strat9_kernel/memory/

frame.rs

//! Physical frame allocator abstraction.
//!
//! ## MetaSlot (per-frame metadata, issue #38)
//!
//! Each 4 KiB physical frame has a **dedicated 64-byte [`MetaSlot`]**
//! in a separate contiguous array (initialized by [`init_metadata_array`]). Buddy
//! free-list [`FreeListLink`] nodes, reference counts, purpose flags,
//! [`meta_guard`] bits, a per-allocation **generation** counter, and an optional
//! [`FrameMetaVtable`] live here : **not** in the mapped page bytes, so mappings
//! see a pristine payload.
//!
//! ## Revue / invariants (issue #38)
//!
//! - **Pas de métadonnées dans la charge utile** : les liens buddy sont dans
//!   [`FreeListLink`], jamais écrits comme « faux pointeurs » dans les 4 KiB mappés.
//! - **`generation`** : incrémentée uniquement par [`MetaSlot::note_new_allocation_epoch`]
//!   après un `CAS` réussi dans [`FrameAllocOptions::allocate`]. Ne pas utiliser
//!   [`MetaSlot::set_generation`] sauf bootstrap/tests : sinon les schémas « généalogiques »
//!   deviennent incohérents.
//! - **`meta_guard::POISONED` vs `frame_flags::POISONED`** : deux espaces (bits dédiés
//!   `guard` vs flags logiques). Pour marquer une frame corrompue, préférer
//!   [`MetaSlot::mark_poisoned`] qui pose les deux.
//! - **`vtable_ref` / `try_vtable_ref`** : bits `0` → défaut ; bits non alignés ou invalides
//!   → défaut (pas d’UB). Les pointeurs alignés doivent désigner une [`FrameMetaVtable`] `'static`
//!   valide lorsqu’ils sont enregistrés par le noyau.
//! - **Cache order-0** : `buddy::alloc(0)` peut servir depuis le cache local ; le chemin
//!   [`FrameAllocOptions::allocate`] applique quand même le CAS + epoch sur la même frame.

use crate::{memory::boot_alloc::BootAllocator, sync::IrqDisabledToken};
use core::{
    mem::{self, offset_of},
    ptr,
    sync::atomic::{AtomicU32, AtomicU64, AtomicU8, Ordering},
};
use x86_64::PhysAddr;

// ==============================================================================
// FrameAllocOptions  (Asterinas OSTD pattern)
// ==============================================================================
//
// DESIGN NOTES : why this wrapper exists:
//
//  * In Asterinas OSTD, `FrameAllocOptions::new()` defaults to `zeroed: true`.
//    This means callers can never accidentally hand out a frame that still holds
//    data from a previous lifetime.  The only way to skip zeroing is an
//    explicit `.zeroed(false)` call at the site that *knows* it is safe to do
//    so (e.g. a frame that will be fully overwritten before any read).
//
//  * The critical failure mode we are fixing:
//    `BuddyFrameAllocator::allocate_frame` (used by `OffsetPageTable` when it
//    needs a new intermediate page-table node) was returning raw, unzeroed
//    frames.  A freshly-split buddy block can contain bytes left behind by the
//    slab allocator (POISON_BYTE = 0xDE) or by whatever previously lived in
//    that memory.  The CPU page-table walker reads all 512 entries of every
//    intermediate node it traverses.  A random non-zero entry is decoded as a
//    valid PTE pointing to an arbitrary physical address : which explains why
//    RIP (the first fetch address the CPU tries after entering Ring 3) changes
//    on every boot.
//
//  * The `flags` field mirrors OSTD's per-frame metadata: we stamp the purpose
//    (kernel / user / page-table) into `FrameMeta::flags` atomically using
//    `Ordering::Release` so that any CPU that later reads the frame through
//    `get_meta` observes the correct flags.
//
//  * Refcount state machine (OSTD-style, fully enforced):
//
//    `buddy.rs` maintains the invariant: free-list frame ⟹ refcount == REFCOUNT_UNUSED.
//    `mark_block_free()` stamps REFCOUNT_UNUSED; `mark_block_allocated()` leaves it
//    untouched.  `FrameAllocOptions::allocate()` performs CAS(REFCOUNT_UNUSED -> 1)
//    as a fail-fast corruption check before publishing the frame as live:
//
//       buddy alloc ▶ optional zero ▶ set flags ▶ CAS(UNUSED -> 1) ▶ live

/// Sentinel refcount for a frame that is in the buddy free list.
///
/// Mirrors `REF_COUNT_UNUSED` in Asterinas OSTD `meta.rs`.
///
/// `buddy.rs` stamps this value in `mark_block_free()` and leaves it intact in
/// `mark_block_allocated()`.  `FrameAllocOptions::allocate()` performs
/// `CAS(REFCOUNT_UNUSED -> 1)` to atomically claim the frame and detect any
/// double-free / free-list corruption.
pub const REFCOUNT_UNUSED: u32 = u32::MAX;

/// Options controlling how a physical frame is allocated.
///
/// The default configuration (`FrameAllocOptions::new()`) produces a
/// **zeroed** frame.  Callers that need a non-zeroed frame (e.g. DMA buffers
/// that are immediately filled by hardware, or frames that will be fully
/// overwritten before any read) must explicitly call `.zeroed(false)`.
///
/// # Example
///
/// ```ignore
/// // Allocate a zeroed page-table frame (the safe default).
/// let frame = FrameAllocOptions::new()
///     .purpose(FramePurpose::PageTable)
///     .allocate(token)?;
///
/// // Allocate a user-data frame without zeroing (caller guarantees it will
/// // be fully overwritten, e.g. by an ELF segment load).
/// let frame = FrameAllocOptions::new()
///     .zeroed(false)
///     .purpose(FramePurpose::UserData)
///     .allocate(token)?;
/// ```
pub struct FrameAllocOptions {
    /// Whether the frame content should be zeroed before being returned.
    ///
    /// Defaults to `true`.  Setting this to `false` is only safe when the
    /// caller guarantees the frame will be fully written before any read.
    zeroed: bool,
    /// The logical purpose of the frame, encoded as `frame_flags` bits.
    purpose_flags: u32,
}

/// Describes the intended purpose of an allocated frame.
///
/// Purpose is written into `FrameMeta::flags` with `Ordering::Release` so
/// that any concurrent reader of the metadata (e.g. a TLB-shootdown handler
/// deciding whether a frame holds a page-table node) sees a consistent view.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum FramePurpose {
    /// Frame will hold a kernel page-table node (PML4/PDPT/PD/PT).
    ///
    /// These frames MUST be zeroed : unzeroed page-table nodes are the primary
    /// source of non-deterministic RIP at Ring 3 transition.
    PageTable,
    /// Frame belongs to kernel address-space (e.g. heap, stack, metadata).
    KernelData,
    /// Frame belongs to a user-space address-space (anonymous or file-backed).
    UserData,
    /// Caller-managed; raw flags are passed through unchanged.
    Custom(u32),
}

impl FramePurpose {
    fn to_flags(self) -> u32 {
        match self {
            // Page-table frames are always kernel-owned.
            Self::PageTable => frame_flags::KERNEL | frame_flags::ALLOCATED,
            Self::KernelData => frame_flags::KERNEL | frame_flags::ALLOCATED,
            Self::UserData => frame_flags::USER | frame_flags::ALLOCATED | frame_flags::MOVABLE,
            Self::Custom(f) => f | frame_flags::ALLOCATED,
        }
    }

    /// Returns `true` if this purpose requires zeroing regardless of the
    /// `zeroed` option.  Page-table nodes must always be zeroed.
    pub fn requires_zero(self) -> bool {
        matches!(self, Self::PageTable)
    }
}

impl Default for FrameAllocOptions {
    fn default() -> Self {
        Self::new()
    }
}

impl FrameAllocOptions {
    /// Creates allocation options with safe defaults:
    ///  - `zeroed = true`
    ///  - purpose = `KernelData`
    pub fn new() -> Self {
        Self {
            zeroed: true,
            purpose_flags: FramePurpose::KernelData.to_flags(),
        }
    }

    /// Override the zero-initialisation policy.
    ///
    /// # Safety contract (enforced by convention, not the type system)
    ///
    /// If `zeroed` is set to `false`, the caller MUST fully overwrite every
    /// byte of the frame before allowing any other CPU or subsystem to read it.
    /// Violating this rule is a memory-safety hazard: stale bytes in an
    /// intermediate page-table node cause the CPU to follow arbitrary PTEs.
    pub fn zeroed(mut self, zeroed: bool) -> Self {
        self.zeroed = zeroed;
        self
    }

    /// Set the intended purpose of the frame.
    ///
    /// `PageTable` purpose forces zeroing even if `.zeroed(false)` was called.
    pub fn purpose(mut self, p: FramePurpose) -> Self {
        self.purpose_flags = p.to_flags();
        // Page-table nodes must always be zeroed : override any caller setting.
        if p.requires_zero() {
            self.zeroed = true;
        }
        self
    }

    /// Allocate a single 4 KiB frame according to the configured options.
    ///
    /// The allocation path is:
    ///
    /// 1. Ask the buddy allocator for an order-0 frame (exclusive ownership is
    ///    guaranteed by the buddy's own bitmap + free-list discipline).
    /// 2. Optionally zero the 4 KiB frame contents via the HHDM.
    /// 3. Stamp `FrameMeta::flags` with the purpose flags using `Release`
    ///    ordering.
    /// 4. Store `refcount = 1` with `Release` ordering so any later `Acquire`
    ///    load of the refcount observes the fully-initialised metadata and
    ///    (if zeroed) zeroed content.
    ///
    /// # Sentinel handoff: `CAS(REFCOUNT_UNUSED -> 1)`
    ///
    /// `buddy.rs` maintains the invariant that every frame on the free list has
    /// `refcount == REFCOUNT_UNUSED`.  `mark_block_allocated()` leaves this
    /// sentinel intact, so the frame arriving here still carries `REFCOUNT_UNUSED`.
    ///
    /// The CAS atomically claims the frame and acts as a fail-fast corruption
    /// check: if the same frame appears twice in the buddy free list (double-free
    /// or metadata corruption), the second allocation attempt will observe a
    /// refcount of `1` (set by the first allocation) and panic immediately rather
    /// than silently aliasing memory.
    pub fn allocate(self, token: &IrqDisabledToken) -> Result<PhysFrame, AllocError> {
        let migratetype = if self.purpose_flags & frame_flags::MOVABLE != 0 {
            crate::memory::zone::Migratetype::Movable
        } else {
            crate::memory::zone::Migratetype::Unmovable
        };

        // Step 1 : exclusive frame from the buddy allocator.
        let frame = crate::memory::buddy::alloc_migratetype(token, 0, migratetype)?;
        let phys = frame.start_address.as_u64();

        // SAFETY: `get_meta` panics only if `phys` is out-of-bounds, which
        // would be a buddy-level invariant violation (it returned an address
        // beyond the metadata array).  That is a kernel bug, not UB here.
        let meta = get_meta(frame.start_address);

        // Step 2 : zero the frame content if required.
        //
        // The zeroing MUST happen before the `Release` store of `refcount = 1`
        // (step 4) so that any thread performing an `Acquire` load of the
        // refcount and then reading frame bytes observes zeros.
        //
        // For `FramePurpose::PageTable` this is unconditional: the CPU's
        // page-table walker reads all 512 entries of every intermediate node it
        // visits.  Stale non-zero bytes would be decoded as valid PTEs pointing
        // to arbitrary physical addresses, producing a non-deterministic RIP on
        // Ring 3 entry (the root cause of the original bug).
        //
        // SAFETY: `phys_to_virt(phys)` is a valid HHDM address covering exactly
        // `PAGE_SIZE` bytes.  The buddy allocator guarantees we have exclusive
        // ownership of these bytes for the duration of this function.
        if self.zeroed {
            unsafe {
                ptr::write_bytes(
                    crate::memory::phys_to_virt(phys) as *mut u8,
                    0,
                    PAGE_SIZE as usize,
                );
            }
        }

        // Step 3 : stamp purpose flags with `Release` ordering.
        //
        // Any reader that subsequently loads `refcount` with `Acquire` (step 4)
        // is guaranteed to observe these flags as well.
        meta.flags.store(self.purpose_flags, Ordering::Release);
        meta.set_order(0);

        // Step 4 : claim the frame and publish it as live.
        //
        // CAS(REFCOUNT_UNUSED -> 1): atomically transitions the frame from the
        // buddy free-list sentinel to a live, exclusively-owned frame.  The
        // `AcqRel` success ordering ensures steps 2 and 3 happen-before any
        // `Acquire` load of this refcount by another CPU, and also observes
        // the buddy's `Release` store of REFCOUNT_UNUSED.
        //
        // Failure means the frame's refcount was not REFCOUNT_UNUSED : either
        // the frame is still live (double-alloc) or the buddy free list is
        // corrupt (double-free).  Both are kernel bugs; panic immediately.
        meta.cas_refcount(REFCOUNT_UNUSED, 1)
            .unwrap_or_else(|actual| {
                panic!(
                    "buddy corruption: frame {:#x} refcount is {:#x} (expected REFCOUNT_UNUSED); \
                 double-free or free-list corruption",
                    phys, actual,
                )
            });

        // New live epoch: default vtable, clear guard bits, bump generation (issue #38).
        meta.note_new_allocation_epoch();

        Ok(frame)
    }
}

pub const PAGE_SIZE: u64 = 4096;
pub const FRAME_META_ALIGN: usize = 64;
pub const FRAME_META_SIZE: usize = 64;
pub const FRAME_META_LINK_NONE: u64 = u64::MAX;

/// Guard bits stored in [`MetaSlot::guard`] (issue #38 : extensible without touching page bytes).
///
/// Distinct from [`frame_flags::POISONED`] (logical frame state in `flags`).
pub mod meta_guard {
    /// No guard condition asserted.
    pub const NONE: u32 = 0;
    /// Frame must not be exposed as a userspace mapping (kernel / debug).
    pub const KERNEL_ONLY: u32 = 1 << 0;
    /// Slot marked poisoned after detected corruption (never recycle blindly).
    pub const POISONED: u32 = 1 << 31;
}

/// Persistent flags stored in [`MetaSlot`] / [`FrameMeta`].
pub mod frame_flags {
    /// The frame is allocated
    pub const ALLOCATED: u32 = 1 << 8;
    /// The frame is free.
    pub const FREE: u32 = 1 << 9;
    /// The frame is reserved for the kernel.
    pub const KERNEL: u32 = 1 << 10;
    /// The frame belongs to user space.
    pub const USER: u32 = 1 << 11;
    /// The frame is poisoned and must not be recycled as-is.
    pub const POISONED: u32 = 1 << 12;
    /// The frame belongs to a movable page class.
    pub const MOVABLE: u32 = 1 << 13;
    /// Frame eligible for copy-on-write.
    pub const COW: u32 = 1 << 0;
    /// Shared frame of type DLL, never COW.
    pub const DLL: u32 = 1 << 1;
    /// Anonymous frame.
    pub const ANONYMOUS: u32 = 1 << 2;
    /// Frame is pinned for DMA transfer : buddy must not recycle.
    pub const DMA: u32 = 1 << 14;
}

/// Buddy free-list link storage (intrusive list nodes live in [`MetaSlot`], not in frame bytes).
///
/// `AtomicU64` matches the rest of the metadata slot’s atomic story and keeps the public
/// [`MetaSlot`] API safe if list helpers are ever used without the buddy spinlock. Today
/// `buddy.rs` mutates these fields only while holding the global buddy lock, so plain
/// `Cell<u64>` would suffice for ordering; that would be a micro-optimization if profiling shows
/// hot contention here.
#[repr(C)]
pub struct FreeListLink {
    pub(crate) next: AtomicU64,
    pub(crate) prev: AtomicU64,
}

impl FreeListLink {
    pub const fn new() -> Self {
        Self {
            next: AtomicU64::new(FRAME_META_LINK_NONE),
            prev: AtomicU64::new(FRAME_META_LINK_NONE),
        }
    }
}

/// Custom vtable for frame-type-specific behavior (DMA teardown, device hooks, …).
///
/// Store a pointer as `u64` in [`MetaSlot::vtable`]; `0` selects [`DEFAULT_FRAME_META_VTABLE`].
#[repr(C)]
pub struct FrameMetaVtable {
    /// Called when the last shared reference to the frame is dropped (`refcount` → 0 path).
    ///
    /// # When it runs
    /// Invoked by [`release_owned_block`] **once** for the head frame of a block,
    /// immediately after the last ownership reference is dropped and before any
    /// per-page `on_unmap` hooks.  It does **not** run for individual page unmappings
    /// that leave the block pinned (e.g. one task unmapping while another still holds a pin).
    ///
    /// # Constraints
    /// Invoked with IRQs **disabled** and without the buddy zone lock held.
    /// MUST be: allocation-free, lock-free, and infallible.
    pub on_last_ref: Option<fn(PhysAddr)>,
    /// Called once per 4 KiB page when a mapping block is released to the allocator (unmap path).
    ///
    /// # When it runs
    /// Invoked by [`release_owned_block`] **before** the buddy allocator decides whether
    /// to recycle or quarantine the block.  It therefore runs even for poisoned frames
    /// that will be quarantined and never reused.
    ///
    /// # Constraints
    /// Invoked with IRQs **disabled** and the buddy zone lock held on the caller's CPU.
    /// MUST be:
    ///   - allocation-free (no heap, no buddy);
    ///   - lock-free (no spinlocks that might be held by the interrupted CPU);
    ///   - infallible (no panic, no unwrap).
    pub on_unmap: Option<fn(PhysAddr)>,
    /// Reserved for future hooks (keeps struct at 64 bytes for [`FRAME_META_SIZE`]).
    pub reserved: [u64; 6],
}

/// Default vtable used when [`MetaSlot::vtable`] is `0`.
pub static DEFAULT_FRAME_META_VTABLE: FrameMetaVtable = FrameMetaVtable {
    on_last_ref: None,
    on_unmap: None,
    reserved: [0; 6],
};

const _: () = assert!(mem::size_of::<FrameMetaVtable>() == FRAME_META_SIZE);

/// 64-byte cache-line metadata for one physical frame (issue #38).
///
/// Layout: free-list links + flags + refcount + optional vtable + generation + reserved tail
/// for future guard bits / generational references without touching the page payload.
///
/// Use plain `#[repr(C)]` (not `align(64)` on the struct): `align(64)` would pad the **type
/// size** to a multiple of 64 and can inflate `size_of` to 128. The metadata **array** is
/// still allocated with [`FRAME_META_ALIGN`] so each slot stays cache-line aligned.
///
/// Field order matters: `vtable` immediately follows `free_link` so `AtomicU64` stays
/// 8-byte aligned without hidden padding after `refcount` (which would inflate the struct
/// to 72 bytes).
#[repr(C)]
pub struct MetaSlot {
    pub free_link: FreeListLink,
    /// `*const FrameMetaVtable` as bits; `0` means [`DEFAULT_FRAME_META_VTABLE`].
    pub vtable: AtomicU64,
    pub flags: AtomicU32,
    pub order: AtomicU8,
    /// Padding so `refcount` stays 4-byte aligned; if `order` widens or new fields are added,
    /// re-check [`META_SLOT_REFCOUNT_BYTE_OFFSET`] / [`MetaSlot::REFCOUNT_BYTE_OFFSET`].
    _reserved0: [u8; 3],
    pub refcount: AtomicU32,
    /// Bumps each time the frame is successfully claimed from the buddy free list
    /// (see [`MetaSlot::note_new_allocation_epoch`]).
    pub generation: AtomicU32,
    /// Kernel-owned guard bits ([`meta_guard`]); independent of `frame_flags`.
    pub guard: AtomicU32,
    /// Low 16 bits: owner CPU id hint (issue #38); upper bits reserved / NUMA placeholder.
    pub meta_aux: AtomicU32,
    pub _reserved_tail: [u8; 16],
}

/// Byte offset of [`MetaSlot::refcount`] from the start of each metadata slot (layout contract).
///
/// Re-exported as [`crate::memory::META_SLOT_REFCOUNT_BYTE_OFFSET`]. Equals [`MetaSlot::REFCOUNT_BYTE_OFFSET`].
pub const META_SLOT_REFCOUNT_BYTE_OFFSET: usize = offset_of!(MetaSlot, refcount);

/// Backwards-compatible name for [`MetaSlot`].
pub type FrameMeta = MetaSlot;

impl MetaSlot {
    /// Empty metadata for boot-time array initialization.
    pub const fn new() -> Self {
        Self {
            free_link: FreeListLink::new(),
            vtable: AtomicU64::new(0),
            flags: AtomicU32::new(0),
            order: AtomicU8::new(0),
            _reserved0: [0; 3],
            refcount: AtomicU32::new(0),
            generation: AtomicU32::new(0),
            guard: AtomicU32::new(0),
            meta_aux: AtomicU32::new(0),
            _reserved_tail: [0; 16],
        }
    }

    /// Reset vtable/guard when returning a frame to the buddy free list (`buddy::set_block_meta`).
    ///
    /// Preserves [`meta_guard::POISONED`] so poisoned frames are not silently « healed » on free.
    #[inline]
    pub fn reset_with_free_list_meta(&self) {
        self.set_vtable_bits(0);
        let poison = self.get_guard() & meta_guard::POISONED;
        self.guard.store(poison, Ordering::Release);
    }

    #[inline]
    pub fn meta_aux_load(&self) -> u32 {
        self.meta_aux.load(Ordering::Relaxed)
    }

    #[inline]
    pub fn meta_aux_store(&self, v: u32) {
        // CPU-id hint : no happens-before relationship needed; Relaxed is sufficient.
        self.meta_aux.store(v, Ordering::Relaxed);
    }

    /// Byte offset of `refcount` from the start of [`MetaSlot`] (same as [`META_SLOT_REFCOUNT_BYTE_OFFSET`]).
    pub const REFCOUNT_BYTE_OFFSET: usize = META_SLOT_REFCOUNT_BYTE_OFFSET;

    /// After a successful `CAS(REFCOUNT_UNUSED → 1)` in [`FrameAllocOptions::allocate`],
    /// start a new metadata epoch: default vtable, clear guards, bump generation.
    ///
    /// The generation bump uses [`Ordering::Release`] so another CPU that later
    /// [`Acquire`]-loads [`Self::generation`] or pairs with the refcount hand-off sees this
    /// epoch for genealogical use-after-free checks. [`Ordering::Relaxed`] would be enough only
    /// if all such checks ran on the allocating CPU with no cross-CPU visibility requirement.
    #[inline]
    pub fn note_new_allocation_epoch(&self) {
        self.set_vtable_bits(0);
        self.guard.store(meta_guard::NONE, Ordering::Release);
        self.generation.fetch_add(1, Ordering::Release);
    }

    #[inline]
    pub fn get_guard(&self) -> u32 {
        self.guard.load(Ordering::Acquire)
    }

    #[inline]
    pub fn set_guard(&self, bits: u32) {
        self.guard.store(bits, Ordering::Release);
    }

    #[inline]
    pub fn fetch_or_guard(&self, bits: u32) -> u32 {
        self.guard.fetch_or(bits, Ordering::AcqRel)
    }

    /// Returns `true` if [`meta_guard::POISONED`] is set.
    #[inline]
    pub fn is_guard_poisoned(&self) -> bool {
        self.get_guard() & meta_guard::POISONED != 0
    }

    /// Marks both [`meta_guard::POISONED`] and [`frame_flags::POISONED`] (corruption / audit path).
    #[inline]
    pub fn mark_poisoned(&self) {
        self.fetch_or_guard(meta_guard::POISONED);
        self.set_flags(self.get_flags() | frame_flags::POISONED);
    }

    /// `(generation, guard_bits, vtable_bits)` for serial / shell diagnostics.
    #[inline]
    pub fn debug_snapshot(&self) -> (u32, u32, u64) {
        (self.generation(), self.get_guard(), self.vtable_bits())
    }

    /// Raw vtable pointer bits (`0` = default).
    #[inline]
    pub fn vtable_bits(&self) -> u64 {
        self.vtable.load(Ordering::Acquire)
    }

    /// Install a custom vtable pointer (must point to a `'static` [`FrameMetaVtable`]).
    #[inline]
    pub fn set_vtable_bits(&self, bits: u64) {
        self.vtable.store(bits, Ordering::Release);
    }

    /// Resolved vtable reference (`0` bits map to [`DEFAULT_FRAME_META_VTABLE`]).
    ///
    /// Misaligned or otherwise invalid non-zero pointer bits fall back to the default vtable
    /// (same as [`Self::try_vtable_ref`] returning `None`).
    pub fn vtable_ref(&self) -> &'static FrameMetaVtable {
        self.try_vtable_ref().unwrap_or(&DEFAULT_FRAME_META_VTABLE)
    }

    /// Like [`Self::vtable_ref`], but returns `None` if non-zero vtable bits are not aligned
    /// to a [`FrameMetaVtable`] pointer (8-byte aligned).
    pub fn try_vtable_ref(&self) -> Option<&'static FrameMetaVtable> {
        let bits = self.vtable_bits();
        if bits == 0 {
            return Some(&DEFAULT_FRAME_META_VTABLE);
        }
        #[cfg(debug_assertions)]
        debug_assert_eq!(
            bits & 7,
            0,
            "MetaSlot::vtable_bits must be 8-byte aligned (got {bits:#x})"
        );
        if bits & 7 != 0 {
            return None;
        }
        // `bits` is non-zero (checked above) and 8-byte aligned, so the pointer is non-null.
        let ptr = bits as *const FrameMetaVtable;
        // SAFETY: aligned, non-null; must point to a `'static` vtable when registered by the kernel.
        unsafe { Some(&*ptr) }
    }

    /// Loads the allocation generation with [`Ordering::Acquire`], pairing with the
    /// [`Ordering::Release`] bump in [`Self::note_new_allocation_epoch`] for cross-CPU checks.
    #[inline]
    pub fn generation(&self) -> u32 {
        self.generation.load(Ordering::Acquire)
    }

    /// Overwrites the generation counter : **only** for boot-time init or tests.
    ///
    /// Normal allocations bump generation via [`MetaSlot::note_new_allocation_epoch`].
    /// Arbitrary values break « generational » use-after-free checks.
    #[inline]
    pub fn set_generation(&self, g: u32) {
        self.generation.store(g, Ordering::Release);
    }

    #[inline]
    pub fn next(&self) -> u64 {
        self.free_link.next.load(Ordering::Acquire)
    }

    #[inline]
    pub fn set_next(&self, next: u64) {
        self.free_link.next.store(next, Ordering::Release);
    }

    #[inline]
    pub fn prev(&self) -> u64 {
        self.free_link.prev.load(Ordering::Acquire)
    }

    #[inline]
    pub fn set_prev(&self, prev: u64) {
        self.free_link.prev.store(prev, Ordering::Release);
    }

    #[inline]
    pub fn inc_ref(&self) {
        self.refcount.fetch_add(1, Ordering::Relaxed);
    }

    #[inline]
    pub fn dec_ref(&self) -> u32 {
        self.refcount.fetch_sub(1, Ordering::Release)
    }

    #[inline]
    pub fn get_refcount(&self) -> u32 {
        self.refcount.load(Ordering::Acquire)
    }

    #[inline]
    pub fn set_flags(&self, flags: u32) {
        self.flags.store(flags, Ordering::Release);
    }

    #[inline]
    pub fn get_flags(&self) -> u32 {
        self.flags.load(Ordering::Acquire)
    }

    /// Atomic OR: set bits without racing with concurrent readers.
    #[inline]
    pub fn or_flags(&self, bits: u32) {
        self.flags.fetch_or(bits, Ordering::AcqRel);
    }

    /// Atomic AND: clear bits without racing with concurrent readers.
    #[inline]
    pub fn and_flags(&self, bits: u32) {
        self.flags.fetch_and(bits, Ordering::AcqRel);
    }

    #[inline]
    pub fn get_order(&self) -> u8 {
        self.order.load(Ordering::Acquire)
    }

    #[inline]
    pub fn set_order(&self, order: u8) {
        self.order.store(order, Ordering::Release);
    }

    #[inline]
    pub fn set_refcount(&self, count: u32) {
        self.refcount.store(count, Ordering::Release);
    }

    #[inline]
    pub fn cas_refcount(&self, expect: u32, new: u32) -> Result<u32, u32> {
        self.refcount
            .compare_exchange(expect, new, Ordering::AcqRel, Ordering::Acquire)
    }

    #[inline]
    pub fn reset_refcount(&self) {
        self.set_refcount(0);
    }

    #[inline]
    pub fn is_cow(&self) -> bool {
        self.get_flags() & frame_flags::COW != 0
    }

    #[inline]
    pub fn is_dll(&self) -> bool {
        self.get_flags() & frame_flags::DLL != 0
    }
}

const _: () = {
    assert!(mem::size_of::<MetaSlot>() == FRAME_META_SIZE);
    // Stride is `FRAME_META_SIZE`; the backing array is allocated with `FRAME_META_ALIGN`
    // so each index maps to a cache-line-aligned slot even if `align_of::<MetaSlot>()` is 8.
    assert!(mem::align_of::<MetaSlot>() <= FRAME_META_SIZE);
    // `_reserved0` exists only to pad `order`+tail to 4 bytes before `refcount`; changing field
    // sizes or order requires updating `META_SLOT_REFCOUNT_BYTE_OFFSET` and this assert.
    assert!(META_SLOT_REFCOUNT_BYTE_OFFSET == 32);
};

/// The metadata array size for `ram_size` bytes, rounded up to the nearest page since each frame
/// has a dedicated metadata entry.

/// @param ram_size Total RAM size to be covered by the metadata (in bytes).
///
pub const fn metadata_size_for(ram_size: u64) -> u64 {
    let frames = (ram_size / PAGE_SIZE) + if ram_size % PAGE_SIZE == 0 { 0 } else { 1 };
    frames * FRAME_META_SIZE as u64
}

static METADATA_BASE_VIRT: AtomicU64 = AtomicU64::new(0);
static METADATA_FRAME_COUNT: AtomicU64 = AtomicU64::new(0);

/// Initialize the global metadata array for all physical frames.
pub fn init_metadata_array(total_ram: u64, boot_alloc: &mut BootAllocator) {
    let frame_count = (total_ram / PAGE_SIZE) + if total_ram % PAGE_SIZE == 0 { 0 } else { 1 };
    if frame_count == 0 {
        METADATA_BASE_VIRT.store(0, Ordering::Release);
        METADATA_FRAME_COUNT.store(0, Ordering::Release);
        return;
    }

    let bytes = metadata_size_for(total_ram) as usize;
    let phys = boot_alloc
        .try_alloc_accessible(bytes, FRAME_META_ALIGN)
        .unwrap_or_else(|| {
            panic!(
                "frame metadata: boot allocator could not reserve {} bytes (align {}) for {} frames : out of early boot memory",
                bytes, FRAME_META_ALIGN, frame_count
            )
        });
    let virt = crate::memory::phys_to_virt(phys.as_u64()) as *mut MetaSlot;

    for idx in 0..frame_count as usize {
        // SAFETY: le bloc a été réservé par le boot allocator avec un alignement
        // compatible `MetaSlot` et une taille suffisante pour tout le tableau.
        unsafe {
            ptr::write(virt.add(idx), MetaSlot::new());
        }
    }

    METADATA_FRAME_COUNT.store(frame_count, Ordering::Release);
    METADATA_BASE_VIRT.store(virt as u64, Ordering::Release);
}

/// Get the [`MetaSlot`] for a given physical frame (same as [`get_meta_slot`]).
#[inline]
pub fn get_meta(phys: PhysAddr) -> &'static MetaSlot {
    get_meta_slot(phys)
}

/// `(generation, guard_bits, vtable_bits)` for debugging (e.g. `serial_println!`).
#[inline]
pub fn frame_meta_debug_snapshot(phys: PhysAddr) -> (u32, u32, u64) {
    get_meta_slot(phys).debug_snapshot()
}

/// Returns `true` if [`MetaSlot::generation`] matches `expected` for `phys` (epoch check for use-after-free guards).
#[inline]
pub fn meta_generation_matches(phys: PhysAddr, expected: u32) -> bool {
    get_meta_slot(phys).generation() == expected
}

/// Returns `true` if any page in `[phys, phys + 2^order * PAGE_SIZE)` has [`MetaSlot::is_guard_poisoned`].
///
/// Returns `false` when the metadata array is not yet initialized (early-boot
/// guard: called from `free_list_push` / `alloc_from_zone` during buddy setup
/// before `init_metadata_array` has run).
pub fn block_phys_has_poison_guard(frame_phys: u64, order: u8) -> bool {
    if METADATA_BASE_VIRT.load(Ordering::Acquire) == 0 {
        return false; // metadata not yet initialized : no poison possible
    }
    let n = 1u64 << order;
    for i in 0..n {
        let p = PhysAddr::new(frame_phys + i * PAGE_SIZE);
        if get_meta_slot(p).is_guard_poisoned() {
            return true;
        }
    }
    false
}

/// Invokes `on_unmap` from the frame vtable, if any (issue #38 : unmap / release path).
pub fn invoke_vtable_on_unmap(phys: PhysAddr) {
    let m = get_meta_slot(phys);
    let Some(vt) = m.try_vtable_ref() else {
        return;
    };
    if let Some(f) = vt.on_unmap {
        f(phys);
    }
}

/// Invokes `on_last_ref` from the frame vtable, if any (last refcount drop).
pub fn invoke_vtable_on_last_ref(phys: PhysAddr) {
    let m = get_meta_slot(phys);
    let Some(vt) = m.try_vtable_ref() else {
        return;
    };
    if let Some(f) = vt.on_last_ref {
        f(phys);
    }
}

/// Preferred name matching the frame metadata design (issue #38).
pub fn get_meta_slot(phys: PhysAddr) -> &'static MetaSlot {
    let base = METADATA_BASE_VIRT.load(Ordering::Acquire);
    let frame_count = METADATA_FRAME_COUNT.load(Ordering::Acquire);
    assert!(base != 0, "frame metadata array is not initialized");

    let pfn = phys.as_u64() / PAGE_SIZE;
    assert!(pfn < frame_count, "frame metadata access out of bounds");

    let byte_offset = pfn as usize * FRAME_META_SIZE;
    // SAFETY: le tableau global couvre au moins `frame_count` entrées et reste
    // vivant pendant toute la durée du noyau.
    unsafe { &*((base as usize + byte_offset) as *const MetaSlot) }
}

/// Physical frame (4KB aligned physical memory)
#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)]
pub struct PhysFrame {
    pub start_address: PhysAddr,
}

/// Performs the phys frame containing address operation.
impl PhysFrame {
    /// Create a PhysFrame containing the given physical address
    pub fn containing_address(addr: PhysAddr) -> Self {
        PhysFrame {
            start_address: PhysAddr::new(addr.as_u64() & !0xFFF),
        }
    }

    /// Create a PhysFrame from a 4KB-aligned address
    pub fn from_start_address(addr: PhysAddr) -> Result<Self, ()> {
        if addr.is_aligned(4096u64) {
            Ok(PhysFrame {
                start_address: addr,
            })
        } else {
            Err(())
        }
    }

    /// Create an inclusive range of frames
    pub fn range_inclusive(start: PhysFrame, end: PhysFrame) -> FrameRangeInclusive {
        FrameRangeInclusive { start, end }
    }
}

/// Iterator over an inclusive range of physical frames
pub struct FrameRangeInclusive {
    pub start: PhysFrame,
    pub end: PhysFrame,
}

/// Performs the iterator operation for FrameRangeInclusive.
impl Iterator for FrameRangeInclusive {
    type Item = PhysFrame;

    /// Performs the next operation.
    fn next(&mut self) -> Option<Self::Item> {
        if self.start <= self.end {
            let frame = self.start;
            self.start.start_address += 4096u64;
            Some(frame)
        } else {
            None
        }
    }
}

/// Frame allocation errors
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum AllocError {
    /// No memory available
    OutOfMemory,
    /// Invalid order (> MAX_ORDER)
    InvalidOrder,
    /// Invalid address alignment
    InvalidAddress,
}

/// Frame allocator trait
pub trait FrameAllocator {
    /// Allocate `2^order` contiguous frames.
    ///
    /// Le token interdit les appels depuis un contexte où le verrou global de
    /// l'allocateur pourrait être ré-entré par interruption.
    fn alloc(&mut self, order: u8, token: &IrqDisabledToken) -> Result<PhysFrame, AllocError>;

    /// Free `2^order` contiguous frames starting at frame.
    fn free(&mut self, frame: PhysFrame, order: u8, token: &IrqDisabledToken);

    /// Allocate a single frame (convenience method)
    fn alloc_frame(&mut self, token: &IrqDisabledToken) -> Result<PhysFrame, AllocError> {
        self.alloc(0, token)
    }
}