dpuVendor: Build a multiarch ipu manifest

The dpu-operator by default expects a multiarch vsp image.

In the past, this code just worked, because we would build and push the
image immediately before deploying the operator on the respective
architecture, so the most recent image would be the proper arch.

This is flimsy, and can cause errors later on if we need to pull the
image again. Instead, as a workaround while waiting for multiarch IPU
vsp build support, build a manifest of the images during the second
phase of deployment (iso-cluster dpu).

This code makes a (potentially bad) assumption that the host side
deployment has already run, and built/pushed the x86_64 vsp.

Signed-off-by: Salvatore Daniele <[email protected]>

Author: SalDaniele

dpuVendor.py

@@ -72,6 +72,11 @@ def build_push(self, h: host.Host, imgReg: ImageRegistry) -> str:
         vsp_image = self.vsp_image_name(imgReg)
         h.run_or_die(f"podman tag intel-ipuplugin:latest {vsp_image}")
         h.run_or_die(f"podman push {vsp_image}")
+        # WA to ensure multiarch vsp image manifest is available
+        # push images with both the name expected by the dpu operator (so we can proceed with deploying host side)
+        # and the name expected by the manifest that we will build during the IPU deployment step
+        h.run_or_die(f"podman tag {vsp_image} {vsp_image}-{self.name_suffix}")
+        h.run_or_die(f"podman push {vsp_image}-{self.name_suffix}")
         return vsp_image
 
     def start(self, vsp_image: str, client: K8sClient) -> None:

extraConfigDpu.py

@@ -179,8 +179,19 @@ def ExtraConfigDpu(cc: ClustersConfig, cfg: ExtraConfigArgs, futures: dict[str,
         # Build on the ACC since an aarch based server is needed for the build
         # (the Dockerfile needs to be fixed to allow layered multi-arch build
         # by removing the calls to pip)
-        vendor_plugin.build_push(acc, imgReg)
-        # vendor_plugin.start(vendor_plugin.vsp_image_name(imgReg), client)
+        vsp_img = vendor_plugin.build_push(acc, imgReg)
+
+        # As a workaround while waiting for properly multiarch build support, we can create a manifest to ensure both host and dpu can deploy the vsp with the same image.
+        # Note that this makes the assumption that the host deployment has already been run and the latest ipu plugin image is already locally available in the registry.
+        # Without these assumptions, this will not work as expected
+        manifest = f"{vsp_img}-manifest"
+        lh.run(f"buildah manifest rm {manifest}")
+        lh.run_or_die(f"buildah manifest create {manifest}")
+        lh.run_or_die(f"podman pull {vsp_img}-x86_64")
+        lh.run_or_die(f"podman pull {vsp_img}-aarch64")
+        lh.run_or_die(f"buildah manifest add {manifest} {vsp_img}-x86_64")
+        lh.run_or_die(f"buildah manifest add {manifest} {vsp_img}-aarch64")
+        lh.run_or_die(f"buildah manifest push --all {manifest} docker://{vsp_img}")
 
     git_repo_setup(repo, repo_wipe=False, url=DPU_OPERATOR_REPO)
     if cfg.rebuild_dpu_operators_images: