Tier 2: per-CPU sched stats, NUMA-aware scheduling, init numa

- CpuStats: add context_switches and steals AtomicU64 counters,
  remove redundant per_cpu field from CpuStatsData
- context/switch.rs: increment per-CPU switches at context switch,
  increment steals at work-steal; add NUMA vruntime bonus (1/8 for
  exact-CPU match, 1/16 for same-node)
- context/mod.rs: least_loaded_cpu() now NUMA-aware, prefers same-node
  CPUs (accepts <=1 extra queued context vs cross-node best)
- scheme/sys/sched.rs: new kernel handler exposing per-CPU scheduler
  stats (switches, steals, queue_depth) via /scheme/sys/sched
- startup/mod.rs: call numa::init_default() during boot (was dead code)

src/context/mod.rs

@@ -127,8 +127,11 @@ pub fn run_contexts(token: LockToken<'_, L0>) -> MutexGuard<'_, L1, RunContextDa
 
 fn least_loaded_cpu() -> LogicalCpuId {
     let current_cpu = crate::cpu_id();
+    let topo = crate::numa::topology();
     let mut best_cpu = current_cpu;
     let mut best_depth = usize::MAX;
+    let mut best_local_cpu = current_cpu;
+    let mut best_local_depth = usize::MAX;
 
     for raw_id in 0..crate::cpu_count() {
         let cpu_id = LogicalCpuId::new(raw_id);
@@ -144,9 +147,18 @@ fn least_loaded_cpu() -> LogicalCpuId {
             best_depth = depth;
             best_cpu = cpu_id;
         }
+
+        if topo.same_node(current_cpu, cpu_id) && depth < best_local_depth {
+            best_local_depth = depth;
+            best_local_cpu = cpu_id;
+        }
     }
 
-    best_cpu
+    if best_local_depth < usize::MAX && best_local_depth <= best_depth + 1 {
+        best_local_cpu
+    } else {
+        best_cpu
+    }
 }
 
 pub fn init(token: &mut CleanLockToken) {

src/context/switch.rs

@@ -162,6 +162,7 @@ pub fn switch(token: &mut CleanLockToken) -> SwitchResult {
 
     let percpu = PercpuBlock::current();
     cpu_stats::add_context_switch();
+    percpu.stats.context_switches.fetch_add(1, Ordering::Relaxed);
 
     //set PIT Interrupt counter to 0, giving each process same amount of PIT ticks
     percpu.switch_internals.pit_ticks.set(0);
@@ -424,6 +425,9 @@ fn steal_work(
                 if let UpdateResult::CanSwitch = sw {
                     assign_context_to_cpu(&mut context_guard, cpu_id);
                     SCHED_STEAL_COUNT.fetch_add(1, Ordering::Relaxed);
+                    if let Some(thief) = get_percpu_block(cpu_id) {
+                        thief.stats.steals.fetch_add(1, Ordering::Relaxed);
+                    }
                     return Some(context_guard);
                 }
 
@@ -695,6 +699,10 @@ fn pick_next_from_queues(
                             let mut vruntime = guard.vruntime;
                             if guard.last_cpu == Some(cpu_id) {
                                 vruntime = vruntime.saturating_sub(vruntime / 8);
+                            } else if let Some(last_cpu) = guard.last_cpu {
+                                if crate::numa::topology().same_node(cpu_id, last_cpu) {
+                                    vruntime = vruntime.saturating_sub(vruntime / 16);
+                                }
                             }
                             drop(guard);
                             if vruntime < min_vruntime {
@@ -870,6 +878,10 @@ fn pick_next_from_global_queues(
                             let mut vruntime = guard.vruntime;
                             if guard.last_cpu == Some(cpu_id) {
                                 vruntime = vruntime.saturating_sub(vruntime / 8);
+                            } else if let Some(last_cpu) = guard.last_cpu {
+                                if crate::numa::topology().same_node(cpu_id, last_cpu) {
+                                    vruntime = vruntime.saturating_sub(vruntime / 16);
+                                }
                             }
                             drop(guard);
                             if vruntime < min_vruntime {

src/cpu_stats.rs

@@ -28,18 +28,14 @@ pub enum CpuState {
 /// Statistics for the CPUs.
 #[derive(Debug, Default)]
 pub struct CpuStats {
-    /// Number of ticks spent on userspace contexts
     user: AtomicU64,
-    /// Number of ticks spent on Niced userspace contexts
     nice: AtomicU64,
-    /// Number of ticks spent on kernel contexts
     kernel: AtomicU64,
-    /// Number of ticks spent idle
     idle: AtomicU64,
-    /// Number of times the CPU handled an interrupt
     irq: AtomicU64,
-    /// Current state of the CPU
     state: AtomicU8,
+    pub context_switches: AtomicU64,
+    pub steals: AtomicU64,
 }
 
 impl CpuStats {
@@ -51,21 +47,20 @@ impl CpuStats {
             idle: AtomicU64::new(0),
             irq: AtomicU64::new(0),
             state: AtomicU8::new(0),
+            context_switches: AtomicU64::new(0),
+            steals: AtomicU64::new(0),
         }
     }
 }
 
 pub struct CpuStatsData {
-    /// Number of ticks spent on userspace contexts
     pub user: u64,
-    /// Number of ticks spent on Niced userspace contexts
     pub nice: u64,
-    /// Number of ticks spent on kernel contexts
     pub kernel: u64,
-    /// Number of ticks spent idle
     pub idle: u64,
-    /// Number of times the CPU handled an interrupt
     pub irq: u64,
+    pub context_switches: u64,
+    pub steals: u64,
 }
 
 impl CpuStats {
@@ -128,6 +123,8 @@ impl From<&CpuStats> for CpuStatsData {
             kernel: val.kernel.load(Ordering::Relaxed),
             idle: val.idle.load(Ordering::Relaxed),
             irq: val.irq.load(Ordering::Relaxed),
+            context_switches: val.context_switches.load(Ordering::Relaxed),
+            steals: val.steals.load(Ordering::Relaxed),
         }
     }
 }

src/scheme/sys/mod.rs

@@ -34,6 +34,7 @@ mod iostat;
 mod irq;
 mod log;
 mod msr;
+mod sched;
 mod stat;
 mod syscall;
 mod uname;
@@ -115,6 +116,7 @@ const FILES: &[(&str, Kind)] = &[
     #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
     ("spurious_irq", Rd(interrupt::irq::spurious_irq_resource)),
     ("stat", Rd(stat::resource)),
+    ("sched", Rd(sched::resource)),
     // Disabled because the debugger is inherently unsafe and probably will break the system.
     /*
     ("trigger_debugger", Rd(|token| unsafe {

src/scheme/sys/sched.rs

@@ -0,0 +1,46 @@
+use core::fmt::Write as _;
+
+use crate::{
+    percpu::{get_all_stats, get_percpu_block},
+    sync::CleanLockToken,
+    syscall::error::Result,
+};
+use alloc::{string::String, vec::Vec};
+
+pub fn resource(_token: &mut CleanLockToken) -> Result<Vec<u8>> {
+    let stats = get_all_stats();
+    let mut out = String::new();
+    let mut total_switches: u64 = 0;
+    let mut total_steals: u64 = 0;
+
+    for (id, stat) in &stats {
+        let queue_depth = get_percpu_block(*id)
+            .map(|p| {
+                p.sched.take_lock();
+                let d = unsafe { p.sched.queues().iter().map(|q| q.len()).sum::<usize>() };
+                p.sched.release_lock();
+                d
+            })
+            .unwrap_or(0);
+
+        let _ = writeln!(
+            &mut out,
+            "cpu{} switches {} steals {} queue_depth {}",
+            id.get(),
+            stat.context_switches,
+            stat.steals,
+            queue_depth,
+        );
+
+        total_switches += stat.context_switches;
+        total_steals += stat.steals;
+    }
+
+    let _ = writeln!(
+        &mut out,
+        "total switches {} steals {}",
+        total_switches, total_steals,
+    );
+
+    Ok(out.into_bytes())
+}

src/startup/mod.rs

@@ -159,7 +159,8 @@ pub(crate) fn kmain(bootstrap: Bootstrap) -> ! {
     //Initialize the first context, stored in kernel/src/context/mod.rs
     context::init(&mut token);
 
-    //Initialize global schemes, such as `acpi:`.
+    crate::numa::init_default();
+
     scheme::init_globals();
 
     debug!("BSP: {} CPUs", crate::cpu_count());