Mercurial > hg > truffle
annotate src/share/vm/gc_implementation/shared/mutableNUMASpace.cpp @ 391:ab4a7734b9c4
6753547: NUMA allocator: Invalid chunk size computation during adaptive resizing
Summary: The per-lgrp chuck size can be incorrectly computed (causing an assertion failure) because of the non-associativity of the floating point operations. The fix is to rearrange the operations.
Reviewed-by: ysr
| author | iveresov |
|---|---|
| date | Mon, 06 Oct 2008 20:59:16 -0700 |
| parents | 06df86c2ec37 |
| children | 85f1b9537f70 |
| rev | line source |
|---|---|
| 0 | 1 |
| 2 /* | |
| 196 | 3 * Copyright 2006-2008 Sun Microsystems, Inc. All Rights Reserved. |
| 0 | 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
| 5 * | |
| 6 * This code is free software; you can redistribute it and/or modify it | |
| 7 * under the terms of the GNU General Public License version 2 only, as | |
| 8 * published by the Free Software Foundation. | |
| 9 * | |
| 10 * This code is distributed in the hope that it will be useful, but WITHOUT | |
| 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
| 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
| 13 * version 2 for more details (a copy is included in the LICENSE file that | |
| 14 * accompanied this code). | |
| 15 * | |
| 16 * You should have received a copy of the GNU General Public License version | |
| 17 * 2 along with this work; if not, write to the Free Software Foundation, | |
| 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
| 19 * | |
| 20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
| 21 * CA 95054 USA or visit www.sun.com if you need additional information or | |
| 22 * have any questions. | |
| 23 * | |
| 24 */ | |
| 25 | |
| 26 # include "incls/_precompiled.incl" | |
| 27 # include "incls/_mutableNUMASpace.cpp.incl" | |
| 28 | |
| 29 | |
| 30 MutableNUMASpace::MutableNUMASpace() { | |
| 31 _lgrp_spaces = new (ResourceObj::C_HEAP) GrowableArray<LGRPSpace*>(0, true); | |
| 32 _page_size = os::vm_page_size(); | |
| 33 _adaptation_cycles = 0; | |
| 34 _samples_count = 0; | |
| 35 update_layout(true); | |
| 36 } | |
| 37 | |
| 38 MutableNUMASpace::~MutableNUMASpace() { | |
| 39 for (int i = 0; i < lgrp_spaces()->length(); i++) { | |
| 40 delete lgrp_spaces()->at(i); | |
| 41 } | |
| 42 delete lgrp_spaces(); | |
| 43 } | |
| 44 | |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
45 #ifndef PRODUCT |
| 0 | 46 void MutableNUMASpace::mangle_unused_area() { |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
47 // This method should do nothing. |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
48 // It can be called on a numa space during a full compaction. |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
49 } |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
50 void MutableNUMASpace::mangle_unused_area_complete() { |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
51 // This method should do nothing. |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
52 // It can be called on a numa space during a full compaction. |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
53 } |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
54 void MutableNUMASpace::mangle_region(MemRegion mr) { |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
55 // This method should do nothing because numa spaces are not mangled. |
| 0 | 56 } |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
57 void MutableNUMASpace::set_top_for_allocations(HeapWord* v) { |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
58 assert(false, "Do not mangle MutableNUMASpace's"); |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
59 } |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
60 void MutableNUMASpace::set_top_for_allocations() { |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
61 // This method should do nothing. |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
62 } |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
63 void MutableNUMASpace::check_mangled_unused_area(HeapWord* limit) { |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
64 // This method should do nothing. |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
65 } |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
66 void MutableNUMASpace::check_mangled_unused_area_complete() { |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
67 // This method should do nothing. |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
68 } |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
69 #endif // NOT_PRODUCT |
| 0 | 70 |
| 71 // There may be unallocated holes in the middle chunks | |
| 72 // that should be filled with dead objects to ensure parseability. | |
| 73 void MutableNUMASpace::ensure_parsability() { | |
| 74 for (int i = 0; i < lgrp_spaces()->length(); i++) { | |
| 75 LGRPSpace *ls = lgrp_spaces()->at(i); | |
| 76 MutableSpace *s = ls->space(); | |
|
144
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
77 if (s->top() < top()) { // For all spaces preceeding the one containing top() |
| 0 | 78 if (s->free_in_words() > 0) { |
| 79 SharedHeap::fill_region_with_object(MemRegion(s->top(), s->end())); | |
|
144
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
80 size_t area_touched_words = pointer_delta(s->end(), s->top()); |
| 0 | 81 #ifndef ASSERT |
| 82 if (!ZapUnusedHeapArea) { | |
| 83 area_touched_words = MIN2((size_t)align_object_size(typeArrayOopDesc::header_size(T_INT)), | |
| 84 area_touched_words); | |
| 85 } | |
| 86 #endif | |
| 141 | 87 if (!os::numa_has_static_binding()) { |
| 88 MemRegion invalid; | |
| 89 HeapWord *crossing_start = (HeapWord*)round_to((intptr_t)s->top(), os::vm_page_size()); | |
| 90 HeapWord *crossing_end = (HeapWord*)round_to((intptr_t)(s->top() + area_touched_words), | |
| 91 os::vm_page_size()); | |
| 92 if (crossing_start != crossing_end) { | |
| 93 // If object header crossed a small page boundary we mark the area | |
| 94 // as invalid rounding it to a page_size(). | |
| 95 HeapWord *start = MAX2((HeapWord*)round_down((intptr_t)s->top(), page_size()), s->bottom()); | |
| 96 HeapWord *end = MIN2((HeapWord*)round_to((intptr_t)(s->top() + area_touched_words), page_size()), | |
| 97 s->end()); | |
| 98 invalid = MemRegion(start, end); | |
| 99 } | |
| 100 | |
| 101 ls->add_invalid_region(invalid); | |
| 0 | 102 } |
| 103 } | |
| 104 } else { | |
| 141 | 105 if (!os::numa_has_static_binding()) { |
| 0 | 106 #ifdef ASSERT |
| 107 MemRegion invalid(s->top(), s->end()); | |
| 108 ls->add_invalid_region(invalid); | |
| 141 | 109 #else |
| 110 if (ZapUnusedHeapArea) { | |
| 111 MemRegion invalid(s->top(), s->end()); | |
| 112 ls->add_invalid_region(invalid); | |
|
144
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
113 } else { |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
114 return; |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
115 } |
| 0 | 116 #endif |
|
144
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
117 } else { |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
118 return; |
| 141 | 119 } |
| 0 | 120 } |
| 121 } | |
| 122 } | |
| 123 | |
| 124 size_t MutableNUMASpace::used_in_words() const { | |
| 125 size_t s = 0; | |
| 126 for (int i = 0; i < lgrp_spaces()->length(); i++) { | |
| 127 s += lgrp_spaces()->at(i)->space()->used_in_words(); | |
| 128 } | |
| 129 return s; | |
| 130 } | |
| 131 | |
| 132 size_t MutableNUMASpace::free_in_words() const { | |
| 133 size_t s = 0; | |
| 134 for (int i = 0; i < lgrp_spaces()->length(); i++) { | |
| 135 s += lgrp_spaces()->at(i)->space()->free_in_words(); | |
| 136 } | |
| 137 return s; | |
| 138 } | |
| 139 | |
| 140 | |
| 141 size_t MutableNUMASpace::tlab_capacity(Thread *thr) const { | |
| 142 guarantee(thr != NULL, "No thread"); | |
| 143 int lgrp_id = thr->lgrp_id(); | |
|
268
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
144 if (lgrp_id == -1) { |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
145 // This case can occur after the topology of the system has |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
146 // changed. Thread can change their location, the new home |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
147 // group will be determined during the first allocation |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
148 // attempt. For now we can safely assume that all spaces |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
149 // have equal size because the whole space will be reinitialized. |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
150 if (lgrp_spaces()->length() > 0) { |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
151 return capacity_in_bytes() / lgrp_spaces()->length(); |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
152 } else { |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
153 assert(false, "There should be at least one locality group"); |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
154 return 0; |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
155 } |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
156 } |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
157 // That's the normal case, where we know the locality group of the thread. |
| 0 | 158 int i = lgrp_spaces()->find(&lgrp_id, LGRPSpace::equals); |
| 159 if (i == -1) { | |
| 160 return 0; | |
| 161 } | |
| 162 return lgrp_spaces()->at(i)->space()->capacity_in_bytes(); | |
| 163 } | |
| 164 | |
| 165 size_t MutableNUMASpace::unsafe_max_tlab_alloc(Thread *thr) const { | |
|
268
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
166 // Please see the comments for tlab_capacity(). |
| 0 | 167 guarantee(thr != NULL, "No thread"); |
| 168 int lgrp_id = thr->lgrp_id(); | |
|
268
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
169 if (lgrp_id == -1) { |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
170 if (lgrp_spaces()->length() > 0) { |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
171 return free_in_bytes() / lgrp_spaces()->length(); |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
172 } else { |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
173 assert(false, "There should be at least one locality group"); |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
174 return 0; |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
175 } |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
176 } |
| 0 | 177 int i = lgrp_spaces()->find(&lgrp_id, LGRPSpace::equals); |
| 178 if (i == -1) { | |
| 179 return 0; | |
| 180 } | |
| 181 return lgrp_spaces()->at(i)->space()->free_in_bytes(); | |
| 182 } | |
| 183 | |
|
373
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
184 |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
185 size_t MutableNUMASpace::capacity_in_words(Thread* thr) const { |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
186 guarantee(thr != NULL, "No thread"); |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
187 int lgrp_id = thr->lgrp_id(); |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
188 if (lgrp_id == -1) { |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
189 if (lgrp_spaces()->length() > 0) { |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
190 return capacity_in_words() / lgrp_spaces()->length(); |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
191 } else { |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
192 assert(false, "There should be at least one locality group"); |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
193 return 0; |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
194 } |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
195 } |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
196 int i = lgrp_spaces()->find(&lgrp_id, LGRPSpace::equals); |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
197 if (i == -1) { |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
198 return 0; |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
199 } |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
200 return lgrp_spaces()->at(i)->space()->capacity_in_words(); |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
201 } |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
202 |
| 0 | 203 // Check if the NUMA topology has changed. Add and remove spaces if needed. |
| 204 // The update can be forced by setting the force parameter equal to true. | |
| 205 bool MutableNUMASpace::update_layout(bool force) { | |
| 206 // Check if the topology had changed. | |
| 207 bool changed = os::numa_topology_changed(); | |
| 208 if (force || changed) { | |
| 209 // Compute lgrp intersection. Add/remove spaces. | |
| 210 int lgrp_limit = (int)os::numa_get_groups_num(); | |
| 211 int *lgrp_ids = NEW_C_HEAP_ARRAY(int, lgrp_limit); | |
| 212 int lgrp_num = (int)os::numa_get_leaf_groups(lgrp_ids, lgrp_limit); | |
| 213 assert(lgrp_num > 0, "There should be at least one locality group"); | |
| 214 // Add new spaces for the new nodes | |
| 215 for (int i = 0; i < lgrp_num; i++) { | |
| 216 bool found = false; | |
| 217 for (int j = 0; j < lgrp_spaces()->length(); j++) { | |
| 218 if (lgrp_spaces()->at(j)->lgrp_id() == lgrp_ids[i]) { | |
| 219 found = true; | |
| 220 break; | |
| 221 } | |
| 222 } | |
| 223 if (!found) { | |
| 224 lgrp_spaces()->append(new LGRPSpace(lgrp_ids[i])); | |
| 225 } | |
| 226 } | |
| 227 | |
| 228 // Remove spaces for the removed nodes. | |
| 229 for (int i = 0; i < lgrp_spaces()->length();) { | |
| 230 bool found = false; | |
| 231 for (int j = 0; j < lgrp_num; j++) { | |
| 232 if (lgrp_spaces()->at(i)->lgrp_id() == lgrp_ids[j]) { | |
| 233 found = true; | |
| 234 break; | |
| 235 } | |
| 236 } | |
| 237 if (!found) { | |
| 238 delete lgrp_spaces()->at(i); | |
| 239 lgrp_spaces()->remove_at(i); | |
| 240 } else { | |
| 241 i++; | |
| 242 } | |
| 243 } | |
| 244 | |
| 245 FREE_C_HEAP_ARRAY(int, lgrp_ids); | |
| 246 | |
| 247 if (changed) { | |
| 248 for (JavaThread *thread = Threads::first(); thread; thread = thread->next()) { | |
| 249 thread->set_lgrp_id(-1); | |
| 250 } | |
| 251 } | |
| 252 return true; | |
| 253 } | |
| 254 return false; | |
| 255 } | |
| 256 | |
| 257 // Bias region towards the first-touching lgrp. Set the right page sizes. | |
| 141 | 258 void MutableNUMASpace::bias_region(MemRegion mr, int lgrp_id) { |
| 0 | 259 HeapWord *start = (HeapWord*)round_to((intptr_t)mr.start(), page_size()); |
| 260 HeapWord *end = (HeapWord*)round_down((intptr_t)mr.end(), page_size()); | |
| 261 if (end > start) { | |
| 262 MemRegion aligned_region(start, end); | |
| 263 assert((intptr_t)aligned_region.start() % page_size() == 0 && | |
| 264 (intptr_t)aligned_region.byte_size() % page_size() == 0, "Bad alignment"); | |
| 265 assert(region().contains(aligned_region), "Sanity"); | |
| 141 | 266 // First we tell the OS which page size we want in the given range. The underlying |
| 267 // large page can be broken down if we require small pages. | |
| 0 | 268 os::realign_memory((char*)aligned_region.start(), aligned_region.byte_size(), page_size()); |
| 141 | 269 // Then we uncommit the pages in the range. |
| 270 os::free_memory((char*)aligned_region.start(), aligned_region.byte_size()); | |
| 271 // And make them local/first-touch biased. | |
| 272 os::numa_make_local((char*)aligned_region.start(), aligned_region.byte_size(), lgrp_id); | |
| 0 | 273 } |
| 274 } | |
| 275 | |
| 276 // Free all pages in the region. | |
| 277 void MutableNUMASpace::free_region(MemRegion mr) { | |
| 278 HeapWord *start = (HeapWord*)round_to((intptr_t)mr.start(), page_size()); | |
| 279 HeapWord *end = (HeapWord*)round_down((intptr_t)mr.end(), page_size()); | |
| 280 if (end > start) { | |
| 281 MemRegion aligned_region(start, end); | |
| 282 assert((intptr_t)aligned_region.start() % page_size() == 0 && | |
| 283 (intptr_t)aligned_region.byte_size() % page_size() == 0, "Bad alignment"); | |
| 284 assert(region().contains(aligned_region), "Sanity"); | |
| 285 os::free_memory((char*)aligned_region.start(), aligned_region.byte_size()); | |
| 286 } | |
| 287 } | |
| 288 | |
| 289 // Update space layout. Perform adaptation. | |
| 290 void MutableNUMASpace::update() { | |
| 291 if (update_layout(false)) { | |
| 292 // If the topology has changed, make all chunks zero-sized. | |
|
268
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
293 // And clear the alloc-rate statistics. |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
294 // In future we may want to handle this more gracefully in order |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
295 // to avoid the reallocation of the pages as much as possible. |
| 0 | 296 for (int i = 0; i < lgrp_spaces()->length(); i++) { |
|
268
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
297 LGRPSpace *ls = lgrp_spaces()->at(i); |
|
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
298 MutableSpace *s = ls->space(); |
| 0 | 299 s->set_end(s->bottom()); |
| 300 s->set_top(s->bottom()); | |
|
268
d6340ab4105b
6723228: NUMA allocator: assert(lgrp_id != -1, "No lgrp_id set")
iveresov
parents:
263
diff
changeset
|
301 ls->clear_alloc_rate(); |
| 0 | 302 } |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
303 // A NUMA space is never mangled |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
304 initialize(region(), |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
305 SpaceDecorator::Clear, |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
306 SpaceDecorator::DontMangle); |
| 0 | 307 } else { |
| 308 bool should_initialize = false; | |
| 141 | 309 if (!os::numa_has_static_binding()) { |
| 310 for (int i = 0; i < lgrp_spaces()->length(); i++) { | |
| 311 if (!lgrp_spaces()->at(i)->invalid_region().is_empty()) { | |
| 312 should_initialize = true; | |
| 313 break; | |
| 314 } | |
| 0 | 315 } |
| 316 } | |
| 317 | |
| 318 if (should_initialize || | |
| 319 (UseAdaptiveNUMAChunkSizing && adaptation_cycles() < samples_count())) { | |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
320 // A NUMA space is never mangled |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
321 initialize(region(), |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
322 SpaceDecorator::Clear, |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
323 SpaceDecorator::DontMangle); |
| 0 | 324 } |
| 325 } | |
| 326 | |
| 327 if (NUMAStats) { | |
| 328 for (int i = 0; i < lgrp_spaces()->length(); i++) { | |
| 329 lgrp_spaces()->at(i)->accumulate_statistics(page_size()); | |
| 330 } | |
| 331 } | |
| 332 | |
| 333 scan_pages(NUMAPageScanRate); | |
| 334 } | |
| 335 | |
| 336 // Scan pages. Free pages that have smaller size or wrong placement. | |
| 337 void MutableNUMASpace::scan_pages(size_t page_count) | |
| 338 { | |
| 339 size_t pages_per_chunk = page_count / lgrp_spaces()->length(); | |
| 340 if (pages_per_chunk > 0) { | |
| 341 for (int i = 0; i < lgrp_spaces()->length(); i++) { | |
| 342 LGRPSpace *ls = lgrp_spaces()->at(i); | |
| 343 ls->scan_pages(page_size(), pages_per_chunk); | |
| 344 } | |
| 345 } | |
| 346 } | |
| 347 | |
| 348 // Accumulate statistics about the allocation rate of each lgrp. | |
| 349 void MutableNUMASpace::accumulate_statistics() { | |
| 350 if (UseAdaptiveNUMAChunkSizing) { | |
| 351 for (int i = 0; i < lgrp_spaces()->length(); i++) { | |
| 352 lgrp_spaces()->at(i)->sample(); | |
| 353 } | |
| 354 increment_samples_count(); | |
| 355 } | |
| 356 | |
| 357 if (NUMAStats) { | |
| 358 for (int i = 0; i < lgrp_spaces()->length(); i++) { | |
| 359 lgrp_spaces()->at(i)->accumulate_statistics(page_size()); | |
| 360 } | |
| 361 } | |
| 362 } | |
| 363 | |
| 364 // Get the current size of a chunk. | |
| 365 // This function computes the size of the chunk based on the | |
| 366 // difference between chunk ends. This allows it to work correctly in | |
| 367 // case the whole space is resized and during the process of adaptive | |
| 368 // chunk resizing. | |
| 369 size_t MutableNUMASpace::current_chunk_size(int i) { | |
| 370 HeapWord *cur_end, *prev_end; | |
| 371 if (i == 0) { | |
| 372 prev_end = bottom(); | |
| 373 } else { | |
| 374 prev_end = lgrp_spaces()->at(i - 1)->space()->end(); | |
| 375 } | |
| 376 if (i == lgrp_spaces()->length() - 1) { | |
| 377 cur_end = end(); | |
| 378 } else { | |
| 379 cur_end = lgrp_spaces()->at(i)->space()->end(); | |
| 380 } | |
| 381 if (cur_end > prev_end) { | |
| 382 return pointer_delta(cur_end, prev_end, sizeof(char)); | |
| 383 } | |
| 384 return 0; | |
| 385 } | |
| 386 | |
| 387 // Return the default chunk size by equally diving the space. | |
| 388 // page_size() aligned. | |
| 389 size_t MutableNUMASpace::default_chunk_size() { | |
| 390 return base_space_size() / lgrp_spaces()->length() * page_size(); | |
| 391 } | |
| 392 | |
| 393 // Produce a new chunk size. page_size() aligned. | |
|
391
ab4a7734b9c4
6753547: NUMA allocator: Invalid chunk size computation during adaptive resizing
iveresov
parents:
373
diff
changeset
|
394 // This function is expected to be called on sequence of i's from 0 to |
|
ab4a7734b9c4
6753547: NUMA allocator: Invalid chunk size computation during adaptive resizing
iveresov
parents:
373
diff
changeset
|
395 // lgrp_spaces()->length(). |
| 0 | 396 size_t MutableNUMASpace::adaptive_chunk_size(int i, size_t limit) { |
| 397 size_t pages_available = base_space_size(); | |
| 398 for (int j = 0; j < i; j++) { | |
| 399 pages_available -= round_down(current_chunk_size(j), page_size()) / page_size(); | |
| 400 } | |
| 401 pages_available -= lgrp_spaces()->length() - i - 1; | |
| 402 assert(pages_available > 0, "No pages left"); | |
| 403 float alloc_rate = 0; | |
| 404 for (int j = i; j < lgrp_spaces()->length(); j++) { | |
| 405 alloc_rate += lgrp_spaces()->at(j)->alloc_rate()->average(); | |
| 406 } | |
| 407 size_t chunk_size = 0; | |
| 408 if (alloc_rate > 0) { | |
| 409 LGRPSpace *ls = lgrp_spaces()->at(i); | |
|
391
ab4a7734b9c4
6753547: NUMA allocator: Invalid chunk size computation during adaptive resizing
iveresov
parents:
373
diff
changeset
|
410 chunk_size = (size_t)(ls->alloc_rate()->average() / alloc_rate * pages_available) * page_size(); |
| 0 | 411 } |
| 412 chunk_size = MAX2(chunk_size, page_size()); | |
| 413 | |
| 414 if (limit > 0) { | |
| 415 limit = round_down(limit, page_size()); | |
| 416 if (chunk_size > current_chunk_size(i)) { | |
| 417 chunk_size = MIN2((off_t)chunk_size, (off_t)current_chunk_size(i) + (off_t)limit); | |
| 418 } else { | |
| 419 chunk_size = MAX2((off_t)chunk_size, (off_t)current_chunk_size(i) - (off_t)limit); | |
| 420 } | |
| 421 } | |
| 422 assert(chunk_size <= pages_available * page_size(), "Chunk size out of range"); | |
| 423 return chunk_size; | |
| 424 } | |
| 425 | |
| 426 | |
| 427 // Return the bottom_region and the top_region. Align them to page_size() boundary. | |
| 428 // |------------------new_region---------------------------------| | |
| 429 // |----bottom_region--|---intersection---|------top_region------| | |
| 430 void MutableNUMASpace::select_tails(MemRegion new_region, MemRegion intersection, | |
| 431 MemRegion* bottom_region, MemRegion *top_region) { | |
| 432 // Is there bottom? | |
| 433 if (new_region.start() < intersection.start()) { // Yes | |
| 434 // Try to coalesce small pages into a large one. | |
| 435 if (UseLargePages && page_size() >= os::large_page_size()) { | |
| 436 HeapWord* p = (HeapWord*)round_to((intptr_t) intersection.start(), os::large_page_size()); | |
| 437 if (new_region.contains(p) | |
| 438 && pointer_delta(p, new_region.start(), sizeof(char)) >= os::large_page_size()) { | |
| 439 if (intersection.contains(p)) { | |
| 440 intersection = MemRegion(p, intersection.end()); | |
| 441 } else { | |
| 442 intersection = MemRegion(p, p); | |
| 443 } | |
| 444 } | |
| 445 } | |
| 446 *bottom_region = MemRegion(new_region.start(), intersection.start()); | |
| 447 } else { | |
| 448 *bottom_region = MemRegion(); | |
| 449 } | |
| 450 | |
| 451 // Is there top? | |
| 452 if (intersection.end() < new_region.end()) { // Yes | |
| 453 // Try to coalesce small pages into a large one. | |
| 454 if (UseLargePages && page_size() >= os::large_page_size()) { | |
| 455 HeapWord* p = (HeapWord*)round_down((intptr_t) intersection.end(), os::large_page_size()); | |
| 456 if (new_region.contains(p) | |
| 457 && pointer_delta(new_region.end(), p, sizeof(char)) >= os::large_page_size()) { | |
| 458 if (intersection.contains(p)) { | |
| 459 intersection = MemRegion(intersection.start(), p); | |
| 460 } else { | |
| 461 intersection = MemRegion(p, p); | |
| 462 } | |
| 463 } | |
| 464 } | |
| 465 *top_region = MemRegion(intersection.end(), new_region.end()); | |
| 466 } else { | |
| 467 *top_region = MemRegion(); | |
| 468 } | |
| 469 } | |
| 470 | |
| 471 // Try to merge the invalid region with the bottom or top region by decreasing | |
| 472 // the intersection area. Return the invalid_region aligned to the page_size() | |
| 473 // boundary if it's inside the intersection. Return non-empty invalid_region | |
| 474 // if it lies inside the intersection (also page-aligned). | |
| 475 // |------------------new_region---------------------------------| | |
| 476 // |----------------|-------invalid---|--------------------------| | |
| 477 // |----bottom_region--|---intersection---|------top_region------| | |
| 478 void MutableNUMASpace::merge_regions(MemRegion new_region, MemRegion* intersection, | |
| 479 MemRegion *invalid_region) { | |
| 480 if (intersection->start() >= invalid_region->start() && intersection->contains(invalid_region->end())) { | |
| 481 *intersection = MemRegion(invalid_region->end(), intersection->end()); | |
| 482 *invalid_region = MemRegion(); | |
| 483 } else | |
| 484 if (intersection->end() <= invalid_region->end() && intersection->contains(invalid_region->start())) { | |
| 485 *intersection = MemRegion(intersection->start(), invalid_region->start()); | |
| 486 *invalid_region = MemRegion(); | |
| 487 } else | |
| 488 if (intersection->equals(*invalid_region) || invalid_region->contains(*intersection)) { | |
| 489 *intersection = MemRegion(new_region.start(), new_region.start()); | |
| 490 *invalid_region = MemRegion(); | |
| 491 } else | |
| 492 if (intersection->contains(invalid_region)) { | |
| 493 // That's the only case we have to make an additional bias_region() call. | |
| 494 HeapWord* start = invalid_region->start(); | |
| 495 HeapWord* end = invalid_region->end(); | |
| 496 if (UseLargePages && page_size() >= os::large_page_size()) { | |
| 497 HeapWord *p = (HeapWord*)round_down((intptr_t) start, os::large_page_size()); | |
| 498 if (new_region.contains(p)) { | |
| 499 start = p; | |
| 500 } | |
| 501 p = (HeapWord*)round_to((intptr_t) end, os::large_page_size()); | |
| 502 if (new_region.contains(end)) { | |
| 503 end = p; | |
| 504 } | |
| 505 } | |
| 506 if (intersection->start() > start) { | |
| 507 *intersection = MemRegion(start, intersection->end()); | |
| 508 } | |
| 509 if (intersection->end() < end) { | |
| 510 *intersection = MemRegion(intersection->start(), end); | |
| 511 } | |
| 512 *invalid_region = MemRegion(start, end); | |
| 513 } | |
| 514 } | |
| 515 | |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
516 void MutableNUMASpace::initialize(MemRegion mr, |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
517 bool clear_space, |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
518 bool mangle_space) { |
| 0 | 519 assert(clear_space, "Reallocation will destory data!"); |
| 520 assert(lgrp_spaces()->length() > 0, "There should be at least one space"); | |
| 521 | |
| 522 MemRegion old_region = region(), new_region; | |
| 523 set_bottom(mr.start()); | |
| 524 set_end(mr.end()); | |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
525 // Must always clear the space |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
526 clear(SpaceDecorator::DontMangle); |
| 0 | 527 |
| 528 // Compute chunk sizes | |
| 529 size_t prev_page_size = page_size(); | |
| 530 set_page_size(UseLargePages ? os::large_page_size() : os::vm_page_size()); | |
| 531 HeapWord* rounded_bottom = (HeapWord*)round_to((intptr_t) bottom(), page_size()); | |
| 532 HeapWord* rounded_end = (HeapWord*)round_down((intptr_t) end(), page_size()); | |
| 533 size_t base_space_size_pages = pointer_delta(rounded_end, rounded_bottom, sizeof(char)) / page_size(); | |
| 534 | |
| 535 // Try small pages if the chunk size is too small | |
| 536 if (base_space_size_pages / lgrp_spaces()->length() == 0 | |
| 537 && page_size() > (size_t)os::vm_page_size()) { | |
| 538 set_page_size(os::vm_page_size()); | |
| 539 rounded_bottom = (HeapWord*)round_to((intptr_t) bottom(), page_size()); | |
| 540 rounded_end = (HeapWord*)round_down((intptr_t) end(), page_size()); | |
| 541 base_space_size_pages = pointer_delta(rounded_end, rounded_bottom, sizeof(char)) / page_size(); | |
| 542 } | |
| 543 guarantee(base_space_size_pages / lgrp_spaces()->length() > 0, "Space too small"); | |
| 544 set_base_space_size(base_space_size_pages); | |
| 545 | |
| 546 // Handle space resize | |
| 547 MemRegion top_region, bottom_region; | |
| 548 if (!old_region.equals(region())) { | |
| 549 new_region = MemRegion(rounded_bottom, rounded_end); | |
| 550 MemRegion intersection = new_region.intersection(old_region); | |
| 551 if (intersection.start() == NULL || | |
| 552 intersection.end() == NULL || | |
| 553 prev_page_size > page_size()) { // If the page size got smaller we have to change | |
| 554 // the page size preference for the whole space. | |
| 555 intersection = MemRegion(new_region.start(), new_region.start()); | |
| 556 } | |
| 557 select_tails(new_region, intersection, &bottom_region, &top_region); | |
| 141 | 558 bias_region(bottom_region, lgrp_spaces()->at(0)->lgrp_id()); |
| 559 bias_region(top_region, lgrp_spaces()->at(lgrp_spaces()->length() - 1)->lgrp_id()); | |
| 0 | 560 } |
| 561 | |
| 562 // Check if the space layout has changed significantly? | |
| 563 // This happens when the space has been resized so that either head or tail | |
| 564 // chunk became less than a page. | |
| 565 bool layout_valid = UseAdaptiveNUMAChunkSizing && | |
| 566 current_chunk_size(0) > page_size() && | |
| 567 current_chunk_size(lgrp_spaces()->length() - 1) > page_size(); | |
| 568 | |
| 569 | |
| 570 for (int i = 0; i < lgrp_spaces()->length(); i++) { | |
| 571 LGRPSpace *ls = lgrp_spaces()->at(i); | |
| 572 MutableSpace *s = ls->space(); | |
| 573 old_region = s->region(); | |
| 574 | |
| 575 size_t chunk_byte_size = 0, old_chunk_byte_size = 0; | |
| 576 if (i < lgrp_spaces()->length() - 1) { | |
| 577 if (!UseAdaptiveNUMAChunkSizing || | |
| 578 (UseAdaptiveNUMAChunkSizing && NUMAChunkResizeWeight == 0) || | |
| 579 samples_count() < AdaptiveSizePolicyReadyThreshold) { | |
| 580 // No adaptation. Divide the space equally. | |
| 581 chunk_byte_size = default_chunk_size(); | |
| 582 } else | |
| 583 if (!layout_valid || NUMASpaceResizeRate == 0) { | |
| 584 // Fast adaptation. If no space resize rate is set, resize | |
| 585 // the chunks instantly. | |
| 586 chunk_byte_size = adaptive_chunk_size(i, 0); | |
| 587 } else { | |
| 588 // Slow adaptation. Resize the chunks moving no more than | |
| 589 // NUMASpaceResizeRate bytes per collection. | |
| 590 size_t limit = NUMASpaceResizeRate / | |
| 591 (lgrp_spaces()->length() * (lgrp_spaces()->length() + 1) / 2); | |
| 592 chunk_byte_size = adaptive_chunk_size(i, MAX2(limit * (i + 1), page_size())); | |
| 593 } | |
| 594 | |
| 595 assert(chunk_byte_size >= page_size(), "Chunk size too small"); | |
| 596 assert(chunk_byte_size <= capacity_in_bytes(), "Sanity check"); | |
| 597 } | |
| 598 | |
| 599 if (i == 0) { // Bottom chunk | |
| 600 if (i != lgrp_spaces()->length() - 1) { | |
| 601 new_region = MemRegion(bottom(), rounded_bottom + (chunk_byte_size >> LogHeapWordSize)); | |
| 602 } else { | |
| 603 new_region = MemRegion(bottom(), end()); | |
| 604 } | |
| 605 } else | |
| 606 if (i < lgrp_spaces()->length() - 1) { // Middle chunks | |
| 607 MutableSpace *ps = lgrp_spaces()->at(i - 1)->space(); | |
| 608 new_region = MemRegion(ps->end(), | |
| 609 ps->end() + (chunk_byte_size >> LogHeapWordSize)); | |
| 610 } else { // Top chunk | |
| 611 MutableSpace *ps = lgrp_spaces()->at(i - 1)->space(); | |
| 612 new_region = MemRegion(ps->end(), end()); | |
| 613 } | |
| 614 guarantee(region().contains(new_region), "Region invariant"); | |
| 615 | |
| 616 | |
| 617 // The general case: | |
| 618 // |---------------------|--invalid---|--------------------------| | |
| 619 // |------------------new_region---------------------------------| | |
| 620 // |----bottom_region--|---intersection---|------top_region------| | |
| 621 // |----old_region----| | |
| 622 // The intersection part has all pages in place we don't need to migrate them. | |
| 623 // Pages for the top and bottom part should be freed and then reallocated. | |
| 624 | |
| 625 MemRegion intersection = old_region.intersection(new_region); | |
| 626 | |
| 627 if (intersection.start() == NULL || intersection.end() == NULL) { | |
| 628 intersection = MemRegion(new_region.start(), new_region.start()); | |
| 629 } | |
| 630 | |
| 141 | 631 if (!os::numa_has_static_binding()) { |
| 632 MemRegion invalid_region = ls->invalid_region().intersection(new_region); | |
| 633 // Invalid region is a range of memory that could've possibly | |
| 634 // been allocated on the other node. That's relevant only on Solaris where | |
| 635 // there is no static memory binding. | |
| 636 if (!invalid_region.is_empty()) { | |
| 637 merge_regions(new_region, &intersection, &invalid_region); | |
| 638 free_region(invalid_region); | |
| 639 ls->set_invalid_region(MemRegion()); | |
| 640 } | |
| 0 | 641 } |
| 141 | 642 |
| 0 | 643 select_tails(new_region, intersection, &bottom_region, &top_region); |
| 141 | 644 |
| 645 if (!os::numa_has_static_binding()) { | |
| 646 // If that's a system with the first-touch policy then it's enough | |
| 647 // to free the pages. | |
| 648 free_region(bottom_region); | |
| 649 free_region(top_region); | |
| 650 } else { | |
| 651 // In a system with static binding we have to change the bias whenever | |
| 652 // we reshape the heap. | |
| 653 bias_region(bottom_region, ls->lgrp_id()); | |
| 654 bias_region(top_region, ls->lgrp_id()); | |
| 655 } | |
| 0 | 656 |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
657 // Clear space (set top = bottom) but never mangle. |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
658 s->initialize(new_region, SpaceDecorator::Clear, SpaceDecorator::DontMangle); |
| 0 | 659 |
| 660 set_adaptation_cycles(samples_count()); | |
| 661 } | |
| 662 } | |
| 663 | |
| 664 // Set the top of the whole space. | |
| 665 // Mark the the holes in chunks below the top() as invalid. | |
| 666 void MutableNUMASpace::set_top(HeapWord* value) { | |
| 667 bool found_top = false; | |
|
190
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
668 for (int i = 0; i < lgrp_spaces()->length();) { |
| 0 | 669 LGRPSpace *ls = lgrp_spaces()->at(i); |
| 670 MutableSpace *s = ls->space(); | |
| 671 HeapWord *top = MAX2((HeapWord*)round_down((intptr_t)s->top(), page_size()), s->bottom()); | |
| 672 | |
| 673 if (s->contains(value)) { | |
|
190
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
674 // Check if setting the chunk's top to a given value would create a hole less than |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
675 // a minimal object; assuming that's not the last chunk in which case we don't care. |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
676 if (i < lgrp_spaces()->length() - 1) { |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
677 size_t remainder = pointer_delta(s->end(), value); |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
678 const size_t minimal_object_size = oopDesc::header_size(); |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
679 if (remainder < minimal_object_size && remainder > 0) { |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
680 // Add a filler object of a minimal size, it will cross the chunk boundary. |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
681 SharedHeap::fill_region_with_object(MemRegion(value, minimal_object_size)); |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
682 value += minimal_object_size; |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
683 assert(!s->contains(value), "Should be in the next chunk"); |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
684 // Restart the loop from the same chunk, since the value has moved |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
685 // to the next one. |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
686 continue; |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
687 } |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
688 } |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
689 |
| 141 | 690 if (!os::numa_has_static_binding() && top < value && top < s->end()) { |
| 0 | 691 ls->add_invalid_region(MemRegion(top, value)); |
| 692 } | |
| 693 s->set_top(value); | |
| 694 found_top = true; | |
| 695 } else { | |
| 696 if (found_top) { | |
| 697 s->set_top(s->bottom()); | |
| 698 } else { | |
| 141 | 699 if (!os::numa_has_static_binding() && top < s->end()) { |
| 700 ls->add_invalid_region(MemRegion(top, s->end())); | |
| 701 } | |
| 702 s->set_top(s->end()); | |
| 0 | 703 } |
| 704 } | |
|
190
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
705 i++; |
| 0 | 706 } |
| 707 MutableSpace::set_top(value); | |
| 708 } | |
| 709 | |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
710 void MutableNUMASpace::clear(bool mangle_space) { |
| 0 | 711 MutableSpace::set_top(bottom()); |
| 712 for (int i = 0; i < lgrp_spaces()->length(); i++) { | |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
713 // Never mangle NUMA spaces because the mangling will |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
714 // bind the memory to a possibly unwanted lgroup. |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
190
diff
changeset
|
715 lgrp_spaces()->at(i)->space()->clear(SpaceDecorator::DontMangle); |
| 0 | 716 } |
| 717 } | |
| 718 | |
| 141 | 719 /* |
| 720 Linux supports static memory binding, therefore the most part of the | |
| 721 logic dealing with the possible invalid page allocation is effectively | |
| 722 disabled. Besides there is no notion of the home node in Linux. A | |
| 723 thread is allowed to migrate freely. Although the scheduler is rather | |
| 724 reluctant to move threads between the nodes. We check for the current | |
| 725 node every allocation. And with a high probability a thread stays on | |
| 726 the same node for some time allowing local access to recently allocated | |
| 727 objects. | |
| 728 */ | |
| 729 | |
| 0 | 730 HeapWord* MutableNUMASpace::allocate(size_t size) { |
| 141 | 731 Thread* thr = Thread::current(); |
| 732 int lgrp_id = thr->lgrp_id(); | |
| 733 if (lgrp_id == -1 || !os::numa_has_group_homing()) { | |
| 0 | 734 lgrp_id = os::numa_get_group_id(); |
| 141 | 735 thr->set_lgrp_id(lgrp_id); |
| 0 | 736 } |
| 737 | |
| 738 int i = lgrp_spaces()->find(&lgrp_id, LGRPSpace::equals); | |
| 739 | |
| 740 // It is possible that a new CPU has been hotplugged and | |
| 741 // we haven't reshaped the space accordingly. | |
| 742 if (i == -1) { | |
| 743 i = os::random() % lgrp_spaces()->length(); | |
| 744 } | |
| 745 | |
|
373
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
746 LGRPSpace* ls = lgrp_spaces()->at(i); |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
747 MutableSpace *s = ls->space(); |
| 0 | 748 HeapWord *p = s->allocate(size); |
| 749 | |
|
144
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
750 if (p != NULL) { |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
751 size_t remainder = s->free_in_words(); |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
752 if (remainder < (size_t)oopDesc::header_size() && remainder > 0) { |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
753 s->set_top(s->top() - size); |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
754 p = NULL; |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
755 } |
| 0 | 756 } |
| 757 if (p != NULL) { | |
| 758 if (top() < s->top()) { // Keep _top updated. | |
| 759 MutableSpace::set_top(s->top()); | |
| 760 } | |
| 761 } | |
| 141 | 762 // Make the page allocation happen here if there is no static binding.. |
| 763 if (p != NULL && !os::numa_has_static_binding()) { | |
| 0 | 764 for (HeapWord *i = p; i < p + size; i += os::vm_page_size() >> LogHeapWordSize) { |
| 765 *(int*)i = 0; | |
| 766 } | |
| 767 } | |
|
373
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
768 if (p == NULL) { |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
769 ls->set_allocation_failed(); |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
770 } |
| 0 | 771 return p; |
| 772 } | |
| 773 | |
| 774 // This version is lock-free. | |
| 775 HeapWord* MutableNUMASpace::cas_allocate(size_t size) { | |
| 141 | 776 Thread* thr = Thread::current(); |
| 777 int lgrp_id = thr->lgrp_id(); | |
| 778 if (lgrp_id == -1 || !os::numa_has_group_homing()) { | |
| 0 | 779 lgrp_id = os::numa_get_group_id(); |
| 141 | 780 thr->set_lgrp_id(lgrp_id); |
| 0 | 781 } |
| 782 | |
| 783 int i = lgrp_spaces()->find(&lgrp_id, LGRPSpace::equals); | |
| 784 // It is possible that a new CPU has been hotplugged and | |
| 785 // we haven't reshaped the space accordingly. | |
| 786 if (i == -1) { | |
| 787 i = os::random() % lgrp_spaces()->length(); | |
| 788 } | |
|
373
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
789 LGRPSpace *ls = lgrp_spaces()->at(i); |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
790 MutableSpace *s = ls->space(); |
| 0 | 791 HeapWord *p = s->cas_allocate(size); |
|
144
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
792 if (p != NULL) { |
|
190
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
793 size_t remainder = pointer_delta(s->end(), p + size); |
|
144
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
794 if (remainder < (size_t)oopDesc::header_size() && remainder > 0) { |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
795 if (s->cas_deallocate(p, size)) { |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
796 // We were the last to allocate and created a fragment less than |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
797 // a minimal object. |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
798 p = NULL; |
|
190
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
799 } else { |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
800 guarantee(false, "Deallocation should always succeed"); |
|
144
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
801 } |
| 0 | 802 } |
| 803 } | |
| 804 if (p != NULL) { | |
| 805 HeapWord* cur_top, *cur_chunk_top = p + size; | |
| 806 while ((cur_top = top()) < cur_chunk_top) { // Keep _top updated. | |
| 807 if (Atomic::cmpxchg_ptr(cur_chunk_top, top_addr(), cur_top) == cur_top) { | |
| 808 break; | |
| 809 } | |
| 810 } | |
| 811 } | |
| 812 | |
| 141 | 813 // Make the page allocation happen here if there is no static binding. |
| 814 if (p != NULL && !os::numa_has_static_binding() ) { | |
| 0 | 815 for (HeapWord *i = p; i < p + size; i += os::vm_page_size() >> LogHeapWordSize) { |
| 816 *(int*)i = 0; | |
| 817 } | |
| 818 } | |
|
373
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
819 if (p == NULL) { |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
820 ls->set_allocation_failed(); |
|
06df86c2ec37
6740923: NUMA allocator: Ensure the progress of adaptive chunk resizing
iveresov
parents:
269
diff
changeset
|
821 } |
| 0 | 822 return p; |
| 823 } | |
| 824 | |
| 825 void MutableNUMASpace::print_short_on(outputStream* st) const { | |
| 826 MutableSpace::print_short_on(st); | |
| 827 st->print(" ("); | |
| 828 for (int i = 0; i < lgrp_spaces()->length(); i++) { | |
| 829 st->print("lgrp %d: ", lgrp_spaces()->at(i)->lgrp_id()); | |
| 830 lgrp_spaces()->at(i)->space()->print_short_on(st); | |
| 831 if (i < lgrp_spaces()->length() - 1) { | |
| 832 st->print(", "); | |
| 833 } | |
| 834 } | |
| 835 st->print(")"); | |
| 836 } | |
| 837 | |
| 838 void MutableNUMASpace::print_on(outputStream* st) const { | |
| 839 MutableSpace::print_on(st); | |
| 840 for (int i = 0; i < lgrp_spaces()->length(); i++) { | |
| 841 LGRPSpace *ls = lgrp_spaces()->at(i); | |
| 842 st->print(" lgrp %d", ls->lgrp_id()); | |
| 843 ls->space()->print_on(st); | |
| 844 if (NUMAStats) { | |
|
144
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
845 for (int i = 0; i < lgrp_spaces()->length(); i++) { |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
846 lgrp_spaces()->at(i)->accumulate_statistics(page_size()); |
|
e3729351c946
6697534: Premature GC and invalid lgrp selection with NUMA-aware allocator.
iveresov
parents:
141
diff
changeset
|
847 } |
| 0 | 848 st->print(" local/remote/unbiased/uncommitted: %dK/%dK/%dK/%dK, large/small pages: %d/%d\n", |
| 849 ls->space_stats()->_local_space / K, | |
| 850 ls->space_stats()->_remote_space / K, | |
| 851 ls->space_stats()->_unbiased_space / K, | |
| 852 ls->space_stats()->_uncommited_space / K, | |
| 853 ls->space_stats()->_large_pages, | |
| 854 ls->space_stats()->_small_pages); | |
| 855 } | |
| 856 } | |
| 857 } | |
| 858 | |
|
190
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
859 void MutableNUMASpace::verify(bool allow_dirty) { |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
860 // This can be called after setting an arbitary value to the space's top, |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
861 // so an object can cross the chunk boundary. We ensure the parsablity |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
862 // of the space and just walk the objects in linear fashion. |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
863 ensure_parsability(); |
|
d1635bf93939
6711930: NUMA allocator: ParOld can create a hole less than minimal object size in the lgrp chunk
iveresov
parents:
144
diff
changeset
|
864 MutableSpace::verify(allow_dirty); |
| 0 | 865 } |
| 866 | |
| 867 // Scan pages and gather stats about page placement and size. | |
| 868 void MutableNUMASpace::LGRPSpace::accumulate_statistics(size_t page_size) { | |
| 869 clear_space_stats(); | |
| 870 char *start = (char*)round_to((intptr_t) space()->bottom(), page_size); | |
| 871 char* end = (char*)round_down((intptr_t) space()->end(), page_size); | |
| 872 if (start < end) { | |
| 873 for (char *p = start; p < end;) { | |
| 874 os::page_info info; | |
| 875 if (os::get_page_info(p, &info)) { | |
| 876 if (info.size > 0) { | |
| 877 if (info.size > (size_t)os::vm_page_size()) { | |
| 878 space_stats()->_large_pages++; | |
| 879 } else { | |
| 880 space_stats()->_small_pages++; | |
| 881 } | |
| 882 if (info.lgrp_id == lgrp_id()) { | |
| 883 space_stats()->_local_space += info.size; | |
| 884 } else { | |
| 885 space_stats()->_remote_space += info.size; | |
| 886 } | |
| 887 p += info.size; | |
| 888 } else { | |
| 889 p += os::vm_page_size(); | |
| 890 space_stats()->_uncommited_space += os::vm_page_size(); | |
| 891 } | |
| 892 } else { | |
| 893 return; | |
| 894 } | |
| 895 } | |
| 896 } | |
| 897 space_stats()->_unbiased_space = pointer_delta(start, space()->bottom(), sizeof(char)) + | |
| 898 pointer_delta(space()->end(), end, sizeof(char)); | |
| 899 | |
| 900 } | |
| 901 | |
| 902 // Scan page_count pages and verify if they have the right size and right placement. | |
| 903 // If invalid pages are found they are freed in hope that subsequent reallocation | |
| 904 // will be more successful. | |
| 905 void MutableNUMASpace::LGRPSpace::scan_pages(size_t page_size, size_t page_count) | |
| 906 { | |
| 907 char* range_start = (char*)round_to((intptr_t) space()->bottom(), page_size); | |
| 908 char* range_end = (char*)round_down((intptr_t) space()->end(), page_size); | |
| 909 | |
| 910 if (range_start > last_page_scanned() || last_page_scanned() >= range_end) { | |
| 911 set_last_page_scanned(range_start); | |
| 912 } | |
| 913 | |
| 914 char *scan_start = last_page_scanned(); | |
| 915 char* scan_end = MIN2(scan_start + page_size * page_count, range_end); | |
| 916 | |
| 917 os::page_info page_expected, page_found; | |
| 918 page_expected.size = page_size; | |
| 919 page_expected.lgrp_id = lgrp_id(); | |
| 920 | |
| 921 char *s = scan_start; | |
| 922 while (s < scan_end) { | |
| 923 char *e = os::scan_pages(s, (char*)scan_end, &page_expected, &page_found); | |
| 924 if (e == NULL) { | |
| 925 break; | |
| 926 } | |
| 927 if (e != scan_end) { | |
| 928 if ((page_expected.size != page_size || page_expected.lgrp_id != lgrp_id()) | |
| 929 && page_expected.size != 0) { | |
| 930 os::free_memory(s, pointer_delta(e, s, sizeof(char))); | |
| 931 } | |
| 932 page_expected = page_found; | |
| 933 } | |
| 934 s = e; | |
| 935 } | |
| 936 | |
| 937 set_last_page_scanned(scan_end); | |
| 938 } |