8342382: Implement JEP 522: G1 GC: Improve Throughput by Reducing Synchronization

src/hotspot/share/gc/g1/g1BarrierSet.hpp

@@ -51,7 +51,7 @@ class G1CardTable;
 // removes the need for any fine-grained (per mutator write) synchronization between
 // them, keeping the write barrier simple.
 //
-// The refinement threads mark cards in the the current collection set specially on the
+// The refinement threads mark cards in the current collection set specially on the
 // card table - this is fine wrt to synchronization with the mutator, because at
 // most the mutator will overwrite it again if there is a race, as G1 will scan the
 // entire card either way during the GC pause.

src/hotspot/share/gc/g1/g1CardTable.inline.hpp

@@ -44,6 +44,12 @@ inline bool G1CardTable::mark_clean_as_from_remset(CardValue* card) {
 }
 
 // Returns bits from a where mask is 0, and bits from b where mask is 1.
+//
+// Example:
+// a      = 0xAAAAAAAA
+// b      = 0xBBBBBBBB
+// mask   = 0xFF00FF00
+// result = 0xBBAABBAA
 inline size_t blend(size_t a, size_t b, size_t mask) {
   return a ^ ((a ^ b) & mask);
 }

src/hotspot/share/gc/g1/g1CardTableClaimTable.cpp

@@ -42,7 +42,7 @@ G1CardTableClaimTable::~G1CardTableClaimTable() {
   FREE_C_HEAP_ARRAY(uint, _card_claims);
 }
 
-void G1CardTableClaimTable::initialize(size_t max_reserved_regions) {
+void G1CardTableClaimTable::initialize(uint max_reserved_regions) {
   assert(_card_claims == nullptr, "Must not be initialized twice");
   _card_claims = NEW_C_HEAP_ARRAY(uint, max_reserved_regions, mtGC);
   _max_reserved_regions = max_reserved_regions;

src/hotspot/share/gc/g1/g1CardTableClaimTable.hpp

@@ -41,7 +41,7 @@ class G1HeapRegionClosure;
 // Claiming works on full region (all cards in region) or a range of contiguous cards
 // (chunk). Chunk size is given at construction time.
 class G1CardTableClaimTable : public CHeapObj<mtGC> {
-  size_t _max_reserved_regions;
+  uint _max_reserved_regions;
 
   // Card table iteration claim values for every heap region, from 0 (completely unclaimed)
   // to (>=) G1HeapRegion::CardsPerRegion (completely claimed).
@@ -57,7 +57,7 @@ class G1CardTableClaimTable : public CHeapObj<mtGC> {
   ~G1CardTableClaimTable();
 
   // Allocates the data structure and initializes the claims to unclaimed.
-  void initialize(size_t max_reserved_regions);
+  void initialize(uint max_reserved_regions);
 
   void reset_all_claims_to_unclaimed();
   void reset_all_claims_to_claimed();

src/hotspot/share/gc/g1/g1ConcurrentRefine.cpp

@@ -125,14 +125,14 @@
 G1ConcurrentRefineWorkState::G1ConcurrentRefineWorkState(uint max_reserved_regions) :
   _state(State::Idle),
   _refine_work_epoch(0),
-  _sweep_state(new G1CardTableClaimTable(G1CollectedHeap::get_chunks_per_region_for_merge())),
+  _sweep_table(new G1CardTableClaimTable(G1CollectedHeap::get_chunks_per_region_for_merge())),
   _stats()
 {
-  _sweep_state->initialize(max_reserved_regions);
+  _sweep_table->initialize(max_reserved_regions);
 }
 
 G1ConcurrentRefineWorkState::~G1ConcurrentRefineWorkState() {
-  delete _sweep_state;
+  delete _sweep_table;
 }
 
 void G1ConcurrentRefineWorkState::set_state_start_time() {
@@ -264,24 +264,34 @@
   return advance_state(State::SnapshotHeap);
 }
 
-void G1ConcurrentRefineWorkState::snapshot_heap() {
-  assert_state(State::SnapshotHeap);
+void G1ConcurrentRefineWorkState::snapshot_heap(bool concurrent) {
+  if (concurrent) {
+    assert_state(State::SnapshotHeap);
 
-  set_state_start_time();
+    set_state_start_time();
+  } else {
+    assert(is_in_progress() && _state < State::SnapshotHeap, "Must be before %s but is %s", state_name(State::SnapshotHeap), state_name(_state));
+  }
 
-  snapshot_heap_into(_sweep_state);
+  snapshot_heap_into(_sweep_table);
 
-  advance_state(State::SweepRT);
+  if (concurrent) {
+    advance_state(State::SweepRT);
+  }
 }
 
-bool G1ConcurrentRefineWorkState::sweep_rt_step() {
+void G1ConcurrentRefineWorkState::sweep_rt_start() {
   assert_state(State::SweepRT);
 
   set_state_start_time();
+}
+
+bool G1ConcurrentRefineWorkState::sweep_rt_step() {
+  assert_state(State::SweepRT);
 
   G1ConcurrentRefine* cr = G1CollectedHeap::heap()->concurrent_refine();
 
-  G1ConcurrentRefineWorkTask task(_sweep_state, &_stats, cr->num_threads_wanted());
+  G1ConcurrentRefineWorkTask task(_sweep_table, &_stats, cr->num_threads_wanted());
   cr->run_with_refinement_workers(&task);
 
   if (task.sweep_completed()) {
@@ -332,18 +342,16 @@
   return has_sweep_rt_work;
 }
 
-void G1ConcurrentRefineWorkState::snapshot_heap_into(G1CardTableClaimTable* sweep_state) {
-  sweep_state->reset_all_claims_to_claimed();
+void G1ConcurrentRefineWorkState::snapshot_heap_into(G1CardTableClaimTable* sweep_table) {
+  // G1CollectedHeap::heap_region_iterate() will only visit committed regions. In the
+  // state table 
+  sweep_table->reset_all_claims_to_claimed();
 
   class SnapshotRegionsClosure : public G1HeapRegionClosure {
-    G1CardTableClaimTable* _sweep_state;
+    G1CardTableClaimTable* _sweep_table;
 
   public:
-    size_t _num_clean;
-    size_t _num_dirty;
-    size_t _num_to_cset;
-
-    SnapshotRegionsClosure(G1CardTableClaimTable* sweep_state) : G1HeapRegionClosure(), _sweep_state(sweep_state), _num_clean(0), _num_dirty(0), _num_to_cset(0) { }
+    SnapshotRegionsClosure(G1CardTableClaimTable* sweep_table) : G1HeapRegionClosure(), _sweep_table(sweep_table) { }
 
     bool do_heap_region(G1HeapRegion* r) override {
       if (!r->is_free()) {
@@ -352,11 +360,11 @@
         // that were allocated before the handshake; the handshake makes such
         // regions' metadata visible to all threads, and we do not care about
         // humongous regions that were allocated afterwards.
-        _sweep_state->reset_to_unclaimed(r->hrm_index());
+        _sweep_table->reset_to_unclaimed(r->hrm_index());
       }
       return false;
     }
-  } cl(sweep_state);
+  } cl(sweep_table);
   G1CollectedHeap::heap()->heap_region_iterate(&cl);
 }
 
@@ -391,14 +399,31 @@
   return _thread_control.initialize(this);
 }
 
+G1ConcurrentRefineWorkState& G1ConcurrentRefine::refine_state_for_merge() {
+  bool has_sweep_claims = refine_state().complete(false);
+  if (has_sweep_claims) {
+    log_debug(gc, refine)("Continue existing work");
+  } else {
+    // Refinement has been interrupted without having a snapshot. There may
+    // be a mix of already swapped and not-swapped card tables assigned to threads,
+    // so they might have already dirtied the swapped card tables.
+    // Conservatively scan all (non-free, non-committed) region's card tables,
+    // creating the snapshot right now.
+    log_debug(gc, refine)("Create work from scratch");
+
+    refine_state().snapshot_heap(false /* concurrent */);
+  }
+  return refine_state();
+}
+
 void G1ConcurrentRefine::run_with_refinement_workers(WorkerTask* task) {
   _thread_control.run_task(task, num_threads_wanted());
 }
 
 void G1ConcurrentRefine::notify_region_reclaimed(G1HeapRegion* r) {
   assert_at_safepoint();
   if (_refine_state.is_in_progress()) {
-    _refine_state.sweep_state()->claim_all_cards(r->hrm_index());
+    _refine_state.sweep_table()->claim_all_cards(r->hrm_index());
   }
 }
 
@@ -517,7 +542,7 @@
   size_t sampled_code_root_rs_length() const { return _sampled_code_root_rs_length; }
 };
 
-// Adjust the target length (in regions) of the young gen, based on the the
+// Adjust the target length (in regions) of the young gen, based on the
 // current length of the remembered sets.
 //
 // At the end of the GC G1 determines the length of the young gen based on

src/hotspot/share/gc/g1/g1ConcurrentRefine.hpp

@@ -113,7 +113,7 @@ class G1ConcurrentRefineWorkState {
   size_t _refine_work_epoch;
 
   // Current heap snapshot.
-  G1CardTableClaimTable* _sweep_state;
+  G1CardTableClaimTable* _sweep_table;
 
   // Start times for all states.
   Ticks _state_start[static_cast<uint>(State::Last)];
@@ -140,14 +140,15 @@ class G1ConcurrentRefineWorkState {
   bool swap_global_card_table();
   bool swap_java_threads_ct();
   bool swap_gc_threads_ct();
-  void snapshot_heap();
+  void snapshot_heap(bool concurrent = true);
+  void sweep_rt_start();
   bool sweep_rt_step();
 
   bool complete(bool concurrent, bool print_log = true);
 
-  static void snapshot_heap_into(G1CardTableClaimTable* sweep_state);
+  static void snapshot_heap_into(G1CardTableClaimTable* sweep_table);
 
-  G1CardTableClaimTable* sweep_state() { return _sweep_state; }
+  G1CardTableClaimTable* sweep_table() { return _sweep_table; }
   G1ConcurrentRefineStats* stats() { return &_stats; }
   void reset_stats();
 
@@ -231,6 +232,8 @@ class G1ConcurrentRefine : public CHeapObj<mtGC> {
 
   G1ConcurrentRefineWorkState& refine_state() { return _refine_state; }
 
+  G1ConcurrentRefineWorkState& refine_state_for_merge();
+
   void run_with_refinement_workers(WorkerTask* task);
 
   void notify_region_reclaimed(G1HeapRegion* r);

src/hotspot/share/gc/g1/g1ConcurrentRefineThread.cpp

@@ -177,6 +177,8 @@ void G1ConcurrentRefineThread::do_refinement() {
 
   // 5. Sweep refinement table until done
   bool interrupted_by_gc = false;
+
+  state.sweep_rt_start();
   while (true) {
     bool completed = state.sweep_rt_step();
 

src/hotspot/share/gc/g1/g1RemSet.cpp

@@ -90,7 +90,7 @@
 class G1RemSetScanState : public CHeapObj<mtGC> {
   class G1DirtyRegions;
 
-  G1CardTableClaimTable _card_state;
+  G1CardTableClaimTable _card_claim_table;
   // The complete set of regions which card table needs to be cleared at the end
   // of GC because we scribbled over these card table entries.
   //
@@ -142,11 +142,11 @@ class G1RemSetScanState : public CHeapObj<mtGC> {
     }
 
     void add_dirty_region(uint region) {
-    if (_contains[region]) {
-      return;
-    }
+      if (_contains[region]) {
+        return;
+      }
 
-    bool marked_as_dirty = Atomic::cmpxchg(&_contains[region], false, true) == false;
+      bool marked_as_dirty = Atomic::cmpxchg(&_contains[region], false, true) == false;
       if (marked_as_dirty) {
         uint allocated = Atomic::fetch_then_add(&_cur_idx, 1u);
         _buffer[allocated] = region;
@@ -234,7 +234,7 @@ class G1ClearCardTableTask : public G1AbstractSubTask {
 
 public:
   G1RemSetScanState() :
-    _card_state(G1CollectedHeap::get_chunks_per_region_for_scan()),
+    _card_claim_table(G1CollectedHeap::get_chunks_per_region_for_scan()),
     _all_dirty_regions(nullptr),
     _next_dirty_regions(nullptr),
     _scan_top(nullptr) { }
@@ -243,8 +243,8 @@ class G1ClearCardTableTask : public G1AbstractSubTask {
     FREE_C_HEAP_ARRAY(HeapWord*, _scan_top);
   }
 
-  void initialize(size_t max_reserved_regions) {
-    _card_state.initialize(max_reserved_regions);
+  void initialize(uint max_reserved_regions) {
+    _card_claim_table.initialize(max_reserved_regions);
     _scan_top = NEW_C_HEAP_ARRAY(HeapWord*, max_reserved_regions, mtGC);
   }
 
@@ -253,7 +253,7 @@ class G1ClearCardTableTask : public G1AbstractSubTask {
   // become used during the collection these values must be valid
   // for those regions as well.
   void prepare() {
-    size_t max_reserved_regions = _card_state.max_reserved_regions();
+    size_t max_reserved_regions = _card_claim_table.max_reserved_regions();
 
     for (size_t i = 0; i < max_reserved_regions; i++) {
       clear_scan_top((uint)i);
@@ -268,7 +268,7 @@ class G1ClearCardTableTask : public G1AbstractSubTask {
     // regions.
     //assert(_next_dirty_regions->size() == 0, "next dirty regions must be empty");
 
-    _card_state.reset_all_claims_to_unclaimed();
+    _card_claim_table.reset_all_claims_to_unclaimed();
   }
 
   void complete_evac_phase(bool merge_dirty_regions) {
@@ -331,7 +331,7 @@ class G1ClearCardTableTask : public G1AbstractSubTask {
   }
 
   bool has_cards_to_scan(uint region) {
-    return _card_state.has_unclaimed_cards(region);
+    return _card_claim_table.has_unclaimed_cards(region);
   }
 
   void add_dirty_region(uint const region) {
@@ -367,7 +367,7 @@ class G1ClearCardTableTask : public G1AbstractSubTask {
   }
 
   G1CardTableChunkClaimer claimer(uint region_idx) {
-    return G1CardTableChunkClaimer(&_card_state, region_idx);
+    return G1CardTableChunkClaimer(&_card_claim_table, region_idx);
   }
 };
 
@@ -827,20 +827,20 @@ class MergeRefinementTableTask : public WorkerTask {
       while (claim.has_next()) {
         size_t const start_idx = region_card_base_idx + claim.value();
 
-        size_t* card_cur_card = (size_t*)card_table->byte_for_index(start_idx);
+        size_t* card_cur_word = (size_t*)card_table->byte_for_index(start_idx);
 
         size_t* refinement_cur_card = (size_t*)refinement_table->byte_for_index(start_idx);
         size_t* const refinement_end_card = refinement_cur_card + claim.size() / (sizeof(size_t) / sizeof(G1CardTable::CardValue));
 
-        for (; refinement_cur_card < refinement_end_card; ++refinement_cur_card, ++card_cur_card) {
+        for (; refinement_cur_card < refinement_end_card; ++refinement_cur_card, ++card_cur_word) {
           size_t value = *refinement_cur_card;
           *refinement_cur_card = G1CardTable::WordAllClean;
           // Dirty is "0", so we need to logically-and here. This is also safe
           // for all other possible values in the card table; at this point this
           // can be either g1_dirty_card or g1_to_cset_card which will both be
           // scanned.
-          size_t new_value = *card_cur_card & value;
-          *card_cur_card = new_value;
+          size_t new_value = *card_cur_word & value;
+          *card_cur_word = new_value;
         }
       }
 
@@ -1230,30 +1230,10 @@ class G1MergeHeapRootsTask : public WorkerTask {
 static void merge_refinement_table() {
   G1CollectedHeap* g1h = G1CollectedHeap::heap();
 
-  G1CardTableClaimTable* claim;
-  G1CardTableClaimTable constructed(G1CollectedHeap::get_chunks_per_region_for_merge());
-
-  G1ConcurrentRefineWorkState& state = g1h->concurrent_refine()->refine_state();
-  bool has_sweep_claims = state.complete(false);
-  if (has_sweep_claims) {
-    log_debug(gc, refine)("Continue existing work");
-    claim = state.sweep_state();
-  } else {
-    // Refinement has been interrupted without having a snapshot. There may
-    // be a mix of already swapped and not-swapped card tables assigned to threads,
-    // so they might have already dirtied the swapped card tables.
-    // Conservatively scan all (non-free, non-committed) region's card tables,
-    // creating the snapshot right now.
-    log_debug(gc, refine)("Create work from scratch");
-
-    constructed.initialize(g1h->max_reserved_regions());
-    G1ConcurrentRefineWorkState::snapshot_heap_into(&constructed);
-    claim = &constructed;
-  }
-
+  G1ConcurrentRefineWorkState& state = g1h->concurrent_refine()->refine_state_for_merge();
   WorkerThreads* workers = g1h->workers();
 
-  MergeRefinementTableTask cl(claim, workers->active_workers());
+  MergeRefinementTableTask cl(state.sweep_table(), workers->active_workers());
   log_debug(gc, ergo)("Running %s using %u workers", cl.name(), workers->active_workers());
   workers->run_task(&cl);
 }