d_00[bit_i][bit_j] << 2U) - distrib_nr_child; - /* Consider worse distance to best */ - if (distance_to_best[1] > distance_to_best[0]) + /* 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] > distance_to_best[0]) + 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] > distance_to_best[0]) + 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. @@ -1167,6 +1348,27 @@ void ja_node_sum_distribution_2d(enum ja_recompact mode, _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 @@ -1181,7 +1383,8 @@ int ja_node_recompact(enum ja_recompact mode, 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) + struct cds_ja_inode_flag **nullify_node_flag_ptr, + int level) { unsigned int new_type_index; struct cds_ja_inode *new_node; @@ -1193,64 +1396,35 @@ int ja_node_recompact(enum ja_recompact mode, 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: - if (old_type->type_class == RCU_JA_POOL) { - /* - * For pool type, try redistributing - * into a different distribution of same - * size if we have not reached limits. - */ - if (shadow_node->nr_child + 1 > old_type->max_child) { - new_type_index = old_type_index + 1; - } else if (shadow_node->nr_child + 1 < old_type->min_child) { - new_type_index = old_type_index - 1; - } else { - new_type_index = old_type_index; - } - } else { - new_type_index = old_type_index; - } + 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 { - if (old_type->type_class == RCU_JA_POOL) { - /* - * For pool type, try redistributing - * into a different distribution of same - * size if we have not reached limits. - */ - if (shadow_node->nr_child + 1 > old_type->max_child) { - new_type_index = old_type_index + 1; - } else { - new_type_index = old_type_index; - } - } else { - new_type_index = old_type_index + 1; - } + 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: - if (old_type_index == 0) { - new_type_index = NODE_INDEX_NULL; - } else { - if (old_type->type_class == RCU_JA_POOL) { - /* - * For pool type, try redistributing - * into a different distribution of same - * size if we have not reached limits. - */ - if (shadow_node->nr_child - 1 < old_type->min_child) { - new_type_index = old_type_index - 1; - } else { - new_type_index = old_type_index; - } - } else { - new_type_index = old_type_index - 1; - } - } + 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); @@ -1261,7 +1435,7 @@ retry: /* for fallback */ 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); + new_node = alloc_cds_ja_node(ja, new_type); if (!new_node) return -ENOMEM; @@ -1306,9 +1480,9 @@ retry: /* for fallback */ } dbg_printf("Recompact inherit lock from %p\n", shadow_node); - new_shadow_node = rcuja_shadow_set(ja->ht, new_node_flag, shadow_node, ja); + new_shadow_node = rcuja_shadow_set(ja->ht, new_node_flag, shadow_node, ja, level); if (!new_shadow_node) { - free(new_node); + free_cds_ja_node(ja, new_node); return -ENOMEM; } if (fallback) @@ -1389,12 +1563,10 @@ retry: /* for fallback */ 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); @@ -1434,7 +1606,8 @@ skip_copy: 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]); + if (ja_debug_counters()) + uatomic_inc(&ja->node_fallback_count_distribution[new_shadow_node->nr_child]); } /* Return pointer to new recompacted node through old_node_flag_ptr */ @@ -1449,7 +1622,7 @@ skip_copy: * This synchronizes removal with re-add of that node. */ if (new_type_index == NODE_INDEX_NULL) - flags = RCUJA_SHADOW_CLEAR_FREE_LOCK; + flags |= RCUJA_SHADOW_CLEAR_FREE_LOCK; ret = rcuja_shadow_clear(ja->ht, old_node_flag, shadow_node, flags); assert(!ret); @@ -1473,7 +1646,7 @@ fallback_toosmall: * node within a pool has unused entries. It should * therefore _never_ be too small. */ - //TODO assert(0); + assert(0); /* Fall-through */ case JA_RECOMPACT_ADD_NEXT: @@ -1500,7 +1673,8 @@ fallback_toosmall: } else { new_type_index++; dbg_printf("Add fallback to type %d\n", new_type_index); - uatomic_inc(&ja->nr_fallback); + if (ja_debug_counters()) + uatomic_inc(&ja->nr_fallback); fallback = 1; goto retry; } @@ -1544,7 +1718,8 @@ 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) + struct cds_ja_shadow_node *shadow_node, + int level) { int ret; unsigned int type_index; @@ -1563,12 +1738,12 @@ int ja_node_set_nth(struct cds_ja *ja, case -ENOSPC: /* 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); + shadow_node, node_flag, n, child_node_flag, NULL, level); break; case -ERANGE: /* Node needs to be recompacted. */ ret = ja_node_recompact(JA_RECOMPACT_ADD_SAME, ja, type_index, type, node, - shadow_node, node_flag, n, child_node_flag, NULL); + shadow_node, node_flag, n, child_node_flag, NULL, level); break; } return ret; @@ -1583,7 +1758,7 @@ 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) + uint8_t n, int level) { int ret; unsigned int type_index; @@ -1601,41 +1776,184 @@ int ja_node_clear_ptr(struct cds_ja *ja, /* Should try recompaction. */ ret = ja_node_recompact(JA_RECOMPACT_DEL, ja, type_index, type, node, shadow_node, parent_node_flag_ptr, n, NULL, - node_flag_ptr); + node_flag_ptr, level); } return ret; } -struct cds_hlist_head cds_ja_lookup(struct cds_ja *ja, uint64_t key) +struct cds_ja_node *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 head; + if (caa_unlikely(key > ja->key_max || key == UINT64_MAX)) + return NULL; tree_depth = ja->tree_depth; node_flag = rcu_dereference(ja->root); /* level 0: root node */ if (!ja_node_ptr(node_flag)) - return head; + return NULL; 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, NULL, 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 head; + return NULL; } /* Last level lookup succeded. We got an actual match. */ - head.next = (struct cds_hlist_node *) node_flag; - return head; + return (struct cds_ja_node *) node_flag; +} + +static +struct cds_ja_node *cds_ja_lookup_inequality(struct cds_ja *ja, uint64_t key, + uint64_t *result_key, enum ja_lookup_inequality mode) +{ + int tree_depth, level; + struct cds_ja_inode_flag *node_flag, *cur_node_depth[JA_MAX_DEPTH]; + uint8_t cur_key[JA_MAX_DEPTH]; + uint64_t _result_key = 0; + enum ja_direction dir; + + switch (mode) { + case JA_LOOKUP_BE: + case JA_LOOKUP_AE: + if (caa_unlikely(key > ja->key_max || key == UINT64_MAX)) + return NULL; + break; + default: + return NULL; + } + + memset(cur_node_depth, 0, sizeof(cur_node_depth)); + memset(cur_key, 0, sizeof(cur_key)); + tree_depth = ja->tree_depth; + node_flag = rcu_dereference(ja->root); + cur_node_depth[0] = node_flag; + + /* level 0: root node */ + if (!ja_node_ptr(node_flag)) + return NULL; + + for (level = 1; level < tree_depth; level++) { + uint8_t iter_key; + + iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - level - 1))); + node_flag = ja_node_get_nth(node_flag, NULL, iter_key); + if (!ja_node_ptr(node_flag)) + break; + cur_key[level - 1] = iter_key; + cur_node_depth[level] = node_flag; + dbg_printf("cds_ja_lookup_inequality iter key lookup %u finds node_flag %p\n", + (unsigned int) iter_key, node_flag); + } + + if (level == tree_depth) { + /* Last level lookup succeded. We got an equal match. */ + if (result_key) + *result_key = key; + return (struct cds_ja_node *) node_flag; + } + + /* + * Find highest value left/right of current node. + * Current node is cur_node_depth[level]. + * Start at current level. If we cannot find any key left/right + * of ours, go one level up, seek highest value left/right of + * current (recursively), and when we find one, get the + * rightmost/leftmost child of its rightmost/leftmost child + * (recursively). + */ + switch (mode) { + case JA_LOOKUP_BE: + dir = JA_LEFT; + break; + case JA_LOOKUP_AE: + dir = JA_RIGHT; + break; + default: + assert(0); + } + for (; level > 0; level--) { + uint8_t iter_key; + + iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (tree_depth - level - 1))); + node_flag = ja_node_get_leftright(cur_node_depth[level - 1], + iter_key, &cur_key[level - 1], dir); + dbg_printf("cds_ja_lookup_inequality find sibling from %u at %u finds node_flag %p\n", + (unsigned int) iter_key, (unsigned int) cur_key[level - 1], + node_flag); + /* If found left/right sibling, find rightmost/leftmost child. */ + if (ja_node_ptr(node_flag)) + break; + } + + if (!level) { + /* Reached the root and could not find a left/right sibling. */ + return NULL; + } + + level++; + + /* + * From this point, we are guaranteed to be able to find a + * "below than"/"above than" match. ja_attach_node() and + * ja_detach_node() both guarantee that it is not possible for a + * lookup to reach a dead-end. + */ + + /* + * Find rightmost/leftmost child of rightmost/leftmost child + * (recursively). + */ + switch (mode) { + case JA_LOOKUP_BE: + dir = JA_RIGHTMOST; + break; + case JA_LOOKUP_AE: + dir = JA_LEFTMOST; + break; + default: + assert(0); + } + for (; level < tree_depth; level++) { + node_flag = ja_node_get_minmax(node_flag, &cur_key[level - 1], dir); + dbg_printf("cds_ja_lookup_inequality find minmax at %u finds node_flag %p\n", + (unsigned int) cur_key[level - 1], + node_flag); + if (!ja_node_ptr(node_flag)) + break; + } + + assert(level == tree_depth); + + if (result_key) { + for (level = 1; level < tree_depth; level++) { + _result_key |= ((uint64_t) cur_key[level - 1]) + << (JA_BITS_PER_BYTE * (tree_depth - level - 1)); + } + *result_key = _result_key; + } + return (struct cds_ja_node *) node_flag; +} + +struct cds_ja_node *cds_ja_lookup_below_equal(struct cds_ja *ja, + uint64_t key, uint64_t *result_key) +{ + dbg_printf("cds_ja_lookup_below_equal key %" PRIu64 "\n", key); + return cds_ja_lookup_inequality(ja, key, result_key, JA_LOOKUP_BE); +} + +struct cds_ja_node *cds_ja_lookup_above_equal(struct cds_ja *ja, + uint64_t key, uint64_t *result_key) +{ + dbg_printf("cds_ja_lookup_above_equal key %" PRIu64 "\n", key); + return cds_ja_lookup_inequality(ja, key, result_key, JA_LOOKUP_AE); } /* @@ -1649,47 +1967,52 @@ struct cds_hlist_head cds_ja_lookup(struct cds_ja *ja, uint64_t key) * parent lock (if needed). Then we can proceed to create the new * branch. Publish the new branch, and release locks. * TODO: we currently always take the parent lock even when not needed. + * + * ja_attach_node() ensures that a lookup will _never_ see a branch that + * leads to a dead-end: before attaching a branch, the entire content of + * the new branch is populated, thus creating a cluster, before + * attaching the cluster to the rest of the tree, thus making it visible + * to lookups. */ static 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 **node_flag_ptr, - struct cds_ja_inode_flag *node_flag, - struct cds_ja_inode_flag *parent_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; - 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_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_flag); - 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_flag); + parent_attach_node_flag); if (!parent_shadow_node) { ret = -EAGAIN; goto unlock_shadow; } } - if (node_flag_ptr && ja_node_ptr(*node_flag_ptr)) { + 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 @@ -1699,6 +2022,27 @@ int ja_attach_node(struct cds_ja *ja, goto unlock_parent; } + /* + * Perform a lookup query to handle the case where + * old_node_flag_ptr is NULL. We cannot use it to check if the + * node has been populated between RCU lookup and mutex + * acquisition. + */ + if (!old_node_flag_ptr) { + uint8_t iter_key; + struct cds_ja_inode_flag *lookup_node_flag; + struct cds_ja_inode_flag **lookup_node_flag_ptr; + + iter_key = (uint8_t) (key >> (JA_BITS_PER_BYTE * (ja->tree_depth - level))); + lookup_node_flag = ja_node_get_nth(attach_node_flag, + &lookup_node_flag_ptr, + iter_key); + if (lookup_node_flag) { + ret = -EEXIST; + goto unlock_parent; + } + } + if (attach_node_flag_ptr && ja_node_ptr(*attach_node_flag_ptr) != ja_node_ptr(attach_node_flag)) { /* @@ -1711,48 +2055,56 @@ int ja_attach_node(struct cds_ja *ja, } /* 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; + iter_node_flag = (struct cds_ja_inode_flag *) child_node; - 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); - if (ret) + NULL, i); + if (ret) { + dbg_printf("branch creation error %d\n", 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); - if (ret) + shadow_node, level - 1); + if (ret) { + dbg_printf("branch publish error %d\n", 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, *attach_node_flag_ptr); - rcu_assign_pointer(*attach_node_flag_ptr, iter_node_flag); - /* Success */ ret = 0; @@ -1783,9 +2135,9 @@ end: } /* - * 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. + * Lock the parent containing the pointer to list of duplicates, and add + * node to this list. 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 @@ -1793,111 +2145,165 @@ int ja_chain_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_hlist_head *head, + struct cds_ja_node *last_node, struct cds_ja_node *node) { struct cds_ja_shadow_node *shadow_node; - int ret = 0; + struct cds_ja_node *iter_node; + int ret = 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)) { + /* + * Ensure that previous node is still there at end of list. + */ + iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); + if ((struct cds_ja_node *) ja_node_ptr(*node_flag_ptr) != iter_node) { + ret = -EAGAIN; + goto end; + } + cds_ja_for_each_duplicate(iter_node) { + if (iter_node == last_node) + found = 1; + } + if (!found) { ret = -EAGAIN; goto end; } - cds_hlist_add_head_rcu(&node->list, head); + /* + * Add node to tail of list to ensure that RCU traversals will + * always see either the prior node or the newly added if + * executed concurrently with a sequence of add followed by del + * on the same key. Safe against concurrent RCU read traversals. + */ + node->next = NULL; + rcu_assign_pointer(last_node->next, node); end: rcuja_shadow_unlock(shadow_node); 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 *node, + struct cds_ja_node **unique_node_ret) { unsigned int tree_depth, i; - struct cds_ja_inode_flag **attach_node_flag_ptr, - **node_flag_ptr; - struct cds_ja_inode_flag *node_flag, + struct cds_ja_inode_flag *attach_node_flag, *parent_node_flag, *parent2_node_flag, - *attach_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 || key == UINT64_MAX)) { return -EINVAL; } tree_depth = ja->tree_depth; retry: dbg_printf("cds_ja_add attempt: key %" PRIu64 ", node %p\n", - key, new_node); + key, node); parent2_node_flag = NULL; parent_node_flag = (struct cds_ja_inode_flag *) &ja->root; /* Use root ptr address as key for mutex */ - attach_node_flag_ptr = &ja->root; - attach_node_flag = rcu_dereference(ja->root); - node_flag_ptr = &ja->root; + parent_node_flag_ptr = NULL; node_flag = rcu_dereference(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_add iter attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n", - attach_node_flag_ptr, node_flag_ptr, node_flag); - if (!ja_node_ptr(node_flag)) { - ret = ja_attach_node(ja, attach_node_flag_ptr, - attach_node_flag, - 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, - &attach_node_flag_ptr, - &attach_node_flag, &node_flag_ptr, iter_key); - dbg_printf("cds_ja_add iter key lookup %u finds node_flag %p attach_node_flag_ptr %p node_flag_ptr %p\n", - (unsigned int) iter_key, node_flag, - attach_node_flag_ptr, - 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 attach_node_flag_ptr %p node_flag_ptr %p node_flag %p\n", - attach_node_flag_ptr, node_flag_ptr, node_flag); + 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, - node_flag_ptr, parent_node_flag, - parent2_node_flag, key, i, new_node); + parent_attach_node_flag, + node_flag_ptr, + node_flag, + key, i, node); } else { + struct cds_ja_node *iter_node, *last_node = NULL; + + if (unique_node_ret) { + *unique_node_ret = (struct cds_ja_node *) ja_node_ptr(node_flag); + return -EEXIST; + } + + /* Find last duplicate */ + iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); + cds_ja_for_each_duplicate_rcu(iter_node) + last_node = iter_node; + + 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, - parent_node_flag, - node_flag_ptr, - node_flag, - (struct cds_hlist_head *) attach_node_flag_ptr, - new_node); + parent_attach_node_flag, + attach_node_flag_ptr, + attach_node_flag, + last_node, + node); } if (ret == -EAGAIN || ret == -EEXIST) goto retry; -end: + return ret; } +int cds_ja_add(struct cds_ja *ja, uint64_t key, + struct cds_ja_node *node) +{ + return _cds_ja_add(ja, key, node, NULL); +} + +struct cds_ja_node *cds_ja_add_unique(struct cds_ja *ja, uint64_t key, + struct cds_ja_node *node) +{ + int ret; + struct cds_ja_node *ret_node; + + ret = _cds_ja_add(ja, key, node, &ret_node); + if (ret == -EEXIST) + return ret_node; + else + return 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 @@ -1908,6 +2314,11 @@ end: * 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. + * + * ja_detach_node() ensures that a lookup will _never_ see a branch that + * leads to a dead-end: when removing branch, it makes sure to perform + * the "cut" at the highest node that has only one child, effectively + * replacing it with a NULL pointer. */ static int ja_detach_node(struct cds_ja *ja, @@ -2039,7 +2450,7 @@ int ja_detach_node(struct cds_ja *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); + n, nr_branch - 1); if (ret) goto end; @@ -2062,8 +2473,7 @@ int ja_unchain_node(struct cds_ja *ja, struct cds_ja_node *node) { struct cds_ja_shadow_node *shadow_node; - struct cds_hlist_node *hlist_node; - struct cds_hlist_head hlist_head; + struct cds_ja_node *iter_node, **iter_node_ptr, **prev_node_ptr = NULL; int ret = 0, count = 0, found = 0; shadow_node = rcuja_shadow_lookup_lock(ja->ht, parent_node_flag); @@ -2073,28 +2483,30 @@ int ja_unchain_node(struct cds_ja *ja, 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). + * Find the previous node's next pointer pointing to our node, + * so we can update it. 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). No need for RCU + * traversal here since we hold the lock on the parent. */ - 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; - } + iter_node_ptr = (struct cds_ja_node **) node_flag_ptr; + iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); + cds_ja_for_each_duplicate(iter_node) { count++; - if (hlist_node == &node->list) + if (iter_node == node) { + prev_node_ptr = iter_node_ptr; found++; + } + iter_node_ptr = &iter_node->next; } assert(found <= 1); if (!found || count == 1) { ret = -EAGAIN; goto end; } - cds_hlist_del_rcu(&node->list); + CMM_STORE_SHARED(*prev_node_ptr, node->next); /* * Validate that we indeed removed the node from linked list. */ @@ -2120,7 +2532,7 @@ int cds_ja_del(struct cds_ja *ja, uint64_t key, int nr_snapshot; int ret; - if (caa_unlikely(key > ja->key_max)) + if (caa_unlikely(key > ja->key_max || key == UINT64_MAX)) return -EINVAL; tree_depth = ja->tree_depth; @@ -2152,10 +2564,10 @@ retry: snapshot_ptr[nr_snapshot] = prev_node_flag_ptr; snapshot[nr_snapshot++] = node_flag; node_flag = ja_node_get_nth(node_flag, - &prev_node_flag_ptr, - NULL, &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); @@ -2169,22 +2581,17 @@ retry: key); return -ENOENT; } else { - struct cds_hlist_head hlist_head; - struct cds_hlist_node *hlist_node; - struct cds_ja_node *entry, *match = NULL; + struct cds_ja_node *iter_node, *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; + iter_node = (struct cds_ja_node *) ja_node_ptr(node_flag); + cds_ja_for_each_duplicate_rcu(iter_node) { + dbg_printf("cds_ja_del: compare %p with iter_node %p\n", node, iter_node); + if (iter_node == node) + match = iter_node; count++; } + if (!match) { dbg_printf("cds_ja_del: no node match for node %p key %" PRIu64 "\n", node, key); return -ENOENT; @@ -2255,12 +2662,11 @@ struct cds_ja *_cds_ja_new(unsigned int key_bits, */ root_shadow_node = rcuja_shadow_set(ja->ht, (struct cds_ja_inode_flag *) &ja->root, - NULL, ja); + NULL, ja, 0); if (!root_shadow_node) { ret = -ENOMEM; goto ht_node_error; } - root_shadow_node->level = 0; return ja; @@ -2274,141 +2680,76 @@ 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; - 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_rcu(entry, pos, &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; - uint8_t v; - - ja_linear_node_get_ith_pos(type, node, j, &v, &iter); - if (!iter) - continue; - head.next = (struct cds_hlist_node *) iter; - cds_hlist_for_each_entry_rcu(entry, pos, &head, list) { - flavor->update_call_rcu(&entry->head, free_node_cb); - } - } - } - break; - } - case RCU_JA_NULL: - break; - case RCU_JA_PIGEON: - { - uint8_t nr_child; - unsigned int i; - - nr_child = shadow_node->nr_child; - 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; - - 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_rcu(entry, pos, &head, list) { - flavor->update_call_rcu(&entry->head, free_node_cb); - } - } - break; - } - default: - assert(0); - } -} - static -void print_debug_fallback_distribution(void) +void print_debug_fallback_distribution(struct cds_ja *ja) { int i; fprintf(stderr, "Fallback node distribution:\n"); for (i = 0; i < JA_ENTRY_PER_NODE; i++) { - if (!node_fallback_count_distribution[i]) + if (!ja->node_fallback_count_distribution[i]) continue; fprintf(stderr, " %3u: %4lu\n", - i, node_fallback_count_distribution[i]); + i, ja->node_fallback_count_distribution[i]); + } +} + +static +int ja_final_checks(struct cds_ja *ja) +{ + double fallback_ratio; + unsigned long na, nf, nr_fallback; + int ret = 0; + + if (!ja_debug_counters()) + return 0; + + fallback_ratio = (double) uatomic_read(&ja->nr_fallback); + fallback_ratio /= (double) uatomic_read(&ja->nr_nodes_allocated); + nr_fallback = uatomic_read(&ja->nr_fallback); + if (nr_fallback) + fprintf(stderr, + "[warning] RCU Judy Array used %lu fallback node(s) (ratio: %g)\n", + uatomic_read(&ja->nr_fallback), + fallback_ratio); + + na = uatomic_read(&ja->nr_nodes_allocated); + nf = uatomic_read(&ja->nr_nodes_freed); + dbg_printf("Nodes allocated: %lu, Nodes freed: %lu.\n", na, nf); + if (nr_fallback) + print_debug_fallback_distribution(ja); + + if (na != nf) { + fprintf(stderr, "[error] Judy array leaked %ld nodes. Allocated: %lu, freed: %lu.\n", + (long) na - nf, na, nf); + ret = -1; } + return ret; } /* - * There should be no more concurrent add to the judy array while it is - * being destroyed (ensured by the caller). + * There should be no more concurrent add, delete, nor look-up performed + * on the Judy array while it is being destroyed (ensured by the + * caller). */ -int cds_ja_destroy(struct cds_ja *ja, - void (*free_node_cb)(struct rcu_head *head)) +int cds_ja_destroy(struct cds_ja *ja) { + 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, - free_node_cb); + RCUJA_SHADOW_CLEAR_FREE_NODE | RCUJA_SHADOW_CLEAR_FREE_LOCK); + flavor->thread_offline(); 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)); - print_debug_fallback_distribution(); + + /* Wait for in-flight call_rcu free to complete. */ + flavor->barrier(); + + flavor->thread_online(); + ret = ja_final_checks(ja); free(ja); - return 0; + return ret; }