1 /* 2 * Copyright (c) 1998, 2019, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/vmSymbols.hpp" 27 #include "logging/log.hpp" 28 #include "logging/logStream.hpp" 29 #include "jfr/jfrEvents.hpp" 30 #include "memory/allocation.inline.hpp" 31 #include "memory/metaspaceShared.hpp" 32 #include "memory/padded.hpp" 33 #include "memory/resourceArea.hpp" 34 #include "memory/universe.hpp" 35 #include "oops/markWord.hpp" 36 #include "oops/oop.inline.hpp" 37 #include "runtime/atomic.hpp" 38 #include "runtime/biasedLocking.hpp" 39 #include "runtime/handles.inline.hpp" 40 #include "runtime/interfaceSupport.inline.hpp" 41 #include "runtime/mutexLocker.hpp" 42 #include "runtime/objectMonitor.hpp" 43 #include "runtime/objectMonitor.inline.hpp" 44 #include "runtime/osThread.hpp" 45 #include "runtime/safepointVerifiers.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 #include "runtime/stubRoutines.hpp" 48 #include "runtime/synchronizer.hpp" 49 #include "runtime/thread.inline.hpp" 50 #include "runtime/timer.hpp" 51 #include "runtime/vframe.hpp" 52 #include "runtime/vmThread.hpp" 53 #include "utilities/align.hpp" 54 #include "utilities/dtrace.hpp" 55 #include "utilities/events.hpp" 56 #include "utilities/preserveException.hpp" 57 58 // The "core" versions of monitor enter and exit reside in this file. 59 // The interpreter and compilers contain specialized transliterated 60 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(), 61 // for instance. If you make changes here, make sure to modify the 62 // interpreter, and both C1 and C2 fast-path inline locking code emission. 63 // 64 // ----------------------------------------------------------------------------- 65 66 #ifdef DTRACE_ENABLED 67 68 // Only bother with this argument setup if dtrace is available 69 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. 70 71 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ 72 char* bytes = NULL; \ 73 int len = 0; \ 74 jlong jtid = SharedRuntime::get_java_tid(thread); \ 75 Symbol* klassname = ((oop)(obj))->klass()->name(); \ 76 if (klassname != NULL) { \ 77 bytes = (char*)klassname->bytes(); \ 78 len = klassname->utf8_length(); \ 79 } 80 81 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ 82 { \ 83 if (DTraceMonitorProbes) { \ 84 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 85 HOTSPOT_MONITOR_WAIT(jtid, \ 86 (uintptr_t)(monitor), bytes, len, (millis)); \ 87 } \ 88 } 89 90 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY 91 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL 92 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED 93 94 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ 95 { \ 96 if (DTraceMonitorProbes) { \ 97 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ 98 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \ 99 (uintptr_t)(monitor), bytes, len); \ 100 } \ 101 } 102 103 #else // ndef DTRACE_ENABLED 104 105 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} 106 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} 107 108 #endif // ndef DTRACE_ENABLED 109 110 // This exists only as a workaround of dtrace bug 6254741 111 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) { 112 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); 113 return 0; 114 } 115 116 #define NINFLATIONLOCKS 256 117 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS]; 118 119 // global list of blocks of monitors 120 PaddedObjectMonitor* ObjectSynchronizer::g_block_list = NULL; 121 122 struct ObjectMonitorListGlobals { 123 char _pad_prefix[OM_CACHE_LINE_SIZE]; 124 // These are highly shared list related variables. 125 // To avoid false-sharing they need to be the sole occupants of a cache line. 126 127 // Global ObjectMonitor free list. Newly allocated and deflated 128 // ObjectMonitors are prepended here. 129 ObjectMonitor* _free_list; 130 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(ObjectMonitor*)); 131 132 // Global ObjectMonitor in-use list. When a JavaThread is exiting, 133 // ObjectMonitors on its per-thread in-use list are prepended here. 134 ObjectMonitor* _in_use_list; 135 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(ObjectMonitor*)); 136 137 int _free_count; // # on free_list 138 DEFINE_PAD_MINUS_SIZE(3, OM_CACHE_LINE_SIZE, sizeof(int)); 139 140 int _in_use_count; // # on in_use_list 141 DEFINE_PAD_MINUS_SIZE(4, OM_CACHE_LINE_SIZE, sizeof(int)); 142 143 int _population; // # Extant -- in circulation 144 DEFINE_PAD_MINUS_SIZE(5, OM_CACHE_LINE_SIZE, sizeof(int)); 145 }; 146 static ObjectMonitorListGlobals om_list_globals; 147 148 #define CHAINMARKER (cast_to_oop<intptr_t>(-1)) 149 150 151 // =====================> Spin-lock functions 152 153 // ObjectMonitors are not lockable outside of this file. We use spin-locks 154 // implemented using a bit in the _next_om field instead of the heavier 155 // weight locking mechanisms for faster list management. 156 157 #define OM_LOCK_BIT 0x1 158 159 // Return true if the ObjectMonitor is locked. 160 // Otherwise returns false. 161 static bool is_locked(ObjectMonitor* om) { 162 return ((intptr_t)om->next_om() & OM_LOCK_BIT) == OM_LOCK_BIT; 163 } 164 165 // Mark an ObjectMonitor* with OM_LOCK_BIT and return it. 166 static ObjectMonitor* mark_om_ptr(ObjectMonitor* om) { 167 return (ObjectMonitor*)((intptr_t)om | OM_LOCK_BIT); 168 } 169 170 // Return the unmarked next field in an ObjectMonitor. Note: the next 171 // field may or may not have been marked with OM_LOCK_BIT originally. 172 static ObjectMonitor* unmarked_next(ObjectMonitor* om) { 173 return (ObjectMonitor*)((intptr_t)om->next_om() & ~OM_LOCK_BIT); 174 } 175 176 // Try to lock an ObjectMonitor. Returns true if locking was successful. 177 // Otherwise returns false. 178 static bool try_om_lock(ObjectMonitor* om) { 179 // Get current next field without any OM_LOCK_BIT value. 180 ObjectMonitor* next = unmarked_next(om); 181 if (om->try_set_next_om(next, mark_om_ptr(next)) != next) { 182 return false; // Cannot lock the ObjectMonitor. 183 } 184 return true; 185 } 186 187 // Lock an ObjectMonitor. 188 static void om_lock(ObjectMonitor* om) { 189 while (true) { 190 if (try_om_lock(om)) { 191 return; 192 } 193 } 194 } 195 196 // Unlock an ObjectMonitor. 197 static void om_unlock(ObjectMonitor* om) { 198 ObjectMonitor* next = om->next_om(); 199 guarantee(((intptr_t)next & OM_LOCK_BIT) == OM_LOCK_BIT, "next=" INTPTR_FORMAT 200 " must have OM_LOCK_BIT=%x set.", p2i(next), OM_LOCK_BIT); 201 202 next = (ObjectMonitor*)((intptr_t)next & ~OM_LOCK_BIT); // Clear OM_LOCK_BIT. 203 om->set_next_om(next); 204 } 205 206 // Get the list head after locking it. Returns the list head or NULL 207 // if the list is empty. 208 static ObjectMonitor* get_list_head_locked(ObjectMonitor** list_p) { 209 while (true) { 210 ObjectMonitor* mid = Atomic::load(list_p); 211 if (mid == NULL) { 212 return NULL; // The list is empty. 213 } 214 if (try_om_lock(mid)) { 215 if (Atomic::load(list_p) != mid) { 216 // The list head changed before we could lock it so we have to retry. 217 om_unlock(mid); 218 continue; 219 } 220 return mid; 221 } 222 } 223 } 224 225 #undef OM_LOCK_BIT 226 227 228 // =====================> List Management functions 229 230 // Prepend a list of ObjectMonitors to the specified *list_p. 'tail' is 231 // the last ObjectMonitor in the list and there are 'count' on the list. 232 // Also updates the specified *count_p. 233 static void prepend_list_to_common(ObjectMonitor* list, ObjectMonitor* tail, 234 int count, ObjectMonitor** list_p, 235 int* count_p) { 236 while (true) { 237 ObjectMonitor* cur = Atomic::load(list_p); 238 // Prepend list to *list_p. 239 if (!try_om_lock(tail)) { 240 // Failed to lock tail due to a list walker so try it all again. 241 continue; 242 } 243 tail->set_next_om(cur); // tail now points to cur (and unlocks tail) 244 if (cur == NULL) { 245 // No potential race with takers or other prependers since 246 // *list_p is empty. 247 if (Atomic::cmpxchg(list_p, cur, list) == cur) { 248 // Successfully switched *list_p to the list value. 249 Atomic::add(count_p, count); 250 break; 251 } 252 // Implied else: try it all again 253 } else { 254 if (!try_om_lock(cur)) { 255 continue; // failed to lock cur so try it all again 256 } 257 // We locked cur so try to switch *list_p to the list value. 258 if (Atomic::cmpxchg(list_p, cur, list) != cur) { 259 // The list head has changed so unlock cur and try again: 260 om_unlock(cur); 261 continue; 262 } 263 Atomic::add(count_p, count); 264 om_unlock(cur); 265 break; 266 } 267 } 268 } 269 270 // Prepend a newly allocated block of ObjectMonitors to g_block_list and 271 // om_list_globals._free_list. Also updates om_list_globals._population 272 // and om_list_globals._free_count. 273 void ObjectSynchronizer::prepend_block_to_lists(PaddedObjectMonitor* new_blk) { 274 // First we handle g_block_list: 275 while (true) { 276 PaddedObjectMonitor* cur = Atomic::load(&g_block_list); 277 // Prepend new_blk to g_block_list. The first ObjectMonitor in 278 // a block is reserved for use as linkage to the next block. 279 new_blk[0].set_next_om(cur); 280 if (Atomic::cmpxchg(&g_block_list, cur, new_blk) == cur) { 281 // Successfully switched g_block_list to the new_blk value. 282 Atomic::add(&om_list_globals._population, _BLOCKSIZE - 1); 283 break; 284 } 285 // Implied else: try it all again 286 } 287 288 // Second we handle om_list_globals._free_list: 289 prepend_list_to_common(new_blk + 1, &new_blk[_BLOCKSIZE - 1], _BLOCKSIZE - 1, 290 &om_list_globals._free_list, &om_list_globals._free_count); 291 } 292 293 // Prepend a list of ObjectMonitors to om_list_globals._free_list. 294 // 'tail' is the last ObjectMonitor in the list and there are 'count' 295 // on the list. Also updates om_list_globals._free_count. 296 static void prepend_list_to_global_free_list(ObjectMonitor* list, 297 ObjectMonitor* tail, int count) { 298 prepend_list_to_common(list, tail, count, &om_list_globals._free_list, 299 &om_list_globals._free_count); 300 } 301 302 // Prepend a list of ObjectMonitors to om_list_globals._in_use_list. 303 // 'tail' is the last ObjectMonitor in the list and there are 'count' 304 // on the list. Also updates om_list_globals._in_use_list. 305 static void prepend_list_to_global_in_use_list(ObjectMonitor* list, 306 ObjectMonitor* tail, int count) { 307 prepend_list_to_common(list, tail, count, &om_list_globals._in_use_list, 308 &om_list_globals._in_use_count); 309 } 310 311 // Prepend an ObjectMonitor to the specified list. Also updates 312 // the specified counter. 313 static void prepend_to_common(ObjectMonitor* m, ObjectMonitor** list_p, 314 int* count_p) { 315 while (true) { 316 om_lock(m); // Lock m so we can safely update its next field. 317 ObjectMonitor* cur = NULL; 318 // Lock the list head to guard against races with a list walker 319 // thread: 320 if ((cur = get_list_head_locked(list_p)) != NULL) { 321 // List head is now locked so we can safely switch it. 322 m->set_next_om(cur); // m now points to cur (and unlocks m) 323 Atomic::store(list_p, m); // Switch list head to unlocked m. 324 om_unlock(cur); 325 break; 326 } 327 // The list is empty so try to set the list head. 328 assert(cur == NULL, "cur must be NULL: cur=" INTPTR_FORMAT, p2i(cur)); 329 m->set_next_om(cur); // m now points to NULL (and unlocks m) 330 if (Atomic::cmpxchg(list_p, cur, m) == cur) { 331 // List head is now unlocked m. 332 break; 333 } 334 // Implied else: try it all again 335 } 336 Atomic::inc(count_p); 337 } 338 339 // Prepend an ObjectMonitor to a per-thread om_free_list. 340 // Also updates the per-thread om_free_count. 341 static void prepend_to_om_free_list(Thread* self, ObjectMonitor* m) { 342 prepend_to_common(m, &self->om_free_list, &self->om_free_count); 343 } 344 345 // Prepend an ObjectMonitor to a per-thread om_in_use_list. 346 // Also updates the per-thread om_in_use_count. 347 static void prepend_to_om_in_use_list(Thread* self, ObjectMonitor* m) { 348 prepend_to_common(m, &self->om_in_use_list, &self->om_in_use_count); 349 } 350 351 // Take an ObjectMonitor from the start of the specified list. Also 352 // decrements the specified counter. Returns NULL if none are available. 353 static ObjectMonitor* take_from_start_of_common(ObjectMonitor** list_p, 354 int* count_p) { 355 ObjectMonitor* take = NULL; 356 // Lock the list head to guard against races with a list walker 357 // thread: 358 if ((take = get_list_head_locked(list_p)) == NULL) { 359 return NULL; // None are available. 360 } 361 ObjectMonitor* next = unmarked_next(take); 362 // Switch locked list head to next (which unlocks the list head, but 363 // leaves take locked): 364 Atomic::store(list_p, next); 365 Atomic::dec(count_p); 366 // Unlock take, but leave the next value for any lagging list 367 // walkers. It will get cleaned up when take is prepended to 368 // the in-use list: 369 om_unlock(take); 370 return take; 371 } 372 373 // Take an ObjectMonitor from the start of the om_list_globals._free_list. 374 // Also updates om_list_globals._free_count. Returns NULL if none are 375 // available. 376 static ObjectMonitor* take_from_start_of_global_free_list() { 377 return take_from_start_of_common(&om_list_globals._free_list, 378 &om_list_globals._free_count); 379 } 380 381 // Take an ObjectMonitor from the start of a per-thread free-list. 382 // Also updates om_free_count. Returns NULL if none are available. 383 static ObjectMonitor* take_from_start_of_om_free_list(Thread* self) { 384 return take_from_start_of_common(&self->om_free_list, &self->om_free_count); 385 } 386 387 388 // =====================> Quick functions 389 390 // The quick_* forms are special fast-path variants used to improve 391 // performance. In the simplest case, a "quick_*" implementation could 392 // simply return false, in which case the caller will perform the necessary 393 // state transitions and call the slow-path form. 394 // The fast-path is designed to handle frequently arising cases in an efficient 395 // manner and is just a degenerate "optimistic" variant of the slow-path. 396 // returns true -- to indicate the call was satisfied. 397 // returns false -- to indicate the call needs the services of the slow-path. 398 // A no-loitering ordinance is in effect for code in the quick_* family 399 // operators: safepoints or indefinite blocking (blocking that might span a 400 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 401 // entry. 402 // 403 // Consider: An interesting optimization is to have the JIT recognize the 404 // following common idiom: 405 // synchronized (someobj) { .... ; notify(); } 406 // That is, we find a notify() or notifyAll() call that immediately precedes 407 // the monitorexit operation. In that case the JIT could fuse the operations 408 // into a single notifyAndExit() runtime primitive. 409 410 bool ObjectSynchronizer::quick_notify(oopDesc* obj, Thread* self, bool all) { 411 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 412 assert(self->is_Java_thread(), "invariant"); 413 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant"); 414 NoSafepointVerifier nsv; 415 if (obj == NULL) return false; // slow-path for invalid obj 416 const markWord mark = obj->mark(); 417 418 if (mark.has_locker() && self->is_lock_owned((address)mark.locker())) { 419 // Degenerate notify 420 // stack-locked by caller so by definition the implied waitset is empty. 421 return true; 422 } 423 424 if (mark.has_monitor()) { 425 ObjectMonitor* const mon = mark.monitor(); 426 assert(mon->object() == obj, "invariant"); 427 if (mon->owner() != self) return false; // slow-path for IMS exception 428 429 if (mon->first_waiter() != NULL) { 430 // We have one or more waiters. Since this is an inflated monitor 431 // that we own, we can transfer one or more threads from the waitset 432 // to the entrylist here and now, avoiding the slow-path. 433 if (all) { 434 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, self); 435 } else { 436 DTRACE_MONITOR_PROBE(notify, mon, obj, self); 437 } 438 int free_count = 0; 439 do { 440 mon->INotify(self); 441 ++free_count; 442 } while (mon->first_waiter() != NULL && all); 443 OM_PERFDATA_OP(Notifications, inc(free_count)); 444 } 445 return true; 446 } 447 448 // biased locking and any other IMS exception states take the slow-path 449 return false; 450 } 451 452 453 // The LockNode emitted directly at the synchronization site would have 454 // been too big if it were to have included support for the cases of inflated 455 // recursive enter and exit, so they go here instead. 456 // Note that we can't safely call AsyncPrintJavaStack() from within 457 // quick_enter() as our thread state remains _in_Java. 458 459 bool ObjectSynchronizer::quick_enter(oop obj, Thread* self, 460 BasicLock * lock) { 461 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 462 assert(self->is_Java_thread(), "invariant"); 463 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant"); 464 NoSafepointVerifier nsv; 465 if (obj == NULL) return false; // Need to throw NPE 466 const markWord mark = obj->mark(); 467 468 if (mark.has_monitor()) { 469 ObjectMonitor* const m = mark.monitor(); 470 assert(m->object() == obj, "invariant"); 471 Thread* const owner = (Thread *) m->_owner; 472 473 // Lock contention and Transactional Lock Elision (TLE) diagnostics 474 // and observability 475 // Case: light contention possibly amenable to TLE 476 // Case: TLE inimical operations such as nested/recursive synchronization 477 478 if (owner == self) { 479 m->_recursions++; 480 return true; 481 } 482 483 // This Java Monitor is inflated so obj's header will never be 484 // displaced to this thread's BasicLock. Make the displaced header 485 // non-NULL so this BasicLock is not seen as recursive nor as 486 // being locked. We do this unconditionally so that this thread's 487 // BasicLock cannot be mis-interpreted by any stack walkers. For 488 // performance reasons, stack walkers generally first check for 489 // Biased Locking in the object's header, the second check is for 490 // stack-locking in the object's header, the third check is for 491 // recursive stack-locking in the displaced header in the BasicLock, 492 // and last are the inflated Java Monitor (ObjectMonitor) checks. 493 lock->set_displaced_header(markWord::unused_mark()); 494 495 if (owner == NULL && m->try_set_owner_from(NULL, self) == NULL) { 496 assert(m->_recursions == 0, "invariant"); 497 return true; 498 } 499 } 500 501 // Note that we could inflate in quick_enter. 502 // This is likely a useful optimization 503 // Critically, in quick_enter() we must not: 504 // -- perform bias revocation, or 505 // -- block indefinitely, or 506 // -- reach a safepoint 507 508 return false; // revert to slow-path 509 } 510 511 // ----------------------------------------------------------------------------- 512 // Monitor Enter/Exit 513 // The interpreter and compiler assembly code tries to lock using the fast path 514 // of this algorithm. Make sure to update that code if the following function is 515 // changed. The implementation is extremely sensitive to race condition. Be careful. 516 517 void ObjectSynchronizer::enter(Handle obj, BasicLock* lock, TRAPS) { 518 if (UseBiasedLocking) { 519 if (!SafepointSynchronize::is_at_safepoint()) { 520 BiasedLocking::revoke(obj, THREAD); 521 } else { 522 BiasedLocking::revoke_at_safepoint(obj); 523 } 524 } 525 526 markWord mark = obj->mark(); 527 assert(!mark.has_bias_pattern(), "should not see bias pattern here"); 528 529 if (mark.is_neutral()) { 530 // Anticipate successful CAS -- the ST of the displaced mark must 531 // be visible <= the ST performed by the CAS. 532 lock->set_displaced_header(mark); 533 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) { 534 return; 535 } 536 // Fall through to inflate() ... 537 } else if (mark.has_locker() && 538 THREAD->is_lock_owned((address)mark.locker())) { 539 assert(lock != mark.locker(), "must not re-lock the same lock"); 540 assert(lock != (BasicLock*)obj->mark().value(), "don't relock with same BasicLock"); 541 lock->set_displaced_header(markWord::from_pointer(NULL)); 542 return; 543 } 544 545 // The object header will never be displaced to this lock, 546 // so it does not matter what the value is, except that it 547 // must be non-zero to avoid looking like a re-entrant lock, 548 // and must not look locked either. 549 lock->set_displaced_header(markWord::unused_mark()); 550 inflate(THREAD, obj(), inflate_cause_monitor_enter)->enter(THREAD); 551 } 552 553 void ObjectSynchronizer::exit(oop object, BasicLock* lock, TRAPS) { 554 markWord mark = object->mark(); 555 // We cannot check for Biased Locking if we are racing an inflation. 556 assert(mark == markWord::INFLATING() || 557 !mark.has_bias_pattern(), "should not see bias pattern here"); 558 559 markWord dhw = lock->displaced_header(); 560 if (dhw.value() == 0) { 561 // If the displaced header is NULL, then this exit matches up with 562 // a recursive enter. No real work to do here except for diagnostics. 563 #ifndef PRODUCT 564 if (mark != markWord::INFLATING()) { 565 // Only do diagnostics if we are not racing an inflation. Simply 566 // exiting a recursive enter of a Java Monitor that is being 567 // inflated is safe; see the has_monitor() comment below. 568 assert(!mark.is_neutral(), "invariant"); 569 assert(!mark.has_locker() || 570 THREAD->is_lock_owned((address)mark.locker()), "invariant"); 571 if (mark.has_monitor()) { 572 // The BasicLock's displaced_header is marked as a recursive 573 // enter and we have an inflated Java Monitor (ObjectMonitor). 574 // This is a special case where the Java Monitor was inflated 575 // after this thread entered the stack-lock recursively. When a 576 // Java Monitor is inflated, we cannot safely walk the Java 577 // Monitor owner's stack and update the BasicLocks because a 578 // Java Monitor can be asynchronously inflated by a thread that 579 // does not own the Java Monitor. 580 ObjectMonitor* m = mark.monitor(); 581 assert(((oop)(m->object()))->mark() == mark, "invariant"); 582 assert(m->is_entered(THREAD), "invariant"); 583 } 584 } 585 #endif 586 return; 587 } 588 589 if (mark == markWord::from_pointer(lock)) { 590 // If the object is stack-locked by the current thread, try to 591 // swing the displaced header from the BasicLock back to the mark. 592 assert(dhw.is_neutral(), "invariant"); 593 if (object->cas_set_mark(dhw, mark) == mark) { 594 return; 595 } 596 } 597 598 // We have to take the slow-path of possible inflation and then exit. 599 inflate(THREAD, object, inflate_cause_vm_internal)->exit(true, THREAD); 600 } 601 602 // ----------------------------------------------------------------------------- 603 // Class Loader support to workaround deadlocks on the class loader lock objects 604 // Also used by GC 605 // complete_exit()/reenter() are used to wait on a nested lock 606 // i.e. to give up an outer lock completely and then re-enter 607 // Used when holding nested locks - lock acquisition order: lock1 then lock2 608 // 1) complete_exit lock1 - saving recursion count 609 // 2) wait on lock2 610 // 3) when notified on lock2, unlock lock2 611 // 4) reenter lock1 with original recursion count 612 // 5) lock lock2 613 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 614 intx ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { 615 if (UseBiasedLocking) { 616 BiasedLocking::revoke(obj, THREAD); 617 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 618 } 619 620 ObjectMonitor* monitor = inflate(THREAD, obj(), inflate_cause_vm_internal); 621 622 return monitor->complete_exit(THREAD); 623 } 624 625 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 626 void ObjectSynchronizer::reenter(Handle obj, intx recursions, TRAPS) { 627 if (UseBiasedLocking) { 628 BiasedLocking::revoke(obj, THREAD); 629 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 630 } 631 632 ObjectMonitor* monitor = inflate(THREAD, obj(), inflate_cause_vm_internal); 633 634 monitor->reenter(recursions, THREAD); 635 } 636 // ----------------------------------------------------------------------------- 637 // JNI locks on java objects 638 // NOTE: must use heavy weight monitor to handle jni monitor enter 639 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { 640 // the current locking is from JNI instead of Java code 641 if (UseBiasedLocking) { 642 BiasedLocking::revoke(obj, THREAD); 643 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 644 } 645 THREAD->set_current_pending_monitor_is_from_java(false); 646 inflate(THREAD, obj(), inflate_cause_jni_enter)->enter(THREAD); 647 THREAD->set_current_pending_monitor_is_from_java(true); 648 } 649 650 // NOTE: must use heavy weight monitor to handle jni monitor exit 651 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { 652 if (UseBiasedLocking) { 653 Handle h_obj(THREAD, obj); 654 BiasedLocking::revoke(h_obj, THREAD); 655 obj = h_obj(); 656 } 657 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 658 659 ObjectMonitor* monitor = inflate(THREAD, obj, inflate_cause_jni_exit); 660 // If this thread has locked the object, exit the monitor. We 661 // intentionally do not use CHECK here because we must exit the 662 // monitor even if an exception is pending. 663 if (monitor->check_owner(THREAD)) { 664 monitor->exit(true, THREAD); 665 } 666 } 667 668 // ----------------------------------------------------------------------------- 669 // Internal VM locks on java objects 670 // standard constructor, allows locking failures 671 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool do_lock) { 672 _dolock = do_lock; 673 _thread = thread; 674 _thread->check_for_valid_safepoint_state(); 675 _obj = obj; 676 677 if (_dolock) { 678 ObjectSynchronizer::enter(_obj, &_lock, _thread); 679 } 680 } 681 682 ObjectLocker::~ObjectLocker() { 683 if (_dolock) { 684 ObjectSynchronizer::exit(_obj(), &_lock, _thread); 685 } 686 } 687 688 689 // ----------------------------------------------------------------------------- 690 // Wait/Notify/NotifyAll 691 // NOTE: must use heavy weight monitor to handle wait() 692 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 693 if (UseBiasedLocking) { 694 BiasedLocking::revoke(obj, THREAD); 695 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 696 } 697 if (millis < 0) { 698 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 699 } 700 ObjectMonitor* monitor = inflate(THREAD, obj(), inflate_cause_wait); 701 702 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); 703 monitor->wait(millis, true, THREAD); 704 705 // This dummy call is in place to get around dtrace bug 6254741. Once 706 // that's fixed we can uncomment the following line, remove the call 707 // and change this function back into a "void" func. 708 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 709 return dtrace_waited_probe(monitor, obj, THREAD); 710 } 711 712 void ObjectSynchronizer::wait_uninterruptibly(Handle obj, jlong millis, TRAPS) { 713 if (UseBiasedLocking) { 714 BiasedLocking::revoke(obj, THREAD); 715 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 716 } 717 if (millis < 0) { 718 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 719 } 720 inflate(THREAD, obj(), inflate_cause_wait)->wait(millis, false, THREAD); 721 } 722 723 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 724 if (UseBiasedLocking) { 725 BiasedLocking::revoke(obj, THREAD); 726 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 727 } 728 729 markWord mark = obj->mark(); 730 if (mark.has_locker() && THREAD->is_lock_owned((address)mark.locker())) { 731 return; 732 } 733 inflate(THREAD, obj(), inflate_cause_notify)->notify(THREAD); 734 } 735 736 // NOTE: see comment of notify() 737 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 738 if (UseBiasedLocking) { 739 BiasedLocking::revoke(obj, THREAD); 740 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 741 } 742 743 markWord mark = obj->mark(); 744 if (mark.has_locker() && THREAD->is_lock_owned((address)mark.locker())) { 745 return; 746 } 747 inflate(THREAD, obj(), inflate_cause_notify)->notifyAll(THREAD); 748 } 749 750 // ----------------------------------------------------------------------------- 751 // Hash Code handling 752 // 753 // Performance concern: 754 // OrderAccess::storestore() calls release() which at one time stored 0 755 // into the global volatile OrderAccess::dummy variable. This store was 756 // unnecessary for correctness. Many threads storing into a common location 757 // causes considerable cache migration or "sloshing" on large SMP systems. 758 // As such, I avoided using OrderAccess::storestore(). In some cases 759 // OrderAccess::fence() -- which incurs local latency on the executing 760 // processor -- is a better choice as it scales on SMP systems. 761 // 762 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for 763 // a discussion of coherency costs. Note that all our current reference 764 // platforms provide strong ST-ST order, so the issue is moot on IA32, 765 // x64, and SPARC. 766 // 767 // As a general policy we use "volatile" to control compiler-based reordering 768 // and explicit fences (barriers) to control for architectural reordering 769 // performed by the CPU(s) or platform. 770 771 struct SharedGlobals { 772 char _pad_prefix[OM_CACHE_LINE_SIZE]; 773 // These are highly shared mostly-read variables. 774 // To avoid false-sharing they need to be the sole occupants of a cache line. 775 volatile int stw_random; 776 volatile int stw_cycle; 777 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int) * 2); 778 // Hot RW variable -- Sequester to avoid false-sharing 779 volatile int hc_sequence; 780 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 781 }; 782 783 static SharedGlobals GVars; 784 static int _forceMonitorScavenge = 0; // Scavenge required and pending 785 786 static markWord read_stable_mark(oop obj) { 787 markWord mark = obj->mark(); 788 if (!mark.is_being_inflated()) { 789 return mark; // normal fast-path return 790 } 791 792 int its = 0; 793 for (;;) { 794 markWord mark = obj->mark(); 795 if (!mark.is_being_inflated()) { 796 return mark; // normal fast-path return 797 } 798 799 // The object is being inflated by some other thread. 800 // The caller of read_stable_mark() must wait for inflation to complete. 801 // Avoid live-lock 802 // TODO: consider calling SafepointSynchronize::do_call_back() while 803 // spinning to see if there's a safepoint pending. If so, immediately 804 // yielding or blocking would be appropriate. Avoid spinning while 805 // there is a safepoint pending. 806 // TODO: add inflation contention performance counters. 807 // TODO: restrict the aggregate number of spinners. 808 809 ++its; 810 if (its > 10000 || !os::is_MP()) { 811 if (its & 1) { 812 os::naked_yield(); 813 } else { 814 // Note that the following code attenuates the livelock problem but is not 815 // a complete remedy. A more complete solution would require that the inflating 816 // thread hold the associated inflation lock. The following code simply restricts 817 // the number of spinners to at most one. We'll have N-2 threads blocked 818 // on the inflationlock, 1 thread holding the inflation lock and using 819 // a yield/park strategy, and 1 thread in the midst of inflation. 820 // A more refined approach would be to change the encoding of INFLATING 821 // to allow encapsulation of a native thread pointer. Threads waiting for 822 // inflation to complete would use CAS to push themselves onto a singly linked 823 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 824 // and calling park(). When inflation was complete the thread that accomplished inflation 825 // would detach the list and set the markword to inflated with a single CAS and 826 // then for each thread on the list, set the flag and unpark() the thread. 827 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease 828 // wakes at most one thread whereas we need to wake the entire list. 829 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1); 830 int YieldThenBlock = 0; 831 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant"); 832 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant"); 833 Thread::muxAcquire(gInflationLocks + ix, "gInflationLock"); 834 while (obj->mark() == markWord::INFLATING()) { 835 // Beware: NakedYield() is advisory and has almost no effect on some platforms 836 // so we periodically call self->_ParkEvent->park(1). 837 // We use a mixed spin/yield/block mechanism. 838 if ((YieldThenBlock++) >= 16) { 839 Thread::current()->_ParkEvent->park(1); 840 } else { 841 os::naked_yield(); 842 } 843 } 844 Thread::muxRelease(gInflationLocks + ix); 845 } 846 } else { 847 SpinPause(); // SMP-polite spinning 848 } 849 } 850 } 851 852 // hashCode() generation : 853 // 854 // Possibilities: 855 // * MD5Digest of {obj,stw_random} 856 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function. 857 // * A DES- or AES-style SBox[] mechanism 858 // * One of the Phi-based schemes, such as: 859 // 2654435761 = 2^32 * Phi (golden ratio) 860 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ; 861 // * A variation of Marsaglia's shift-xor RNG scheme. 862 // * (obj ^ stw_random) is appealing, but can result 863 // in undesirable regularity in the hashCode values of adjacent objects 864 // (objects allocated back-to-back, in particular). This could potentially 865 // result in hashtable collisions and reduced hashtable efficiency. 866 // There are simple ways to "diffuse" the middle address bits over the 867 // generated hashCode values: 868 869 static inline intptr_t get_next_hash(Thread* self, oop obj) { 870 intptr_t value = 0; 871 if (hashCode == 0) { 872 // This form uses global Park-Miller RNG. 873 // On MP system we'll have lots of RW access to a global, so the 874 // mechanism induces lots of coherency traffic. 875 value = os::random(); 876 } else if (hashCode == 1) { 877 // This variation has the property of being stable (idempotent) 878 // between STW operations. This can be useful in some of the 1-0 879 // synchronization schemes. 880 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3; 881 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random; 882 } else if (hashCode == 2) { 883 value = 1; // for sensitivity testing 884 } else if (hashCode == 3) { 885 value = ++GVars.hc_sequence; 886 } else if (hashCode == 4) { 887 value = cast_from_oop<intptr_t>(obj); 888 } else { 889 // Marsaglia's xor-shift scheme with thread-specific state 890 // This is probably the best overall implementation -- we'll 891 // likely make this the default in future releases. 892 unsigned t = self->_hashStateX; 893 t ^= (t << 11); 894 self->_hashStateX = self->_hashStateY; 895 self->_hashStateY = self->_hashStateZ; 896 self->_hashStateZ = self->_hashStateW; 897 unsigned v = self->_hashStateW; 898 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 899 self->_hashStateW = v; 900 value = v; 901 } 902 903 value &= markWord::hash_mask; 904 if (value == 0) value = 0xBAD; 905 assert(value != markWord::no_hash, "invariant"); 906 return value; 907 } 908 909 intptr_t ObjectSynchronizer::FastHashCode(Thread* self, oop obj) { 910 if (UseBiasedLocking) { 911 // NOTE: many places throughout the JVM do not expect a safepoint 912 // to be taken here, in particular most operations on perm gen 913 // objects. However, we only ever bias Java instances and all of 914 // the call sites of identity_hash that might revoke biases have 915 // been checked to make sure they can handle a safepoint. The 916 // added check of the bias pattern is to avoid useless calls to 917 // thread-local storage. 918 if (obj->mark().has_bias_pattern()) { 919 // Handle for oop obj in case of STW safepoint 920 Handle hobj(self, obj); 921 // Relaxing assertion for bug 6320749. 922 assert(Universe::verify_in_progress() || 923 !SafepointSynchronize::is_at_safepoint(), 924 "biases should not be seen by VM thread here"); 925 BiasedLocking::revoke(hobj, JavaThread::current()); 926 obj = hobj(); 927 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 928 } 929 } 930 931 // hashCode() is a heap mutator ... 932 // Relaxing assertion for bug 6320749. 933 assert(Universe::verify_in_progress() || DumpSharedSpaces || 934 !SafepointSynchronize::is_at_safepoint(), "invariant"); 935 assert(Universe::verify_in_progress() || DumpSharedSpaces || 936 self->is_Java_thread() , "invariant"); 937 assert(Universe::verify_in_progress() || DumpSharedSpaces || 938 ((JavaThread *)self)->thread_state() != _thread_blocked, "invariant"); 939 940 ObjectMonitor* monitor = NULL; 941 markWord temp, test; 942 intptr_t hash; 943 markWord mark = read_stable_mark(obj); 944 945 // object should remain ineligible for biased locking 946 assert(!mark.has_bias_pattern(), "invariant"); 947 948 if (mark.is_neutral()) { // if this is a normal header 949 hash = mark.hash(); 950 if (hash != 0) { // if it has a hash, just return it 951 return hash; 952 } 953 hash = get_next_hash(self, obj); // get a new hash 954 temp = mark.copy_set_hash(hash); // merge the hash into header 955 // try to install the hash 956 test = obj->cas_set_mark(temp, mark); 957 if (test == mark) { // if the hash was installed, return it 958 return hash; 959 } 960 // Failed to install the hash. It could be that another thread 961 // installed the hash just before our attempt or inflation has 962 // occurred or... so we fall thru to inflate the monitor for 963 // stability and then install the hash. 964 } else if (mark.has_monitor()) { 965 monitor = mark.monitor(); 966 temp = monitor->header(); 967 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 968 hash = temp.hash(); 969 if (hash != 0) { // if it has a hash, just return it 970 return hash; 971 } 972 // Fall thru so we only have one place that installs the hash in 973 // the ObjectMonitor. 974 } else if (self->is_lock_owned((address)mark.locker())) { 975 // This is a stack lock owned by the calling thread so fetch the 976 // displaced markWord from the BasicLock on the stack. 977 temp = mark.displaced_mark_helper(); 978 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 979 hash = temp.hash(); 980 if (hash != 0) { // if it has a hash, just return it 981 return hash; 982 } 983 // WARNING: 984 // The displaced header in the BasicLock on a thread's stack 985 // is strictly immutable. It CANNOT be changed in ANY cases. 986 // So we have to inflate the stack lock into an ObjectMonitor 987 // even if the current thread owns the lock. The BasicLock on 988 // a thread's stack can be asynchronously read by other threads 989 // during an inflate() call so any change to that stack memory 990 // may not propagate to other threads correctly. 991 } 992 993 // Inflate the monitor to set the hash. 994 monitor = inflate(self, obj, inflate_cause_hash_code); 995 // Load ObjectMonitor's header/dmw field and see if it has a hash. 996 mark = monitor->header(); 997 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 998 hash = mark.hash(); 999 if (hash == 0) { // if it does not have a hash 1000 hash = get_next_hash(self, obj); // get a new hash 1001 temp = mark.copy_set_hash(hash); // merge the hash into header 1002 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1003 uintptr_t v = Atomic::cmpxchg((volatile uintptr_t*)monitor->header_addr(), mark.value(), temp.value()); 1004 test = markWord(v); 1005 if (test != mark) { 1006 // The attempt to update the ObjectMonitor's header/dmw field 1007 // did not work. This can happen if another thread managed to 1008 // merge in the hash just before our cmpxchg(). 1009 // If we add any new usages of the header/dmw field, this code 1010 // will need to be updated. 1011 hash = test.hash(); 1012 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value()); 1013 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash"); 1014 } 1015 } 1016 // We finally get the hash. 1017 return hash; 1018 } 1019 1020 // Deprecated -- use FastHashCode() instead. 1021 1022 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { 1023 return FastHashCode(Thread::current(), obj()); 1024 } 1025 1026 1027 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, 1028 Handle h_obj) { 1029 if (UseBiasedLocking) { 1030 BiasedLocking::revoke(h_obj, thread); 1031 assert(!h_obj->mark().has_bias_pattern(), "biases should be revoked by now"); 1032 } 1033 1034 assert(thread == JavaThread::current(), "Can only be called on current thread"); 1035 oop obj = h_obj(); 1036 1037 markWord mark = read_stable_mark(obj); 1038 1039 // Uncontended case, header points to stack 1040 if (mark.has_locker()) { 1041 return thread->is_lock_owned((address)mark.locker()); 1042 } 1043 // Contended case, header points to ObjectMonitor (tagged pointer) 1044 if (mark.has_monitor()) { 1045 ObjectMonitor* monitor = mark.monitor(); 1046 return monitor->is_entered(thread) != 0; 1047 } 1048 // Unlocked case, header in place 1049 assert(mark.is_neutral(), "sanity check"); 1050 return false; 1051 } 1052 1053 // Be aware of this method could revoke bias of the lock object. 1054 // This method queries the ownership of the lock handle specified by 'h_obj'. 1055 // If the current thread owns the lock, it returns owner_self. If no 1056 // thread owns the lock, it returns owner_none. Otherwise, it will return 1057 // owner_other. 1058 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership 1059 (JavaThread *self, Handle h_obj) { 1060 // The caller must beware this method can revoke bias, and 1061 // revocation can result in a safepoint. 1062 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 1063 assert(self->thread_state() != _thread_blocked, "invariant"); 1064 1065 // Possible mark states: neutral, biased, stack-locked, inflated 1066 1067 if (UseBiasedLocking && h_obj()->mark().has_bias_pattern()) { 1068 // CASE: biased 1069 BiasedLocking::revoke(h_obj, self); 1070 assert(!h_obj->mark().has_bias_pattern(), 1071 "biases should be revoked by now"); 1072 } 1073 1074 assert(self == JavaThread::current(), "Can only be called on current thread"); 1075 oop obj = h_obj(); 1076 markWord mark = read_stable_mark(obj); 1077 1078 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. 1079 if (mark.has_locker()) { 1080 return self->is_lock_owned((address)mark.locker()) ? 1081 owner_self : owner_other; 1082 } 1083 1084 // CASE: inflated. Mark (tagged pointer) points to an ObjectMonitor. 1085 // The Object:ObjectMonitor relationship is stable as long as we're 1086 // not at a safepoint. 1087 if (mark.has_monitor()) { 1088 void* owner = mark.monitor()->_owner; 1089 if (owner == NULL) return owner_none; 1090 return (owner == self || 1091 self->is_lock_owned((address)owner)) ? owner_self : owner_other; 1092 } 1093 1094 // CASE: neutral 1095 assert(mark.is_neutral(), "sanity check"); 1096 return owner_none; // it's unlocked 1097 } 1098 1099 // FIXME: jvmti should call this 1100 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) { 1101 if (UseBiasedLocking) { 1102 if (SafepointSynchronize::is_at_safepoint()) { 1103 BiasedLocking::revoke_at_safepoint(h_obj); 1104 } else { 1105 BiasedLocking::revoke(h_obj, JavaThread::current()); 1106 } 1107 assert(!h_obj->mark().has_bias_pattern(), "biases should be revoked by now"); 1108 } 1109 1110 oop obj = h_obj(); 1111 address owner = NULL; 1112 1113 markWord mark = read_stable_mark(obj); 1114 1115 // Uncontended case, header points to stack 1116 if (mark.has_locker()) { 1117 owner = (address) mark.locker(); 1118 } 1119 1120 // Contended case, header points to ObjectMonitor (tagged pointer) 1121 else if (mark.has_monitor()) { 1122 ObjectMonitor* monitor = mark.monitor(); 1123 assert(monitor != NULL, "monitor should be non-null"); 1124 owner = (address) monitor->owner(); 1125 } 1126 1127 if (owner != NULL) { 1128 // owning_thread_from_monitor_owner() may also return NULL here 1129 return Threads::owning_thread_from_monitor_owner(t_list, owner); 1130 } 1131 1132 // Unlocked case, header in place 1133 // Cannot have assertion since this object may have been 1134 // locked by another thread when reaching here. 1135 // assert(mark.is_neutral(), "sanity check"); 1136 1137 return NULL; 1138 } 1139 1140 // Visitors ... 1141 1142 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { 1143 PaddedObjectMonitor* block = Atomic::load(&g_block_list); 1144 while (block != NULL) { 1145 assert(block->object() == CHAINMARKER, "must be a block header"); 1146 for (int i = _BLOCKSIZE - 1; i > 0; i--) { 1147 ObjectMonitor* mid = (ObjectMonitor *)(block + i); 1148 oop object = (oop)mid->object(); 1149 if (object != NULL) { 1150 // Only process with closure if the object is set. 1151 closure->do_monitor(mid); 1152 } 1153 } 1154 // unmarked_next() is not needed with g_block_list (no locking 1155 // used with block linkage _next_om fields). 1156 block = (PaddedObjectMonitor*)block->next_om(); 1157 } 1158 } 1159 1160 static bool monitors_used_above_threshold() { 1161 int population = Atomic::load(&om_list_globals._population); 1162 if (population == 0) { 1163 return false; 1164 } 1165 if (MonitorUsedDeflationThreshold > 0) { 1166 int monitors_used = population - Atomic::load(&om_list_globals._free_count); 1167 int monitor_usage = (monitors_used * 100LL) / population; 1168 return monitor_usage > MonitorUsedDeflationThreshold; 1169 } 1170 return false; 1171 } 1172 1173 // Returns true if MonitorBound is set (> 0) and if the specified 1174 // cnt is > MonitorBound. Otherwise returns false. 1175 static bool is_MonitorBound_exceeded(const int cnt) { 1176 const int mx = MonitorBound; 1177 return mx > 0 && cnt > mx; 1178 } 1179 1180 bool ObjectSynchronizer::is_cleanup_needed() { 1181 if (monitors_used_above_threshold()) { 1182 // Too many monitors in use. 1183 return true; 1184 } 1185 return needs_monitor_scavenge(); 1186 } 1187 1188 bool ObjectSynchronizer::needs_monitor_scavenge() { 1189 if (Atomic::load(&_forceMonitorScavenge) == 1) { 1190 log_info(monitorinflation)("Monitor scavenge needed, triggering safepoint cleanup."); 1191 return true; 1192 } 1193 return false; 1194 } 1195 1196 void ObjectSynchronizer::oops_do(OopClosure* f) { 1197 // We only scan the global used list here (for moribund threads), and 1198 // the thread-local monitors in Thread::oops_do(). 1199 global_used_oops_do(f); 1200 } 1201 1202 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) { 1203 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1204 list_oops_do(Atomic::load(&om_list_globals._in_use_list), f); 1205 } 1206 1207 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) { 1208 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1209 list_oops_do(thread->om_in_use_list, f); 1210 } 1211 1212 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, OopClosure* f) { 1213 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1214 // The oops_do() phase does not overlap with monitor deflation 1215 // so no need to lock ObjectMonitors for the list traversal. 1216 for (ObjectMonitor* mid = list; mid != NULL; mid = unmarked_next(mid)) { 1217 if (mid->object() != NULL) { 1218 f->do_oop((oop*)mid->object_addr()); 1219 } 1220 } 1221 } 1222 1223 1224 // ----------------------------------------------------------------------------- 1225 // ObjectMonitor Lifecycle 1226 // ----------------------- 1227 // Inflation unlinks monitors from om_list_globals._free_list or a per-thread 1228 // free list and associates them with objects. Deflation -- which occurs at 1229 // STW-time -- disassociates idle monitors from objects. 1230 // Such scavenged monitors are returned to the om_list_globals._free_list. 1231 // 1232 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. 1233 // 1234 // Lifecycle: 1235 // -- unassigned and on the om_list_globals._free_list 1236 // -- unassigned and on a per-thread free list 1237 // -- assigned to an object. The object is inflated and the mark refers 1238 // to the ObjectMonitor. 1239 1240 1241 // Constraining monitor pool growth via MonitorBound ... 1242 // 1243 // If MonitorBound is not set (<= 0), MonitorBound checks are disabled. 1244 // 1245 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the 1246 // the rate of scavenging is driven primarily by GC. As such, we can find 1247 // an inordinate number of monitors in circulation. 1248 // To avoid that scenario we can artificially induce a STW safepoint 1249 // if the pool appears to be growing past some reasonable bound. 1250 // Generally we favor time in space-time tradeoffs, but as there's no 1251 // natural back-pressure on the # of extant monitors we need to impose some 1252 // type of limit. Beware that if MonitorBound is set to too low a value 1253 // we could just loop. In addition, if MonitorBound is set to a low value 1254 // we'll incur more safepoints, which are harmful to performance. 1255 // See also: GuaranteedSafepointInterval 1256 // 1257 // If MonitorBound is set, the boundry applies to 1258 // (om_list_globals._population - om_list_globals._free_count) 1259 // i.e., if there are not enough ObjectMonitors on the global free list, 1260 // then a safepoint deflation is induced. Picking a good MonitorBound value 1261 // is non-trivial. 1262 1263 static void InduceScavenge(Thread* self, const char * Whence) { 1264 // Induce STW safepoint to trim monitors 1265 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. 1266 // More precisely, trigger a cleanup safepoint as the number 1267 // of active monitors passes the specified threshold. 1268 // TODO: assert thread state is reasonable 1269 1270 if (Atomic::xchg(&_forceMonitorScavenge, 1) == 0) { 1271 VMThread::check_for_forced_cleanup(); 1272 } 1273 } 1274 1275 ObjectMonitor* ObjectSynchronizer::om_alloc(Thread* self) { 1276 // A large MAXPRIVATE value reduces both list lock contention 1277 // and list coherency traffic, but also tends to increase the 1278 // number of ObjectMonitors in circulation as well as the STW 1279 // scavenge costs. As usual, we lean toward time in space-time 1280 // tradeoffs. 1281 const int MAXPRIVATE = 1024; 1282 NoSafepointVerifier nsv; 1283 1284 stringStream ss; 1285 for (;;) { 1286 ObjectMonitor* m; 1287 1288 // 1: try to allocate from the thread's local om_free_list. 1289 // Threads will attempt to allocate first from their local list, then 1290 // from the global list, and only after those attempts fail will the 1291 // thread attempt to instantiate new monitors. Thread-local free lists 1292 // improve allocation latency, as well as reducing coherency traffic 1293 // on the shared global list. 1294 m = take_from_start_of_om_free_list(self); 1295 if (m != NULL) { 1296 guarantee(m->object() == NULL, "invariant"); 1297 prepend_to_om_in_use_list(self, m); 1298 return m; 1299 } 1300 1301 // 2: try to allocate from the global om_list_globals._free_list 1302 // If we're using thread-local free lists then try 1303 // to reprovision the caller's free list. 1304 if (Atomic::load(&om_list_globals._free_list) != NULL) { 1305 // Reprovision the thread's om_free_list. 1306 // Use bulk transfers to reduce the allocation rate and heat 1307 // on various locks. 1308 for (int i = self->om_free_provision; --i >= 0;) { 1309 ObjectMonitor* take = take_from_start_of_global_free_list(); 1310 if (take == NULL) { 1311 break; // No more are available. 1312 } 1313 guarantee(take->object() == NULL, "invariant"); 1314 take->Recycle(); 1315 om_release(self, take, false); 1316 } 1317 self->om_free_provision += 1 + (self->om_free_provision / 2); 1318 if (self->om_free_provision > MAXPRIVATE) self->om_free_provision = MAXPRIVATE; 1319 1320 if (is_MonitorBound_exceeded(Atomic::load(&om_list_globals._population) - 1321 Atomic::load(&om_list_globals._free_count))) { 1322 // Not enough ObjectMonitors on the global free list. 1323 // We can't safely induce a STW safepoint from om_alloc() as our thread 1324 // state may not be appropriate for such activities and callers may hold 1325 // naked oops, so instead we defer the action. 1326 InduceScavenge(self, "om_alloc"); 1327 } 1328 continue; 1329 } 1330 1331 // 3: allocate a block of new ObjectMonitors 1332 // Both the local and global free lists are empty -- resort to malloc(). 1333 // In the current implementation ObjectMonitors are TSM - immortal. 1334 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want 1335 // each ObjectMonitor to start at the beginning of a cache line, 1336 // so we use align_up(). 1337 // A better solution would be to use C++ placement-new. 1338 // BEWARE: As it stands currently, we don't run the ctors! 1339 assert(_BLOCKSIZE > 1, "invariant"); 1340 size_t neededsize = sizeof(PaddedObjectMonitor) * _BLOCKSIZE; 1341 PaddedObjectMonitor* temp; 1342 size_t aligned_size = neededsize + (OM_CACHE_LINE_SIZE - 1); 1343 void* real_malloc_addr = NEW_C_HEAP_ARRAY(char, aligned_size, mtInternal); 1344 temp = (PaddedObjectMonitor*)align_up(real_malloc_addr, OM_CACHE_LINE_SIZE); 1345 (void)memset((void *) temp, 0, neededsize); 1346 1347 // Format the block. 1348 // initialize the linked list, each monitor points to its next 1349 // forming the single linked free list, the very first monitor 1350 // will points to next block, which forms the block list. 1351 // The trick of using the 1st element in the block as g_block_list 1352 // linkage should be reconsidered. A better implementation would 1353 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 1354 1355 for (int i = 1; i < _BLOCKSIZE; i++) { 1356 temp[i].set_next_om((ObjectMonitor*)&temp[i + 1]); 1357 } 1358 1359 // terminate the last monitor as the end of list 1360 temp[_BLOCKSIZE - 1].set_next_om((ObjectMonitor*)NULL); 1361 1362 // Element [0] is reserved for global list linkage 1363 temp[0].set_object(CHAINMARKER); 1364 1365 // Consider carving out this thread's current request from the 1366 // block in hand. This avoids some lock traffic and redundant 1367 // list activity. 1368 1369 prepend_block_to_lists(temp); 1370 } 1371 } 1372 1373 // Place "m" on the caller's private per-thread om_free_list. 1374 // In practice there's no need to clamp or limit the number of 1375 // monitors on a thread's om_free_list as the only non-allocation time 1376 // we'll call om_release() is to return a monitor to the free list after 1377 // a CAS attempt failed. This doesn't allow unbounded #s of monitors to 1378 // accumulate on a thread's free list. 1379 // 1380 // Key constraint: all ObjectMonitors on a thread's free list and the global 1381 // free list must have their object field set to null. This prevents the 1382 // scavenger -- deflate_monitor_list() -- from reclaiming them while we 1383 // are trying to release them. 1384 1385 void ObjectSynchronizer::om_release(Thread* self, ObjectMonitor* m, 1386 bool from_per_thread_alloc) { 1387 guarantee(m->header().value() == 0, "invariant"); 1388 guarantee(m->object() == NULL, "invariant"); 1389 NoSafepointVerifier nsv; 1390 1391 stringStream ss; 1392 guarantee((m->is_busy() | m->_recursions) == 0, "freeing in-use monitor: " 1393 "%s, recursions=" INTX_FORMAT, m->is_busy_to_string(&ss), 1394 m->_recursions); 1395 // _next_om is used for both per-thread in-use and free lists so 1396 // we have to remove 'm' from the in-use list first (as needed). 1397 if (from_per_thread_alloc) { 1398 // Need to remove 'm' from om_in_use_list. 1399 ObjectMonitor* mid = NULL; 1400 ObjectMonitor* next = NULL; 1401 1402 // This list walk can only race with another list walker since 1403 // deflation can only happen at a safepoint so we don't have to 1404 // worry about an ObjectMonitor being removed from this list 1405 // while we are walking it. 1406 1407 // Lock the list head to avoid racing with another list walker. 1408 if ((mid = get_list_head_locked(&self->om_in_use_list)) == NULL) { 1409 fatal("thread=" INTPTR_FORMAT " in-use list must not be empty.", p2i(self)); 1410 } 1411 next = unmarked_next(mid); 1412 if (m == mid) { 1413 // First special case: 1414 // 'm' matches mid, is the list head and is locked. Switch the list 1415 // head to next which unlocks the list head, but leaves the extracted 1416 // mid locked: 1417 Atomic::store(&self->om_in_use_list, next); 1418 } else if (m == next) { 1419 // Second special case: 1420 // 'm' matches next after the list head and we already have the list 1421 // head locked so set mid to what we are extracting: 1422 mid = next; 1423 // Lock mid to prevent races with a list walker: 1424 om_lock(mid); 1425 // Update next to what follows mid (if anything): 1426 next = unmarked_next(mid); 1427 // Switch next after the list head to new next which unlocks the 1428 // list head, but leaves the extracted mid locked: 1429 self->om_in_use_list->set_next_om(next); 1430 } else { 1431 // We have to search the list to find 'm'. 1432 om_unlock(mid); // unlock the list head 1433 guarantee(next != NULL, "thread=" INTPTR_FORMAT ": om_in_use_list=" INTPTR_FORMAT 1434 " is too short.", p2i(self), p2i(self->om_in_use_list)); 1435 // Our starting anchor is next after the list head which is the 1436 // last ObjectMonitor we checked: 1437 ObjectMonitor* anchor = next; 1438 while ((mid = unmarked_next(anchor)) != NULL) { 1439 if (m == mid) { 1440 // We found 'm' on the per-thread in-use list so extract it. 1441 om_lock(anchor); // Lock the anchor so we can safely modify it. 1442 // Update next to what follows mid (if anything): 1443 next = unmarked_next(mid); 1444 // Switch next after the anchor to new next which unlocks the 1445 // anchor, but leaves the extracted mid locked: 1446 anchor->set_next_om(next); 1447 break; 1448 } else { 1449 anchor = mid; 1450 } 1451 } 1452 } 1453 1454 if (mid == NULL) { 1455 // Reached end of the list and didn't find 'm' so: 1456 fatal("thread=" INTPTR_FORMAT " must find m=" INTPTR_FORMAT "on om_in_use_list=" 1457 INTPTR_FORMAT, p2i(self), p2i(m), p2i(self->om_in_use_list)); 1458 } 1459 1460 // At this point mid is disconnected from the in-use list so 1461 // its lock no longer has any effects on the in-use list. 1462 Atomic::dec(&self->om_in_use_count); 1463 // Unlock mid, but leave the next value for any lagging list 1464 // walkers. It will get cleaned up when mid is prepended to 1465 // the thread's free list: 1466 om_unlock(mid); 1467 } 1468 1469 prepend_to_om_free_list(self, m); 1470 } 1471 1472 // Return ObjectMonitors on a moribund thread's free and in-use 1473 // lists to the appropriate global lists. The ObjectMonitors on the 1474 // per-thread in-use list may still be in use by other threads. 1475 // 1476 // We currently call om_flush() from Threads::remove() before the 1477 // thread has been excised from the thread list and is no longer a 1478 // mutator. This means that om_flush() cannot run concurrently with 1479 // a safepoint and interleave with deflate_idle_monitors(). In 1480 // particular, this ensures that the thread's in-use monitors are 1481 // scanned by a GC safepoint, either via Thread::oops_do() (before 1482 // om_flush() is called) or via ObjectSynchronizer::oops_do() (after 1483 // om_flush() is called). 1484 1485 void ObjectSynchronizer::om_flush(Thread* self) { 1486 // Process the per-thread in-use list first to be consistent. 1487 int in_use_count = 0; 1488 ObjectMonitor* in_use_list = NULL; 1489 ObjectMonitor* in_use_tail = NULL; 1490 NoSafepointVerifier nsv; 1491 1492 // This function can race with a list walker thread so we lock the 1493 // list head to prevent confusion. 1494 if ((in_use_list = get_list_head_locked(&self->om_in_use_list)) != NULL) { 1495 // At this point, we have locked the in-use list head so a racing 1496 // thread cannot come in after us. However, a racing thread could 1497 // be ahead of us; we'll detect that and delay to let it finish. 1498 // 1499 // The thread is going away, however the ObjectMonitors on the 1500 // om_in_use_list may still be in-use by other threads. Link 1501 // them to in_use_tail, which will be linked into the global 1502 // in-use list (om_list_globals._in_use_list) below. 1503 // 1504 // Account for the in-use list head before the loop since it is 1505 // already locked (by this thread): 1506 in_use_tail = in_use_list; 1507 in_use_count++; 1508 for (ObjectMonitor* cur_om = unmarked_next(in_use_list); cur_om != NULL; cur_om = unmarked_next(cur_om)) { 1509 if (is_locked(cur_om)) { 1510 // cur_om is locked so there must be a racing walker thread ahead 1511 // of us so we'll give it a chance to finish. 1512 while (is_locked(cur_om)) { 1513 os::naked_short_sleep(1); 1514 } 1515 } 1516 in_use_tail = cur_om; 1517 in_use_count++; 1518 } 1519 guarantee(in_use_tail != NULL, "invariant"); 1520 int l_om_in_use_count = Atomic::load(&self->om_in_use_count); 1521 assert(l_om_in_use_count == in_use_count, "in-use counts don't match: " 1522 "l_om_in_use_count=%d, in_use_count=%d", l_om_in_use_count, in_use_count); 1523 Atomic::store(&self->om_in_use_count, 0); 1524 // Clear the in-use list head (which also unlocks it): 1525 Atomic::store(&self->om_in_use_list, (ObjectMonitor*)NULL); 1526 om_unlock(in_use_list); 1527 } 1528 1529 int free_count = 0; 1530 ObjectMonitor* free_list = NULL; 1531 ObjectMonitor* free_tail = NULL; 1532 // This function can race with a list walker thread so we lock the 1533 // list head to prevent confusion. 1534 if ((free_list = get_list_head_locked(&self->om_free_list)) != NULL) { 1535 // At this point, we have locked the free list head so a racing 1536 // thread cannot come in after us. However, a racing thread could 1537 // be ahead of us; we'll detect that and delay to let it finish. 1538 // 1539 // The thread is going away. Set 'free_tail' to the last per-thread free 1540 // monitor which will be linked to om_list_globals._free_list below. 1541 // 1542 // Account for the free list head before the loop since it is 1543 // already locked (by this thread): 1544 free_tail = free_list; 1545 free_count++; 1546 for (ObjectMonitor* s = unmarked_next(free_list); s != NULL; s = unmarked_next(s)) { 1547 if (is_locked(s)) { 1548 // s is locked so there must be a racing walker thread ahead 1549 // of us so we'll give it a chance to finish. 1550 while (is_locked(s)) { 1551 os::naked_short_sleep(1); 1552 } 1553 } 1554 free_tail = s; 1555 free_count++; 1556 guarantee(s->object() == NULL, "invariant"); 1557 stringStream ss; 1558 guarantee(!s->is_busy(), "must be !is_busy: %s", s->is_busy_to_string(&ss)); 1559 } 1560 guarantee(free_tail != NULL, "invariant"); 1561 int l_om_free_count = Atomic::load(&self->om_free_count); 1562 assert(l_om_free_count == free_count, "free counts don't match: " 1563 "l_om_free_count=%d, free_count=%d", l_om_free_count, free_count); 1564 Atomic::store(&self->om_free_count, 0); 1565 Atomic::store(&self->om_free_list, (ObjectMonitor*)NULL); 1566 om_unlock(free_list); 1567 } 1568 1569 if (free_tail != NULL) { 1570 prepend_list_to_global_free_list(free_list, free_tail, free_count); 1571 } 1572 1573 if (in_use_tail != NULL) { 1574 prepend_list_to_global_in_use_list(in_use_list, in_use_tail, in_use_count); 1575 } 1576 1577 LogStreamHandle(Debug, monitorinflation) lsh_debug; 1578 LogStreamHandle(Info, monitorinflation) lsh_info; 1579 LogStream* ls = NULL; 1580 if (log_is_enabled(Debug, monitorinflation)) { 1581 ls = &lsh_debug; 1582 } else if ((free_count != 0 || in_use_count != 0) && 1583 log_is_enabled(Info, monitorinflation)) { 1584 ls = &lsh_info; 1585 } 1586 if (ls != NULL) { 1587 ls->print_cr("om_flush: jt=" INTPTR_FORMAT ", free_count=%d" 1588 ", in_use_count=%d" ", om_free_provision=%d", 1589 p2i(self), free_count, in_use_count, self->om_free_provision); 1590 } 1591 } 1592 1593 static void post_monitor_inflate_event(EventJavaMonitorInflate* event, 1594 const oop obj, 1595 ObjectSynchronizer::InflateCause cause) { 1596 assert(event != NULL, "invariant"); 1597 assert(event->should_commit(), "invariant"); 1598 event->set_monitorClass(obj->klass()); 1599 event->set_address((uintptr_t)(void*)obj); 1600 event->set_cause((u1)cause); 1601 event->commit(); 1602 } 1603 1604 // Fast path code shared by multiple functions 1605 void ObjectSynchronizer::inflate_helper(oop obj) { 1606 markWord mark = obj->mark(); 1607 if (mark.has_monitor()) { 1608 assert(ObjectSynchronizer::verify_objmon_isinpool(mark.monitor()), "monitor is invalid"); 1609 assert(mark.monitor()->header().is_neutral(), "monitor must record a good object header"); 1610 return; 1611 } 1612 inflate(Thread::current(), obj, inflate_cause_vm_internal); 1613 } 1614 1615 ObjectMonitor* ObjectSynchronizer::inflate(Thread* self, 1616 oop object, const InflateCause cause) { 1617 // Inflate mutates the heap ... 1618 // Relaxing assertion for bug 6320749. 1619 assert(Universe::verify_in_progress() || 1620 !SafepointSynchronize::is_at_safepoint(), "invariant"); 1621 1622 EventJavaMonitorInflate event; 1623 1624 for (;;) { 1625 const markWord mark = object->mark(); 1626 assert(!mark.has_bias_pattern(), "invariant"); 1627 1628 // The mark can be in one of the following states: 1629 // * Inflated - just return 1630 // * Stack-locked - coerce it to inflated 1631 // * INFLATING - busy wait for conversion to complete 1632 // * Neutral - aggressively inflate the object. 1633 // * BIASED - Illegal. We should never see this 1634 1635 // CASE: inflated 1636 if (mark.has_monitor()) { 1637 ObjectMonitor* inf = mark.monitor(); 1638 markWord dmw = inf->header(); 1639 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1640 assert(inf->object() == object, "invariant"); 1641 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1642 return inf; 1643 } 1644 1645 // CASE: inflation in progress - inflating over a stack-lock. 1646 // Some other thread is converting from stack-locked to inflated. 1647 // Only that thread can complete inflation -- other threads must wait. 1648 // The INFLATING value is transient. 1649 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1650 // We could always eliminate polling by parking the thread on some auxiliary list. 1651 if (mark == markWord::INFLATING()) { 1652 read_stable_mark(object); 1653 continue; 1654 } 1655 1656 // CASE: stack-locked 1657 // Could be stack-locked either by this thread or by some other thread. 1658 // 1659 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1660 // to install INFLATING into the mark word. We originally installed INFLATING, 1661 // allocated the objectmonitor, and then finally STed the address of the 1662 // objectmonitor into the mark. This was correct, but artificially lengthened 1663 // the interval in which INFLATED appeared in the mark, thus increasing 1664 // the odds of inflation contention. 1665 // 1666 // We now use per-thread private objectmonitor free lists. 1667 // These list are reprovisioned from the global free list outside the 1668 // critical INFLATING...ST interval. A thread can transfer 1669 // multiple objectmonitors en-mass from the global free list to its local free list. 1670 // This reduces coherency traffic and lock contention on the global free list. 1671 // Using such local free lists, it doesn't matter if the om_alloc() call appears 1672 // before or after the CAS(INFLATING) operation. 1673 // See the comments in om_alloc(). 1674 1675 LogStreamHandle(Trace, monitorinflation) lsh; 1676 1677 if (mark.has_locker()) { 1678 ObjectMonitor* m = om_alloc(self); 1679 // Optimistically prepare the objectmonitor - anticipate successful CAS 1680 // We do this before the CAS in order to minimize the length of time 1681 // in which INFLATING appears in the mark. 1682 m->Recycle(); 1683 m->_Responsible = NULL; 1684 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class 1685 1686 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark); 1687 if (cmp != mark) { 1688 om_release(self, m, true); 1689 continue; // Interference -- just retry 1690 } 1691 1692 // We've successfully installed INFLATING (0) into the mark-word. 1693 // This is the only case where 0 will appear in a mark-word. 1694 // Only the singular thread that successfully swings the mark-word 1695 // to 0 can perform (or more precisely, complete) inflation. 1696 // 1697 // Why do we CAS a 0 into the mark-word instead of just CASing the 1698 // mark-word from the stack-locked value directly to the new inflated state? 1699 // Consider what happens when a thread unlocks a stack-locked object. 1700 // It attempts to use CAS to swing the displaced header value from the 1701 // on-stack BasicLock back into the object header. Recall also that the 1702 // header value (hash code, etc) can reside in (a) the object header, or 1703 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1704 // header in an ObjectMonitor. The inflate() routine must copy the header 1705 // value from the BasicLock on the owner's stack to the ObjectMonitor, all 1706 // the while preserving the hashCode stability invariants. If the owner 1707 // decides to release the lock while the value is 0, the unlock will fail 1708 // and control will eventually pass from slow_exit() to inflate. The owner 1709 // will then spin, waiting for the 0 value to disappear. Put another way, 1710 // the 0 causes the owner to stall if the owner happens to try to 1711 // drop the lock (restoring the header from the BasicLock to the object) 1712 // while inflation is in-progress. This protocol avoids races that might 1713 // would otherwise permit hashCode values to change or "flicker" for an object. 1714 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable. 1715 // 0 serves as a "BUSY" inflate-in-progress indicator. 1716 1717 1718 // fetch the displaced mark from the owner's stack. 1719 // The owner can't die or unwind past the lock while our INFLATING 1720 // object is in the mark. Furthermore the owner can't complete 1721 // an unlock on the object, either. 1722 markWord dmw = mark.displaced_mark_helper(); 1723 // Catch if the object's header is not neutral (not locked and 1724 // not marked is what we care about here). 1725 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1726 1727 // Setup monitor fields to proper values -- prepare the monitor 1728 m->set_header(dmw); 1729 1730 // Optimization: if the mark.locker stack address is associated 1731 // with this thread we could simply set m->_owner = self. 1732 // Note that a thread can inflate an object 1733 // that it has stack-locked -- as might happen in wait() -- directly 1734 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1735 m->set_owner_from(NULL, mark.locker()); 1736 m->set_object(object); 1737 // TODO-FIXME: assert BasicLock->dhw != 0. 1738 1739 // Must preserve store ordering. The monitor state must 1740 // be stable at the time of publishing the monitor address. 1741 guarantee(object->mark() == markWord::INFLATING(), "invariant"); 1742 object->release_set_mark(markWord::encode(m)); 1743 1744 // Hopefully the performance counters are allocated on distinct cache lines 1745 // to avoid false sharing on MP systems ... 1746 OM_PERFDATA_OP(Inflations, inc()); 1747 if (log_is_enabled(Trace, monitorinflation)) { 1748 ResourceMark rm(self); 1749 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark=" 1750 INTPTR_FORMAT ", type='%s'", p2i(object), 1751 object->mark().value(), object->klass()->external_name()); 1752 } 1753 if (event.should_commit()) { 1754 post_monitor_inflate_event(&event, object, cause); 1755 } 1756 return m; 1757 } 1758 1759 // CASE: neutral 1760 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1761 // If we know we're inflating for entry it's better to inflate by swinging a 1762 // pre-locked ObjectMonitor pointer into the object header. A successful 1763 // CAS inflates the object *and* confers ownership to the inflating thread. 1764 // In the current implementation we use a 2-step mechanism where we CAS() 1765 // to inflate and then CAS() again to try to swing _owner from NULL to self. 1766 // An inflateTry() method that we could call from enter() would be useful. 1767 1768 // Catch if the object's header is not neutral (not locked and 1769 // not marked is what we care about here). 1770 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 1771 ObjectMonitor* m = om_alloc(self); 1772 // prepare m for installation - set monitor to initial state 1773 m->Recycle(); 1774 m->set_header(mark); 1775 m->set_object(object); 1776 m->_Responsible = NULL; 1777 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class 1778 1779 if (object->cas_set_mark(markWord::encode(m), mark) != mark) { 1780 m->set_header(markWord::zero()); 1781 m->set_object(NULL); 1782 m->Recycle(); 1783 om_release(self, m, true); 1784 m = NULL; 1785 continue; 1786 // interference - the markword changed - just retry. 1787 // The state-transitions are one-way, so there's no chance of 1788 // live-lock -- "Inflated" is an absorbing state. 1789 } 1790 1791 // Hopefully the performance counters are allocated on distinct 1792 // cache lines to avoid false sharing on MP systems ... 1793 OM_PERFDATA_OP(Inflations, inc()); 1794 if (log_is_enabled(Trace, monitorinflation)) { 1795 ResourceMark rm(self); 1796 lsh.print_cr("inflate(neutral): object=" INTPTR_FORMAT ", mark=" 1797 INTPTR_FORMAT ", type='%s'", p2i(object), 1798 object->mark().value(), object->klass()->external_name()); 1799 } 1800 if (event.should_commit()) { 1801 post_monitor_inflate_event(&event, object, cause); 1802 } 1803 return m; 1804 } 1805 } 1806 1807 1808 // We maintain a list of in-use monitors for each thread. 1809 // 1810 // deflate_thread_local_monitors() scans a single thread's in-use list, while 1811 // deflate_idle_monitors() scans only a global list of in-use monitors which 1812 // is populated only as a thread dies (see om_flush()). 1813 // 1814 // These operations are called at all safepoints, immediately after mutators 1815 // are stopped, but before any objects have moved. Collectively they traverse 1816 // the population of in-use monitors, deflating where possible. The scavenged 1817 // monitors are returned to the global monitor free list. 1818 // 1819 // Beware that we scavenge at *every* stop-the-world point. Having a large 1820 // number of monitors in-use could negatively impact performance. We also want 1821 // to minimize the total # of monitors in circulation, as they incur a small 1822 // footprint penalty. 1823 // 1824 // Perversely, the heap size -- and thus the STW safepoint rate -- 1825 // typically drives the scavenge rate. Large heaps can mean infrequent GC, 1826 // which in turn can mean large(r) numbers of ObjectMonitors in circulation. 1827 // This is an unfortunate aspect of this design. 1828 1829 // Deflate a single monitor if not in-use 1830 // Return true if deflated, false if in-use 1831 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, 1832 ObjectMonitor** free_head_p, 1833 ObjectMonitor** free_tail_p) { 1834 bool deflated; 1835 // Normal case ... The monitor is associated with obj. 1836 const markWord mark = obj->mark(); 1837 guarantee(mark == markWord::encode(mid), "should match: mark=" 1838 INTPTR_FORMAT ", encoded mid=" INTPTR_FORMAT, mark.value(), 1839 markWord::encode(mid).value()); 1840 // Make sure that mark.monitor() and markWord::encode() agree: 1841 guarantee(mark.monitor() == mid, "should match: monitor()=" INTPTR_FORMAT 1842 ", mid=" INTPTR_FORMAT, p2i(mark.monitor()), p2i(mid)); 1843 const markWord dmw = mid->header(); 1844 guarantee(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1845 1846 if (mid->is_busy()) { 1847 // Easy checks are first - the ObjectMonitor is busy so no deflation. 1848 deflated = false; 1849 } else { 1850 // Deflate the monitor if it is no longer being used 1851 // It's idle - scavenge and return to the global free list 1852 // plain old deflation ... 1853 if (log_is_enabled(Trace, monitorinflation)) { 1854 ResourceMark rm; 1855 log_trace(monitorinflation)("deflate_monitor: " 1856 "object=" INTPTR_FORMAT ", mark=" 1857 INTPTR_FORMAT ", type='%s'", p2i(obj), 1858 mark.value(), obj->klass()->external_name()); 1859 } 1860 1861 // Restore the header back to obj 1862 obj->release_set_mark(dmw); 1863 mid->clear(); 1864 1865 assert(mid->object() == NULL, "invariant: object=" INTPTR_FORMAT, 1866 p2i(mid->object())); 1867 1868 // Move the deflated ObjectMonitor to the working free list 1869 // defined by free_head_p and free_tail_p. 1870 if (*free_head_p == NULL) *free_head_p = mid; 1871 if (*free_tail_p != NULL) { 1872 // We append to the list so the caller can use mid->_next_om 1873 // to fix the linkages in its context. 1874 ObjectMonitor* prevtail = *free_tail_p; 1875 // Should have been cleaned up by the caller: 1876 // Note: Should not have to lock prevtail here since we're at a 1877 // safepoint and ObjectMonitors on the local free list should 1878 // not be accessed in parallel. 1879 #ifdef ASSERT 1880 ObjectMonitor* l_next_om = prevtail->next_om(); 1881 #endif 1882 assert(l_next_om == NULL, "must be NULL: _next_om=" INTPTR_FORMAT, p2i(l_next_om)); 1883 prevtail->set_next_om(mid); 1884 } 1885 *free_tail_p = mid; 1886 // At this point, mid->_next_om still refers to its current 1887 // value and another ObjectMonitor's _next_om field still 1888 // refers to this ObjectMonitor. Those linkages have to be 1889 // cleaned up by the caller who has the complete context. 1890 deflated = true; 1891 } 1892 return deflated; 1893 } 1894 1895 // Walk a given monitor list, and deflate idle monitors. 1896 // The given list could be a per-thread list or a global list. 1897 // 1898 // In the case of parallel processing of thread local monitor lists, 1899 // work is done by Threads::parallel_threads_do() which ensures that 1900 // each Java thread is processed by exactly one worker thread, and 1901 // thus avoid conflicts that would arise when worker threads would 1902 // process the same monitor lists concurrently. 1903 // 1904 // See also ParallelSPCleanupTask and 1905 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and 1906 // Threads::parallel_java_threads_do() in thread.cpp. 1907 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** list_p, 1908 int* count_p, 1909 ObjectMonitor** free_head_p, 1910 ObjectMonitor** free_tail_p) { 1911 ObjectMonitor* cur_mid_in_use = NULL; 1912 ObjectMonitor* mid = NULL; 1913 ObjectMonitor* next = NULL; 1914 int deflated_count = 0; 1915 1916 // This list walk executes at a safepoint and does not race with any 1917 // other list walkers. 1918 1919 for (mid = Atomic::load(list_p); mid != NULL; mid = next) { 1920 next = unmarked_next(mid); 1921 oop obj = (oop) mid->object(); 1922 if (obj != NULL && deflate_monitor(mid, obj, free_head_p, free_tail_p)) { 1923 // Deflation succeeded and already updated free_head_p and 1924 // free_tail_p as needed. Finish the move to the local free list 1925 // by unlinking mid from the global or per-thread in-use list. 1926 if (cur_mid_in_use == NULL) { 1927 // mid is the list head so switch the list head to next: 1928 Atomic::store(list_p, next); 1929 } else { 1930 // Switch cur_mid_in_use's next field to next: 1931 cur_mid_in_use->set_next_om(next); 1932 } 1933 // At this point mid is disconnected from the in-use list. 1934 deflated_count++; 1935 Atomic::dec(count_p); 1936 // mid is current tail in the free_head_p list so NULL terminate it: 1937 mid->set_next_om(NULL); 1938 } else { 1939 cur_mid_in_use = mid; 1940 } 1941 } 1942 return deflated_count; 1943 } 1944 1945 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) { 1946 counters->n_in_use = 0; // currently associated with objects 1947 counters->n_in_circulation = 0; // extant 1948 counters->n_scavenged = 0; // reclaimed (global and per-thread) 1949 counters->per_thread_scavenged = 0; // per-thread scavenge total 1950 counters->per_thread_times = 0.0; // per-thread scavenge times 1951 } 1952 1953 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) { 1954 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1955 bool deflated = false; 1956 1957 ObjectMonitor* free_head_p = NULL; // Local SLL of scavenged monitors 1958 ObjectMonitor* free_tail_p = NULL; 1959 elapsedTimer timer; 1960 1961 if (log_is_enabled(Info, monitorinflation)) { 1962 timer.start(); 1963 } 1964 1965 // Note: the thread-local monitors lists get deflated in 1966 // a separate pass. See deflate_thread_local_monitors(). 1967 1968 // For moribund threads, scan om_list_globals._in_use_list 1969 int deflated_count = 0; 1970 if (Atomic::load(&om_list_globals._in_use_list) != NULL) { 1971 // Update n_in_circulation before om_list_globals._in_use_count is 1972 // updated by deflation. 1973 Atomic::add(&counters->n_in_circulation, 1974 Atomic::load(&om_list_globals._in_use_count)); 1975 1976 deflated_count = deflate_monitor_list(&om_list_globals._in_use_list, 1977 &om_list_globals._in_use_count, 1978 &free_head_p, &free_tail_p); 1979 Atomic::add(&counters->n_in_use, Atomic::load(&om_list_globals._in_use_count)); 1980 } 1981 1982 if (free_head_p != NULL) { 1983 // Move the deflated ObjectMonitors back to the global free list. 1984 guarantee(free_tail_p != NULL && deflated_count > 0, "invariant"); 1985 #ifdef ASSERT 1986 ObjectMonitor* l_next_om = free_tail_p->next_om(); 1987 #endif 1988 assert(l_next_om == NULL, "must be NULL: _next_om=" INTPTR_FORMAT, p2i(l_next_om)); 1989 prepend_list_to_global_free_list(free_head_p, free_tail_p, deflated_count); 1990 Atomic::add(&counters->n_scavenged, deflated_count); 1991 } 1992 timer.stop(); 1993 1994 LogStreamHandle(Debug, monitorinflation) lsh_debug; 1995 LogStreamHandle(Info, monitorinflation) lsh_info; 1996 LogStream* ls = NULL; 1997 if (log_is_enabled(Debug, monitorinflation)) { 1998 ls = &lsh_debug; 1999 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) { 2000 ls = &lsh_info; 2001 } 2002 if (ls != NULL) { 2003 ls->print_cr("deflating global idle monitors, %3.7f secs, %d monitors", timer.seconds(), deflated_count); 2004 } 2005 } 2006 2007 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) { 2008 // Report the cumulative time for deflating each thread's idle 2009 // monitors. Note: if the work is split among more than one 2010 // worker thread, then the reported time will likely be more 2011 // than a beginning to end measurement of the phase. 2012 log_info(safepoint, cleanup)("deflating per-thread idle monitors, %3.7f secs, monitors=%d", counters->per_thread_times, counters->per_thread_scavenged); 2013 2014 if (log_is_enabled(Debug, monitorinflation)) { 2015 // exit_globals()'s call to audit_and_print_stats() is done 2016 // at the Info level and not at a safepoint. 2017 ObjectSynchronizer::audit_and_print_stats(false /* on_exit */); 2018 } else if (log_is_enabled(Info, monitorinflation)) { 2019 log_info(monitorinflation)("global_population=%d, global_in_use_count=%d, " 2020 "global_free_count=%d", 2021 Atomic::load(&om_list_globals._population), 2022 Atomic::load(&om_list_globals._in_use_count), 2023 Atomic::load(&om_list_globals._free_count)); 2024 } 2025 2026 Atomic::store(&_forceMonitorScavenge, 0); // Reset 2027 2028 OM_PERFDATA_OP(Deflations, inc(counters->n_scavenged)); 2029 OM_PERFDATA_OP(MonExtant, set_value(counters->n_in_circulation)); 2030 2031 GVars.stw_random = os::random(); 2032 GVars.stw_cycle++; 2033 } 2034 2035 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) { 2036 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 2037 2038 ObjectMonitor* free_head_p = NULL; // Local SLL of scavenged monitors 2039 ObjectMonitor* free_tail_p = NULL; 2040 elapsedTimer timer; 2041 2042 if (log_is_enabled(Info, safepoint, cleanup) || 2043 log_is_enabled(Info, monitorinflation)) { 2044 timer.start(); 2045 } 2046 2047 // Update n_in_circulation before om_in_use_count is updated by deflation. 2048 Atomic::add(&counters->n_in_circulation, Atomic::load(&thread->om_in_use_count)); 2049 2050 int deflated_count = deflate_monitor_list(&thread->om_in_use_list, &thread->om_in_use_count, &free_head_p, &free_tail_p); 2051 Atomic::add(&counters->n_in_use, Atomic::load(&thread->om_in_use_count)); 2052 2053 if (free_head_p != NULL) { 2054 // Move the deflated ObjectMonitors back to the global free list. 2055 guarantee(free_tail_p != NULL && deflated_count > 0, "invariant"); 2056 #ifdef ASSERT 2057 ObjectMonitor* l_next_om = free_tail_p->next_om(); 2058 #endif 2059 assert(l_next_om == NULL, "must be NULL: _next_om=" INTPTR_FORMAT, p2i(l_next_om)); 2060 prepend_list_to_global_free_list(free_head_p, free_tail_p, deflated_count); 2061 Atomic::add(&counters->n_scavenged, deflated_count); 2062 Atomic::add(&counters->per_thread_scavenged, deflated_count); 2063 } 2064 2065 timer.stop(); 2066 counters->per_thread_times += timer.seconds(); 2067 2068 LogStreamHandle(Debug, monitorinflation) lsh_debug; 2069 LogStreamHandle(Info, monitorinflation) lsh_info; 2070 LogStream* ls = NULL; 2071 if (log_is_enabled(Debug, monitorinflation)) { 2072 ls = &lsh_debug; 2073 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) { 2074 ls = &lsh_info; 2075 } 2076 if (ls != NULL) { 2077 ls->print_cr("jt=" INTPTR_FORMAT ": deflating per-thread idle monitors, %3.7f secs, %d monitors", p2i(thread), timer.seconds(), deflated_count); 2078 } 2079 } 2080 2081 // Monitor cleanup on JavaThread::exit 2082 2083 // Iterate through monitor cache and attempt to release thread's monitors 2084 // Gives up on a particular monitor if an exception occurs, but continues 2085 // the overall iteration, swallowing the exception. 2086 class ReleaseJavaMonitorsClosure: public MonitorClosure { 2087 private: 2088 TRAPS; 2089 2090 public: 2091 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 2092 void do_monitor(ObjectMonitor* mid) { 2093 if (mid->owner() == THREAD) { 2094 (void)mid->complete_exit(CHECK); 2095 } 2096 } 2097 }; 2098 2099 // Release all inflated monitors owned by THREAD. Lightweight monitors are 2100 // ignored. This is meant to be called during JNI thread detach which assumes 2101 // all remaining monitors are heavyweight. All exceptions are swallowed. 2102 // Scanning the extant monitor list can be time consuming. 2103 // A simple optimization is to add a per-thread flag that indicates a thread 2104 // called jni_monitorenter() during its lifetime. 2105 // 2106 // Instead of No_Savepoint_Verifier it might be cheaper to 2107 // use an idiom of the form: 2108 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 2109 // <code that must not run at safepoint> 2110 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 2111 // Since the tests are extremely cheap we could leave them enabled 2112 // for normal product builds. 2113 2114 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 2115 assert(THREAD == JavaThread::current(), "must be current Java thread"); 2116 NoSafepointVerifier nsv; 2117 ReleaseJavaMonitorsClosure rjmc(THREAD); 2118 ObjectSynchronizer::monitors_iterate(&rjmc); 2119 THREAD->clear_pending_exception(); 2120 } 2121 2122 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 2123 switch (cause) { 2124 case inflate_cause_vm_internal: return "VM Internal"; 2125 case inflate_cause_monitor_enter: return "Monitor Enter"; 2126 case inflate_cause_wait: return "Monitor Wait"; 2127 case inflate_cause_notify: return "Monitor Notify"; 2128 case inflate_cause_hash_code: return "Monitor Hash Code"; 2129 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 2130 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 2131 default: 2132 ShouldNotReachHere(); 2133 } 2134 return "Unknown"; 2135 } 2136 2137 //------------------------------------------------------------------------------ 2138 // Debugging code 2139 2140 u_char* ObjectSynchronizer::get_gvars_addr() { 2141 return (u_char*)&GVars; 2142 } 2143 2144 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() { 2145 return (u_char*)&GVars.hc_sequence; 2146 } 2147 2148 size_t ObjectSynchronizer::get_gvars_size() { 2149 return sizeof(SharedGlobals); 2150 } 2151 2152 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() { 2153 return (u_char*)&GVars.stw_random; 2154 } 2155 2156 // This function can be called at a safepoint or it can be called when 2157 // we are trying to exit the VM. When we are trying to exit the VM, the 2158 // list walker functions can run in parallel with the other list 2159 // operations so spin-locking is used for safety. 2160 // 2161 // Calls to this function can be added in various places as a debugging 2162 // aid; pass 'true' for the 'on_exit' parameter to have in-use monitor 2163 // details logged at the Info level and 'false' for the 'on_exit' 2164 // parameter to have in-use monitor details logged at the Trace level. 2165 // 2166 void ObjectSynchronizer::audit_and_print_stats(bool on_exit) { 2167 assert(on_exit || SafepointSynchronize::is_at_safepoint(), "invariant"); 2168 2169 LogStreamHandle(Debug, monitorinflation) lsh_debug; 2170 LogStreamHandle(Info, monitorinflation) lsh_info; 2171 LogStreamHandle(Trace, monitorinflation) lsh_trace; 2172 LogStream* ls = NULL; 2173 if (log_is_enabled(Trace, monitorinflation)) { 2174 ls = &lsh_trace; 2175 } else if (log_is_enabled(Debug, monitorinflation)) { 2176 ls = &lsh_debug; 2177 } else if (log_is_enabled(Info, monitorinflation)) { 2178 ls = &lsh_info; 2179 } 2180 assert(ls != NULL, "sanity check"); 2181 2182 // Log counts for the global and per-thread monitor lists: 2183 int chk_om_population = log_monitor_list_counts(ls); 2184 int error_cnt = 0; 2185 2186 ls->print_cr("Checking global lists:"); 2187 2188 // Check om_list_globals._population: 2189 if (Atomic::load(&om_list_globals._population) == chk_om_population) { 2190 ls->print_cr("global_population=%d equals chk_om_population=%d", 2191 Atomic::load(&om_list_globals._population), chk_om_population); 2192 } else { 2193 // With fine grained locks on the monitor lists, it is possible for 2194 // log_monitor_list_counts() to return a value that doesn't match 2195 // om_list_globals._population. So far a higher value has been 2196 // seen in testing so something is being double counted by 2197 // log_monitor_list_counts(). 2198 ls->print_cr("WARNING: global_population=%d is not equal to " 2199 "chk_om_population=%d", 2200 Atomic::load(&om_list_globals._population), chk_om_population); 2201 } 2202 2203 // Check om_list_globals._in_use_list and om_list_globals._in_use_count: 2204 chk_global_in_use_list_and_count(ls, &error_cnt); 2205 2206 // Check om_list_globals._free_list and om_list_globals._free_count: 2207 chk_global_free_list_and_count(ls, &error_cnt); 2208 2209 ls->print_cr("Checking per-thread lists:"); 2210 2211 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 2212 // Check om_in_use_list and om_in_use_count: 2213 chk_per_thread_in_use_list_and_count(jt, ls, &error_cnt); 2214 2215 // Check om_free_list and om_free_count: 2216 chk_per_thread_free_list_and_count(jt, ls, &error_cnt); 2217 } 2218 2219 if (error_cnt == 0) { 2220 ls->print_cr("No errors found in monitor list checks."); 2221 } else { 2222 log_error(monitorinflation)("found monitor list errors: error_cnt=%d", error_cnt); 2223 } 2224 2225 if ((on_exit && log_is_enabled(Info, monitorinflation)) || 2226 (!on_exit && log_is_enabled(Trace, monitorinflation))) { 2227 // When exiting this log output is at the Info level. When called 2228 // at a safepoint, this log output is at the Trace level since 2229 // there can be a lot of it. 2230 log_in_use_monitor_details(ls); 2231 } 2232 2233 ls->flush(); 2234 2235 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt); 2236 } 2237 2238 // Check a free monitor entry; log any errors. 2239 void ObjectSynchronizer::chk_free_entry(JavaThread* jt, ObjectMonitor* n, 2240 outputStream * out, int *error_cnt_p) { 2241 stringStream ss; 2242 if (n->is_busy()) { 2243 if (jt != NULL) { 2244 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2245 ": free per-thread monitor must not be busy: %s", p2i(jt), 2246 p2i(n), n->is_busy_to_string(&ss)); 2247 } else { 2248 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor " 2249 "must not be busy: %s", p2i(n), n->is_busy_to_string(&ss)); 2250 } 2251 *error_cnt_p = *error_cnt_p + 1; 2252 } 2253 if (n->header().value() != 0) { 2254 if (jt != NULL) { 2255 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2256 ": free per-thread monitor must have NULL _header " 2257 "field: _header=" INTPTR_FORMAT, p2i(jt), p2i(n), 2258 n->header().value()); 2259 } else { 2260 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor " 2261 "must have NULL _header field: _header=" INTPTR_FORMAT, 2262 p2i(n), n->header().value()); 2263 } 2264 *error_cnt_p = *error_cnt_p + 1; 2265 } 2266 if (n->object() != NULL) { 2267 if (jt != NULL) { 2268 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2269 ": free per-thread monitor must have NULL _object " 2270 "field: _object=" INTPTR_FORMAT, p2i(jt), p2i(n), 2271 p2i(n->object())); 2272 } else { 2273 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor " 2274 "must have NULL _object field: _object=" INTPTR_FORMAT, 2275 p2i(n), p2i(n->object())); 2276 } 2277 *error_cnt_p = *error_cnt_p + 1; 2278 } 2279 } 2280 2281 // Lock the next ObjectMonitor for traversal and unlock the current 2282 // ObjectMonitor. Returns the next ObjectMonitor if there is one. 2283 // Otherwise returns NULL (after unlocking the current ObjectMonitor). 2284 // This function is used by the various list walker functions to 2285 // safely walk a list without allowing an ObjectMonitor to be moved 2286 // to another list in the middle of a walk. 2287 static ObjectMonitor* lock_next_for_traversal(ObjectMonitor* cur) { 2288 assert(is_locked(cur), "cur=" INTPTR_FORMAT " must be locked", p2i(cur)); 2289 ObjectMonitor* next = unmarked_next(cur); 2290 if (next == NULL) { // Reached the end of the list. 2291 om_unlock(cur); 2292 return NULL; 2293 } 2294 om_lock(next); // Lock next before unlocking current to keep 2295 om_unlock(cur); // from being by-passed by another thread. 2296 return next; 2297 } 2298 2299 // Check the global free list and count; log the results of the checks. 2300 void ObjectSynchronizer::chk_global_free_list_and_count(outputStream * out, 2301 int *error_cnt_p) { 2302 int chk_om_free_count = 0; 2303 ObjectMonitor* cur = NULL; 2304 if ((cur = get_list_head_locked(&om_list_globals._free_list)) != NULL) { 2305 // Marked the global free list head so process the list. 2306 while (true) { 2307 chk_free_entry(NULL /* jt */, cur, out, error_cnt_p); 2308 chk_om_free_count++; 2309 2310 cur = lock_next_for_traversal(cur); 2311 if (cur == NULL) { 2312 break; 2313 } 2314 } 2315 } 2316 int l_free_count = Atomic::load(&om_list_globals._free_count); 2317 if (l_free_count == chk_om_free_count) { 2318 out->print_cr("global_free_count=%d equals chk_om_free_count=%d", 2319 l_free_count, chk_om_free_count); 2320 } else { 2321 // With fine grained locks on om_list_globals._free_list, it 2322 // is possible for an ObjectMonitor to be prepended to 2323 // om_list_globals._free_list after we started calculating 2324 // chk_om_free_count so om_list_globals._free_count may not 2325 // match anymore. 2326 out->print_cr("WARNING: global_free_count=%d is not equal to " 2327 "chk_om_free_count=%d", l_free_count, chk_om_free_count); 2328 } 2329 } 2330 2331 // Check the global in-use list and count; log the results of the checks. 2332 void ObjectSynchronizer::chk_global_in_use_list_and_count(outputStream * out, 2333 int *error_cnt_p) { 2334 int chk_om_in_use_count = 0; 2335 ObjectMonitor* cur = NULL; 2336 if ((cur = get_list_head_locked(&om_list_globals._in_use_list)) != NULL) { 2337 // Marked the global in-use list head so process the list. 2338 while (true) { 2339 chk_in_use_entry(NULL /* jt */, cur, out, error_cnt_p); 2340 chk_om_in_use_count++; 2341 2342 cur = lock_next_for_traversal(cur); 2343 if (cur == NULL) { 2344 break; 2345 } 2346 } 2347 } 2348 int l_in_use_count = Atomic::load(&om_list_globals._in_use_count); 2349 if (l_in_use_count == chk_om_in_use_count) { 2350 out->print_cr("global_in_use_count=%d equals chk_om_in_use_count=%d", 2351 l_in_use_count, chk_om_in_use_count); 2352 } else { 2353 // With fine grained locks on the monitor lists, it is possible for 2354 // an exiting JavaThread to put its in-use ObjectMonitors on the 2355 // global in-use list after chk_om_in_use_count is calculated above. 2356 out->print_cr("WARNING: global_in_use_count=%d is not equal to chk_om_in_use_count=%d", 2357 l_in_use_count, chk_om_in_use_count); 2358 } 2359 } 2360 2361 // Check an in-use monitor entry; log any errors. 2362 void ObjectSynchronizer::chk_in_use_entry(JavaThread* jt, ObjectMonitor* n, 2363 outputStream * out, int *error_cnt_p) { 2364 if (n->header().value() == 0) { 2365 if (jt != NULL) { 2366 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2367 ": in-use per-thread monitor must have non-NULL _header " 2368 "field.", p2i(jt), p2i(n)); 2369 } else { 2370 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use global monitor " 2371 "must have non-NULL _header field.", p2i(n)); 2372 } 2373 *error_cnt_p = *error_cnt_p + 1; 2374 } 2375 if (n->object() == NULL) { 2376 if (jt != NULL) { 2377 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2378 ": in-use per-thread monitor must have non-NULL _object " 2379 "field.", p2i(jt), p2i(n)); 2380 } else { 2381 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use global monitor " 2382 "must have non-NULL _object field.", p2i(n)); 2383 } 2384 *error_cnt_p = *error_cnt_p + 1; 2385 } 2386 const oop obj = (oop)n->object(); 2387 const markWord mark = obj->mark(); 2388 if (!mark.has_monitor()) { 2389 if (jt != NULL) { 2390 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2391 ": in-use per-thread monitor's object does not think " 2392 "it has a monitor: obj=" INTPTR_FORMAT ", mark=" 2393 INTPTR_FORMAT, p2i(jt), p2i(n), p2i(obj), mark.value()); 2394 } else { 2395 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use global " 2396 "monitor's object does not think it has a monitor: obj=" 2397 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n), 2398 p2i(obj), mark.value()); 2399 } 2400 *error_cnt_p = *error_cnt_p + 1; 2401 } 2402 ObjectMonitor* const obj_mon = mark.monitor(); 2403 if (n != obj_mon) { 2404 if (jt != NULL) { 2405 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2406 ": in-use per-thread monitor's object does not refer " 2407 "to the same monitor: obj=" INTPTR_FORMAT ", mark=" 2408 INTPTR_FORMAT ", obj_mon=" INTPTR_FORMAT, p2i(jt), 2409 p2i(n), p2i(obj), mark.value(), p2i(obj_mon)); 2410 } else { 2411 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use global " 2412 "monitor's object does not refer to the same monitor: obj=" 2413 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon=" 2414 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon)); 2415 } 2416 *error_cnt_p = *error_cnt_p + 1; 2417 } 2418 } 2419 2420 // Check the thread's free list and count; log the results of the checks. 2421 void ObjectSynchronizer::chk_per_thread_free_list_and_count(JavaThread *jt, 2422 outputStream * out, 2423 int *error_cnt_p) { 2424 int chk_om_free_count = 0; 2425 ObjectMonitor* cur = NULL; 2426 if ((cur = get_list_head_locked(&jt->om_free_list)) != NULL) { 2427 // Marked the per-thread free list head so process the list. 2428 while (true) { 2429 chk_free_entry(jt, cur, out, error_cnt_p); 2430 chk_om_free_count++; 2431 2432 cur = lock_next_for_traversal(cur); 2433 if (cur == NULL) { 2434 break; 2435 } 2436 } 2437 } 2438 int l_om_free_count = Atomic::load(&jt->om_free_count); 2439 if (l_om_free_count == chk_om_free_count) { 2440 out->print_cr("jt=" INTPTR_FORMAT ": om_free_count=%d equals " 2441 "chk_om_free_count=%d", p2i(jt), l_om_free_count, chk_om_free_count); 2442 } else { 2443 out->print_cr("ERROR: jt=" INTPTR_FORMAT ": om_free_count=%d is not " 2444 "equal to chk_om_free_count=%d", p2i(jt), l_om_free_count, 2445 chk_om_free_count); 2446 *error_cnt_p = *error_cnt_p + 1; 2447 } 2448 } 2449 2450 // Check the thread's in-use list and count; log the results of the checks. 2451 void ObjectSynchronizer::chk_per_thread_in_use_list_and_count(JavaThread *jt, 2452 outputStream * out, 2453 int *error_cnt_p) { 2454 int chk_om_in_use_count = 0; 2455 ObjectMonitor* cur = NULL; 2456 if ((cur = get_list_head_locked(&jt->om_in_use_list)) != NULL) { 2457 // Marked the per-thread in-use list head so process the list. 2458 while (true) { 2459 chk_in_use_entry(jt, cur, out, error_cnt_p); 2460 chk_om_in_use_count++; 2461 2462 cur = lock_next_for_traversal(cur); 2463 if (cur == NULL) { 2464 break; 2465 } 2466 } 2467 } 2468 int l_om_in_use_count = Atomic::load(&jt->om_in_use_count); 2469 if (l_om_in_use_count == chk_om_in_use_count) { 2470 out->print_cr("jt=" INTPTR_FORMAT ": om_in_use_count=%d equals " 2471 "chk_om_in_use_count=%d", p2i(jt), l_om_in_use_count, 2472 chk_om_in_use_count); 2473 } else { 2474 out->print_cr("ERROR: jt=" INTPTR_FORMAT ": om_in_use_count=%d is not " 2475 "equal to chk_om_in_use_count=%d", p2i(jt), l_om_in_use_count, 2476 chk_om_in_use_count); 2477 *error_cnt_p = *error_cnt_p + 1; 2478 } 2479 } 2480 2481 // Log details about ObjectMonitors on the in-use lists. The 'BHL' 2482 // flags indicate why the entry is in-use, 'object' and 'object type' 2483 // indicate the associated object and its type. 2484 void ObjectSynchronizer::log_in_use_monitor_details(outputStream * out) { 2485 stringStream ss; 2486 if (Atomic::load(&om_list_globals._in_use_count) > 0) { 2487 out->print_cr("In-use global monitor info:"); 2488 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)"); 2489 out->print_cr("%18s %s %18s %18s", 2490 "monitor", "BHL", "object", "object type"); 2491 out->print_cr("================== === ================== =================="); 2492 ObjectMonitor* cur = NULL; 2493 if ((cur = get_list_head_locked(&om_list_globals._in_use_list)) != NULL) { 2494 // Marked the global in-use list head so process the list. 2495 while (true) { 2496 const oop obj = (oop) cur->object(); 2497 const markWord mark = cur->header(); 2498 ResourceMark rm; 2499 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(cur), 2500 cur->is_busy() != 0, mark.hash() != 0, cur->owner() != NULL, 2501 p2i(obj), obj->klass()->external_name()); 2502 if (cur->is_busy() != 0) { 2503 out->print(" (%s)", cur->is_busy_to_string(&ss)); 2504 ss.reset(); 2505 } 2506 out->cr(); 2507 2508 cur = lock_next_for_traversal(cur); 2509 if (cur == NULL) { 2510 break; 2511 } 2512 } 2513 } 2514 } 2515 2516 out->print_cr("In-use per-thread monitor info:"); 2517 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)"); 2518 out->print_cr("%18s %18s %s %18s %18s", 2519 "jt", "monitor", "BHL", "object", "object type"); 2520 out->print_cr("================== ================== === ================== =================="); 2521 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 2522 ObjectMonitor* cur = NULL; 2523 if ((cur = get_list_head_locked(&jt->om_in_use_list)) != NULL) { 2524 // Marked the global in-use list head so process the list. 2525 while (true) { 2526 const oop obj = (oop) cur->object(); 2527 const markWord mark = cur->header(); 2528 ResourceMark rm; 2529 out->print(INTPTR_FORMAT " " INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT 2530 " %s", p2i(jt), p2i(cur), cur->is_busy() != 0, 2531 mark.hash() != 0, cur->owner() != NULL, p2i(obj), 2532 obj->klass()->external_name()); 2533 if (cur->is_busy() != 0) { 2534 out->print(" (%s)", cur->is_busy_to_string(&ss)); 2535 ss.reset(); 2536 } 2537 out->cr(); 2538 2539 cur = lock_next_for_traversal(cur); 2540 if (cur == NULL) { 2541 break; 2542 } 2543 } 2544 } 2545 } 2546 2547 out->flush(); 2548 } 2549 2550 // Log counts for the global and per-thread monitor lists and return 2551 // the population count. 2552 int ObjectSynchronizer::log_monitor_list_counts(outputStream * out) { 2553 int pop_count = 0; 2554 out->print_cr("%18s %10s %10s %10s", 2555 "Global Lists:", "InUse", "Free", "Total"); 2556 out->print_cr("================== ========== ========== =========="); 2557 int l_in_use_count = Atomic::load(&om_list_globals._in_use_count); 2558 int l_free_count = Atomic::load(&om_list_globals._free_count); 2559 out->print_cr("%18s %10d %10d %10d", "", l_in_use_count, 2560 l_free_count, Atomic::load(&om_list_globals._population)); 2561 pop_count += l_in_use_count + l_free_count; 2562 2563 out->print_cr("%18s %10s %10s %10s", 2564 "Per-Thread Lists:", "InUse", "Free", "Provision"); 2565 out->print_cr("================== ========== ========== =========="); 2566 2567 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 2568 int l_om_in_use_count = Atomic::load(&jt->om_in_use_count); 2569 int l_om_free_count = Atomic::load(&jt->om_free_count); 2570 out->print_cr(INTPTR_FORMAT " %10d %10d %10d", p2i(jt), 2571 l_om_in_use_count, l_om_free_count, jt->om_free_provision); 2572 pop_count += l_om_in_use_count + l_om_free_count; 2573 } 2574 return pop_count; 2575 } 2576 2577 #ifndef PRODUCT 2578 2579 // Check if monitor belongs to the monitor cache 2580 // The list is grow-only so it's *relatively* safe to traverse 2581 // the list of extant blocks without taking a lock. 2582 2583 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 2584 PaddedObjectMonitor* block = Atomic::load(&g_block_list); 2585 while (block != NULL) { 2586 assert(block->object() == CHAINMARKER, "must be a block header"); 2587 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) { 2588 address mon = (address)monitor; 2589 address blk = (address)block; 2590 size_t diff = mon - blk; 2591 assert((diff % sizeof(PaddedObjectMonitor)) == 0, "must be aligned"); 2592 return 1; 2593 } 2594 // unmarked_next() is not needed with g_block_list (no locking 2595 // used with block linkage _next_om fields). 2596 block = (PaddedObjectMonitor*)block->next_om(); 2597 } 2598 return 0; 2599 } 2600 2601 #endif