Aggregate live bytes info in on_gc_finished (#1292)

This PR moves the call to `aggregate_live_bytes_in_last_gc` from the
`Release` work packet to `on_gc_finished`. Spaces may do release work in
parallel with the `Release` work packet, and
`aggregate_live_bytes_in_last_gc` accesses reserved pages of spaces.
This leads to a data race. We observed the reserved pages in the
collected live bytes stats to be larger than the actual reserved pages
at the end of a GC, as some pages were not released yet.

Author: qinsoon

src/scheduler/gc_work.rs

@@ -153,15 +153,6 @@ impl<C: GCWorkContext + 'static> GCWork<C::VM> for Release<C> {
             let result = w.designated_work.push(Box::new(ReleaseCollector));
             debug_assert!(result.is_ok());
         }
-
-        if *mmtk.get_options().count_live_bytes_in_gc {
-            let live_bytes = mmtk
-                .scheduler
-                .worker_group
-                .get_and_clear_worker_live_bytes();
-            *mmtk.state.live_bytes_in_last_gc.borrow_mut() =
-                mmtk.aggregate_live_bytes_in_last_gc(live_bytes);
-        }
     }
 }
 

src/scheduler/scheduler.rs

@@ -554,7 +554,15 @@ impl<VM: VMBinding> GCWorkScheduler<VM> {
         probe!(mmtk, gc_end);
 
         if *mmtk.get_options().count_live_bytes_in_gc {
-            for (space_name, &stats) in mmtk.state.live_bytes_in_last_gc.borrow().iter() {
+            // Aggregate the live bytes
+            let live_bytes = mmtk
+                .scheduler
+                .worker_group
+                .get_and_clear_worker_live_bytes();
+            let mut live_bytes_in_last_gc = mmtk.state.live_bytes_in_last_gc.borrow_mut();
+            *live_bytes_in_last_gc = mmtk.aggregate_live_bytes_in_last_gc(live_bytes);
+            // Logging
+            for (space_name, &stats) in live_bytes_in_last_gc.iter() {
                 info!(
                     "{} = {} pages ({:.1}% live)",
                     space_name,