Re: [PATCH] ntfs: use page allocation for resident attribute inline data

["Darrick J. Wong" <djwong@kernel.org>]

On Wed, Apr 22, 2026 at 11:35:32PM +0900, Namjae Jeon wrote:
> On Wed, Apr 22, 2026 at 9:55 PM Matthew Wilcox <willy@infradead.org> wrote:
> >
> > On Wed, Apr 22, 2026 at 07:46:27PM +0900, Namjae Jeon wrote:
> > > The current kmemdup() based allocation for IOMAP_INLINE can result in
> > > inline_data pointer having a non-zero page offset. This causes
> > > iomap_inline_data_valid() to fail the check:
> > >
> > >     iomap->length <= PAGE_SIZE - offset_in_page(iomap->inline_data)
> > >
> > > and triggers the kernel BUG at fs/iomap/buffered-io.c:1061.
> >
> > Hang on, hang on, hang on.
> >
> > First, maybe this check is too strict.  I'm sure it's true for EROFS,
> > but I don't see why it should be true for everybody.  Perhaps we should
> > delete this check or relax it?
> I agree that the current check might be unnecessarily strict for
> general cases. So I will prepare another patch to remove this trap for
> further discussion with iomap maintainers.
> >
> > Second, why are you calling kmemdup() to begin with?  This seems
> > entirely pointless; the iomap code is going to call memcpy() on it.
> > You're supposed to just be pointing into your data structures.
> In the initial implementation of NTFS with iomap, I pointed directly
> to the internal data structures. However, I encountered this BUG_ON
> trap during testing, so I switched to page allocation to avoid it.
> Then, during the review process for the NTFS series, I changed it to
> kmemdup() without much thought. If this BUG_ON trap can be removed, I
> could have simply pointed to the internal data structures as you said.

I think the check is wrong.  We rely on the filesystem to point
iomap::inline_data to kernel memory that is at least iomap::length bytes
in size.  If that crosses a PAGE_SIZE boundary that's fine, so long as
the caller actually mapped that much memory.  IOWs, if you have an
iomap:

{pos = 0, inline_data = 0xB0000, length = 32768, ...}

then we trust that you really did map all of the MDA text mode memory
and that memcpy'ing 100 bytes to pos 4090 is ok.

(Perhaps this is a relic of the bs<=ps days?)

--D