[bug#67245] [PATCH] store: Use a non-blocking socket for store connections.

[Ludovic Courtès]

Hi Christopher,

Christopher Baines <mail@cbaines.net> skribis:

> For some applications, it's important to do this here rather than just making
> the socket non-blocking after the connection is established because there can
> be I/O on the socket that will block during the handshake.
>
> I've noticed this blocking during the handshake causing issues in the build
> coordinator for example.
>
> * guix/store.scm (open-unix-domain-socket, open-inet-socket): Pass
> SOCK_NONBLOCK when calling socket.
>
> Change-Id: I8225762b78448bc1f7b698c8de5d736e13f577bf

I feel we should really discuss on Guix + Fibers since we’ve apparently
been going through the exact same set of issues.  :-)

(The other thing that comes to mind is the resource pool!)

> +++ b/guix/store.scm
> @@ -460,7 +460,9 @@ (define (open-unix-domain-socket file)
>  '&store-connection-error' upon error."
>    (let ((s (with-fluids ((%default-port-encoding #f))
>               ;; This trick allows use of the `scm_c_read' optimization.
> -             (socket PF_UNIX (logior SOCK_STREAM SOCK_CLOEXEC) 0)))
> +             (socket PF_UNIX
> +                     (logior SOCK_STREAM SOCK_CLOEXEC SOCK_NONBLOCK)
> +                     0)))

We cannot do this here because callers have to be prepared to deal with
non-blocking sockets, and that’s not the case in Guix itself.

In Cuirass, I have this:

--8<---------------cut here---------------start------------->8---
(define (non-blocking-port port)
  "Make PORT non-blocking and return it."
  (let ((flags (fcntl port F_GETFL)))
    (when (zero? (logand O_NONBLOCK flags))
      (fcntl port F_SETFL (logior O_NONBLOCK flags)))
    port))

(define (ensure-non-blocking-store-connection store)
  "Mark the file descriptor that backs STORE, a <store-connection>, as
O_NONBLOCK."
  (match (store-connection-socket store)
    ((? file-port? port)
     (non-blocking-port port))
    (_ #f)))

(define-syntax-rule (with-store/non-blocking store exp ...)
  "Like 'with-store', bind STORE to a connection to the store, but ensure that
said connection is non-blocking (O_NONBLOCK).  Evaluate EXP... in that
context."
  (with-store store
    (ensure-non-blocking-store-connection store)
    (let ()
      exp ...)))
--8<---------------cut here---------------end--------------->8---

Then ‘with-store/non-blocking’ is used in fiberized context where I know
this is fine.

I think it’ll have to remain this way until Guix itself is fiberized or
something.

Does that make sense?

Ludo’.