feat: P0-P6 kernel scheduler + relibc threading comprehensive implementation

P0-P2: Barrier SMP, sigmask/pthread_kill races, robust mutexes, RT scheduling, POSIX sched API P3: PerCpuSched struct, per-CPU wiring, work stealing, load balancing, initial placement P4: 64-shard futex table, REQUEUE, PI futexes (LOCK_PI/UNLOCK_PI/TRYLOCK_PI), robust futexes, vruntime tracking, min-vruntime SCHED_OTHER selection P5: setpriority/getpriority, pthread_setaffinity_np, pthread_setname_np, pthread_setschedparam (Redox) P6: Cache-affine scheduling (last_cpu + vruntime bonus), NUMA topology kernel hints + numad userspace daemon Stability fixes: make_consistent stores 0 (dead TID fix), cond.rs error propagation, SPIN_COUNT adaptive spinning, Sys::open &str fix, PI futex CAS race, proc.rs lock ordering, barrier destroy Patches: 33 kernel + 58 relibc patches, all tracked in recipes Docs: KERNEL-SCHEDULER-MULTITHREAD-IMPROVEMENT-PLAN.md updated, SCHEDULER-REVIEW-FINAL.md created Architecture: NUMA topology parsing stays userspace (numad daemon), kernel stores lightweight NumaTopology hints
2026-04-30 18:21:48 +01:00
parent 55d00c3a24
commit 34360e1e4f
70 changed files with 15268 additions and 10 deletions
@@ -0,0 +1,70 @@
+diff --git a/src/scheme/proc.rs b/src/scheme/proc.rs
+--- a/src/scheme/proc.rs
+++ b/src/scheme/proc.rs
+@@ -145,8 +145,9 @@ enum ContextHandle {
+     // TODO: Remove this once openat is implemented, or allow openat-via-dup via e.g. the top-level
+     // directory.
+     OpenViaDup,
+    Priority,
+     SchedAffinity,
+     SchedPolicy,
+     Name,
+ 
+     MmapMinAddr(Arc<AddrSpaceWrapper>),
+@@ -160,6 +161,17 @@ pub struct ProcScheme;
+ static NEXT_ID: AtomicUsize = AtomicUsize::new(1);
+ static HANDLES: RwLock<L1, HashMap<usize, Handle>> =
+     RwLock::new(HashMap::with_hasher(DefaultHashBuilder::new()));
+
+const NICE_MIN: i32 = -20;
+const NICE_MAX: i32 = 19;
+
+fn nice_to_kernel_prio(nice: i32) -> usize {
+    (nice.saturating_add(20)).clamp(0, 39) as usize
+}
+
+fn kernel_prio_to_nice(prio: usize) -> i32 {
+    (prio.min(39) as i32) - 20
+}
+ 
+ #[cfg(feature = "debugger")]
+ #[allow(dead_code)]
+ pub fn foreach_addrsp(
+@@ -253,6 +265,7 @@ impl ProcScheme {
+             "sighandler" => (ContextHandle::Sighandler, false),
+             "start" => (ContextHandle::Start, false),
+             "open_via_dup" => (ContextHandle::OpenViaDup, false),
+            "priority" => (ContextHandle::Priority, false),
+             "mmap-min-addr" => (
+                 ContextHandle::MmapMinAddr(Arc::clone(
+                     context
+@@ -1191,6 +1204,17 @@ impl ContextHandle {
+ 
+                 Ok(size_of_val(&mask))
+             }
+            Self::Priority => {
+                let nice = unsafe { buf.read_exact::<i32>()? };
+                if !(NICE_MIN..=NICE_MAX).contains(&nice) {
+                    return Err(Error::new(EINVAL));
+                }
+
+                context
+                    .write(token.token())
+                    .set_sched_other_prio(nice_to_kernel_prio(nice));
+
+                Ok(size_of::<i32>())
+            }
+             Self::SchedPolicy => {
+                 if buf.len() != 2 {
+                     return Err(Error::new(EINVAL));
+@@ -1522,6 +1546,10 @@ impl ContextHandle {
+ 
+                 buf.copy_exactly(crate::cpu_set::mask_as_bytes(&mask))?;
+                 Ok(size_of_val(&mask))
+            }
+            ContextHandle::Priority => {
+                let nice = kernel_prio_to_nice(context.read(token.token()).prio);
+                buf.copy_common_bytes_from_slice(&nice.to_ne_bytes())
+             }
+             ContextHandle::SchedPolicy => {
+                 let context = context.read(token.token());