Hello.
Today socket-connect option :nodelay if specified accepts T or NIL. And whatever value is passed, socket-connect signals the UNSUPPORTED exception on implementations where it is not possible to control the socket no-delay property.
This dooms any usocket-using code which wishes to use :nodelay to be broken on implementations where it is not implemented (e.g. CLISP, CMUCL, ...).
I think in most cases users would prefer "no delay if possible, otherwise just work as you can" behavior. For example drakma. On the implementations where it is possible, it will have maximum performance by specifying no-delay. But on other implementations it is desirable to work on usual socket.
To support this in backward-compatible fashion I propose to introduce another value for the :nodelay option. :nodelay :if-supported. The meaning is obvious.
Please see the patch attached. It is tested (by changing drakma to use :nodelay :if-supported) on all the implementations available to me: Allegro, CCL, CLISP, SBCL.
Can't test on ECL and ABCL because of other obstacles (the last public release of ECL can't comile babel, and ABCL can't load cffi).
Best regards, - Anton
Anton,
I committed your patch to usocket's trunk, thanks a lot!
-Hans
On Fri, Jan 20, 2012 at 11:23 PM, Anton Vodonosov avodonosov@yandex.ru wrote:
Hello.
Today socket-connect option :nodelay if specified accepts T or NIL. And whatever value is passed, socket-connect signals the UNSUPPORTED exception on implementations where it is not possible to control the socket no-delay property.
This dooms any usocket-using code which wishes to use :nodelay to be broken on implementations where it is not implemented (e.g. CLISP, CMUCL, ...).
I think in most cases users would prefer "no delay if possible, otherwise just work as you can" behavior. For example drakma. On the implementations where it is possible, it will have maximum performance by specifying no-delay. But on other implementations it is desirable to work on usual socket.
To support this in backward-compatible fashion I propose to introduce another value for the :nodelay option. :nodelay :if-supported. The meaning is obvious.
Please see the patch attached. It is tested (by changing drakma to use :nodelay :if-supported) on all the implementations available to me: Allegro, CCL, CLISP, SBCL.
Can't test on ECL and ABCL because of other obstacles (the last public release of ECL can't comile babel, and ABCL can't load cffi).
Best regards,
- Anton
-
Anton Vodonosov -
Hans Hübner