epoll 是为处理大量句柄而改进的poll，在UDT中也有支持。UDT使用了内核提供的epoll，主要是epoll_create，epoll_wait，epoll_ctl，UDT定义了CEPollDesc这个结构来管理epoll的描述符和套接字。
struct CEPollDesc
{
int m_iID; // epoll ID
std::set<UDTSOCKET> m_sUDTSocksOut; // set of UDT sockets waiting for write events
std::set<UDTSOCKET> m_sUDTSocksIn; // set of UDT sockets waiting for read events
std::set<UDTSOCKET> m_sUDTSocksEx; // set of UDT sockets waiting for exceptions
int m_iLocalID; // local system epoll ID
std::set<SYSSOCKET> m_sLocals; // set of local (non-UDT) descriptors
std::set<UDTSOCKET> m_sUDTWrites; // UDT sockets ready for write
std::set<UDTSOCKET> m_sUDTReads; // UDT sockets ready for read
std::set<UDTSOCKET> m_sUDTExcepts; // UDT sockets with exceptions (connection broken, etc.)
};
特别要提醒的是，当对端socket连接中断的时候，也是在m_sUDTReads里的
UDT还实现了一个类来进行各项操作，实现的有
create():创建一个epoll，调用了epoll_create
add_usock():添加一个UDT套接字到epoll
add_ssock():添加一个系统套接字到epoll，调用了epoll_ctl
remove_usock():从epoll中移除一个UDT套接字
remove_ssock():从epoll中移除一个系统套接字，调用了epoll_ctl
wait():等待epoll事件或者超时，调用了epoll_wait
release():关闭并释放一个epoll
UDT里对epoll的调用是四段式的，比如add_usock这里，调用顺序是epoll_add_usock()->CUDT::epoll_add_usock()->s_UDTUnited.epoll_add_usock()->CEPoll::add_usock
int epoll_add_usock(int eid, UDTSOCKET u, const int* events)
return CUDT::epoll_add_usock(eid, u, events);
}
int CUDT::epoll_add_usock(const int eid, const UDTSOCKET u, const int* events)
try
{
return s_UDTUnited.epoll_add_usock(eid, u, events);
}
catch (CUDTException e)
s_UDTUnited.setError(new CUDTException(e));
return ERROR;
catch (...)
s_UDTUnited.setError(new CUDTException(-1, 0, 0));
int CUDTUnited::epoll_add_usock(const int eid, const UDTSOCKET u, const int* events)
CUDTSocket* s = locate(u);
int ret = -1;
if (NULL != s)
ret = m_EPoll.add_usock(eid, u, events);
s->m_pUDT->addEPoll(eid);
else
throw CUDTException(5, 4);
return ret;
int CEPoll::add_usock(const int eid, const UDTSOCKET& u, const int* events)
CGuard pg(m_EPollLock);
map<int, CEPollDesc>::iterator p = m_mPolls.find(eid);
if (p == m_mPolls.end())
throw CUDTException(5, 13);
if (!events || (*events & UDT_EPOLL_IN)) //UDT_EPOLL_IN 和UDT_EPOLL_OUT、UDT_EPOLL_ERROR分别是0x1, 0x4, 0x8，可以进行&运算
p->second.m_sUDTSocksIn.insert(u);
if (!events || (*events & UDT_EPOLL_OUT))
p->second.m_sUDTSocksOut.insert(u);
return 0;
UDT命名空间提供给应用程序调用接口，可成为API层，API层调用CUDT API，这一层主要做错误处理，捕捉下面两层抛出的错误保存起来交给应用程序使用，CUDT API调用CUDTUnited的实现，如果是UDT SOCKET的socket()，bind()，listen()等，到这层也就结束了，不过epoll要多个第四层，再调用CEPoll的实现。现在来看看CUDTUnited和CEPoll的实现。
CUDTSocket* s = locate(u);
调用CUDTUnited::locate()，根据SocketID，也就是UDT Socket handle在CUDTUnited的std::map<UDTSOCKET, CUDTSocket*> m_Sockets中找到对应的CUDTSocket结构，如果找不到，抛出错误，找到了就调用CEPoll的add_usock实现，根据传的event参数，将socket导入相应的队列。之后调用s->m_pUDT->addEPoll(eid)
void CUDT::addEPoll(const int eid)
CGuard::enterCS(s_UDTUnited.m_EPoll.m_EPollLock); //这种通过类来实现加锁解锁的，我第一次见，还挺方便。
m_sPollID.insert(eid);
CGuard::leaveCS(s_UDTUnited.m_EPoll.m_EPollLock);
if (!m_bConnected || m_bBroken || m_bClosing)
return;
if (((UDT_STREAM == m_iSockType) && (m_pRcvBuffer->getRcvDataSize() > 0)) ||
((UDT_DGRAM == m_iSockType) && (m_pRcvBuffer->getRcvMsgNum() > 0)))
s_UDTUnited.m_EPoll.update_events(m_SocketID, m_sPollID, UDT_EPOLL_IN, true);
if (m_iSndBufSize > m_pSndBuffer->getCurrBufSize())
s_UDTUnited.m_EPoll.update_events(m_SocketID, m_sPollID, UDT_EPOLL_OUT, true);
这里已经开始更新m_sUDTWrites，m_sUDTReads，通过update_events()，如前所述，update_events()也是CEPoll的成员函数。
int CEPoll::update_events(const UDTSOCKET& uid, std::set<int>& eids, int events, bool enable)
map<int, CEPollDesc>::iterator p;
vector<int> lost;
for (set<int>::iterator i = eids.begin(); i != eids.end(); ++ i)
p = m_mPolls.find(*i);
if (p == m_mPolls.end())
{
lost.push_back(*i);
}
else
if ((events & UDT_EPOLL_IN) != 0)
update_epoll_sets(uid, p->second.m_sUDTSocksIn, p->second.m_sUDTReads, enable); //UDT_EPOLL_IN 和UDT_EPOLL_OUT、UDT_EPOLL_ERROR分别是0x1, 0x4, 0x8，可以进行&运算
if ((events & UDT_EPOLL_OUT) != 0)
update_epoll_sets(uid, p->second.m_sUDTSocksOut, p->second.m_sUDTWrites, enable);
if ((events & UDT_EPOLL_ERR) != 0)
update_epoll_sets(uid, p->second.m_sUDTSocksEx, p->second.m_sUDTExcepts, enable);
for (vector<int>::iterator i = lost.begin(); i != lost.end(); ++ i)
eids.erase(*i);
void update_epoll_sets(const UDTSOCKET& uid, const set<UDTSOCKET>& watch, set<UDTSOCKET>& result, bool enable)
if (enable && (watch.find(uid) != watch.end()))
result.insert(uid);
else if (!enable)
result.erase(uid);
最后说下wait函数的实现，一样是四段式，就跳过前面三段，直接看第四段
int CEPoll::wait(const int eid, set<UDTSOCKET>* readfds, set<UDTSOCKET>* writefds, int64_t msTimeOut, set<SYSSOCKET>* lrfds, set<SYSSOCKET>* lwfds)
// if all fields is NULL and waiting time is infinite, then this would be a deadlock 都空的的话，会死锁，抛出异常
if (!readfds && !writefds && !lrfds && lwfds && (msTimeOut < 0))
throw CUDTException(5, 3, 0);
// Clear these sets in case the app forget to do it. 清空
if (readfds) readfds->clear();
if (writefds) writefds->clear();
if (lrfds) lrfds->clear();
if (lwfds) lwfds->clear();
int total = 0;
int64_t entertime = CTimer::getTime();
while (true)
CGuard::enterCS(m_EPollLock);
map<int, CEPollDesc>::iterator p = m_mPolls.find(eid);
CGuard::leaveCS(m_EPollLock);
throw CUDTException(5, 13);
if (p->second.m_sUDTSocksIn.empty() && p->second.m_sUDTSocksOut.empty() && p->second.m_sLocals.empty() && (msTimeOut < 0))
// no socket is being monitored, this may be a deadlock 都空的的话，会死锁，抛出异常
throw CUDTException(5, 3);
// Sockets with exceptions are returned to both read and write sets. 把m_sUDTReads和m_sUDTExcepts都读到readfds里
if ((NULL != readfds) && (!p->second.m_sUDTReads.empty() || !p->second.m_sUDTExcepts.empty()))
*readfds = p->second.m_sUDTReads;
for (set<UDTSOCKET>::const_iterator i = p->second.m_sUDTExcepts.begin(); i != p->second.m_sUDTExcepts.end(); ++ i)
readfds->insert(*i);
total += p->second.m_sUDTReads.size() + p->second.m_sUDTExcepts.size();
if ((NULL != writefds) && (!p->second.m_sUDTWrites.empty() || !p->second.m_sUDTExcepts.empty())) //把m_sUDTWrites和m_sUDTExcepts都读到writefds里
*writefds = p->second.m_sUDTWrites;
writefds->insert(*i);
total += p->second.m_sUDTWrites.size() + p->second.m_sUDTExcepts.size();
if (lrfds || lwfds) //读系统套接字
#ifdef LINUX
const int max_events = p->second.m_sLocals.size();
epoll_event ev[max_events];
int nfds = ::epoll_wait(p->second.m_iLocalID, ev, max_events, 0);
for (int i = 0; i < nfds; ++ i)
{
if ((NULL != lrfds) && (ev[i].events & EPOLLIN))
{
lrfds->insert(ev[i].data.fd);
++ total;
}
if ((NULL != lwfds) && (ev[i].events & EPOLLOUT))
{
lwfds->insert(ev[i].data.fd);
}
#else
//currently "select" is used for all non-Linux platforms.
//faster approaches can be applied for specific systems in the future.
//"select" has a limitation on the number of sockets
fd_set readfds;
fd_set writefds;
FD_ZERO(&readfds);
FD_ZERO(&writefds);
for (set<SYSSOCKET>::const_iterator i = p->second.m_sLocals.begin(); i != p->second.m_sLocals.end(); ++ i)
if (lrfds)
FD_SET(*i, &readfds);
if (lwfds)
FD_SET(*i, &writefds);
timeval tv;
tv.tv_sec = 0;
tv.tv_usec = 0;
if (::select(0, &readfds, &writefds, NULL, &tv) > 0)
for (set<SYSSOCKET>::const_iterator i = p->second.m_sLocals.begin(); i != p->second.m_sLocals.end(); ++ i)
if (lrfds && FD_ISSET(*i, &readfds))
{
lrfds->insert(*i);
++ total;
}
if (lwfds && FD_ISSET(*i, &writefds))
lwfds->insert(*i);
#endif
CGuard::leaveCS(m_EPollLock);
if (total > 0)
return total;
if ((msTimeOut >= 0) && (int64_t(CTimer::getTime() - entertime) >= msTimeOut * 1000LL))
throw CUDTException(6, 3, 0);
CTimer::waitForEvent();
![linux-svn卸载与安装[图]](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACwAAAAAAQABAAACAkQBADs=)