RE: [PATCH 1/1] scripts: kstack_erase: use relative stackleak plugin path

[Jaihind Yadav]

Hi Nathan And Nicolas ,

Thanks for the feedback.

I took another look at this and experimented with a different approach for external modules:

diff --git a/scripts/Makefile.kstack_erase b/scripts/Makefile.kstack_erase
index ee7e4ef7b892..xxxxxxxxxxxx 100644
--- a/scripts/Makefile.kstack_erase
+++ b/scripts/Makefile.kstack_erase
@@ -5,6 +5,10 @@ kstack-erase-cflags-y += -fplugin=$(objtree)/scripts/gcc-plugins/stackleak_plugin.so
 kstack-erase-cflags-y += -fplugin-arg-stackleak_plugin-track-min-size=$(CONFIG_KSTACK_ERASE_TRACK_MIN_SIZE)
 kstack-erase-cflags-y += -fplugin-arg-stackleak_plugin-arch=$(SRCARCH)
 kstack-erase-cflags-$(CONFIG_GCC_PLUGIN_STACKLEAK_VERBOSE) += -fplugin-arg-stackleak_plugin-verbose
+ifneq ($(KBUILD_EXTMOD),)
+# Avoid embedding absolute -fplugin paths into external module DWARF metadata.
+kstack-erase-cflags-y += -gno-record-gcc-switches
+endif
 DISABLE_KSTACK_ERASE := -fplugin-arg-stackleak_plugin-disable
 endif

In my testing, this prevents the absolute stackleak plugin path specified via

    -fplugin=$(objtree)/scripts/gcc-plugins/stackleak_plugin.so

from being recorded in the DWARF information of out-of-tree modules, while avoiding the use of platform-specific utilities such as realpath/readlink --relative-to.

I agree this does not solve the broader issue for all GCC plugins, but it appears to address the specific KSTACK_ERASE case that originally motivated the discussion.

Would this direction be more acceptable than the previous relative-path approach, or would suppressing GCC switch recording for external modules be considered undesirable from a debugging-information perspective?

Thanks,
Jaihind

-----Original Message-----
From: Nathan Chancellor <nathan@xxxxxxxxxx> 
Sent: Monday, August 3, 2026 11:42 PM
To: Nicolas Schier <nsc@xxxxxxxxxx>
Cc: Jaihind Yadav <jaihindy@xxxxxxxxxxxxxxxx>; kees@xxxxxxxxxx; linux-kbuild@xxxxxxxxxxxxxxx; linux-hardening@xxxxxxxxxxxxxxx; gustavoars@xxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
Subject: Re: [PATCH 1/1] scripts: kstack_erase: use relative stackleak plugin path

WARNING: This email originated from outside of Qualcomm. Please be wary of any links or attachments, and do not enable macros.

On Fri, Jul 24, 2026 at 10:18:57PM +0200, Nicolas Schier wrote:
> On Wed, Jul 22, 2026 at 07:00:43PM +0000, Jaihind Yadav wrote:
> >
> > Hi Nicolas Schier,
> >
> > Thanks for the suggestion. I tested the proposed change:
> >
> > -kstack-erase-cflags-y += 
> > -fplugin=$(objtree)/scripts/gcc-plugins/stackleak_plugin.so
> > +kstack-erase-cflags-y += 
> > +-fplugin=scripts/gcc-plugins/stackleak_plugin.so
> >
> >
> > Unfortunately, it does not work for the out-of-tree module builds in 
> > my environment.
>
> Thanks for checking; and yes I forgot out-of-tree modules.  When 
> building them, Kbuild changes to their respective output directory so 
> my suggestion had to fail for oot kmods.

Technically, this is only a problem for out of tree kernel modules, right? When building in-tree kernel modules, objtree is always '.', so we will never have the output directory path in the debug information (at least based on my brief testing).

> > Given that the proposed relative path breaks out-of-tree module 
> > builds, could you please take a look at the patch I posted? If the 
> > approach looks reasonable, I would appreciate your review and 
> > consideration for merging it.
>
> I am not sure, if we can take your path, as 'readlink --relative-to' 
> is only available in coreutils/Linux but coreutils are not a 
> documented requirement for Linux builds (cp. 
> Documentation/process/changes.rst),
> thus this would probably kill Linux builds on BSDs and Darwin/MacOS.
>
> But let me think about that a few days;  and perhaps someone else 
> comes up with an idea.

Yeah, I don't really like 'realpath --relative-to' either, it feels a bit like a hack since it will only avoid including a full path if the module output directory includes part of the build directory. For
example:

  O=/mnt/build/kernel
  MO=/mnt/build/extmod

$ realpath -m --relative-to /mnt/build/extmod /mnt/build/kernel ../kernel

vs.

  O=/home/nathan/build/kernel
  MO=/mnt/build/extmod

$ realpath -m --relative-to /mnt/build/extmod /home/nathan/build/kernel ../../../home/nathan/build/kernel

Furthermore, this patch only fixes KSTACK_ERASE but we have other GCC plugins that presumably have the same problem.

I do not have a good solution for this either but since it only affects external kernel modules, I am not particularly inclined to think too hard about it.

--
Cheers,
Nathan