> And if we want to compare NetBSD's ntp performance to other
> implementations then shouldnt' we use the same API and the same number
> of kernel crossings? 

re: same API:

(1) convince others the use the 'good' API.  8-)

(2) the 'get timestamp' and the existing API's 'get timestamp and set
timestamp' are pretty much identical in terms of operation, as seen
from the user process.

(3) in the short term, i.e. with no third-party source API-use
modification, the compatibility API would do the right thing.


re: kernel entries and performance:

(1) the _only_ extra kernel entry if you're using the 'good' API
should be at program startup, i.e. when timestamping is enabled on the
line.

(2) There would be _no change_ in the number of kernel entries if
you're using the compatibility API.  Just an extra function call (or
so) in the kernel.

(3) If using the 'normal' API, the 'run-time overhead' (i.e. excluding
the extra ioctl at startup to set up the line) would/should be _lower_
than if you were using the broken API, because you wouldn't need to
set the 'enable timestamping' flag on each call.



cgd