On Mon, Dec 01, 2014 at 03:03:50PM +0000, Peter Maydell wrote: > On 1 December 2014 at 14:02, Richard W.M. Jones <rjones@xxxxxxxxxx> wrote: > > OK I fixed it ... > > > > This patch to qemu fixes the problem for me. I will post it to > > qemu-devel shortly. [In reply to your other comment about wasting memory on machines with small amounts of memory] How about we just do this on aarch64? Isn't it the case that only aarch64 has 64K pages, and aarch64 machines presumably all have loads of RAM. > Really this is a workaround for a bug in the (guest) kernel. > The booting ABI documentation says nothing about having to keep > the DTB outside the tail end of the page occupied by the initrd, > so it's a kernel bug that it fails to work in this situation > (or a kernel documentation bug that it doesn't say that the > last part-page of the initrd can't be shared with anything, > if you prefer). What's actually happening is the kernels frees initrd memory just after unpacking the initramfs (arch/arm64/mm/init.c:free_initrd_mem). That happens to be some time before the device tree is parsed for virtio-mmio devices. The act of freeing the initrd memory deletes a bit of the device tree (or probably allows that memory to be overwritten by a future allocation somewhere) with predictable consequences. So I don't know how to fix this easily in the kernel. Perhaps: - move the initrd free later? - don't free the initrd memory (wasting much more than my qemu patch) - don't free the last page of initrd? (this might have been the intention of Yalin Wang's patch?) TBH it does sound like a documentation bug - the kernel requires that the initrd occupies exclusively whole pages. IA64 has a similar-but-different problem with the kernel possibly overlapping the initrd because EFI and the kernel might differ about page size (see arch/ia64/mm/init.c:free_initrd_mem). This leads me to think that we may need to align the start of the initrd as well. Rich. -- Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones Read my programming and virtualization blog: http://rwmj.wordpress.com virt-df lists disk usage of guests without needing to install any software inside the virtual machine. Supports Linux and Windows. http://people.redhat.com/~rjones/virt-df/ _______________________________________________ kvmarm mailing list kvmarm@xxxxxxxxxxxxxxxxxxxxx https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
