Sync main with reliability fixes - #14

…rlink-docker-plugin into 2-light-version-no-gpu

use just dind

Signed-off-by: Diego Ciangottini <dciangot@host-10-25-127-213.wired.local>

Signed-off-by: Diego Ciangottini <diego.ciangottini@pg.infn.it>

enable listen on socket

…n into 2-light-version-no-gpu

Signed-off-by: Giulio Bianchini <giulio.bianchini@hotmail.it>

…llback

  Create.go
  - Do not touch the ResponseWriter after the create handler has returned. The
    asynchronous container-creation goroutine called HandleErrorAndRemoveData,
    which writes to w long after the response was sent; that races with the http
    server and can corrupt the next response on a keep-alive connection. Split
    the cleanup out into cleanupPodData, which never touches w, and use it from
    the goroutine. HandleErrorAndRemoveData stays for the request goroutine.
  - Recover panics in that goroutine. net/http does not recover panics in bare
    goroutines spawned by a handler, so any nil deref there took the whole
    plugin down along with the bookkeeping of every other running pod.
  - Bound the mesh_ready wait. When mesh.sh fails the sentinel is never written,
    and the poll loop in containers_command.sh spun forever: the pod stayed in
    Waiting with no diagnostics and its DIND stayed claimed until a human
    deleted it. The loop now times out after 300s, dumps the overlay container
    logs and exits non-zero, and the docker exec itself runs under a 10m context
    so it cannot leak the goroutine for any other reason. Script output is
    logged on failure.
  - Select the network-overlay container by name instead of assuming it is
    containers[0]. The startup path was gated on the mesh annotation, a weaker
    condition than the one under which the overlay is actually built; when the
    two disagreed a workload container was started as the overlay and every
    other container waited on a sentinel nothing would write. It also panicked
    on an empty container list. The mismatch is now logged as a warning and the
    containers start directly.
  - Scope envVars and fpgaArgs to a single container. Declared outside the loop,
    every container inherited the -e/-v flags and --device entries of the
    containers processed before it.

  Status.go
  - Do not fail a whole status batch because one pod has no DIND container.
    interLink asks for every pod on the node in one call, so a 404 for a single
    missing DIND left all the other pods unreconciled. That pod is now skipped
    with a warning, which is also safer than inventing a terminal status: the
    DIND is legitimately absent between the claim and the rename. An all-skipped
    response marshals to [] rather than null.

  DindHandler.go
  - Tear the DIND down before returning its subnet to the pool. The failure
    paths freed the subnet while the container and its network still existed, so
    the next network create for that range failed with "Pool overlaps with other
    one on this address space" — permanently, since the reaper cannot remove a
    network whose container is alive. rollbackDind removes container, then
    network, then frees the subnet, and deliberately leaks the subnet with a
    loud error if the network cannot be removed.
  - Check the exit code of docker network create. go-execute returns a nil error
    for a non-zero exit, so a rejected create was treated as a success and only
    surfaced later as an unrelated docker run failure.

  AMDHandler.go
  - Track FPGA assignees as a list instead of a single ContainerID, so sharing
    under FPGA_DISABLE_BOOKKEEPING keeps every assignee and Release can find
    each one instead of only the last. Assign/GetAvailableFPGAs no longer
    double-lock FPGASpecsMutex (added assignLocked/getAvailableFPGAsLocked), the
    fallback sort runs on a copy rather than reordering the shared list under no
    lock, and the same physical device is no longer returned twice in one call.

  Create_test.go
  - Unit tests for selectNetworkOverlay, including the no-overlay regression and
    the empty-list panic.

Brings the reliability work from the light branch onto main without adopting
its removal of GPU support. pkg/docker/gpustrategies/ is kept unchanged, and
the GPU wiring the branch had deleted is restored on top of the merged code:
the gpuManager init in cmd/main.go, SidecarHandler.GpuManager, the
nvidia.com/gpu request block in Create.go, and the GPU release in Delete.go.

CGO_ENABLED stays at 1 in Makefile and .goreleaser.yaml: the branch could set
it to 0 only because it had dropped go-nvml, which does not build without cgo.

What the merge brings in:

- The ResponseWriter is no longer written after the create handler returned,
  and the asynchronous container-creation goroutine now recovers panics
  instead of taking the whole plugin down with it.
- A pod whose DIND container is missing is skipped in the status response
  rather than failing the whole batch with a 404, which used to leave every
  other pod on the node unreconciled.
- envVars, fpgaArgs and gpuArgs are scoped to a single container; they
  previously leaked from one container to the next.
- DIND subnet pool: rollback on failed creates, an orphan-network reaper, and
  the missing exit-code check on docker network create.
- The wait for the mesh_ready sentinel is bounded, and the network-overlay
  container is selected by name instead of assumed to be containers[0].
- FPGA assignment tracks every assignee, no longer double-locks its mutex, and
  no longer sorts the shared list without holding one.