Re: [PATCH v1] mm/gup: remove (VM_)BUG_ONs

[David Hildenbrand <david@redhat.com>]

On 06.06.25 13:04, Lorenzo Stoakes wrote:
> On Fri, Jun 06, 2025 at 12:28:28PM +0200, David Hildenbrand wrote:
>> On 06.06.25 12:19, Lorenzo Stoakes wrote:
>>> On Fri, Jun 06, 2025 at 12:13:27PM +0200, Michal Hocko wrote:
>>>> On Fri 06-06-25 11:01:18, David Hildenbrand wrote:
>>>>> On 06.06.25 10:31, Michal Hocko wrote:
>>>> [...]
>>>>>> Turning them into VM_WARN_ON
>>>>>> should be reasonably safe as they are not enabled in production
>>>>>> environment anyway so we cannot really rely on those. Having them in
>>>>>> WARN form would be still useful for debugging and those that really need
>>>>>> a crash dump while debugging can achieve the same result.
>>>>>
>>>>> One question is if we should be VM_WARN_ON vs. VM_WARN_ON_ONCE ...
>>>>
>>>> *WARN_ONCE ha a very limited use from code paths which are generally
>>>> shared by many callers. You just see one and then nothing. Some time ago
>>>> we have discussed an option to have _ONCE per call trace but I haven't
>>>> see any follow up.
>>>>
>>>> Anyway starting without _ONCE seems like safer option because we are not
>>>> losing potentially useful debugging information. Afterall this is
>>>> debugging only thing. But no strong position on my side.
>>>>
>>>>> VM_BUG_ON is essentially a "once" thing so far, but then, we don't continue
>>>>> ... so probably most should be _ONCE.
>>>>>
>>>>>>
>>>>>> So while I agree that many of them could be dropped or made more clear
>>>>>> those could be dealt with after a mass move. An advantage of this would
>>>>>> be that we can drop VM_BUG_ON* and stop new instances from being added.
>>>>>
>>>>> As a first step we could probably just #define them to go to the
>>>>> VM_WARN_ON_* variants and see what happens.
>>>>
>>>> You meand VM_BUG_ON expand to VM_WARN_ON by default?
>>>
>>> Sorry to interject in the conversation, but to boldly throw my two English pence
>>> into the mix:
>>>
>>> As the "king of churn" (TM) you'll not be surprised to hear that I'm in favour
>>> of us just doing a big patch and convert all VM_BUG_ON() -> VM_WARN_ON_ONCE()
>>> and remove VM_BUG_ON*().
>>>
>>> Pull the band-aid off... I think better than a #define if this indeed what you
>>> meant David.
>>>
>>> But of course, you'd expect me to have this opinion ;)
>>
>> See my reply to Michal regarding keeping VM_BUG_ON() until we actually
>> decided what the right cleanup is.
> 
> Sure, but to me the concept of VM_BUG_ON() is surely fundamentally broken - if
> BUG_ON() means 'stop everything we're going to corrupt' then it makes no sense
> to add a '...but only if CONFIG_DEBUG_VM is set' in there.
> 
> So to me the only assessment needed is 'do we want to warn on this or not?'.

Well, when done carefully, it would be when reworking a VM_BUG_ON:

(a) Should this really only be checked with DEBUG_VM or should this
     actually be a WARN_ON_ONCE() + recovery
(b) Does this check even still make sense in current code, or were we
     just extra careful initially.
(c) Do we even understand why it is checked or should we add a comment.
(d) Should we use one of the _PAGE / _FOLIO / _VMA etc. variants instead
     or even add new ones.

One could argue that the same is true for any other VM_WARN_ON ... but 
my point from the beginning was that if we're already touching them, why 
not spend some extra time and do it properly ..

... but yeah, 600 instances are a bit much.

:)

But yeah, VM_BUG_ON is easier to replace than BUG_ON, because ... not 
even Fedora is running with CONFIG_DEBUG_VM anymore, so these checks are 
mostly only there in actual debug kernels.

So agreed, let's move forward with a simple conversion.

> 
> And as you say, really WARN_ON_ONCE() seems appropriate, because nearly always
> we will get flooded with useless information.
> 
> I think this being debug code gives us a lot of leeway here.
> 
> After the big change, we can always revisit individual cases and see if the
> warning is valid at all.
> 
>>
>> I don't have a very strong opinion on any of this ... as long as VM_BUG_ON()
>> goes away one way or the other.
> 
> Am happy to come up with the churn-meister version of this patch and take the
> heat :P

I assume with "this patch" you mean "a patch that gets rid of VM_BUG_ON 
completely", because I want this patch here (that started the 
discussion) to go in first.

Fascinating how you are always looking for work :P

Thanks!

-- 
Cheers,

David / dhildenb