Re: [PATCH] mm/debug: Use BUILD_BUG_ON_INVALID() for VIRTUAL_BUG_ON()

[David Hildenbrand]

On 07.06.25 18:21, Tal Zussman wrote:
> On Sat, Jun 7, 2025 at 3:59 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
> > On 07.06.25 09:09, Tal Zussman wrote:
> > > This allows the compiler to validate the condition even with
> > > CONFIG_DEBUG_VIRTUAL disabled, and aligns VIRTUAL_BUG_ON() with the
> > > other macros in mmdebug.h.
> >
> > In the light of recent discussions, I think we should get rid of
> > VIRTUAL_BUG_ON completely.
> >
> > There are only a hand full of callers, and I am preety sure for most of
> > them VM_WARN_ON is a suitable replacement.
>
> Makes sense. However, all of the callers (except for vmalloc) are already
> gated by CONFIG_DEBUG_VIRTUAL, which doesn't depend on CONFIG_DEBUG_VM, so
> using VM_WARN_ON_ONCE() would break DEBUG_VIRTUAL on its own.

It should either be folded or made dependent on DEBUG_VM

> Perhaps it makes sense to convert the non-vmalloc callers to WARN_ON_ONCE()
> instead so DEBUG_VIRTUAL still works.
> The vmalloc caller would then become>
>          if (IS_ENABLED(CONFIG_DEBUG_VIRTUAL))
>                  WARN_ON_ONCE(...);
>
> as opposed to VM_WARN_ON_ONCE(), in order to maintain the existing
> DEBUG_VIRTUAL behavior.

From a quick glimpse, I am not even sure Fedora/RHEL enable it in the 
debug config (and if so, probably it's not enabled by mistake)

I think we can make it depend on DEBUG_VM.

> Alternatively, DEBUG_VIRTUAL could be folded into DEBUG_VM, but that seems
> like a slightly more invasive change...

I'd start with making it depend on DEBUG_VM.

If there is good reason to not fold it, then all VIRTUAL_BUG_ON should 
be renamed to VIRTUAL_WARN_ON and get defined as VM_WARN_ON.

If there is good reason to fold it, then all VIRTUAL_BUG_ON should be 
replaced by VM_WARN_ON.

The only question is what the performance overhead is. I mean, 
performance with DEBUG_VM is already bad (debug kernels), so not sure if 
we really care that much.

--
Cheers,

David / dhildenb