cyber@ono-sendai.com wrote:

> I'm still catching up on this, how will this affect other cobalt models?

I think this could occur on all cobalt models.
(at least Qube2 and RaQ2)

I guess this could also happen all R5k CPU machines with
tlp(4) or pcn(4) etc. (which can transmit packets via DMA
from unaligned mbufs directly without memcpy).
Actually the same panic happened on my R5000 O2 (sgimips)
with tlp(4) and pcn(4) on its PCI slot.

The problem might be caused by sosend_loan() in sys/kern/uipc_socket.c
which doesn't care virtual index on KVA mappings and creates a bunch of
virtual cache aliases, but I have not tracked the real problem yet.

Maybe the ideal fix is to add some hint to pmap(9) APIs which
indicates virtual index of PA to be mapped to KVA, but it will
take long time.
---
Izumi Tsutsui