[openib-general] Re: [RFC] [PATCH] user_mad: Support RMPP on send side

[Hal Rosenstock]

On Mon, 2005-05-23 at 10:29, Roland Dreier wrote:
>     Hal> It appears to me that there is nothing that enforces the man
>     Hal> page behavior above in Linux so this is a "convention". Is
>     Hal> this something we can take advantage of or do we need to
>     Hal> require the additional call to get the buffer length ?
> 
> Something I was thinking a little while ago was that the character
> device interface for userspace MADs might have been a mistake.  Having
> a new datagram socket family would solve this problem quite nicely --
> a consumer could pass MSG_PEEK and/or MSG_TRUNC to recv() to get the
> true length of a datagram.  

One could always add an ioctl to the character device to do this if it
is unacceptable to change the convention for this character device.

Changing over to a socket based interface appears to be a longer term
approach. Does this have other advantages over the character device
based approach ? Are sockets more hot swap friendly ?

> However it's a little tricky to see a
> really clean way for userspace to use this to handle arbitrary sized
> RMPP receives.

Not sure what you are referring to here. Wouldn't the returned size be
used for the size on the subsequent read ? Isn't the issue still the
same whether it is socket or character device based: how is a read with
a too short length handled ?

-- Hal