svn commit: r271782 - in head/sys/dev: bge bxe e1000 fxp nfe

[Gleb Smirnoff]

On Thu, Sep 18, 2014 at 12:18:24PM -0700, Adrian Chadd wrote:
A> On 18 September 2014 12:11, Gleb Smirnoff <glebius at freebsd.org> wrote:
A> > On Thu, Sep 18, 2014 at 03:56:14PM +0000, Gleb Smirnoff wrote:
A> > T> Author: glebius
A> > T> Date: Thu Sep 18 15:56:14 2014
A> > T> New Revision: 271782
A> > T> URL: http://svnweb.freebsd.org/changeset/base/271782
A> > T>
A> > T> Log:
A> > T>   - Use if_inc_counter() to increment various counters.
A> > T>   - Do not ever set a counter to a value. For those counters
A> > T>     that we don't increment, but return directly from hardware
A> > T>     create cases in if_get_counter() method.
A> >
A> > Note that this patch was done mechanically. However, with new API
A> > all statistics handling in all drivers can be improved. It is no
A> > longer required to poll hardware counters and write down their
A> > values into ifnet. Hardware counters now can (and should!) be read
A> > on demand at the time when either sysctl(2) or network stack
A> > asks for a counter.
A> >
A> > Of course such change to drivers can't be committed without
A> > testing on hardware, that's why I decided to be conservative.
A> >
A> > I'll convert several drivers as an example soon.
A> 
A> Something that recently popped up at home with my RSS testing is that
A> the hardware counters may actually reflect the packets that the MAC
A> has seen, rather than the packets the driver side has actually seen.
A> 
A> It looks like with flow control disabled, the counters used for
A> ixgbe(4) reflect the per-queue and global MAC RX counters, before a
A> saturated receive thread drops some frames. Before, the interface
A> counters reflected what the driver managed to see, not necessarily
A> what was received by the NIC.
A> 
A> What should we do about that?

Of course we should report that as ierrors, being silent on them
is a disaster for a sysadmin.

-- 
Totus tuus, Glebius.