ax-task 0.8.1

//! Thread construction data kept independent from an operating system.

use alloc::{sync::Arc, vec, vec::Vec};

use crate::{
    runtime::{
        resource::{
            AddressSpaceHandle, AddressSpaceToken, ExecutionContextHandle, StackHandle, TlsHandle,
        },
        service::{SchedulerTickCpuTime, SchedulerTickGate, SchedulerTickTaskWork},
        task_runtime,
    },
    sched::{CpuId, SchedulePolicy},
    thread::{SchedulerTickWork, TaskError, ThreadHandle, ThreadId},
};

/// Runtime-owned resources whose lifetime follows one thread.
#[repr(C)]
#[derive(Debug, Eq, PartialEq)]
pub struct ThreadResources {
    context: ExecutionContextHandle,
    stack: StackHandle,
    tls: TlsHandle,
    address_space: AddressSpaceToken,
}

impl ThreadResources {
    /// Empty resources for pure scheduler models.
    pub const NONE: Self = Self {
        context: ExecutionContextHandle::NONE,
        stack: StackHandle::NONE,
        tls: TlsHandle::NONE,
        address_space: AddressSpaceToken::NONE,
    };

    /// Creates a complete runtime resource bundle from uniquely owned handles.
    ///
    /// # Safety
    ///
    /// Every non-empty handle must be live, belong to the currently installed
    /// [`crate::runtime::TaskRuntime`], and have its unique destruction right
    /// transferred into this bundle. The caller must not construct another
    /// owning bundle from the same scalar handles.
    pub const unsafe fn new(
        context: ExecutionContextHandle,
        stack: StackHandle,
        tls: TlsHandle,
        address_space: AddressSpaceToken,
    ) -> Self {
        Self {
            context,
            stack,
            tls,
            address_space,
        }
    }

    /// Returns the execution context.
    pub const fn context(&self) -> ExecutionContextHandle {
        self.context
    }
    /// Returns the guarded stack allocation.
    pub const fn stack(&self) -> StackHandle {
        self.stack
    }
    /// Returns the TLS allocation.
    pub const fn tls(&self) -> TlsHandle {
        self.tls
    }
    /// Returns the address-space handle.
    pub const fn address_space(&self) -> AddressSpaceHandle {
        self.address_space.handle()
    }

    pub(crate) fn replace_address_space(
        &mut self,
        address_space: AddressSpaceToken,
    ) -> AddressSpaceToken {
        core::mem::replace(&mut self.address_space, address_space)
    }

    pub(crate) fn take_address_space(&mut self) -> AddressSpaceToken {
        core::mem::replace(&mut self.address_space, AddressSpaceToken::NONE)
    }

    /// Releases thread-private resources and returns the independent active-mm
    /// ownership token.
    ///
    /// The registry calls this only after switch tail has cleared physical CPU
    /// ownership. Context, TLS, and stack destruction are consequently
    /// one-way operations with no retry state. The address-space token has a
    /// separate active-CPU lifetime and is handed to that reclaim protocol
    /// instead of retaining already-dead thread resources.
    pub(crate) fn release(mut self) -> AddressSpaceToken {
        if !self.context.is_none() {
            task_runtime::destroy_context(self.context);
            self.context = ExecutionContextHandle::NONE;
        }

        if !self.tls.is_none() {
            task_runtime::deallocate_tls(self.tls);
            self.tls = TlsHandle::NONE;
        }

        if !self.stack.is_none() {
            task_runtime::deallocate_stack(self.stack);
            self.stack = StackHandle::NONE;
        }

        core::mem::replace(&mut self.address_space, AddressSpaceToken::NONE)
    }
}

/// Why a running thread relinquished its execution context.
///
/// The value crosses the OS extension callback boundary, so its numeric layout
/// is stable and may also be written directly to allocation-free trace records.
#[repr(u32)]
#[derive(Clone, Copy, Debug, Eq, PartialEq)]
pub enum SwitchReason {
    /// A scheduler request selected a more urgent or otherwise eligible thread.
    Preempted = 1,
    /// The thread voluntarily yielded its current service position.
    Yield     = 2,
    /// The thread committed a park or another blocking operation.
    Blocked   = 3,
    /// The thread terminated and will never become runnable again.
    Exited    = 4,
    /// CPU affinity or balancing moved the thread away from this CPU.
    Migrated  = 5,
}

/// CPU affinity expressed against one [`crate::runtime::TaskSystem`] topology.
#[derive(Clone, Debug, Eq, PartialEq)]
pub struct CpuSet {
    words: Vec<usize>,
    topology_len: usize,
    // Mirrors Linux task_struct::nr_cpus_allowed so scheduler class decisions
    // do not repeatedly derive affinity cardinality from the mask.
    allowed_count: usize,
}

impl CpuSet {
    const BITS_PER_WORD: usize = usize::BITS as usize;

    /// Creates a set that permits every CPU in a topology.
    pub fn all(cpu_count: usize) -> Self {
        let mut words = vec![usize::MAX; cpu_count.div_ceil(Self::BITS_PER_WORD)];
        if let Some(last) = words.last_mut()
            && !cpu_count.is_multiple_of(Self::BITS_PER_WORD)
        {
            *last = (1usize << (cpu_count % Self::BITS_PER_WORD)) - 1;
        }
        Self {
            words,
            topology_len: cpu_count,
            allowed_count: cpu_count,
        }
    }

    pub(crate) fn try_all(cpu_count: usize) -> Result<Self, super::TaskError> {
        let mut words =
            crate::thread::allocation::try_vec(cpu_count.div_ceil(Self::BITS_PER_WORD))?;
        words.resize(cpu_count.div_ceil(Self::BITS_PER_WORD), usize::MAX);
        if let Some(last) = words.last_mut()
            && !cpu_count.is_multiple_of(Self::BITS_PER_WORD)
        {
            *last = (1usize << (cpu_count % Self::BITS_PER_WORD)) - 1;
        }
        Ok(Self {
            words,
            topology_len: cpu_count,
            allowed_count: cpu_count,
        })
    }

    /// Creates an empty CPU set for a topology.
    pub fn empty(cpu_count: usize) -> Self {
        Self {
            words: vec![0; cpu_count.div_ceil(Self::BITS_PER_WORD)],
            topology_len: cpu_count,
            allowed_count: 0,
        }
    }

    /// Enables one CPU if it is represented by this set.
    pub fn insert(&mut self, cpu: CpuId) -> bool {
        let index = cpu.as_usize();
        if index >= self.topology_len {
            return false;
        }
        let mask = 1usize << (index % Self::BITS_PER_WORD);
        let word = &mut self.words[index / Self::BITS_PER_WORD];
        let changed = *word & mask == 0;
        *word |= mask;
        if changed {
            self.allowed_count += 1;
        }
        changed
    }

    /// Disables one CPU if it is represented by this set.
    pub fn remove(&mut self, cpu: CpuId) -> bool {
        let index = cpu.as_usize();
        if index >= self.topology_len {
            return false;
        }
        let mask = 1usize << (index % Self::BITS_PER_WORD);
        let word = &mut self.words[index / Self::BITS_PER_WORD];
        let changed = *word & mask != 0;
        *word &= !mask;
        if changed {
            self.allowed_count -= 1;
        }
        changed
    }

    pub(crate) fn clear(&mut self) {
        self.words.fill(0);
        self.allowed_count = 0;
    }

    /// Tests whether a CPU is allowed.
    pub fn contains(&self, cpu: CpuId) -> bool {
        let index = cpu.as_usize();
        index < self.topology_len
            && self.words[index / Self::BITS_PER_WORD] & (1usize << (index % Self::BITS_PER_WORD))
                != 0
    }

    /// Returns the number of CPUs represented by the set.
    pub fn topology_len(&self) -> usize {
        self.topology_len
    }

    /// Returns the number of CPUs selected by this set.
    pub(crate) fn count(&self) -> usize {
        self.allowed_count
    }

    /// Iterates selected CPUs in ascending logical-ID order.
    pub fn iter(&self) -> impl Iterator<Item = CpuId> + '_ {
        (0..self.topology_len)
            .map(|index| CpuId::new(index as u32))
            .filter(|cpu| self.contains(*cpu))
    }

    /// Returns the only allowed CPU when migration is impossible.
    pub(crate) fn sole_cpu(&self) -> Option<CpuId> {
        if self.allowed_count != 1 {
            return None;
        }
        let (word_index, word) = self
            .words
            .iter()
            .copied()
            .enumerate()
            .find(|(_, word)| *word != 0)?;
        let index = word_index * Self::BITS_PER_WORD + word.trailing_zeros() as usize;
        (index < self.topology_len).then_some(CpuId::new(index as u32))
    }

    /// Returns whether a runnable thread can leave its current allowed CPU.
    pub(crate) fn is_migration_capable(&self) -> bool {
        self.allowed_count > 1
    }

    /// Returns whether this set permits every CPU selected by `required`.
    pub fn covers(&self, required: &Self) -> bool {
        self.topology_len == required.topology_len
            && self
                .words
                .iter()
                .zip(&required.words)
                .all(|(allowed, is_required)| allowed & is_required == *is_required)
    }

    pub(crate) fn copy_from_set(&mut self, source: &Self) -> Result<(), TaskError> {
        if self.topology_len != source.topology_len {
            return Err(TaskError::InvalidConfiguration);
        }
        self.words.copy_from_slice(&source.words);
        self.allowed_count = source.allowed_count;
        Ok(())
    }

    /// Returns the first CPU in the intersection that satisfies `accepts`.
    ///
    /// This is the `cpumask_any_and()` primitive used by cpupri/cpudl: the
    /// intersection is formed a machine word at a time rather than scanning
    /// every logical CPU.
    pub(crate) fn first_intersection(
        &self,
        other: &Self,
        mut accepts: impl FnMut(CpuId) -> bool,
    ) -> Option<CpuId> {
        if self.topology_len != other.topology_len {
            return None;
        }
        for (word_index, (left, right)) in self.words.iter().zip(&other.words).enumerate() {
            let mut candidates = left & right;
            while candidates != 0 {
                let bit = candidates.trailing_zeros() as usize;
                candidates &= candidates - 1;
                let index = word_index * Self::BITS_PER_WORD + bit;
                if index >= self.topology_len {
                    break;
                }
                let cpu = CpuId::new(index as u32);
                if accepts(cpu) {
                    return Some(cpu);
                }
            }
        }
        None
    }

    pub(crate) fn word(&self, word_index: usize) -> usize {
        self.words.get(word_index).copied().unwrap_or(0)
    }
}

/// OS-owned callbacks attached to a thread without exposing OS types.
#[repr(C)]
#[derive(Debug)]
pub struct ThreadExtensionOps {
    /// Invoked after the incoming thread becomes current. The runtime value
    /// is the rq-charged total before its new execution interval, allowing OS
    /// accounting to use the switch boundary without querying the registry.
    pub on_switch_in: unsafe extern "Rust" fn(
        data: usize,
        thread: ThreadId,
        policy: SchedulePolicy,
        charged_runtime_ns: u64,
    ),
    /// Invoked after the thread stops being the current execution context.
    pub on_switch_out: unsafe extern "Rust" fn(data: usize, thread: ThreadId, reason: SwitchReason),
    /// Invoked in task context after the thread exits.
    pub on_exit: unsafe extern "Rust" fn(data: usize, thread: ThreadId),
    /// Invoked in task context for requested Deadline overrun notification.
    pub on_deadline_overrun: unsafe extern "Rust" fn(data: usize, thread: ThreadId),
    /// Releases the OS-owned extension data in task or reaper context.
    pub drop: unsafe extern "Rust" fn(data: usize),
}

/// Bounded OS hook invoked when the owner changes a running thread's base policy.
pub type RunningPolicyAppliedHook = unsafe extern "Rust" fn(
    data: usize,
    thread: ThreadId,
    base_policy: SchedulePolicy,
    observed_ns: u64,
);

/// Opaque OS-specific data attached to a thread.
#[derive(Debug)]
pub struct ThreadExtension {
    data: usize,
    ops: &'static ThreadExtensionOps,
    running_policy_applied_hook: Option<RunningPolicyAppliedHook>,
    scheduler_tick_cpu_time: Option<Arc<SchedulerTickCpuTime>>,
    scheduler_tick_work: Option<SchedulerTickWork>,
}

impl ThreadExtension {
    /// Creates an extension from opaque data and a static callback table.
    ///
    /// # Safety
    ///
    /// `data` must satisfy every callback contract in `ops`, and the owning OS
    /// must ensure callbacks do not allocate, block, or re-enter the scheduler
    /// when invoked as switch hooks. Task-context callbacks must return to the
    /// dedicated service thread; abandoning that stack leaves their explicit
    /// in-flight lifetime claim closed to prevent use-after-free.
    pub const unsafe fn new(data: usize, ops: &'static ThreadExtensionOps) -> Self {
        Self {
            data,
            ops,
            running_policy_applied_hook: None,
            scheduler_tick_cpu_time: None,
            scheduler_tick_work: None,
        }
    }

    /// Attaches IRQ-safe user/system CPU-time sampling to this thread.
    ///
    /// The scheduler retains the capability and charges it directly from each
    /// periodic tick. No OS callback or deferred task work runs in hard IRQ.
    pub fn with_scheduler_tick_cpu_time(mut self, accounting: Arc<SchedulerTickCpuTime>) -> Self {
        self.scheduler_tick_cpu_time = Some(accounting);
        self
    }

    /// Adds a bounded callback for base-policy changes applied to a running thread.
    ///
    /// The callback runs after the scheduler releases the thread-state lock.
    /// The current CPU still owns the scheduler baton, so the callback is
    /// serialized with switch hooks for the same thread. Queued and inactive
    /// base-policy changes are observed through the policy snapshot passed to
    /// the next switch-in instead. PI donation does not change this value.
    ///
    /// # Safety
    ///
    /// `callback` must interpret `data` according to this extension, remain
    /// valid for its complete lifetime, and perform only bounded operations.
    /// It must not allocate, block, or re-enter the scheduler.
    pub unsafe fn with_running_policy_applied_hook(
        mut self,
        callback: RunningPolicyAppliedHook,
    ) -> Self {
        self.running_policy_applied_hook = Some(callback);
        self
    }

    /// Adds task-context work gated by scheduler tick interest.
    ///
    /// The scheduler hard-IRQ path only publishes a typed deferred-work record.
    /// The callback runs later on the dedicated task-work service thread.
    ///
    /// # Safety
    ///
    /// `callback` must interpret `data` according to this extension, remain
    /// valid for its complete lifetime, and return normally to the task-work
    /// service. The callback may use task-context synchronization but must not
    /// retain the borrowed extension data after it returns. It may return
    /// [`SchedulerTickWorkDisposition::Retry`] only after a transient conflict
    /// and before publishing any accounting, timer, or signal state.
    pub unsafe fn with_scheduler_tick_work(
        mut self,
        gate: Arc<SchedulerTickGate>,
        callback: SchedulerTickTaskWork,
    ) -> Self {
        self.scheduler_tick_work = Some(SchedulerTickWork::new(gate, callback));
        self
    }

    /// Returns the opaque OS-owned value.
    pub const fn data(&self) -> usize {
        self.data
    }

    /// Returns the callback table used as the extension type identity.
    pub const fn ops(&self) -> &'static ThreadExtensionOps {
        self.ops
    }

    /// Clones the IRQ-safe CPU-time sampling capability.
    ///
    /// Thread creation retains this capability alongside the extension.
    pub fn scheduler_tick_cpu_time(&self) -> Option<Arc<SchedulerTickCpuTime>> {
        self.scheduler_tick_cpu_time.as_ref().map(Arc::clone)
    }

    pub(crate) const fn as_view(&self) -> ThreadExtensionView {
        ThreadExtensionView {
            data: self.data,
            ops: self.ops,
            running_policy_applied_hook: self.running_policy_applied_hook,
        }
    }

    pub(crate) fn scheduler_tick_work(&self) -> Option<SchedulerTickWork> {
        self.scheduler_tick_work.clone()
    }
}

impl Drop for ThreadExtension {
    fn drop(&mut self) {
        // SAFETY: construction transfers the unique callback-data destruction
        // right into this non-cloneable owner.
        unsafe { (self.ops.drop)(self.data) };
    }
}

/// Copy-only borrowed identity for an installed OS extension.
#[derive(Clone, Copy, Debug)]
pub struct ThreadExtensionView {
    data: usize,
    ops: &'static ThreadExtensionOps,
    running_policy_applied_hook: Option<RunningPolicyAppliedHook>,
}

/// Extension identity borrowed for exactly as long as a strong thread handle.
///
/// This wrapper deliberately does not expose its copyable internal view. The
/// strong handle borrowed by the wrapper prevents the registry reaper from
/// destroying the extension while its opaque data is being inspected.
#[derive(Debug)]
pub struct ThreadExtensionBorrow<'thread> {
    view: ThreadExtensionView,
    _thread: &'thread ThreadHandle,
}

impl<'thread> ThreadExtensionBorrow<'thread> {
    pub(crate) const fn new(view: ThreadExtensionView, thread: &'thread ThreadHandle) -> Self {
        Self {
            view,
            _thread: thread,
        }
    }

    /// Returns the borrowed opaque data value.
    pub const fn data(&self) -> usize {
        self.view.data()
    }

    /// Returns the callback table used as the extension type identity.
    pub const fn ops(&self) -> &'static ThreadExtensionOps {
        self.view.ops()
    }
}

/// Owned extension lease used when the caller has no pre-existing handle.
///
/// Keeping this value alive pins both the thread header and the registry record,
/// so current-thread helpers cannot return data that becomes stale immediately
/// after their temporary lookup handle is dropped.
#[derive(Debug)]
pub struct ThreadExtensionLease {
    view: ThreadExtensionView,
    thread: ThreadHandle,
}

impl ThreadExtensionLease {
    pub(crate) const fn new(view: ThreadExtensionView, thread: ThreadHandle) -> Self {
        Self { view, thread }
    }

    /// Returns the generation-bearing identity pinned by this lease.
    pub fn thread_id(&self) -> ThreadId {
        self.thread.id()
    }

    /// Returns the leased opaque data value.
    pub const fn data(&self) -> usize {
        self.view.data()
    }

    /// Returns the callback table used as the extension type identity.
    pub const fn ops(&self) -> &'static ThreadExtensionOps {
        self.view.ops()
    }
}

impl ThreadExtensionView {
    /// Returns the borrowed opaque data value.
    pub const fn data(self) -> usize {
        self.data
    }

    /// Returns the callback table used as the extension type identity.
    pub const fn ops(self) -> &'static ThreadExtensionOps {
        self.ops
    }

    pub(crate) unsafe fn notify_running_policy_applied(
        self,
        thread: ThreadId,
        base_policy: SchedulePolicy,
        observed_ns: u64,
    ) {
        if let Some(callback) = self.running_policy_applied_hook {
            unsafe { callback(self.data, thread, base_policy, observed_ns) };
        }
    }
}

/// Validated inputs used to create a scheduler thread record.
#[derive(Debug)]
pub struct ThreadSpec {
    pub(crate) execution: Option<Arc<crate::thread::execution::ThreadExecution>>,
    policy: SchedulePolicy,
    affinity: Option<CpuSet>,
    // Runtime resources must be dropped before the extension that owns their
    // address-space and entry metadata, including on fallback destruction.
    resources: ThreadResources,
    extension: Option<ThreadExtension>,
}

impl ThreadSpec {
    /// Creates a thread specification with full topology affinity.
    pub const fn new(policy: SchedulePolicy) -> Self {
        Self {
            execution: None,
            policy,
            affinity: None,
            resources: ThreadResources::NONE,
            extension: None,
        }
    }

    /// Restricts the thread to an explicit CPU set.
    pub fn with_affinity(mut self, affinity: CpuSet) -> Self {
        self.affinity = Some(affinity);
        self
    }

    /// Attaches OS-specific state.
    pub fn with_extension(mut self, extension: ThreadExtension) -> Self {
        self.extension = Some(extension);
        self
    }

    /// Associates a complete runtime resource bundle with the thread.
    ///
    /// # Safety
    ///
    /// `resources` must satisfy [`ThreadResources::new`] and must be consumed by
    /// exactly this specification and its eventual scheduler record.
    pub unsafe fn with_resources(mut self, resources: ThreadResources) -> Self {
        self.resources = resources;
        self
    }

    /// Returns the base scheduling policy.
    pub const fn policy(&self) -> SchedulePolicy {
        self.policy
    }

    /// Returns explicit affinity, if one was supplied.
    pub fn affinity(&self) -> Option<&CpuSet> {
        self.affinity.as_ref()
    }

    pub(crate) fn take_affinity(&mut self) -> Option<CpuSet> {
        self.affinity.take()
    }
    pub(crate) fn resources(&self) -> &ThreadResources {
        &self.resources
    }
    pub(crate) fn extension(&self) -> Option<&ThreadExtension> {
        self.extension.as_ref()
    }

    pub(crate) fn into_owned_parts(mut self) -> (Option<ThreadExtension>, ThreadResources) {
        let extension = self.extension.take();
        let resources = core::mem::replace(&mut self.resources, ThreadResources::NONE);
        (extension, resources)
    }
}