crosvm: add vhost_net virtio device

vhost_net is an in-kernel implementation of a virtio_net device. This device
implementation is responsible only for negotiating features and setting up
vhost_net to point to the virtqueues in the guest.

Signed-off-by: Stephen Barber <smbarber@chromium.org>

BUG=chromium:738639
TEST=boot with networking enabled and ping host

Change-Id: I9192765b9348528a8a7a7f273e9522c2373d22fb
Reviewed-on: https://chromium-review.googlesource.com/538106
Commit-Ready: Stephen Barber <smbarber@chromium.org>
Tested-by: Stephen Barber <smbarber@chromium.org>
Reviewed-by: Zach Reizner <zachr@chromium.org>

6 files changed, 450 insertions, 1 deletions

diff --git a/Cargo.toml b/Cargo.toml
index b24346d..1c1e60b 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -15,6 +15,11 @@ x86_64 = { path = "x86_64" }
 kernel_loader = { path = "kernel_loader" }
 libc = "0.2.21"
 byteorder = "1"
+syscall_defines = { path = "syscall_defines" }
+net_sys = { path = "net_sys" }
+net_util = { path = "net_util" }
+vhost = { path = "vhost" }
+virtio_sys = { path = "virtio_sys" }
 
 [dependencies.clap]
 version = "*"
diff --git a/net_device.policy b/net_device.policy
new file mode 100644
index 0000000..f1e0515
--- /dev/null
+++ b/net_device.policy
@@ -0,0 +1,35 @@
+close: 1
+exit_group: 1
+futex: 1
+# Whitelist vhost_net ioctls only.
+# arg1 == VHOST_GET_FEATURES ||
+# arg1 == VHOST_SET_FEATURES ||
+# arg1 == VHOST_SET_OWNER ||
+# arg1 == VHOST_RESET_OWNER ||
+# arg1 == VHOST_SET_MEM_TABLE ||
+# arg1 == VHOST_SET_LOG_BASE ||
+# arg1 == VHOST_SET_LOG_FD ||
+# arg1 == VHOST_SET_VRING_NUM ||
+# arg1 == VHOST_SET_VRING_ADDR ||
+# arg1 == VHOST_SET_VRING_BASE ||
+# arg1 == VHOST_GET_VRING_BASE ||
+# arg1 == VHOST_SET_VRING_KICK ||
+# arg1 == VHOST_SET_VRING_CALL ||
+# arg1 == VHOST_SET_VRING_ERR ||
+# arg1 == VHOST_NET_SET_BACKEND
+ioctl: arg1 == 0x8008af00 || arg1 == 0x4008af00 || arg1 == 0x0000af01 || arg1 == 0x0000af02 || arg1 == 0x4008af03 || arg1 == 0x4008af04 || arg1 == 0x4004af07 || arg1 == 0x4008af10 || arg1 == 0x4028af11 || arg1 == 0x4008af12 || arg1 == 0xc008af12 || arg1 == 0x4008af20 || arg1 == 0x4008af21 || arg1 == 0x4008af22 || arg1 == 0x4008af30
+# Disallow mmap with PROT_EXEC set.  The syntax here doesn't allow bit
+# negation, thus the manually negated mask constant.
+mmap: arg2 in 0xfffffffb
+mprotect: arg2 in 0xfffffffb
+munmap: 1
+poll: 1
+read: 1
+recvfrom: 1
+sched_getaffinity: 1
+set_robust_list: 1
+sigaltstack: 1
+# Disallow clone's other than new threads.
+# arg0 is flags. Because kernel.
+clone: arg0 & 0x00010000
+write: 1
diff --git a/src/hw/virtio/mod.rs b/src/hw/virtio/mod.rs
index 7335ac1..a7298bb 100644
--- a/src/hw/virtio/mod.rs
+++ b/src/hw/virtio/mod.rs
@@ -7,10 +7,12 @@
 mod queue;
 mod mmio;
 mod block;
+mod net;
 
 pub use self::queue::*;
 pub use self::mmio::*;
 pub use self::block::*;
+pub use self::net::*;
 
 const DEVICE_ACKNOWLEDGE: u32 = 0x01;
 const DEVICE_DRIVER: u32 = 0x02;
@@ -18,6 +20,7 @@ const DEVICE_DRIVER_OK: u32 = 0x04;
 const DEVICE_FEATURES_OK: u32 = 0x08;
 const DEVICE_FAILED: u32 = 0x80;
 
+const TYPE_NET: u32 = 1;
 const TYPE_BLOCK: u32 = 2;
 
 const INTERRUPT_STATUS_USED_RING: u32 = 0x1;
diff --git a/src/hw/virtio/net.rs b/src/hw/virtio/net.rs
new file mode 100644
index 0000000..731d579
--- /dev/null
+++ b/src/hw/virtio/net.rs
@@ -0,0 +1,340 @@
+// Copyright 2017 The Chromium OS Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+use std::mem;
+use std::net::Ipv4Addr;
+use std::os::raw::*;
+use std::os::unix::io::{AsRawFd, RawFd};
+use std::sync::Arc;
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::thread::spawn;
+
+use net_sys;
+use net_util::{Tap, Error as TapError};
+use sys_util::{Error as SysError, EventFd, GuestMemory, Poller};
+use vhost::{VhostNet, Error as VhostError};
+use virtio_sys::{vhost, virtio_net};
+use virtio_sys::virtio_net::virtio_net_hdr_mrg_rxbuf;
+
+use super::{VirtioDevice, Queue, INTERRUPT_STATUS_USED_RING, TYPE_NET};
+
+const QUEUE_SIZE: u16 = 256;
+const QUEUE_SIZES: &'static [u16] = &[QUEUE_SIZE, QUEUE_SIZE];
+
+#[derive(Debug)]
+pub enum NetError {
+    /// Creating kill eventfd failed.
+    CreateKillEventFd(SysError),
+    /// Cloning kill eventfd failed.
+    CloneKillEventFd(SysError),
+    /// Open tap device failed.
+    TapOpen(TapError),
+    /// Setting tap IP failed.
+    TapSetIp(TapError),
+    /// Setting tap netmask failed.
+    TapSetNetmask(TapError),
+    /// Setting tap interface offload flags failed.
+    TapSetOffload(TapError),
+    /// Setting vnet header size failed.
+    TapSetVnetHdrSize(TapError),
+    /// Enabling tap interface failed.
+    TapEnable(TapError),
+    /// Open vhost-net device failed.
+    VhostOpen(VhostError),
+    /// Set owner failed.
+    VhostSetOwner(VhostError),
+    /// Get features failed.
+    VhostGetFeatures(VhostError),
+    /// Set features failed.
+    VhostSetFeatures(VhostError),
+    /// Set mem table failed.
+    VhostSetMemTable(VhostError),
+    /// Set vring num failed.
+    VhostSetVringNum(VhostError),
+    /// Set vring addr failed.
+    VhostSetVringAddr(VhostError),
+    /// Set vring base failed.
+    VhostSetVringBase(VhostError),
+    /// Set vring call failed.
+    VhostSetVringCall(VhostError),
+    /// Set vring kick failed.
+    VhostSetVringKick(VhostError),
+    /// Net set backend failed.
+    VhostNetSetBackend(VhostError),
+    /// Failed to create vhost eventfd.
+    VhostIrqCreate(SysError),
+    /// Failed to read vhost eventfd.
+    VhostIrqRead(SysError),
+    /// Error while polling for events.
+    PollError(SysError),
+}
+
+struct Worker {
+    queues: Vec<Queue>,
+    tap: Tap,
+    vhost_net: VhostNet,
+    vhost_interrupt: EventFd,
+    interrupt_status: Arc<AtomicUsize>,
+    interrupt_evt: EventFd,
+    acked_features: u64,
+}
+
+impl Worker {
+    fn signal_used_queue(&self) {
+        self.interrupt_status
+            .fetch_or(INTERRUPT_STATUS_USED_RING as usize, Ordering::SeqCst);
+        self.interrupt_evt.write(1).unwrap();
+    }
+
+    fn run(&mut self, queue_evts: Vec<EventFd>, kill_evt: EventFd) -> Result<(), NetError> {
+        // Preliminary setup for vhost net.
+        self.vhost_net.set_owner().map_err(NetError::VhostSetOwner)?;
+
+        let avail_features = self.vhost_net
+            .get_features()
+            .map_err(NetError::VhostGetFeatures)?;
+
+        let features: c_ulonglong = self.acked_features & avail_features;
+        self.vhost_net
+            .set_features(features)
+            .map_err(NetError::VhostSetFeatures)?;
+
+        self.vhost_net
+            .set_mem_table()
+            .map_err(NetError::VhostSetMemTable)?;
+
+        for (queue_index, ref queue) in self.queues.iter().enumerate() {
+            self.vhost_net
+                .set_vring_num(queue_index, queue.max_size)
+                .map_err(NetError::VhostSetVringNum)?;
+
+            self.vhost_net
+                .set_vring_addr(QUEUE_SIZES[queue_index],
+                                queue.actual_size(),
+                                queue_index,
+                                0,
+                                queue.desc_table,
+                                queue.used_ring,
+                                queue.avail_ring,
+                                None)
+                .map_err(NetError::VhostSetVringAddr)?;
+            self.vhost_net
+                .set_vring_base(queue_index, 0)
+                .map_err(NetError::VhostSetVringBase)?;
+            self.vhost_net
+                .set_vring_call(queue_index, &self.vhost_interrupt)
+                .map_err(NetError::VhostSetVringCall)?;
+            self.vhost_net
+                .set_vring_kick(queue_index, &queue_evts[queue_index])
+                .map_err(NetError::VhostSetVringKick)?;
+            self.vhost_net
+                .net_set_backend(queue_index, &self.tap)
+                .map_err(NetError::VhostNetSetBackend)?;
+        }
+
+        const VHOST_IRQ: u32 = 1;
+        const KILL: u32 = 2;
+
+        let mut poller = Poller::new(2);
+
+        'poll: loop {
+            let tokens =
+                match poller.poll(&[(VHOST_IRQ, &self.vhost_interrupt), (KILL, &kill_evt)]) {
+                    Ok(v) => v,
+                    Err(e) => return Err(NetError::PollError(e))
+                };
+
+            let mut needs_interrupt = false;
+            for &token in tokens {
+                match token {
+                    VHOST_IRQ => {
+                        needs_interrupt = true;
+                        self.vhost_interrupt.read().map_err(NetError::VhostIrqRead)?;
+                    }
+                    KILL => break 'poll,
+                    _ => unreachable!(),
+                }
+            }
+            if needs_interrupt {
+                self.signal_used_queue();
+            }
+        }
+        Ok(())
+    }
+}
+
+pub struct Net {
+    workers_kill_evt: Option<EventFd>,
+    kill_evt: EventFd,
+    tap: Option<Tap>,
+    vhost_net: Option<VhostNet>,
+    vhost_interrupt: Option<EventFd>,
+    avail_features: u64,
+    acked_features: u64,
+}
+
+impl Net {
+    /// Create a new virtio network device with the given IP address and
+    /// netmask.
+    pub fn new(ip_addr: Ipv4Addr,
+               netmask: Ipv4Addr,
+               mem: &GuestMemory) -> Result<Net, NetError> {
+        let kill_evt = EventFd::new().map_err(NetError::CreateKillEventFd)?;
+
+        let tap = Tap::new().map_err(NetError::TapOpen)?;
+        tap.set_ip_addr(ip_addr).map_err(NetError::TapSetIp)?;
+        tap.set_netmask(netmask).map_err(NetError::TapSetNetmask)?;
+
+        // Set offload flags to match the virtio features below.
+        tap.set_offload(net_sys::TUN_F_CSUM |
+                        net_sys::TUN_F_UFO |
+                        net_sys::TUN_F_TSO4 |
+                        net_sys::TUN_F_TSO6)
+            .map_err(NetError::TapSetOffload)?;
+
+        // We declare VIRTIO_NET_F_MRG_RXBUF, so set the vnet hdr size to match.
+        let vnet_hdr_size = mem::size_of::<virtio_net_hdr_mrg_rxbuf>() as i32;
+        tap.set_vnet_hdr_size(vnet_hdr_size)
+            .map_err(NetError::TapSetVnetHdrSize)?;
+
+        tap.enable().map_err(NetError::TapEnable)?;
+        let vhost_net = VhostNet::new(mem).map_err(NetError::VhostOpen)?;
+
+        let avail_features =
+            1 << virtio_net::VIRTIO_NET_F_GUEST_CSUM |
+            1 << virtio_net::VIRTIO_NET_F_CSUM |
+            1 << virtio_net::VIRTIO_NET_F_GUEST_TSO4 |
+            1 << virtio_net::VIRTIO_NET_F_GUEST_UFO |
+            1 << virtio_net::VIRTIO_NET_F_HOST_TSO4 |
+            1 << virtio_net::VIRTIO_NET_F_HOST_UFO |
+            1 << virtio_net::VIRTIO_NET_F_MRG_RXBUF |
+            1 << vhost::VIRTIO_RING_F_INDIRECT_DESC |
+            1 << vhost::VIRTIO_RING_F_EVENT_IDX |
+            1 << vhost::VIRTIO_F_NOTIFY_ON_EMPTY |
+            1 << vhost::VIRTIO_F_VERSION_1;
+
+        Ok(Net {
+               workers_kill_evt: Some(kill_evt.try_clone().map_err(NetError::CloneKillEventFd)?),
+               kill_evt: kill_evt,
+               tap: Some(tap),
+               vhost_net: Some(vhost_net),
+               vhost_interrupt: Some(EventFd::new().map_err(NetError::VhostIrqCreate)?),
+               avail_features: avail_features,
+               acked_features: 0u64,
+           })
+    }
+}
+
+impl Drop for Net {
+    fn drop(&mut self) {
+        // Only kill the child if it claimed its eventfd.
+        if self.workers_kill_evt.is_none() {
+            // Ignore the result because there is nothing we can do about it.
+            let _ = self.kill_evt.write(1);
+        }
+    }
+}
+
+impl VirtioDevice for Net {
+    fn keep_fds(&self) -> Vec<RawFd> {
+        let mut keep_fds = Vec::new();
+
+        if let Some(ref tap) = self.tap {
+            keep_fds.push(tap.as_raw_fd());
+        }
+
+        if let Some(ref vhost_net) = self.vhost_net {
+            keep_fds.push(vhost_net.as_raw_fd());
+        }
+
+        if let Some(ref vhost_interrupt) = self.vhost_interrupt {
+            keep_fds.push(vhost_interrupt.as_raw_fd());
+        }
+
+        if let Some(ref workers_kill_evt) = self.workers_kill_evt {
+            keep_fds.push(workers_kill_evt.as_raw_fd());
+        }
+
+        keep_fds
+    }
+
+    fn device_type(&self) -> u32 {
+        TYPE_NET
+    }
+
+    fn queue_max_sizes(&self) -> &[u16] {
+        QUEUE_SIZES
+    }
+
+    fn features(&self, page: u32) -> u32 {
+        match page {
+            0 => self.avail_features as u32,
+            1 => (self.avail_features >> 32) as u32,
+            _ => {
+                warn!("net: virtio net got request for features page: {}", page);
+                0u32
+            }
+        }
+    }
+
+    fn ack_features(&mut self, page: u32, value: u32) {
+        let mut v = match page {
+            0 => value as u64,
+            1 => (value as u64) << 32,
+            _ => {
+                warn!("net: virtio net device cannot ack unknown feature page: {}",
+                         page);
+                0u64
+            }
+        };
+
+        // 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!("net: virtio net got unknown feature ack: {:x}", v);
+
+            // Don't count these features as acked.
+            v &= !unrequested_features;
+        }
+        self.acked_features |= v;
+    }
+
+    fn activate(&mut self,
+                _: GuestMemory,
+                interrupt_evt: EventFd,
+                status: Arc<AtomicUsize>,
+                queues: Vec<Queue>,
+                queue_evts: Vec<EventFd>) {
+        if queues.len() != 2 || queue_evts.len() != 2 {
+            error!("net: expected 2 queues, got {}", queues.len());
+            return;
+        }
+
+        if let Some(vhost_net) = self.vhost_net.take() {
+            if let Some(tap) = self.tap.take() {
+                if let Some(vhost_interrupt) = self.vhost_interrupt.take() {
+                    if let Some(kill_evt) = self.workers_kill_evt.take() {
+                        let acked_features = self.acked_features;
+                        spawn(move || {
+                            let mut worker = Worker {
+                                queues: queues,
+                                tap: tap,
+                                vhost_net: vhost_net,
+                                vhost_interrupt: vhost_interrupt,
+                                interrupt_status: status,
+                                interrupt_evt: interrupt_evt,
+                                acked_features: acked_features,
+                            };
+                            let result = worker.run(queue_evts, kill_evt);
+                            if let Err(e) = result {
+                                error!("net worker thread exited with error: {:?}",
+                                       e);
+                            }
+                        });
+                    }
+                }
+            }
+        }
+    }
+}
diff --git a/src/hw/virtio/queue.rs b/src/hw/virtio/queue.rs
index 1cdb9c0..e1f2207 100644
--- a/src/hw/virtio/queue.rs
+++ b/src/hw/virtio/queue.rs
@@ -198,7 +198,9 @@ impl Queue {
         }
     }
 
-    fn actual_size(&self) -> u16 {
+    /// Return the actual size of the queue, as the driver may not set up a
+    /// queue as big as the device allows.
+    pub fn actual_size(&self) -> u16 {
         min(self.size, self.max_size)
     }
 
diff --git a/src/main.rs b/src/main.rs
index b7c6cca..1e95a19 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -12,11 +12,16 @@ extern crate x86_64;
 extern crate kernel_loader;
 extern crate byteorder;
 #[macro_use] extern crate sys_util;
+extern crate net_sys;
+extern crate net_util;
+extern crate vhost;
+extern crate virtio_sys;
 
 use std::ffi::{CString, CStr};
 use std::fmt;
 use std::fs::File;
 use std::io::{stdin, stdout};
+use std::net;
 use std::path::{Path, PathBuf};
 use std::ptr;
 use std::string::String;
@@ -45,9 +50,14 @@ enum Error {
     Disk(std::io::Error),
     BlockDeviceNew(sys_util::Error),
     BlockDeviceRootSetup(sys_util::Error),
+    NetDeviceNew(hw::virtio::NetError),
+    NetDeviceRootSetup(sys_util::Error),
+    MacAddressNeedsNetConfig,
+    NetMissingConfig,
     DeviceJail(io_jail::Error),
     DevicePivotRoot(io_jail::Error),
     RegisterBlock(device_manager::Error),
+    RegisterNet(device_manager::Error),
     Cmdline(kernel_cmdline::Error),
     RegisterIoevent(sys_util::Error),
     RegisterIrqfd(sys_util::Error),
@@ -89,8 +99,15 @@ impl fmt::Display for Error {
                 write!(f, "failed to create root directory for a block device: {:?}", e)
             }
             &Error::RegisterBlock(ref e) => write!(f, "error registering block device: {:?}", e),
+            &Error::NetDeviceNew(ref e) => write!(f, "failed to set up networking: {:?}", e),
+            &Error::NetDeviceRootSetup(ref e) => {
+                write!(f, "failed to create root directory for a net device: {:?}", e)
+            }
+            &Error::MacAddressNeedsNetConfig => write!(f, "MAC address can only be specified when host IP and netmask are provided"),
+            &Error::NetMissingConfig => write!(f, "networking requires both host IP and netmask specified"),
             &Error::DeviceJail(ref e) => write!(f, "failed to jail device: {:?}", e),
             &Error::DevicePivotRoot(ref e) => write!(f, "failed to pivot root device: {:?}", e),
+            &Error::RegisterNet(ref e) => write!(f, "error registering net device: {:?}", e),
             &Error::Cmdline(ref e) => write!(f, "the given kernel command line was invalid: {}", e),
             &Error::RegisterIoevent(ref e) => write!(f, "error registering ioevent: {:?}", e),
             &Error::RegisterIrqfd(ref e) => write!(f, "error registering irqfd: {:?}", e),
@@ -114,6 +131,9 @@ struct Config {
     memory: Option<usize>,
     kernel_image: File,
     params: Option<String>,
+    host_ip: Option<net::Ipv4Addr>,
+    netmask: Option<net::Ipv4Addr>,
+    mac_address: Option<String>,
     socket_path: Option<String>,
     multiprocess: bool,
     warn_unknown_ports: bool,
@@ -185,6 +205,15 @@ fn wait_all_children() -> bool {
 }
 
 fn run_config(cfg: Config) -> Result<()> {
+    if cfg.mac_address.is_some() &&
+       (cfg.netmask.is_none() || cfg.host_ip.is_none()) {
+        return Err(Error::MacAddressNeedsNetConfig);
+    }
+
+    if cfg.netmask.is_some() != cfg.host_ip.is_some() {
+        return Err(Error::NetMissingConfig);
+    }
+
     let socket = if let Some(ref socket_path) = cfg.socket_path {
         Some(ControlSocketRecv::new(socket_path)
                  .map_err(|e| Error::Socket(e))?)
@@ -226,6 +255,26 @@ fn run_config(cfg: Config) -> Result<()> {
                       .map_err(Error::RegisterBlock)?;
     }
 
+    // We checked above that if the IP is defined, then the netmask is, too.
+    let net_root = TempDir::new(&PathBuf::from("/tmp/net_root"))
+        .map_err(Error::NetDeviceRootSetup)?;
+    if let Some(host_ip) = cfg.host_ip {
+        if let Some(netmask) = cfg.netmask {
+            let net_box = Box::new(hw::virtio::Net::new(host_ip, netmask, &guest_mem)
+                                   .map_err(|e| Error::NetDeviceNew(e))?);
+            let jail = if cfg.multiprocess {
+                let net_root_path = net_root.as_path().unwrap(); // Won't fail if new succeeded.
+
+                Some(create_base_minijail(net_root_path, Path::new("net_device.policy"))?)
+            }
+            else {
+                None
+            };
+
+            device_manager.register_mmio(net_box, jail, &mut cmdline).map_err(Error::RegisterNet)?;
+        }
+    }
+
     if let Some(params) = cfg.params {
         cmdline
             .insert_str(params)
@@ -586,6 +635,18 @@ fn main() {
                                  .short("u")
                                  .long("multiprocess")
                                  .help("run the devices in a child process"))
+                        .arg(Arg::with_name("host_ip")
+                                 .long("host_ip")
+                                 .value_name("HOST_IP")
+                                 .help("IP address to assign to host tap interface"))
+                        .arg(Arg::with_name("netmask")
+                                 .long("netmask")
+                                 .value_name("NETMASK")
+                                 .help("netmask for VM subnet"))
+                        .arg(Arg::with_name("mac")
+                                 .long("mac")
+                                 .value_name("MAC")
+                                 .help("mac address for VM"))
                         .arg(Arg::with_name("socket")
                                  .short("s")
                                  .long("socket")
@@ -624,6 +685,9 @@ fn main() {
                     .expect("Expected kernel image path to be valid"),
                 params: matches.value_of("params").map(|s| s.to_string()),
                 multiprocess: matches.is_present("multiprocess"),
+                host_ip: matches.value_of("host_ip").and_then(|v| v.parse().ok()),
+                netmask: matches.value_of("netmask").and_then(|v| v.parse().ok()),
+                mac_address: matches.value_of("mac").map(|s| s.to_string()),
                 socket_path: matches.value_of("socket").map(|s| s.to_string()),
                 warn_unknown_ports: matches.is_present("warn-unknown-ports"),
             };