[PATCH 00/12] ARM: use adr_l/ldr_l macros for PC-relative references

[Ard Biesheuvel]

On Tue, 15 Sep 2020 at 22:32, Nick Desaulniers <ndesaulniers at google.com> wrote:
>
> On Mon, Sep 14, 2020 at 2:57 AM Ard Biesheuvel <ardb at kernel.org> wrote:
> >
> > This is a respin of the adr_l/ldr_l code I wrote some years ago in the
> > context of my KASLR proof of concept for 32-bit ARM.
> >
> > A new use case came up, in the form of Clang, which does not implement
> > the 'adrl' pseudo-instruction in its assembler, and so for PC-relative
> > references that don't fit into a ARM adr instruction, we need something
> > else. Patch #2 addresses an actual Clang build issue of this nature, by
> > replacing an occurrence of adrl with adr_l.
> >
> > I have included my existing cleanup patches that were built on top of the
> > adr_l macro, which replace several occurrences of open coded arithmetic to
> > calculate runtime addresses based on link time virtual addresses stored
> > in literals.
> >
> > Note that all of these patches with the exception of #2 were reviewed or
> > acked by Nico before, but given that this was a while ago (and the fact
> > that neither of us work for Linaro anymore), I have dropped these. Note
> > that only patch #1 deviates significantly from the last version that I
> > sent out, the remaining ones were just freshened up (and their commit
> > logs slightly expanded).
>
> Tested-by: Nick Desaulniers <ndesaulniers at google.com>
>
> Thanks for the series, Ard.  I was able to compile and boot the
> following with this series (and the fixup to 01/12 applied):
>
> $ ARCH=arm CROSS_COMPILE=arm-linux-gnueabihf- make defconfig
> $ ARCH=arm CROSS_COMPILE=arm-linux-gnueabihf- make LLVM=1 defconfig
> $ ARCH=arm CROSS_COMPILE=arm-linux-gnueabihf- make LLVM=1 aspeed_g5_defconfig
> $ ARCH=arm CROSS_COMPILE=arm-linux-gnueabihf- make LLVM=1 multi_v5_defconfig
>
> (So ARM v7/GCC, ARM v5,6,7/LLVM).  (Technically, the v6 is not
> booting, but it's not related to this series and fails to boot without
> the series as well.  Our CI on -next is red for an unrelated issue
> which is masking the regression).
>

Excellent, thanks for testing. Do you have any coverage for Thumb2
builds as well? (CONFIG_THUMB2_KERNEL=y)

> I was also able to build+boot the defconfig with v5 and v7 with some
> configs disabled and a few hacks with LLVM_IAS=1.  This series allowed
> me to get further in the build/testing, and I have a few new bugs to
> go chase.  If anyone's interested:
> https://github.com/ClangBuiltLinux/linux/issues/1154
> https://github.com/ClangBuiltLinux/linux/issues/1155
> so we're still a handful of bugs away from LLVM_IAS=1 with ARCH=arm,
> but we're steadily chipping away at it, and this series is a big help.
> Lest it look like there's only kernel fixes in this area, Jian's
> https://reviews.llvm.org/D69411 recently was a big help, specifically
> for ARCH=arm LLVM_IAS=1.
> --
> Thanks,
> ~Nick Desaulniers