Sorry for the late update about this. I can confirm RM7000 O2 systems experience sporadic failures from 4.9 onwards, while they don't in 4.8.
An R5000 system with 256MB of memory runs fine. I have had a look at the 4.8->4.9 changes, and nothing seems suspicious. So I need to check which one of these things is triggering the problems on the RM7000: - physical memory larger than 256MB (thought it should not have any effect on the O2 kernels with 4KB pages) - the gcc 3 -> gcc 4 switch, which might produce a sequence of instructions which would in turn make the RM7000 unhappy. In particular, I wonder if gcc4 may trigger RM7000 errata #32 when generating switch() jump tables. Miod
