不可预测的文件描述符泄漏

Question

Okay, so I'm investigating which seems to be an FD leak.

It's a linux CppCms application written in C++11, I test it with a tools that send approximately 250 requests/seconds.

The tests is during 30 minutes, and sometimes I get the lighttpd (server.c.1446) [note] sockets disabled, out-of-fds error.

I've started investigating FDs using watch "sudo ls /proc/<lighttpd>/fd/ | wc -l; echo /;sudo ls /proc/<myApp>/fd/ | wc -l;echo /; sysctl fs.file-nr"

Results seems to say that everything is fine until it's not. FDs are stable (50 - 100) and randomly get up to the lighttpd max-fd number. If my application was involved, I think it will happened earlier (not after 200000 successful requests)

I've used valgrind to try to see anything and I've found this :

==5647== Open AF_UNIX socket 6: <unknown>
==5647==    at 0x8E379EA: socketpair (syscall-template.S:84)
==5647==    by 0x97F8001: booster::aio::socket_pair(booster::aio::stream_socket&, booster::aio::stream_socket&) (stream_socket.cpp:570)
==5647==    by 0x7312766: cppcms::service::setup_exit_handling() (service.cpp:378)
==5647==    by 0x73135C7: cppcms::service::run() (service.cpp:572)
==5647==    by 0x43F56E: MainMngr::runServer() (mainmngr.cpp:46)
==5647==    by 0x443A95: main (main.cpp:22)
==5647== 
==5647== Open AF_UNIX socket 5: <unknown>
==5647==    at 0x8E379EA: socketpair (syscall-template.S:84)
==5647==    by 0x97F8001: booster::aio::socket_pair(booster::aio::stream_socket&, booster::aio::stream_socket&) (stream_socket.cpp:570)
==5647==    by 0x7312766: cppcms::service::setup_exit_handling() (service.cpp:378)
==5647==    by 0x73135C7: cppcms::service::run() (service.cpp:572)
==5647==    by 0x43F56E: MainMngr::runServer() (mainmngr.cpp:46)
==5647==    by 0x443A95: main (main.cpp:22)

I'm not able to say if it's a real file descriptor leak, but it seems so.

Does my code have a problem ? Does CppCms have a problem ? Have you ever encounter such problem ? Have you any tips to detect who is incrementing the file descriptor handle from 100 to 8000 ?

Answer 1

Sorry. There is a regression in lighttpd 1.4.43 with mod_cgi and CGI POST requests. Description of problem and patch is available in https://redmine.lighttpd.net/issues/2771