Dropped vs. missed packets in the ixgbe driver

[Manish Vachharajani]

My co-founder, John, just pointed out the problem.

The MPC register on ixgbe is clear on read.  stats.mpc[i] correctly
accumulates the misses, but missed_rx gets set to 0 on any interval
where there are no misses and subsequently, if_errors gets set to 0
(assuming no crcerrs or rlecs.)  I believe the correct fix is in the
patch below:

diff --git a/fbsd/ixgbe-1.7.4/ixgbe.c b/fbsd/ixgbe-1.7.4/ixgbe.c
index f1fa728..9cbaec6 100644
--- a/fbsd/ixgbe-1.7.4/ixgbe.c
+++ b/fbsd/ixgbe-1.7.4/ixgbe.c
@@ -4294,7 +4294,7 @@ ixgbe_update_stats_counters(struct adapter *adapter)
 {
        struct ifnet   *ifp = adapter->ifp;;
        struct ixgbe_hw *hw = &adapter->hw;
-       u32  missed_rx = 0, bprc, lxon, lxoff, total;
+       u32  missed_rx = 0, missed_rx_cum = 0, bprc, lxon, lxoff, total;

        adapter->stats.crcerrs += IXGBE_READ_REG(hw, IXGBE_CRCERRS);

@@ -4303,6 +4303,7 @@ ixgbe_update_stats_counters(struct adapter *adapter)
                mp = IXGBE_READ_REG(hw, IXGBE_MPC(i));
                missed_rx += mp;
                adapter->stats.mpc[i] += mp;
+               missed_rx_cum += adapter->stats.mpc[i];
                adapter->stats.rnbc[i] += IXGBE_READ_REG(hw, IXGBE_RNBC(i));
        }

(Note that it may not apply since I've pulled the driver out into a
separate directory structure)

We need the missed_rx_cum that is the total number of rx misses seen
because missed_rx is decremented from gprc to work around a hardware
issue and so needs to be the misses seen on this call to the function.

Manish


On Thu, Aug 20, 2009 at 11:23 AM, Manish
Vachharajani<manishv at lineratesystems.com> wrote:
> I noticed the bogus XON, XOFF numbers.  I'm glad to see it will be
> fixed so I can cross it off my todo list.  :)  I don't think the issue
> is related though, but you never know.
>
> Barney pointed out that missed_rx in the ixgbe_update_stats_counters
> function accumulates the missed packet registers into the missed_rx
> field.  At the end of the function, these are summed into
> ifp->if_ierrors which should be reported under Ierrs by netstat -idh.
> However that count is zero as reported via netstat.  The stats printfs
> activated via sysctl dev.ix.#.stats=1 prints stats.mpc[0].  This
> number is large (around 400k on the machine I've been using to find
> the problem).  If you look at the code, missed_rx and mpc[i] are
> updated with the same variables.  Given that missed_rx is unsigned, I
> don't understand how the number fails to make it into ifp->if_ierrors,
> but I have yet to dive into debugging this seriously.
>
> Initially, I thought the fix was simple since I didn't see the
> missed_rx getting added to ierrors.  But now, I am not sure where the
> problem is.  Hopefully, it will be more obvious to you than it is to
> me.  I'm sure it is something simple that I am missing.
>
> Manish
>
> On Thu, Aug 20, 2009 at 11:08 AM, Jack Vogel<jfvogel at gmail.com> wrote:
>> I've been looking at the code due to another problem of bogus flow control
>> numbers, and there are some changes needed, things that should be 82599 vs
>> 82598 and were not seperated properly. Will be forthcoming. Not sure if it
>> has any relevance to this, but its possible.
>>
>> Jack
>>
>>
>> On Thu, Aug 20, 2009 at 9:53 AM, Manish Vachharajani
>> <manishv at lineratesystems.com> wrote:
>>>
>>> Oh whoops, sorry didn't see that.  So the plot thickens.  Why don't
>>> these errors show up in the netstat output I forwarded originally?
>>> Ierrs was 0 but the dmesg output clearly shows missed packets.  Any
>>> thoughts on what is going on?  Looking at the code, missed_rx should
>>> certainly get counted in the ierrors field as you said.
>>>
>>> Is the Ierrs in the netstat output some other counter?  If so, how do
>>> I get the if_ierrors variable from the command line?
>>>
>>> Manish
>>>
>>> On Thu, Aug 20, 2009 at 6:49 AM, Barney Cordoba<barney_cordoba at yahoo.com>
>>> wrote:
>>> >
>>> >
>>> > --- On Wed, 8/19/09, Manish Vachharajani <manishv at lineratesystems.com>
>>> > wrote:
>>> >
>>> >> From: Manish Vachharajani <manishv at lineratesystems.com>
>>> >> Subject: Re: Dropped vs. missed packets in the ixgbe driver
>>> >> To: "Barney Cordoba" <barney_cordoba at yahoo.com>
>>> >> Cc: freebsd-net at freebsd.org
>>> >> Date: Wednesday, August 19, 2009, 2:46 PM
>>> >> Agreed, the errors are reported but
>>> >> missed packets are not.  The
>>> >> question is, is the correct fix to just add stats.mpc[0] to
>>> >> if_ierrors
>>> >> in that line or to add it to if_iqdrops.  The fix is
>>> >> easy once we
>>> >> agree on what the correct behavior is.
>>> >>
>>> >> Manish
>>> >>
>>> >> > Barney wrote:
>>> >> >
>>> >> > if you look in ixgbe_update_stats_counters at the
>>> >> bottom:
>>> >> >
>>> >> >        ifp->if_ierrors = missed_rx +
>>> >> adapter->stats.crcerrs +
>>> >> >                adapter->stats.rlec;
>>> >> >
>>> >> > the errors are added in.
>>> >> >
>>> >> > BC
>>> >
>>> > Huh? missed_rx are the missed packets. So they are counted.
>>> >
>>> > BC
>>> >
>>> >
>>> >
>>> >
>>> >
>>>
>>>
>>>
>>> --
>>> Manish Vachharajani
>
>>> _______________________________________________
>>> freebsd-net at freebsd.org mailing list
>>> http://lists.freebsd.org/mailman/listinfo/freebsd-net
>>> To unsubscribe, send any mail to "freebsd-net-unsubscribe at freebsd.org"
>>
>>
>
> --
> Manish Vachharajani
>