/* Note: this isn't an assoc_update. The key must not already exist to call this */
int assoc_insert(struct default_engine *engine, uint32_t hash, hash_item *it)
{
unsigned int oldbucket;
assert(assoc_find(engine, hash, item_get_key(it), it->nkey) == 0); /* shouldn't have duplicately named things defined */
// inserting actual hash_item to appropriate assoc_t
if (engine->assoc.expanding &&
(oldbucket = (hash & hashmask(engine->assoc.hashpower - 1))) >= engine->assoc.expand_bucket)
{
it->h_next = engine->assoc.old_hashtable[oldbucket];
engine->assoc.old_hashtable[oldbucket] = it;
} else {
it->h_next = engine->assoc.primary_hashtable[hash & hashmask(engine->assoc.hashpower)];
engine->assoc.primary_hashtable[hash & hashmask(engine->assoc.hashpower)] = it;
}
engine->assoc.hash_items++;
if (! engine->assoc.expanding && engine->assoc.hash_items > (hashsize(engine->assoc.hashpower) * 3) / 2) {
assoc_expand(engine);
}
MEMCACHED_ASSOC_INSERT(item_get_key(it), it->nkey, engine->assoc.hash_items);
return 1;
}
/* Note: this isn't an assoc_update. The key must not already exist to call this */
int assoc_insert(struct default_engine *engine, uint32_t hash, hash_item *it)
{
struct assoc *assoc = &engine->assoc;
uint32_t bucket = GET_HASH_BUCKET(hash, assoc->hashmask);
uint32_t tabidx;
assert(assoc_find(engine, hash, item_get_key(it), it->nkey) == 0); /* shouldn't have duplicately named things defined */
if (assoc->infotable[bucket].curpower != assoc->rootpower &&
assoc->infotable[bucket].refcount == 0) {
redistribute(engine, bucket);
}
tabidx = GET_HASH_TABIDX(hash, assoc->hashpower,
hashmask(assoc->infotable[bucket].curpower));
// inserting actual hash_item to appropriate assoc_t
it->h_next = assoc->roottable[tabidx].hashtable[bucket];
assoc->roottable[tabidx].hashtable[bucket] = it;
assoc->hash_items++;
if (assoc->hash_items > (hashsize(assoc->hashpower + assoc->rootpower) * 3) / 2) {
assoc_expand(engine);
}
MEMCACHED_ASSOC_INSERT(item_get_key(it), it->nkey, assoc->hash_items);
return 1;
}
/* Note: this isn't an assoc_update. The key must not already exist to call this */
int assoc_insert(item *it) {
uint32_t hv;
unsigned int oldbucket;
assert(assoc_find(ITEM_key(it), it->nkey) == 0); /* shouldn't have duplicately named things defined */
hv = hash(ITEM_key(it), it->nkey, 0);
if (expanding &&
(oldbucket = (hv & hashmask(hashpower - 1))) >= expand_bucket)
{
it->h_next = old_hashtable[oldbucket];
old_hashtable[oldbucket] = it;
} else {
it->h_next = primary_hashtable[hv & hashmask(hashpower)];
primary_hashtable[hv & hashmask(hashpower)] = it;
}
hash_items++;
if (! expanding && hash_items > (hashsize(hashpower) * 3) / 2) {
assoc_expand();
}
MEMCACHED_ASSOC_INSERT(ITEM_key(it), it->nkey, hash_items);
return 1;
}
static void *assoc_maintenance_thread(void *arg) {
while (do_run_maintenance_thread) {
int ii = 0;
/* Lock the cache, and bulk move multiple buckets to the new
* hash table. */
item_lock_global();
mutex_lock(&cache_lock);
for (ii = 0; ii < hash_bulk_move && expanding; ++ii) {
item *it, *next;
int bucket;
for (it = old_hashtable[expand_bucket]; NULL != it; it = next) {
next = it->h_next;
bucket = hash(ITEM_key(it), it->nkey, 0) & hashmask(hashpower);
it->h_next = primary_hashtable[bucket];
primary_hashtable[bucket] = it;
}
old_hashtable[expand_bucket] = NULL;
expand_bucket++;
if (expand_bucket == hashsize(hashpower - 1)) {
expanding = false;
free(old_hashtable);
STATS_LOCK();
stats.hash_bytes -= hashsize(hashpower - 1) * sizeof(void *);
stats.hash_is_expanding = 0;
STATS_UNLOCK();
if (settings.verbose > 1)
fprintf(stderr, "Hash table expansion done\n");
}
}
mutex_unlock(&cache_lock);
item_unlock_global();
if (!expanding) {
/* finished expanding. tell all threads to use fine-grained locks */
switch_item_lock_type(ITEM_LOCK_GRANULAR);
slabs_rebalancer_resume();
/* We are done expanding.. just wait for next invocation */
mutex_lock(&cache_lock);
started_expanding = false;
pthread_cond_wait(&maintenance_cond, &cache_lock);
/* Before doing anything, tell threads to use a global lock */
mutex_unlock(&cache_lock);
slabs_rebalancer_pause();
switch_item_lock_type(ITEM_LOCK_GLOBAL);
mutex_lock(&cache_lock);
assoc_expand();
mutex_unlock(&cache_lock);
}
}
return NULL;
}
/* Note: this isn't an assoc_update. The key must not already exist to call this */
int assoc_insert(struct persistent_engine *engine, uint32_t hash, hash_item *it) {
unsigned int oldbucket;
assert(assoc_find(engine, hash, item_get_key(&it->item), it->item.nkey) == 0); /* shouldn't have duplicately named things defined */
if (engine->assoc.expanding &&
(oldbucket = (hash & hashmask(engine->assoc.hashpower - 1))) >= engine->assoc.expand_bucket)
{
it->h_next = engine->assoc.old_hashtable[oldbucket];
engine->assoc.old_hashtable[oldbucket] = it;
} else {
it->h_next = engine->assoc.primary_hashtable[hash & hashmask(engine->assoc.hashpower)];
engine->assoc.primary_hashtable[hash & hashmask(engine->assoc.hashpower)] = it;
}
engine->assoc.hash_items++;
if (! engine->assoc.expanding && engine->assoc.hash_items > (hashsize(engine->assoc.hashpower) * 3) / 2) {
assoc_expand(engine);
}
return 1;
}
static void *assoc_maintenance_thread(void *arg) {
mutex_lock(&maintenance_lock);
while (do_run_maintenance_thread) {
int ii = 0;
/* There is only one expansion thread, so no need to global lock. */
for (ii = 0; ii < hash_bulk_move && expanding; ++ii) {
item *it, *next;
int bucket;
void *item_lock = NULL;
/* bucket = hv & hashmask(hashpower) =>the bucket of hash table
* is the lowest N bits of the hv, and the bucket of item_locks is
* also the lowest M bits of hv, and N is greater than M.
* So we can process expanding with only one item_lock. cool! */
if ((item_lock = item_trylock(expand_bucket))) {
for (it = old_hashtable[expand_bucket]; NULL != it; it = next) {
next = it->h_next;
bucket = hash(ITEM_key(it), it->nkey) & hashmask(hashpower);
it->h_next = primary_hashtable[bucket];
primary_hashtable[bucket] = it;
}
old_hashtable[expand_bucket] = NULL;
expand_bucket++;
if (expand_bucket == hashsize(hashpower - 1)) {
expanding = false;
free(old_hashtable);
STATS_LOCK();
stats.hash_bytes -= hashsize(hashpower - 1) * sizeof(void *);
stats.hash_is_expanding = 0;
STATS_UNLOCK();
if (settings.verbose > 1)
fprintf(stderr, "Hash table expansion done\n");
}
} else {
usleep(10*1000);
}
if (item_lock) {
item_trylock_unlock(item_lock);
item_lock = NULL;
}
}
if (!expanding) {
/* We are done expanding.. just wait for next invocation */
started_expanding = false;
pthread_cond_wait(&maintenance_cond, &maintenance_lock);
/* assoc_expand() swaps out the hash table entirely, so we need
* all threads to not hold any references related to the hash
* table while this happens.
* This is instead of a more complex, possibly slower algorithm to
* allow dynamic hash table expansion without causing significant
* wait times.
*/
pause_threads(PAUSE_ALL_THREADS);
assoc_expand();
pause_threads(RESUME_ALL_THREADS);
}
}
return NULL;
}
//启动扩容线程，扩容线程在main函数中会启动，启动运行一遍之后会阻塞在条件变量maintenance_cond上面,
//当插入元素超过规定，唤醒条件变量，再次运行
static void *assoc_maintenance_thread(void *arg) {
//do_run_maintenance_thread是全局变量，初始值为1
//在stop_assoc_maintenance_thread函数中会被赋值0，用来终止迁移线程
while (do_run_maintenance_thread) {
int ii = 0;
/* Lock the cache, and bulk move multiple buckets to the new
* hash table. */
item_lock_global();
mutex_lock(&cache_lock);
//hash_bulk_move用来控制每次迁移，移动多少个桶的item。默认是一个.
//如果expanding为true才会进入循环体，所以迁移线程刚创建的时候，并不会进入循环体
for (ii = 0; ii < hash_bulk_move && expanding; ++ii) {
item *it, *next;
int bucket;
//遍历旧哈希表中由expand_bucket指明的桶,将该桶的所有item迁移到新扩容的哈希表中
for (it = old_hashtable[expand_bucket]; NULL != it; it = next) {
next = it->h_next;
bucket = hash(ITEM_key(it), it->nkey, 0) & hashmask(hashpower);
it->h_next = primary_hashtable[bucket];
primary_hashtable[bucket] = it;
}
old_hashtable[expand_bucket] = NULL;
//迁移完一个桶，接着把expand_bucket指向下一个待迁移的桶
expand_bucket++;
//数据迁移完毕
if (expand_bucket == hashsize(hashpower - 1)) {
expanding = false;
free(old_hashtable);
STATS_LOCK();
stats.hash_bytes -= hashsize(hashpower - 1) * sizeof(void *);
stats.hash_is_expanding = 0;
STATS_UNLOCK();
if (settings.verbose > 1)
fprintf(stderr, "Hash table expansion done\n");
}
}
//释放锁
mutex_unlock(&cache_lock);
item_unlock_global();
//不再迁移数据
if (!expanding) {
/* finished expanding. tell all threads to use fine-grained locks */
switch_item_lock_type(ITEM_LOCK_GRANULAR);
slabs_rebalancer_resume();
/* We are done expanding.. just wait for next invocation */
mutex_lock(&cache_lock);
started_expanding = false;
//挂起迁移线程，直到worker线程插入数据后发现item数量已经到了1.5倍哈希表大小，
//此时调用worker线程调用assoc_start_expand函数，该函数会调用pthread_cond_signal
//唤醒迁移线程
pthread_cond_wait(&maintenance_cond, &cache_lock);
/* Before doing anything, tell threads to use a global lock */
mutex_unlock(&cache_lock);
slabs_rebalancer_pause();
switch_item_lock_type(ITEM_LOCK_GLOBAL);
mutex_lock(&cache_lock);
assoc_expand();//???
mutex_unlock(&cache_lock);
}
}
return NULL;
}
/* 扩容线程函数，扩容策略如下:
* 扩容线程在main函数中创建，在assoc_insert后发现item数目大于哈希表容量1.5倍，唤醒扩容线程。
* 扩容线程先创建一个2倍容量的新哈希表，然后进行把数据从旧哈希表迁移到新哈希表。
* 迁移从旧表索引0开始，每次迁移一个桶(可以增加迁移粒度，但由于迁移需要加锁，可能导致work线程获取锁的等待时间增加)，
* 迁移完成后释放旧表。
*/
static void *assoc_maintenance_thread(void *arg) {
mutex_lock(&maintenance_lock);
while (do_run_maintenance_thread) {
int ii = 0;
/* There is only one expansion thread, so no need to global lock. */
//如果expanding为true才会进入循环体，所以迁移线程刚创建的时候，并不会进入循环体
for (ii = 0; ii < hash_bulk_move && expanding; ++ii) {
item *it, *next;
int bucket;
void *item_lock = NULL;
/* bucket = hv & hashmask(hashpower) =>the bucket of hash table
* is the lowest N bits of the hv, and the bucket of item_locks is
* also the lowest M bits of hv, and N is greater than M.
* So we can process expanding with only one item_lock. cool! */
//获取单个桶锁
if ((item_lock = item_trylock(expand_bucket))) {
//迁移一个桶中所有item
for (it = old_hashtable[expand_bucket]; NULL != it; it = next) {
next = it->h_next;
//重新计算哈希值
bucket = hash(ITEM_key(it), it->nkey) & hashmask(hashpower);
it->h_next = primary_hashtable[bucket];
primary_hashtable[bucket] = it;
}
old_hashtable[expand_bucket] = NULL;
expand_bucket++;
//迁移完成
if (expand_bucket == hashsize(hashpower - 1)) {
expanding = false; //将迁移标志设置0
free(old_hashtable); //释放旧表
STATS_LOCK();
stats.hash_bytes -= hashsize(hashpower - 1) * sizeof(void *);
stats.hash_is_expanding = 0;
STATS_UNLOCK();
if (settings.verbose > 1)
fprintf(stderr, "Hash table expansion done\n");
}
} else {
usleep(10*1000);
}
if (item_lock) {
item_trylock_unlock(item_lock);
item_lock = NULL;
}
}
if (!expanding) {
/* We are done expanding.. just wait for next invocation */
//不需要迁移，挂起迁移线程，直到worker线程插入数据后发现item数量已经到了1.5倍哈希表大小，
//此时调用worker线程调用assoc_start_expand函数，该函数会调用pthread_cond_signal
//唤醒迁移线程
started_expanding = false;
pthread_cond_wait(&maintenance_cond, &maintenance_lock);
/* assoc_expand() swaps out the hash table entirely, so we need
* all threads to not hold any references related to the hash
* table while this happens.
* This is instead of a more complex, possibly slower algorithm to
* allow dynamic hash table expansion without causing significant
* wait times.
*/
pause_threads(PAUSE_ALL_THREADS);
assoc_expand(); //申请更大的哈希表,并将expanding设置为true
pause_threads(RESUME_ALL_THREADS);
}
}
return NULL;
}
static void *assoc_maintenance_thread(void *arg) {
mutex_lock(&maintenance_lock);
//do_run_maintenance_thread是全局变量，初始值为1，在stop_assoc_maintenance_thread
//函数中会被赋值0，终止迁移线程
while (do_run_maintenance_thread) {
int ii = 0;
/* There is only one expansion thread, so no need to global lock. */
for (ii = 0; ii < hash_bulk_move && expanding; ++ii) {
item *it, *next;
int bucket;
void *item_lock = NULL;
/* bucket = hv & hashmask(hashpower) =>the bucket of hash table
* is the lowest N bits of the hv, and the bucket of item_locks is
* also the lowest M bits of hv, and N is greater than M.
* So we can process expanding with only one item_lock. cool! */
//hash_bulk_move用来控制每次迁移，移动多少个桶的item。默认是一个.
//如果expanding为true才会进入循环体，所以迁移线程刚创建的时候，并不会进入循环体
if ((item_lock = item_trylock(expand_bucket))) {
//在assoc_expand函数中expand_bucket会被赋值0
//遍历旧哈希表中由expand_bucket指明的桶,将该桶的所有item
//迁移到新哈希表中。
for (it = old_hashtable[expand_bucket]; NULL != it; it = next) {
next = it->h_next;
//重新计算新的哈希值,得到其在新哈希表的位置
bucket = hash(ITEM_key(it), it->nkey) & hashmask(hashpower);
//将这个item插入到新哈希表中
it->h_next = primary_hashtable[bucket];
primary_hashtable[bucket] = it;
}
//不需要清空旧桶。直接将冲突链的链头赋值为NULL即可
old_hashtable[expand_bucket] = NULL;
//迁移完一个桶，接着把expand_bucket指向下一个待迁移的桶
expand_bucket++;
if (expand_bucket == hashsize(hashpower - 1)) {//全部数据迁移完毕
expanding = false;//将扩展标志设置为false
free(old_hashtable);
STATS_LOCK();
stats.hash_bytes -= hashsize(hashpower - 1) * sizeof(void *);
stats.hash_is_expanding = 0;
STATS_UNLOCK();
if (settings.verbose > 1)
fprintf(stderr, "Hash table expansion done\n");
}
} else {
usleep(10*1000);
}
if (item_lock) {
//遍历完hash_bulk_move个桶的所有item后，就释放锁
item_trylock_unlock(item_lock);
item_lock = NULL;
}
}
if (!expanding) {//不需要迁移数据(了)
/* We are done expanding.. just wait for next invocation */
started_expanding = false;//重置
//挂起迁移线程，直到worker线程插入数据后发现item数量已经到了1.5倍哈希表大小，
//此时调用worker线程调用assoc_start_expand函数，该函数会调用pthread_cond_signal
//唤醒迁移线程
pthread_cond_wait(&maintenance_cond, &maintenance_lock);
/* assoc_expand() swaps out the hash table entirely, so we need
* all threads to not hold any references related to the hash
* table while this happens.
* This is instead of a more complex, possibly slower algorithm to
* allow dynamic hash table expansion without causing significant
* wait times.
*/
pause_threads(PAUSE_ALL_THREADS);
assoc_expand();//申请更大的哈希表,并将expanding设置为true
pause_threads(RESUME_ALL_THREADS);
}
}
return NULL;
}
//数据迁移线程回调函数
static void *assoc_maintenance_thread(void *arg) {
//do_run_maintenance_thread 是全局变量，初始值为1，在stop_assoc_mainternance_thread
//函数中会被赋值0，之中迁移线程
while (do_run_maintenance_thread) {
int ii = 0;
/* Lock the cache, and bulk move multiple buckets to the new
* hash table. */
//上锁
item_lock_global();//锁上全局级别的锁，全部的item都在全局锁的控制之下
//锁住哈希表里面的item。不然别的线程对哈希表进行增删操作时，会出现
//数据不一致的情况.在item.c的do_item_link和do_item_unlink可以看到
//其内部也会锁住cache_lock锁.
mutex_lock(&cache_lock);
//进行item迁移
for (ii = 0; ii < hash_bulk_move && expanding; ++ii) {
item *it, *next;
int bucket;
for (it = old_hashtable[expand_bucket]; NULL != it; it = next) {
next = it->h_next;
bucket = hash(ITEM_key(it), it->nkey) & hashmask(hashpower);
it->h_next = primary_hashtable[bucket];
primary_hashtable[bucket] = it;
}
old_hashtable[expand_bucket] = NULL;
expand_bucket++;
if (expand_bucket == hashsize(hashpower - 1)) {
expanding = false;
free(old_hashtable);
STATS_LOCK();
stats.hash_bytes -= hashsize(hashpower - 1) * sizeof(void *);
stats.hash_is_expanding = 0;
STATS_UNLOCK();
if (settings.verbose > 1)
fprintf(stderr, "Hash table expansion done\n");
}
}
//遍历完就释放锁
mutex_unlock(&cache_lock);
item_unlock_global();
//不需要迁移数据了
if (!expanding) {
/*
迁移线程为什么要这么迂回曲折地切换workers线程的锁类型呢？直接修改所有线程的LIBEVENT_THREAD结构的item_lock_type
成员变量不就行了吗？
这主要是因为迁移线程不知道worker线程此刻在干些什么。如果worker线程正在访问item，并抢占了段级别锁。此时你把worker
线程的锁切换到全局锁，等worker线程解锁的时候就会解全局锁(参考前面的item_lock和item_unlock代码)，这样程序就崩溃了。
所以不能迁移线程去切换，只能迁移线程通知worker线程，然后worker线程自己去切换。当然是要worker线程忙完了手头上的事情
后，才会去修改切换的。所以迁移线程在通知完所有的worker线程后，会调用wait_for_thread_registration函数休眠等待所有的
worker线程都切换到指定的锁类型后才醒来。
*/
/* finished expanding. tell all threads to use fine-grained locks */
//进入到这里，说明已经不需要迁移数据(停止扩展了)。
//告诉所有的workers线程，访问item时，切换到段级别的锁。
//会阻塞到所有workers线程都切换到段级别的锁
switch_item_lock_type(ITEM_LOCK_GRANULAR);
slabs_rebalancer_resume();
/* We are done expanding.. just wait for next invocation */
mutex_lock(&cache_lock);
// 重置
started_expanding = false;
//挂起迁移线程，直到worker线程插入数据后发现item数量已经到了1.5被哈希表大小，
//此时调用worker线程调用assoc_start_expand函数，该函数会调用pthread_cond_signal唤醒迁移线程
pthread_cond_wait(&maintenance_cond, &cache_lock);
/* Before doing anything, tell threads to use a global lock */
mutex_unlock(&cache_lock);
slabs_rebalancer_pause();
//从maintenance_cond条件变量中醒来，说明又要开始扩展哈希表和迁移数据了。
//迁移线程在迁移一个桶的数据时是锁上全局级别的锁.
//此时workers线程不能使用段级别的锁，而是要使用全局级别的锁，
//所有的workers线程和迁移线程一起，争抢全局级别的锁.
//哪个线程抢到了，才有权利访问item.
//下面一行代码就是通知所有的workers线程，把你们访问item的锁切换
//到全局级别的锁。switch_item_lock_type会通过条件变量休眠等待，
//直到，所有的workers线程都切换到全局级别的锁，才会醒来过
switch_item_lock_type(ITEM_LOCK_GLOBAL);
mutex_lock(&cache_lock);
//申请更大的哈希表，并将expanding设置为true
assoc_expand();
mutex_unlock(&cache_lock);
}
}
return NULL;
}
