8260044: Parallel GC: Concurrent allocation after heap expansion may cause unnecessary full gc

8260045: Parallel GC: Waiting on ExpandHeap_lock may cause "expansion storm"

Loop to retry allocation if expand succeeds.  Treat space available after obtaining expand lock as expand success.

Reviewed-by: tschatzl, iwalulya, sjohanss

Author: Kim Barrett

src/hotspot/share/gc/parallel/mutableSpace.cpp

@@ -215,6 +215,15 @@ bool MutableSpace::cas_deallocate(HeapWord *obj, size_t size) {
   return Atomic::cmpxchg(top_addr(), expected_top, obj) == expected_top;
 }
 
+// Only used by oldgen allocation.
+bool MutableSpace::needs_expand(size_t word_size) const {
+  assert_lock_strong(ExpandHeap_lock);
+  // Holding the lock means end is stable.  So while top may be advancing
+  // via concurrent allocations, there is no need to order the reads of top
+  // and end here, unlike in cas_allocate.
+  return pointer_delta(end(), top()) < word_size;
+}
+
 void MutableSpace::oop_iterate(OopIterateClosure* cl) {
   HeapWord* obj_addr = bottom();
   HeapWord* t = top();

src/hotspot/share/gc/parallel/mutableSpace.hpp

@@ -142,6 +142,11 @@ class MutableSpace: public CHeapObj<mtGC> {
   virtual HeapWord* cas_allocate(size_t word_size);
   // Optional deallocation. Used in NUMA-allocator.
   bool cas_deallocate(HeapWord *obj, size_t size);
+  // Return true if this space needs to be expanded in order to satisfy an
+  // allocation request of the indicated size.  Concurrent allocations and
+  // resizes may change the result of a later call.  Used by oldgen allocator.
+  // precondition: holding ExpandHeap_lock
+  bool needs_expand(size_t word_size) const;
 
   // Iteration.
   void oop_iterate(OopIterateClosure* cl);

src/hotspot/share/gc/parallel/psOldGen.cpp

@@ -178,19 +178,31 @@ void PSOldGen::object_iterate_block(ObjectClosure* cl, size_t block_index) {
   }
 }
 
-HeapWord* PSOldGen::expand_and_cas_allocate(size_t word_size) {
-  expand(word_size*HeapWordSize);
+bool PSOldGen::expand_for_allocate(size_t word_size) {
+  assert(word_size > 0, "allocating zero words?");
+  bool result = true;
+  {
+    MutexLocker x(ExpandHeap_lock);
+    // Avoid "expand storms" by rechecking available space after obtaining
+    // the lock, because another thread may have already made sufficient
+    // space available.  If insufficient space available, that will remain
+    // true until we expand, since we have the lock.  Other threads may take
+    // the space we need before we can allocate it, regardless of whether we
+    // expand.  That's okay, we'll just try expanding again.
+    if (object_space()->needs_expand(word_size)) {
+      result = expand(word_size*HeapWordSize);
+    }
+  }
   if (GCExpandToAllocateDelayMillis > 0) {
     os::naked_sleep(GCExpandToAllocateDelayMillis);
   }
-  return cas_allocate_noexpand(word_size);
+  return result;
 }
 
-void PSOldGen::expand(size_t bytes) {
-  if (bytes == 0) {
-    return;
-  }
-  MutexLocker x(ExpandHeap_lock);
+bool PSOldGen::expand(size_t bytes) {
+  assert_lock_strong(ExpandHeap_lock);
+  assert_locked_or_safepoint(Heap_lock);
+  assert(bytes > 0, "precondition");
   const size_t alignment = virtual_space()->alignment();
   size_t aligned_bytes  = align_up(bytes, alignment);
   size_t aligned_expand_bytes = align_up(MinHeapDeltaBytes, alignment);
@@ -200,13 +212,11 @@ void PSOldGen::expand(size_t bytes) {
     // providing a page per lgroup. Alignment is larger or equal to the page size.
     aligned_expand_bytes = MAX2(aligned_expand_bytes, alignment * os::numa_get_groups_num());
   }
-  if (aligned_bytes == 0){
-    // The alignment caused the number of bytes to wrap.  An expand_by(0) will
-    // return true with the implication that and expansion was done when it
-    // was not.  A call to expand implies a best effort to expand by "bytes"
-    // but not a guarantee.  Align down to give a best effort.  This is likely
-    // the most that the generation can expand since it has some capacity to
-    // start with.
+  if (aligned_bytes == 0) {
+    // The alignment caused the number of bytes to wrap.  A call to expand
+    // implies a best effort to expand by "bytes" but not a guarantee.  Align
+    // down to give a best effort.  This is likely the most that the generation
+    // can expand since it has some capacity to start with.
     aligned_bytes = align_down(bytes, alignment);
   }
 
@@ -224,14 +234,13 @@ void PSOldGen::expand(size_t bytes) {
   if (success && GCLocker::is_active_and_needs_gc()) {
     log_debug(gc)("Garbage collection disabled, expanded heap instead");
   }
+  return success;
 }
 
 bool PSOldGen::expand_by(size_t bytes) {
   assert_lock_strong(ExpandHeap_lock);
   assert_locked_or_safepoint(Heap_lock);
-  if (bytes == 0) {
-    return true;  // That's what virtual_space()->expand_by(0) would return
-  }
+  assert(bytes > 0, "precondition");
   bool result = virtual_space()->expand_by(bytes);
   if (result) {
     if (ZapUnusedHeapArea) {
@@ -268,7 +277,7 @@ bool PSOldGen::expand_to_reserved() {
   assert_lock_strong(ExpandHeap_lock);
   assert_locked_or_safepoint(Heap_lock);
 
-  bool result = true;
+  bool result = false;
   const size_t remaining_bytes = virtual_space()->uncommitted_size();
   if (remaining_bytes > 0) {
     result = expand_by(remaining_bytes);
@@ -323,10 +332,10 @@ void PSOldGen::resize(size_t desired_free_space) {
   }
   if (new_size > current_size) {
     size_t change_bytes = new_size - current_size;
+    MutexLocker x(ExpandHeap_lock);
     expand(change_bytes);
   } else {
     size_t change_bytes = current_size - new_size;
-    // shrink doesn't grab this lock, expand does. Is that right?
     MutexLocker x(ExpandHeap_lock);
     shrink(change_bytes);
   }

src/hotspot/share/gc/parallel/psOldGen.hpp

@@ -79,8 +79,8 @@ class PSOldGen : public CHeapObj<mtGC> {
     return res;
   }
 
-  HeapWord* expand_and_cas_allocate(size_t word_size);
-  void expand(size_t bytes);
+  bool expand_for_allocate(size_t word_size);
+  bool expand(size_t bytes);
   bool expand_by(size_t bytes);
   bool expand_to_reserved();
 
@@ -135,8 +135,12 @@ class PSOldGen : public CHeapObj<mtGC> {
   void resize(size_t desired_free_space);
 
   HeapWord* allocate(size_t word_size) {
-    HeapWord* res = cas_allocate_noexpand(word_size);
-    return (res == NULL) ? expand_and_cas_allocate(word_size) : res;
+    HeapWord* res;
+    do {
+      res = cas_allocate_noexpand(word_size);
+      // Retry failed allocation if expand succeeds.
+    } while ((res == nullptr) && expand_for_allocate(word_size));
+    return res;
   }
 
   // Iteration.