Print this page
rev 2896 : 6484965: G1: piggy-back liveness accounting phase on marking
Summary: Remove the separate counting phase of concurrent marking by tracking the amount of marked bytes and the cards spanned by marked objects in marking task/worker thread local data structures, which are updated as individual objects are marked.
Reviewed-by: brutisso
| Split |
Close |
| Expand all |
| Collapse all |
--- old/src/share/vm/gc_implementation/g1/concurrentMark.inline.hpp
+++ new/src/share/vm/gc_implementation/g1/concurrentMark.inline.hpp
1 1 /*
2 2 * Copyright (c) 2001, 2011, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
|
↓ open down ↓ |
20 lines elided |
↑ open up ↑ |
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #ifndef SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_INLINE_HPP
26 26 #define SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_INLINE_HPP
27 27
28 28 #include "gc_implementation/g1/concurrentMark.hpp"
29 29 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
30 30
31 +// Counts the given memory region in the given task/worker
32 +// counting data structures.
33 +inline void ConcurrentMark::count_region(MemRegion mr, HeapRegion* hr,
34 + size_t* marked_bytes_array,
35 + BitMap* task_card_bm) {
36 + G1CollectedHeap* g1h = _g1h;
37 + HeapWord* start = mr.start();
38 + HeapWord* last = mr.last();
39 + size_t region_size = mr.byte_size();
40 + size_t index = hr->hrs_index();
41 +
42 + assert(!hr->continuesHumongous(), "should not be HC region");
43 + assert(hr == g1h->heap_region_containing(start), "sanity");
44 + assert(hr == g1h->heap_region_containing(mr.last()), "sanity");
45 + assert(marked_bytes_array != NULL, "pre-condition");
46 + assert(task_card_bm != NULL, "pre-condition");
47 +
48 + // Add to the task local marked bytes for this region.
49 + marked_bytes_array[index] += region_size;
50 +
51 + // Below, the term "card num" means the result of shifting an address
52 + // by the card shift -- address 0 corresponds to card number 0. One
53 + // must subtract the card num of the bottom of the heap to obtain a
54 + // card table index.
55 +
56 + intptr_t start_card_num = intptr_t(uintptr_t(start) >> CardTableModRefBS::card_shift);
57 + intptr_t last_card_num = intptr_t(uintptr_t(last) >> CardTableModRefBS::card_shift);
58 +
59 + BitMap::idx_t start_idx = start_card_num - heap_bottom_card_num();
60 + BitMap::idx_t last_idx = last_card_num - heap_bottom_card_num();
61 +
62 + // The card bitmap is task/worker specific => no need to use 'par' routines.
63 + // Set bits in the inclusive bit range [start_idx, last_idx].
64 + for (BitMap::idx_t i = start_idx; i <= last_idx; i += 1) {
65 + task_card_bm->set_bit(i);
66 + }
67 +}
68 +
69 +// Counts the given memory region in the ask/worker counting
70 +// data structures for the given worker id.
71 +inline void ConcurrentMark::count_region(MemRegion mr, int worker_i) {
72 + size_t* marked_bytes_array = count_marked_bytes_array_for(worker_i);
73 + BitMap* task_card_bm = count_card_bitmap_for(worker_i);
74 + HeapWord* addr = mr.start();
75 + HeapRegion* hr = _g1h->heap_region_containing(addr);
76 + count_region(mr, hr, marked_bytes_array, task_card_bm);
77 +}
78 +
79 +// Counts the given object in the given task/worker counting data structures.
80 +inline void ConcurrentMark::count_object(oop obj,
81 + HeapRegion* hr,
82 + size_t* marked_bytes_array,
83 + BitMap* task_card_bm) {
84 + MemRegion mr((HeapWord*)obj, obj->size());
85 + count_region(mr, hr, marked_bytes_array, task_card_bm);
86 +}
87 +
88 +// Counts the given object in the task/worker counting data
89 +// structures for the given worker id.
90 +inline void ConcurrentMark::count_object(oop obj, HeapRegion* hr, int worker_i) {
91 + size_t* marked_bytes_array = count_marked_bytes_array_for(worker_i);
92 + BitMap* task_card_bm = count_card_bitmap_for(worker_i);
93 + HeapWord* addr = (HeapWord*) obj;
94 + count_object(obj, hr, marked_bytes_array, task_card_bm);
95 +}
96 +
97 +// Attempts to mark the given object and, if successful, counts
98 +// the object in the given task/worker counting structures.
99 +inline bool ConcurrentMark::par_mark_and_count(oop obj,
100 + HeapRegion* hr,
101 + size_t* marked_bytes_array,
102 + BitMap* task_card_bm) {
103 + HeapWord* addr = (HeapWord*)obj;
104 + if (_nextMarkBitMap->parMark(addr)) {
105 + // Update the task specific count data for the object.
106 + count_object(obj, hr, marked_bytes_array, task_card_bm);
107 + return true;
108 + }
109 + return false;
110 +}
111 +
112 +// Attempts to mark the given object and, if successful, counts
113 +// the object in the task/worker counting structures for the
114 +// given worker id.
115 +inline bool ConcurrentMark::par_mark_and_count(oop obj,
116 + HeapRegion* hr,
117 + int worker_i) {
118 + HeapWord* addr = (HeapWord*)obj;
119 + if (_nextMarkBitMap->parMark(addr)) {
120 + // Update the task specific count data for the object.
121 + count_object(obj, hr, worker_i);
122 + return true;
123 + }
124 + return false;
125 +}
126 +
127 +// As above - but we don't know the heap region containing the
128 +// object and so have to supply it.
129 +inline bool ConcurrentMark::par_mark_and_count(oop obj, int worker_i) {
130 + HeapWord* addr = (HeapWord*)obj;
131 + HeapRegion* hr = _g1h->heap_region_containing(addr);
132 + return par_mark_and_count(obj, hr, worker_i);
133 +}
134 +
135 +// Unconditionally mark the given object, and unconditinally count
136 +// the object in the counting structures for worker id 0.
137 +// Should *not* be called from parallel code.
138 +inline bool ConcurrentMark::mark_and_count(oop obj, HeapRegion* hr) {
139 + HeapWord* addr = (HeapWord*)obj;
140 + _nextMarkBitMap->mark(addr);
141 + // Update the task specific count data for the object.
142 + count_object(obj, hr, 0 /* worker_i */);
143 + return true;
144 +}
145 +
146 +// As above - but we don't have the heap region containing the
147 +// object, so we have to supply it.
148 +inline bool ConcurrentMark::mark_and_count(oop obj) {
149 + HeapWord* addr = (HeapWord*)obj;
150 + HeapRegion* hr = _g1h->heap_region_containing(addr);
151 + return mark_and_count(obj, hr);
152 +}
153 +
31 154 inline void CMTask::push(oop obj) {
32 155 HeapWord* objAddr = (HeapWord*) obj;
33 156 assert(_g1h->is_in_g1_reserved(objAddr), "invariant");
34 157 assert(!_g1h->is_on_master_free_list(
35 158 _g1h->heap_region_containing((HeapWord*) objAddr)), "invariant");
36 159 assert(!_g1h->is_obj_ill(obj), "invariant");
37 160 assert(_nextMarkBitMap->isMarked(objAddr), "invariant");
38 161
39 162 if (_cm->verbose_high()) {
40 163 gclog_or_tty->print_cr("[%d] pushing "PTR_FORMAT, _task_id, (void*) obj);
41 164 }
42 165
43 166 if (!_task_queue->push(obj)) {
44 167 // The local task queue looks full. We need to push some entries
45 168 // to the global stack.
46 169
47 170 if (_cm->verbose_medium()) {
48 171 gclog_or_tty->print_cr("[%d] task queue overflow, "
49 172 "moving entries to the global stack",
50 173 _task_id);
51 174 }
52 175 move_entries_to_global_stack();
53 176
54 177 // this should succeed since, even if we overflow the global
55 178 // stack, we should have definitely removed some entries from the
56 179 // local queue. So, there must be space on it.
57 180 bool success = _task_queue->push(obj);
58 181 assert(success, "invariant");
59 182 }
60 183
61 184 statsOnly( int tmp_size = _task_queue->size();
62 185 if (tmp_size > _local_max_size) {
63 186 _local_max_size = tmp_size;
64 187 }
65 188 ++_local_pushes );
66 189 }
67 190
68 191 // This determines whether the method below will check both the local
69 192 // and global fingers when determining whether to push on the stack a
70 193 // gray object (value 1) or whether it will only check the global one
71 194 // (value 0). The tradeoffs are that the former will be a bit more
72 195 // accurate and possibly push less on the stack, but it might also be
73 196 // a little bit slower.
74 197
75 198 #define _CHECK_BOTH_FINGERS_ 1
76 199
|
↓ open down ↓ |
36 lines elided |
↑ open up ↑ |
77 200 inline void CMTask::deal_with_reference(oop obj) {
78 201 if (_cm->verbose_high()) {
79 202 gclog_or_tty->print_cr("[%d] we're dealing with reference = "PTR_FORMAT,
80 203 _task_id, (void*) obj);
81 204 }
82 205
83 206 ++_refs_reached;
84 207
85 208 HeapWord* objAddr = (HeapWord*) obj;
86 209 assert(obj->is_oop_or_null(true /* ignore mark word */), "Error");
87 - if (_g1h->is_in_g1_reserved(objAddr)) {
210 + if (_g1h->is_in_g1_reserved(objAddr)) {
88 211 assert(obj != NULL, "null check is implicit");
89 212 if (!_nextMarkBitMap->isMarked(objAddr)) {
90 213 // Only get the containing region if the object is not marked on the
91 214 // bitmap (otherwise, it's a waste of time since we won't do
92 215 // anything with it).
93 216 HeapRegion* hr = _g1h->heap_region_containing_raw(obj);
94 217 if (!hr->obj_allocated_since_next_marking(obj)) {
95 218 if (_cm->verbose_high()) {
96 219 gclog_or_tty->print_cr("[%d] "PTR_FORMAT" is not considered marked",
97 220 _task_id, (void*) obj);
98 221 }
99 222
100 223 // we need to mark it first
101 - if (_nextMarkBitMap->parMark(objAddr)) {
224 + if (_cm->par_mark_and_count(obj, hr, _marked_bytes_array, _card_bm)) {
102 225 // No OrderAccess:store_load() is needed. It is implicit in the
103 - // CAS done in parMark(objAddr) above
226 + // CAS done in CMBitMap::parMark() call in the routine above.
104 227 HeapWord* global_finger = _cm->finger();
105 228
106 229 #if _CHECK_BOTH_FINGERS_
107 230 // we will check both the local and global fingers
108 231
109 232 if (_finger != NULL && objAddr < _finger) {
110 233 if (_cm->verbose_high()) {
111 234 gclog_or_tty->print_cr("[%d] below the local finger ("PTR_FORMAT"), "
112 235 "pushing it", _task_id, _finger);
113 236 }
114 237 push(obj);
115 238 } else if (_curr_region != NULL && objAddr < _region_limit) {
116 239 // do nothing
117 240 } else if (objAddr < global_finger) {
118 241 // Notice that the global finger might be moving forward
119 242 // concurrently. This is not a problem. In the worst case, we
120 243 // mark the object while it is above the global finger and, by
121 244 // the time we read the global finger, it has moved forward
122 245 // passed this object. In this case, the object will probably
123 246 // be visited when a task is scanning the region and will also
124 247 // be pushed on the stack. So, some duplicate work, but no
125 248 // correctness problems.
126 249
127 250 if (_cm->verbose_high()) {
128 251 gclog_or_tty->print_cr("[%d] below the global finger "
129 252 "("PTR_FORMAT"), pushing it",
130 253 _task_id, global_finger);
131 254 }
132 255 push(obj);
133 256 } else {
134 257 // do nothing
135 258 }
136 259 #else // _CHECK_BOTH_FINGERS_
137 260 // we will only check the global finger
138 261
139 262 if (objAddr < global_finger) {
140 263 // see long comment above
141 264
142 265 if (_cm->verbose_high()) {
143 266 gclog_or_tty->print_cr("[%d] below the global finger "
144 267 "("PTR_FORMAT"), pushing it",
145 268 _task_id, global_finger);
146 269 }
147 270 push(obj);
148 271 }
149 272 #endif // _CHECK_BOTH_FINGERS_
150 273 }
151 274 }
152 275 }
153 276 }
154 277 }
155 278
156 279 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_INLINE_HPP
|
↓ open down ↓ |
43 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX