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 ListGlobals { 123 char _pad_prefix[DEFAULT_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, DEFAULT_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, DEFAULT_CACHE_LINE_SIZE, sizeof(ObjectMonitor*)); 136 137 int free_count; // # on free_list 138 DEFINE_PAD_MINUS_SIZE(3, DEFAULT_CACHE_LINE_SIZE, sizeof(int)); 139 140 int in_use_count; // # on in_use_list 141 DEFINE_PAD_MINUS_SIZE(4, DEFAULT_CACHE_LINE_SIZE, sizeof(int)); 142 143 int population; // # Extant -- in circulation 144 DEFINE_PAD_MINUS_SIZE(5, DEFAULT_CACHE_LINE_SIZE, sizeof(int)); 145 }; 146 static ListGlobals LVars; 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)Atomic::load(&om->_next_om) & OM_LOCK_BIT) == OM_LOCK_BIT; 163 } 164 165 // Mark an ObjectMonitor* with OM_LOCK_BIT and return it. 166 // Note: the om parameter may or may not have been marked originally. 167 static ObjectMonitor* mark_om_ptr(ObjectMonitor* om) { 168 return (ObjectMonitor*)((intptr_t)om | OM_LOCK_BIT); 169 } 170 171 // Try to lock an ObjectMonitor. Returns true if locking was successful. 172 // Otherwise returns false. 173 static bool try_om_lock(ObjectMonitor* om) { 174 // Get current next field without any OM_LOCK_BIT value. 175 ObjectMonitor* next = (ObjectMonitor*)((intptr_t)Atomic::load(&om->_next_om) & ~OM_LOCK_BIT); 176 if (Atomic::cmpxchg(&om->_next_om, next, mark_om_ptr(next)) != next) { 177 return false; // Cannot lock the ObjectMonitor. 178 } 179 return true; 180 } 181 182 // Lock an ObjectMonitor. 183 static void om_lock(ObjectMonitor* om) { 184 while (true) { 185 if (try_om_lock(om)) { 186 return; 187 } 188 } 189 } 190 191 // Unlock an ObjectMonitor. 192 static void om_unlock(ObjectMonitor* om) { 193 ObjectMonitor* next = Atomic::load(&om->_next_om); 194 guarantee(((intptr_t)next & OM_LOCK_BIT) == OM_LOCK_BIT, "next=" INTPTR_FORMAT 195 " must have OM_LOCK_BIT=%x set.", p2i(next), OM_LOCK_BIT); 196 197 next = (ObjectMonitor*)((intptr_t)next & ~OM_LOCK_BIT); // Clear OM_LOCK_BIT. 198 Atomic::store(&om->_next_om, next); 199 } 200 201 // Get the list head after locking it. Returns the list head or NULL 202 // if the list is empty. 203 static ObjectMonitor* get_list_head_locked(ObjectMonitor** list_p) { 204 while (true) { 205 ObjectMonitor* mid = Atomic::load(list_p); 206 if (mid == NULL) { 207 return NULL; // The list is empty. 208 } 209 if (try_om_lock(mid)) { 210 if (Atomic::load(list_p) != mid) { 211 // The list head changed so we have to retry. 212 om_unlock(mid); 213 continue; 214 } 215 return mid; 216 } 217 } 218 } 219 220 // Return the unmarked next field in an ObjectMonitor. Note: the next 221 // field may or may not have been marked with OM_LOCK_BIT originally. 222 static ObjectMonitor* unmarked_next(ObjectMonitor* om) { 223 return (ObjectMonitor*)((intptr_t)Atomic::load(&om->_next_om) & ~OM_LOCK_BIT); 224 } 225 226 #undef OM_LOCK_BIT 227 228 229 // =====================> List Management functions 230 231 // Set the next field in an ObjectMonitor to the specified value. 232 static void set_next(ObjectMonitor* om, ObjectMonitor* value) { 233 Atomic::store(&om->_next_om, value); 234 } 235 236 // Prepend a list of ObjectMonitors to the specified *list_p. 'tail' is 237 // the last ObjectMonitor in the list and there are 'count' on the list. 238 // Also updates the specified *count_p. 239 static void prepend_list_to_common(ObjectMonitor* list, ObjectMonitor* tail, 240 int count, ObjectMonitor** list_p, 241 int* count_p) { 242 while (true) { 243 ObjectMonitor* cur = Atomic::load(list_p); 244 // Prepend list to *list_p. 245 if (!try_om_lock(tail)) { 246 continue; // failed to lock tail so try it all again 247 } 248 set_next(tail, cur); // tail now points to cur (and unlocks tail) 249 if (cur == NULL) { 250 // No potential race with takers or other prependers since 251 // *list_p is empty. 252 if (Atomic::cmpxchg(list_p, cur, list) == cur) { 253 // Successfully switched *list_p to the list value. 254 Atomic::add(count_p, count); 255 break; 256 } 257 // Implied else: try it all again 258 } else { 259 if (!try_om_lock(cur)) { 260 continue; // failed to lock cur so try it all again 261 } 262 // We locked cur so try to switch *list_p to the list value. 263 if (Atomic::cmpxchg(list_p, cur, list) != cur) { 264 // The list head has changed so unlock cur and try again: 265 om_unlock(cur); 266 continue; 267 } 268 Atomic::add(count_p, count); 269 om_unlock(cur); 270 break; 271 } 272 } 273 } 274 275 // Prepend a newly allocated block of ObjectMonitors to g_block_list and 276 // LVars.free_list. Also updates LVars.population and LVars.free_count. 277 void ObjectSynchronizer::prepend_block_to_lists(PaddedObjectMonitor* new_blk) { 278 // First we handle g_block_list: 279 while (true) { 280 PaddedObjectMonitor* cur = Atomic::load(&g_block_list); 281 // Prepend new_blk to g_block_list. The first ObjectMonitor in 282 // a block is reserved for use as linkage to the next block. 283 Atomic::store(&new_blk[0]._next_om, cur); 284 if (Atomic::cmpxchg(&g_block_list, cur, new_blk) == cur) { 285 // Successfully switched g_block_list to the new_blk value. 286 Atomic::add(&LVars.population, _BLOCKSIZE - 1); 287 break; 288 } 289 // Implied else: try it all again 290 } 291 292 // Second we handle LVars.free_list: 293 prepend_list_to_common(new_blk + 1, &new_blk[_BLOCKSIZE - 1], _BLOCKSIZE - 1, 294 &LVars.free_list, &LVars.free_count); 295 } 296 297 // Prepend a list of ObjectMonitors to LVars.free_list. 'tail' is the last 298 // ObjectMonitor in the list and there are 'count' on the list. Also 299 // updates LVars.free_count. 300 static void prepend_list_to_global_free_list(ObjectMonitor* list, 301 ObjectMonitor* tail, int count) { 302 prepend_list_to_common(list, tail, count, &LVars.free_list, &LVars.free_count); 303 } 304 305 // Prepend a list of ObjectMonitors to LVars.in_use_list. 'tail' is the last 306 // ObjectMonitor in the list and there are 'count' on the list. Also 307 // updates LVars.in_use_list. 308 static void prepend_list_to_global_in_use_list(ObjectMonitor* list, 309 ObjectMonitor* tail, int count) { 310 prepend_list_to_common(list, tail, count, &LVars.in_use_list, &LVars.in_use_count); 311 } 312 313 // Prepend an ObjectMonitor to the specified list. Also updates 314 // the specified counter. 315 static void prepend_to_common(ObjectMonitor* m, ObjectMonitor** list_p, 316 int* count_p) { 317 while (true) { 318 om_lock(m); // Lock m so we can safely update its next field. 319 ObjectMonitor* cur = NULL; 320 // Lock the list head to guard against races with a list walker 321 // thread: 322 if ((cur = get_list_head_locked(list_p)) != NULL) { 323 // List head is now locked so we can safely switch it. 324 set_next(m, cur); // m now points to cur (and unlocks m) 325 Atomic::store(list_p, m); // Switch list head to unlocked m. 326 om_unlock(cur); 327 break; 328 } 329 // The list is empty so try to set the list head. 330 assert(cur == NULL, "cur must be NULL: cur=" INTPTR_FORMAT, p2i(cur)); 331 set_next(m, cur); // m now points to NULL (and unlocks m) 332 if (Atomic::cmpxchg(list_p, cur, m) == cur) { 333 // List head is now unlocked m. 334 break; 335 } 336 // Implied else: try it all again 337 } 338 Atomic::inc(count_p); 339 } 340 341 // Prepend an ObjectMonitor to a per-thread om_free_list. 342 // Also updates the per-thread om_free_count. 343 static void prepend_to_om_free_list(Thread* self, ObjectMonitor* m) { 344 prepend_to_common(m, &self->om_free_list, &self->om_free_count); 345 } 346 347 // Prepend an ObjectMonitor to a per-thread om_in_use_list. 348 // Also updates the per-thread om_in_use_count. 349 static void prepend_to_om_in_use_list(Thread* self, ObjectMonitor* m) { 350 prepend_to_common(m, &self->om_in_use_list, &self->om_in_use_count); 351 } 352 353 // Take an ObjectMonitor from the start of the specified list. Also 354 // decrements the specified counter. Returns NULL if none are available. 355 static ObjectMonitor* take_from_start_of_common(ObjectMonitor** list_p, 356 int* count_p) { 357 ObjectMonitor* take = NULL; 358 // Lock the list head to guard against races with a list walker 359 // thread: 360 if ((take = get_list_head_locked(list_p)) == NULL) { 361 return NULL; // None are available. 362 } 363 ObjectMonitor* next = unmarked_next(take); 364 // Switch locked list head to next (which unlocks the list head, but 365 // leaves take locked): 366 Atomic::store(list_p, next); 367 Atomic::dec(count_p); 368 // Unlock take, but leave the next value for any lagging list 369 // walkers. It will get cleaned up when take is prepended to 370 // the in-use list: 371 om_unlock(take); 372 return take; 373 } 374 375 // Take an ObjectMonitor from the start of the LVars.free_list. Also 376 // updates LVars.free_count. Returns NULL if none are available. 377 static ObjectMonitor* take_from_start_of_global_free_list() { 378 return take_from_start_of_common(&LVars.free_list, &LVars.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 && Atomic::replace_if_null(&(m->_owner), self)) { 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[DEFAULT_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, DEFAULT_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, DEFAULT_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 with block linkage _next_om fields). 1156 block = (PaddedObjectMonitor*)Atomic::load(&block->_next_om); 1157 } 1158 } 1159 1160 static bool monitors_used_above_threshold() { 1161 if (Atomic::load(&LVars.population) == 0) { 1162 return false; 1163 } 1164 if (MonitorUsedDeflationThreshold > 0) { 1165 int monitors_used = Atomic::load(&LVars.population) - Atomic::load(&LVars.free_count); 1166 int monitor_usage = (monitors_used * 100LL) / Atomic::load(&LVars.population); 1167 return monitor_usage > MonitorUsedDeflationThreshold; 1168 } 1169 return false; 1170 } 1171 1172 // Returns true if MonitorBound is set (> 0) and if the specified 1173 // cnt is > MonitorBound. Otherwise returns false. 1174 static bool is_MonitorBound_exceeded(const int cnt) { 1175 const int mx = MonitorBound; 1176 return mx > 0 && cnt > mx; 1177 } 1178 1179 bool ObjectSynchronizer::is_cleanup_needed() { 1180 if (monitors_used_above_threshold()) { 1181 // Too many monitors in use. 1182 return true; 1183 } 1184 return needs_monitor_scavenge(); 1185 } 1186 1187 bool ObjectSynchronizer::needs_monitor_scavenge() { 1188 if (Atomic::load(&_forceMonitorScavenge) == 1) { 1189 log_info(monitorinflation)("Monitor scavenge needed, triggering safepoint cleanup."); 1190 return true; 1191 } 1192 return false; 1193 } 1194 1195 void ObjectSynchronizer::oops_do(OopClosure* f) { 1196 // We only scan the global used list here (for moribund threads), and 1197 // the thread-local monitors in Thread::oops_do(). 1198 global_used_oops_do(f); 1199 } 1200 1201 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) { 1202 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1203 list_oops_do(Atomic::load(&LVars.in_use_list), f); 1204 } 1205 1206 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) { 1207 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1208 list_oops_do(thread->om_in_use_list, f); 1209 } 1210 1211 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, OopClosure* f) { 1212 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1213 // The oops_do() phase does not overlap with monitor deflation 1214 // so no need to lock ObjectMonitors for the list traversal. 1215 for (ObjectMonitor* mid = list; mid != NULL; mid = unmarked_next(mid)) { 1216 if (mid->object() != NULL) { 1217 f->do_oop((oop*)mid->object_addr()); 1218 } 1219 } 1220 } 1221 1222 1223 // ----------------------------------------------------------------------------- 1224 // ObjectMonitor Lifecycle 1225 // ----------------------- 1226 // Inflation unlinks monitors from LVars.free_list or a per-thread free 1227 // list and associates them with objects. Deflation -- which occurs at 1228 // STW-time -- disassociates idle monitors from objects. 1229 // Such scavenged monitors are returned to the LVars.free_list. 1230 // 1231 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. 1232 // 1233 // Lifecycle: 1234 // -- unassigned and on the LVars.free_list 1235 // -- unassigned and on a per-thread free list 1236 // -- assigned to an object. The object is inflated and the mark refers 1237 // to the ObjectMonitor. 1238 1239 1240 // Constraining monitor pool growth via MonitorBound ... 1241 // 1242 // If MonitorBound is not set (<= 0), MonitorBound checks are disabled. 1243 // 1244 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the 1245 // the rate of scavenging is driven primarily by GC. As such, we can find 1246 // an inordinate number of monitors in circulation. 1247 // To avoid that scenario we can artificially induce a STW safepoint 1248 // if the pool appears to be growing past some reasonable bound. 1249 // Generally we favor time in space-time tradeoffs, but as there's no 1250 // natural back-pressure on the # of extant monitors we need to impose some 1251 // type of limit. Beware that if MonitorBound is set to too low a value 1252 // we could just loop. In addition, if MonitorBound is set to a low value 1253 // we'll incur more safepoints, which are harmful to performance. 1254 // See also: GuaranteedSafepointInterval 1255 // 1256 // If MonitorBound is set, the boundry applies to 1257 // (LVars.population - LVars.free_count) 1258 // i.e., if there are not enough ObjectMonitors on the global free list, 1259 // then a safepoint deflation is induced. Picking a good MonitorBound value 1260 // is non-trivial. 1261 1262 static void InduceScavenge(Thread* self, const char * Whence) { 1263 // Induce STW safepoint to trim monitors 1264 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. 1265 // More precisely, trigger a cleanup safepoint as the number 1266 // of active monitors passes the specified threshold. 1267 // TODO: assert thread state is reasonable 1268 1269 if (Atomic::xchg(&_forceMonitorScavenge, 1) == 0) { 1270 VMThread::check_for_forced_cleanup(); 1271 } 1272 } 1273 1274 ObjectMonitor* ObjectSynchronizer::om_alloc(Thread* self) { 1275 // A large MAXPRIVATE value reduces both list lock contention 1276 // and list coherency traffic, but also tends to increase the 1277 // number of ObjectMonitors in circulation as well as the STW 1278 // scavenge costs. As usual, we lean toward time in space-time 1279 // tradeoffs. 1280 const int MAXPRIVATE = 1024; 1281 1282 stringStream ss; 1283 for (;;) { 1284 ObjectMonitor* m; 1285 1286 // 1: try to allocate from the thread's local om_free_list. 1287 // Threads will attempt to allocate first from their local list, then 1288 // from the global list, and only after those attempts fail will the 1289 // thread attempt to instantiate new monitors. Thread-local free lists 1290 // improve allocation latency, as well as reducing coherency traffic 1291 // on the shared global list. 1292 m = take_from_start_of_om_free_list(self); 1293 if (m != NULL) { 1294 guarantee(m->object() == NULL, "invariant"); 1295 prepend_to_om_in_use_list(self, m); 1296 return m; 1297 } 1298 1299 // 2: try to allocate from the global LVars.free_list 1300 // CONSIDER: use muxTry() instead of muxAcquire(). 1301 // If the muxTry() fails then drop immediately into case 3. 1302 // If we're using thread-local free lists then try 1303 // to reprovision the caller's free list. 1304 if (Atomic::load(&LVars.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(&LVars.population) - Atomic::load(&LVars.free_count))) { 1321 // Not enough ObjectMonitors on the global free list. 1322 // We can't safely induce a STW safepoint from om_alloc() as our thread 1323 // state may not be appropriate for such activities and callers may hold 1324 // naked oops, so instead we defer the action. 1325 InduceScavenge(self, "om_alloc"); 1326 } 1327 continue; 1328 } 1329 1330 // 3: allocate a block of new ObjectMonitors 1331 // Both the local and global free lists are empty -- resort to malloc(). 1332 // In the current implementation ObjectMonitors are TSM - immortal. 1333 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want 1334 // each ObjectMonitor to start at the beginning of a cache line, 1335 // so we use align_up(). 1336 // A better solution would be to use C++ placement-new. 1337 // BEWARE: As it stands currently, we don't run the ctors! 1338 assert(_BLOCKSIZE > 1, "invariant"); 1339 size_t neededsize = sizeof(PaddedObjectMonitor) * _BLOCKSIZE; 1340 PaddedObjectMonitor* temp; 1341 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1); 1342 void* real_malloc_addr = NEW_C_HEAP_ARRAY(char, aligned_size, mtInternal); 1343 temp = (PaddedObjectMonitor*)align_up(real_malloc_addr, DEFAULT_CACHE_LINE_SIZE); 1344 (void)memset((void *) temp, 0, neededsize); 1345 1346 // Format the block. 1347 // initialize the linked list, each monitor points to its next 1348 // forming the single linked free list, the very first monitor 1349 // will points to next block, which forms the block list. 1350 // The trick of using the 1st element in the block as g_block_list 1351 // linkage should be reconsidered. A better implementation would 1352 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 1353 1354 for (int i = 1; i < _BLOCKSIZE; i++) { 1355 Atomic::store(&temp[i]._next_om, (ObjectMonitor*)&temp[i + 1]); 1356 } 1357 1358 // terminate the last monitor as the end of list 1359 Atomic::store(&temp[_BLOCKSIZE - 1]._next_om, (ObjectMonitor*)NULL); 1360 1361 // Element [0] is reserved for global list linkage 1362 temp[0].set_object(CHAINMARKER); 1363 1364 // Consider carving out this thread's current request from the 1365 // block in hand. This avoids some lock traffic and redundant 1366 // list activity. 1367 1368 prepend_block_to_lists(temp); 1369 } 1370 } 1371 1372 // Place "m" on the caller's private per-thread om_free_list. 1373 // In practice there's no need to clamp or limit the number of 1374 // monitors on a thread's om_free_list as the only non-allocation time 1375 // we'll call om_release() is to return a monitor to the free list after 1376 // a CAS attempt failed. This doesn't allow unbounded #s of monitors to 1377 // accumulate on a thread's free list. 1378 // 1379 // Key constraint: all ObjectMonitors on a thread's free list and the global 1380 // free list must have their object field set to null. This prevents the 1381 // scavenger -- deflate_monitor_list() -- from reclaiming them while we 1382 // are trying to release them. 1383 1384 void ObjectSynchronizer::om_release(Thread* self, ObjectMonitor* m, 1385 bool from_per_thread_alloc) { 1386 guarantee(m->header().value() == 0, "invariant"); 1387 guarantee(m->object() == NULL, "invariant"); 1388 stringStream ss; 1389 guarantee((m->is_busy() | m->_recursions) == 0, "freeing in-use monitor: " 1390 "%s, recursions=" INTX_FORMAT, m->is_busy_to_string(&ss), 1391 m->_recursions); 1392 // _next_om is used for both per-thread in-use and free lists so 1393 // we have to remove 'm' from the in-use list first (as needed). 1394 if (from_per_thread_alloc) { 1395 // Need to remove 'm' from om_in_use_list. 1396 ObjectMonitor* cur_mid_in_use = NULL; 1397 ObjectMonitor* mid = NULL; 1398 ObjectMonitor* next = NULL; 1399 bool extracted = false; 1400 1401 // We use the simpler lock-mid-as-we-go protocol to prevent races 1402 // with a list walker thread since there are no parallel list 1403 // deletions (deflations happen at a safepoint). 1404 if ((mid = get_list_head_locked(&self->om_in_use_list)) == NULL) { 1405 fatal("thread=" INTPTR_FORMAT " in-use list must not be empty.", p2i(self)); 1406 } 1407 next = unmarked_next(mid); 1408 while (true) { 1409 if (m == mid) { 1410 // We found 'm' on the per-thread in-use list so try to extract it. 1411 if (cur_mid_in_use == NULL) { 1412 // mid is the list head and it is locked. Switch the list head 1413 // to next which unlocks the list head, but leaves mid locked: 1414 Atomic::store(&self->om_in_use_list, next); 1415 } else { 1416 // mid is locked. Switch cur_mid_in_use's next field to next 1417 // which is safe because we have no parallel list deletions, 1418 // but we leave mid locked: 1419 set_next(cur_mid_in_use, next); 1420 } 1421 // At this point mid is disconnected from the in-use list so 1422 // its lock no longer has any effects on the in-use list. 1423 extracted = true; 1424 Atomic::dec(&self->om_in_use_count); 1425 // Unlock mid, but leave the next value for any lagging list 1426 // walkers. It will get cleaned up when mid is prepended to 1427 // the thread's free list: 1428 om_unlock(mid); 1429 break; 1430 } else { 1431 om_unlock(mid); 1432 cur_mid_in_use = mid; 1433 } 1434 // All the list management is done so move on to the next one: 1435 mid = next; 1436 if (mid == NULL) { 1437 // Reached end of the list and didn't find m so: 1438 fatal("must find m=" INTPTR_FORMAT "on om_in_use_list=" INTPTR_FORMAT, 1439 p2i(m), p2i(self->om_in_use_list)); 1440 } 1441 // Lock mid so we can possibly extract it: 1442 om_lock(mid); 1443 next = unmarked_next(mid); 1444 } 1445 } 1446 1447 prepend_to_om_free_list(self, m); 1448 } 1449 1450 // Return ObjectMonitors on a moribund thread's free and in-use 1451 // lists to the appropriate global lists. The ObjectMonitors on the 1452 // per-thread in-use list may still be in use by other threads. 1453 // 1454 // We currently call om_flush() from Threads::remove() before the 1455 // thread has been excised from the thread list and is no longer a 1456 // mutator. This means that om_flush() cannot run concurrently with 1457 // a safepoint and interleave with deflate_idle_monitors(). In 1458 // particular, this ensures that the thread's in-use monitors are 1459 // scanned by a GC safepoint, either via Thread::oops_do() (before 1460 // om_flush() is called) or via ObjectSynchronizer::oops_do() (after 1461 // om_flush() is called). 1462 1463 void ObjectSynchronizer::om_flush(Thread* self) { 1464 // Process the per-thread in-use list first to be consistent. 1465 int in_use_count = 0; 1466 ObjectMonitor* in_use_list = NULL; 1467 ObjectMonitor* in_use_tail = NULL; 1468 1469 // This function can race with a list walker thread so we lock the 1470 // list head to prevent confusion. 1471 if ((in_use_list = get_list_head_locked(&self->om_in_use_list)) != NULL) { 1472 // At this point, we have locked the in-use list head so a racing 1473 // thread cannot come in after us. However, a racing thread could 1474 // be ahead of us; we'll detect that and delay to let it finish. 1475 // 1476 // The thread is going away, however the ObjectMonitors on the 1477 // om_in_use_list may still be in-use by other threads. Link 1478 // them to in_use_tail, which will be linked into the global 1479 // in-use list (LVars.in_use_list) below. 1480 // 1481 // Account for the in-use list head before the loop since it is 1482 // already locked (by this thread): 1483 in_use_tail = in_use_list; 1484 in_use_count++; 1485 for (ObjectMonitor* cur_om = unmarked_next(in_use_list); cur_om != NULL; cur_om = unmarked_next(cur_om)) { 1486 if (is_locked(cur_om)) { 1487 // cur_om is locked so there must be a racing walker thread ahead 1488 // of us so we'll give it a chance to finish. 1489 while (is_locked(cur_om)) { 1490 os::naked_short_sleep(1); 1491 } 1492 } 1493 in_use_tail = cur_om; 1494 in_use_count++; 1495 } 1496 guarantee(in_use_tail != NULL, "invariant"); 1497 int l_om_in_use_count = Atomic::load(&self->om_in_use_count); 1498 assert(l_om_in_use_count == in_use_count, "in-use counts don't match: " 1499 "l_om_in_use_count=%d, in_use_count=%d", l_om_in_use_count, in_use_count); 1500 Atomic::store(&self->om_in_use_count, 0); 1501 // Clear the in-use list head (which also unlocks it): 1502 Atomic::store(&self->om_in_use_list, (ObjectMonitor*)NULL); 1503 om_unlock(in_use_list); 1504 } 1505 1506 int free_count = 0; 1507 ObjectMonitor* free_list = NULL; 1508 ObjectMonitor* free_tail = NULL; 1509 // This function can race with a list walker thread so we lock the 1510 // list head to prevent confusion. 1511 if ((free_list = get_list_head_locked(&self->om_free_list)) != NULL) { 1512 // At this point, we have locked the free list head so a racing 1513 // thread cannot come in after us. However, a racing thread could 1514 // be ahead of us; we'll detect that and delay to let it finish. 1515 // 1516 // The thread is going away. Set 'free_tail' to the last per-thread free 1517 // monitor which will be linked to LVars.free_list below. 1518 // 1519 // Account for the free list head before the loop since it is 1520 // already locked (by this thread): 1521 free_tail = free_list; 1522 free_count++; 1523 for (ObjectMonitor* s = unmarked_next(free_list); s != NULL; s = unmarked_next(s)) { 1524 if (is_locked(s)) { 1525 // s is locked so there must be a racing walker thread ahead 1526 // of us so we'll give it a chance to finish. 1527 while (is_locked(s)) { 1528 os::naked_short_sleep(1); 1529 } 1530 } 1531 free_tail = s; 1532 free_count++; 1533 guarantee(s->object() == NULL, "invariant"); 1534 stringStream ss; 1535 guarantee(!s->is_busy(), "must be !is_busy: %s", s->is_busy_to_string(&ss)); 1536 } 1537 guarantee(free_tail != NULL, "invariant"); 1538 int l_om_free_count = Atomic::load(&self->om_free_count); 1539 assert(l_om_free_count == free_count, "free counts don't match: " 1540 "l_om_free_count=%d, free_count=%d", l_om_free_count, free_count); 1541 Atomic::store(&self->om_free_count, 0); 1542 Atomic::store(&self->om_free_list, (ObjectMonitor*)NULL); 1543 om_unlock(free_list); 1544 } 1545 1546 if (free_tail != NULL) { 1547 prepend_list_to_global_free_list(free_list, free_tail, free_count); 1548 } 1549 1550 if (in_use_tail != NULL) { 1551 prepend_list_to_global_in_use_list(in_use_list, in_use_tail, in_use_count); 1552 } 1553 1554 LogStreamHandle(Debug, monitorinflation) lsh_debug; 1555 LogStreamHandle(Info, monitorinflation) lsh_info; 1556 LogStream* ls = NULL; 1557 if (log_is_enabled(Debug, monitorinflation)) { 1558 ls = &lsh_debug; 1559 } else if ((free_count != 0 || in_use_count != 0) && 1560 log_is_enabled(Info, monitorinflation)) { 1561 ls = &lsh_info; 1562 } 1563 if (ls != NULL) { 1564 ls->print_cr("om_flush: jt=" INTPTR_FORMAT ", free_count=%d" 1565 ", in_use_count=%d" ", om_free_provision=%d", 1566 p2i(self), free_count, in_use_count, self->om_free_provision); 1567 } 1568 } 1569 1570 static void post_monitor_inflate_event(EventJavaMonitorInflate* event, 1571 const oop obj, 1572 ObjectSynchronizer::InflateCause cause) { 1573 assert(event != NULL, "invariant"); 1574 assert(event->should_commit(), "invariant"); 1575 event->set_monitorClass(obj->klass()); 1576 event->set_address((uintptr_t)(void*)obj); 1577 event->set_cause((u1)cause); 1578 event->commit(); 1579 } 1580 1581 // Fast path code shared by multiple functions 1582 void ObjectSynchronizer::inflate_helper(oop obj) { 1583 markWord mark = obj->mark(); 1584 if (mark.has_monitor()) { 1585 assert(ObjectSynchronizer::verify_objmon_isinpool(mark.monitor()), "monitor is invalid"); 1586 assert(mark.monitor()->header().is_neutral(), "monitor must record a good object header"); 1587 return; 1588 } 1589 inflate(Thread::current(), obj, inflate_cause_vm_internal); 1590 } 1591 1592 ObjectMonitor* ObjectSynchronizer::inflate(Thread* self, 1593 oop object, const InflateCause cause) { 1594 // Inflate mutates the heap ... 1595 // Relaxing assertion for bug 6320749. 1596 assert(Universe::verify_in_progress() || 1597 !SafepointSynchronize::is_at_safepoint(), "invariant"); 1598 1599 EventJavaMonitorInflate event; 1600 1601 for (;;) { 1602 const markWord mark = object->mark(); 1603 assert(!mark.has_bias_pattern(), "invariant"); 1604 1605 // The mark can be in one of the following states: 1606 // * Inflated - just return 1607 // * Stack-locked - coerce it to inflated 1608 // * INFLATING - busy wait for conversion to complete 1609 // * Neutral - aggressively inflate the object. 1610 // * BIASED - Illegal. We should never see this 1611 1612 // CASE: inflated 1613 if (mark.has_monitor()) { 1614 ObjectMonitor* inf = mark.monitor(); 1615 markWord dmw = inf->header(); 1616 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1617 assert(inf->object() == object, "invariant"); 1618 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1619 return inf; 1620 } 1621 1622 // CASE: inflation in progress - inflating over a stack-lock. 1623 // Some other thread is converting from stack-locked to inflated. 1624 // Only that thread can complete inflation -- other threads must wait. 1625 // The INFLATING value is transient. 1626 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1627 // We could always eliminate polling by parking the thread on some auxiliary list. 1628 if (mark == markWord::INFLATING()) { 1629 read_stable_mark(object); 1630 continue; 1631 } 1632 1633 // CASE: stack-locked 1634 // Could be stack-locked either by this thread or by some other thread. 1635 // 1636 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1637 // to install INFLATING into the mark word. We originally installed INFLATING, 1638 // allocated the objectmonitor, and then finally STed the address of the 1639 // objectmonitor into the mark. This was correct, but artificially lengthened 1640 // the interval in which INFLATED appeared in the mark, thus increasing 1641 // the odds of inflation contention. 1642 // 1643 // We now use per-thread private objectmonitor free lists. 1644 // These list are reprovisioned from the global free list outside the 1645 // critical INFLATING...ST interval. A thread can transfer 1646 // multiple objectmonitors en-mass from the global free list to its local free list. 1647 // This reduces coherency traffic and lock contention on the global free list. 1648 // Using such local free lists, it doesn't matter if the om_alloc() call appears 1649 // before or after the CAS(INFLATING) operation. 1650 // See the comments in om_alloc(). 1651 1652 LogStreamHandle(Trace, monitorinflation) lsh; 1653 1654 if (mark.has_locker()) { 1655 ObjectMonitor* m = om_alloc(self); 1656 // Optimistically prepare the objectmonitor - anticipate successful CAS 1657 // We do this before the CAS in order to minimize the length of time 1658 // in which INFLATING appears in the mark. 1659 m->Recycle(); 1660 m->_Responsible = NULL; 1661 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class 1662 1663 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark); 1664 if (cmp != mark) { 1665 om_release(self, m, true); 1666 continue; // Interference -- just retry 1667 } 1668 1669 // We've successfully installed INFLATING (0) into the mark-word. 1670 // This is the only case where 0 will appear in a mark-word. 1671 // Only the singular thread that successfully swings the mark-word 1672 // to 0 can perform (or more precisely, complete) inflation. 1673 // 1674 // Why do we CAS a 0 into the mark-word instead of just CASing the 1675 // mark-word from the stack-locked value directly to the new inflated state? 1676 // Consider what happens when a thread unlocks a stack-locked object. 1677 // It attempts to use CAS to swing the displaced header value from the 1678 // on-stack BasicLock back into the object header. Recall also that the 1679 // header value (hash code, etc) can reside in (a) the object header, or 1680 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1681 // header in an ObjectMonitor. The inflate() routine must copy the header 1682 // value from the BasicLock on the owner's stack to the ObjectMonitor, all 1683 // the while preserving the hashCode stability invariants. If the owner 1684 // decides to release the lock while the value is 0, the unlock will fail 1685 // and control will eventually pass from slow_exit() to inflate. The owner 1686 // will then spin, waiting for the 0 value to disappear. Put another way, 1687 // the 0 causes the owner to stall if the owner happens to try to 1688 // drop the lock (restoring the header from the BasicLock to the object) 1689 // while inflation is in-progress. This protocol avoids races that might 1690 // would otherwise permit hashCode values to change or "flicker" for an object. 1691 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable. 1692 // 0 serves as a "BUSY" inflate-in-progress indicator. 1693 1694 1695 // fetch the displaced mark from the owner's stack. 1696 // The owner can't die or unwind past the lock while our INFLATING 1697 // object is in the mark. Furthermore the owner can't complete 1698 // an unlock on the object, either. 1699 markWord dmw = mark.displaced_mark_helper(); 1700 // Catch if the object's header is not neutral (not locked and 1701 // not marked is what we care about here). 1702 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1703 1704 // Setup monitor fields to proper values -- prepare the monitor 1705 m->set_header(dmw); 1706 1707 // Optimization: if the mark.locker stack address is associated 1708 // with this thread we could simply set m->_owner = self. 1709 // Note that a thread can inflate an object 1710 // that it has stack-locked -- as might happen in wait() -- directly 1711 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1712 m->set_owner(mark.locker()); 1713 m->set_object(object); 1714 // TODO-FIXME: assert BasicLock->dhw != 0. 1715 1716 // Must preserve store ordering. The monitor state must 1717 // be stable at the time of publishing the monitor address. 1718 guarantee(object->mark() == markWord::INFLATING(), "invariant"); 1719 object->release_set_mark(markWord::encode(m)); 1720 1721 // Hopefully the performance counters are allocated on distinct cache lines 1722 // to avoid false sharing on MP systems ... 1723 OM_PERFDATA_OP(Inflations, inc()); 1724 if (log_is_enabled(Trace, monitorinflation)) { 1725 ResourceMark rm(self); 1726 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark=" 1727 INTPTR_FORMAT ", type='%s'", p2i(object), 1728 object->mark().value(), object->klass()->external_name()); 1729 } 1730 if (event.should_commit()) { 1731 post_monitor_inflate_event(&event, object, cause); 1732 } 1733 return m; 1734 } 1735 1736 // CASE: neutral 1737 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 1738 // If we know we're inflating for entry it's better to inflate by swinging a 1739 // pre-locked ObjectMonitor pointer into the object header. A successful 1740 // CAS inflates the object *and* confers ownership to the inflating thread. 1741 // In the current implementation we use a 2-step mechanism where we CAS() 1742 // to inflate and then CAS() again to try to swing _owner from NULL to self. 1743 // An inflateTry() method that we could call from enter() would be useful. 1744 1745 // Catch if the object's header is not neutral (not locked and 1746 // not marked is what we care about here). 1747 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 1748 ObjectMonitor* m = om_alloc(self); 1749 // prepare m for installation - set monitor to initial state 1750 m->Recycle(); 1751 m->set_header(mark); 1752 m->set_object(object); 1753 m->_Responsible = NULL; 1754 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class 1755 1756 if (object->cas_set_mark(markWord::encode(m), mark) != mark) { 1757 m->set_header(markWord::zero()); 1758 m->set_object(NULL); 1759 m->Recycle(); 1760 om_release(self, m, true); 1761 m = NULL; 1762 continue; 1763 // interference - the markword changed - just retry. 1764 // The state-transitions are one-way, so there's no chance of 1765 // live-lock -- "Inflated" is an absorbing state. 1766 } 1767 1768 // Hopefully the performance counters are allocated on distinct 1769 // cache lines to avoid false sharing on MP systems ... 1770 OM_PERFDATA_OP(Inflations, inc()); 1771 if (log_is_enabled(Trace, monitorinflation)) { 1772 ResourceMark rm(self); 1773 lsh.print_cr("inflate(neutral): object=" INTPTR_FORMAT ", mark=" 1774 INTPTR_FORMAT ", type='%s'", p2i(object), 1775 object->mark().value(), object->klass()->external_name()); 1776 } 1777 if (event.should_commit()) { 1778 post_monitor_inflate_event(&event, object, cause); 1779 } 1780 return m; 1781 } 1782 } 1783 1784 1785 // We maintain a list of in-use monitors for each thread. 1786 // 1787 // deflate_thread_local_monitors() scans a single thread's in-use list, while 1788 // deflate_idle_monitors() scans only a global list of in-use monitors which 1789 // is populated only as a thread dies (see om_flush()). 1790 // 1791 // These operations are called at all safepoints, immediately after mutators 1792 // are stopped, but before any objects have moved. Collectively they traverse 1793 // the population of in-use monitors, deflating where possible. The scavenged 1794 // monitors are returned to the global monitor free list. 1795 // 1796 // Beware that we scavenge at *every* stop-the-world point. Having a large 1797 // number of monitors in-use could negatively impact performance. We also want 1798 // to minimize the total # of monitors in circulation, as they incur a small 1799 // footprint penalty. 1800 // 1801 // Perversely, the heap size -- and thus the STW safepoint rate -- 1802 // typically drives the scavenge rate. Large heaps can mean infrequent GC, 1803 // which in turn can mean large(r) numbers of ObjectMonitors in circulation. 1804 // This is an unfortunate aspect of this design. 1805 1806 // Deflate a single monitor if not in-use 1807 // Return true if deflated, false if in-use 1808 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, 1809 ObjectMonitor** free_head_p, 1810 ObjectMonitor** free_tail_p) { 1811 bool deflated; 1812 // Normal case ... The monitor is associated with obj. 1813 const markWord mark = obj->mark(); 1814 guarantee(mark == markWord::encode(mid), "should match: mark=" 1815 INTPTR_FORMAT ", encoded mid=" INTPTR_FORMAT, mark.value(), 1816 markWord::encode(mid).value()); 1817 // Make sure that mark.monitor() and markWord::encode() agree: 1818 guarantee(mark.monitor() == mid, "should match: monitor()=" INTPTR_FORMAT 1819 ", mid=" INTPTR_FORMAT, p2i(mark.monitor()), p2i(mid)); 1820 const markWord dmw = mid->header(); 1821 guarantee(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1822 1823 if (mid->is_busy()) { 1824 // Easy checks are first - the ObjectMonitor is busy so no deflation. 1825 deflated = false; 1826 } else { 1827 // Deflate the monitor if it is no longer being used 1828 // It's idle - scavenge and return to the global free list 1829 // plain old deflation ... 1830 if (log_is_enabled(Trace, monitorinflation)) { 1831 ResourceMark rm; 1832 log_trace(monitorinflation)("deflate_monitor: " 1833 "object=" INTPTR_FORMAT ", mark=" 1834 INTPTR_FORMAT ", type='%s'", p2i(obj), 1835 mark.value(), obj->klass()->external_name()); 1836 } 1837 1838 // Restore the header back to obj 1839 obj->release_set_mark(dmw); 1840 mid->clear(); 1841 1842 assert(mid->object() == NULL, "invariant: object=" INTPTR_FORMAT, 1843 p2i(mid->object())); 1844 1845 // Move the deflated ObjectMonitor to the working free list 1846 // defined by free_head_p and free_tail_p. The working list is 1847 // local so no need for a memory barrier. 1848 if (*free_head_p == NULL) *free_head_p = mid; 1849 if (*free_tail_p != NULL) { 1850 // We append to the list so the caller can use mid->_next_om 1851 // to fix the linkages in its context. 1852 ObjectMonitor* prevtail = *free_tail_p; 1853 // Should have been cleaned up by the caller: 1854 // Note: Should not have to lock prevtail here since we're at a 1855 // safepoint and ObjectMonitors on the local free list should 1856 // not be accessed in parallel. 1857 #ifdef ASSERT 1858 ObjectMonitor* l_next_om = Atomic::load(&prevtail->_next_om); 1859 #endif 1860 assert(l_next_om == NULL, "must be NULL: _next_om=" INTPTR_FORMAT, p2i(l_next_om)); 1861 set_next(prevtail, mid); 1862 } 1863 *free_tail_p = mid; 1864 // At this point, mid->_next_om still refers to its current 1865 // value and another ObjectMonitor's _next_om field still 1866 // refers to this ObjectMonitor. Those linkages have to be 1867 // cleaned up by the caller who has the complete context. 1868 deflated = true; 1869 } 1870 return deflated; 1871 } 1872 1873 // Walk a given monitor list, and deflate idle monitors. 1874 // The given list could be a per-thread list or a global list. 1875 // 1876 // In the case of parallel processing of thread local monitor lists, 1877 // work is done by Threads::parallel_threads_do() which ensures that 1878 // each Java thread is processed by exactly one worker thread, and 1879 // thus avoid conflicts that would arise when worker threads would 1880 // process the same monitor lists concurrently. 1881 // 1882 // See also ParallelSPCleanupTask and 1883 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and 1884 // Threads::parallel_java_threads_do() in thread.cpp. 1885 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** list_p, 1886 int* count_p, 1887 ObjectMonitor** free_head_p, 1888 ObjectMonitor** free_tail_p) { 1889 ObjectMonitor* cur_mid_in_use = NULL; 1890 ObjectMonitor* mid = NULL; 1891 ObjectMonitor* next = NULL; 1892 int deflated_count = 0; 1893 1894 // We use the simpler lock-mid-as-we-go protocol to prevent races 1895 // with a list walker thread since this caller is the only one doing 1896 // deletions on this list during the safepoint. 1897 if ((mid = get_list_head_locked(list_p)) == NULL) { 1898 return 0; // The list is empty so nothing to deflate. 1899 } 1900 next = unmarked_next(mid); 1901 1902 while (true) { 1903 oop obj = (oop) mid->object(); 1904 if (obj != NULL && deflate_monitor(mid, obj, free_head_p, free_tail_p)) { 1905 // Deflation succeeded and already updated free_head_p and 1906 // free_tail_p as needed. Finish the move to the local free list 1907 // by unlinking mid from the global or per-thread in-use list. 1908 if (cur_mid_in_use == NULL) { 1909 // mid is the list head and it is locked. Switch the list head 1910 // to next which unlocks the list head, but leaves mid locked: 1911 Atomic::store(list_p, next); 1912 } else { 1913 // mid is locked. Switch cur_mid_in_use's next field to next 1914 // which is safe because we have no parallel list deletions, 1915 // but we leave mid locked: 1916 set_next(cur_mid_in_use, next); 1917 } 1918 // At this point mid is disconnected from the in-use list so 1919 // its lock no longer has any effects on the in-use list. 1920 deflated_count++; 1921 Atomic::dec(count_p); 1922 // mid is current tail in the free_head_p list so NULL terminate it 1923 // (which also unlocks it): 1924 set_next(mid, NULL); 1925 } else { 1926 om_unlock(mid); 1927 cur_mid_in_use = mid; 1928 } 1929 // All the list management is done so move on to the next one: 1930 mid = next; 1931 if (mid == NULL) { 1932 break; // Reached end of the list so nothing more to deflate. 1933 } 1934 // Lock mid so we can possibly deflate it: 1935 om_lock(mid); 1936 next = unmarked_next(mid); 1937 } 1938 return deflated_count; 1939 } 1940 1941 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) { 1942 counters->n_in_use = 0; // currently associated with objects 1943 counters->n_in_circulation = 0; // extant 1944 counters->n_scavenged = 0; // reclaimed (global and per-thread) 1945 counters->per_thread_scavenged = 0; // per-thread scavenge total 1946 counters->per_thread_times = 0.0; // per-thread scavenge times 1947 OrderAccess::storestore(); // flush inits for worker threads 1948 } 1949 1950 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) { 1951 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1952 bool deflated = false; 1953 1954 ObjectMonitor* free_head_p = NULL; // Local SLL of scavenged monitors 1955 ObjectMonitor* free_tail_p = NULL; 1956 elapsedTimer timer; 1957 1958 if (log_is_enabled(Info, monitorinflation)) { 1959 timer.start(); 1960 } 1961 1962 // Note: the thread-local monitors lists get deflated in 1963 // a separate pass. See deflate_thread_local_monitors(). 1964 1965 // For moribund threads, scan LVars.in_use_list 1966 int deflated_count = 0; 1967 if (Atomic::load(&LVars.in_use_list) != NULL) { 1968 // Update n_in_circulation before LVars.in_use_count is updated by deflation. 1969 Atomic::add(&counters->n_in_circulation, Atomic::load(&LVars.in_use_count)); 1970 1971 deflated_count = deflate_monitor_list(&LVars.in_use_list, &LVars.in_use_count, &free_head_p, &free_tail_p); 1972 Atomic::add(&counters->n_in_use, Atomic::load(&LVars.in_use_count)); 1973 } 1974 1975 if (free_head_p != NULL) { 1976 // Move the deflated ObjectMonitors back to the global free list. 1977 // The working list is local so no need for a memory barrier. 1978 guarantee(free_tail_p != NULL && deflated_count > 0, "invariant"); 1979 #ifdef ASSERT 1980 ObjectMonitor* l_next_om = Atomic::load(&free_tail_p->_next_om); 1981 #endif 1982 assert(l_next_om == NULL, "must be NULL: _next_om=" INTPTR_FORMAT, p2i(l_next_om)); 1983 prepend_list_to_global_free_list(free_head_p, free_tail_p, deflated_count); 1984 Atomic::add(&counters->n_scavenged, deflated_count); 1985 } 1986 timer.stop(); 1987 1988 LogStreamHandle(Debug, monitorinflation) lsh_debug; 1989 LogStreamHandle(Info, monitorinflation) lsh_info; 1990 LogStream* ls = NULL; 1991 if (log_is_enabled(Debug, monitorinflation)) { 1992 ls = &lsh_debug; 1993 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) { 1994 ls = &lsh_info; 1995 } 1996 if (ls != NULL) { 1997 ls->print_cr("deflating global idle monitors, %3.7f secs, %d monitors", timer.seconds(), deflated_count); 1998 } 1999 } 2000 2001 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) { 2002 // Report the cumulative time for deflating each thread's idle 2003 // monitors. Note: if the work is split among more than one 2004 // worker thread, then the reported time will likely be more 2005 // than a beginning to end measurement of the phase. 2006 log_info(safepoint, cleanup)("deflating per-thread idle monitors, %3.7f secs, monitors=%d", counters->per_thread_times, counters->per_thread_scavenged); 2007 2008 if (log_is_enabled(Debug, monitorinflation)) { 2009 // exit_globals()'s call to audit_and_print_stats() is done 2010 // at the Info level and not at a safepoint. 2011 ObjectSynchronizer::audit_and_print_stats(false /* on_exit */); 2012 } else if (log_is_enabled(Info, monitorinflation)) { 2013 log_info(monitorinflation)("global_population=%d, global_in_use_count=%d, " 2014 "global_free_count=%d", Atomic::load(&LVars.population), 2015 Atomic::load(&LVars.in_use_count), Atomic::load(&LVars.free_count)); 2016 } 2017 2018 Atomic::store(&_forceMonitorScavenge, 0); // Reset 2019 2020 OM_PERFDATA_OP(Deflations, inc(counters->n_scavenged)); 2021 OM_PERFDATA_OP(MonExtant, set_value(counters->n_in_circulation)); 2022 2023 GVars.stw_random = os::random(); 2024 GVars.stw_cycle++; 2025 } 2026 2027 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) { 2028 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 2029 2030 ObjectMonitor* free_head_p = NULL; // Local SLL of scavenged monitors 2031 ObjectMonitor* free_tail_p = NULL; 2032 elapsedTimer timer; 2033 2034 if (log_is_enabled(Info, safepoint, cleanup) || 2035 log_is_enabled(Info, monitorinflation)) { 2036 timer.start(); 2037 } 2038 2039 // Update n_in_circulation before om_in_use_count is updated by deflation. 2040 Atomic::add(&counters->n_in_circulation, Atomic::load(&thread->om_in_use_count)); 2041 2042 int deflated_count = deflate_monitor_list(&thread->om_in_use_list, &thread->om_in_use_count, &free_head_p, &free_tail_p); 2043 Atomic::add(&counters->n_in_use, Atomic::load(&thread->om_in_use_count)); 2044 2045 if (free_head_p != NULL) { 2046 // Move the deflated ObjectMonitors back to the global free list. 2047 // The working list is local so no need for a memory barrier. 2048 guarantee(free_tail_p != NULL && deflated_count > 0, "invariant"); 2049 #ifdef ASSERT 2050 ObjectMonitor* l_next_om = Atomic::load(&free_tail_p->_next_om); 2051 #endif 2052 assert(l_next_om == NULL, "must be NULL: _next_om=" INTPTR_FORMAT, p2i(l_next_om)); 2053 prepend_list_to_global_free_list(free_head_p, free_tail_p, deflated_count); 2054 Atomic::add(&counters->n_scavenged, deflated_count); 2055 Atomic::add(&counters->per_thread_scavenged, deflated_count); 2056 } 2057 2058 timer.stop(); 2059 // Safepoint logging cares about cumulative per_thread_times and 2060 // we'll capture most of the cost, but not the muxRelease() which 2061 // should be cheap. 2062 counters->per_thread_times += timer.seconds(); 2063 2064 LogStreamHandle(Debug, monitorinflation) lsh_debug; 2065 LogStreamHandle(Info, monitorinflation) lsh_info; 2066 LogStream* ls = NULL; 2067 if (log_is_enabled(Debug, monitorinflation)) { 2068 ls = &lsh_debug; 2069 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) { 2070 ls = &lsh_info; 2071 } 2072 if (ls != NULL) { 2073 ls->print_cr("jt=" INTPTR_FORMAT ": deflating per-thread idle monitors, %3.7f secs, %d monitors", p2i(thread), timer.seconds(), deflated_count); 2074 } 2075 } 2076 2077 // Monitor cleanup on JavaThread::exit 2078 2079 // Iterate through monitor cache and attempt to release thread's monitors 2080 // Gives up on a particular monitor if an exception occurs, but continues 2081 // the overall iteration, swallowing the exception. 2082 class ReleaseJavaMonitorsClosure: public MonitorClosure { 2083 private: 2084 TRAPS; 2085 2086 public: 2087 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 2088 void do_monitor(ObjectMonitor* mid) { 2089 if (mid->owner() == THREAD) { 2090 (void)mid->complete_exit(CHECK); 2091 } 2092 } 2093 }; 2094 2095 // Release all inflated monitors owned by THREAD. Lightweight monitors are 2096 // ignored. This is meant to be called during JNI thread detach which assumes 2097 // all remaining monitors are heavyweight. All exceptions are swallowed. 2098 // Scanning the extant monitor list can be time consuming. 2099 // A simple optimization is to add a per-thread flag that indicates a thread 2100 // called jni_monitorenter() during its lifetime. 2101 // 2102 // Instead of No_Savepoint_Verifier it might be cheaper to 2103 // use an idiom of the form: 2104 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 2105 // <code that must not run at safepoint> 2106 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 2107 // Since the tests are extremely cheap we could leave them enabled 2108 // for normal product builds. 2109 2110 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 2111 assert(THREAD == JavaThread::current(), "must be current Java thread"); 2112 NoSafepointVerifier nsv; 2113 ReleaseJavaMonitorsClosure rjmc(THREAD); 2114 ObjectSynchronizer::monitors_iterate(&rjmc); 2115 THREAD->clear_pending_exception(); 2116 } 2117 2118 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 2119 switch (cause) { 2120 case inflate_cause_vm_internal: return "VM Internal"; 2121 case inflate_cause_monitor_enter: return "Monitor Enter"; 2122 case inflate_cause_wait: return "Monitor Wait"; 2123 case inflate_cause_notify: return "Monitor Notify"; 2124 case inflate_cause_hash_code: return "Monitor Hash Code"; 2125 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 2126 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 2127 default: 2128 ShouldNotReachHere(); 2129 } 2130 return "Unknown"; 2131 } 2132 2133 //------------------------------------------------------------------------------ 2134 // Debugging code 2135 2136 u_char* ObjectSynchronizer::get_gvars_addr() { 2137 return (u_char*)&GVars; 2138 } 2139 2140 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() { 2141 return (u_char*)&GVars.hc_sequence; 2142 } 2143 2144 size_t ObjectSynchronizer::get_gvars_size() { 2145 return sizeof(SharedGlobals); 2146 } 2147 2148 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() { 2149 return (u_char*)&GVars.stw_random; 2150 } 2151 2152 // This function can be called at a safepoint or it can be called when 2153 // we are trying to exit the VM. When we are trying to exit the VM, the 2154 // list walker functions can run in parallel with the other list 2155 // operations so spin-locking is used for safety. 2156 // 2157 // Calls to this function can be added in various places as a debugging 2158 // aid; pass 'true' for the 'on_exit' parameter to have in-use monitor 2159 // details logged at the Info level and 'false' for the 'on_exit' 2160 // parameter to have in-use monitor details logged at the Trace level. 2161 // 2162 void ObjectSynchronizer::audit_and_print_stats(bool on_exit) { 2163 assert(on_exit || SafepointSynchronize::is_at_safepoint(), "invariant"); 2164 2165 LogStreamHandle(Debug, monitorinflation) lsh_debug; 2166 LogStreamHandle(Info, monitorinflation) lsh_info; 2167 LogStreamHandle(Trace, monitorinflation) lsh_trace; 2168 LogStream* ls = NULL; 2169 if (log_is_enabled(Trace, monitorinflation)) { 2170 ls = &lsh_trace; 2171 } else if (log_is_enabled(Debug, monitorinflation)) { 2172 ls = &lsh_debug; 2173 } else if (log_is_enabled(Info, monitorinflation)) { 2174 ls = &lsh_info; 2175 } 2176 assert(ls != NULL, "sanity check"); 2177 2178 // Log counts for the global and per-thread monitor lists: 2179 int chk_om_population = log_monitor_list_counts(ls); 2180 int error_cnt = 0; 2181 2182 ls->print_cr("Checking global lists:"); 2183 2184 // Check LVars.population: 2185 if (Atomic::load(&LVars.population) == chk_om_population) { 2186 ls->print_cr("global_population=%d equals chk_om_population=%d", 2187 Atomic::load(&LVars.population), chk_om_population); 2188 } else { 2189 // With fine grained locks on the monitor lists, it is possible for 2190 // log_monitor_list_counts() to return a value that doesn't match 2191 // LVars.population. So far a higher value has been seen in testing 2192 // so something is being double counted by log_monitor_list_counts(). 2193 ls->print_cr("WARNING: global_population=%d is not equal to " 2194 "chk_om_population=%d", Atomic::load(&LVars.population), chk_om_population); 2195 } 2196 2197 // Check LVars.in_use_list and LVars.in_use_count: 2198 chk_global_in_use_list_and_count(ls, &error_cnt); 2199 2200 // Check LVars.free_list and LVars.free_count: 2201 chk_global_free_list_and_count(ls, &error_cnt); 2202 2203 ls->print_cr("Checking per-thread lists:"); 2204 2205 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 2206 // Check om_in_use_list and om_in_use_count: 2207 chk_per_thread_in_use_list_and_count(jt, ls, &error_cnt); 2208 2209 // Check om_free_list and om_free_count: 2210 chk_per_thread_free_list_and_count(jt, ls, &error_cnt); 2211 } 2212 2213 if (error_cnt == 0) { 2214 ls->print_cr("No errors found in monitor list checks."); 2215 } else { 2216 log_error(monitorinflation)("found monitor list errors: error_cnt=%d", error_cnt); 2217 } 2218 2219 if ((on_exit && log_is_enabled(Info, monitorinflation)) || 2220 (!on_exit && log_is_enabled(Trace, monitorinflation))) { 2221 // When exiting this log output is at the Info level. When called 2222 // at a safepoint, this log output is at the Trace level since 2223 // there can be a lot of it. 2224 log_in_use_monitor_details(ls); 2225 } 2226 2227 ls->flush(); 2228 2229 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt); 2230 } 2231 2232 // Check a free monitor entry; log any errors. 2233 void ObjectSynchronizer::chk_free_entry(JavaThread* jt, ObjectMonitor* n, 2234 outputStream * out, int *error_cnt_p) { 2235 stringStream ss; 2236 if (n->is_busy()) { 2237 if (jt != NULL) { 2238 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2239 ": free per-thread monitor must not be busy: %s", p2i(jt), 2240 p2i(n), n->is_busy_to_string(&ss)); 2241 } else { 2242 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor " 2243 "must not be busy: %s", p2i(n), n->is_busy_to_string(&ss)); 2244 } 2245 *error_cnt_p = *error_cnt_p + 1; 2246 } 2247 if (n->header().value() != 0) { 2248 if (jt != NULL) { 2249 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2250 ": free per-thread monitor must have NULL _header " 2251 "field: _header=" INTPTR_FORMAT, p2i(jt), p2i(n), 2252 n->header().value()); 2253 } else { 2254 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor " 2255 "must have NULL _header field: _header=" INTPTR_FORMAT, 2256 p2i(n), n->header().value()); 2257 } 2258 *error_cnt_p = *error_cnt_p + 1; 2259 } 2260 if (n->object() != NULL) { 2261 if (jt != NULL) { 2262 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2263 ": free per-thread monitor must have NULL _object " 2264 "field: _object=" INTPTR_FORMAT, p2i(jt), p2i(n), 2265 p2i(n->object())); 2266 } else { 2267 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor " 2268 "must have NULL _object field: _object=" INTPTR_FORMAT, 2269 p2i(n), p2i(n->object())); 2270 } 2271 *error_cnt_p = *error_cnt_p + 1; 2272 } 2273 } 2274 2275 // Lock the next ObjectMonitor for traversal and unlock the current 2276 // ObjectMonitor. Returns the next ObjectMonitor if there is one. 2277 // Otherwise returns NULL (after unlocking the current ObjectMonitor). 2278 // This function is used by the various list walker functions to 2279 // safely walk a list without allowing an ObjectMonitor to be moved 2280 // to another list in the middle of a walk. 2281 static ObjectMonitor* lock_next_for_traversal(ObjectMonitor* cur) { 2282 assert(is_locked(cur), "cur=" INTPTR_FORMAT " must be locked", p2i(cur)); 2283 ObjectMonitor* next = unmarked_next(cur); 2284 if (next == NULL) { // Reached the end of the list. 2285 om_unlock(cur); 2286 return NULL; 2287 } 2288 om_lock(next); // Lock next before unlocking current to keep 2289 om_unlock(cur); // from being by-passed by another thread. 2290 return next; 2291 } 2292 2293 // Check the global free list and count; log the results of the checks. 2294 void ObjectSynchronizer::chk_global_free_list_and_count(outputStream * out, 2295 int *error_cnt_p) { 2296 int chk_om_free_count = 0; 2297 ObjectMonitor* cur = NULL; 2298 if ((cur = get_list_head_locked(&LVars.free_list)) != NULL) { 2299 // Marked the global free list head so process the list. 2300 while (true) { 2301 chk_free_entry(NULL /* jt */, cur, out, error_cnt_p); 2302 chk_om_free_count++; 2303 2304 cur = lock_next_for_traversal(cur); 2305 if (cur == NULL) { 2306 break; 2307 } 2308 } 2309 } 2310 int l_free_count = Atomic::load(&LVars.free_count); 2311 if (l_free_count == chk_om_free_count) { 2312 out->print_cr("global_free_count=%d equals chk_om_free_count=%d", 2313 l_free_count, chk_om_free_count); 2314 } else { 2315 // With fine grained locks on LVars.free_list, it is possible for an 2316 // ObjectMonitor to be prepended to LVars.free_list after we started 2317 // calculating chk_om_free_count so LVars.free_count may not 2318 // match anymore. 2319 out->print_cr("WARNING: global_free_count=%d is not equal to " 2320 "chk_om_free_count=%d", l_free_count, chk_om_free_count); 2321 } 2322 } 2323 2324 // Check the global in-use list and count; log the results of the checks. 2325 void ObjectSynchronizer::chk_global_in_use_list_and_count(outputStream * out, 2326 int *error_cnt_p) { 2327 int chk_om_in_use_count = 0; 2328 ObjectMonitor* cur = NULL; 2329 if ((cur = get_list_head_locked(&LVars.in_use_list)) != NULL) { 2330 // Marked the global in-use list head so process the list. 2331 while (true) { 2332 chk_in_use_entry(NULL /* jt */, cur, out, error_cnt_p); 2333 chk_om_in_use_count++; 2334 2335 cur = lock_next_for_traversal(cur); 2336 if (cur == NULL) { 2337 break; 2338 } 2339 } 2340 } 2341 int l_in_use_count = Atomic::load(&LVars.in_use_count); 2342 if (l_in_use_count == chk_om_in_use_count) { 2343 out->print_cr("global_in_use_count=%d equals chk_om_in_use_count=%d", 2344 l_in_use_count, chk_om_in_use_count); 2345 } else { 2346 // With fine grained locks on the monitor lists, it is possible for 2347 // an exiting JavaThread to put its in-use ObjectMonitors on the 2348 // global in-use list after chk_om_in_use_count is calculated above. 2349 out->print_cr("WARNING: global_in_use_count=%d is not equal to chk_om_in_use_count=%d", 2350 l_in_use_count, chk_om_in_use_count); 2351 } 2352 } 2353 2354 // Check an in-use monitor entry; log any errors. 2355 void ObjectSynchronizer::chk_in_use_entry(JavaThread* jt, ObjectMonitor* n, 2356 outputStream * out, int *error_cnt_p) { 2357 if (n->header().value() == 0) { 2358 if (jt != NULL) { 2359 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2360 ": in-use per-thread monitor must have non-NULL _header " 2361 "field.", p2i(jt), p2i(n)); 2362 } else { 2363 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use global monitor " 2364 "must have non-NULL _header field.", p2i(n)); 2365 } 2366 *error_cnt_p = *error_cnt_p + 1; 2367 } 2368 if (n->object() == NULL) { 2369 if (jt != NULL) { 2370 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2371 ": in-use per-thread monitor must have non-NULL _object " 2372 "field.", p2i(jt), p2i(n)); 2373 } else { 2374 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use global monitor " 2375 "must have non-NULL _object field.", p2i(n)); 2376 } 2377 *error_cnt_p = *error_cnt_p + 1; 2378 } 2379 const oop obj = (oop)n->object(); 2380 const markWord mark = obj->mark(); 2381 if (!mark.has_monitor()) { 2382 if (jt != NULL) { 2383 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2384 ": in-use per-thread monitor's object does not think " 2385 "it has a monitor: obj=" INTPTR_FORMAT ", mark=" 2386 INTPTR_FORMAT, p2i(jt), p2i(n), p2i(obj), mark.value()); 2387 } else { 2388 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use global " 2389 "monitor's object does not think it has a monitor: obj=" 2390 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n), 2391 p2i(obj), mark.value()); 2392 } 2393 *error_cnt_p = *error_cnt_p + 1; 2394 } 2395 ObjectMonitor* const obj_mon = mark.monitor(); 2396 if (n != obj_mon) { 2397 if (jt != NULL) { 2398 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2399 ": in-use per-thread monitor's object does not refer " 2400 "to the same monitor: obj=" INTPTR_FORMAT ", mark=" 2401 INTPTR_FORMAT ", obj_mon=" INTPTR_FORMAT, p2i(jt), 2402 p2i(n), p2i(obj), mark.value(), p2i(obj_mon)); 2403 } else { 2404 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use global " 2405 "monitor's object does not refer to the same monitor: obj=" 2406 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon=" 2407 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon)); 2408 } 2409 *error_cnt_p = *error_cnt_p + 1; 2410 } 2411 } 2412 2413 // Check the thread's free list and count; log the results of the checks. 2414 void ObjectSynchronizer::chk_per_thread_free_list_and_count(JavaThread *jt, 2415 outputStream * out, 2416 int *error_cnt_p) { 2417 int chk_om_free_count = 0; 2418 ObjectMonitor* cur = NULL; 2419 if ((cur = get_list_head_locked(&jt->om_free_list)) != NULL) { 2420 // Marked the per-thread free list head so process the list. 2421 while (true) { 2422 chk_free_entry(jt, cur, out, error_cnt_p); 2423 chk_om_free_count++; 2424 2425 cur = lock_next_for_traversal(cur); 2426 if (cur == NULL) { 2427 break; 2428 } 2429 } 2430 } 2431 int l_om_free_count = Atomic::load(&jt->om_free_count); 2432 if (l_om_free_count == chk_om_free_count) { 2433 out->print_cr("jt=" INTPTR_FORMAT ": om_free_count=%d equals " 2434 "chk_om_free_count=%d", p2i(jt), l_om_free_count, chk_om_free_count); 2435 } else { 2436 out->print_cr("ERROR: jt=" INTPTR_FORMAT ": om_free_count=%d is not " 2437 "equal to chk_om_free_count=%d", p2i(jt), l_om_free_count, 2438 chk_om_free_count); 2439 *error_cnt_p = *error_cnt_p + 1; 2440 } 2441 } 2442 2443 // Check the thread's in-use list and count; log the results of the checks. 2444 void ObjectSynchronizer::chk_per_thread_in_use_list_and_count(JavaThread *jt, 2445 outputStream * out, 2446 int *error_cnt_p) { 2447 int chk_om_in_use_count = 0; 2448 ObjectMonitor* cur = NULL; 2449 if ((cur = get_list_head_locked(&jt->om_in_use_list)) != NULL) { 2450 // Marked the per-thread in-use list head so process the list. 2451 while (true) { 2452 chk_in_use_entry(jt, cur, out, error_cnt_p); 2453 chk_om_in_use_count++; 2454 2455 cur = lock_next_for_traversal(cur); 2456 if (cur == NULL) { 2457 break; 2458 } 2459 } 2460 } 2461 int l_om_in_use_count = Atomic::load(&jt->om_in_use_count); 2462 if (l_om_in_use_count == chk_om_in_use_count) { 2463 out->print_cr("jt=" INTPTR_FORMAT ": om_in_use_count=%d equals " 2464 "chk_om_in_use_count=%d", p2i(jt), l_om_in_use_count, 2465 chk_om_in_use_count); 2466 } else { 2467 out->print_cr("ERROR: jt=" INTPTR_FORMAT ": om_in_use_count=%d is not " 2468 "equal to chk_om_in_use_count=%d", p2i(jt), l_om_in_use_count, 2469 chk_om_in_use_count); 2470 *error_cnt_p = *error_cnt_p + 1; 2471 } 2472 } 2473 2474 // Log details about ObjectMonitors on the in-use lists. The 'BHL' 2475 // flags indicate why the entry is in-use, 'object' and 'object type' 2476 // indicate the associated object and its type. 2477 void ObjectSynchronizer::log_in_use_monitor_details(outputStream * out) { 2478 stringStream ss; 2479 if (Atomic::load(&LVars.in_use_count) > 0) { 2480 out->print_cr("In-use global monitor info:"); 2481 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)"); 2482 out->print_cr("%18s %s %18s %18s", 2483 "monitor", "BHL", "object", "object type"); 2484 out->print_cr("================== === ================== =================="); 2485 ObjectMonitor* cur = NULL; 2486 if ((cur = get_list_head_locked(&LVars.in_use_list)) != NULL) { 2487 // Marked the global in-use list head so process the list. 2488 while (true) { 2489 const oop obj = (oop) cur->object(); 2490 const markWord mark = cur->header(); 2491 ResourceMark rm; 2492 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(cur), 2493 cur->is_busy() != 0, mark.hash() != 0, cur->owner() != NULL, 2494 p2i(obj), obj->klass()->external_name()); 2495 if (cur->is_busy() != 0) { 2496 out->print(" (%s)", cur->is_busy_to_string(&ss)); 2497 ss.reset(); 2498 } 2499 out->cr(); 2500 2501 cur = lock_next_for_traversal(cur); 2502 if (cur == NULL) { 2503 break; 2504 } 2505 } 2506 } 2507 } 2508 2509 out->print_cr("In-use per-thread monitor info:"); 2510 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)"); 2511 out->print_cr("%18s %18s %s %18s %18s", 2512 "jt", "monitor", "BHL", "object", "object type"); 2513 out->print_cr("================== ================== === ================== =================="); 2514 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 2515 ObjectMonitor* cur = NULL; 2516 if ((cur = get_list_head_locked(&jt->om_in_use_list)) != NULL) { 2517 // Marked the global in-use list head so process the list. 2518 while (true) { 2519 const oop obj = (oop) cur->object(); 2520 const markWord mark = cur->header(); 2521 ResourceMark rm; 2522 out->print(INTPTR_FORMAT " " INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT 2523 " %s", p2i(jt), p2i(cur), cur->is_busy() != 0, 2524 mark.hash() != 0, cur->owner() != NULL, p2i(obj), 2525 obj->klass()->external_name()); 2526 if (cur->is_busy() != 0) { 2527 out->print(" (%s)", cur->is_busy_to_string(&ss)); 2528 ss.reset(); 2529 } 2530 out->cr(); 2531 2532 cur = lock_next_for_traversal(cur); 2533 if (cur == NULL) { 2534 break; 2535 } 2536 } 2537 } 2538 } 2539 2540 out->flush(); 2541 } 2542 2543 // Log counts for the global and per-thread monitor lists and return 2544 // the population count. 2545 int ObjectSynchronizer::log_monitor_list_counts(outputStream * out) { 2546 int pop_count = 0; 2547 out->print_cr("%18s %10s %10s %10s", 2548 "Global Lists:", "InUse", "Free", "Total"); 2549 out->print_cr("================== ========== ========== =========="); 2550 int l_in_use_count = Atomic::load(&LVars.in_use_count); 2551 int l_free_count = Atomic::load(&LVars.free_count); 2552 out->print_cr("%18s %10d %10d %10d", "", l_in_use_count, 2553 l_free_count, Atomic::load(&LVars.population)); 2554 pop_count += l_in_use_count + l_free_count; 2555 2556 out->print_cr("%18s %10s %10s %10s", 2557 "Per-Thread Lists:", "InUse", "Free", "Provision"); 2558 out->print_cr("================== ========== ========== =========="); 2559 2560 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 2561 int l_om_in_use_count = Atomic::load(&jt->om_in_use_count); 2562 int l_om_free_count = Atomic::load(&jt->om_free_count); 2563 out->print_cr(INTPTR_FORMAT " %10d %10d %10d", p2i(jt), 2564 l_om_in_use_count, l_om_free_count, jt->om_free_provision); 2565 pop_count += l_om_in_use_count + l_om_free_count; 2566 } 2567 return pop_count; 2568 } 2569 2570 #ifndef PRODUCT 2571 2572 // Check if monitor belongs to the monitor cache 2573 // The list is grow-only so it's *relatively* safe to traverse 2574 // the list of extant blocks without taking a lock. 2575 2576 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 2577 PaddedObjectMonitor* block = Atomic::load(&g_block_list); 2578 while (block != NULL) { 2579 assert(block->object() == CHAINMARKER, "must be a block header"); 2580 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) { 2581 address mon = (address)monitor; 2582 address blk = (address)block; 2583 size_t diff = mon - blk; 2584 assert((diff % sizeof(PaddedObjectMonitor)) == 0, "must be aligned"); 2585 return 1; 2586 } 2587 // unmarked_next() is not needed with g_block_list (no locking 2588 // used with with block linkage _next_om fields). 2589 block = (PaddedObjectMonitor*)Atomic::load(&block->_next_om); 2590 } 2591 return 0; 2592 } 2593 2594 #endif