feat(xen): update xenclient and xenplatform to the latest structure (#433)

Cargo.lock

@@ -107,6 +107,12 @@ dependencies = [
  "windows-targets",
 ]
 
+[[package]]
+name = "bit-vec"
+version = "0.8.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5e764a1d40d510daf35e07be9eb06e75770908c27d411ee6c92109c9840eaaf7"
+
 [[package]]
 name = "bitflags"
 version = "2.6.0"
@@ -267,6 +273,8 @@ dependencies = [
 name = "krata-xenclient"
 version = "0.0.23"
 dependencies = [
+ "async-trait",
+ "bit-vec",
  "env_logger",
  "indexmap",
  "krata-xencall",

Cargo.toml

@@ -17,8 +17,8 @@ repository = "https://github.com/edera-dev/krata"
 
 [workspace.dependencies]
 async-trait = "0.1.83"
+bit-vec = "0.8.0"
 byteorder = "1"
-c2rust-bitfields = "0.19.0"
 elf = "0.7.4"
 env_logger = "0.11.5"
 flate2 = "1.0"

crates/xen/xenclient/Cargo.toml

@@ -9,6 +9,8 @@ edition = "2021"
 resolver = "2"
 
 [dependencies]
+async-trait = { workspace = true }
+bit-vec = { workspace = true }
 indexmap = { workspace = true }
 log = { workspace = true }
 krata-xencall = { path = "../xencall", version = "^0.0.23" }

crates/xen/xenclient/examples/boot.rs

@@ -1,15 +1,15 @@
+use std::sync::Arc;
 use std::{env, process};
 use tokio::fs;
 use uuid::Uuid;
 use xenclient::error::Result;
-use xenclient::{DomainConfig, XenClient};
-use xenplatform::domain::BaseDomainConfig;
-
-#[cfg(target_arch = "x86_64")]
-type RuntimePlatform = xenplatform::x86pv::X86PvPlatform;
-
-#[cfg(not(target_arch = "x86_64"))]
-type RuntimePlatform = xenplatform::unsupported::UnsupportedPlatform;
+use xenclient::tx::channel::ChannelDeviceConfig;
+use xenclient::{config::DomainConfig, XenClient};
+use xenplatform::domain::{
+    KernelFormat, PlatformDomainConfig, PlatformKernelConfig, PlatformOptions,
+    PlatformResourcesConfig,
+};
+use xenplatform::RuntimePlatformType;
 
 #[tokio::main]
 async fn main() -> Result<()> {
@@ -22,32 +22,31 @@ async fn main() -> Result<()> {
     }
     let kernel_image_path = args.get(1).expect("argument not specified");
     let initrd_path = args.get(2).expect("argument not specified");
-    let client = XenClient::new(0, RuntimePlatform::new()).await?;
-    let config = DomainConfig {
-        base: BaseDomainConfig {
+    let client = XenClient::new().await?;
+
+    let mut config = DomainConfig::new();
+    config.platform(PlatformDomainConfig {
         uuid: Uuid::new_v4(),
-            max_vcpus: 1,
-            target_vcpus: 1,
-            max_mem_mb: 512,
-            target_mem_mb: 512,
-            enable_iommu: true,
-            kernel: fs::read(&kernel_image_path).await?,
-            initrd: fs::read(&initrd_path).await?,
+        platform: RuntimePlatformType::Pv,
+        kernel: PlatformKernelConfig {
+            data: Arc::new(fs::read(&kernel_image_path).await?),
+            format: KernelFormat::ElfCompressed,
             cmdline: "earlyprintk=xen earlycon=xen console=hvc0 init=/init".to_string(),
-            owner_domid: 0,
+            initrd: Some(Arc::new(fs::read(&initrd_path).await?)),
         },
-        backend_domid: 0,
-        name: "xenclient-test".to_string(),
-        swap_console_backend: None,
-        disks: vec![],
-        channels: vec![],
-        vifs: vec![],
-        pcis: vec![],
-        filesystems: vec![],
-        extra_keys: vec![],
-        extra_rw_paths: vec![],
-    };
-    let created = client.create(&config).await?;
-    println!("created domain {}", created.domid);
+        resources: PlatformResourcesConfig {
+            max_vcpus: 1,
+            assigned_vcpus: 1,
+            max_memory_mb: 512,
+            assigned_memory_mb: 512,
+        },
+        options: PlatformOptions { iommu: true },
+    });
+    config.name("xenclient-test");
+    let mut channel = ChannelDeviceConfig::new();
+    channel.default_console().backend_initialized();
+    config.add_channel(channel);
+    let created = client.create(config).await?;
+    println!("created domain {}", created.platform.domid);
     Ok(())
 }

crates/xen/xenclient/examples/boot_speed.rs

@@ -0,0 +1,66 @@
+use std::sync::Arc;
+use std::{env, process};
+use tokio::fs;
+use uuid::Uuid;
+use xenclient::config::{DomainConfig, DomainResult};
+use xenclient::error::Result;
+use xenclient::tx::channel::ChannelDeviceConfig;
+use xenclient::XenClient;
+use xenplatform::domain::{
+    KernelFormat, PlatformDomainConfig, PlatformKernelConfig, PlatformOptions,
+    PlatformResourcesConfig,
+};
+use xenplatform::elfloader::ElfImageLoader;
+use xenplatform::RuntimePlatformType;
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    env_logger::init();
+
+    let args: Vec<String> = env::args().collect();
+    if args.len() != 2 {
+        println!("usage: boot-speed <kernel-image>");
+        process::exit(1);
+    }
+    let kernel_path = args.get(1).expect("argument not specified");
+    let kernel = Arc::new(fs::read(kernel_path).await?);
+    let kernel = ElfImageLoader::load(kernel)?.into_elf_bytes();
+    let client = XenClient::new().await?;
+
+    for i in 0..5u32 {
+        let start = std::time::Instant::now();
+        let domain = create_domain(&client, kernel.clone(), i).await?;
+        let end = std::time::Instant::now();
+        let duration = end - start;
+        println!("boot setup time: {:?}", duration);
+        client.destroy(domain.platform.domid).await?;
+    }
+    Ok(())
+}
+
+async fn create_domain(client: &XenClient, kernel: Arc<Vec<u8>>, i: u32) -> Result<DomainResult> {
+    let mut config = DomainConfig::new();
+    config.platform(PlatformDomainConfig {
+        uuid: Uuid::new_v4(),
+        platform: RuntimePlatformType::Pv,
+        kernel: PlatformKernelConfig {
+            data: kernel,
+            format: KernelFormat::ElfUncompressed,
+            cmdline: "earlyprintk=xen earlycon=xen console=hvc0 init=/init".to_string(),
+            initrd: None,
+        },
+        resources: PlatformResourcesConfig {
+            max_vcpus: 1,
+            assigned_vcpus: 1,
+            max_memory_mb: 512,
+            assigned_memory_mb: 512,
+        },
+        options: PlatformOptions { iommu: true },
+    });
+    config.name(format!("xenboot-{}", i));
+    config.start(false);
+    let mut channel = ChannelDeviceConfig::new();
+    channel.default_console().backend_initialized();
+    config.add_channel(channel);
+    client.create(config).await
+}

crates/xen/xenclient/src/config.rs

@@ -0,0 +1,185 @@
+use std::collections::HashMap;
+
+use xencall::XenCall;
+pub use xenplatform::domain::PlatformDomainConfig;
+use xenplatform::domain::PlatformDomainInfo;
+
+use crate::{
+    error::Result,
+    tx::{
+        channel::ChannelDeviceConfig,
+        fs9p::Fs9pDeviceConfig,
+        pci::PciRootDeviceConfig,
+        vbd::VbdDeviceConfig,
+        vif::VifDeviceConfig,
+        {BlockDeviceResult, DeviceResult},
+    },
+};
+
+pub struct DomainConfig {
+    platform: Option<PlatformDomainConfig>,
+    name: Option<String>,
+    backend_domid: u32,
+    channels: Vec<ChannelDeviceConfig>,
+    vifs: Vec<VifDeviceConfig>,
+    vbds: Vec<VbdDeviceConfig>,
+    fs9ps: Vec<Fs9pDeviceConfig>,
+    pci: Option<PciRootDeviceConfig>,
+    extra_keys: HashMap<String, String>,
+    extra_rw_paths: Vec<String>,
+    start: bool,
+}
+
+impl Default for DomainConfig {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl DomainConfig {
+    pub fn new() -> Self {
+        Self {
+            platform: None,
+            name: None,
+            backend_domid: 0,
+            channels: Vec::new(),
+            vifs: Vec::new(),
+            vbds: Vec::new(),
+            fs9ps: Vec::new(),
+            pci: None,
+            extra_keys: HashMap::new(),
+            extra_rw_paths: Vec::new(),
+            start: true,
+        }
+    }
+
+    pub fn platform(&mut self, platform: PlatformDomainConfig) -> &mut Self {
+        self.platform = Some(platform);
+        self
+    }
+
+    pub fn get_platform(&self) -> &Option<PlatformDomainConfig> {
+        &self.platform
+    }
+
+    pub fn name(&mut self, name: impl AsRef<str>) -> &mut Self {
+        self.name = Some(name.as_ref().to_string());
+        self
+    }
+
+    pub fn get_name(&self) -> &Option<String> {
+        &self.name
+    }
+
+    pub fn backend_domid(&mut self, backend_domid: u32) -> &mut Self {
+        self.backend_domid = backend_domid;
+        self
+    }
+
+    pub fn get_backend_domid(&self) -> u32 {
+        self.backend_domid
+    }
+
+    pub fn add_channel(&mut self, channel: ChannelDeviceConfig) -> &mut Self {
+        self.channels.push(channel);
+        self
+    }
+
+    pub fn get_channels(&self) -> &Vec<ChannelDeviceConfig> {
+        &self.channels
+    }
+
+    pub fn add_vif(&mut self, vif: VifDeviceConfig) -> &mut Self {
+        self.vifs.push(vif);
+        self
+    }
+
+    pub fn get_vifs(&self) -> &Vec<VifDeviceConfig> {
+        &self.vifs
+    }
+
+    pub fn add_vbd(&mut self, vbd: VbdDeviceConfig) -> &mut Self {
+        self.vbds.push(vbd);
+        self
+    }
+
+    pub fn get_vbds(&self) -> &Vec<VbdDeviceConfig> {
+        &self.vbds
+    }
+
+    pub fn add_fs9p(&mut self, fs9p: Fs9pDeviceConfig) -> &mut Self {
+        self.fs9ps.push(fs9p);
+        self
+    }
+
+    pub fn get_fs9ps(&self) -> &Vec<Fs9pDeviceConfig> {
+        &self.fs9ps
+    }
+
+    pub fn pci(&mut self, pci: PciRootDeviceConfig) -> &mut Self {
+        self.pci = Some(pci);
+        self
+    }
+
+    pub fn get_pci(&self) -> &Option<PciRootDeviceConfig> {
+        &self.pci
+    }
+
+    pub fn add_extra_key(&mut self, key: impl AsRef<str>, value: impl ToString) -> &mut Self {
+        self.extra_keys
+            .insert(key.as_ref().to_string(), value.to_string());
+        self
+    }
+
+    pub fn get_extra_keys(&self) -> &HashMap<String, String> {
+        &self.extra_keys
+    }
+
+    pub fn add_rw_path(&mut self, path: impl AsRef<str>) -> &mut Self {
+        self.extra_rw_paths.push(path.as_ref().to_string());
+        self
+    }
+
+    pub fn get_rw_paths(&self) -> &Vec<String> {
+        &self.extra_rw_paths
+    }
+
+    pub fn start(&mut self, start: bool) -> &mut Self {
+        self.start = start;
+        self
+    }
+
+    pub fn get_start(&self) -> bool {
+        self.start
+    }
+
+    pub fn done(self) -> Self {
+        self
+    }
+
+    pub(crate) async fn prepare(
+        &mut self,
+        domid: u32,
+        call: &XenCall,
+        platform: &PlatformDomainInfo,
+    ) -> Result<()> {
+        if let Some(pci) = self.pci.as_mut() {
+            pci.prepare(domid, call).await?;
+        }
+
+        for channel in &mut self.channels {
+            channel.prepare(platform).await?;
+        }
+
+        Ok(())
+    }
+}
+
+pub struct DomainResult {
+    pub platform: PlatformDomainInfo,
+    pub channels: Vec<DeviceResult>,
+    pub vifs: Vec<DeviceResult>,
+    pub vbds: Vec<BlockDeviceResult>,
+    pub fs9ps: Vec<DeviceResult>,
+    pub pci: Option<DeviceResult>,
+}

crates/xen/xenclient/src/devalloc.rs

@@ -0,0 +1,79 @@
+use bit_vec::BitVec;
+
+const DEVICE_COUNT: usize = 4096;
+const BYTE_COUNT: usize = DEVICE_COUNT / 8;
+
+pub struct DeviceIdAllocator {
+    states: BitVec,
+    cursor: u32,
+}
+
+impl Default for DeviceIdAllocator {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl DeviceIdAllocator {
+    pub fn new() -> Self {
+        Self {
+            states: BitVec::from_elem(DEVICE_COUNT, false),
+            cursor: 0,
+        }
+    }
+
+    pub fn deserialize(bytes: &[u8]) -> Option<Self> {
+        if bytes.len() != BYTE_COUNT + 4 {
+            return None;
+        }
+
+        let cursor = bytes[0] as u32
+            | ((bytes[1] as u32) << 8)
+            | ((bytes[2] as u32) << 16)
+            | ((bytes[3] as u32) << 24);
+        let slice = &bytes[4..BYTE_COUNT + 4];
+        if slice.len() != BYTE_COUNT {
+            return None;
+        }
+        let states = BitVec::from_bytes(slice);
+
+        Some(Self { states, cursor })
+    }
+
+    pub fn allocate(&mut self) -> Option<u32> {
+        let start = self.cursor;
+        loop {
+            let id = self.cursor;
+            let value = self.states.get(self.cursor as usize)?;
+
+            self.cursor = (self.cursor + 1) % DEVICE_COUNT as u32;
+
+            if !value {
+                self.states.set(id as usize, true);
+                return Some(id);
+            }
+
+            if self.cursor == start {
+                return None;
+            }
+        }
+    }
+
+    pub fn release(&mut self, id: u32) {
+        self.states.set(id as usize, false);
+    }
+
+    pub fn count_free(&mut self) -> u32 {
+        self.states.count_zeros() as u32
+    }
+
+    pub fn serialize(&mut self) -> Vec<u8> {
+        let mut bytes = Vec::with_capacity(BYTE_COUNT + 4);
+        bytes.push((self.cursor & 0xff) as u8);
+        bytes.push(((self.cursor >> 8) & 0xff) as u8);
+        bytes.push(((self.cursor >> 16) & 0xff) as u8);
+        bytes.push(((self.cursor >> 24) & 0xff) as u8);
+        bytes.extend_from_slice(&self.states.to_bytes());
+        bytes
+    }
+}

crates/xen/xenclient/src/devstate.rs

@@ -0,0 +1,131 @@
+use std::time::Duration;
+
+use tokio::{
+    select,
+    time::{sleep, timeout},
+};
+use xenstore::{XsdClient, XsdInterface};
+
+use crate::error::{Error, Result};
+
+pub struct DeviceLocator {
+    pub frontend_domid: u32,
+    pub backend_domid: u32,
+    pub frontend_type: String,
+    pub backend_type: String,
+    pub device_id: u64,
+}
+
+impl DeviceLocator {
+    pub fn new(
+        frontend_domid: u32,
+        backend_domid: u32,
+        frontend_type: String,
+        backend_type: String,
+        device_id: u64,
+    ) -> Self {
+        DeviceLocator {
+            frontend_domid,
+            backend_domid,
+            frontend_type,
+            backend_type,
+            device_id,
+        }
+    }
+
+    pub fn frontend_state_path(&self) -> String {
+        format!(
+            "/local/domain/{}/device/{}/{}/state",
+            self.frontend_domid, self.frontend_type, self.device_id
+        )
+    }
+
+    pub fn backend_state_path(&self) -> String {
+        format!(
+            "/local/domain/{}/backend/{}/{}/{}/state",
+            self.backend_domid, self.backend_type, self.frontend_domid, self.device_id
+        )
+    }
+}
+
+pub struct DeviceStateWaiter {
+    devices: Vec<DeviceLocator>,
+    xsd: XsdClient,
+}
+
+impl DeviceStateWaiter {
+    pub fn new(xsd: XsdClient) -> Self {
+        DeviceStateWaiter {
+            devices: vec![],
+            xsd,
+        }
+    }
+
+    pub fn add_device(&mut self, device: DeviceLocator) -> &mut DeviceStateWaiter {
+        self.devices.push(device);
+        self
+    }
+
+    async fn check_states(xsd: &XsdClient, state_paths: &[String], desired: u32) -> Result<bool> {
+        let mut ready = 0;
+        for state_path in state_paths {
+            let Some(state_text) = xsd.read_string(state_path).await? else {
+                return Err(Error::DevStateWaitError(format!(
+                    "state path '{}' did not exist",
+                    state_path
+                )));
+            };
+
+            let Some(state_value) = state_text.parse::<u32>().ok() else {
+                return Err(Error::DevStateWaitError(format!(
+                    "state path '{}' did not have a valid value",
+                    state_path
+                )));
+            };
+
+            if state_value > desired {
+                return Err(Error::DevStateWaitError(format!(
+                    "state path '{}' had a state of {} which is greater than {}",
+                    state_path, state_value, desired
+                )));
+            }
+
+            if state_value == desired {
+                ready += 1;
+            }
+        }
+        Ok(ready == state_paths.len())
+    }
+
+    async fn do_wait(self, desired: u32) -> Result<()> {
+        let mut watch = self.xsd.create_multi_watch().await?;
+        let mut state_paths = Vec::new();
+        for device in self.devices {
+            let state_path = device.backend_state_path();
+            self.xsd.bind_watch_id(watch.id, &state_path).await?;
+            state_paths.push(state_path);
+        }
+
+        loop {
+            if DeviceStateWaiter::check_states(&self.xsd, &state_paths, desired).await? {
+                break;
+            }
+
+            select! {
+                _update = watch.receiver.recv() => {},
+                _timeout = sleep(Duration::from_millis(250)) => {},
+            }
+        }
+        Ok(())
+    }
+
+    pub async fn wait(self, desired: u32, deadline: Duration) -> Result<()> {
+        if let Some(err) = timeout(deadline, self.do_wait(desired)).await.err() {
+            return Err(Error::DevStateWaitError(format!(
+                "took too long for devices to be ready: {}",
+                err
+            )));
+        }
+        Ok(())
+    }
+}

crates/xen/xenclient/src/error.rs

@@ -34,6 +34,8 @@ pub enum Error {
     RegexError(#[from] regex::Error),
     #[error("error: {0}")]
     GenericError(String),
+    #[error("parameter missing: {0}")]
+    ParameterMissing(&'static str),
     #[error("failed to parse int: {0}")]
     ParseIntError(#[from] std::num::ParseIntError),
     #[error("invalid pci bdf string")]
@@ -42,6 +44,12 @@ pub enum Error {
     PciDeviceNotAssignable(PciBdf),
     #[error("xen platform error: {0}")]
     XenPlatform(#[from] xenplatform::error::Error),
+    #[error("invalid block index")]
+    InvalidBlockIdx,
+    #[error("device state wait error: {0}")]
+    DevStateWaitError(String),
+    #[error("device ids exhausted")]
+    DevIdExhausted,
 }
 
 pub type Result<T> = std::result::Result<T, Error>;

crates/xen/xenclient/src/lib.rs

@@ -1,12 +1,11 @@
 pub mod error;
 
-use crate::error::{Error, Result};
+use config::{DomainConfig, DomainResult};
+use error::{Error, Result};
 use log::{debug, trace};
-use pci::PciBdf;
 use tokio::time::timeout;
-use tx::ClientTransaction;
-use xenplatform::boot::BootSetupPlatform;
-use xenplatform::domain::{BaseDomainConfig, BaseDomainManager, CreatedDomain};
+use tx::{DeviceConfig, XenTransaction};
+use xenplatform::domain::{PlatformDomainInfo, PlatformDomainManager};
 
 use std::path::PathBuf;
 use std::str::FromStr;
@@ -15,109 +14,26 @@ use std::time::Duration;
 use xencall::XenCall;
 use xenstore::{XsdClient, XsdInterface};
 
+pub mod config;
+pub mod devalloc;
+pub mod devstate;
 pub mod pci;
 pub mod tx;
+pub mod util;
 
 #[derive(Clone)]
-pub struct XenClient<P: BootSetupPlatform> {
+pub struct XenClient {
     pub store: XsdClient,
     pub call: XenCall,
-    domain_manager: Arc<BaseDomainManager<P>>,
-}
-
-#[derive(Clone, Debug)]
-pub struct BlockDeviceRef {
-    pub path: String,
-    pub major: u32,
-    pub minor: u32,
-}
-
-#[derive(Clone, Debug)]
-pub struct DomainDisk {
-    pub vdev: String,
-    pub block: BlockDeviceRef,
-    pub writable: bool,
-}
-
-#[derive(Clone, Debug)]
-pub struct DomainFilesystem {
-    pub path: String,
-    pub tag: String,
-}
-
-#[derive(Clone, Debug)]
-pub struct DomainNetworkInterface {
-    pub mac: String,
-    pub mtu: u32,
-    pub bridge: Option<String>,
-    pub script: Option<String>,
-}
-
-#[derive(Clone, Debug)]
-pub struct DomainChannel {
-    pub typ: String,
-    pub initialized: bool,
-}
-
-#[derive(Clone, Debug)]
-pub struct DomainEventChannel {
-    pub name: String,
-}
-
-#[derive(Clone, Debug, Default, Eq, PartialEq)]
-pub enum DomainPciRdmReservePolicy {
-    Invalid,
-    #[default]
-    Strict,
-    Relaxed,
-}
-
-impl DomainPciRdmReservePolicy {
-    pub fn to_option_str(&self) -> &str {
-        match self {
-            DomainPciRdmReservePolicy::Invalid => "-1",
-            DomainPciRdmReservePolicy::Strict => "0",
-            DomainPciRdmReservePolicy::Relaxed => "1",
-        }
-    }
-}
-
-#[derive(Clone, Debug)]
-pub struct DomainPciDevice {
-    pub bdf: PciBdf,
-    pub permissive: bool,
-    pub msi_translate: bool,
-    pub power_management: bool,
-    pub rdm_reserve_policy: DomainPciRdmReservePolicy,
-}
-
-#[derive(Clone, Debug)]
-pub struct DomainConfig {
-    pub base: BaseDomainConfig,
-    pub backend_domid: u32,
-    pub name: String,
-    pub disks: Vec<DomainDisk>,
-    pub swap_console_backend: Option<String>,
-    pub channels: Vec<DomainChannel>,
-    pub vifs: Vec<DomainNetworkInterface>,
-    pub filesystems: Vec<DomainFilesystem>,
-    pub pcis: Vec<DomainPciDevice>,
-    pub extra_keys: Vec<(String, String)>,
-    pub extra_rw_paths: Vec<String>,
-}
-
-#[derive(Debug)]
-pub struct CreatedChannel {
-    pub ring_ref: u64,
-    pub evtchn: u32,
+    domain_manager: Arc<PlatformDomainManager>,
 }
 
 #[allow(clippy::too_many_arguments)]
-impl<P: BootSetupPlatform> XenClient<P> {
-    pub async fn new(current_domid: u32, platform: P) -> Result<XenClient<P>> {
+impl XenClient {
+    pub async fn new() -> Result<XenClient> {
         let store = XsdClient::open().await?;
-        let call: XenCall = XenCall::open(current_domid)?;
-        let domain_manager = BaseDomainManager::new(call.clone(), platform).await?;
+        let call: XenCall = XenCall::open(0)?;
+        let domain_manager = PlatformDomainManager::new(call.clone()).await?;
         Ok(XenClient {
             store,
             call,
@@ -125,29 +41,47 @@ impl<P: BootSetupPlatform> XenClient<P> {
         })
     }
 
-    pub async fn create(&self, config: &DomainConfig) -> Result<CreatedDomain> {
-        let created = self.domain_manager.create(config.base.clone()).await?;
-        match self.init(created.domid, config, &created).await {
-            Ok(_) => Ok(created),
+    pub async fn create(&self, config: DomainConfig) -> Result<DomainResult> {
+        let platform = config
+            .get_platform()
+            .as_ref()
+            .ok_or_else(|| Error::ParameterMissing("platform"))?
+            .clone();
+        let platform = self.domain_manager.create(platform).await?;
+        match self.init(platform.domid, config, &platform).await {
+            Ok(result) => Ok(result),
             Err(err) => {
                 // ignore since destroying a domain is best-effort when an error occurs
-                let _ = self.domain_manager.destroy(created.domid).await;
+                let _ = self.domain_manager.destroy(platform.domid).await;
                 Err(err)
             }
         }
     }
 
-    pub async fn transaction(&self, domid: u32, backend_domid: u32) -> Result<ClientTransaction> {
-        ClientTransaction::new(&self.store, domid, backend_domid).await
+    pub async fn transaction(&self, domid: u32, backend_domid: u32) -> Result<XenTransaction> {
+        XenTransaction::new(&self.store, domid, backend_domid).await
     }
 
-    async fn init(&self, domid: u32, config: &DomainConfig, created: &CreatedDomain) -> Result<()> {
+    async fn init(
+        &self,
+        domid: u32,
+        mut config: DomainConfig,
+        created: &PlatformDomainInfo,
+    ) -> Result<DomainResult> {
         trace!("xenclient init domid={} domain={:?}", domid, created);
-        let transaction = self.transaction(domid, config.backend_domid).await?;
+        let platform_config = config
+            .get_platform()
+            .as_ref()
+            .ok_or_else(|| Error::ParameterMissing("platform"))?;
+        loop {
+            let transaction = self.transaction(domid, config.get_backend_domid()).await?;
             transaction
-            .add_domain_declaration(&config.name, &config.base, created)
+                .add_domain_declaration(config.get_name().clone(), platform_config, created)
                 .await?;
-        transaction.commit().await?;
+            if transaction.maybe_commit().await? {
+                break;
+            }
+        }
         if !self
             .store
             .introduce_domain(domid, created.store_mfn, created.store_evtchn)
@@ -155,57 +89,69 @@ impl<P: BootSetupPlatform> XenClient<P> {
         {
             return Err(Error::IntroduceDomainFailed);
         }
-        let transaction = self.transaction(domid, config.backend_domid).await?;
-        transaction
-            .add_channel_device(
-                created,
-                0,
-                &DomainChannel {
-                    typ: config
-                        .swap_console_backend
-                        .clone()
-                        .unwrap_or("xenconsoled".to_string())
-                        .to_string(),
-                    initialized: true,
-                },
-            )
-            .await?;
+        config.prepare(domid, &self.call, created).await?;
+        let mut channels;
+        let mut vifs;
+        let mut vbds;
+        let mut fs9ps;
+        let mut pci_result;
+        loop {
+            let transaction = self.transaction(domid, config.get_backend_domid()).await?;
 
-        for (index, channel) in config.channels.iter().enumerate() {
-            transaction
-                .add_channel_device(created, index + 1, channel)
-                .await?;
+            channels = Vec::new();
+            for channel in config.get_channels() {
+                let result = channel.add_to_transaction(&transaction).await?;
+                channels.push(result);
             }
 
-        for (index, disk) in config.disks.iter().enumerate() {
-            transaction.add_vbd_device(index, disk).await?;
+            vifs = Vec::new();
+            for vif in config.get_vifs() {
+                let result = vif.add_to_transaction(&transaction).await?;
+                vifs.push(result);
             }
 
-        for (index, filesystem) in config.filesystems.iter().enumerate() {
-            transaction.add_9pfs_device(index, filesystem).await?;
+            vbds = Vec::new();
+            for vbd in config.get_vbds() {
+                let result = vbd.add_to_transaction(&transaction).await?;
+                vbds.push(result);
             }
 
-        for (index, vif) in config.vifs.iter().enumerate() {
-            transaction.add_vif_device(index, vif).await?;
+            fs9ps = Vec::new();
+            for fs9p in config.get_fs9ps() {
+                let result = fs9p.add_to_transaction(&transaction).await?;
+                fs9ps.push(result);
             }
 
-        for (index, pci) in config.pcis.iter().enumerate() {
-            transaction
-                .add_pci_device(&self.call, index, config.pcis.len(), pci)
-                .await?;
+            pci_result = None;
+            if let Some(pci) = config.get_pci().as_ref() {
+                pci_result = Some(pci.add_to_transaction(&transaction).await?);
             }
 
-        for (key, value) in &config.extra_keys {
-            transaction.write_key(key, value).await?;
+            for (key, value) in config.get_extra_keys() {
+                transaction.write(key, value, None).await?;
             }
 
-        for key in &config.extra_rw_paths {
-            transaction.add_rw_path(key).await?;
+            for rw_path in config.get_rw_paths() {
+                transaction.add_rw_path(rw_path).await?;
             }
 
-        transaction.commit().await?;
+            if transaction.maybe_commit().await? {
+                break;
+            }
+        }
+
+        if config.get_start() {
             self.call.unpause_domain(domid).await?;
-        Ok(())
+        }
+
+        Ok(DomainResult {
+            platform: created.clone(),
+            channels,
+            vifs,
+            vbds,
+            fs9ps,
+            pci: pci_result,
+        })
     }
 
     pub async fn destroy(&self, domid: u32) -> Result<()> {
@@ -251,13 +197,37 @@ impl<P: BootSetupPlatform> XenClient<P> {
         }
 
         for backend in &backend_paths {
+            self.destroy_backend(backend).await?;
+        }
+
+        let tx = self.store.transaction().await?;
+        let mut backend_removals: Vec<String> = Vec::new();
+        backend_removals.extend_from_slice(backend_paths.as_slice());
+        if let Some(backend) = console_backend_path {
+            backend_removals.push(backend);
+        }
+        for path in &backend_removals {
+            let path = PathBuf::from(path);
+            let parent = path.parent().ok_or(Error::PathParentNotFound)?;
+            tx.rm(parent.to_str().ok_or(Error::PathStringConversion)?)
+                .await?;
+        }
+        if let Some(vm_path) = vm_path {
+            tx.rm(&vm_path).await?;
+        }
+        tx.rm(&dom_path).await?;
+        tx.commit().await?;
+        Ok(())
+    }
+
+    async fn destroy_backend(&self, backend: &str) -> Result<()> {
         let state_path = format!("{}/state", backend);
         let mut watch = self.store.create_watch(&state_path).await?;
         let online_path = format!("{}/online", backend);
         let tx = self.store.transaction().await?;
         let state = tx.read_string(&state_path).await?.unwrap_or(String::new());
         if state.is_empty() {
-                break;
+            return Ok(());
         }
         tx.write_string(&online_path, "0").await?;
         if !state.is_empty() && u32::from_str(&state).unwrap_or(0) != 6 {
@@ -284,25 +254,37 @@ impl<P: BootSetupPlatform> XenClient<P> {
             }
             count += 1;
         }
+        self.store.rm(backend).await?;
+        Ok(())
     }
 
-        let tx = self.store.transaction().await?;
-        let mut backend_removals: Vec<String> = Vec::new();
-        backend_removals.extend_from_slice(backend_paths.as_slice());
-        if let Some(backend) = console_backend_path {
-            backend_removals.push(backend);
+    pub async fn destroy_device(
+        &self,
+        category: &str,
+        domid: u32,
+        devid: u64,
+        blkid: Option<u32>,
+    ) -> Result<()> {
+        let dom_path = self.store.get_domain_path(domid).await?;
+        let device_path = format!("{}/device/{}/{}", dom_path, category, devid);
+        if let Some(backend_path) = self
+            .store
+            .read_string(format!("{}/backend", device_path).as_str())
+            .await?
+        {
+            self.destroy_backend(&backend_path).await?;
         }
-        for path in &backend_removals {
-            let path = PathBuf::from(path);
-            let parent = path.parent().ok_or(Error::PathParentNotFound)?;
-            tx.rm(parent.to_str().ok_or(Error::PathStringConversion)?)
-                .await?;
+        self.destroy_backend(&device_path).await?;
+        loop {
+            let tx = self.transaction(domid, 0).await?;
+            tx.release_devid(devid).await?;
+            if let Some(blkid) = blkid {
+                tx.release_blkid(blkid).await?;
+            }
+            if tx.maybe_commit().await? {
+                break;
             }
-        if let Some(vm_path) = vm_path {
-            tx.rm(&vm_path).await?;
         }
-        tx.rm(&dom_path).await?;
-        tx.commit().await?;
         Ok(())
     }
 }

crates/xen/xenclient/src/tx.rs

@@ -1,582 +0,0 @@
-use indexmap::IndexMap;
-use xencall::{sys::DOMCTL_DEV_RDM_RELAXED, XenCall};
-use xenplatform::{
-    domain::{BaseDomainConfig, CreatedDomain},
-    sys::XEN_PAGE_SHIFT,
-};
-use xenstore::{
-    XsPermission, XsdClient, XsdInterface, XsdTransaction, XS_PERM_NONE, XS_PERM_READ,
-    XS_PERM_READ_WRITE,
-};
-
-use crate::{
-    error::{Error, Result},
-    pci::XenPciBackend,
-    DomainChannel, DomainDisk, DomainFilesystem, DomainNetworkInterface, DomainPciDevice,
-    DomainPciRdmReservePolicy,
-};
-
-pub struct ClientTransaction {
-    tx: XsdTransaction,
-    abort: bool,
-    domid: u32,
-    dom_path: String,
-    backend_domid: u32,
-    backend_dom_path: String,
-}
-
-impl ClientTransaction {
-    pub async fn new(store: &XsdClient, domid: u32, backend_domid: u32) -> Result<Self> {
-        let backend_dom_path = store.get_domain_path(0).await?;
-        let dom_path = store.get_domain_path(domid).await?;
-        Ok(ClientTransaction {
-            tx: store.transaction().await?,
-            abort: true,
-            domid,
-            dom_path,
-            backend_domid,
-            backend_dom_path,
-        })
-    }
-
-    pub async fn add_domain_declaration(
-        &self,
-        name: impl AsRef<str>,
-        base: &BaseDomainConfig,
-        created: &CreatedDomain,
-    ) -> Result<()> {
-        let vm_path = format!("/vm/{}", base.uuid);
-        let ro_perm = &[
-            XsPermission {
-                id: 0,
-                perms: XS_PERM_NONE,
-            },
-            XsPermission {
-                id: self.domid,
-                perms: XS_PERM_READ,
-            },
-        ];
-
-        let no_perm = &[XsPermission {
-            id: 0,
-            perms: XS_PERM_NONE,
-        }];
-
-        let rw_perm = &[XsPermission {
-            id: self.domid,
-            perms: XS_PERM_READ_WRITE,
-        }];
-
-        self.tx.rm(&self.dom_path).await?;
-        self.tx.mknod(&self.dom_path, ro_perm).await?;
-
-        self.tx.rm(&vm_path).await?;
-        self.tx.mknod(&vm_path, ro_perm).await?;
-
-        self.tx.mknod(&vm_path, no_perm).await?;
-        self.tx
-            .mknod(format!("{}/device", vm_path).as_str(), no_perm)
-            .await?;
-
-        self.tx
-            .write_string(format!("{}/vm", self.dom_path).as_str(), &vm_path)
-            .await?;
-
-        self.tx
-            .mknod(format!("{}/cpu", self.dom_path).as_str(), ro_perm)
-            .await?;
-        self.tx
-            .mknod(format!("{}/memory", self.dom_path).as_str(), ro_perm)
-            .await?;
-
-        self.tx
-            .mknod(format!("{}/control", self.dom_path).as_str(), ro_perm)
-            .await?;
-
-        self.tx
-            .mknod(
-                format!("{}/control/shutdown", self.dom_path).as_str(),
-                rw_perm,
-            )
-            .await?;
-        self.tx
-            .mknod(
-                format!("{}/control/feature-poweroff", self.dom_path).as_str(),
-                rw_perm,
-            )
-            .await?;
-        self.tx
-            .mknod(
-                format!("{}/control/feature-reboot", self.dom_path).as_str(),
-                rw_perm,
-            )
-            .await?;
-        self.tx
-            .mknod(
-                format!("{}/control/feature-suspend", self.dom_path).as_str(),
-                rw_perm,
-            )
-            .await?;
-        self.tx
-            .mknod(format!("{}/control/sysrq", self.dom_path).as_str(), rw_perm)
-            .await?;
-
-        self.tx
-            .mknod(format!("{}/data", self.dom_path).as_str(), rw_perm)
-            .await?;
-        self.tx
-            .mknod(format!("{}/drivers", self.dom_path).as_str(), rw_perm)
-            .await?;
-        self.tx
-            .mknod(format!("{}/feature", self.dom_path).as_str(), rw_perm)
-            .await?;
-        self.tx
-            .mknod(format!("{}/attr", self.dom_path).as_str(), rw_perm)
-            .await?;
-        self.tx
-            .mknod(format!("{}/error", self.dom_path).as_str(), rw_perm)
-            .await?;
-
-        self.tx
-            .write_string(format!("{}/uuid", vm_path).as_str(), &base.uuid.to_string())
-            .await?;
-        self.tx
-            .write_string(format!("{}/name", self.dom_path).as_str(), name.as_ref())
-            .await?;
-        self.tx
-            .write_string(format!("{}/name", vm_path).as_str(), name.as_ref())
-            .await?;
-
-        self.tx
-            .write_string(format!("{}/image/os_type", vm_path).as_str(), "linux")
-            .await?;
-        self.tx
-            .write_string(format!("{}/image/cmdline", vm_path).as_str(), &base.cmdline)
-            .await?;
-        self.tx
-            .write_string(
-                format!("{}/memory/static-max", self.dom_path).as_str(),
-                &(base.max_mem_mb * 1024).to_string(),
-            )
-            .await?;
-        self.tx
-            .write_string(
-                format!("{}/memory/target", self.dom_path).as_str(),
-                &(base.target_mem_mb * 1024).to_string(),
-            )
-            .await?;
-        self.tx
-            .write_string(format!("{}/memory/videoram", self.dom_path).as_str(), "0")
-            .await?;
-        self.tx
-            .write_string(
-                format!("{}/domid", self.dom_path).as_str(),
-                &created.domid.to_string(),
-            )
-            .await?;
-        self.tx
-            .write_string(format!("{}/type", self.dom_path).as_str(), "PV")
-            .await?;
-        self.tx
-            .write_string(
-                format!("{}/store/port", self.dom_path).as_str(),
-                &created.store_evtchn.to_string(),
-            )
-            .await?;
-        self.tx
-            .write_string(
-                format!("{}/store/ring-ref", self.dom_path).as_str(),
-                &created.store_mfn.to_string(),
-            )
-            .await?;
-        for i in 0..base.max_vcpus {
-            let path = format!("{}/cpu/{}", self.dom_path, i);
-            self.tx.mkdir(&path).await?;
-            self.tx.set_perms(&path, ro_perm).await?;
-            let path = format!("{}/cpu/{}/availability", self.dom_path, i);
-            self.tx
-                .write_string(
-                    &path,
-                    if i < base.target_vcpus {
-                        "online"
-                    } else {
-                        "offline"
-                    },
-                )
-                .await?;
-            self.tx.set_perms(&path, ro_perm).await?;
-        }
-        Ok(())
-    }
-
-    pub async fn write_key(&self, key: impl AsRef<str>, value: impl AsRef<str>) -> Result<()> {
-        self.tx
-            .write_string(
-                &format!("{}/{}", self.dom_path, key.as_ref()),
-                value.as_ref(),
-            )
-            .await?;
-        Ok(())
-    }
-
-    pub async fn add_rw_path(&self, key: impl AsRef<str>) -> Result<()> {
-        let rw_perm = &[XsPermission {
-            id: self.domid,
-            perms: XS_PERM_READ_WRITE,
-        }];
-
-        self.tx
-            .mknod(&format!("{}/{}", self.dom_path, key.as_ref()), rw_perm)
-            .await?;
-        Ok(())
-    }
-
-    pub async fn add_device(
-        &self,
-        typ: impl AsRef<str>,
-        id: u64,
-        frontend_items: Vec<(&str, String)>,
-        backend_items: Vec<(&str, String)>,
-    ) -> Result<()> {
-        let console_zero = typ.as_ref() == "console" && id == 0;
-
-        let frontend_path = if console_zero {
-            format!("{}/console", self.dom_path)
-        } else {
-            format!("{}/device/{}/{}", self.dom_path, typ.as_ref(), id)
-        };
-        let backend_path = format!(
-            "{}/backend/{}/{}/{}",
-            self.backend_dom_path,
-            typ.as_ref(),
-            self.domid,
-            id
-        );
-
-        let mut backend_items: Vec<(&str, String)> = backend_items.clone();
-        let mut frontend_items: Vec<(&str, String)> = frontend_items.clone();
-        backend_items.push(("frontend", frontend_path.clone()));
-        frontend_items.push(("backend", backend_path.clone()));
-        let frontend_perms = &[
-            XsPermission {
-                id: self.domid,
-                perms: XS_PERM_NONE,
-            },
-            XsPermission {
-                id: self.backend_domid,
-                perms: XS_PERM_READ,
-            },
-        ];
-
-        let backend_perms = &[
-            XsPermission {
-                id: self.backend_domid,
-                perms: XS_PERM_NONE,
-            },
-            XsPermission {
-                id: self.domid,
-                perms: XS_PERM_READ,
-            },
-        ];
-
-        self.tx.mknod(&frontend_path, frontend_perms).await?;
-        for (p, value) in &frontend_items {
-            let path = format!("{}/{}", frontend_path, *p);
-            self.tx.write_string(&path, value).await?;
-            if !console_zero {
-                self.tx.set_perms(&path, frontend_perms).await?;
-            }
-        }
-        self.tx.mknod(&backend_path, backend_perms).await?;
-        for (p, value) in &backend_items {
-            let path = format!("{}/{}", backend_path, *p);
-            self.tx.write_string(&path, value).await?;
-        }
-        Ok(())
-    }
-
-    pub async fn add_vbd_device(&self, index: usize, disk: &DomainDisk) -> Result<()> {
-        let id = (202 << 8) | (index << 4) as u64;
-        let backend_items: Vec<(&str, String)> = vec![
-            ("frontend-id", self.domid.to_string()),
-            ("online", "1".to_string()),
-            ("removable", "0".to_string()),
-            ("bootable", "1".to_string()),
-            ("state", "1".to_string()),
-            ("dev", disk.vdev.to_string()),
-            ("type", "phy".to_string()),
-            ("mode", if disk.writable { "w" } else { "r" }.to_string()),
-            ("device-type", "disk".to_string()),
-            ("discard-enable", "0".to_string()),
-            ("specification", "xen".to_string()),
-            ("physical-device-path", disk.block.path.to_string()),
-            (
-                "physical-device",
-                format!("{:02x}:{:02x}", disk.block.major, disk.block.minor),
-            ),
-        ];
-
-        let frontend_items: Vec<(&str, String)> = vec![
-            ("backend-id", self.backend_domid.to_string()),
-            ("state", "1".to_string()),
-            ("virtual-device", id.to_string()),
-            ("device-type", "disk".to_string()),
-            ("trusted", "1".to_string()),
-            ("protocol", "x86_64-abi".to_string()),
-        ];
-
-        self.add_device("vbd", id, frontend_items, backend_items)
-            .await?;
-        Ok(())
-    }
-
-    pub async fn add_vif_device(&self, index: usize, vif: &DomainNetworkInterface) -> Result<()> {
-        let id = 20 + index as u64;
-        let mut backend_items: Vec<(&str, String)> = vec![
-            ("frontend-id", self.domid.to_string()),
-            ("online", "1".to_string()),
-            ("state", "1".to_string()),
-            ("mac", vif.mac.to_string()),
-            ("mtu", vif.mtu.to_string()),
-            ("type", "vif".to_string()),
-            ("handle", id.to_string()),
-        ];
-
-        if vif.bridge.is_some() {
-            backend_items.extend_from_slice(&[("bridge", vif.bridge.clone().unwrap())]);
-        }
-
-        if vif.script.is_some() {
-            backend_items.extend_from_slice(&[
-                ("script", vif.script.clone().unwrap()),
-                ("hotplug-status", "".to_string()),
-            ]);
-        } else {
-            backend_items.extend_from_slice(&[
-                ("script", "".to_string()),
-                ("hotplug-status", "connected".to_string()),
-            ]);
-        }
-
-        let frontend_items: Vec<(&str, String)> = vec![
-            ("backend-id", self.backend_domid.to_string()),
-            ("state", "1".to_string()),
-            ("mac", vif.mac.to_string()),
-            ("trusted", "1".to_string()),
-            ("mtu", vif.mtu.to_string()),
-        ];
-
-        self.add_device("vif", id, frontend_items, backend_items)
-            .await?;
-        Ok(())
-    }
-
-    pub async fn add_9pfs_device(&self, index: usize, filesystem: &DomainFilesystem) -> Result<()> {
-        let id = 90 + index as u64;
-        let backend_items: Vec<(&str, String)> = vec![
-            ("frontend-id", self.domid.to_string()),
-            ("online", "1".to_string()),
-            ("state", "1".to_string()),
-            ("path", filesystem.path.to_string()),
-            ("security-model", "none".to_string()),
-        ];
-
-        let frontend_items: Vec<(&str, String)> = vec![
-            ("backend-id", self.backend_domid.to_string()),
-            ("state", "1".to_string()),
-            ("tag", filesystem.tag.to_string()),
-        ];
-
-        self.add_device("9pfs", id, frontend_items, backend_items)
-            .await?;
-        Ok(())
-    }
-
-    pub async fn add_channel_device(
-        &self,
-        domain: &CreatedDomain,
-        index: usize,
-        channel: &DomainChannel,
-    ) -> Result<()> {
-        let port = domain.console_evtchn;
-        let ring = domain.console_mfn;
-
-        let mut backend_items = vec![
-            ("frontend-id", self.domid.to_string()),
-            ("online", "1".to_string()),
-            ("protocol", "vt100".to_string()),
-        ];
-
-        let mut frontend_items = vec![
-            ("backend-id", self.backend_domid.to_string()),
-            ("limit", "1048576".to_string()),
-            ("output", "pty".to_string()),
-            ("tty", "".to_string()),
-        ];
-
-        frontend_items.push(("type", channel.typ.clone()));
-        backend_items.push(("type", channel.typ.clone()));
-
-        if index == 0 {
-            if channel.typ != "xenconsoled" {
-                frontend_items.push(("state", "1".to_string()));
-            }
-
-            frontend_items
-                .extend_from_slice(&[("port", port.to_string()), ("ring-ref", ring.to_string())]);
-        } else {
-            frontend_items.extend_from_slice(&[
-                ("state", "1".to_string()),
-                ("protocol", "vt100".to_string()),
-            ]);
-        }
-
-        if channel.initialized {
-            backend_items.push(("state", "4".to_string()));
-        } else {
-            backend_items.push(("state", "1".to_string()));
-        }
-
-        self.add_device("console", index as u64, frontend_items, backend_items)
-            .await?;
-        Ok(())
-    }
-
-    pub async fn add_pci_device(
-        &self,
-        call: &XenCall,
-        index: usize,
-        device_count: usize,
-        device: &DomainPciDevice,
-    ) -> Result<()> {
-        let backend = XenPciBackend::new();
-        if !backend.is_assigned(&device.bdf).await? {
-            return Err(Error::PciDeviceNotAssignable(device.bdf));
-        }
-        let resources = backend.read_resources(&device.bdf).await?;
-        for resource in resources {
-            if resource.is_bar_io() {
-                call.ioport_permission(
-                    self.domid,
-                    resource.start as u32,
-                    resource.size() as u32,
-                    true,
-                )
-                .await?;
-            } else {
-                call.iomem_permission(
-                    self.domid,
-                    resource.start >> XEN_PAGE_SHIFT,
-                    (resource.size() + (XEN_PAGE_SHIFT - 1)) >> XEN_PAGE_SHIFT,
-                    true,
-                )
-                .await?;
-            }
-        }
-
-        if let Some(irq) = backend.read_irq(&device.bdf).await? {
-            let irq = call.map_pirq(self.domid, irq as isize, None).await?;
-            call.irq_permission(self.domid, irq, true).await?;
-        }
-
-        backend.reset(&device.bdf).await?;
-
-        call.assign_device(
-            self.domid,
-            device.bdf.encode(),
-            if device.rdm_reserve_policy == DomainPciRdmReservePolicy::Relaxed {
-                DOMCTL_DEV_RDM_RELAXED
-            } else {
-                0
-            },
-        )
-        .await?;
-
-        if device.permissive {
-            backend.enable_permissive(&device.bdf).await?;
-        }
-
-        let id = 60;
-
-        if index == 0 {
-            let backend_items: Vec<(&str, String)> = vec![
-                ("frontend-id", self.domid.to_string()),
-                ("online", "1".to_string()),
-                ("state", "1".to_string()),
-                ("num_devs", device_count.to_string()),
-            ];
-
-            let frontend_items: Vec<(&str, String)> = vec![
-                ("backend-id", self.backend_domid.to_string()),
-                ("state", "1".to_string()),
-            ];
-
-            self.add_device("pci", id, frontend_items, backend_items)
-                .await?;
-        }
-
-        let backend_path = format!(
-            "{}/backend/{}/{}/{}",
-            self.backend_dom_path, "pci", self.domid, id
-        );
-
-        self.tx
-            .write_string(
-                format!("{}/key-{}", backend_path, index),
-                &device.bdf.to_string(),
-            )
-            .await?;
-        self.tx
-            .write_string(
-                format!("{}/dev-{}", backend_path, index),
-                &device.bdf.to_string(),
-            )
-            .await?;
-
-        if let Some(vdefn) = device.bdf.vdefn {
-            self.tx
-                .write_string(
-                    format!("{}/vdefn-{}", backend_path, index),
-                    &format!("{:#x}", vdefn),
-                )
-                .await?;
-        }
-
-        let mut options = IndexMap::new();
-        options.insert("permissive", if device.permissive { "1" } else { "0" });
-        options.insert("rdm_policy", device.rdm_reserve_policy.to_option_str());
-        options.insert("msitranslate", if device.msi_translate { "1" } else { "0" });
-        options.insert(
-            "power_mgmt",
-            if device.power_management { "1" } else { "0" },
-        );
-        let options = options
-            .into_iter()
-            .map(|(key, value)| format!("{}={}", key, value))
-            .collect::<Vec<_>>()
-            .join(",");
-
-        self.tx
-            .write_string(format!("{}/opts-{}", backend_path, index), &options)
-            .await?;
-        Ok(())
-    }
-
-    pub async fn commit(mut self) -> Result<()> {
-        self.abort = false;
-        self.tx.commit().await?;
-        Ok(())
-    }
-}
-
-impl Drop for ClientTransaction {
-    fn drop(&mut self) {
-        if !self.abort {
-            return;
-        }
-        let tx = self.tx.clone();
-        tokio::task::spawn(async move {
-            let _ = tx.abort().await;
-        });
-    }
-}

crates/xen/xenclient/src/tx/channel.rs

@@ -0,0 +1,89 @@
+use xenplatform::domain::PlatformDomainInfo;
+
+use super::{DeviceConfig, DeviceDescription, DeviceResult, XenTransaction};
+use crate::error::{Error, Result};
+
+pub struct ChannelDeviceConfig {
+    backend_type: String,
+    default_console: bool,
+    default_console_options: Option<(u32, u64)>,
+    backend_initialized: bool,
+}
+
+impl Default for ChannelDeviceConfig {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl ChannelDeviceConfig {
+    pub fn new() -> Self {
+        Self {
+            backend_type: "console".to_string(),
+            default_console: false,
+            default_console_options: None,
+            backend_initialized: false,
+        }
+    }
+
+    pub fn backend_type(&mut self, backend_type: impl AsRef<str>) -> &mut Self {
+        self.backend_type = backend_type.as_ref().to_string();
+        self
+    }
+
+    pub fn default_console(&mut self) -> &mut Self {
+        self.default_console = true;
+        self
+    }
+
+    pub fn backend_initialized(&mut self) -> &mut Self {
+        self.backend_initialized = true;
+        self
+    }
+
+    pub fn done(self) -> Self {
+        self
+    }
+
+    pub async fn prepare(&mut self, platform: &PlatformDomainInfo) -> Result<()> {
+        if self.default_console {
+            self.default_console_options = Some((platform.console_evtchn, platform.console_mfn));
+        }
+        Ok(())
+    }
+}
+
+#[async_trait::async_trait]
+impl DeviceConfig for ChannelDeviceConfig {
+    type Result = DeviceResult;
+
+    async fn add_to_transaction(&self, tx: &XenTransaction) -> Result<DeviceResult> {
+        let id = tx.assign_next_devid().await?;
+        let mut device = DeviceDescription::new("console", &self.backend_type);
+        device
+            .add_backend_bool("online", true)
+            .add_backend_item("protocol", "vt100")
+            .add_backend_item("type", &self.backend_type)
+            .add_backend_item("state", if self.backend_initialized { 4 } else { 1 });
+
+        if self.default_console {
+            device.special_frontend_path("console");
+            let (port, ring_ref) = self
+                .default_console_options
+                .as_ref()
+                .ok_or_else(|| Error::ParameterMissing("default_console_options"))?;
+            device
+                .add_frontend_item("port", port)
+                .add_frontend_item("ring-ref", ring_ref);
+        }
+
+        device
+            .add_frontend_item("limit", 1048576)
+            .add_frontend_item("output", "pty")
+            .add_frontend_item("tty", "")
+            .add_frontend_item("type", &self.backend_type)
+            .add_frontend_item("state", 1);
+        tx.add_device(id, device).await?;
+        Ok(DeviceResult { id })
+    }
+}

crates/xen/xenclient/src/tx/fs9p.rs

@@ -0,0 +1,78 @@
+use super::{DeviceConfig, DeviceDescription, DeviceResult, XenTransaction};
+use crate::error::{Error, Result};
+
+pub struct Fs9pDeviceConfig {
+    backend_type: String,
+    security_model: String,
+    path: Option<String>,
+    tag: Option<String>,
+}
+
+impl Default for Fs9pDeviceConfig {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl Fs9pDeviceConfig {
+    pub fn new() -> Self {
+        Self {
+            backend_type: "9pfs".to_string(),
+            security_model: "none".to_string(),
+            path: None,
+            tag: None,
+        }
+    }
+
+    pub fn backend_type(&mut self, backend_type: impl AsRef<str>) -> &mut Self {
+        self.backend_type = backend_type.as_ref().to_string();
+        self
+    }
+
+    pub fn security_model(&mut self, security_model: impl AsRef<str>) -> &mut Self {
+        self.security_model = security_model.as_ref().to_string();
+        self
+    }
+
+    pub fn path(&mut self, path: impl AsRef<str>) -> &mut Self {
+        self.path = Some(path.as_ref().to_string());
+        self
+    }
+
+    pub fn tag(&mut self, tag: impl AsRef<str>) -> &mut Self {
+        self.tag = Some(tag.as_ref().to_string());
+        self
+    }
+
+    pub fn done(self) -> Self {
+        self
+    }
+}
+
+#[async_trait::async_trait]
+impl DeviceConfig for Fs9pDeviceConfig {
+    type Result = DeviceResult;
+
+    async fn add_to_transaction(&self, tx: &XenTransaction) -> Result<DeviceResult> {
+        let id = tx.assign_next_devid().await?;
+        let path = self
+            .path
+            .as_ref()
+            .ok_or_else(|| Error::ParameterMissing("path"))?;
+        let tag = self
+            .tag
+            .as_ref()
+            .ok_or_else(|| Error::ParameterMissing("tag"))?;
+        let mut device = DeviceDescription::new("9pfs", &self.backend_type);
+        device
+            .add_backend_bool("online", true)
+            .add_backend_item("state", 1)
+            .add_backend_item("path", path)
+            .add_backend_item("security_model", &self.security_model);
+        device
+            .add_frontend_item("state", 1)
+            .add_frontend_item("tag", tag);
+        tx.add_device(id, device).await?;
+        Ok(DeviceResult { id })
+    }
+}

crates/xen/xenclient/src/tx/mod.rs

@@ -0,0 +1,425 @@
+pub mod channel;
+pub mod fs9p;
+pub mod pci;
+pub mod vbd;
+pub mod vif;
+
+use crate::{
+    devalloc::DeviceIdAllocator,
+    error::{Error, Result},
+};
+use std::{collections::HashMap, sync::Arc};
+use tokio::sync::Mutex;
+use xenplatform::domain::{PlatformDomainConfig, PlatformDomainInfo};
+use xenstore::{
+    XsPermission, XsdClient, XsdInterface, XsdTransaction, XS_PERM_NONE, XS_PERM_READ,
+    XS_PERM_READ_WRITE,
+};
+
+pub struct XenTransaction {
+    frontend_domid: u32,
+    frontend_dom_path: String,
+    backend_domid: u32,
+    backend_dom_path: String,
+    blkalloc: Arc<Mutex<DeviceIdAllocator>>,
+    devalloc: Arc<Mutex<DeviceIdAllocator>>,
+    tx: XsdTransaction,
+    abort: bool,
+}
+
+impl XenTransaction {
+    pub async fn new(store: &XsdClient, frontend_domid: u32, backend_domid: u32) -> Result<Self> {
+        let frontend_dom_path = store.get_domain_path(frontend_domid).await?;
+        let backend_dom_path = store.get_domain_path(backend_domid).await?;
+        let tx = store.transaction().await?;
+
+        let devalloc = XenTransaction::load_id_allocator(&tx, "devid", &frontend_dom_path).await?;
+        let blkalloc = XenTransaction::load_id_allocator(&tx, "blkid", &frontend_dom_path).await?;
+
+        Ok(XenTransaction {
+            frontend_domid,
+            frontend_dom_path,
+            backend_domid,
+            backend_dom_path,
+            tx,
+            devalloc: Arc::new(Mutex::new(devalloc)),
+            blkalloc: Arc::new(Mutex::new(blkalloc)),
+            abort: true,
+        })
+    }
+
+    async fn load_id_allocator(
+        tx: &XsdTransaction,
+        allocator_type: &str,
+        frontend_dom_path: &str,
+    ) -> Result<DeviceIdAllocator> {
+        let state = tx
+            .read(format!(
+                "{}/{}-alloc-state",
+                frontend_dom_path, allocator_type
+            ))
+            .await?;
+        let allocator = state
+            .and_then(|state| DeviceIdAllocator::deserialize(&state))
+            .unwrap_or_else(DeviceIdAllocator::new);
+        Ok(allocator)
+    }
+
+    pub async fn assign_next_devid(&self) -> Result<u64> {
+        self.devalloc
+            .lock()
+            .await
+            .allocate()
+            .ok_or(Error::DevIdExhausted)
+            .map(|x| x as u64)
+    }
+
+    pub async fn assign_next_blkidx(&self) -> Result<u32> {
+        self.blkalloc
+            .lock()
+            .await
+            .allocate()
+            .ok_or(Error::DevIdExhausted)
+    }
+
+    pub async fn release_devid(&self, devid: u64) -> Result<()> {
+        self.devalloc.lock().await.release(devid as u32);
+        Ok(())
+    }
+
+    pub async fn release_blkid(&self, blkid: u32) -> Result<()> {
+        self.blkalloc.lock().await.release(blkid);
+        Ok(())
+    }
+
+    pub async fn write(
+        &self,
+        key: impl AsRef<str>,
+        value: impl AsRef<str>,
+        perms: Option<&[XsPermission]>,
+    ) -> Result<()> {
+        let path = format!("{}/{}", self.frontend_dom_path, key.as_ref());
+        if let Some(perms) = perms {
+            self.tx.mknod(&path, perms).await?;
+        }
+
+        // empty string is written by mknod, if perms is set we can skip it.
+        if perms.is_none() || perms.is_some() && !value.as_ref().is_empty() {
+            self.tx.write_string(path, value.as_ref()).await?;
+        }
+        Ok(())
+    }
+
+    pub async fn add_domain_declaration(
+        &self,
+        name: Option<impl AsRef<str>>,
+        platform: &PlatformDomainConfig,
+        created: &PlatformDomainInfo,
+    ) -> Result<()> {
+        let vm_path = format!("/vm/{}", platform.uuid);
+        let ro_perm = &[
+            XsPermission {
+                id: 0,
+                perms: XS_PERM_NONE,
+            },
+            XsPermission {
+                id: self.frontend_domid,
+                perms: XS_PERM_READ,
+            },
+        ];
+
+        let no_perm = &[XsPermission {
+            id: 0,
+            perms: XS_PERM_NONE,
+        }];
+
+        let rw_perm = &[XsPermission {
+            id: self.frontend_domid,
+            perms: XS_PERM_READ_WRITE,
+        }];
+
+        self.tx.rm(&self.frontend_dom_path).await?;
+        self.tx.mknod(&self.frontend_dom_path, ro_perm).await?;
+
+        self.tx.rm(&vm_path).await?;
+        self.tx.mknod(&vm_path, no_perm).await?;
+        self.tx
+            .write_string(format!("{}/uuid", vm_path), &platform.uuid.to_string())
+            .await?;
+
+        self.write("vm", &vm_path, None).await?;
+        self.write("cpu", "", Some(ro_perm)).await?;
+        self.write("memory", "", Some(ro_perm)).await?;
+        self.write("control", "", Some(ro_perm)).await?;
+        self.write("control/shutdown", "", Some(rw_perm)).await?;
+        self.write("control/feature-poweroff", "", Some(rw_perm))
+            .await?;
+        self.write("control/feature-reboot", "", Some(rw_perm))
+            .await?;
+        self.write("control/feature-suspend", "", Some(rw_perm))
+            .await?;
+        self.write("control/sysrq", "", Some(rw_perm)).await?;
+        self.write("data", "", Some(rw_perm)).await?;
+        self.write("drivers", "", Some(rw_perm)).await?;
+        self.write("feature", "", Some(rw_perm)).await?;
+        self.write("attr", "", Some(rw_perm)).await?;
+        self.write("error", "", Some(rw_perm)).await?;
+        self.write("uuid", platform.uuid.to_string(), Some(ro_perm))
+            .await?;
+        if let Some(name) = name {
+            self.write("name", name.as_ref(), Some(ro_perm)).await?;
+        }
+        self.write(
+            "memory/static-max",
+            (platform.resources.max_memory_mb * 1024).to_string(),
+            None,
+        )
+        .await?;
+        self.write(
+            "memory/target",
+            (platform.resources.assigned_memory_mb * 1024).to_string(),
+            None,
+        )
+        .await?;
+        self.write("memory/videoram", "0", None).await?;
+        self.write("domid", self.frontend_domid.to_string(), None)
+            .await?;
+        self.write("type", "PV", None).await?;
+        self.write("store/port", created.store_evtchn.to_string(), None)
+            .await?;
+        self.write("store/ring-ref", created.store_mfn.to_string(), None)
+            .await?;
+        for i in 0..platform.resources.max_vcpus {
+            let path = format!("{}/cpu/{}", self.frontend_dom_path, i);
+            self.tx.mkdir(&path).await?;
+            self.tx.set_perms(&path, ro_perm).await?;
+            let path = format!("{}/cpu/{}/availability", self.frontend_dom_path, i);
+            self.tx
+                .write_string(
+                    &path,
+                    if i < platform.resources.assigned_vcpus {
+                        "online"
+                    } else {
+                        "offline"
+                    },
+                )
+                .await?;
+            self.tx.set_perms(&path, ro_perm).await?;
+        }
+        Ok(())
+    }
+
+    pub async fn add_device(&self, id: u64, device: DeviceDescription) -> Result<()> {
+        let frontend_path = if let Some(ref special_frontend_path) = device.special_frontend_path {
+            format!("{}/{}", self.frontend_dom_path, special_frontend_path)
+        } else {
+            format!(
+                "{}/device/{}/{}",
+                self.frontend_dom_path, device.frontend_type, id
+            )
+        };
+        let backend_path = format!(
+            "{}/backend/{}/{}/{}",
+            self.backend_dom_path, device.backend_type, self.frontend_domid, id
+        );
+
+        let frontend_perms = &[
+            XsPermission {
+                id: self.frontend_domid,
+                perms: XS_PERM_READ_WRITE,
+            },
+            XsPermission {
+                id: self.backend_domid,
+                perms: XS_PERM_READ,
+            },
+        ];
+
+        let backend_perms = &[
+            XsPermission {
+                id: self.backend_domid,
+                perms: XS_PERM_READ_WRITE,
+            },
+            XsPermission {
+                id: self.frontend_domid,
+                perms: XS_PERM_READ,
+            },
+        ];
+
+        self.tx.mknod(&frontend_path, frontend_perms).await?;
+        self.tx.mknod(&backend_path, backend_perms).await?;
+
+        for (key, value) in &device.backend_items {
+            let path = format!("{}/{}", backend_path, key);
+            self.tx.write_string(&path, value).await?;
+        }
+
+        self.tx
+            .write_string(format!("{}/frontend", backend_path), &frontend_path)
+            .await?;
+        self.tx
+            .write_string(
+                format!("{}/frontend-id", backend_path),
+                &self.frontend_domid.to_string(),
+            )
+            .await?;
+        for (key, value) in &device.frontend_items {
+            let path = format!("{}/{}", frontend_path, key);
+            self.tx.write_string(&path, value).await?;
+            if device.special_frontend_path.is_none() {
+                self.tx.set_perms(&path, frontend_perms).await?;
+            }
+        }
+        self.tx
+            .write_string(format!("{}/backend", frontend_path), &backend_path)
+            .await?;
+        self.tx
+            .write_string(
+                format!("{}/backend-id", frontend_path),
+                &self.backend_domid.to_string(),
+            )
+            .await?;
+        Ok(())
+    }
+
+    pub async fn add_rw_path(&self, key: impl AsRef<str>) -> Result<()> {
+        let rw_perm = &[XsPermission {
+            id: self.frontend_domid,
+            perms: XS_PERM_READ_WRITE,
+        }];
+
+        self.tx
+            .mknod(
+                &format!("{}/{}", self.frontend_dom_path, key.as_ref()),
+                rw_perm,
+            )
+            .await?;
+        Ok(())
+    }
+
+    async fn before_commit(&self) -> Result<()> {
+        let devid_allocator_state = self.devalloc.lock().await.serialize();
+        let blkid_allocator_state = self.blkalloc.lock().await.serialize();
+        self.tx
+            .write(
+                format!("{}/devid-alloc-state", self.frontend_dom_path),
+                devid_allocator_state,
+            )
+            .await?;
+        self.tx
+            .write(
+                format!("{}/blkid-alloc-state", self.frontend_dom_path),
+                blkid_allocator_state,
+            )
+            .await?;
+        Ok(())
+    }
+
+    pub async fn maybe_commit(mut self) -> Result<bool> {
+        self.abort = false;
+        self.before_commit().await?;
+        Ok(self.tx.maybe_commit().await?)
+    }
+
+    pub async fn commit(mut self) -> Result<()> {
+        self.abort = false;
+        self.before_commit().await?;
+        self.tx.commit().await?;
+        Ok(())
+    }
+}
+
+impl Drop for XenTransaction {
+    fn drop(&mut self) {
+        if !self.abort {
+            return;
+        }
+        let tx = self.tx.clone();
+        tokio::task::spawn(async move {
+            let _ = tx.abort().await;
+        });
+    }
+}
+
+pub struct DeviceDescription {
+    frontend_type: String,
+    backend_type: String,
+    special_frontend_path: Option<String>,
+    frontend_items: HashMap<String, String>,
+    backend_items: HashMap<String, String>,
+}
+
+impl DeviceDescription {
+    pub fn new(frontend_type: impl AsRef<str>, backend_type: impl AsRef<str>) -> Self {
+        Self {
+            frontend_type: frontend_type.as_ref().to_string(),
+            backend_type: backend_type.as_ref().to_string(),
+            special_frontend_path: None,
+            frontend_items: HashMap::new(),
+            backend_items: HashMap::new(),
+        }
+    }
+
+    pub fn special_frontend_path(&mut self, path: impl AsRef<str>) -> &mut Self {
+        self.special_frontend_path = Some(path.as_ref().to_string());
+        self
+    }
+
+    pub fn add_frontend_item(&mut self, key: impl AsRef<str>, value: impl ToString) -> &mut Self {
+        self.frontend_items
+            .insert(key.as_ref().to_string(), value.to_string());
+        self
+    }
+
+    pub fn add_backend_item(&mut self, key: impl AsRef<str>, value: impl ToString) -> &mut Self {
+        self.backend_items
+            .insert(key.as_ref().to_string(), value.to_string());
+        self
+    }
+
+    pub fn add_frontend_bool(&mut self, key: impl AsRef<str>, value: bool) -> &mut Self {
+        self.add_frontend_item(key, if value { "1" } else { "0" })
+    }
+
+    pub fn add_backend_bool(&mut self, key: impl AsRef<str>, value: bool) -> &mut Self {
+        self.add_backend_item(key, if value { "1" } else { "0" })
+    }
+
+    pub fn done(self) -> Self {
+        self
+    }
+}
+
+#[derive(Clone, Debug)]
+pub struct DeviceResult {
+    pub id: u64,
+}
+
+#[derive(Clone, Debug)]
+pub struct BlockDeviceResult {
+    pub id: u64,
+    pub idx: u32,
+}
+
+#[async_trait::async_trait]
+pub trait DeviceConfig {
+    type Result;
+
+    async fn add_to_transaction(&self, tx: &XenTransaction) -> Result<Self::Result>;
+}
+
+#[derive(Clone, Debug)]
+pub struct BlockDeviceRef {
+    pub path: String,
+    pub major: u32,
+    pub minor: u32,
+}
+
+impl BlockDeviceRef {
+    pub fn new(path: impl AsRef<str>, major: u32, minor: u32) -> Self {
+        Self {
+            path: path.as_ref().to_string(),
+            major,
+            minor,
+        }
+    }
+}

crates/xen/xenclient/src/tx/pci.rs

@@ -0,0 +1,194 @@
+use super::{DeviceConfig, DeviceDescription, DeviceResult, XenTransaction};
+use crate::{
+    error::{Error, Result},
+    pci::{PciBdf, XenPciBackend},
+};
+use indexmap::IndexMap;
+use xencall::{sys::DOMCTL_DEV_RDM_RELAXED, XenCall};
+use xenplatform::sys::XEN_PAGE_SHIFT;
+
+#[derive(Clone, Debug, Default, Eq, PartialEq)]
+pub enum PciRdmReservePolicy {
+    Invalid,
+    #[default]
+    Strict,
+    Relaxed,
+}
+
+impl PciRdmReservePolicy {
+    pub fn to_option_str(&self) -> &str {
+        match self {
+            PciRdmReservePolicy::Invalid => "-1",
+            PciRdmReservePolicy::Strict => "0",
+            PciRdmReservePolicy::Relaxed => "1",
+        }
+    }
+}
+
+pub struct PciDeviceConfig {
+    bdf: PciBdf,
+    rdm_reserve_policy: PciRdmReservePolicy,
+    permissive: bool,
+    msi_translate: bool,
+    power_management: bool,
+}
+
+pub struct PciRootDeviceConfig {
+    backend_type: String,
+    devices: Vec<PciDeviceConfig>,
+}
+
+impl PciDeviceConfig {
+    pub fn new(bdf: PciBdf) -> Self {
+        Self {
+            bdf,
+            rdm_reserve_policy: PciRdmReservePolicy::Strict,
+            permissive: false,
+            msi_translate: false,
+            power_management: false,
+        }
+    }
+
+    pub fn rdm_reserve_policy(&mut self, rdm_reserve_policy: PciRdmReservePolicy) -> &mut Self {
+        self.rdm_reserve_policy = rdm_reserve_policy;
+        self
+    }
+
+    pub fn permissive(&mut self, permissive: bool) -> &mut Self {
+        self.permissive = permissive;
+        self
+    }
+
+    pub fn msi_translate(&mut self, msi_translate: bool) -> &mut Self {
+        self.msi_translate = msi_translate;
+        self
+    }
+
+    pub fn power_management(&mut self, power_management: bool) -> &mut Self {
+        self.power_management = power_management;
+        self
+    }
+
+    pub fn done(self) -> Self {
+        self
+    }
+}
+
+impl Default for PciRootDeviceConfig {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl PciRootDeviceConfig {
+    pub fn new() -> Self {
+        Self {
+            backend_type: "pci".to_string(),
+            devices: Vec::new(),
+        }
+    }
+
+    pub fn backend_type(&mut self, backend_type: impl AsRef<str>) -> &mut Self {
+        self.backend_type = backend_type.as_ref().to_string();
+        self
+    }
+
+    pub fn add_device(&mut self, device: PciDeviceConfig) -> &mut Self {
+        self.devices.push(device);
+        self
+    }
+
+    pub async fn prepare(&self, domid: u32, call: &XenCall) -> Result<()> {
+        for device in &self.devices {
+            let backend = XenPciBackend::new();
+            if !backend.is_assigned(&device.bdf).await? {
+                return Err(Error::PciDeviceNotAssignable(device.bdf));
+            }
+            let resources = backend.read_resources(&device.bdf).await?;
+            for resource in resources {
+                if resource.is_bar_io() {
+                    call.ioport_permission(
+                        domid,
+                        resource.start as u32,
+                        resource.size() as u32,
+                        true,
+                    )
+                    .await?;
+                } else {
+                    call.iomem_permission(
+                        domid,
+                        resource.start >> XEN_PAGE_SHIFT,
+                        (resource.size() + (XEN_PAGE_SHIFT - 1)) >> XEN_PAGE_SHIFT,
+                        true,
+                    )
+                    .await?;
+                }
+            }
+
+            if let Some(irq) = backend.read_irq(&device.bdf).await? {
+                let irq = call.map_pirq(domid, irq as isize, None).await?;
+                call.irq_permission(domid, irq, true).await?;
+            }
+
+            backend.reset(&device.bdf).await?;
+
+            call.assign_device(
+                domid,
+                device.bdf.encode(),
+                if device.rdm_reserve_policy == PciRdmReservePolicy::Relaxed {
+                    DOMCTL_DEV_RDM_RELAXED
+                } else {
+                    0
+                },
+            )
+            .await?;
+
+            if device.permissive {
+                backend.enable_permissive(&device.bdf).await?;
+            }
+        }
+        Ok(())
+    }
+
+    pub fn done(self) -> Self {
+        self
+    }
+}
+
+#[async_trait::async_trait]
+impl DeviceConfig for PciRootDeviceConfig {
+    type Result = DeviceResult;
+
+    async fn add_to_transaction(&self, tx: &XenTransaction) -> Result<DeviceResult> {
+        let id = tx.assign_next_devid().await?;
+        let mut device = DeviceDescription::new("pci", &self.backend_type);
+        device
+            .add_backend_bool("online", true)
+            .add_backend_item("state", 1)
+            .add_backend_item("num_devs", self.devices.len());
+
+        for (index, pci) in self.devices.iter().enumerate() {
+            let mut options = IndexMap::new();
+            options.insert("permissive", if pci.permissive { "1" } else { "0" });
+            options.insert("rdm_policy", pci.rdm_reserve_policy.to_option_str());
+            options.insert("msitranslate", if pci.msi_translate { "1" } else { "0" });
+            let options = options
+                .into_iter()
+                .map(|(key, value)| format!("{}={}", key, value))
+                .collect::<Vec<_>>()
+                .join(",");
+            device
+                .add_backend_item(format!("key-{}", index), pci.bdf.to_string())
+                .add_backend_item(format!("dev-{}", index), pci.bdf.to_string())
+                .add_backend_item(format!("opts-{}", index), options);
+
+            if let Some(vdefn) = pci.bdf.vdefn {
+                device.add_backend_item(format!("vdefn-{}", index), format!("{:#x}", vdefn));
+            }
+        }
+
+        device.add_frontend_item("state", 1);
+        tx.add_device(id, device).await?;
+        Ok(DeviceResult { id })
+    }
+}

crates/xen/xenclient/src/tx/vbd.rs

@@ -0,0 +1,131 @@
+use super::{BlockDeviceRef, BlockDeviceResult, DeviceConfig, DeviceDescription, XenTransaction};
+use crate::{
+    error::{Error, Result},
+    util::vbd_blkidx_to_disk_name,
+};
+
+pub struct VbdDeviceConfig {
+    backend_type: String,
+    removable: bool,
+    bootable: bool,
+    writable: bool,
+    discard: bool,
+    trusted: bool,
+    block_device: Option<BlockDeviceRef>,
+}
+
+impl Default for VbdDeviceConfig {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl VbdDeviceConfig {
+    pub fn new() -> Self {
+        Self {
+            backend_type: "vbd".to_string(),
+            removable: false,
+            bootable: true,
+            writable: false,
+            discard: false,
+            trusted: true,
+            block_device: None,
+        }
+    }
+
+    pub fn backend_type(&mut self, backend_type: impl AsRef<str>) -> &mut Self {
+        self.backend_type = backend_type.as_ref().to_string();
+        self
+    }
+
+    pub fn removable(&mut self, removable: bool) -> &mut Self {
+        self.removable = removable;
+        self
+    }
+
+    pub fn bootable(&mut self, bootable: bool) -> &mut Self {
+        self.bootable = bootable;
+        self
+    }
+
+    pub fn writable(&mut self, writable: bool) -> &mut Self {
+        self.writable = writable;
+        self
+    }
+
+    pub fn discard(&mut self, discard: bool) -> &mut Self {
+        self.discard = discard;
+        self
+    }
+
+    pub fn trusted(&mut self, trusted: bool) -> &mut Self {
+        self.trusted = trusted;
+        self
+    }
+
+    pub fn block_device(&mut self, block_device: BlockDeviceRef) -> &mut Self {
+        self.block_device = Some(block_device);
+        self
+    }
+
+    pub fn done(self) -> Self {
+        self
+    }
+}
+
+#[async_trait::async_trait]
+impl DeviceConfig for VbdDeviceConfig {
+    type Result = BlockDeviceResult;
+
+    async fn add_to_transaction(&self, tx: &XenTransaction) -> Result<BlockDeviceResult> {
+        let id = tx.assign_next_devid().await?;
+        let idx = tx.assign_next_blkidx().await?;
+        let vdev = vbd_blkidx_to_disk_name(idx)?;
+        let block_device = self
+            .block_device
+            .as_ref()
+            .ok_or_else(|| Error::ParameterMissing("block device"))?;
+
+        let mut device = DeviceDescription::new("vbd", &self.backend_type);
+        device
+            .add_backend_item("online", 1)
+            .add_backend_bool("removable", self.removable)
+            .add_backend_bool("bootable", self.bootable)
+            .add_backend_item("type", "phy")
+            .add_backend_item("device-type", "disk")
+            .add_backend_item("discard-enable", self.discard)
+            .add_backend_item("specification", "xen")
+            .add_backend_item("physical-device-path", &block_device.path)
+            .add_backend_item("mode", if self.writable { "w" } else { "r" })
+            .add_backend_item(
+                "physical-device",
+                format!("{:02x}:{:02x}", block_device.major, block_device.minor),
+            )
+            .add_backend_item("dev", &vdev)
+            .add_backend_item("state", 1);
+
+        // we should use standard virtual-device support for first few block devices.
+        // the kernel warns when you use ext for indexes 5 or less, due to
+        // potential id overlapping.
+        let (vdev, vd_key) = if idx <= 5 {
+            // shift by 4 as partition count is 16
+            ((202 << 8) | (idx as u64 * 16u64), "virtual-device")
+        } else {
+            // this is silly but 256 is the number of partitions
+            // multiply the index by that to get the actual id
+            ((1u64 << 28u64) + (idx as u64) * 256, "virtual-device-ext")
+        };
+
+        device
+            .add_frontend_item(vd_key, vdev)
+            .add_frontend_item("state", 1)
+            .add_frontend_item("device-type", "disk")
+            .add_frontend_bool("trusted", self.trusted)
+            .add_frontend_item("protocol", "x86_64-abi")
+            .add_frontend_item("x-index", idx);
+
+        tx.add_device(id, device).await?;
+
+        Ok(BlockDeviceResult { id, idx })
+    }
+}

crates/xen/xenclient/src/tx/vif.rs

@@ -0,0 +1,112 @@
+use super::{DeviceConfig, DeviceDescription, DeviceResult, XenTransaction};
+use crate::error::{Error, Result};
+
+pub struct VifDeviceConfig {
+    backend_type: String,
+    mac: Option<String>,
+    mtu: Option<u32>,
+    script: Option<String>,
+    bridge: Option<String>,
+    trusted: bool,
+}
+
+impl Default for VifDeviceConfig {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl VifDeviceConfig {
+    pub fn new() -> Self {
+        Self {
+            backend_type: "vif".to_string(),
+            mac: None,
+            mtu: None,
+            script: None,
+            bridge: None,
+            trusted: true,
+        }
+    }
+
+    pub fn backend_type(&mut self, backend_type: impl AsRef<str>) -> &mut Self {
+        self.backend_type = backend_type.as_ref().to_string();
+        self
+    }
+
+    pub fn mac(&mut self, mac: impl AsRef<str>) -> &mut Self {
+        self.mac = Some(mac.as_ref().to_string());
+        self
+    }
+
+    pub fn mtu(&mut self, mtu: u32) -> &mut Self {
+        self.mtu = Some(mtu);
+        self
+    }
+
+    pub fn script(&mut self, script: impl AsRef<str>) -> &mut Self {
+        self.script = Some(script.as_ref().to_string());
+        self
+    }
+
+    pub fn bridge(&mut self, bridge: impl AsRef<str>) -> &mut Self {
+        self.bridge = Some(bridge.as_ref().to_string());
+        self
+    }
+
+    pub fn trusted(&mut self, trusted: bool) -> &mut Self {
+        self.trusted = trusted;
+        self
+    }
+
+    pub fn done(self) -> Self {
+        self
+    }
+}
+
+#[async_trait::async_trait]
+impl DeviceConfig for VifDeviceConfig {
+    type Result = DeviceResult;
+
+    async fn add_to_transaction(&self, tx: &XenTransaction) -> Result<DeviceResult> {
+        let id = tx.assign_next_devid().await?;
+        let mac = self
+            .mac
+            .as_ref()
+            .ok_or_else(|| Error::ParameterMissing("mac address"))?;
+        let mtu = self
+            .mtu
+            .ok_or_else(|| Error::ParameterMissing("mtu"))?
+            .to_string();
+        let mut device = DeviceDescription::new("vif", &self.backend_type);
+        device
+            .add_backend_item("online", 1)
+            .add_backend_item("state", 1)
+            .add_backend_item("mac", mac)
+            .add_backend_item("mtu", &mtu)
+            .add_backend_item("type", "vif")
+            .add_backend_item("handle", id);
+
+        if let Some(bridge) = self.bridge.as_ref() {
+            device.add_backend_item("bridge", bridge);
+        }
+
+        if let Some(script) = self.script.as_ref() {
+            device
+                .add_backend_item("script", script)
+                .add_backend_item("hotplug-status", "");
+        } else {
+            device
+                .add_backend_item("script", "")
+                .add_backend_item("hotplug-status", "connected");
+        }
+
+        device
+            .add_frontend_item("state", 1)
+            .add_frontend_item("mac", mac)
+            .add_frontend_item("mtu", &mtu)
+            .add_frontend_bool("trusted", self.trusted);
+
+        tx.add_device(id, device.done()).await?;
+        Ok(DeviceResult { id })
+    }
+}

crates/xen/xenclient/src/util.rs

@@ -0,0 +1,21 @@
+use crate::error::{Error, Result};
+
+pub fn vbd_blkidx_to_disk_name(blkid: u32) -> Result<String> {
+    let mut name = "xvd".to_string();
+    let mut suffix = String::new();
+    let mut n = blkid;
+    loop {
+        let c = (n % 26) as u8;
+        let c = b'a' + c;
+        let c = char::from_u32(c as u32).ok_or(Error::InvalidBlockIdx)?;
+        suffix.push(c);
+        if n >= 26 {
+            n = (n / 26) - 1;
+            continue;
+        } else {
+            break;
+        }
+    }
+    name.push_str(&suffix.chars().rev().collect::<String>());
+    Ok(name)
+}

crates/xen/xenplatform/src/boot.rs

@@ -8,16 +8,9 @@ use crate::{
     error::{Error, Result},
     mem::PhysicalPages,
     sys::XEN_PAGE_SHIFT,
+    ImageLoader, PlatformKernelConfig, PlatformResourcesConfig,
 };
 
-pub struct BootSetup<I: BootImageLoader, P: BootSetupPlatform> {
-    pub call: XenCall,
-    pub domid: u32,
-    pub platform: P,
-    pub image_loader: I,
-    pub dtb: Option<Vec<u8>>,
-}
-
 #[derive(Debug, Default, Clone)]
 pub struct DomainSegment {
     pub vstart: u64,
@@ -42,7 +35,7 @@ pub struct BootDomain {
     pub phys: PhysicalPages,
     pub store_evtchn: u32,
     pub store_mfn: u64,
-    pub initrd_segment: DomainSegment,
+    pub initrd_segment: Option<DomainSegment>,
     pub console_evtchn: u32,
     pub console_mfn: u64,
     pub cmdline: String,
@@ -142,129 +135,8 @@ impl BootDomain {
     }
 }
 
-impl<I: BootImageLoader, P: BootSetupPlatform> BootSetup<I, P> {
-    pub fn new(
-        call: XenCall,
-        domid: u32,
-        platform: P,
-        image_loader: I,
-        dtb: Option<Vec<u8>>,
-    ) -> BootSetup<I, P> {
-        BootSetup {
-            call,
-            domid,
-            platform,
-            image_loader,
-            dtb,
-        }
-    }
-
-    pub async fn initialize(
-        &mut self,
-        initrd: &[u8],
-        target_mem_mb: u64,
-        max_mem_mb: u64,
-        max_vcpus: u32,
-        cmdline: &str,
-    ) -> Result<BootDomain> {
-        let target_pages = target_mem_mb << (20 - self.platform.page_shift());
-        let total_pages = max_mem_mb << (20 - self.platform.page_shift());
-        let image_info = self.image_loader.parse(self.platform.hvm()).await?;
-        let mut domain = BootDomain {
-            domid: self.domid,
-            call: self.call.clone(),
-            virt_alloc_end: 0,
-            virt_pgtab_end: 0,
-            pfn_alloc_end: 0,
-            total_pages,
-            target_pages,
-            page_size: self.platform.page_size(),
-            image_info,
-            console_evtchn: 0,
-            console_mfn: 0,
-            max_vcpus,
-            phys: PhysicalPages::new(self.call.clone(), self.domid, self.platform.page_shift()),
-            initrd_segment: DomainSegment::default(),
-            store_evtchn: 0,
-            store_mfn: 0,
-            cmdline: cmdline.to_string(),
-        };
-
-        self.platform.initialize_early(&mut domain).await?;
-
-        let mut initrd_segment = if !domain.image_info.unmapped_initrd {
-            Some(domain.alloc_module(initrd).await?)
-        } else {
-            None
-        };
-
-        let mut kernel_segment = if self.platform.needs_early_kernel() {
-            Some(self.load_kernel_segment(&mut domain).await?)
-        } else {
-            None
-        };
-
-        self.platform.initialize_memory(&mut domain).await?;
-        domain.virt_alloc_end = domain.image_info.virt_base;
-
-        if kernel_segment.is_none() {
-            kernel_segment = Some(self.load_kernel_segment(&mut domain).await?);
-        }
-
-        if domain.image_info.unmapped_initrd {
-            initrd_segment = Some(domain.alloc_module(initrd).await?);
-        }
-
-        domain.initrd_segment =
-            initrd_segment.ok_or(Error::MemorySetupFailed("initrd_segment missing"))?;
-
-        self.platform.alloc_magic_pages(&mut domain).await?;
-
-        domain.store_evtchn = self.call.evtchn_alloc_unbound(self.domid, 0).await?;
-
-        let _kernel_segment =
-            kernel_segment.ok_or(Error::MemorySetupFailed("kernel_segment missing"))?;
-
-        Ok(domain)
-    }
-
-    pub async fn boot(&mut self, domain: &mut BootDomain) -> Result<()> {
-        let domain_info = self.call.get_domain_info(self.domid).await?;
-        let shared_info_frame = domain_info.shared_info_frame;
-        self.platform.setup_page_tables(domain).await?;
-        self.platform
-            .setup_start_info(domain, shared_info_frame)
-            .await?;
-        self.platform.setup_hypercall_page(domain).await?;
-        self.platform.bootlate(domain).await?;
-        self.platform
-            .setup_shared_info(domain, shared_info_frame)
-            .await?;
-        self.platform.vcpu(domain).await?;
-        domain.phys.unmap_all()?;
-        self.platform.gnttab_seed(domain).await?;
-        Ok(())
-    }
-
-    async fn load_kernel_segment(&mut self, domain: &mut BootDomain) -> Result<DomainSegment> {
-        let kernel_segment = domain
-            .alloc_segment(
-                domain.image_info.virt_kstart,
-                domain.image_info.virt_kend - domain.image_info.virt_kstart,
-            )
-            .await?;
-        let kernel_segment_ptr = kernel_segment.addr as *mut u8;
-        let kernel_segment_slice =
-            unsafe { slice::from_raw_parts_mut(kernel_segment_ptr, kernel_segment.size as usize) };
-        self.image_loader
-            .load(&domain.image_info, kernel_segment_slice)
-            .await?;
-        Ok(kernel_segment)
-    }
-}
-
 #[async_trait::async_trait]
-pub trait BootSetupPlatform: Clone {
+pub trait BootSetupPlatform {
     fn create_domain(&self, enable_iommu: bool) -> CreateDomain;
     fn page_size(&self) -> u64;
     fn page_shift(&self) -> u64;
@@ -304,6 +176,135 @@ pub trait BootSetupPlatform: Clone {
     async fn vcpu(&mut self, domain: &mut BootDomain) -> Result<()>;
 
     async fn setup_hypercall_page(&mut self, domain: &mut BootDomain) -> Result<()>;
+
+    async fn initialize_internal(
+        &mut self,
+        domid: u32,
+        call: XenCall,
+        image_loader: &ImageLoader,
+        domain: &mut BootDomain,
+        kernel: &PlatformKernelConfig,
+    ) -> Result<()> {
+        self.initialize_early(domain).await?;
+
+        let mut initrd_segment = if !domain.image_info.unmapped_initrd && kernel.initrd.is_some() {
+            Some(domain.alloc_module(kernel.initrd.as_ref().unwrap()).await?)
+        } else {
+            None
+        };
+
+        let mut kernel_segment = if self.needs_early_kernel() {
+            Some(self.load_kernel_segment(image_loader, domain).await?)
+        } else {
+            None
+        };
+
+        self.initialize_memory(domain).await?;
+        domain.virt_alloc_end = domain.image_info.virt_base;
+
+        if kernel_segment.is_none() {
+            kernel_segment = Some(self.load_kernel_segment(image_loader, domain).await?);
+        }
+
+        if domain.image_info.unmapped_initrd && kernel.initrd.is_some() {
+            initrd_segment = Some(domain.alloc_module(kernel.initrd.as_ref().unwrap()).await?);
+        }
+
+        domain.initrd_segment = initrd_segment;
+        self.alloc_magic_pages(domain).await?;
+        domain.store_evtchn = call.evtchn_alloc_unbound(domid, 0).await?;
+        let _kernel_segment =
+            kernel_segment.ok_or(Error::MemorySetupFailed("kernel_segment missing"))?;
+        Ok(())
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    async fn initialize(
+        &mut self,
+        domid: u32,
+        call: XenCall,
+        image_loader: &ImageLoader,
+        kernel: &PlatformKernelConfig,
+        resources: &PlatformResourcesConfig,
+    ) -> Result<BootDomain> {
+        let target_pages = resources.assigned_memory_mb << (20 - self.page_shift());
+        let total_pages = resources.max_memory_mb << (20 - self.page_shift());
+        let image_info = image_loader.parse(self.hvm()).await?;
+        let mut domain = BootDomain {
+            domid,
+            call: call.clone(),
+            virt_alloc_end: 0,
+            virt_pgtab_end: 0,
+            pfn_alloc_end: 0,
+            total_pages,
+            target_pages,
+            page_size: self.page_size(),
+            image_info,
+            console_evtchn: 0,
+            console_mfn: 0,
+            max_vcpus: resources.max_vcpus,
+            phys: PhysicalPages::new(call.clone(), domid, self.page_shift()),
+            initrd_segment: None,
+            store_evtchn: 0,
+            store_mfn: 0,
+            cmdline: kernel.cmdline.clone(),
+        };
+        match self
+            .initialize_internal(domid, call, image_loader, &mut domain, kernel)
+            .await
+        {
+            Ok(_) => Ok(domain),
+            Err(error) => {
+                domain.phys.unmap_all()?;
+                Err(error)
+            }
+        }
+    }
+
+    async fn boot_internal(
+        &mut self,
+        call: XenCall,
+        domid: u32,
+        domain: &mut BootDomain,
+    ) -> Result<()> {
+        let domain_info = call.get_domain_info(domid).await?;
+        let shared_info_frame = domain_info.shared_info_frame;
+        self.setup_page_tables(domain).await?;
+        self.setup_start_info(domain, shared_info_frame).await?;
+        self.setup_hypercall_page(domain).await?;
+        self.bootlate(domain).await?;
+        self.setup_shared_info(domain, shared_info_frame).await?;
+        self.vcpu(domain).await?;
+        self.gnttab_seed(domain).await?;
+        domain.phys.unmap_all()?;
+        Ok(())
+    }
+
+    async fn boot(&mut self, domid: u32, call: XenCall, domain: &mut BootDomain) -> Result<()> {
+        let result = self.boot_internal(call, domid, domain).await;
+        domain.phys.unmap_all()?;
+        result
+    }
+
+    async fn load_kernel_segment(
+        &mut self,
+        image_loader: &ImageLoader,
+        domain: &mut BootDomain,
+    ) -> Result<DomainSegment> {
+        let kernel_segment = domain
+            .alloc_segment(
+                domain.image_info.virt_kstart,
+                domain.image_info.virt_kend - domain.image_info.virt_kstart,
+            )
+            .await?;
+        let kernel_segment_ptr = kernel_segment.addr as *mut u8;
+        let kernel_segment_slice =
+            unsafe { slice::from_raw_parts_mut(kernel_segment_ptr, kernel_segment.size as usize) };
+        image_loader
+            .load(&domain.image_info, kernel_segment_slice)
+            .await?;
+        Ok(kernel_segment)
+    }
 }
 
 #[async_trait::async_trait]

crates/xen/xenplatform/src/domain.rs

@@ -1,9 +1,10 @@
 use std::sync::Arc;
 
 use crate::{
-    boot::{BootSetup, BootSetupPlatform},
-    elfloader::ElfImageLoader,
+    boot::BootDomain, elfloader::ElfImageLoader, error::Error, ImageLoader, RuntimePlatform,
+    RuntimePlatformType,
 };
+use log::warn;
 use uuid::Uuid;
 use xencall::XenCall;
 
@@ -11,42 +12,92 @@ use crate::error::Result;
 
 pub const XEN_EXTRA_MEMORY_KB: u64 = 2048;
 
-pub struct BaseDomainManager<P: BootSetupPlatform> {
+pub struct PlatformDomainManager {
     call: XenCall,
-    pub platform: Arc<P>,
 }
 
-impl<P: BootSetupPlatform> BaseDomainManager<P> {
-    pub async fn new(call: XenCall, platform: P) -> Result<BaseDomainManager<P>> {
-        Ok(BaseDomainManager {
-            call,
-            platform: Arc::new(platform),
-        })
+impl PlatformDomainManager {
+    pub async fn new(call: XenCall) -> Result<PlatformDomainManager> {
+        Ok(PlatformDomainManager { call })
     }
 
-    pub async fn create(&self, config: BaseDomainConfig) -> Result<CreatedDomain> {
-        let mut domain = self.platform.create_domain(config.enable_iommu);
+    fn max_memory_kb(resources: &PlatformResourcesConfig) -> u64 {
+        (resources.max_memory_mb * 1024) + XEN_EXTRA_MEMORY_KB
+    }
+
+    async fn create_base_domain(
+        &self,
+        config: &PlatformDomainConfig,
+        platform: &RuntimePlatform,
+    ) -> Result<u32> {
+        let mut domain = platform.create_domain(config.options.iommu);
         domain.handle = config.uuid.into_bytes();
-        domain.max_vcpus = config.max_vcpus;
+        domain.max_vcpus = config.resources.max_vcpus;
         let domid = self.call.create_domain(domain).await?;
-        self.call.set_max_vcpus(domid, config.max_vcpus).await?;
+        Ok(domid)
+    }
+
+    async fn configure_domain_resources(
+        &self,
+        domid: u32,
+        config: &PlatformDomainConfig,
+    ) -> Result<()> {
         self.call
-            .set_max_mem(domid, (config.max_mem_mb * 1024) + XEN_EXTRA_MEMORY_KB)
+            .set_max_vcpus(domid, config.resources.max_vcpus)
             .await?;
-        let loader = ElfImageLoader::load_file_kernel(&config.kernel)?;
-        let platform = (*self.platform).clone();
-        let mut boot = BootSetup::new(self.call.clone(), domid, platform, loader, None);
-        let mut domain = boot
-            .initialize(
-                &config.initrd,
-                config.target_mem_mb,
-                config.max_mem_mb,
-                config.max_vcpus,
-                &config.cmdline,
+        self.call
+            .set_max_mem(
+                domid,
+                PlatformDomainManager::max_memory_kb(&config.resources),
             )
             .await?;
-        boot.boot(&mut domain).await?;
-        Ok(CreatedDomain {
+        Ok(())
+    }
+
+    async fn create_internal(
+        &self,
+        domid: u32,
+        config: &PlatformDomainConfig,
+        mut platform: RuntimePlatform,
+    ) -> Result<BootDomain> {
+        self.configure_domain_resources(domid, config).await?;
+        let kernel = config.kernel.clone();
+        let loader = tokio::task::spawn_blocking(move || match kernel.format {
+            KernelFormat::ElfCompressed => ElfImageLoader::load(kernel.data),
+            KernelFormat::ElfUncompressed => Ok(ElfImageLoader::new(kernel.data)),
+        })
+        .await
+        .map_err(Error::AsyncJoinError)??;
+        let loader = ImageLoader::Elf(loader);
+        let mut domain = platform
+            .initialize(
+                domid,
+                self.call.clone(),
+                &loader,
+                &config.kernel,
+                &config.resources,
+            )
+            .await?;
+        platform.boot(domid, self.call.clone(), &mut domain).await?;
+        Ok(domain)
+    }
+
+    pub async fn create(&self, config: PlatformDomainConfig) -> Result<PlatformDomainInfo> {
+        let platform = config.platform.create();
+        let domid = self.create_base_domain(&config, &platform).await?;
+        let domain = match self.create_internal(domid, &config, platform).await {
+            Ok(domain) => domain,
+            Err(error) => {
+                if let Err(destroy_fail) = self.call.destroy_domain(domid).await {
+                    warn!(
+                        "failed to destroy failed domain {}: {}",
+                        domid, destroy_fail
+                    );
+                }
+                return Err(error);
+            }
+        };
+        Ok(PlatformDomainInfo {
             domid,
             store_evtchn: domain.store_evtchn,
             store_mfn: domain.store_mfn,
@@ -62,21 +113,43 @@ impl<P: BootSetupPlatform> BaseDomainManager<P> {
 }
 
 #[derive(Clone, Debug)]
-pub struct BaseDomainConfig {
+pub struct PlatformDomainConfig {
     pub uuid: Uuid,
-    pub owner_domid: u32,
-    pub max_vcpus: u32,
-    pub target_vcpus: u32,
-    pub max_mem_mb: u64,
-    pub target_mem_mb: u64,
-    pub kernel: Vec<u8>,
-    pub initrd: Vec<u8>,
-    pub cmdline: String,
-    pub enable_iommu: bool,
+    pub platform: RuntimePlatformType,
+    pub resources: PlatformResourcesConfig,
+    pub kernel: PlatformKernelConfig,
+    pub options: PlatformOptions,
 }
 
 #[derive(Clone, Debug)]
-pub struct CreatedDomain {
+pub struct PlatformKernelConfig {
+    pub data: Arc<Vec<u8>>,
+    pub format: KernelFormat,
+    pub initrd: Option<Arc<Vec<u8>>>,
+    pub cmdline: String,
+}
+
+#[derive(Clone, Debug)]
+pub struct PlatformResourcesConfig {
+    pub max_vcpus: u32,
+    pub assigned_vcpus: u32,
+    pub max_memory_mb: u64,
+    pub assigned_memory_mb: u64,
+}
+
+#[derive(Clone, Debug)]
+pub struct PlatformOptions {
+    pub iommu: bool,
+}
+
+#[derive(Clone, Debug)]
+pub enum KernelFormat {
+    ElfUncompressed,
+    ElfCompressed,
+}
+
+#[derive(Clone, Debug)]
+pub struct PlatformDomainInfo {
     pub domid: u32,
     pub store_evtchn: u32,
     pub store_mfn: u64,

crates/xen/xenplatform/src/elfloader.rs

@@ -19,6 +19,12 @@ use std::mem::size_of;
 use std::sync::Arc;
 use xz2::bufread::XzDecoder;
 
+const ELF_MAGIC: &[u8] = &[
+    0x7f, 0x45, 0x4c, 0x46, 0x02, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+];
+const GZIP_MAGIC: &[u8] = &[0x1f, 0x8b];
+const XZ_MAGIC: &[u8] = &[0xfd, 0x37, 0x7a, 0x58];
+
 #[derive(Clone)]
 pub struct ElfImageLoader {
     data: Arc<Vec<u8>>,
@@ -60,22 +66,40 @@ fn xen_note_value_as_u64(endian: AnyEndian, value: &[u8]) -> Option<u64> {
 }
 
 impl ElfImageLoader {
-    pub fn new(data: Vec<u8>) -> ElfImageLoader {
-        ElfImageLoader {
-            data: Arc::new(data),
-        }
+    pub fn new(data: Arc<Vec<u8>>) -> ElfImageLoader {
+        ElfImageLoader { data }
     }
 
     pub fn load_gz(data: &[u8]) -> Result<ElfImageLoader> {
         let buff = BufReader::new(data);
         let image = ElfImageLoader::read_one_stream(&mut GzDecoder::new(buff))?;
-        Ok(ElfImageLoader::new(image))
+        Ok(ElfImageLoader::new(Arc::new(image)))
     }
 
     pub fn load_xz(data: &[u8]) -> Result<ElfImageLoader> {
         let buff = BufReader::new(data);
         let image = ElfImageLoader::read_one_stream(&mut XzDecoder::new(buff))?;
-        Ok(ElfImageLoader::new(image))
+        Ok(ElfImageLoader::new(Arc::new(image)))
+    }
+
+    pub fn load(data: Arc<Vec<u8>>) -> Result<ElfImageLoader> {
+        if data.len() >= 16 && find_iter(&data[0..15], ELF_MAGIC).next().is_some() {
+            return Ok(ElfImageLoader::new(data));
+        }
+
+        for start in find_iter(&data, GZIP_MAGIC) {
+            if let Ok(elf) = ElfImageLoader::load_gz(&data[start..]) {
+                return Ok(elf);
+            }
+        }
+
+        for start in find_iter(&data, XZ_MAGIC) {
+            if let Ok(elf) = ElfImageLoader::load_xz(&data[start..]) {
+                return Ok(elf);
+            }
+        }
+
+        Err(Error::ElfCompressionUnknown)
     }
 
     fn read_one_stream(read: &mut dyn Read) -> Result<Vec<u8>> {
@@ -101,36 +125,11 @@ impl ElfImageLoader {
         Ok(result)
     }
 
-    pub fn load_file_gz(path: &str) -> Result<ElfImageLoader> {
-        let file = std::fs::read(path)?;
-        ElfImageLoader::load_gz(file.as_slice())
-    }
-
-    pub fn load_file_xz(path: &str) -> Result<ElfImageLoader> {
-        let file = std::fs::read(path)?;
-        ElfImageLoader::load_xz(file.as_slice())
-    }
-
-    pub fn load_file_kernel(data: &[u8]) -> Result<ElfImageLoader> {
-        for start in find_iter(data, &[0x1f, 0x8b]) {
-            if let Ok(elf) = ElfImageLoader::load_gz(&data[start..]) {
-                return Ok(elf);
-            }
-        }
-
-        for start in find_iter(data, &[0xfd, 0x37, 0x7a, 0x58]) {
-            if let Ok(elf) = ElfImageLoader::load_xz(&data[start..]) {
-                return Ok(elf);
-            }
-        }
-
-        Err(Error::ElfCompressionUnknown)
-    }
-
     fn parse_sync(&self, hvm: bool) -> Result<BootImageInfo> {
         let elf = ElfBytes::<AnyEndian>::minimal_parse(self.data.as_slice())?;
-        let headers = elf.section_headers().ok_or(Error::ElfInvalidImage)?;
-        let mut linux_notes: HashMap<u64, Vec<u8>> = HashMap::new();
+        let headers = elf
+            .section_headers()
+            .ok_or(Error::ElfInvalidImage("section headers missing"))?;
         let mut xen_notes: HashMap<u64, ElfNoteValue> = HashMap::new();
 
         for header in headers {
@@ -140,10 +139,9 @@ impl ElfImageLoader {
 
             let notes = elf.section_data_as_notes(&header)?;
             for note in notes {
-                if let Note::Unknown(note) = note {
-                    if note.name == "Linux" {
-                        linux_notes.insert(note.n_type, note.desc.to_vec());
-                    }
+                let Note::Unknown(note) = note else {
+                    continue;
+                };
 
                 if note.name == "Xen" {
                     for typ in XEN_ELFNOTE_TYPES {
@@ -159,15 +157,9 @@ impl ElfImageLoader {
 
                         xen_notes.insert(typ.id, ElfNoteValue { value });
                     }
-                        continue;
                 }
             }
         }
-        }
-
-        if linux_notes.is_empty() {
-            return Err(Error::ElfInvalidImage);
-        }
 
         if xen_notes.is_empty() {
             return Err(Error::ElfXenSupportMissing);
@@ -175,27 +167,27 @@ impl ElfImageLoader {
 
         let paddr_offset = xen_notes
             .get(&XEN_ELFNOTE_PADDR_OFFSET)
-            .ok_or(Error::ElfInvalidImage)?
+            .ok_or(Error::ElfXenNoteMissing("PADDR_OFFSET"))?
             .value;
         let virt_base = xen_notes
             .get(&XEN_ELFNOTE_VIRT_BASE)
-            .ok_or(Error::ElfInvalidImage)?
+            .ok_or(Error::ElfXenNoteMissing("VIRT_BASE"))?
             .value;
         let entry = xen_notes
             .get(&XEN_ELFNOTE_ENTRY)
-            .ok_or(Error::ElfInvalidImage)?
+            .ok_or(Error::ElfXenNoteMissing("ENTRY"))?
             .value;
         let virt_hypercall = xen_notes
             .get(&XEN_ELFNOTE_HYPERCALL_PAGE)
-            .ok_or(Error::ElfInvalidImage)?
+            .ok_or(Error::ElfXenNoteMissing("HYPERCALL_PAGE"))?
             .value;
         let init_p2m = xen_notes
             .get(&XEN_ELFNOTE_INIT_P2M)
-            .ok_or(Error::ElfInvalidImage)?
+            .ok_or(Error::ElfXenNoteMissing("INIT_P2M"))?
             .value;
         let mod_start_pfn = xen_notes
             .get(&XEN_ELFNOTE_MOD_START_PFN)
-            .ok_or(Error::ElfInvalidImage)?
+            .ok_or(Error::ElfXenNoteMissing("MOD_START_PFN"))?
             .value;
 
         let phys32_entry = xen_notes.get(&XEN_ELFNOTE_PHYS32_ENTRY).map(|x| x.value);
@@ -203,7 +195,9 @@ impl ElfImageLoader {
         let mut start: u64 = u64::MAX;
         let mut end: u64 = 0;
 
-        let segments = elf.segments().ok_or(Error::ElfInvalidImage)?;
+        let segments = elf
+            .segments()
+            .ok_or(Error::ElfInvalidImage("segments missing"))?;
 
         for header in segments {
             if (header.p_type != PT_LOAD) || (header.p_flags & (PF_R | PF_W | PF_X)) == 0 {
@@ -221,7 +215,9 @@ impl ElfImageLoader {
         }
 
         if paddr_offset != u64::MAX && virt_base == u64::MAX {
-            return Err(Error::ElfInvalidImage);
+            return Err(Error::ElfInvalidImage(
+                "paddr_offset specified, but virt_base is not specified",
+            ));
         }
 
         let virt_offset = virt_base - paddr_offset;
@@ -247,8 +243,13 @@ impl ElfImageLoader {
         };
         Ok(image_info)
     }
+
+    pub fn into_elf_bytes(self) -> Arc<Vec<u8>> {
+        self.data
+    }
 }
 
+#[derive(Debug)]
 struct ElfNoteValue {
     value: u64,
 }
@@ -262,7 +263,9 @@ impl BootImageLoader for ElfImageLoader {
 
     async fn load(&self, image_info: &BootImageInfo, dst: &mut [u8]) -> Result<()> {
         let elf = ElfBytes::<AnyEndian>::minimal_parse(self.data.as_slice())?;
-        let segments = elf.segments().ok_or(Error::ElfInvalidImage)?;
+        let segments = elf
+            .segments()
+            .ok_or(Error::ElfInvalidImage("segments missing"))?;
 
         debug!(
             "load dst={:#x} segments={}",

crates/xen/xenplatform/src/error.rs

@@ -28,10 +28,14 @@ pub enum Error {
     PopulatePhysmapFailed(usize, usize, usize),
     #[error("unknown elf compression method")]
     ElfCompressionUnknown,
-    #[error("expected elf image format not found")]
-    ElfInvalidImage,
+    #[error("elf image format invalid: {0}")]
+    ElfInvalidImage(&'static str),
+    #[error("elf linux image not found")]
+    ElfNotLinux,
     #[error("provided elf image does not contain xen support")]
     ElfXenSupportMissing,
+    #[error("provided elf image does not contain xen note {0}")]
+    ElfXenNoteMissing(&'static str),
     #[error("regex error: {0}")]
     RegexError(#[from] regex::Error),
     #[error("error: {0}")]

crates/xen/xenplatform/src/lib.rs

@@ -4,9 +4,104 @@ pub mod error;
 pub mod mem;
 pub mod sys;
 
+use boot::{BootDomain, BootImageInfo, BootImageLoader, BootSetupPlatform};
+use domain::{PlatformKernelConfig, PlatformResourcesConfig};
+use elfloader::ElfImageLoader;
+use error::Result;
+use unsupported::UnsupportedPlatform;
+use xencall::{sys::CreateDomain, XenCall};
+
 use crate::error::Error;
 
 pub mod domain;
 pub mod unsupported;
 #[cfg(target_arch = "x86_64")]
 pub mod x86pv;
+
+#[derive(Clone)]
+pub enum ImageLoader {
+    Elf(ElfImageLoader),
+}
+
+impl ImageLoader {
+    async fn parse(&self, hvm: bool) -> Result<BootImageInfo> {
+        match self {
+            ImageLoader::Elf(elf) => elf.parse(hvm).await,
+        }
+    }
+
+    async fn load(&self, image_info: &BootImageInfo, dst: &mut [u8]) -> Result<()> {
+        match self {
+            ImageLoader::Elf(elf) => elf.load(image_info, dst).await,
+        }
+    }
+}
+
+#[derive(Clone, Debug, PartialEq, PartialOrd, Eq, Ord)]
+pub enum RuntimePlatformType {
+    Unsupported,
+    #[cfg(target_arch = "x86_64")]
+    Pv,
+}
+
+impl RuntimePlatformType {
+    pub fn create(&self) -> RuntimePlatform {
+        match self {
+            RuntimePlatformType::Unsupported => {
+                RuntimePlatform::Unsupported(UnsupportedPlatform::new())
+            }
+            #[cfg(target_arch = "x86_64")]
+            RuntimePlatformType::Pv => RuntimePlatform::Pv(x86pv::X86PvPlatform::new()),
+        }
+    }
+}
+
+#[allow(clippy::large_enum_variant)]
+pub enum RuntimePlatform {
+    Unsupported(UnsupportedPlatform),
+    #[cfg(target_arch = "x86_64")]
+    Pv(x86pv::X86PvPlatform),
+}
+
+impl RuntimePlatform {
+    #[allow(clippy::too_many_arguments)]
+    pub async fn initialize(
+        &mut self,
+        domid: u32,
+        call: XenCall,
+        image_loader: &ImageLoader,
+        kernel: &PlatformKernelConfig,
+        resources: &PlatformResourcesConfig,
+    ) -> Result<BootDomain> {
+        match self {
+            RuntimePlatform::Unsupported(unsupported) => {
+                unsupported
+                    .initialize(domid, call, image_loader, kernel, resources)
+                    .await
+            }
+            #[cfg(target_arch = "x86_64")]
+            RuntimePlatform::Pv(pv) => {
+                pv.initialize(domid, call, image_loader, kernel, resources)
+                    .await
+            }
+        }
+    }
+
+    pub async fn boot(&mut self, domid: u32, call: XenCall, domain: &mut BootDomain) -> Result<()> {
+        match self {
+            RuntimePlatform::Unsupported(unsupported) => {
+                unsupported.boot(domid, call, domain).await
+            }
+            #[cfg(target_arch = "x86_64")]
+            RuntimePlatform::Pv(pv) => pv.boot(domid, call, domain).await,
+        }
+    }
+
+    pub fn create_domain(&self, enable_iommu: bool) -> CreateDomain {
+        match self {
+            RuntimePlatform::Unsupported(unsupported) => unsupported.create_domain(enable_iommu),
+            #[cfg(target_arch = "x86_64")]
+            RuntimePlatform::Pv(pv) => pv.create_domain(enable_iommu),
+        }
+    }
+}

crates/xen/xenplatform/src/x86pv.rs

@@ -9,8 +9,8 @@ use log::{debug, trace};
 use nix::errno::Errno;
 use slice_copy::copy;
 use xencall::sys::{
-    x8664VcpuGuestContext, CreateDomain, E820Entry, VcpuGuestContextAny, E820_MAX, E820_RAM,
-    E820_UNUSABLE, MMUEXT_PIN_L4_TABLE, XEN_DOMCTL_CDF_IOMMU,
+    x8664VcpuGuestContext, CreateDomain, VcpuGuestContextAny, MMUEXT_PIN_L4_TABLE,
+    XEN_DOMCTL_CDF_IOMMU,
 };
 
 use crate::{
@@ -282,52 +282,6 @@ impl X86PvPlatform {
         self.table.mappings[m] = map;
         Ok(m)
     }
-
-    fn e820_sanitize(
-        &self,
-        mut source: Vec<E820Entry>,
-        map_limit_kb: u64,
-    ) -> Result<Vec<E820Entry>> {
-        let mut e820 = vec![E820Entry::default(); E820_MAX as usize];
-
-        for entry in &mut source {
-            if entry.addr > 0x100000 {
-                continue;
-            }
-
-            // entries under 1MB should be removed.
-            entry.typ = 0;
-            entry.size = 0;
-            entry.addr = u64::MAX;
-        }
-
-        let mut lowest = u64::MAX;
-        let mut highest = 0;
-
-        for entry in &source {
-            if entry.typ == E820_RAM || entry.typ == E820_UNUSABLE || entry.typ == 0 {
-                continue;
-            }
-
-            lowest = if entry.addr < lowest {
-                entry.addr
-            } else {
-                lowest
-            };
-
-            highest = if entry.addr + entry.size > highest {
-                entry.addr + entry.size
-            } else {
-                highest
-            }
-        }
-
-        e820[0].addr = 0;
-        e820[0].size = map_limit_kb << 10;
-        e820[0].typ = E820_RAM;
-
-        Ok(e820)
-    }
 }
 
 #[async_trait::async_trait]
@@ -690,8 +644,10 @@ impl BootSetupPlatform for X86PvPlatform {
             (*info).store_mfn = domain.phys.p2m[xenstore_segment.pfn as usize];
             (*info).console.mfn = domain.console_mfn;
             (*info).console.evtchn = domain.console_evtchn;
-            (*info).mod_start = domain.initrd_segment.vstart;
-            (*info).mod_len = domain.initrd_segment.size;
+            if let Some(ref initrd_segment) = domain.initrd_segment {
+                (*info).mod_start = initrd_segment.vstart;
+                (*info).mod_len = initrd_segment.size;
+            }
             for (i, c) in domain.cmdline.chars().enumerate() {
                 (*info).cmdline[i] = c as c_char;
             }
@@ -715,12 +671,6 @@ impl BootSetupPlatform for X86PvPlatform {
         domain.phys.unmap(pg_pfn)?;
         domain.phys.unmap(p2m_segment.pfn)?;
 
-        let map = domain.call.get_memory_map(E820_MAX).await?;
-        let mem_mb = domain.total_pages >> (20 - self.page_shift());
-        let mem_kb = mem_mb * 1024;
-        let e820 = self.e820_sanitize(map, mem_kb)?;
-        domain.call.set_memory_map(domain.domid, e820).await?;
-
         domain
             .call
             .mmuext(domain.domid, MMUEXT_PIN_L4_TABLE, pg_mfn, 0)