Re: [PATCH v3 1/2] libbpf: show error info about missing ".BTF" section

From: Quentin Monnet
Date: Thu Jan 05 2023 - 09:58:03 EST


2023-01-03 15:46 UTC-0800 ~ Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx>
> On Tue, Jan 3, 2023 at 7:03 AM Quentin Monnet <quentin@xxxxxxxxxxxxx> wrote:
>>
>> 2022-12-20 16:13 UTC-0800 ~ Andrii Nakryiko <andrii.nakryiko@xxxxxxxxx>
>>> On Tue, Dec 20, 2022 at 3:34 AM Leo Yan <leo.yan@xxxxxxxxxx> wrote:
>>>>
>>>> On Tue, Dec 20, 2022 at 09:31:14AM +0800, Changbin Du wrote:
>>>>
>>>> [...]
>>>>
>>>>>>> Now will print below info:
>>>>>>> libbpf: failed to find '.BTF' ELF section in /home/changbin/work/linux/vmlinux
>>>>>>
>>>>>> Recently I encountered the same issue, it could be caused by:
>>>>>> either missing to install tool pahole or missing to enable kernel
>>>>>> configuration CONFIG_DEBUG_INFO_BTF.
>>>>>>
>>>>>> Could we give explict info for reasoning failure? Like:
>>>>>>
>>>>>> "libbpf: failed to find '.BTF' ELF section in /home/changbin/work/linux/vmlinux,
>>>>>> please install pahole and enable CONFIG_DEBUG_INFO_BTF=y for kernel building".
>>>>>>
>>>>> This is vmlinux special information and similar tips are removed from
>>>>> patch V2. libbpf is common for all ELFs.
>>>>
>>>> Okay, I see. Sorry for noise.
>>>>
>>>>>>> Error: failed to load BTF from /home/changbin/work/linux/vmlinux: No such file or directory
>>>>>>
>>>>>> This log is confusing when we can find vmlinux file but without BTF
>>>>>> section. Consider to use a separate patch to detect vmlinux not
>>>>>> found case and print out "No such file or directory"?
>>>>>>
>>>>> I think it's already there. If the file doesn't exist, open will fail.
>>>>
>>>> [...]
>>>>
>>>>>>> @@ -990,6 +990,7 @@ static struct btf *btf_parse_elf(const char *path, struct btf *base_btf,
>>>>>>> err = 0;
>>>>>>>
>>>>>>> if (!btf_data) {
>>>>>>> + pr_warn("failed to find '%s' ELF section in %s\n", BTF_ELF_SEC, path);
>>>>>>> err = -ENOENT;
>>>>
>>>> btf_parse_elf() returns -ENOENT when ELF file doesn't contain BTF
>>>> section, therefore, bpftool dumps error string "No such file or
>>>> directory". It's confused that actually vmlinux is existed.
>>>>
>>>> I am wondering if we can use error -LIBBPF_ERRNO__FORMAT (or any
>>>> better choice?) to replace -ENOENT at here, this can avoid bpftool to
>>>> outputs "No such file or directory" in this case.
>>>
>>> The only really meaningful error code would be -ESRCH, which
>>> strerror() will translate to "No such process", which is also
>>> completely confusing.
>>>
>>> In general, I always found these strerror() messages extremely
>>> unhelpful and confusing. I wonder if we should make an effort to
>>> actually emit symbolic names of errors instead (literally, "-ENOENT"
>>> in this case). This is all tooling for engineers, I find -ENOENT or
>>> -ESRCH much more meaningful as an error message, compared to "No such
>>> file" seemingly human-readable interpretation.
>>>
>>> Quenting, what do you think about the above proposal for bpftool? We
>>> can have some libbpf helper internally and do it in libbpf error
>>> messages as well and just reuse the logic in bpftool, perhaps?
>>
>> Apologies for the delay.
>> What you're proposing is to replace all messages currently looking like
>> this:
>>
>> $ bpftool prog
>> Error: can't get next program: Operation not permitted
>>
>> by:
>>
>> $ bpftool prog
>> Error: can't get next program: -EPERM
>>
>> Do I understand correctly?
>
> yep, that's what I had in mind
>
>>
>> I think the strerror() messages are helpful in some occasions (they
>> _are_ more human-friendly to many users), but it's also true that
>> they're not always precise. With bpftool, "Invalid argument" is a
>> classic when the program doesn't load, and may lead to confusion with
>> the args passed to bpftool on the command line. Then there are the other
>> corner cases like the one discussed in this thread. So, why not.
>
> maybe the right approach would be to have both symbolic error name and
> its human-readable representation, so for example above
>
> Error: can't get next program: [-EPERM] Operation not permitted
>
> or something like that? And if error value is unknown, just keep it as
> integer: "[-5555]" ?
That would be great, we'd have both the error name for savvy users and
the (more or less accurate) interpretation for others.

>> If we do change, yeah I'd rather have as much of this handling in libbpf
>> itself, and then adjust bpftool to handle the remaining cases, for
>> consistency.
>
> we can teach libbpf_strerror_r() to do this and if bpftool is going to
> use it consistently then it would get the benefit automatically
Sounds good to me.