move xen libraries to xen/

xen/xenclient/Cargo.toml

@@ -0,0 +1,37 @@
+[package]
+name = "xenclient"
+version = "0.0.1"
+edition = "2021"
+resolver = "2"
+
+[dependencies]
+libc = "0.2"
+elf = "0.7.4"
+flate2 = "1.0"
+xz2 = "0.1"
+memchr = "2"
+slice-copy = "0.3.0"
+log = "0.4.20"
+
+[dependencies.xencall]
+path = "../xencall"
+
+[dependencies.xenstore]
+path = "../xenstore"
+
+[dependencies.xenevtchn]
+path = "../xenevtchn"
+
+[dependencies.uuid]
+version = "1.6.1"
+features = ["v4"]
+
+[dev-dependencies]
+env_logger = "0.10.1"
+
+[lib]
+path = "src/lib.rs"
+
+[[example]]
+name = "xenclient-boot"
+path = "examples/boot.rs"

xen/xenclient/examples/boot.rs

@@ -0,0 +1,30 @@
+use std::{env, process};
+use xenclient::{DomainConfig, XenClient, XenClientError};
+
+fn main() -> Result<(), XenClientError> {
+    env_logger::init();
+
+    let args: Vec<String> = env::args().collect();
+    if args.len() != 3 {
+        println!("usage: boot <kernel-image> <initrd>");
+        process::exit(1);
+    }
+    let kernel_image_path = args.get(1).expect("argument not specified");
+    let initrd_path = args.get(2).expect("argument not specified");
+    let mut client = XenClient::open()?;
+    let config = DomainConfig {
+        backend_domid: 0,
+        name: "xenclient-test",
+        max_vcpus: 1,
+        mem_mb: 512,
+        kernel_path: kernel_image_path.as_str(),
+        initrd_path: initrd_path.as_str(),
+        cmdline: "debug elevator=noop",
+        disks: vec![],
+        filesystems: vec![],
+        extra_keys: vec![],
+    };
+    let domid = client.create(&config)?;
+    println!("created domain {}", domid);
+    Ok(())
+}

xen/xenclient/src/boot.rs

@@ -0,0 +1,394 @@
+use crate::mem::PhysicalPages;
+use crate::sys::{GrantEntry, XEN_PAGE_SHIFT};
+use crate::XenClientError;
+use libc::munmap;
+use log::debug;
+use slice_copy::copy;
+
+use std::ffi::c_void;
+use std::slice;
+use xencall::XenCall;
+
+pub trait BootImageLoader {
+    fn parse(&self) -> Result<BootImageInfo, XenClientError>;
+    fn load(&self, image_info: &BootImageInfo, dst: &mut [u8]) -> Result<(), XenClientError>;
+}
+
+pub const XEN_UNSET_ADDR: u64 = -1i64 as u64;
+
+#[derive(Debug)]
+pub struct BootImageInfo {
+    pub start: u64,
+    pub virt_base: u64,
+    pub virt_kstart: u64,
+    pub virt_kend: u64,
+    pub virt_hypercall: u64,
+    pub virt_entry: u64,
+    pub virt_p2m_base: u64,
+    pub unmapped_initrd: bool,
+}
+
+pub struct BootSetup<'a> {
+    pub(crate) call: &'a XenCall,
+    pub phys: PhysicalPages<'a>,
+    pub(crate) domid: u32,
+    pub(crate) virt_alloc_end: u64,
+    pub(crate) pfn_alloc_end: u64,
+    pub(crate) virt_pgtab_end: u64,
+    pub(crate) total_pages: u64,
+}
+
+#[derive(Debug)]
+pub struct DomainSegment {
+    pub(crate) vstart: u64,
+    vend: u64,
+    pub pfn: u64,
+    pub(crate) addr: u64,
+    pub(crate) size: u64,
+    pub(crate) pages: u64,
+}
+
+#[derive(Debug)]
+pub struct BootState {
+    pub kernel_segment: DomainSegment,
+    pub start_info_segment: DomainSegment,
+    pub xenstore_segment: DomainSegment,
+    pub console_segment: DomainSegment,
+    pub boot_stack_segment: DomainSegment,
+    pub p2m_segment: DomainSegment,
+    pub page_table_segment: DomainSegment,
+    pub image_info: BootImageInfo,
+    pub shared_info_frame: u64,
+    pub initrd_segment: DomainSegment,
+    pub store_evtchn: u32,
+    pub console_evtchn: u32,
+}
+
+impl BootSetup<'_> {
+    pub fn new(call: &XenCall, domid: u32) -> BootSetup {
+        BootSetup {
+            call,
+            phys: PhysicalPages::new(call, domid),
+            domid,
+            virt_alloc_end: 0,
+            pfn_alloc_end: 0,
+            virt_pgtab_end: 0,
+            total_pages: 0,
+        }
+    }
+
+    fn initialize_memory(
+        &mut self,
+        arch: &mut dyn ArchBootSetup,
+        total_pages: u64,
+    ) -> Result<(), XenClientError> {
+        self.call.set_address_size(self.domid, 64)?;
+        arch.meminit(self, total_pages)?;
+        Ok(())
+    }
+
+    pub fn initialize(
+        &mut self,
+        arch: &mut dyn ArchBootSetup,
+        image_loader: &dyn BootImageLoader,
+        initrd: &[u8],
+        max_vcpus: u32,
+        mem_mb: u64,
+    ) -> Result<BootState, XenClientError> {
+        debug!(
+            "BootSetup initialize max_vcpus={:?} mem_mb={:?}",
+            max_vcpus, mem_mb
+        );
+
+        let total_pages = mem_mb << (20 - arch.page_shift());
+        self.initialize_memory(arch, total_pages)?;
+
+        let image_info = image_loader.parse()?;
+        debug!("BootSetup initialize image_info={:?}", image_info);
+        self.virt_alloc_end = image_info.virt_base;
+        let kernel_segment = self.load_kernel_segment(arch, image_loader, &image_info)?;
+        let mut p2m_segment: Option<DomainSegment> = None;
+        if image_info.virt_p2m_base >= image_info.virt_base
+            || (image_info.virt_p2m_base & ((1 << arch.page_shift()) - 1)) != 0
+        {
+            p2m_segment = Some(arch.alloc_p2m_segment(self, &image_info)?);
+        }
+        let start_info_segment = self.alloc_page(arch)?;
+        let xenstore_segment = self.alloc_page(arch)?;
+        let console_segment = self.alloc_page(arch)?;
+        let page_table_segment = arch.alloc_page_tables(self, &image_info)?;
+        let boot_stack_segment = self.alloc_page(arch)?;
+
+        if self.virt_pgtab_end > 0 {
+            self.alloc_padding_pages(arch, self.virt_pgtab_end)?;
+        }
+
+        let mut initrd_segment: Option<DomainSegment> = None;
+        if !image_info.unmapped_initrd {
+            initrd_segment = Some(self.alloc_module(arch, initrd)?);
+        }
+        if p2m_segment.is_none() {
+            let mut segment = arch.alloc_p2m_segment(self, &image_info)?;
+            segment.vstart = image_info.virt_p2m_base;
+            p2m_segment = Some(segment);
+        }
+        let p2m_segment = p2m_segment.unwrap();
+
+        if image_info.unmapped_initrd {
+            initrd_segment = Some(self.alloc_module(arch, initrd)?);
+        }
+
+        let initrd_segment = initrd_segment.unwrap();
+        let store_evtchn = self.call.evtchn_alloc_unbound(self.domid, 0)?;
+        let console_evtchn = self.call.evtchn_alloc_unbound(self.domid, 0)?;
+        let state = BootState {
+            kernel_segment,
+            start_info_segment,
+            xenstore_segment,
+            console_segment,
+            boot_stack_segment,
+            p2m_segment,
+            page_table_segment,
+            image_info,
+            initrd_segment,
+            store_evtchn,
+            console_evtchn,
+            shared_info_frame: 0,
+        };
+        debug!("BootSetup initialize state={:?}", state);
+        Ok(state)
+    }
+
+    pub fn boot(
+        &mut self,
+        arch: &mut dyn ArchBootSetup,
+        state: &mut BootState,
+        cmdline: &str,
+    ) -> Result<(), XenClientError> {
+        let domain_info = self.call.get_domain_info(self.domid)?;
+        let shared_info_frame = domain_info.shared_info_frame;
+        state.shared_info_frame = shared_info_frame;
+        arch.setup_page_tables(self, state)?;
+        arch.setup_start_info(self, state, cmdline)?;
+        arch.setup_hypercall_page(self, &state.image_info)?;
+        arch.bootlate(self, state)?;
+        arch.setup_shared_info(self, state.shared_info_frame)?;
+        arch.vcpu(self, state)?;
+        self.phys.unmap_all()?;
+        self.gnttab_seed(state)?;
+        Ok(())
+    }
+
+    fn gnttab_seed(&mut self, state: &mut BootState) -> Result<(), XenClientError> {
+        let console_gfn = self.phys.p2m[state.console_segment.pfn as usize];
+        let xenstore_gfn = self.phys.p2m[state.xenstore_segment.pfn as usize];
+        let addr = self
+            .call
+            .mmap(0, 1 << XEN_PAGE_SHIFT)
+            .ok_or(XenClientError::new("failed to mmap for resource"))?;
+        self.call.map_resource(self.domid, 1, 0, 0, 1, addr)?;
+        let entries = unsafe { slice::from_raw_parts_mut(addr as *mut GrantEntry, 2) };
+        entries[0].flags = 1 << 0;
+        entries[0].domid = 0;
+        entries[0].frame = console_gfn as u32;
+        entries[1].flags = 1 << 0;
+        entries[1].domid = 0;
+        entries[1].frame = xenstore_gfn as u32;
+        unsafe {
+            let result = munmap(addr as *mut c_void, 1 << XEN_PAGE_SHIFT);
+            if result != 0 {
+                return Err(XenClientError::new("failed to unmap resource"));
+            }
+        }
+        Ok(())
+    }
+
+    fn load_kernel_segment(
+        &mut self,
+        arch: &mut dyn ArchBootSetup,
+        image_loader: &dyn BootImageLoader,
+        image_info: &BootImageInfo,
+    ) -> Result<DomainSegment, XenClientError> {
+        let kernel_segment = self.alloc_segment(
+            arch,
+            image_info.virt_kstart,
+            image_info.virt_kend - image_info.virt_kstart,
+        )?;
+        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(image_info, kernel_segment_slice)?;
+        Ok(kernel_segment)
+    }
+
+    pub(crate) fn round_up(addr: u64, mask: u64) -> u64 {
+        addr | mask
+    }
+
+    pub(crate) fn bits_to_mask(bits: u64) -> u64 {
+        (1 << bits) - 1
+    }
+
+    pub(crate) fn alloc_segment(
+        &mut self,
+        arch: &mut dyn ArchBootSetup,
+        start: u64,
+        size: u64,
+    ) -> Result<DomainSegment, XenClientError> {
+        if start > 0 {
+            self.alloc_padding_pages(arch, start)?;
+        }
+
+        let page_size: u32 = (1i64 << XEN_PAGE_SHIFT) as u32;
+        let pages = (size + page_size as u64 - 1) / page_size as u64;
+        let start = self.virt_alloc_end;
+
+        let mut segment = DomainSegment {
+            vstart: start,
+            vend: 0,
+            pfn: self.pfn_alloc_end,
+            addr: 0,
+            size,
+            pages,
+        };
+
+        self.chk_alloc_pages(arch, pages)?;
+
+        let ptr = self.phys.pfn_to_ptr(segment.pfn, pages)?;
+        segment.addr = ptr;
+        let slice = unsafe {
+            slice::from_raw_parts_mut(ptr as *mut u8, (pages * page_size as u64) as usize)
+        };
+        slice.fill(0);
+        segment.vend = self.virt_alloc_end;
+        debug!(
+            "BootSetup alloc_segment {:#x} -> {:#x} (pfn {:#x} + {:#x} pages)",
+            start, segment.vend, segment.pfn, pages
+        );
+        Ok(segment)
+    }
+
+    fn alloc_page(
+        &mut self,
+        arch: &mut dyn ArchBootSetup,
+    ) -> Result<DomainSegment, XenClientError> {
+        let start = self.virt_alloc_end;
+        let pfn = self.pfn_alloc_end;
+
+        self.chk_alloc_pages(arch, 1)?;
+        debug!("BootSetup alloc_page {:#x} (pfn {:#x})", start, pfn);
+        Ok(DomainSegment {
+            vstart: start,
+            vend: (start + arch.page_size()) - 1,
+            pfn,
+            addr: 0,
+            size: 0,
+            pages: 1,
+        })
+    }
+
+    fn alloc_module(
+        &mut self,
+        arch: &mut dyn ArchBootSetup,
+        buffer: &[u8],
+    ) -> Result<DomainSegment, XenClientError> {
+        let segment = self.alloc_segment(arch, 0, buffer.len() as u64)?;
+        let slice = unsafe { slice::from_raw_parts_mut(segment.addr as *mut u8, buffer.len()) };
+        copy(slice, buffer);
+        Ok(segment)
+    }
+
+    fn alloc_padding_pages(
+        &mut self,
+        arch: &mut dyn ArchBootSetup,
+        boundary: u64,
+    ) -> Result<(), XenClientError> {
+        if (boundary & (arch.page_size() - 1)) != 0 {
+            return Err(XenClientError::new(
+                format!("segment boundary isn't page aligned: {:#x}", boundary).as_str(),
+            ));
+        }
+
+        if boundary < self.virt_alloc_end {
+            return Err(XenClientError::new(
+                format!("segment boundary too low: {:#x})", boundary).as_str(),
+            ));
+        }
+        let pages = (boundary - self.virt_alloc_end) / arch.page_size();
+        self.chk_alloc_pages(arch, pages)?;
+        Ok(())
+    }
+
+    fn chk_alloc_pages(
+        &mut self,
+        arch: &mut dyn ArchBootSetup,
+        pages: u64,
+    ) -> Result<(), XenClientError> {
+        if pages > self.total_pages
+            || self.pfn_alloc_end > self.total_pages
+            || pages > self.total_pages - self.pfn_alloc_end
+        {
+            return Err(XenClientError::new(
+                format!(
+                    "segment too large: pages={} total_pages={} pfn_alloc_end={}",
+                    pages, self.total_pages, self.pfn_alloc_end
+                )
+                .as_str(),
+            ));
+        }
+
+        self.pfn_alloc_end += pages;
+        self.virt_alloc_end += pages * arch.page_size();
+        Ok(())
+    }
+}
+
+pub trait ArchBootSetup {
+    fn page_size(&mut self) -> u64;
+    fn page_shift(&mut self) -> u64;
+
+    fn alloc_p2m_segment(
+        &mut self,
+        setup: &mut BootSetup,
+        image_info: &BootImageInfo,
+    ) -> Result<DomainSegment, XenClientError>;
+
+    fn alloc_page_tables(
+        &mut self,
+        setup: &mut BootSetup,
+        image_info: &BootImageInfo,
+    ) -> Result<DomainSegment, XenClientError>;
+
+    fn setup_page_tables(
+        &mut self,
+        setup: &mut BootSetup,
+        state: &mut BootState,
+    ) -> Result<(), XenClientError>;
+
+    fn setup_start_info(
+        &mut self,
+        setup: &mut BootSetup,
+        state: &BootState,
+        cmdline: &str,
+    ) -> Result<(), XenClientError>;
+
+    fn setup_shared_info(
+        &mut self,
+        setup: &mut BootSetup,
+        shared_info_frame: u64,
+    ) -> Result<(), XenClientError>;
+
+    fn setup_hypercall_page(
+        &mut self,
+        setup: &mut BootSetup,
+        image_info: &BootImageInfo,
+    ) -> Result<(), XenClientError>;
+
+    fn meminit(&mut self, setup: &mut BootSetup, total_pages: u64) -> Result<(), XenClientError>;
+    fn bootlate(
+        &mut self,
+        setup: &mut BootSetup,
+        state: &mut BootState,
+    ) -> Result<(), XenClientError>;
+    fn vcpu(&mut self, setup: &mut BootSetup, state: &mut BootState) -> Result<(), XenClientError>;
+}

xen/xenclient/src/elfloader.rs

@@ -0,0 +1,331 @@
+use crate::boot::{BootImageInfo, BootImageLoader, XEN_UNSET_ADDR};
+use crate::sys::{
+    XEN_ELFNOTE_ENTRY, XEN_ELFNOTE_HYPERCALL_PAGE, XEN_ELFNOTE_INIT_P2M, XEN_ELFNOTE_MOD_START_PFN,
+    XEN_ELFNOTE_PADDR_OFFSET, XEN_ELFNOTE_TYPES, XEN_ELFNOTE_VIRT_BASE,
+};
+use crate::XenClientError;
+use elf::abi::{PF_R, PF_W, PF_X, PT_LOAD, SHT_NOTE};
+use elf::endian::AnyEndian;
+use elf::note::Note;
+use elf::{ElfBytes, ParseError};
+use flate2::bufread::GzDecoder;
+use log::debug;
+use memchr::memmem::find_iter;
+use slice_copy::copy;
+use std::collections::HashMap;
+use std::ffi::{FromVecWithNulError, IntoStringError};
+use std::io::{BufReader, Read};
+use std::mem::size_of;
+use xz2::bufread::XzDecoder;
+
+impl From<ParseError> for XenClientError {
+    fn from(value: ParseError) -> Self {
+        XenClientError::new(value.to_string().as_str())
+    }
+}
+
+impl From<FromVecWithNulError> for XenClientError {
+    fn from(value: FromVecWithNulError) -> Self {
+        XenClientError::new(value.to_string().as_str())
+    }
+}
+
+impl From<IntoStringError> for XenClientError {
+    fn from(value: IntoStringError) -> Self {
+        XenClientError::new(value.to_string().as_str())
+    }
+}
+
+pub struct ElfImageLoader {
+    data: Vec<u8>,
+}
+
+fn xen_note_value_as_u64(endian: AnyEndian, value: &[u8]) -> Option<u64> {
+    let bytes = value.to_vec();
+    match value.len() {
+        1 => {
+            let bytes: Option<[u8; size_of::<u8>()]> = bytes.try_into().ok();
+            Some(match endian {
+                AnyEndian::Little => u8::from_le_bytes(bytes?),
+                AnyEndian::Big => u8::from_be_bytes(bytes?),
+            } as u64)
+        }
+        2 => {
+            let bytes: Option<[u8; size_of::<u16>()]> = bytes.try_into().ok();
+            Some(match endian {
+                AnyEndian::Little => u16::from_le_bytes(bytes?),
+                AnyEndian::Big => u16::from_be_bytes(bytes?),
+            } as u64)
+        }
+        4 => {
+            let bytes: Option<[u8; size_of::<u32>()]> = bytes.try_into().ok();
+            Some(match endian {
+                AnyEndian::Little => u32::from_le_bytes(bytes?),
+                AnyEndian::Big => u32::from_be_bytes(bytes?),
+            } as u64)
+        }
+        8 => {
+            let bytes: Option<[u8; size_of::<u64>()]> = bytes.try_into().ok();
+            Some(match endian {
+                AnyEndian::Little => u64::from_le_bytes(bytes?),
+                AnyEndian::Big => u64::from_be_bytes(bytes?),
+            })
+        }
+        _ => None,
+    }
+}
+
+impl ElfImageLoader {
+    pub fn new(data: Vec<u8>) -> ElfImageLoader {
+        ElfImageLoader { data }
+    }
+
+    pub fn load_file(path: &str) -> Result<ElfImageLoader, XenClientError> {
+        let data = std::fs::read(path)?;
+        Ok(ElfImageLoader::new(data))
+    }
+
+    pub fn load_gz(data: &[u8]) -> Result<ElfImageLoader, XenClientError> {
+        let buff = BufReader::new(data);
+        let image = ElfImageLoader::read_one_stream(&mut GzDecoder::new(buff))?;
+        Ok(ElfImageLoader::new(image))
+    }
+
+    pub fn load_xz(data: &[u8]) -> Result<ElfImageLoader, XenClientError> {
+        let buff = BufReader::new(data);
+        let image = ElfImageLoader::read_one_stream(&mut XzDecoder::new(buff))?;
+        Ok(ElfImageLoader::new(image))
+    }
+
+    fn read_one_stream(read: &mut dyn Read) -> Result<Vec<u8>, XenClientError> {
+        let mut result: Vec<u8> = Vec::new();
+        let mut buffer = [0u8; 8192];
+
+        loop {
+            match read.read(&mut buffer) {
+                Ok(size) => {
+                    if size == 0 {
+                        break;
+                    }
+                    result.extend_from_slice(&buffer[0..size])
+                }
+                Err(error) => {
+                    if !result.is_empty() {
+                        break;
+                    }
+                    return Err(XenClientError::from(error));
+                }
+            }
+        }
+        Ok(result)
+    }
+
+    pub fn load_file_gz(path: &str) -> Result<ElfImageLoader, XenClientError> {
+        let file = std::fs::read(path)?;
+        ElfImageLoader::load_gz(file.as_slice())
+    }
+
+    pub fn load_file_xz(path: &str) -> Result<ElfImageLoader, XenClientError> {
+        let file = std::fs::read(path)?;
+        ElfImageLoader::load_xz(file.as_slice())
+    }
+
+    pub fn load_file_kernel(path: &str) -> Result<ElfImageLoader, XenClientError> {
+        let file = std::fs::read(path)?;
+
+        for start in find_iter(file.as_slice(), &[0x1f, 0x8b]) {
+            if let Ok(elf) = ElfImageLoader::load_gz(&file[start..]) {
+                return Ok(elf);
+            }
+        }
+
+        for start in find_iter(file.as_slice(), &[0xfd, 0x37, 0x7a, 0x58]) {
+            if let Ok(elf) = ElfImageLoader::load_xz(&file[start..]) {
+                return Ok(elf);
+            }
+        }
+
+        Err(XenClientError::new(
+            "Unable to parse kernel image: unknown compression type",
+        ))
+    }
+}
+
+struct ElfNoteValue {
+    value: u64,
+}
+
+impl BootImageLoader for ElfImageLoader {
+    fn parse(&self) -> Result<BootImageInfo, XenClientError> {
+        let elf = ElfBytes::<AnyEndian>::minimal_parse(self.data.as_slice())?;
+        let headers = elf.section_headers().ok_or(XenClientError::new(
+            "Unable to parse kernel image: section headers not found.",
+        ))?;
+        let mut linux_notes: HashMap<u64, Vec<u8>> = HashMap::new();
+        let mut xen_notes: HashMap<u64, ElfNoteValue> = HashMap::new();
+
+        for header in headers {
+            if header.sh_type != SHT_NOTE {
+                continue;
+            }
+
+            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());
+                    }
+
+                    if note.name == "Xen" {
+                        for typ in XEN_ELFNOTE_TYPES {
+                            if typ.id != note.n_type {
+                                continue;
+                            }
+
+                            let value = if !typ.is_string {
+                                xen_note_value_as_u64(elf.ehdr.endianness, note.desc).unwrap_or(0)
+                            } else {
+                                0
+                            };
+
+                            xen_notes.insert(typ.id, ElfNoteValue { value });
+                        }
+                        continue;
+                    }
+                }
+            }
+        }
+
+        if linux_notes.is_empty() {
+            return Err(XenClientError::new(
+                "Provided kernel does not appear to be a Linux kernel image.",
+            ));
+        }
+
+        if xen_notes.is_empty() {
+            return Err(XenClientError::new(
+                "Provided kernel does not have Xen support.",
+            ));
+        }
+
+        let paddr_offset = xen_notes
+            .get(&XEN_ELFNOTE_PADDR_OFFSET)
+            .ok_or(XenClientError::new(
+                "Unable to find paddr_offset note in kernel.",
+            ))?
+            .value;
+        let virt_base = xen_notes
+            .get(&XEN_ELFNOTE_VIRT_BASE)
+            .ok_or(XenClientError::new(
+                "Unable to find virt_base note in kernel.",
+            ))?
+            .value;
+        let entry = xen_notes
+            .get(&XEN_ELFNOTE_ENTRY)
+            .ok_or(XenClientError::new("Unable to find entry note in kernel."))?
+            .value;
+        let virt_hypercall = xen_notes
+            .get(&XEN_ELFNOTE_HYPERCALL_PAGE)
+            .ok_or(XenClientError::new(
+                "Unable to find hypercall_page note in kernel.",
+            ))?
+            .value;
+        let init_p2m = xen_notes
+            .get(&XEN_ELFNOTE_INIT_P2M)
+            .ok_or(XenClientError::new(
+                "Unable to find init_p2m note in kernel.",
+            ))?
+            .value;
+        let mod_start_pfn = xen_notes
+            .get(&XEN_ELFNOTE_MOD_START_PFN)
+            .ok_or(XenClientError::new(
+                "Unable to find mod_start_pfn note in kernel.",
+            ))?
+            .value;
+
+        let mut start: u64 = u64::MAX;
+        let mut end: u64 = 0;
+
+        let segments = elf.segments().ok_or(XenClientError::new(
+            "Unable to parse kernel image: segments not found.",
+        ))?;
+
+        for header in segments {
+            if (header.p_type != PT_LOAD) || (header.p_flags & (PF_R | PF_W | PF_X)) == 0 {
+                continue;
+            }
+            let paddr = header.p_paddr;
+            let memsz = header.p_memsz;
+            if start > paddr {
+                start = paddr;
+            }
+
+            if end < paddr + memsz {
+                end = paddr + memsz;
+            }
+        }
+
+        if paddr_offset != XEN_UNSET_ADDR && virt_base == XEN_UNSET_ADDR {
+            return Err(XenClientError::new(
+                "Unable to load kernel image: paddr_offset set but virt_base is unset.",
+            ));
+        }
+
+        let virt_offset = virt_base - paddr_offset;
+        let virt_kstart = start + virt_offset;
+        let virt_kend = end + virt_offset;
+        let virt_entry = entry;
+
+        let image_info = BootImageInfo {
+            start,
+            virt_base,
+            virt_kstart,
+            virt_kend,
+            virt_hypercall,
+            virt_entry,
+            virt_p2m_base: init_p2m,
+            unmapped_initrd: mod_start_pfn != 0,
+        };
+        Ok(image_info)
+    }
+
+    fn load(&self, image_info: &BootImageInfo, dst: &mut [u8]) -> Result<(), XenClientError> {
+        let elf = ElfBytes::<AnyEndian>::minimal_parse(self.data.as_slice())?;
+        let segments = elf.segments().ok_or(XenClientError::new(
+            "Unable to parse kernel image: segments not found.",
+        ))?;
+
+        debug!(
+            "ElfImageLoader load dst={:#x} segments={}",
+            dst.as_ptr() as u64,
+            segments.len()
+        );
+        for header in segments {
+            let paddr = header.p_paddr;
+            let filesz = header.p_filesz;
+            let memsz = header.p_memsz;
+            let base_offset = paddr - image_info.start;
+            let data = elf.segment_data(&header)?;
+            let segment_dst = &mut dst[base_offset as usize..];
+            let copy_slice = &data[0..filesz as usize];
+            debug!(
+                "ElfImageLoader load copy hdr={:?} dst={:#x} len={}",
+                header,
+                copy_slice.as_ptr() as u64,
+                copy_slice.len()
+            );
+            copy(segment_dst, copy_slice);
+            if memsz - filesz > 0 {
+                let remaining = &mut segment_dst[filesz as usize..(memsz - filesz) as usize];
+                debug!(
+                    "ElfImageLoader load fill_zero hdr={:?} dst={:#x} len={}",
+                    header.p_offset,
+                    remaining.as_ptr() as u64,
+                    remaining.len()
+                );
+                remaining.fill(0);
+            }
+        }
+        Ok(())
+    }
+}

xen/xenclient/src/lib.rs

@@ -0,0 +1,646 @@
+pub mod boot;
+pub mod elfloader;
+pub mod mem;
+pub mod sys;
+mod x86;
+
+use crate::boot::BootSetup;
+use crate::elfloader::ElfImageLoader;
+use crate::x86::X86BootSetup;
+use log::{trace, warn};
+use std::error::Error;
+use std::fmt::{Display, Formatter};
+use std::fs::{read, File, OpenOptions};
+use std::path::PathBuf;
+use std::str::FromStr;
+use std::string::FromUtf8Error;
+use std::thread;
+use std::time::Duration;
+use uuid::Uuid;
+use xencall::sys::CreateDomain;
+use xencall::{XenCall, XenCallError};
+use xenevtchn::EventChannelError;
+use xenstore::bus::XsdBusError;
+use xenstore::client::{
+    XsPermission, XsdClient, XsdInterface, XS_PERM_NONE, XS_PERM_READ, XS_PERM_READ_WRITE,
+};
+
+pub struct XenClient {
+    pub store: XsdClient,
+    call: XenCall,
+}
+
+#[derive(Debug)]
+pub struct XenClientError {
+    message: String,
+}
+
+impl XenClientError {
+    pub fn new(msg: &str) -> XenClientError {
+        XenClientError {
+            message: msg.to_string(),
+        }
+    }
+}
+
+impl Display for XenClientError {
+    fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}", self.message)
+    }
+}
+
+impl Error for XenClientError {
+    fn description(&self) -> &str {
+        &self.message
+    }
+}
+
+impl From<std::io::Error> for XenClientError {
+    fn from(value: std::io::Error) -> Self {
+        XenClientError::new(value.to_string().as_str())
+    }
+}
+
+impl From<XsdBusError> for XenClientError {
+    fn from(value: XsdBusError) -> Self {
+        XenClientError::new(value.to_string().as_str())
+    }
+}
+
+impl From<XenCallError> for XenClientError {
+    fn from(value: XenCallError) -> Self {
+        XenClientError::new(value.to_string().as_str())
+    }
+}
+
+impl From<FromUtf8Error> for XenClientError {
+    fn from(value: FromUtf8Error) -> Self {
+        XenClientError::new(value.to_string().as_str())
+    }
+}
+
+impl From<EventChannelError> for XenClientError {
+    fn from(value: EventChannelError) -> Self {
+        XenClientError::new(value.to_string().as_str())
+    }
+}
+
+#[derive(Debug)]
+pub struct BlockDeviceRef {
+    pub path: String,
+    pub major: u32,
+    pub minor: u32,
+}
+
+#[derive(Debug)]
+pub struct DomainDisk<'a> {
+    pub vdev: &'a str,
+    pub block: &'a BlockDeviceRef,
+    pub writable: bool,
+}
+
+#[derive(Debug)]
+pub struct DomainFilesystem<'a> {
+    pub path: &'a str,
+    pub tag: &'a str,
+}
+
+#[derive(Debug)]
+pub struct DomainConfig<'a> {
+    pub backend_domid: u32,
+    pub name: &'a str,
+    pub max_vcpus: u32,
+    pub mem_mb: u64,
+    pub kernel_path: &'a str,
+    pub initrd_path: &'a str,
+    pub cmdline: &'a str,
+    pub disks: Vec<DomainDisk<'a>>,
+    pub filesystems: Vec<DomainFilesystem<'a>>,
+    pub extra_keys: Vec<(String, String)>,
+}
+
+impl XenClient {
+    pub fn open() -> Result<XenClient, XenClientError> {
+        let store = XsdClient::open()?;
+        let call = XenCall::open()?;
+        Ok(XenClient { store, call })
+    }
+
+    pub fn create(&mut self, config: &DomainConfig) -> Result<u32, XenClientError> {
+        let domain = CreateDomain {
+            max_vcpus: config.max_vcpus,
+            ..Default::default()
+        };
+        let domid = self.call.create_domain(domain)?;
+        match self.init(domid, &domain, config) {
+            Ok(_) => Ok(domid),
+            Err(err) => {
+                // ignore since destroying a domain is best
+                // effort when an error occurs
+                let _ = self.destroy(domid);
+                Err(err)
+            }
+        }
+    }
+
+    pub fn destroy(&mut self, domid: u32) -> Result<(), XenClientError> {
+        if let Err(err) = self.destroy_store(domid) {
+            warn!("failed to destroy store for domain {}: {}", domid, err);
+        }
+        self.call.destroy_domain(domid)?;
+        Ok(())
+    }
+
+    fn destroy_store(&mut self, domid: u32) -> Result<(), XenClientError> {
+        let dom_path = self.store.get_domain_path(domid)?;
+        let vm_path = self.store.read_string(&format!("{}/vm", dom_path))?;
+        if vm_path.is_empty() {
+            return Err(XenClientError::new(
+                "cannot destroy domain that doesn't exist",
+            ));
+        }
+
+        let mut backend_paths: Vec<String> = Vec::new();
+        let console_frontend_path = format!("{}/console", dom_path);
+        let console_backend_path = self
+            .store
+            .read_string_optional(format!("{}/backend", console_frontend_path).as_str())?;
+
+        for device_category in self
+            .store
+            .list_any(format!("{}/device", dom_path).as_str())?
+        {
+            for device_id in self
+                .store
+                .list_any(format!("{}/device/{}", dom_path, device_category).as_str())?
+            {
+                let device_path = format!("{}/device/{}/{}", dom_path, device_category, device_id);
+                let backend_path = self
+                    .store
+                    .read_string(format!("{}/backend", device_path).as_str())?;
+                backend_paths.push(backend_path);
+            }
+        }
+
+        for backend in &backend_paths {
+            let state_path = format!("{}/state", backend);
+            let online_path = format!("{}/online", backend);
+            let mut tx = self.store.transaction()?;
+            let state = tx.read_string(&state_path)?;
+            if state.is_empty() {
+                break;
+            }
+            tx.write_string(&online_path, "0")?;
+            if !state.is_empty() && u32::from_str(&state).unwrap_or(0) != 6 {
+                tx.write_string(&state_path, "5")?;
+            }
+            tx.commit()?;
+
+            let mut count: u32 = 0;
+            loop {
+                if count >= 100 {
+                    warn!("unable to safely destroy backend: {}", backend);
+                    break;
+                }
+                let state = self.store.read_string(&state_path)?;
+                let state = i64::from_str(&state).unwrap_or(-1);
+                if state == 6 {
+                    break;
+                }
+                thread::sleep(Duration::from_millis(100));
+                count += 1;
+            }
+        }
+
+        let mut tx = self.store.transaction()?;
+        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(XenClientError::new("unable to get parent of backend path"))?;
+            tx.rm(parent
+                .to_str()
+                .ok_or(XenClientError::new("unable to convert parent to string"))?)?;
+        }
+        tx.rm(&vm_path)?;
+        tx.rm(&dom_path)?;
+        tx.commit()?;
+        Ok(())
+    }
+
+    fn init(
+        &mut self,
+        domid: u32,
+        domain: &CreateDomain,
+        config: &DomainConfig,
+    ) -> Result<(), XenClientError> {
+        trace!(
+            "XenClient init domid={} domain={:?} config={:?}",
+            domid,
+            domain,
+            config
+        );
+        let backend_dom_path = self.store.get_domain_path(0)?;
+        let dom_path = self.store.get_domain_path(domid)?;
+        let uuid_string = Uuid::from_bytes(domain.handle).to_string();
+        let vm_path = format!("/vm/{}", uuid_string);
+
+        let ro_perm = &[
+            XsPermission {
+                id: 0,
+                perms: XS_PERM_NONE,
+            },
+            XsPermission {
+                id: domid,
+                perms: XS_PERM_READ,
+            },
+        ];
+
+        let rw_perm = &[XsPermission {
+            id: domid,
+            perms: XS_PERM_READ_WRITE,
+        }];
+
+        let no_perm = &[XsPermission {
+            id: 0,
+            perms: XS_PERM_NONE,
+        }];
+
+        {
+            let mut tx = self.store.transaction()?;
+
+            tx.rm(dom_path.as_str())?;
+            tx.mknod(dom_path.as_str(), ro_perm)?;
+
+            tx.rm(vm_path.as_str())?;
+            tx.mknod(vm_path.as_str(), ro_perm)?;
+
+            tx.mknod(vm_path.as_str(), no_perm)?;
+            tx.mknod(format!("{}/device", vm_path).as_str(), no_perm)?;
+
+            tx.write_string(format!("{}/vm", dom_path).as_str(), &vm_path)?;
+
+            tx.mknod(format!("{}/cpu", dom_path).as_str(), ro_perm)?;
+            tx.mknod(format!("{}/memory", dom_path).as_str(), ro_perm)?;
+
+            tx.mknod(format!("{}/control", dom_path).as_str(), ro_perm)?;
+
+            tx.mknod(format!("{}/control/shutdown", dom_path).as_str(), rw_perm)?;
+            tx.mknod(
+                format!("{}/control/feature-poweroff", dom_path).as_str(),
+                rw_perm,
+            )?;
+            tx.mknod(
+                format!("{}/control/feature-reboot", dom_path).as_str(),
+                rw_perm,
+            )?;
+            tx.mknod(
+                format!("{}/control/feature-suspend", dom_path).as_str(),
+                rw_perm,
+            )?;
+            tx.mknod(format!("{}/control/sysrq", dom_path).as_str(), rw_perm)?;
+
+            tx.mknod(format!("{}/data", dom_path).as_str(), rw_perm)?;
+            tx.mknod(format!("{}/drivers", dom_path).as_str(), rw_perm)?;
+            tx.mknod(format!("{}/feature", dom_path).as_str(), rw_perm)?;
+            tx.mknod(format!("{}/attr", dom_path).as_str(), rw_perm)?;
+            tx.mknod(format!("{}/error", dom_path).as_str(), rw_perm)?;
+
+            tx.write_string(
+                format!("{}/uuid", vm_path).as_str(),
+                &Uuid::from_bytes(domain.handle).to_string(),
+            )?;
+            tx.write_string(format!("{}/name", dom_path).as_str(), config.name)?;
+            tx.write_string(format!("{}/name", vm_path).as_str(), config.name)?;
+
+            for (key, value) in &config.extra_keys {
+                tx.write_string(format!("{}/{}", dom_path, key).as_str(), value)?;
+            }
+
+            tx.commit()?;
+        }
+
+        self.call.set_max_vcpus(domid, config.max_vcpus)?;
+        self.call.set_max_mem(domid, config.mem_mb * 1024)?;
+        let image_loader = ElfImageLoader::load_file_kernel(config.kernel_path)?;
+
+        let console_evtchn: u32;
+        let xenstore_evtchn: u32;
+        let console_mfn: u64;
+        let xenstore_mfn: u64;
+
+        {
+            let mut boot = BootSetup::new(&self.call, domid);
+            let mut arch = X86BootSetup::new();
+            let initrd = read(config.initrd_path)?;
+            let mut state = boot.initialize(
+                &mut arch,
+                &image_loader,
+                initrd.as_slice(),
+                config.max_vcpus,
+                config.mem_mb,
+            )?;
+            boot.boot(&mut arch, &mut state, config.cmdline)?;
+            console_evtchn = state.console_evtchn;
+            xenstore_evtchn = state.store_evtchn;
+            console_mfn = boot.phys.p2m[state.console_segment.pfn as usize];
+            xenstore_mfn = boot.phys.p2m[state.xenstore_segment.pfn as usize];
+        }
+
+        {
+            let mut tx = self.store.transaction()?;
+            tx.write_string(format!("{}/image/os_type", vm_path).as_str(), "linux")?;
+            tx.write_string(
+                format!("{}/image/kernel", vm_path).as_str(),
+                config.kernel_path,
+            )?;
+            tx.write_string(
+                format!("{}/image/ramdisk", vm_path).as_str(),
+                config.initrd_path,
+            )?;
+            tx.write_string(
+                format!("{}/image/cmdline", vm_path).as_str(),
+                config.cmdline,
+            )?;
+
+            tx.write_string(
+                format!("{}/memory/static-max", dom_path).as_str(),
+                &(config.mem_mb * 1024).to_string(),
+            )?;
+            tx.write_string(
+                format!("{}/memory/target", dom_path).as_str(),
+                &(config.mem_mb * 1024).to_string(),
+            )?;
+            tx.write_string(format!("{}/memory/videoram", dom_path).as_str(), "0")?;
+            tx.write_string(format!("{}/domid", dom_path).as_str(), &domid.to_string())?;
+            tx.write_string(
+                format!("{}/store/port", dom_path).as_str(),
+                &xenstore_evtchn.to_string(),
+            )?;
+            tx.write_string(
+                format!("{}/store/ring-ref", dom_path).as_str(),
+                &xenstore_mfn.to_string(),
+            )?;
+            for i in 0..config.max_vcpus {
+                let path = format!("{}/cpu/{}", dom_path, i);
+                tx.mkdir(&path)?;
+                tx.set_perms(&path, ro_perm)?;
+                let path = format!("{}/cpu/{}/availability", dom_path, i);
+                tx.write_string(&path, "online")?;
+                tx.set_perms(&path, ro_perm)?;
+            }
+            tx.commit()?;
+        }
+        if !self
+            .store
+            .introduce_domain(domid, xenstore_mfn, xenstore_evtchn)?
+        {
+            return Err(XenClientError::new("failed to introduce domain"));
+        }
+        self.console_device_add(
+            &dom_path,
+            &backend_dom_path,
+            config.backend_domid,
+            domid,
+            console_evtchn,
+            console_mfn,
+        )?;
+        for (index, disk) in config.disks.iter().enumerate() {
+            self.disk_device_add(
+                &dom_path,
+                &backend_dom_path,
+                config.backend_domid,
+                domid,
+                index,
+                disk,
+            )?;
+        }
+        for (index, filesystem) in config.filesystems.iter().enumerate() {
+            self.fs_9p_device_add(
+                &dom_path,
+                &backend_dom_path,
+                config.backend_domid,
+                domid,
+                index,
+                filesystem,
+            )?;
+        }
+        self.call.unpause_domain(domid)?;
+        Ok(())
+    }
+
+    fn disk_device_add(
+        &mut self,
+        dom_path: &str,
+        backend_dom_path: &str,
+        backend_domid: u32,
+        domid: u32,
+        index: usize,
+        disk: &DomainDisk,
+    ) -> Result<(), XenClientError> {
+        let id = (202 << 8) | (index << 4) as u64;
+        let backend_items: Vec<(&str, String)> = vec![
+            ("frontend-id", 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", 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.device_add(
+            "vbd",
+            id,
+            dom_path,
+            backend_dom_path,
+            backend_domid,
+            domid,
+            frontend_items,
+            backend_items,
+        )?;
+        Ok(())
+    }
+
+    fn console_device_add(
+        &mut self,
+        dom_path: &str,
+        backend_dom_path: &str,
+        backend_domid: u32,
+        domid: u32,
+        port: u32,
+        mfn: u64,
+    ) -> Result<(), XenClientError> {
+        let backend_entries = vec![
+            ("frontend-id", domid.to_string()),
+            ("online", "1".to_string()),
+            ("state", "1".to_string()),
+            ("protocol", "vt100".to_string()),
+        ];
+
+        let frontend_entries = vec![
+            ("backend-id", backend_domid.to_string()),
+            ("limit", "1048576".to_string()),
+            ("type", "xenconsoled".to_string()),
+            ("output", "pty".to_string()),
+            ("tty", "".to_string()),
+            ("port", port.to_string()),
+            ("ring-ref", mfn.to_string()),
+        ];
+
+        self.device_add(
+            "console",
+            0,
+            dom_path,
+            backend_dom_path,
+            backend_domid,
+            domid,
+            frontend_entries,
+            backend_entries,
+        )?;
+        Ok(())
+    }
+
+    fn fs_9p_device_add(
+        &mut self,
+        dom_path: &str,
+        backend_dom_path: &str,
+        backend_domid: u32,
+        domid: u32,
+        index: usize,
+        filesystem: &DomainFilesystem,
+    ) -> Result<(), XenClientError> {
+        let id = 90 + index as u64;
+        let backend_items: Vec<(&str, String)> = vec![
+            ("frontend-id", 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", backend_domid.to_string()),
+            ("state", "1".to_string()),
+            ("tag", filesystem.tag.to_string()),
+        ];
+
+        self.device_add(
+            "9pfs",
+            id,
+            dom_path,
+            backend_dom_path,
+            backend_domid,
+            domid,
+            frontend_items,
+            backend_items,
+        )?;
+        Ok(())
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    fn device_add(
+        &mut self,
+        typ: &str,
+        id: u64,
+        dom_path: &str,
+        backend_dom_path: &str,
+        backend_domid: u32,
+        domid: u32,
+        frontend_items: Vec<(&str, String)>,
+        backend_items: Vec<(&str, String)>,
+    ) -> Result<(), XenClientError> {
+        let console_zero = typ == "console" && id == 0;
+
+        let frontend_path = if console_zero {
+            format!("{}/console", dom_path)
+        } else {
+            format!("{}/device/{}/{}", dom_path, typ, id)
+        };
+        let backend_path = format!("{}/backend/{}/{}/{}", backend_dom_path, typ, 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: domid,
+                perms: XS_PERM_NONE,
+            },
+            XsPermission {
+                id: backend_domid,
+                perms: XS_PERM_READ,
+            },
+        ];
+
+        let backend_perms = &[
+            XsPermission {
+                id: backend_domid,
+                perms: XS_PERM_NONE,
+            },
+            XsPermission {
+                id: domid,
+                perms: XS_PERM_READ,
+            },
+        ];
+
+        let mut tx = self.store.transaction()?;
+        tx.mknod(&frontend_path, frontend_perms)?;
+        for (p, value) in &frontend_items {
+            let path = format!("{}/{}", frontend_path, *p);
+            tx.write_string(&path, value)?;
+            if !console_zero {
+                tx.set_perms(&path, frontend_perms)?;
+            }
+        }
+        tx.mknod(&backend_path, backend_perms)?;
+        for (p, value) in &backend_items {
+            let path = format!("{}/{}", backend_path, *p);
+            tx.write_string(&path, value)?;
+        }
+        tx.commit()?;
+        Ok(())
+    }
+
+    pub fn open_console(&mut self, domid: u32) -> Result<(File, File), XenClientError> {
+        let dom_path = self.store.get_domain_path(domid)?;
+        let console_tty_path = format!("{}/console/tty", dom_path);
+        let tty = self
+            .store
+            .read_string_optional(&console_tty_path)?
+            .unwrap_or("".to_string());
+        if tty.is_empty() {
+            return Err(XenClientError::new(&format!(
+                "domain {} does not have a tty",
+                domid
+            )));
+        }
+        let read = OpenOptions::new().read(true).write(false).open(&tty)?;
+        let write = OpenOptions::new().read(false).write(true).open(&tty)?;
+        Ok((read, write))
+    }
+}

xen/xenclient/src/mem.rs

@@ -0,0 +1,187 @@
+use crate::sys::{XEN_PAGE_SHIFT, XEN_PAGE_SIZE};
+use crate::XenClientError;
+use libc::munmap;
+use log::debug;
+use std::ffi::c_void;
+
+use crate::x86::X86_PAGE_SHIFT;
+use xencall::sys::MmapEntry;
+use xencall::XenCall;
+
+#[derive(Debug)]
+pub struct PhysicalPage {
+    pfn: u64,
+    ptr: u64,
+    count: u64,
+}
+
+pub struct PhysicalPages<'a> {
+    domid: u32,
+    pub(crate) p2m: Vec<u64>,
+    call: &'a XenCall,
+    pages: Vec<PhysicalPage>,
+}
+
+impl PhysicalPages<'_> {
+    pub fn new(call: &XenCall, domid: u32) -> PhysicalPages {
+        PhysicalPages {
+            domid,
+            p2m: Vec::new(),
+            call,
+            pages: Vec::new(),
+        }
+    }
+
+    pub fn load_p2m(&mut self, p2m: Vec<u64>) {
+        self.p2m = p2m;
+    }
+
+    pub fn p2m_size(&mut self) -> u64 {
+        self.p2m.len() as u64
+    }
+
+    pub fn pfn_to_ptr(&mut self, pfn: u64, count: u64) -> Result<u64, XenClientError> {
+        for page in &self.pages {
+            if pfn >= page.pfn + page.count {
+                continue;
+            }
+
+            if count > 0 {
+                if (pfn + count) <= page.pfn {
+                    continue;
+                }
+
+                if pfn < page.pfn || (pfn + count) > page.pfn + page.count {
+                    return Err(XenClientError::new("request overlaps allocated block"));
+                }
+            } else {
+                if pfn < page.pfn {
+                    continue;
+                }
+
+                if pfn >= page.pfn + page.count {
+                    continue;
+                }
+            }
+
+            return Ok(page.ptr + ((pfn - page.pfn) << X86_PAGE_SHIFT));
+        }
+
+        if count == 0 {
+            return Err(XenClientError::new(
+                "allocation is only allowed when a size is given",
+            ));
+        }
+
+        self.pfn_alloc(pfn, count)
+    }
+
+    fn pfn_alloc(&mut self, pfn: u64, count: u64) -> Result<u64, XenClientError> {
+        let mut entries = vec![MmapEntry::default(); count as usize];
+        for (i, entry) in entries.iter_mut().enumerate() {
+            entry.mfn = self.p2m[pfn as usize + i];
+        }
+        let chunk_size = 1 << XEN_PAGE_SHIFT;
+        let num_per_entry = chunk_size >> XEN_PAGE_SHIFT;
+        let num = num_per_entry * count as usize;
+        let mut pfns = vec![u64::MAX; num];
+        for i in 0..count as usize {
+            for j in 0..num_per_entry {
+                pfns[i * num_per_entry + j] = entries[i].mfn + j as u64;
+            }
+        }
+
+        let actual_mmap_len = (num as u64) << XEN_PAGE_SHIFT;
+        let addr = self
+            .call
+            .mmap(0, actual_mmap_len)
+            .ok_or(XenClientError::new("failed to mmap address"))?;
+        debug!("mapped {:#x} foreign bytes at {:#x}", actual_mmap_len, addr);
+        let result = self.call.mmap_batch(self.domid, num as u64, addr, pfns)?;
+        if result != 0 {
+            return Err(XenClientError::new("mmap_batch call failed"));
+        }
+        let page = PhysicalPage {
+            pfn,
+            ptr: addr,
+            count,
+        };
+        debug!(
+            "alloc_pfn {:#x}+{:#x} at {:#x}",
+            page.pfn, page.count, page.ptr
+        );
+        self.pages.push(page);
+        Ok(addr)
+    }
+
+    pub fn map_foreign_pages(&mut self, mfn: u64, size: u64) -> Result<u64, XenClientError> {
+        let num = ((size + XEN_PAGE_SIZE - 1) >> XEN_PAGE_SHIFT) as usize;
+        let mut pfns = vec![u64::MAX; num];
+        for (i, item) in pfns.iter_mut().enumerate().take(num) {
+            *item = mfn + i as u64;
+        }
+
+        let actual_mmap_len = (num as u64) << XEN_PAGE_SHIFT;
+        let addr = self
+            .call
+            .mmap(0, actual_mmap_len)
+            .ok_or(XenClientError::new("failed to mmap address"))?;
+        debug!("mapped {:#x} foreign bytes at {:#x}", actual_mmap_len, addr);
+        let result = self.call.mmap_batch(self.domid, num as u64, addr, pfns)?;
+        if result != 0 {
+            return Err(XenClientError::new("mmap_batch call failed"));
+        }
+        let page = PhysicalPage {
+            pfn: u64::MAX,
+            ptr: addr,
+            count: num as u64,
+        };
+        debug!(
+            "alloc_mfn {:#x}+{:#x} at {:#x}",
+            page.pfn, page.count, page.ptr
+        );
+        self.pages.push(page);
+        Ok(addr)
+    }
+
+    pub fn unmap_all(&mut self) -> Result<(), XenClientError> {
+        for page in &self.pages {
+            unsafe {
+                let err = munmap(
+                    page.ptr as *mut c_void,
+                    (page.count << X86_PAGE_SHIFT) as usize,
+                );
+                if err != 0 {
+                    return Err(XenClientError::new("failed to munmap all pages"));
+                }
+            }
+        }
+        self.pages.clear();
+        Ok(())
+    }
+
+    pub fn unmap(&mut self, pfn: u64) -> Result<(), XenClientError> {
+        let page = self.pages.iter().enumerate().find(|(_, x)| x.pfn == pfn);
+        if page.is_none() {
+            return Err(XenClientError::new("unable to find page to unmap"));
+        }
+        let (i, page) = page.unwrap();
+
+        unsafe {
+            let err = munmap(
+                page.ptr as *mut c_void,
+                (page.count << X86_PAGE_SHIFT) as usize,
+            );
+            debug!(
+                "unmapped {:#x} foreign bytes at {:#x}",
+                (page.count << X86_PAGE_SHIFT) as usize,
+                page.ptr
+            );
+            if err != 0 {
+                return Err(XenClientError::new("failed to munmap page"));
+            }
+            self.pages.remove(i);
+        }
+        Ok(())
+    }
+}

xen/xenclient/src/sys.rs

@@ -0,0 +1,130 @@
+pub const XEN_ELFNOTE_INFO: u64 = 0;
+pub const XEN_ELFNOTE_ENTRY: u64 = 1;
+pub const XEN_ELFNOTE_HYPERCALL_PAGE: u64 = 2;
+pub const XEN_ELFNOTE_VIRT_BASE: u64 = 3;
+pub const XEN_ELFNOTE_PADDR_OFFSET: u64 = 4;
+pub const XEN_ELFNOTE_XEN_VERSION: u64 = 5;
+pub const XEN_ELFNOTE_GUEST_OS: u64 = 6;
+pub const XEN_ELFNOTE_GUEST_VERSION: u64 = 7;
+pub const XEN_ELFNOTE_LOADER: u64 = 8;
+pub const XEN_ELFNOTE_PAE_MODE: u64 = 9;
+pub const XEN_ELFNOTE_FEATURES: u64 = 10;
+pub const XEN_ELFNOTE_BSD_SYMTAB: u64 = 11;
+pub const XEN_ELFNOTE_HV_START_LOW: u64 = 12;
+pub const XEN_ELFNOTE_L1_MFN_VALID: u64 = 13;
+pub const XEN_ELFNOTE_SUSPEND_CANCEL: u64 = 14;
+pub const XEN_ELFNOTE_INIT_P2M: u64 = 15;
+pub const XEN_ELFNOTE_MOD_START_PFN: u64 = 16;
+pub const XEN_ELFNOTE_SUPPORTED_FEATURES: u64 = 17;
+pub const XEN_ELFNOTE_PHYS32_ENTRY: u64 = 18;
+
+#[derive(Copy, Clone)]
+pub struct ElfNoteXenType {
+    pub id: u64,
+    pub name: &'static str,
+    pub is_string: bool,
+}
+
+pub const XEN_ELFNOTE_TYPES: &[ElfNoteXenType] = &[
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_ENTRY,
+        name: "ENTRY",
+        is_string: false,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_HYPERCALL_PAGE,
+        name: "HYPERCALL_PAGE",
+        is_string: false,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_VIRT_BASE,
+        name: "VIRT_BASE",
+        is_string: false,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_INIT_P2M,
+        name: "INIT_P2M",
+        is_string: false,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_PADDR_OFFSET,
+        name: "PADDR_OFFSET",
+        is_string: false,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_HV_START_LOW,
+        name: "HV_START_LOW",
+        is_string: false,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_XEN_VERSION,
+        name: "XEN_VERSION",
+        is_string: true,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_GUEST_OS,
+        name: "GUEST_OS",
+        is_string: true,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_GUEST_VERSION,
+        name: "GUEST_VERSION",
+        is_string: true,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_LOADER,
+        name: "LOADER",
+        is_string: true,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_PAE_MODE,
+        name: "PAE_MODE",
+        is_string: true,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_FEATURES,
+        name: "FEATURES",
+        is_string: true,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_SUPPORTED_FEATURES,
+        name: "SUPPORTED_FEATURES",
+        is_string: false,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_BSD_SYMTAB,
+        name: "BSD_SYMTAB",
+        is_string: true,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_SUSPEND_CANCEL,
+        name: "SUSPEND_CANCEL",
+        is_string: false,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_MOD_START_PFN,
+        name: "MOD_START_PFN",
+        is_string: false,
+    },
+    ElfNoteXenType {
+        id: XEN_ELFNOTE_PHYS32_ENTRY,
+        name: "PHYS32_ENTRY",
+        is_string: false,
+    },
+];
+
+pub const XEN_PAGE_SHIFT: u64 = 12;
+pub const XEN_PAGE_SIZE: u64 = 1 << XEN_PAGE_SHIFT;
+pub const XEN_PAGE_MASK: u64 = !(XEN_PAGE_SIZE - 1);
+pub const SUPERPAGE_BATCH_SIZE: u64 = 512;
+pub const SUPERPAGE_2MB_SHIFT: u64 = 9;
+pub const SUPERPAGE_2MB_NR_PFNS: u64 = 1u64 << SUPERPAGE_2MB_SHIFT;
+pub const VGCF_IN_KERNEL: u64 = 1 << 2;
+pub const VGCF_ONLINE: u64 = 1 << 5;
+
+#[repr(C)]
+pub struct GrantEntry {
+    pub flags: u16,
+    pub domid: u16,
+    pub frame: u32,
+}

xen/xenclient/src/x86.rs

@@ -0,0 +1,638 @@
+use crate::boot::{
+    ArchBootSetup, BootImageInfo, BootSetup, BootState, DomainSegment, XEN_UNSET_ADDR,
+};
+use crate::sys::{
+    SUPERPAGE_2MB_NR_PFNS, SUPERPAGE_2MB_SHIFT, SUPERPAGE_BATCH_SIZE, VGCF_IN_KERNEL, VGCF_ONLINE,
+    XEN_PAGE_SHIFT,
+};
+use crate::XenClientError;
+use libc::c_char;
+use log::{debug, trace};
+use slice_copy::copy;
+use std::cmp::{max, min};
+use std::mem::size_of;
+use std::slice;
+use xencall::sys::{VcpuGuestContext, MMUEXT_PIN_L4_TABLE};
+
+pub const X86_PAGE_SHIFT: u64 = 12;
+pub const X86_PAGE_SIZE: u64 = 1 << X86_PAGE_SHIFT;
+pub const X86_VIRT_BITS: u64 = 48;
+pub const X86_VIRT_MASK: u64 = (1 << X86_VIRT_BITS) - 1;
+pub const X86_PGTABLE_LEVELS: u64 = 4;
+pub const X86_PGTABLE_LEVEL_SHIFT: u64 = 9;
+
+#[repr(C)]
+#[derive(Debug, Clone, Default)]
+pub struct PageTableMappingLevel {
+    pub from: u64,
+    pub to: u64,
+    pub pfn: u64,
+    pub pgtables: usize,
+}
+
+#[repr(C)]
+#[derive(Debug, Clone, Default)]
+pub struct PageTableMapping {
+    pub area: PageTableMappingLevel,
+    pub levels: [PageTableMappingLevel; X86_PGTABLE_LEVELS as usize],
+}
+
+pub const X86_PAGE_TABLE_MAX_MAPPINGS: usize = 2;
+
+#[repr(C)]
+#[derive(Debug, Clone, Default)]
+pub struct PageTable {
+    pub mappings_count: usize,
+    pub mappings: [PageTableMapping; X86_PAGE_TABLE_MAX_MAPPINGS],
+}
+
+#[repr(C)]
+#[derive(Debug)]
+pub struct StartInfoConsole {
+    pub mfn: u64,
+    pub evtchn: u32,
+}
+
+pub const MAX_GUEST_CMDLINE: usize = 1024;
+
+#[repr(C)]
+#[derive(Debug)]
+pub struct StartInfo {
+    pub magic: [c_char; 32],
+    pub nr_pages: u64,
+    pub shared_info: u64,
+    pub flags: u32,
+    pub store_mfn: u64,
+    pub store_evtchn: u32,
+    pub console: StartInfoConsole,
+    pub pt_base: u64,
+    pub nr_pt_frames: u64,
+    pub mfn_list: u64,
+    pub mod_start: u64,
+    pub mod_len: u64,
+    pub cmdline: [c_char; MAX_GUEST_CMDLINE],
+    pub first_p2m_pfn: u64,
+    pub nr_p2m_frames: u64,
+}
+
+pub const X86_GUEST_MAGIC: &str = "xen-3.0-x86_64";
+
+#[repr(C)]
+#[derive(Debug)]
+pub struct ArchVcpuInfo {
+    pub cr2: u64,
+    pub pad: u64,
+}
+
+#[repr(C)]
+#[derive(Debug)]
+pub struct VcpuInfoTime {
+    pub version: u32,
+    pub pad0: u32,
+    pub tsc_timestamp: u64,
+    pub system_time: u64,
+    pub tsc_to_system_mul: u32,
+    pub tsc_shift: i8,
+    pub flags: u8,
+    pub pad1: [u8; 2],
+}
+
+#[repr(C)]
+#[derive(Debug)]
+pub struct VcpuInfo {
+    pub evtchn_upcall_pending: u8,
+    pub evtchn_upcall_mask: u8,
+    pub evtchn_pending_sel: u64,
+    pub arch_vcpu_info: ArchVcpuInfo,
+    pub vcpu_info_time: VcpuInfoTime,
+}
+
+#[repr(C)]
+#[derive(Debug)]
+pub struct SharedInfo {
+    pub vcpu_info: [VcpuInfo; 32],
+    pub evtchn_pending: [u64; u64::BITS as usize],
+    pub evtchn_mask: [u64; u64::BITS as usize],
+    pub wc_version: u32,
+    pub wc_sec: u32,
+    pub wc_nsec: u32,
+    pub wc_sec_hi: u32,
+
+    // arch shared info
+    pub max_pfn: u64,
+    pub pfn_to_mfn_frame_list_list: u64,
+    pub nmi_reason: u64,
+    pub p2m_cr3: u64,
+    pub p2m_vaddr: u64,
+    pub p2m_generation: u64,
+}
+
+pub struct X86BootSetup {
+    table: PageTable,
+}
+
+#[derive(Debug)]
+struct VmemRange {
+    start: u64,
+    end: u64,
+    _flags: u32,
+    _nid: u32,
+}
+
+impl X86BootSetup {
+    pub fn new() -> X86BootSetup {
+        X86BootSetup {
+            table: PageTable::default(),
+        }
+    }
+
+    const PAGE_PRESENT: u64 = 0x001;
+    const PAGE_RW: u64 = 0x002;
+    const PAGE_USER: u64 = 0x004;
+    const PAGE_ACCESSED: u64 = 0x020;
+    const PAGE_DIRTY: u64 = 0x040;
+    fn get_pg_prot(&mut self, l: usize, pfn: u64) -> u64 {
+        let prot = [
+            X86BootSetup::PAGE_PRESENT | X86BootSetup::PAGE_RW | X86BootSetup::PAGE_ACCESSED,
+            X86BootSetup::PAGE_PRESENT
+                | X86BootSetup::PAGE_RW
+                | X86BootSetup::PAGE_ACCESSED
+                | X86BootSetup::PAGE_DIRTY
+                | X86BootSetup::PAGE_USER,
+            X86BootSetup::PAGE_PRESENT
+                | X86BootSetup::PAGE_RW
+                | X86BootSetup::PAGE_ACCESSED
+                | X86BootSetup::PAGE_DIRTY
+                | X86BootSetup::PAGE_USER,
+            X86BootSetup::PAGE_PRESENT
+                | X86BootSetup::PAGE_RW
+                | X86BootSetup::PAGE_ACCESSED
+                | X86BootSetup::PAGE_DIRTY
+                | X86BootSetup::PAGE_USER,
+        ];
+
+        let prot = prot[l];
+        if l > 0 {
+            return prot;
+        }
+
+        for m in 0..self.table.mappings_count {
+            let map = &self.table.mappings[m];
+            let pfn_s = map.levels[(X86_PGTABLE_LEVELS - 1) as usize].pfn;
+            let pfn_e = map.area.pgtables as u64 + pfn_s;
+            if pfn >= pfn_s && pfn < pfn_e {
+                return prot & !X86BootSetup::PAGE_RW;
+            }
+        }
+        prot
+    }
+
+    fn count_page_tables(
+        &mut self,
+        setup: &mut BootSetup,
+        from: u64,
+        to: u64,
+        pfn: u64,
+    ) -> Result<usize, XenClientError> {
+        debug!("counting pgtables from={} to={} pfn={}", from, to, pfn);
+        if self.table.mappings_count == X86_PAGE_TABLE_MAX_MAPPINGS {
+            return Err(XenClientError::new("too many mappings"));
+        }
+
+        let m = self.table.mappings_count;
+
+        let pfn_end = pfn + ((to - from) >> X86_PAGE_SHIFT);
+        if pfn_end >= setup.phys.p2m_size() {
+            return Err(XenClientError::new("not enough memory for initial mapping"));
+        }
+
+        for idx in 0..self.table.mappings_count {
+            if from < self.table.mappings[idx].area.to && to > self.table.mappings[idx].area.from {
+                return Err(XenClientError::new("overlapping mappings"));
+            }
+        }
+        let mut map = PageTableMapping::default();
+        map.area.from = from & X86_VIRT_MASK;
+        map.area.to = to & X86_VIRT_MASK;
+
+        for l in (0usize..X86_PGTABLE_LEVELS as usize).rev() {
+            map.levels[l].pfn = setup.pfn_alloc_end + map.area.pgtables as u64;
+            if l as u64 == X86_PGTABLE_LEVELS - 1 {
+                if self.table.mappings_count == 0 {
+                    map.levels[l].from = 0;
+                    map.levels[l].to = X86_VIRT_MASK;
+                    map.levels[l].pgtables = 1;
+                    map.area.pgtables += 1;
+                }
+                continue;
+            }
+
+            let bits = X86_PAGE_SHIFT + (l + 1) as u64 * X86_PGTABLE_LEVEL_SHIFT;
+            let mask = BootSetup::bits_to_mask(bits);
+            map.levels[l].from = map.area.from & !mask;
+            map.levels[l].to = map.area.to | mask;
+
+            for cmp in &mut self.table.mappings[0..self.table.mappings_count] {
+                if cmp.levels[l].from == cmp.levels[l].to {
+                    continue;
+                }
+
+                if map.levels[l].from >= cmp.levels[l].from && map.levels[l].to <= cmp.levels[l].to
+                {
+                    map.levels[l].from = 0;
+                    map.levels[l].to = 0;
+                    break;
+                }
+
+                if map.levels[l].from >= cmp.levels[l].from
+                    && map.levels[l].from <= cmp.levels[l].to
+                {
+                    map.levels[l].from = cmp.levels[l].to + 1;
+                }
+
+                if map.levels[l].to >= cmp.levels[l].from && map.levels[l].to <= cmp.levels[l].to {
+                    map.levels[l].to = cmp.levels[l].from - 1;
+                }
+            }
+
+            if map.levels[l].from < map.levels[l].to {
+                map.levels[l].pgtables =
+                    (((map.levels[l].to - map.levels[l].from) >> bits) + 1) as usize;
+            }
+
+            debug!(
+                "BootSetup count_pgtables {:#x}/{}: {:#x} -> {:#x}, {} tables",
+                mask, bits, map.levels[l].from, map.levels[l].to, map.levels[l].pgtables
+            );
+            map.area.pgtables += map.levels[l].pgtables;
+        }
+        self.table.mappings[m] = map;
+        Ok(m)
+    }
+}
+
+impl ArchBootSetup for X86BootSetup {
+    fn page_size(&mut self) -> u64 {
+        X86_PAGE_SIZE
+    }
+
+    fn page_shift(&mut self) -> u64 {
+        X86_PAGE_SHIFT
+    }
+
+    fn alloc_p2m_segment(
+        &mut self,
+        setup: &mut BootSetup,
+        image_info: &BootImageInfo,
+    ) -> Result<DomainSegment, XenClientError> {
+        let mut p2m_alloc_size =
+            ((setup.phys.p2m_size() * 8) + X86_PAGE_SIZE - 1) & !(X86_PAGE_SIZE - 1);
+        let from = image_info.virt_p2m_base;
+        let to = from + p2m_alloc_size - 1;
+        let m = self.count_page_tables(setup, from, to, setup.pfn_alloc_end)?;
+
+        let pgtables: usize;
+        {
+            let map = &mut self.table.mappings[m];
+            map.area.pfn = setup.pfn_alloc_end;
+            for lvl_idx in 0..4 {
+                map.levels[lvl_idx].pfn += p2m_alloc_size >> X86_PAGE_SHIFT;
+            }
+            pgtables = map.area.pgtables;
+        }
+        self.table.mappings_count += 1;
+        p2m_alloc_size += (pgtables << X86_PAGE_SHIFT) as u64;
+        let p2m_segment = setup.alloc_segment(self, 0, p2m_alloc_size)?;
+        Ok(p2m_segment)
+    }
+
+    fn alloc_page_tables(
+        &mut self,
+        setup: &mut BootSetup,
+        image_info: &BootImageInfo,
+    ) -> Result<DomainSegment, XenClientError> {
+        let mut extra_pages = 1;
+        extra_pages += (512 * 1024) / X86_PAGE_SIZE;
+        let mut pages = extra_pages;
+
+        let mut try_virt_end: u64;
+        let mut m: usize;
+        loop {
+            try_virt_end = BootSetup::round_up(
+                setup.virt_alloc_end + pages * X86_PAGE_SIZE,
+                BootSetup::bits_to_mask(22),
+            );
+            m = self.count_page_tables(setup, image_info.virt_base, try_virt_end, 0)?;
+            pages = self.table.mappings[m].area.pgtables as u64 + extra_pages;
+            if setup.virt_alloc_end + pages * X86_PAGE_SIZE <= try_virt_end + 1 {
+                break;
+            }
+        }
+
+        self.table.mappings[m].area.pfn = 0;
+        self.table.mappings_count += 1;
+        setup.virt_pgtab_end = try_virt_end + 1;
+        let size = self.table.mappings[m].area.pgtables as u64 * X86_PAGE_SIZE;
+        let segment = setup.alloc_segment(self, 0, size)?;
+        debug!(
+            "BootSetup alloc_page_tables table={:?} segment={:?}",
+            self.table, segment
+        );
+        Ok(segment)
+    }
+
+    fn setup_page_tables(
+        &mut self,
+        setup: &mut BootSetup,
+        state: &mut BootState,
+    ) -> Result<(), XenClientError> {
+        let p2m_guest = unsafe {
+            slice::from_raw_parts_mut(
+                state.p2m_segment.addr as *mut u64,
+                setup.phys.p2m_size() as usize,
+            )
+        };
+        copy(p2m_guest, &setup.phys.p2m);
+
+        for l in (0usize..X86_PGTABLE_LEVELS as usize).rev() {
+            for m1 in 0usize..self.table.mappings_count {
+                let map1 = &self.table.mappings[m1];
+                let from = map1.levels[l].from;
+                let to = map1.levels[l].to;
+                let pg_ptr = setup.phys.pfn_to_ptr(map1.levels[l].pfn, 0)? as *mut u64;
+                for m2 in 0usize..self.table.mappings_count {
+                    let map2 = &self.table.mappings[m2];
+                    let lvl = if l > 0 {
+                        &map2.levels[l - 1]
+                    } else {
+                        &map2.area
+                    };
+
+                    if l > 0 && lvl.pgtables == 0 {
+                        continue;
+                    }
+
+                    if lvl.from >= to || lvl.to <= from {
+                        continue;
+                    }
+
+                    let p_s = (max(from, lvl.from) - from)
+                        >> (X86_PAGE_SHIFT + l as u64 * X86_PGTABLE_LEVEL_SHIFT);
+                    let p_e = (min(to, lvl.to) - from)
+                        >> (X86_PAGE_SHIFT + l as u64 * X86_PGTABLE_LEVEL_SHIFT);
+                    let rhs = X86_PAGE_SHIFT as usize + l * X86_PGTABLE_LEVEL_SHIFT as usize;
+                    let mut pfn = ((max(from, lvl.from) - lvl.from) >> rhs) + lvl.pfn;
+
+                    debug!(
+                        "BootSetup setup_page_tables lvl={} map_1={} map_2={} pfn={:#x} p_s={:#x} p_e={:#x}",
+                        l, m1, m2, pfn, p_s, p_e
+                    );
+
+                    let pg = unsafe { slice::from_raw_parts_mut(pg_ptr, (p_e + 1) as usize) };
+                    for p in p_s..p_e + 1 {
+                        let prot = self.get_pg_prot(l, pfn);
+                        let pfn_paddr = setup.phys.p2m[pfn as usize] << X86_PAGE_SHIFT;
+                        let value = pfn_paddr | prot;
+                        pg[p as usize] = value;
+                        pfn += 1;
+                    }
+                }
+            }
+        }
+        Ok(())
+    }
+
+    fn setup_start_info(
+        &mut self,
+        setup: &mut BootSetup,
+        state: &BootState,
+        cmdline: &str,
+    ) -> Result<(), XenClientError> {
+        let ptr = setup.phys.pfn_to_ptr(state.start_info_segment.pfn, 1)?;
+        let byte_slice =
+            unsafe { slice::from_raw_parts_mut(ptr as *mut u8, X86_PAGE_SIZE as usize) };
+        byte_slice.fill(0);
+        let info = ptr as *mut StartInfo;
+        unsafe {
+            for (i, c) in X86_GUEST_MAGIC.chars().enumerate() {
+                (*info).magic[i] = c as c_char;
+            }
+            (*info).magic[X86_GUEST_MAGIC.len()] = 0 as c_char;
+            (*info).nr_pages = setup.total_pages;
+            (*info).shared_info = state.shared_info_frame << X86_PAGE_SHIFT;
+            (*info).pt_base = state.page_table_segment.vstart;
+            (*info).nr_pt_frames = self.table.mappings[0].area.pgtables as u64;
+            (*info).mfn_list = state.p2m_segment.vstart;
+            (*info).first_p2m_pfn = state.p2m_segment.pfn;
+            (*info).nr_p2m_frames = state.p2m_segment.pages;
+            (*info).flags = 0;
+            (*info).store_evtchn = state.store_evtchn;
+            (*info).store_mfn = setup.phys.p2m[state.xenstore_segment.pfn as usize];
+            (*info).console.mfn = setup.phys.p2m[state.console_segment.pfn as usize];
+            (*info).console.evtchn = state.console_evtchn;
+            (*info).mod_start = state.initrd_segment.vstart;
+            (*info).mod_len = state.initrd_segment.size;
+            for (i, c) in cmdline.chars().enumerate() {
+                (*info).cmdline[i] = c as c_char;
+            }
+            (*info).cmdline[MAX_GUEST_CMDLINE - 1] = 0;
+            trace!("BootSetup setup_start_info start_info={:?}", *info);
+        }
+        Ok(())
+    }
+
+    fn setup_shared_info(
+        &mut self,
+        setup: &mut BootSetup,
+        shared_info_frame: u64,
+    ) -> Result<(), XenClientError> {
+        let info = setup
+            .phys
+            .map_foreign_pages(shared_info_frame, X86_PAGE_SIZE)?
+            as *mut SharedInfo;
+        unsafe {
+            let size = size_of::<SharedInfo>();
+            let info_as_buff = slice::from_raw_parts_mut(info as *mut u8, size);
+            info_as_buff.fill(0);
+            for i in 0..32 {
+                (*info).vcpu_info[i].evtchn_upcall_mask = 1;
+            }
+            trace!("BootSetup setup_shared_info shared_info={:?}", *info);
+        }
+        Ok(())
+    }
+
+    fn setup_hypercall_page(
+        &mut self,
+        setup: &mut BootSetup,
+        image_info: &BootImageInfo,
+    ) -> Result<(), XenClientError> {
+        if image_info.virt_hypercall == XEN_UNSET_ADDR {
+            return Ok(());
+        }
+
+        let pfn = (image_info.virt_hypercall - image_info.virt_base) >> X86_PAGE_SHIFT;
+        let mfn = setup.phys.p2m[pfn as usize];
+        setup.call.hypercall_init(setup.domid, mfn)?;
+        Ok(())
+    }
+
+    fn meminit(&mut self, setup: &mut BootSetup, total_pages: u64) -> Result<(), XenClientError> {
+        setup.call.claim_pages(setup.domid, total_pages)?;
+        let mut vmemranges: Vec<VmemRange> = Vec::new();
+        let stub = VmemRange {
+            start: 0,
+            end: total_pages << XEN_PAGE_SHIFT,
+            _flags: 0,
+            _nid: 0,
+        };
+        vmemranges.push(stub);
+        let mut p2m_size: u64 = 0;
+        let mut total: u64 = 0;
+        for range in &vmemranges {
+            total += (range.end - range.start) >> XEN_PAGE_SHIFT;
+            p2m_size = p2m_size.max(range.end >> XEN_PAGE_SHIFT);
+        }
+
+        if total != total_pages {
+            return Err(XenClientError::new(
+                "page count mismatch while calculating pages",
+            ));
+        }
+
+        setup.total_pages = total;
+
+        let mut p2m = vec![u64::MAX; p2m_size as usize];
+        for range in &vmemranges {
+            let mut extents_init = vec![0u64; SUPERPAGE_BATCH_SIZE as usize];
+            let pages = (range.end - range.start) >> XEN_PAGE_SHIFT;
+            let pfn_base = range.start >> XEN_PAGE_SHIFT;
+
+            for pfn in pfn_base..pfn_base + pages {
+                p2m[pfn as usize] = pfn;
+            }
+
+            let mut super_pages = pages >> SUPERPAGE_2MB_SHIFT;
+            let mut pfn_base_idx: u64 = pfn_base;
+            while super_pages > 0 {
+                let count = super_pages.min(SUPERPAGE_BATCH_SIZE);
+                super_pages -= count;
+
+                let mut j: usize = 0;
+                let mut pfn: u64 = pfn_base_idx;
+                loop {
+                    if pfn >= pfn_base_idx + (count << SUPERPAGE_2MB_SHIFT) {
+                        break;
+                    }
+                    extents_init[j] = p2m[pfn as usize];
+                    pfn += SUPERPAGE_2MB_NR_PFNS;
+                    j += 1;
+                }
+
+                let extents_init_slice = extents_init.as_slice();
+                let extents = setup.call.populate_physmap(
+                    setup.domid,
+                    count,
+                    SUPERPAGE_2MB_SHIFT as u32,
+                    0,
+                    &extents_init_slice[0usize..count as usize],
+                )?;
+
+                pfn = pfn_base_idx;
+                for mfn in extents {
+                    for k in 0..SUPERPAGE_2MB_NR_PFNS {
+                        p2m[pfn as usize] = mfn + k;
+                        pfn += 1;
+                    }
+                }
+                pfn_base_idx = pfn;
+            }
+
+            let mut j = pfn_base_idx - pfn_base;
+            loop {
+                if j >= pages {
+                    break;
+                }
+
+                let allocsz = (1024 * 1024).min(pages - j);
+                let p2m_idx = (pfn_base + j) as usize;
+                let p2m_end_idx = p2m_idx + allocsz as usize;
+                let input_extent_starts = &p2m[p2m_idx..p2m_end_idx];
+                let result =
+                    setup
+                        .call
+                        .populate_physmap(setup.domid, allocsz, 0, 0, input_extent_starts)?;
+
+                if result.len() != allocsz as usize {
+                    return Err(XenClientError::new(
+                        format!(
+                            "failed to populate physmap: wanted={} received={} input_extents={}",
+                            allocsz,
+                            result.len(),
+                            input_extent_starts.len()
+                        )
+                        .as_str(),
+                    ));
+                }
+
+                for (i, item) in result.iter().enumerate() {
+                    let p = (pfn_base + j + i as u64) as usize;
+                    let m = *item;
+                    p2m[p] = m;
+                }
+                j += allocsz;
+            }
+        }
+
+        setup.phys.load_p2m(p2m);
+        setup.call.claim_pages(setup.domid, 0)?;
+        Ok(())
+    }
+
+    fn bootlate(
+        &mut self,
+        setup: &mut BootSetup,
+        state: &mut BootState,
+    ) -> Result<(), XenClientError> {
+        let pg_pfn = state.page_table_segment.pfn;
+        let pg_mfn = setup.phys.p2m[pg_pfn as usize];
+        setup.phys.unmap(pg_pfn)?;
+        setup.phys.unmap(state.p2m_segment.pfn)?;
+        setup
+            .call
+            .mmuext(setup.domid, MMUEXT_PIN_L4_TABLE, pg_mfn, 0)?;
+        Ok(())
+    }
+
+    fn vcpu(&mut self, setup: &mut BootSetup, state: &mut BootState) -> Result<(), XenClientError> {
+        let pg_pfn = state.page_table_segment.pfn;
+        let pg_mfn = setup.phys.p2m[pg_pfn as usize];
+        let mut vcpu = VcpuGuestContext::default();
+        vcpu.user_regs.rip = state.image_info.virt_entry;
+        vcpu.user_regs.rsp =
+            state.image_info.virt_base + (state.boot_stack_segment.pfn + 1) * self.page_size();
+        vcpu.user_regs.rsi =
+            state.image_info.virt_base + (state.start_info_segment.pfn) * self.page_size();
+        vcpu.user_regs.rflags = 1 << 9;
+        vcpu.debugreg[6] = 0xffff0ff0;
+        vcpu.debugreg[7] = 0x00000400;
+        vcpu.flags = VGCF_IN_KERNEL | VGCF_ONLINE;
+        let cr3_pfn = pg_mfn;
+        debug!(
+            "cr3: pfn {:#x} mfn {:#x}",
+            state.page_table_segment.pfn, cr3_pfn
+        );
+        vcpu.ctrlreg[3] = cr3_pfn << 12;
+        vcpu.user_regs.ds = 0x0;
+        vcpu.user_regs.es = 0x0;
+        vcpu.user_regs.fs = 0x0;
+        vcpu.user_regs.gs = 0x0;
+        vcpu.user_regs.ss = 0xe02b;
+        vcpu.user_regs.cs = 0xe033;
+        vcpu.kernel_ss = vcpu.user_regs.ss as u64;
+        vcpu.kernel_sp = vcpu.user_regs.rsp;
+        debug!("vcpu context: {:?}", vcpu);
+        setup.call.set_vcpu_context(setup.domid, 0, &vcpu)?;
+        Ok(())
+    }
+}