Re: [PATCH] efi/libstub: Indicate clang the relocation mode for arm64

[Ard Biesheuvel]

(adding Arnd and Will to cc, who are likely to have an opinion as to
which GCC is the oldest we need to support for arm64)

On 10 May 2017 at 20:47, Matthias Kaehlcke <mka@xxxxxxxxxxxx> wrote:
> El Wed, May 10, 2017 at 09:05:28PM +0200 Ard Biesheuvel ha dit:
>
>>
>>
>> > On 10 May 2017, at 20:38, Matthias Kaehlcke <mka@xxxxxxxxxxxx> wrote:
>> >
>> > Hoi Ard,
>> >
>> > El Wed, May 10, 2017 at 08:51:44AM +0100 Ard Biesheuvel ha dit:
>> >
>> >> On 9 May 2017 at 22:49, Matthias Kaehlcke <mka@xxxxxxxxxxxx> wrote:
>> >>> El Tue, May 09, 2017 at 01:50:36PM -0700 Greg Hackmann ha dit:
>> >>>
>> >>>> On 05/09/2017 12:36 PM, Matthias Kaehlcke wrote:
>> >>>>> From: Greg Hackmann <ghackmann@xxxxxxxxxx>
>> >>>>>
>> >>>>> Without any extra guidance, clang will generate libstub with either
>> >>>>> absolute or relative ELF relocations. Use the right combination of
>> >>>>> -fpic and -fno-pic on different files to avoid this.
>> >>>>>
>> >>>>> Signed-off-by: Greg Hackmann <ghackmann@xxxxxxxxxx>
>> >>>>> Signed-off-by: Bernhard RosenkrÃnzer <Bernhard.Rosenkranzer@xxxxxxxxxx>
>> >>>>> Signed-off-by: Matthias Kaehlcke <mka@xxxxxxxxxxxx>
>> >>>>> ---
>> >>>>> drivers/firmware/efi/libstub/Makefile | 6 ++++++
>> >>>>> 1 file changed, 6 insertions(+)
>> >>>>>
>> >>>>> diff --git a/drivers/firmware/efi/libstub/Makefile b/drivers/firmware/efi/libstub/Makefile
>> >>>>> index f7425960f6a5..ccbaaf4d8650 100644
>> >>>>> --- a/drivers/firmware/efi/libstub/Makefile
>> >>>>> +++ b/drivers/firmware/efi/libstub/Makefile
>> >>>>> @@ -11,6 +11,9 @@ cflags-$(CONFIG_X86)               += -m$(BITS) -D__KERNEL__ -O2 \
>> >>>>>                               -mno-mmx -mno-sse
>> >>>>>
>> >>>>> cflags-$(CONFIG_ARM64)              := $(subst -pg,,$(KBUILD_CFLAGS))
>> >>>>> +ifeq ($(cc-name),clang)
>> >>>>> +cflags-$(CONFIG_ARM64)              += -fpic
>> >>>>> +endif
>> >>>>> cflags-$(CONFIG_ARM)                := $(subst -pg,,$(KBUILD_CFLAGS)) \
>> >>>>>                               -fno-builtin -fpic -mno-single-pic-base
>> >>>>>
>> >>>>> @@ -38,6 +41,9 @@ $(obj)/lib-%.o: $(srctree)/lib/%.c FORCE
>> >>>>>
>> >>>>> lib-$(CONFIG_EFI_ARMSTUB)   += arm-stub.o fdt.o string.o random.o \
>> >>>>>                               $(patsubst %.c,lib-%.o,$(arm-deps))
>> >>>>> +ifeq ($(cc-name),clang)
>> >>>>> +CFLAGS_arm64-stub.o            += -fno-pic
>> >>>>> +endif
>> >>>>>
>> >>>>> lib-$(CONFIG_ARM)           += arm32-stub.o
>> >>>>> lib-$(CONFIG_ARM64)         += arm64-stub.o
>> >>>>>
>> >>>>
>> >>>> NAK.
>> >>>>
>> >>>> This patch was labeled "HACK:" in our experimental tree.  There's no
>> >>>> rhyme or reason to why this combination of -f[no-]pic flags
>> >>>> generates code without problematic relocations.  It's inherently
>> >>>> fragile, and was only intended as a temporary workaround until I (or
>> >>>> someone more familiar with EFI) got a chance to revisit the problem.
>> >>>>
>> >>>> Unless the gcc CFLAGS are also an artifact of "mess with -f[no-]pic
>> >>>> until the compiler generates what you want", this doesn't belong
>> >>>> upstream.
>> >>>
>> >>> Sorry, I didn't realize it is that bad of a hack. Unfortunately I'm
>> >>> not very familiar with EFI either.
>> >>>
>> >>> I saw Ard did some work in this code related with relocation, maybe he
>> >>> can provide a pointer towards a better solution.
>> >>>
>> >>
>> >> This is a known issue. The problem is that generic AArch64 small model
>> >> code is mostly position independent already, due to its use of
>> >> adrp/add pairs to generate symbol references with a +/- 4 GB range.
>> >> Building the same code with -fpic will result in GOT entries to be
>> >> generated, which carry absolute addresses, so this achieves the exact
>> >> opposite of what we want.
>> >>
>> >> The reason for the GOT entries is that GCC (and Clang, apparently)
>> >> infer from the -fpic flag that you are building objects that will be
>> >> linked into a shared library, to which ELF symbol preemption rules
>> >> apply that stipulate that a symbol in the main executable supersedes a
>> >> symbol under the same name in the shared library, and that the shared
>> >> library should update all its internal references to the main
>> >> executable's version of the symbol. The easiest way (but certainly not
>> >> the only way) to achieve that is to indirect all internal symbol
>> >> references via GOT entries, which can be made to refer to another
>> >> symbol by updating a single value.
>> >>
>> >> The workaround I used is to use hidden visibility, using a #pragma.
>> >> (There is a -fvisibility=hidden command line option as well, but this
>> >> is a weaker form that does not apply to extern declarations, only to
>> >> definitions). So if you add
>> >>
>> >> #pragma GCC visibility push(hidden)
>> >>
>> >> at the beginning of arm64-stub.c (and perhaps to one or two other
>> >> files that contain externally visible symbol declarations these days),
>> >> you should be able to compile the entire EFI stub with -fpic. Note
>> >> that making those externally visible symbols 'static' where possible
>> >> would solve the problem as well, but this triggers another issue in
>> >> the 32-bit ARM stub.
>> >>
>> >> In my opinion, the correct fix would be to make -fpie (as opposed to
>> >> -fpic) imply hidden visibility, given that PIE executables don't
>> >> export symbols in the first place, and so the preemption rules do not
>> >> apply. It is worth a try whether -fpie works as expected in this case
>> >> on Clang, but the last time I tried it on GCC, it behaved exactly like
>> >> -fpic.
>> >
>> > Thanks a lot for the detailed description and your suggestions!
>> >
>> > A clang build with -fpie for the EFI stub succeeds without complaints
>> > about GOT entries. I will send out an updated patch (with -fpie only
>> > for clang) later.
>> >
>>
>> Good! I never liked the visibility hack, which is why I never upstreamed it.
>>
>> Could you please check how recent GCC behaves?
>
> I tried GCC v4.9.4 and v6.3.1, both build the EFI stub with -fpie
> without errors.
>
> Are you suggesting to use -fpie for both clang and GCC? Do you know
> what the minimum required GCC version is for building an arm64 kernel?

Yes. Up until now, we have been relying on the position independent
nature of small model code, but it would be better to specify it
explicitly, so if -fpie gives us mostly identical code and does not
need visibility hacks, I would prefer to add it for all compilers and
not have an exception only for Clang. Note that the same applies to
the entire kernel when built in KASLR mode, so it would also be good
to know our options here.

Arnd, Will, what is the oldest GCC version we claim to support for arm64?