svn commit: r278640 - head/sys/netgraph

Julian Elischer julian at freebsd.org
Tue Feb 17 07:07:37 UTC 2015


On 2/16/15 11:26 AM, Gleb Smirnoff wrote:
> On Mon, Feb 16, 2015 at 03:10:50AM -0800, Julian Elischer wrote:
> J> On 2/13/15 4:05 PM, Gleb Smirnoff wrote:
> J> >    Hi!
> J> >
> J> > On Thu, Feb 12, 2015 at 10:20:34PM +0000, Gleb Smirnoff wrote:
> J> > T> Author: glebius
> J> > T> Date: Thu Feb 12 22:20:34 2015
> J> > T> New Revision: 278640
> J> > T> URL: https://svnweb.freebsd.org/changeset/base/278640
> J> > T>
> J> > T> Log:
> J> > T>   Revise default limit for maximum of netgraph data items.
> J> > T>   With modern internet speeds the limit can be reached even
> J> > T>   on a single L2TP link.
> J> >
> J> > Actually any ng_item of data type requires an mbuf to be connected
> J> > to it, and thus I suggest to use mbuf limits to drive ng_item limits.
> J> >
> J> > If we got an item leak with mbufs being properly freed, then we've
> J> > got a bug to fix, and any limit won't work in long run time. I never
> J> > evidenced such a bug, but there complaints on hitting limit at
> J> > traffic bursts. As said, with previous 512 item limit I experienced
> J> > that even on a laptop.
> J> >
> J> > Any objections on removing the limit, guys?
> J> maybe derived from it, but not the same number.
> J> maybe change the systctl to define the relationship?
>
> I guess you want to have ng_items limit smaller than mbuf limit?
> How smaller?
>
> What problem are we actually guarding against putting this limit?

general paranoia, and remember 3rd parties write netgraph nodes.
I just would limit to say 50% by default and have a sysctl for 1-100 %

so netgraph will not lose all mbufs




More information about the svn-src-head mailing list