49 return -1; |
50 return -1; |
50 } |
51 } |
51 |
52 |
52 double gc_eff1 = hr1->gc_efficiency(); |
53 double gc_eff1 = hr1->gc_efficiency(); |
53 double gc_eff2 = hr2->gc_efficiency(); |
54 double gc_eff2 = hr2->gc_efficiency(); |
|
55 |
54 if (gc_eff1 > gc_eff2) { |
56 if (gc_eff1 > gc_eff2) { |
55 return -1; |
57 return -1; |
56 } if (gc_eff1 < gc_eff2) { |
58 } if (gc_eff1 < gc_eff2) { |
57 return 1; |
59 return 1; |
58 } else { |
60 } else { |
59 return 0; |
61 return 0; |
60 } |
62 } |
61 } |
63 } |
62 |
64 |
63 static int order_regions(HeapRegion** hr1p, HeapRegion** hr2p) { |
65 // Determine collection set candidates: For all regions determine whether they |
64 return order_regions(*hr1p, *hr2p); |
66 // should be a collection set candidates, calculate their efficiency, sort and |
65 } |
67 // return them as G1CollectionSetCandidates instance. |
66 |
68 // Threads calculate the GC efficiency of the regions they get to process, and |
67 CollectionSetChooser::CollectionSetChooser() : |
69 // put them into some work area unsorted. At the end the array is sorted and |
68 // The line below is the worst bit of C++ hackery I've ever written |
70 // copied into the G1CollectionSetCandidates instance; the caller will be the new |
69 // (Detlefs, 11/23). You should think of it as equivalent to |
71 // owner of this object. |
70 // "_regions(100, true)": initialize the growable array and inform it |
72 class G1BuildCandidateRegionsTask : public AbstractGangTask { |
71 // that it should allocate its elem array(s) on the C heap. |
73 |
72 // |
74 // Work area for building the set of collection set candidates. Contains references |
73 // The first argument, however, is actually a comma expression |
75 // to heap regions with their GC efficiencies calculated. To reduce contention |
74 // (set_allocation_type(this, C_HEAP), 100). The purpose of the |
76 // on claiming array elements, worker threads claim parts of this array in chunks; |
75 // set_allocation_type() call is to replace the default allocation |
77 // Array elements may be NULL as threads might not get enough regions to fill |
76 // type for embedded objects STACK_OR_EMBEDDED with C_HEAP. It will |
78 // up their chunks completely. |
77 // allow to pass the assert in GenericGrowableArray() which checks |
79 // Final sorting will remove them. |
78 // that a growable array object must be on C heap if elements are. |
80 class G1BuildCandidateArray : public StackObj { |
79 // |
81 |
80 // Note: containing object is allocated on C heap since it is CHeapObj. |
82 uint const _max_size; |
81 // |
83 uint const _chunk_size; |
82 _regions((ResourceObj::set_allocation_type((address) &_regions, |
84 |
83 ResourceObj::C_HEAP), |
85 HeapRegion** _data; |
84 100), true /* C_Heap */), |
86 |
85 _front(0), _end(0), _first_par_unreserved_idx(0), |
87 uint volatile _cur_claim_idx; |
86 _region_live_threshold_bytes(0), _remaining_reclaimable_bytes(0) { |
88 |
87 _region_live_threshold_bytes = mixed_gc_live_threshold_bytes(); |
89 // Calculates the maximum array size that will be used. |
88 } |
90 static uint required_array_size(uint num_regions, uint num_workers, uint chunk_size) { |
89 |
91 uint const max_waste = num_workers * chunk_size; |
90 #ifndef PRODUCT |
92 // The array should be aligned with respect to chunk_size. |
91 void CollectionSetChooser::verify() { |
93 uint const aligned_num_regions = ((num_regions + chunk_size - 1) / chunk_size) * chunk_size; |
92 guarantee(_end <= regions_length(), "_end: %u regions length: %u", _end, regions_length()); |
94 |
93 guarantee(_front <= _end, "_front: %u _end: %u", _front, _end); |
95 return aligned_num_regions + max_waste; |
94 uint index = 0; |
96 } |
95 size_t sum_of_reclaimable_bytes = 0; |
97 |
96 while (index < _front) { |
98 public: |
97 guarantee(regions_at(index) == NULL, |
99 G1BuildCandidateArray(uint max_num_regions, uint num_workers, uint chunk_size) : |
98 "all entries before _front should be NULL"); |
100 _max_size(required_array_size(max_num_regions, num_workers, chunk_size)), |
99 index += 1; |
101 _chunk_size(chunk_size), |
100 } |
102 _data(NEW_C_HEAP_ARRAY(HeapRegion*, _max_size, mtGC)), |
101 HeapRegion *prev = NULL; |
103 _cur_claim_idx(0) { |
102 while (index < _end) { |
104 for (uint i = 0; i < _max_size; i++) { |
103 HeapRegion *curr = regions_at(index++); |
105 _data[i] = NULL; |
104 guarantee(curr != NULL, "Regions in _regions array cannot be NULL"); |
106 } |
105 guarantee(!curr->is_young(), "should not be young!"); |
107 } |
106 guarantee(!curr->is_pinned(), |
108 |
107 "Pinned region should not be in collection set (index %u)", curr->hrm_index()); |
109 ~G1BuildCandidateArray() { |
108 if (prev != NULL) { |
110 FREE_C_HEAP_ARRAY(HeapRegion*, _data); |
109 guarantee(order_regions(prev, curr) != 1, |
111 } |
110 "GC eff prev: %1.4f GC eff curr: %1.4f", |
112 |
111 prev->gc_efficiency(), curr->gc_efficiency()); |
113 // Claim a new chunk, returning its bounds [from, to[. |
112 } |
114 void claim_chunk(uint& from, uint& to) { |
113 sum_of_reclaimable_bytes += curr->reclaimable_bytes(); |
115 uint result = Atomic::add(_chunk_size, &_cur_claim_idx); |
114 prev = curr; |
116 assert(_max_size > result - 1, |
115 } |
117 "Array too small, is %u should be %u with chunk size %u.", |
116 guarantee(sum_of_reclaimable_bytes == _remaining_reclaimable_bytes, |
118 _max_size, result, _chunk_size); |
117 "reclaimable bytes inconsistent, " |
119 from = result - _chunk_size; |
118 "remaining: " SIZE_FORMAT " sum: " SIZE_FORMAT, |
120 to = result; |
119 _remaining_reclaimable_bytes, sum_of_reclaimable_bytes); |
121 } |
120 } |
122 |
121 #endif // !PRODUCT |
123 // Set element in array. |
122 |
124 void set(uint idx, HeapRegion* hr) { |
123 void CollectionSetChooser::sort_regions() { |
125 assert(idx < _max_size, "Index %u out of bounds %u", idx, _max_size); |
124 // First trim any unused portion of the top in the parallel case. |
126 assert(_data[idx] == NULL, "Value must not have been set."); |
125 if (_first_par_unreserved_idx > 0) { |
127 _data[idx] = hr; |
126 assert(_first_par_unreserved_idx <= regions_length(), |
128 } |
127 "Or we didn't reserved enough length"); |
129 |
128 regions_trunc_to(_first_par_unreserved_idx); |
130 void sort_and_copy_into(HeapRegion** dest, uint num_regions) { |
129 } |
131 if (_cur_claim_idx == 0) { |
130 _regions.sort(order_regions); |
132 return; |
131 assert(_end <= regions_length(), "Requirement"); |
133 } |
132 #ifdef ASSERT |
134 for (uint i = _cur_claim_idx; i < _max_size; i++) { |
133 for (uint i = 0; i < _end; i++) { |
135 assert(_data[i] == NULL, "must be"); |
134 assert(regions_at(i) != NULL, "Should be true by sorting!"); |
136 } |
135 } |
137 QuickSort::sort(_data, _cur_claim_idx, order_regions, true); |
136 #endif // ASSERT |
138 for (uint i = num_regions; i < _max_size; i++) { |
137 if (log_is_enabled(Trace, gc, liveness)) { |
139 assert(_data[i] == NULL, "must be"); |
138 G1PrintRegionLivenessInfoClosure cl("Post-Sorting"); |
140 } |
139 for (uint i = 0; i < _end; ++i) { |
141 for (uint i = 0; i < num_regions; i++) { |
140 HeapRegion* r = regions_at(i); |
142 dest[i] = _data[i]; |
141 cl.do_heap_region(r); |
143 } |
142 } |
144 } |
143 } |
145 }; |
144 verify(); |
146 |
145 } |
147 // Per-region closure. In addition to determining whether a region should be |
146 |
148 // added to the candidates, and calculating those regions' gc efficiencies, also |
147 void CollectionSetChooser::add_region(HeapRegion* hr) { |
149 // gather additional statistics. |
148 assert(!hr->is_pinned(), |
150 class G1BuildCandidateRegionsClosure : public HeapRegionClosure { |
149 "Pinned region shouldn't be added to the collection set (index %u)", hr->hrm_index()); |
151 G1BuildCandidateArray* _array; |
150 assert(hr->is_old(), "should be old but is %s", hr->get_type_str()); |
152 |
151 assert(hr->rem_set()->is_complete(), |
153 uint _cur_chunk_idx; |
152 "Trying to add region %u to the collection set with incomplete remembered set", hr->hrm_index()); |
154 uint _cur_chunk_end; |
153 _regions.append(hr); |
155 |
154 _end++; |
156 uint _regions_added; |
155 _remaining_reclaimable_bytes += hr->reclaimable_bytes(); |
157 size_t _reclaimable_bytes_added; |
156 hr->calc_gc_efficiency(); |
158 |
157 } |
159 void add_region(HeapRegion* hr) { |
158 |
160 if (_cur_chunk_idx == _cur_chunk_end) { |
159 void CollectionSetChooser::push(HeapRegion* hr) { |
161 _array->claim_chunk(_cur_chunk_idx, _cur_chunk_end); |
160 assert(hr != NULL, "Can't put back a NULL region"); |
162 } |
161 assert(_front >= 1, "Too many regions have been put back"); |
163 assert(_cur_chunk_idx < _cur_chunk_end, "Must be"); |
162 _front--; |
164 |
163 regions_at_put(_front, hr); |
165 hr->calc_gc_efficiency(); |
164 _remaining_reclaimable_bytes += hr->reclaimable_bytes(); |
166 _array->set(_cur_chunk_idx, hr); |
165 } |
167 |
166 |
168 _cur_chunk_idx++; |
167 void CollectionSetChooser::prepare_for_par_region_addition(uint n_threads, |
169 |
168 uint n_regions, |
170 _regions_added++; |
169 uint chunk_size) { |
171 _reclaimable_bytes_added += hr->reclaimable_bytes(); |
170 _first_par_unreserved_idx = 0; |
172 } |
171 uint max_waste = n_threads * chunk_size; |
173 |
172 // it should be aligned with respect to chunk_size |
174 bool should_add(HeapRegion* hr) { return CollectionSetChooser::should_add(hr); } |
173 uint aligned_n_regions = (n_regions + chunk_size - 1) / chunk_size * chunk_size; |
175 |
174 assert(aligned_n_regions % chunk_size == 0, "should be aligned"); |
176 public: |
175 regions_at_put_grow(aligned_n_regions + max_waste - 1, NULL); |
177 G1BuildCandidateRegionsClosure(G1BuildCandidateArray* array) : |
176 } |
178 _array(array), |
177 |
179 _cur_chunk_idx(0), |
178 uint CollectionSetChooser::claim_array_chunk(uint chunk_size) { |
180 _cur_chunk_end(0), |
179 uint res = (uint) Atomic::add((jint) chunk_size, |
181 _regions_added(0), |
180 (volatile jint*) &_first_par_unreserved_idx); |
182 _reclaimable_bytes_added(0) { } |
181 assert(regions_length() > res + chunk_size - 1, |
183 |
182 "Should already have been expanded"); |
184 bool do_heap_region(HeapRegion* r) { |
183 return res - chunk_size; |
185 // We will skip any region that's currently used as an old GC |
184 } |
186 // alloc region (we should not consider those for collection |
185 |
187 // before we fill them up). |
186 void CollectionSetChooser::set_region(uint index, HeapRegion* hr) { |
188 if (should_add(r) && !G1CollectedHeap::heap()->is_old_gc_alloc_region(r)) { |
187 assert(regions_at(index) == NULL, "precondition"); |
189 add_region(r); |
188 assert(hr->is_old(), "should be old but is %s", hr->get_type_str()); |
190 } else if (r->is_old()) { |
189 regions_at_put(index, hr); |
191 // Keep remembered sets for humongous regions, otherwise clean out remembered |
190 hr->calc_gc_efficiency(); |
192 // sets for old regions. |
191 } |
193 r->rem_set()->clear(true /* only_cardset */); |
192 |
194 } else { |
193 void CollectionSetChooser::update_totals(uint region_num, |
195 assert(r->is_archive() || !r->is_old() || !r->rem_set()->is_tracked(), |
194 size_t reclaimable_bytes) { |
196 "Missed to clear unused remembered set of region %u (%s) that is %s", |
195 // Only take the lock if we actually need to update the totals. |
197 r->hrm_index(), r->get_type_str(), r->rem_set()->get_state_str()); |
196 if (region_num > 0) { |
198 } |
197 assert(reclaimable_bytes > 0, "invariant"); |
199 return false; |
198 // We could have just used atomics instead of taking the |
200 } |
199 // lock. However, we currently don't have an atomic add for size_t. |
201 |
200 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); |
202 uint regions_added() const { return _regions_added; } |
201 _end += region_num; |
203 size_t reclaimable_bytes_added() const { return _reclaimable_bytes_added; } |
202 _remaining_reclaimable_bytes += reclaimable_bytes; |
204 }; |
203 } else { |
205 |
204 assert(reclaimable_bytes == 0, "invariant"); |
|
205 } |
|
206 } |
|
207 |
|
208 void CollectionSetChooser::iterate(HeapRegionClosure* cl) { |
|
209 for (uint i = _front; i < _end; i++) { |
|
210 HeapRegion* r = regions_at(i); |
|
211 if (cl->do_heap_region(r)) { |
|
212 cl->set_incomplete(); |
|
213 break; |
|
214 } |
|
215 } |
|
216 } |
|
217 |
|
218 void CollectionSetChooser::clear() { |
|
219 _regions.clear(); |
|
220 _front = 0; |
|
221 _end = 0; |
|
222 _remaining_reclaimable_bytes = 0; |
|
223 } |
|
224 |
|
225 class ParKnownGarbageHRClosure: public HeapRegionClosure { |
|
226 G1CollectedHeap* _g1h; |
|
227 CSetChooserParUpdater _cset_updater; |
|
228 |
|
229 public: |
|
230 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, |
|
231 uint chunk_size) : |
|
232 _g1h(G1CollectedHeap::heap()), |
|
233 _cset_updater(hrSorted, true /* parallel */, chunk_size) { } |
|
234 |
|
235 bool do_heap_region(HeapRegion* r) { |
|
236 // We will skip any region that's currently used as an old GC |
|
237 // alloc region (we should not consider those for collection |
|
238 // before we fill them up). |
|
239 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) { |
|
240 _cset_updater.add_region(r); |
|
241 } else if (r->is_old()) { |
|
242 // Keep remembered sets for humongous regions, otherwise clean out remembered |
|
243 // sets for old regions. |
|
244 r->rem_set()->clear(true /* only_cardset */); |
|
245 } else { |
|
246 assert(r->is_archive() || !r->is_old() || !r->rem_set()->is_tracked(), |
|
247 "Missed to clear unused remembered set of region %u (%s) that is %s", |
|
248 r->hrm_index(), r->get_type_str(), r->rem_set()->get_state_str()); |
|
249 } |
|
250 return false; |
|
251 } |
|
252 }; |
|
253 |
|
254 class ParKnownGarbageTask: public AbstractGangTask { |
|
255 CollectionSetChooser* _hrSorted; |
|
256 uint _chunk_size; |
|
257 G1CollectedHeap* _g1h; |
206 G1CollectedHeap* _g1h; |
258 HeapRegionClaimer _hrclaimer; |
207 HeapRegionClaimer _hrclaimer; |
259 |
208 |
|
209 uint volatile _num_regions_added; |
|
210 size_t volatile _reclaimable_bytes_added; |
|
211 |
|
212 G1BuildCandidateArray _result; |
|
213 |
|
214 void update_totals(uint num_regions, size_t reclaimable_bytes) { |
|
215 if (num_regions > 0) { |
|
216 assert(reclaimable_bytes > 0, "invariant"); |
|
217 Atomic::add(num_regions, &_num_regions_added); |
|
218 Atomic::add(reclaimable_bytes, &_reclaimable_bytes_added); |
|
219 } else { |
|
220 assert(reclaimable_bytes == 0, "invariant"); |
|
221 } |
|
222 } |
|
223 |
260 public: |
224 public: |
261 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) : |
225 G1BuildCandidateRegionsTask(uint max_num_regions, uint chunk_size, uint num_workers) : |
262 AbstractGangTask("ParKnownGarbageTask"), |
226 AbstractGangTask("G1 Build Candidate Regions"), |
263 _hrSorted(hrSorted), _chunk_size(chunk_size), |
227 _g1h(G1CollectedHeap::heap()), |
264 _g1h(G1CollectedHeap::heap()), _hrclaimer(n_workers) {} |
228 _hrclaimer(num_workers), |
|
229 _num_regions_added(0), |
|
230 _reclaimable_bytes_added(0), |
|
231 _result(max_num_regions, chunk_size, num_workers) { } |
265 |
232 |
266 void work(uint worker_id) { |
233 void work(uint worker_id) { |
267 ParKnownGarbageHRClosure par_known_garbage_cl(_hrSorted, _chunk_size); |
234 G1BuildCandidateRegionsClosure cl(&_result); |
268 _g1h->heap_region_par_iterate_from_worker_offset(&par_known_garbage_cl, &_hrclaimer, worker_id); |
235 _g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hrclaimer, worker_id); |
|
236 update_totals(cl.regions_added(), cl.reclaimable_bytes_added()); |
|
237 } |
|
238 |
|
239 G1CollectionSetCandidates* get_sorted_candidates() { |
|
240 HeapRegion** regions = NEW_C_HEAP_ARRAY(HeapRegion*, _num_regions_added, mtGC); |
|
241 _result.sort_and_copy_into(regions, _num_regions_added); |
|
242 return new G1CollectionSetCandidates(regions, |
|
243 _num_regions_added, |
|
244 _reclaimable_bytes_added); |
269 } |
245 } |
270 }; |
246 }; |
271 |
247 |
272 uint CollectionSetChooser::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const { |
248 uint CollectionSetChooser::calculate_work_chunk_size(uint num_workers, uint num_regions) { |
273 assert(n_workers > 0, "Active gc workers should be greater than 0"); |
249 assert(num_workers > 0, "Active gc workers should be greater than 0"); |
274 const uint overpartition_factor = 4; |
250 return MAX2(num_regions / num_workers, 1U); |
275 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U); |
251 } |
276 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size); |
252 |
277 } |
253 bool CollectionSetChooser::should_add(HeapRegion* hr) { |
278 |
|
279 bool CollectionSetChooser::region_occupancy_low_enough_for_evac(size_t live_bytes) { |
|
280 return live_bytes < mixed_gc_live_threshold_bytes(); |
|
281 } |
|
282 |
|
283 bool CollectionSetChooser::should_add(HeapRegion* hr) const { |
|
284 return !hr->is_young() && |
254 return !hr->is_young() && |
285 !hr->is_pinned() && |
255 !hr->is_pinned() && |
286 region_occupancy_low_enough_for_evac(hr->live_bytes()) && |
256 region_occupancy_low_enough_for_evac(hr->live_bytes()) && |
287 hr->rem_set()->is_complete(); |
257 hr->rem_set()->is_complete(); |
288 } |
258 } |
289 |
259 |
290 void CollectionSetChooser::rebuild(WorkGang* workers, uint n_regions) { |
260 G1CollectionSetCandidates* CollectionSetChooser::build(WorkGang* workers, uint max_num_regions) { |
291 clear(); |
261 uint num_workers = workers->active_workers(); |
292 |
262 uint chunk_size = calculate_work_chunk_size(num_workers, max_num_regions); |
293 uint n_workers = workers->active_workers(); |
263 |
294 |
264 G1BuildCandidateRegionsTask cl(max_num_regions, chunk_size, num_workers); |
295 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions); |
265 workers->run_task(&cl, num_workers); |
296 prepare_for_par_region_addition(n_workers, n_regions, chunk_size); |
266 |
297 |
267 G1CollectionSetCandidates* result = cl.get_sorted_candidates(); |
298 ParKnownGarbageTask par_known_garbage_task(this, chunk_size, n_workers); |
268 result->verify(); |
299 workers->run_task(&par_known_garbage_task); |
269 return result; |
300 |
270 } |
301 sort_regions(); |
|
302 } |
|