背景

对于同一服务可能存在多次调用的情况,然而每次调用都需要建立一次tcp连接导致大量重复工作的同时还增加了连接超时或连接错误的概率,为了减少tcp连接次数最大限度的提高连接利用率,需要能够重复利用每个tcp连接。

原理

  • HTTP1.1与HTTP2.0支持对于一次TCP连接建立的通道重复使用。
  • HTTP2.0支持多路复用
  • CURL支持对HTTP1.1和HTTP2.0已建立连接的复用,如果旧连接已失效则主动关闭旧连接,如果连接有效则尝试使用已有连接传输数据。关键代码如下:
// php/ext/url/interface.c
/* {{{ proto bool curl_exec(resource ch)
   Perform a cURL session */
PHP_FUNCTION(curl_exec)
{
    CURLcode    error;
    zval        *zid;
    php_curl    *ch;

    if (zend_parse_parameters(ZEND_NUM_ARGS() TSRMLS_CC, "r", &zid) == FAILURE) {
        return;
    }

    ZEND_FETCH_RESOURCE(ch, php_curl *, &zid, -1, le_curl_name, le_curl);

    _php_curl_verify_handlers(ch, 1 TSRMLS_CC);

    _php_curl_cleanup_handle(ch);

   // 调用CURL方法
    error = curl_easy_perform(ch->cp);
    SAVE_CURL_ERROR(ch, error);
    /* CURLE_PARTIAL_FILE is returned by HEAD requests */
    if (error != CURLE_OK && error != CURLE_PARTIAL_FILE) {
        if (ch->handlers->write->buf.len > 0) {
            smart_str_free(&ch->handlers->write->buf);
        }
        RETURN_FALSE;
    }

    if (ch->handlers->std_err) {
        php_stream  *stream;
        stream = (php_stream*)zend_fetch_resource(&ch->handlers->std_err TSRMLS_CC, -1, NULL, NULL, 2, php_file_le_stream(), php_file_le_pstream());
        if (stream) {
            php_stream_flush(stream);
        }
    }

    if (ch->handlers->write->method == PHP_CURL_RETURN && ch->handlers->write->buf.len > 0) {
        smart_str_0(&ch->handlers->write->buf);
        RETURN_STRINGL(ch->handlers->write->buf.c, ch->handlers->write->buf.len, 1);
    }

    /* flush the file handle, so any remaining data is synched to disk */
    if (ch->handlers->write->method == PHP_CURL_FILE && ch->handlers->write->fp) {
        fflush(ch->handlers->write->fp);
    }
    if (ch->handlers->write_header->method == PHP_CURL_FILE && ch->handlers->write_header->fp) {
        fflush(ch->handlers->write_header->fp);
    }

    if (ch->handlers->write->method == PHP_CURL_RETURN) {
        RETURN_EMPTY_STRING();
    } else {
        RETURN_TRUE;
    }
}
/* }}} */


// curl/lib/url.c line 4328

  // 主动关闭已失效的连接
  prune_dead_connections(data);

  /*************************************************************
   * Check the current list of connections to see if we can
   * re-use an already existing one or if we have to create a
   * new one.
   *************************************************************/

  /* reuse_fresh is TRUE if we are told to use a new connection by force, but
     we only acknowledge this option if this is not a re-used connection
     already (which happens due to follow-location or during a HTTP
     authentication phase). */
  if(data->set.reuse_fresh && !data->state.this_is_a_follow)
    reuse = FALSE;
  else
    // 从已存在的链接中查找出可以复用的连接(如果是不支持多路复用且正在使用中的连接会被忽略)
    reuse = ConnectionExists(data, conn, &conn_temp, &force_reuse, &waitpipe);

  /* If we found a reusable connection, we may still want to
     open a new connection if we are pipelining. */
  if(reuse && !force_reuse && IsPipeliningPossible(data, conn_temp)) {
    size_t pipelen = conn_temp->send_pipe.size + conn_temp->recv_pipe.size;
    if(pipelen > 0) {
      infof(data, "Found connection %ld, with requests in the pipe (%zu)\n",
            conn_temp->connection_id, pipelen);

      if(conn_temp->bundle->num_connections < max_host_connections &&
         data->state.conn_cache->num_connections < max_total_connections) {
        /* We want a new connection anyway */
        reuse = FALSE;

        infof(data, "We can reuse, but we want a new connection anyway\n");
      }
    }
  }

  if(reuse) {
    /*
     * We already have a connection for this, we got the former connection
     * in the conn_temp variable and thus we need to cleanup the one we
     * just allocated before we can move along and use the previously
     * existing one.
     */
    conn_temp->inuse = TRUE; /* mark this as being in use so that no other
                                handle in a multi stack may nick it */
    reuse_conn(conn, conn_temp);
    free(conn);          /* we don't need this anymore */
    conn = conn_temp;
    *in_connect = conn;

    infof(data, "Re-using existing connection! (#%ld) with %s %s\n",
          conn->connection_id,
          conn->bits.proxy?"proxy":"host",
          conn->socks_proxy.host.name ? conn->socks_proxy.host.dispname :
          conn->http_proxy.host.name ? conn->http_proxy.host.dispname :
                                       conn->host.dispname);
  }
  else {
    /* We have decided that we want a new connection. However, we may not
       be able to do that if we have reached the limit of how many
       connections we are allowed to open. */
    struct connectbundle *bundle = NULL;

    if(conn->handler->flags & PROTOPT_ALPN_NPN) {
      /* The protocol wants it, so set the bits if enabled in the easy handle
         (default) */
      if(data->set.ssl_enable_alpn)
        conn->bits.tls_enable_alpn = TRUE;
      if(data->set.ssl_enable_npn)
        conn->bits.tls_enable_npn = TRUE;
    }

    if(waitpipe)
      /* There is a connection that *might* become usable for pipelining
         "soon", and we wait for that */
      connections_available = FALSE;
    else
      bundle = Curl_conncache_find_bundle(conn, data->state.conn_cache);

    if(max_host_connections > 0 && bundle &&
       (bundle->num_connections >= max_host_connections)) {
      struct connectdata *conn_candidate;

      /* The bundle is full. Let's see if we can kill a connection. */
      conn_candidate = find_oldest_idle_connection_in_bundle(data, bundle);

      if(conn_candidate) {
        /* Set the connection's owner correctly, then kill it */
        conn_candidate->data = data;
        (void)Curl_disconnect(conn_candidate, /* dead_connection */ FALSE);
      }
      else {
        infof(data, "No more connections allowed to host: %d\n",
              max_host_connections);
        connections_available = FALSE;
      }
    }

    if(connections_available &&
       (max_total_connections > 0) &&
       (data->state.conn_cache->num_connections >= max_total_connections)) {
      struct connectdata *conn_candidate;

      /* The cache is full. Let's see if we can kill a connection. */
      conn_candidate = Curl_conncache_oldest_idle(data);

      if(conn_candidate) {
        /* Set the connection's owner correctly, then kill it */
        conn_candidate->data = data;
        (void)Curl_disconnect(conn_candidate, /* dead_connection */ FALSE);
      }
      else {
        infof(data, "No connections available in cache\n");
        connections_available = FALSE;
      }
    }

    if(!connections_available) {
      infof(data, "No connections available.\n");

      conn_free(conn);
      *in_connect = NULL;

      result = CURLE_NO_CONNECTION_AVAILABLE;
      goto out;
    }
    else {
      /*
       * This is a brand new connection, so let's store it in the connection
       * cache of ours!
       */
      Curl_conncache_add_conn(data->state.conn_cache, conn);
    }

#if defined(USE_NTLM)
    /* If NTLM is requested in a part of this connection, make sure we don't
       assume the state is fine as this is a fresh connection and NTLM is
       connection based. */
    if((data->state.authhost.picked & (CURLAUTH_NTLM | CURLAUTH_NTLM_WB)) &&
       data->state.authhost.done) {
      infof(data, "NTLM picked AND auth done set, clear picked!\n");
      data->state.authhost.picked = CURLAUTH_NONE;
      data->state.authhost.done = FALSE;
    }

    if((data->state.authproxy.picked & (CURLAUTH_NTLM | CURLAUTH_NTLM_WB)) &&
       data->state.authproxy.done) {
      infof(data, "NTLM-proxy picked AND auth done set, clear picked!\n");
      data->state.authproxy.picked = CURLAUTH_NONE;
      data->state.authproxy.done = FALSE;
    }
#endif
  }

// curl/lib/multi.c
/*
 * This function scans the connection cache for half-open/dead connections,
 * closes and removes them.
 * The cleanup is done at most once per second.
 */
static void prune_dead_connections(struct Curl_easy *data)
{
  struct curltime now = Curl_now();
  time_t elapsed = Curl_timediff(now, data->state.conn_cache->last_cleanup);

  if(elapsed >= 1000L) {
    Curl_conncache_foreach(data, data->state.conn_cache, data,
                           call_disconnect_if_dead);
    data->state.conn_cache->last_cleanup = now;
  }
}

PHP实现

class Curl
{
    protected $ch = null;
    protected $errorCode = 0;
    protected $errorMsg = '';
    protected $curlInfo = array();
    protected $verbose = null;
    private static $instance = null;

    public function getLastErrorCode()
    {
        return $this->errorCode;
    }

    public function getLastErrorMsg()
    {
        return $this->errorMsg;
    }

    public function getLastCurlInfo()
    {
        return $this->curlInfo;
    }

    private function __construct()
    {
        $this->ch = curl_init();
    }

    /*
     * 单例模式防止被clone
     */
    private function __clone(){
        throw new CurlException('The Curl library can\'t be cloned');
    }

    /*
     * 使用单例模式调用
     */
    public static function getInstance(){
        if(!self::$instance instanceof self){
            self::$instance = new self();
        }
        return self::$instance;
    }

    /**
     * curl以get的方式访问
     * @param $url
     * @param int $timeout
     * @param array $params get请求的参数,可以在url中直接带参数,也可以在这里传
     * @param array $headers 支持['Accept' => 'application/json']和['Accept: application/json']两种方式
     * @return mixed
     */
    public function get($url, $timeout = 3, $params = [], $headers = [])
    {
        $url = $this->buildQuery($url,$params);
        $this->setGeneralOption($url,$timeout,$headers);
        $result = $this->execute();
        return $result;
    }

    /**
     * curl以post的方式访问
     * @param $url
     * @param array $params
     * @param array $headers 支持['Accept' => 'application/json']和['Accept: application/json']两种方式
     * @param bool $withHttpBuildQuery
     * @param int $timeout
     * @return mixed
     */
    public function post($url, $params = [], $headers = [], $withHttpBuildQuery = true, $timeout=3)
    {
        if ($withHttpBuildQuery) {
            if (!is_array($params)) {
                $params = [$params];
            }
            $params = http_build_query($params);
        }
        
        curl_setopt($this->ch, CURLOPT_POST, 1);
        curl_setopt($this->ch, CURLOPT_POSTFIELDS, $params);
        $this->setGeneralOption($url,$timeout,$headers);
        $result = $this->execute();
        return $result;
    }

    /**
     * curl以HTTP2.0 get的方式访问
     * @param string $url 请求URL
     * @param int $timeout 超时时间,单位秒
     * @param array $params get请求的参数,可以在url中直接带参数,也可以在这里传
     * @param array $headers 支持['Accept' => 'application/json']和['Accept: application/json']两种方式
     * @return mixed
     */
    public function get2($url, $timeout = 3, $params = [], $headers = [])
    {
        $url = $this->buildQuery($url,$params);
        $this->setGeneralOption($url,$timeout,$headers,CURL_HTTP_VERSION_2_0);
        $result = $this->execute();
        return $result;
    }

    /**
     * curl以HTTP2.0 post的方式访问
     * @param string $url 请求URL
     * @param array $params
     * @param array $headers 支持['Accept' => 'application/json']和['Accept: application/json']两种方式
     * @param bool $withHttpBuildQuery
     * @param int $timeout 超时时间,单位秒
     * @return mixed
     */
    public function post2($url, $params = [], $headers = [], $withHttpBuildQuery = true, $timeout=3)
    {
        if ($withHttpBuildQuery) {
            if (!is_array($params)) {
                $params = [$params];
            }
            $params = http_build_query($params);
        }

        curl_setopt($this->ch, CURLOPT_POST, 1);
        curl_setopt($this->ch, CURLOPT_POSTFIELDS, $params);
        $this->setGeneralOption($url,$timeout,$headers,CURL_HTTP_VERSION_2_0);
        $result = $this->execute();
        return $result;
    }

    /**
     * 实例销毁前主动关闭所有连接
     */
    public function __destruct()
    {
        $this->close();
    }

    /**
     * 关闭所有连接
     * Description: 这一步在php-fpm中可以省略,实例结束后php-fpm的垃圾回收机制会关闭
     */
    public function close()
    {
        if (is_resource($this->ch)) {
            curl_close($this->ch);
            $this->ch = null;
        }
    }

    /**
     * 拼接请求URL
     * @param string $url 请求URL
     * @param array $params 待拼接参数
     * @return string
     */
    protected function buildQuery($url,$params)
    {
        if (!$params) {
            return $url;
        }

        if (strpos($url, '?') === false) {
            $url .= '?';
        } else {
            $url .= '&';
        }

        $url .= http_build_query($params);
        return $url;
    }

    /**
     * 设置通用curl配置
     * @param string $url 请求URL
     * @param int $timeout 超时时间,单位秒
     * @param array $headers 请求header
     * @param int $httpVersion 使用的http协议,默认为1.1
     */
    protected function setGeneralOption($url,$timeout,$headers=array(),$httpVersion=CURL_HTTP_VERSION_1_1)
    {
        curl_setopt($this->ch, CURLOPT_URL, $url);
        curl_setopt($this->ch, CURLOPT_SSL_VERIFYPEER, true); //让CURL支持HTTPS访问
        curl_setopt($this->ch, CURLOPT_SSL_VERIFYHOST, 2);
        curl_setopt($this->ch, CURLOPT_RETURNTRANSFER, 1);
        curl_setopt($this->ch, CURLOPT_TIMEOUT, $timeout);
        curl_setopt($this->ch, CURLOPT_HTTP_VERSION, $httpVersion);
        // 启用debug获取更详细的连接信息,与CURLOPT_HEADER互斥
        curl_setopt($this->ch, CURLOPT_VERBOSE, 1);
        $this->verbose = fopen('php://temp', 'w+');
        curl_setopt($this->ch, CURLOPT_STDERR, $this->verbose);

        if ($headers && is_array($headers)) {
            $realHeader = [];
            foreach ($headers as $key => $val) {
                if (is_string($key)) {
                    $realHeader[] = $key. ': '. $val;
                } else {
                    $realHeader[] = $val;
                }
            }
            curl_setopt($this->ch, CURLOPT_HTTPHEADER, $realHeader);
        }
    }

    /**
     * 执行请求
     * @return mixed
     */
    protected function execute()
    {
        $result = curl_exec($this->ch);

        // 记录详细的debug信息
        $this->curlInfo = curl_getinfo($this->ch);
        rewind($this->verbose);
        $this->curlInfo['verbose'] = stream_get_contents($this->verbose);
        $this->verbose = null;

        if ($result === false) {
            $this->errorCode = curl_errno($this->ch);
            $this->errorMsg = curl_error($this->ch);
            $this->curlInfo['error_code'] = $this->errorCode;
            $this->curlInfo['error_message'] = $this->errorMsg;
        }

        curl_reset($this->ch);
        return $result;
    }
}

class CurlException extends \Exception {}

拓展

  • 由于PHP-FPM的回收机制,一次请求结束后CURL的资源将会被回收,这意味着这次请求建立的TCP连接将会被关闭,在这种情况下就无法达到垮请求复用的目的。因此可以利用独立进程的方式来维护已建立的TCP连接专门负责CURL的请求。
  • 对于HTTP2.0而言,由于支持多路复用,因此对于一个域名的请求建立一次tcp连接后可以支持同时多个请求的处理(HTTP1.1一个tcp连接同时只支持一个请求,如果第二个请求同时到达则CURL将建立新的tcp连接以便完成请求),利用这一特性使用独立进程配合协程可以达到对于单一场景的curl高并发的支撑。
  • 同理除PHP外可扩展到其他语言。