Re: [PATCH v7 0/5] LoongArch: Support toolchain with new relocation types

From: Xi Ruoyao <xry111@xry111.site>
To: Huacai Chen <chenhuacai@kernel.org>
Cc: Jinyang He <hejinyang@loongson.cn>,
	WANG Xuerui <kernel@xen0n.name>,
	loongarch@lists.linux.dev, LKML <linux-kernel@vger.kernel.org>,
	Youling Tang <tangyouling@loongson.cn>
Subject: Re: [PATCH v7 0/5] LoongArch: Support toolchain with new relocation types
Date: Wed, 31 Aug 2022 23:14:55 +0800	[thread overview]
Message-ID: <97291c0fe5a660c844475ff019c8db6af77ecf86.camel@xry111.site> (raw)
In-Reply-To: <CAAhV-H4N_XvmP9KA1M5crU44kHr33MZUVSsMY4Ugu5wQSv_LOQ@mail.gmail.com>

On Wed, 2022-08-31 at 22:40 +0800, Huacai Chen wrote:
> On Wed, Aug 31, 2022 at 4:09 PM Xi Ruoyao <xry111@xry111.site> wrote:
> > 
> > On Wed, 2022-08-31 at 14:58 +0800, Jinyang He wrote:
> > > That's right. Also I am wondering why new toolchain produce .got* in
> > > kernel. It's unneeded. In the past, gcc create la.global and parsed
> > > to la.pcrel by gas, and kernel works well. Now it seems we lost this
> > > feature in gcc. I checked the x86 asm code just now. And some info
> > > follows,
> > > 
> > > LoongArch64, ./net/ipv4/udp_diag.s, *have reloc hint*
> > >           pcalau12i       $r4,%got_pc_hi20(udplite_table)
> > >           ld.d    $r4,$r4,%got_pc_lo12(udplite_table)
> > >           b       udp_dump
> > > 
> > > x86_64, ./net/ipv4/udp_diag.s
> > >           movq    $udplite_table, %rdi
> > >           jmp     udp_dump
> > > 
> > > It seems related to -fno-PIE and -cmodel=kernel on x86_64.
> > > Hope new gcc with this feature now.
> > 
> > On x86_64 -mcmodel=kernel means "all code and data are located in [-
> > 2GiB, 0) range.  We actually don't strictly require a "high" range as
> > we're mostly a PIC-friendly architecture: note that we use a
> > pcalau12i/addi.d pair for PIC addressing in [PC-2GiB, PC+2GiB, and a
> > lu12i.w/addi.d pair for "non-PIC" addressing in [-2GiB, 2GiB), both are
> > 2-insn sequence.
> > 
> > If we can put the main kernel image and the modules in one 2GiB VA
> > range, we can avoid GOT completely.  But it's not possible for now
> > because main kernel image is loaded in XKPRANGE but the modules are in
> > XKVRANGE.  So the best we can achieve before implementing
> > CONFIG_RELOCATION is using GOT in modules, and avoid GOT in the main
> > kernel image (with a new code model in GCC, which will benefit both the
> > kernel and statically linked executables).

> Emmm, can you implement this new code model in the near future?

I have a plan to make our toolchain addressing the symbols better:

(1) https://sourceware.org/pipermail/binutils/2022-August/122682.html. 
This change will allow the linker to link a main executable image
(dynamically linked or statically linked, PIE or non-PIE, kernel or
userspace) with R_LARCH_COPY instead of GOT.  (Note that R_LARCH_COPY
will not show up in the kernel because we don't link to shared objects,
but GOT will be gone.)

(2) Change GCC to stop using GOT unless -fPIC.  (Technically it's a one-
line change.)

(3) In kernel, for main kernel image the default of toolchain will be
good enough (no GOT).  For modules we have two options:

  (a) get rid of XKPRANGE.
  (b) force -mcmodel=extreme globally.
  (c) use -Wl,nocopyreloc to produce GOT.

(a) is the best, the performance of (b) and (c) will be worse than (a).
I'm not sure which one in (b) and (c) is better, but as (a) will be the
final solution we can just choose one in (b) and (c) "randomly" for now.

I don't want to add a new code model now, because if (1) works fine
we'll not need a new code model.  (1) is also the most tricky step in
the plan (I've sent the patch but not sure if it's completely correct),
(2) and (3) should be trivial.
-- 
Xi Ruoyao <xry111@xry111.site>
School of Aerospace Science and Technology, Xidian University