Re: [PATCH v4 0/4] LoongArch: Support new relocation types

[Jinyang He <hejinyang@loongson.cn>]

On 08/01/2022 05:55 PM, Xi Ruoyao wrote:

> On Mon, 2022-08-01 at 10:34 +0800, Huacai Chen wrote:
>> Hi, all,
>>
>> On Mon, Aug 1, 2022 at 10:16 AM Youling Tang <tangyouling@loongson.cn>
>> wrote:
>>> Hi, Ruoyao
>>>
>>> On 07/30/2022 10:52 AM, Xi Ruoyao wrote:
>>>> On Sat, 2022-07-30 at 10:24 +0800, Xi Ruoyao wrote:
>>>>> On Sat, 2022-07-30 at 01:55 +0800, Xi Ruoyao wrote:
>>>>>> On Fri, 2022-07-29 at 20:19 +0800, Youling Tang wrote:
>>>>>>
>>>>>>> On 07/29/2022 07:45 PM, Xi Ruoyao wrote:
>>>>>>>> Hmm... The problem is the "addresses" of per-cpu symbols
>>>>>>>> are
>>>>>>>> faked: they
>>>>>>>> are actually offsets from $r21.  So we can't just load
>>>>>>>> such an
>>>>>>>> offset
>>>>>>>> with PCALA addressing.
>>>>>>>>
>>>>>>>> It looks like we'll need to introduce an attribute for GCC
>>>>>>>> to
>>>>>>>> make
>>>>>>>> an
>>>>>>>> variable "must be addressed via GOT", and add the
>>>>>>>> attribute into
>>>>>>>> PER_CPU_ATTRIBUTES.
>>>>>>> Yes, we need a GCC attribute to specify the per-cpu
>>>>>>> variable.
>>>>>> GCC patch adding "addr_global" attribute for LoongArch:
>>>>>> https://gcc.gnu.org/pipermail/gcc-patches/2022-July/599064.html
>>>>>>
>>>>>> An experiment to use it:
>>>>>> https://github.com/xry111/linux/commit/c1d5d70
>>>>> Correction: https://github.com/xry111/linux/commit/c1d5d708
>>>>>
>>>>> It seems 7-bit SHA is not enough for kernel repo.
>>>> If addr_global is rejected or not implemented (for example,
>>>> building the
>>>> kernel with GCC 12), *I expect* the following hack to work (I've
>>>> not
>>>> tested it because I'm AFK now).  Using visibility in kernel seems
>>>> strange, but I think it may make some sense because the modules
>>>> are some
>>>> sort of similar to an ELF shared object being dlopen()'ed, and our
>>>> way
>>>> to inject per-CPU symbols is analog to ELF interposition.
>>>>
>>>> arch/loongarch/include/asm/percpu.h:
>>>>
>>>>     #if !__has_attribute(__addr_global__) && defined(MODULE)
>>>>     /* Magically remove "static" for per-CPU variables.  */
>>>>     # define ARCH_NEEDS_WEAK_PER_CPU
>>>>     /* Force GOT-relocation for per-CPU variables.  */
>>>>     # define PER_CPU_ATTRIBUTES
>>>> __attribute__((__visibility__("default")))
>>>>     #endif
>>>>
>>>> arch/loongarch/Makefile:
>>>>
>>>>     # Hack for per-CPU variables, see PER_CPU_ATTRIBUTES in
>>>>     # include/asm/percpu.h
>>>>     if (call gcc-does-not-support-addr-global)
>>>>       KBUILD_CFLAGS_MODULE += -fPIC -fvisibility=hidden
>>>>     endif
>>>>
>>> Using the old toolchain (GCC 12) can successfully load the
>>> nf_tables.ko
>>> module after applying the above patch.
>> I don't like such a hack..., can we consider using old relocation
>> types when building by old toolchains?
>
> I don't like the hack too.  I only developed it as an intellectual game.
>
> We need to consider multiple combinations:
>
> (1) Old GCC + old Binutils.  We need -mla-local-with-abs for
> KBUILD_CFLAGS_MODULE.
>
> (2) Old GCC + new Binutils.  We need -mla-local-with-abs for
> KBUILD_CFLAGS_MODULE, *and* adding the support for
> R_LARCH_ABS{_HI20,_LO12,64_LO20,64_HI12} in the kernel module loader.
>
> (3) New GCC + old Binutils.  As new GCC should support our new attribute
> (I now intend to send V2 patch to gcc-patches using "movable" as the
> attribute name), no special action is needed.
>
> Basically, we need:
>
> (1) Handle R_LARCH_ABS{_HI20,_LO12,64_LO20,64_HI12} in the kernel module
> loader.
> (2) Add -Wa,-mla-local-with-abs into KBUILD_CFLAGS_MODULE if GCC version
> is <= 12.

Actually, I really hope kernel image is in the XKVRANGE, rather
than being in XKPRANGE. So that we can limit kernel and modules
be in 4GB range. I think it will make all work normally. :-(