Re: [Buildroot] [PATCH] package/libbpf: Don't remove bpf.h on host

[Arnout Vandecappelle <arnout@mind.be>]

On 13/02/2023 13:24, Lang Daniel wrote:
> Hi Arnout, Francis,
[snip]
>> Therefore, we believe that the solution is to use -isystem instead of -I in the
>> HOSTCC setting of the kernel. This is indeed what is already done for uboot:
>>
>> 	HOSTCC="$(HOSTCC) $(subst -I/,-isystem /,$(subst -I /,-isystem
>> /,$(HOST_CFLAGS)))" \
>>
>> Note that there is a risk that this by itself goes wrong as well. We tried at
>> some point to use -isystem instead of -I in HOST_CFLAGS:
>>
>> commit 6f8162cf8c1abef7e0a4771fe0d6b26a28f5c2b6
>> Author: David Raeman <draeman@bbn.com>
>> Date:   Mon Jul 25 21:52:26 2016
>>
>>       package/Makefile.in should grab HOST_DIR headers using -isystem instead of -I.
>>
>>
>> but this was reverted a copule of days later:
>>
>> commit 255b6f80d395ef048f46cfcf75dba690c56af657
>> Author: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
>> Date:   Sat Jul 30 18:10:18 2016
>>
>>       Revert "package/Makefile.in should grab HOST_DIR headers using -isystem
>> instead of -I."
>>
>> (Unfortunately, the commit has no further explanation of what went wrong).
>>
>>    So, could you (Daniel and Francis) try to apply this patch and replace -I with
>> -isystem like is done for U-Boot, and do a bunch of kernel builds to see if it
>> breaks anything?
> 
> I didn't apply the patch for package/Makefile.in as that would change the behavour
> for all packages.

  Yeah, that's the thing that turned out not to work before.

> I did however copy the logic used in the U-Boot.
> 
> -HOSTCC="$(HOSTCC) $(HOST_CFLAGS) $(HOST_LDFLAGS)" \
> +HOSTCC="$(HOSTCC) $(subst -I/,-isystem /,$(subst -I /,-isystem /,$(HOST_CFLAGS))) $(HOST_LDFLAGS)" \

  That was indeed what I meant.

> 
> And applied this patch to keep bpf.h
> 
> I tested the following config:
> 
> BR2_aarch64=y
> BR2_TOOLCHAIN_EXTERNAL=y
> BR2_TOOLCHAIN_EXTERNAL_DOWNLOAD=y
> BR2_TOOLCHAIN_EXTERNAL_BOOTLIN=y
> BR2_TOOLCHAIN_EXTERNAL_BOOTLIN_AARCH64_GLIBC_STABLE=y
> BR2_LINUX_KERNEL=y
> BR2_LINUX_KERNEL_CUSTOM_VERSION=y
> BR2_LINUX_KERNEL_CUSTOM_VERSION_VALUE="{VERSION}"
> BR2_LINUX_KERNEL_USE_ARCH_DEFAULT_CONFIG=y
> BR2_LINUX_KERNEL_CONFIG_FRAGMENT_FILES="pahole-kernel.config"
> BR2_LINUX_KERNEL_NEEDS_HOST_OPENSSL=y
> BR2_LINUX_KERNEL_NEEDS_HOST_PAHOLE=y
> 
> and with a fragment for linux:
> 
> CONFIG_BPF_SYSCALL=y
> CONFIG_BPF_UNPRIV_DEFAULT_OFF=n
> CONFIG_DEBUG_INFO_REDUCED=n
> CONFIG_DEBUG_INFO_COMPRESSED=n
> CONFIG_DEBUG_INFO_BTF=y
> CONFIG_SYSTEM_TRUSTED_KEYRING=y
> 
> where VERSION is one of:
> 5.2.21 5.3.18 5.4.231 5.5.19 5.6.19 5.7.19 5.8.18 5.9.16 5.10.167 5.11.22
> 5.12.19 5.13.19 5.14.21 5.15.93 5.16.20 5.17.15 5.18.19 5.19.17 6.0.19 6.1.11
> 
> Version 5.2, as far as I could work it out, is the version that introduced
> the pahole dependency when CONFIG_DEBUG_INFO_BTF is set.
> 
> None-LTS versions after 5.10 and before 6.0 fail with:
> 
>    LD      vmlinux.o
>    MODPOST vmlinux.symvers
>    MODINFO modules.builtin.modinfo
>    GEN     modules.builtin
>    LD      .tmp_vmlinux.btf
>    BTF     .btf.vmlinux.bin.o
>    LD      .tmp_vmlinux.kallsyms1
>    KSYMS   .tmp_vmlinux.kallsyms1.S
>    AS      .tmp_vmlinux.kallsyms1.S
>    LD      .tmp_vmlinux.kallsyms2
>    KSYMS   .tmp_vmlinux.kallsyms2.S
>    AS      .tmp_vmlinux.kallsyms2.S
>    LD      vmlinux
>    BTFIDS  vmlinux
> FAILED: load BTF from vmlinux: Invalid argument
> make[2]: *** [Makefile:1177: vmlinux] Error 255
> make[1]: *** [package/pkg-generic.mk:293: /home/d.lang/ws/other/buildroot/output/build/linux-5.11.22/.stamp_built] Error 2
> make: *** [Makefile:82: _all] Error 2
> 
> But that should be unrelated to the change in package/libbpf and HOSTCC.
> I will try to work out the reason behind that error.

  Since you said "non-LTS versions", I guess it's an upstream issue that got 
fixed in 6.1 and applied to 6.0.x and 5.15.x?


> Should I create a patch to change HOSTCC or is additional testing required?

  Yes please! And include your extensive test coverage in the commit message.

  Regards,
  Arnout

> 
>>
>>    For now, I've marked this patch as Changes Requested (it can only be applied
>> if the bpf.h problem is solved in some other way), and marked Francis's patch
>> [1] as Superseded.
>>
>>    Regards,
>>    Arnout
>>
>>
>> [1]
>> https://patchwork.ozlabs.org/project/buildroot/patch/20220610165441.84812-2-flaniel@linux.microsoft.com/
>>
>>
>>>
>>> Sadly, I tested your patch and I hit the problem described in the comment
>>> regarding redefinition of bpf_core_relo_kind.
>>>
>>> Did you also hit it or you were able to build the kernel with
>>> BR2_LINUX_KERNEL_NEEDS_HOST_PAHOLE?
> 
> It seems like I was tested with a too recent kernel and originally preparing
> this patch and therefore didn't hit this error.
> 
>>>
>>>
>>> Best regards and thank you in advance.
>>>
>>>
>>> _______________________________________________
>>> buildroot mailing list
>>> buildroot@buildroot.org
>>> https://lists.buildroot.org/mailman/listinfo/buildroot
_______________________________________________
buildroot mailing list
buildroot@buildroot.org
https://lists.buildroot.org/mailman/listinfo/buildroot