fix(xunix): improve handling of gpu library mounts

integration/gpu_test.go

@@ -9,6 +9,7 @@ import (
 	"testing"
 	"time"
 
+	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 
 	"github.com/coder/envbox/integration/integrationtest"
@@ -41,8 +42,7 @@ func TestDocker_Nvidia(t *testing.T) {
 		)
 
 		// Assert that we can run nvidia-smi in the inner container.
-		_, err := execContainerCmd(ctx, t, ctID, "docker", "exec", "workspace_cvm", "nvidia-smi")
-		require.NoError(t, err, "failed to run nvidia-smi in the inner container")
+		assertInnerNvidiaSMI(ctx, t, ctID)
 	})
 
 	t.Run("Redhat", func(t *testing.T) {
@@ -52,16 +52,23 @@ func TestDocker_Nvidia(t *testing.T) {
 
 		// Start the envbox container.
 		ctID := startEnvboxCmd(ctx, t, integrationtest.RedhatImage, "root",
-			"-v", "/usr/lib/x86_64-linux-gnu:/var/coder/usr/lib64",
+			"-v", "/usr/lib/x86_64-linux-gnu:/var/coder/usr/lib",
 			"--env", "CODER_ADD_GPU=true",
-			"--env", "CODER_USR_LIB_DIR=/var/coder/usr/lib64",
+			"--env", "CODER_USR_LIB_DIR=/var/coder/usr/lib",
 			"--runtime=nvidia",
 			"--gpus=all",
 		)
 
 		// Assert that we can run nvidia-smi in the inner container.
-		_, err := execContainerCmd(ctx, t, ctID, "docker", "exec", "workspace_cvm", "nvidia-smi")
-		require.NoError(t, err, "failed to run nvidia-smi in the inner container")
+		assertInnerNvidiaSMI(ctx, t, ctID)
+
+		// Make sure dnf still works. This checks for a regression due to
+		// gpuExtraRegex matching `libglib.so` in the outer container.
+		// This had a dependency on `libpcre.so.3` which would cause dnf to fail.
+		out, err := execContainerCmd(ctx, t, ctID, "docker", "exec", "workspace_cvm", "dnf")
+		if !assert.NoError(t, err, "failed to run dnf in the inner container") {
+			t.Logf("dnf output:\n%s", strings.TrimSpace(out))
+		}
 	})
 
 	t.Run("InnerUsrLibDirOverride", func(t *testing.T) {
@@ -79,11 +86,58 @@ func TestDocker_Nvidia(t *testing.T) {
 			"--gpus=all",
 		)
 
-		// Assert that the libraries end up in the expected location in the inner container.
-		out, err := execContainerCmd(ctx, t, ctID, "docker", "exec", "workspace_cvm", "ls", "-l", "/usr/lib/coder")
+		// Assert that the libraries end up in the expected location in the inner
+		// container.
+		out, err := execContainerCmd(ctx, t, ctID, "docker", "exec", "workspace_cvm", "ls", "-1", "/usr/lib/coder")
 		require.NoError(t, err, "inner usr lib dir override failed")
 		require.Regexp(t, `(?i)(libgl|nvidia|vulkan|cuda)`, out)
 	})
+
+	t.Run("EmptyHostUsrLibDir", func(t *testing.T) {
+		t.Parallel()
+		ctx, cancel := context.WithCancel(context.Background())
+		t.Cleanup(cancel)
+		emptyUsrLibDir := t.TempDir()
+
+		// Start the envbox container.
+		ctID := startEnvboxCmd(ctx, t, integrationtest.UbuntuImage, "root",
+			"-v", emptyUsrLibDir+":/var/coder/usr/lib",
+			"--env", "CODER_ADD_GPU=true",
+			"--env", "CODER_USR_LIB_DIR=/var/coder/usr/lib",
+			"--runtime=nvidia",
+			"--gpus=all",
+		)
+
+		ofs := outerFiles(ctx, t, ctID, "/usr/lib/x86_64-linux-gnu/libnv*")
+		// Assert invariant: the outer container has the files we expect.
+		require.NotEmpty(t, ofs, "failed to list outer container files")
+		// Assert that expected files are available in the inner container.
+		assertInnerFiles(ctx, t, ctID, "/usr/lib/x86_64-linux-gnu/libnv*", ofs...)
+		assertInnerNvidiaSMI(ctx, t, ctID)
+	})
+
+	t.Run("CUDASample", func(t *testing.T) {
+		t.Parallel()
+
+		ctx, cancel := context.WithCancel(context.Background())
+		t.Cleanup(cancel)
+
+		// Start the envbox container.
+		ctID := startEnvboxCmd(ctx, t, integrationtest.CUDASampleImage, "root",
+			"-v", "/usr/lib/x86_64-linux-gnu:/var/coder/usr/lib",
+			"--env", "CODER_ADD_GPU=true",
+			"--env", "CODER_USR_LIB_DIR=/var/coder/usr/lib",
+			"--runtime=nvidia",
+			"--gpus=all",
+		)
+
+		// Assert that we can run nvidia-smi in the inner container.
+		assertInnerNvidiaSMI(ctx, t, ctID)
+
+		// Assert that /tmp/vectorAdd runs successfully in the inner container.
+		_, err := execContainerCmd(ctx, t, ctID, "docker", "exec", "workspace_cvm", "/tmp/vectorAdd")
+		require.NoError(t, err, "failed to run /tmp/vectorAdd in the inner container")
+	})
 }
 
 // dockerRuntimes returns the list of container runtimes available on the host.
@@ -101,6 +155,49 @@ func dockerRuntimes(t *testing.T) []string {
 	return strings.Split(raw, "\n")
 }
 
+// outerFiles returns the list of files in the outer container matching the
+// given pattern. It does this by running `ls -1` in the outer container.
+func outerFiles(ctx context.Context, t *testing.T, containerID, pattern string) []string {
+	t.Helper()
+	// We need to use /bin/sh -c to avoid the shell interpreting the glob.
+	out, err := execContainerCmd(ctx, t, containerID, "/bin/sh", "-c", "ls -1 "+pattern)
+	require.NoError(t, err, "failed to list outer container files")
+	files := strings.Split(strings.TrimSpace(out), "\n")
+	slices.Sort(files)
+	return files
+}
+
+// assertInnerFiles checks that all the files matching the given pattern exist in the
+// inner container.
+func assertInnerFiles(ctx context.Context, t *testing.T, containerID, pattern string, expected ...string) {
+	t.Helper()
+
+	// Get the list of files in the inner container.
+	// We need to use /bin/sh -c to avoid the shell interpreting the glob.
+	out, err := execContainerCmd(ctx, t, containerID, "docker", "exec", "workspace_cvm", "/bin/sh", "-c", "ls -1 "+pattern)
+	require.NoError(t, err, "failed to list inner container files")
+	innerFiles := strings.Split(strings.TrimSpace(out), "\n")
+
+	// Check that the expected files exist in the inner container.
+	missingFiles := make([]string, 0)
+	for _, expectedFile := range expected {
+		if !slices.Contains(innerFiles, expectedFile) {
+			missingFiles = append(missingFiles, expectedFile)
+		}
+	}
+	require.Empty(t, missingFiles, "missing files in inner container: %s", strings.Join(missingFiles, ", "))
+}
+
+// assertInnerNvidiaSMI checks that nvidia-smi runs successfully in the inner
+// container.
+func assertInnerNvidiaSMI(ctx context.Context, t *testing.T, containerID string) {
+	t.Helper()
+	// Assert that we can run nvidia-smi in the inner container.
+	out, err := execContainerCmd(ctx, t, containerID, "docker", "exec", "workspace_cvm", "nvidia-smi")
+	require.NoError(t, err, "failed to run nvidia-smi in the inner container")
+	require.Contains(t, out, "NVIDIA-SMI", "nvidia-smi output does not contain NVIDIA-SMI")
+}
+
 // startEnvboxCmd starts the envbox container with the given arguments.
 // Ideally we would use ory/dockertest for this, but it doesn't support
 // specifying the runtime. We have alternatively used the docker client library,

integration/integrationtest/docker.go

@@ -42,6 +42,9 @@ const (
 	UbuntuImage = "gcr.io/coder-dev-1/sreya/ubuntu-coder"
 	// Redhat UBI9 image as of 2025-03-05
 	RedhatImage = "registry.access.redhat.com/ubi9/ubi:9.5"
+	// CUDASampleImage is a CUDA sample image from NVIDIA's container registry.
+	// It contains a binary /tmp/vectorAdd which can be run to test the CUDA setup.
+	CUDASampleImage = "nvcr.io/nvidia/k8s/cuda-sample:vectoradd-cuda10.2"
 
 	// RegistryImage is used to assert that we add certs
 	// correctly to the docker daemon when pulling an image

xunix/gpu.go

@@ -6,6 +6,7 @@ import (
 	"os"
 	"path/filepath"
 	"regexp"
+	"slices"
 	"sort"
 	"strings"
 
@@ -17,9 +18,9 @@ import (
 )
 
 var (
-	gpuMountRegex     = regexp.MustCompile("(?i)(nvidia|vulkan|cuda)")
-	gpuExtraRegex     = regexp.MustCompile("(?i)(libgl|nvidia|vulkan|cuda)")
-	gpuEnvRegex       = regexp.MustCompile("(?i)nvidia")
+	gpuMountRegex     = regexp.MustCompile(`(?i)(nvidia|vulkan|cuda)`)
+	gpuExtraRegex     = regexp.MustCompile(`(?i)(libgl(e|sx|\.)|nvidia|vulkan|cuda)`)
+	gpuEnvRegex       = regexp.MustCompile(`(?i)nvidia`)
 	sharedObjectRegex = regexp.MustCompile(`\.so(\.[0-9\.]+)?$`)
 )
 
@@ -39,6 +40,7 @@ func GPUEnvs(ctx context.Context) []string {
 
 func GPUs(ctx context.Context, log slog.Logger, usrLibDir string) ([]Device, []mount.MountPoint, error) {
 	var (
+		afs     = GetFS(ctx)
 		mounter = Mounter(ctx)
 		devices = []Device{}
 		binds   = []mount.MountPoint{}
@@ -64,6 +66,22 @@ func GPUs(ctx context.Context, log slog.Logger, usrLibDir string) ([]Device, []m
 
 			// If it's not in /dev treat it as a bind mount.
 			binds = append(binds, m)
+			// We also want to find any symlinks that point to the target.
+			// This is important for the nvidia driver as it mounts the driver
+			// files with the driver version appended to the end, and creates
+			// symlinks that point to the actual files.
+			links, err := SameDirSymlinks(afs, m.Path)
+			if err != nil {
+				log.Error(ctx, "find symlinks", slog.F("path", m.Path), slog.Error(err))
+			} else {
+				for _, link := range links {
+					log.Debug(ctx, "found symlink", slog.F("link", link), slog.F("target", m.Path))
+					binds = append(binds, mount.MountPoint{
+						Path: link,
+						Opts: []string{"ro"},
+					})
+				}
+			}
 		}
 	}
 
@@ -104,7 +122,11 @@ func usrLibGPUs(ctx context.Context, log slog.Logger, usrLibDir string) ([]mount
 				return nil
 			}
 
-			if !sharedObjectRegex.MatchString(path) || !gpuExtraRegex.MatchString(path) {
+			if !gpuExtraRegex.MatchString(path) {
+				return nil
+			}
+
+			if !sharedObjectRegex.MatchString(path) {
 				return nil
 			}
 
@@ -176,6 +198,68 @@ func recursiveSymlinks(afs FS, mountpoint string, path string) ([]string, error)
 	return paths, nil
 }
 
+// SameDirSymlinks returns all links in the same directory as `target` that
+// point to target, either indirectly or directly. Only symlinks in the same
+// directory as `target` are considered.
+func SameDirSymlinks(afs FS, target string) ([]string, error) {
+	var (
+		found         = make([]string, 0)
+		maxIterations = 10 // arbitrary upper limit to prevent infinite loops
+	)
+	for range maxIterations {
+		foundThisTime := false
+		fis, err := afero.ReadDir(afs, filepath.Dir(target))
+		if err != nil {
+			return nil, xerrors.Errorf("read dir %q: %w", filepath.Dir(target), err)
+		}
+		for _, fi := range fis {
+			// Ignore the target itself.
+			if fi.Name() == filepath.Base(target) {
+				continue
+			}
+			// Ignore non-symlinks.
+			if fi.Mode()&os.ModeSymlink == 0 {
+				continue
+			}
+			// Get the target of the symlink.
+			link, err := afs.Readlink(filepath.Join(filepath.Dir(target), fi.Name()))
+			if err != nil {
+				return nil, xerrors.Errorf("readlink %q: %w", fi.Name(), err)
+			}
+			// Make the link absolute.
+			if !filepath.IsAbs(link) {
+				link = filepath.Join(filepath.Dir(target), link)
+			}
+			// Ignore symlinks that point outside of target's directory.
+			if filepath.Dir(link) != filepath.Dir(target) {
+				continue
+			}
+
+			// Check if the symlink points to to the target, or if it points
+			// to one of the symlinks we've already found.
+			if link != target {
+				if !slices.Contains(found, link) {
+					continue
+				}
+			}
+
+			// Have we already seen this target?
+			fullPath := filepath.Join(filepath.Dir(target), fi.Name())
+			if slices.Contains(found, fullPath) {
+				continue
+			}
+
+			found = append(found, filepath.Join(filepath.Dir(target), fi.Name()))
+			foundThisTime = true
+		}
+		// If we didn't find any symlinks this time, we're done.
+		if !foundThisTime {
+			break
+		}
+	}
+	return found, nil
+}
+
 // TryUnmountProcGPUDrivers unmounts any GPU-related mounts under /proc as it causes
 // issues when creating any container in some cases. Errors encountered while
 // unmounting are treated as non-fatal.