Re: [PATCH v2] x86/Kconfig: Fix allyesconfig

From: Andrew Morton
Date: Wed Apr 16 2025 - 21:28:39 EST


On Wed, 16 Apr 2025 17:17:00 -0700 Guenter Roeck <linux@xxxxxxxxxxxx> wrote:

> On 4/16/25 17:03, Andrew Morton wrote:
> > On Wed, 16 Apr 2025 16:05:59 -0700 Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
> >
> >> 64-bit allyesconfig builds fail with
> >>
> >> x86_64-linux-ld: kernel image bigger than KERNEL_IMAGE_SIZE
> >>
> >> Bisect points to commit 6f110a5e4f99 ("Disable SLUB_TINY for build
> >> testing") as the responsible commit. Reverting that patch does indeed
> >> fix the problem. Further analysis shows that disabling SLUB_TINY enables
> >> KASAN, and that KASAN is responsible for the image size increase.
> >>
> >> Solve the build problem by disabling KASAN for test builds.
> >>
> >
> > Excluding KASAN from COMPILE_TEST builds is regrettable.
> >
> > Can we address this some other way? One way might be to alter or
> > disable the KERNEL_IMAGE_SIZE check if COMPILE_TEST? That will be sad
> > for anyone who tries to boot a COMPILE_TEST kernel, but who the heck
> > does that?
>
> I tried increasing the limit. It didn't work. With the RFC I sent earlier
> I made it dependent on allmodconfig, but Linus said I should just disable
> it for test builds (which was the cases anyway until commit 6f110a5e4f99).
>
> Personally I don't have a preference either way. I can also do nothing and
> stop testing allyesconfig. That would help reducing the load on my testbed,
> so I would be all for it.

How about this?



--- a/arch/x86/kernel/vmlinux.lds.S~a
+++ a/arch/x86/kernel/vmlinux.lds.S
@@ -466,10 +466,19 @@ SECTIONS
}

/*
- * The ASSERT() sink to . is intentional, for binutils 2.14 compatibility:
+ * COMPILE_TEST kernels can be large - CONFIG_KASAN, for example, can cause
+ * this. Let's assume that nobody will be running a COMPILE_TEST kernel and
+ * let's assert that fuller build coverage is more valuable than being able to
+ * run a COMPILE_TEST kernel.
+ */
+#ifndef CONFIG_COMPILE_TEST
+/*
+/*
+ * The ASSERT() sync to . is intentional, for binutils 2.14 compatibility:
*/
. = ASSERT((_end - LOAD_OFFSET <= KERNEL_IMAGE_SIZE),
"kernel image bigger than KERNEL_IMAGE_SIZE");
+#endif

/* needed for Clang - see arch/x86/entry/entry.S */
PROVIDE(__ref_stack_chk_guard = __stack_chk_guard);
_


(contains gratuitous s/sink/sync/)


I'd like to add

#else
WARN((_end - LOAD_OFFSET <= KERNEL_IMAGE_SIZE),
"kernel image bigger than KERNEL_IMAGE_SIZE - kernel probably will not work");
#endif /* CONFIG_COMPILE_TEST */

but I lack the patience to figure out how to do that.