1/*
2 * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25#include "precompiled.hpp"
26#include "gc/g1/collectionSetChooser.hpp"
27#include "gc/g1/g1CollectedHeap.inline.hpp"
28#include "gc/shared/space.inline.hpp"
29#include "runtime/atomic.hpp"
30
31// Even though we don't use the GC efficiency in our heuristics as
32// much as we used to, we still order according to GC efficiency. This
33// will cause regions with a lot of live objects and large RSets to
34// end up at the end of the array. Given that we might skip collecting
35// the last few old regions, if after a few mixed GCs the remaining
36// have reclaimable bytes under a certain threshold, the hope is that
37// the ones we'll skip are ones with both large RSets and a lot of
38// live objects, not the ones with just a lot of live objects if we
39// ordered according to the amount of reclaimable bytes per region.
40static int order_regions(HeapRegion* hr1, HeapRegion* hr2) {
41  if (hr1 == NULL) {
42    if (hr2 == NULL) {
43      return 0;
44    } else {
45      return 1;
46    }
47  } else if (hr2 == NULL) {
48    return -1;
49  }
50
51  double gc_eff1 = hr1->gc_efficiency();
52  double gc_eff2 = hr2->gc_efficiency();
53  if (gc_eff1 > gc_eff2) {
54    return -1;
55  } if (gc_eff1 < gc_eff2) {
56    return 1;
57  } else {
58    return 0;
59  }
60}
61
62static int order_regions(HeapRegion** hr1p, HeapRegion** hr2p) {
63  return order_regions(*hr1p, *hr2p);
64}
65
66CollectionSetChooser::CollectionSetChooser() :
67  // The line below is the worst bit of C++ hackery I've ever written
68  // (Detlefs, 11/23).  You should think of it as equivalent to
69  // "_regions(100, true)": initialize the growable array and inform it
70  // that it should allocate its elem array(s) on the C heap.
71  //
72  // The first argument, however, is actually a comma expression
73  // (set_allocation_type(this, C_HEAP), 100). The purpose of the
74  // set_allocation_type() call is to replace the default allocation
75  // type for embedded objects STACK_OR_EMBEDDED with C_HEAP. It will
76  // allow to pass the assert in GenericGrowableArray() which checks
77  // that a growable array object must be on C heap if elements are.
78  //
79  // Note: containing object is allocated on C heap since it is CHeapObj.
80  //
81  _regions((ResourceObj::set_allocation_type((address) &_regions,
82                                             ResourceObj::C_HEAP),
83                  100), true /* C_Heap */),
84    _front(0), _end(0), _first_par_unreserved_idx(0),
85    _region_live_threshold_bytes(0), _remaining_reclaimable_bytes(0) {
86  _region_live_threshold_bytes =
87    HeapRegion::GrainBytes * (size_t) G1MixedGCLiveThresholdPercent / 100;
88}
89
90#ifndef PRODUCT
91void CollectionSetChooser::verify() {
92  guarantee(_end <= regions_length(), "_end: %u regions length: %u", _end, regions_length());
93  guarantee(_front <= _end, "_front: %u _end: %u", _front, _end);
94  uint index = 0;
95  size_t sum_of_reclaimable_bytes = 0;
96  while (index < _front) {
97    guarantee(regions_at(index) == NULL,
98              "all entries before _front should be NULL");
99    index += 1;
100  }
101  HeapRegion *prev = NULL;
102  while (index < _end) {
103    HeapRegion *curr = regions_at(index++);
104    guarantee(curr != NULL, "Regions in _regions array cannot be NULL");
105    guarantee(!curr->is_young(), "should not be young!");
106    guarantee(!curr->is_pinned(),
107              "Pinned region should not be in collection set (index %u)", curr->hrm_index());
108    if (prev != NULL) {
109      guarantee(order_regions(prev, curr) != 1,
110                "GC eff prev: %1.4f GC eff curr: %1.4f",
111                prev->gc_efficiency(), curr->gc_efficiency());
112    }
113    sum_of_reclaimable_bytes += curr->reclaimable_bytes();
114    prev = curr;
115  }
116  guarantee(sum_of_reclaimable_bytes == _remaining_reclaimable_bytes,
117            "reclaimable bytes inconsistent, "
118            "remaining: " SIZE_FORMAT " sum: " SIZE_FORMAT,
119            _remaining_reclaimable_bytes, sum_of_reclaimable_bytes);
120}
121#endif // !PRODUCT
122
123void CollectionSetChooser::sort_regions() {
124  // First trim any unused portion of the top in the parallel case.
125  if (_first_par_unreserved_idx > 0) {
126    assert(_first_par_unreserved_idx <= regions_length(),
127           "Or we didn't reserved enough length");
128    regions_trunc_to(_first_par_unreserved_idx);
129  }
130  _regions.sort(order_regions);
131  assert(_end <= regions_length(), "Requirement");
132#ifdef ASSERT
133  for (uint i = 0; i < _end; i++) {
134    assert(regions_at(i) != NULL, "Should be true by sorting!");
135  }
136#endif // ASSERT
137  if (log_is_enabled(Trace, gc, liveness)) {
138    G1PrintRegionLivenessInfoClosure cl("Post-Sorting");
139    for (uint i = 0; i < _end; ++i) {
140      HeapRegion* r = regions_at(i);
141      cl.doHeapRegion(r);
142    }
143  }
144  verify();
145}
146
147void CollectionSetChooser::add_region(HeapRegion* hr) {
148  assert(!hr->is_pinned(),
149         "Pinned region shouldn't be added to the collection set (index %u)", hr->hrm_index());
150  assert(!hr->is_young(), "should not be young!");
151  _regions.append(hr);
152  _end++;
153  _remaining_reclaimable_bytes += hr->reclaimable_bytes();
154  hr->calc_gc_efficiency();
155}
156
157void CollectionSetChooser::push(HeapRegion* hr) {
158  assert(hr != NULL, "Can't put back a NULL region");
159  assert(_front >= 1, "Too many regions have been put back");
160  _front--;
161  regions_at_put(_front, hr);
162  _remaining_reclaimable_bytes += hr->reclaimable_bytes();
163}
164
165void CollectionSetChooser::prepare_for_par_region_addition(uint n_threads,
166                                                           uint n_regions,
167                                                           uint chunk_size) {
168  _first_par_unreserved_idx = 0;
169  uint max_waste = n_threads * chunk_size;
170  // it should be aligned with respect to chunk_size
171  uint aligned_n_regions = (n_regions + chunk_size - 1) / chunk_size * chunk_size;
172  assert(aligned_n_regions % chunk_size == 0, "should be aligned");
173  regions_at_put_grow(aligned_n_regions + max_waste - 1, NULL);
174}
175
176uint CollectionSetChooser::claim_array_chunk(uint chunk_size) {
177  uint res = (uint) Atomic::add((jint) chunk_size,
178                                (volatile jint*) &_first_par_unreserved_idx);
179  assert(regions_length() > res + chunk_size - 1,
180         "Should already have been expanded");
181  return res - chunk_size;
182}
183
184void CollectionSetChooser::set_region(uint index, HeapRegion* hr) {
185  assert(regions_at(index) == NULL, "precondition");
186  assert(!hr->is_young(), "should not be young!");
187  regions_at_put(index, hr);
188  hr->calc_gc_efficiency();
189}
190
191void CollectionSetChooser::update_totals(uint region_num,
192                                         size_t reclaimable_bytes) {
193  // Only take the lock if we actually need to update the totals.
194  if (region_num > 0) {
195    assert(reclaimable_bytes > 0, "invariant");
196    // We could have just used atomics instead of taking the
197    // lock. However, we currently don't have an atomic add for size_t.
198    MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
199    _end += region_num;
200    _remaining_reclaimable_bytes += reclaimable_bytes;
201  } else {
202    assert(reclaimable_bytes == 0, "invariant");
203  }
204}
205
206void CollectionSetChooser::clear() {
207  _regions.clear();
208  _front = 0;
209  _end = 0;
210  _remaining_reclaimable_bytes = 0;
211}
212
213class ParKnownGarbageHRClosure: public HeapRegionClosure {
214  G1CollectedHeap* _g1h;
215  CSetChooserParUpdater _cset_updater;
216
217public:
218  ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
219                           uint chunk_size) :
220    _g1h(G1CollectedHeap::heap()),
221    _cset_updater(hrSorted, true /* parallel */, chunk_size) { }
222
223  bool doHeapRegion(HeapRegion* r) {
224    // Do we have any marking information for this region?
225    if (r->is_marked()) {
226      // We will skip any region that's currently used as an old GC
227      // alloc region (we should not consider those for collection
228      // before we fill them up).
229      if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
230        _cset_updater.add_region(r);
231      }
232    }
233    return false;
234  }
235};
236
237class ParKnownGarbageTask: public AbstractGangTask {
238  CollectionSetChooser* _hrSorted;
239  uint _chunk_size;
240  G1CollectedHeap* _g1;
241  HeapRegionClaimer _hrclaimer;
242
243public:
244  ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) :
245      AbstractGangTask("ParKnownGarbageTask"),
246      _hrSorted(hrSorted), _chunk_size(chunk_size),
247      _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {}
248
249  void work(uint worker_id) {
250    ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size);
251    _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer);
252  }
253};
254
255uint CollectionSetChooser::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const {
256  assert(n_workers > 0, "Active gc workers should be greater than 0");
257  const uint overpartition_factor = 4;
258  const uint min_chunk_size = MAX2(n_regions / n_workers, 1U);
259  return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size);
260}
261
262void CollectionSetChooser::rebuild(WorkGang* workers, uint n_regions) {
263  clear();
264
265  uint n_workers = workers->active_workers();
266
267  uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions);
268  prepare_for_par_region_addition(n_workers, n_regions, chunk_size);
269
270  ParKnownGarbageTask par_known_garbage_task(this, chunk_size, n_workers);
271  workers->run_task(&par_known_garbage_task);
272
273  sort_regions();
274}
275