drm/xe/dma_buf: stop relying on placement in unmap

The is_vram() is checking the current placement, however if we consider
exported VRAM with dynamic dma-buf, it looks possible for the xe driver
to async evict the memory, notifying the importer, however importer does
not have to call unmap_attachment() immediately, but rather just as
"soon as possible", like when the dma-resv idles. Following from this we
would then pipeline the move, attaching the fence to the manager, and
then update the current placement. But when the unmap_attachment() runs
at some later point we might see that is_vram() is now false, and take
the complete wrong path when dma-unmapping the sg, leading to
explosions.

To fix this check if the sgl was mapping a struct page.

v2:
  - The attachment can be mapped multiple times it seems, so we can't
    really rely on encoding something in the attachment->priv. Instead
    see if the page_link has an encoded struct page. For vram we expect
    this to be NULL.

Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/4563


Fixes: dd08ebf6 ("drm/xe: Introduce a new DRM driver for Intel GPUs")
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: <stable@vger.kernel.org> # v6.8+
Acked-by: Christian König <christian.koenig@amd.com>
Link: https://lore.kernel.org/r/20250410162716.159403-2-matthew.auld@intel.com


(cherry picked from commit d755887f)
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>