kernel: pmm with overlapping check

This commit is contained in:
janis 2026-08-05 21:34:00 +02:00
parent 86d80132be
commit ae06f75cc2
Signed by: janis
SSH key fingerprint: SHA256:bB1qbbqmDXZNT0KKD5c2Dfjg53JGhj7B3CFcLIzSqq8
3 changed files with 503 additions and 91 deletions

View file

@ -8,6 +8,7 @@
ptr_cast_slice,
likely_unlikely,
int_roundings,
slice_ptr_get,
never_type
)]
#![cfg_attr(test, feature(custom_test_frameworks))]

View file

@ -27,15 +27,15 @@ extern "C" fn _start() -> ! {
kernel::limine::init_limine_boot_info();
kernel::serial_println!(
"HHDM offset: 0x{:#x}",
kernel::memory::HHDM_BASE.get().unwrap()
);
// kernel::serial_println!(
// "HHDM offset: 0x{:#x}",
// kernel::memory::HHDM_BASE.get().unwrap()
// );
kernel::serial_println!(
"Memory map: {:#?}",
kernel::limine::MEMMAP_REQUEST.entries()
);
// kernel::serial_println!(
// "Memory map: {:#?}",
// kernel::limine::MEMMAP_REQUEST.entries()
// );
GDT.load();
IDT.load();

View file

@ -4,10 +4,11 @@ use core::{
fmt::Debug,
num::{NonZero, NonZeroUsize},
ptr::NonNull,
range::Range,
};
use bit_field::BitField;
use rbtree::RBTree;
use rbtree::{RBTree, UnsafeNode};
use crate::{serial_println, sync::OnceLock, x86_64::PAGE_SIZE};
@ -22,7 +23,7 @@ pub trait VirtAddrTranslationExt: Sized {
}
}
#[derive(Clone, Copy, PartialEq, Eq)]
#[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
pub struct PhyAddr(pub u64);
impl PhyAddr {
@ -119,11 +120,6 @@ impl VirtAddr {
}
}
pub struct PageChunk {
pub next: PhyAddr,
pub count: usize,
}
pub struct PhysicalMemoryManager {
/// On amd64 platforms, there are at most 2^40 pages of addressable physical
/// memory. We maintain a binary tree of free page chunks for each
@ -165,20 +161,21 @@ impl Debug for PhysicalMemoryManager {
impl PhysicalMemoryManager {
pub fn from_memory_map(memory_map: &[crate::boot::MemoryRegion]) -> Self {
let usable_regions = memory_map
.iter()
.filter(|region| region.region_type.is_usable());
let mut pmm = PhysicalMemoryManager {
buddies: [(); 40].map(|_| RBTree::default()),
};
let buddies = [(); 40].map(|_| RBTree::default());
let mut free_usable_regions = free_usable_regions_from_memory_map(memory_map);
let mut pmm = PhysicalMemoryManager { buddies };
serial_println!("Freeing usable regions: {:#?}", free_usable_regions);
usable_regions.for_each(|region| {
let idx = region.start / crate::x86_64::PAGE_SIZE as u64;
let count = region.length / crate::x86_64::PAGE_SIZE as u64;
pmm.free_region(idx as usize, count as usize);
});
for region in free_usable_regions
.drain()
.map(|p| unsafe { p.cast::<PageHeader>().as_ref() })
{
let (idx, count) = region.index_and_count();
pmm.free_region(idx, count);
}
pmm
}
@ -226,6 +223,8 @@ impl PhysicalMemoryManager {
unsafe { chunk.as_ref().phy().map(|phy| (phy, 1 << bin)) }
}
// TODO: grow and shrink
fn free_region(&mut self, page_index: usize, mut count: usize) {
serial_println!(
"Freeing region: page_index = {:#x}, count = {:#x}",
@ -286,7 +285,7 @@ impl PhysicalMemoryManager {
);
let None = self.buddies[bin].insert_node(node_ptr) else {
panic!(
"Attempted to free a region that is already free: page_index = {}, count = {}",
"Attempted to free a region that is already free: page_index = {:#x}, count = {:#x}",
page_index, bit
)
};
@ -294,10 +293,6 @@ impl PhysicalMemoryManager {
}
}
pub struct PhysicalMemoryAllocator {
tree: RBTree<PhysicalPageNode>,
}
#[repr(C)]
#[derive(Debug)]
struct PageHeader {
@ -305,11 +300,79 @@ struct PageHeader {
count: u64,
}
unsafe impl UnsafeNode for PageHeader {
type Key = PhysicalPageNodeKey;
fn left(&self) -> Option<NonNull<Self>> {
self.node.left().map(Self::unsafe_from_node)
}
fn right(&self) -> Option<NonNull<Self>> {
self.node.right().map(Self::unsafe_from_node)
}
fn parent(&self) -> Option<NonNull<Self>> {
self.node.parent().map(Self::unsafe_from_node)
}
fn key(&self) -> &Self::Key {
self.node.key()
}
fn color(&self) -> rbtree::Color {
self.node.color()
}
fn set_left(&self, left: Option<NonNull<Self>>) {
self.node.set_left(left.map(Self::unsafe_into_node))
}
fn set_right(&self, right: Option<NonNull<Self>>) {
self.node.set_right(right.map(Self::unsafe_into_node))
}
fn set_parent(&self, parent: Option<NonNull<Self>>) {
self.node.set_parent(parent.map(Self::unsafe_into_node))
}
fn set_color(&self, color: rbtree::Color) {
self.node.set_color(color)
}
}
impl PageHeader {
fn phy(&self) -> Option<PhyAddr> {
PhyAddr::from_hhdm_virt(self)
}
fn unsafe_from_node(node: NonNull<PhysicalPageNode>) -> NonNull<Self> {
node.cast::<Self>()
}
fn unsafe_into_node(ptr: NonNull<Self>) -> NonNull<PhysicalPageNode> {
ptr.cast::<PhysicalPageNode>()
}
fn contains(&self, phy: PhyAddr) -> bool {
let start = self.phy().expect("PageHeader is not in HHDM");
let end = start.page_add(self.count as usize);
phy >= start && phy < end
}
fn page_range(&self) -> Range<usize> {
let (start, count) = self.index_and_count();
(start..start + count).into()
}
fn index_and_count(&self) -> (usize, usize) {
let start = self.phy().expect("PageHeader is not in HHDM").page_index() as usize;
let count = self.count as usize;
(start, count)
}
fn new_from_page_idx_and_count(page_idx: usize, count: usize) -> NonNull<Self> {
let phy = PhyAddr(page_idx as u64 * PAGE_SIZE as u64);
let virt = phy.as_hhdm_virt();
@ -326,72 +389,112 @@ impl PageHeader {
}
}
impl Debug for PhysicalMemoryAllocator {
fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
f.debug_struct("PhysicalMemoryAllocator")
.field_with("tree", |f| {
let iter = self.tree.iter();
write!(f, "[")?;
for key in iter {
let header = unsafe { (&raw const *key).cast::<PageHeader>().read_volatile() };
fn free_usable_regions_from_memory_map(
memory_map: &[crate::boot::MemoryRegion],
) -> RBTree<PageHeader> {
let mut usable_regions = memory_map
.iter()
.copied()
.filter(|region| region.region_type.is_usable())
.fold(RBTree::new(), |mut tree, region| {
let idx = region.start as usize / crate::x86_64::PAGE_SIZE;
let count = region.length as usize / crate::x86_64::PAGE_SIZE;
if let Some(phy) = header.phy() {
writeln!(
f,
"({:?}..{:?}), ",
phy,
phy.page_add(header.count as usize)
)?;
}
}
write!(f, "]")
})
.finish()
}
}
let node_ptr = PageHeader::new_from_page_idx_and_count(idx, count);
let None = tree.insert_node(node_ptr) else {
panic!(
"Failed to insert node for page_index = {}, count = {}",
idx, count
);
};
tree
});
impl PhysicalMemoryAllocator {
pub fn from_memory_map(memory_map: &[crate::boot::MemoryRegion]) -> Self {
let mapped_non_usable_regions = memory_map
.iter()
.filter(|region| region.region_type.mapped_non_usable());
let usable_regions = memory_map
.iter()
.copied()
.filter(|region| region.region_type.is_usable());
for region in mapped_non_usable_regions {
let idx = region.start as usize / crate::x86_64::PAGE_SIZE;
let count = region.length as usize / crate::x86_64::PAGE_SIZE;
let end = idx + count;
let mut pmm = PhysicalMemoryAllocator {
tree: RBTree::new(),
};
while let Some(overlapping) = usable_regions
.range(
PhysicalPageNodeKey::from_page_index(idx)
..PhysicalPageNodeKey::from_page_index(idx + count),
)
.map(|p| unsafe {
p.phy()
.unwrap()
.as_hhdm_virt()
.as_ptr::<PageHeader>()
.as_ref()
.unwrap()
})
.find(|header| {
header.contains(PhyAddr::from_page_index(idx))
|| header.contains(PhyAddr::from_page_index(idx + count - 1))
})
.inspect(|header| {
usable_regions
.remove(header.node.key())
.expect("node exists");
})
{
// We found an overlapping region.
usable_regions.for_each(|region| {
let idx = region.start / crate::x86_64::PAGE_SIZE as u64;
let count = region.length / crate::x86_64::PAGE_SIZE as u64;
serial_println!("Found overlapping region: {:?}", overlapping);
pmm.free_region(idx as usize, count as usize);
});
let Range {
start: overlap_idx,
end: overlap_end,
} = overlapping.page_range();
let overlap_count = overlap_end - overlap_idx;
pmm
}
pub fn free_region(&mut self, page_index: usize, count: usize) {
serial_println!(
"Freeing region: page_index = {}, count = {}",
page_index,
count
);
let node_ptr =
PageHeader::new_from_page_idx_and_count(page_index, count).cast::<PhysicalPageNode>();
let None = self.tree.insert_node(node_ptr) else {
panic!(
"Failed to insert node for page_index = {}, count = {}",
page_index, count
);
// check if our region is a suffix or prefix of the overlapping region
if overlap_idx == idx {
if let Some(suffix_count) = overlap_count.checked_sub(count) {
let None = usable_regions.insert_node(PageHeader::new_from_page_idx_and_count(
overlap_idx + count,
suffix_count,
)) else {
panic!("pmm: double free")
};
}
} else if overlap_end == end {
if let Some(suffix_count) = overlap_count.checked_sub(count) {
let None = usable_regions.insert_node(PageHeader::new_from_page_idx_and_count(
overlap_idx,
suffix_count,
)) else {
panic!("pmm: double free")
};
}
} else {
// region is in the middle of the overlapping region
if let Some(prefix_count) = idx.checked_sub(overlap_idx) {
let None = usable_regions.insert_node(PageHeader::new_from_page_idx_and_count(
overlap_idx,
prefix_count,
)) else {
panic!("pmm: double free")
};
}
if let Some(suffix_count) = overlap_end.checked_sub(end) {
let None = usable_regions
.insert_node(PageHeader::new_from_page_idx_and_count(end, suffix_count))
else {
panic!("pmm: double free")
};
}
}
}
}
usable_regions
}
#[derive(Debug, Default)]
@ -629,7 +732,315 @@ unsafe impl rbtree::UnsafeNode for PhysicalPageNode {
}
}
pub mod slab {
//! A slab allocator
use core::{
alloc::Layout,
cell::Cell,
hint::{cold_path, unlikely},
num::NonZero,
ptr::NonNull,
};
const UNLINKED: NonNull<()> = unsafe { NonNull::new_unchecked(!0 as *mut ()) };
use alloc::alloc::Allocator;
use crate::x86_64::PAGE_SIZE;
pub struct Slab<A: Allocator + Clone> {
/// Size and alignment of each element in the slab.
element_size: usize,
/// Pointer to the first chunk in the slab.
head: Option<NonNull<SlabChunk<A>>>,
alloc: A,
}
struct SlabChunk<A: Allocator + Clone> {
/// Pointer to the next chunk in the slab.
next: Option<NonNull<SlabChunk<A>>>,
/// Pointer to the slab that owns this chunk.
slab: NonNull<Slab<A>>,
/// Linked list of free elements in this slab. When this is `None`, the
/// slab is full.
free: Cell<Option<NonNull<ChunkSlot>>>,
/// Number of outstanding allocations from this slab. When this reaches
/// zero, the slab can be freed.
count: Cell<usize>,
}
struct ChunkSlot(Option<NonNull<Self>>);
enum SlotResult {
Some(NonNull<u8>),
Last(NonNull<u8>),
None,
}
impl<A: Allocator + Clone> SlabChunk<A> {
fn pop_free_slot(&self) -> SlotResult {
let Some(slot) = self.free.get() else {
return SlotResult::None;
};
self.free.set(unsafe { slot.as_ref() }.next());
self.count.update(|count| count + 1);
match self.free.get() {
Some(_) => SlotResult::Some(slot.cast()),
None => SlotResult::Last(slot.cast()),
}
}
/// returns `true` if the slab is now empty and can be freed
fn push_free_slot(&self, slot: NonNull<u8>) -> bool {
let slot = slot.cast::<ChunkSlot>();
let next = self.free.get();
unsafe { slot.as_ptr().write(ChunkSlot(next)) };
self.free.set(Some(slot));
self.count.update(|count| count - 1);
self.count.get() == 0
}
}
impl ChunkSlot {
fn next(&self) -> Option<NonNull<Self>> {
self.0
}
}
impl<A: Allocator + Clone> Slab<A> {
fn new(element_size: usize, alloc: A) -> Self {
assert!(
element_size.is_power_of_two(),
"element_size must be a power of two"
);
Self {
element_size,
head: None,
alloc,
}
}
fn first_slot_offset(&self) -> usize {
foundation::mem::align_up(core::mem::size_of::<SlabChunk<A>>(), self.element_size)
}
fn count_and_layout(&self) -> (usize, Layout) {
let (count, size, align) = {
let one_page_count = (PAGE_SIZE - self.first_slot_offset()) / self.element_size;
if one_page_count < 3 {
let count = 3;
let size =
(self.first_slot_offset() + count * self.element_size).next_power_of_two();
assert!(size.is_multiple_of(PAGE_SIZE));
assert!(size >= PAGE_SIZE);
assert!(size.is_multiple_of(self.element_size));
(count, size, size)
} else {
let count = one_page_count;
let size = PAGE_SIZE;
(count, size, self.element_size)
}
};
(count, unsafe {
Layout::from_size_align_unchecked(size, align)
})
}
fn alloc_chunk(&mut self) -> NonNull<SlabChunk<A>> {
// we want to limit chunks to 1 page unless the element size is so
// large that we can fit fewer than 3 elements in a page.
let (count, layout) = self.count_and_layout();
let Some(bytes) = self.alloc.allocate(layout).ok() else {
panic!()
};
let chunk = bytes.as_non_null_ptr().cast::<SlabChunk<A>>();
unsafe {
let first_slot = chunk
.as_ptr()
.byte_add(self.first_slot_offset())
.cast::<ChunkSlot>();
for i in 0..(count - 1) {
let chunk = first_slot.byte_add(i * self.element_size);
let next = first_slot.byte_add((i + 1) * self.element_size);
chunk.write(ChunkSlot(Some(NonNull::new_unchecked(next))));
}
first_slot
.byte_add((count - 1) * self.element_size)
.write(ChunkSlot(None));
chunk.write(SlabChunk {
next: self.head,
slab: NonNull::from(self),
free: Cell::new(Some(NonNull::new_unchecked(first_slot))),
count: Cell::new(0),
});
}
chunk
}
#[cold]
fn alloc_chunk_cold(&mut self) -> NonNull<SlabChunk<A>> {
self.alloc_chunk()
}
fn alloc_slot(&mut self) -> NonNull<[u8]> {
let mut chunk = match self.head {
Some(chunk) => chunk,
None => {
let chunk = self.alloc_chunk_cold();
self.head = Some(chunk);
chunk
}
};
let chunk = unsafe { chunk.as_mut() };
let ptr = match chunk.pop_free_slot() {
SlotResult::Some(non_null) => non_null,
SlotResult::Last(non_null) => {
self.head = chunk.next.replace(UNLINKED.cast());
non_null
}
SlotResult::None => {
panic!("SlabChunk is full, but it is still the head of the slab");
}
};
ptr.cast_slice(self.element_size)
}
fn free_slot(&mut self, slot: NonNull<u8>) {
let (_, layout) = self.count_and_layout();
let mut chunk = slot
.map_addr(|addr| unsafe {
NonZero::new_unchecked(foundation::mem::align_down(addr.get(), layout.align()))
})
.cast::<SlabChunk<A>>();
let chunk = unsafe { chunk.as_mut() };
let linked = chunk.next != Some(UNLINKED.cast());
if chunk.push_free_slot(slot) {
if linked {
let mut head = self.head.expect("chunk is linked, so head exists");
while let Some(next) = unsafe { head.as_ref().next } {
if next == chunk.into() {
unsafe { head.as_mut().next = chunk.next };
break;
}
head = next;
}
}
unsafe {
self.alloc
.deallocate(NonNull::from_mut(chunk).cast(), layout)
};
} else if !linked {
chunk.next = self.head;
self.head = Some(chunk.into());
}
}
}
const SLAB_ALLOCATOR_BUCKETS: usize = 8;
pub struct SlabAllocator<A: Allocator + Clone> {
/// Slabs for each power-of-two from 16 bytes to 2048 bytes (inclusive).
slabs: [Slab<A>; SLAB_ALLOCATOR_BUCKETS],
alloc: A,
}
impl<A: Allocator + Clone> SlabAllocator<A> {
pub fn new(alloc: A) -> Self {
let slabs = [
Slab::new(16, alloc.clone()),
Slab::new(32, alloc.clone()),
Slab::new(64, alloc.clone()),
Slab::new(128, alloc.clone()),
Slab::new(256, alloc.clone()),
Slab::new(512, alloc.clone()),
Slab::new(1024, alloc.clone()),
Slab::new(2048, alloc.clone()),
];
Self { slabs, alloc }
}
fn slab_index_for_size(size: usize) -> Option<usize> {
if size == 0 {
return None;
}
// the smallest slab is 16 bytes
let size = size.max(16);
// get the index of the slab by calculating the log2 of the size and
// subtracting 4 (since 2^4 = 16)
let index =
(size.next_power_of_two().trailing_zeros() - 16usize.trailing_zeros()) as usize;
// we have 8 slabs, so the index must be less than 8
if index < SLAB_ALLOCATOR_BUCKETS {
Some(index)
} else {
None
}
}
pub fn alloc(&mut self, layout: Layout) -> Option<NonNull<[u8]>> {
if unlikely(layout.size() == 0) {
return Some(NonNull::dangling().cast_slice(0));
}
let size = layout.size().max(layout.align());
match Self::slab_index_for_size(size) {
Some(slab_index) => {
Some(unsafe { self.slabs.get_unchecked_mut(slab_index).alloc_slot() })
}
None => {
// allocate directly from the backing allocator
self.alloc.allocate(layout).ok()
}
}
}
pub fn dealloc(&mut self, ptr: NonNull<u8>, layout: Layout) {
if unlikely(layout.size() == 0) {
return;
}
let size = layout.size().max(layout.align());
match Self::slab_index_for_size(size) {
Some(slab_index) => unsafe {
self.slabs.get_unchecked_mut(slab_index).free_slot(ptr)
},
None => {
// deallocate directly to the backing allocator
unsafe { self.alloc.deallocate(ptr, layout) }
}
}
}
}
}
pub mod bump {
//! A bump allocator inspired by / taken from the `stumpalo` crate
use core::{
alloc::{Allocator, Layout},
cell::Cell,