On Wed, Jan 06, 2021 at 10:36:38PM +0000, Alexander Lobakin wrote:
> From: Kees Cook <keescook@xxxxxxxxxxxx>
> Date: Wed, 6 Jan 2021 14:07:07 -0800
>
> > On Wed, Jan 06, 2021 at 08:08:19PM +0000, Alexander Lobakin wrote:
> >> Discard GNU attributes at link time as kernel doesn't use it at all.
> >> Solves a dozen of the following ld warnings (one per every file):
> >>
> >> mips-alpine-linux-musl-ld: warning: orphan section `.gnu.attributes'
> >> from `arch/mips/kernel/head.o' being placed in section
> >> `.gnu.attributes'
> >> mips-alpine-linux-musl-ld: warning: orphan section `.gnu.attributes'
> >> from `init/main.o' being placed in section `.gnu.attributes'
> >>
> >> Misc: sort DISCARDS section entries alphabetically.
> >
> > Hmm, I wonder what is causing the appearance of .eh_frame? With help I
> > tracked down all the causes of this on x86, arm, and arm64, so that's
> > why it's not in the asm-generic DISCARDS section. I suspect this could
> > be cleaned up for mips too?
>
> I could take a look and hunt it down. Could you please give some refs on
> what were the causes and solutions for the mentioned architectures?
Sure! Here are the ones I could find again:
34b4a5c54c42 ("arm64/kernel: Remove needless Call Frame Information annotations")
6e0a66d10c5b ("arm64/build: Remove .eh_frame* sections due to unwind tables")
d1c0272bc1c0 ("x86/boot/compressed: Remove, discard, or assert for unwanted sections")
> > Similarly for .gnu.attributes. What is generating that? (Or, more
> > specifically, why is it both being generated AND discarded?)
>
> On my setup, GNU Attributes consist of MIPS FP type (soft) and
> (if I'm correct) MIPS GNU Hash tables.
Ah, right, the soft-float markings sound correct to discard, IIUC.
> By the way. I've built the kernel with LLVM stack (and found several
> subjects for more patches) and, besides '.got', also got a fistful
> of '.data..compoundliteral*' symbols (drivers/mtd/nand/spi/,
> net/ipv6/ etc). Where should they be placed (rodata, rwdata, ...)
> or they are anomalies of some kind and should be fixed somehow?
Ah yeah, I've seen this before:
https://lore.kernel.org/lkml/202010051345.2Q0cvqdM-lkp@xxxxxxxxx/
https://lore.kernel.org/lkml/CAKwvOd=s53vUELe311VSjxt2_eQd+RGNCf__n+cV+R=PQ_CdXQ@xxxxxxxxxxxxxx/
And it looks like LTO trips over it too:
https://lore.kernel.org/lkml/20201211184633.3213045-3-samitolvanen@xxxxxxxxxx/
So I think the correct solution is to follow Sami's patch and add it to
vmlinux.lds.h:
-#define DATA_MAIN .data .data.[0-9a-zA-Z_]* .data..LPBX*
+#define DATA_MAIN .data .data.[0-9a-zA-Z_]* .data..L* .data..compoundliteral*
...
-#define RODATA_MAIN .rodata .rodata.[0-9a-zA-Z_]*
-#define BSS_MAIN .bss .bss.[0-9a-zA-Z_]*
+#define RODATA_MAIN .rodata .rodata.[0-9a-zA-Z_]* .rodata..L*
+#define BSS_MAIN .bss .bss.[0-9a-zA-Z_]* .bss..compoundliteral*
Can you include a patch for this in your series?
Thanks!
--
Kees Cook
