On Sat, 26 Jan 2019 at 15:33, Heinrich Schuchardt <xypron.glpk@xxxxxx> wrote: > > On 1/26/19 1:34 PM, Alexander Graf wrote: > > > > > >> Am 26.01.2019 um 13:28 schrieb Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx>: > >> > >>> On Sat, 26 Jan 2019 at 13:27, Heinrich Schuchardt <xypron.glpk@xxxxxx> wrote: > >>> > >>>> On 1/26/19 11:22 AM, Ard Biesheuvel wrote: > >>>> The UEFI spec revision 2.7 errata A section 8.4 has the following to > >>>> say about the virtual memory runtime services: > >>>> > >>>> "This section contains function definitions for the virtual memory > >>>> support that may be optionally used by an operating system at runtime. > >>>> If an operating system chooses to make EFI runtime service calls in a > >>>> virtual addressing mode instead of the flat physical mode, then the > >>>> operating system must use the services in this section to switch the > >>>> EFI runtime services from flat physical addressing to virtual > >>>> addressing." > >>>> > >>>> So it is pretty clear that calling SetVirtualAddressMap() is entirely > >>>> optional, and so there is no point in doing so unless it achieves > >>>> anything useful for us. > >>>> > >>>> This is not the case for 64-bit ARM. The native mapping used by the OS > >>>> is arbitrarily converted into another permutation of userland addresses > >>>> (i.e., bits [63:48] cleared), and the runtime code could easily deal > >>>> with the original layout in exactly the same way as it deals with the > >>>> converted layout. However, due to constraints related to page size > >>>> differences if the OS is not running with 4k pages, and related to > >>>> systems that may expose the individual sections of PE/COFF runtime > >>>> modules as different memory regions, creating the virtual layout is a > >>>> bit fiddly, and requires us to sort the memory map and reason about > >>>> adjacent regions with identical memory types etc etc. > >>>> > >>>> So the obvious fix is to stop calling SetVirtualAddressMap() altogether > >>>> on arm64 systems. However, to avoid surprises, which are notoriously > >>>> hard to diagnose when it comes to OS<->firmware interactions, let's > >>>> start by making it an opt-out feature, and implement support for the > >>>> 'efi=novamap' kernel command line parameter on ARM and arm64 systems. > >>>> > >>>> (Note that 32-bit ARM generally does require SetVirtualAddressMap() to be > >>>> used, given that the physical memory map and the kernel virtual address > >>>> map are not guaranteed to be non-overlapping like on arm64. However, > >>>> having support for efi=novamap,noruntime on 32-bit ARM, combined with > >>>> the recently proposed support for earlycon=efi, is likely to be useful > >>>> to diagnose boot issues on such systems if they have no accessible serial > >>>> port) > >>>> > >>> > >>> NAK > >>> > >> > >> Excuse me? > >> > >>> This patch breaks EFI booting with any known U-Boot release. > >>> > >> > >> It does if you pass 'efi=novmap'. Otherwise, it works fine. > > It think it would be helpful to add this information to the commit message. > Did you read the commit message? > If it is strictly opt-in, I have no concern. > Thanks.
