e1000: Count packets dropped for want of a receive descriptor

[rcm]

Summary

The igb class, 82575 through I354, has the same per-queue RQDPC register
as igc(4): the count of packets the hardware dropped because a receive
ring was full. The driver never read it, so such drops were invisible;
MPC counts packet buffer overflows, not descriptor exhaustion.

Read it for each queue with the other statistics on those parts and
account the drops per queue as dev.em.N.queue_rx_M.rx_drops and in total
as the interface's input queue drop counter (netstat -d, Idrop). On this class
the register clears on read. Earlier parts have no such register and are unchanged.

MFC after: 2 weeks
Sponsored by: Rubicon Communications, LLC ("Netgate")

[kbowling]

This is trickier because i210/i211 aren't read1clear. It may be worth considering delta tracking and resetting the value on init here?

IOV initialization remaps PF queue 0 to sc->pool (sys/dev/e1000/if_igb_iov.c:2088). Reading RQDPC(i) reads a VF’s counter while missing the PF’s drops. Use sc->rx_queues[i].rxr.me. On I350, the wrong read also clears that VF’s counter, per erratum 3. Gate the new sysctl (sys/dev/e1000/if_em.c:7870) with !sc->vf_ifp: VF statistics bypass this accumulation, so it currently exposes an always-zero counter.