efi_loader/RISC-V: misaligned load when running grubriscv64.efi

[Heinrich Schuchardt]

Am 30. Juli 2020 20:31:47 MESZ schrieb Atish Patra <atishp at atishpatra.org>:
>On Thu, Jul 30, 2020 at 4:04 AM Heinrich Schuchardt
><xypron.glpk at gmx.de> wrote:
>>
>> On 30.07.20 12:16, Sean Anderson wrote:
>> > On 7/30/20 6:03 AM, Heinrich Schuchardt wrote:
>> >> Dear Sean,
>> >>
>> >> when trying to run grubriscv64.efi from the
>> >> trini/u-boot-gitlab-ci-runner:bionic-20200526-18Jun2020 Docker
>image on
>> >> a MAIXDUINO the relocations are not naturally aligned:
>> >>
>> >> lib/efi_loader/efi_image_loader.c(133) efi_loader_relocate():
>> >>
>> >>     efi_reloc 000000008030a000, offset 0x101e, type 10
>> >>
>> >> Here we are trying to change an u64 at 0x8030B01E:
>> >>
>> >>     uint64_t *x64 = efi_reloc + offset;
>> >>     *x64 += (uint64_t)delta;
>> >>
>> >> This leads to an exception in function efi_loader_relocate():
>> >>
>> >>     Unhandled exception: Load address misaligned
>> >>     EPC: 00000000805a95ac RA: 00000000805a953a TVAL:
>000000008030b01e
>> >>     EPC: 000000008001c5ac RA: 000000008001c53a reloc
>> >>
>> >> The GRUB image is available here:
>> >>
>> >>
>https://gist.github.com/xypron/522a91962248e9c3888d8554cb61ad2c/raw/b959661626b38a738673a9efb2f398b2fabd5c77/grubriscv64.efi
>> >>
>> >> On QEMU the GRUB image is executed without problems:
>> >>
>> >> https://gitlab.denx.de/u-boot/custodians/u-boot-efi/-/jobs/132919
>> >>
>> >> The UEFI specification requires for the ARM platform that
>unaligned
>> >> support is enabled. This is why we have implemented function
>> >> allow_unaligned().
>> >>
>> >> On RISC-V we have not yet implemented allow_unaligned() yet. Is
>there a
>> >> way to switch RISCV64 CPUs especially the Kendryte K210 into a
>mode
>> >> supporting unaligned access?
>> >
>> > AFAIK RISC-V has no requirement that un-aligned loads/stores
>complete. I
>> > believe the recommended solution is to install a trap handler which
>> > completes the un-aligned load through a series of aligned loads and
>then
>> > returns back to the application. For an example of such an
>> > implementation, check out arch/riscv/kernel/traps_misaligned.c in
>Linux.
>> > This may be too complex for U-Boot, so perhaps you can simply
>disallow
>> > unaligned accesses?
>> >
>> > --Sean
>> >
>>
>> Working around the problem inside U-Boot is easy (just some memcpy()
>> calls) but the GRUB image itself also makes unaligned accesses:
>>
>> Unhandled exception: Load address misaligned
>> EPC: 000000008030b004 RA: 00000000805a4eca TVAL: 000000008030b02e
>> EPC: 000000007fd7e004 RA: 0000000080017eca reloc
>>
>> UEFI image [0x000000008030a000:0x0000000080433fff] pc=0x1004
>>
>> This is what I found in "RISC-V Unprivileged ISA V20191213"
>>
>> "Loads  and  stores  where  the  effective  address  is  not 
>naturally
>> aligned to the referenced datatype (i.e., on a four-byte boundary for
>> 32-bit accesses, and a two-byte boundary for 16-bit accesses) have
>> behavior dependent on the EEI. An EEI may guarantee that misaligned
>> loads and stores are fully supported, and so the software running
>inside
>> the execution environment will never experience a contained or fatal
>> address-misaligned trap."
>>
>> @Leif
>> Should GRUB be built with -mstrict-align for RISC-V?
>>
>
>That shouldn't be necessary. Any real board with an MMU that can boot
>Linux needs
>a SBI provider such as OpenSBI. OpenSBI already implements a misaligned
>handler.
>
>Are we planning to support EFI booting for NoMMU platforms ? As per my
>understanding
>runtime services need to be mapped through kernel page tables.
>

My interest is to have an affordable hardware platform where I can test U-Boot's UEFI sub-system on RISC-V.

With 6 MiB usable RAM. (2 MiB reserved for AI) we probably won't get further than running GRUB.

Can OpenSBI be built for the Kendryte K210 SoC? What is the size of OpenSBI?

Best regards

Heinrich

>> @Ard
>> How about the EFI part of the Linux kernel?
>>
>> Best regards
>>
>> Heinrich