nvme: Build the page list in the existing dma buffer

nvme: Build the page list in the existing dma buffer

Commit 01f2736cc905d ("nvme: Pass large I/O requests as PRP lists")
introduced multi-page requests using the NVMe PRP mechanism. To store the
list and "first page to write to" hints, it added fields to the NVMe
namespace struct.

Unfortunately, that struct resides in fseg which is read-only at runtime.
While KVM ignores the read-only part and allows writes, real hardware and
TCG adhere to the semantics and ignore writes to the fseg region. The net
effect of that is that reads and writes were always happening on address 0,
unless they went through the bounce buffer logic.

This patch builds the PRP maintenance data in the existing "dma bounce
buffer" and only builds it when needed.

Fixes: 01f2736cc905d ("nvme: Pass large I/O requests as PRP lists")
Reported-by: Matt DeVillier <matt.devillier@gmail.com>
Signed-off-by: Alexander Graf <graf@amazon.com>
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Reviewed-by: Alexander Graf <graf@amazon.com>