When a bound evhttp socket triggers for an accept with no free file descriptors, it fails with EMFILE or ENFILE, the event isn’t cleared and thus continues to call the accept handler until the connection is dropped. This seems like a denial of service vulnerability.

I’ve been thinking about how to handle this elegantly, and I’m wondering if it makes sense for the calling program to allocate libevent a certain number of file descriptors. Then, libevent can do something smart (allow incoming connections to pile up until we have a free fd? disallow keep-alives if we’re low on fds?).

Thoughts?

Brian

_______________________________________________
Libevent-users mailing list
Libevent-users <at> monkey.org
http://monkeymail.org/mailman/listinfo/libevent-users

I’m seeing interesting behavior. When evhttp_make_request fails (which I can make happen by restricting the number of fds available), it seems that sometimes the callback on the connection is called, and sometimes it’s not. Is there a best practice for what to do when this fails? I don’t care if the callback is called, but I need to put cleanup logic either in there or in my failure handler (right now it gets called twice – sometimes – which causes crashes).

Thanks!
Brian

_______________________________________________
Libevent-users mailing list
Libevent-users <at> monkey.org
http://monkeymail.org/mailman/listinfo/libevent-users

.hmmessage P { margin:0px; padding:0px } body.hmmessage { FONT-SIZE: 10pt; FONT-FAMILY:Tahoma } .hmmessage P { margin:0px; padding:0px } body.hmmessage { FONT-SIZE: 10pt; FONT-FAMILY:Tahoma } Hi,
I'm new to libevent. I'm having some issues with my server program on multi-processor machines. The same code seemed to work fine on a uniprocessor machine. The machine is running FC8, kernel 2.6.23.1-42.

The errors I'm getting are:
[err] event_queue_remove: 0x9bcb0c4(fd 293) not on queue 8
and sometimes
event_process_active: Assertion `activeq != ((void *)0)' failed

The server consists of a main dispatch thread running evhttp and multiple worker threads, each with their own instance of libevent.
The main dispatch thread gets an http request, sticks it onto a worker thread's queue, and then writes a byte to the pipe of the worker thread to notify it.
The worker thread will respond to the http request.

Here's the skeleton of my program

main() {
   httpd = evhttp_start("0.0.0.0", 8080);
   evhttp_set_gencb(httpd, generic_handler, NULL);
   event_dispatch();
   evhttp_free(httpd);
}

generic_handler(evhttp_request *req, void *arg)
{
    pick worker thread 'w'
    stick request on worker thread queue
    write(thread[w].send_fd, "", 1);
}

Each thread's initialization includes:
thread->base = event_init();
event_set(thread->event, thread->receive_fd, EV_READ | EV_PERSIST, libevent_process, thread);
event_base_set(thread->base, thread->event);
event_add(thread->event, 0);

libevent_process(int fd, short which, void *arg)
{
    if (read(fd, buf, 1) != 1) error;
   
    if item on queue, respond then free item
}

Could you let me know which assumptions I'm violating here?
Thanks,
- Jad

_______________________________________________
Libevent-users mailing list
Libevent-users <at> monkey.org
http://monkeymail.org/mailman/listinfo/libevent-users