Set no_new_privs for container processes

Sources/Containerization/LinuxProcessConfiguration.swift

@@ -372,6 +372,8 @@ public struct LinuxProcessConfiguration: Sendable {
     public var user: ContainerizationOCI.User = .init()
     /// The rlimits for the container process.
     public var rlimits: [LinuxRLimit] = []
+    /// Prevents the process from gaining additional privileges.
+    public var noNewPrivileges: Bool = false
     /// The Linux capabilities for the container process.
     public var capabilities: LinuxCapabilities = .allCapabilities
     /// Whether to allocate a pseudo terminal for the process. If you'd like interactive
@@ -437,6 +439,7 @@ public struct LinuxProcessConfiguration: Sendable {
             args: self.arguments,
             cwd: self.workingDirectory,
             env: self.environmentVariables,
+            noNewPrivileges: self.noNewPrivileges,
             capabilities: self.capabilities.toOCI(),
             user: self.user,
             rlimits: self.rlimits.map { $0.toOCI() },

vminitd/Sources/LCShim/include/syscall.h

@@ -23,6 +23,8 @@ int CZ_pivot_root(const char *new_root, const char *put_old);
 
 int CZ_set_sub_reaper();
 
+int CZ_set_no_new_privs();
+
 #ifndef SYS_pidfd_open
 #define SYS_pidfd_open 434
 #endif

vminitd/Sources/LCShim/syscall.c

@@ -26,6 +26,8 @@ int CZ_pivot_root(const char *new_root, const char *put_old) {
 
 int CZ_set_sub_reaper() { return prctl(PR_SET_CHILD_SUBREAPER, 1); }
 
+int CZ_set_no_new_privs() { return prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0); }
+
 int CZ_pidfd_open(pid_t pid, unsigned int flags) {
   // Musl doesn't have pidfd_open.
   return syscall(SYS_pidfd_open, pid, flags);

vminitd/Sources/vmexec/ExecCommand.swift

@@ -133,18 +133,23 @@ struct ExecCommand: ParsableCommand {
             try App.applyCloseExecOnFDs()
             try App.setRLimits(rlimits: process.rlimits)
 
-            // Prepare capabilities (before user change)
+            // Prepare capabilities (before user change).
             let preparedCaps = try App.prepareCapabilities(capabilities: process.capabilities ?? ContainerizationOCI.LinuxCapabilities())
 
             // Change stdio to be owned by the requested user.
             try App.fixStdioPerms(user: process.user)
 
-            // Set uid, gid, and supplementary groups
+            // Set uid, gid, and supplementary groups.
             try App.setPermissions(user: process.user)
 
-            // Finish capabilities (after user change)
+            // Finish capabilities (after user change).
             try App.finishCapabilities(preparedCaps)
 
+            // Set no_new_privs (after user/capability changes).
+            if process.noNewPrivileges {
+                try App.setNoNewPrivs()
+            }
+
             try App.exec(process: process, currentEnv: process.env)
         } else {  // parent process
             // Send our child's pid to our parent before we exit.

vminitd/Sources/vmexec/RunCommand.swift

@@ -167,7 +167,7 @@ struct RunCommand: ParsableCommand {
 
         try App.setRLimits(rlimits: process.rlimits)
 
-        // Prepare capabilities (before user change)
+        // Prepare capabilities (before user change).
         let preparedCaps = try App.prepareCapabilities(capabilities: process.capabilities ?? ContainerizationOCI.LinuxCapabilities())
 
         // Change stdio to be owned by the requested user.
@@ -176,9 +176,14 @@ struct RunCommand: ParsableCommand {
         // Set uid, gid, and supplementary groups.
         try App.setPermissions(user: process.user)
 
-        // Finish capabilities (after user change)
+        // Finish capabilities (after user change).
         try App.finishCapabilities(preparedCaps)
 
+        // Set no_new_privs (after user/capability changes).
+        if process.noNewPrivileges {
+            try App.setNoNewPrivs()
+        }
+
         // Finally execve the container process.
         try App.exec(process: process, currentEnv: process.env)
     }

vminitd/Sources/vmexec/vmexec.swift

@@ -106,6 +106,12 @@ extension App {
         fatalError("execvpe failed")
     }
 
+    static func setNoNewPrivs() throws {
+        guard CZ_set_no_new_privs() == 0 else {
+            throw App.Errno(stage: "prctl(PR_SET_NO_NEW_PRIVS)")
+        }
+    }
+
     static func setPermissions(user: ContainerizationOCI.User) throws {
         if user.additionalGids.count > 0 {
             guard setgroups(user.additionalGids.count, user.additionalGids) == 0 else {