From 8bf0a72bce91081f0f6e7c29ff5584e834487657 Mon Sep 17 00:00:00 2001
From: Ashwin Naren <arihant2math@gmail.com>
Date: Sun, 21 Dec 2025 17:02:19 -0800
Subject: [PATCH] fix futex based thread joining

---
 src/process/clone.rs     | 18 ++++++++++++++++--
 src/process/exit.rs      | 12 ++++++++++++
 src/process/mod.rs       |  3 +++
 src/process/threading.rs | 35 ++++++++++++++++++++---------------
 4 files changed, 51 insertions(+), 17 deletions(-)

diff --git a/src/process/clone.rs b/src/process/clone.rs
index 4d5abc4..681ab56 100644
--- a/src/process/clone.rs
+++ b/src/process/clone.rs
@@ -1,3 +1,4 @@
+use crate::memory::uaccess::copy_to_user;
 use crate::{
     process::{TASK_LIST, Task, TaskState},
     sched::{self, current_task},
@@ -44,8 +45,8 @@ bitflags! {
 pub async fn sys_clone(
     flags: u32,
     newsp: UA,
-    _arent_tidptr: UA,
-    _child_tidptr: UA,
+    parent_tidptr: UA,
+    child_tidptr: UA,
     tls: usize,
 ) -> Result<usize> {
     let flags = CloneFlags::from_bits_truncate(flags);
@@ -146,6 +147,11 @@ pub async fn sys_clone(
             state: Arc::new(SpinLock::new(TaskState::Runnable)),
             last_run: SpinLock::new(None),
             robust_list: SpinLock::new(None),
+            child_tid_ptr: SpinLock::new(if !child_tidptr.is_null() {
+                Some(child_tidptr.cast::<u32>())
+            } else {
+                None
+            }),
         }
     };
 
@@ -157,5 +163,13 @@ pub async fn sys_clone(
 
     sched::insert_task(Arc::new(new_task));
 
+    // Honour CLONE_*SETTID semantics for the parent and (shared-VM) child.
+    if flags.contains(CloneFlags::CLONE_PARENT_SETTID) && !parent_tidptr.is_null() {
+        copy_to_user(parent_tidptr.cast::<u32>(), tid.value() as u32).await?;
+    }
+    if flags.contains(CloneFlags::CLONE_CHILD_SETTID) && !child_tidptr.is_null() {
+        copy_to_user(child_tidptr.cast::<u32>(), tid.value() as u32).await?;
+    }
+
     Ok(tid.value() as _)
 }
diff --git a/src/process/exit.rs b/src/process/exit.rs
index b586313..373181c 100644
--- a/src/process/exit.rs
+++ b/src/process/exit.rs
@@ -1,3 +1,4 @@
+use crate::process::threading::futex_wake_addr;
 use crate::sched::current_task;
 use alloc::vec::Vec;
 use libkernel::error::Result;
@@ -101,6 +102,17 @@ pub fn sys_exit_group(exit_code: usize) -> Result<usize> {
 
 pub fn sys_exit(exit_code: usize) -> Result<usize> {
     let task = current_task();
+
+    // Honour CLONE_CHILD_CLEARTID: clear the user TID word and futex-wake any waiters.
+    if let Some(ptr) = task.child_tid_ptr.lock_save_irq().take() {
+        unsafe {
+            (ptr.value() as *mut u32).write_volatile(0);
+        }
+
+        // Wake any thread waiting on this futex.
+        futex_wake_addr(ptr, 1);
+    }
+
     let process = Arc::clone(&task.process);
     let mut thread_lock = process.threads.lock_save_irq();
 
diff --git a/src/process/mod.rs b/src/process/mod.rs
index ac928dd..f7a5255 100644
--- a/src/process/mod.rs
+++ b/src/process/mod.rs
@@ -181,6 +181,7 @@ pub struct Task {
     pub last_run: SpinLock<Option<Instant>>,
     pub state: Arc<SpinLock<TaskState>>,
     pub robust_list: SpinLock<Option<TUA<RobustListHead>>>,
+    pub child_tid_ptr: SpinLock<Option<TUA<u32>>>,
 }
 
 impl Task {
@@ -214,6 +215,7 @@ impl Task {
             fd_table: Arc::new(SpinLock::new(FileDescriptorTable::new())),
             last_run: SpinLock::new(None),
             robust_list: SpinLock::new(None),
+            child_tid_ptr: SpinLock::new(None),
         }
     }
 
@@ -241,6 +243,7 @@ impl Task {
             )),
             last_run: SpinLock::new(None),
             robust_list: SpinLock::new(None),
+            child_tid_ptr: SpinLock::new(None),
         }
     }
 
diff --git a/src/process/threading.rs b/src/process/threading.rs
index fdc881b..3dea397 100644
--- a/src/process/threading.rs
+++ b/src/process/threading.rs
@@ -74,6 +74,25 @@ const FUTEX_WAIT_BITSET: i32 = 9;
 const FUTEX_WAKE_BITSET: i32 = 10;
 const FUTEX_PRIVATE_FLAG: i32 = 128;
 
+/// Wake up to `nr` waiters sleeping on the futex word located at `uaddr`.
+/// Returns the number of tasks actually woken.
+pub fn futex_wake_addr(uaddr: TUA<u32>, nr: usize) -> usize {
+    let mut woke = 0;
+
+    if let Some(table) = FUTEX_TABLE.get()
+        && let Some(waitq_arc) = table.lock_save_irq().get(&uaddr).cloned()
+    {
+        let mut waitq = waitq_arc.lock_save_irq();
+        for _ in 0..nr {
+            waitq.wakeups = waitq.wakeups.saturating_add(1);
+            waitq.wakers.wake_one();
+            woke += 1;
+        }
+    }
+
+    woke
+}
+
 pub async fn sys_futex(
     uaddr: TUA<u32>,
     op: i32,
@@ -132,21 +151,7 @@ pub async fn sys_futex(
 
         FUTEX_WAKE | FUTEX_WAKE_BITSET => {
             let nr_wake = val as usize;
-            let mut woke = 0;
-
-            if let Some(table) = FUTEX_TABLE.get()
-                && let Some(waitq_arc) = table.lock_save_irq().get(&uaddr).cloned()
-            {
-                let mut waitq = waitq_arc.lock_save_irq();
-                for _ in 0..nr_wake {
-                    // Record a pending wake-up and attempt to wake a single waiter.
-                    waitq.wakeups = waitq.wakeups.saturating_add(1);
-                    waitq.wakers.wake_one();
-                    woke += 1;
-                }
-            }
-
-            Ok(woke)
+            Ok(futex_wake_addr(uaddr, nr_wake))
         }
 
         _ => Err(KernelError::NotSupported),