On Wed, Oct 18, 2023 at 08:18:41AM -0700, Keith Busch wrote:

From: Keith Busch <kbusch@kernel.org>

Map user metadata buffers directly instead of maintaining a complicated
copy buffer.

Now that the bio tracks the metadata through its bip, nvme doesn't need
special metadata handling, callbacks, or additional fields in the pdu.
This greatly simplifies passthrough handling and avoids a "might_fault"
copy_to_user in the completion path. This also creates pdu space to
track the original request separately from its bio, further simplifying
polling without relying on special iouring fields.

The downside is that nvme requires the metadata buffer be physically
contiguous, so user space will need to utilize huge pages if the buffer
needs to span multiple pages. In practice, metadata payload sizes are a
small fraction of the main payload, so this shouldn't be a problem.

We can't just remove the old path.  We might still need bounce
buffering to due misalignment and/or because it is notcontiguous.
Same as we have a direct map and a copy path for data.