summaryrefslogtreecommitdiffhomepage
path: root/runsc/boot/vfs.go
diff options
context:
space:
mode:
Diffstat (limited to 'runsc/boot/vfs.go')
-rw-r--r--runsc/boot/vfs.go212
1 files changed, 112 insertions, 100 deletions
diff --git a/runsc/boot/vfs.go b/runsc/boot/vfs.go
index 147c901c4..8eeb43e79 100644
--- a/runsc/boot/vfs.go
+++ b/runsc/boot/vfs.go
@@ -22,22 +22,21 @@ import (
specs "github.com/opencontainers/runtime-spec/specs-go"
"gvisor.dev/gvisor/pkg/abi/linux"
+ "gvisor.dev/gvisor/pkg/context"
"gvisor.dev/gvisor/pkg/fspath"
+ "gvisor.dev/gvisor/pkg/log"
"gvisor.dev/gvisor/pkg/sentry/devices/memdev"
- "gvisor.dev/gvisor/pkg/sentry/fs"
+ "gvisor.dev/gvisor/pkg/sentry/fs/user"
"gvisor.dev/gvisor/pkg/sentry/fsimpl/devpts"
"gvisor.dev/gvisor/pkg/sentry/fsimpl/devtmpfs"
"gvisor.dev/gvisor/pkg/sentry/fsimpl/gofer"
"gvisor.dev/gvisor/pkg/sentry/fsimpl/proc"
"gvisor.dev/gvisor/pkg/sentry/fsimpl/sys"
"gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs"
- "gvisor.dev/gvisor/pkg/syserror"
-
- "gvisor.dev/gvisor/pkg/context"
- "gvisor.dev/gvisor/pkg/log"
"gvisor.dev/gvisor/pkg/sentry/kernel"
"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
"gvisor.dev/gvisor/pkg/sentry/vfs"
+ "gvisor.dev/gvisor/pkg/syserror"
)
func registerFilesystems(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials) error {
@@ -95,69 +94,14 @@ func setupContainerVFS2(ctx context.Context, conf *Config, mntr *containerMounte
return fmt.Errorf("failed to setupFS: %w", err)
}
procArgs.MountNamespaceVFS2 = mns
- return setExecutablePathVFS2(ctx, procArgs)
-}
-
-func setExecutablePathVFS2(ctx context.Context, procArgs *kernel.CreateProcessArgs) error {
- exe := procArgs.Argv[0]
-
- // Absolute paths can be used directly.
- if path.IsAbs(exe) {
- procArgs.Filename = exe
- return nil
- }
-
- // Paths with '/' in them should be joined to the working directory, or
- // to the root if working directory is not set.
- if strings.IndexByte(exe, '/') > 0 {
- if !path.IsAbs(procArgs.WorkingDirectory) {
- return fmt.Errorf("working directory %q must be absolute", procArgs.WorkingDirectory)
- }
- procArgs.Filename = path.Join(procArgs.WorkingDirectory, exe)
- return nil
- }
-
- // Paths with a '/' are relative to the CWD.
- if strings.IndexByte(exe, '/') > 0 {
- procArgs.Filename = path.Join(procArgs.WorkingDirectory, exe)
- return nil
- }
- // Otherwise, We must lookup the name in the paths, starting from the
- // root directory.
- root := procArgs.MountNamespaceVFS2.Root()
- defer root.DecRef()
-
- paths := fs.GetPath(procArgs.Envv)
- creds := procArgs.Credentials
-
- for _, p := range paths {
- binPath := path.Join(p, exe)
- pop := &vfs.PathOperation{
- Root: root,
- Start: root,
- Path: fspath.Parse(binPath),
- FollowFinalSymlink: true,
- }
- opts := &vfs.OpenOptions{
- FileExec: true,
- Flags: linux.O_RDONLY,
- }
- dentry, err := root.Mount().Filesystem().VirtualFilesystem().OpenAt(ctx, creds, pop, opts)
- if err == syserror.ENOENT || err == syserror.EACCES {
- // Didn't find it here.
- continue
- }
- if err != nil {
- return err
- }
- dentry.DecRef()
-
- procArgs.Filename = binPath
- return nil
+ // Resolve the executable path from working dir and environment.
+ resolved, err := user.ResolveExecutablePath(ctx, procArgs)
+ if err != nil {
+ return err
}
-
- return fmt.Errorf("executable %q not found in $PATH=%q", exe, strings.Join(paths, ":"))
+ procArgs.Filename = resolved
+ return nil
}
func (c *containerMounter) setupVFS2(ctx context.Context, conf *Config, procArgs *kernel.CreateProcessArgs) (*vfs.MountNamespace, error) {
@@ -192,7 +136,7 @@ func (c *containerMounter) setupVFS2(ctx context.Context, conf *Config, procArgs
func (c *containerMounter) createMountNamespaceVFS2(ctx context.Context, conf *Config, creds *auth.Credentials) (*vfs.MountNamespace, error) {
fd := c.fds.remove()
- opts := strings.Join(p9MountOptions(fd, conf.FileAccess, true /* vfs2 */), ",")
+ opts := strings.Join(p9MountData(fd, conf.FileAccess, true /* vfs2 */), ",")
log.Infof("Mounting root over 9P, ioFD: %d", fd)
mns, err := c.k.VFS().NewMountNamespace(ctx, creds, "", gofer.Name, &vfs.GetFilesystemOptions{Data: opts})
@@ -216,8 +160,9 @@ func (c *containerMounter) mountSubmountsVFS2(ctx context.Context, conf *Config,
}
}
- // TODO(gvisor.dev/issue/1487): implement mountTmp from fs.go.
-
+ if err := c.mountTmpVFS2(ctx, conf, creds, mns); err != nil {
+ return fmt.Errorf(`mount submount "\tmp": %w`, err)
+ }
return nil
}
@@ -235,7 +180,7 @@ func (c *containerMounter) prepareMountsVFS2() ([]mountAndFD, error) {
fd := -1
// Only bind mounts use host FDs; see
// containerMounter.getMountNameAndOptionsVFS2.
- if m.Type == bind || isBindMount(m) {
+ if m.Type == bind {
fd = c.fds.remove()
}
mounts = append(mounts, mountAndFD{
@@ -255,8 +200,6 @@ func (c *containerMounter) prepareMountsVFS2() ([]mountAndFD, error) {
return mounts, nil
}
-// TODO(gvisor.dev/issue/1487): Implement submount options similar to the VFS1
-// version.
func (c *containerMounter) mountSubmountVFS2(ctx context.Context, conf *Config, mns *vfs.MountNamespace, creds *auth.Credentials, submount *mountAndFD) error {
root := mns.Root()
defer root.DecRef()
@@ -265,12 +208,11 @@ func (c *containerMounter) mountSubmountVFS2(ctx context.Context, conf *Config,
Start: root,
Path: fspath.Parse(submount.Destination),
}
-
- fsName, options, useOverlay, err := c.getMountNameAndOptionsVFS2(conf, submount)
+ fsName, opts, err := c.getMountNameAndOptionsVFS2(conf, submount)
if err != nil {
return fmt.Errorf("mountOptions failed: %w", err)
}
- if fsName == "" {
+ if len(fsName) == 0 {
// Filesystem is not supported (e.g. cgroup), just skip it.
return nil
}
@@ -278,17 +220,6 @@ func (c *containerMounter) mountSubmountVFS2(ctx context.Context, conf *Config,
if err := c.makeSyntheticMount(ctx, submount.Destination, root, creds); err != nil {
return err
}
-
- opts := &vfs.MountOptions{
- GetFilesystemOptions: vfs.GetFilesystemOptions{
- Data: strings.Join(options, ","),
- },
- InternalMount: true,
- }
-
- // All writes go to upper, be paranoid and make lower readonly.
- opts.ReadOnly = useOverlay
-
if err := c.k.VFS().MountAt(ctx, creds, "", target, fsName, opts); err != nil {
return fmt.Errorf("failed to mount %q (type: %s): %w, opts: %v", submount.Destination, submount.Type, err, opts)
}
@@ -298,17 +229,13 @@ func (c *containerMounter) mountSubmountVFS2(ctx context.Context, conf *Config,
// getMountNameAndOptionsVFS2 retrieves the fsName, opts, and useOverlay values
// used for mounts.
-func (c *containerMounter) getMountNameAndOptionsVFS2(conf *Config, m *mountAndFD) (string, []string, bool, error) {
+func (c *containerMounter) getMountNameAndOptionsVFS2(conf *Config, m *mountAndFD) (string, *vfs.MountOptions, error) {
var (
- fsName string
- opts []string
- useOverlay bool
+ fsName string
+ data []string
)
- if isBindMount(m.Mount) {
- m.Type = bind
- }
-
+ // Find filesystem name and FS specific data field.
switch m.Type {
case devpts.Name, devtmpfs.Name, proc.Name, sys.Name:
fsName = m.Type
@@ -318,21 +245,46 @@ func (c *containerMounter) getMountNameAndOptionsVFS2(conf *Config, m *mountAndF
fsName = m.Type
var err error
- opts, err = parseAndFilterOptions(m.Options, tmpfsAllowedOptions...)
+ data, err = parseAndFilterOptions(m.Options, tmpfsAllowedData...)
if err != nil {
- return "", nil, false, err
+ return "", nil, err
}
case bind:
fsName = gofer.Name
- opts = p9MountOptions(m.fd, c.getMountAccessType(m.Mount), true /* vfs2 */)
- // If configured, add overlay to all writable mounts.
- useOverlay = conf.Overlay && !mountFlags(m.Options).ReadOnly
+ data = p9MountData(m.fd, c.getMountAccessType(m.Mount), true /* vfs2 */)
default:
log.Warningf("ignoring unknown filesystem type %q", m.Type)
}
- return fsName, opts, useOverlay, nil
+
+ opts := &vfs.MountOptions{
+ GetFilesystemOptions: vfs.GetFilesystemOptions{
+ Data: strings.Join(data, ","),
+ },
+ InternalMount: true,
+ }
+
+ for _, o := range m.Options {
+ switch o {
+ case "rw":
+ opts.ReadOnly = false
+ case "ro":
+ opts.ReadOnly = true
+ case "noatime":
+ opts.Flags.NoATime = true
+ case "noexec":
+ opts.Flags.NoExec = true
+ default:
+ log.Warningf("ignoring unknown mount option %q", o)
+ }
+ }
+
+ if conf.Overlay {
+ // All writes go to upper, be paranoid and make lower readonly.
+ opts.ReadOnly = true
+ }
+ return fsName, opts, nil
}
func (c *containerMounter) makeSyntheticMount(ctx context.Context, currentPath string, root vfs.VirtualDentry, creds *auth.Credentials) error {
@@ -361,3 +313,63 @@ func (c *containerMounter) makeSyntheticMount(ctx context.Context, currentPath s
}
return nil
}
+
+// mountTmpVFS2 mounts an internal tmpfs at '/tmp' if it's safe to do so.
+// Technically we don't have to mount tmpfs at /tmp, as we could just rely on
+// the host /tmp, but this is a nice optimization, and fixes some apps that call
+// mknod in /tmp. It's unsafe to mount tmpfs if:
+// 1. /tmp is mounted explicitly: we should not override user's wish
+// 2. /tmp is not empty: mounting tmpfs would hide existing files in /tmp
+//
+// Note that when there are submounts inside of '/tmp', directories for the
+// mount points must be present, making '/tmp' not empty anymore.
+func (c *containerMounter) mountTmpVFS2(ctx context.Context, conf *Config, creds *auth.Credentials, mns *vfs.MountNamespace) error {
+ for _, m := range c.mounts {
+ // m.Destination has been cleaned, so it's to use equality here.
+ if m.Destination == "/tmp" {
+ log.Debugf(`Explict "/tmp" mount found, skipping internal tmpfs, mount: %+v`, m)
+ return nil
+ }
+ }
+
+ root := mns.Root()
+ defer root.DecRef()
+ pop := vfs.PathOperation{
+ Root: root,
+ Start: root,
+ Path: fspath.Parse("/tmp"),
+ }
+ // TODO(gvisor.dev/issue/2782): Use O_PATH when available.
+ statx, err := c.k.VFS().StatAt(ctx, creds, &pop, &vfs.StatOptions{})
+ switch err {
+ case nil:
+ // Found '/tmp' in filesystem, check if it's empty.
+ if linux.FileMode(statx.Mode).FileType() != linux.ModeDirectory {
+ // Not a dir?! Leave it be.
+ return nil
+ }
+ if statx.Nlink > 2 {
+ // If more than "." and ".." is found, skip internal tmpfs to prevent
+ // hiding existing files.
+ log.Infof(`Skipping internal tmpfs mount for "/tmp" because it's not empty`)
+ return nil
+ }
+ log.Infof(`Mounting internal tmpfs on top of empty "/tmp"`)
+ fallthrough
+
+ case syserror.ENOENT:
+ // No '/tmp' found (or fallthrough from above). It's safe to mount internal
+ // tmpfs.
+ tmpMount := specs.Mount{
+ Type: tmpfs.Name,
+ Destination: "/tmp",
+ // Sticky bit is added to prevent accidental deletion of files from
+ // another user. This is normally done for /tmp.
+ Options: []string{"mode=01777"},
+ }
+ return c.mountSubmountVFS2(ctx, conf, mns, creds, &mountAndFD{Mount: tmpMount})
+
+ default:
+ return fmt.Errorf(`stating "/tmp" inside container: %w`, err)
+ }
+}