qlxgb: enter the network epoch in qla_rcv_isr

[mark_cloudbsd.org]

Summary

tcp_lro_flush_all() asserts the network epoch before it looks at the queue. qla_isr() is an INTR_TYPE_NET thread and already runs inside that epoch. When the status ring is not drained, it enqueues qla_rcv(), and that task calls the same qla_rcv_isr() with no epoch held. The assert fires even when the queue is empty.

Enter the epoch for the body of qla_rcv_isr(), after the stop_rcv return, so both callers are covered. A nested enter from the interrupt thread is valid.

Test Plan

I installed this patch on a server; used iperf to test; then tested services running on that machine. As this is a "quick fix" with what limited testing harness I have, it looks good for me, but if anyone has anything more elaborate, it would be welcomed.So far so good with what I have, I got 9.4gb with iperf for my consumer grade 10gb equipment.

[mark_cloudbsd.org]

Updated to contain more context

[crest_freebsd_rlwinm.de]

Looks correct to me.

• net epoch enter/exit is balanced
• nested epoch enter/exit is allowed
• nothing between enter and exit sleeps

I think this driver has been broken since at least rGa6b55ee6be15. Without this patch the driver doesn't just trip over an assertion on a debug kernel. On a production kernel it proceeds without entering the net epoch. This should be MFCed to at least 14-stable and 15-stable. Given how long it took for someone to report (and fix) this system crashing driver bug an errata is probably not required, but would be nice to have if releng has the capacity.