svn commit: r234834 - in head/sys: contrib/pf/net net netinet netinet/ipfw ofed/drivers/infiniband/ulp/ipoib

[Alexander V. Chernikov]

On 30.04.2012 15:36, Gleb Smirnoff wrote:
> On Mon, Apr 30, 2012 at 01:48:36PM +0200, Luigi Rizzo wrote:
> L>  On Mon, Apr 30, 2012 at 10:22:23AM +0000, Alexander V. Chernikov wrote:
> L>  >  Author: melifaro
> L>  >  Date: Mon Apr 30 10:22:23 2012
> L>  >  New Revision: 234834
> L>  >  URL: http://svn.freebsd.org/changeset/base/234834
> L>  >
> L>  >  Log:
> L>  >    Move several enums and structures required for L2 filtering from ip_fw_private.h to ip_fw.h.
> L>
> L>  I would be really grateful if you could revert this back and discuss
> L>  what you wanted to achieve with this change other than saving one
> L>  entry in the list of includes.
Changing something inside ip_fw_private.h (for example, locking  change) 
requires changes in several totally unrelated subsystems, which is 
clearly bad.
> L>
> L>  As clearly mentioned in the commit logs
> L>
> L>      http://svnweb.freebsd.org/base?view=revision&revision=200580
Maybe there are some other possibilities documenting preferred layout 
other than commit log? Searching 2+yrs commit history is not the best 
way of finding information.

> L>
> L>  when i did the last revision of the ipfw+dummynet code i tried
> L>  to put a strong separation between what is visible in userland
> L>  (ip_fw.h and ip_dummynet.h) and kernel specific stuff.
> L>  This way changes in the kernel code do not need to affect userland,
> L>  modify installed headers and so on.
> L>
> L>  This is why kernel-specific definitions were put in private files.
Unfortunately, it is not so obvious (at least for me "ip_fw_ 
_private__.h" looks much like "ipfw private headers, used by ipfw 
subsystem only", not "kernel ipfw specific stuff").
> L>  We may discuss on the filename, ip_fw_kernel.h may be a better fit,
Personally I prefer glebius@ suggestion with ip_fw_var.h for such common 
headers.
> L>  but merging back kernel and userland defs is a bad design decision.
So should style(9) be updated with "_KERNEL define is bad" line to make 
newcomers aware of this "easy" way ? :)
> L>
> L>  20-30 years ago there were good reasons to use a single header
> L>  for all sorts of definitions: user-only, kernel-only, and kernel-userland API.
> L>  Machines were slow, disks were small, portability was not a big deal.
> L>
> L>  These days none of these conditions apply and keeping things
> L>  separate helps maintainance and avoid accidental pollution of
> L>  definitions and their misuse.
> L>
> L>  Besides, keep in mind that ipfw and dummynet are meant to work
> L>  on multiple platforms so this change is causing portability troubles.
>
> Can we split ip_fw_private.h to ip_fw_private.h, and ip_fw_var.h? The
> former is really private, and the latter is for other kernel modules.