Mon, 17 May 2010 00:47:28 -0700
6948539: CMS+UseCompressedOops: placement of cms_free bit interferes with promoted object link
Summary: When using compressed oops, use compressed promoted pointers in b63:b31 of the mark word, so as not to interfere with the CMS "freeness bit" at b7. Updated mark-word layout documentation.
Reviewed-by: minqi, poonam, jmasa, coleenp
1 /*
2 * Copyright 1997-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 // The markOop describes the header of an object.
26 //
27 // Note that the mark is not a real oop but just a word.
28 // It is placed in the oop hierarchy for historical reasons.
29 //
30 // Bit-format of an object header (most significant first, big endian layout below):
31 //
32 // 32 bits:
33 // --------
34 // hash:25 ------------>| age:4 biased_lock:1 lock:2 (normal object)
35 // JavaThread*:23 epoch:2 age:4 biased_lock:1 lock:2 (biased object)
36 // size:32 ------------------------------------------>| (CMS free block)
37 // PromotedObject*:29 ---------->| promo_bits:3 ----->| (CMS promoted object)
38 //
39 // 64 bits:
40 // --------
41 // unused:25 hash:31 -->| unused:1 age:4 biased_lock:1 lock:2 (normal object)
42 // JavaThread*:54 epoch:2 unused:1 age:4 biased_lock:1 lock:2 (biased object)
43 // PromotedObject*:61 --------------------->| promo_bits:3 ----->| (CMS promoted object)
44 // size:64 ----------------------------------------------------->| (CMS free block)
45 //
46 // unused:25 hash:31 -->| cms_free:1 age:4 biased_lock:1 lock:2 (COOPs && normal object)
47 // JavaThread*:54 epoch:2 cms_free:1 age:4 biased_lock:1 lock:2 (COOPs && biased object)
48 // narrowOop:32 unused:24 cms_free:1 unused:4 promo_bits:3 ----->| (COOPs && CMS promoted object)
49 // unused:21 size:35 -->| cms_free:1 unused:7 ------------------>| (COOPs && CMS free block)
50 //
51 // - hash contains the identity hash value: largest value is
52 // 31 bits, see os::random(). Also, 64-bit vm's require
53 // a hash value no bigger than 32 bits because they will not
54 // properly generate a mask larger than that: see library_call.cpp
55 // and c1_CodePatterns_sparc.cpp.
56 //
57 // - the biased lock pattern is used to bias a lock toward a given
58 // thread. When this pattern is set in the low three bits, the lock
59 // is either biased toward a given thread or "anonymously" biased,
60 // indicating that it is possible for it to be biased. When the
61 // lock is biased toward a given thread, locking and unlocking can
62 // be performed by that thread without using atomic operations.
63 // When a lock's bias is revoked, it reverts back to the normal
64 // locking scheme described below.
65 //
66 // Note that we are overloading the meaning of the "unlocked" state
67 // of the header. Because we steal a bit from the age we can
68 // guarantee that the bias pattern will never be seen for a truly
69 // unlocked object.
70 //
71 // Note also that the biased state contains the age bits normally
72 // contained in the object header. Large increases in scavenge
73 // times were seen when these bits were absent and an arbitrary age
74 // assigned to all biased objects, because they tended to consume a
75 // significant fraction of the eden semispaces and were not
76 // promoted promptly, causing an increase in the amount of copying
77 // performed. The runtime system aligns all JavaThread* pointers to
78 // a very large value (currently 128 bytes (32bVM) or 256 bytes (64bVM))
79 // to make room for the age bits & the epoch bits (used in support of
80 // biased locking), and for the CMS "freeness" bit in the 64bVM (+COOPs).
81 //
82 // [JavaThread* | epoch | age | 1 | 01] lock is biased toward given thread
83 // [0 | epoch | age | 1 | 01] lock is anonymously biased
84 //
85 // - the two lock bits are used to describe three states: locked/unlocked and monitor.
86 //
87 // [ptr | 00] locked ptr points to real header on stack
88 // [header | 0 | 01] unlocked regular object header
89 // [ptr | 10] monitor inflated lock (header is wapped out)
90 // [ptr | 11] marked used by markSweep to mark an object
91 // not valid at any other time
92 //
93 // We assume that stack/thread pointers have the lowest two bits cleared.
95 class BasicLock;
96 class ObjectMonitor;
97 class JavaThread;
99 class markOopDesc: public oopDesc {
100 private:
101 // Conversion
102 uintptr_t value() const { return (uintptr_t) this; }
104 public:
105 // Constants
106 enum { age_bits = 4,
107 lock_bits = 2,
108 biased_lock_bits = 1,
109 max_hash_bits = BitsPerWord - age_bits - lock_bits - biased_lock_bits,
110 hash_bits = max_hash_bits > 31 ? 31 : max_hash_bits,
111 cms_bits = LP64_ONLY(1) NOT_LP64(0),
112 epoch_bits = 2
113 };
115 // The biased locking code currently requires that the age bits be
116 // contiguous to the lock bits. Class data sharing would prefer the
117 // hash bits to be lower down to provide more random hash codes for
118 // shared read-only symbolOop objects, because these objects' mark
119 // words are set to their own address with marked_value in the lock
120 // bit, and using lower bits would make their identity hash values
121 // more random. However, the performance decision was made in favor
122 // of the biased locking code.
124 enum { lock_shift = 0,
125 biased_lock_shift = lock_bits,
126 age_shift = lock_bits + biased_lock_bits,
127 cms_shift = age_shift + age_bits,
128 hash_shift = cms_shift + cms_bits,
129 epoch_shift = hash_shift
130 };
132 enum { lock_mask = right_n_bits(lock_bits),
133 lock_mask_in_place = lock_mask << lock_shift,
134 biased_lock_mask = right_n_bits(lock_bits + biased_lock_bits),
135 biased_lock_mask_in_place= biased_lock_mask << lock_shift,
136 biased_lock_bit_in_place = 1 << biased_lock_shift,
137 age_mask = right_n_bits(age_bits),
138 age_mask_in_place = age_mask << age_shift,
139 epoch_mask = right_n_bits(epoch_bits),
140 epoch_mask_in_place = epoch_mask << epoch_shift,
141 cms_mask = right_n_bits(cms_bits),
142 cms_mask_in_place = cms_mask << cms_shift
143 #ifndef _WIN64
144 ,hash_mask = right_n_bits(hash_bits),
145 hash_mask_in_place = (address_word)hash_mask << hash_shift
146 #endif
147 };
149 // Alignment of JavaThread pointers encoded in object header required by biased locking
150 enum { biased_lock_alignment = 2 << (epoch_shift + epoch_bits)
151 };
153 #ifdef _WIN64
154 // These values are too big for Win64
155 const static uintptr_t hash_mask = right_n_bits(hash_bits);
156 const static uintptr_t hash_mask_in_place =
157 (address_word)hash_mask << hash_shift;
158 #endif
160 enum { locked_value = 0,
161 unlocked_value = 1,
162 monitor_value = 2,
163 marked_value = 3,
164 biased_lock_pattern = 5
165 };
167 enum { no_hash = 0 }; // no hash value assigned
169 enum { no_hash_in_place = (address_word)no_hash << hash_shift,
170 no_lock_in_place = unlocked_value
171 };
173 enum { max_age = age_mask };
175 enum { max_bias_epoch = epoch_mask };
177 // Biased Locking accessors.
178 // These must be checked by all code which calls into the
179 // ObjectSynchronizer and other code. The biasing is not understood
180 // by the lower-level CAS-based locking code, although the runtime
181 // fixes up biased locks to be compatible with it when a bias is
182 // revoked.
183 bool has_bias_pattern() const {
184 return (mask_bits(value(), biased_lock_mask_in_place) == biased_lock_pattern);
185 }
186 JavaThread* biased_locker() const {
187 assert(has_bias_pattern(), "should not call this otherwise");
188 return (JavaThread*) ((intptr_t) (mask_bits(value(), ~(biased_lock_mask_in_place | age_mask_in_place | epoch_mask_in_place))));
189 }
190 // Indicates that the mark has the bias bit set but that it has not
191 // yet been biased toward a particular thread
192 bool is_biased_anonymously() const {
193 return (has_bias_pattern() && (biased_locker() == NULL));
194 }
195 // Indicates epoch in which this bias was acquired. If the epoch
196 // changes due to too many bias revocations occurring, the biases
197 // from the previous epochs are all considered invalid.
198 int bias_epoch() const {
199 assert(has_bias_pattern(), "should not call this otherwise");
200 return (mask_bits(value(), epoch_mask_in_place) >> epoch_shift);
201 }
202 markOop set_bias_epoch(int epoch) {
203 assert(has_bias_pattern(), "should not call this otherwise");
204 assert((epoch & (~epoch_mask)) == 0, "epoch overflow");
205 return markOop(mask_bits(value(), ~epoch_mask_in_place) | (epoch << epoch_shift));
206 }
207 markOop incr_bias_epoch() {
208 return set_bias_epoch((1 + bias_epoch()) & epoch_mask);
209 }
210 // Prototype mark for initialization
211 static markOop biased_locking_prototype() {
212 return markOop( biased_lock_pattern );
213 }
215 // lock accessors (note that these assume lock_shift == 0)
216 bool is_locked() const {
217 return (mask_bits(value(), lock_mask_in_place) != unlocked_value);
218 }
219 bool is_unlocked() const {
220 return (mask_bits(value(), biased_lock_mask_in_place) == unlocked_value);
221 }
222 bool is_marked() const {
223 return (mask_bits(value(), lock_mask_in_place) == marked_value);
224 }
225 bool is_neutral() const { return (mask_bits(value(), biased_lock_mask_in_place) == unlocked_value); }
227 // Special temporary state of the markOop while being inflated.
228 // Code that looks at mark outside a lock need to take this into account.
229 bool is_being_inflated() const { return (value() == 0); }
231 // Distinguished markword value - used when inflating over
232 // an existing stacklock. 0 indicates the markword is "BUSY".
233 // Lockword mutators that use a LD...CAS idiom should always
234 // check for and avoid overwriting a 0 value installed by some
235 // other thread. (They should spin or block instead. The 0 value
236 // is transient and *should* be short-lived).
237 static markOop INFLATING() { return (markOop) 0; } // inflate-in-progress
239 // Should this header be preserved during GC?
240 inline bool must_be_preserved(oop obj_containing_mark) const;
241 inline bool must_be_preserved_with_bias(oop obj_containing_mark) const;
243 // Should this header (including its age bits) be preserved in the
244 // case of a promotion failure during scavenge?
245 // Note that we special case this situation. We want to avoid
246 // calling BiasedLocking::preserve_marks()/restore_marks() (which
247 // decrease the number of mark words that need to be preserved
248 // during GC) during each scavenge. During scavenges in which there
249 // is no promotion failure, we actually don't need to call the above
250 // routines at all, since we don't mutate and re-initialize the
251 // marks of promoted objects using init_mark(). However, during
252 // scavenges which result in promotion failure, we do re-initialize
253 // the mark words of objects, meaning that we should have called
254 // these mark word preservation routines. Currently there's no good
255 // place in which to call them in any of the scavengers (although
256 // guarded by appropriate locks we could make one), but the
257 // observation is that promotion failures are quite rare and
258 // reducing the number of mark words preserved during them isn't a
259 // high priority.
260 inline bool must_be_preserved_for_promotion_failure(oop obj_containing_mark) const;
261 inline bool must_be_preserved_with_bias_for_promotion_failure(oop obj_containing_mark) const;
263 // Should this header be preserved during a scavenge where CMS is
264 // the old generation?
265 // (This is basically the same body as must_be_preserved_for_promotion_failure(),
266 // but takes the klassOop as argument instead)
267 inline bool must_be_preserved_for_cms_scavenge(klassOop klass_of_obj_containing_mark) const;
268 inline bool must_be_preserved_with_bias_for_cms_scavenge(klassOop klass_of_obj_containing_mark) const;
270 // WARNING: The following routines are used EXCLUSIVELY by
271 // synchronization functions. They are not really gc safe.
272 // They must get updated if markOop layout get changed.
273 markOop set_unlocked() const {
274 return markOop(value() | unlocked_value);
275 }
276 bool has_locker() const {
277 return ((value() & lock_mask_in_place) == locked_value);
278 }
279 BasicLock* locker() const {
280 assert(has_locker(), "check");
281 return (BasicLock*) value();
282 }
283 bool has_monitor() const {
284 return ((value() & monitor_value) != 0);
285 }
286 ObjectMonitor* monitor() const {
287 assert(has_monitor(), "check");
288 // Use xor instead of &~ to provide one extra tag-bit check.
289 return (ObjectMonitor*) (value() ^ monitor_value);
290 }
291 bool has_displaced_mark_helper() const {
292 return ((value() & unlocked_value) == 0);
293 }
294 markOop displaced_mark_helper() const {
295 assert(has_displaced_mark_helper(), "check");
296 intptr_t ptr = (value() & ~monitor_value);
297 return *(markOop*)ptr;
298 }
299 void set_displaced_mark_helper(markOop m) const {
300 assert(has_displaced_mark_helper(), "check");
301 intptr_t ptr = (value() & ~monitor_value);
302 *(markOop*)ptr = m;
303 }
304 markOop copy_set_hash(intptr_t hash) const {
305 intptr_t tmp = value() & (~hash_mask_in_place);
306 tmp |= ((hash & hash_mask) << hash_shift);
307 return (markOop)tmp;
308 }
309 // it is only used to be stored into BasicLock as the
310 // indicator that the lock is using heavyweight monitor
311 static markOop unused_mark() {
312 return (markOop) marked_value;
313 }
314 // the following two functions create the markOop to be
315 // stored into object header, it encodes monitor info
316 static markOop encode(BasicLock* lock) {
317 return (markOop) lock;
318 }
319 static markOop encode(ObjectMonitor* monitor) {
320 intptr_t tmp = (intptr_t) monitor;
321 return (markOop) (tmp | monitor_value);
322 }
323 static markOop encode(JavaThread* thread, int age, int bias_epoch) {
324 intptr_t tmp = (intptr_t) thread;
325 assert(UseBiasedLocking && ((tmp & (epoch_mask_in_place | age_mask_in_place | biased_lock_mask_in_place)) == 0), "misaligned JavaThread pointer");
326 assert(age <= max_age, "age too large");
327 assert(bias_epoch <= max_bias_epoch, "bias epoch too large");
328 return (markOop) (tmp | (bias_epoch << epoch_shift) | (age << age_shift) | biased_lock_pattern);
329 }
331 // used to encode pointers during GC
332 markOop clear_lock_bits() { return markOop(value() & ~lock_mask_in_place); }
334 // age operations
335 markOop set_marked() { return markOop((value() & ~lock_mask_in_place) | marked_value); }
337 int age() const { return mask_bits(value() >> age_shift, age_mask); }
338 markOop set_age(int v) const {
339 assert((v & ~age_mask) == 0, "shouldn't overflow age field");
340 return markOop((value() & ~age_mask_in_place) | (((intptr_t)v & age_mask) << age_shift));
341 }
342 markOop incr_age() const { return age() == max_age ? markOop(this) : set_age(age() + 1); }
344 // hash operations
345 intptr_t hash() const {
346 return mask_bits(value() >> hash_shift, hash_mask);
347 }
349 bool has_no_hash() const {
350 return hash() == no_hash;
351 }
353 // Prototype mark for initialization
354 static markOop prototype() {
355 return markOop( no_hash_in_place | no_lock_in_place );
356 }
358 // Helper function for restoration of unmarked mark oops during GC
359 static inline markOop prototype_for_object(oop obj);
361 // Debugging
362 void print_on(outputStream* st) const;
364 // Prepare address of oop for placement into mark
365 inline static markOop encode_pointer_as_mark(void* p) { return markOop(p)->set_marked(); }
367 // Recover address of oop from encoded form used in mark
368 inline void* decode_pointer() { if (UseBiasedLocking && has_bias_pattern()) return NULL; return clear_lock_bits(); }
370 // see the definition in markOop.cpp for the gory details
371 bool should_not_be_cached() const;
373 // These markOops indicate cms free chunk blocks and not objects.
374 // In 64 bit, the markOop is set to distinguish them from oops.
375 // These are defined in 32 bit mode for vmStructs.
376 const static uintptr_t cms_free_chunk_pattern = 0x1;
378 // Constants for the size field.
379 enum { size_shift = cms_shift + cms_bits,
380 size_bits = 35 // need for compressed oops 32G
381 };
382 // These values are too big for Win64
383 const static uintptr_t size_mask = LP64_ONLY(right_n_bits(size_bits))
384 NOT_LP64(0);
385 const static uintptr_t size_mask_in_place =
386 (address_word)size_mask << size_shift;
388 #ifdef _LP64
389 static markOop cms_free_prototype() {
390 return markOop(((intptr_t)prototype() & ~cms_mask_in_place) |
391 ((cms_free_chunk_pattern & cms_mask) << cms_shift));
392 }
393 uintptr_t cms_encoding() const {
394 return mask_bits(value() >> cms_shift, cms_mask);
395 }
396 bool is_cms_free_chunk() const {
397 return is_neutral() &&
398 (cms_encoding() & cms_free_chunk_pattern) == cms_free_chunk_pattern;
399 }
401 size_t get_size() const { return (size_t)(value() >> size_shift); }
402 static markOop set_size_and_free(size_t size) {
403 assert((size & ~size_mask) == 0, "shouldn't overflow size field");
404 return markOop(((intptr_t)cms_free_prototype() & ~size_mask_in_place) |
405 (((intptr_t)size & size_mask) << size_shift));
406 }
407 #endif // _LP64
408 };