On 07/22/2014 12:32 AM, Chris Dickinson wrote: > Hi all, > > I'm currently working on a feature for Node that would allow Node to > apply backpressure to incoming socket connections -- specifically, > letting them pool in the kernel if JS is currently busy. > In the course of implementing this feature, I found that it was > impossible to keep libuv from accept()'ing the socket from the Node > side. Specifically, uv__server_io calls accept() before the > connection_cb() [1]. In order to let the client apply backpressure / > keep sockets waiting in the kernel queue, it would be necessary to move > the accept() call & associated logic into uv_accept, > so that the connection_cb() can trigger accept() itself. Is this a sane > course of action? The only caveat I saw was that the emfile trick[2] > would have to be moved / otherwise accounted for somehow. > > Thanks for your feedback! > Chris > > [1]: > https://github.com/joyent/libuv/blob/06c60e9662fea752d77237857bbf2c9ea5889a00/src/unix/stream.c#L468 > [2]: > https://github.com/joyent/libuv/blob/06c60e9662fea752d77237857bbf2c9ea5889a00/src/unix/stream.c#L415 >
I guess some uv_listen_stop would also work for you, right? The idea being that you stop listening for new incoming connections for a while (for example to mitigate EMFILE). We do have some plans, which we'll put together and share after libuv 0.12 is released, but in a nutshell uv_listen would be going away, along with the current incarnation of uv_accept. The new uv_accept will be a request, which completes when a new connection is accepted, gives you the fd, and you can then create a handle with it. There are some details to be polished, but that's the idea. So, havin libuv 0.12 around the corner, and considering we already have plans to change uv_listen / uv_accept, I'm not sure I'd like to change the behavior to something else in the meantime. Cheers, -- Saúl Ibarra Corretgé bettercallsaghul.com
signature.asc
Description: OpenPGP digital signature
