[Bug 192013] [xen] [pf] pf performance very bad in xen when tso enabled

[bugzilla-noreply at freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=192013

Dag-Erling Smørgrav <des at FreeBSD.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
 Attachment #144852|0                           |1
        is obsolete|                            |
             Status|Needs Triage                |In Discussion
           Assignee|freebsd-bugs at FreeBSD.org    |des at FreeBSD.org

--- Comment #2 from Dag-Erling Smørgrav <des at FreeBSD.org> ---
Created attachment 144883
  --> https://bugs.freebsd.org/bugzilla/attachment.cgi?id=144883&action=edit
More conservative patch

When you say “TSO is disabled”, I assume you mean on the virtual interface
(xnN) in the guest?

I'm not comfortable removing this block of code without a clear understanding
of why it's there.  It looks to me like the intent is to ensure that all
packets have a valid checksum before they enter pf.  Are we certain that pf
won't break in some subtle way if that assumption is violated?

BTW, the patch does not remove the equivalent code in pf_check6_out(), but that
code is broken anyway (as the comment points out).

A more conservative fix is to keep calculating the checksum before pf_test
*unless* hardware checksum offloading is enabled.

-- 
You are receiving this mail because:
You are the assignee for the bug.