diff --git a/virtio-devices/src/balloon.rs b/virtio-devices/src/balloon.rs
new file mode 100644
index 000000000..8287be218
--- /dev/null
+++ b/virtio-devices/src/balloon.rs
@@ -0,0 +1,593 @@
+// Copyright (c) 2020 Ant Financial
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use super::Error as DeviceError;
+use super::{
+    ActivateError, ActivateResult, DeviceEventT, Queue, VirtioDevice, VirtioDeviceType,
+    VIRTIO_F_VERSION_1,
+};
+use crate::vm_memory::GuestMemory;
+use crate::{VirtioInterrupt, VirtioInterruptType};
+use libc::EFD_NONBLOCK;
+use std::cmp;
+use std::fs::File;
+use std::io::{self, Write};
+use std::mem::size_of;
+use std::os::unix::io::{AsRawFd, FromRawFd};
+use std::result;
+use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
+use std::sync::mpsc;
+use std::sync::{Arc, Mutex};
+use std::thread;
+use vm_memory::{
+    Address, ByteValued, Bytes, GuestAddress, GuestAddressSpace, GuestMemoryAtomic,
+    GuestMemoryError, GuestMemoryMmap, GuestMemoryRegion,
+};
+use vm_migration::{Migratable, MigratableError, Pausable, Snapshottable, Transportable};
+use vmm_sys_util::eventfd::EventFd;
+
+const QUEUE_SIZE: u16 = 128;
+const NUM_QUEUES: usize = 2;
+const QUEUE_SIZES: &[u16] = &[QUEUE_SIZE; NUM_QUEUES];
+
+// Get resize event.
+const RESIZE_EVENT: DeviceEventT = 0;
+// New descriptors are pending on the virtio queue.
+const INFLATE_QUEUE_EVENT: DeviceEventT = 1;
+// New descriptors are pending on the virtio queue.
+const DEFLATE_QUEUE_EVENT: DeviceEventT = 2;
+// The device has been dropped.
+const KILL_EVENT: DeviceEventT = 3;
+// The device should be paused.
+const PAUSE_EVENT: DeviceEventT = 4;
+
+// Page shift in the host.
+const PAGE_SHIFT: u32 = 12;
+
+// Size of a PFN in the balloon interface.
+const VIRTIO_BALLOON_PFN_SHIFT: u64 = 12;
+
+#[derive(Debug)]
+pub enum Error {
+    // Guest gave us bad memory addresses.
+    GuestMemory(GuestMemoryError),
+    // Guest gave us a write only descriptor that protocol says to read from.
+    UnexpectedWriteOnlyDescriptor,
+    // Guest sent us invalid request.
+    InvalidRequest,
+    // Madvise fail.
+    MadviseFail(std::io::Error),
+    // Failed to EventFd write.
+    EventFdWriteFail(std::io::Error),
+    // Failed to EventFd try_clone.
+    EventFdTryCloneFail(std::io::Error),
+    // Failed to MpscRecv.
+    MpscRecvFail(mpsc::RecvError),
+    // Resize invalid argument
+    ResizeInval(String),
+    // process_queue got wrong ev_type
+    ProcessQueueWrongEvType(u16),
+    // Fail tp signal
+    FailedSignal(io::Error),
+}
+
+// Got from include/uapi/linux/virtio_balloon.h
+#[repr(C, packed)]
+#[derive(Copy, Clone, Debug, Default)]
+struct VirtioBalloonConfig {
+    // Number of pages host wants Guest to give up.
+    num_pages: u32,
+    // Number of pages we've actually got in balloon.
+    actual: u32,
+}
+
+// Safe because it only has data and has no implicit padding.
+unsafe impl ByteValued for VirtioBalloonConfig {}
+
+struct VirtioBalloonResizeReceiver {
+    size: Arc<AtomicU64>,
+    tx: mpsc::Sender<Result<(), Error>>,
+    evt: EventFd,
+}
+
+impl VirtioBalloonResizeReceiver {
+    fn get_size(&self) -> u64 {
+        self.size.load(Ordering::SeqCst)
+    }
+
+    fn send(&self, r: Result<(), Error>) -> Result<(), mpsc::SendError<Result<(), Error>>> {
+        self.tx.send(r)
+    }
+}
+
+struct VirtioBalloonResize {
+    size: Arc<AtomicU64>,
+    tx: mpsc::Sender<Result<(), Error>>,
+    rx: mpsc::Receiver<Result<(), Error>>,
+    evt: EventFd,
+}
+
+impl VirtioBalloonResize {
+    pub fn new() -> io::Result<Self> {
+        let (tx, rx) = mpsc::channel();
+
+        Ok(Self {
+            size: Arc::new(AtomicU64::new(0)),
+            tx,
+            rx,
+            evt: EventFd::new(EFD_NONBLOCK)?,
+        })
+    }
+
+    pub fn get_receiver(&self) -> Result<VirtioBalloonResizeReceiver, Error> {
+        Ok(VirtioBalloonResizeReceiver {
+            size: self.size.clone(),
+            tx: self.tx.clone(),
+            evt: self.evt.try_clone().map_err(Error::EventFdTryCloneFail)?,
+        })
+    }
+
+    pub fn work(&self, size: u64) -> Result<(), Error> {
+        self.size.store(size, Ordering::SeqCst);
+        self.evt.write(1).map_err(Error::EventFdWriteFail)?;
+        self.rx.recv().map_err(Error::MpscRecvFail)?
+    }
+}
+
+struct BalloonEpollHandler {
+    config: Arc<Mutex<VirtioBalloonConfig>>,
+    resize_receiver: VirtioBalloonResizeReceiver,
+    queues: Vec<Queue>,
+    mem: GuestMemoryAtomic<GuestMemoryMmap>,
+    interrupt_cb: Arc<dyn VirtioInterrupt>,
+    inflate_queue_evt: EventFd,
+    deflate_queue_evt: EventFd,
+    kill_evt: EventFd,
+    pause_evt: EventFd,
+}
+
+impl BalloonEpollHandler {
+    fn signal(
+        &self,
+        int_type: &VirtioInterruptType,
+        queue: Option<&Queue>,
+    ) -> result::Result<(), Error> {
+        self.interrupt_cb.trigger(int_type, queue).map_err(|e| {
+            error!("Failed to signal used queue: {:?}", e);
+            Error::FailedSignal(e)
+        })
+    }
+
+    fn process_queue(&mut self, ev_type: u16) -> result::Result<(), Error> {
+        let queue_index = match ev_type {
+            INFLATE_QUEUE_EVENT => 0,
+            DEFLATE_QUEUE_EVENT => 1,
+            _ => return Err(Error::ProcessQueueWrongEvType(ev_type)),
+        };
+
+        let mut used_desc_heads = [0; QUEUE_SIZE as usize];
+        let mut used_count = 0;
+        let mem = self.mem.memory();
+        for avail_desc in self.queues[queue_index].iter(&mem) {
+            used_desc_heads[used_count] = avail_desc.index;
+            used_count += 1;
+
+            // The head contains the request type which MUST be readable.
+            if avail_desc.is_write_only() {
+                error!("The head contains the request type is not right");
+                return Err(Error::UnexpectedWriteOnlyDescriptor);
+            }
+            if avail_desc.len as usize % size_of::<u32>() != 0 {
+                error!("the request size {} is not right", avail_desc.len);
+                return Err(Error::InvalidRequest);
+            }
+
+            let mut offset = 0u64;
+            while offset < avail_desc.len as u64 {
+                let pfn: u32 = mem
+                    .read_obj(GuestAddress(avail_desc.addr.0 + offset))
+                    .map_err(Error::GuestMemory)?;
+                offset += size_of::<u32>() as u64;
+
+                let pa = pfn << VIRTIO_BALLOON_PFN_SHIFT;
+                if let Some(region) = mem.find_region(GuestAddress(pa as u64)) {
+                    let addr = region.as_ptr() as u64 + pa as u64 - region.start_addr().raw_value();
+                    let advice = match ev_type {
+                        INFLATE_QUEUE_EVENT => libc::MADV_DONTNEED,
+                        DEFLATE_QUEUE_EVENT => libc::MADV_WILLNEED,
+                        _ => return Err(Error::ProcessQueueWrongEvType(ev_type)),
+                    };
+                    // Need unsafe to do syscall madvise
+                    let res = unsafe {
+                        libc::madvise(
+                            addr as *mut libc::c_void,
+                            (1 << PAGE_SHIFT) as libc::size_t,
+                            advice,
+                        )
+                    };
+                    if res != 0 {
+                        return Err(Error::MadviseFail(io::Error::last_os_error()));
+                    }
+                } else {
+                    error!("Address {} is not available", pa);
+                    return Err(Error::InvalidRequest);
+                }
+            }
+        }
+
+        for &desc_index in &used_desc_heads[..used_count] {
+            self.queues[queue_index].add_used(&mem, desc_index, 0);
+        }
+        if used_count > 0 {
+            self.signal(&VirtioInterruptType::Queue, Some(&self.queues[queue_index]))?;
+        }
+
+        Ok(())
+    }
+
+    fn run(&mut self, paused: Arc<AtomicBool>) -> result::Result<(), DeviceError> {
+        // Create the epoll file descriptor
+        let epoll_fd = epoll::create(true).map_err(DeviceError::EpollCreateFd)?;
+        // Use 'File' to enforce closing on 'epoll_fd'
+        let epoll_file = unsafe { File::from_raw_fd(epoll_fd) };
+
+        // Add events
+        epoll::ctl(
+            epoll_file.as_raw_fd(),
+            epoll::ControlOptions::EPOLL_CTL_ADD,
+            self.resize_receiver.evt.as_raw_fd(),
+            epoll::Event::new(epoll::Events::EPOLLIN, u64::from(RESIZE_EVENT)),
+        )
+        .map_err(DeviceError::EpollCtl)?;
+
+        epoll::ctl(
+            epoll_file.as_raw_fd(),
+            epoll::ControlOptions::EPOLL_CTL_ADD,
+            self.inflate_queue_evt.as_raw_fd(),
+            epoll::Event::new(epoll::Events::EPOLLIN, u64::from(INFLATE_QUEUE_EVENT)),
+        )
+        .map_err(DeviceError::EpollCtl)?;
+
+        epoll::ctl(
+            epoll_file.as_raw_fd(),
+            epoll::ControlOptions::EPOLL_CTL_ADD,
+            self.deflate_queue_evt.as_raw_fd(),
+            epoll::Event::new(epoll::Events::EPOLLIN, u64::from(DEFLATE_QUEUE_EVENT)),
+        )
+        .map_err(DeviceError::EpollCtl)?;
+
+        epoll::ctl(
+            epoll_file.as_raw_fd(),
+            epoll::ControlOptions::EPOLL_CTL_ADD,
+            self.kill_evt.as_raw_fd(),
+            epoll::Event::new(epoll::Events::EPOLLIN, u64::from(KILL_EVENT)),
+        )
+        .map_err(DeviceError::EpollCtl)?;
+
+        epoll::ctl(
+            epoll_file.as_raw_fd(),
+            epoll::ControlOptions::EPOLL_CTL_ADD,
+            self.pause_evt.as_raw_fd(),
+            epoll::Event::new(epoll::Events::EPOLLIN, u64::from(PAUSE_EVENT)),
+        )
+        .map_err(DeviceError::EpollCtl)?;
+
+        const EPOLL_EVENTS_LEN: usize = 100;
+        let mut events = vec![epoll::Event::new(epoll::Events::empty(), 0); EPOLL_EVENTS_LEN];
+
+        'epoll: loop {
+            let num_events = match epoll::wait(epoll_file.as_raw_fd(), -1, &mut events[..]) {
+                Ok(res) => res,
+                Err(e) => {
+                    if e.kind() == io::ErrorKind::Interrupted {
+                        // It's well defined from the epoll_wait() syscall
+                        // documentation that the epoll loop can be interrupted
+                        // before any of the requested events occurred or the
+                        // timeout expired. In both those cases, epoll_wait()
+                        // returns an error of type EINTR, but this should not
+                        // be considered as a regular error. Instead it is more
+                        // appropriate to retry, by calling into epoll_wait().
+                        continue;
+                    }
+                    return Err(DeviceError::EpollWait(e));
+                }
+            };
+
+            for event in events.iter().take(num_events) {
+                let ev_type = event.data as u16;
+
+                match ev_type {
+                    RESIZE_EVENT => {
+                        if let Err(e) = self.resize_receiver.evt.read() {
+                            return Err(DeviceError::EpollHander(format!(
+                                "Failed to get resize event: {:?}",
+                                e
+                            )));
+                        }
+                        let mut signal_error = false;
+                        let r = {
+                            let mut config = self.config.lock().unwrap();
+                            config.num_pages =
+                                (self.resize_receiver.get_size() >> PAGE_SHIFT) as u32;
+                            if let Err(e) = self.signal(&VirtioInterruptType::Config, None) {
+                                signal_error = true;
+                                Err(e)
+                            } else {
+                                Ok(())
+                            }
+                        };
+                        if let Err(e) = &r {
+                            // This error will send back to resize caller.
+                            error!("Handle resize event get error: {:?}", e);
+                        }
+                        if let Err(e) = self.resize_receiver.send(r) {
+                            return Err(DeviceError::EpollHander(format!(
+                                "Sending \"resize\" generated error: {:?}",
+                                e
+                            )));
+                        }
+                        if signal_error {
+                            return Err(DeviceError::EpollHander(String::from("Signal get error")));
+                        }
+                    }
+                    INFLATE_QUEUE_EVENT => {
+                        if let Err(e) = self.inflate_queue_evt.read() {
+                            return Err(DeviceError::EpollHander(format!(
+                                "Failed to get inflate queue event: {:?}",
+                                e
+                            )));
+                        } else if let Err(e) = self.process_queue(ev_type) {
+                            return Err(DeviceError::EpollHander(format!(
+                                "Failed to signal used inflate queue: {:?}",
+                                e
+                            )));
+                        }
+                    }
+                    DEFLATE_QUEUE_EVENT => {
+                        if let Err(e) = self.deflate_queue_evt.read() {
+                            return Err(DeviceError::EpollHander(format!(
+                                "Failed to get deflate queue event: {:?}",
+                                e
+                            )));
+                        } else if let Err(e) = self.process_queue(ev_type) {
+                            return Err(DeviceError::EpollHander(format!(
+                                "Failed to signal used deflate queue: {:?}",
+                                e
+                            )));
+                        }
+                    }
+                    KILL_EVENT => {
+                        debug!("kill_evt received, stopping epoll loop");
+                        break 'epoll;
+                    }
+                    PAUSE_EVENT => {
+                        debug!("PAUSE_EVENT received, pausing virtio-balloon epoll loop");
+                        // We loop here to handle spurious park() returns.
+                        // Until we have not resumed, the paused boolean will
+                        // be true.
+                        while paused.load(Ordering::SeqCst) {
+                            thread::park();
+                        }
+                    }
+                    _ => {
+                        return Err(DeviceError::EpollHander(String::from(
+                            "Unknown event for virtio-mem",
+                        )));
+                    }
+                }
+            }
+        }
+
+        Ok(())
+    }
+}
+
+// Virtio device for exposing entropy to the guest OS through virtio.
+pub struct Balloon {
+    id: String,
+    resize: VirtioBalloonResize,
+    kill_evt: Option<EventFd>,
+    pause_evt: Option<EventFd>,
+    avail_features: u64,
+    pub acked_features: u64,
+    config: Arc<Mutex<VirtioBalloonConfig>>,
+    queue_evts: Option<Vec<EventFd>>,
+    interrupt_cb: Option<Arc<dyn VirtioInterrupt>>,
+    epoll_threads: Option<Vec<thread::JoinHandle<result::Result<(), DeviceError>>>>,
+    paused: Arc<AtomicBool>,
+}
+
+impl Balloon {
+    // Create a new virtio-balloon.
+    pub fn new(id: String) -> io::Result<Self> {
+        let avail_features = 1u64 << VIRTIO_F_VERSION_1;
+
+        let config = VirtioBalloonConfig::default();
+
+        Ok(Balloon {
+            id,
+            resize: VirtioBalloonResize::new()?,
+            kill_evt: None,
+            pause_evt: None,
+            avail_features,
+            acked_features: 0u64,
+            config: Arc::new(Mutex::new(config)),
+            queue_evts: None,
+            interrupt_cb: None,
+            epoll_threads: None,
+            paused: Arc::new(AtomicBool::new(false)),
+        })
+    }
+
+    pub fn resize(&self, size: u64) -> Result<(), Error> {
+        self.resize.work(size)
+    }
+}
+
+impl Drop for Balloon {
+    fn drop(&mut self) {
+        if let Some(kill_evt) = self.kill_evt.take() {
+            // Ignore the result because there is nothing we can do about it.
+            let _ = kill_evt.write(1);
+        }
+    }
+}
+
+impl VirtioDevice for Balloon {
+    fn device_type(&self) -> u32 {
+        VirtioDeviceType::TYPE_BALLOON as u32
+    }
+
+    fn queue_max_sizes(&self) -> &[u16] {
+        QUEUE_SIZES
+    }
+
+    fn features(&self) -> u64 {
+        self.avail_features
+    }
+
+    fn ack_features(&mut self, value: u64) {
+        let mut v = value;
+        // Check if the guest is ACK'ing a feature that we didn't claim to have.
+        let unrequested_features = v & !self.avail_features;
+        if unrequested_features != 0 {
+            warn!("Received acknowledge request for unknown feature.");
+
+            // Don't count these features as acked.
+            v &= !unrequested_features;
+        }
+        self.acked_features |= v;
+    }
+
+    fn read_config(&self, offset: u64, mut data: &mut [u8]) {
+        let config = self.config.lock().unwrap();
+        let config_slice = config.as_slice();
+        let config_len = config_slice.len() as u64;
+        if offset >= config_len {
+            error!("Failed to read config space");
+            return;
+        }
+        if let Some(end) = offset.checked_add(data.len() as u64) {
+            // This write can't fail, offset and end are checked against config_len.
+            data.write_all(&config_slice[offset as usize..cmp::min(end, config_len) as usize])
+                .unwrap();
+        }
+    }
+
+    fn write_config(&mut self, _offset: u64, _data: &[u8]) {
+        warn!("virtio-balloon device configuration is read-only");
+    }
+
+    fn activate(
+        &mut self,
+        mem: GuestMemoryAtomic<GuestMemoryMmap>,
+        interrupt_cb: Arc<dyn VirtioInterrupt>,
+        queues: Vec<Queue>,
+        mut queue_evts: Vec<EventFd>,
+    ) -> ActivateResult {
+        if queues.len() != NUM_QUEUES || queue_evts.len() != NUM_QUEUES {
+            error!(
+                "Cannot perform activate. Expected {} queue(s), got {}",
+                NUM_QUEUES,
+                queues.len()
+            );
+            return Err(ActivateError::BadActivate);
+        }
+
+        let (self_kill_evt, kill_evt) = EventFd::new(EFD_NONBLOCK)
+            .and_then(|e| Ok((e.try_clone()?, e)))
+            .map_err(|e| {
+                error!("failed creating kill EventFd pair: {}", e);
+                ActivateError::BadActivate
+            })?;
+        self.kill_evt = Some(self_kill_evt);
+
+        let (self_pause_evt, pause_evt) = EventFd::new(EFD_NONBLOCK)
+            .and_then(|e| Ok((e.try_clone()?, e)))
+            .map_err(|e| {
+                error!("failed creating pause EventFd pair: {}", e);
+                ActivateError::BadActivate
+            })?;
+        self.pause_evt = Some(self_pause_evt);
+
+        self.interrupt_cb = Some(interrupt_cb.clone());
+
+        let mut tmp_queue_evts: Vec<EventFd> = Vec::new();
+        for queue_evt in queue_evts.iter() {
+            // Save the queue EventFD as we need to return it on reset
+            // but clone it to pass into the thread.
+            tmp_queue_evts.push(queue_evt.try_clone().map_err(|e| {
+                error!("failed to clone queue EventFd: {}", e);
+                ActivateError::BadActivate
+            })?);
+        }
+        self.queue_evts = Some(tmp_queue_evts);
+
+        let mut handler = BalloonEpollHandler {
+            config: self.config.clone(),
+            resize_receiver: self.resize.get_receiver().map_err(|e| {
+                error!("failed to clone resize EventFd: {:?}", e);
+                ActivateError::BadActivate
+            })?,
+            queues,
+            mem,
+            interrupt_cb,
+            inflate_queue_evt: queue_evts.remove(0),
+            deflate_queue_evt: queue_evts.remove(0),
+            kill_evt,
+            pause_evt,
+        };
+
+        let paused = self.paused.clone();
+        let mut epoll_threads = Vec::new();
+        thread::Builder::new()
+            .name("virtio_balloon".to_string())
+            .spawn(move || handler.run(paused))
+            .map(|thread| epoll_threads.push(thread))
+            .map_err(|e| {
+                error!("failed to clone virtio-balloon epoll thread: {}", e);
+                ActivateError::BadActivate
+            })?;
+        self.epoll_threads = Some(epoll_threads);
+
+        Ok(())
+    }
+
+    fn reset(&mut self) -> Option<(Arc<dyn VirtioInterrupt>, Vec<EventFd>)> {
+        // We first must resume the virtio thread if it was paused.
+        if self.pause_evt.take().is_some() {
+            self.resume().ok()?;
+        }
+
+        if let Some(kill_evt) = self.kill_evt.take() {
+            // Ignore the result because there is nothing we can do about it.
+            let _ = kill_evt.write(1);
+        }
+
+        // Return the interrupt and queue EventFDs
+        Some((
+            self.interrupt_cb.take().unwrap(),
+            self.queue_evts.take().unwrap(),
+        ))
+    }
+}
+
+virtio_pausable!(Balloon);
+impl Snapshottable for Balloon {
+    fn id(&self) -> String {
+        self.id.clone()
+    }
+}
+impl Transportable for Balloon {}
+impl Migratable for Balloon {}
diff --git a/virtio-devices/src/lib.rs b/virtio-devices/src/lib.rs
index e543ab894..d3f5c07a6 100644
--- a/virtio-devices/src/lib.rs
+++ b/virtio-devices/src/lib.rs
@@ -29,6 +29,7 @@ use std::io;
 
 #[macro_use]
 mod device;
+pub mod balloon;
 pub mod block;
 mod console;
 mod iommu;
@@ -41,6 +42,7 @@ pub mod transport;
 pub mod vhost_user;
 pub mod vsock;
 
+pub use self::balloon::*;
 pub use self::block::*;
 pub use self::console::*;
 pub use self::device::*;