[openib-general] [PATCHv3][RFC] kDAPL: use cm timers instead of own

Talpey, Thomas Thomas.Talpey at netapp.com
Wed Jun 1 19:39:23 PDT 2005


At 05:55 PM 6/1/2005, Tom Duffy wrote:
>I was looking at some of the ULPs.  SDP sets the timeout to 20, without
>explanation.  DAPL has a #define (set to 20 saying it is 4 sec in a
>comment).  simple example?  Set to 20, no reason.  To me, it ends up
>just being this opaque value.  Maybe some pre#defines in CM would make
>it clearer.

Constant values are never correct in networking.

The main reason for a timeout is because there isn't a cancel.
Cancel is hard (impossible) because the operation is proceeding
partly in hardware. The most portable thing for an upper layer
to do with a timeout is to specify a "large" value. How big is large
is open to debate.

I think it is less of an upper layer issue and more of a user
environment one. If there's a human with a ^C key, timeouts
will tend to be smaller. If there is kernel mediation (e.g. an NFS
mount point), timeouts can be long.

Tom.



More information about the general mailing list