Re: [PATCH v3 2/3] QIOChannelSocket: Implement io_async_write & io_async_flush

[Peter Xu <peterx@redhat.com>]

On Wed, Sep 29, 2021 at 04:36:10PM -0300, Leonardo Bras Soares Passos wrote:
> On Tue, Sep 28, 2021 at 7:45 PM Peter Xu <peterx@redhat.com> wrote:
> >
> > On Wed, Sep 22, 2021 at 07:24:22PM -0300, Leonardo Bras wrote:
> > > +static void qio_channel_socket_async_flush(QIOChannel *ioc,
> > > +                                           Error **errp)
> > > +{
> > > +    QIOChannelSocket *sioc = QIO_CHANNEL_SOCKET(ioc);
> > > +    struct msghdr msg = {};
> > > +    struct pollfd pfd;
> > > +    struct sock_extended_err *serr;
> > > +    struct cmsghdr *cm;
> > > +    char control[CMSG_SPACE(sizeof(int) * SOCKET_MAX_FDS)];
> > > +    int ret;
> > > +
> > > +    memset(control, 0, CMSG_SPACE(sizeof(int) * SOCKET_MAX_FDS));
> > > +    msg.msg_control = control;
> > > +    msg.msg_controllen = sizeof(control);
> > > +
> > > +    while (sioc->async_sent < sioc->async_queued) {
> > > +        ret = recvmsg(sioc->fd, &msg, MSG_ERRQUEUE);
> > > +        if (ret < 0) {
> > > +            if (errno == EAGAIN) {
> > > +                /* Nothing on errqueue, wait */
> > > +                pfd.fd = sioc->fd;
> > > +                pfd.events = 0;
> > > +                ret = poll(&pfd, 1, 250);
> > > +                if (ret == 0) {
> > > +                    /*
> > > +                     * Timeout : After 250ms without receiving any zerocopy
> > > +                     * notification, consider all data as sent.
> > > +                     */
> > > +                    break;
> >
> > After a timeout, we'll break the while loop and continue parsing an invalid
> > msg [1].  Is that what we want?
> 
> No, the point here was returning from flush if this (long) timeout
> happened, as in
> "if asso long has passed, there must be no pending send", which I
> agree is quite bad,
> but it was all I could think to avoid an infinite loop here if
> something goes wrong.

IMHO it's the same when we write() to a socket but the buffer is always full,
we'll simply block there until it has some space.  I don't know what we can do
here besides infinite loop on the timeout - we shouldn't eat the cpu all, but
we should still wait?

-- 
Peter Xu