Re: [PATCH v3 bpf] kbuild: fix dependencies for DEBUG_INFO_BTF

[Slava Bacherikov]

01.04.2020 20:46, Andrii Nakryiko ÐÐÑÐÑ:
> On Wed, Apr 1, 2020 at 7:38 AM Slava Bacherikov <slava@xxxxxxxxxxxx> wrote:
>>
>>
>>
>> 01.04.2020 17:20, Slava Bacherikov wrotes:
>>> Currently turning on DEBUG_INFO_SPLIT when DEBUG_INFO_BTF is also
>>> enabled will produce invalid btf file, since gen_btf function in
>>> link-vmlinux.sh script doesn't handle *.dwo files.
>>>
>>> Enabling DEBUG_INFO_REDUCED will also produce invalid btf file, and
>>> using GCC_PLUGIN_RANDSTRUCT with BTF makes no sense.
>>>
>>> Signed-off-by: Slava Bacherikov <slava@xxxxxxxxxxxx>
>>> Reported-by: Jann Horn <jannh@xxxxxxxxxx>
>>> Reported-by: Liu Yiding <liuyd.fnst@xxxxxxxxxxxxxx>
>>> Acked-by: KP Singh <kpsingh@xxxxxxxxxx>
>>> Acked-by: Andrii Nakryiko <andriin@xxxxxx>
>>> Fixes: e83b9f55448a ("kbuild: add ability to generate BTF type info for vmlinux")
>>> ---
>>>  lib/Kconfig.debug | 4 +++-
>>>  1 file changed, 3 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug
>>> index f61d834e02fe..b94227be2d62 100644
>>> --- a/lib/Kconfig.debug
>>> +++ b/lib/Kconfig.debug
>>> @@ -222,7 +222,9 @@ config DEBUG_INFO_DWARF4
>>>
>>>  config DEBUG_INFO_BTF
>>>       bool "Generate BTF typeinfo"
>>> -     depends on DEBUG_INFO
>>> +     depends on DEBUG_INFO || COMPILE_TEST
>> I had to add this, since DEBUG_INFO which depends on:
>>
>>         DEBUG_KERNEL && !COMPILE_TEST
>>
>> would block DEBUG_INFO_BTF when COMPILE_TEST is turned on.
>>
> 
> Sorry if I'm being dense here. But what's the point in enabling
> DEBUG_INFO_BTF if there is no *valid* DWARF info available for
> DWARF-to-BTF conversion?

As I mention in [0] there is no point in having `!GCC_PLUGIN_RANDSTRUCT
|| COMPILE_TEST` without `DEBUG_INFO || COMPILE_TEST`, since without it
COMPILE_TEST would block DEBUG_INFO_BTF and because of that
GCC_PLUGIN_RANDSTRUCT would be never blocked by BTF.

As far as I understood from [1] main point for all these these things
with COMPILE_TEST is to be able to check if kernel could be compiled
with all these options (e.g. check syntax, build scripts, etc).

I can rollback `DEBUG_INFO || COMPILE_TEST`, but in that case there is
no point in `GCC_PLUGIN_RANDSTRUCT || COMPILE_TEST` since COMPILE_TEST
in that case will not affect anything here, regardless from it's value.

[0]:
https://lore.kernel.org/bpf/202004010029.167BA4AA1F@keescook/T/#m2f493902d6aed09d30e5c4144a0164459386339d
[1]:
https://lore.kernel.org/bpf/202004010029.167BA4AA1F@keescook/T/#m8f25fab3476c9619249fee9ae692acb98c02cdc7
> 
> 
>> In that case allyesconfig will emit both:
>>
>> CONFIG_DEBUG_INFO_BTF=y
>> CONFIG_GCC_PLUGIN_RANDSTRUCT=y
> 
> Which I thought is exactly what we wanted to avoid. Not sure what's
> the point of compiling kernel (even if it's the one that is not
> supposed to ever run) that apriori has broken BTF? If it was
> acceptable to not have DEBUG_INFO for COMPILE_TEST, why it's not
> acceptable to not have DEBUG_INFO_BTF in that situation as well?
> 
>>
>>
>>
>>> +     depends on !DEBUG_INFO_SPLIT && !DEBUG_INFO_REDUCED
>>> +     depends on !GCC_PLUGIN_RANDSTRUCT || COMPILE_TEST
>>>       help
>>>         Generate deduplicated BTF type information from DWARF debug info.
>>>         Turning this on expects presence of pahole tool, which will convert
>>>