mirrors/krata
1
0
Fork 0
mirror of https://github.com/edera-dev/krata.git synced 2025-08-02 12:50:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

rearrange code to move most x86 specifics to the x86 module

Browse Source
This commit is contained in:
Alex Zenla 2024-01-17 06:27:12 -08:00
parent 15ba27e573
commit ee32cbee16
No known key found for this signature in database
GPG Key ID: 067B238899B51269
3 changed files with 672 additions and 557 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

657
xenclient/src/boot.rs
View File

@ -1,22 +1,12 @@
use crate::mem::PhysicalPages; use crate::mem::PhysicalPages;
use crate::sys::{ use crate::sys::{GrantEntry, XEN_PAGE_SHIFT};
GrantEntry, SUPERPAGE_2MB_NR_PFNS, SUPERPAGE_2MB_SHIFT, SUPERPAGE_BATCH_SIZE, VGCF_IN_KERNEL,
VGCF_ONLINE, XEN_PAGE_SHIFT,
};
use crate::x86::{
PageTable, PageTableMapping, SharedInfo, StartInfo, MAX_GUEST_CMDLINE, X86_GUEST_MAGIC,
X86_PAGE_SHIFT, X86_PAGE_SIZE, X86_PAGE_TABLE_MAX_MAPPINGS, X86_PGTABLE_LEVELS,
X86_PGTABLE_LEVEL_SHIFT, X86_VIRT_MASK,
};
use crate::XenClientError; use crate::XenClientError;
use libc::{c_char, munmap}; use libc::munmap;
use log::{debug, trace}; use log::debug;
use slice_copy::copy; use slice_copy::copy;
use std::cmp::{max, min};
use std::ffi::c_void; use std::ffi::c_void;
use std::mem::size_of;
use std::slice; use std::slice;
use xencall::sys::{VcpuGuestContext, MMUEXT_PIN_L4_TABLE};
use xencall::XenCall; use xencall::XenCall;
pub trait BootImageLoader { pub trait BootImageLoader {
@ -39,31 +29,23 @@ pub struct BootImageInfo {
} }
pub struct BootSetup<'a> { pub struct BootSetup<'a> {
call: &'a XenCall, pub(crate) call: &'a XenCall,
pub phys: PhysicalPages<'a>, pub phys: PhysicalPages<'a>,
domid: u32, pub(crate) domid: u32,
virt_alloc_end: u64, pub(crate) virt_alloc_end: u64,
pfn_alloc_end: u64, pub(crate) pfn_alloc_end: u64,
virt_pgtab_end: u64, pub(crate) virt_pgtab_end: u64,
total_pages: u64, pub(crate) total_pages: u64,
} }
#[derive(Debug)] #[derive(Debug)]
pub struct DomainSegment { pub struct DomainSegment {
vstart: u64, pub(crate) vstart: u64,
_vend: u64, vend: u64,
pub pfn: u64, pub pfn: u64,
addr: u64, pub(crate) addr: u64,
size: u64, pub(crate) size: u64,
pages: u64, pub(crate) pages: u64,
}
#[derive(Debug)]
struct VmemRange {
start: u64,
end: u64,
_flags: u32,
_nid: u32,
} }
#[derive(Debug)] #[derive(Debug)]
@ -75,7 +57,6 @@ pub struct BootState {
pub boot_stack_segment: DomainSegment, pub boot_stack_segment: DomainSegment,
pub p2m_segment: DomainSegment, pub p2m_segment: DomainSegment,
pub page_table_segment: DomainSegment, pub page_table_segment: DomainSegment,
pub page_table: PageTable,
pub image_info: BootImageInfo, pub image_info: BootImageInfo,
pub shared_info_frame: u64, pub shared_info_frame: u64,
pub initrd_segment: DomainSegment, pub initrd_segment: DomainSegment,
@ -96,133 +77,19 @@ impl BootSetup<'_> {
} }
} }
fn initialize_memory(&mut self, total_pages: u64) -> Result<(), XenClientError> { fn initialize_memory(
&mut self,
arch: &mut dyn ArchBootSetup,
total_pages: u64,
) -> Result<(), XenClientError> {
self.call.set_address_size(self.domid, 64)?; self.call.set_address_size(self.domid, 64)?;
arch.meminit(self, 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);
debug!("BootSetup initialize_memory vmemranges: {:?}", vmemranges);
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",
));
}
debug!("BootSetup initialize_memory total_pages={}", total);
self.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 = self.call.populate_physmap(
self.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 =
self.call
.populate_physmap(self.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;
}
}
self.phys.load_p2m(p2m);
Ok(())
}
fn setup_hypercall_page(&mut self, 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 = self.phys.p2m[pfn as usize];
self.call.hypercall_init(self.domid, mfn)?;
Ok(()) Ok(())
} }
pub fn initialize( pub fn initialize(
&mut self, &mut self,
arch: &mut dyn ArchBootSetup,
image_loader: &dyn BootImageLoader, image_loader: &dyn BootImageLoader,
initrd: &[u8], initrd: &[u8],
max_vcpus: u32, max_vcpus: u32,
@ -233,43 +100,42 @@ impl BootSetup<'_> {
max_vcpus, mem_mb max_vcpus, mem_mb
); );
let total_pages = mem_mb << (20 - X86_PAGE_SHIFT); let total_pages = mem_mb << (20 - arch.page_shift());
self.initialize_memory(total_pages)?; self.initialize_memory(arch, total_pages)?;
let image_info = image_loader.parse()?; let image_info = image_loader.parse()?;
debug!("BootSetup initialize image_info={:?}", image_info); debug!("BootSetup initialize image_info={:?}", image_info);
self.virt_alloc_end = image_info.virt_base; self.virt_alloc_end = image_info.virt_base;
let kernel_segment = self.load_kernel_segment(image_loader, &image_info)?; let kernel_segment = self.load_kernel_segment(arch, image_loader, &image_info)?;
let mut p2m_segment: Option<DomainSegment> = None; let mut p2m_segment: Option<DomainSegment> = None;
let mut page_table = PageTable::default();
if image_info.virt_p2m_base >= image_info.virt_base if image_info.virt_p2m_base >= image_info.virt_base
|| (image_info.virt_p2m_base & ((1 << X86_PAGE_SHIFT) - 1)) != 0 || (image_info.virt_p2m_base & ((1 << arch.page_shift()) - 1)) != 0
{ {
p2m_segment = Some(self.alloc_p2m_segment(&mut page_table, &image_info)?); p2m_segment = Some(arch.alloc_p2m_segment(self, &image_info)?);
} }
let start_info_segment = self.alloc_page()?; let start_info_segment = self.alloc_page(arch)?;
let xenstore_segment = self.alloc_page()?; let xenstore_segment = self.alloc_page(arch)?;
let console_segment = self.alloc_page()?; let console_segment = self.alloc_page(arch)?;
let page_table_segment = self.alloc_page_tables(&mut page_table, &image_info)?; let page_table_segment = arch.alloc_page_tables(self, &image_info)?;
let boot_stack_segment = self.alloc_page()?; let boot_stack_segment = self.alloc_page(arch)?;
if self.virt_pgtab_end > 0 { if self.virt_pgtab_end > 0 {
self.alloc_padding_pages(self.virt_pgtab_end)?; self.alloc_padding_pages(arch, self.virt_pgtab_end)?;
} }
let mut initrd_segment: Option<DomainSegment> = None; let mut initrd_segment: Option<DomainSegment> = None;
if !image_info.unmapped_initrd { if !image_info.unmapped_initrd {
initrd_segment = Some(self.alloc_module(initrd)?); initrd_segment = Some(self.alloc_module(arch, initrd)?);
} }
if p2m_segment.is_none() { if p2m_segment.is_none() {
let mut segment = self.alloc_p2m_segment(&mut page_table, &image_info)?; let mut segment = arch.alloc_p2m_segment(self, &image_info)?;
segment.vstart = image_info.virt_p2m_base; segment.vstart = image_info.virt_p2m_base;
p2m_segment = Some(segment); p2m_segment = Some(segment);
} }
let p2m_segment = p2m_segment.unwrap(); let p2m_segment = p2m_segment.unwrap();
if image_info.unmapped_initrd { if image_info.unmapped_initrd {
initrd_segment = Some(self.alloc_module(initrd)?); initrd_segment = Some(self.alloc_module(arch, initrd)?);
} }
let initrd_segment = initrd_segment.unwrap(); let initrd_segment = initrd_segment.unwrap();
@ -283,7 +149,6 @@ impl BootSetup<'_> {
boot_stack_segment, boot_stack_segment,
p2m_segment, p2m_segment,
page_table_segment, page_table_segment,
page_table,
image_info, image_info,
initrd_segment, initrd_segment,
store_evtchn, store_evtchn,
@ -294,48 +159,21 @@ impl BootSetup<'_> {
Ok(state) Ok(state)
} }
pub fn boot(&mut self, state: &mut BootState, cmdline: &str) -> Result<(), XenClientError> { 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 domain_info = self.call.get_domain_info(self.domid)?;
let shared_info_frame = domain_info.shared_info_frame; let shared_info_frame = domain_info.shared_info_frame;
state.shared_info_frame = shared_info_frame; state.shared_info_frame = shared_info_frame;
self.setup_page_tables(state)?; arch.setup_page_tables(self, state)?;
self.setup_start_info(state, cmdline)?; arch.setup_start_info(self, state, cmdline)?;
self.setup_hypercall_page(&state.image_info)?; arch.setup_hypercall_page(self, &state.image_info)?;
arch.bootlate(self, state)?;
let pg_pfn = state.page_table_segment.pfn; arch.setup_shared_info(self, state.shared_info_frame)?;
self.phys.unmap(pg_pfn)?; arch.vcpu(self, state)?;
self.phys.unmap(state.p2m_segment.pfn)?;
let pg_mfn = self.phys.p2m[pg_pfn as usize];
self.call
.mmuext(self.domid, MMUEXT_PIN_L4_TABLE, pg_mfn, 0)?;
self.setup_shared_info(state.shared_info_frame)?;
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) * X86_PAGE_SIZE;
vcpu.user_regs.rsi =
state.image_info.virt_base + (state.start_info_segment.pfn) * X86_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);
self.call.set_vcpu_context(self.domid, 0, &vcpu)?;
self.phys.unmap_all()?; self.phys.unmap_all()?;
self.gnttab_seed(state)?; self.gnttab_seed(state)?;
Ok(()) Ok(())
@ -365,161 +203,14 @@ impl BootSetup<'_> {
Ok(()) Ok(())
} }
fn setup_page_tables(&mut self, state: &mut BootState) -> Result<(), XenClientError> {
let p2m_guest = unsafe {
slice::from_raw_parts_mut(
state.p2m_segment.addr as *mut u64,
self.phys.p2m_size() as usize,
)
};
copy(p2m_guest, &self.phys.p2m);
for l in (0usize..X86_PGTABLE_LEVELS as usize).rev() {
for m1 in 0usize..state.page_table.mappings_count {
let map1 = &state.page_table.mappings[m1];
let from = map1.levels[l].from;
let to = map1.levels[l].to;
let pg_ptr = self.phys.pfn_to_ptr(map1.levels[l].pfn, 0)? as *mut u64;
for m2 in 0usize..state.page_table.mappings_count {
let map2 = &state.page_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, &state.page_table);
let pfn_paddr = self.phys.p2m[pfn as usize] << X86_PAGE_SHIFT;
let value = pfn_paddr | prot;
pg[p as usize] = value;
pfn += 1;
}
}
}
}
Ok(())
}
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, table: &PageTable) -> u64 {
let prot = [
BootSetup::PAGE_PRESENT | BootSetup::PAGE_RW | BootSetup::PAGE_ACCESSED,
BootSetup::PAGE_PRESENT
| BootSetup::PAGE_RW
| BootSetup::PAGE_ACCESSED
| BootSetup::PAGE_DIRTY
| BootSetup::PAGE_USER,
BootSetup::PAGE_PRESENT
| BootSetup::PAGE_RW
| BootSetup::PAGE_ACCESSED
| BootSetup::PAGE_DIRTY
| BootSetup::PAGE_USER,
BootSetup::PAGE_PRESENT
| BootSetup::PAGE_RW
| BootSetup::PAGE_ACCESSED
| BootSetup::PAGE_DIRTY
| BootSetup::PAGE_USER,
];
let prot = prot[l];
if l > 0 {
return prot;
}
for m in 0..table.mappings_count {
let map = &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 & !BootSetup::PAGE_RW;
}
}
prot
}
fn setup_start_info(&mut self, state: &BootState, cmdline: &str) -> Result<(), XenClientError> {
let ptr = self.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 = self.total_pages;
(*info).shared_info = state.shared_info_frame << X86_PAGE_SHIFT;
(*info).pt_base = state.page_table_segment.vstart;
(*info).nr_pt_frames = state.page_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 = self.phys.p2m[state.xenstore_segment.pfn as usize];
(*info).console.mfn = self.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, shared_info_frame: u64) -> Result<(), XenClientError> {
let info = self
.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 load_kernel_segment( fn load_kernel_segment(
&mut self, &mut self,
arch: &mut dyn ArchBootSetup,
image_loader: &dyn BootImageLoader, image_loader: &dyn BootImageLoader,
image_info: &BootImageInfo, image_info: &BootImageInfo,
) -> Result<DomainSegment, XenClientError> { ) -> Result<DomainSegment, XenClientError> {
let kernel_segment = self.alloc_segment( let kernel_segment = self.alloc_segment(
arch,
image_info.virt_kstart, image_info.virt_kstart,
image_info.virt_kend - image_info.virt_kstart, image_info.virt_kend - image_info.virt_kstart,
)?; )?;
@ -530,161 +221,22 @@ impl BootSetup<'_> {
Ok(kernel_segment) Ok(kernel_segment)
} }
fn count_page_tables( pub(crate) fn round_up(addr: u64, mask: u64) -> u64 {
&mut self,
table: &mut PageTable,
from: u64,
to: u64,
pfn: u64,
) -> Result<usize, XenClientError> {
debug!("counting pgtables from={} to={} pfn={}", from, to, pfn);
if table.mappings_count == X86_PAGE_TABLE_MAX_MAPPINGS {
return Err(XenClientError::new("too many mappings"));
}
let m = table.mappings_count;
let pfn_end = pfn + ((to - from) >> X86_PAGE_SHIFT);
if pfn_end >= self.phys.p2m_size() {
return Err(XenClientError::new("not enough memory for initial mapping"));
}
for idx in 0..table.mappings_count {
if from < table.mappings[idx].area.to && to > 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 = self.pfn_alloc_end + map.area.pgtables as u64;
if l as u64 == X86_PGTABLE_LEVELS - 1 {
if 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 table.mappings[0..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;
}
table.mappings[m] = map;
Ok(m)
}
fn alloc_p2m_segment(
&mut self,
page_table: &mut PageTable,
image_info: &BootImageInfo,
) -> Result<DomainSegment, XenClientError> {
let mut p2m_alloc_size =
((self.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(page_table, from, to, self.pfn_alloc_end)?;
let pgtables: usize;
{
let map = &mut page_table.mappings[m];
map.area.pfn = self.pfn_alloc_end;
for lvl_idx in 0..4 {
map.levels[lvl_idx].pfn += p2m_alloc_size >> X86_PAGE_SHIFT;
}
pgtables = map.area.pgtables;
}
page_table.mappings_count += 1;
p2m_alloc_size += (pgtables << X86_PAGE_SHIFT) as u64;
let p2m_segment = self.alloc_segment(0, p2m_alloc_size)?;
Ok(p2m_segment)
}
fn round_up(addr: u64, mask: u64) -> u64 {
addr | mask addr | mask
} }
fn bits_to_mask(bits: u64) -> u64 { pub(crate) fn bits_to_mask(bits: u64) -> u64 {
(1 << bits) - 1 (1 << bits) - 1
} }
fn alloc_page_tables( pub(crate) fn alloc_segment(
&mut self, &mut self,
table: &mut PageTable, arch: &mut dyn ArchBootSetup,
image_info: &BootImageInfo, start: u64,
size: u64,
) -> Result<DomainSegment, XenClientError> { ) -> 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(
self.virt_alloc_end + pages * X86_PAGE_SIZE,
BootSetup::bits_to_mask(22),
);
m = self.count_page_tables(table, image_info.virt_base, try_virt_end, 0)?;
pages = table.mappings[m].area.pgtables as u64 + extra_pages;
if self.virt_alloc_end + pages * X86_PAGE_SIZE <= try_virt_end + 1 {
break;
}
}
table.mappings[m].area.pfn = 0;
table.mappings_count += 1;
self.virt_pgtab_end = try_virt_end + 1;
let segment =
self.alloc_segment(0, table.mappings[m].area.pgtables as u64 * X86_PAGE_SIZE)?;
debug!(
"BootSetup alloc_page_tables table={:?} segment={:?}",
table, segment
);
Ok(segment)
}
fn alloc_segment(&mut self, start: u64, size: u64) -> Result<DomainSegment, XenClientError> {
if start > 0 { if start > 0 {
self.alloc_padding_pages(start)?; self.alloc_padding_pages(arch, start)?;
} }
let page_size: u32 = (1i64 << XEN_PAGE_SHIFT) as u32; let page_size: u32 = (1i64 << XEN_PAGE_SHIFT) as u32;
@ -693,14 +245,14 @@ impl BootSetup<'_> {
let mut segment = DomainSegment { let mut segment = DomainSegment {
vstart: start, vstart: start,
_vend: 0, vend: 0,
pfn: self.pfn_alloc_end, pfn: self.pfn_alloc_end,
addr: 0, addr: 0,
size, size,
pages, pages,
}; };
self.chk_alloc_pages(pages)?; self.chk_alloc_pages(arch, pages)?;
let ptr = self.phys.pfn_to_ptr(segment.pfn, pages)?; let ptr = self.phys.pfn_to_ptr(segment.pfn, pages)?;
segment.addr = ptr; segment.addr = ptr;
@ -708,23 +260,26 @@ impl BootSetup<'_> {
slice::from_raw_parts_mut(ptr as *mut u8, (pages * page_size as u64) as usize) slice::from_raw_parts_mut(ptr as *mut u8, (pages * page_size as u64) as usize)
}; };
slice.fill(0); slice.fill(0);
segment._vend = self.virt_alloc_end; segment.vend = self.virt_alloc_end;
debug!( debug!(
"BootSetup alloc_segment {:#x} -> {:#x} (pfn {:#x} + {:#x} pages)", "BootSetup alloc_segment {:#x} -> {:#x} (pfn {:#x} + {:#x} pages)",
start, segment._vend, segment.pfn, pages start, segment.vend, segment.pfn, pages
); );
Ok(segment) Ok(segment)
} }
fn alloc_page(&mut self) -> Result<DomainSegment, XenClientError> { fn alloc_page(
&mut self,
arch: &mut dyn ArchBootSetup,
) -> Result<DomainSegment, XenClientError> {
let start = self.virt_alloc_end; let start = self.virt_alloc_end;
let pfn = self.pfn_alloc_end; let pfn = self.pfn_alloc_end;
self.chk_alloc_pages(1)?; self.chk_alloc_pages(arch, 1)?;
debug!("BootSetup alloc_page {:#x} (pfn {:#x})", start, pfn); debug!("BootSetup alloc_page {:#x} (pfn {:#x})", start, pfn);
Ok(DomainSegment { Ok(DomainSegment {
vstart: start, vstart: start,
_vend: (start + X86_PAGE_SIZE) - 1, vend: (start + arch.page_size()) - 1,
pfn, pfn,
addr: 0, addr: 0,
size: 0, size: 0,
@ -732,15 +287,23 @@ impl BootSetup<'_> {
}) })
} }
fn alloc_module(&mut self, buffer: &[u8]) -> Result<DomainSegment, XenClientError> { fn alloc_module(
let segment = self.alloc_segment(0, buffer.len() as u64)?; &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()) }; let slice = unsafe { slice::from_raw_parts_mut(segment.addr as *mut u8, buffer.len()) };
copy(slice, buffer); copy(slice, buffer);
Ok(segment) Ok(segment)
} }
fn alloc_padding_pages(&mut self, boundary: u64) -> Result<(), XenClientError> { fn alloc_padding_pages(
if (boundary & (X86_PAGE_SIZE - 1)) != 0 { &mut self,
arch: &mut dyn ArchBootSetup,
boundary: u64,
) -> Result<(), XenClientError> {
if (boundary & (arch.page_size() - 1)) != 0 {
return Err(XenClientError::new( return Err(XenClientError::new(
format!("segment boundary isn't page aligned: {:#x}", boundary).as_str(), format!("segment boundary isn't page aligned: {:#x}", boundary).as_str(),
)); ));
@ -751,12 +314,16 @@ impl BootSetup<'_> {
format!("segment boundary too low: {:#x})", boundary).as_str(), format!("segment boundary too low: {:#x})", boundary).as_str(),
)); ));
} }
let pages = (boundary - self.virt_alloc_end) / X86_PAGE_SIZE; let pages = (boundary - self.virt_alloc_end) / arch.page_size();
self.chk_alloc_pages(pages)?; self.chk_alloc_pages(arch, pages)?;
Ok(()) Ok(())
} }
fn chk_alloc_pages(&mut self, pages: u64) -> Result<(), XenClientError> { fn chk_alloc_pages(
&mut self,
arch: &mut dyn ArchBootSetup,
pages: u64,
) -> Result<(), XenClientError> {
if pages > self.total_pages if pages > self.total_pages
|| self.pfn_alloc_end > self.total_pages || self.pfn_alloc_end > self.total_pages
|| pages > self.total_pages - self.pfn_alloc_end || pages > self.total_pages - self.pfn_alloc_end
@ -771,7 +338,57 @@ impl BootSetup<'_> {
} }
self.pfn_alloc_end += pages; self.pfn_alloc_end += pages;
self.virt_alloc_end += pages * X86_PAGE_SIZE; self.virt_alloc_end += pages * arch.page_size();
Ok(()) 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>;
}

50
xenclient/src/lib.rs
View File

@ -8,6 +8,7 @@ mod x86;
use crate::boot::BootSetup; use crate::boot::BootSetup;
use crate::create::DomainConfig; use crate::create::DomainConfig;
use crate::elfloader::ElfImageLoader; use crate::elfloader::ElfImageLoader;
use crate::x86::X86BootSetup;
use std::error::Error; use std::error::Error;
use std::fmt::{Display, Formatter}; use std::fmt::{Display, Formatter};
use std::fs::read; use std::fs::read;
@ -148,10 +149,7 @@ impl XenClient {
format!("{}/uuid", vm_path).as_str(), format!("{}/uuid", vm_path).as_str(),
&Uuid::from_bytes(domain.handle).to_string(), &Uuid::from_bytes(domain.handle).to_string(),
)?; )?;
tx.write_string( tx.write_string(format!("{}/name", vm_path).as_str(), "mycelium")?;
format!("{}/name", vm_path).as_str(),
"mycelium",
)?;
tx.write_string(format!("{}/type", libxl_path).as_str(), "pv")?; tx.write_string(format!("{}/type", libxl_path).as_str(), "pv")?;
tx.commit()?; tx.commit()?;
} }
@ -167,14 +165,16 @@ impl XenClient {
{ {
let mut boot = BootSetup::new(&self.call, domid); let mut boot = BootSetup::new(&self.call, domid);
let mut arch = X86BootSetup::new();
let initrd = read(config.initrd_path.as_str())?; let initrd = read(config.initrd_path.as_str())?;
let mut state = boot.initialize( let mut state = boot.initialize(
&mut arch,
&image_loader, &image_loader,
initrd.as_slice(), initrd.as_slice(),
config.max_vcpus, config.max_vcpus,
config.mem_mb, config.mem_mb,
)?; )?;
boot.boot(&mut state, config.cmdline.as_str())?; boot.boot(&mut arch, &mut state, config.cmdline.as_str())?;
console_evtchn = state.console_evtchn; console_evtchn = state.console_evtchn;
xenstore_evtchn = state.store_evtchn; xenstore_evtchn = state.store_evtchn;
console_mfn = boot.phys.p2m[state.console_segment.pfn as usize]; console_mfn = boot.phys.p2m[state.console_segment.pfn as usize];
@ -183,10 +183,7 @@ impl XenClient {
{ {
let mut tx = self.store.transaction()?; let mut tx = self.store.transaction()?;
tx.write_string( tx.write_string(format!("{}/image/os_type", vm_path).as_str(), "linux")?;
format!("{}/image/os_type", vm_path).as_str(),
"linux",
)?;
tx.write_string( tx.write_string(
format!("{}/image/kernel", vm_path).as_str(), format!("{}/image/kernel", vm_path).as_str(),
&config.kernel_path, &config.kernel_path,
@ -208,10 +205,7 @@ impl XenClient {
format!("{}/memory/target", dom_path).as_str(), format!("{}/memory/target", dom_path).as_str(),
&(config.mem_mb * 1024).to_string(), &(config.mem_mb * 1024).to_string(),
)?; )?;
tx.write_string( tx.write_string(format!("{}/memory/videoram", dom_path).as_str(), "0")?;
format!("{}/memory/videoram", dom_path).as_str(),
"0",
)?;
tx.write_string(format!("{}/domid", dom_path).as_str(), &domid.to_string())?; tx.write_string(format!("{}/domid", dom_path).as_str(), &domid.to_string())?;
tx.write_string( tx.write_string(
format!("{}/store/port", dom_path).as_str(), format!("{}/store/port", dom_path).as_str(),
@ -252,32 +246,14 @@ impl XenClient {
format!("{}/frontend-id", backend_path).as_str(), format!("{}/frontend-id", backend_path).as_str(),
&domid.to_string(), &domid.to_string(),
)?; )?;
tx.write_string( tx.write_string(format!("{}/online", backend_path).as_str(), "1")?;
format!("{}/online", backend_path).as_str(),
"1",
)?;
tx.write_string(format!("{}/state", backend_path).as_str(), "1")?; tx.write_string(format!("{}/state", backend_path).as_str(), "1")?;
tx.write_string( tx.write_string(format!("{}/protocol", backend_path).as_str(), "vt100")?;
format!("{}/protocol", backend_path).as_str(),
"vt100",
)?;
tx.write_string( tx.write_string(format!("{}/backend-id", frontend_path).as_str(), "0")?;
format!("{}/backend-id", frontend_path).as_str(), tx.write_string(format!("{}/limit", frontend_path).as_str(), "1048576")?;
"0", tx.write_string(format!("{}/type", frontend_path).as_str(), "xenconsoled")?;
)?; tx.write_string(format!("{}/output", frontend_path).as_str(), "pty")?;
tx.write_string(
format!("{}/limit", frontend_path).as_str(),
"1048576",
)?;
tx.write_string(
format!("{}/type", frontend_path).as_str(),
"xenconsoled",
)?;
tx.write_string(
format!("{}/output", frontend_path).as_str(),
"pty",
)?;
tx.write_string(format!("{}/tty", frontend_path).as_str(), "")?; tx.write_string(format!("{}/tty", frontend_path).as_str(), "")?;
tx.write_string( tx.write_string(
format!("{}/port", frontend_path).as_str(), format!("{}/port", frontend_path).as_str(),

522
xenclient/src/x86.rs
View File

@ -1,4 +1,18 @@
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 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_SHIFT: u64 = 12;
pub const X86_PAGE_SIZE: u64 = 1 << X86_PAGE_SHIFT; pub const X86_PAGE_SIZE: u64 = 1 << X86_PAGE_SHIFT;
@ -112,3 +126,511 @@ pub struct SharedInfo {
pub p2m_vaddr: u64, pub p2m_vaddr: u64,
pub p2m_generation: 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 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 meminit(&mut self, setup: &mut BootSetup, total_pages: u64) -> Result<(), XenClientError> {
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);
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 page_shift(&mut self) -> u64 {
X86_PAGE_SHIFT
}
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(())
}
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(())
}
}