[ovs-discuss] branch-2.3: rx/tx counter overflow

[Ben Pfaff]

On Tue, Oct 28, 2014 at 07:32:36PM +0400, Andrey Korolyov wrote:
> On Tue, Oct 28, 2014 at 6:31 PM, Ben Pfaff <blp at nicira.com> wrote:
> > On Tue, Oct 28, 2014 at 03:37:58PM +0400, Andrey Korolyov wrote:
> >> On Mon, Oct 27, 2014 at 8:19 PM, Ben Pfaff <blp at nicira.com> wrote:
> >> > On Mon, Oct 27, 2014 at 9:15 AM, Andrey Korolyov <andrey at xdel.ru> wrote:
> >> >> On Mon, Oct 27, 2014 at 7:15 PM, Ben Pfaff <blp at nicira.com> wrote:
> >> >>> On Mon, Oct 27, 2014 at 08:12:28PM +0400, Andrey Korolyov wrote:
> >> >>>> On Mon, Oct 27, 2014 at 7:09 PM, Ben Pfaff <blp at nicira.com> wrote:
> >> >>>> > On Mon, Oct 27, 2014 at 08:05:01PM +0400, Andrey Korolyov wrote:
> >> >>>> >> On Mon, Oct 27, 2014 at 7:04 PM, Ben Pfaff <blp at nicira.com> wrote:
> >> >>>> >> > On Mon, Oct 27, 2014 at 07:55:01PM +0400, Andrey Korolyov wrote:
> >> >>>> >> >> ovs-ofctl dump-ports currently reporting values not large than u32 in
> >> >>>> >> >> the mentioned branch. lib/ofp-util.c has no regressions at a glance,
> >> >>>> >> >> probably truncation going in the different (not so obvious) way.
> >> >>>> >> >
> >> >>>> >> > Are you using a 64-bit kernel?  There is some unavoidable truncation
> >> >>>> >> > with 32-bit kernels.
> >> >>>> >>
> >> >>>> >> Yes, of course.
> >> >>>> >
> >> >>>> > Thanks.  I guess you must have previously seen 64-bit values with some
> >> >>>> > earlier version.  Do you know what the most recent version was?
> >> >>>>
> >> >>>> For 0fe1d7f39de9836fea01c560a6fdbfd1405096ea it is clearly positive
> >> >>>> that it reports 64-bit counter values (branch-2.1). Had not tested
> >> >>>> against other revisions yet. Are you suggesting that the counter
> >> >>>> behavior with 32b limit is currently taken as a right one?
> >> >>>
> >> >>> I am trying to narrow down the range of commits that could have caused
> >> >>> the problem.  "git bisect" would be the ideal way to do it, if you are
> >> >>> willing and able to try it.
> >> >>
> >> >> Heh, okay. I`m signing over bisection, please give me some time :)
> >> >
> >> > Thanks a lot.
> >>
> >>
> >> The bad cast introduced by 04c881eb6441fff2e91c9b9e23502bc554c0f437.
> >
> > That patch only adds assignments of 64-bit integers to 64-bit integers.
> > No casts or conversions are involved.
> >
> > Looking more closely, I think the problem here is that even 64-bit
> > kernels always pass IFLA_STATS to userspace using 32-bit integers.
> > Usersspace needs to look at IFLA_STATS64, instead, when it is present,
> > but there is currently no code to do that.
> 
> 
> Thanks Ben. How big is an inaccuracy impact from calculation in a
> legacy (regarding to this patch) manner with enabled offloads, if you
> have such information?

It would depend on the size of the packets being offloaded.  A 64 kB TCP
packet would be offloaded into over 40 1500-byte on-wire frames, so in
the worst case the packet count could be off by a factor of about 40.