Re: [GIT PULL] VFIO updates for v6.17-rc1

[Alex Williamson <alex.williamson@redhat.com>]

On Mon, 4 Aug 2025 16:55:09 -0700
Linus Torvalds <torvalds@linux-foundation.org> wrote:

> On Mon, 4 Aug 2025 at 15:22, Alex Williamson <alex.williamson@redhat.com> wrote:
> >
> > Li Zhe (6):
> >       mm: introduce num_pages_contiguous()  
> 
> WHY?
> 
> There is exactly *ONE* user, why the heck do we introduce this
> completely pointless "helper" function, and put it in a core header
> file like it was worth it?

There was discussion here[1] where David Hildenbrand and Jason
Gunthorpe suggested this should be in common code and I believe there
was some intent that this would get reused.  I took this as
endorsement from mm folks.  This can certainly be pulled back into
subsystem code.

> And it's not like that code is some kind of work of art that we want
> to expose everybody to *anyway*. It's written in a particularly stupid
> way that means that it's *way* more expensive than it needs to be.
> 
> And then it's made "inline" despite the code generation being
> horrible, which makes it all entirely pointless.
> 
> Yes, I'm grumpy. This pull request came in late, I'm already
> traveling, and then I look at it and it just makes me *angry* at how
> bad that code is, and how annoying it is.

Sorry, I usually try to get in later during the first week to let the
dust settle a bit from the bigger subsystems, I guess I'm running a
little behind this cycle.  We'll get it fixed and I'll resend.  Thanks,

Alex

> My builds are already slower than usual because they happen on my
> laptop while traveling, I do *not* need to see this kind of absolutely
> disgusting code that does stupid things that make the build even
> slower.
> 
> So I refuse to pull this kind of crap.
> 
> If you insist on making my build slower and exposing these kinds of
> helper functions, they had better be *good* helper functions.
> 
> Hint: absolutely nobody cares about "the pages crossed a sparsemem
> border. If your driver cares about the number of contiguous pages, it
> might as well say "yeah, they are contiguous, but they are in
> different sparsemem chunks, so we'll break here too".
> 
> And at that point all you care about is 'struct page' being
> contiguous, instead of doing that disgusting 'nth_page'.
> 
> And then - since there is only *one* single user - you don't put it in
> the most central header file that EVERYBODY ELSE cares about.
> 
> And you absolutely don't do it if it generates garbage code for no good reason!
> 
>             Linus
> 

[1]https://lore.kernel.org/all/20250703111216.GG904431@ziepe.ca/