em network issues

[Gleb Smirnoff]

On Thu, Oct 19, 2006 at 12:18:16PM -0700, Jeremy Chadwick wrote:
J> > A bit more helpfull, but unfortunately not much is a datapoint saying no
J> > problems April 3rd and watchdog timeouts after September 28 RELENG_6. I
J> > know, probably too vague to be of any use, but there it is.
J> 
J> Someone else has already discussed the date of the commit which
J> supposedly broke this.  Here is the exact post in the exact thread
J> discussing this:
J> 
J> http://lists.freebsd.org/pipermail/freebsd-stable/2006-October/029094.html

Yes, this merge was awesome in its volume. However, it is suspected
that this part of merge causes the problem:

  o a significant performance improvements. the interrupt handler
    schedules work to a private taskqueue. the em_rxeof() function
    runs lockless.
    rev. 1.98 - 1.101 by scottl.

To check whether this is true or not, one needs to build kernel
with em(4) static in the kernel and with DEVICE_POLLING option.
One shouldn't turn polling(4) on em(4), but option must be present
in kernel config. In this case the driver will use interrupt driven
module, but with old style interrupt handler, that doesn't make use
of taskqueue.

I'd appreciate if people who are observing the problem will report
whether adding DEVICE_POLLING option to kernel config helps them
or not. This will help to tell whether the problem is in the above
quote or in the import of new versions from vendor.

-- 
Totus tuus, Glebius.
GLEBIUS-RIPN GLEB-RIPE