fix(core): add O_DIRECT for NFS

[nevermarine]

Description

NFS: use O_DIRECT and qemu-img cache=none with fallback

• NFS detection
  Added util.IsPathOnNFS(path) (Linux: statfs + NFS_SUPER_MAGIC; other OS: always false) so we can detect when the destination is on NFS.

• Importer write path (io.Copy)
  In streamDataToFile (pkg/importer/util.go), when the destination is on NFS we first open the file with O_DIRECT (util.OpenFileOrBlockDeviceWithDirectIO on Linux) and use io.Copy. If opening with O_DIRECT fails, we fall back to a normal open and copy. Added OpenFileOrBlockDeviceWithDirectIO in pkg/util (Linux implementation + stub on other platforms).

• qemu-img convert
  For NFS destinations, ConvertToFormatStream now runs qemu-img with -t none -T none. If that fails, it retries with -t writeback -T writeback. The internal helper is convertTo(format, src, dest, preallocate, cacheMode); the public API is a single ConvertToFormatStream(..., useDirectIO bool).

• Where it’s used
  Data processor convert phase and uploadserver clone flow call IsPathOnNFS on the destination and pass the result as useDirectIO into ConvertToFormatStream. NFS is checked once per flow; O_DIRECT and cache=none are tried first, with fallback only when open or convert fails.

• Logging
  NFS detection, O_DIRECT usage, and fallbacks are logged at klog.V(1).

Why do we need it, and what problem does it solve?

When CDI imports or clones disk images onto NFS-backed PVCs, the importer writes through the page cache (normal io.Copy and qemu-img with default cache mode). On NFS this can lead to high memory use and OOMs because data is buffered in the client’s page cache before being sent over the network, and qemu-img also uses the page cache when writing the converted image. To reduce memory pressure and improve behaviour on NFS, we should bypass the page cache where the filesystem allows it: use O_DIRECT for the importer’s file writes and qemu-img convert -t none -T none so that both source and target use direct I/O. This PR adds NFS detection (via statfs and NFS_SUPER_MAGIC) and, when the destination is on NFS, tries O_DIRECT for io.Copy and cache=none for qemu-img convert, with an automatic fallback to the previous behaviour if either fails (e.g. if the NFS server or mount does not support direct I/O). After merging, imports and clones to NFS volumes use less memory and are less likely to hit OOM, while non-NFS and unsupported NFS setups keep working as before.

What is the expected result?

• NFS destination

  1. Create a PVC backed by NFS (or use an existing NFS PVC).
  2. Import or clone a disk image into that PVC (e.g. HTTP import, registry, or clone from another PVC).
  3. In importer logs at -v=1 you should see messages such as: destination on NFS, attempt to write with O_DIRECT; and for convert: “Destination on NFS, running qemu-img convert with -t none -T none”.
  4. Import/clone completes successfully and the VM can use the disk. No OOM from the importer pod during the write/convert phase.

• Fallback when direct I/O is not supported

  1. Use an NFS mount (or other filesystem) that does not support O_DIRECT / cache=none for the target file.
  2. Run the same import or clone.
  3. Logs at -v=1 should show fallback messages (e.g. “O_DIRECT open failed … falling back to normal write” and/or “qemu-img convert with -t none -T none failed … falling back to writeback”).
  4. Import/clone still completes using the fallback path (normal write and/or writeback cache).

• Non-NFS destination

  1. Import or clone to a non-NFS volume (e.g. local storage, block, or other non-NFS filesystem).
  2. Behaviour is unchanged: no NFS-specific messages, no O_DIRECT or cache=none for the importer; existing logic (including existing cache-mode handling) applies.
  3. Import/clone completes as before.

Checklist

• The code is covered by unit tests.
• e2e tests passed.
• Documentation updated according to the changes.
• Changes were tested in the Kubernetes cluster manually.

Changelog entries

section: core
type: fix
summary: "Fixed a potential `OOMKill` during virtual disk creation on NFS."

[nevermarine]

/backport v1.5.2

[deckhouse-BOaTswain]

Failure: cherry-pick commit 82246e3 to the branch release-1.5 failed. See Job for details.