PHP与Memcached服务器交互的分布式实现源码分析

前段时间，因为一个项目的关系，研究了php通过调用memcache和memcached PECL扩展库的接口存储到分布式缓存服务器的机制，在此做我根据他们各自的源码进行分析，希望能对这方面感兴趣的人有些帮助。
本篇文章我会针对php和memcache扩展库的交互根据源码展开分析。
PHP调用memcache的接口通常会是如下过程：

<?php$mmc = new Memcache();$mmc->addServer('node1', 11211);$mmc->addServer('node2', 11211, MemcacheConfig::MEMCACHE_PERSISTENT, 2);$mmc->set('key', 'value');echo $mmc->get('key');$mmc->delete('key');

短短几行代码，一个缓存key的生命周期就已经完整层现。从Memcache的初始化，到addServer添加两个服务器节点，接着set一个key到服务器上，然后get到这个key输出，最后delete这个key。在这个生命周期里，Memcache在底层究竟做了哪些事情，保证了数据存储服务器的均匀分布，数据的完整性？
接下来，我会根据上述生命周期的顺序，循序渐进的分析（由于主题是分布式算法的分析，所以接下来不相干的代码我会略去，很多分析我会直接备注在源码上）。


1. Memcache的初始化
对应PHP的代码：

$mmc = new Memcache();

对应C的代码：// Memcache类对应的方法名已经实际在c中实现过程的函数名，在接下来的分析中会用到。忽略不会分析到的方法。

static zend_function_entry php_memcache_class_functions[] = {PHP_FALIAS(addserver, memcache_add_server, NULL)PHP_FALIAS(set, memcache_set, NULL)PHP_FALIAS(get, memcache_get, NULL)PHP_FALIAS(delete, memcache_delete, NULL)......};PHP_MINIT_FUNCTION(memcache){// 初始化Memcache类实体，给类定在php空间中的调用名称以及类所拥有的方法zend_class_entry memcache_class_entry;INIT_CLASS_ENTRY(memcache_class_entry, "Memcache", php_memcache_class_functions);memcache_class_entry_ptr = zend_register_internal_class(&memcache_class_entry TSRMLS_CC);......}

以上过程是在Module Initialization的环节已经做好，在new的过程中，并无其余处理。
2. 添加缓存服务器，使之成为分布式存储

对应PHP的代码：

$mmc->addServer('node1', 11211);$mmc->addServer('node2', 11211, MemcacheConfig::MEMCACHE_PERSISTENT, 2);

由上面的php_memcache_class_functions结构可以看出，addServer方法对应的是memcache_add_server函数，因此对应C的代码：

PHP_FUNCTION(memcache_add_server){zval **connection, *mmc_object = getThis(), *failure_callback = NULL;// 整个Memcache中最重要的一个结构mmc_pool_tmmc_pool_t *pool;// 当前新添服务器的结构变量mmc_t *mmc;......// 如果pool之前没有初始化过，则初始化if (zend_hash_find(Z_OBJPROP_P(mmc_object), "connection", sizeof("connection"), (void **) &connection) == FAILURE) {// 调用mmp_pool_new完成初始化pool = mmc_pool_new(TSRMLS_C);......}else {......}//将新增服务器添加到pool中mmc_pool_add(pool, mmc, weight);RETURN_TRUE;}

来看下mmc_pool_t结构的定义：

typedef struct mmc_pool {mmc_t **servers; // 所有服务器的状态int num_servers; // 服务器数量mmc_t **requests; // 根据get的array key请求顺序返回的服务器数组状态int compress_threshold; // 待存储的数据压缩的下限值double min_compress_savings; // 待存储的数据最小的压缩百分比zend_bool in_free; // 标记该pool是否被释放mmc_hash_t *hash; // hash策略容器void *hash_state; // hash函数} mmc_pool_t;

然后我们看下mmc_hash_t的结构，再接下去的分析中会用到：// 结构定义中包含了四种抽象函数，作为基本结构，用于定义子结构

typedef struct mmc_hash {mmc_hash_create_state create_state; // 创建hash策略状态，主要是接纳了hash函数算法mmc_hash_free_state free_state; // 释放hash策略状态mmc_hash_find_server find_server; // 根据key和分布式算法定位到某台服务器mmc_hash_add_server add_server; // 根据hash策略、算法以及权重值添加服务器资源} mmc_hash_t;

接着我们追踪memcache_add_server函数中的mmc_pool_new函数调用方法：

typedef struct mmc_hash {mmc_hash_create_state create_state; // 创建hash策略状态，主要是接纳了hash函数算法mmc_hash_free_state free_state; // 释放hash策略状态mmc_hash_find_server find_server; // 根据key和分布式算法定位到某台服务器mmc_hash_add_server add_server; // 根据hash策略、算法以及权重值添加服务器资源} mmc_hash_t;

现在初始化hash算法已经逐渐显露，继续追踪mmc_pool_init_hash函数:

static void mmc_pool_init_hash(mmc_pool_t *pool TSRMLS_DC) /* {{{ */{mmc_hash_function hash;// 初始化hash函数// 根据php.ini中的memcache.hash_strategy配置选择hash存储策略，默认为标准hash存储策略switch (MEMCACHE_G(hash_strategy)) {case MMC_CONSISTENT_HASH:pool->hash = &mmc_consistent_hash;// 采用持久化hash存储策略break;default:pool->hash = &mmc_standard_hash;// 采用标准hash存储策略}

// 根据php.ini中的memcache.hash_function配置选择hash函数，默认为crc32算法

switch (MEMCACHE_G(hash_function)) {case MMC_HASH_FNV1A:hash = &mmc_hash_fnv1a; // 采用fnv1a算法break;default:hash = &mmc_hash_crc32; // 采用crc32算法}// hash策略中根据选择的hash函数创建对应的状态pool->hash_state = pool->hash->create_state(hash);}

根据上面的两个switch可以知道，在create_state的时候，是有两种策略选择的可能性，接着传入的hash参数也存在两种可能性，这里我先分析标准hash存储策略，以及对应的两种hash算法，然后再分析持久化hash策略。
先看下mmc_consistent_hash结构：// 根据mmc_hash_t的定义包含了四种具体函数实现

mmc_hash_t mmc_standard_hash = {mmc_standard_create_state,mmc_standard_free_state,mmc_standard_find_server,mmc_standard_add_server};

由上可知，pool->hash->create_state的函数调用实际是对mmc_standard_create_state的函数调用，继续看mmc_standard_create_state函数代码的实现：

// hash策略状态typedef struct mmc_standard_state {int num_servers; // 服务器数量mmc_t **buckets; // 哈希桶，和权重值相关int num_buckets; // 哈系桶的数量mmc_hash_function hash; // hash算法} mmc_standard_state_t;void *mmc_standard_create_state(mmc_hash_function hash) /* {{{ */{// 初始化状态mmc_standard_state_t *state = emalloc(sizeof(mmc_standard_state_t));memset(state, 0, sizeof(mmc_standard_state_t));// 选择的hash函数赋给hash属性state->hash = hash;return state;}

crc的算法实现：

static unsigned int mmc_hash_crc32(const char *key, int key_len) /* CRC32 hash {{{ */{unsigned int crc = ~0;int z;for (z=0; z<key_len; z++) {CRC32(crc, key[z]);}return ~crc;}

有关CRC32再深入的实现可以参考Cyclic redundancy check

然后来看看fnv算法实现：

/* 32 bit magic FNV-1a prime and init */#define FNV_32_PRIME 0x01000193#define FNV_32_INIT 0x811c9dc5static unsigned int mmc_hash_fnv1a(const char *key, int key_len) /* FNV-1a hash {{{ */{unsigned int hval = FNV_32_INIT;int z;for (z=0; z<key_len; z++) {hval ^= (unsigned int)key[z];hval *= FNV_32_PRIME;}return hval;}

具体fnv算法的深入实现可以参考Fowler–Noll–Vo hash function

最后我们看看mmc_consistent_hash结构：

mmc_hash_t mmc_consistent_hash = {mmc_consistent_create_state,mmc_consistent_free_state,mmc_consistent_find_server,mmc_consistent_add_server};

一样是四个函数，看下对应的create_state中的mmc_consistent_create_state的实现：

/* number of precomputed buckets, should be power of 2 */#define MMC_CONSISTENT_BUCKETS 1024typedef struct mmc_consistent_point {mmc_t *server; // 服务器状态unsigned int point; // 对应的指针} mmc_consistent_point_t;typedef struct mmc_consistent_state {int num_servers; // 服务器数量mmc_consistent_point_t *points; // 持久化服务器指针int num_points; // 指针数量mmc_t *buckets[MMC_CONSISTENT_BUCKETS]; // 哈希桶int buckets_populated; //标记哈希桶是否计算过mmc_hash_function hash; // hash函数} mmc_consistent_state_t;void *mmc_consistent_create_state(mmc_hash_function hash) /* {{{ */{// 初始化statemmc_consistent_state_t *state = emalloc(sizeof(mmc_consistent_state_t));memset(state, 0, sizeof(mmc_consistent_state_t));// 将hash函数赋值给hash属性state->hash = hash;return state;}

至此，memcache_add_server中mmc_pool_new函数流程结束，接着来看mmc_pool_add函数：

void mmc_pool_add(mmc_pool_t *pool, mmc_t *mmc, unsigned int weight) /* {{{ */{/* add server and a preallocated request pointer */if (pool->num_servers) {pool->servers = erealloc(pool->servers, sizeof(mmc_t *) * (pool->num_servers + 1));pool->requests = erealloc(pool->requests, sizeof(mmc_t *) * (pool->num_servers + 1));}else {pool->servers = emalloc(sizeof(mmc_t *));pool->requests = emalloc(sizeof(mmc_t *));}pool->servers[pool->num_servers] = mmc;pool->num_servers++;// 根据pool状态，当前要添加的服务器状态和权重调用add_server函数pool->hash->add_server(pool->hash_state, mmc, weight);}

由上面的说明可知add_server在标准hash模式下对应mmc_standard_add_server函数：

void mmc_standard_add_server(void *s, mmc_t *mmc, unsigned int weight) /* {{{ */{mmc_standard_state_t *state = s;int i;// 哈希桶初始化或重新分配相应的权重数值对应的空间if (state->num_buckets) {state->buckets = erealloc(state->buckets, sizeof(mmc_t *) * (state->num_buckets + weight));}else {state->buckets = emalloc(sizeof(mmc_t *) * (weight));}// 在某个区间内为哈希桶赋予服务器状态for (i=0; i<weight; i++) {buckets[state->num_buckets + i] = mmc;}state->num_buckets += weight;state->num_servers++;}

在持久化hash模式下，对应的是mmc_consistent_add_server函数：

#define MMC_CONSISTENT_POINTS 160 /* points per server */void mmc_consistent_add_server(void *s, mmc_t *mmc, unsigned int weight) /* {{{ */{mmc_consistent_state_t *state = s;int i, key_len, points = weight * MMC_CONSISTENT_POINTS;/* buffer for "host:port-i\0" */char *key = emalloc(strlen(mmc->host) + MAX_LENGTH_OF_LONG * 2 + 3);/* add weight * MMC_CONSISTENT_POINTS number of points for this server */state->points = erealloc(state->points, sizeof(mmc_consistent_point_t) * (state->num_points + points));// 将区块内的server赋予当前服务器状态，point赋予hash函数处理后的值for (i=0; i<points; i++) {key_len = sprintf(key, "%s:%d-%d", mmc->host, mmc->port, i);state->points[state->num_points + i].server = mmc;state->points[state->num_points + i].point = state->hash(key, key_len);MMC_DEBUG(("mmc_consistent_add_server: key %s, point %lu", key, state->points[state->num_points + i].point));}state->num_points += points;state->num_servers++;// 新增加服务器后需重新计算buckets顺序state->buckets_populated = 0;efree(key);}

以上代码有持久化hash算法的赋值实现，具体深入的了解请看Consistent hashing和国内大侠charlee翻译的小日本的文章memcached全面剖析–PDF总结篇。
Consistent hashing 算法最大的特点是当你的缓存服务器数量变更的时候，它能够最大化的保留原有的缓存不变，而不需要重新分布原有缓存的服务器位置。
至此，整个memcache_add_server流程结束。
3. 向缓存服务器保存数据

对应PHP的代码：

$mmc->set('key', 'value');

由上面的分析可知，set方法对应的是memcache_set函数：

/* {{{ proto bool memcache_set( object memcache, string key, mixed var [, int flag [, int expire ] ] )Sets the value of an item. Item may exist or not */PHP_FUNCTION(memcache_set){// Memcache对象中的add，set和replace皆会走该函数php_mmc_store(INTERNAL_FUNCTION_PARAM_PASSTHRU, "set", sizeof("set") - 1);}

看php_mmc_store函数：

static void php_mmc_store(INTERNAL_FUNCTION_PARAMETERS, char *command, int command_len) /* {{{ */{mmc_pool_t *pool;......// 获得poolif (!mmc_get_pool(mmc_object, &pool TSRMLS_CC) || !pool->num_servers) {RETURN_FALSE;}// 对不同的存储的值类型进行不同的处理switch (Z_TYPE_P(value)) {// 字符串类型case IS_STRING:result = mmc_pool_store(pool, command, command_len, key_tmp, key_tmp_len, flags, expire, Z_STRVAL_P(value), Z_STRLEN_P(value) TSRMLS_CC);break;// 长整型，浮点型，布尔型case IS_LONG:case IS_DOUBLE:case IS_BOOL: {......result = mmc_pool_store(pool, command, command_len, key_tmp, key_tmp_len, flags, expire, Z_STRVAL(value_copy), Z_STRLEN(value_copy) TSRMLS_CC);zval_dtor(&value_copy);break;}// 默认为数组类型default: {......result = mmc_pool_store(pool, command, command_len, key_tmp, key_tmp_len, flags, expire, buf.c, buf.len TSRMLS_CC);}}......}

由上代码可以看出，存储数据主要是交由mmc_pool_store处理：

int mmc_pool_store(mmc_pool_t *pool, const char *command, int command_len, const char *key, int key_len, int flags, int expire, const char *value, int value_len TSRMLS_DC) /* {{{ */{/* 该省略过程处理数据压缩，处理待发送的请求数据 */......// 通过key确定待保存的服务器while (result < 0 && (mmc = mmc_pool_find(pool, key, key_len TSRMLS_CC)) != NULL) {// 向缓存服务器发送请求，保存数据if ((result = mmc_server_store(mmc, request, request_len TSRMLS_CC)) < 0) {mmc_server_failure(mmc TSRMLS_CC);}}if (key_copy != NULL) {efree(key_copy);}if (data != NULL) {efree(data);}efree(request);return result;}

接着我们看下mmc_pool_find是处理的

#define mmc_pool_find(pool, key, key_len) \pool->hash->find_server(pool->hash_state, key, key_len)

原来是再次多态调用了find_server函数，由之前的分析可以得知find_server在标准hash模式中的函数为mmc_standard_find_server，在持久化hash模式中的函数为mmc_consistent_find_server，一样先看

mmc_standard_find_servermmc_t *mmc_standard_find_server(void *s, const char *key, int key_len TSRMLS_DC) /* {{{ */{mmc_standard_state_t *state = s;mmc_t *mmc;if (state->num_servers > 1) {// 用设定的hash函数算法，找到对应的服务器unsigned int hash = mmc_hash(state, key, key_len), i;mmc = state->buckets[hash % state->num_buckets];// 如果获取到的服务器状态有问题，则重新hash遍历寻找到可用的缓存服务器为止 for (i=0; !mmc_open(mmc, 0, NULL, NULL TSRMLS_CC) && MEMCACHE_G(allow_failover) && i<MEMCACHE_G(max_failover_attempts); i++) {char *next_key = emalloc(key_len + MAX_LENGTH_OF_LONG + 1);int next_len = sprintf(next_key, "%d%s", i+1, key);MMC_DEBUG(("mmc_standard_find_server: failed to connect to server '%s:%d' status %d, trying next", mmc->host, mmc->port, mmc->status));hash += mmc_hash(state, next_key, next_len);mmc = state->buckets[hash % state->num_buckets];efree(next_key);}}else {mmc = state->buckets[0];mmc_open(mmc, 0, NULL, NULL TSRMLS_CC);}return mmc->status != MMC_STATUS_FAILED ? mmc : NULL;}

再看

mmc_consistent_find_servermmc_t *mmc_consistent_find_server(void *s, const char *key, int key_len TSRMLS_DC) /* {{{ */{mmc_consistent_state_t *state = s;mmc_t *mmc;if (state->num_servers > 1) {unsigned int i, hash = state->hash(key, key_len);// 如果哈希桶没有进行过排序，则进行圆环排序操作if (!state->buckets_populated) {mmc_consistent_populate_buckets(state);}mmc = state->buckets[hash % MMC_CONSISTENT_BUCKETS];// 如果获取到的服务器状态有问题，则重新hash遍历寻找到可用的缓存服务器为止 for (i=0; !mmc_open(mmc, 0, NULL, NULL TSRMLS_CC) && MEMCACHE_G(allow_failover) && i<MEMCACHE_G(max_failover_attempts); i++) {char *next_key = emalloc(key_len + MAX_LENGTH_OF_LONG + 1);int next_len = sprintf(next_key, "%s-%d", key, i);MMC_DEBUG(("mmc_consistent_find_server: failed to connect to server '%s:%d' status %d, trying next", mmc->host, mmc->port, mmc->status));hash = state->hash(next_key, next_len);mmc = state->buckets[hash % MMC_CONSISTENT_BUCKETS];efree(next_key);}}else {mmc = state->points[0].server;mmc_open(mmc, 0, NULL, NULL TSRMLS_CC);}return mmc->status != MMC_STATUS_FAILED ? mmc : NULL;}// 持久化哈希算法的核心部分static void mmc_consistent_populate_buckets(mmc_consistent_state_t *state) /* {{{ */{unsigned int z, step = 0xffffffff / MMC_CONSISTENT_BUCKETS;qsort((void *)state->points, state->num_points, sizeof(mmc_consistent_point_t), mmc_consistent_compare);for (z=0; z<MMC_CONSISTENT_BUCKETS; z++) {state->buckets[z] = mmc_consistent_find(state, step * z);}state->buckets_populated = 1;}static int mmc_consistent_compare(const void *a, const void *b) /* {{{ */{if (((mmc_consistent_point_t *)a)->point < ((mmc_consistent_point_t *)b)->point) {return -1;}if (((mmc_consistent_point_t *)a)->point > ((mmc_consistent_point_t *)b)->point) {return 1;}return 0;}static mmc_t *mmc_consistent_find(mmc_consistent_state_t *state, unsigned int point) /* {{{ */{int lo = 0, hi = state->num_points - 1, mid;while (1) {/* point is outside interval or lo >= hi, wrap-around */if (point <= state->points[lo].point || point > state->points[hi].point) {return state->points[lo].server;}/* test middle point */mid = lo + (hi - lo) / 2;MMC_DEBUG(("mmc_consistent_find: lo %d, hi %d, mid %d, point %u, midpoint %u", lo, hi, mid, point, state->points[mid].point));/* perfect match */if (point <= state->points[mid].point && point > (mid ? state->points[mid-1].point : 0)) {return state->points[mid].server;}/* too low, go up */if (state->points[mid].point < point) {lo = mid + 1;}else {hi = mid - 1;}}}

至此，memcache_set过程结束。

4. 向缓存服务器获得已保存的数据

对应PHP的代码：

echo $mmc->get('key');

由上面的分析可知，get方法对应的是memcache_get函数：

PHP_FUNCTION(memcache_get){......// 获得poolif (!mmc_get_pool(mmc_object, &pool TSRMLS_CC) || !pool->num_servers) {RETURN_FALSE;}// 当key不为数组的情况下处理if (Z_TYPE_P(zkey) != IS_ARRAY) {// 检查key的合法性if (mmc_prepare_key(zkey, key, &key_len TSRMLS_CC) == MMC_OK) {// 获取key获取valueif (mmc_exec_retrieval_cmd(pool, key, key_len, &return_value, flags TSRMLS_CC) < 0) {zval_dtor(return_value);RETVAL_FALSE;}}else {RETVAL_FALSE;}// 为数组的情况下处理} else if (zend_hash_num_elements(Z_ARRVAL_P(zkey))){//根据数据key获取数组值if (mmc_exec_retrieval_cmd_multi(pool, zkey, &return_value, flags TSRMLS_CC) < 0) {zval_dtor(return_value);RETVAL_FALSE;}} else {RETVAL_FALSE;}}

接着看mmc_exec_retrieval_cmd和mmc_exec_retrieval_cmd_multi函数：

int mmc_exec_retrieval_cmd(mmc_pool_t *pool, const char *key, int key_len, zval **return_value, zval *return_flags TSRMLS_DC) /* {{{ */{mmc_t *mmc;char *command, *value;int result = -1, command_len, response_len, value_len, flags = 0;MMC_DEBUG(("mmc_exec_retrieval_cmd: key '%s'", key));command_len = spprintf(&command, 0, "get %s", key);// 遍历寻找到key对应的value值while (result < 0 && (mmc = mmc_pool_find(pool, key, key_len TSRMLS_CC)) != NULL) {......}if (return_flags != NULL) {zval_dtor(return_flags);ZVAL_LONG(return_flags, flags);}efree(command);return result;}static int mmc_exec_retrieval_cmd_multi(mmc_pool_t *pool, zval *keys, zval **return_value, zval *return_flags TSRMLS_DC) /* {{{ */{......do {result_status = num_requests = 0;zend_hash_internal_pointer_reset_ex(Z_ARRVAL_P(keys), &pos);// 遍历key得到所有key对应的服务器资源存入pool->requests中while (zend_hash_get_current_data_ex(Z_ARRVAL_P(keys), (void **)&zkey, &pos) == SUCCESS) {if (mmc_prepare_key(*zkey, key, &key_len TSRMLS_CC) == MMC_OK) {/* schedule key if first round or if missing from result */if ((!i || !zend_hash_exists(Z_ARRVAL_PP(return_value), key, key_len)) &&// 根据key寻找到服务器(mmc = mmc_pool_find(pool, key, key_len TSRMLS_CC)) != NULL) {if (!(mmc->outbuf.len)) {smart_str_appendl(&(mmc->outbuf), "get", sizeof("get")-1);pool->requests[num_requests++] = mmc;}smart_str_appendl(&(mmc->outbuf), " ", 1);smart_str_appendl(&(mmc->outbuf), key, key_len);MMC_DEBUG(("mmc_exec_retrieval_cmd_multi: scheduled key '%s' for '%s:%d' request length '%d'", key, mmc->host, mmc->port, mmc->outbuf.len));}}zend_hash_move_forward_ex(Z_ARRVAL_P(keys), &pos);}......} while (result_status < 0 && MEMCACHE_G(allow_failover) && i++ < MEMCACHE_G(max_failover_attempts));......return result_status;}

由上可见分布式hash的核心函数皆为mmc_pool_find，首先找到key对应的服务器资源，然后根据服务器资源请求数据。
至此，memcache_get的过程结束。
5.向缓存服务器删除已保存的数据
对应的php代码：

$mmc->delete('key');

由之前的分析可知，delete对应的为

memcache_delete：/* {{{ proto bool memcache_delete( object memcache, string key [, int expire ])Deletes existing item */PHP_FUNCTION(memcache_delete){mmc_t *mmc;mmc_pool_t *pool;int result = -1, key_len;zval *mmc_object = getThis();char *key;long time = 0;char key_tmp[MMC_KEY_MAX_SIZE];unsigned int key_tmp_len;if (mmc_object == NULL) {if (zend_parse_parameters(ZEND_NUM_ARGS() TSRMLS_CC, "Os|l", &mmc_object, memcache_class_entry_ptr, &key, &key_len, &time) == FAILURE) {return;}}else {if (zend_parse_parameters(ZEND_NUM_ARGS() TSRMLS_CC, "s|l", &key, &key_len, &time) == FAILURE) {return;}}if (!mmc_get_pool(mmc_object, &pool TSRMLS_CC) || !pool->num_servers) {RETURN_FALSE;}if (mmc_prepare_key_ex(key, key_len, key_tmp, &key_tmp_len TSRMLS_CC) != MMC_OK) {RETURN_FALSE;}// 先获得服务器资源while (result < 0 && (mmc = mmc_pool_find(pool, key_tmp, key_tmp_len TSRMLS_CC)) != NULL) {// 根据资源向缓存服务器发送请求删除存储的数据 if ((result = mmc_delete(mmc, key_tmp, key_tmp_len, time TSRMLS_CC)) < 0) {mmc_server_failure(mmc TSRMLS_CC);}}if (result > 0) {RETURN_TRUE;}RETURN_FALSE;}/* }}} */

至此，memcache_delete过程结束。

来自：http://blog.liubijian.com/php_memcache_code_analysis.html