diff --git a/authfs/fd_server/src/main.rs b/authfs/fd_server/src/main.rs
index 12f013c..32ffc0a 100644
--- a/authfs/fd_server/src/main.rs
+++ b/authfs/fd_server/src/main.rs
@@ -32,6 +32,7 @@
 use std::convert::TryInto;
 use std::fs::File;
 use std::io;
+use std::os::raw;
 use std::os::unix::fs::FileExt;
 use std::os::unix::io::{AsRawFd, FromRawFd};
 
@@ -292,7 +293,12 @@
     Ok((fd, FdConfig::ReadWrite(file)))
 }
 
-fn parse_args() -> Result<BTreeMap<i32, FdConfig>> {
+struct Args {
+    fd_pool: BTreeMap<i32, FdConfig>,
+    ready_fd: Option<File>,
+}
+
+fn parse_args() -> Result<Args> {
     #[rustfmt::skip]
     let matches = clap::App::new("fd_server")
         .arg(clap::Arg::with_name("ro-fds")
@@ -303,6 +309,9 @@
              .long("rw-fds")
              .multiple(true)
              .number_of_values(1))
+        .arg(clap::Arg::with_name("ready-fd")
+            .long("ready-fd")
+            .takes_value(true))
         .get_matches();
 
     let mut fd_pool = BTreeMap::new();
@@ -318,8 +327,13 @@
             fd_pool.insert(fd, config);
         }
     }
-
-    Ok(fd_pool)
+    let ready_fd = if let Some(arg) = matches.value_of("ready-fd") {
+        let fd = arg.parse::<i32>()?;
+        Some(fd_to_file(fd)?)
+    } else {
+        None
+    };
+    Ok(Args { fd_pool, ready_fd })
 }
 
 fn main() -> Result<()> {
@@ -327,16 +341,22 @@
         android_logger::Config::default().with_tag("fd_server").with_min_level(log::Level::Debug),
     );
 
-    let fd_pool = parse_args()?;
+    let args = parse_args()?;
 
-    let mut service = FdService::new_binder(fd_pool).as_binder();
+    let mut service = FdService::new_binder(args.fd_pool).as_binder();
+    let mut ready_notifier = ReadyNotifier(args.ready_fd);
+
     debug!("fd_server is starting as a rpc service.");
     // SAFETY: Service ownership is transferring to the server and won't be valid afterward.
     // Plus the binder objects are threadsafe.
+    // RunRpcServerCallback does not retain a reference to ready_callback, and only ever
+    // calls it with the param we provide during the lifetime of ready_notifier.
     let retval = unsafe {
-        binder_rpc_unstable_bindgen::RunRpcServer(
+        binder_rpc_unstable_bindgen::RunRpcServerCallback(
             service.as_native_mut() as *mut binder_rpc_unstable_bindgen::AIBinder,
             RPC_SERVICE_PORT,
+            Some(ReadyNotifier::ready_callback),
+            ready_notifier.as_void_ptr(),
         )
     };
     if retval {
@@ -346,3 +366,25 @@
         bail!("Premature termination of RPC server");
     }
 }
+
+struct ReadyNotifier(Option<File>);
+
+impl ReadyNotifier {
+    fn notify(&mut self) {
+        debug!("fd_server is ready");
+        // Close the ready-fd if we were given one to signal our readiness.
+        drop(self.0.take());
+    }
+
+    fn as_void_ptr(&mut self) -> *mut raw::c_void {
+        self as *mut _ as *mut raw::c_void
+    }
+
+    unsafe extern "C" fn ready_callback(param: *mut raw::c_void) {
+        // SAFETY: This is only ever called by RunRpcServerCallback, within the lifetime of the
+        // ReadyNotifier, with param taking the value returned by as_void_ptr (so a properly aligned
+        // non-null pointer to an initialized instance).
+        let ready_notifier = param as *mut Self;
+        ready_notifier.as_mut().unwrap().notify()
+    }
+}
diff --git a/authfs/service/src/authfs.rs b/authfs/service/src/authfs.rs
index 5601738..6d87243 100644
--- a/authfs/service/src/authfs.rs
+++ b/authfs/service/src/authfs.rs
@@ -82,8 +82,11 @@
             &config.outputFdAnnotations,
             debuggable,
         )?;
-        wait_until_authfs_ready(&mountpoint).map_err(|e| {
-            debug!("Wait for authfs: {:?}", child.wait());
+        wait_until_authfs_ready(&child, &mountpoint).map_err(|e| {
+            match child.wait() {
+                Ok(status) => debug!("Wait for authfs: {}", status),
+                Err(e) => warn!("Failed to wait for child: {}", e),
+            }
             e
         })?;
 
@@ -144,13 +147,18 @@
     SharedChild::spawn(&mut command).context("Spawn authfs")
 }
 
-fn wait_until_authfs_ready(mountpoint: &OsStr) -> Result<()> {
+fn wait_until_authfs_ready(child: &SharedChild, mountpoint: &OsStr) -> Result<()> {
     let start_time = Instant::now();
     loop {
         if is_fuse(mountpoint)? {
             break;
         }
+        if let Some(exit_status) = child.try_wait()? {
+            // If the child has exited, we will never become ready.
+            bail!("Child has exited: {}", exit_status);
+        }
         if start_time.elapsed() > AUTHFS_SETUP_TIMEOUT_SEC {
+            let _ = child.kill();
             bail!("Time out mounting authfs");
         }
         sleep(AUTHFS_SETUP_POLL_INTERVAL_MS);
diff --git a/authfs/service/src/main.rs b/authfs/service/src/main.rs
index af8c7f9..890e108 100644
--- a/authfs/service/src/main.rs
+++ b/authfs/service/src/main.rs
@@ -59,7 +59,7 @@
         create_dir(&mountpoint).map_err(|e| {
             new_binder_exception(
                 ExceptionCode::SERVICE_SPECIFIC,
-                format!("Cannot create mount directory {:?}: {}", &mountpoint, e),
+                format!("Cannot create mount directory {:?}: {:?}", &mountpoint, e),
             )
         })?;
 
@@ -109,7 +109,7 @@
     Ok(())
 }
 
-fn main() -> Result<()> {
+fn try_main() -> Result<()> {
     let debuggable = env!("TARGET_BUILD_VARIANT") != "user";
     let log_level = if debuggable { log::Level::Trace } else { log::Level::Info };
     android_logger::init_once(
@@ -128,3 +128,10 @@
     ProcessState::join_thread_pool();
     bail!("Unexpected exit after join_thread_pool")
 }
+
+fn main() {
+    if let Err(e) = try_main() {
+        error!("failed with {:?}", e);
+        std::process::exit(1);
+    }
+}
diff --git a/authfs/src/main.rs b/authfs/src/main.rs
index c85d801..ecb0e68 100644
--- a/authfs/src/main.rs
+++ b/authfs/src/main.rs
@@ -28,6 +28,7 @@
 //! e.g. /mountpoint/42.
 
 use anyhow::{bail, Context, Result};
+use log::error;
 use std::collections::BTreeMap;
 use std::convert::TryInto;
 use std::fs::File;
@@ -325,7 +326,7 @@
     Ok(file_pool)
 }
 
-fn main() -> Result<()> {
+fn try_main() -> Result<()> {
     let args = Args::from_args();
 
     let log_level = if args.debug { log::Level::Debug } else { log::Level::Info };
@@ -337,3 +338,10 @@
     fusefs::loop_forever(file_pool, &args.mount_point, &args.extra_options)?;
     bail!("Unexpected exit after the handler loop")
 }
+
+fn main() {
+    if let Err(e) = try_main() {
+        error!("failed with {:?}", e);
+        std::process::exit(1);
+    }
+}
diff --git a/binder_common/Android.bp b/binder_common/Android.bp
index 789a891..9e6d590 100644
--- a/binder_common/Android.bp
+++ b/binder_common/Android.bp
@@ -9,6 +9,7 @@
     edition: "2018",
     rustlibs: [
         "libbinder_rs",
+        "liblazy_static",
     ],
     apex_available: [
         "com.android.compos",
diff --git a/binder_common/lazy_service.rs b/binder_common/lazy_service.rs
new file mode 100644
index 0000000..a2b85db
--- /dev/null
+++ b/binder_common/lazy_service.rs
@@ -0,0 +1,77 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+//! Rust API for lazy (aka dynamic) AIDL services.
+//! See https://source.android.com/devices/architecture/aidl/dynamic-aidl.
+
+use binder::public_api::force_lazy_services_persist;
+use lazy_static::lazy_static;
+use std::sync::Mutex;
+
+// TODO(b/200924402): Move this class to libbinder_rs once the infrastructure needed exists.
+
+/// An RAII object to ensure a server of lazy services is not killed. During the lifetime of any of
+/// these objects the service manager will not not kill the current process even if none of its
+/// lazy services are in use.
+#[must_use]
+#[derive(Debug)]
+pub struct LazyServiceGuard {
+    // Prevent construction outside this module.
+    _private: (),
+}
+
+lazy_static! {
+    // Count of how many LazyServiceGuard objects are in existence.
+    static ref GUARD_COUNT: Mutex<u64> = Mutex::new(0);
+}
+
+impl LazyServiceGuard {
+    /// Create a new LazyServiceGuard to prevent the service manager prematurely killing this
+    /// process.
+    pub fn new() -> Self {
+        let mut count = GUARD_COUNT.lock().unwrap();
+        *count += 1;
+        if *count == 1 {
+            // It's important that we make this call with the mutex held, to make sure
+            // that multiple calls (e.g. if the count goes 1 -> 0 -> 1) are correctly
+            // sequenced. (That also means we can't just use an AtomicU64.)
+            force_lazy_services_persist(true);
+        }
+        Self { _private: () }
+    }
+}
+
+impl Drop for LazyServiceGuard {
+    fn drop(&mut self) {
+        let mut count = GUARD_COUNT.lock().unwrap();
+        *count -= 1;
+        if *count == 0 {
+            force_lazy_services_persist(false);
+        }
+    }
+}
+
+impl Clone for LazyServiceGuard {
+    fn clone(&self) -> Self {
+        Self::new()
+    }
+}
+
+impl Default for LazyServiceGuard {
+    fn default() -> Self {
+        Self::new()
+    }
+}
diff --git a/binder_common/lib.rs b/binder_common/lib.rs
index 54cb80e..055688a 100644
--- a/binder_common/lib.rs
+++ b/binder_common/lib.rs
@@ -16,6 +16,8 @@
 
 //! Common items useful for binder clients and/or servers.
 
+pub mod lazy_service;
+
 use binder::public_api::{ExceptionCode, Status};
 use std::ffi::CString;
 
diff --git a/compos/common/compos_client.rs b/compos/common/compos_client.rs
index e68deb8..79d8354 100644
--- a/compos/common/compos_client.rs
+++ b/compos/common/compos_client.rs
@@ -46,11 +46,8 @@
 
 /// This owns an instance of the CompOS VM.
 pub struct VmInstance {
-    #[allow(dead_code)] // Prevent service manager from killing the dynamic service
-    service: Strong<dyn IVirtualizationService>,
     #[allow(dead_code)] // Keeps the VM alive even if we don`t touch it
     vm: Strong<dyn IVirtualMachine>,
-    #[allow(dead_code)] // TODO: Do we need this?
     cid: i32,
 }
 
@@ -113,7 +110,7 @@
 
         let cid = vm_state.wait_until_ready()?;
 
-        Ok(VmInstance { service, vm, cid })
+        Ok(VmInstance { vm, cid })
     }
 
     /// Create and return an RPC Binder connection to the Comp OS service in the VM.
@@ -129,6 +126,7 @@
 
     /// Return the CID of the VM.
     pub fn cid(&self) -> i32 {
+        // TODO: Do we actually need/use this?
         self.cid
     }
 }
diff --git a/compos/src/compsvc_main.rs b/compos/src/compsvc_main.rs
index 388e79b..d0c920a 100644
--- a/compos/src/compsvc_main.rs
+++ b/compos/src/compsvc_main.rs
@@ -43,7 +43,14 @@
 /// The CID representing the host VM
 const VMADDR_CID_HOST: u32 = 2;
 
-fn main() -> Result<()> {
+fn main() {
+    if let Err(e) = try_main() {
+        error!("failed with {:?}", e);
+        std::process::exit(1);
+    }
+}
+
+fn try_main() -> Result<()> {
     let args = clap::App::new("compsvc")
         .arg(clap::Arg::with_name("log_to_stderr").long("log_to_stderr"))
         .get_matches();
diff --git a/compos/src/pvm_exec.rs b/compos/src/pvm_exec.rs
index fdd9c57..8ce27db 100644
--- a/compos/src/pvm_exec.rs
+++ b/compos/src/pvm_exec.rs
@@ -32,8 +32,11 @@
 use clap::{value_t, App, Arg};
 use log::{debug, error, warn};
 use minijail::Minijail;
-use nix::fcntl::{fcntl, FcntlArg::F_GETFD};
-use std::os::unix::io::RawFd;
+use nix::fcntl::{fcntl, FcntlArg::F_GETFD, OFlag};
+use nix::unistd::pipe2;
+use std::fs::File;
+use std::io::Read;
+use std::os::unix::io::{AsRawFd, FromRawFd, RawFd};
 use std::path::Path;
 use std::process::exit;
 
@@ -69,7 +72,11 @@
     }
 }
 
-fn spawn_fd_server(fd_annotation: &FdAnnotation, debuggable: bool) -> Result<Minijail> {
+fn spawn_fd_server(
+    fd_annotation: &FdAnnotation,
+    ready_file: File,
+    debuggable: bool,
+) -> Result<Minijail> {
     let mut inheritable_fds = if debuggable {
         vec![1, 2] // inherit/redirect stdout/stderr for debugging
     } else {
@@ -87,6 +94,10 @@
         args.push(fd.to_string());
         inheritable_fds.push(*fd);
     }
+    let ready_fd = ready_file.as_raw_fd();
+    args.push("--ready-fd".to_string());
+    args.push(ready_fd.to_string());
+    inheritable_fds.push(ready_fd);
 
     let jail = Minijail::new()?;
     let _pid = jail.run(Path::new(FD_SERVER_BIN), &inheritable_fds, &args)?;
@@ -153,12 +164,31 @@
     Ok(Config { args, fd_annotation: FdAnnotation { input_fds, output_fds }, cid, debuggable })
 }
 
+fn create_pipe() -> Result<(File, File)> {
+    let (raw_read, raw_write) = pipe2(OFlag::O_CLOEXEC)?;
+    // SAFETY: We are the sole owners of these fds as they were just created.
+    let read_fd = unsafe { File::from_raw_fd(raw_read) };
+    let write_fd = unsafe { File::from_raw_fd(raw_write) };
+    Ok((read_fd, write_fd))
+}
+
+fn wait_for_fd_server_ready(mut ready_fd: File) -> Result<()> {
+    let mut buffer = [0];
+    // When fd_server is ready it closes its end of the pipe. And if it exits, the pipe is also
+    // closed. Either way this read will return 0 bytes at that point, and there's no point waiting
+    // any longer.
+    let _ = ready_fd.read(&mut buffer).context("Waiting for fd_server to be ready")?;
+    debug!("fd_server is ready");
+    Ok(())
+}
+
 fn try_main() -> Result<()> {
     // 1. Parse the command line arguments for collect execution data.
     let Config { args, fd_annotation, cid, debuggable } = parse_args()?;
 
     // 2. Spawn and configure a fd_server to serve remote read/write requests.
-    let fd_server_jail = spawn_fd_server(&fd_annotation, debuggable)?;
+    let (ready_read_fd, ready_write_fd) = create_pipe()?;
+    let fd_server_jail = spawn_fd_server(&fd_annotation, ready_write_fd, debuggable)?;
     let fd_server_lifetime = scopeguard::guard(fd_server_jail, |fd_server_jail| {
         if let Err(e) = fd_server_jail.kill() {
             if !matches!(e, minijail::Error::Killed(_)) {
@@ -171,10 +201,12 @@
     let result = if cid == VMADDR_CID_ANY {
         // Sentinel value that indicates we should use composd
         let composd = get_composd()?;
+        wait_for_fd_server_ready(ready_read_fd)?;
         composd.compile(&args, &fd_annotation)
     } else {
         // Call directly into the VM
         let compos_vm = get_rpc_binder(cid)?;
+        wait_for_fd_server_ready(ready_read_fd)?;
         compos_vm.compile(&args, &fd_annotation)
     };
     let result = result.context("Binder call failed")?;
@@ -190,7 +222,7 @@
     // Be explicit about the lifetime, which should last at least until the task is finished.
     drop(fd_server_lifetime);
 
-    if result.exitCode > 0 {
+    if result.exitCode != 0 {
         error!("remote execution failed with exit code {}", result.exitCode);
         exit(result.exitCode as i32);
     }
diff --git a/virtualizationservice/src/aidl.rs b/virtualizationservice/src/aidl.rs
index 76c3a16..a072060 100644
--- a/virtualizationservice/src/aidl.rs
+++ b/virtualizationservice/src/aidl.rs
@@ -19,13 +19,11 @@
 use crate::payload::add_microdroid_images;
 use crate::{Cid, FIRST_GUEST_CID, SYSPROP_LAST_CID};
 
-use binder_common::new_binder_exception;
+use ::binder::unstable_api::AsNative;
 use android_os_permissions_aidl::aidl::android::os::IPermissionController;
-use android_system_virtualizationservice::aidl::android::system::virtualizationservice::IVirtualMachine::{
-    BnVirtualMachine, IVirtualMachine,
-};
 use android_system_virtualizationservice::aidl::android::system::virtualizationservice::{
     DiskImage::DiskImage,
+    IVirtualMachine::{BnVirtualMachine, IVirtualMachine},
     IVirtualMachineCallback::IVirtualMachineCallback,
     IVirtualizationService::IVirtualizationService,
     PartitionType::PartitionType,
@@ -36,20 +34,24 @@
     VirtualMachineState::VirtualMachineState,
 };
 use android_system_virtualizationservice::binder::{
-    self, force_lazy_services_persist, BinderFeatures, ExceptionCode, Interface, ParcelFileDescriptor, Status, Strong, ThreadState,
+    self, BinderFeatures, ExceptionCode, Interface, ParcelFileDescriptor, Status, Strong,
+    ThreadState,
 };
-use android_system_virtualmachineservice::aidl::android::system::virtualmachineservice::IVirtualMachineService::{
-    VM_BINDER_SERVICE_PORT, VM_STREAM_SERVICE_PORT, BnVirtualMachineService, IVirtualMachineService,
+use android_system_virtualmachineservice::aidl::android::system::virtualmachineservice::{
+    IVirtualMachineService::{
+        BnVirtualMachineService, IVirtualMachineService, VM_BINDER_SERVICE_PORT,
+        VM_STREAM_SERVICE_PORT,
+    },
 };
 use anyhow::{anyhow, bail, Context, Result};
-use ::binder::unstable_api::AsNative;
+use binder_common::{lazy_service::LazyServiceGuard, new_binder_exception};
 use disk::QcowFile;
-use idsig::{V4Signature, HashAlgorithm};
-use log::{debug, error, warn, info};
+use idsig::{HashAlgorithm, V4Signature};
+use log::{debug, error, info, warn};
 use microdroid_payload_config::VmPayloadConfig;
 use rustutils::system_properties;
 use std::convert::TryInto;
-use std::fs::{File, OpenOptions, create_dir};
+use std::fs::{create_dir, File, OpenOptions};
 use std::io::{Error, ErrorKind, Write};
 use std::num::NonZeroU32;
 use std::os::unix::io::{FromRawFd, IntoRawFd};
@@ -557,11 +559,13 @@
 #[derive(Debug)]
 struct VirtualMachine {
     instance: Arc<VmInstance>,
+    /// Keeps our service process running as long as this VM instance exists.
+    lazy_service_guard: LazyServiceGuard,
 }
 
 impl VirtualMachine {
     fn create(instance: Arc<VmInstance>) -> Strong<dyn IVirtualMachine> {
-        let binder = VirtualMachine { instance };
+        let binder = VirtualMachine { instance, lazy_service_guard: Default::default() };
         BnVirtualMachine::new_binder(binder, BinderFeatures::default())
     }
 }
@@ -714,19 +718,12 @@
     /// Store a strong VM reference.
     fn debug_hold_vm(&mut self, vm: Strong<dyn IVirtualMachine>) {
         self.debug_held_vms.push(vm);
-        // Make sure our process is not shut down while we hold the VM reference
-        // on behalf of the caller.
-        force_lazy_services_persist(true);
     }
 
     /// Retrieve and remove a strong VM reference.
     fn debug_drop_vm(&mut self, cid: i32) -> Option<Strong<dyn IVirtualMachine>> {
         let pos = self.debug_held_vms.iter().position(|vm| vm.getCid() == Ok(cid))?;
         let vm = self.debug_held_vms.swap_remove(pos);
-        if self.debug_held_vms.is_empty() {
-            // Once we no longer hold any VM references it is ok for our process to be shut down.
-            force_lazy_services_persist(false);
-        }
         Some(vm)
     }
 }