X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=rculfhash.c;h=d7d107f3788f3a70cec57be5380ae1ff1bbff173;hp=c08447bb1e1e60d84d79667635f2e0e5f3786528;hb=a59f39055b5ecb77b68cf78b9839aa9e8e4ec332;hpb=996ff57cc0490bc4ae26de70e9ebe620ff18515f diff --git a/rculfhash.c b/rculfhash.c index c08447b..d7d107f 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 @@ -33,24 +34,31 @@ * 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 bucket nodes is kept. These bucket 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. + * 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. * - Split-counters are used to keep track of the number of @@ -70,32 +78,131 @@ * (not visible to lookups anymore) before the RCU read-side critical * section held across removal ends. Furthermore, this ensures that * the node with "removed" flag set is removed from the linked-list - * before its memory is reclaimed. Only the thread which removal - * successfully set the "removed" flag (with a cmpxchg) into a node's - * next pointer is considered to have succeeded its removal (and thus - * owns the node to reclaim). Because we garbage-collect starting from - * an invariant node (the start-of-bucket bucket node) up to the - * "removed" node (or find a reverse-hash that is higher), we are sure - * that a successful traversal of the chain leads to a chain that is - * present in the linked-list (the start node is never removed) and - * that is does not contain the "removed" node anymore, even if - * concurrent delete/add operations are changing the structure of the - * list concurrently. - * - The add operation performs gargage collection of buckets if it - * 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 "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). - * - synchronzie_rcu is used to garbage-collect the old bucket node table. - * - The per-order bucket 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. + * + * Ordering Guarantees: + * + * To discuss these guarantees, we first define "read" operation as any + * of the the basic cds_lfht_lookup, cds_lfht_next_duplicate, + * cds_lfht_first, cds_lfht_next operation, as well as + * cds_lfht_add_unique (failure). + * + * We define "read traversal" operation as any of the following + * group of operations + * - cds_lfht_lookup followed by iteration with cds_lfht_next_duplicate + * (and/or cds_lfht_next, although less common). + * - cds_lfht_add_unique (failure) followed by iteration with + * cds_lfht_next_duplicate (and/or cds_lfht_next, although less + * common). + * - cds_lfht_first followed iteration with cds_lfht_next (and/or + * cds_lfht_next_duplicate, although less common). + * + * We define "write" operations as any of cds_lfht_add, cds_lfht_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: * @@ -118,12 +225,12 @@ * 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: - * + * * bits reverse * 0 000 000 * 4 100 001 @@ -133,10 +240,10 @@ * 5 101 101 * 3 011 110 * 7 111 111 - * - * This shows the nodes in order of non-reversed bits, linked by + * + * This shows the nodes in order of non-reversed bits, linked by * reversed-bit order. - * + * * order bits reverse * 0 0 000 000 * 1 | 1 001 100 <- @@ -149,34 +256,34 @@ */ #define _LGPL_SOURCE +#define _GNU_SOURCE #include #include #include #include #include #include +#include +#include #include "config.h" -#include +#include "compat-getcpu.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 - /* * 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 @@ -186,13 +293,8 @@ /* * 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 bucket nodes to touch per thread to parallelize grow/shrink. @@ -200,24 +302,22 @@ #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 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 BUCKET_FLAG (1UL << 1) -#define FLAGS_MASK ((1UL << 2) - 1) +#define REMOVAL_OWNER_FLAG (1UL << 2) +#define FLAGS_MASK ((1UL << 3) - 1) /* Value of the end pointer. Should not interact with flags. */ #define END_VALUE NULL @@ -236,63 +336,6 @@ struct ht_items_count { unsigned long add, del; } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); -/* - * rcu_level: Contains the per order-index-level bucket node table. The - * size of each bucket node table is half the number of hashes contained - * in this order (except for order 0). The minimum allocation size - * parameter allows combining the bucket node arrays of the lowermost - * levels to improve cache locality for small index orders. - */ -struct rcu_level { - /* Note: manually update allocation length when adding a field */ - struct cds_lfht_node nodes[0]; -}; - -/* - * rcu_table: Contains the size and desired new size if a resize - * operation is in progress, as well as the statically-sized array of - * rcu_level pointers. - */ -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]; -}; - -/* - * cds_lfht: Top-level data structure representing a lock-free hash - * table. Defined in the implementation file to make it be an opaque - * cookie to users. - */ -struct cds_lfht { - struct rcu_table t; - unsigned long min_alloc_order; - unsigned long min_alloc_size; - 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 *split_count; /* split item count */ -}; - /* * rcu_resize_work: Contains arguments passed to RCU worker thread * responsible for performing lazy resize. @@ -315,15 +358,6 @@ struct partition_resize_work { unsigned long start, unsigned long len); }; -static -void _cds_lfht_add(struct cds_lfht *ht, - cds_lfht_match_fct match, - const void *key, - unsigned long size, - struct cds_lfht_node *node, - struct cds_lfht_iter *unique_ret, - int bucket); - /* * Algorithm to reverse bits in a word by lookup table, extended to * 64-bit words. @@ -332,7 +366,7 @@ void _cds_lfht_add(struct cds_lfht *ht, * Originally from Public Domain. */ -static const uint8_t BitReverseTable256[256] = +static const uint8_t BitReverseTable256[256] = { #define R2(n) (n), (n) + 2*64, (n) + 1*64, (n) + 3*64 #define R4(n) R2(n), R2((n) + 2*16), R2((n) + 1*16), R2((n) + 3*16) @@ -349,27 +383,29 @@ 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) | - ((uint32_t) bit_reverse_u8(v >> 8) << 16) | - ((uint32_t) bit_reverse_u8(v >> 16) << 8) | + return ((uint32_t) bit_reverse_u8(v) << 24) | + ((uint32_t) bit_reverse_u8(v >> 8) << 16) | + ((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) | - ((uint64_t) bit_reverse_u8(v >> 8) << 48) | + return ((uint64_t) bit_reverse_u8(v) << 56) | + ((uint64_t) bit_reverse_u8(v >> 8) << 48) | ((uint64_t) bit_reverse_u8(v >> 16) << 40) | ((uint64_t) bit_reverse_u8(v >> 24) << 32) | - ((uint64_t) bit_reverse_u8(v >> 32) << 24) | - ((uint64_t) bit_reverse_u8(v >> 40) << 16) | + ((uint64_t) bit_reverse_u8(v >> 32) << 24) | + ((uint64_t) bit_reverse_u8(v >> 40) << 16) | ((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) @@ -392,7 +428,7 @@ unsigned int fls_u32(uint32_t x) { int r; - asm("bsrl %1,%0\n\t" + __asm__ ("bsrl %1,%0\n\t" "jnz 1f\n\t" "movl $-1,%0\n\t" "1:\n\t" @@ -408,7 +444,7 @@ unsigned int fls_u64(uint64_t x) { long r; - asm("bsrq %1,%0\n\t" + __asm__ ("bsrq %1,%0\n\t" "jnz 1f\n\t" "movq $-1,%0\n\t" "1:\n\t" @@ -487,7 +523,7 @@ 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) return fls_u32(x); @@ -500,7 +536,7 @@ 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; @@ -512,26 +548,14 @@ 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 { \ - if (ptr) { \ - memset(ptr, 0x42, sizeof(*(ptr))); \ - free(ptr); \ - } \ - } while (0) -#else -#define poison_free(ptr) free(ptr) -#endif - static void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int growth); @@ -541,6 +565,7 @@ void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, static long nr_cpus_mask = -1; static long split_count_mask = -1; +static int split_count_order = -1; #if defined(HAVE_SYSCONF) static void ht_init_nr_cpus_mask(void) @@ -556,7 +581,7 @@ static void ht_init_nr_cpus_mask(void) * round up number of CPUs to next power of two, so we * can use & for modulo. */ - maxcpus = 1UL << get_count_order_ulong(maxcpus); + maxcpus = 1UL << cds_lfht_get_count_order_ulong(maxcpus); nr_cpus_mask = maxcpus - 1; } #else /* #if defined(HAVE_SYSCONF) */ @@ -569,20 +594,21 @@ static void ht_init_nr_cpus_mask(void) static void alloc_split_items_count(struct cds_lfht *ht) { - struct ht_items_count *count; - 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); } assert(split_count_mask >= 0); if (ht->flags & CDS_LFHT_ACCOUNTING) { - ht->split_count = calloc(split_count_mask + 1, sizeof(*count)); + ht->split_count = calloc(split_count_mask + 1, + sizeof(struct ht_items_count)); assert(ht->split_count); } else { ht->split_count = NULL; @@ -595,52 +621,46 @@ void free_split_items_count(struct cds_lfht *ht) poison_free(ht->split_count); } -#if defined(HAVE_SCHED_GETCPU) static int ht_get_split_count_index(unsigned long hash) { int cpu; assert(split_count_mask >= 0); - cpu = sched_getcpu(); + cpu = urcu_sched_getcpu(); if (caa_unlikely(cpu < 0)) return hash & split_count_mask; else return cpu & split_count_mask; } -#else /* #if defined(HAVE_SCHED_GETCPU) */ -static -int ht_get_split_count_index(unsigned long hash) -{ - return hash & split_count_mask; -} -#endif /* #else #if defined(HAVE_SCHED_GETCPU) */ static void ht_count_add(struct cds_lfht *ht, unsigned long size, unsigned long hash) { unsigned long split_count; int index; + long count; if (caa_unlikely(!ht->split_count)) return; index = ht_get_split_count_index(hash); split_count = uatomic_add_return(&ht->split_count[index].add, 1); - if (caa_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { - long count; - - dbg_printf("add split count %lu\n", split_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)); - } - } + if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1))) + return; + /* Only if number of add multiple of 1UL << COUNT_COMMIT_ORDER */ + + 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 */ + + 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 @@ -648,32 +668,34 @@ void ht_count_del(struct cds_lfht *ht, unsigned long size, unsigned long hash) { unsigned long split_count; int index; + long count; 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_unlikely(!(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1)))) { - long count; - - dbg_printf("del split count %lu\n", split_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) * (split_count_mask + 1)) - return; - cds_lfht_resize_lazy_count(ht, size, - count >> (CHAIN_LEN_TARGET - 1)); - } - } + if (caa_likely(split_count & ((1UL << COUNT_COMMIT_ORDER) - 1))) + return; + /* Only if number of deletes multiple of 1UL << COUNT_COMMIT_ORDER */ + + 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 @@ -688,14 +710,39 @@ 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_grow(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 @@ -710,12 +757,6 @@ int is_removed(struct cds_lfht_node *node) return ((unsigned long) node) & REMOVED_FLAG; } -static -struct cds_lfht_node *flag_removed(struct cds_lfht_node *node) -{ - return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG); -} - static int is_bucket(struct cds_lfht_node *node) { @@ -728,6 +769,24 @@ struct cds_lfht_node *flag_bucket(struct cds_lfht_node *node) return (struct cds_lfht_node *) (((unsigned long) node) | BUCKET_FLAG); } +static +int is_removal_owner(struct cds_lfht_node *node) +{ + return ((unsigned long) node) & REMOVAL_OWNER_FLAG; +} + +static +struct cds_lfht_node *flag_removal_owner(struct cds_lfht_node *node) +{ + return (struct cds_lfht_node *) (((unsigned long) node) | REMOVAL_OWNER_FLAG); +} + +static +struct cds_lfht_node *flag_removed_or_removal_owner(struct cds_lfht_node *node) +{ + return (struct cds_lfht_node *) (((unsigned long) node) | REMOVED_FLAG | REMOVAL_OWNER_FLAG); +} + static struct cds_lfht_node *get_end(void) { @@ -756,28 +815,34 @@ unsigned long _uatomic_xchg_monotonic_increase(unsigned long *ptr, } static +void cds_lfht_alloc_bucket_table(struct cds_lfht *ht, unsigned long order) +{ + return ht->mm->alloc_bucket_table(ht, order); +} + +/* + * 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); +} + +static inline +struct cds_lfht_node *bucket_at(struct cds_lfht *ht, unsigned long index) +{ + return ht->bucket_at(ht, index); +} + +static inline struct cds_lfht_node *lookup_bucket(struct cds_lfht *ht, unsigned long size, unsigned long hash) { - unsigned long index, order; - assert(size > 0); - index = hash & (size - 1); - - if (index < ht->min_alloc_size) { - dbg_printf("lookup hash %lu index %lu order 0 aridx 0\n", - hash, index); - return &ht->t.tbl[0]->nodes[index]; - } - /* - * 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); - dbg_printf("lookup hash %lu index %lu order %lu aridx %lu\n", - hash, index, order, index & ((1UL << (order - 1)) - 1)); - return &ht->t.tbl[order]->nodes[index & ((1UL << (order - 1)) - 1)]; + return bucket_at(ht, hash & (size - 1)); } /* @@ -790,13 +855,16 @@ void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *nod 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 = bucket; /* We can always skip the bucket node initially */ iter = rcu_dereference(iter_prev->next); assert(!is_removed(iter)); + assert(!is_removal_owner(iter)); assert(iter_prev->reverse_hash <= node->reverse_hash); /* * We should never be called with bucket (start of chain) @@ -817,13 +885,13 @@ void _cds_lfht_gc_bucket(struct cds_lfht_node *bucket, struct cds_lfht_node *nod iter = next; } assert(!is_removed(iter)); + 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->next, iter, new_next); } - return; } static @@ -838,8 +906,10 @@ int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, return -ENOENT; assert(!is_removed(old_node)); + assert(!is_removal_owner(old_node)); assert(!is_bucket(old_node)); assert(!is_removed(new_node)); + assert(!is_removal_owner(new_node)); assert(!is_bucket(new_node)); assert(new_node != old_node); for (;;) { @@ -851,9 +921,15 @@ int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, */ return -ENOENT; } - assert(!is_bucket(old_next)); - assert(new_node != clear_flag(old_next)); - new_node->next = clear_flag(old_next); + assert(old_next == clear_flag(old_next)); + assert(new_node != old_next); + /* + * 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 @@ -863,9 +939,14 @@ 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->next, - old_next, flag_removed(new_node)); + old_next, flag_removed_or_removal_owner(new_node)); if (ret_next == old_next) break; /* We performed the replacement. */ old_next = ret_next; @@ -879,7 +960,7 @@ int _cds_lfht_replace(struct cds_lfht *ht, unsigned long size, 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->next))); + assert(is_removed(CMM_LOAD_SHARED(old_node->next))); return 0; } @@ -889,6 +970,7 @@ 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, @@ -902,7 +984,8 @@ void _cds_lfht_add(struct cds_lfht *ht, assert(!is_bucket(node)); assert(!is_removed(node)); - bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash)); + assert(!is_removal_owner(node)); + bucket = lookup_bucket(ht, size, hash); for (;;) { uint32_t chain_len = 0; @@ -940,8 +1023,8 @@ 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, match, key, &d_iter); if (!d_iter.node) @@ -962,7 +1045,9 @@ 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 (!bucket_flag) node->next = clear_flag(iter); @@ -982,6 +1067,7 @@ void _cds_lfht_add(struct cds_lfht *ht, gc_node: assert(!is_removed(iter)); + assert(!is_removal_owner(iter)); if (is_bucket(iter)) new_next = flag_bucket(clear_flag(next)); else @@ -998,10 +1084,9 @@ end: static int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, - struct cds_lfht_node *node, - int bucket_removal) + struct cds_lfht_node *node) { - struct cds_lfht_node *bucket, *next, *old; + struct cds_lfht_node *bucket, *next; if (!node) /* Return -ENOENT if asked to delete NULL node */ return -ENOENT; @@ -1009,20 +1094,30 @@ int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, /* logically delete the node */ assert(!is_bucket(node)); assert(!is_removed(node)); - old = rcu_dereference(node->next); - do { - struct cds_lfht_node *new_next; + assert(!is_removal_owner(node)); - next = old; - if (caa_unlikely(is_removed(next))) - return -ENOENT; - if (bucket_removal) - assert(is_bucket(next)); - else - assert(!is_bucket(next)); - new_next = flag_removed(next); - old = uatomic_cmpxchg(&node->next, next, new_next); - } while (old != next); + /* + * We are first checking if the node had previously been + * logically removed (this check is not atomic with setting the + * logical removal flag). Return -ENOENT if the node had + * previously been removed. + */ + next = 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. */ /* @@ -1033,8 +1128,24 @@ int _cds_lfht_del(struct cds_lfht *ht, unsigned long size, bucket = lookup_bucket(ht, size, bit_reverse_ulong(node->reverse_hash)); _cds_lfht_gc_bucket(bucket, node); - assert(is_removed(rcu_dereference(node->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 @@ -1042,9 +1153,9 @@ 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; } @@ -1054,11 +1165,15 @@ void partition_resize_helper(struct cds_lfht *ht, unsigned long i, void (*fct)(struct cds_lfht *ht, unsigned long i, unsigned long start, unsigned long len)) { - unsigned long partition_len; + unsigned long partition_len, start = 0; struct partition_resize_work *work; int thread, ret; unsigned long nr_threads; + assert(nr_cpus_mask != -1); + if (nr_cpus_mask < 0 || len < 2 * MIN_PARTITION_PER_THREAD) + goto fallback; + /* * Note: nr_cpus_mask + 1 is always power of 2. * We spawn just the number of threads we need to satisfy the minimum @@ -1070,9 +1185,12 @@ 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); + if (!work) { + dbg_printf("error allocating for resize, single-threading\n"); + goto fallback; + } for (thread = 0; thread < nr_threads; thread++) { work[thread].ht = ht; work[thread].i = i; @@ -1081,6 +1199,17 @@ void partition_resize_helper(struct cds_lfht *ht, unsigned long i, work[thread].fct = fct; ret = pthread_create(&(work[thread].thread_id), ht->resize_attr, partition_resize_thread, &work[thread]); + if (ret == EAGAIN) { + /* + * Out of resources: wait and join the threads + * we've created, then handle leftovers. + */ + dbg_printf("error spawning for resize, single-threading\n"); + start = work[thread].start; + len -= start; + nr_threads = thread; + break; + } assert(!ret); } for (thread = 0; thread < nr_threads; thread++) { @@ -1088,6 +1217,18 @@ void partition_resize_helper(struct cds_lfht *ht, unsigned long i, assert(!ret); } free(work); + + /* + * A pthread_create failure above will either lead in us having + * no threads to join or starting at a non-zero offset, + * fallback to single thread processing of leftovers. + */ + if (start == 0 && nr_threads > 0) + return; +fallback: + ht->flavor->thread_online(); + fct(ht, i, start, len); + ht->flavor->thread_offline(); } /* @@ -1105,34 +1246,26 @@ 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); - assert(i > ht->min_alloc_order); - ht->cds_lfht_rcu_read_lock(); - for (j = start; j < start + len; j++) { - struct cds_lfht_node *new_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, (1UL << (i - 1)) + j); - new_node->reverse_hash = - bit_reverse_ulong((1UL << (i - 1)) + j); - _cds_lfht_add(ht, NULL, NULL, 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 void init_table_populate(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(); - init_table_populate_partition(ht, i, 0, len); - ht->cds_lfht_rcu_thread_offline(); - return; - } partition_resize_helper(ht, i, len, init_table_populate_partition); } @@ -1144,7 +1277,7 @@ 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 > ht->min_alloc_order); + assert(first_order > MIN_TABLE_ORDER); for (i = first_order; i <= last_order; i++) { unsigned long len; @@ -1152,11 +1285,10 @@ void init_table(struct cds_lfht *ht, 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) < (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 bucket nodes reverse hash values for a level and @@ -1168,7 +1300,7 @@ void init_table(struct cds_lfht *ht, * Update table size. */ cmm_smp_wmb(); /* populate data before RCU size */ - CMM_STORE_SHARED(ht->t.size, 1UL << i); + CMM_STORE_SHARED(ht->size, 1UL << i); dbg_printf("init new size: %lu\n", 1UL << i); if (CMM_LOAD_SHARED(ht->in_progress_destroy)) @@ -1193,8 +1325,8 @@ void init_table(struct cds_lfht *ht, * removed nodes have been garbage-collected (unlinked) before call_rcu is * 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 @@ -1205,58 +1337,57 @@ 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); - assert(i > ht->min_alloc_order); - ht->cds_lfht_rcu_read_lock(); - for (j = start; j < start + len; j++) { - struct cds_lfht_node *fini_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, (1UL << (i - 1)) + j); - fini_node->reverse_hash = - bit_reverse_ulong((1UL << (i - 1)) + j); - (void) _cds_lfht_del(ht, 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 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(); - remove_table_partition(ht, i, 0, len); - ht->cds_lfht_rcu_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 > ht->min_alloc_order); + assert(first_order > MIN_TABLE_ORDER); for (i = last_order; i >= first_order; i--) { unsigned long len; len = 1UL << (i - 1); - dbg_printf("fini order %lu len: %lu\n", i, len); + dbg_printf("fini order %ld 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 @@ -1264,9 +1395,9 @@ void fini_table(struct cds_lfht *ht, * 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 bucket nodes about to be removed. @@ -1276,16 +1407,16 @@ void fini_table(struct cds_lfht *ht, */ 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); } } @@ -1293,91 +1424,117 @@ static void cds_lfht_create_bucket(struct cds_lfht *ht, unsigned long size) { struct cds_lfht_node *prev, *node; - unsigned long order, len, i, j; + unsigned long order, len, i; - ht->t.tbl[0] = calloc(1, ht->min_alloc_size * sizeof(struct cds_lfht_node)); - assert(ht->t.tbl[0]); + cds_lfht_alloc_bucket_table(ht, 0); - dbg_printf("create bucket: order %lu index %lu hash %lu\n", 0, 0, 0); - ht->t.tbl[0]->nodes[0].next = flag_bucket(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); - if (order <= ht->min_alloc_order) { - ht->t.tbl[order] = (struct rcu_level *) (ht->t.tbl[0]->nodes + len); - } else { - ht->t.tbl[order] = calloc(1, len * sizeof(struct cds_lfht_node)); - assert(ht->t.tbl[order]); - } + cds_lfht_alloc_bucket_table(ht, 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; - } + 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 bucket: order %lu index %lu hash %lu\n", - order, j, j + len); + 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_bucket(node->next)); - node->reverse_hash = bit_reverse_ulong(j + len); prev->next = flag_bucket(node); } } } struct cds_lfht *_cds_lfht_new(unsigned long init_size, - unsigned long min_alloc_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_alloc_size must be power of two */ - if (!min_alloc_size || (min_alloc_size & (min_alloc_size - 1))) + /* 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))) return NULL; - min_alloc_size = max(min_alloc_size, MIN_TABLE_SIZE); - init_size = max(init_size, min_alloc_size); - ht = calloc(1, sizeof(struct cds_lfht)); + + /* + * 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; + + 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); + assert(ht->mm == mm); + assert(ht->bucket_at == mm->bucket_at); + ht->flags = flags; - 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; + ht->flavor = flavor; ht->resize_attr = attr; alloc_split_items_count(ht); /* this mutex should not nest in read-side C.S. */ pthread_mutex_init(&ht->resize_mutex, NULL); - order = get_count_order_ulong(init_size); - ht->t.resize_target = 1UL << order; - ht->min_alloc_size = min_alloc_size; - ht->min_alloc_order = get_count_order_ulong(min_alloc_size); + order = cds_lfht_get_count_order_ulong(init_size); + ht->resize_target = 1UL << order; cds_lfht_create_bucket(ht, 1UL << order); - ht->t.size = 1UL << order; + ht->size = 1UL << order; return ht; } @@ -1390,7 +1547,7 @@ void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash, reverse_hash = bit_reverse_ulong(hash); - size = rcu_dereference(ht->t.size); + size = rcu_dereference(ht->size); bucket = lookup_bucket(ht, size, hash); /* We can always skip the bucket node initially */ node = rcu_dereference(bucket->next); @@ -1414,7 +1571,7 @@ void cds_lfht_lookup(struct cds_lfht *ht, unsigned long hash, } node = clear_flag(next); } - assert(!node || !is_bucket(rcu_dereference(node->next))); + assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next))); iter->node = node; iter->next = next; } @@ -1447,7 +1604,7 @@ void cds_lfht_next_duplicate(struct cds_lfht *ht, cds_lfht_match_fct match, } node = clear_flag(next); } - assert(!node || !is_bucket(rcu_dereference(node->next))); + assert(!node || !is_bucket(CMM_LOAD_SHARED(node->next))); iter->node = node; iter->next = next; } @@ -1469,21 +1626,18 @@ void cds_lfht_next(struct cds_lfht *ht, struct cds_lfht_iter *iter) } node = clear_flag(next); } - assert(!node || !is_bucket(rcu_dereference(node->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 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); } @@ -1492,9 +1646,9 @@ void cds_lfht_add(struct cds_lfht *ht, unsigned long hash, { unsigned long size; - node->reverse_hash = bit_reverse_ulong((unsigned long) hash); - size = rcu_dereference(ht->t.size); - _cds_lfht_add(ht, NULL, NULL, size, node, NULL, 0); + 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); } @@ -1507,9 +1661,9 @@ struct cds_lfht_node *cds_lfht_add_unique(struct cds_lfht *ht, unsigned long size; struct cds_lfht_iter iter; - node->reverse_hash = bit_reverse_ulong((unsigned long) hash); - size = rcu_dereference(ht->t.size); - _cds_lfht_add(ht, match, key, 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, hash); return iter.node; @@ -1524,10 +1678,10 @@ struct cds_lfht_node *cds_lfht_add_replace(struct cds_lfht *ht, unsigned long size; struct cds_lfht_iter iter; - node->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, match, key, size, node, &iter, 0); + _cds_lfht_add(ht, hash, match, key, size, node, &iter, 0); if (iter.node == node) { ht_count_add(ht, size, hash); return NULL; @@ -1538,30 +1692,48 @@ 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, hash; + unsigned long size; int ret; - size = rcu_dereference(ht->t.size); - ret = _cds_lfht_del(ht, size, iter->node, 0); + size = rcu_dereference(ht->size); + ret = _cds_lfht_del(ht, size, node); if (!ret) { - hash = bit_reverse_ulong(iter->node->reverse_hash); + 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_bucket(struct cds_lfht *ht) { @@ -1569,36 +1741,30 @@ int cds_lfht_delete_bucket(struct cds_lfht *ht) unsigned long order, i, size; /* Check that the table is empty */ - node = &ht->t.tbl[0]->nodes[0]; + node = bucket_at(ht, 0); do { 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 bucket */ - for (order = 0; order < get_count_order_ulong(size) + 1; order++) { - unsigned long len; + 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)); + } - 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_bucket(ht->t.tbl[order]->nodes[i].next)); - } + for (order = cds_lfht_get_count_order_ulong(size); (long)order >= 0; order--) + cds_lfht_free_bucket_table(ht, order); - if (order == ht->min_alloc_order) - poison_free(ht->t.tbl[0]); - else if (order > ht->min_alloc_order) - poison_free(ht->t.tbl[order]); - /* Nothing to delete for order < ht->min_alloc_order */ - } return 0; } @@ -1608,31 +1774,46 @@ int cds_lfht_delete_bucket(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 */ + if (was_online) + ht->flavor->thread_online(); ret = cds_lfht_delete_bucket(ht); if (ret) return ret; free_split_items_count(ht); if (attr) *attr = ht->resize_attr; + ret = pthread_mutex_destroy(&ht->resize_mutex); + if (ret) + ret = -EBUSY; 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; - unsigned long nr_bucket = 0; + unsigned long nr_bucket = 0, nr_removed = 0; *approx_before = 0; if (ht->split_count) { @@ -1645,15 +1826,14 @@ void cds_lfht_count_nodes(struct cds_lfht *ht, } *count = 0; - *removed = 0; /* Count non-bucket nodes in the table */ - node = &ht->t.tbl[0]->nodes[0]; + node = bucket_at(ht, 0); do { next = rcu_dereference(node->next); if (is_removed(next)) { if (!is_bucket(next)) - (*removed)++; + (nr_removed)++; else (nr_bucket)++; } else if (!is_bucket(next)) @@ -1662,6 +1842,7 @@ void cds_lfht_count_nodes(struct cds_lfht *ht, (nr_bucket)++; node = clear_flag(next); } while (!is_end(node)); + dbg_printf("number of logically removed nodes: %lu\n", nr_removed); dbg_printf("number of bucket nodes: %lu\n", nr_bucket); *approx_after = 0; if (ht->split_count) { @@ -1681,8 +1862,8 @@ 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); @@ -1696,9 +1877,9 @@ void _do_cds_lfht_shrink(struct cds_lfht *ht, { unsigned long old_order, new_order; - new_size = max(new_size, ht->min_alloc_size); - old_order = get_count_order_ulong(old_size); - new_order = get_count_order_ulong(new_size); + new_size = max(new_size, MIN_TABLE_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); @@ -1721,42 +1902,60 @@ 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_grow(struct cds_lfht *ht, unsigned long new_size) { - return _uatomic_xchg_monotonic_increase(&ht->t.resize_target, new_size); + return _uatomic_xchg_monotonic_increase(&ht->resize_target, new_size); } static void resize_target_update_count(struct cds_lfht *ht, unsigned long count) { - count = max(count, ht->min_alloc_size); - uatomic_set(&ht->t.resize_target, count); + count = max(count, MIN_TABLE_SIZE); + 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 @@ -1766,11 +1965,11 @@ 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); @@ -1783,7 +1982,7 @@ void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht) /* Store resize_target before read resize_initiated */ cmm_smp_mb(); - if (!CMM_LOAD_SHARED(ht->t.resize_initiated)) { + 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)) { @@ -1791,9 +1990,14 @@ void __cds_lfht_resize_lazy_launch(struct cds_lfht *ht) 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); } } @@ -1802,6 +2006,7 @@ void cds_lfht_resize_lazy_grow(struct cds_lfht *ht, unsigned long size, int grow { 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; @@ -1819,7 +2024,8 @@ void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, { if (!(ht->flags & CDS_LFHT_AUTO_RESIZE)) return; - count = max(count, ht->min_alloc_size); + 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 */ @@ -1829,7 +2035,7 @@ void cds_lfht_resize_lazy_count(struct cds_lfht *ht, unsigned long size, for (;;) { unsigned long s; - s = uatomic_cmpxchg(&ht->t.resize_target, size, count); + s = uatomic_cmpxchg(&ht->resize_target, size, count); if (s == size) break; /* no resize needed */ if (s > size)