On Thu, Jan 21, 2016 at 02:43:15PM +0900, AKASHI Takahiro wrote: > On 01/20/2016 11:59 PM, Ard Biesheuvel wrote: > >On 20 January 2016 at 13:36, Mark Rutland <mark.rutland at arm.com> wrote: > >>Ard, Ganapatrao, the below is something we need to consider for the > >>combination of the NUMA & kexec approaches. It only becomes a problem > >>if/when we preserve DT memory nodes in the presence of EFI, though it > >>would be nice to not box ourselves into a corner. > >> > >>On Wed, Jan 20, 2016 at 12:02:58PM +0000, Mark Rutland wrote: > >>>On Wed, Jan 20, 2016 at 02:25:07PM +0900, AKASHI Takahiro wrote: > >>>>On 01/19/2016 11:01 PM, Mark Rutland wrote: > >>>>>For NUMA topology in !ACPI kernels, we might need to also retain and > >>>>>parse memory nodes, but only for toplogy information. The kernel would > >>>>>still only use memory as described by the EFI memory map. > >>>>> > >>>>>There's a horrible edge case I've spotted if performing a chain of > >>>>>cross-endian kexecs: LE -> BE -> LE, as the BE kernel would have to > >>>>>respect the EFI memory map so as to avoid corrupting it for the > >>>>>subsequent LE kernel. Other than this I believe everything should just > >>>>>work. > >>>> > >>>>BE kernel doesn't support UEFI yet and cannot access UEFI memmap table. So, > >>>>for LE -> BE, we don't use a dtb generated from /sys/firmware/fdt (or /proc/device-tree) > >>>>(as in the case of LE -> LE) and require users to provide a dtb file explicitly. > >>> > >>>As I mentioned above, the problem exists when memory nodes also exist > >>>(for describing NUMA topology). In that case the BE kernel would try to > >>>use the information from the memory nodes. > >>> > >>>>For BE -> LE, BE kernel doesn't know wther UEFI memmap table is available or not > >>>>and so use the same (explicitly-provided) dtb (as LE -> LE in !UEFI) > >>> > >>>See above. The problem I imagine is: > >>> > >>>LE kernel - uses EFI mmap, takes NUMA information from DT memory nodes > >>> > >>> v kexec > >>> > >>>BE kernel - uses DT memory nodes > >>> - clobbers EFI runtime regions as it sees them as available > >>> > >>> v kexec > >>> > >>>LE kernel - uses EFI mmap, takes NUMA information from DT memory nodes > >>> - tries to call EFI runtime services, and explodes. > >> > >>I'm not really sure what the best approach is here, but I thought that > >>it would be good to raise awareness of the edge-case. > >> > > > >I think we should simply allow the BE kernel to deal with a UEFI > >memory map. It only involves a bit of byte swapping (which I already > >implemented at some point) > > Just from my curiosity, > will runtime services be also available on BE kernel with LE uefi? It may be possible to implement that (I recall that Ard had a go), but that's far more complicated than simply supporting the EFI memory map, as you need separate (endian-swapped) page tables and other data structures, lose the ability to handle exceptions, etc. All that's suggested above is supporting the memory map. Thanks, Mark.
