MQTT 开源代理mosquitto的网络层封装相当sucks
最近学习MQTT协议,选择了当前比较流行的MQTT Broker “mosquitto”,但是在阅读代码过程中发现其网络底层库封装的相当差劲。
对于MQTT协议的变长头长度的读取上,基本上采取每次一个byte的方式进行读取判断,对于系统调用read的高代价来讲,真的是相当的浪费,也难怪其不能作为高并发的服务器进行处理。
当然mosquitto需要优化的地方还很多:
1. 使用poll而不是使用epoll (可能是处于跨平台考虑,如果linux下可以使用epoll替换),同时的就是刚才提到的 byte 读取网络数据
2. 订阅树的管理上,对于大量的请求断开或者重练效率比较低
3. 空闲空间管理机制优化和数据包发送方式的修改
4. 内存管理上malloc new 没有使用mem pool机制,在大并发情况下,内存管理容易出现问题
5. 锁遍地飞,如果采用reactor_
但是从另一个方面讲,mosquitto作为开源的实现,思路上还是比较清晰,为mqtt服务器开发提供了比较完备的参考,这也就是它的价值所在了。
#ifdef WITH_BROKER
int _mosquitto_packet_read(struct mosquitto_db *db, struct mosquitto *mosq)
#else
int _mosquitto_packet_read(struct mosquitto *mosq)
#endif
{
uint8_t byte;
ssize_t read_length;
int rc = 0;
if(!mosq) return MOSQ_ERR_INVAL;
if(mosq->sock == INVALID_SOCKET) return MOSQ_ERR_NO_CONN;
if(mosq->state == mosq_cs_connect_pending){
return MOSQ_ERR_SUCCESS;
}
/* This gets called if pselect() indicates that there is network data
* available - ie. at least one byte. What we do depends on what data we
* already have.
* If we've not got a command, attempt to read one and save it. This should
* always work because it's only a single byte.
* Then try to read the remaining length. This may fail because it is may
* be more than one byte - will need to save data pending next read if it
* does fail.
* Then try to read the remaining payload, where 'payload' here means the
* combined variable header and actual payload. This is the most likely to
* fail due to longer length, so save current data and current position.
* After all data is read, send to _mosquitto_handle_packet() to deal with.
* Finally, free the memory and reset everything to starting conditions.
*/
if(!mosq->in_packet.command){
read_length = _mosquitto_net_read(mosq, &byte, 1);
if(read_length == 1){
mosq->in_packet.command = byte;
#ifdef WITH_BROKER
# ifdef WITH_SYS_TREE
g_bytes_received++;
# endif
/* Clients must send CONNECT as their first command. */
if(!(mosq->bridge) && mosq->state == mosq_cs_new && (byte&0xF0) != CONNECT) return MOSQ_ERR_PROTOCOL;
#endif
}else{
if(read_length == 0) return MOSQ_ERR_CONN_LOST; /* EOF */
#ifdef WIN32
errno = WSAGetLastError();
#endif
if(errno == EAGAIN || errno == COMPAT_EWOULDBLOCK){
return MOSQ_ERR_SUCCESS;
}else{
switch(errno){
case COMPAT_ECONNRESET:
return MOSQ_ERR_CONN_LOST;
default:
return MOSQ_ERR_ERRNO;
}
}
}
}
/* remaining_count is the number of bytes that the remaining_length
* parameter occupied in this incoming packet. We don't use it here as such
* (it is used when allocating an outgoing packet), but we must be able to
* determine whether all of the remaining_length parameter has been read.
* remaining_count has three states here:
* 0 means that we haven't read any remaining_length bytes
* <0 means we have read some remaining_length bytes but haven't finished
* >0 means we have finished reading the remaining_length bytes.
*/
if(mosq->in_packet.remaining_count <= 0){
do{
read_length = _mosquitto_net_read(mosq, &byte, 1);
if(read_length == 1){
mosq->in_packet.remaining_count--;
/* Max 4 bytes length for remaining length as defined by protocol.
* Anything more likely means a broken/malicious client.
*/
if(mosq->in_packet.remaining_count < -4) return MOSQ_ERR_PROTOCOL;
#if defined(WITH_BROKER) && defined(WITH_SYS_TREE)
g_bytes_received++;
#endif
mosq->in_packet.remaining_length += (byte & 127) * mosq->in_packet.remaining_mult;
mosq->in_packet.remaining_mult *= 128;
}else{
if(read_length == 0) return MOSQ_ERR_CONN_LOST; /* EOF */
#ifdef WIN32
errno = WSAGetLastError();
#endif
if(errno == EAGAIN || errno == COMPAT_EWOULDBLOCK){
return MOSQ_ERR_SUCCESS;
}else{
switch(errno){
case COMPAT_ECONNRESET:
return MOSQ_ERR_CONN_LOST;
default:
return MOSQ_ERR_ERRNO;
}
}
}
}while((byte & 128) != 0);
/* We have finished reading remaining_length, so make remaining_count
* positive. */
mosq->in_packet.remaining_count *= -1;
if(mosq->in_packet.remaining_length > 0){
mosq->in_packet.payload = _mosquitto_malloc(mosq->in_packet.remaining_length*sizeof(uint8_t));
if(!mosq->in_packet.payload) return MOSQ_ERR_NOMEM;
mosq->in_packet.to_process = mosq->in_packet.remaining_length;
}
}
while(mosq->in_packet.to_process>0){
read_length = _mosquitto_net_read(mosq, &(mosq->in_packet.payload[mosq->in_packet.pos]), mosq->in_packet.to_process);
if(read_length > 0){
#if defined(WITH_BROKER) && defined(WITH_SYS_TREE)
g_bytes_received += read_length;
#endif
mosq->in_packet.to_process -= read_length;
mosq->in_packet.pos += read_length;
}else{
#ifdef WIN32
errno = WSAGetLastError();
#endif
if(errno == EAGAIN || errno == COMPAT_EWOULDBLOCK){
if(mosq->in_packet.to_process > 1000){
/* Update last_msg_in time if more than 1000 bytes left to
* receive. Helps when receiving large messages.
* This is an arbitrary limit, but with some consideration.
* If a client can't send 1000 bytes in a second it
* probably shouldn't be using a 1 second keep alive. */
pthread_mutex_lock(&mosq->msgtime_mutex);
mosq->last_msg_in = mosquitto_time();
pthread_mutex_unlock(&mosq->msgtime_mutex);
}
return MOSQ_ERR_SUCCESS;
}else{
switch(errno){
case COMPAT_ECONNRESET:
return MOSQ_ERR_CONN_LOST;
default:
return MOSQ_ERR_ERRNO;
}
}
}
}
/* All data for this packet is read. */
mosq->in_packet.pos = 0;
#ifdef WITH_BROKER
# ifdef WITH_SYS_TREE
g_msgs_received++;
if(((mosq->in_packet.command)&0xF5) == PUBLISH){
g_pub_msgs_received++;
}
# endif
rc = mqtt3_packet_handle(db, mosq);
#else
rc = _mosquitto_packet_handle(mosq);
#endif
/* Free data and reset values */
_mosquitto_packet_cleanup(&mosq->in_packet);
pthread_mutex_lock(&mosq->msgtime_mutex);
mosq->last_msg_in = mosquitto_time();
pthread_mutex_unlock(&mosq->msgtime_mutex);
return rc;
}