On Mon, Nov 09, 2009 at 03:53:21PM -0600, Scott Wood wrote: > On Fri, Nov 06, 2009 at 10:29:44AM +0100, Joakim Tjernlund wrote: > > > > With this, the kernel hangs after "Mount-cache hash table entries: 512". > > > > > > Somewhat surprising result. I didn't expect you would even hit this > > > condition now as we haven't enabled the use of dcbX insn yet. > > > The only thing I can think of is the you hit the 0x00f0 due to other > > > dcbX insn use and the kernel managed to fixup this in the page fault > > > handler > > > by pure luck before. > > It's bizarre -- it still happens even if I revert the branch to FixupDAR. > However, if I comment out the final "b 151b", it stops happening. It also > stops happening sometimes depending on where I stick printk()s to debug > this. > > So it may be an unrelated issue that just got perturbed somehow.
OK, figured it out. The fixup code pushed things around so that in syscall_exit_cont, SRR0/SRR1 were being loaded immediately prior to a page boundary, with the rfi after the page boundary. On crossing the boundary, we take an ITLB miss (which goes from possibility to certainty with the CPU15 workaround), and SRR0/SRR1 get clobbered. I suppose we'll need to find all places where we do rfi with the MMU enabled, and ensure acceptable alignment. :-( Either that, or require that the kernel text be pinned, though that does not interact well with CPU15. -Scott _______________________________________________ Linuxppc-dev mailing list Linuxppc-dev@lists.ozlabs.org https://lists.ozlabs.org/listinfo/linuxppc-dev
