evpp ioevent benchmark closes only one endpoint of each socketpair

[OneBucket126]

Summary

Each FdChannel owns and closes g_pipes[2*i], but the peer endpoint g_pipes[2*i+1], used for writes, has no owner and is never closed.

Affected code

• Repository commit: 477033f938fd47dfecde43c82257cd286d9fa38e
• File: benchmark/ioevent/evpp/evpp_ioevent_bench.cc
• Original scanner location: line 106

Evidence

Source: benchmark/ioevent/evpp/evpp_ioevent_bench.cc lines 104-123

 104:     g_pipes.resize(2 * numPipes);
 105:     for (int i = 0; i < numPipes; ++i) {
 106:         if (evutil_socketpair(AF_UNIX, SOCK_STREAM, 0, &g_pipes[i * 2]) == -1) {
 107:             perror("pipe");
 108:         }
 109: 
 110:         if (evutil_make_socket_nonblocking(g_pipes[i*2]) < 0 ||
 111:             evutil_make_socket_nonblocking(g_pipes[i*2 + 1]) < 0) {
 112:         }
 113: 
 114:     }
 115: 
 116:     EventLoop loop;
 117:     g_loop = &loop;
 118: 
 119:     for (int i = 0; i < numPipes; ++i) {
 120:         FdChannel* channel = new FdChannel(&loop, g_pipes[i * 2], true, false);
 121:         channel->SetReadCallback(std::bind(readCallback, channel->fd(), i));
 122:         channel->AttachToLoop();
 123:         g_channels.push_back(channel);

Source: benchmark/ioevent/evpp/evpp_ioevent_bench.cc lines 140-146

 140:     for (auto it = g_channels.begin();
 141:          it != g_channels.end(); ++it) {
 142:         (*it)->DisableAllEvent();
 143:         (*it)->Close();
 144:         delete *it;
 145:     }
 146: }

Impact

The benchmark retains one file descriptor per socketpair until exit and can exhaust descriptor limits at larger pipe counts.