Add mount namespaces to linux sandbox

Cargo.toml

@@ -29,6 +29,11 @@ name = "fs"
 path = "tests/fs.rs"
 harness = false
 
+[[test]]
+name = "fs_without_landlock"
+path = "tests/fs_without_landlock.rs"
+harness = false
+
 [[test]]
 name = "full_env"
 path = "tests/full_env.rs"

src/linux/mod.rs

@@ -3,10 +3,9 @@
 //! This module implements sandboxing on Linux based on the Landlock LSM,
 //! namespaces, and seccomp.
 
-use std::fs;
-use std::io::Error as IoError;
+use std::collections::HashMap;
+use std::path::PathBuf;
 
-use bitflags::bitflags;
 use landlock::{
     make_bitflags, Access, AccessFs, Compatible, PathBeneath, PathFd, Ruleset, RulesetAttr,
     RulesetCreated, RulesetCreatedAttr, RulesetStatus, ABI as LANDLOCK_ABI,
@@ -16,19 +15,42 @@ use crate::error::{Error, Result};
 use crate::linux::seccomp::NetworkFilter;
 use crate::{Exception, Sandbox};
 
+mod namespaces;
 mod seccomp;
 
 /// Minimum landlock ABI version.
 const ABI: LANDLOCK_ABI = LANDLOCK_ABI::V1;
 
 /// Linux sandboxing.
 pub struct LinuxSandbox {
+    bind_mounts: HashMap<PathBuf, libc::c_ulong>,
     env_exceptions: Vec<String>,
     landlock: RulesetCreated,
     allow_networking: bool,
     full_env: bool,
 }
 
+impl LinuxSandbox {
+    /// Add or modify a bind mount.
+    ///
+    /// This will add a new bind mount with the specified permission if it does
+    /// not exist already.
+    ///
+    /// If the bind mount already exists, it will *ADD* the additional
+    /// permissions.
+    fn update_bind_mount(&mut self, path: PathBuf, write: bool, execute: bool) {
+        let flags = self.bind_mounts.entry(path).or_insert(libc::MS_RDONLY | libc::MS_NOEXEC);
+
+        if write {
+            *flags &= !libc::MS_RDONLY;
+        }
+
+        if execute {
+            *flags &= !libc::MS_NOEXEC;
+        }
+    }
+}
+
 impl Sandbox for LinuxSandbox {
     fn new() -> Result<Self> {
         // Setup landlock filtering.
@@ -38,14 +60,38 @@ impl Sandbox for LinuxSandbox {
             .create()?;
         landlock.as_mut().set_no_new_privs(true);
 
-        Ok(Self { landlock, env_exceptions: Vec::new(), allow_networking: false, full_env: false })
+        Ok(Self {
+            landlock,
+            allow_networking: false,
+            full_env: false,
+            env_exceptions: Default::default(),
+            bind_mounts: Default::default(),
+        })
     }
 
     fn add_exception(&mut self, exception: Exception) -> Result<&mut Self> {
-        let (path, access) = match exception {
-            Exception::Read(path) => (path, make_bitflags!(AccessFs::{ ReadFile | ReadDir })),
-            Exception::Write(path) => (path, AccessFs::from_write(ABI)),
-            Exception::ExecuteAndRead(path) => (path, AccessFs::from_read(ABI)),
+        let (path_fd, access) = match exception {
+            Exception::Read(path) => {
+                let path_fd = PathFd::new(&path)?;
+
+                self.update_bind_mount(path, false, false);
+
+                (path_fd, make_bitflags!(AccessFs::{ ReadFile | ReadDir }))
+            },
+            Exception::Write(path) => {
+                let path_fd = PathFd::new(&path)?;
+
+                self.update_bind_mount(path, true, false);
+
+                (path_fd, AccessFs::from_write(ABI))
+            },
+            Exception::ExecuteAndRead(path) => {
+                let path_fd = PathFd::new(&path)?;
+
+                self.update_bind_mount(path, false, true);
+
+                (path_fd, AccessFs::from_read(ABI))
+            },
             Exception::Environment(key) => {
                 self.env_exceptions.push(key);
                 return Ok(self);
@@ -60,7 +106,7 @@ impl Sandbox for LinuxSandbox {
             },
         };
 
-        let rule = PathBeneath::new(PathFd::new(path)?, access);
+        let rule = PathBeneath::new(path_fd, access);
 
         self.landlock.as_mut().add_rule(rule)?;
 
@@ -73,12 +119,23 @@ impl Sandbox for LinuxSandbox {
             crate::restrict_env_variables(&self.env_exceptions);
         }
 
-        // Clear abstract namespace by entering a new user namespace.
-        let _ = create_user_namespace(false);
+        // Setup namespaces.
+        let namespace_result =
+            namespaces::create_namespaces(self.allow_networking, self.bind_mounts);
 
-        // Create network namespace.
+        // Setup seccomp network filter.
         if !self.allow_networking {
-            restrict_networking()?;
+            let seccomp_result = NetworkFilter::apply();
+
+            // Propagate failure if neither seccomp nor namespaces could isolate networking.
+            if let (Err(_), Err(err)) = (&namespace_result, seccomp_result) {
+                return Err(err);
+            }
+        }
+
+        // Use landlock only if namespaces failed.
+        if namespace_result.is_ok() {
+            return Ok(());
         }
 
         // Apply landlock rules.
@@ -92,99 +149,3 @@ impl Sandbox for LinuxSandbox {
         }
     }
 }
-
-/// Restrict networking using seccomp and namespaces.
-fn restrict_networking() -> Result<()> {
-    // Create network namespace.
-    let result = create_user_namespace(true).and_then(|_| unshare(Namespaces::NETWORK));
-
-    // Apply seccomp network filter.
-    let seccomp_result = NetworkFilter::apply();
-    result.or(seccomp_result)
-}
-
-/// Create a new user namespace.
-///
-/// If the `become_root` flag is set, then the current user will be mapped to
-/// UID 0 inside the namespace. Otherwise the current user will be mapped to its
-/// UID of the parent namespace.
-fn create_user_namespace(become_root: bool) -> Result<()> {
-    // Get the current UID/GID.
-    let uid = unsafe { libc::geteuid() };
-    let gid = unsafe { libc::getegid() };
-
-    // Create the namespace.
-    unshare(Namespaces::USER)?;
-
-    // Map the UID and GID.
-    let uid_map = if become_root { format!("0 {uid} 1\n") } else { format!("{uid} {uid} 1\n") };
-    let gid_map = if become_root { format!("0 {gid} 1\n") } else { format!("{gid} {gid} 1\n") };
-    fs::write("/proc/self/uid_map", uid_map.as_bytes())?;
-    fs::write("/proc/self/setgroups", b"deny")?;
-    fs::write("/proc/self/gid_map", gid_map.as_bytes())?;
-
-    Ok(())
-}
-
-/// Enter a namespace.
-fn unshare(namespaces: Namespaces) -> Result<()> {
-    let result = unsafe { libc::unshare(namespaces.bits()) };
-    if result == 0 {
-        Ok(())
-    } else {
-        Err(IoError::last_os_error().into())
-    }
-}
-
-bitflags! {
-    /// Unshare system call namespace flags.
-    #[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
-    struct Namespaces: libc::c_int {
-        /// Unshare the file descriptor table, so that the calling process no longer
-        /// shares its file descriptors with any other process.
-        const FILES = libc::CLONE_FILES;
-        /// Unshare filesystem attributes, so that the calling process no longer shares
-        /// its root directory, current directory, or umask attributes with any other process.
-        const FS = libc::CLONE_FS;
-        /// Unshare the cgroup namespace.
-        const CGROUP = libc::CLONE_NEWCGROUP;
-        /// Unshare the IPC namespace, so that the calling process has a private copy of
-        /// the IPC namespace which is not shared with any other process. Specifying
-        /// this flag automatically implies [`Namespaces::SYSVSEM`] as well.
-        const IPC = libc::CLONE_NEWIPC;
-        /// Unshare the network namespace, so that the calling process is moved into a
-        /// new network namespace which is not shared with any previously existing process.
-        const NETWORK = libc::CLONE_NEWNET;
-        /// Unshare the mount namespace, so that the calling process has a private copy
-        /// of its namespace which is not shared with any other process. Specifying this
-        /// flag automatically implies [`Namespaces::FS`] as well.
-        const MOUNT = libc::CLONE_NEWNS;
-        /// Unshare the PID namespace, so that the calling process has a new PID
-        /// namespace for its children which is not shared with any previously existing
-        /// process. The calling process is **not** moved into the new namespace. The
-        /// first child created by the calling process will have the process ID 1 and
-        /// will assume the role of init in the new namespace. Specifying this flag
-        /// automatically implies [`libc::CLONE_THREAD`] as well.
-        const PID = libc::CLONE_NEWPID;
-        /// Unshare the time namespace, so that the calling process has a new time
-        /// namespace for its children which is not shared with any previously existing
-        /// process. The calling process is **not** moved into the new namespace.
-        const TIME = 0x80;
-        /// Unshare the user namespace, so that the calling process is moved into a new
-        /// user namespace which is not shared with any previously existing process. The
-        /// caller obtains a full set of capabilities in the new namespace.
-        ///
-        /// Requires that the calling process is not threaded; specifying this flag
-        /// automatically implies [`libc::CLONE_THREAD`] and [`Namespaces::FS`] as well.
-        const USER = libc::CLONE_NEWUSER;
-        /// Unshare the UTS IPC namespace, so that the calling process has a private
-        /// copy of the UTS namespace which is not shared with any other process.
-        const UTS = libc::CLONE_NEWUTS;
-        /// Unshare System V semaphore adjustment (semadj) values, so that the calling
-        /// process has a new empty semadj list that is not shared with any other
-        /// process. If this is the last process that has a reference to the process's
-        /// current semadj list, then the adjustments in that list are applied to the
-        /// corresponding semaphores
-        const SYSVSEM = libc::CLONE_SYSVSEM;
-    }
-}