Re: BUG and WARN kernel log levels

[Kees Cook]

On Mon, Aug 15, 2016 at 12:00 PM, Joe Perches <joe@xxxxxxxxxxx> wrote:
> On Mon, 2016-08-15 at 11:53 -0700, Kees Cook wrote:
>> Hi,
>>
>> So, I noticed that asm-gemeric/bug.h defines BUG() without a log level:
>>
>> #ifndef HAVE_ARCH_BUG
>> #define BUG() do { \
>>        printk("BUG: failure at %s:%d/%s()!\n", __FILE__, __LINE__, __func__); \
>>
>> Seems like it should have one?
>>
>> Also, I think we might want to examine WARN() a bit... it doesn't have
>> a log level either, but only a fraction of callers set one:
>>
>> $ git grep -E 'WARN(_TAINT|)(_RATELIMIT|_ONCE|)\([^\)]' | grep -v KERN_ | wc -l
>> 2735
>>
>> $ git grep -E 'WARN(_TAINT|)(_RATELIMIT|_ONCE|)\([^\)]' | grep KERN_ | wc -l
>> 77
>>
>> If I'm reading checkpatch.pl correctly, it doesn't warn about missing
>> log levels on WARN calls, but I think it should.
>>
>> How do you think is best to clean this up?
>>
>> Mainly, I'd like to add a format string to BUG, or introduce a new
>> BUGish call that takes a format...
>
> I once suggested something similar awhile ago.
> https://lkml.org/lkml/2008/7/8/261
>
> I think it's best to remove any KERN_<LEVEL> from the use of
> all the WARN variants and add it to the WARN definitions.

Yeah, that's what I was thinking too. It does mean that the format
needs to be a const char string, though. I haven't checked all of them
but most seem to be. It's an easy fix to move them to "%s" as needed,
though.

> Same with BUG.

Yeah, though for full effect, it needs to be plumbed into each
architecture's BUG handler. Mainly what I don't like is that if I do
this on x86:

pr_err("eek, terrible thing\n");
BUG();

I get a dmesg that read:

eek, terrible thing
=== cut here ===
...traceback, etc

I'd like the "eek" part to be inside the "cut here". And I'd like BUGs
to be able to be more verbose.

-Kees

-- 
Kees Cook
Nexus Security