Introduce VMSpace, and allow VMSpace to be set lazily (#802)

This PR allows users to set the start address and size for VM space, and
also allows VM space to be initialised lazily.

* Add `VMSpace` which wraps the normal `ImmortalSpace`. The policy is
mostly the same as `ImmortalSpace` except that the space may be
uninitialized and can be initialized later.
* Add option `vm_space_start` and `vm_space_size`.
* Modify the heap start constant for 32 bits, so JikesRVM's VM space
does not overlap with our heap range.
* Fix a subtraction overflow issue in `align_allocation`.

Cargo.toml

@@ -45,6 +45,7 @@ itertools = "0.10.5"
 sys-info = "0.9"
 regex = "1.7.0"
 static_assertions = "1.1.0"
+delegate = "0.9.0"
 
 [dev-dependencies]
 rand = "0.8.5"

docs/tutorial/code/mygc_semispace/global.rs

@@ -154,6 +154,10 @@ impl<VM: VMBinding> Plan for MyGC<VM> {
     fn base(&self) -> &BasePlan<VM> {
         &self.common.base
     }
+
+    fn base_mut(&mut self) -> &mut BasePlan<Self::VM> {
+        &mut self.common.base
+    }
     // ANCHOR_END: plan_base
 
     // Add

src/memory_manager.rs

@@ -85,6 +85,11 @@ pub fn mmtk_init<VM: VMBinding>(builder: &MMTKBuilder) -> Box<MMTK<VM>> {
     Box::new(mmtk)
 }
 
+#[cfg(feature = "vm_space")]
+pub fn lazy_init_vm_space<VM: VMBinding>(mmtk: &'static mut MMTK<VM>, start: Address, size: usize) {
+    mmtk.plan.base_mut().vm_space.lazy_initialize(start, size);
+}
+
 /// Request MMTk to create a mutator for the given thread. The ownership
 /// of returned boxed mutator is transferred to the binding, and the binding needs to take care of its
 /// lifetime. For performance reasons, A VM should store the returned mutator in a thread local storage

src/plan/generational/copying/global.rs

@@ -148,6 +148,10 @@ impl<VM: VMBinding> Plan for GenCopy<VM> {
         &self.gen.common.base
     }
 
+    fn base_mut(&mut self) -> &mut BasePlan<Self::VM> {
+        &mut self.gen.common.base
+    }
+
     fn common(&self) -> &CommonPlan<VM> {
         &self.gen.common
     }

src/plan/generational/immix/global.rs

@@ -178,6 +178,10 @@ impl<VM: VMBinding> Plan for GenImmix<VM> {
         &self.gen.common.base
     }
 
+    fn base_mut(&mut self) -> &mut BasePlan<Self::VM> {
+        &mut self.gen.common.base
+    }
+
     fn common(&self) -> &CommonPlan<VM> {
         &self.gen.common
     }

src/plan/global.rs

@@ -8,6 +8,8 @@ use crate::plan::Mutator;
 use crate::policy::immortalspace::ImmortalSpace;
 use crate::policy::largeobjectspace::LargeObjectSpace;
 use crate::policy::space::{PlanCreateSpaceArgs, Space};
+#[cfg(feature = "vm_space")]
+use crate::policy::vmspace::VMSpace;
 use crate::scheduler::*;
 use crate::util::alloc::allocators::AllocatorSelector;
 #[cfg(feature = "analysis")]
@@ -176,6 +178,7 @@ pub trait Plan: 'static + Sync + Downcast {
     }
 
     fn base(&self) -> &BasePlan<Self::VM>;
+    fn base_mut(&mut self) -> &mut BasePlan<Self::VM>;
     fn schedule_collection(&'static self, _scheduler: &GCWorkScheduler<Self::VM>);
     fn common(&self) -> &CommonPlan<Self::VM> {
         panic!("Common Plan not handled!")
@@ -423,29 +426,7 @@ pub struct BasePlan<VM: VMBinding> {
     ///     the VM space.
     #[cfg(feature = "vm_space")]
     #[trace]
-    pub vm_space: ImmortalSpace<VM>,
-}
-
-#[cfg(feature = "vm_space")]
-pub fn create_vm_space<VM: VMBinding>(args: &mut CreateSpecificPlanArgs<VM>) -> ImmortalSpace<VM> {
-    use crate::util::constants::LOG_BYTES_IN_MBYTE;
-    let boot_segment_bytes = *args.global_args.options.vm_space_size;
-    debug_assert!(boot_segment_bytes > 0);
-
-    use crate::util::conversions::raw_align_up;
-    use crate::util::heap::layout::vm_layout_constants::BYTES_IN_CHUNK;
-    let boot_segment_mb = raw_align_up(boot_segment_bytes, BYTES_IN_CHUNK) >> LOG_BYTES_IN_MBYTE;
-
-    let space = ImmortalSpace::new_vm_space(args.get_space_args(
-        "boot",
-        false,
-        VMRequest::fixed_size(boot_segment_mb),
-    ));
-
-    // The space is mapped externally by the VM. We need to update our mmapper to mark the range as mapped.
-    space.ensure_mapped();
-
-    space
+    pub vm_space: VMSpace<VM>,
 }
 
 /// Args needed for creating any plan. This includes a set of contexts from MMTK or global. This
@@ -520,7 +501,7 @@ impl<VM: VMBinding> BasePlan<VM> {
                 VMRequest::discontiguous(),
             )),
             #[cfg(feature = "vm_space")]
-            vm_space: create_vm_space(&mut args),
+            vm_space: VMSpace::new(&mut args),
 
             initialized: AtomicBool::new(false),
             trigger_gc_when_heap_is_full: AtomicBool::new(true),

src/plan/immix/global.rs

@@ -117,6 +117,10 @@ impl<VM: VMBinding> Plan for Immix<VM> {
         &self.common.base
     }
 
+    fn base_mut(&mut self) -> &mut BasePlan<Self::VM> {
+        &mut self.common.base
+    }
+
     fn common(&self) -> &CommonPlan<VM> {
         &self.common
     }

src/plan/markcompact/global.rs

@@ -61,6 +61,10 @@ impl<VM: VMBinding> Plan for MarkCompact<VM> {
         &self.common.base
     }
 
+    fn base_mut(&mut self) -> &mut BasePlan<Self::VM> {
+        &mut self.common.base
+    }
+
     fn common(&self) -> &CommonPlan<VM> {
         &self.common
     }

src/plan/marksweep/global.rs

@@ -87,6 +87,10 @@ impl<VM: VMBinding> Plan for MarkSweep<VM> {
         &self.common.base
     }
 
+    fn base_mut(&mut self) -> &mut BasePlan<Self::VM> {
+        &mut self.common.base
+    }
+
     fn common(&self) -> &CommonPlan<VM> {
         &self.common
     }

src/plan/mod.rs

@@ -26,6 +26,8 @@ pub use global::AllocationSemantics;
 pub(crate) use global::GcStatus;
 pub use global::Plan;
 pub(crate) use global::PlanTraceObject;
+#[cfg(feature = "vm_space")] // This is used for creating VM space
+pub(crate) use global::{CreateGeneralPlanArgs, CreateSpecificPlanArgs};
 
 mod mutator_context;
 pub use mutator_context::Mutator;

src/plan/nogc/global.rs

@@ -53,6 +53,10 @@ impl<VM: VMBinding> Plan for NoGC<VM> {
         &self.base
     }
 
+    fn base_mut(&mut self) -> &mut BasePlan<Self::VM> {
+        &mut self.base
+    }
+
     fn prepare(&mut self, _tls: VMWorkerThread) {
         unreachable!()
     }

src/plan/pageprotect/global.rs

@@ -77,6 +77,10 @@ impl<VM: VMBinding> Plan for PageProtect<VM> {
         &self.common.base
     }
 
+    fn base_mut(&mut self) -> &mut BasePlan<Self::VM> {
+        &mut self.common.base
+    }
+
     fn common(&self) -> &CommonPlan<VM> {
         &self.common
     }

src/plan/semispace/global.rs

@@ -129,6 +129,10 @@ impl<VM: VMBinding> Plan for SemiSpace<VM> {
         &self.common.base
     }
 
+    fn base_mut(&mut self) -> &mut BasePlan<Self::VM> {
+        &mut self.common.base
+    }
+
     fn common(&self) -> &CommonPlan<VM> {
         &self.common
     }

src/plan/sticky/immix/global.rs

@@ -67,6 +67,10 @@ impl<VM: VMBinding> Plan for StickyImmix<VM> {
         self.immix.base()
     }
 
+    fn base_mut(&mut self) -> &mut crate::plan::global::BasePlan<Self::VM> {
+        self.immix.base_mut()
+    }
+
     fn generational(
         &self,
     ) -> Option<&dyn crate::plan::generational::global::GenerationalPlan<VM = Self::VM>> {

src/policy/immortalspace.rs

@@ -120,15 +120,6 @@ impl<VM: VMBinding> crate::policy::gc_work::PolicyTraceObject<VM> for ImmortalSp
 
 impl<VM: VMBinding> ImmortalSpace<VM> {
     pub fn new(args: crate::policy::space::PlanCreateSpaceArgs<VM>) -> Self {
-        Self::new_inner(args, false)
-    }
-
-    #[cfg(feature = "vm_space")]
-    pub fn new_vm_space(args: crate::policy::space::PlanCreateSpaceArgs<VM>) -> Self {
-        Self::new_inner(args, true)
-    }
-
-    pub fn new_inner(args: crate::policy::space::PlanCreateSpaceArgs<VM>, vm_space: bool) -> Self {
         let vm_map = args.vm_map;
         let is_discontiguous = args.vmrequest.is_discontiguous();
         let common = CommonSpace::new(args.into_policy_args(
@@ -144,7 +135,26 @@ impl<VM: VMBinding> ImmortalSpace<VM> {
                 MonotonePageResource::new_contiguous(common.start, common.extent, vm_map)
             },
             common,
-            vm_space,
+            vm_space: false,
+        }
+    }
+
+    #[cfg(feature = "vm_space")]
+    pub fn new_vm_space(
+        args: crate::policy::space::PlanCreateSpaceArgs<VM>,
+        start: Address,
+        size: usize,
+    ) -> Self {
+        assert!(!args.vmrequest.is_discontiguous());
+        ImmortalSpace {
+            mark_state: MarkState::new(),
+            pr: MonotonePageResource::new_contiguous(start, size, args.vm_map),
+            common: CommonSpace::new(args.into_policy_args(
+                false,
+                true,
+                metadata::extract_side_metadata(&[*VM::VMObjectModel::LOCAL_MARK_BIT_SPEC]),
+            )),
+            vm_space: true,
         }
     }
 

src/policy/mod.rs

@@ -27,3 +27,5 @@ pub mod largeobjectspace;
 pub mod lockfreeimmortalspace;
 pub mod markcompactspace;
 pub mod marksweepspace;
+#[cfg(feature = "vm_space")]
+pub mod vmspace;

src/policy/sft_map.rs

@@ -120,8 +120,13 @@ mod space_map {
                 assert!(old.name() == EMPTY_SFT_NAME || old.name() == space.name());
                 // Make sure the range is in the space
                 let space_start = Self::index_to_space_start(index);
-                assert!(start >= space_start);
-                assert!(start + bytes <= space_start + MAX_SPACE_EXTENT);
+                // FIXME: Curerntly skip the check for the last space. The following works fine for MMTk internal spaces,
+                // but the VM space is an exception. Any address after the last space is considered as the last space,
+                // based on our indexing function. In that case, we cannot assume the end of the region is within the last space (with MAX_SPACE_EXTENT).
+                if index != Self::TABLE_SIZE - 1 {
+                    assert!(start >= space_start);
+                    assert!(start + bytes <= space_start + MAX_SPACE_EXTENT);
+                }
             }
             *mut_self.sft.get_unchecked_mut(index) = space;
         }
@@ -512,8 +517,9 @@ mod sparse_chunk_map {
                 // in which case, we still set SFT map again.
                 debug_assert!(
                     old == EMPTY_SFT_NAME || new == EMPTY_SFT_NAME || old == new,
-                    "attempt to overwrite a non-empty chunk {} in SFT map (from {} to {})",
+                    "attempt to overwrite a non-empty chunk {} ({}) in SFT map (from {} to {})",
                     chunk,
+                    crate::util::conversions::chunk_index_to_address(chunk),
                     old,
                     new
                 );

src/policy/space.rs

@@ -490,10 +490,8 @@ impl<VM: VMBinding> CommonSpace<VM> {
                 top: _top,
             } => (_extent, _top),
             VMRequest::Fixed {
-                extent: _extent,
-                top: _top,
-                ..
-            } => (_extent, _top),
+                extent: _extent, ..
+            } => (_extent, false),
             _ => unreachable!(),
         };
 
@@ -524,7 +522,22 @@ impl<VM: VMBinding> CommonSpace<VM> {
         // FIXME
         rtn.descriptor = SpaceDescriptor::create_descriptor_from_heap_range(start, start + extent);
         // VM.memory.setHeapRange(index, start, start.plus(extent));
-        args.plan_args.vm_map.insert(start, extent, rtn.descriptor);
+
+        // We only initialize our vm map if the range of the space is in our available heap range. For normally spaces,
+        // they are definitely in our heap range. But for VM space, a runtime could give us an arbitrary range. We only
+        // insert into our vm map if the range overlaps with our heap.
+        {
+            use crate::util::heap::layout;
+            let overlap =
+                Address::range_intersection(&(start..start + extent), &layout::available_range());
+            if !overlap.is_empty() {
+                args.plan_args.vm_map.insert(
+                    overlap.start,
+                    overlap.end - overlap.start,
+                    rtn.descriptor,
+                );
+            }
+        }
 
         // For contiguous space, we know its address range so we reserve metadata memory for its range.
         if rtn