Re: [PATCH] powerpc: Fix kernel-jump address for ppc64 wrapper boot

From: Oliver O'Halloran
Date: Tue Jun 08 2021 - 03:29:52 EST


On Tue, Jun 8, 2021 at 4:33 PM He Ying <heying24@xxxxxxxxxx> wrote:
>
> Hello,
>
> 在 2021/6/8 13:26, Oliver O'Halloran 写道:
> > On Fri, Jun 4, 2021 at 7:39 PM He Ying <heying24@xxxxxxxxxx> wrote:
> >> From "64-bit PowerPC ELF Application Binary Interface Supplement 1.9",
> >> we know that the value of a function pointer in a language like C is
> >> the address of the function descriptor and the first doubleword
> >> of the function descriptor contains the address of the entry point
> >> of the function.
> >>
> >> So, when we want to jump to an address (e.g. addr) to execute for
> >> PPC-elf64abi, we should assign the address of addr *NOT* addr itself
> >> to the function pointer or system will jump to the wrong address.
> > How have you tested this?
>
> I tested ppc64-elf big-endian. I changed the Kconfig so that ppc64
> big-endian
>
> selects PPC64_WRAPPER_BOOT. I used qemu to run the cuImage and found
>
> the problem. It made me confused. By applying this patch, I found it works.
>
> I thought it works for ppc64le too. So I upstream this patch.
>
> >
> > IIRC the 64bit wrapper is only used for ppc64le builds. For that case
> > the current code is work because the LE ABI (ABIv2) doesn't use
> > function descriptors. I think even for a BE kernel we need the current
> > behaviour because the vmlinux's entry point is screwed up (i.e.
> > doesn't point a descriptor) and tools in the wild (probably kexec)
> > expect it to be screwed up.
>
> Yes, you're right. PPC64_WRAPPER_BOOT is only used for ppc64le builds
> currently.
>
> LE ABI (ABI v2) doesn't use function descriptors. Is that right? I don't
> test that. If so,
>
> this patch should be dropped. But why does ppc64 have different ABIs? So
> strange.

Yeah, it is strange. When LE support was added the toolchain team took
the opportunity to revamp the ABI since BE and LE binaries were never
going to be compatible. IIRC there is a slight performance advantage
to using v2 since function descriptors added an extra load when
performing a non-local function call. I think.

> If the wrapper is built to ppc64be, my patch is tested right. The entry
> point in the ELF
>
> header is always right so you can assign the header->e_entry to the
> function pointer
>
> and then jump to the entry by calling the function. But in the ppc
> wrapper, the address
>
> is intialized to 0 or malloced to be an address later. In this
> situation, I think my patch
>
> should be right for ppc64be.

Yeah maybe it's fine. I just have some memories of running into some
bizzare edge case at some point. It might have been the entrypoint of
the zImage rather than the vmlinux which had (has?) that problem.