Re: [PATCH 3/7] [HACK] x86: crypto: fix link error with LTO

From: Arnd Bergmann
Date: Fri Feb 02 2018 - 17:18:28 EST


On Fri, Feb 2, 2018 at 8:49 PM, Nicolas Pitre <nicolas.pitre@xxxxxxxxxx> wrote:
> On Fri, 2 Feb 2018, Arnd Bergmann wrote:
>
>> crypto_it_tab and the other symbols like it are defined in
>> crypto/aes_generic.c and exported for loadable modules. When
>> building with LTO and CONFIG_TRIM_UNUSED_KSYMS, the exports
>> are eliminated, since kbuild fails to take the users in
>> the arch/x86/crypto/aes-i586-asm_32.S assembler file into
>> account.
>>
>> This adds an ugly workaround by adding a reference to each symbol
>> into aes_glue.c, which gets linked together with the assembler
>> file.
>>
>> We obviously want to fix the CONFIG_TRIM_UNUSED_KSYMS logic
>> to do the right thing here instead, but I couldn't come up
>> with a good fix, so I use this instead to get a clean build
>> for testing.
>
> Could you give me the .config you used for this? So far I can't
> reproduce, and guessing with LTO takes lots of time.

I've uploaded my randconfig file to

https://pastebin.com/raw/iLk4sSpw

>> This fix only works most of the time, but I still ran into
>> some cases where combining an .S with a .c file did not
>> produce the correct .ko file, as the lto linker apparently
>> did not expect that kind of input. 'nm' on the file after
>> 'ld -r' showed only the contents of the assembler file, and
>> after the lto-ld stage, only the contents of the .c file
>> are there.
>
> 'ld -r' does not support mixed (LTO and non-LTO) objects.

Ok, that would explain it. I checked that CONFIG_THIN_ARCHIVES
is set here, but I see clearly that the module is linked using

x86_64-linux-ld -m elf_x86_64 -r -o arch/x86/crypto/aes-x86_64.o
arch/x86/crypto/aes-x86_64-asm_64.o arch/x86/crypto/aes_glue.o ;
scripts/mod/modpost arch/x86/crypto/aes-x86_64.o

Arnd