KASSERT_WARN for asserting malloc(M_WAITOK) not in a non-sleepable thread

[Justin Hibbits]

On Thu, Sep 25, 2014 at 11:16 AM, Davide Italiano <davide at freebsd.org> wrote:
> On Thu, Sep 25, 2014 at 11:09 AM, Ian Lepore <ian at freebsd.org> wrote:
>> On Thu, 2014-09-25 at 10:51 -0700, Davide Italiano wrote:
>>> On Thu, Sep 25, 2014 at 9:14 AM, Adrian Chadd <adrian at freebsd.org> wrote:
>>> > Hi,
>>> >
>>> > Please bring in KASSERT_WARN().
>>> >
>>> > I'm grown up enough to use KASSERT_WARN() along with handling the
>>> > invariant check myself in code. Having KASSERT_WARN() means I can add
>>> > in this rather than printf()s or device_printf()'s with various knobs
>>> > to remove it.
>>> >
>>> > (This is absolutely _not_ the "should KASSERT() optionally just log"
>>> > argument. I'm not going to get into that a second time.)
>>> >
>>> >
>>>
>>> If you put a KASSERT() inside your code -- probably you should be
>>> careful enough to put that iff you're sure that it should be always
>>> verified. No exceptions.
>>> People tend to be very lazy (including me). I don't expect everybody
>>> diligently upgrading KASSERT_WARN to KASSERT. So KASSERT_WARN start
>>> becoming more and more widespread, and people realize all of these
>>> need to be upgraded to KASSERT or removed. This generally happens
>>> after years. Yet. Another. Crusade.
>>> There's a lot of work in the kernel to remove old/wrong/naive  KPI
>>> from the kernel. jhb@ is looking at timeout()-> callout() conversion.
>>> I'm personally looking at dev_clone() removal. There are a lot of
>>> other examples.
>>> Adding KASSERT_WARN is a step backward, not a step forward, IMHO.
>>> That said, if you want to pollute the kernel, fine. I expressed my
>>> opinion, and I'm personally not happy about this, but I never stated
>>> I'm gonna stop you from doing that.
>>>
>>> Thanks,
>>>
>>> --
>>
>> IMO, this entire argument is ridiculous.  Some conditions are so insane
>> that you've got to stop immediately rather than make things worse.
>> Other conditions indicate problems, but the code can recover or
>> otherwise continue to operate safely.  Trying to define every possible
>> anomalous condition as either fatal or not worth mentioning is insane.
>>
>> Everyone is free to write code such as
>>
>> #ifdef INVARIANTS
>>   if (some_condition)
>>     printf("whatever warning\n");
>> #endif
>>
>> So let's be clear here:  the objections are to spelling that code
>> sequence KASSERT_WARN.  If you object, please explain what's wrong with
>> that spelling and how you would prefer it to be spelled.
>>
>> -- Ian
>>
>>
>
> Take the assert out of the name. Call it DEBUG_WARN, or something else
> if you like.
> assert as a pretty *clear* and specific semantic, no need to mess
> around with it.
>
> Thanks,
>
> --
> Davide

I like my bikeshed a nice royal blue.  At a previous job we used
ASSERT and VERIFY macros.  VERIFY was comparable to this (warn if
condition not met, don't panic), so how about KVERIFY() (I'll also
support KWARN, but I think KVERIFY() conveys a better message by
name).

- Justin