Fix LOS object enumeration

src/policy/compressor/compressorspace.rs

@@ -405,7 +405,7 @@ impl<VM: VMBinding> CompressorSpace<VM> {
     pub fn after_compact(&self, worker: &mut GCWorker<VM>, los: &LargeObjectSpace<VM>) {
         self.pr.reset_allocator();
         // Update references from the LOS to Compressor too.
-        los.enumerate_objects(&mut object_enum::ClosureObjectEnumerator::<_, VM>::new(
+        los.enumerate_to_space_objects(&mut object_enum::ClosureObjectEnumerator::<_, VM>::new(
             &mut |o: ObjectReference| {
                 self.update_references(worker, o);
             },

src/policy/largeobjectspace.rs

@@ -219,21 +219,39 @@ impl<VM: VMBinding> Space<VM> for LargeObjectSpace<VM> {
     }
 
     fn enumerate_objects(&self, enumerator: &mut dyn ObjectEnumerator) {
-        self.treadmill.enumerate_objects(enumerator);
+        // `MMTK::enumerate_objects` is not allowed during GC, so the collection nursery and the
+        // from space must be empty.  In `ConcurrentImmix`, mutators may run during GC and call
+        // `MMTK::enumerate_objects`.  It has undefined behavior according to the current API, so
+        // the assertion failure is expected.
+        assert!(
+            self.treadmill.is_collect_nursery_empty(),
+            "Collection nursery is not empty"
+        );
+        assert!(
+            self.treadmill.is_from_space_empty(),
+            "From-space is not empty"
+        );
+
+        // Visit objects in the allocation nursery and the to-space, which contain young and old
+        // objects, respectively, during mutator time.
+        self.treadmill.enumerate_objects(enumerator, false);
     }
 
     fn clear_side_log_bits(&self) {
-        let mut enumator = ClosureObjectEnumerator::<_, VM>::new(|object| {
+        let mut enumerator = ClosureObjectEnumerator::<_, VM>::new(|object| {
             VM::VMObjectModel::GLOBAL_LOG_BIT_SPEC.clear::<VM>(object, Ordering::SeqCst);
         });
-        self.treadmill.enumerate_objects(&mut enumator);
+        // Visit all objects.  It can be ordered arbitrarily with `Self::Release` which sweeps dead
+        // objects (removing them from the treadmill) and clears their unlog bits, too.
+        self.treadmill.enumerate_objects(&mut enumerator, true);
     }
 
     fn set_side_log_bits(&self) {
-        let mut enumator = ClosureObjectEnumerator::<_, VM>::new(|object| {
+        let mut enumerator = ClosureObjectEnumerator::<_, VM>::new(|object| {
             VM::VMObjectModel::GLOBAL_LOG_BIT_SPEC.mark_as_unlogged::<VM>(object, Ordering::SeqCst);
         });
-        self.treadmill.enumerate_objects(&mut enumator);
+        // Visit all objects.
+        self.treadmill.enumerate_objects(&mut enumerator, true);
     }
 }
 
@@ -297,10 +315,16 @@ impl<VM: VMBinding> LargeObjectSpace<VM> {
     }
 
     pub fn release(&mut self, full_heap: bool) {
+        // We swapped the allocation nursery and the collection nursery when GC starts, and we don't
+        // add objects to the allocation nursery during GC.  It should have remained empty during
+        // the whole GC.
+        debug_assert!(self.treadmill.is_alloc_nursery_empty());
+
         self.sweep_large_pages(true);
-        debug_assert!(self.treadmill.is_nursery_empty());
+        debug_assert!(self.treadmill.is_collect_nursery_empty());
         if full_heap {
             self.sweep_large_pages(false);
+            debug_assert!(self.treadmill.is_from_space_empty());
         }
     }
 
@@ -364,12 +388,22 @@ impl<VM: VMBinding> LargeObjectSpace<VM> {
                 sweep(object);
             }
         } else {
-            for object in self.treadmill.collect() {
+            for object in self.treadmill.collect_mature() {
                 sweep(object)
             }
         }
     }
 
+    /// Enumerate objects in the to-space.  It is a workaround for Compressor which currently needs
+    /// to enumerate reachable objects for during reference forwarding.
+    pub(crate) fn enumerate_to_space_objects(&self, enumerator: &mut dyn ObjectEnumerator) {
+        // This function is intended to enumerate objects in the to-space.
+        // The alloc nursery should have remained empty during the GC.
+        debug_assert!(self.treadmill.is_alloc_nursery_empty());
+        // We only need to visit the to_space, which contains all objects determined to be live.
+        self.treadmill.enumerate_objects(enumerator, false);
+    }
+
     /// Allocate an object
     pub fn allocate_pages(
         &self,

src/util/treadmill.rs

@@ -6,111 +6,165 @@ use crate::util::ObjectReference;
 
 use super::object_enum::ObjectEnumerator;
 
+/// A data structure for recording objects in the LOS.
+///
+/// All operations are protected by a single mutex [`TreadMill::sync`].
 pub struct TreadMill {
-    from_space: Mutex<HashSet<ObjectReference>>,
-    to_space: Mutex<HashSet<ObjectReference>>,
-    collect_nursery: Mutex<HashSet<ObjectReference>>,
-    alloc_nursery: Mutex<HashSet<ObjectReference>>,
+    sync: Mutex<TreadMillSync>,
+}
+
+/// The synchronized part of [`TreadMill`]
+#[derive(Default)]
+struct TreadMillSync {
+    /// The from-space.  During GC, it contains old objects with unknown liveness.
+    from_space: HashSet<ObjectReference>,
+    /// The to-space.  During mutator time, it contains old objects; during GC, it contains objects
+    /// determined to be live.
+    to_space: HashSet<ObjectReference>,
+    /// The collection nursery.  During GC, it contains young objects with unknown liveness.
+    collect_nursery: HashSet<ObjectReference>,
+    /// The allocation nursery.  During mutator time, it contains young objects; during GC, it
+    /// remains empty.
+    alloc_nursery: HashSet<ObjectReference>,
 }
 
 impl std::fmt::Debug for TreadMill {
     fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        let sync = self.sync.lock().unwrap();
         f.debug_struct("TreadMill")
-            .field("from", &self.from_space.lock().unwrap())
-            .field("to", &self.to_space.lock().unwrap())
-            .field("collect_nursery", &self.collect_nursery.lock().unwrap())
-            .field("alloc_nursery", &self.alloc_nursery.lock().unwrap())
+            .field("from_space", &sync.from_space)
+            .field("to_space", &sync.to_space)
+            .field("collect_nursery", &sync.collect_nursery)
+            .field("alloc_nursery", &sync.alloc_nursery)
             .finish()
     }
 }
 
 impl TreadMill {
     pub fn new() -> Self {
         TreadMill {
-            from_space: Mutex::new(HashSet::new()),
-            to_space: Mutex::new(HashSet::new()),
-            collect_nursery: Mutex::new(HashSet::new()),
-            alloc_nursery: Mutex::new(HashSet::new()),
+            sync: Mutex::new(Default::default()),
         }
     }
 
+    /// Add an object to the treadmill.
+    ///
+    /// New objects are normally added to `alloc_nursery`.  But when allocating as live (e.g. when
+    /// concurrent marking is active), we directly add into the `to_space`.
     pub fn add_to_treadmill(&self, object: ObjectReference, nursery: bool) {
+        let mut sync = self.sync.lock().unwrap();
         if nursery {
-            trace!("Adding {} to nursery", object);
-            self.alloc_nursery.lock().unwrap().insert(object);
+            trace!("Adding {} to alloc_nursery", object);
+            sync.alloc_nursery.insert(object);
         } else {
             trace!("Adding {} to to_space", object);
-            self.to_space.lock().unwrap().insert(object);
+            sync.to_space.insert(object);
         }
     }
 
-    pub fn collect_nursery(&self) -> Vec<ObjectReference> {
-        let mut guard = self.collect_nursery.lock().unwrap();
-        let vals = guard.iter().copied().collect();
-        guard.clear();
-        drop(guard);
-        vals
+    /// Take all objects from the `collect_nursery`.  This is called during sweeping at which time
+    /// all unreachable young objects are in the collection nursery.
+    pub fn collect_nursery(&self) -> impl IntoIterator<Item = ObjectReference> {
+        let mut sync = self.sync.lock().unwrap();
+        std::mem::take(&mut sync.collect_nursery)
     }
 
-    pub fn collect(&self) -> Vec<ObjectReference> {
-        let mut guard = self.from_space.lock().unwrap();
-        let vals = guard.iter().copied().collect();
-        guard.clear();
-        drop(guard);
-        vals
+    /// Take all objects from the `from_space`.  This is called during sweeping at which time all
+    /// unreachable old objects are in the from-space.
+    pub fn collect_mature(&self) -> impl IntoIterator<Item = ObjectReference> {
+        let mut sync = self.sync.lock().unwrap();
+        std::mem::take(&mut sync.from_space)
     }
 
+    /// Move an object to `to_space`.  Called when an object is determined to be reachable.
     pub fn copy(&self, object: ObjectReference, is_in_nursery: bool) {
+        let mut sync = self.sync.lock().unwrap();
         if is_in_nursery {
-            let mut guard = self.collect_nursery.lock().unwrap();
             debug_assert!(
-                guard.contains(&object),
+                sync.collect_nursery.contains(&object),
                 "copy source object ({}) must be in collect_nursery",
                 object
             );
-            guard.remove(&object);
+            sync.collect_nursery.remove(&object);
         } else {
-            let mut guard = self.from_space.lock().unwrap();
             debug_assert!(
-                guard.contains(&object),
+                sync.from_space.contains(&object),
                 "copy source object ({}) must be in from_space",
                 object
             );
-            guard.remove(&object);
+            sync.from_space.remove(&object);
         }
-        self.to_space.lock().unwrap().insert(object);
+        sync.to_space.insert(object);
     }
 
+    /// Return true if the to-space is empty.
     pub fn is_to_space_empty(&self) -> bool {
-        self.to_space.lock().unwrap().is_empty()
+        let sync = self.sync.lock().unwrap();
+        sync.to_space.is_empty()
     }
 
+    /// Return true if the from-space is empty.
     pub fn is_from_space_empty(&self) -> bool {
-        self.from_space.lock().unwrap().is_empty()
+        let sync = self.sync.lock().unwrap();
+        sync.from_space.is_empty()
+    }
+
+    /// Return true if the allocation nursery is empty.
+    pub fn is_alloc_nursery_empty(&self) -> bool {
+        let sync = self.sync.lock().unwrap();
+        sync.alloc_nursery.is_empty()
     }
 
-    pub fn is_nursery_empty(&self) -> bool {
-        self.collect_nursery.lock().unwrap().is_empty()
+    /// Return true if the collection nursery is empty.
+    pub fn is_collect_nursery_empty(&self) -> bool {
+        let sync = self.sync.lock().unwrap();
+        sync.collect_nursery.is_empty()
     }
 
+    /// Flip object sets.
+    ///
+    /// It will flip the allocation nursery and the collection nursery.
+    ///
+    /// If `full_heap` is true, it will also flip the from-space and the to-space.
     pub fn flip(&mut self, full_heap: bool) {
-        swap(&mut self.alloc_nursery, &mut self.collect_nursery);
+        let sync = self.sync.get_mut().unwrap();
+        swap(&mut sync.alloc_nursery, &mut sync.collect_nursery);
         trace!("Flipped alloc_nursery and collect_nursery");
         if full_heap {
-            swap(&mut self.from_space, &mut self.to_space);
+            swap(&mut sync.from_space, &mut sync.to_space);
             trace!("Flipped from_space and to_space");
         }
     }
 
-    pub(crate) fn enumerate_objects(&self, enumerator: &mut dyn ObjectEnumerator) {
-        let mut visit_objects = |set: &Mutex<HashSet<ObjectReference>>| {
-            let set = set.lock().unwrap();
+    /// Enumerate objects.
+    ///
+    /// Objects in the allocation nursery and the to-spaces are always enumerated.  They include all
+    /// objects during mutator time, and objects determined to be live during a GC.
+    ///
+    /// If `all` is true, it will enumerate the collection nursery and the from-space, too.
+    pub(crate) fn enumerate_objects(&self, enumerator: &mut dyn ObjectEnumerator, all: bool) {
+        let sync = self.sync.lock().unwrap();
+        let mut enumerated = 0usize;
+        let mut visit_objects = |set: &HashSet<ObjectReference>| {
             for object in set.iter() {
                 enumerator.visit_object(*object);
+                enumerated += 1;
             }
         };
-        visit_objects(&self.alloc_nursery);
-        visit_objects(&self.to_space);
+        visit_objects(&sync.alloc_nursery);
+        visit_objects(&sync.to_space);
+
+        if all {
+            visit_objects(&sync.collect_nursery);
+            visit_objects(&sync.from_space);
+        }
+
+        debug!("Enumerated {enumerated} objects in LOS.  all: {all}.  from_space: {fs}, to_space: {ts}, collect_nursery: {cn}, alloc_nursery: {an}",
+            fs=sync.from_space.len(),
+            ts=sync.to_space.len(),
+            cn=sync.collect_nursery.len(),
+            an=sync.alloc_nursery.len(),
+        );
     }
 }