Re: [PATCH v6 3/7] Adjust symbol ordering in text output section
From: Rong Xu
Date: Tue Nov 12 2024 - 00:39:23 EST
I compared the System.map files from Klara Modin. The linker script is
doing what I expected: relocating the unlikely executed functions to the
beginning of the .text section.
However, the problem is with the _stext symbol. It belongs to the
.text section, so
it is positioned after the unlikely (or hot) functions. But it really
needs to be
the start of the text section.
I checked all vmlinux.lds.S in arch/, I found that most archs
explicitly assign _stext to the same address as _text, with the
following 3 exceptions:
arch/sh/kernel/vmlinux.lds.S
arch/mips/kernel/vmlinux.lds.S
arch/sparc/kernel/vmlinux.lds.S
Note that we already partially handled arch/sparc/kernel/vmlinux.lds.S
for sparc64.
But we need to handle sparc32 also.
Additionally, the boot/compressed/vmlinux.lds.S also the TEXT_TEXT
template. However,
I presume these files do not generate the .text.unlikely. or
.text.hot.* sections.
I sent the following patch to Klara because I don't have an
environment to build and test.
====================
diff --git a/arch/mips/kernel/vmlinux.lds.S b/arch/mips/kernel/vmlinux.lds.S
index 9ff55cb80a64..5f130af44247 100644
--- a/arch/mips/kernel/vmlinux.lds.S
+++ b/arch/mips/kernel/vmlinux.lds.S
@@ -61,6 +61,7 @@ SECTIONS
/* read-only */
_text = .; /* Text and read-only data */
.text : {
+ _stext = .;
TEXT_TEXT
SCHED_TEXT
LOCK_TEXT
======================
If Klara confirms the fix, I will send the patch for review.
Thanks,
-Rong
On Mon, Nov 11, 2024 at 2:39 PM Rong Xu <xur@xxxxxxxxxx> wrote:
>
> In the new System.map, we have:
> ffffffff81112400 T _stext
>
> This looks wrong. It should point to the beginning of the text, like
> ffffffff81100400 T _stext
>
> I'll do some debugging on this.
>
> -Rong
>
> On Mon, Nov 11, 2024 at 1:32 PM Klara Modin <klarasmodin@xxxxxxxxx> wrote:
> >
> > On 2024-11-11 21:43, Rong Xu wrote:
> > > Thanks for reporting this issue!
> > >
> > > I'm assuming your kernel build enables dead code elimination and
> > > uses the --ffunction-sections compiler flag. Without this patch, all
> > > the functions
> > > -- I think there are only .text.unlikely.* and .text.* are grouped
> > > together in the
> > > final vmlinux. This patch modifies the linker script to place
> > > .text.unlikely.* functions
> > > before .text.* functions. I've examined arch/mips/kernel/vmlinux.lds.S, and
> > > haven't found any obvious issue.
> > >
> > > Can you send me the following?
> > > (1) the kernel build command
> > > (2) System.map without the patch
> > > (3) System.map with the patch
> > >
> > > Best regards,
> > >
> > > -Rong
> > >
> > I don't set -ffunction-sections explicitly but it seems to be used when
> > I look at the .cmd files. The build command is nothing fancy, I just set
> > ARCH=mips CROSS_COMPILE=mips64-unknown-linux-gnuabin32- and build with
> > make -j24.
> >
> > I've attached the System.map, built on next-20241111 as well as it with
> > this series reverted.
> >
> > Regards,
> > Klara Modin