Re: [tcpm] poll for adoption of draft-ananth-persist-02

[John Heffner <[e-mail]>]

On Thu, Apr 1, 2010 at 3:53 PM, Mark Allman <[e-mail]> wrote:
>> There are other aspects too, it's also important for the application
>> to control the duration that the connection spends in persist state,
>> this is planned to be accomplished via a socket option.
>
> Why?
>
> I mean, sure, an app may want to not let a connection hang around
> forever.  But, can't it just ABORT whenever it wants?

I suspect one reason this has become an issue is that the sockets
interface does not provide ABORT semantics, so it's not possible on
most operating systems to write an application that is able to avoid
tying up kernel memory.  I'm not sure this is an IETF issue, though.

I think a good approach for an operating system implementor is to (1)
provide ABORT via a socket fd so individual applications can manage
their own connections, and (2) provide a system-wide nanny so admins
can provide some degree of safety even if applications aren't written
so carefully (obviously, most aren't).

(And, not to sound like a broken record, the receiver's current window
size should probably have nothing to do with whether the connection is
aborted.)

  -John