feat(nvsnap): import gpushare, checkpoint support for GPU memory shared between processes

.github/workflows/nvsnap-gpushare.yml

@@ -0,0 +1,53 @@
+# SPDX-FileCopyrightText: Copyright (c) NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Compile nvsnap's gpushare C code (libnvsnap_gpushare.so, nvsnap-gpu-suspend
+# and its GPU tests) for amd64 and arm64 with warnings as errors. The Bazel row
+# for nvsnap covers its Go code only, so without this job a change to the C
+# code is first compiled when the agent base image is built. Runs only when the
+# C code, its build or this workflow changes; no GPU is needed to compile.
+
+name: nvsnap gpushare
+
+on:
+  push:
+    branches: [main, 'release-**']
+    paths: &paths
+      - 'src/compute-plane-services/nvsnap/docker/agent/gpushare/**'
+      - 'src/compute-plane-services/nvsnap/tests/gpushare/**'
+      - 'src/compute-plane-services/nvsnap/docker/agent/Dockerfile.base'
+      - 'src/compute-plane-services/nvsnap/scripts/check-gpushare-build.sh'
+      - '.github/workflows/nvsnap-gpushare.yml'
+  pull_request:
+    branches: [main, 'release-**']
+    paths: *paths
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: nvsnap-gpushare-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  build:
+    name: build gpushare (${{ matrix.platform }})
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        platform: [linux/amd64, linux/arm64]
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          persist-credentials: false
+
+      - name: Set up QEMU
+        if: matrix.platform == 'linux/arm64'
+        uses: docker/setup-qemu-action@c7c53464625b32c7a7e944ae62b3e17d2b600130 # v3.7.0
+
+      - name: Compile with warnings as errors
+        env:
+          PLATFORMS: ${{ matrix.platform }}
+        run: src/compute-plane-services/nvsnap/scripts/check-gpushare-build.sh

src/compute-plane-services/nvsnap/Makefile

@@ -9,7 +9,7 @@ CONTAINER_TOOL ?= docker
 GO_FLAGS ?= -ldflags="-s -w -X main.version=$(VERSION)"
 BUILD_DIR ?= bin
 
-.PHONY: all build agent nvsnap-server clean test test-coverage test-checkpoint test-server deploy install-crds generate
+.PHONY: all build agent nvsnap-server clean test test-coverage test-checkpoint test-server deploy install-crds generate check-gpushare
 
 all: build
 
@@ -70,6 +70,10 @@ test:
 
 # Run all unit tests with coverage profile (used by CI).
 # Coverage output lands in coverage/coverage.out for Codecov upload.
+# Compile the gpushare C code for amd64 and arm64 (Docker, no GPU); same check as CI.
+check-gpushare:
+	./scripts/check-gpushare-build.sh
+
 test-coverage:
 	mkdir -p coverage
 	go test -v -race -coverprofile=coverage/coverage.out -covermode=atomic $(GO_TEST_PKGS)

src/compute-plane-services/nvsnap/docker/agent/Dockerfile.base

@@ -22,6 +22,16 @@ RUN gcc -O2 -o /src/nvsnap-cuda-checkpoint /src/nvsnap-cuda-checkpoint.c \
     strings /src/nvsnap-cuda-checkpoint | grep -q "nvsnap-cuda-checkpoint" \
         || (echo "provenance marker missing" && exit 1)
 
+# gpushare: libnvsnap_gpushare.so, an LD_PRELOAD shim that releases and
+# re-maps GPU memory shared between a workload's processes (NCCL P2P and NVLS,
+# CUDA IPC, page-locked host memory) around a CUDA checkpoint, and
+# nvsnap-gpu-suspend, which drives it. Needs CUDA 13 headers
+# (CUcheckpointGpuPair). Built on ubuntu 22.04: the shim is loaded into
+# workload containers, which need glibc >= 2.35.
+FROM nvidia/cuda:13.0.3-devel-ubuntu22.04 AS gpushare-builder
+COPY gpushare /src/gpushare
+RUN make -C /src/gpushare
+
 FROM ubuntu:22.04 AS criu-builder
 
 # Install build dependencies
@@ -165,6 +175,9 @@ COPY --from=cuda-cli-builder /src/nvsnap-cuda-checkpoint /criu-bundle/cuda-check
 COPY cuda-checkpoint-wrapper.sh /criu-bundle/cuda-checkpoint
 RUN chmod +x /criu-bundle/cuda-checkpoint
 
+# gpushare (see gpushare-builder above).
+COPY --from=gpushare-builder /src/gpushare/libnvsnap_gpushare.so /src/gpushare/nvsnap-gpu-suspend /criu-bundle/
+
 # Make everything executable
 RUN chmod +x /criu-bundle/*
 

src/compute-plane-services/nvsnap/docker/agent/gpushare/Makefile

@@ -0,0 +1,24 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# libnvsnap_gpushare.so: LD_PRELOAD shim that releases and re-maps GPU memory
+# shared between processes around a CUDA checkpoint (see gpushare.h).
+# nvsnap-gpu-suspend: drives it and the CUDA checkpoint API.
+# Needs CUDA >= 13 headers (cuda.h: CUcheckpointGpuPair for --gpu-map).
+
+CC ?= gcc
+CUDA_INC ?= /usr/local/cuda/include
+CFLAGS ?= -O2 -Wall -Wextra
+
+all: libnvsnap_gpushare.so nvsnap-gpu-suspend
+
+libnvsnap_gpushare.so: gpushare.c gpushare.h
+	$(CC) $(CFLAGS) -fPIC -shared -I. -I$(CUDA_INC) -o $@ gpushare.c -ldl -lpthread
+
+nvsnap-gpu-suspend: nvsnap-gpu-suspend.c gpushare.h
+	$(CC) $(CFLAGS) -I. -I$(CUDA_INC) -o $@ nvsnap-gpu-suspend.c -ldl -lpthread
+
+clean:
+	rm -f libnvsnap_gpushare.so nvsnap-gpu-suspend
+
+.PHONY: all clean