• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2017 The Chromium OS Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file.
4 
5 use std;
6 use std::cmp::min;
7 use std::convert::TryFrom;
8 use std::error::Error as StdError;
9 use std::ffi::CStr;
10 use std::fmt::{self, Display};
11 use std::fs::{File, OpenOptions};
12 use std::io::{self, stdin, Read};
13 use std::net::Ipv4Addr;
14 use std::os::unix::io::{AsRawFd, FromRawFd, RawFd};
15 use std::os::unix::net::UnixStream;
16 use std::path::{Path, PathBuf};
17 use std::str;
18 use std::sync::{Arc, Barrier};
19 use std::thread;
20 use std::thread::JoinHandle;
21 use std::time::{Duration, SystemTime, UNIX_EPOCH};
22 
23 use libc::{self, c_int, gid_t, uid_t};
24 
25 use audio_streams::DummyStreamSource;
26 use devices::virtio::{self, VirtioDevice};
27 use devices::{self, HostBackendDeviceProvider, PciDevice, VirtioPciDevice, XhciController};
28 use io_jail::{self, Minijail};
29 use kvm::*;
30 use libcras::CrasClient;
31 use msg_socket::{MsgError, MsgReceiver, MsgSender, MsgSocket};
32 use net_util::{Error as NetError, MacAddress, Tap};
33 use qcow::{self, ImageType, QcowFile};
34 use rand_ish::SimpleRng;
35 use remain::sorted;
36 use resources::{Alloc, SystemAllocator};
37 use sync::{Condvar, Mutex};
38 use sys_util::net::{UnixSeqpacket, UnixSeqpacketListener, UnlinkUnixSeqpacketListener};
39 
40 use sys_util::{
41     self, block_signal, clear_signal, drop_capabilities, error, flock, get_blocked_signals,
42     get_group_id, get_user_id, getegid, geteuid, info, register_signal_handler, set_cpu_affinity,
43     validate_raw_fd, warn, EventFd, FlockOperation, GuestAddress, GuestMemory, Killable,
44     MemoryMapping, PollContext, PollToken, Protection, SignalFd, Terminal, TimerFd, WatchingEvents,
45     SIGRTMIN,
46 };
47 use vhost;
48 use vm_control::{
49     BalloonControlCommand, BalloonControlRequestSocket, BalloonControlResponseSocket,
50     DiskControlCommand, DiskControlRequestSocket, DiskControlResponseSocket, DiskControlResult,
51     UsbControlSocket, VmControlResponseSocket, VmMemoryControlRequestSocket,
52     VmMemoryControlResponseSocket, VmMemoryRequest, VmMemoryResponse, VmRunMode,
53 };
54 
55 use crate::{Config, DiskOption, Executable, TouchDeviceOption};
56 
57 use arch::{self, LinuxArch, RunnableLinuxVm, VirtioDeviceStub, VmComponents, VmImage};
58 
59 #[cfg(any(target_arch = "arm", target_arch = "aarch64"))]
60 use aarch64::AArch64 as Arch;
61 #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
62 use x86_64::X8664arch as Arch;
63 
64 #[cfg(feature = "gpu-forward")]
65 use render_node_forward::*;
66 #[cfg(not(feature = "gpu-forward"))]
67 type RenderNodeHost = ();
68 
69 #[sorted]
70 #[derive(Debug)]
71 pub enum Error {
72     AddGpuDeviceMemory(sys_util::Error),
73     AddPmemDeviceMemory(sys_util::Error),
74     AllocateGpuDeviceAddress,
75     AllocatePmemDeviceAddress(resources::Error),
76     BalloonDeviceNew(virtio::BalloonError),
77     BlockDeviceNew(sys_util::Error),
78     BlockSignal(sys_util::signal::Error),
79     BuildVm(<Arch as LinuxArch>::Error),
80     ChownTpmStorage(sys_util::Error),
81     CloneEventFd(sys_util::Error),
82     CreateCrasClient(libcras::Error),
83     CreateEventFd(sys_util::Error),
84     CreatePollContext(sys_util::Error),
85     CreateSignalFd(sys_util::SignalFdError),
86     CreateSocket(io::Error),
87     CreateTapDevice(NetError),
88     CreateTimerFd(sys_util::Error),
89     CreateTpmStorage(PathBuf, io::Error),
90     CreateUsbProvider(devices::usb::host_backend::error::Error),
91     DetectImageType(qcow::Error),
92     DeviceJail(io_jail::Error),
93     DevicePivotRoot(io_jail::Error),
94     Disk(io::Error),
95     DiskImageLock(sys_util::Error),
96     DropCapabilities(sys_util::Error),
97     InputDeviceNew(virtio::InputError),
98     InputEventsOpen(std::io::Error),
99     InvalidFdPath,
100     InvalidWaylandPath,
101     IoJail(io_jail::Error),
102     LoadKernel(Box<dyn StdError>),
103     NetDeviceNew(virtio::NetError),
104     OpenAndroidFstab(PathBuf, io::Error),
105     OpenBios(PathBuf, io::Error),
106     OpenInitrd(PathBuf, io::Error),
107     OpenKernel(PathBuf, io::Error),
108     OpenVinput(PathBuf, io::Error),
109     P9DeviceNew(virtio::P9Error),
110     PivotRootDoesntExist(&'static str),
111     PmemDeviceImageTooBig,
112     PmemDeviceNew(sys_util::Error),
113     PollContextAdd(sys_util::Error),
114     PollContextDelete(sys_util::Error),
115     QcowDeviceCreate(qcow::Error),
116     ReadLowmemAvailable(io::Error),
117     ReadLowmemMargin(io::Error),
118     RegisterBalloon(arch::DeviceRegistrationError),
119     RegisterBlock(arch::DeviceRegistrationError),
120     RegisterGpu(arch::DeviceRegistrationError),
121     RegisterNet(arch::DeviceRegistrationError),
122     RegisterP9(arch::DeviceRegistrationError),
123     RegisterRng(arch::DeviceRegistrationError),
124     RegisterSignalHandler(sys_util::Error),
125     RegisterWayland(arch::DeviceRegistrationError),
126     ReserveGpuMemory(sys_util::MmapError),
127     ReserveMemory(sys_util::Error),
128     ReservePmemMemory(sys_util::MmapError),
129     ResetTimerFd(sys_util::Error),
130     RngDeviceNew(virtio::RngError),
131     SettingGidMap(io_jail::Error),
132     SettingUidMap(io_jail::Error),
133     SignalFd(sys_util::SignalFdError),
134     SpawnVcpu(io::Error),
135     TimerFd(sys_util::Error),
136     ValidateRawFd(sys_util::Error),
137     VhostNetDeviceNew(virtio::vhost::Error),
138     VhostVsockDeviceNew(virtio::vhost::Error),
139     VirtioPciDev(sys_util::Error),
140     WaylandDeviceNew(sys_util::Error),
141 }
142 
143 impl Display for Error {
144     #[remain::check]
fmt(&self, f: &mut fmt::Formatter) -> fmt::Result145     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
146         use self::Error::*;
147 
148         #[sorted]
149         match self {
150             AddGpuDeviceMemory(e) => write!(f, "failed to add gpu device memory: {}", e),
151             AddPmemDeviceMemory(e) => write!(f, "failed to add pmem device memory: {}", e),
152             AllocateGpuDeviceAddress => write!(f, "failed to allocate gpu device guest address"),
153             AllocatePmemDeviceAddress(e) => {
154                 write!(f, "failed to allocate memory for pmem device: {}", e)
155             }
156             BalloonDeviceNew(e) => write!(f, "failed to create balloon: {}", e),
157             BlockDeviceNew(e) => write!(f, "failed to create block device: {}", e),
158             BlockSignal(e) => write!(f, "failed to block signal: {}", e),
159             BuildVm(e) => write!(f, "The architecture failed to build the vm: {}", e),
160             ChownTpmStorage(e) => write!(f, "failed to chown tpm storage: {}", e),
161             CloneEventFd(e) => write!(f, "failed to clone eventfd: {}", e),
162             CreateCrasClient(e) => write!(f, "failed to create cras client: {}", e),
163             CreateEventFd(e) => write!(f, "failed to create eventfd: {}", e),
164             CreatePollContext(e) => write!(f, "failed to create poll context: {}", e),
165             CreateSignalFd(e) => write!(f, "failed to create signalfd: {}", e),
166             CreateSocket(e) => write!(f, "failed to create socket: {}", e),
167             CreateTapDevice(e) => write!(f, "failed to create tap device: {}", e),
168             CreateTimerFd(e) => write!(f, "failed to create timerfd: {}", e),
169             CreateTpmStorage(p, e) => {
170                 write!(f, "failed to create tpm storage dir {}: {}", p.display(), e)
171             }
172             CreateUsbProvider(e) => write!(f, "failed to create usb provider: {}", e),
173             DetectImageType(e) => write!(f, "failed to detect disk image type: {}", e),
174             DeviceJail(e) => write!(f, "failed to jail device: {}", e),
175             DevicePivotRoot(e) => write!(f, "failed to pivot root device: {}", e),
176             Disk(e) => write!(f, "failed to load disk image: {}", e),
177             DiskImageLock(e) => write!(f, "failed to lock disk image: {}", e),
178             DropCapabilities(e) => write!(f, "failed to drop process capabilities: {}", e),
179             InputDeviceNew(e) => write!(f, "failed to set up input device: {}", e),
180             InputEventsOpen(e) => write!(f, "failed to open event device: {}", e),
181             InvalidFdPath => write!(f, "failed parsing a /proc/self/fd/*"),
182             InvalidWaylandPath => write!(f, "wayland socket path has no parent or file name"),
183             IoJail(e) => write!(f, "{}", e),
184             LoadKernel(e) => write!(f, "failed to load kernel: {}", e),
185             NetDeviceNew(e) => write!(f, "failed to set up virtio networking: {}", e),
186             OpenAndroidFstab(p, e) => write!(
187                 f,
188                 "failed to open android fstab file {}: {}",
189                 p.display(),
190                 e
191             ),
192             OpenBios(p, e) => write!(f, "failed to open bios {}: {}", p.display(), e),
193             OpenInitrd(p, e) => write!(f, "failed to open initrd {}: {}", p.display(), e),
194             OpenKernel(p, e) => write!(f, "failed to open kernel image {}: {}", p.display(), e),
195             OpenVinput(p, e) => write!(f, "failed to open vinput device {}: {}", p.display(), e),
196             P9DeviceNew(e) => write!(f, "failed to create 9p device: {}", e),
197             PivotRootDoesntExist(p) => write!(f, "{} doesn't exist, can't jail devices.", p),
198             PmemDeviceImageTooBig => {
199                 write!(f, "failed to create pmem device: pmem device image too big")
200             }
201             PmemDeviceNew(e) => write!(f, "failed to create pmem device: {}", e),
202             PollContextAdd(e) => write!(f, "failed to add fd to poll context: {}", e),
203             PollContextDelete(e) => write!(f, "failed to remove fd from poll context: {}", e),
204             QcowDeviceCreate(e) => write!(f, "failed to read qcow formatted file {}", e),
205             ReadLowmemAvailable(e) => write!(
206                 f,
207                 "failed to read /sys/kernel/mm/chromeos-low_mem/available: {}",
208                 e
209             ),
210             ReadLowmemMargin(e) => write!(
211                 f,
212                 "failed to read /sys/kernel/mm/chromeos-low_mem/margin: {}",
213                 e
214             ),
215             RegisterBalloon(e) => write!(f, "error registering balloon device: {}", e),
216             RegisterBlock(e) => write!(f, "error registering block device: {}", e),
217             RegisterGpu(e) => write!(f, "error registering gpu device: {}", e),
218             RegisterNet(e) => write!(f, "error registering net device: {}", e),
219             RegisterP9(e) => write!(f, "error registering 9p device: {}", e),
220             RegisterRng(e) => write!(f, "error registering rng device: {}", e),
221             RegisterSignalHandler(e) => write!(f, "error registering signal handler: {}", e),
222             RegisterWayland(e) => write!(f, "error registering wayland device: {}", e),
223             ReserveGpuMemory(e) => write!(f, "failed to reserve gpu memory: {}", e),
224             ReserveMemory(e) => write!(f, "failed to reserve memory: {}", e),
225             ReservePmemMemory(e) => write!(f, "failed to reserve pmem memory: {}", e),
226             ResetTimerFd(e) => write!(f, "failed to reset timerfd: {}", e),
227             RngDeviceNew(e) => write!(f, "failed to set up rng: {}", e),
228             SettingGidMap(e) => write!(f, "error setting GID map: {}", e),
229             SettingUidMap(e) => write!(f, "error setting UID map: {}", e),
230             SignalFd(e) => write!(f, "failed to read signal fd: {}", e),
231             SpawnVcpu(e) => write!(f, "failed to spawn VCPU thread: {}", e),
232             TimerFd(e) => write!(f, "failed to read timer fd: {}", e),
233             ValidateRawFd(e) => write!(f, "failed to validate raw fd: {}", e),
234             VhostNetDeviceNew(e) => write!(f, "failed to set up vhost networking: {}", e),
235             VhostVsockDeviceNew(e) => write!(f, "failed to set up virtual socket device: {}", e),
236             VirtioPciDev(e) => write!(f, "failed to create virtio pci dev: {}", e),
237             WaylandDeviceNew(e) => write!(f, "failed to create wayland device: {}", e),
238         }
239     }
240 }
241 
242 impl From<io_jail::Error> for Error {
from(err: io_jail::Error) -> Self243     fn from(err: io_jail::Error) -> Self {
244         Error::IoJail(err)
245     }
246 }
247 
248 impl std::error::Error for Error {}
249 
250 type Result<T> = std::result::Result<T, Error>;
251 
252 enum TaggedControlSocket {
253     Vm(VmControlResponseSocket),
254     VmMemory(VmMemoryControlResponseSocket),
255 }
256 
257 impl AsRef<UnixSeqpacket> for TaggedControlSocket {
as_ref(&self) -> &UnixSeqpacket258     fn as_ref(&self) -> &UnixSeqpacket {
259         use self::TaggedControlSocket::*;
260         match &self {
261             Vm(ref socket) => socket,
262             VmMemory(ref socket) => socket,
263         }
264     }
265 }
266 
267 impl AsRawFd for TaggedControlSocket {
as_raw_fd(&self) -> RawFd268     fn as_raw_fd(&self) -> RawFd {
269         self.as_ref().as_raw_fd()
270     }
271 }
272 
create_base_minijail(root: &Path, seccomp_policy: &Path) -> Result<Minijail>273 fn create_base_minijail(root: &Path, seccomp_policy: &Path) -> Result<Minijail> {
274     // All child jails run in a new user namespace without any users mapped,
275     // they run as nobody unless otherwise configured.
276     let mut j = Minijail::new().map_err(Error::DeviceJail)?;
277     j.namespace_pids();
278     j.namespace_user();
279     j.namespace_user_disable_setgroups();
280     // Don't need any capabilities.
281     j.use_caps(0);
282     // Create a new mount namespace with an empty root FS.
283     j.namespace_vfs();
284     j.enter_pivot_root(root).map_err(Error::DevicePivotRoot)?;
285     // Run in an empty network namespace.
286     j.namespace_net();
287     // Apply the block device seccomp policy.
288     j.no_new_privs();
289     // Use TSYNC only for the side effect of it using SECCOMP_RET_TRAP, which will correctly kill
290     // the entire device process if a worker thread commits a seccomp violation.
291     j.set_seccomp_filter_tsync();
292     #[cfg(debug_assertions)]
293     j.log_seccomp_filter_failures();
294     j.parse_seccomp_filters(seccomp_policy)
295         .map_err(Error::DeviceJail)?;
296     j.use_seccomp_filter();
297     // Don't do init setup.
298     j.run_as_init();
299     Ok(j)
300 }
301 
simple_jail(cfg: &Config, policy: &str) -> Result<Option<Minijail>>302 fn simple_jail(cfg: &Config, policy: &str) -> Result<Option<Minijail>> {
303     if cfg.sandbox {
304         let pivot_root: &str = option_env!("DEFAULT_PIVOT_ROOT").unwrap_or("/var/empty");
305         // A directory for a jailed device's pivot root.
306         let root_path = Path::new(pivot_root);
307         if !root_path.exists() {
308             return Err(Error::PivotRootDoesntExist(pivot_root));
309         }
310         let policy_path: PathBuf = cfg.seccomp_policy_dir.join(policy);
311         Ok(Some(create_base_minijail(root_path, &policy_path)?))
312     } else {
313         Ok(None)
314     }
315 }
316 
317 type DeviceResult<T = VirtioDeviceStub> = std::result::Result<T, Error>;
318 
create_block_device( cfg: &Config, disk: &DiskOption, disk_device_socket: DiskControlResponseSocket, ) -> DeviceResult319 fn create_block_device(
320     cfg: &Config,
321     disk: &DiskOption,
322     disk_device_socket: DiskControlResponseSocket,
323 ) -> DeviceResult {
324     // Special case '/proc/self/fd/*' paths. The FD is already open, just use it.
325     let raw_image: File = if disk.path.parent() == Some(Path::new("/proc/self/fd")) {
326         // Safe because we will validate |raw_fd|.
327         unsafe { File::from_raw_fd(raw_fd_from_path(&disk.path)?) }
328     } else {
329         OpenOptions::new()
330             .read(true)
331             .write(!disk.read_only)
332             .open(&disk.path)
333             .map_err(Error::Disk)?
334     };
335     // Lock the disk image to prevent other crosvm instances from using it.
336     let lock_op = if disk.read_only {
337         FlockOperation::LockShared
338     } else {
339         FlockOperation::LockExclusive
340     };
341     flock(&raw_image, lock_op, true).map_err(Error::DiskImageLock)?;
342 
343     let image_type = qcow::detect_image_type(&raw_image).map_err(Error::DetectImageType)?;
344     let dev = match image_type {
345         ImageType::Raw => {
346             // Access as a raw block device.
347             let dev = virtio::Block::new(raw_image, disk.read_only, Some(disk_device_socket))
348                 .map_err(Error::BlockDeviceNew)?;
349             Box::new(dev) as Box<dyn VirtioDevice>
350         }
351         ImageType::Qcow2 => {
352             // Valid qcow header present
353             let qcow_image = QcowFile::from(raw_image).map_err(Error::QcowDeviceCreate)?;
354             let dev = virtio::Block::new(qcow_image, disk.read_only, Some(disk_device_socket))
355                 .map_err(Error::BlockDeviceNew)?;
356             Box::new(dev) as Box<dyn VirtioDevice>
357         }
358     };
359 
360     Ok(VirtioDeviceStub {
361         dev,
362         jail: simple_jail(&cfg, "block_device.policy")?,
363     })
364 }
365 
create_rng_device(cfg: &Config) -> DeviceResult366 fn create_rng_device(cfg: &Config) -> DeviceResult {
367     let dev = virtio::Rng::new().map_err(Error::RngDeviceNew)?;
368 
369     Ok(VirtioDeviceStub {
370         dev: Box::new(dev),
371         jail: simple_jail(&cfg, "rng_device.policy")?,
372     })
373 }
374 
375 #[cfg(feature = "tpm")]
create_tpm_device(cfg: &Config) -> DeviceResult376 fn create_tpm_device(cfg: &Config) -> DeviceResult {
377     use std::ffi::CString;
378     use std::fs;
379     use std::process;
380     use sys_util::chown;
381 
382     let tpm_storage: PathBuf;
383     let mut tpm_jail = simple_jail(&cfg, "tpm_device.policy")?;
384 
385     match &mut tpm_jail {
386         Some(jail) => {
387             // Create a tmpfs in the device's root directory for tpm
388             // simulator storage. The size is 20*1024, or 20 KB.
389             jail.mount_with_data(
390                 Path::new("none"),
391                 Path::new("/"),
392                 "tmpfs",
393                 (libc::MS_NOSUID | libc::MS_NODEV | libc::MS_NOEXEC) as usize,
394                 "size=20480",
395             )?;
396 
397             let crosvm_ids = add_crosvm_user_to_jail(jail, "tpm")?;
398 
399             let pid = process::id();
400             let tpm_pid_dir = format!("/run/vm/tpm.{}", pid);
401             tpm_storage = Path::new(&tpm_pid_dir).to_owned();
402             fs::create_dir_all(&tpm_storage)
403                 .map_err(|e| Error::CreateTpmStorage(tpm_storage.to_owned(), e))?;
404             let tpm_pid_dir_c = CString::new(tpm_pid_dir).expect("no nul bytes");
405             chown(&tpm_pid_dir_c, crosvm_ids.uid, crosvm_ids.gid)
406                 .map_err(Error::ChownTpmStorage)?;
407 
408             jail.mount_bind(&tpm_storage, &tpm_storage, true)?;
409         }
410         None => {
411             // Path used inside cros_sdk which does not have /run/vm.
412             tpm_storage = Path::new("/tmp/tpm-simulator").to_owned();
413         }
414     }
415 
416     let dev = virtio::Tpm::new(tpm_storage);
417 
418     Ok(VirtioDeviceStub {
419         dev: Box::new(dev),
420         jail: tpm_jail,
421     })
422 }
423 
create_single_touch_device(cfg: &Config, single_touch_spec: &TouchDeviceOption) -> DeviceResult424 fn create_single_touch_device(cfg: &Config, single_touch_spec: &TouchDeviceOption) -> DeviceResult {
425     let socket = create_input_socket(&single_touch_spec.path).map_err(|e| {
426         error!("failed configuring virtio single touch: {:?}", e);
427         e
428     })?;
429 
430     let dev = virtio::new_single_touch(socket, single_touch_spec.width, single_touch_spec.height)
431         .map_err(Error::InputDeviceNew)?;
432     Ok(VirtioDeviceStub {
433         dev: Box::new(dev),
434         jail: simple_jail(&cfg, "input_device.policy")?,
435     })
436 }
437 
create_trackpad_device(cfg: &Config, trackpad_spec: &TouchDeviceOption) -> DeviceResult438 fn create_trackpad_device(cfg: &Config, trackpad_spec: &TouchDeviceOption) -> DeviceResult {
439     let socket = create_input_socket(&trackpad_spec.path).map_err(|e| {
440         error!("failed configuring virtio trackpad: {}", e);
441         e
442     })?;
443 
444     let dev = virtio::new_trackpad(socket, trackpad_spec.width, trackpad_spec.height)
445         .map_err(Error::InputDeviceNew)?;
446 
447     Ok(VirtioDeviceStub {
448         dev: Box::new(dev),
449         jail: simple_jail(&cfg, "input_device.policy")?,
450     })
451 }
452 
create_mouse_device(cfg: &Config, mouse_socket: &Path) -> DeviceResult453 fn create_mouse_device(cfg: &Config, mouse_socket: &Path) -> DeviceResult {
454     let socket = create_input_socket(&mouse_socket).map_err(|e| {
455         error!("failed configuring virtio mouse: {}", e);
456         e
457     })?;
458 
459     let dev = virtio::new_mouse(socket).map_err(Error::InputDeviceNew)?;
460 
461     Ok(VirtioDeviceStub {
462         dev: Box::new(dev),
463         jail: simple_jail(&cfg, "input_device.policy")?,
464     })
465 }
466 
create_keyboard_device(cfg: &Config, keyboard_socket: &Path) -> DeviceResult467 fn create_keyboard_device(cfg: &Config, keyboard_socket: &Path) -> DeviceResult {
468     let socket = create_input_socket(&keyboard_socket).map_err(|e| {
469         error!("failed configuring virtio keyboard: {}", e);
470         e
471     })?;
472 
473     let dev = virtio::new_keyboard(socket).map_err(Error::InputDeviceNew)?;
474 
475     Ok(VirtioDeviceStub {
476         dev: Box::new(dev),
477         jail: simple_jail(&cfg, "input_device.policy")?,
478     })
479 }
480 
create_vinput_device(cfg: &Config, dev_path: &Path) -> DeviceResult481 fn create_vinput_device(cfg: &Config, dev_path: &Path) -> DeviceResult {
482     let dev_file = OpenOptions::new()
483         .read(true)
484         .write(true)
485         .open(dev_path)
486         .map_err(|e| Error::OpenVinput(dev_path.to_owned(), e))?;
487 
488     let dev = virtio::new_evdev(dev_file).map_err(Error::InputDeviceNew)?;
489 
490     Ok(VirtioDeviceStub {
491         dev: Box::new(dev),
492         jail: simple_jail(&cfg, "input_device.policy")?,
493     })
494 }
495 
create_balloon_device(cfg: &Config, socket: BalloonControlResponseSocket) -> DeviceResult496 fn create_balloon_device(cfg: &Config, socket: BalloonControlResponseSocket) -> DeviceResult {
497     let dev = virtio::Balloon::new(socket).map_err(Error::BalloonDeviceNew)?;
498 
499     Ok(VirtioDeviceStub {
500         dev: Box::new(dev),
501         jail: simple_jail(&cfg, "balloon_device.policy")?,
502     })
503 }
504 
create_tap_net_device(cfg: &Config, tap_fd: RawFd) -> DeviceResult505 fn create_tap_net_device(cfg: &Config, tap_fd: RawFd) -> DeviceResult {
506     // Safe because we ensure that we get a unique handle to the fd.
507     let tap = unsafe {
508         Tap::from_raw_fd(validate_raw_fd(tap_fd).map_err(Error::ValidateRawFd)?)
509             .map_err(Error::CreateTapDevice)?
510     };
511 
512     let dev = virtio::Net::from(tap).map_err(Error::NetDeviceNew)?;
513 
514     Ok(VirtioDeviceStub {
515         dev: Box::new(dev),
516         jail: simple_jail(&cfg, "net_device.policy")?,
517     })
518 }
519 
create_net_device( cfg: &Config, host_ip: Ipv4Addr, netmask: Ipv4Addr, mac_address: MacAddress, mem: &GuestMemory, ) -> DeviceResult520 fn create_net_device(
521     cfg: &Config,
522     host_ip: Ipv4Addr,
523     netmask: Ipv4Addr,
524     mac_address: MacAddress,
525     mem: &GuestMemory,
526 ) -> DeviceResult {
527     let dev = if cfg.vhost_net {
528         let dev =
529             virtio::vhost::Net::<Tap, vhost::Net<Tap>>::new(host_ip, netmask, mac_address, mem)
530                 .map_err(Error::VhostNetDeviceNew)?;
531         Box::new(dev) as Box<dyn VirtioDevice>
532     } else {
533         let dev =
534             virtio::Net::<Tap>::new(host_ip, netmask, mac_address).map_err(Error::NetDeviceNew)?;
535         Box::new(dev) as Box<dyn VirtioDevice>
536     };
537 
538     let policy = if cfg.vhost_net {
539         "vhost_net_device.policy"
540     } else {
541         "net_device.policy"
542     };
543 
544     Ok(VirtioDeviceStub {
545         dev,
546         jail: simple_jail(&cfg, policy)?,
547     })
548 }
549 
550 #[cfg(feature = "gpu")]
create_gpu_device( cfg: &Config, exit_evt: &EventFd, gpu_device_socket: VmMemoryControlRequestSocket, gpu_socket: virtio::resource_bridge::ResourceResponseSocket, wayland_socket_path: &Path, ) -> DeviceResult551 fn create_gpu_device(
552     cfg: &Config,
553     exit_evt: &EventFd,
554     gpu_device_socket: VmMemoryControlRequestSocket,
555     gpu_socket: virtio::resource_bridge::ResourceResponseSocket,
556     wayland_socket_path: &Path,
557 ) -> DeviceResult {
558     let jailed_wayland_path = Path::new("/wayland-0");
559 
560     let dev = virtio::Gpu::new(
561         exit_evt.try_clone().map_err(Error::CloneEventFd)?,
562         Some(gpu_device_socket),
563         Some(gpu_socket),
564         if cfg.sandbox {
565             &jailed_wayland_path
566         } else {
567             wayland_socket_path
568         },
569     );
570 
571     let jail = match simple_jail(&cfg, "gpu_device.policy")? {
572         Some(mut jail) => {
573             // Create a tmpfs in the device's root directory so that we can bind mount the
574             // dri directory into it.  The size=67108864 is size=64*1024*1024 or size=64MB.
575             jail.mount_with_data(
576                 Path::new("none"),
577                 Path::new("/"),
578                 "tmpfs",
579                 (libc::MS_NOSUID | libc::MS_NODEV | libc::MS_NOEXEC) as usize,
580                 "size=67108864",
581             )?;
582 
583             // Device nodes required for DRM.
584             let sys_dev_char_path = Path::new("/sys/dev/char");
585             jail.mount_bind(sys_dev_char_path, sys_dev_char_path, false)?;
586             let sys_devices_path = Path::new("/sys/devices");
587             jail.mount_bind(sys_devices_path, sys_devices_path, false)?;
588             let drm_dri_path = Path::new("/dev/dri");
589             jail.mount_bind(drm_dri_path, drm_dri_path, false)?;
590 
591             // Libraries that are required when mesa drivers are dynamically loaded.
592             let lib_path = Path::new("/lib64");
593             jail.mount_bind(lib_path, lib_path, false)?;
594             let usr_lib_path = Path::new("/usr/lib64");
595             jail.mount_bind(usr_lib_path, usr_lib_path, false)?;
596 
597             // Bind mount the wayland socket into jail's root. This is necessary since each
598             // new wayland context must open() the socket.
599             jail.mount_bind(wayland_socket_path, jailed_wayland_path, true)?;
600 
601             add_crosvm_user_to_jail(&mut jail, "gpu")?;
602 
603             Some(jail)
604         }
605         None => None,
606     };
607 
608     Ok(VirtioDeviceStub {
609         dev: Box::new(dev),
610         jail,
611     })
612 }
613 
create_wayland_device( cfg: &Config, socket_path: &Path, socket: VmMemoryControlRequestSocket, resource_bridge: Option<virtio::resource_bridge::ResourceRequestSocket>, ) -> DeviceResult614 fn create_wayland_device(
615     cfg: &Config,
616     socket_path: &Path,
617     socket: VmMemoryControlRequestSocket,
618     resource_bridge: Option<virtio::resource_bridge::ResourceRequestSocket>,
619 ) -> DeviceResult {
620     let wayland_socket_dir = socket_path.parent().ok_or(Error::InvalidWaylandPath)?;
621     let wayland_socket_name = socket_path.file_name().ok_or(Error::InvalidWaylandPath)?;
622     let jailed_wayland_dir = Path::new("/wayland");
623     let jailed_wayland_path = jailed_wayland_dir.join(wayland_socket_name);
624 
625     let dev = virtio::Wl::new(
626         if cfg.sandbox {
627             &jailed_wayland_path
628         } else {
629             socket_path
630         },
631         socket,
632         resource_bridge,
633     )
634     .map_err(Error::WaylandDeviceNew)?;
635 
636     let jail = match simple_jail(&cfg, "wl_device.policy")? {
637         Some(mut jail) => {
638             // Create a tmpfs in the device's root directory so that we can bind mount the wayland
639             // socket directory into it. The size=67108864 is size=64*1024*1024 or size=64MB.
640             jail.mount_with_data(
641                 Path::new("none"),
642                 Path::new("/"),
643                 "tmpfs",
644                 (libc::MS_NOSUID | libc::MS_NODEV | libc::MS_NOEXEC) as usize,
645                 "size=67108864",
646             )?;
647 
648             // Bind mount the wayland socket's directory into jail's root. This is necessary since
649             // each new wayland context must open() the socket. If the wayland socket is ever
650             // destroyed and remade in the same host directory, new connections will be possible
651             // without restarting the wayland device.
652             jail.mount_bind(wayland_socket_dir, jailed_wayland_dir, true)?;
653 
654             add_crosvm_user_to_jail(&mut jail, "Wayland")?;
655 
656             Some(jail)
657         }
658         None => None,
659     };
660 
661     Ok(VirtioDeviceStub {
662         dev: Box::new(dev),
663         jail,
664     })
665 }
666 
create_vhost_vsock_device(cfg: &Config, cid: u64, mem: &GuestMemory) -> DeviceResult667 fn create_vhost_vsock_device(cfg: &Config, cid: u64, mem: &GuestMemory) -> DeviceResult {
668     let dev = virtio::vhost::Vsock::new(cid, mem).map_err(Error::VhostVsockDeviceNew)?;
669 
670     Ok(VirtioDeviceStub {
671         dev: Box::new(dev),
672         jail: simple_jail(&cfg, "vhost_vsock_device.policy")?,
673     })
674 }
675 
create_9p_device(cfg: &Config, chronos: Ids, src: &Path, tag: &str) -> DeviceResult676 fn create_9p_device(cfg: &Config, chronos: Ids, src: &Path, tag: &str) -> DeviceResult {
677     let (jail, root) = match simple_jail(&cfg, "9p_device.policy")? {
678         Some(mut jail) => {
679             //  The shared directory becomes the root of the device's file system.
680             let root = Path::new("/");
681             jail.mount_bind(src, root, true)?;
682 
683             // Set the uid/gid for the jailed process, and give a basic id map. This
684             // is required for the above bind mount to work.
685             jail.change_uid(chronos.uid);
686             jail.change_gid(chronos.gid);
687             jail.uidmap(&format!("{0} {0} 1", chronos.uid))
688                 .map_err(Error::SettingUidMap)?;
689             jail.gidmap(&format!("{0} {0} 1", chronos.gid))
690                 .map_err(Error::SettingGidMap)?;
691 
692             (Some(jail), root)
693         }
694         None => {
695             // There's no bind mount so we tell the server to treat the source directory as the
696             // root.
697             (None, src)
698         }
699     };
700 
701     let dev = virtio::P9::new(root, tag).map_err(Error::P9DeviceNew)?;
702 
703     Ok(VirtioDeviceStub {
704         dev: Box::new(dev),
705         jail,
706     })
707 }
708 
create_pmem_device( cfg: &Config, vm: &mut Vm, resources: &mut SystemAllocator, disk: &DiskOption, index: usize, ) -> DeviceResult709 fn create_pmem_device(
710     cfg: &Config,
711     vm: &mut Vm,
712     resources: &mut SystemAllocator,
713     disk: &DiskOption,
714     index: usize,
715 ) -> DeviceResult {
716     let fd = OpenOptions::new()
717         .read(true)
718         .write(!disk.read_only)
719         .open(&disk.path)
720         .map_err(Error::Disk)?;
721 
722     let image_size = {
723         let metadata = std::fs::metadata(&disk.path).map_err(Error::Disk)?;
724         metadata.len()
725     };
726 
727     let protection = {
728         if disk.read_only {
729             Protection::read()
730         } else {
731             Protection::read_write()
732         }
733     };
734 
735     let memory_mapping = {
736         // Conversion from u64 to usize may fail on 32bit system.
737         let image_size = usize::try_from(image_size).map_err(|_| Error::PmemDeviceImageTooBig)?;
738 
739         MemoryMapping::from_fd_offset_protection(&fd, image_size, 0, protection)
740             .map_err(Error::ReservePmemMemory)?
741     };
742 
743     let mapping_address = resources
744         .device_allocator()
745         .allocate_with_align(
746             image_size,
747             Alloc::PmemDevice(index),
748             format!("pmem_disk_image_{}", index),
749             // Linux kernel requires pmem namespaces to be 128 MiB aligned.
750             128 * 1024 * 1024, /* 128 MiB */
751         )
752         .map_err(Error::AllocatePmemDeviceAddress)?;
753 
754     vm.add_device_memory(
755         GuestAddress(mapping_address),
756         memory_mapping,
757         /* read_only = */ disk.read_only,
758         /* log_dirty_pages = */ false,
759     )
760     .map_err(Error::AddPmemDeviceMemory)?;
761 
762     let dev = virtio::Pmem::new(fd, GuestAddress(mapping_address), image_size)
763         .map_err(Error::PmemDeviceNew)?;
764 
765     Ok(VirtioDeviceStub {
766         dev: Box::new(dev) as Box<dyn VirtioDevice>,
767         /// TODO(jstaron) Create separate device policy for pmem_device.
768         jail: simple_jail(&cfg, "block_device.policy")?,
769     })
770 }
771 
772 // gpu_device_socket is not used when GPU support is disabled.
773 #[cfg_attr(not(feature = "gpu"), allow(unused_variables))]
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>>774 fn create_virtio_devices(
775     cfg: &Config,
776     mem: &GuestMemory,
777     vm: &mut Vm,
778     resources: &mut SystemAllocator,
779     _exit_evt: &EventFd,
780     wayland_device_socket: VmMemoryControlRequestSocket,
781     gpu_device_socket: VmMemoryControlRequestSocket,
782     balloon_device_socket: BalloonControlResponseSocket,
783     disk_device_sockets: &mut Vec<DiskControlResponseSocket>,
784 ) -> DeviceResult<Vec<VirtioDeviceStub>> {
785     let mut devs = Vec::new();
786 
787     for disk in &cfg.disks {
788         let disk_device_socket = disk_device_sockets.remove(0);
789         devs.push(create_block_device(cfg, disk, disk_device_socket)?);
790     }
791 
792     for (index, pmem_disk) in cfg.pmem_devices.iter().enumerate() {
793         devs.push(create_pmem_device(cfg, vm, resources, pmem_disk, index)?);
794     }
795 
796     devs.push(create_rng_device(cfg)?);
797 
798     #[cfg(feature = "tpm")]
799     {
800         if cfg.software_tpm {
801             devs.push(create_tpm_device(cfg)?);
802         }
803     }
804 
805     if let Some(single_touch_spec) = &cfg.virtio_single_touch {
806         devs.push(create_single_touch_device(cfg, single_touch_spec)?);
807     }
808 
809     if let Some(trackpad_spec) = &cfg.virtio_trackpad {
810         devs.push(create_trackpad_device(cfg, trackpad_spec)?);
811     }
812 
813     if let Some(mouse_socket) = &cfg.virtio_mouse {
814         devs.push(create_mouse_device(cfg, mouse_socket)?);
815     }
816 
817     if let Some(keyboard_socket) = &cfg.virtio_keyboard {
818         devs.push(create_keyboard_device(cfg, keyboard_socket)?);
819     }
820 
821     for dev_path in &cfg.virtio_input_evdevs {
822         devs.push(create_vinput_device(cfg, dev_path)?);
823     }
824 
825     devs.push(create_balloon_device(cfg, balloon_device_socket)?);
826 
827     // We checked above that if the IP is defined, then the netmask is, too.
828     for tap_fd in &cfg.tap_fd {
829         devs.push(create_tap_net_device(cfg, *tap_fd)?);
830     }
831 
832     if let (Some(host_ip), Some(netmask), Some(mac_address)) =
833         (cfg.host_ip, cfg.netmask, cfg.mac_address)
834     {
835         devs.push(create_net_device(cfg, host_ip, netmask, mac_address, mem)?);
836     }
837 
838     #[cfg_attr(not(feature = "gpu"), allow(unused_mut))]
839     let mut resource_bridge_wl_socket = None::<virtio::resource_bridge::ResourceRequestSocket>;
840 
841     #[cfg(feature = "gpu")]
842     {
843         if cfg.gpu {
844             if let Some(wayland_socket_path) = &cfg.wayland_socket_path {
845                 let (wl_socket, gpu_socket) =
846                     virtio::resource_bridge::pair().map_err(Error::CreateSocket)?;
847                 resource_bridge_wl_socket = Some(wl_socket);
848 
849                 devs.push(create_gpu_device(
850                     cfg,
851                     _exit_evt,
852                     gpu_device_socket,
853                     gpu_socket,
854                     wayland_socket_path,
855                 )?);
856             }
857         }
858     }
859 
860     if let Some(wayland_socket_path) = cfg.wayland_socket_path.as_ref() {
861         devs.push(create_wayland_device(
862             cfg,
863             wayland_socket_path,
864             wayland_device_socket,
865             resource_bridge_wl_socket,
866         )?);
867     }
868 
869     if let Some(cid) = cfg.cid {
870         devs.push(create_vhost_vsock_device(cfg, cid, mem)?);
871     }
872 
873     let chronos = get_chronos_ids();
874 
875     for (src, tag) in &cfg.shared_dirs {
876         devs.push(create_9p_device(cfg, chronos, src, tag)?);
877     }
878 
879     Ok(devs)
880 }
881 
create_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>, usb_provider: HostBackendDeviceProvider, ) -> DeviceResult<Vec<(Box<dyn PciDevice>, Option<Minijail>)>>882 fn create_devices(
883     cfg: &Config,
884     mem: &GuestMemory,
885     vm: &mut Vm,
886     resources: &mut SystemAllocator,
887     exit_evt: &EventFd,
888     wayland_device_socket: VmMemoryControlRequestSocket,
889     gpu_device_socket: VmMemoryControlRequestSocket,
890     balloon_device_socket: BalloonControlResponseSocket,
891     disk_device_sockets: &mut Vec<DiskControlResponseSocket>,
892     usb_provider: HostBackendDeviceProvider,
893 ) -> DeviceResult<Vec<(Box<dyn PciDevice>, Option<Minijail>)>> {
894     let stubs = create_virtio_devices(
895         &cfg,
896         mem,
897         vm,
898         resources,
899         exit_evt,
900         wayland_device_socket,
901         gpu_device_socket,
902         balloon_device_socket,
903         disk_device_sockets,
904     )?;
905 
906     let mut pci_devices = Vec::new();
907 
908     for stub in stubs {
909         let dev = VirtioPciDevice::new(mem.clone(), stub.dev).map_err(Error::VirtioPciDev)?;
910         let dev = Box::new(dev) as Box<dyn PciDevice>;
911         pci_devices.push((dev, stub.jail));
912     }
913 
914     if cfg.cras_audio {
915         let mut server = Box::new(CrasClient::new().map_err(Error::CreateCrasClient)?);
916         if cfg.cras_capture {
917             server.enable_cras_capture();
918         }
919         let cras_audio = devices::Ac97Dev::new(mem.clone(), server);
920 
921         pci_devices.push((
922             Box::new(cras_audio),
923             simple_jail(&cfg, "cras_audio_device.policy")?,
924         ));
925     }
926 
927     if cfg.null_audio {
928         let server = Box::new(DummyStreamSource::new());
929         let null_audio = devices::Ac97Dev::new(mem.clone(), server);
930 
931         pci_devices.push((
932             Box::new(null_audio),
933             simple_jail(&cfg, "null_audio_device.policy")?,
934         ));
935     }
936     // Create xhci controller.
937     let usb_controller = Box::new(XhciController::new(mem.clone(), usb_provider));
938     pci_devices.push((usb_controller, simple_jail(&cfg, "xhci.policy")?));
939 
940     Ok(pci_devices)
941 }
942 
943 #[derive(Copy, Clone)]
944 struct Ids {
945     uid: uid_t,
946     gid: gid_t,
947 }
948 
get_chronos_ids() -> Ids949 fn get_chronos_ids() -> Ids {
950     let chronos_user_group = CStr::from_bytes_with_nul(b"chronos\0").unwrap();
951 
952     let chronos_uid = match get_user_id(&chronos_user_group) {
953         Ok(u) => u,
954         Err(e) => {
955             warn!("falling back to current user id for 9p: {}", e);
956             geteuid()
957         }
958     };
959 
960     let chronos_gid = match get_group_id(&chronos_user_group) {
961         Ok(u) => u,
962         Err(e) => {
963             warn!("falling back to current group id for 9p: {}", e);
964             getegid()
965         }
966     };
967 
968     Ids {
969         uid: chronos_uid,
970         gid: chronos_gid,
971     }
972 }
973 
974 // Set the uid/gid for the jailed process and give a basic id map. This is
975 // required for bind mounts to work.
add_crosvm_user_to_jail(jail: &mut Minijail, feature: &str) -> Result<Ids>976 fn add_crosvm_user_to_jail(jail: &mut Minijail, feature: &str) -> Result<Ids> {
977     let crosvm_user_group = CStr::from_bytes_with_nul(b"crosvm\0").unwrap();
978 
979     let crosvm_uid = match get_user_id(&crosvm_user_group) {
980         Ok(u) => u,
981         Err(e) => {
982             warn!("falling back to current user id for {}: {}", feature, e);
983             geteuid()
984         }
985     };
986 
987     let crosvm_gid = match get_group_id(&crosvm_user_group) {
988         Ok(u) => u,
989         Err(e) => {
990             warn!("falling back to current group id for {}: {}", feature, e);
991             getegid()
992         }
993     };
994 
995     jail.change_uid(crosvm_uid);
996     jail.change_gid(crosvm_gid);
997     jail.uidmap(&format!("{0} {0} 1", crosvm_uid))
998         .map_err(Error::SettingUidMap)?;
999     jail.gidmap(&format!("{0} {0} 1", crosvm_gid))
1000         .map_err(Error::SettingGidMap)?;
1001 
1002     Ok(Ids {
1003         uid: crosvm_uid,
1004         gid: crosvm_gid,
1005     })
1006 }
1007 
raw_fd_from_path(path: &Path) -> Result<RawFd>1008 fn raw_fd_from_path(path: &Path) -> Result<RawFd> {
1009     if !path.is_file() {
1010         return Err(Error::InvalidFdPath);
1011     }
1012     let raw_fd = path
1013         .file_name()
1014         .and_then(|fd_osstr| fd_osstr.to_str())
1015         .and_then(|fd_str| fd_str.parse::<c_int>().ok())
1016         .ok_or(Error::InvalidFdPath)?;
1017     validate_raw_fd(raw_fd).map_err(Error::ValidateRawFd)
1018 }
1019 
create_input_socket(path: &Path) -> Result<UnixStream>1020 fn create_input_socket(path: &Path) -> Result<UnixStream> {
1021     if path.parent() == Some(Path::new("/proc/self/fd")) {
1022         // Safe because we will validate |raw_fd|.
1023         unsafe { Ok(UnixStream::from_raw_fd(raw_fd_from_path(path)?)) }
1024     } else {
1025         UnixStream::connect(path).map_err(Error::InputEventsOpen)
1026     }
1027 }
1028 
setup_vcpu_signal_handler() -> Result<()>1029 fn setup_vcpu_signal_handler() -> Result<()> {
1030     unsafe {
1031         extern "C" fn handle_signal() {}
1032         // Our signal handler does nothing and is trivially async signal safe.
1033         register_signal_handler(SIGRTMIN() + 0, handle_signal)
1034             .map_err(Error::RegisterSignalHandler)?;
1035     }
1036     block_signal(SIGRTMIN() + 0).map_err(Error::BlockSignal)?;
1037     Ok(())
1038 }
1039 
1040 #[derive(Default)]
1041 struct VcpuRunMode {
1042     mtx: Mutex<VmRunMode>,
1043     cvar: Condvar,
1044 }
1045 
1046 impl VcpuRunMode {
set_and_notify(&self, new_mode: VmRunMode)1047     fn set_and_notify(&self, new_mode: VmRunMode) {
1048         *self.mtx.lock() = new_mode;
1049         self.cvar.notify_all();
1050     }
1051 }
1052 
run_vcpu( vcpu: Vcpu, cpu_id: u32, vcpu_affinity: Vec<usize>, start_barrier: Arc<Barrier>, io_bus: devices::Bus, mmio_bus: devices::Bus, exit_evt: EventFd, requires_kvmclock_ctrl: bool, run_mode_arc: Arc<VcpuRunMode>, ) -> Result<JoinHandle<()>>1053 fn run_vcpu(
1054     vcpu: Vcpu,
1055     cpu_id: u32,
1056     vcpu_affinity: Vec<usize>,
1057     start_barrier: Arc<Barrier>,
1058     io_bus: devices::Bus,
1059     mmio_bus: devices::Bus,
1060     exit_evt: EventFd,
1061     requires_kvmclock_ctrl: bool,
1062     run_mode_arc: Arc<VcpuRunMode>,
1063 ) -> Result<JoinHandle<()>> {
1064     thread::Builder::new()
1065         .name(format!("crosvm_vcpu{}", cpu_id))
1066         .spawn(move || {
1067             if vcpu_affinity.len() != 0 {
1068                 if let Err(e) = set_cpu_affinity(vcpu_affinity) {
1069                     error!("Failed to set CPU affinity: {}", e);
1070                 }
1071             }
1072 
1073             let mut sig_ok = true;
1074             match get_blocked_signals() {
1075                 Ok(mut v) => {
1076                     v.retain(|&x| x != SIGRTMIN() + 0);
1077                     if let Err(e) = vcpu.set_signal_mask(&v) {
1078                         error!(
1079                             "Failed to set the KVM_SIGNAL_MASK for vcpu {} : {}",
1080                             cpu_id, e
1081                         );
1082                         sig_ok = false;
1083                     }
1084                 }
1085                 Err(e) => {
1086                     error!(
1087                         "Failed to retrieve signal mask for vcpu {} : {}",
1088                         cpu_id, e
1089                     );
1090                     sig_ok = false;
1091                 }
1092             };
1093 
1094             start_barrier.wait();
1095 
1096             if sig_ok {
1097                 'vcpu_loop: loop {
1098                     let mut interrupted_by_signal = false;
1099                     match vcpu.run() {
1100                         Ok(VcpuExit::IoIn { port, mut size }) => {
1101                             let mut data = [0; 8];
1102                             if size > data.len() {
1103                                 error!("unsupported IoIn size of {} bytes", size);
1104                                 size = data.len();
1105                             }
1106                             io_bus.read(port as u64, &mut data[..size]);
1107                             if let Err(e) = vcpu.set_data(&data[..size]) {
1108                                 error!("failed to set return data for IoIn: {}", e);
1109                             }
1110                         }
1111                         Ok(VcpuExit::IoOut {
1112                             port,
1113                             mut size,
1114                             data,
1115                         }) => {
1116                             if size > data.len() {
1117                                 error!("unsupported IoOut size of {} bytes", size);
1118                                 size = data.len();
1119                             }
1120                             io_bus.write(port as u64, &data[..size]);
1121                         }
1122                         Ok(VcpuExit::MmioRead { address, size }) => {
1123                             let mut data = [0; 8];
1124                             mmio_bus.read(address, &mut data[..size]);
1125                             // Setting data for mmio can not fail.
1126                             let _ = vcpu.set_data(&data[..size]);
1127                         }
1128                         Ok(VcpuExit::MmioWrite {
1129                             address,
1130                             size,
1131                             data,
1132                         }) => {
1133                             mmio_bus.write(address, &data[..size]);
1134                         }
1135                         Ok(VcpuExit::Hlt) => break,
1136                         Ok(VcpuExit::Shutdown) => break,
1137                         Ok(VcpuExit::SystemEvent(_, _)) => break,
1138                         Ok(r) => warn!("unexpected vcpu exit: {:?}", r),
1139                         Err(e) => match e.errno() {
1140                             libc::EINTR => interrupted_by_signal = true,
1141                             libc::EAGAIN => {}
1142                             _ => {
1143                                 error!("vcpu hit unknown error: {}", e);
1144                                 break;
1145                             }
1146                         },
1147                     }
1148 
1149                     if interrupted_by_signal {
1150                         // Try to clear the signal that we use to kick VCPU if it is pending before
1151                         // attempting to handle pause requests.
1152                         if let Err(e) = clear_signal(SIGRTMIN() + 0) {
1153                             error!("failed to clear pending signal: {}", e);
1154                             break;
1155                         }
1156                         let mut run_mode_lock = run_mode_arc.mtx.lock();
1157                         loop {
1158                             match *run_mode_lock {
1159                                 VmRunMode::Running => break,
1160                                 VmRunMode::Suspending => {
1161                                     // On KVM implementations that use a paravirtualized clock (e.g.
1162                                     // x86), a flag must be set to indicate to the guest kernel that
1163                                     // a VCPU was suspended. The guest kernel will use this flag to
1164                                     // prevent the soft lockup detection from triggering when this
1165                                     // VCPU resumes, which could happen days later in realtime.
1166                                     if requires_kvmclock_ctrl {
1167                                         if let Err(e) = vcpu.kvmclock_ctrl() {
1168                                             error!("failed to signal to kvm that vcpu {} is being suspended: {}", cpu_id, e);
1169                                         }
1170                                     }
1171                                 }
1172                                 VmRunMode::Exiting => break 'vcpu_loop,
1173                             }
1174                             // Give ownership of our exclusive lock to the condition variable that
1175                             // will block. When the condition variable is notified, `wait` will
1176                             // unblock and return a new exclusive lock.
1177                             run_mode_lock = run_mode_arc.cvar.wait(run_mode_lock);
1178                         }
1179                     }
1180                 }
1181             }
1182             exit_evt
1183                 .write(1)
1184                 .expect("failed to signal vcpu exit eventfd");
1185         })
1186         .map_err(Error::SpawnVcpu)
1187 }
1188 
1189 // Reads the contents of a file and converts the space-separated fields into a Vec of u64s.
1190 // Returns an error if any of the fields fail to parse.
file_fields_to_u64<P: AsRef<Path>>(path: P) -> io::Result<Vec<u64>>1191 fn file_fields_to_u64<P: AsRef<Path>>(path: P) -> io::Result<Vec<u64>> {
1192     let mut file = File::open(path)?;
1193 
1194     let mut buf = [0u8; 32];
1195     let count = file.read(&mut buf)?;
1196 
1197     let content =
1198         str::from_utf8(&buf[..count]).map_err(|e| io::Error::new(io::ErrorKind::InvalidData, e))?;
1199     content
1200         .trim()
1201         .split_whitespace()
1202         .map(|x| {
1203             x.parse::<u64>()
1204                 .map_err(|e| io::Error::new(io::ErrorKind::InvalidData, e))
1205         })
1206         .collect()
1207 }
1208 
1209 // Reads the contents of a file and converts them into a u64, and if there
1210 // are multiple fields it only returns the first one.
file_to_u64<P: AsRef<Path>>(path: P) -> io::Result<u64>1211 fn file_to_u64<P: AsRef<Path>>(path: P) -> io::Result<u64> {
1212     file_fields_to_u64(path)?
1213         .into_iter()
1214         .next()
1215         .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidData, "empty file"))
1216 }
1217 
run_config(cfg: Config) -> Result<()>1218 pub fn run_config(cfg: Config) -> Result<()> {
1219     if cfg.sandbox {
1220         // Printing something to the syslog before entering minijail so that libc's syslogger has a
1221         // chance to open files necessary for its operation, like `/etc/localtime`. After jailing,
1222         // access to those files will not be possible.
1223         info!("crosvm entering multiprocess mode");
1224     }
1225 
1226     let (usb_control_socket, usb_provider) =
1227         HostBackendDeviceProvider::new().map_err(Error::CreateUsbProvider)?;
1228     // Masking signals is inherently dangerous, since this can persist across clones/execs. Do this
1229     // before any jailed devices have been spawned, so that we can catch any of them that fail very
1230     // quickly.
1231     let sigchld_fd = SignalFd::new(libc::SIGCHLD).map_err(Error::CreateSignalFd)?;
1232 
1233     let initrd_image = if let Some(initrd_path) = &cfg.initrd_path {
1234         Some(File::open(initrd_path).map_err(|e| Error::OpenInitrd(initrd_path.clone(), e))?)
1235     } else {
1236         None
1237     };
1238 
1239     let vm_image = match cfg.executable_path {
1240         Some(Executable::Kernel(ref kernel_path)) => VmImage::Kernel(
1241             File::open(kernel_path).map_err(|e| Error::OpenKernel(kernel_path.to_path_buf(), e))?,
1242         ),
1243         Some(Executable::Bios(ref bios_path)) => VmImage::Bios(
1244             File::open(bios_path).map_err(|e| Error::OpenBios(bios_path.to_path_buf(), e))?,
1245         ),
1246         _ => panic!("Did not receive a bios or kernel, should be impossible."),
1247     };
1248 
1249     let components = VmComponents {
1250         memory_size: (cfg.memory.unwrap_or(256) << 20) as u64,
1251         vcpu_count: cfg.vcpu_count.unwrap_or(1),
1252         vcpu_affinity: cfg.vcpu_affinity.clone(),
1253         vm_image,
1254         android_fstab: cfg
1255             .android_fstab
1256             .as_ref()
1257             .map(|x| File::open(x).map_err(|e| Error::OpenAndroidFstab(x.to_path_buf(), e)))
1258             .map_or(Ok(None), |v| v.map(Some))?,
1259         initrd_image,
1260         extra_kernel_params: cfg.params.clone(),
1261         wayland_dmabuf: cfg.wayland_dmabuf,
1262     };
1263 
1264     let control_server_socket = match &cfg.socket_path {
1265         Some(path) => Some(UnlinkUnixSeqpacketListener(
1266             UnixSeqpacketListener::bind(path).map_err(Error::CreateSocket)?,
1267         )),
1268         None => None,
1269     };
1270 
1271     let mut control_sockets = Vec::new();
1272     let (wayland_host_socket, wayland_device_socket) =
1273         msg_socket::pair::<VmMemoryResponse, VmMemoryRequest>().map_err(Error::CreateSocket)?;
1274     control_sockets.push(TaggedControlSocket::VmMemory(wayland_host_socket));
1275     // Balloon gets a special socket so balloon requests can be forwarded from the main process.
1276     let (balloon_host_socket, balloon_device_socket) =
1277         msg_socket::pair::<BalloonControlCommand, ()>().map_err(Error::CreateSocket)?;
1278 
1279     // Create one control socket per disk.
1280     let mut disk_device_sockets = Vec::new();
1281     let mut disk_host_sockets = Vec::new();
1282     let disk_count = cfg.disks.len();
1283     for _ in 0..disk_count {
1284         let (disk_host_socket, disk_device_socket) =
1285             msg_socket::pair::<DiskControlCommand, DiskControlResult>()
1286                 .map_err(Error::CreateSocket)?;
1287         disk_host_sockets.push(disk_host_socket);
1288         disk_device_sockets.push(disk_device_socket);
1289     }
1290 
1291     let (gpu_host_socket, gpu_device_socket) =
1292         msg_socket::pair::<VmMemoryResponse, VmMemoryRequest>().map_err(Error::CreateSocket)?;
1293     control_sockets.push(TaggedControlSocket::VmMemory(gpu_host_socket));
1294 
1295     let sandbox = cfg.sandbox;
1296     let linux = Arch::build_vm(
1297         components,
1298         cfg.split_irqchip,
1299         &cfg.serial_parameters,
1300         |mem, vm, sys_allocator, exit_evt| {
1301             create_devices(
1302                 &cfg,
1303                 mem,
1304                 vm,
1305                 sys_allocator,
1306                 exit_evt,
1307                 wayland_device_socket,
1308                 gpu_device_socket,
1309                 balloon_device_socket,
1310                 &mut disk_device_sockets,
1311                 usb_provider,
1312             )
1313         },
1314     )
1315     .map_err(Error::BuildVm)?;
1316 
1317     let _render_node_host = ();
1318     #[cfg(feature = "gpu-forward")]
1319     let (_render_node_host, linux) = {
1320         // Rebinds linux as mutable.
1321         let mut linux = linux;
1322 
1323         // Reserve memory range for GPU buffer allocation in advance to bypass region count
1324         // limitation. We use mremap/MAP_FIXED later to make sure GPU buffers fall into this range.
1325         let gpu_mmap =
1326             MemoryMapping::new_protection(RENDER_NODE_HOST_SIZE as usize, Protection::none())
1327                 .map_err(Error::ReserveGpuMemory)?;
1328 
1329         // Put the non-accessible memory map into device memory so that no other devices use that
1330         // guest address space.
1331         let gpu_addr = linux
1332             .resources
1333             .device_allocator()
1334             .allocate(
1335                 RENDER_NODE_HOST_SIZE,
1336                 Alloc::GpuRenderNode,
1337                 "gpu_render_node".to_string(),
1338             )
1339             .map_err(|_| Error::AllocateGpuDeviceAddress)?;
1340 
1341         let host = RenderNodeHost::start(&gpu_mmap, gpu_addr, linux.vm.get_memory().clone());
1342 
1343         // Makes the gpu memory accessible at allocated address.
1344         linux
1345             .vm
1346             .add_device_memory(
1347                 GuestAddress(gpu_addr),
1348                 gpu_mmap,
1349                 /* read_only = */ false,
1350                 /* log_dirty_pages = */ false,
1351             )
1352             .map_err(Error::AddGpuDeviceMemory)?;
1353         (host, linux)
1354     };
1355 
1356     run_control(
1357         linux,
1358         control_server_socket,
1359         control_sockets,
1360         balloon_host_socket,
1361         &disk_host_sockets,
1362         usb_control_socket,
1363         sigchld_fd,
1364         _render_node_host,
1365         sandbox,
1366     )
1367 }
1368 
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, _render_node_host: RenderNodeHost, sandbox: bool, ) -> Result<()>1369 fn run_control(
1370     mut linux: RunnableLinuxVm,
1371     control_server_socket: Option<UnlinkUnixSeqpacketListener>,
1372     mut control_sockets: Vec<TaggedControlSocket>,
1373     balloon_host_socket: BalloonControlRequestSocket,
1374     disk_host_sockets: &[DiskControlRequestSocket],
1375     usb_control_socket: UsbControlSocket,
1376     sigchld_fd: SignalFd,
1377     _render_node_host: RenderNodeHost,
1378     sandbox: bool,
1379 ) -> Result<()> {
1380     // Paths to get the currently available memory and the low memory threshold.
1381     const LOWMEM_MARGIN: &str = "/sys/kernel/mm/chromeos-low_mem/margin";
1382     const LOWMEM_AVAILABLE: &str = "/sys/kernel/mm/chromeos-low_mem/available";
1383 
1384     // The amount of additional memory to claim back from the VM whenever the system is
1385     // low on memory.
1386     const ONE_GB: u64 = (1 << 30);
1387 
1388     let max_balloon_memory = match linux.vm.get_memory().memory_size() {
1389         // If the VM has at least 1.5 GB, the balloon driver can consume all but the last 1 GB.
1390         n if n >= (ONE_GB / 2) * 3 => n - ONE_GB,
1391         // Otherwise, if the VM has at least 500MB the balloon driver will consume at most
1392         // half of it.
1393         n if n >= (ONE_GB / 2) => n / 2,
1394         // Otherwise, the VM is too small for us to take memory away from it.
1395         _ => 0,
1396     };
1397     let mut current_balloon_memory: u64 = 0;
1398     let balloon_memory_increment: u64 = max_balloon_memory / 16;
1399 
1400     #[derive(PollToken)]
1401     enum Token {
1402         Exit,
1403         Stdin,
1404         ChildSignal,
1405         CheckAvailableMemory,
1406         LowMemory,
1407         LowmemTimer,
1408         VmControlServer,
1409         VmControl { index: usize },
1410     }
1411 
1412     let stdin_handle = stdin();
1413     let stdin_lock = stdin_handle.lock();
1414     stdin_lock
1415         .set_raw_mode()
1416         .expect("failed to set terminal raw mode");
1417 
1418     let poll_ctx = PollContext::new().map_err(Error::CreatePollContext)?;
1419     poll_ctx
1420         .add(&linux.exit_evt, Token::Exit)
1421         .map_err(Error::PollContextAdd)?;
1422     if let Err(e) = poll_ctx.add(&stdin_handle, Token::Stdin) {
1423         warn!("failed to add stdin to poll context: {}", e);
1424     }
1425     poll_ctx
1426         .add(&sigchld_fd, Token::ChildSignal)
1427         .map_err(Error::PollContextAdd)?;
1428 
1429     if let Some(socket_server) = &control_server_socket {
1430         poll_ctx
1431             .add(socket_server, Token::VmControlServer)
1432             .map_err(Error::PollContextAdd)?;
1433     }
1434     for (index, socket) in control_sockets.iter().enumerate() {
1435         poll_ctx
1436             .add(socket.as_ref(), Token::VmControl { index })
1437             .map_err(Error::PollContextAdd)?;
1438     }
1439 
1440     // Watch for low memory notifications and take memory back from the VM.
1441     let low_mem = File::open("/dev/chromeos-low-mem").ok();
1442     if let Some(low_mem) = &low_mem {
1443         poll_ctx
1444             .add(low_mem, Token::LowMemory)
1445             .map_err(Error::PollContextAdd)?;
1446     } else {
1447         warn!("Unable to open low mem indicator, maybe not a chrome os kernel");
1448     }
1449 
1450     // Used to rate limit balloon requests.
1451     let mut lowmem_timer = TimerFd::new().map_err(Error::CreateTimerFd)?;
1452     poll_ctx
1453         .add(&lowmem_timer, Token::LowmemTimer)
1454         .map_err(Error::PollContextAdd)?;
1455 
1456     // Used to check whether it's ok to start giving memory back to the VM.
1457     let mut freemem_timer = TimerFd::new().map_err(Error::CreateTimerFd)?;
1458     poll_ctx
1459         .add(&freemem_timer, Token::CheckAvailableMemory)
1460         .map_err(Error::PollContextAdd)?;
1461 
1462     // Used to add jitter to timer values so that we don't have a thundering herd problem when
1463     // multiple VMs are running.
1464     let mut simple_rng = SimpleRng::new(
1465         SystemTime::now()
1466             .duration_since(UNIX_EPOCH)
1467             .expect("time went backwards")
1468             .subsec_nanos() as u64,
1469     );
1470 
1471     if sandbox {
1472         // Before starting VCPUs, in case we started with some capabilities, drop them all.
1473         drop_capabilities().map_err(Error::DropCapabilities)?;
1474     }
1475 
1476     let mut vcpu_handles = Vec::with_capacity(linux.vcpus.len());
1477     let vcpu_thread_barrier = Arc::new(Barrier::new(linux.vcpus.len() + 1));
1478     let run_mode_arc = Arc::new(VcpuRunMode::default());
1479     setup_vcpu_signal_handler()?;
1480     for (cpu_id, vcpu) in linux.vcpus.into_iter().enumerate() {
1481         let handle = run_vcpu(
1482             vcpu,
1483             cpu_id as u32,
1484             linux.vcpu_affinity.clone(),
1485             vcpu_thread_barrier.clone(),
1486             linux.io_bus.clone(),
1487             linux.mmio_bus.clone(),
1488             linux.exit_evt.try_clone().map_err(Error::CloneEventFd)?,
1489             linux.vm.check_extension(Cap::KvmclockCtrl),
1490             run_mode_arc.clone(),
1491         )?;
1492         vcpu_handles.push(handle);
1493     }
1494     vcpu_thread_barrier.wait();
1495 
1496     'poll: loop {
1497         let events = {
1498             match poll_ctx.wait() {
1499                 Ok(v) => v,
1500                 Err(e) => {
1501                     error!("failed to poll: {}", e);
1502                     break;
1503                 }
1504             }
1505         };
1506 
1507         let mut vm_control_indices_to_remove = Vec::new();
1508         for event in events.iter_readable() {
1509             match event.token() {
1510                 Token::Exit => {
1511                     info!("vcpu requested shutdown");
1512                     break 'poll;
1513                 }
1514                 Token::Stdin => {
1515                     let mut out = [0u8; 64];
1516                     match stdin_lock.read_raw(&mut out[..]) {
1517                         Ok(0) => {
1518                             // Zero-length read indicates EOF. Remove from pollables.
1519                             let _ = poll_ctx.delete(&stdin_handle);
1520                         }
1521                         Err(e) => {
1522                             warn!("error while reading stdin: {}", e);
1523                             let _ = poll_ctx.delete(&stdin_handle);
1524                         }
1525                         Ok(count) => {
1526                             if let Some(ref stdio_serial) = linux.stdio_serial {
1527                                 stdio_serial
1528                                     .lock()
1529                                     .queue_input_bytes(&out[..count])
1530                                     .expect("failed to queue bytes into serial port");
1531                             }
1532                         }
1533                     }
1534                 }
1535                 Token::ChildSignal => {
1536                     // Print all available siginfo structs, then exit the loop.
1537                     while let Some(siginfo) = sigchld_fd.read().map_err(Error::SignalFd)? {
1538                         let pid = siginfo.ssi_pid;
1539                         let pid_label = match linux.pid_debug_label_map.get(&pid) {
1540                             Some(label) => format!("{} (pid {})", label, pid),
1541                             None => format!("pid {}", pid),
1542                         };
1543                         error!(
1544                             "child {} died: signo {}, status {}, code {}",
1545                             pid_label, siginfo.ssi_signo, siginfo.ssi_status, siginfo.ssi_code
1546                         );
1547                     }
1548                     break 'poll;
1549                 }
1550                 Token::CheckAvailableMemory => {
1551                     // Acknowledge the timer.
1552                     freemem_timer.wait().map_err(Error::TimerFd)?;
1553                     if current_balloon_memory == 0 {
1554                         // Nothing to see here.
1555                         if let Err(e) = freemem_timer.clear() {
1556                             warn!("unable to clear available memory check timer: {}", e);
1557                         }
1558                         continue;
1559                     }
1560 
1561                     // Otherwise see if we can free up some memory.
1562                     let margin = file_to_u64(LOWMEM_MARGIN).map_err(Error::ReadLowmemMargin)?;
1563                     let available =
1564                         file_to_u64(LOWMEM_AVAILABLE).map_err(Error::ReadLowmemAvailable)?;
1565 
1566                     // `available` and `margin` are specified in MB while `balloon_memory_increment` is in
1567                     // bytes.  So to correctly compare them we need to turn the increment value into MB.
1568                     if available >= margin + 2 * (balloon_memory_increment >> 20) {
1569                         current_balloon_memory =
1570                             if current_balloon_memory >= balloon_memory_increment {
1571                                 current_balloon_memory - balloon_memory_increment
1572                             } else {
1573                                 0
1574                             };
1575                         let command = BalloonControlCommand::Adjust {
1576                             num_bytes: current_balloon_memory,
1577                         };
1578                         if let Err(e) = balloon_host_socket.send(&command) {
1579                             warn!("failed to send memory value to balloon device: {}", e);
1580                         }
1581                     }
1582                 }
1583                 Token::LowMemory => {
1584                     if let Some(low_mem) = &low_mem {
1585                         let old_balloon_memory = current_balloon_memory;
1586                         current_balloon_memory = min(
1587                             current_balloon_memory + balloon_memory_increment,
1588                             max_balloon_memory,
1589                         );
1590                         if current_balloon_memory != old_balloon_memory {
1591                             let command = BalloonControlCommand::Adjust {
1592                                 num_bytes: current_balloon_memory,
1593                             };
1594                             if let Err(e) = balloon_host_socket.send(&command) {
1595                                 warn!("failed to send memory value to balloon device: {}", e);
1596                             }
1597                         }
1598 
1599                         // Stop polling the lowmem device until the timer fires.
1600                         poll_ctx.delete(low_mem).map_err(Error::PollContextDelete)?;
1601 
1602                         // Add some jitter to the timer so that if there are multiple VMs running
1603                         // they don't all start ballooning at exactly the same time.
1604                         let lowmem_dur = Duration::from_millis(1000 + simple_rng.rng() % 200);
1605                         lowmem_timer
1606                             .reset(lowmem_dur, None)
1607                             .map_err(Error::ResetTimerFd)?;
1608 
1609                         // Also start a timer to check when we can start giving memory back.  Do the
1610                         // first check after a minute (with jitter) and subsequent checks after
1611                         // every 30 seconds (with jitter).
1612                         let freemem_dur = Duration::from_secs(60 + simple_rng.rng() % 12);
1613                         let freemem_int = Duration::from_secs(30 + simple_rng.rng() % 6);
1614                         freemem_timer
1615                             .reset(freemem_dur, Some(freemem_int))
1616                             .map_err(Error::ResetTimerFd)?;
1617                     }
1618                 }
1619                 Token::LowmemTimer => {
1620                     // Acknowledge the timer.
1621                     lowmem_timer.wait().map_err(Error::TimerFd)?;
1622 
1623                     if let Some(low_mem) = &low_mem {
1624                         // Start polling the lowmem device again.
1625                         poll_ctx
1626                             .add(low_mem, Token::LowMemory)
1627                             .map_err(Error::PollContextAdd)?;
1628                     }
1629                 }
1630                 Token::VmControlServer => {
1631                     if let Some(socket_server) = &control_server_socket {
1632                         match socket_server.accept() {
1633                             Ok(socket) => {
1634                                 poll_ctx
1635                                     .add(
1636                                         &socket,
1637                                         Token::VmControl {
1638                                             index: control_sockets.len(),
1639                                         },
1640                                     )
1641                                     .map_err(Error::PollContextAdd)?;
1642                                 control_sockets
1643                                     .push(TaggedControlSocket::Vm(MsgSocket::new(socket)));
1644                             }
1645                             Err(e) => error!("failed to accept socket: {}", e),
1646                         }
1647                     }
1648                 }
1649                 Token::VmControl { index } => {
1650                     if let Some(socket) = control_sockets.get(index) {
1651                         match socket {
1652                             TaggedControlSocket::Vm(socket) => match socket.recv() {
1653                                 Ok(request) => {
1654                                     let mut run_mode_opt = None;
1655                                     let response = request.execute(
1656                                         &mut run_mode_opt,
1657                                         &balloon_host_socket,
1658                                         disk_host_sockets,
1659                                         &usb_control_socket,
1660                                     );
1661                                     if let Err(e) = socket.send(&response) {
1662                                         error!("failed to send VmResponse: {}", e);
1663                                     }
1664                                     if let Some(run_mode) = run_mode_opt {
1665                                         info!("control socket changed run mode to {}", run_mode);
1666                                         match run_mode {
1667                                             VmRunMode::Exiting => {
1668                                                 break 'poll;
1669                                             }
1670                                             other => {
1671                                                 run_mode_arc.set_and_notify(other);
1672                                                 for handle in &vcpu_handles {
1673                                                     let _ = handle.kill(SIGRTMIN() + 0);
1674                                                 }
1675                                             }
1676                                         }
1677                                     }
1678                                 }
1679                                 Err(e) => {
1680                                     if let MsgError::BadRecvSize { actual: 0, .. } = e {
1681                                         vm_control_indices_to_remove.push(index);
1682                                     } else {
1683                                         error!("failed to recv VmRequest: {}", e);
1684                                     }
1685                                 }
1686                             },
1687                             TaggedControlSocket::VmMemory(socket) => match socket.recv() {
1688                                 Ok(request) => {
1689                                     let response =
1690                                         request.execute(&mut linux.vm, &mut linux.resources);
1691                                     if let Err(e) = socket.send(&response) {
1692                                         error!("failed to send VmMemoryControlResponse: {}", e);
1693                                     }
1694                                 }
1695                                 Err(e) => {
1696                                     if let MsgError::BadRecvSize { actual: 0, .. } = e {
1697                                         vm_control_indices_to_remove.push(index);
1698                                     } else {
1699                                         error!("failed to recv VmMemoryControlRequest: {}", e);
1700                                     }
1701                                 }
1702                             },
1703                         }
1704                     }
1705                 }
1706             }
1707         }
1708 
1709         for event in events.iter_hungup() {
1710             match event.token() {
1711                 Token::Exit => {}
1712                 Token::Stdin => {
1713                     let _ = poll_ctx.delete(&stdin_handle);
1714                 }
1715                 Token::ChildSignal => {}
1716                 Token::CheckAvailableMemory => {}
1717                 Token::LowMemory => {}
1718                 Token::LowmemTimer => {}
1719                 Token::VmControlServer => {}
1720                 Token::VmControl { index } => {
1721                     // It's possible more data is readable and buffered while the socket is hungup,
1722                     // so don't delete the socket from the poll context until we're sure all the
1723                     // data is read.
1724                     match control_sockets
1725                         .get(index)
1726                         .map(|s| s.as_ref().get_readable_bytes())
1727                     {
1728                         Some(Ok(0)) | Some(Err(_)) => vm_control_indices_to_remove.push(index),
1729                         Some(Ok(x)) => info!("control index {} has {} bytes readable", index, x),
1730                         _ => {}
1731                     }
1732                 }
1733             }
1734         }
1735 
1736         // Sort in reverse so the highest indexes are removed first. This removal algorithm
1737         // preserved correct indexes as each element is removed.
1738         vm_control_indices_to_remove.sort_unstable_by(|a, b| b.cmp(a));
1739         vm_control_indices_to_remove.dedup();
1740         for index in vm_control_indices_to_remove {
1741             control_sockets.swap_remove(index);
1742             if let Some(socket) = control_sockets.get(index) {
1743                 poll_ctx
1744                     .modify(
1745                         socket,
1746                         WatchingEvents::empty().set_read(),
1747                         Token::VmControl { index },
1748                     )
1749                     .map_err(Error::PollContextAdd)?;
1750             }
1751         }
1752     }
1753 
1754     // VCPU threads MUST see the VmRunMode flag, otherwise they may re-enter the VM.
1755     run_mode_arc.set_and_notify(VmRunMode::Exiting);
1756     for handle in vcpu_handles {
1757         match handle.kill(SIGRTMIN() + 0) {
1758             Ok(_) => {
1759                 if let Err(e) = handle.join() {
1760                     error!("failed to join vcpu thread: {:?}", e);
1761                 }
1762             }
1763             Err(e) => error!("failed to kill vcpu thread: {}", e),
1764         }
1765     }
1766 
1767     stdin_lock
1768         .set_canon_mode()
1769         .expect("failed to restore canonical mode for terminal");
1770 
1771     Ok(())
1772 }
1773