Re: [PATCH 2/4] mm: perform VMA allocation, freeing, duplication in mm

[Suren Baghdasaryan <surenb@google.com>]

On Fri, Apr 25, 2025 at 6:55 AM Liam R. Howlett <Liam.Howlett@oracle.com> wrote:
>
> * Lorenzo Stoakes <lorenzo.stoakes@oracle.com> [250425 06:40]:
> > On Thu, Apr 24, 2025 at 08:15:26PM -0700, Kees Cook wrote:
> > >
> > >
> > > On April 24, 2025 2:15:27 PM PDT, Lorenzo Stoakes <lorenzo.stoakes@oracle.com> wrote:
> > > >+static void vm_area_init_from(const struct vm_area_struct *src,
> > > >+                        struct vm_area_struct *dest)
> > > >+{
> > > >+  dest->vm_mm = src->vm_mm;
> > > >+  dest->vm_ops = src->vm_ops;
> > > >+  dest->vm_start = src->vm_start;
> > > >+  dest->vm_end = src->vm_end;
> > > >+  dest->anon_vma = src->anon_vma;
> > > >+  dest->vm_pgoff = src->vm_pgoff;
> > > >+  dest->vm_file = src->vm_file;
> > > >+  dest->vm_private_data = src->vm_private_data;
> > > >+  vm_flags_init(dest, src->vm_flags);
> > > >+  memcpy(&dest->vm_page_prot, &src->vm_page_prot,
> > > >+         sizeof(dest->vm_page_prot));
> > > >+  /*
> > > >+   * src->shared.rb may be modified concurrently when called from
> > > >+   * dup_mmap(), but the clone will reinitialize it.
> > > >+   */
> > > >+  data_race(memcpy(&dest->shared, &src->shared, sizeof(dest->shared)));
> > > >+  memcpy(&dest->vm_userfaultfd_ctx, &src->vm_userfaultfd_ctx,
> > > >+         sizeof(dest->vm_userfaultfd_ctx));
> > > >+#ifdef CONFIG_ANON_VMA_NAME
> > > >+  dest->anon_name = src->anon_name;
> > > >+#endif
> > > >+#ifdef CONFIG_SWAP
> > > >+  memcpy(&dest->swap_readahead_info, &src->swap_readahead_info,
> > > >+         sizeof(dest->swap_readahead_info));
> > > >+#endif
> > > >+#ifdef CONFIG_NUMA
> > > >+  dest->vm_policy = src->vm_policy;
> > > >+#endif
> > > >+}
> > >
> > > I know you're doing a big cut/paste here, but why in the world is this function written this way? Why not just:
> > >
> > > *dest = *src;
> > >
> > > And then do any one-off cleanups?
> >
> > Yup I find it odd, and error prone to be honest. We'll end up with uninitialised
> > state for some fields if we miss them here, seems unwise...
> >
> > Presumably for performance?
> >
> > This is, as you say, me simply propagating what exists, but I do wonder.
>
> Two things come to mind:
>
> 1. How ctors are done.  (v3 of Suren's RCU safe patch series, willy made
> a comment.. I think)
>
> 2. Some race that Vlastimil came up with the copy and the RCU safeness.
> IIRC it had to do with the ordering of the setting of things?
>
> Also, looking at it again...
>
> How is it safe to do dest->anon_name = src->anon_name?  Isn't that ref
> counted?

dest->anon_name = src->anon_name is fine here because right after
vm_area_init_from() we call dup_anon_vma_name() which will bump up the
refcount. I don't recall why this is done this way but now looking at
it I wonder if I could call dup_anon_vma_name() directly instead of
this assignment. Might be just an overlooked legacy from the time we
memcpy'd the entire structure. I'll need to double-check.

>
> Pretty sure it's okay, but Suren would know for sure on all of this.
>
> Suren, maybe you could send a patch with comments on this stuff?

Yeah, I think I need to add some comments in this code for
clarification. We do not copy the entire vm_area_struct because we
have to preserve vma->vm_refcnt field of the dest vma. Since these
structures are allocated from a cache with SLAB_TYPESAFE_BY_RCU,
another thread might be concurrently checking the state of the dest
object by reading dest->vm_refcnt. Therefore it's important here not
to override the vm_refcnt. Changelog in
https://lore.kernel.org/all/20250213224655.1680278-18-surenb@google.com/
touches on it but a comment in the code would be indeed helpful. Will
add it but will wait for Lorenzo's refactoring to land into linux-mm
first to avoid adding merge conflicts.

>
> Thanks,
> Liam