RedBear-OS/local/patches/kernel/P13-priority-inheritance.patch

--- a/src/sync/mcs.rs
+++ b/src/sync/mcs.rs
@@ -4,7 +4,7 @@
 //! word, eliminating cache-line bouncing under contention. FIFO ordering
 //! guarantees fairness. O(1) cache-line transfers on unlock.
 
-use core::sync::atomic::{AtomicBool, AtomicPtr, Ordering};
+use core::sync::atomic::{AtomicBool, AtomicPtr, AtomicU32, Ordering};
 use core::{hint, ptr};
 
 use crate::percpu::PercpuBlock;
@@ -27,12 +27,16 @@
 /// Raw MCS spinlock primitive.
 pub struct McsRawLock {
     tail: AtomicPtr<McsNode>,
+    /// CPU ID of the current lock holder (for priority inheritance).
+    /// `u32::MAX` means no holder.
+    holder_cpu: AtomicU32,
 }
 
 impl McsRawLock {
     pub const fn new() -> Self {
         Self {
             tail: AtomicPtr::new(ptr::null_mut()),
+            holder_cpu: AtomicU32::new(u32::MAX),
         }
     }
 
@@ -42,21 +46,37 @@
         node.locked.store(true, Ordering::Relaxed);
         let prev = self.tail.swap((node as *const McsNode).cast_mut(), Ordering::AcqRel);
         if prev.is_null() {
+            // Uncontended — record ourselves as holder
+            let cpu_id = PercpuBlock::current().cpu_id.get();
+            self.holder_cpu.store(cpu_id, Ordering::Release);
             return false;
         }
         unsafe {
             (*prev).next.store((node as *const McsNode).cast_mut(), Ordering::Release);
         }
         let percpu = PercpuBlock::current();
+        let mut donated = false;
         while node.locked.load(Ordering::Acquire) {
             percpu.maybe_handle_tlb_shootdown();
+            // Donate priority to the lock holder once per acquisition
+            if !donated {
+                self.maybe_donate_priority(percpu);
+                donated = true;
+            }
             hint::spin_loop();
         }
+        // We now hold the lock
+        self.holder_cpu.store(percpu.cpu_id.get(), Ordering::Release);
         true
     }
 
     #[inline]
     pub fn release(&self, node: &McsNode) {
+        // Clear priority inheritance donation — we no longer hold the lock
+        PercpuBlock::current().pi_donated_prio.store(u32::MAX, Ordering::Relaxed);
+        // Clear holder CPU
+        self.holder_cpu.store(u32::MAX, Ordering::Release);
+
         let next = node.next.load(Ordering::Acquire);
         if next.is_null() {
             if self
@@ -84,13 +104,43 @@
     pub fn try_acquire(&self, node: &McsNode) -> bool {
         node.next.store(ptr::null_mut(), Ordering::Relaxed);
         node.locked.store(true, Ordering::Relaxed);
-        self.tail
+        let ok = self
+            .tail
             .compare_exchange(
                 ptr::null_mut(),
                 (node as *const McsNode).cast_mut(),
                 Ordering::AcqRel,
                 Ordering::Acquire,
             )
-            .is_ok()
+            .is_ok();
+        if ok {
+            let cpu_id = PercpuBlock::current().cpu_id.get();
+            self.holder_cpu.store(cpu_id, Ordering::Release);
+        }
+        ok
+    }
+
+    /// Donate current CPU's context priority to the lock holder's CPU.
+    /// Reads priority from PercpuBlock::current_prio (cached by the scheduler)
+    /// to avoid acquiring any lock in the MCS spin loop.
+    fn maybe_donate_priority(&self, my_percpu: &PercpuBlock) {
+        let holder_cpu_id = self.holder_cpu.load(Ordering::Relaxed);
+        if holder_cpu_id == u32::MAX {
+            return;
+        }
+        // Read our own priority from the per-CPU cache (set by scheduler,
+        // no lock required).
+        let my_prio = my_percpu.current_prio.get();
+        // Look up holder's PercpuBlock
+        let holder_percpu = crate::percpu::get_for_cpu(
+            crate::cpu_set::LogicalCpuId::new(holder_cpu_id),
+        );
+        if let Some(holder) = holder_percpu {
+            let current_donated = holder.pi_donated_prio.load(Ordering::Relaxed);
+            // Donate if our priority is higher (lower number)
+            if (my_prio as u32) < current_donated {
+                holder.pi_donated_prio.store(my_prio as u32, Ordering::Relaxed);
+            }
+        }
     }
 }
--- a/src/context/switch.rs
+++ b/src/context/switch.rs
@@ -16,6 +16,7 @@
 use core::{
     cell::{Cell, RefCell},
     mem,
+    sync::atomic::Ordering,
 };
 use syscall::PtraceFlags;
 
@@ -442,6 +443,8 @@
             // Is this context runnable on this CPU?
             let sw = unsafe { update_runnable(&mut next_context_guard, cpu_id, switch_time) };
             if let UpdateResult::CanSwitch = sw {
+                // Cache the new context's priority for MCS lock priority donation.
+                percpu.current_prio.set(next_context_guard.prio);
                 next_context_guard_opt = Some(next_context_guard);
                 balance[i] -= SCHED_PRIO_TO_WEIGHT[20];
                 break 'priority;
@@ -466,7 +469,10 @@
         // Send the old process to the back of the line (if it is still runnable)
         let prev_ctx = WeakContextRef(Arc::downgrade(&prev_context_lock));
         if prev_context_guard.status.is_runnable() {
-            let prio = prev_context_guard.prio;
+            let raw_prio = prev_context_guard.prio;
+            let prio = percpu.effective_prio(raw_prio);
+            // Clear PI donation — previous context is being re-queued
+            percpu.pi_donated_prio.store(u32::MAX, Ordering::Relaxed);
             contexts_list[prio].push_back(prev_ctx);
         } else {
             idle_contexts(token.token()).push_back(prev_ctx);
@@ -478,7 +484,8 @@
         return Ok(Some(next_context_guard));
     } else {
         if !was_idle && !Arc::ptr_eq(&prev_context_lock, &idle_context) {
-            // We switch into the idle context
+            // Switching to idle context — cache lowest priority
+            percpu.current_prio.set(39);
             Ok(Some(unsafe { idle_context.write_arc() }))
         } else {
             // We found no other process to run.