[openib-general] rdma_listen() backlog parameter
Caitlin Bestler
caitlinb at broadcom.com
Thu Nov 17 16:31:17 PST 2005
-----Original Message-----
From: Sean Hefty [mailto:mshefty at ichips.intel.com]
Sent: Thu 11/17/2005 4:24 PM
To: Caitlin Bestler
Cc: Steve Wise; openib-general at openib.org
Subject: Re: [openib-general] rdma_listen() backlog parameter
> Agreed, although I would think of it in terms of how many
> connection requests the ULP is willing to accept and then
> secondly as to what resources the CM/driver/device needs.
I re-added the backlog parameter to the kernel CMA. From userspace, the backlog
is the maximum number of connection request events that will be queued for the
user. From the kernel, the backlog is the number of resources needed by the
underlying device or transport.
- Sean
------ start reply
Even in kernel sticking with the "largest number of connection
requests I'm willing to handle" provides for transport neutral
semantics.
The "number of resources" assumes we know what those resources
actually are. For example, over IP, you have two distinct set
of resources: the TCP connection backlog and the number of open
connections on which MPA mode is being negotiated. Those differences
can be swept under the rug if a backlog of "N" means that the CM
should have sufficient resources to generate N connection requests,
and promises not to inflict the N+1st request on the application.
I suppose a case could be made for this being a range (from I want
at least X but no more than Y), but there's a lot of tradition
behind having a single number to represent the connection backlog.
More information about the general
mailing list