iflib sysctl cleanup

[shurd]

Summary

While working on an iflib.4 manpage, various issues with
the sysctl OIDs exposed by iflib came up. This patch has fixes for
most of them, but there's still a couple open questions.

Don't increment txq_drain_encapfail twice on all errors other than ENOBUFS
Add comments regarding txq_drain_encapfail and encap_txd_encap_fail this needs further review
Increment rxd_flush whenever ctx->isc_rxd_flush() is called
Increment tx_encap whenever ctx->isc_txd_encap() is called
Increment rx_intr_enables whenever IFDI_RX_QUEUE_INTR_ENABLE() is called
Increment tx_frees in more places where tx mbufs are freed and not cloned
Increment txq->ift_pullups whenever m_pullup() is called. This requires adding a txq argument to collapse_pkthdr()
Increment txq->ift_mbuf_defrag whenever m_defrag() is called

[shurd]

Remove bnxt.4 changes from review.

[shurd]

sys/net/iflib.c

846 ↗

(On Diff #46736)

There's an open question here of the netmap refill should increment the fl_refills* values.

3480 ↗

(On Diff #46736)

Not sure why this doesn't count against encap_txd_encap_fail.

[gallatin]

sys/net/iflib.c

3055 ↗

(On Diff #46737)

Totally not your fault, but looking at the review reminded me that this code existed. This collapse_pkthdr() thing is broken (derefs m_next->m_next after m_next has been freed via m, which is the same pointer). I don't recall any other driver checking for a pkthdr with no data, and I don't think the stack will ever do that to a driver, so maybe we should just remove the check, or replace it with an error return or even a panic.. It seems clear that this code is not tested, so if we ever did encounter such a case, we'd go off the rails.

3480 ↗

(On Diff #46736)

I think all the error returns should count against encap_txd_encap_fail.

Eg, maybe set an retval to replace the return values in the existing error returns, and then always jump to a new encap_failed: label after defrag_failed than increments encap_txd_ecnap_fail and returns?

[shurd]

Address feedback.