Re: [PATCH v3 2/5] mm: userfaultfd: don't pass around both mm and vma

[Peter Xu <peterx@redhat.com>]

On Tue, Mar 07, 2023 at 01:44:05AM +0000, Nadav Amit wrote:
> 
> 
> > On Mar 6, 2023, at 5:03 PM, Peter Xu <peterx@redhat.com> wrote:
> > 
> > !! External Email
> > 
> > On Mon, Mar 06, 2023 at 02:50:21PM -0800, Axel Rasmussen wrote:
> >> Quite a few userfaultfd functions took both mm and vma pointers as
> >> arguments. Since the mm is trivially accessible via vma->vm_mm, there's
> >> no reason to pass both; it just needlessly extends the already long
> >> argument list.
> >> 
> >> Get rid of the mm pointer, where possible, to shorten the argument list.
> >> 
> >> Signed-off-by: Axel Rasmussen <axelrasmussen@google.com>
> > 
> > Acked-by: Peter Xu <peterx@redhat.com>
> > 
> > One nit below:
> > 
> >> @@ -6277,7 +6276,7 @@ int hugetlb_mfill_atomic_pte(struct mm_struct *dst_mm,
> >>              folio_in_pagecache = true;
> >>      }
> >> 
> >> -     ptl = huge_pte_lock(h, dst_mm, dst_pte);
> >> +     ptl = huge_pte_lock(h, dst_vma->vm_mm, dst_pte);
> >> 
> >>      ret = -EIO;
> >>      if (folio_test_hwpoison(folio))
> >> @@ -6319,9 +6318,9 @@ int hugetlb_mfill_atomic_pte(struct mm_struct *dst_mm,
> >>      if (wp_copy)
> >>              _dst_pte = huge_pte_mkuffd_wp(_dst_pte);
> >> 
> >> -     set_huge_pte_at(dst_mm, dst_addr, dst_pte, _dst_pte);
> >> +     set_huge_pte_at(dst_vma->vm_mm, dst_addr, dst_pte, _dst_pte);
> >> 
> >> -     hugetlb_count_add(pages_per_huge_page(h), dst_mm);
> >> +     hugetlb_count_add(pages_per_huge_page(h), dst_vma->vm_mm);
> > 
> > When vm_mm referenced multiple times (say, >=3?), let's still cache it in a
> > temp var?
> > 
> > I'm not sure whether compiler is smart enough to already do that with a
> > reg, even if so it may slightly improve readability too, imho, by avoiding
> > the multiple but same indirection for the reader.
> 
> I am not sure if you referred to this code specifically or in general.

In general.  IIRC there're more than one such case in this patch.

> I once looked into it, and the compiler is really stupid in this regard
> and super conservative when it comes to aliasing. Even if you use
> “restrict” keyword or “__pure” or “__const” function attributes, in
> certain cases (function calls to other compilation units, or inline
> assembly - I don’t remember) the compiler might ignore them. Worse, llvm
> and gcc are inconsistent.
> 
> From code-generated perspective, I did not see a clear cut that benefits
> caching over not. From performance perspective the impact is negligible. I
> mention all of that because I thought it matters too, but it mostly does
> not.
> 
> That’s all to say that in most cases, I think that whatever makes the code
> more readable should be preferred. I think that you are correct in saying
> that “caching” it will make the code more readable, but performance-wise
> it is probably meaningless.

Thanks for the knowledge.  I would suspect no matter how the output layout
of the compilers will be the difference will be small.  I prefer it more
for readability as you said but not strongly either way.

-- 
Peter Xu