7023069: G1: Introduce symmetric locking in the slow allocation path

7023151: G1: refactor the code that operates on _cur_alloc_region to be re-used for allocs by the GC threads
7018286: G1: humongous allocation attempts should take the GC locker into account
Summary: First, this change replaces the asymmetric locking scheme in the G1 slow alloc path by a summetric one. Second, it factors out the code that operates on _cur_alloc_region so that it can be re-used for allocations by the GC threads in the future.
Reviewed-by: stefank, brutisso, johnc

src/share/vm/gc_implementation/g1/g1AllocRegion.cpp

+/*
+ * Copyright (c) 2011, Oracle and/or its affiliates. All rights reserved.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This code is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 only, as
+ * published by the Free Software Foundation.
+ *
+ * This code is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+ * version 2 for more details (a copy is included in the LICENSE file that
+ * accompanied this code).
+ *
+ * You should have received a copy of the GNU General Public License version
+ * 2 along with this work; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
+ * or visit www.oracle.com if you need additional information or have any
+ * questions.
+ *
+ */
+
+#include "precompiled.hpp"
+#include "gc_implementation/g1/g1AllocRegion.inline.hpp"
+#include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
+
+G1CollectedHeap* G1AllocRegion::_g1h = NULL;
+HeapRegion* G1AllocRegion::_dummy_region = NULL;
+
+void G1AllocRegion::setup(G1CollectedHeap* g1h, HeapRegion* dummy_region) {
+  assert(_dummy_region == NULL, "should be set once");
+  assert(dummy_region != NULL, "pre-condition");
+  assert(dummy_region->free() == 0, "pre-condition");
+
+  // Make sure that any allocation attempt on this region will fail
+  // and will not trigger any asserts.
+  assert(allocate(dummy_region, 1, false) == NULL, "should fail");
+  assert(par_allocate(dummy_region, 1, false) == NULL, "should fail");
+  assert(allocate(dummy_region, 1, true) == NULL, "should fail");
+  assert(par_allocate(dummy_region, 1, true) == NULL, "should fail");
+
+  _g1h = g1h;
+  _dummy_region = dummy_region;
+}
+
+void G1AllocRegion::fill_up_remaining_space(HeapRegion* alloc_region,
+                                            bool bot_updates) {
+  assert(alloc_region != NULL && alloc_region != _dummy_region,
+         "pre-condition");
+
+  // Other threads might still be trying to allocate using a CAS out
+  // of the region we are trying to retire, as they can do so without
+  // holding the lock. So, we first have to make sure that noone else
+  // can allocate out of it by doing a maximal allocation. Even if our
+  // CAS attempt fails a few times, we'll succeed sooner or later
+  // given that failed CAS attempts mean that the region is getting
+  // closed to being full.
+  size_t free_word_size = alloc_region->free() / HeapWordSize;
+
+  // This is the minimum free chunk we can turn into a dummy
+  // object. If the free space falls below this, then noone can
+  // allocate in this region anyway (all allocation requests will be
+  // of a size larger than this) so we won't have to perform the dummy
+  // allocation.
+  size_t min_word_size_to_fill = CollectedHeap::min_fill_size();
+
+  while (free_word_size >= min_word_size_to_fill) {
+    HeapWord* dummy = par_allocate(alloc_region, free_word_size, bot_updates);
+    if (dummy != NULL) {
+      // If the allocation was successful we should fill in the space.
+      CollectedHeap::fill_with_object(dummy, free_word_size);
+      alloc_region->set_pre_dummy_top(dummy);
+      break;
+    }
+
+    free_word_size = alloc_region->free() / HeapWordSize;
+    // It's also possible that someone else beats us to the
+    // allocation and they fill up the region. In that case, we can
+    // just get out of the loop.
+  }
+  assert(alloc_region->free() / HeapWordSize < min_word_size_to_fill,
+         "post-condition");
+}
+
+void G1AllocRegion::retire(bool fill_up) {
+  assert(_alloc_region != NULL, ar_ext_msg(this, "not initialized properly"));
+
+  trace("retiring");
+  HeapRegion* alloc_region = _alloc_region;
+  if (alloc_region != _dummy_region) {
+    // We never have to check whether the active region is empty or not,
+    // and potentially free it if it is, given that it's guaranteed that
+    // it will never be empty.
+    assert(!alloc_region->is_empty(),
+           ar_ext_msg(this, "the alloc region should never be empty"));
+
+    if (fill_up) {
+      fill_up_remaining_space(alloc_region, _bot_updates);
+    }
+
+    assert(alloc_region->used() >= _used_bytes_before,
+           ar_ext_msg(this, "invariant"));
+    size_t allocated_bytes = alloc_region->used() - _used_bytes_before;
+    retire_region(alloc_region, allocated_bytes);
+    _used_bytes_before = 0;
+    _alloc_region = _dummy_region;
+  }
+  trace("retired");
+}
+
+HeapWord* G1AllocRegion::new_alloc_region_and_allocate(size_t word_size,
+                                                       bool force) {
+  assert(_alloc_region == _dummy_region, ar_ext_msg(this, "pre-condition"));
+  assert(_used_bytes_before == 0, ar_ext_msg(this, "pre-condition"));
+
+  trace("attempting region allocation");
+  HeapRegion* new_alloc_region = allocate_new_region(word_size, force);
+  if (new_alloc_region != NULL) {
+    new_alloc_region->reset_pre_dummy_top();
+    // Need to do this before the allocation
+    _used_bytes_before = new_alloc_region->used();
+    HeapWord* result = allocate(new_alloc_region, word_size, _bot_updates);
+    assert(result != NULL, ar_ext_msg(this, "the allocation should succeeded"));
+
+    OrderAccess::storestore();
+    // Note that we first perform the allocation and then we store the
+    // region in _alloc_region. This is the reason why an active region
+    // can never be empty.
+    _alloc_region = new_alloc_region;
+    trace("region allocation successful");
+    return result;
+  } else {
+    trace("region allocation failed");
+    return NULL;
+  }
+  ShouldNotReachHere();
+}
+
+void G1AllocRegion::fill_in_ext_msg(ar_ext_msg* msg, const char* message) {
+  msg->append("[%s] %s b: %s r: "PTR_FORMAT" u: "SIZE_FORMAT,
+              _name, message, BOOL_TO_STR(_bot_updates),
+              _alloc_region, _used_bytes_before);
+}
+
+void G1AllocRegion::init() {
+  trace("initializing");
+  assert(_alloc_region == NULL && _used_bytes_before == 0,
+         ar_ext_msg(this, "pre-condition"));
+  assert(_dummy_region != NULL, "should have been set");
+  _alloc_region = _dummy_region;
+  trace("initialized");
+}
+
+HeapRegion* G1AllocRegion::release() {
+  trace("releasing");
+  HeapRegion* alloc_region = _alloc_region;
+  retire(false /* fill_up */);
+  assert(_alloc_region == _dummy_region, "post-condition of retire()");
+  _alloc_region = NULL;
+  trace("released");
+  return (alloc_region == _dummy_region) ? NULL : alloc_region;
+}
+
+#if G1_ALLOC_REGION_TRACING
+void G1AllocRegion::trace(const char* str, size_t word_size, HeapWord* result) {
+  // All the calls to trace that set either just the size or the size
+  // and the result are considered part of level 2 tracing and are
+  // skipped during level 1 tracing.
+  if ((word_size == 0 && result == NULL) || (G1_ALLOC_REGION_TRACING > 1)) {
+    const size_t buffer_length = 128;
+    char hr_buffer[buffer_length];
+    char rest_buffer[buffer_length];
+
+    HeapRegion* alloc_region = _alloc_region;
+    if (alloc_region == NULL) {
+      jio_snprintf(hr_buffer, buffer_length, "NULL");
+    } else if (alloc_region == _dummy_region) {
+      jio_snprintf(hr_buffer, buffer_length, "DUMMY");
+    } else {
+      jio_snprintf(hr_buffer, buffer_length,
+                   HR_FORMAT, HR_FORMAT_PARAMS(alloc_region));
+    }
+
+    if (G1_ALLOC_REGION_TRACING > 1) {
+      if (result != NULL) {
+        jio_snprintf(rest_buffer, buffer_length, SIZE_FORMAT" "PTR_FORMAT,
+                     word_size, result);
+      } else if (word_size != 0) {
+        jio_snprintf(rest_buffer, buffer_length, SIZE_FORMAT, word_size);
+      } else {
+        jio_snprintf(rest_buffer, buffer_length, "");
+      }
+    } else {
+      jio_snprintf(rest_buffer, buffer_length, "");
+    }
+
+    tty->print_cr("[%s] %s : %s %s", _name, hr_buffer, str, rest_buffer);
+  }
+}
+#endif // G1_ALLOC_REGION_TRACING
+
+G1AllocRegion::G1AllocRegion(const char* name,
+                             bool bot_updates)
+  : _name(name), _bot_updates(bot_updates),
+    _alloc_region(NULL), _used_bytes_before(0) { }
+

src/share/vm/gc_implementation/g1/g1AllocRegion.hpp

+/*
+ * Copyright (c) 2011, Oracle and/or its affiliates. All rights reserved.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This code is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 only, as
+ * published by the Free Software Foundation.
+ *
+ * This code is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+ * version 2 for more details (a copy is included in the LICENSE file that
+ * accompanied this code).
+ *
+ * You should have received a copy of the GNU General Public License version
+ * 2 along with this work; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
+ * or visit www.oracle.com if you need additional information or have any
+ * questions.
+ *
+ */
+
+#ifndef SHARE_VM_GC_IMPLEMENTATION_G1_G1ALLOCREGION_HPP
+#define SHARE_VM_GC_IMPLEMENTATION_G1_G1ALLOCREGION_HPP
+
+#include "gc_implementation/g1/heapRegion.hpp"
+
+class G1CollectedHeap;
+
+// 0 -> no tracing, 1 -> basic tracing, 2 -> basic + allocation tracing
+#define G1_ALLOC_REGION_TRACING 0
+
+class ar_ext_msg;
+
+// A class that holds a region that is active in satisfying allocation
+// requests, potentially issued in parallel. When the active region is
+// full it will be retired it replaced with a new one. The
+// implementation assumes that fast-path allocations will be lock-free
+// and a lock will need to be taken when the active region needs to be
+// replaced.
+
+class G1AllocRegion VALUE_OBJ_CLASS_SPEC {
+  friend class ar_ext_msg;
+
+private:
+  // The active allocating region we are currently allocating out
+  // of. The invariant is that if this object is initialized (i.e.,
+  // init() has been called and release() has not) then _alloc_region
+  // is either an active allocating region or the dummy region (i.e.,
+  // it can never be NULL) and this object can be used to satisfy
+  // allocation requests. If this object is not initialized
+  // (i.e. init() has not been called or release() has been called)
+  // then _alloc_region is NULL and this object should not be used to
+  // satisfy allocation requests (it was done this way to force the
+  // correct use of init() and release()).
+  HeapRegion* _alloc_region;
+
+  // When we set up a new active region we save its used bytes in this
+  // field so that, when we retire it, we can calculate how much space
+  // we allocated in it.
+  size_t _used_bytes_before;
+
+  // Specifies whether the allocate calls will do BOT updates or not.
+  bool _bot_updates;
+
+  // Useful for debugging and tracing.
+  const char* _name;
+
+  // A dummy region (i.e., it's been allocated specially for this
+  // purpose and it is not part of the heap) that is full (i.e., top()
+  // == end()). When we don't have a valid active region we make
+  // _alloc_region point to this. This allows us to skip checking
+  // whether the _alloc_region is NULL or not.
+  static HeapRegion* _dummy_region;
+
+  // Some of the methods below take a bot_updates parameter. Its value
+  // should be the same as the _bot_updates field. The idea is that
+  // the parameter will be a constant for a particular alloc region
+  // and, given that these methods will be hopefully inlined, the
+  // compiler should compile out the test.
+
+  // Perform a non-MT-safe allocation out of the given region.
+  static inline HeapWord* allocate(HeapRegion* alloc_region,
+                                   size_t word_size,
+                                   bool bot_updates);
+
+  // Perform a MT-safe allocation out of the given region.
+  static inline HeapWord* par_allocate(HeapRegion* alloc_region,
+                                       size_t word_size,
+                                       bool bot_updates);
+
+  // Ensure that the region passed as a parameter has been filled up
+  // so that noone else can allocate out of it any more.
+  static void fill_up_remaining_space(HeapRegion* alloc_region,
+                                      bool bot_updates);
+
+  // Retire the active allocating region. If fill_up is true then make
+  // sure that the region is full before we retire it so that noone
+  // else can allocate out of it.
+  void retire(bool fill_up);
+
+  // Allocate a new active region and use it to perform a word_size
+  // allocation. The force parameter will be passed on to
+  // G1CollectedHeap::allocate_new_alloc_region() and tells it to try
+  // to allocate a new region even if the max has been reached.
+  HeapWord* new_alloc_region_and_allocate(size_t word_size, bool force);
+
+  void fill_in_ext_msg(ar_ext_msg* msg, const char* message);
+
+protected:
+  // For convenience as subclasses use it.
+  static G1CollectedHeap* _g1h;
+
+  virtual HeapRegion* allocate_new_region(size_t word_size, bool force) = 0;
+  virtual void retire_region(HeapRegion* alloc_region,
+                             size_t allocated_bytes) = 0;
+
+  G1AllocRegion(const char* name, bool bot_updates);
+
+public:
+  static void setup(G1CollectedHeap* g1h, HeapRegion* dummy_region);
+
+  HeapRegion* get() const {
+    // Make sure that the dummy region does not escape this class.
+    return (_alloc_region == _dummy_region) ? NULL : _alloc_region;
+  }
+
+  // The following two are the building blocks for the allocation method.
+
+  // First-level allocation: Should be called without holding a
+  // lock. It will try to allocate lock-free out of the active region,
+  // or return NULL if it was unable to.
+  inline HeapWord* attempt_allocation(size_t word_size, bool bot_updates);
+
+  // Second-level allocation: Should be called while holding a
+  // lock. It will try to first allocate lock-free out of the active
+  // region or, if it's unable to, it will try to replace the active
+  // alloc region with a new one. We require that the caller takes the
+  // appropriate lock before calling this so that it is easier to make
+  // it conform to its locking protocol.
+  inline HeapWord* attempt_allocation_locked(size_t word_size,
+                                             bool bot_updates);
+
+  // Should be called to allocate a new region even if the max of this
+  // type of regions has been reached. Should only be called if other
+  // allocation attempts have failed and we are not holding a valid
+  // active region.
+  inline HeapWord* attempt_allocation_force(size_t word_size,
+                                            bool bot_updates);
+
+  // Should be called before we start using this object.
+  void init();
+
+  // Should be called when we want to release the active region which
+  // is returned after it's been retired.
+  HeapRegion* release();
+
+#if G1_ALLOC_REGION_TRACING
+  void trace(const char* str, size_t word_size = 0, HeapWord* result = NULL);
+#else // G1_ALLOC_REGION_TRACING
+  void trace(const char* str, size_t word_size = 0, HeapWord* result = NULL) { }
+#endif // G1_ALLOC_REGION_TRACING
+};
+
+class ar_ext_msg : public err_msg {
+public:
+  ar_ext_msg(G1AllocRegion* alloc_region, const char *message) : err_msg("") {
+    alloc_region->fill_in_ext_msg(this, message);
+  }
+};
+
+#endif // SHARE_VM_GC_IMPLEMENTATION_G1_G1ALLOCREGION_HPP

src/share/vm/gc_implementation/g1/g1AllocRegion.inline.hpp

+/*
+ * Copyright (c) 2011, Oracle and/or its affiliates. All rights reserved.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This code is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 only, as
+ * published by the Free Software Foundation.
+ *
+ * This code is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+ * version 2 for more details (a copy is included in the LICENSE file that
+ * accompanied this code).
+ *
+ * You should have received a copy of the GNU General Public License version
+ * 2 along with this work; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
+ * or visit www.oracle.com if you need additional information or have any
+ * questions.
+ *
+ */
+
+#ifndef SHARE_VM_GC_IMPLEMENTATION_G1_G1ALLOCREGION_INLINE_HPP
+#define SHARE_VM_GC_IMPLEMENTATION_G1_G1ALLOCREGION_INLINE_HPP
+
+#include "gc_implementation/g1/g1AllocRegion.hpp"
+
+inline HeapWord* G1AllocRegion::allocate(HeapRegion* alloc_region,
+                                         size_t word_size,
+                                         bool bot_updates) {
+  assert(alloc_region != NULL, err_msg("pre-condition"));
+
+  if (!bot_updates) {
+    return alloc_region->allocate_no_bot_updates(word_size);
+  } else {
+    return alloc_region->allocate(word_size);
+  }
+}
+
+inline HeapWord* G1AllocRegion::par_allocate(HeapRegion* alloc_region,
+                                             size_t word_size,
+                                             bool bot_updates) {
+  assert(alloc_region != NULL, err_msg("pre-condition"));
+  assert(!alloc_region->is_empty(), err_msg("pre-condition"));
+
+  if (!bot_updates) {
+    return alloc_region->par_allocate_no_bot_updates(word_size);
+  } else {
+    return alloc_region->par_allocate(word_size);
+  }
+}
+
+inline HeapWord* G1AllocRegion::attempt_allocation(size_t word_size,
+                                                   bool bot_updates) {
+  assert(bot_updates == _bot_updates, ar_ext_msg(this, "pre-condition"));
+
+  HeapRegion* alloc_region = _alloc_region;
+  assert(alloc_region != NULL, ar_ext_msg(this, "not initialized properly"));
+
+  HeapWord* result = par_allocate(alloc_region, word_size, bot_updates);
+  if (result != NULL) {
+    trace("alloc", word_size, result);
+    return result;
+  }
+  trace("alloc failed", word_size);
+  return NULL;
+}
+
+inline HeapWord* G1AllocRegion::attempt_allocation_locked(size_t word_size,
+                                                          bool bot_updates) {
+  // First we have to tedo the allocation, assuming we're holding the
+  // appropriate lock, in case another thread changed the region while
+  // we were waiting to get the lock.
+  HeapWord* result = attempt_allocation(word_size, bot_updates);
+  if (result != NULL) {
+    return result;
+  }
+
+  retire(true /* fill_up */);
+  result = new_alloc_region_and_allocate(word_size, false /* force */);
+  if (result != NULL) {
+    trace("alloc locked (second attempt)", word_size, result);
+    return result;
+  }
+  trace("alloc locked failed", word_size);
+  return NULL;
+}
+
+inline HeapWord* G1AllocRegion::attempt_allocation_force(size_t word_size,
+                                                         bool bot_updates) {
+  assert(bot_updates == _bot_updates, ar_ext_msg(this, "pre-condition"));
+  assert(_alloc_region != NULL, ar_ext_msg(this, "not initialized properly"));
+
+  trace("forcing alloc");
+  HeapWord* result = new_alloc_region_and_allocate(word_size, true /* force */);
+  if (result != NULL) {
+    trace("alloc forced", word_size, result);
+    return result;
+  }
+  trace("alloc forced failed", word_size);
+  return NULL;
+}
+
+#endif // SHARE_VM_GC_IMPLEMENTATION_G1_G1ALLOCREGION_INLINE_HPP

src/share/vm/gc_implementation/g1/g1CollectedHeap.hpp

@@ -26,6 +26,7 @@
 #define SHARE_VM_GC_IMPLEMENTATION_G1_G1COLLECTEDHEAP_HPP
 
 #include "gc_implementation/g1/concurrentMark.hpp"
+#include "gc_implementation/g1/g1AllocRegion.hpp"
 #include "gc_implementation/g1/g1RemSet.hpp"
 #include "gc_implementation/g1/heapRegionSets.hpp"
 #include "gc_implementation/parNew/parGCAllocBuffer.hpp"
@@ -128,6 +129,15 @@ public:
   void          print();
 };
 
+class MutatorAllocRegion : public G1AllocRegion {
+protected:
+  virtual HeapRegion* allocate_new_region(size_t word_size, bool force);
+  virtual void retire_region(HeapRegion* alloc_region, size_t allocated_bytes);
+public:
+  MutatorAllocRegion()
+    : G1AllocRegion("Mutator Alloc Region", false /* bot_updates */) { }
+};
+
 class RefineCardTableEntryClosure;
 class G1CollectedHeap : public SharedHeap {
   friend class VM_G1CollectForAllocation;
@@ -135,6 +145,7 @@ class G1CollectedHeap : public SharedHeap {
   friend class VM_G1CollectFull;
   friend class VM_G1IncCollectionPause;
   friend class VMStructs;
+  friend class MutatorAllocRegion;
 
   // Closures used in implementation.
   friend class G1ParCopyHelper;
@@ -197,12 +208,15 @@ private:
   // The sequence of all heap regions in the heap.
   HeapRegionSeq* _hrs;
 
-  // The region from which normal-sized objects are currently being
-  // allocated.  May be NULL.
-  HeapRegion* _cur_alloc_region;
+  // Alloc region used to satisfy mutator allocation requests.
+  MutatorAllocRegion _mutator_alloc_region;
+
+  // It resets the mutator alloc region before new allocations can take place.
+  void init_mutator_alloc_region();
+
+  // It releases the mutator alloc region.
+  void release_mutator_alloc_region();
 
-  // Postcondition: cur_alloc_region == NULL.
-  void abandon_cur_alloc_region();
   void abandon_gc_alloc_regions();
 
   // The to-space memory regions into which objects are being copied during
@@ -360,27 +374,21 @@ protected:
   G1CollectorPolicy* _g1_policy;
 
   // This is the second level of trying to allocate a new region. If
-  // new_region_work didn't find a region in the free_list, this call
-  // will check whether there's anything available in the
-  // secondary_free_list and/or wait for more regions to appear in that
-  // list, if _free_regions_coming is set.
+  // new_region() didn't find a region on the free_list, this call will
+  // check whether there's anything available on the
+  // secondary_free_list and/or wait for more regions to appear on
+  // that list, if _free_regions_coming is set.
   HeapRegion* new_region_try_secondary_free_list();
 
   // Try to allocate a single non-humongous HeapRegion sufficient for
   // an allocation of the given word_size. If do_expand is true,
   // attempt to expand the heap if necessary to satisfy the allocation
   // request.
-  HeapRegion* new_region_work(size_t word_size, bool do_expand);
+  HeapRegion* new_region(size_t word_size, bool do_expand);
 
-  // Try to allocate a new region to be used for allocation by a
-  // mutator thread. Attempt to expand the heap if no region is
+  // Try to allocate a new region to be used for allocation by
+  // a GC thread. It will try to expand the heap if no region is
   // available.
-  HeapRegion* new_alloc_region(size_t word_size) {
-    return new_region_work(word_size, false /* do_expand */);
-  }
-
-  // Try to allocate a new region to be used for allocation by a GC
-  // thread. Attempt to expand the heap if no region is available.
   HeapRegion* new_gc_alloc_region(int purpose, size_t word_size);
 
   // Attempt to satisfy a humongous allocation request of the given
@@ -415,10 +423,6 @@ protected:
   // * All non-TLAB allocation requests should go to mem_allocate()
   //   and mem_allocate() should never be called with is_tlab == true.
   //
-  // * If the GC locker is active we currently stall until we can
-  //   allocate a new young region. This will be changed in the
-  //   near future (see CR 6994056).
-  //
   // * If either call cannot satisfy the allocation request using the
   //   current allocating region, they will try to get a new one. If
   //   this fails, they will attempt to do an evacuation pause and
@@ -441,122 +445,38 @@ protected:
                                  bool   is_tlab, /* expected to be false */
                                  bool*  gc_overhead_limit_was_exceeded);
 
-  // The following methods, allocate_from_cur_allocation_region(),
-  // attempt_allocation(), attempt_allocation_locked(),
-  // replace_cur_alloc_region_and_allocate(),
-  // attempt_allocation_slow(), and attempt_allocation_humongous()
-  // have very awkward pre- and post-conditions with respect to
-  // locking:
-  //
-  // If they are called outside a safepoint they assume the caller
-  // holds the Heap_lock when it calls them. However, on exit they
-  // will release the Heap_lock if they return a non-NULL result, but
-  // keep holding the Heap_lock if they return a NULL result. The
-  // reason for this is that we need to dirty the cards that span
-  // allocated blocks on young regions to avoid having to take the
-  // slow path of the write barrier (for performance reasons we don't
-  // update RSets for references whose source is a young region, so we
-  // don't need to look at dirty cards on young regions). But, doing
-  // this card dirtying while holding the Heap_lock can be a
-  // scalability bottleneck, especially given that some allocation
-  // requests might be of non-trivial size (and the larger the region
-  // size is, the fewer allocations requests will be considered
-  // humongous, as the humongous size limit is a fraction of the
-  // region size). So, when one of these calls succeeds in allocating
-  // a block it does the card dirtying after it releases the Heap_lock
-  // which is why it will return without holding it.
-  //
-  // The above assymetry is the reason why locking / unlocking is done
-  // explicitly (i.e., with Heap_lock->lock() and
-  // Heap_lock->unlocked()) instead of using MutexLocker and
-  // MutexUnlocker objects. The latter would ensure that the lock is
-  // unlocked / re-locked at every possible exit out of the basic
-  // block. However, we only want that action to happen in selected
-  // places.
-  //
-  // Further, if the above methods are called during a safepoint, then
-  // naturally there's no assumption about the Heap_lock being held or
-  // there's no attempt to unlock it. The parameter at_safepoint
-  // indicates whether the call is made during a safepoint or not (as
-  // an optimization, to avoid reading the global flag with
-  // SafepointSynchronize::is_at_safepoint()).
-  //
-  // The methods share these parameters:
-  //
-  // * word_size     : the size of the allocation request in words
-  // * at_safepoint  : whether the call is done at a safepoint; this
-  //                   also determines whether a GC is permitted
-  //                   (at_safepoint == false) or not (at_safepoint == true)
-  // * do_dirtying   : whether the method should dirty the allocated
-  //                   block before returning
-  //
-  // They all return either the address of the block, if they
-  // successfully manage to allocate it, or NULL.
-
-  // It tries to satisfy an allocation request out of the current
-  // alloc region, which is passed as a parameter. It assumes that the
-  // caller has checked that the current alloc region is not NULL.
-  // Given that the caller has to check the current alloc region for
-  // at least NULL, it might as well pass it as the first parameter so
-  // that the method doesn't have to read it from the
-  // _cur_alloc_region field again. It is called from both
-  // attempt_allocation() and attempt_allocation_locked() and the
-  // with_heap_lock parameter indicates whether the caller was holding
-  // the heap lock when it called it or not.
-  inline HeapWord* allocate_from_cur_alloc_region(HeapRegion* cur_alloc_region,
-                                                  size_t word_size,
-                                                  bool with_heap_lock);
-
-  // First-level of allocation slow path: it attempts to allocate out
-  // of the current alloc region in a lock-free manner using a CAS. If
-  // that fails it takes the Heap_lock and calls
-  // attempt_allocation_locked() for the second-level slow path.
-  inline HeapWord* attempt_allocation(size_t word_size);
-
-  // Second-level of allocation slow path: while holding the Heap_lock
-  // it tries to allocate out of the current alloc region and, if that
-  // fails, tries to allocate out of a new current alloc region.
-  inline HeapWord* attempt_allocation_locked(size_t word_size);
-
-  // It assumes that the current alloc region has been retired and
-  // tries to allocate a new one. If it's successful, it performs the
-  // allocation out of the new current alloc region and updates
-  // _cur_alloc_region. Normally, it would try to allocate a new
-  // region if the young gen is not full, unless can_expand is true in
-  // which case it would always try to allocate a new region.
-  HeapWord* replace_cur_alloc_region_and_allocate(size_t word_size,
-                                                  bool at_safepoint,
-                                                  bool do_dirtying,
-                                                  bool can_expand);
-
-  // Third-level of allocation slow path: when we are unable to
-  // allocate a new current alloc region to satisfy an allocation
-  // request (i.e., when attempt_allocation_locked() fails). It will
-  // try to do an evacuation pause, which might stall due to the GC
-  // locker, and retry the allocation attempt when appropriate.
-  HeapWord* attempt_allocation_slow(size_t word_size);
-
-  // The method that tries to satisfy a humongous allocation
-  // request. If it cannot satisfy it it will try to do an evacuation
-  // pause to perhaps reclaim enough space to be able to satisfy the
-  // allocation request afterwards.
+  // The following three methods take a gc_count_before_ret
+  // parameter which is used to return the GC count if the method
+  // returns NULL. Given that we are required to read the GC count
+  // while holding the Heap_lock, and these paths will take the
+  // Heap_lock at some point, it's easier to get them to read the GC
+  // count while holding the Heap_lock before they return NULL instead
+  // of the caller (namely: mem_allocate()) having to also take the
+  // Heap_lock just to read the GC count.
+
+  // First-level mutator allocation attempt: try to allocate out of
+  // the mutator alloc region without taking the Heap_lock. This
+  // should only be used for non-humongous allocations.
+  inline HeapWord* attempt_allocation(size_t word_size,
+                                      unsigned int* gc_count_before_ret);
+
+  // Second-level mutator allocation attempt: take the Heap_lock and
+  // retry the allocation attempt, potentially scheduling a GC
+  // pause. This should only be used for non-humongous allocations.
+  HeapWord* attempt_allocation_slow(size_t word_size,
+                                    unsigned int* gc_count_before_ret);
+
+  // Takes the Heap_lock and attempts a humongous allocation. It can
+  // potentially schedule a GC pause.
   HeapWord* attempt_allocation_humongous(size_t word_size,
-                                         bool at_safepoint);
+                                         unsigned int* gc_count_before_ret);
 
-  // It does the common work when we are retiring the current alloc region.
-  inline void retire_cur_alloc_region_common(HeapRegion* cur_alloc_region);
-
-  // It retires the current alloc region, which is passed as a
-  // parameter (since, typically, the caller is already holding on to
-  // it). It sets _cur_alloc_region to NULL.
-  void retire_cur_alloc_region(HeapRegion* cur_alloc_region);
-
-  // It attempts to do an allocation immediately before or after an
-  // evacuation pause and can only be called by the VM thread. It has
-  // slightly different assumptions that the ones before (i.e.,
-  // assumes that the current alloc region has been retired).
+  // Allocation attempt that should be called during safepoints (e.g.,
+  // at the end of a successful GC). expect_null_mutator_alloc_region
+  // specifies whether the mutator alloc region is expected to be NULL
+  // or not.
   HeapWord* attempt_allocation_at_safepoint(size_t word_size,
-                                            bool expect_null_cur_alloc_region);
+                                       bool expect_null_mutator_alloc_region);
 
   // It dirties the cards that cover the block so that so that the post
   // write barrier never queues anything when updating objects on this
@@ -583,6 +503,12 @@ protected:
   // GC pause.
   void  retire_alloc_region(HeapRegion* alloc_region, bool par);
 
+  // These two methods are the "callbacks" from the G1AllocRegion class.
+
+  HeapRegion* new_mutator_alloc_region(size_t word_size, bool force);
+  void retire_mutator_alloc_region(HeapRegion* alloc_region,
+                                   size_t allocated_bytes);
+
   // - if explicit_gc is true, the GC is for a System.gc() or a heap
   //   inspection request and should collect the entire heap
   // - if clear_all_soft_refs is true, all soft references should be
@@ -1027,6 +953,9 @@ public:
   // The number of regions available for "regular" expansion.
   size_t expansion_regions() { return _expansion_regions; }
 
+  void verify_dirty_young_list(HeapRegion* head) PRODUCT_RETURN;
+  void verify_dirty_young_regions() PRODUCT_RETURN;
+
   // verify_region_sets() performs verification over the region
   // lists. It will be compiled in the product code to be used when
   // necessary (i.e., during heap verification).

src/share/vm/gc_implementation/g1/g1CollectedHeap.inline.hpp

@@ -27,6 +27,7 @@
 
 #include "gc_implementation/g1/concurrentMark.hpp"
 #include "gc_implementation/g1/g1CollectedHeap.hpp"
+#include "gc_implementation/g1/g1AllocRegion.inline.hpp"
 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
 #include "utilities/taskqueue.hpp"
@@ -59,131 +60,23 @@ inline bool G1CollectedHeap::obj_in_cs(oop obj) {
   return r != NULL && r->in_collection_set();
 }
 
-// See the comment in the .hpp file about the locking protocol and
-// assumptions of this method (and other related ones).
 inline HeapWord*
-G1CollectedHeap::allocate_from_cur_alloc_region(HeapRegion* cur_alloc_region,
-                                                size_t word_size,
-                                                bool with_heap_lock) {
-  assert_not_at_safepoint();
-  assert(with_heap_lock == Heap_lock->owned_by_self(),
-         "with_heap_lock and Heap_lock->owned_by_self() should be a tautology");
-  assert(cur_alloc_region != NULL, "pre-condition of the method");
-  assert(cur_alloc_region->is_young(),
-         "we only support young current alloc regions");
-  assert(!isHumongous(word_size), "allocate_from_cur_alloc_region() "
-         "should not be used for humongous allocations");
-  assert(!cur_alloc_region->isHumongous(), "Catch a regression of this bug.");
-
-  assert(!cur_alloc_region->is_empty(),
-         err_msg("region ["PTR_FORMAT","PTR_FORMAT"] should not be empty",
-                 cur_alloc_region->bottom(), cur_alloc_region->end()));
-  HeapWord* result = cur_alloc_region->par_allocate_no_bot_updates(word_size);
-  if (result != NULL) {
-    assert(is_in(result), "result should be in the heap");
-
-    if (with_heap_lock) {
-      Heap_lock->unlock();
-    }
-    assert_heap_not_locked();
-    // Do the dirtying after we release the Heap_lock.
-    dirty_young_block(result, word_size);
-    return result;
-  }
-
-  if (with_heap_lock) {
-    assert_heap_locked();
-  } else {
-    assert_heap_not_locked();
-  }
-  return NULL;
-}
-
-// See the comment in the .hpp file about the locking protocol and
-// assumptions of this method (and other related ones).
-inline HeapWord*
-G1CollectedHeap::attempt_allocation(size_t word_size) {
+G1CollectedHeap::attempt_allocation(size_t word_size,
+                                    unsigned int* gc_count_before_ret) {
   assert_heap_not_locked_and_not_at_safepoint();
-  assert(!isHumongous(word_size), "attempt_allocation() should not be called "
-         "for humongous allocation requests");
-
-  HeapRegion* cur_alloc_region = _cur_alloc_region;
-  if (cur_alloc_region != NULL) {
-    HeapWord* result = allocate_from_cur_alloc_region(cur_alloc_region,
-                                                   word_size,
-                                                   false /* with_heap_lock */);
-    assert_heap_not_locked();
-    if (result != NULL) {
-      return result;
-    }
-  }
+  assert(!isHumongous(word_size), "attempt_allocation() should not "
+         "be called for humongous allocation requests");
 
-  // Our attempt to allocate lock-free failed as the current
-  // allocation region is either NULL or full. So, we'll now take the
-  // Heap_lock and retry.
-  Heap_lock->lock();
-
-  HeapWord* result = attempt_allocation_locked(word_size);
-  if (result != NULL) {
-    assert_heap_not_locked();
-    return result;
+  HeapWord* result = _mutator_alloc_region.attempt_allocation(word_size,
+                                                      false /* bot_updates */);
+  if (result == NULL) {
+    result = attempt_allocation_slow(word_size, gc_count_before_ret);
   }
-
-  assert_heap_locked();
-  return NULL;
-}
-
-inline void
-G1CollectedHeap::retire_cur_alloc_region_common(HeapRegion* cur_alloc_region) {
-  assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
-  assert(cur_alloc_region != NULL && cur_alloc_region == _cur_alloc_region,
-         "pre-condition of the call");
-  assert(cur_alloc_region->is_young(),
-         "we only support young current alloc regions");
-
-  // The region is guaranteed to be young
-  g1_policy()->add_region_to_incremental_cset_lhs(cur_alloc_region);
-  _summary_bytes_used += cur_alloc_region->used();
-  _cur_alloc_region = NULL;
-}
-
-inline HeapWord*
-G1CollectedHeap::attempt_allocation_locked(size_t word_size) {
-  assert_heap_locked_and_not_at_safepoint();
-  assert(!isHumongous(word_size), "attempt_allocation_locked() "
-         "should not be called for humongous allocation requests");
-
-  // First, reread the current alloc region and retry the allocation
-  // in case somebody replaced it while we were waiting to get the
-  // Heap_lock.
-  HeapRegion* cur_alloc_region = _cur_alloc_region;
-  if (cur_alloc_region != NULL) {
-    HeapWord* result = allocate_from_cur_alloc_region(
-                                                  cur_alloc_region, word_size,
-                                                  true /* with_heap_lock */);
-    if (result != NULL) {
   assert_heap_not_locked();
-      return result;
-    }
-
-    // We failed to allocate out of the current alloc region, so let's
-    // retire it before getting a new one.
-    retire_cur_alloc_region(cur_alloc_region);
-  }
-
-  assert_heap_locked();
-  // Try to get a new region and allocate out of it
-  HeapWord* result = replace_cur_alloc_region_and_allocate(word_size,
-                                                     false, /* at_safepoint */
-                                                     true,  /* do_dirtying */
-                                                     false  /* can_expand */);
   if (result != NULL) {
-    assert_heap_not_locked();
-    return result;
+    dirty_young_block(result, word_size);
   }
-
-  assert_heap_locked();
-  return NULL;
+  return result;
 }
 
 // It dirties the cards that cover the block so that so that the post

src/share/vm/gc_implementation/g1/heapRegion.cpp

@@ -360,6 +360,7 @@ void HeapRegion::hr_clear(bool par, bool clear_space) {
   set_young_index_in_cset(-1);
   uninstall_surv_rate_group();
   set_young_type(NotYoung);
+  reset_pre_dummy_top();
 
   if (!par) {
     // If this is parallel, this will be done later.
@@ -923,11 +924,11 @@ void G1OffsetTableContigSpace::set_saved_mark() {
     ContiguousSpace::set_saved_mark();
     OrderAccess::storestore();
     _gc_time_stamp = curr_gc_time_stamp;
-    // The following fence is to force a flush of the writes above, but
-    // is strictly not needed because when an allocating worker thread
-    // calls set_saved_mark() it does so under the ParGCRareEvent_lock;
-    // when the lock is released, the write will be flushed.
-    // OrderAccess::fence();
+    // No need to do another barrier to flush the writes above. If
+    // this is called in parallel with other threads trying to
+    // allocate into the region, the caller should call this while
+    // holding a lock and when the lock is released the writes will be
+    // flushed.
   }
 }
 

src/share/vm/gc_implementation/g1/heapRegion.hpp

@@ -149,6 +149,13 @@ class G1OffsetTableContigSpace: public ContiguousSpace {
   G1BlockOffsetArrayContigSpace _offsets;
   Mutex _par_alloc_lock;
   volatile unsigned _gc_time_stamp;
+  // When we need to retire an allocation region, while other threads
+  // are also concurrently trying to allocate into it, we typically
+  // allocate a dummy object at the end of the region to ensure that
+  // no more allocations can take place in it. However, sometimes we
+  // want to know where the end of the last "real" object we allocated
+  // into the region was and this is what this keeps track.
+  HeapWord* _pre_dummy_top;
 
  public:
   // Constructor.  If "is_zeroed" is true, the MemRegion "mr" may be
@@ -163,6 +170,17 @@ class G1OffsetTableContigSpace: public ContiguousSpace {
   virtual void set_saved_mark();
   void reset_gc_time_stamp() { _gc_time_stamp = 0; }
 
+  // See the comment above in the declaration of _pre_dummy_top for an
+  // explanation of what it is.
+  void set_pre_dummy_top(HeapWord* pre_dummy_top) {
+    assert(is_in(pre_dummy_top) && pre_dummy_top <= top(), "pre-condition");
+    _pre_dummy_top = pre_dummy_top;
+  }
+  HeapWord* pre_dummy_top() {
+    return (_pre_dummy_top == NULL) ? top() : _pre_dummy_top;
+  }
+  void reset_pre_dummy_top() { _pre_dummy_top = NULL; }
+
   virtual void initialize(MemRegion mr, bool clear_space, bool mangle_space);
   virtual void clear(bool mangle_space);
 

src/share/vm/gc_implementation/g1/heapRegion.inline.hpp

@@ -38,15 +38,8 @@ inline HeapWord* G1OffsetTableContigSpace::allocate(size_t size) {
 // this is used for larger LAB allocations only.
 inline HeapWord* G1OffsetTableContigSpace::par_allocate(size_t size) {
   MutexLocker x(&_par_alloc_lock);
-  // This ought to be just "allocate", because of the lock above, but that
-  // ContiguousSpace::allocate asserts that either the allocating thread
-  // holds the heap lock or it is the VM thread and we're at a safepoint.
-  // The best I (dld) could figure was to put a field in ContiguousSpace
-  // meaning "locking at safepoint taken care of", and set/reset that
-  // here.  But this will do for now, especially in light of the comment
-  // above.  Perhaps in the future some lock-free manner of keeping the
-  // coordination.
-  HeapWord* res = ContiguousSpace::par_allocate(size);
+  // Given that we take the lock no need to use par_allocate() here.
+  HeapWord* res = ContiguousSpace::allocate(size);
   if (res != NULL) {
     _offsets.alloc_block(res, size);
   }

src/share/vm/memory/cardTableModRefBS.hpp

@@ -382,6 +382,11 @@ public:
     return (addr_for(pcard) == p);
   }
 
+  HeapWord* align_to_card_boundary(HeapWord* p) {
+    jbyte* pcard = byte_for(p + card_size_in_words - 1);
+    return addr_for(pcard);
+  }
+
   // The kinds of precision a CardTableModRefBS may offer.
   enum PrecisionStyle {
     Precise,

src/share/vm/memory/space.cpp

@@ -818,9 +818,14 @@ size_t ContiguousSpace::block_size(const HeapWord* p) const {
 // This version requires locking.
 inline HeapWord* ContiguousSpace::allocate_impl(size_t size,
                                                 HeapWord* const end_value) {
+  // In G1 there are places where a GC worker can allocates into a
+  // region using this serial allocation code without being prone to a
+  // race with other GC workers (we ensure that no other GC worker can
+  // access the same region at the same time). So the assert below is
+  // too strong in the case of G1.
   assert(Heap_lock->owned_by_self() ||
          (SafepointSynchronize::is_at_safepoint() &&
-          Thread::current()->is_VM_thread()),
+                               (Thread::current()->is_VM_thread() || UseG1GC)),
          "not locked");
   HeapWord* obj = top();
   if (pointer_delta(end_value, obj) >= size) {