X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=rculfhash.c;h=7d39388f494700a1b06a5df885d5a5350cdf42a8;hp=8644f9c3f23ed184117eb2f74264cd9df121f19d;hb=e53ab1eb49ad61eddcc9481bd6a6fbaf1b257439;hpb=ff0d69de620eab8a49e02724efadf0278516e20d diff --git a/rculfhash.c b/rculfhash.c index 8644f9c..7d39388 100644 --- a/rculfhash.c +++ b/rculfhash.c @@ -4,6 +4,7 @@ * Userspace RCU library - Lock-Free Resizable RCU Hash Table * * Copyright 2010-2011 - Mathieu Desnoyers + * Copyright 2011 - Lai Jiangshan * * This library is free software; you can redistribute it and/or * modify it under the terms of the GNU Lesser General Public @@ -33,27 +34,34 @@ * implementation: * * - RCU read-side critical section allows readers to perform hash - * table lookups and use the returned objects safely by delaying - * memory reclaim of a grace period. + * table lookups, as well as traversals, and use the returned objects + * safely by allowing memory reclaim to take place only after a grace + * period. * - Add and remove operations are lock-free, and do not need to * allocate memory. They need to be executed within RCU read-side * critical section to ensure the objects they read are valid and to * deal with the cmpxchg ABA problem. * - add and add_unique operations are supported. add_unique checks if - * the node key already exists in the hash table. It ensures no key - * duplicata exists. - * - The resize operation executes concurrently with add/remove/lookup. + * the node key already exists in the hash table. It ensures not to + * populate a duplicate key if the node key already exists in the hash + * table. + * - The resize operation executes concurrently with + * add/add_unique/add_replace/remove/lookup/traversal. * - Hash table nodes are contained within a split-ordered list. This * list is ordered by incrementing reversed-bits-hash value. - * - An index of dummy nodes is kept. These dummy nodes are the hash - * table "buckets", and they are also chained together in the - * split-ordered list, which allows recursive expansion. - * - The resize operation for small tables only allows expanding the hash table. - * It is triggered automatically by detecting long chains in the add - * operation. + * - An index of bucket nodes is kept. These bucket nodes are the hash + * table "buckets". These buckets are internal nodes that allow to + * perform a fast hash lookup, similarly to a skip list. These + * buckets are chained together in the split-ordered list, which + * allows recursive expansion by inserting new buckets between the + * existing buckets. The split-ordered list allows adding new buckets + * between existing buckets as the table needs to grow. + * - The resize operation for small tables only allows expanding the + * hash table. It is triggered automatically by detecting long chains + * in the add operation. * - The resize operation for larger tables (and available through an * API) allows both expanding and shrinking the hash table. - * - Per-CPU Split-counters are used to keep track of the number of + * - Split-counters are used to keep track of the number of * nodes within the hash table for automatic resize triggering. * - Resize operation initiated by long chain detection is executed by a * call_rcu thread, which keeps lock-freedom of add and remove. @@ -70,37 +78,136 @@ * (not visible to lookups anymore) before the RCU read-side critical * section held across removal ends. Furthermore, this ensures that * the node with "removed" flag set is removed from the linked-list - * before its memory is reclaimed. Only the thread which removal - * successfully set the "removed" flag (with a cmpxchg) into a node's - * next pointer is considered to have succeeded its removal (and thus - * owns the node to reclaim). Because we garbage-collect starting from - * an invariant node (the start-of-bucket dummy node) up to the - * "removed" node (or find a reverse-hash that is higher), we are sure - * that a successful traversal of the chain leads to a chain that is - * present in the linked-list (the start node is never removed) and - * that is does not contain the "removed" node anymore, even if - * concurrent delete/add operations are changing the structure of the - * list concurrently. - * - The add operation performs gargage collection of buckets if it - * encounters nodes with removed flag set in the bucket where it wants - * to add its new node. This ensures lock-freedom of add operation by + * before its memory is reclaimed. After setting the "removal" flag, + * only the thread which removal is the first to set the "removal + * owner" flag (with an xchg) into a node's next pointer is considered + * to have succeeded its removal (and thus owns the node to reclaim). + * Because we garbage-collect starting from an invariant node (the + * start-of-bucket bucket node) up to the "removed" node (or find a + * reverse-hash that is higher), we are sure that a successful + * traversal of the chain leads to a chain that is present in the + * linked-list (the start node is never removed) and that it does not + * contain the "removed" node anymore, even if concurrent delete/add + * operations are changing the structure of the list concurrently. + * - The add operations perform garbage collection of buckets if they + * encounter nodes with removed flag set in the bucket where they want + * to add their new node. This ensures lock-freedom of add operation by * helping the remover unlink nodes from the list rather than to wait * for it do to so. - * - A RCU "order table" indexed by log2(hash index) is copied and - * expanded by the resize operation. This order table allows finding - * the "dummy node" tables. - * - There is one dummy node table per hash index order. The size of - * each dummy node table is half the number of hashes contained in - * this order (except for order 0). - * - synchronzie_rcu is used to garbage-collect the old dummy node table. - * - The per-order dummy node tables contain a compact version of the - * hash table nodes. These tables are invariant after they are - * populated into the hash table. + * - There are three memory backends for the hash table buckets: the + * "order table", the "chunks", and the "mmap". + * - These bucket containers contain a compact version of the hash table + * nodes. + * - The RCU "order table": + * - has a first level table indexed by log2(hash index) which is + * copied and expanded by the resize operation. This order table + * allows finding the "bucket node" tables. + * - There is one bucket node table per hash index order. The size of + * each bucket node table is half the number of hashes contained in + * this order (except for order 0). + * - The RCU "chunks" is best suited for close interaction with a page + * allocator. It uses a linear array as index to "chunks" containing + * each the same number of buckets. + * - The RCU "mmap" memory backend uses a single memory map to hold + * all buckets. + * - synchronize_rcu is used to garbage-collect the old bucket node table. * - * Dummy node tables: + * Ordering Guarantees: * - * hash table hash table the last all dummy node tables - * order size dummy node 0 1 2 3 4 5 6(index) + * To discuss these guarantees, we first define "read" operation as any + * of the the basic cds_lfht_lookup, cds_lfht_next_duplicate, + * cds_lfht_first, cds_lfht_next operation, as well as + * cds_lfht_add_unique (failure). + * + * We define "read traversal" operation as any of the following + * group of operations + * - cds_lfht_lookup followed by iteration with cds_lfht_next_duplicate + * (and/or cds_lfht_next, although less common). + * - cds_lfht_add_unique (failure) followed by iteration with + * cds_lfht_next_duplicate (and/or cds_lfht_next, although less + * common). + * - cds_lfht_first followed iteration with cds_lfht_next (and/or + * cds_lfht_next_duplicate, although less common). + * + * We define "write" operations as any of cds_lfht_add, cds_lfht_replace, + * cds_lfht_add_unique (success), cds_lfht_add_replace, cds_lfht_del. + * + * When cds_lfht_add_unique succeeds (returns the node passed as + * parameter), it acts as a "write" operation. When cds_lfht_add_unique + * fails (returns a node different from the one passed as parameter), it + * acts as a "read" operation. A cds_lfht_add_unique failure is a + * cds_lfht_lookup "read" operation, therefore, any ordering guarantee + * referring to "lookup" imply any of "lookup" or cds_lfht_add_unique + * (failure). + * + * We define "prior" and "later" node as nodes observable by reads and + * read traversals respectively before and after a write or sequence of + * write operations. + * + * Hash-table operations are often cascaded, for example, the pointer + * returned by a cds_lfht_lookup() might be passed to a cds_lfht_next(), + * whose return value might in turn be passed to another hash-table + * operation. This entire cascaded series of operations must be enclosed + * by a pair of matching rcu_read_lock() and rcu_read_unlock() + * operations. + * + * The following ordering guarantees are offered by this hash table: + * + * A.1) "read" after "write": if there is ordering between a write and a + * later read, then the read is guaranteed to see the write or some + * later write. + * A.2) "read traversal" after "write": given that there is dependency + * ordering between reads in a "read traversal", if there is + * ordering between a write and the first read of the traversal, + * then the "read traversal" is guaranteed to see the write or + * some later write. + * B.1) "write" after "read": if there is ordering between a read and a + * later write, then the read will never see the write. + * B.2) "write" after "read traversal": given that there is dependency + * ordering between reads in a "read traversal", if there is + * ordering between the last read of the traversal and a later + * write, then the "read traversal" will never see the write. + * C) "write" while "read traversal": if a write occurs during a "read + * traversal", the traversal may, or may not, see the write. + * D.1) "write" after "write": if there is ordering between a write and + * a later write, then the later write is guaranteed to see the + * effects of the first write. + * D.2) Concurrent "write" pairs: The system will assign an arbitrary + * order to any pair of concurrent conflicting writes. + * Non-conflicting writes (for example, to different keys) are + * unordered. + * E) If a grace period separates a "del" or "replace" operation + * and a subsequent operation, then that subsequent operation is + * guaranteed not to see the removed item. + * F) Uniqueness guarantee: given a hash table that does not contain + * duplicate items for a given key, there will only be one item in + * the hash table after an arbitrary sequence of add_unique and/or + * add_replace operations. Note, however, that a pair of + * concurrent read operations might well access two different items + * with that key. + * G.1) If a pair of lookups for a given key are ordered (e.g. by a + * memory barrier), then the second lookup will return the same + * node as the previous lookup, or some later node. + * G.2) A "read traversal" that starts after the end of a prior "read + * traversal" (ordered by memory barriers) is guaranteed to see the + * same nodes as the previous traversal, or some later nodes. + * G.3) Concurrent "read" pairs: concurrent reads are unordered. For + * example, if a pair of reads to the same key run concurrently + * with an insertion of that same key, the reads remain unordered + * regardless of their return values. In other words, you cannot + * rely on the values returned by the reads to deduce ordering. + * + * Progress guarantees: + * + * * Reads are wait-free. These operations always move forward in the + * hash table linked list, and this list has no loop. + * * Writes are lock-free. Any retry loop performed by a write operation + * is triggered by progress made within another update operation. + * + * Bucket node tables: + * + * hash table hash table the last all bucket node tables + * order size bucket node 0 1 2 3 4 5 6(index) * table size * 0 1 1 1 * 1 2 1 1 1 @@ -110,17 +217,17 @@ * 5 32 16 1 1 2 4 8 16 * 6 64 32 1 1 2 4 8 16 32 * - * When growing/shrinking, we only focus on the last dummy node table + * When growing/shrinking, we only focus on the last bucket node table * which size is (!order ? 1 : (1 << (order -1))). * * Example for growing/shrinking: - * grow hash table from order 5 to 6: init the index=6 dummy node table - * shrink hash table from order 6 to 5: fini the index=6 dummy node table + * grow hash table from order 5 to 6: init the index=6 bucket node table + * shrink hash table from order 6 to 5: fini the index=6 bucket node table * * A bit of ascii art explanation: * - * Order index is the off-by-one compare to the actual power of 2 because - * we use index 0 to deal with the 0 special-case. + * The order index is the off-by-one compared to the actual power of 2 + * because we use index 0 to deal with the 0 special-case. * * This shows the nodes for a small table ordered by reversed bits: * @@ -149,128 +256,98 @@ */ #define _LGPL_SOURCE +#define _GNU_SOURCE #include #include #include #include #include #include +#include #include "config.h" #include #include +#include #include #include #include #include +#include #include #include -#ifdef DEBUG -#define dbg_printf(fmt, args...) printf("[debug rculfhash] " fmt, ## args) -#else -#define dbg_printf(fmt, args...) -#endif - /* - * Per-CPU split-counters lazily update the global counter each 1024 + * Split-counters lazily update the global counter each 1024 * addition/removal. It automatically keeps track of resize required. * We use the bucket length as indicator for need to expand for small - * tables and machines lacking per-cpu data suppport. + * tables and machines lacking per-cpu data support. */ #define COUNT_COMMIT_ORDER 10 +#define DEFAULT_SPLIT_COUNT_MASK 0xFUL #define CHAIN_LEN_TARGET 1 #define CHAIN_LEN_RESIZE_THRESHOLD 3 /* * Define the minimum table size. */ -#define MIN_TABLE_SIZE 1 - -#if (CAA_BITS_PER_LONG == 32) -#define MAX_TABLE_ORDER 32 -#else -#define MAX_TABLE_ORDER 64 -#endif +#define MIN_TABLE_ORDER 0 +#define MIN_TABLE_SIZE (1UL << MIN_TABLE_ORDER) /* - * Minimum number of dummy nodes to touch per thread to parallelize grow/shrink. + * Minimum number of bucket nodes to touch per thread to parallelize grow/shrink. */ #define MIN_PARTITION_PER_THREAD_ORDER 12 #define MIN_PARTITION_PER_THREAD (1UL << MIN_PARTITION_PER_THREAD_ORDER) -#ifndef min -#define min(a, b) ((a) < (b) ? (a) : (b)) -#endif - -#ifndef max -#define max(a, b) ((a) > (b) ? (a) : (b)) -#endif - /* * The removed flag needs to be updated atomically with the pointer. * It indicates that no node must attach to the node scheduled for * removal, and that node garbage collection must be performed. - * The dummy flag does not require to be updated atomically with the + * The bucket flag does not require to be updated atomically with the * pointer, but it is added as a pointer low bit flag to save space. + * The "removal owner" flag is used to detect which of the "del" + * operation that has set the "removed flag" gets to return the removed + * node to its caller. Note that the replace operation does not need to + * iteract with the "removal owner" flag, because it validates that + * the "removed" flag is not set before performing its cmpxchg. */ #define REMOVED_FLAG (1UL << 0) -#define DUMMY_FLAG (1UL << 1) -#define FLAGS_MASK ((1UL << 2) - 1) +#define BUCKET_FLAG (1UL << 1) +#define REMOVAL_OWNER_FLAG (1UL << 2) +#define FLAGS_MASK ((1UL << 3) - 1) /* Value of the end pointer. Should not interact with flags. */ #define END_VALUE NULL +/* + * ht_items_count: Split-counters counting the number of node addition + * and removal in the table. Only used if the CDS_LFHT_ACCOUNTING flag + * is set at hash table creation. + * + * These are free-running counters, never reset to zero. They count the + * number of add/remove, and trigger every (1 << COUNT_COMMIT_ORDER) + * operations to update the global counter. We choose a power-of-2 value + * for the trigger to deal with 32 or 64-bit overflow of the counter. + */ struct ht_items_count { unsigned long add, del; } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); -struct rcu_level { - /* Note: manually update allocation length when adding a field */ - struct _cds_lfht_node nodes[0]; -}; - -struct rcu_table { - unsigned long size; /* always a power of 2, shared (RCU) */ - unsigned long resize_target; - int resize_initiated; - struct rcu_level *tbl[MAX_TABLE_ORDER]; -}; - -struct cds_lfht { - struct rcu_table t; - cds_lfht_hash_fct hash_fct; - cds_lfht_compare_fct compare_fct; - unsigned long hash_seed; - int flags; - /* - * We need to put the work threads offline (QSBR) when taking this - * mutex, because we use synchronize_rcu within this mutex critical - * section, which waits on read-side critical sections, and could - * therefore cause grace-period deadlock if we hold off RCU G.P. - * completion. - */ - pthread_mutex_t resize_mutex; /* resize mutex: add/del mutex */ - unsigned int in_progress_resize, in_progress_destroy; - void (*cds_lfht_call_rcu)(struct rcu_head *head, - void (*func)(struct rcu_head *head)); - void (*cds_lfht_synchronize_rcu)(void); - void (*cds_lfht_rcu_read_lock)(void); - void (*cds_lfht_rcu_read_unlock)(void); - void (*cds_lfht_rcu_thread_offline)(void); - void (*cds_lfht_rcu_thread_online)(void); - void (*cds_lfht_rcu_register_thread)(void); - void (*cds_lfht_rcu_unregister_thread)(void); - pthread_attr_t *resize_attr; /* Resize threads attributes */ - long count; /* global approximate item count */ - struct ht_items_count *percpu_count; /* per-cpu item count */ -}; - +/* + * rcu_resize_work: Contains arguments passed to RCU worker thread + * responsible for performing lazy resize. + */ struct rcu_resize_work { struct rcu_head head; struct cds_lfht *ht; }; +/* + * partition_resize_work: Contains arguments passed to worker threads + * executing the hash table resize on partitions of the hash table + * assigned to each processor's worker thread. + */ struct partition_resize_work { pthread_t thread_id; struct cds_lfht *ht; @@ -279,13 +356,6 @@ struct partition_resize_work { unsigned long start, unsigned long len); }; -static -void _cds_lfht_add(struct cds_lfht *ht, - unsigned long size, - struct cds_lfht_node *node, - struct cds_lfht_iter *unique_ret, - int dummy); - /* * Algorithm to reverse bits in a word by lookup table, extended to * 64-bit words. @@ -311,7 +381,8 @@ uint8_t bit_reverse_u8(uint8_t v) return BitReverseTable256[v]; } -static __attribute__((unused)) +#if (CAA_BITS_PER_LONG == 32) +static uint32_t bit_reverse_u32(uint32_t v) { return ((uint32_t) bit_reverse_u8(v) << 24) | @@ -319,8 +390,8 @@ uint32_t bit_reverse_u32(uint32_t v) ((uint32_t) bit_reverse_u8(v >> 16) << 8) | ((uint32_t) bit_reverse_u8(v >> 24)); } - -static __attribute__((unused)) +#else +static uint64_t bit_reverse_u64(uint64_t v) { return ((uint64_t) bit_reverse_u8(v) << 56) | @@ -332,6 +403,7 @@ uint64_t bit_reverse_u64(uint64_t v) ((uint64_t) bit_reverse_u8(v >> 48) << 8) | ((uint64_t) bit_reverse_u8(v >> 56)); } +#endif static unsigned long bit_reverse_ulong(unsigned long v) @@ -449,9 +521,9 @@ unsigned int fls_u32(uint32_t x) } #endif -unsigned int fls_ulong(unsigned long x) +unsigned int cds_lfht_fls_ulong(unsigned long x) { -#if (CAA_BITS_PER_lONG == 32) +#if (CAA_BITS_PER_LONG == 32) return fls_u32(x); #else return fls_u64(x); @@ -462,7 +534,7 @@ unsigned int fls_ulong(unsigned long x) * Return the minimum order for which x <= (1UL << order). * Return -1 if x is 0. */ -int get_count_order_u32(uint32_t x) +int cds_lfht_get_count_order_u32(uint32_t x) { if (!x) return -1; @@ -474,181 +546,163 @@ int get_count_order_u32(uint32_t x) * Return the minimum order for which x <= (1UL << order). * Return -1 if x is 0. */ -int get_count_order_ulong(unsigned long x) +int cds_lfht_get_count_order_ulong(unsigned long x) { if (!x) return -1; - return fls_ulong(x - 1); + return cds_lfht_fls_ulong(x - 1); } -#ifdef POISON_FREE -#define poison_free(ptr) \ - do { \ - memset(ptr, 0x42, sizeof(*(ptr))); \ - free(ptr); \ - } while (0) -#else -#define poison_free(ptr) free(ptr) -#endif - static -void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth); - -/* - * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are - * available, then we support hash table item accounting. - * In the unfortunate event the number of CPUs reported would be - * inaccurate, we use modulo arithmetic on the number of CPUs we got. - */ -#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) +void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth); static void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, unsigned long count); static long nr_cpus_mask = -1; +static long split_count_mask = -1; +static int split_count_order = -1; -static -struct ht_items_count *alloc_per_cpu_items_count(void) +#if defined(HAVE_SYSCONF) +static void ht_init_nr_cpus_mask(void) { - struct ht_items_count *count; + long maxcpus; - switch (nr_cpus_mask) { - case -2: - return NULL; - case -1: - { - long maxcpus; + maxcpus = sysconf(_SC_NPROCESSORS_CONF); + if (maxcpus <= 0) { + nr_cpus_mask = -2; + return; + } + /* + * round up number of CPUs to next power of two, so we + * can use & for modulo. + */ + maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus); + nr_cpus_mask = maxcpus - 1; +} +#else /* #if defined(HAVE_SYSCONF) */ +static void ht_init_nr_cpus_mask(void) +{ + nr_cpus_mask = -2; +} +#endif /* #else #if defined(HAVE_SYSCONF) */ - maxcpus = sysconf(_SC_NPROCESSORS_CONF); - if (maxcpus <= 0) { - nr_cpus_mask = -2; - return NULL; - } - /* - * round up number of CPUs to next power of two, so we - * can use & for modulo. - */ - maxcpus = 1UL << get_count_order_ulong(maxcpus); - nr_cpus_mask = maxcpus - 1; +static +void alloc_split_items_count(struct cds_lfht *ht) +{ + if (nr_cpus_mask == -1) { + ht_init_nr_cpus_mask(); + if (nr_cpus_mask < 0) + split_count_mask = DEFAULT_SPLIT_COUNT_MASK; + else + split_count_mask = nr_cpus_mask; + split_count_order = + cds_lfht_get_count_order_ulong(split_count_mask + 1); } - /* Fall-through */ - default: - return calloc(nr_cpus_mask + 1, sizeof(*count)); + + assert(split_count_mask >= 0); + + if (ht->flags & CDS_LFHT_ACCOUNTING) { + ht->split_count = calloc(split_count_mask + 1, + sizeof(struct ht_items_count)); + assert(ht->split_count); + } else { + ht->split_count = NULL; } } static -void free_per_cpu_items_count(struct ht_items_count *count) +void free_split_items_count(struct cds_lfht *ht) { - poison_free(count); + poison_free(ht->split_count); } +#if defined(HAVE_SCHED_GETCPU) static -int ht_get_cpu(void) +int ht_get_split_count_index(unsigned long hash) { int cpu; - assert(nr_cpus_mask >= 0); + assert(split_count_mask >= 0); cpu = sched_getcpu(); - if (unlikely(cpu < 0)) - return cpu; + if (caa_unlikely(cpu < 0)) + return hash & split_count_mask; else - return cpu & nr_cpus_mask; + return cpu & split_count_mask; } - +#else /* #if defined(HAVE_SCHED_GETCPU) */ static -void ht_count_add(struct cds_lfht *ht, unsigned long size) +int ht_get_split_count_index(unsigned long hash) { - unsigned long percpu_count; - int cpu; - - if (unlikely(!ht->percpu_count)) - return; - cpu = ht_get_cpu(); - if (unlikely(cpu < 0)) - return; - percpu_count = uatomic_add_return(&ht->percpu_count[cpu].add, 1); - if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { - long count; - - dbg_printf("add percpu %lu\n", percpu_count); - count = uatomic_add_return(&ht->count, - 1UL << COUNT_COMMIT_ORDER); - /* If power of 2 */ - if (!(count & (count - 1))) { - if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size) - return; - dbg_printf("add set global %ld\n", count); - cds_lfht_resize_lazy_count(ht, size, - count >> (CHAIN_LEN_TARGET - 1)); - } - } + return hash & split_count_mask; } +#endif /* #else #if defined(HAVE_SCHED_GETCPU) */ static -void ht_count_del(struct cds_lfht *ht, unsigned long size) +void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash) { - unsigned long percpu_count; - int cpu; + unsigned long split_count; + int index; + long count; - if (unlikely(!ht->percpu_count)) + if (caa_unlikely(!ht->split_count)) return; - cpu = ht_get_cpu(); - if (unlikely(cpu < 0)) + index = ht_get_split_count_index(hash); + split_count = uatomic_add_return(&ht->split_count[index].add, 1); + if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1))) return; - percpu_count = uatomic_add_return(&ht->percpu_count[cpu].del, 1); - if (unlikely(!(percpu_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { - long count; - - dbg_printf("del percpu %lu\n", percpu_count); - count = uatomic_add_return(&ht->count, - -(1UL << COUNT_COMMIT_ORDER)); - /* If power of 2 */ - if (!(count & (count - 1))) { - if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size) - return; - dbg_printf("del set global %ld\n", count); - /* - * Don't shrink table if the number of nodes is below a - * certain threshold. - */ - if (count < (1UL << COUNT_COMMIT_ORDER) * (nr_cpus_mask + 1)) - return; - cds_lfht_resize_lazy_count(ht, size, - count >> (CHAIN_LEN_TARGET - 1)); - } - } -} + /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */ -#else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ - -static const long nr_cpus_mask = -2; - -static -struct ht_items_count *alloc_per_cpu_items_count(void) -{ - return NULL; -} + dbg_printf("add split count %lu\n", split_count); + count = uatomic_add_return(&ht->count, + 1UL << COUNT_COMMIT_ORDER); + if (caa_likely(count & (count - 1))) + return; + /* Only if global count is power of 2 */ -static -void free_per_cpu_items_count(struct ht_items_count *count) -{ + if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) < size) + return; + dbg_printf("add set global %ld\n", count); + cds_lfht_resize_lazy_count(ht, size, + count >> (CHAIN_LEN_TARGET - 1)); } static -void ht_count_add(struct cds_lfht *ht, unsigned long size) +void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash) { -} + unsigned long split_count; + int index; + long count; -static -void ht_count_del(struct cds_lfht *ht, unsigned long size) -{ -} + if (caa_unlikely(!ht->split_count)) + return; + index = ht_get_split_count_index(hash); + split_count = uatomic_add_return(&ht->split_count[index].del, 1); + if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1))) + return; + /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */ -#endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ + dbg_printf("del split count %lu\n", split_count); + count = uatomic_add_return(&ht->count, + -(1UL << COUNT_COMMIT_ORDER)); + if (caa_likely(count & (count - 1))) + return; + /* Only if global count is power of 2 */ + if ((count >> CHAIN_LEN_RESIZE_THRESHOLD) >= size) + return; + dbg_printf("del set global %ld\n", count); + /* + * Don't shrink table if the number of nodes is below a + * certain threshold. + */ + if (count < (1UL << COUNT_COMMIT_ORDER) * (split_count_mask + 1)) + return; + cds_lfht_resize_lazy_count(ht, size, + count >> (CHAIN_LEN_TARGET - 1)); +} static void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len) @@ -662,14 +716,39 @@ void check_resize(struct cds_lfht *ht, unsigned long size, uint32_t chain_len) * Use bucket-local length for small table expand and for * environments lacking per-cpu data support. */ - if (count >= (1UL << COUNT_COMMIT_ORDER)) + if (count >= (1UL << (COUNT_COMMIT_ORDER + split_count_order))) return; if (chain_len > 100) dbg_printf("WARNING: large chain length: %u.\n", chain_len); - if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) - cds_lfht_resize_lazy(ht, size, - get_count_order_u32(chain_len - (CHAIN_LEN_TARGET - 1))); + if (chain_len >= CHAIN_LEN_RESIZE_THRESHOLD) { + int growth; + + /* + * Ideal growth calculated based on chain length. + */ + growth = cds_lfht_get_count_order_u32(chain_len + - (CHAIN_LEN_TARGET - 1)); + if ((ht->flags & CDS_LFHT_ACCOUNTING) + && (size << growth) + >= (1UL << (COUNT_COMMIT_ORDER + + split_count_order))) { + /* + * If ideal growth expands the hash table size + * beyond the "small hash table" sizes, use the + * maximum small hash table size to attempt + * expanding the hash table. This only applies + * when node accounting is available, otherwise + * the chain length is used to expand the hash + * table in every case. + */ + growth = COUNT_COMMIT_ORDER + split_count_order + - cds_lfht_get_count_order_ulong(size); + if (growth <= 0) + return; + } + cds_lfht_resize_lazy_grow(ht, size, growth); + } } static @@ -685,21 +764,33 @@ int is_removed(struct cds_lfht_node *node) } static -struct cds_lfht_node *flag_removed(struct cds_lfht_node *node) +int is_bucket(struct cds_lfht_node *node) { - return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG); + return ((unsigned long) node) & BUCKET_FLAG; } static -int is_dummy(struct cds_lfht_node *node) +struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node) { - return ((unsigned long) node) & DUMMY_FLAG; + return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG); } static -struct cds_lfht_node *flag_dummy(struct cds_lfht_node *node) +int is_removal_owner(struct cds_lfht_node *node) { - return (struct cds_lfht_node *) (((unsigned long) node) | DUMMY_FLAG); + return ((unsigned long) node) & REMOVAL_OWNER_FLAG; +} + +static +struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node) +{ + return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG); +} + +static +struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node) +{ + return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG); } static @@ -715,7 +806,8 @@ int is_end(struct cds_lfht_node *node) } static -unsigned long _uatomic_max(unsigned long *ptr, unsigned long v) +unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr, + unsigned long v) { unsigned long old1, old2; @@ -725,74 +817,87 @@ unsigned long _uatomic_max(unsigned long *ptr, unsigned long v) if (old2 >= v) return old2; } while ((old1 = uatomic_cmpxchg(ptr, old2, v)) != old2); - return v; + return old2; } static -struct _cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size, - unsigned long hash) +void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order) { - unsigned long index, order; + return ht->mm->alloc_bucket_table(ht, order); +} - assert(size > 0); - index = hash & (size - 1); - /* - * equivalent to get_count_order_ulong(index + 1), but optimizes - * away the non-existing 0 special-case for - * get_count_order_ulong. - */ - order = fls_ulong(index); +/* + * cds_lfht_free_bucket_table() should be called with decreasing order. + * When cds_lfht_free_bucket_table(0) is called, it means the whole + * lfht is destroyed. + */ +static +void cds_lfht_free_bucket_table(struct cds_lfht *ht, unsigned long order) +{ + return ht->mm->free_bucket_table(ht, order); +} - dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n", - hash, index, order, index & (!order ? 0 : ((1UL << (order - 1)) - 1))); +static inline +struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index) +{ + return ht->bucket_at(ht, index); +} - return &ht->t.tbl[order]->nodes[index & (!order ? 0 : ((1UL << (order - 1)) - 1))]; +static inline +struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size, + unsigned long hash) +{ + assert(size > 0); + return bucket_at(ht, hash & (size - 1)); } /* * Remove all logically deleted nodes from a bucket up to a certain node key. */ static -void _cds_lfht_gc_bucket(struct cds_lfht_node *dummy, struct cds_lfht_node *node) +void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *node) { struct cds_lfht_node *iter_prev, *iter, *next, *new_next; - assert(!is_dummy(dummy)); - assert(!is_removed(dummy)); - assert(!is_dummy(node)); + assert(!is_bucket(bucket)); + assert(!is_removed(bucket)); + assert(!is_removal_owner(bucket)); + assert(!is_bucket(node)); assert(!is_removed(node)); + assert(!is_removal_owner(node)); for (;;) { - iter_prev = dummy; - /* We can always skip the dummy node initially */ - iter = rcu_dereference(iter_prev->p.next); + iter_prev = bucket; + /* We can always skip the bucket node initially */ + iter = rcu_dereference(iter_prev->next); assert(!is_removed(iter)); - assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); + assert(!is_removal_owner(iter)); + assert(iter_prev->reverse_hash <= node->reverse_hash); /* - * We should never be called with dummy (start of chain) + * We should never be called with bucket (start of chain) * and logically removed node (end of path compression * marker) being the actual same node. This would be a * bug in the algorithm implementation. */ - assert(dummy != node); + assert(bucket != node); for (;;) { - if (unlikely(is_end(iter))) + if (caa_unlikely(is_end(iter))) return; - if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) + if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash)) return; - next = rcu_dereference(clear_flag(iter)->p.next); - if (likely(is_removed(next))) + next = rcu_dereference(clear_flag(iter)->next); + if (caa_likely(is_removed(next))) break; iter_prev = clear_flag(iter); iter = next; } assert(!is_removed(iter)); - if (is_dummy(iter)) - new_next = flag_dummy(clear_flag(next)); + assert(!is_removal_owner(iter)); + if (is_bucket(iter)) + new_next = flag_bucket(clear_flag(next)); else new_next = clear_flag(next); - (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); + (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next); } - return; } static @@ -801,16 +906,17 @@ int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, struct cds_lfht_node *old_next, struct cds_lfht_node *new_node) { - struct cds_lfht_node *dummy, *ret_next; - struct _cds_lfht_node *lookup; + struct cds_lfht_node *bucket, *ret_next; if (!old_node) /* Return -ENOENT if asked to replace NULL node */ return -ENOENT; assert(!is_removed(old_node)); - assert(!is_dummy(old_node)); + assert(!is_removal_owner(old_node)); + assert(!is_bucket(old_node)); assert(!is_removed(new_node)); - assert(!is_dummy(new_node)); + assert(!is_removal_owner(new_node)); + assert(!is_bucket(new_node)); assert(new_node != old_node); for (;;) { /* Insert after node to be replaced */ @@ -821,9 +927,15 @@ int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, */ return -ENOENT; } - assert(!is_dummy(old_next)); - assert(new_node != clear_flag(old_next)); - new_node->p.next = clear_flag(old_next); + assert(old_next == clear_flag(old_next)); + assert(new_node != old_next); + /* + * REMOVAL_OWNER flag is _NEVER_ set before the REMOVED + * flag. It is either set atomically at the same time + * (replace) or after (del). + */ + assert(!is_removal_owner(old_next)); + new_node->next = old_next; /* * Here is the whole trick for lock-free replace: we add * the replacement node _after_ the node we want to @@ -833,9 +945,14 @@ int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, * next pointer, they will either skip the old node due * to the removal flag and see the new node, or use * the old node, but will not see the new one. + * This is a replacement of a node with another node + * that has the same value: we are therefore not + * removing a value from the hash table. We set both the + * REMOVED and REMOVAL_OWNER flags atomically so we own + * the node after successful cmpxchg. */ - ret_next = uatomic_cmpxchg(&old_node->p.next, - old_next, flag_removed(new_node)); + ret_next = uatomic_cmpxchg(&old_node->next, + old_next, flag_removed_or_removal_owner(new_node)); if (ret_next == old_next) break; /* We performed the replacement. */ old_next = ret_next; @@ -846,11 +963,10 @@ int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, * lookup for the node, and remove it (along with any other * logically removed node) if found. */ - lookup = lookup_bucket(ht, size, bit_reverse_ulong(old_node->p.reverse_hash)); - dummy = (struct cds_lfht_node *) lookup; - _cds_lfht_gc_bucket(dummy, new_node); + bucket = lookup_bucket(ht, size, bit_reverse_ulong(old_node->reverse_hash)); + _cds_lfht_gc_bucket(bucket, new_node); - assert(is_removed(rcu_dereference(old_node->p.next))); + assert(is_removed(CMM_LOAD_SHARED(old_node->next))); return 0; } @@ -860,24 +976,22 @@ int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, */ static void _cds_lfht_add(struct cds_lfht *ht, + unsigned long hash, + cds_lfht_match_fct match, + const void *key, unsigned long size, struct cds_lfht_node *node, struct cds_lfht_iter *unique_ret, - int dummy) + int bucket_flag) { struct cds_lfht_node *iter_prev, *iter, *next, *new_node, *new_next, *return_node; - struct _cds_lfht_node *lookup; + struct cds_lfht_node *bucket; - assert(!is_dummy(node)); + assert(!is_bucket(node)); assert(!is_removed(node)); - if (!size) { - assert(dummy); - assert(!unique_ret); - node->p.next = flag_dummy(get_end()); - return; /* Initial first add (head) */ - } - lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash)); + assert(!is_removal_owner(node)); + bucket = lookup_bucket(ht, size, hash); for (;;) { uint32_t chain_len = 0; @@ -885,28 +999,28 @@ void _cds_lfht_add(struct cds_lfht *ht, * iter_prev points to the non-removed node prior to the * insert location. */ - iter_prev = (struct cds_lfht_node *) lookup; - /* We can always skip the dummy node initially */ - iter = rcu_dereference(iter_prev->p.next); - assert(iter_prev->p.reverse_hash <= node->p.reverse_hash); + iter_prev = bucket; + /* We can always skip the bucket node initially */ + iter = rcu_dereference(iter_prev->next); + assert(iter_prev->reverse_hash <= node->reverse_hash); for (;;) { - if (unlikely(is_end(iter))) + if (caa_unlikely(is_end(iter))) goto insert; - if (likely(clear_flag(iter)->p.reverse_hash > node->p.reverse_hash)) + if (caa_likely(clear_flag(iter)->reverse_hash > node->reverse_hash)) goto insert; - /* dummy node is the first node of the identical-hash-value chain */ - if (dummy && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) + /* bucket node is the first node of the identical-hash-value chain */ + if (bucket_flag && clear_flag(iter)->reverse_hash == node->reverse_hash) goto insert; - next = rcu_dereference(clear_flag(iter)->p.next); - if (unlikely(is_removed(next))) + next = rcu_dereference(clear_flag(iter)->next); + if (caa_unlikely(is_removed(next))) goto gc_node; /* uniquely add */ if (unique_ret - && !is_dummy(next) - && clear_flag(iter)->p.reverse_hash == node->p.reverse_hash) { + && !is_bucket(next) + && clear_flag(iter)->reverse_hash == node->reverse_hash) { struct cds_lfht_iter d_iter = { .node = node, .next = iter, }; /* @@ -915,10 +1029,10 @@ void _cds_lfht_add(struct cds_lfht *ht, * * This semantic ensures no duplicated keys * should ever be observable in the table - * (including observe one node by one node - * by forward iterations) + * (including traversing the table node by + * node by forward iterations) */ - cds_lfht_next_duplicate(ht, &d_iter); + cds_lfht_next_duplicate(ht, match, key, &d_iter); if (!d_iter.node) goto insert; @@ -927,8 +1041,8 @@ void _cds_lfht_add(struct cds_lfht *ht, } /* Only account for identical reverse hash once */ - if (iter_prev->p.reverse_hash != clear_flag(iter)->p.reverse_hash - && !is_dummy(next)) + if (iter_prev->reverse_hash != clear_flag(iter)->reverse_hash + && !is_bucket(next)) check_resize(ht, size, ++chain_len); iter_prev = clear_flag(iter); iter = next; @@ -937,17 +1051,19 @@ void _cds_lfht_add(struct cds_lfht *ht, insert: assert(node != clear_flag(iter)); assert(!is_removed(iter_prev)); + assert(!is_removal_owner(iter_prev)); assert(!is_removed(iter)); + assert(!is_removal_owner(iter)); assert(iter_prev != node); - if (!dummy) - node->p.next = clear_flag(iter); + if (!bucket_flag) + node->next = clear_flag(iter); else - node->p.next = flag_dummy(clear_flag(iter)); - if (is_dummy(iter)) - new_node = flag_dummy(node); + node->next = flag_bucket(clear_flag(iter)); + if (is_bucket(iter)) + new_node = flag_bucket(node); else new_node = node; - if (uatomic_cmpxchg(&iter_prev->p.next, iter, + if (uatomic_cmpxchg(&iter_prev->next, iter, new_node) != iter) { continue; /* retry */ } else { @@ -957,11 +1073,12 @@ void _cds_lfht_add(struct cds_lfht *ht, gc_node: assert(!is_removed(iter)); - if (is_dummy(iter)) - new_next = flag_dummy(clear_flag(next)); + assert(!is_removal_owner(iter)); + if (is_bucket(iter)) + new_next = flag_bucket(clear_flag(next)); else new_next = clear_flag(next); - (void) uatomic_cmpxchg(&iter_prev->p.next, iter, new_next); + (void) uatomic_cmpxchg(&iter_prev->next, iter, new_next); /* retry */ } end: @@ -973,32 +1090,40 @@ end: static int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, - struct cds_lfht_node *node, - int dummy_removal) + struct cds_lfht_node *node) { - struct cds_lfht_node *dummy, *next, *old; - struct _cds_lfht_node *lookup; + struct cds_lfht_node *bucket, *next; if (!node) /* Return -ENOENT if asked to delete NULL node */ return -ENOENT; /* logically delete the node */ - assert(!is_dummy(node)); + assert(!is_bucket(node)); assert(!is_removed(node)); - old = rcu_dereference(node->p.next); - do { - struct cds_lfht_node *new_next; + assert(!is_removal_owner(node)); - next = old; - if (unlikely(is_removed(next))) - return -ENOENT; - if (dummy_removal) - assert(is_dummy(next)); - else - assert(!is_dummy(next)); - new_next = flag_removed(next); - old = uatomic_cmpxchg(&node->p.next, next, new_next); - } while (old != next); + /* + * We are first checking if the node had previously been + * logically removed (this check is not atomic with setting the + * logical removal flag). Return -ENOENT if the node had + * previously been removed. + */ + next = CMM_LOAD_SHARED(node->next); /* next is not dereferenced */ + if (caa_unlikely(is_removed(next))) + return -ENOENT; + assert(!is_bucket(next)); + /* + * The del operation semantic guarantees a full memory barrier + * before the uatomic_or atomic commit of the deletion flag. + */ + cmm_smp_mb__before_uatomic_or(); + /* + * We set the REMOVED_FLAG unconditionally. Note that there may + * be more than one concurrent thread setting this flag. + * Knowing which wins the race will be known after the garbage + * collection phase, stay tuned! + */ + uatomic_or(&node->next, REMOVED_FLAG); /* We performed the (logical) deletion. */ /* @@ -1006,12 +1131,27 @@ int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, * the node, and remove it (along with any other logically removed node) * if found. */ - lookup = lookup_bucket(ht, size, bit_reverse_ulong(node->p.reverse_hash)); - dummy = (struct cds_lfht_node *) lookup; - _cds_lfht_gc_bucket(dummy, node); + bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash)); + _cds_lfht_gc_bucket(bucket, node); - assert(is_removed(rcu_dereference(node->p.next))); - return 0; + assert(is_removed(CMM_LOAD_SHARED(node->next))); + /* + * Last phase: atomically exchange node->next with a version + * having "REMOVAL_OWNER_FLAG" set. If the returned node->next + * pointer did _not_ have "REMOVAL_OWNER_FLAG" set, we now own + * the node and win the removal race. + * It is interesting to note that all "add" paths are forbidden + * to change the next pointer starting from the point where the + * REMOVED_FLAG is set, so here using a read, followed by a + * xchg() suffice to guarantee that the xchg() will ever only + * set the "REMOVAL_OWNER_FLAG" (or change nothing if the flag + * was already set). + */ + if (!is_removal_owner(uatomic_xchg(&node->next, + flag_removal_owner(node->next)))) + return 0; + else + return -ENOENT; } static @@ -1019,9 +1159,9 @@ void *partition_resize_thread(void *arg) { struct partition_resize_work *work = arg; - work->ht->cds_lfht_rcu_register_thread(); + work->ht->flavor->register_thread(); work->fct(work->ht, work->i, work->start, work->len); - work->ht->cds_lfht_rcu_unregister_thread(); + work->ht->flavor->unregister_thread(); return NULL; } @@ -1047,7 +1187,7 @@ void partition_resize_helper(struct cds_lfht *ht, unsigned long i, } else { nr_threads = 1; } - partition_len = len >> get_count_order_ulong(nr_threads); + partition_len = len >> cds_lfht_get_count_order_ulong(nr_threads); work = calloc(nr_threads, sizeof(*work)); assert(work); for (thread = 0; thread < nr_threads; thread++) { @@ -1076,27 +1216,26 @@ void partition_resize_helper(struct cds_lfht *ht, unsigned long i, * many worker threads, based on the number of CPUs available in the system. * This should therefore take care of not having the expand lagging behind too * many concurrent insertion threads by using the scheduler's ability to - * schedule dummy node population fairly with insertions. + * schedule bucket node population fairly with insertions. */ static void init_table_populate_partition(struct cds_lfht *ht, unsigned long i, unsigned long start, unsigned long len) { - unsigned long j; + unsigned long j, size = 1UL << (i - 1); - ht->cds_lfht_rcu_read_lock(); - for (j = start; j < start + len; j++) { - struct cds_lfht_node *new_node = - (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j]; + assert(i > MIN_TABLE_ORDER); + ht->flavor->read_lock(); + for (j = size + start; j < size + start + len; j++) { + struct cds_lfht_node *new_node = bucket_at(ht, j); - dbg_printf("init populate: i %lu j %lu hash %lu\n", - i, j, !i ? 0 : (1UL << (i - 1)) + j); - new_node->p.reverse_hash = - bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); - _cds_lfht_add(ht, !i ? 0 : (1UL << (i - 1)), - new_node, NULL, 1); + assert(j >= size && j < (size << 1)); + dbg_printf("init populate: order %lu index %lu hash %lu\n", + i, j, j); + new_node->reverse_hash = bit_reverse_ulong(j); + _cds_lfht_add(ht, j, NULL, NULL, size, new_node, NULL, 1); } - ht->cds_lfht_rcu_read_unlock(); + ht->flavor->read_unlock(); } static @@ -1105,9 +1244,9 @@ void init_table_populate(struct cds_lfht *ht, unsigned long i, { assert(nr_cpus_mask != -1); if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { - ht->cds_lfht_rcu_thread_online(); + ht->flavor->thread_online(); init_table_populate_partition(ht, i, 0, len); - ht->cds_lfht_rcu_thread_offline(); + ht->flavor->thread_offline(); return; } partition_resize_helper(ht, i, len, init_table_populate_partition); @@ -1121,22 +1260,22 @@ void init_table(struct cds_lfht *ht, dbg_printf("init table: first_order %lu last_order %lu\n", first_order, last_order); + assert(first_order > MIN_TABLE_ORDER); for (i = first_order; i <= last_order; i++) { unsigned long len; - len = !i ? 1 : 1UL << (i - 1); + len = 1UL << (i - 1); dbg_printf("init order %lu len: %lu\n", i, len); /* Stop expand if the resize target changes under us */ - if (CMM_LOAD_SHARED(ht->t.resize_target) < (!i ? 1 : (1UL << i))) + if (CMM_LOAD_SHARED(ht->resize_target) < (1UL << i)) break; - ht->t.tbl[i] = calloc(1, len * sizeof(struct _cds_lfht_node)); - assert(ht->t.tbl[i]); + cds_lfht_alloc_bucket_table(ht, i); /* - * Set all dummy nodes reverse hash values for a level and - * link all dummy nodes into the table. + * Set all bucket nodes reverse hash values for a level and + * link all bucket nodes into the table. */ init_table_populate(ht, i, len); @@ -1144,9 +1283,9 @@ void init_table(struct cds_lfht *ht, * Update table size. */ cmm_smp_wmb(); /* populate data before RCU size */ - CMM_STORE_SHARED(ht->t.size, !i ? 1 : (1UL << i)); + CMM_STORE_SHARED(ht->size, 1UL << i); - dbg_printf("init new size: %lu\n", !i ? 1 : (1UL << i)); + dbg_printf("init new size: %lu\n", 1UL << i); if (CMM_LOAD_SHARED(ht->in_progress_destroy)) break; } @@ -1167,10 +1306,10 @@ void init_table(struct cds_lfht *ht, * Concurrent removal and add operations are helping us perform garbage * collection of logically removed nodes. We guarantee that all logically * removed nodes have been garbage-collected (unlinked) before call_rcu is - * invoked to free a hole level of dummy nodes (after a grace period). + * invoked to free a hole level of bucket nodes (after a grace period). * - * Logical removal and garbage collection can therefore be done in batch or on a - * node-per-node basis, as long as the guarantee above holds. + * Logical removal and garbage collection can therefore be done in batch + * or on a node-per-node basis, as long as the guarantee above holds. * * When we reach a certain length, we can split this removal over many worker * threads, based on the number of CPUs available in the system. This should @@ -1181,21 +1320,22 @@ static void remove_table_partition(struct cds_lfht *ht, unsigned long i, unsigned long start, unsigned long len) { - unsigned long j; + unsigned long j, size = 1UL << (i - 1); - ht->cds_lfht_rcu_read_lock(); - for (j = start; j < start + len; j++) { - struct cds_lfht_node *fini_node = - (struct cds_lfht_node *) &ht->t.tbl[i]->nodes[j]; + assert(i > MIN_TABLE_ORDER); + ht->flavor->read_lock(); + for (j = size + start; j < size + start + len; j++) { + struct cds_lfht_node *fini_bucket = bucket_at(ht, j); + struct cds_lfht_node *parent_bucket = bucket_at(ht, j - size); - dbg_printf("remove entry: i %lu j %lu hash %lu\n", - i, j, !i ? 0 : (1UL << (i - 1)) + j); - fini_node->p.reverse_hash = - bit_reverse_ulong(!i ? 0 : (1UL << (i - 1)) + j); - (void) _cds_lfht_del(ht, !i ? 0 : (1UL << (i - 1)), - fini_node, 1); + assert(j >= size && j < (size << 1)); + dbg_printf("remove entry: order %lu index %lu hash %lu\n", + i, j, j); + /* Set the REMOVED_FLAG to freeze the ->next for gc */ + uatomic_or(&fini_bucket->next, REMOVED_FLAG); + _cds_lfht_gc_bucket(parent_bucket, fini_bucket); } - ht->cds_lfht_rcu_read_unlock(); + ht->flavor->read_unlock(); } static @@ -1204,225 +1344,258 @@ void remove_table(struct cds_lfht *ht, unsigned long i, unsigned long len) assert(nr_cpus_mask != -1); if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) { - ht->cds_lfht_rcu_thread_online(); + ht->flavor->thread_online(); remove_table_partition(ht, i, 0, len); - ht->cds_lfht_rcu_thread_offline(); + ht->flavor->thread_offline(); return; } partition_resize_helper(ht, i, len, remove_table_partition); } +/* + * fini_table() is never called for first_order == 0, which is why + * free_by_rcu_order == 0 can be used as criterion to know if free must + * be called. + */ static void fini_table(struct cds_lfht *ht, unsigned long first_order, unsigned long last_order) { long i; - void *free_by_rcu = NULL; + unsigned long free_by_rcu_order = 0; dbg_printf("fini table: first_order %lu last_order %lu\n", first_order, last_order); - assert(first_order > 0); + assert(first_order > MIN_TABLE_ORDER); for (i = last_order; i >= first_order; i--) { unsigned long len; - len = !i ? 1 : 1UL << (i - 1); + len = 1UL << (i - 1); dbg_printf("fini order %lu len: %lu\n", i, len); /* Stop shrink if the resize target changes under us */ - if (CMM_LOAD_SHARED(ht->t.resize_target) > (1UL << (i - 1))) + if (CMM_LOAD_SHARED(ht->resize_target) > (1UL << (i - 1))) break; cmm_smp_wmb(); /* populate data before RCU size */ - CMM_STORE_SHARED(ht->t.size, 1UL << (i - 1)); + CMM_STORE_SHARED(ht->size, 1UL << (i - 1)); /* * We need to wait for all add operations to reach Q.S. (and * thus use the new table for lookups) before we can start - * releasing the old dummy nodes. Otherwise their lookup will + * releasing the old bucket nodes. Otherwise their lookup will * return a logically removed node as insert position. */ - ht->cds_lfht_synchronize_rcu(); - if (free_by_rcu) - free(free_by_rcu); + ht->flavor->update_synchronize_rcu(); + if (free_by_rcu_order) + cds_lfht_free_bucket_table(ht, free_by_rcu_order); /* - * Set "removed" flag in dummy nodes about to be removed. - * Unlink all now-logically-removed dummy node pointers. + * Set "removed" flag in bucket nodes about to be removed. + * Unlink all now-logically-removed bucket node pointers. * Concurrent add/remove operation are helping us doing * the gc. */ remove_table(ht, i, len); - free_by_rcu = ht->t.tbl[i]; + free_by_rcu_order = i; dbg_printf("fini new size: %lu\n", 1UL << i); if (CMM_LOAD_SHARED(ht->in_progress_destroy)) break; } - if (free_by_rcu) { - ht->cds_lfht_synchronize_rcu(); - free(free_by_rcu); + if (free_by_rcu_order) { + ht->flavor->update_synchronize_rcu(); + cds_lfht_free_bucket_table(ht, free_by_rcu_order); } } static -void cds_lfht_create_dummy(struct cds_lfht *ht, unsigned long size) +void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size) { - struct _cds_lfht_node *prev, *node; - unsigned long order, len, i, j; + struct cds_lfht_node *prev, *node; + unsigned long order, len, i; - ht->t.tbl[0] = calloc(1, sizeof(struct _cds_lfht_node)); - assert(ht->t.tbl[0]); + cds_lfht_alloc_bucket_table(ht, 0); - dbg_printf("create dummy: order %lu index %lu hash %lu\n", 0, 0, 0); - ht->t.tbl[0]->nodes[0].next = flag_dummy(get_end()); - ht->t.tbl[0]->nodes[0].reverse_hash = 0; + dbg_printf("create bucket: order 0 index 0 hash 0\n"); + node = bucket_at(ht, 0); + node->next = flag_bucket(get_end()); + node->reverse_hash = 0; - for (order = 1; order < get_count_order_ulong(size) + 1; order++) { + for (order = 1; order < cds_lfht_get_count_order_ulong(size) + 1; order++) { len = 1UL << (order - 1); - ht->t.tbl[order] = calloc(1, len * sizeof(struct _cds_lfht_node)); - assert(ht->t.tbl[order]); - - i = 0; - prev = ht->t.tbl[i]->nodes; - for (j = 0; j < len; j++) { - if (j & (j - 1)) { /* Between power of 2 */ - prev++; - } else if (j) { /* At each power of 2 */ - i++; - prev = ht->t.tbl[i]->nodes; - } + cds_lfht_alloc_bucket_table(ht, order); + + for (i = 0; i < len; i++) { + /* + * Now, we are trying to init the node with the + * hash=(len+i) (which is also a bucket with the + * index=(len+i)) and insert it into the hash table, + * so this node has to be inserted after the bucket + * with the index=(len+i)&(len-1)=i. And because there + * is no other non-bucket node nor bucket node with + * larger index/hash inserted, so the bucket node + * being inserted should be inserted directly linked + * after the bucket node with index=i. + */ + prev = bucket_at(ht, i); + node = bucket_at(ht, len + i); - node = &ht->t.tbl[order]->nodes[j]; - dbg_printf("create dummy: order %lu index %lu hash %lu\n", - order, j, j + len); + dbg_printf("create bucket: order %lu index %lu hash %lu\n", + order, len + i, len + i); + node->reverse_hash = bit_reverse_ulong(len + i); + + /* insert after prev */ + assert(is_bucket(prev->next)); node->next = prev->next; - assert(is_dummy(node->next)); - node->reverse_hash = bit_reverse_ulong(j + len); - prev->next = flag_dummy((struct cds_lfht_node *)node); + prev->next = flag_bucket(node); } } } -struct cds_lfht *_cds_lfht_new(cds_lfht_hash_fct hash_fct, - cds_lfht_compare_fct compare_fct, - unsigned long hash_seed, - unsigned long init_size, +struct cds_lfht *_cds_lfht_new(unsigned long init_size, + unsigned long min_nr_alloc_buckets, + unsigned long max_nr_buckets, int flags, - void (*cds_lfht_call_rcu)(struct rcu_head *head, - void (*func)(struct rcu_head *head)), - void (*cds_lfht_synchronize_rcu)(void), - void (*cds_lfht_rcu_read_lock)(void), - void (*cds_lfht_rcu_read_unlock)(void), - void (*cds_lfht_rcu_thread_offline)(void), - void (*cds_lfht_rcu_thread_online)(void), - void (*cds_lfht_rcu_register_thread)(void), - void (*cds_lfht_rcu_unregister_thread)(void), + const struct cds_lfht_mm_type *mm, + const struct rcu_flavor_struct *flavor, pthread_attr_t *attr) { struct cds_lfht *ht; unsigned long order; + /* min_nr_alloc_buckets must be power of two */ + if (!min_nr_alloc_buckets || (min_nr_alloc_buckets & (min_nr_alloc_buckets - 1))) + return NULL; + /* init_size must be power of two */ - if (init_size && (init_size & (init_size - 1))) + if (!init_size || (init_size & (init_size - 1))) + return NULL; + + /* + * Memory management plugin default. + */ + if (!mm) { + if (CAA_BITS_PER_LONG > 32 + && max_nr_buckets + && max_nr_buckets <= (1ULL << 32)) { + /* + * For 64-bit architectures, with max number of + * buckets small enough not to use the entire + * 64-bit memory mapping space (and allowing a + * fair number of hash table instances), use the + * mmap allocator, which is faster than the + * order allocator. + */ + mm = &cds_lfht_mm_mmap; + } else { + /* + * The fallback is to use the order allocator. + */ + mm = &cds_lfht_mm_order; + } + } + + /* max_nr_buckets == 0 for order based mm means infinite */ + if (mm == &cds_lfht_mm_order && !max_nr_buckets) + max_nr_buckets = 1UL << (MAX_TABLE_ORDER - 1); + + /* max_nr_buckets must be power of two */ + if (!max_nr_buckets || (max_nr_buckets & (max_nr_buckets - 1))) return NULL; - ht = calloc(1, sizeof(struct cds_lfht)); + + min_nr_alloc_buckets = max(min_nr_alloc_buckets, MIN_TABLE_SIZE); + init_size = max(init_size, MIN_TABLE_SIZE); + max_nr_buckets = max(max_nr_buckets, min_nr_alloc_buckets); + init_size = min(init_size, max_nr_buckets); + + ht = mm->alloc_cds_lfht(min_nr_alloc_buckets, max_nr_buckets); assert(ht); - ht->hash_fct = hash_fct; - ht->compare_fct = compare_fct; - ht->hash_seed = hash_seed; - ht->cds_lfht_call_rcu = cds_lfht_call_rcu; - ht->cds_lfht_synchronize_rcu = cds_lfht_synchronize_rcu; - ht->cds_lfht_rcu_read_lock = cds_lfht_rcu_read_lock; - ht->cds_lfht_rcu_read_unlock = cds_lfht_rcu_read_unlock; - ht->cds_lfht_rcu_thread_offline = cds_lfht_rcu_thread_offline; - ht->cds_lfht_rcu_thread_online = cds_lfht_rcu_thread_online; - ht->cds_lfht_rcu_register_thread = cds_lfht_rcu_register_thread; - ht->cds_lfht_rcu_unregister_thread = cds_lfht_rcu_unregister_thread; + assert(ht->mm == mm); + assert(ht->bucket_at == mm->bucket_at); + + ht->flags = flags; + ht->flavor = flavor; ht->resize_attr = attr; - ht->percpu_count = alloc_per_cpu_items_count(); + alloc_split_items_count(ht); /* this mutex should not nest in read-side C.S. */ pthread_mutex_init(&ht->resize_mutex, NULL); - ht->flags = flags; - order = get_count_order_ulong(max(init_size, MIN_TABLE_SIZE)); - ht->t.resize_target = 1UL << order; - cds_lfht_create_dummy(ht, 1UL << order); - ht->t.size = 1UL << order; + order = cds_lfht_get_count_order_ulong(init_size); + ht->resize_target = 1UL << order; + cds_lfht_create_bucket(ht, 1UL << order); + ht->size = 1UL << order; return ht; } -void cds_lfht_lookup(struct cds_lfht *ht, void *key, size_t key_len, +void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash, + cds_lfht_match_fct match, const void *key, struct cds_lfht_iter *iter) { - struct cds_lfht_node *node, *next, *dummy_node; - struct _cds_lfht_node *lookup; - unsigned long hash, reverse_hash, size; + struct cds_lfht_node *node, *next, *bucket; + unsigned long reverse_hash, size; - hash = ht->hash_fct(key, key_len, ht->hash_seed); reverse_hash = bit_reverse_ulong(hash); - size = rcu_dereference(ht->t.size); - lookup = lookup_bucket(ht, size, hash); - dummy_node = (struct cds_lfht_node *) lookup; - /* We can always skip the dummy node initially */ - node = rcu_dereference(dummy_node->p.next); + size = rcu_dereference(ht->size); + bucket = lookup_bucket(ht, size, hash); + /* We can always skip the bucket node initially */ + node = rcu_dereference(bucket->next); node = clear_flag(node); for (;;) { - if (unlikely(is_end(node))) { + if (caa_unlikely(is_end(node))) { node = next = NULL; break; } - if (unlikely(node->p.reverse_hash > reverse_hash)) { + if (caa_unlikely(node->reverse_hash > reverse_hash)) { node = next = NULL; break; } - next = rcu_dereference(node->p.next); - if (likely(!is_removed(next)) - && !is_dummy(next) - && clear_flag(node)->p.reverse_hash == reverse_hash - && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { + next = rcu_dereference(node->next); + assert(node == clear_flag(node)); + if (caa_likely(!is_removed(next)) + && !is_bucket(next) + && node->reverse_hash == reverse_hash + && caa_likely(match(node, key))) { break; } node = clear_flag(next); } - assert(!node || !is_dummy(rcu_dereference(node->p.next))); + assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next))); iter->node = node; iter->next = next; } -void cds_lfht_next_duplicate(struct cds_lfht *ht, struct cds_lfht_iter *iter) +void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match, + const void *key, struct cds_lfht_iter *iter) { struct cds_lfht_node *node, *next; unsigned long reverse_hash; - void *key; - size_t key_len; node = iter->node; - reverse_hash = node->p.reverse_hash; - key = node->key; - key_len = node->key_len; + reverse_hash = node->reverse_hash; next = iter->next; node = clear_flag(next); for (;;) { - if (unlikely(is_end(node))) { + if (caa_unlikely(is_end(node))) { node = next = NULL; break; } - if (unlikely(node->p.reverse_hash > reverse_hash)) { + if (caa_unlikely(node->reverse_hash > reverse_hash)) { node = next = NULL; break; } - next = rcu_dereference(node->p.next); - if (likely(!is_removed(next)) - && !is_dummy(next) - && likely(!ht->compare_fct(node->key, node->key_len, key, key_len))) { + next = rcu_dereference(node->next); + if (caa_likely(!is_removed(next)) + && !is_bucket(next) + && caa_likely(match(node, key))) { break; } node = clear_flag(next); } - assert(!node || !is_dummy(rcu_dereference(node->p.next))); + assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next))); iter->node = node; iter->next = next; } @@ -1433,77 +1606,75 @@ void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter) node = clear_flag(iter->next); for (;;) { - if (unlikely(is_end(node))) { + if (caa_unlikely(is_end(node))) { node = next = NULL; break; } - next = rcu_dereference(node->p.next); - if (likely(!is_removed(next)) - && !is_dummy(next)) { + next = rcu_dereference(node->next); + if (caa_likely(!is_removed(next)) + && !is_bucket(next)) { break; } node = clear_flag(next); } - assert(!node || !is_dummy(rcu_dereference(node->p.next))); + assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next))); iter->node = node; iter->next = next; } void cds_lfht_first(struct cds_lfht *ht, struct cds_lfht_iter *iter) { - struct _cds_lfht_node *lookup; - /* - * Get next after first dummy node. The first dummy node is the + * Get next after first bucket node. The first bucket node is the * first node of the linked list. */ - lookup = &ht->t.tbl[0]->nodes[0]; - iter->next = lookup->next; + iter->next = bucket_at(ht, 0)->next; cds_lfht_next(ht, iter); } -void cds_lfht_add(struct cds_lfht *ht, struct cds_lfht_node *node) +void cds_lfht_add(struct cds_lfht *ht, unsigned long hash, + struct cds_lfht_node *node) { - unsigned long hash, size; - - hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); - node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); + unsigned long size; - size = rcu_dereference(ht->t.size); - _cds_lfht_add(ht, size, node, NULL, 0); - ht_count_add(ht, size); + node->reverse_hash = bit_reverse_ulong(hash); + size = rcu_dereference(ht->size); + _cds_lfht_add(ht, hash, NULL, NULL, size, node, NULL, 0); + ht_count_add(ht, size, hash); } struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, + unsigned long hash, + cds_lfht_match_fct match, + const void *key, struct cds_lfht_node *node) { - unsigned long hash, size; + unsigned long size; struct cds_lfht_iter iter; - hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); - node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); - - size = rcu_dereference(ht->t.size); - _cds_lfht_add(ht, size, node, &iter, 0); + node->reverse_hash = bit_reverse_ulong(hash); + size = rcu_dereference(ht->size); + _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0); if (iter.node == node) - ht_count_add(ht, size); + ht_count_add(ht, size, hash); return iter.node; } struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht, + unsigned long hash, + cds_lfht_match_fct match, + const void *key, struct cds_lfht_node *node) { - unsigned long hash, size; + unsigned long size; struct cds_lfht_iter iter; - hash = ht->hash_fct(node->key, node->key_len, ht->hash_seed); - node->p.reverse_hash = bit_reverse_ulong((unsigned long) hash); - - size = rcu_dereference(ht->t.size); + node->reverse_hash = bit_reverse_ulong(hash); + size = rcu_dereference(ht->size); for (;;) { - _cds_lfht_add(ht, size, node, &iter, 0); + _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0); if (iter.node == node) { - ht_count_add(ht, size); + ht_count_add(ht, size, hash); return NULL; } @@ -1512,62 +1683,79 @@ struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht, } } -int cds_lfht_replace(struct cds_lfht *ht, struct cds_lfht_iter *old_iter, +int cds_lfht_replace(struct cds_lfht *ht, + struct cds_lfht_iter *old_iter, + unsigned long hash, + cds_lfht_match_fct match, + const void *key, struct cds_lfht_node *new_node) { unsigned long size; - size = rcu_dereference(ht->t.size); + new_node->reverse_hash = bit_reverse_ulong(hash); + if (!old_iter->node) + return -ENOENT; + if (caa_unlikely(old_iter->node->reverse_hash != new_node->reverse_hash)) + return -EINVAL; + if (caa_unlikely(!match(old_iter->node, key))) + return -EINVAL; + size = rcu_dereference(ht->size); return _cds_lfht_replace(ht, size, old_iter->node, old_iter->next, new_node); } -int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_iter *iter) +int cds_lfht_del(struct cds_lfht *ht, struct cds_lfht_node *node) { unsigned long size; int ret; - size = rcu_dereference(ht->t.size); - ret = _cds_lfht_del(ht, size, iter->node, 0); - if (!ret) - ht_count_del(ht, size); + size = rcu_dereference(ht->size); + ret = _cds_lfht_del(ht, size, node); + if (!ret) { + unsigned long hash; + + hash = bit_reverse_ulong(node->reverse_hash); + ht_count_del(ht, size, hash); + } return ret; } +int cds_lfht_is_node_deleted(struct cds_lfht_node *node) +{ + return is_removed(CMM_LOAD_SHARED(node->next)); +} + static -int cds_lfht_delete_dummy(struct cds_lfht *ht) +int cds_lfht_delete_bucket(struct cds_lfht *ht) { struct cds_lfht_node *node; - struct _cds_lfht_node *lookup; unsigned long order, i, size; /* Check that the table is empty */ - lookup = &ht->t.tbl[0]->nodes[0]; - node = (struct cds_lfht_node *) lookup; + node = bucket_at(ht, 0); do { - node = clear_flag(node)->p.next; - if (!is_dummy(node)) + node = clear_flag(node)->next; + if (!is_bucket(node)) return -EPERM; assert(!is_removed(node)); + assert(!is_removal_owner(node)); } while (!is_end(node)); /* * size accessed without rcu_dereference because hash table is * being destroyed. */ - size = ht->t.size; - /* Internal sanity check: all nodes left should be dummy */ - for (order = 0; order < get_count_order_ulong(size) + 1; order++) { - unsigned long len; - - len = !order ? 1 : 1UL << (order - 1); - for (i = 0; i < len; i++) { - dbg_printf("delete order %lu i %lu hash %lu\n", - order, i, - bit_reverse_ulong(ht->t.tbl[order]->nodes[i].reverse_hash)); - assert(is_dummy(ht->t.tbl[order]->nodes[i].next)); - } - poison_free(ht->t.tbl[order]); + size = ht->size; + /* Internal sanity check: all nodes left should be buckets */ + for (i = 0; i < size; i++) { + node = bucket_at(ht, i); + dbg_printf("delete bucket: index %lu expected hash %lu hash %lu\n", + i, i, bit_reverse_ulong(node->reverse_hash)); + assert(is_bucket(node->next)); } + + for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--) + cds_lfht_free_bucket_table(ht, order); + return 0; } @@ -1577,70 +1765,80 @@ int cds_lfht_delete_dummy(struct cds_lfht *ht) */ int cds_lfht_destroy(struct cds_lfht *ht, pthread_attr_t **attr) { - int ret; + int ret, was_online; /* Wait for in-flight resize operations to complete */ _CMM_STORE_SHARED(ht->in_progress_destroy, 1); cmm_smp_mb(); /* Store destroy before load resize */ + was_online = ht->flavor->read_ongoing(); + if (was_online) + ht->flavor->thread_offline(); + /* Calling with RCU read-side held is an error. */ + if (ht->flavor->read_ongoing()) { + ret = -EINVAL; + if (was_online) + ht->flavor->thread_online(); + goto end; + } while (uatomic_read(&ht->in_progress_resize)) poll(NULL, 0, 100); /* wait for 100ms */ - ret = cds_lfht_delete_dummy(ht); + if (was_online) + ht->flavor->thread_online(); + ret = cds_lfht_delete_bucket(ht); if (ret) return ret; - free_per_cpu_items_count(ht->percpu_count); + free_split_items_count(ht); if (attr) *attr = ht->resize_attr; poison_free(ht); +end: return ret; } void cds_lfht_count_nodes(struct cds_lfht *ht, long *approx_before, unsigned long *count, - unsigned long *removed, long *approx_after) { struct cds_lfht_node *node, *next; - struct _cds_lfht_node *lookup; - unsigned long nr_dummy = 0; + unsigned long nr_bucket = 0, nr_removed = 0; *approx_before = 0; - if (nr_cpus_mask >= 0) { + if (ht->split_count) { int i; - for (i = 0; i < nr_cpus_mask + 1; i++) { - *approx_before += uatomic_read(&ht->percpu_count[i].add); - *approx_before -= uatomic_read(&ht->percpu_count[i].del); + for (i = 0; i < split_count_mask + 1; i++) { + *approx_before += uatomic_read(&ht->split_count[i].add); + *approx_before -= uatomic_read(&ht->split_count[i].del); } } *count = 0; - *removed = 0; - /* Count non-dummy nodes in the table */ - lookup = &ht->t.tbl[0]->nodes[0]; - node = (struct cds_lfht_node *) lookup; + /* Count non-bucket nodes in the table */ + node = bucket_at(ht, 0); do { - next = rcu_dereference(node->p.next); + next = rcu_dereference(node->next); if (is_removed(next)) { - if (!is_dummy(next)) - (*removed)++; + if (!is_bucket(next)) + (nr_removed)++; else - (nr_dummy)++; - } else if (!is_dummy(next)) + (nr_bucket)++; + } else if (!is_bucket(next)) (*count)++; else - (nr_dummy)++; + (nr_bucket)++; node = clear_flag(next); } while (!is_end(node)); - dbg_printf("number of dummy nodes: %lu\n", nr_dummy); + dbg_printf("number of logically removed nodes: %lu\n", nr_removed); + dbg_printf("number of bucket nodes: %lu\n", nr_bucket); *approx_after = 0; - if (nr_cpus_mask >= 0) { + if (ht->split_count) { int i; - for (i = 0; i < nr_cpus_mask + 1; i++) { - *approx_after += uatomic_read(&ht->percpu_count[i].add); - *approx_after -= uatomic_read(&ht->percpu_count[i].del); + for (i = 0; i < split_count_mask + 1; i++) { + *approx_after += uatomic_read(&ht->split_count[i].add); + *approx_after -= uatomic_read(&ht->split_count[i].del); } } } @@ -1652,8 +1850,8 @@ void _do_cds_lfht_grow(struct cds_lfht *ht, { unsigned long old_order, new_order; - old_order = get_count_order_ulong(old_size); - new_order = get_count_order_ulong(new_size); + old_order = cds_lfht_get_count_order_ulong(old_size); + new_order = cds_lfht_get_count_order_ulong(new_size); dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", old_size, old_order, new_size, new_order); assert(new_size > old_size); @@ -1668,13 +1866,13 @@ void _do_cds_lfht_shrink(struct cds_lfht *ht, unsigned long old_order, new_order; new_size = max(new_size, MIN_TABLE_SIZE); - old_order = get_count_order_ulong(old_size); - new_order = get_count_order_ulong(new_size); + old_order = cds_lfht_get_count_order_ulong(old_size); + new_order = cds_lfht_get_count_order_ulong(new_size); dbg_printf("resize from %lu (order %lu) to %lu (order %lu) buckets\n", old_size, old_order, new_size, new_order); assert(new_size < old_size); - /* Remove and unlink all dummy nodes to remove. */ + /* Remove and unlink all bucket nodes to remove. */ fini_table(ht, new_order + 1, old_order); } @@ -1692,25 +1890,23 @@ void _do_cds_lfht_resize(struct cds_lfht *ht) assert(uatomic_read(&ht->in_progress_resize)); if (CMM_LOAD_SHARED(ht->in_progress_destroy)) break; - ht->t.resize_initiated = 1; - old_size = ht->t.size; - new_size = CMM_LOAD_SHARED(ht->t.resize_target); + ht->resize_initiated = 1; + old_size = ht->size; + new_size = CMM_LOAD_SHARED(ht->resize_target); if (old_size < new_size) _do_cds_lfht_grow(ht, old_size, new_size); else if (old_size > new_size) _do_cds_lfht_shrink(ht, old_size, new_size); - ht->t.resize_initiated = 0; + ht->resize_initiated = 0; /* write resize_initiated before read resize_target */ cmm_smp_mb(); - } while (ht->t.size != CMM_LOAD_SHARED(ht->t.resize_target)); + } while (ht->size != CMM_LOAD_SHARED(ht->resize_target)); } static -unsigned long resize_target_update(struct cds_lfht *ht, unsigned long size, - int growth_order) +unsigned long resize_target_grow(struct cds_lfht *ht, unsigned long new_size) { - return _uatomic_max(&ht->t.resize_target, - size << growth_order); + return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size); } static @@ -1718,18 +1914,36 @@ void resize_target_update_count(struct cds_lfht *ht, unsigned long count) { count = max(count, MIN_TABLE_SIZE); - uatomic_set(&ht->t.resize_target, count); + count = min(count, ht->max_nr_buckets); + uatomic_set(&ht->resize_target, count); } void cds_lfht_resize(struct cds_lfht *ht, unsigned long new_size) { + int was_online; + + was_online = ht->flavor->read_ongoing(); + if (was_online) + ht->flavor->thread_offline(); + /* Calling with RCU read-side held is an error. */ + if (ht->flavor->read_ongoing()) { + static int print_once; + + if (!CMM_LOAD_SHARED(print_once)) + fprintf(stderr, "[error] rculfhash: cds_lfht_resize " + "called with RCU read-side lock held.\n"); + CMM_STORE_SHARED(print_once, 1); + assert(0); + goto end; + } resize_target_update_count(ht, new_size); - CMM_STORE_SHARED(ht->t.resize_initiated, 1); - ht->cds_lfht_rcu_thread_offline(); + CMM_STORE_SHARED(ht->resize_initiated, 1); pthread_mutex_lock(&ht->resize_mutex); _do_cds_lfht_resize(ht); pthread_mutex_unlock(&ht->resize_mutex); - ht->cds_lfht_rcu_thread_online(); +end: + if (was_online) + ht->flavor->thread_online(); } static @@ -1739,26 +1953,24 @@ void do_resize_cb(struct rcu_head *head) caa_container_of(head, struct rcu_resize_work, head); struct cds_lfht *ht = work->ht; - ht->cds_lfht_rcu_thread_offline(); + ht->flavor->thread_offline(); pthread_mutex_lock(&ht->resize_mutex); _do_cds_lfht_resize(ht); pthread_mutex_unlock(&ht->resize_mutex); - ht->cds_lfht_rcu_thread_online(); + ht->flavor->thread_online(); poison_free(work); cmm_smp_mb(); /* finish resize before decrement */ uatomic_dec(&ht->in_progress_resize); } static -void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth) +void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht) { struct rcu_resize_work *work; - unsigned long target_size; - target_size = resize_target_update(ht, size, growth); /* Store resize_target before read resize_initiated */ cmm_smp_mb(); - if (!CMM_LOAD_SHARED(ht->t.resize_initiated) && size < target_size) { + if (!CMM_LOAD_SHARED(ht->resize_initiated)) { uatomic_inc(&ht->in_progress_resize); cmm_smp_mb(); /* increment resize count before load destroy */ if (CMM_LOAD_SHARED(ht->in_progress_destroy)) { @@ -1766,37 +1978,60 @@ void cds_lfht_resize_lazy(struct cds_lfht *ht, unsigned long size, int growth) return; } work = malloc(sizeof(*work)); + if (work == NULL) { + dbg_printf("error allocating resize work, bailing out\n"); + uatomic_dec(&ht->in_progress_resize); + return; + } work->ht = ht; - ht->cds_lfht_call_rcu(&work->head, do_resize_cb); - CMM_STORE_SHARED(ht->t.resize_initiated, 1); + ht->flavor->update_call_rcu(&work->head, do_resize_cb); + CMM_STORE_SHARED(ht->resize_initiated, 1); } } -#if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) +static +void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth) +{ + unsigned long target_size = size << growth; + + target_size = min(target_size, ht->max_nr_buckets); + if (resize_target_grow(ht, target_size) >= target_size) + return; + __cds_lfht_resize_lazy_launch(ht); +} + +/* + * We favor grow operations over shrink. A shrink operation never occurs + * if a grow operation is queued for lazy execution. A grow operation + * cancels any pending shrink lazy execution. + */ static void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, unsigned long count) { - struct rcu_resize_work *work; - if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) return; - resize_target_update_count(ht, count); - /* Store resize_target before read resize_initiated */ - cmm_smp_mb(); - if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) { - uatomic_inc(&ht->in_progress_resize); - cmm_smp_mb(); /* increment resize count before load destroy */ - if (CMM_LOAD_SHARED(ht->in_progress_destroy)) { - uatomic_dec(&ht->in_progress_resize); + count = max(count, MIN_TABLE_SIZE); + count = min(count, ht->max_nr_buckets); + if (count == size) + return; /* Already the right size, no resize needed */ + if (count > size) { /* lazy grow */ + if (resize_target_grow(ht, count) >= count) return; + } else { /* lazy shrink */ + for (;;) { + unsigned long s; + + s = uatomic_cmpxchg(&ht->resize_target, size, count); + if (s == size) + break; /* no resize needed */ + if (s > size) + return; /* growing is/(was just) in progress */ + if (s <= count) + return; /* some other thread do shrink */ + size = s; } - work = malloc(sizeof(*work)); - work->ht = ht; - ht->cds_lfht_call_rcu(&work->head, do_resize_cb); - CMM_STORE_SHARED(ht->t.resize_initiated, 1); } + __cds_lfht_resize_lazy_launch(ht); } - -#endif