src/linux.rs

// Copyright 2017 The Chromium OS Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

use std;
use std::cmp::min;
use std::convert::TryFrom;
use std::error::Error as StdError;
use std::ffi::CStr;
use std::fmt::{self, Display};
use std::fs::{File, OpenOptions};
use std::io::{self, stdin, Read};
use std::mem;
use std::net::Ipv4Addr;
#[cfg(feature = "gpu")]
use std::num::NonZeroU8;
use std::num::ParseIntError;
use std::os::unix::io::{AsRawFd, FromRawFd, RawFd};
use std::os::unix::net::UnixStream;
use std::path::{Path, PathBuf};
use std::ptr;
use std::str;
use std::sync::{Arc, Barrier};
use std::thread;
use std::thread::JoinHandle;
use std::time::{Duration, SystemTime, UNIX_EPOCH};

use libc::{self, c_int, gid_t, uid_t};

use audio_streams::shm_streams::NullShmStreamSource;
#[cfg(feature = "gpu")]
use devices::virtio::EventDevice;
use devices::virtio::{self, VirtioDevice};
use devices::{
    self, HostBackendDeviceProvider, PciDevice, VfioDevice, VfioPciDevice, VirtioPciDevice,
    XhciController,
};
use io_jail::{self, Minijail};
use kvm::*;
use libcras::CrasClient;
use msg_socket::{MsgError, MsgReceiver, MsgSender, MsgSocket};
use net_util::{Error as NetError, MacAddress, Tap};
use rand_ish::SimpleRng;
use remain::sorted;
use resources::{Alloc, MmioType, SystemAllocator};
use sync::{Condvar, Mutex};
use sys_util::net::{UnixSeqpacket, UnixSeqpacketListener, UnlinkUnixSeqpacketListener};

use sys_util::{
    self, block_signal, clear_signal, drop_capabilities, error, flock, get_blocked_signals,
    get_group_id, get_user_id, getegid, geteuid, info, register_rt_signal_handler,
    set_cpu_affinity, validate_raw_fd, warn, EventFd, FlockOperation, GuestAddress, GuestMemory,
    Killable, MemoryMappingArena, PollContext, PollToken, Protection, ScopedEvent, SignalFd,
    Terminal, TimerFd, WatchingEvents, SIGRTMIN,
};
use vhost;
use vm_control::{
    BalloonControlCommand, BalloonControlRequestSocket, BalloonControlResponseSocket,
    DiskControlCommand, DiskControlRequestSocket, DiskControlResponseSocket, DiskControlResult,
    UsbControlSocket, VmControlResponseSocket, VmIrqRequest, VmIrqResponse, VmIrqResponseSocket,
    VmMemoryControlRequestSocket, VmMemoryControlResponseSocket, VmMemoryRequest, VmMemoryResponse,
    VmRunMode,
};

use crate::{Config, DiskOption, Executable, SharedDir, SharedDirKind, TouchDeviceOption};

#[cfg(feature = "gpu")]
use crate::{DEFAULT_TOUCH_DEVICE_HEIGHT, DEFAULT_TOUCH_DEVICE_WIDTH};

use arch::{self, LinuxArch, RunnableLinuxVm, VirtioDeviceStub, VmComponents, VmImage};

#[cfg(any(target_arch = "arm", target_arch = "aarch64"))]
use aarch64::AArch64 as Arch;
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
use x86_64::X8664arch as Arch;

#[sorted]
#[derive(Debug)]
pub enum Error {
    AddGpuDeviceMemory(sys_util::Error),
    AddPmemDeviceMemory(sys_util::Error),
    AllocateGpuDeviceAddress,
    AllocatePmemDeviceAddress(resources::Error),
    BalloonDeviceNew(virtio::BalloonError),
    BlockDeviceNew(sys_util::Error),
    BlockSignal(sys_util::signal::Error),
    BuildVm(<Arch as LinuxArch>::Error),
    ChownTpmStorage(sys_util::Error),
    CloneEventFd(sys_util::Error),
    CreateCrasClient(libcras::Error),
    CreateDiskError(disk::Error),
    CreateEventFd(sys_util::Error),
    CreatePollContext(sys_util::Error),
    CreateSignalFd(sys_util::SignalFdError),
    CreateSocket(io::Error),
    CreateTapDevice(NetError),
    CreateTimerFd(sys_util::Error),
    CreateTpmStorage(PathBuf, io::Error),
    CreateUsbProvider(devices::usb::host_backend::error::Error),
    CreateVfioDevice(devices::vfio::VfioError),
    DeviceJail(io_jail::Error),
    DevicePivotRoot(io_jail::Error),
    Disk(PathBuf, io::Error),
    DiskImageLock(sys_util::Error),
    DropCapabilities(sys_util::Error),
    FsDeviceNew(virtio::fs::Error),
    GetMaxOpenFiles(io::Error),
    InputDeviceNew(virtio::InputError),
    InputEventsOpen(std::io::Error),
    InvalidFdPath,
    InvalidWaylandPath,
    IoJail(io_jail::Error),
    LoadKernel(Box<dyn StdError>),
    MemoryTooLarge,
    NetDeviceNew(virtio::NetError),
    OpenAndroidFstab(PathBuf, io::Error),
    OpenBios(PathBuf, io::Error),
    OpenInitrd(PathBuf, io::Error),
    OpenKernel(PathBuf, io::Error),
    OpenVinput(PathBuf, io::Error),
    P9DeviceNew(virtio::P9Error),
    ParseMaxOpenFiles(ParseIntError),
    PivotRootDoesntExist(&'static str),
    PmemDeviceImageTooBig,
    PmemDeviceNew(sys_util::Error),
    PollContextAdd(sys_util::Error),
    PollContextDelete(sys_util::Error),
    ReadLowmemAvailable(io::Error),
    ReadLowmemMargin(io::Error),
    RegisterBalloon(arch::DeviceRegistrationError),
    RegisterBlock(arch::DeviceRegistrationError),
    RegisterGpu(arch::DeviceRegistrationError),
    RegisterNet(arch::DeviceRegistrationError),
    RegisterP9(arch::DeviceRegistrationError),
    RegisterRng(arch::DeviceRegistrationError),
    RegisterSignalHandler(sys_util::Error),
    RegisterWayland(arch::DeviceRegistrationError),
    ReserveGpuMemory(sys_util::MmapError),
    ReserveMemory(sys_util::Error),
    ReservePmemMemory(sys_util::MmapError),
    ResetTimerFd(sys_util::Error),
    RngDeviceNew(virtio::RngError),
    SettingGidMap(io_jail::Error),
    SettingMaxOpenFiles(io_jail::Error),
    SettingUidMap(io_jail::Error),
    SignalFd(sys_util::SignalFdError),
    SpawnVcpu(io::Error),
    TimerFd(sys_util::Error),
    ValidateRawFd(sys_util::Error),
    VhostNetDeviceNew(virtio::vhost::Error),
    VhostVsockDeviceNew(virtio::vhost::Error),
    VirtioPciDev(sys_util::Error),
    WaylandDeviceNew(sys_util::Error),
}

impl Display for Error {
    #[remain::check]
    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
        use self::Error::*;

        #[sorted]
        match self {
            AddGpuDeviceMemory(e) => write!(f, "failed to add gpu device memory: {}", e),
            AddPmemDeviceMemory(e) => write!(f, "failed to add pmem device memory: {}", e),
            AllocateGpuDeviceAddress => write!(f, "failed to allocate gpu device guest address"),
            AllocatePmemDeviceAddress(e) => {
                write!(f, "failed to allocate memory for pmem device: {}", e)
            }
            BalloonDeviceNew(e) => write!(f, "failed to create balloon: {}", e),
            BlockDeviceNew(e) => write!(f, "failed to create block device: {}", e),
            BlockSignal(e) => write!(f, "failed to block signal: {}", e),
            BuildVm(e) => write!(f, "The architecture failed to build the vm: {}", e),
            ChownTpmStorage(e) => write!(f, "failed to chown tpm storage: {}", e),
            CloneEventFd(e) => write!(f, "failed to clone eventfd: {}", e),
            CreateCrasClient(e) => write!(f, "failed to create cras client: {}", e),
            CreateDiskError(e) => write!(f, "failed to create virtual disk: {}", e),
            CreateEventFd(e) => write!(f, "failed to create eventfd: {}", e),
            CreatePollContext(e) => write!(f, "failed to create poll context: {}", e),
            CreateSignalFd(e) => write!(f, "failed to create signalfd: {}", e),
            CreateSocket(e) => write!(f, "failed to create socket: {}", e),
            CreateTapDevice(e) => write!(f, "failed to create tap device: {}", e),
            CreateTimerFd(e) => write!(f, "failed to create timerfd: {}", e),
            CreateTpmStorage(p, e) => {
                write!(f, "failed to create tpm storage dir {}: {}", p.display(), e)
            }
            CreateUsbProvider(e) => write!(f, "failed to create usb provider: {}", e),
            CreateVfioDevice(e) => write!(f, "Failed to create vfio device {}", e),
            DeviceJail(e) => write!(f, "failed to jail device: {}", e),
            DevicePivotRoot(e) => write!(f, "failed to pivot root device: {}", e),
            Disk(p, e) => write!(f, "failed to load disk image {}: {}", p.display(), e),
            DiskImageLock(e) => write!(f, "failed to lock disk image: {}", e),
            DropCapabilities(e) => write!(f, "failed to drop process capabilities: {}", e),
            FsDeviceNew(e) => write!(f, "failed to create fs device: {}", e),
            GetMaxOpenFiles(e) => write!(f, "failed to get max number of open files: {}", e),
            InputDeviceNew(e) => write!(f, "failed to set up input device: {}", e),
            InputEventsOpen(e) => write!(f, "failed to open event device: {}", e),
            InvalidFdPath => write!(f, "failed parsing a /proc/self/fd/*"),
            InvalidWaylandPath => write!(f, "wayland socket path has no parent or file name"),
            IoJail(e) => write!(f, "{}", e),
            LoadKernel(e) => write!(f, "failed to load kernel: {}", e),
            MemoryTooLarge => write!(f, "requested memory size too large"),
            NetDeviceNew(e) => write!(f, "failed to set up virtio networking: {}", e),
            OpenAndroidFstab(p, e) => write!(
                f,
                "failed to open android fstab file {}: {}",
                p.display(),
                e
            ),
            OpenBios(p, e) => write!(f, "failed to open bios {}: {}", p.display(), e),
            OpenInitrd(p, e) => write!(f, "failed to open initrd {}: {}", p.display(), e),
            OpenKernel(p, e) => write!(f, "failed to open kernel image {}: {}", p.display(), e),
            OpenVinput(p, e) => write!(f, "failed to open vinput device {}: {}", p.display(), e),
            P9DeviceNew(e) => write!(f, "failed to create 9p device: {}", e),
            ParseMaxOpenFiles(e) => write!(f, "failed to parse max number of open files: {}", e),
            PivotRootDoesntExist(p) => write!(f, "{} doesn't exist, can't jail devices.", p),
            PmemDeviceImageTooBig => {
                write!(f, "failed to create pmem device: pmem device image too big")
            }
            PmemDeviceNew(e) => write!(f, "failed to create pmem device: {}", e),
            PollContextAdd(e) => write!(f, "failed to add fd to poll context: {}", e),
            PollContextDelete(e) => write!(f, "failed to remove fd from poll context: {}", e),
            ReadLowmemAvailable(e) => write!(
                f,
                "failed to read /sys/kernel/mm/chromeos-low_mem/available: {}",
                e
            ),
            ReadLowmemMargin(e) => write!(
                f,
                "failed to read /sys/kernel/mm/chromeos-low_mem/margin: {}",
                e
            ),
            RegisterBalloon(e) => write!(f, "error registering balloon device: {}", e),
            RegisterBlock(e) => write!(f, "error registering block device: {}", e),
            RegisterGpu(e) => write!(f, "error registering gpu device: {}", e),
            RegisterNet(e) => write!(f, "error registering net device: {}", e),
            RegisterP9(e) => write!(f, "error registering 9p device: {}", e),
            RegisterRng(e) => write!(f, "error registering rng device: {}", e),
            RegisterSignalHandler(e) => write!(f, "error registering signal handler: {}", e),
            RegisterWayland(e) => write!(f, "error registering wayland device: {}", e),
            ReserveGpuMemory(e) => write!(f, "failed to reserve gpu memory: {}", e),
            ReserveMemory(e) => write!(f, "failed to reserve memory: {}", e),
            ReservePmemMemory(e) => write!(f, "failed to reserve pmem memory: {}", e),
            ResetTimerFd(e) => write!(f, "failed to reset timerfd: {}", e),
            RngDeviceNew(e) => write!(f, "failed to set up rng: {}", e),
            SettingGidMap(e) => write!(f, "error setting GID map: {}", e),
            SettingMaxOpenFiles(e) => write!(f, "error setting max open files: {}", e),
            SettingUidMap(e) => write!(f, "error setting UID map: {}", e),
            SignalFd(e) => write!(f, "failed to read signal fd: {}", e),
            SpawnVcpu(e) => write!(f, "failed to spawn VCPU thread: {}", e),
            TimerFd(e) => write!(f, "failed to read timer fd: {}", e),
            ValidateRawFd(e) => write!(f, "failed to validate raw fd: {}", e),
            VhostNetDeviceNew(e) => write!(f, "failed to set up vhost networking: {}", e),
            VhostVsockDeviceNew(e) => write!(f, "failed to set up virtual socket device: {}", e),
            VirtioPciDev(e) => write!(f, "failed to create virtio pci dev: {}", e),
            WaylandDeviceNew(e) => write!(f, "failed to create wayland device: {}", e),
        }
    }
}

impl From<io_jail::Error> for Error {
    fn from(err: io_jail::Error) -> Self {
        Error::IoJail(err)
    }
}

impl std::error::Error for Error {}

type Result<T> = std::result::Result<T, Error>;

enum TaggedControlSocket {
    Vm(VmControlResponseSocket),
    VmMemory(VmMemoryControlResponseSocket),
    VmIrq(VmIrqResponseSocket),
}

impl AsRef<UnixSeqpacket> for TaggedControlSocket {
    fn as_ref(&self) -> &UnixSeqpacket {
        use self::TaggedControlSocket::*;
        match &self {
            Vm(ref socket) => socket.as_ref(),
            VmMemory(ref socket) => socket.as_ref(),
            VmIrq(ref socket) => socket.as_ref(),
        }
    }
}

impl AsRawFd for TaggedControlSocket {
    fn as_raw_fd(&self) -> RawFd {
        self.as_ref().as_raw_fd()
    }
}

fn get_max_open_files() -> Result<libc::rlim64_t> {
    let mut buf = mem::MaybeUninit::<libc::rlimit64>::zeroed();

    // Safe because this will only modify `buf` and we check the return value.
    let res = unsafe { libc::prlimit64(0, libc::RLIMIT_NOFILE, ptr::null(), buf.as_mut_ptr()) };
    if res == 0 {
        // Safe because the kernel guarantees that the struct is fully initialized.
        let limit = unsafe { buf.assume_init() };
        Ok(limit.rlim_max)
    } else {
        Err(Error::GetMaxOpenFiles(io::Error::last_os_error()))
    }
}

struct SandboxConfig<'a> {
    limit_caps: bool,
    log_failures: bool,
    seccomp_policy: &'a Path,
    uid_map: Option<&'a str>,
    gid_map: Option<&'a str>,
}

fn create_base_minijail(
    root: &Path,
    r_limit: Option<u64>,
    config: Option<&SandboxConfig>,
) -> Result<Minijail> {
    // All child jails run in a new user namespace without any users mapped,
    // they run as nobody unless otherwise configured.
    let mut j = Minijail::new().map_err(Error::DeviceJail)?;

    if let Some(config) = config {
        j.namespace_pids();
        j.namespace_user();
        j.namespace_user_disable_setgroups();
        if config.limit_caps {
            // Don't need any capabilities.
            j.use_caps(0);
        }
        if let Some(uid_map) = config.uid_map {
            j.uidmap(uid_map).map_err(Error::SettingUidMap)?;
        }
        if let Some(gid_map) = config.gid_map {
            j.gidmap(gid_map).map_err(Error::SettingGidMap)?;
        }
        // Run in a new mount namespace.
        j.namespace_vfs();

        // Run in an empty network namespace.
        j.namespace_net();

        // Don't allow the device to gain new privileges.
        j.no_new_privs();

        // By default we'll prioritize using the pre-compiled .bpf over the .policy
        // file (the .bpf is expected to be compiled using "trap" as the failure
        // behavior instead of the default "kill" behavior).
        // Refer to the code comment for the "seccomp-log-failures"
        // command-line parameter for an explanation about why the |log_failures|
        // flag forces the use of .policy files (and the build-time alternative to
        // this run-time flag).
        let bpf_policy_file = config.seccomp_policy.with_extension("bpf");
        if bpf_policy_file.exists() && !config.log_failures {
            j.parse_seccomp_program(&bpf_policy_file)
                .map_err(Error::DeviceJail)?;
        } else {
            // Use TSYNC only for the side effect of it using SECCOMP_RET_TRAP,
            // which will correctly kill the entire device process if a worker
            // thread commits a seccomp violation.
            j.set_seccomp_filter_tsync();
            if config.log_failures {
                j.log_seccomp_filter_failures();
            }
            j.parse_seccomp_filters(&config.seccomp_policy.with_extension("policy"))
                .map_err(Error::DeviceJail)?;
        }
        j.use_seccomp_filter();
        // Don't do init setup.
        j.run_as_init();
    }

    // Only pivot_root if we are not re-using the current root directory.
    if root != Path::new("/") {
        // It's safe to call `namespace_vfs` multiple times.
        j.namespace_vfs();
        j.enter_pivot_root(root).map_err(Error::DevicePivotRoot)?;
    }

    // Most devices don't need to open many fds.
    let limit = if let Some(r) = r_limit { r } else { 1024u64 };
    j.set_rlimit(libc::RLIMIT_NOFILE as i32, limit, limit)
        .map_err(Error::SettingMaxOpenFiles)?;

    Ok(j)
}

fn simple_jail(cfg: &Config, policy: &str) -> Result<Option<Minijail>> {
    if cfg.sandbox {
        let pivot_root: &str = option_env!("DEFAULT_PIVOT_ROOT").unwrap_or("/var/empty");
        // A directory for a jailed device's pivot root.
        let root_path = Path::new(pivot_root);
        if !root_path.exists() {
            return Err(Error::PivotRootDoesntExist(pivot_root));
        }
        let policy_path: PathBuf = cfg.seccomp_policy_dir.join(policy);
        let config = SandboxConfig {
            limit_caps: true,
            log_failures: cfg.seccomp_log_failures,
            seccomp_policy: &policy_path,
            uid_map: None,
            gid_map: None,
        };
        Ok(Some(create_base_minijail(root_path, None, Some(&config))?))
    } else {
        Ok(None)
    }
}

type DeviceResult<T = VirtioDeviceStub> = std::result::Result<T, Error>;

fn create_block_device(
    cfg: &Config,
    disk: &DiskOption,
    disk_device_socket: DiskControlResponseSocket,
) -> DeviceResult {
    // Special case '/proc/self/fd/*' paths. The FD is already open, just use it.
    let raw_image: File = if disk.path.parent() == Some(Path::new("/proc/self/fd")) {
        // Safe because we will validate |raw_fd|.
        unsafe { File::from_raw_fd(raw_fd_from_path(&disk.path)?) }
    } else {
        OpenOptions::new()
            .read(true)
            .write(!disk.read_only)
            .open(&disk.path)
            .map_err(|e| Error::Disk(disk.path.to_path_buf(), e))?
    };
    // Lock the disk image to prevent other crosvm instances from using it.
    let lock_op = if disk.read_only {
        FlockOperation::LockShared
    } else {
        FlockOperation::LockExclusive
    };
    flock(&raw_image, lock_op, true).map_err(Error::DiskImageLock)?;

    let disk_file = disk::create_disk_file(raw_image).map_err(Error::CreateDiskError)?;
    let dev = virtio::Block::new(
        disk_file,
        disk.read_only,
        disk.sparse,
        disk.block_size,
        Some(disk_device_socket),
    )
    .map_err(Error::BlockDeviceNew)?;

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail: simple_jail(&cfg, "block_device")?,
    })
}

fn create_rng_device(cfg: &Config) -> DeviceResult {
    let dev = virtio::Rng::new().map_err(Error::RngDeviceNew)?;

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail: simple_jail(&cfg, "rng_device")?,
    })
}

#[cfg(feature = "tpm")]
fn create_tpm_device(cfg: &Config) -> DeviceResult {
    use std::ffi::CString;
    use std::fs;
    use std::process;
    use sys_util::chown;

    let tpm_storage: PathBuf;
    let mut tpm_jail = simple_jail(&cfg, "tpm_device")?;

    match &mut tpm_jail {
        Some(jail) => {
            // Create a tmpfs in the device's root directory for tpm
            // simulator storage. The size is 20*1024, or 20 KB.
            jail.mount_with_data(
                Path::new("none"),
                Path::new("/"),
                "tmpfs",
                (libc::MS_NOSUID | libc::MS_NODEV | libc::MS_NOEXEC) as usize,
                "size=20480",
            )?;

            let crosvm_ids = add_crosvm_user_to_jail(jail, "tpm")?;

            let pid = process::id();
            let tpm_pid_dir = format!("/run/vm/tpm.{}", pid);
            tpm_storage = Path::new(&tpm_pid_dir).to_owned();
            fs::create_dir_all(&tpm_storage)
                .map_err(|e| Error::CreateTpmStorage(tpm_storage.to_owned(), e))?;
            let tpm_pid_dir_c = CString::new(tpm_pid_dir).expect("no nul bytes");
            chown(&tpm_pid_dir_c, crosvm_ids.uid, crosvm_ids.gid)
                .map_err(Error::ChownTpmStorage)?;

            jail.mount_bind(&tpm_storage, &tpm_storage, true)?;
        }
        None => {
            // Path used inside cros_sdk which does not have /run/vm.
            tpm_storage = Path::new("/tmp/tpm-simulator").to_owned();
        }
    }

    let dev = virtio::Tpm::new(tpm_storage);

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail: tpm_jail,
    })
}

fn create_single_touch_device(cfg: &Config, single_touch_spec: &TouchDeviceOption) -> DeviceResult {
    let socket = single_touch_spec
        .get_path()
        .into_unix_stream()
        .map_err(|e| {
            error!("failed configuring virtio single touch: {:?}", e);
            e
        })?;

    let (width, height) = single_touch_spec.get_size();
    let dev = virtio::new_single_touch(socket, width, height).map_err(Error::InputDeviceNew)?;
    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail: simple_jail(&cfg, "input_device")?,
    })
}

fn create_trackpad_device(cfg: &Config, trackpad_spec: &TouchDeviceOption) -> DeviceResult {
    let socket = trackpad_spec.get_path().into_unix_stream().map_err(|e| {
        error!("failed configuring virtio trackpad: {}", e);
        e
    })?;

    let (width, height) = trackpad_spec.get_size();
    let dev = virtio::new_trackpad(socket, width, height).map_err(Error::InputDeviceNew)?;

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail: simple_jail(&cfg, "input_device")?,
    })
}

fn create_mouse_device<T: IntoUnixStream>(cfg: &Config, mouse_socket: T) -> DeviceResult {
    let socket = mouse_socket.into_unix_stream().map_err(|e| {
        error!("failed configuring virtio mouse: {}", e);
        e
    })?;

    let dev = virtio::new_mouse(socket).map_err(Error::InputDeviceNew)?;

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail: simple_jail(&cfg, "input_device")?,
    })
}

fn create_keyboard_device<T: IntoUnixStream>(cfg: &Config, keyboard_socket: T) -> DeviceResult {
    let socket = keyboard_socket.into_unix_stream().map_err(|e| {
        error!("failed configuring virtio keyboard: {}", e);
        e
    })?;

    let dev = virtio::new_keyboard(socket).map_err(Error::InputDeviceNew)?;

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail: simple_jail(&cfg, "input_device")?,
    })
}

fn create_vinput_device(cfg: &Config, dev_path: &Path) -> DeviceResult {
    let dev_file = OpenOptions::new()
        .read(true)
        .write(true)
        .open(dev_path)
        .map_err(|e| Error::OpenVinput(dev_path.to_owned(), e))?;

    let dev = virtio::new_evdev(dev_file).map_err(Error::InputDeviceNew)?;

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail: simple_jail(&cfg, "input_device")?,
    })
}

fn create_balloon_device(cfg: &Config, socket: BalloonControlResponseSocket) -> DeviceResult {
    let dev = virtio::Balloon::new(socket).map_err(Error::BalloonDeviceNew)?;

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail: simple_jail(&cfg, "balloon_device")?,
    })
}

fn create_tap_net_device(cfg: &Config, tap_fd: RawFd) -> DeviceResult {
    // Safe because we ensure that we get a unique handle to the fd.
    let tap = unsafe {
        Tap::from_raw_fd(validate_raw_fd(tap_fd).map_err(Error::ValidateRawFd)?)
            .map_err(Error::CreateTapDevice)?
    };

    let dev = virtio::Net::from(tap).map_err(Error::NetDeviceNew)?;

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail: simple_jail(&cfg, "net_device")?,
    })
}

fn create_net_device(
    cfg: &Config,
    host_ip: Ipv4Addr,
    netmask: Ipv4Addr,
    mac_address: MacAddress,
    mem: &GuestMemory,
) -> DeviceResult {
    let dev = if cfg.vhost_net {
        let dev =
            virtio::vhost::Net::<Tap, vhost::Net<Tap>>::new(host_ip, netmask, mac_address, mem)
                .map_err(Error::VhostNetDeviceNew)?;
        Box::new(dev) as Box<dyn VirtioDevice>
    } else {
        let dev =
            virtio::Net::<Tap>::new(host_ip, netmask, mac_address).map_err(Error::NetDeviceNew)?;
        Box::new(dev) as Box<dyn VirtioDevice>
    };

    let policy = if cfg.vhost_net {
        "vhost_net_device"
    } else {
        "net_device"
    };

    Ok(VirtioDeviceStub {
        dev,
        jail: simple_jail(&cfg, policy)?,
    })
}

#[cfg(feature = "gpu")]
fn create_gpu_device(
    cfg: &Config,
    exit_evt: &EventFd,
    gpu_device_socket: VmMemoryControlRequestSocket,
    gpu_sockets: Vec<virtio::resource_bridge::ResourceResponseSocket>,
    wayland_socket_path: Option<&PathBuf>,
    x_display: Option<String>,
    event_devices: Vec<EventDevice>,
) -> DeviceResult {
    let jailed_wayland_path = Path::new("/wayland-0");

    let mut display_backends = vec![
        virtio::DisplayBackend::X(x_display),
        virtio::DisplayBackend::Stub,
    ];

    if let Some(socket_path) = wayland_socket_path {
        display_backends.insert(
            0,
            virtio::DisplayBackend::Wayland(if cfg.sandbox {
                Some(jailed_wayland_path.to_owned())
            } else {
                Some(socket_path.to_owned())
            }),
        );
    }

    let dev = virtio::Gpu::new(
        exit_evt.try_clone().map_err(Error::CloneEventFd)?,
        Some(gpu_device_socket),
        NonZeroU8::new(1).unwrap(), // number of scanouts
        gpu_sockets,
        display_backends,
        cfg.gpu_parameters.as_ref().unwrap(),
        event_devices,
    );

    let jail = match simple_jail(&cfg, "gpu_device")? {
        Some(mut jail) => {
            // Create a tmpfs in the device's root directory so that we can bind mount the
            // dri directory into it.  The size=67108864 is size=64*1024*1024 or size=64MB.
            jail.mount_with_data(
                Path::new("none"),
                Path::new("/"),
                "tmpfs",
                (libc::MS_NOSUID | libc::MS_NODEV | libc::MS_NOEXEC) as usize,
                "size=67108864",
            )?;

            // Device nodes required for DRM.
            let sys_dev_char_path = Path::new("/sys/dev/char");
            jail.mount_bind(sys_dev_char_path, sys_dev_char_path, false)?;
            let sys_devices_path = Path::new("/sys/devices");
            jail.mount_bind(sys_devices_path, sys_devices_path, false)?;
            let drm_dri_path = Path::new("/dev/dri");
            jail.mount_bind(drm_dri_path, drm_dri_path, false)?;

            // If the ARM specific devices exist on the host, bind mount them in.
            let mali0_path = Path::new("/dev/mali0");
            if mali0_path.exists() {
                jail.mount_bind(mali0_path, mali0_path, true)?;
            }

            let pvr_sync_path = Path::new("/dev/pvr_sync");
            if pvr_sync_path.exists() {
                jail.mount_bind(pvr_sync_path, pvr_sync_path, true)?;
            }

            // Libraries that are required when mesa drivers are dynamically loaded.
            let lib_dirs = &["/usr/lib", "/usr/lib64", "/lib", "/lib64"];
            for dir in lib_dirs {
                let dir_path = Path::new(dir);
                if dir_path.exists() {
                    jail.mount_bind(dir_path, dir_path, false)?;
                }
            }

            // Bind mount the wayland socket into jail's root. This is necessary since each
            // new wayland context must open() the socket.
            if let Some(path) = wayland_socket_path {
                jail.mount_bind(path, jailed_wayland_path, true)?;
            }

            add_crosvm_user_to_jail(&mut jail, "gpu")?;

            // pvr driver requires read access to /proc/self/task/*/comm.
            let proc_path = Path::new("/proc");
            jail.mount(
                proc_path,
                proc_path,
                "proc",
                (libc::MS_NOSUID | libc::MS_NODEV | libc::MS_NOEXEC | libc::MS_RDONLY) as usize,
            )?;

            Some(jail)
        }
        None => None,
    };

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail,
    })
}

fn create_wayland_device(
    cfg: &Config,
    socket: VmMemoryControlRequestSocket,
    resource_bridge: Option<virtio::resource_bridge::ResourceRequestSocket>,
) -> DeviceResult {
    let wayland_socket_dirs = cfg
        .wayland_socket_paths
        .iter()
        .map(|(_name, path)| path.parent())
        .collect::<Option<Vec<_>>>()
        .ok_or(Error::InvalidWaylandPath)?;

    let dev = virtio::Wl::new(cfg.wayland_socket_paths.clone(), socket, resource_bridge)
        .map_err(Error::WaylandDeviceNew)?;

    let jail = match simple_jail(&cfg, "wl_device")? {
        Some(mut jail) => {
            // Create a tmpfs in the device's root directory so that we can bind mount the wayland
            // socket directory into it. The size=67108864 is size=64*1024*1024 or size=64MB.
            jail.mount_with_data(
                Path::new("none"),
                Path::new("/"),
                "tmpfs",
                (libc::MS_NOSUID | libc::MS_NODEV | libc::MS_NOEXEC) as usize,
                "size=67108864",
            )?;

            // Bind mount the wayland socket's directory into jail's root. This is necessary since
            // each new wayland context must open() the socket. If the wayland socket is ever
            // destroyed and remade in the same host directory, new connections will be possible
            // without restarting the wayland device.
            for dir in &wayland_socket_dirs {
                jail.mount_bind(dir, dir, true)?;
            }
            add_crosvm_user_to_jail(&mut jail, "Wayland")?;

            Some(jail)
        }
        None => None,
    };

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail,
    })
}

fn create_vhost_vsock_device(cfg: &Config, cid: u64, mem: &GuestMemory) -> DeviceResult {
    let dev = virtio::vhost::Vsock::new(cid, mem).map_err(Error::VhostVsockDeviceNew)?;

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail: simple_jail(&cfg, "vhost_vsock_device")?,
    })
}

fn create_fs_device(
    cfg: &Config,
    uid_map: &str,
    gid_map: &str,
    src: &Path,
    tag: &str,
    fs_cfg: virtio::fs::passthrough::Config,
) -> DeviceResult {
    let max_open_files = get_max_open_files()?;
    let j = if cfg.sandbox {
        let seccomp_policy = cfg.seccomp_policy_dir.join("fs_device");
        let config = SandboxConfig {
            limit_caps: false,
            uid_map: Some(uid_map),
            gid_map: Some(gid_map),
            log_failures: cfg.seccomp_log_failures,
            seccomp_policy: &seccomp_policy,
        };
        create_base_minijail(src, Some(max_open_files), Some(&config))?
    } else {
        create_base_minijail(src, Some(max_open_files), None)?
    };

    // TODO(chirantan): Use more than one worker once the kernel driver has been fixed to not panic
    // when num_queues > 1.
    let dev = virtio::fs::Fs::new(tag, 1, fs_cfg).map_err(Error::FsDeviceNew)?;

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail: Some(j),
    })
}

fn create_9p_device(
    cfg: &Config,
    uid_map: &str,
    gid_map: &str,
    src: &Path,
    tag: &str,
) -> DeviceResult {
    let max_open_files = get_max_open_files()?;
    let (jail, root) = if cfg.sandbox {
        let seccomp_policy = cfg.seccomp_policy_dir.join("9p_device");
        let config = SandboxConfig {
            limit_caps: false,
            uid_map: Some(uid_map),
            gid_map: Some(gid_map),
            log_failures: cfg.seccomp_log_failures,
            seccomp_policy: &seccomp_policy,
        };

        let mut jail = create_base_minijail(src, Some(max_open_files), Some(&config))?;
        // We want bind mounts from the parent namespaces to propagate into the 9p server's
        // namespace.
        jail.set_remount_mode(libc::MS_SLAVE);

        //  The shared directory becomes the root of the device's file system.
        let root = Path::new("/");
        (Some(jail), root)
    } else {
        // There's no mount namespace so we tell the server to treat the source directory as the
        // root.
        (None, src)
    };

    let dev = virtio::P9::new(root, tag).map_err(Error::P9DeviceNew)?;

    Ok(VirtioDeviceStub {
        dev: Box::new(dev),
        jail,
    })
}

fn create_pmem_device(
    cfg: &Config,
    vm: &mut Vm,
    resources: &mut SystemAllocator,
    disk: &DiskOption,
    index: usize,
) -> DeviceResult {
    let fd = OpenOptions::new()
        .read(true)
        .write(!disk.read_only)
        .open(&disk.path)
        .map_err(|e| Error::Disk(disk.path.to_path_buf(), e))?;

    let (disk_size, arena_size) = {
        let metadata =
            std::fs::metadata(&disk.path).map_err(|e| Error::Disk(disk.path.to_path_buf(), e))?;
        let disk_len = metadata.len();
        // Linux requires pmem region sizes to be 2 MiB aligned. Linux will fill any partial page
        // at the end of an mmap'd file and won't write back beyond the actual file length, but if
        // we just align the size of the file to 2 MiB then access beyond the last page of the
        // mapped file will generate SIGBUS. So use a memory mapping arena that will provide
        // padding up to 2 MiB.
        let alignment = 2 * 1024 * 1024;
        let align_adjust = if disk_len % alignment != 0 {
            alignment - (disk_len % alignment)
        } else {
            0
        };
        (
            disk_len,
            disk_len
                .checked_add(align_adjust)
                .ok_or(Error::PmemDeviceImageTooBig)?,
        )
    };

    let protection = {
        if disk.read_only {
            Protection::read()
        } else {
            Protection::read_write()
        }
    };

    let arena = {
        // Conversion from u64 to usize may fail on 32bit system.
        let arena_size = usize::try_from(arena_size).map_err(|_| Error::PmemDeviceImageTooBig)?;
        let disk_size = usize::try_from(disk_size).map_err(|_| Error::PmemDeviceImageTooBig)?;

        let mut arena = MemoryMappingArena::new(arena_size).map_err(Error::ReservePmemMemory)?;
        arena
            .add_fd_offset_protection(0, disk_size, &fd, 0, protection)
            .map_err(Error::ReservePmemMemory)?;
        arena
    };

    let mapping_address = resources
        .mmio_allocator(MmioType::High)
        .allocate_with_align(
            arena_size,
            Alloc::PmemDevice(index),
            format!("pmem_disk_image_{}", index),
            // Linux kernel requires pmem namespaces to be 128 MiB aligned.
            128 * 1024 * 1024, /* 128 MiB */
        )
        .map_err(Error::AllocatePmemDeviceAddress)?;

    vm.add_mmap_arena(
        GuestAddress(mapping_address),
        arena,
        /* read_only = */ disk.read_only,
        /* log_dirty_pages = */ false,
    )
    .map_err(Error::AddPmemDeviceMemory)?;

    let dev = virtio::Pmem::new(fd, GuestAddress(mapping_address), arena_size)
        .map_err(Error::PmemDeviceNew)?;

    Ok(VirtioDeviceStub {
        dev: Box::new(dev) as Box<dyn VirtioDevice>,
        jail: simple_jail(&cfg, "pmem_device")?,
    })
}

// gpu_device_socket is not used when GPU support is disabled.
#[cfg_attr(not(feature = "gpu"), allow(unused_variables))]
fn create_virtio_devices(
    cfg: &Config,
    mem: &GuestMemory,
    vm: &mut Vm,
    resources: &mut SystemAllocator,
    _exit_evt: &EventFd,
    wayland_device_socket: VmMemoryControlRequestSocket,
    gpu_device_socket: VmMemoryControlRequestSocket,
    balloon_device_socket: BalloonControlResponseSocket,
    disk_device_sockets: &mut Vec<DiskControlResponseSocket>,
) -> DeviceResult<Vec<VirtioDeviceStub>> {
    let mut devs = Vec::new();

    for disk in &cfg.disks {
        let disk_device_socket = disk_device_sockets.remove(0);
        devs.push(create_block_device(cfg, disk, disk_device_socket)?);
    }

    for (index, pmem_disk) in cfg.pmem_devices.iter().enumerate() {
        devs.push(create_pmem_device(cfg, vm, resources, pmem_disk, index)?);
    }

    devs.push(create_rng_device(cfg)?);

    #[cfg(feature = "tpm")]
    {
        if cfg.software_tpm {
            devs.push(create_tpm_device(cfg)?);
        }
    }

    if let Some(single_touch_spec) = &cfg.virtio_single_touch {
        devs.push(create_single_touch_device(cfg, single_touch_spec)?);
    }

    if let Some(trackpad_spec) = &cfg.virtio_trackpad {
        devs.push(create_trackpad_device(cfg, trackpad_spec)?);
    }

    if let Some(mouse_socket) = &cfg.virtio_mouse {
        devs.push(create_mouse_device(cfg, mouse_socket)?);
    }

    if let Some(keyboard_socket) = &cfg.virtio_keyboard {
        devs.push(create_keyboard_device(cfg, keyboard_socket)?);
    }

    for dev_path in &cfg.virtio_input_evdevs {
        devs.push(create_vinput_device(cfg, dev_path)?);
    }

    devs.push(create_balloon_device(cfg, balloon_device_socket)?);

    // We checked above that if the IP is defined, then the netmask is, too.
    for tap_fd in &cfg.tap_fd {
        devs.push(create_tap_net_device(cfg, *tap_fd)?);
    }

    if let (Some(host_ip), Some(netmask), Some(mac_address)) =
        (cfg.host_ip, cfg.netmask, cfg.mac_address)
    {
        devs.push(create_net_device(cfg, host_ip, netmask, mac_address, mem)?);
    }

    #[cfg_attr(not(feature = "gpu"), allow(unused_mut))]
    let mut resource_bridges = Vec::<virtio::resource_bridge::ResourceResponseSocket>::new();

    if !cfg.wayland_socket_paths.is_empty() {
        #[cfg_attr(not(feature = "gpu"), allow(unused_mut))]
        let mut wl_resource_bridge = None::<virtio::resource_bridge::ResourceRequestSocket>;

        #[cfg(feature = "gpu")]
        {
            if cfg.gpu_parameters.is_some() {
                let (wl_socket, gpu_socket) =
                    virtio::resource_bridge::pair().map_err(Error::CreateSocket)?;
                resource_bridges.push(gpu_socket);
                wl_resource_bridge = Some(wl_socket);
            }
        }

        devs.push(create_wayland_device(
            cfg,
            wayland_device_socket,
            wl_resource_bridge,
        )?);
    }

    #[cfg(feature = "gpu")]
    {
        if cfg.gpu_parameters.is_some() {
            let mut event_devices = Vec::new();
            if cfg.display_window_mouse {
                let (event_device_socket, virtio_dev_socket) =
                    UnixStream::pair().map_err(Error::CreateSocket)?;
                // TODO(nkgold): the width/height here should match the display's height/width. When
                // those settings are available as CLI options, we should use the CLI options here
                // as well.
                let (single_touch_width, single_touch_height) = cfg
                    .virtio_single_touch
                    .as_ref()
                    .map(|single_touch_spec| single_touch_spec.get_size())
                    .unwrap_or((DEFAULT_TOUCH_DEVICE_WIDTH, DEFAULT_TOUCH_DEVICE_HEIGHT));
                let dev = virtio::new_single_touch(
                    virtio_dev_socket,
                    single_touch_width,
                    single_touch_height,
                )
                .map_err(Error::InputDeviceNew)?;
                devs.push(VirtioDeviceStub {
                    dev: Box::new(dev),
                    jail: simple_jail(&cfg, "input_device")?,
                });
                event_devices.push(EventDevice::touchscreen(event_device_socket));
            }
            if cfg.display_window_keyboard {
                let (event_device_socket, virtio_dev_socket) =
                    UnixStream::pair().map_err(Error::CreateSocket)?;
                let dev = virtio::new_keyboard(virtio_dev_socket).map_err(Error::InputDeviceNew)?;
                devs.push(VirtioDeviceStub {
                    dev: Box::new(dev),
                    jail: simple_jail(&cfg, "input_device")?,
                });
                event_devices.push(EventDevice::keyboard(event_device_socket));
            }
            devs.push(create_gpu_device(
                cfg,
                _exit_evt,
                gpu_device_socket,
                resource_bridges,
                // Use the unnamed socket for GPU display screens.
                cfg.wayland_socket_paths.get(""),
                cfg.x_display.clone(),
                event_devices,
            )?);
        }
    }

    if let Some(cid) = cfg.cid {
        devs.push(create_vhost_vsock_device(cfg, cid, mem)?);
    }

    for shared_dir in &cfg.shared_dirs {
        let SharedDir {
            src,
            tag,
            kind,
            uid_map,
            gid_map,
            cfg: fs_cfg,
        } = shared_dir;

        let dev = match kind {
            SharedDirKind::FS => create_fs_device(cfg, uid_map, gid_map, src, tag, fs_cfg.clone())?,
            SharedDirKind::P9 => create_9p_device(cfg, uid_map, gid_map, src, tag)?,
        };
        devs.push(dev);
    }

    Ok(devs)
}

fn create_devices(
    cfg: &Config,
    mem: &GuestMemory,
    vm: &mut Vm,
    resources: &mut SystemAllocator,
    exit_evt: &EventFd,
    control_sockets: &mut Vec<TaggedControlSocket>,
    wayland_device_socket: VmMemoryControlRequestSocket,
    gpu_device_socket: VmMemoryControlRequestSocket,
    balloon_device_socket: BalloonControlResponseSocket,
    disk_device_sockets: &mut Vec<DiskControlResponseSocket>,
    usb_provider: HostBackendDeviceProvider,
) -> DeviceResult<Vec<(Box<dyn PciDevice>, Option<Minijail>)>> {
    let stubs = create_virtio_devices(
        &cfg,
        mem,
        vm,
        resources,
        exit_evt,
        wayland_device_socket,
        gpu_device_socket,
        balloon_device_socket,
        disk_device_sockets,
    )?;

    let mut pci_devices = Vec::new();

    for stub in stubs {
        let (msi_host_socket, msi_device_socket) =
            msg_socket::pair::<VmIrqResponse, VmIrqRequest>().map_err(Error::CreateSocket)?;
        control_sockets.push(TaggedControlSocket::VmIrq(msi_host_socket));
        let dev = VirtioPciDevice::new(mem.clone(), stub.dev, msi_device_socket)
            .map_err(Error::VirtioPciDev)?;
        let dev = Box::new(dev) as Box<dyn PciDevice>;
        pci_devices.push((dev, stub.jail));
    }

    if cfg.cras_audio {
        let mut server = Box::new(CrasClient::new().map_err(Error::CreateCrasClient)?);
        if cfg.cras_capture {
            server.enable_cras_capture();
        }
        let cras_audio = devices::Ac97Dev::new(mem.clone(), server);

        pci_devices.push((
            Box::new(cras_audio),
            simple_jail(&cfg, "cras_audio_device")?,
        ));
    }

    if cfg.null_audio {
        let server = Box::new(NullShmStreamSource::new());
        let null_audio = devices::Ac97Dev::new(mem.clone(), server);

        pci_devices.push((
            Box::new(null_audio),
            simple_jail(&cfg, "null_audio_device")?,
        ));
    }
    // Create xhci controller.
    let usb_controller = Box::new(XhciController::new(mem.clone(), usb_provider));
    pci_devices.push((usb_controller, simple_jail(&cfg, "xhci")?));

    if cfg.vfio.is_some() {
        let (vfio_host_socket_irq, vfio_device_socket_irq) =
            msg_socket::pair::<VmIrqResponse, VmIrqRequest>().map_err(Error::CreateSocket)?;
        control_sockets.push(TaggedControlSocket::VmIrq(vfio_host_socket_irq));

        let (vfio_host_socket_mem, vfio_device_socket_mem) =
            msg_socket::pair::<VmMemoryResponse, VmMemoryRequest>().map_err(Error::CreateSocket)?;
        control_sockets.push(TaggedControlSocket::VmMemory(vfio_host_socket_mem));

        let vfio_path = cfg.vfio.as_ref().unwrap().as_path();
        let vfiodevice =
            VfioDevice::new(vfio_path, vm, mem.clone()).map_err(Error::CreateVfioDevice)?;
        let vfiopcidevice = Box::new(VfioPciDevice::new(
            vfiodevice,
            vfio_device_socket_irq,
            vfio_device_socket_mem,
        ));
        pci_devices.push((vfiopcidevice, simple_jail(&cfg, "vfio_device")?));
    }

    Ok(pci_devices)
}

#[derive(Copy, Clone)]
#[cfg_attr(not(feature = "tpm"), allow(dead_code))]
struct Ids {
    uid: uid_t,
    gid: gid_t,
}

// Set the uid/gid for the jailed process and give a basic id map. This is
// required for bind mounts to work.
fn add_crosvm_user_to_jail(jail: &mut Minijail, feature: &str) -> Result<Ids> {
    let crosvm_user_group = CStr::from_bytes_with_nul(b"crosvm\0").unwrap();

    let crosvm_uid = match get_user_id(&crosvm_user_group) {
        Ok(u) => u,
        Err(e) => {
            warn!("falling back to current user id for {}: {}", feature, e);
            geteuid()
        }
    };

    let crosvm_gid = match get_group_id(&crosvm_user_group) {
        Ok(u) => u,
        Err(e) => {
            warn!("falling back to current group id for {}: {}", feature, e);
            getegid()
        }
    };

    jail.change_uid(crosvm_uid);
    jail.change_gid(crosvm_gid);
    jail.uidmap(&format!("{0} {0} 1", crosvm_uid))
        .map_err(Error::SettingUidMap)?;
    jail.gidmap(&format!("{0} {0} 1", crosvm_gid))
        .map_err(Error::SettingGidMap)?;

    Ok(Ids {
        uid: crosvm_uid,
        gid: crosvm_gid,
    })
}

fn raw_fd_from_path(path: &Path) -> Result<RawFd> {
    if !path.is_file() {
        return Err(Error::InvalidFdPath);
    }
    let raw_fd = path
        .file_name()
        .and_then(|fd_osstr| fd_osstr.to_str())
        .and_then(|fd_str| fd_str.parse::<c_int>().ok())
        .ok_or(Error::InvalidFdPath)?;
    validate_raw_fd(raw_fd).map_err(Error::ValidateRawFd)
}

trait IntoUnixStream {
    fn into_unix_stream(self) -> Result<UnixStream>;
}

impl<'a> IntoUnixStream for &'a Path {
    fn into_unix_stream(self) -> Result<UnixStream> {
        if self.parent() == Some(Path::new("/proc/self/fd")) {
            // Safe because we will validate |raw_fd|.
            unsafe { Ok(UnixStream::from_raw_fd(raw_fd_from_path(self)?)) }
        } else {
            UnixStream::connect(self).map_err(Error::InputEventsOpen)
        }
    }
}
impl<'a> IntoUnixStream for &'a PathBuf {
    fn into_unix_stream(self) -> Result<UnixStream> {
        self.as_path().into_unix_stream()
    }
}

impl IntoUnixStream for UnixStream {
    fn into_unix_stream(self) -> Result<UnixStream> {
        Ok(self)
    }
}

fn setup_vcpu_signal_handler(use_kvm_signals: bool) -> Result<()> {
    if use_kvm_signals {
        unsafe {
            extern "C" fn handle_signal() {}
            // Our signal handler does nothing and is trivially async signal safe.
            register_rt_signal_handler(SIGRTMIN() + 0, handle_signal)
                .map_err(Error::RegisterSignalHandler)?;
        }
        block_signal(SIGRTMIN() + 0).map_err(Error::BlockSignal)?;
    } else {
        unsafe {
            extern "C" fn handle_signal() {
                Vcpu::set_local_immediate_exit(true);
            }
            register_rt_signal_handler(SIGRTMIN() + 0, handle_signal)
                .map_err(Error::RegisterSignalHandler)?;
        }
    }
    Ok(())
}

#[derive(Default)]
struct VcpuRunMode {
    mtx: Mutex<VmRunMode>,
    cvar: Condvar,
}

impl VcpuRunMode {
    fn set_and_notify(&self, new_mode: VmRunMode) {
        *self.mtx.lock() = new_mode;
        self.cvar.notify_all();
    }
}

// Converts a vcpu into a runnable vcpu if possible. On failure, returns `None`.
fn runnable_vcpu(vcpu: Vcpu, use_kvm_signals: bool, cpu_id: u32) -> Option<RunnableVcpu> {
    if use_kvm_signals {
        match get_blocked_signals() {
            Ok(mut v) => {
                v.retain(|&x| x != SIGRTMIN() + 0);
                if let Err(e) = vcpu.set_signal_mask(&v) {
                    error!(
                        "Failed to set the KVM_SIGNAL_MASK for vcpu {} : {}",
                        cpu_id, e
                    );
                    return None;
                }
            }
            Err(e) => {
                error!("Failed to retrieve signal mask for vcpu {} : {}", cpu_id, e);
                return None;
            }
        };
    }

    match vcpu.to_runnable(Some(SIGRTMIN() + 0)) {
        Ok(v) => Some(v),
        Err(e) => {
            error!("Failed to set thread id for vcpu {} : {}", cpu_id, e);
            None
        }
    }
}

#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
fn inject_interrupt(pic: &Arc<Mutex<devices::Pic>>, vcpu: &RunnableVcpu) {
    let mut pic = pic.lock();
    if pic.interrupt_requested() && vcpu.ready_for_interrupt() {
        if let Some(vector) = pic.get_external_interrupt() {
            if let Err(e) = vcpu.interrupt(vector as u32) {
                error!("PIC: failed to inject interrupt to vCPU0: {}", e);
            }
        }
        // The second interrupt request should be handled immediately, so ask
        // vCPU to exit as soon as possible.
        if pic.interrupt_requested() {
            vcpu.request_interrupt_window();
        }
    }
}

#[cfg(any(target_arch = "arm", target_arch = "aarch64"))]
fn inject_interrupt(pic: &Arc<Mutex<devices::Pic>>, vcpu: &RunnableVcpu) {}

fn run_vcpu(
    vcpu: Vcpu,
    cpu_id: u32,
    vcpu_affinity: Vec<usize>,
    start_barrier: Arc<Barrier>,
    io_bus: devices::Bus,
    mmio_bus: devices::Bus,
    split_irqchip: Option<(Arc<Mutex<devices::Pic>>, Arc<Mutex<devices::Ioapic>>)>,
    exit_evt: EventFd,
    requires_kvmclock_ctrl: bool,
    run_mode_arc: Arc<VcpuRunMode>,
    use_kvm_signals: bool,
) -> Result<JoinHandle<()>> {
    thread::Builder::new()
        .name(format!("crosvm_vcpu{}", cpu_id))
        .spawn(move || {
            // The VCPU thread must trigger the `exit_evt` in all paths, and a `ScopedEvent`'s Drop
            // implementation accomplishes that.
            let _scoped_exit_evt = ScopedEvent::from(exit_evt);

            if vcpu_affinity.len() != 0 {
                if let Err(e) = set_cpu_affinity(vcpu_affinity) {
                    error!("Failed to set CPU affinity: {}", e);
                }
            }

            let vcpu = runnable_vcpu(vcpu, use_kvm_signals, cpu_id);

            start_barrier.wait();

            if let Some(vcpu) = vcpu {
                loop {
                    let mut interrupted_by_signal = false;
                    match vcpu.run() {
                        Ok(VcpuExit::IoIn { port, mut size }) => {
                            let mut data = [0; 8];
                            if size > data.len() {
                                error!("unsupported IoIn size of {} bytes", size);
                                size = data.len();
                            }
                            io_bus.read(port as u64, &mut data[..size]);
                            if let Err(e) = vcpu.set_data(&data[..size]) {
                                error!("failed to set return data for IoIn: {}", e);
                            }
                        }
                        Ok(VcpuExit::IoOut {
                            port,
                            mut size,
                            data,
                        }) => {
                            if size > data.len() {
                                error!("unsupported IoOut size of {} bytes", size);
                                size = data.len();
                            }
                            io_bus.write(port as u64, &data[..size]);
                        }
                        Ok(VcpuExit::MmioRead { address, size }) => {
                            let mut data = [0; 8];
                            mmio_bus.read(address, &mut data[..size]);
                            // Setting data for mmio can not fail.
                            let _ = vcpu.set_data(&data[..size]);
                        }
                        Ok(VcpuExit::MmioWrite {
                            address,
                            size,
                            data,
                        }) => {
                            mmio_bus.write(address, &data[..size]);
                        }
                        Ok(VcpuExit::IoapicEoi{vector}) => {
                            if let Some((_, ioapic)) = &split_irqchip {
                                ioapic.lock().end_of_interrupt(vector);
                            } else {
                                panic!("userspace ioapic not found in split irqchip mode, should be impossible.");
                            }
                        },
                        Ok(VcpuExit::Hlt) => break,
                        Ok(VcpuExit::Shutdown) => break,
                        Ok(VcpuExit::FailEntry {
                            hardware_entry_failure_reason,
                        }) => {
                            error!("vcpu hw run failure: {:#x}", hardware_entry_failure_reason);
                            break;
                        },
                        Ok(VcpuExit::SystemEvent(_, _)) => break,
                        Ok(r) => warn!("unexpected vcpu exit: {:?}", r),
                        Err(e) => match e.errno() {
                            libc::EINTR => interrupted_by_signal = true,
                            libc::EAGAIN => {}
                            _ => {
                                error!("vcpu hit unknown error: {}", e);
                                break;
                            }
                        },
                    }

                    if interrupted_by_signal {
                        if use_kvm_signals {
                            // Try to clear the signal that we use to kick VCPU if it is pending before
                            // attempting to handle pause requests.
                            if let Err(e) = clear_signal(SIGRTMIN() + 0) {
                                error!("failed to clear pending signal: {}", e);
                                break;
                            }
                        } else {
                            vcpu.set_immediate_exit(false);
                        }
                        let mut run_mode_lock = run_mode_arc.mtx.lock();
                        loop {
                            match *run_mode_lock {
                                VmRunMode::Running => break,
                                VmRunMode::Suspending => {
                                    // On KVM implementations that use a paravirtualized clock (e.g.
                                    // x86), a flag must be set to indicate to the guest kernel that
                                    // a VCPU was suspended. The guest kernel will use this flag to
                                    // prevent the soft lockup detection from triggering when this
                                    // VCPU resumes, which could happen days later in realtime.
                                    if requires_kvmclock_ctrl {
                                        if let Err(e) = vcpu.kvmclock_ctrl() {
                                            error!("failed to signal to kvm that vcpu {} is being suspended: {}", cpu_id, e);
                                        }
                                    }
                                }
                                VmRunMode::Exiting => return,
                            }
                            // Give ownership of our exclusive lock to the condition variable that
                            // will block. When the condition variable is notified, `wait` will
                            // unblock and return a new exclusive lock.
                            run_mode_lock = run_mode_arc.cvar.wait(run_mode_lock);
                        }
                    }

                    if cpu_id != 0 { continue; }
                    if let Some((pic, _)) = &split_irqchip {
                        inject_interrupt(pic, &vcpu);
                    }
                }
            }
        })
        .map_err(Error::SpawnVcpu)
}

// Reads the contents of a file and converts the space-separated fields into a Vec of u64s.
// Returns an error if any of the fields fail to parse.
fn file_fields_to_u64<P: AsRef<Path>>(path: P) -> io::Result<Vec<u64>> {
    let mut file = File::open(path)?;

    let mut buf = [0u8; 32];
    let count = file.read(&mut buf)?;

    let content =
        str::from_utf8(&buf[..count]).map_err(|e| io::Error::new(io::ErrorKind::InvalidData, e))?;
    content
        .trim()
        .split_whitespace()
        .map(|x| {
            x.parse::<u64>()
                .map_err(|e| io::Error::new(io::ErrorKind::InvalidData, e))
        })
        .collect()
}

// Reads the contents of a file and converts them into a u64, and if there
// are multiple fields it only returns the first one.
fn file_to_u64<P: AsRef<Path>>(path: P) -> io::Result<u64> {
    file_fields_to_u64(path)?
        .into_iter()
        .next()
        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "empty file"))
}

pub fn run_config(cfg: Config) -> Result<()> {
    if cfg.sandbox {
        // Printing something to the syslog before entering minijail so that libc's syslogger has a
        // chance to open files necessary for its operation, like `/etc/localtime`. After jailing,
        // access to those files will not be possible.
        info!("crosvm entering multiprocess mode");
    }

    let (usb_control_socket, usb_provider) =
        HostBackendDeviceProvider::new().map_err(Error::CreateUsbProvider)?;
    // Masking signals is inherently dangerous, since this can persist across clones/execs. Do this
    // before any jailed devices have been spawned, so that we can catch any of them that fail very
    // quickly.
    let sigchld_fd = SignalFd::new(libc::SIGCHLD).map_err(Error::CreateSignalFd)?;

    let initrd_image = if let Some(initrd_path) = &cfg.initrd_path {
        Some(File::open(initrd_path).map_err(|e| Error::OpenInitrd(initrd_path.clone(), e))?)
    } else {
        None
    };

    let vm_image = match cfg.executable_path {
        Some(Executable::Kernel(ref kernel_path)) => VmImage::Kernel(
            File::open(kernel_path).map_err(|e| Error::OpenKernel(kernel_path.to_path_buf(), e))?,
        ),
        Some(Executable::Bios(ref bios_path)) => VmImage::Bios(
            File::open(bios_path).map_err(|e| Error::OpenBios(bios_path.to_path_buf(), e))?,
        ),
        _ => panic!("Did not receive a bios or kernel, should be impossible."),
    };

    let components = VmComponents {
        memory_size: cfg
            .memory
            .unwrap_or(256)
            .checked_mul(1024 * 1024)
            .ok_or(Error::MemoryTooLarge)?,
        vcpu_count: cfg.vcpu_count.unwrap_or(1),
        vcpu_affinity: cfg.vcpu_affinity.clone(),
        vm_image,
        android_fstab: cfg
            .android_fstab
            .as_ref()
            .map(|x| File::open(x).map_err(|e| Error::OpenAndroidFstab(x.to_path_buf(), e)))
            .map_or(Ok(None), |v| v.map(Some))?,
        pstore: cfg.pstore.clone(),
        initrd_image,
        extra_kernel_params: cfg.params.clone(),
        wayland_dmabuf: cfg.wayland_dmabuf,
    };

    let control_server_socket = match &cfg.socket_path {
        Some(path) => Some(UnlinkUnixSeqpacketListener(
            UnixSeqpacketListener::bind(path).map_err(Error::CreateSocket)?,
        )),
        None => None,
    };

    let mut control_sockets = Vec::new();
    let (wayland_host_socket, wayland_device_socket) =
        msg_socket::pair::<VmMemoryResponse, VmMemoryRequest>().map_err(Error::CreateSocket)?;
    control_sockets.push(TaggedControlSocket::VmMemory(wayland_host_socket));
    // Balloon gets a special socket so balloon requests can be forwarded from the main process.
    let (balloon_host_socket, balloon_device_socket) =
        msg_socket::pair::<BalloonControlCommand, ()>().map_err(Error::CreateSocket)?;

    // Create one control socket per disk.
    let mut disk_device_sockets = Vec::new();
    let mut disk_host_sockets = Vec::new();
    let disk_count = cfg.disks.len();
    for _ in 0..disk_count {
        let (disk_host_socket, disk_device_socket) =
            msg_socket::pair::<DiskControlCommand, DiskControlResult>()
                .map_err(Error::CreateSocket)?;
        disk_host_sockets.push(disk_host_socket);
        disk_device_sockets.push(disk_device_socket);
    }

    let (gpu_host_socket, gpu_device_socket) =
        msg_socket::pair::<VmMemoryResponse, VmMemoryRequest>().map_err(Error::CreateSocket)?;
    control_sockets.push(TaggedControlSocket::VmMemory(gpu_host_socket));

    let (ioapic_host_socket, ioapic_device_socket) =
        msg_socket::pair::<VmIrqResponse, VmIrqRequest>().map_err(Error::CreateSocket)?;
    control_sockets.push(TaggedControlSocket::VmIrq(ioapic_host_socket));

    let sandbox = cfg.sandbox;
    let linux = Arch::build_vm(
        components,
        cfg.split_irqchip,
        ioapic_device_socket,
        &cfg.serial_parameters,
        simple_jail(&cfg, "serial")?,
        |mem, vm, sys_allocator, exit_evt| {
            create_devices(
                &cfg,
                mem,
                vm,
                sys_allocator,
                exit_evt,
                &mut control_sockets,
                wayland_device_socket,
                gpu_device_socket,
                balloon_device_socket,
                &mut disk_device_sockets,
                usb_provider,
            )
        },
    )
    .map_err(Error::BuildVm)?;

    run_control(
        linux,
        control_server_socket,
        control_sockets,
        balloon_host_socket,
        &disk_host_sockets,
        usb_control_socket,
        sigchld_fd,
        sandbox,
    )
}

fn run_control(
    mut linux: RunnableLinuxVm,
    control_server_socket: Option<UnlinkUnixSeqpacketListener>,
    mut control_sockets: Vec<TaggedControlSocket>,
    balloon_host_socket: BalloonControlRequestSocket,
    disk_host_sockets: &[DiskControlRequestSocket],
    usb_control_socket: UsbControlSocket,
    sigchld_fd: SignalFd,
    sandbox: bool,
) -> Result<()> {
    // Paths to get the currently available memory and the low memory threshold.
    const LOWMEM_MARGIN: &str = "/sys/kernel/mm/chromeos-low_mem/margin";
    const LOWMEM_AVAILABLE: &str = "/sys/kernel/mm/chromeos-low_mem/available";

    // The amount of additional memory to claim back from the VM whenever the system is
    // low on memory.
    const ONE_GB: u64 = (1 << 30);

    let max_balloon_memory = match linux.vm.get_memory().memory_size() {
        // If the VM has at least 1.5 GB, the balloon driver can consume all but the last 1 GB.
        n if n >= (ONE_GB / 2) * 3 => n - ONE_GB,
        // Otherwise, if the VM has at least 500MB the balloon driver will consume at most
        // half of it.
        n if n >= (ONE_GB / 2) => n / 2,
        // Otherwise, the VM is too small for us to take memory away from it.
        _ => 0,
    };
    let mut current_balloon_memory: u64 = 0;
    let balloon_memory_increment: u64 = max_balloon_memory / 16;

    #[derive(PollToken)]
    enum Token {
        Exit,
        Suspend,
        ChildSignal,
        CheckAvailableMemory,
        LowMemory,
        LowmemTimer,
        VmControlServer,
        VmControl { index: usize },
    }

    stdin()
        .set_raw_mode()
        .expect("failed to set terminal raw mode");

    let poll_ctx = PollContext::build_with(&[
        (&linux.exit_evt, Token::Exit),
        (&linux.suspend_evt, Token::Suspend),
        (&sigchld_fd, Token::ChildSignal),
    ])
    .map_err(Error::PollContextAdd)?;

    if let Some(socket_server) = &control_server_socket {
        poll_ctx
            .add(socket_server, Token::VmControlServer)
            .map_err(Error::PollContextAdd)?;
    }
    for (index, socket) in control_sockets.iter().enumerate() {
        poll_ctx
            .add(socket.as_ref(), Token::VmControl { index })
            .map_err(Error::PollContextAdd)?;
    }

    // Watch for low memory notifications and take memory back from the VM.
    let low_mem = File::open("/dev/chromeos-low-mem").ok();
    if let Some(low_mem) = &low_mem {
        poll_ctx
            .add(low_mem, Token::LowMemory)
            .map_err(Error::PollContextAdd)?;
    } else {
        warn!("Unable to open low mem indicator, maybe not a chrome os kernel");
    }

    // Used to rate limit balloon requests.
    let mut lowmem_timer = TimerFd::new().map_err(Error::CreateTimerFd)?;
    poll_ctx
        .add(&lowmem_timer, Token::LowmemTimer)
        .map_err(Error::PollContextAdd)?;

    // Used to check whether it's ok to start giving memory back to the VM.
    let mut freemem_timer = TimerFd::new().map_err(Error::CreateTimerFd)?;
    poll_ctx
        .add(&freemem_timer, Token::CheckAvailableMemory)
        .map_err(Error::PollContextAdd)?;

    // Used to add jitter to timer values so that we don't have a thundering herd problem when
    // multiple VMs are running.
    let mut simple_rng = SimpleRng::new(
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time went backwards")
            .subsec_nanos() as u64,
    );

    if sandbox {
        // Before starting VCPUs, in case we started with some capabilities, drop them all.
        drop_capabilities().map_err(Error::DropCapabilities)?;
    }

    let mut vcpu_handles = Vec::with_capacity(linux.vcpus.len());
    let vcpu_thread_barrier = Arc::new(Barrier::new(linux.vcpus.len() + 1));
    let run_mode_arc = Arc::new(VcpuRunMode::default());
    let use_kvm_signals = !linux.kvm.check_extension(Cap::ImmediateExit);
    setup_vcpu_signal_handler(use_kvm_signals)?;
    let vcpus = linux.vcpus.split_off(0);
    for (cpu_id, vcpu) in vcpus.into_iter().enumerate() {
        let handle = run_vcpu(
            vcpu,
            cpu_id as u32,
            linux.vcpu_affinity.clone(),
            vcpu_thread_barrier.clone(),
            linux.io_bus.clone(),
            linux.mmio_bus.clone(),
            linux.split_irqchip.clone(),
            linux.exit_evt.try_clone().map_err(Error::CloneEventFd)?,
            linux.vm.check_extension(Cap::KvmclockCtrl),
            run_mode_arc.clone(),
            use_kvm_signals,
        )?;
        vcpu_handles.push(handle);
    }
    vcpu_thread_barrier.wait();

    'poll: loop {
        let events = {
            match poll_ctx.wait() {
                Ok(v) => v,
                Err(e) => {
                    error!("failed to poll: {}", e);
                    break;
                }
            }
        };

        let mut vm_control_indices_to_remove = Vec::new();
        for event in events.iter_readable() {
            match event.token() {
                Token::Exit => {
                    info!("vcpu requested shutdown");
                    break 'poll;
                }
                Token::Suspend => {
                    info!("VM requested suspend");
                    linux.suspend_evt.read().unwrap();
                    run_mode_arc.set_and_notify(VmRunMode::Suspending);
                    for handle in &vcpu_handles {
                        let _ = handle.kill(SIGRTMIN() + 0);
                    }
                }
                Token::ChildSignal => {
                    // Print all available siginfo structs, then exit the loop.
                    while let Some(siginfo) = sigchld_fd.read().map_err(Error::SignalFd)? {
                        let pid = siginfo.ssi_pid;
                        let pid_label = match linux.pid_debug_label_map.get(&pid) {
                            Some(label) => format!("{} (pid {})", label, pid),
                            None => format!("pid {}", pid),
                        };
                        error!(
                            "child {} died: signo {}, status {}, code {}",
                            pid_label, siginfo.ssi_signo, siginfo.ssi_status, siginfo.ssi_code
                        );
                    }
                    break 'poll;
                }
                Token::CheckAvailableMemory => {
                    // Acknowledge the timer.
                    freemem_timer.wait().map_err(Error::TimerFd)?;
                    if current_balloon_memory == 0 {
                        // Nothing to see here.
                        if let Err(e) = freemem_timer.clear() {
                            warn!("unable to clear available memory check timer: {}", e);
                        }
                        continue;
                    }

                    // Otherwise see if we can free up some memory.
                    let margin = file_to_u64(LOWMEM_MARGIN).map_err(Error::ReadLowmemMargin)?;
                    let available =
                        file_to_u64(LOWMEM_AVAILABLE).map_err(Error::ReadLowmemAvailable)?;

                    // `available` and `margin` are specified in MB while `balloon_memory_increment` is in
                    // bytes.  So to correctly compare them we need to turn the increment value into MB.
                    if available >= margin + 2 * (balloon_memory_increment >> 20) {
                        current_balloon_memory =
                            if current_balloon_memory >= balloon_memory_increment {
                                current_balloon_memory - balloon_memory_increment
                            } else {
                                0
                            };
                        let command = BalloonControlCommand::Adjust {
                            num_bytes: current_balloon_memory,
                        };
                        if let Err(e) = balloon_host_socket.send(&command) {
                            warn!("failed to send memory value to balloon device: {}", e);
                        }
                    }
                }
                Token::LowMemory => {
                    if let Some(low_mem) = &low_mem {
                        let old_balloon_memory = current_balloon_memory;
                        current_balloon_memory = min(
                            current_balloon_memory + balloon_memory_increment,
                            max_balloon_memory,
                        );
                        if current_balloon_memory != old_balloon_memory {
                            let command = BalloonControlCommand::Adjust {
                                num_bytes: current_balloon_memory,
                            };
                            if let Err(e) = balloon_host_socket.send(&command) {
                                warn!("failed to send memory value to balloon device: {}", e);
                            }
                        }

                        // Stop polling the lowmem device until the timer fires.
                        poll_ctx.delete(low_mem).map_err(Error::PollContextDelete)?;

                        // Add some jitter to the timer so that if there are multiple VMs running
                        // they don't all start ballooning at exactly the same time.
                        let lowmem_dur = Duration::from_millis(1000 + simple_rng.rng() % 200);
                        lowmem_timer
                            .reset(lowmem_dur, None)
                            .map_err(Error::ResetTimerFd)?;

                        // Also start a timer to check when we can start giving memory back.  Do the
                        // first check after a minute (with jitter) and subsequent checks after
                        // every 30 seconds (with jitter).
                        let freemem_dur = Duration::from_secs(60 + simple_rng.rng() % 12);
                        let freemem_int = Duration::from_secs(30 + simple_rng.rng() % 6);
                        freemem_timer
                            .reset(freemem_dur, Some(freemem_int))
                            .map_err(Error::ResetTimerFd)?;
                    }
                }
                Token::LowmemTimer => {
                    // Acknowledge the timer.
                    lowmem_timer.wait().map_err(Error::TimerFd)?;

                    if let Some(low_mem) = &low_mem {
                        // Start polling the lowmem device again.
                        poll_ctx
                            .add(low_mem, Token::LowMemory)
                            .map_err(Error::PollContextAdd)?;
                    }
                }
                Token::VmControlServer => {
                    if let Some(socket_server) = &control_server_socket {
                        match socket_server.accept() {
                            Ok(socket) => {
                                poll_ctx
                                    .add(
                                        &socket,
                                        Token::VmControl {
                                            index: control_sockets.len(),
                                        },
                                    )
                                    .map_err(Error::PollContextAdd)?;
                                control_sockets
                                    .push(TaggedControlSocket::Vm(MsgSocket::new(socket)));
                            }
                            Err(e) => error!("failed to accept socket: {}", e),
                        }
                    }
                }
                Token::VmControl { index } => {
                    if let Some(socket) = control_sockets.get(index) {
                        match socket {
                            TaggedControlSocket::Vm(socket) => match socket.recv() {
                                Ok(request) => {
                                    let mut run_mode_opt = None;
                                    let response = request.execute(
                                        &mut run_mode_opt,
                                        &balloon_host_socket,
                                        disk_host_sockets,
                                        &usb_control_socket,
                                    );
                                    if let Err(e) = socket.send(&response) {
                                        error!("failed to send VmResponse: {}", e);
                                    }
                                    if let Some(run_mode) = run_mode_opt {
                                        info!("control socket changed run mode to {}", run_mode);
                                        match run_mode {
                                            VmRunMode::Exiting => {
                                                break 'poll;
                                            }
                                            VmRunMode::Running => {
                                                if let VmRunMode::Suspending =
                                                    *run_mode_arc.mtx.lock()
                                                {
                                                    linux.io_bus.notify_resume();
                                                }
                                                run_mode_arc.set_and_notify(VmRunMode::Running);
                                                for handle in &vcpu_handles {
                                                    let _ = handle.kill(SIGRTMIN() + 0);
                                                }
                                            }
                                            other => {
                                                run_mode_arc.set_and_notify(other);
                                                for handle in &vcpu_handles {
                                                    let _ = handle.kill(SIGRTMIN() + 0);
                                                }
                                            }
                                        }
                                    }
                                }
                                Err(e) => {
                                    if let MsgError::BadRecvSize { actual: 0, .. } = e {
                                        vm_control_indices_to_remove.push(index);
                                    } else {
                                        error!("failed to recv VmRequest: {}", e);
                                    }
                                }
                            },
                            TaggedControlSocket::VmMemory(socket) => match socket.recv() {
                                Ok(request) => {
                                    let response =
                                        request.execute(&mut linux.vm, &mut linux.resources);
                                    if let Err(e) = socket.send(&response) {
                                        error!("failed to send VmMemoryControlResponse: {}", e);
                                    }
                                }
                                Err(e) => {
                                    if let MsgError::BadRecvSize { actual: 0, .. } = e {
                                        vm_control_indices_to_remove.push(index);
                                    } else {
                                        error!("failed to recv VmMemoryControlRequest: {}", e);
                                    }
                                }
                            },
                            TaggedControlSocket::VmIrq(socket) => match socket.recv() {
                                Ok(request) => {
                                    let response =
                                        request.execute(&mut linux.vm, &mut linux.resources);
                                    if let Err(e) = socket.send(&response) {
                                        error!("failed to send VmIrqResponse: {}", e);
                                    }
                                }
                                Err(e) => {
                                    if let MsgError::BadRecvSize { actual: 0, .. } = e {
                                        vm_control_indices_to_remove.push(index);
                                    } else {
                                        error!("failed to recv VmIrqRequest: {}", e);
                                    }
                                }
                            },
                        }
                    }
                }
            }
        }

        for event in events.iter_hungup() {
            match event.token() {
                Token::Exit => {}
                Token::Suspend => {}
                Token::ChildSignal => {}
                Token::CheckAvailableMemory => {}
                Token::LowMemory => {}
                Token::LowmemTimer => {}
                Token::VmControlServer => {}
                Token::VmControl { index } => {
                    // It's possible more data is readable and buffered while the socket is hungup,
                    // so don't delete the socket from the poll context until we're sure all the
                    // data is read.
                    match control_sockets
                        .get(index)
                        .map(|s| s.as_ref().get_readable_bytes())
                    {
                        Some(Ok(0)) | Some(Err(_)) => vm_control_indices_to_remove.push(index),
                        Some(Ok(x)) => info!("control index {} has {} bytes readable", index, x),
                        _ => {}
                    }
                }
            }
        }

        // Sort in reverse so the highest indexes are removed first. This removal algorithm
        // preserves correct indexes as each element is removed.
        vm_control_indices_to_remove.sort_unstable_by(|a, b| b.cmp(a));
        vm_control_indices_to_remove.dedup();
        for index in vm_control_indices_to_remove {
            // Delete the socket from the `poll_ctx` synchronously. Otherwise, the kernel will do
            // this automatically when the FD inserted into the `poll_ctx` is closed after this
            // if-block, but this removal can be deferred unpredictably. In some instances where the
            // system is under heavy load, we can even get events returned by `poll_ctx` for an FD
            // that has already been closed. Because the token associated with that spurious event
            // now belongs to a different socket, the control loop will start to interact with
            // sockets that might not be ready to use. This can cause incorrect hangup detection or
            // blocking on a socket that will never be ready. See also: crbug.com/1019986
            if let Some(socket) = control_sockets.get(index) {
                poll_ctx.delete(socket).map_err(Error::PollContextDelete)?;
            }

            // This line implicitly drops the socket at `index` when it gets returned by
            // `swap_remove`. After this line, the socket at `index` is not the one from
            // `vm_control_indices_to_remove`. Because of this socket's change in index, we need to
            // use `poll_ctx.modify` to change the associated index in its `Token::VmControl`.
            control_sockets.swap_remove(index);
            if let Some(socket) = control_sockets.get(index) {
                poll_ctx
                    .modify(
                        socket,
                        WatchingEvents::empty().set_read(),
                        Token::VmControl { index },
                    )
                    .map_err(Error::PollContextAdd)?;
            }
        }
    }

    // VCPU threads MUST see the VmRunMode flag, otherwise they may re-enter the VM.
    run_mode_arc.set_and_notify(VmRunMode::Exiting);
    for handle in vcpu_handles {
        match handle.kill(SIGRTMIN() + 0) {
            Ok(_) => {
                if let Err(e) = handle.join() {
                    error!("failed to join vcpu thread: {:?}", e);
                }
            }
            Err(e) => error!("failed to kill vcpu thread: {}", e),
        }
    }

    // Explicitly drop the VM structure here to allow the devices to clean up before the
    // control sockets are closed when this function exits.
    mem::drop(linux);

    stdin()
        .set_canon_mode()
        .expect("failed to restore canonical mode for terminal");

    Ok(())
}