On Thu, 2016-06-23 at 15:48 +1000, David Gibson wrote: > From: Benjamin Herrenschmidt <b...@kernel.crashing.org> > > This reworks emulation of the various "rfi" variants. I removed > some masking bits that I couldn't make sense of, the only bit that > I am aware we should mask here is POW, the CPU's MSR mask should > take care of the rest.
See I'd rather we didn't boot at all. I just spent hours trying to figure out why my kernel wouldn't boot in qemu on a mac99 model with 970, weird weird things happening inside the device-tree parsing... Until I figured we were losing the 64-bit mode in the MSR. Why ? Because OpenBIOS isn't bolting the hash entries or SLBs for the entire kernel ! So we are taking some exceptions right during the early assembly, precisely between enable_64b_mode and __mmu_off. Now this is really fishy to begin with, there is code in there that will use SRR0/SRR1 and won't expect a fault of any sort... such as __mmu_off itself. The problem in our case was that OpenBIOS using rfi, it only restores 32-bits of the MSR, so we lose the 64-bit flag. Typically that was happening on the call to __cpu_preinit_ppc970 which happens to reside far enough away that it needs a new translation. I wonder if prom_init should "touch" the entire kernel for safety, but in any case, OpenBIOS need that fix urgently. Cheers, Ben.
