[systemd-devel] [PATCH kdbus] handle: Return POLLOUT | POLLWRNORM mask when no messages are pending

[Lennart Poettering]

On Fri, 15.08.14 14:16, Marcel Holtmann (marcel at holtmann.org) wrote:

> 
> Hi Daniel,
> 
> >> To facility the feature of doing an asynchronous sending of messages
> >> when the bus is idle, make sure to return POLLOUT | POLLWRNORM from
> >> kdbus_handle_poll.
> >> 
> >> Signed-off-by: Marcel Holtmann <marcel at holtmann.org>
> >> ---
> >> handle.c | 2 ++
> >> 1 file changed, 2 insertions(+)
> >> 
> >> diff --git a/handle.c b/handle.c
> >> index ac6868133280..fc15d28351b3 100644
> >> --- a/handle.c
> >> +++ b/handle.c
> >> @@ -884,6 +884,8 @@ static unsigned int kdbus_handle_poll(struct file *file,
> >> 		mask |= POLLERR | POLLHUP;
> >> 	else if (!list_empty(&conn->msg_list))
> >> 		mask |= POLLIN | POLLRDNORM;
> >> +	else
> >> +		mask |= POLLOUT | POLLWRNORM;
> > 
> > Hmm, what's your use case here? list_empty(&conn->msg_list) only checks
> > the incoming list of the current connection for pending messages. That
> > doesn't tell you whether the bus is idle, or if your receiving end has
> > messages pending ...
> 
> if you are a good client citizen, then you only send messages when
> POLLOUT gets signaled. That is what this is allowing now.
> 
> Blindly sending messages is never a good idea. You want to poll for
> POLLOUT first. This does not make a big difference for current kernel,
> but it allows future extensions when the clients are well behaving and
> just waiting for POLLOUT. Meaning once the kernel does not signal
> POLLOUT, no new messages will come from the client.
> 
> Current code does not do this POLLOUT before sending a messages, but
> our kdbus client actually does that. It is the right thing to do. And
> we have been doing this with all of our protocols that are using
> asynchronous IO. No point in kdbus being any different.

Well, kdbus keeps per-reciever buffers only, hence signalling on the
kdbus fd when you are able to write is not really possible, since this
information is not bound to the sender fd but only to the receiever of
which there are many... If I understand you correctly you hence want the
kdbus fd to always return EPOLLOUT then, because if a client wants to
send something it can do that at any time?

If that's the case then POLLOUT should really be ORed into the mask
unconditionally, not just in some cases...

So, I can sympathize with what you are trying to do. However, I think
your patch doesn't do the right thing... It should really OR the POLLOUT
into all masks always.

Lennart

-- 
Lennart Poettering, Red Hat