Re: [RFC PATCH] blk-integrity: support arbitrary buffer alignment

[Keith Busch <kbusch@kernel.org>]

On Wed, Nov 05, 2025 at 03:53:09PM +0100, Christoph Hellwig wrote:
> On Wed, Nov 05, 2025 at 07:42:41AM -0700, Keith Busch wrote:
> > This generally does pass 'void *' to contiguous protection data. If the
> > actual protection payload is spread across multiple segments, though, we
> > don't have contiguous data, so we have to bounce it through something.
> > Declaring the union on the stack provides this memory for that unusual
> > case. If the bip segment is big enough, we point the 'void *' directly
> > at it; if not, we point it to the temporary onstack allocation. Using
> > a union ensures that it's definitely big enough for any checksum type.
> 
> The union contains pointers, which are always 8 bytes.  But I guess
> I can see now how this didn't blow up, assuming this was only tested
> on classic PI, as the tuples are just 8 bytes and you just used the
> pointer as storage?  I.e. these aren't supposed to be pointers, but
> values?

Yeah, should have been values, not pointers. My mistake.

We don't hit the condition that needs the copy buffer with the kernel's
auto-pi, so the bug would never get hit in that path.

I have an io_uring test program that sends pi with offsets such that the
field straddles pages. It does not look like it's hitting any of the ref
tag remapping code though, so still looking into that.