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* volatile ObjectSynchronizer::g_block_list = NULL; 121 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false; 122 bool volatile ObjectSynchronizer::_is_special_deflation_requested = false; 123 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0; 124 125 // Global ObjectMonitor free list. Newly allocated and deflated 126 // ObjectMonitors are prepended here. 127 static ObjectMonitor* volatile g_free_list = NULL; 128 // Global ObjectMonitor in-use list. When a JavaThread is exiting, 129 // ObjectMonitors on its per-thread in-use list are prepended here. 130 static ObjectMonitor* volatile g_om_in_use_list = NULL; 131 132 static volatile int g_om_free_count = 0; // # on g_free_list 133 static volatile int g_om_in_use_count = 0; // # on g_om_in_use_list 134 static volatile int g_om_population = 0; // # Extant -- in circulation 135 136 #define CHAINMARKER (cast_to_oop<intptr_t>(-1)) 137 138 139 // =====================> List Management functions 140 141 // Return true if the ObjectMonitor's next field is marked. 142 // Otherwise returns false. 143 static bool is_next_marked(ObjectMonitor* om) { 144 return ((intptr_t)OrderAccess::load_acquire(&om->_next_om) & 0x1) != 0; 145 } 146 147 // Mark an ObjectMonitor* and return it. Note: the om parameter 148 // may or may not have been marked originally. 149 static ObjectMonitor* mark_om_ptr(ObjectMonitor* om) { 150 return (ObjectMonitor*)((intptr_t)om | 0x1); 151 } 152 153 // Mark the next field in an ObjectMonitor. If marking was successful, 154 // then the unmarked next field is returned via parameter and true is 155 // returned. Otherwise false is returned. 156 static bool mark_next(ObjectMonitor* om, ObjectMonitor** next_p) { 157 // Get current next field without any marking value. 158 ObjectMonitor* next = (ObjectMonitor*) 159 ((intptr_t)OrderAccess::load_acquire(&om->_next_om) & ~0x1); 160 if (Atomic::cmpxchg(mark_om_ptr(next), &om->_next_om, next) != next) { 161 return false; // Could not mark the next field or it was already marked. 162 } 163 *next_p = next; 164 return true; 165 } 166 167 // Loop until we mark the next field in an ObjectMonitor. The unmarked 168 // next field is returned. 169 static ObjectMonitor* mark_next_loop(ObjectMonitor* om) { 170 ObjectMonitor* next; 171 while (true) { 172 if (mark_next(om, &next)) { 173 // Marked om's next field so return the unmarked value. 174 return next; 175 } 176 } 177 } 178 179 // Set the next field in an ObjectMonitor to the specified value. 180 // The caller of set_next() must be the same thread that marked the 181 // ObjectMonitor. 182 static void set_next(ObjectMonitor* om, ObjectMonitor* value) { 183 OrderAccess::release_store(&om->_next_om, value); 184 } 185 186 // Mark the next field in the list head ObjectMonitor. If marking was 187 // successful, then the mid and the unmarked next field are returned 188 // via parameter and true is returned. Otherwise false is returned. 189 static bool mark_list_head(ObjectMonitor* volatile * list_p, 190 ObjectMonitor** mid_p, ObjectMonitor** next_p) { 191 while (true) { 192 ObjectMonitor* mid = OrderAccess::load_acquire(list_p); 193 if (mid == NULL) { 194 return false; // The list is empty so nothing to mark. 195 } 196 if (mark_next(mid, next_p)) { 197 if (OrderAccess::load_acquire(list_p) != mid) { 198 // The list head changed so we have to retry. 199 set_next(mid, *next_p); // unmark mid 200 continue; 201 } 202 // We marked next field to guard against races. 203 *mid_p = mid; 204 return true; 205 } 206 } 207 } 208 209 // Return the unmarked next field in an ObjectMonitor. Note: the next 210 // field may or may not have been marked originally. 211 static ObjectMonitor* unmarked_next(ObjectMonitor* om) { 212 return (ObjectMonitor*)((intptr_t)OrderAccess::load_acquire(&om->_next_om) & ~0x1); 213 } 214 215 #if 0 216 // XXX - this is unused 217 // Unmark the next field in an ObjectMonitor. Requires that the next 218 // field be marked. 219 static void unmark_next(ObjectMonitor* om) { 220 ADIM_guarantee(is_next_marked(om), "next field must be marked: next=" INTPTR_FORMAT, p2i(om->_next_om)); 221 222 ObjectMonitor* next = unmarked_next(om); 223 set_next(om, next); 224 } 225 #endif 226 227 volatile int visit_counter = 42; 228 static void chk_for_list_loop(ObjectMonitor* list, int count) { 229 if (!CheckMonitorLists) { 230 return; 231 } 232 int l_visit_counter = Atomic::add(1, &visit_counter); 233 int l_count = 0; 234 ObjectMonitor* prev = NULL; 235 for (ObjectMonitor* mid = list; mid != NULL; mid = unmarked_next(mid)) { 236 if (mid->visit_marker == l_visit_counter) { 237 log_error(monitorinflation)("ERROR: prev=" INTPTR_FORMAT ", l_count=%d" 238 " refers to an ObjectMonitor that has" 239 " already been visited: mid=" INTPTR_FORMAT, 240 p2i(prev), l_count, p2i(mid)); 241 fatal("list=" INTPTR_FORMAT " of %d items has a loop.", p2i(list), count); 242 } 243 mid->visit_marker = l_visit_counter; 244 prev = mid; 245 if (++l_count > count + 1024 * 1024) { 246 fatal("list=" INTPTR_FORMAT " of %d items may have a loop; l_count=%d", 247 p2i(list), count, l_count); 248 } 249 } 250 } 251 252 static void chk_om_not_on_list(ObjectMonitor* om, ObjectMonitor* list, int count) { 253 if (!CheckMonitorLists) { 254 return; 255 } 256 guarantee(list != om, "ERROR: om=" INTPTR_FORMAT " must not be head of the " 257 "list=" INTPTR_FORMAT ", count=%d", p2i(om), p2i(list), count); 258 int l_count = 0; 259 for (ObjectMonitor* mid = list; mid != NULL; mid = unmarked_next(mid)) { 260 if (unmarked_next(mid) == om) { 261 log_error(monitorinflation)("ERROR: mid=" INTPTR_FORMAT ", l_count=%d" 262 " next_om refers to om=" INTPTR_FORMAT, 263 p2i(mid), l_count, p2i(om)); 264 fatal("list=" INTPTR_FORMAT " of %d items has bad next_om value.", 265 p2i(list), count); 266 } 267 if (++l_count > count + 1024 * 1024) { 268 fatal("list=" INTPTR_FORMAT " of %d items may have a loop; l_count=%d", 269 p2i(list), count, l_count); 270 } 271 } 272 } 273 274 static void chk_om_elems_not_on_list(ObjectMonitor* elems, int elems_count, 275 ObjectMonitor* list, int list_count) { 276 if (!CheckMonitorLists) { 277 return; 278 } 279 chk_for_list_loop(elems, elems_count); 280 for (ObjectMonitor* mid = elems; mid != NULL; mid = unmarked_next(mid)) { 281 chk_om_not_on_list(mid, list, list_count); 282 } 283 } 284 285 // Prepend a list of ObjectMonitors to the specified *list_p. 'tail' is 286 // the last ObjectMonitor in the list and there are 'count' on the list. 287 // Also updates the specified *count_p. 288 static void prepend_list_to_common(ObjectMonitor* list, ObjectMonitor* tail, 289 int count, ObjectMonitor* volatile* list_p, 290 volatile int* count_p) { 291 chk_for_list_loop(OrderAccess::load_acquire(list_p), 292 OrderAccess::load_acquire(count_p)); 293 chk_om_elems_not_on_list(list, count, OrderAccess::load_acquire(list_p), 294 OrderAccess::load_acquire(count_p)); 295 while (true) { 296 ObjectMonitor* cur = OrderAccess::load_acquire(list_p); 297 // Prepend list to *list_p. 298 ObjectMonitor* next = NULL; 299 if (!mark_next(tail, &next)) { 300 continue; // failed to mark next field so try it all again 301 } 302 set_next(tail, cur); // tail now points to cur (and unmarks tail) 303 if (cur == NULL) { 304 // No potential race with takers or other prependers since 305 // *list_p is empty. 306 if (Atomic::cmpxchg(list, list_p, cur) == cur) { 307 // Successfully switched *list_p to the list value. 308 Atomic::add(count, count_p); 309 break; 310 } 311 // Implied else: try it all again 312 } else { 313 // Try to mark next field to guard against races: 314 if (!mark_next(cur, &next)) { 315 continue; // failed to mark next field so try it all again 316 } 317 // We marked the next field so try to switch *list_p to the list value. 318 if (Atomic::cmpxchg(list, list_p, cur) != cur) { 319 // The list head has changed so unmark the next field and try again: 320 set_next(cur, next); 321 continue; 322 } 323 Atomic::add(count, count_p); 324 set_next(cur, next); // unmark next field 325 break; 326 } 327 } 328 } 329 330 // Prepend a newly allocated block of ObjectMonitors to g_block_list and 331 // g_free_list. Also updates g_om_population and g_om_free_count. 332 void ObjectSynchronizer::prepend_block_to_lists(PaddedObjectMonitor* new_blk) { 333 // First we handle g_block_list: 334 while (true) { 335 PaddedObjectMonitor* cur = OrderAccess::load_acquire(&g_block_list); 336 // Prepend new_blk to g_block_list. The first ObjectMonitor in 337 // a block is reserved for use as linkage to the next block. 338 OrderAccess::release_store(&new_blk[0]._next_om, cur); 339 if (Atomic::cmpxchg(new_blk, &g_block_list, cur) == cur) { 340 // Successfully switched g_block_list to the new_blk value. 341 Atomic::add(_BLOCKSIZE - 1, &g_om_population); 342 break; 343 } 344 // Implied else: try it all again 345 } 346 347 // Second we handle g_free_list: 348 prepend_list_to_common(new_blk + 1, &new_blk[_BLOCKSIZE - 1], _BLOCKSIZE - 1, 349 &g_free_list, &g_om_free_count); 350 } 351 352 // Prepend a list of ObjectMonitors to g_free_list. 'tail' is the last 353 // ObjectMonitor in the list and there are 'count' on the list. Also 354 // updates g_om_free_count. 355 static void prepend_list_to_g_free_list(ObjectMonitor* list, 356 ObjectMonitor* tail, int count) { 357 prepend_list_to_common(list, tail, count, &g_free_list, &g_om_free_count); 358 } 359 360 // Prepend a list of ObjectMonitors to g_om_in_use_list. 'tail' is the last 361 // ObjectMonitor in the list and there are 'count' on the list. Also 362 // updates g_om_in_use_list. 363 static void prepend_list_to_g_om_in_use_list(ObjectMonitor* list, 364 ObjectMonitor* tail, int count) { 365 prepend_list_to_common(list, tail, count, &g_om_in_use_list, &g_om_in_use_count); 366 } 367 368 // Prepend an ObjectMonitor to the specified list. Also updates 369 // the specified counter. 370 static void prepend_to_common(ObjectMonitor* m, ObjectMonitor* volatile * list_p, 371 int volatile * count_p) { 372 chk_for_list_loop(OrderAccess::load_acquire(list_p), 373 OrderAccess::load_acquire(count_p)); 374 chk_om_not_on_list(m, OrderAccess::load_acquire(list_p), 375 OrderAccess::load_acquire(count_p)); 376 377 while (true) { 378 ObjectMonitor* cur = OrderAccess::load_acquire(list_p); 379 // Prepend ObjectMonitor to *list_p. 380 ObjectMonitor* next = NULL; 381 if (!mark_next(m, &next)) { 382 continue; // failed to mark next field so try it all again 383 } 384 set_next(m, cur); // m now points to cur (and unmarks m) 385 if (cur == NULL) { 386 // No potential race with other prependers since *list_p is empty. 387 if (Atomic::cmpxchg(m, list_p, cur) == cur) { 388 // Successfully switched *list_p to 'm'. 389 Atomic::inc(count_p); 390 break; 391 } 392 // Implied else: try it all again 393 } else { 394 // Try to mark next field to guard against races: 395 if (!mark_next(cur, &next)) { 396 continue; // failed to mark next field so try it all again 397 } 398 // We marked the next field so try to switch *list_p to 'm'. 399 if (Atomic::cmpxchg(m, list_p, cur) != cur) { 400 // The list head has changed so unmark the next field and try again: 401 set_next(cur, next); 402 continue; 403 } 404 Atomic::inc(count_p); 405 set_next(cur, next); // unmark next field 406 break; 407 } 408 } 409 } 410 411 // Prepend an ObjectMonitor to a per-thread om_free_list. 412 // Also updates the per-thread om_free_count. 413 static void prepend_to_om_free_list(Thread* self, ObjectMonitor* m) { 414 prepend_to_common(m, &self->om_free_list, &self->om_free_count); 415 } 416 417 // Prepend an ObjectMonitor to a per-thread om_in_use_list. 418 // Also updates the per-thread om_in_use_count. 419 static void prepend_to_om_in_use_list(Thread* self, ObjectMonitor* m) { 420 prepend_to_common(m, &self->om_in_use_list, &self->om_in_use_count); 421 } 422 423 // Take an ObjectMonitor from the start of the specified list. Also 424 // decrements the specified counter. Returns NULL if none are available. 425 static ObjectMonitor* take_from_start_of_common(ObjectMonitor* volatile * list_p, 426 int volatile * count_p) { 427 chk_for_list_loop(OrderAccess::load_acquire(list_p), 428 OrderAccess::load_acquire(count_p)); 429 430 ObjectMonitor* next = NULL; 431 ObjectMonitor* take = NULL; 432 // Mark the list head to guard against A-B-A race: 433 if (!mark_list_head(list_p, &take, &next)) { 434 return NULL; // None are available. 435 } 436 // Switch marked list head to next (which unmarks the list head, but 437 // leaves take marked): 438 OrderAccess::release_store(list_p, next); 439 Atomic::dec(count_p); 440 // Unmark take, but leave the next value for any lagging list 441 // walkers. It will get cleaned up when take is prepended to 442 // the in-use list: 443 set_next(take, next); 444 return take; 445 } 446 447 // Take an ObjectMonitor from the start of the global free-list. Also 448 // updates g_om_free_count. Returns NULL if none are available. 449 static ObjectMonitor* take_from_start_of_g_free_list() { 450 return take_from_start_of_common(&g_free_list, &g_om_free_count); 451 } 452 453 // Take an ObjectMonitor from the start of a per-thread free-list. 454 // Also updates om_free_count. Returns NULL if none are available. 455 static ObjectMonitor* take_from_start_of_om_free_list(Thread* self) { 456 return take_from_start_of_common(&self->om_free_list, &self->om_free_count); 457 } 458 459 460 // =====================> Quick functions 461 462 // The quick_* forms are special fast-path variants used to improve 463 // performance. In the simplest case, a "quick_*" implementation could 464 // simply return false, in which case the caller will perform the necessary 465 // state transitions and call the slow-path form. 466 // The fast-path is designed to handle frequently arising cases in an efficient 467 // manner and is just a degenerate "optimistic" variant of the slow-path. 468 // returns true -- to indicate the call was satisfied. 469 // returns false -- to indicate the call needs the services of the slow-path. 470 // A no-loitering ordinance is in effect for code in the quick_* family 471 // operators: safepoints or indefinite blocking (blocking that might span a 472 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon 473 // entry. 474 // 475 // Consider: An interesting optimization is to have the JIT recognize the 476 // following common idiom: 477 // synchronized (someobj) { .... ; notify(); } 478 // That is, we find a notify() or notifyAll() call that immediately precedes 479 // the monitorexit operation. In that case the JIT could fuse the operations 480 // into a single notifyAndExit() runtime primitive. 481 482 bool ObjectSynchronizer::quick_notify(oopDesc* obj, Thread* self, bool all) { 483 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 484 assert(self->is_Java_thread(), "invariant"); 485 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant"); 486 NoSafepointVerifier nsv; 487 if (obj == NULL) return false; // slow-path for invalid obj 488 const markWord mark = obj->mark(); 489 490 if (mark.has_locker() && self->is_lock_owned((address)mark.locker())) { 491 // Degenerate notify 492 // stack-locked by caller so by definition the implied waitset is empty. 493 return true; 494 } 495 496 if (mark.has_monitor()) { 497 ObjectMonitor* const mon = mark.monitor(); 498 assert(mon->object() == obj, "invariant"); 499 if (mon->owner() != self) return false; // slow-path for IMS exception 500 501 if (mon->first_waiter() != NULL) { 502 // We have one or more waiters. Since this is an inflated monitor 503 // that we own, we can transfer one or more threads from the waitset 504 // to the entrylist here and now, avoiding the slow-path. 505 if (all) { 506 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, self); 507 } else { 508 DTRACE_MONITOR_PROBE(notify, mon, obj, self); 509 } 510 int free_count = 0; 511 do { 512 mon->INotify(self); 513 ++free_count; 514 } while (mon->first_waiter() != NULL && all); 515 OM_PERFDATA_OP(Notifications, inc(free_count)); 516 } 517 return true; 518 } 519 520 // biased locking and any other IMS exception states take the slow-path 521 return false; 522 } 523 524 525 // The LockNode emitted directly at the synchronization site would have 526 // been too big if it were to have included support for the cases of inflated 527 // recursive enter and exit, so they go here instead. 528 // Note that we can't safely call AsyncPrintJavaStack() from within 529 // quick_enter() as our thread state remains _in_Java. 530 531 bool ObjectSynchronizer::quick_enter(oop obj, Thread* self, 532 BasicLock * lock) { 533 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 534 assert(self->is_Java_thread(), "invariant"); 535 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant"); 536 NoSafepointVerifier nsv; 537 if (obj == NULL) return false; // Need to throw NPE 538 539 while (true) { 540 const markWord mark = obj->mark(); 541 542 if (mark.has_monitor()) { 543 ObjectMonitorHandle omh; 544 if (!omh.save_om_ptr(obj, mark)) { 545 // Lost a race with async deflation so try again. 546 assert(AsyncDeflateIdleMonitors, "sanity check"); 547 continue; 548 } 549 ObjectMonitor* const m = omh.om_ptr(); 550 assert(m->object() == obj, "invariant"); 551 Thread* const owner = (Thread *) m->_owner; 552 553 // Lock contention and Transactional Lock Elision (TLE) diagnostics 554 // and observability 555 // Case: light contention possibly amenable to TLE 556 // Case: TLE inimical operations such as nested/recursive synchronization 557 558 if (owner == self) { 559 m->_recursions++; 560 return true; 561 } 562 563 // This Java Monitor is inflated so obj's header will never be 564 // displaced to this thread's BasicLock. Make the displaced header 565 // non-NULL so this BasicLock is not seen as recursive nor as 566 // being locked. We do this unconditionally so that this thread's 567 // BasicLock cannot be mis-interpreted by any stack walkers. For 568 // performance reasons, stack walkers generally first check for 569 // Biased Locking in the object's header, the second check is for 570 // stack-locking in the object's header, the third check is for 571 // recursive stack-locking in the displaced header in the BasicLock, 572 // and last are the inflated Java Monitor (ObjectMonitor) checks. 573 lock->set_displaced_header(markWord::unused_mark()); 574 575 if (owner == NULL && Atomic::replace_if_null(self, &(m->_owner))) { 576 assert(m->_recursions == 0, "invariant"); 577 return true; 578 } 579 580 if (AsyncDeflateIdleMonitors && 581 Atomic::cmpxchg(self, &m->_owner, DEFLATER_MARKER) == DEFLATER_MARKER) { 582 // The deflation protocol finished the first part (setting owner), 583 // but it failed the second part (making ref_count negative) and 584 // bailed. Or the ObjectMonitor was async deflated and reused. 585 // Acquired the monitor. 586 assert(m->_recursions == 0, "invariant"); 587 return true; 588 } 589 } 590 break; 591 } 592 593 // Note that we could inflate in quick_enter. 594 // This is likely a useful optimization 595 // Critically, in quick_enter() we must not: 596 // -- perform bias revocation, or 597 // -- block indefinitely, or 598 // -- reach a safepoint 599 600 return false; // revert to slow-path 601 } 602 603 // ----------------------------------------------------------------------------- 604 // Monitor Enter/Exit 605 // The interpreter and compiler assembly code tries to lock using the fast path 606 // of this algorithm. Make sure to update that code if the following function is 607 // changed. The implementation is extremely sensitive to race condition. Be careful. 608 609 void ObjectSynchronizer::enter(Handle obj, BasicLock* lock, TRAPS) { 610 if (UseBiasedLocking) { 611 if (!SafepointSynchronize::is_at_safepoint()) { 612 BiasedLocking::revoke(obj, THREAD); 613 } else { 614 BiasedLocking::revoke_at_safepoint(obj); 615 } 616 } 617 618 markWord mark = obj->mark(); 619 assert(!mark.has_bias_pattern(), "should not see bias pattern here"); 620 621 if (mark.is_neutral()) { 622 // Anticipate successful CAS -- the ST of the displaced mark must 623 // be visible <= the ST performed by the CAS. 624 lock->set_displaced_header(mark); 625 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) { 626 return; 627 } 628 // Fall through to inflate() ... 629 } else if (mark.has_locker() && 630 THREAD->is_lock_owned((address)mark.locker())) { 631 assert(lock != mark.locker(), "must not re-lock the same lock"); 632 assert(lock != (BasicLock*)obj->mark().value(), "don't relock with same BasicLock"); 633 lock->set_displaced_header(markWord::from_pointer(NULL)); 634 return; 635 } 636 637 // The object header will never be displaced to this lock, 638 // so it does not matter what the value is, except that it 639 // must be non-zero to avoid looking like a re-entrant lock, 640 // and must not look locked either. 641 lock->set_displaced_header(markWord::unused_mark()); 642 ObjectMonitorHandle omh; 643 inflate(&omh, THREAD, obj(), inflate_cause_monitor_enter); 644 omh.om_ptr()->enter(THREAD); 645 } 646 647 void ObjectSynchronizer::exit(oop object, BasicLock* lock, TRAPS) { 648 markWord mark = object->mark(); 649 // We cannot check for Biased Locking if we are racing an inflation. 650 assert(mark == markWord::INFLATING() || 651 !mark.has_bias_pattern(), "should not see bias pattern here"); 652 653 markWord dhw = lock->displaced_header(); 654 if (dhw.value() == 0) { 655 // If the displaced header is NULL, then this exit matches up with 656 // a recursive enter. No real work to do here except for diagnostics. 657 #ifndef PRODUCT 658 if (mark != markWord::INFLATING()) { 659 // Only do diagnostics if we are not racing an inflation. Simply 660 // exiting a recursive enter of a Java Monitor that is being 661 // inflated is safe; see the has_monitor() comment below. 662 assert(!mark.is_neutral(), "invariant"); 663 assert(!mark.has_locker() || 664 THREAD->is_lock_owned((address)mark.locker()), "invariant"); 665 if (mark.has_monitor()) { 666 // The BasicLock's displaced_header is marked as a recursive 667 // enter and we have an inflated Java Monitor (ObjectMonitor). 668 // This is a special case where the Java Monitor was inflated 669 // after this thread entered the stack-lock recursively. When a 670 // Java Monitor is inflated, we cannot safely walk the Java 671 // Monitor owner's stack and update the BasicLocks because a 672 // Java Monitor can be asynchronously inflated by a thread that 673 // does not own the Java Monitor. 674 ObjectMonitor* m = mark.monitor(); 675 assert(((oop)(m->object()))->mark() == mark, "invariant"); 676 assert(m->is_entered(THREAD), "invariant"); 677 } 678 } 679 #endif 680 return; 681 } 682 683 if (mark == markWord::from_pointer(lock)) { 684 // If the object is stack-locked by the current thread, try to 685 // swing the displaced header from the BasicLock back to the mark. 686 assert(dhw.is_neutral(), "invariant"); 687 if (object->cas_set_mark(dhw, mark) == mark) { 688 return; 689 } 690 } 691 692 // We have to take the slow-path of possible inflation and then exit. 693 ObjectMonitorHandle omh; 694 inflate(&omh, THREAD, object, inflate_cause_vm_internal); 695 omh.om_ptr()->exit(true, THREAD); 696 } 697 698 // ----------------------------------------------------------------------------- 699 // Class Loader support to workaround deadlocks on the class loader lock objects 700 // Also used by GC 701 // complete_exit()/reenter() are used to wait on a nested lock 702 // i.e. to give up an outer lock completely and then re-enter 703 // Used when holding nested locks - lock acquisition order: lock1 then lock2 704 // 1) complete_exit lock1 - saving recursion count 705 // 2) wait on lock2 706 // 3) when notified on lock2, unlock lock2 707 // 4) reenter lock1 with original recursion count 708 // 5) lock lock2 709 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 710 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { 711 if (UseBiasedLocking) { 712 BiasedLocking::revoke(obj, THREAD); 713 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 714 } 715 716 ObjectMonitorHandle omh; 717 inflate(&omh, THREAD, obj(), inflate_cause_vm_internal); 718 intptr_t ret_code = omh.om_ptr()->complete_exit(THREAD); 719 return ret_code; 720 } 721 722 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() 723 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, 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 ObjectMonitorHandle omh; 730 inflate(&omh, THREAD, obj(), inflate_cause_vm_internal); 731 omh.om_ptr()->reenter(recursion, THREAD); 732 } 733 // ----------------------------------------------------------------------------- 734 // JNI locks on java objects 735 // NOTE: must use heavy weight monitor to handle jni monitor enter 736 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { 737 // the current locking is from JNI instead of Java code 738 if (UseBiasedLocking) { 739 BiasedLocking::revoke(obj, THREAD); 740 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 741 } 742 THREAD->set_current_pending_monitor_is_from_java(false); 743 ObjectMonitorHandle omh; 744 inflate(&omh, THREAD, obj(), inflate_cause_jni_enter); 745 omh.om_ptr()->enter(THREAD); 746 THREAD->set_current_pending_monitor_is_from_java(true); 747 } 748 749 // NOTE: must use heavy weight monitor to handle jni monitor exit 750 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { 751 if (UseBiasedLocking) { 752 Handle h_obj(THREAD, obj); 753 BiasedLocking::revoke(h_obj, THREAD); 754 obj = h_obj(); 755 } 756 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 757 758 ObjectMonitorHandle omh; 759 inflate(&omh, THREAD, obj, inflate_cause_jni_exit); 760 ObjectMonitor* monitor = omh.om_ptr(); 761 // If this thread has locked the object, exit the monitor. We 762 // intentionally do not use CHECK here because we must exit the 763 // monitor even if an exception is pending. 764 if (monitor->check_owner(THREAD)) { 765 monitor->exit(true, THREAD); 766 } 767 } 768 769 // ----------------------------------------------------------------------------- 770 // Internal VM locks on java objects 771 // standard constructor, allows locking failures 772 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool do_lock) { 773 _dolock = do_lock; 774 _thread = thread; 775 _thread->check_for_valid_safepoint_state(); 776 _obj = obj; 777 778 if (_dolock) { 779 ObjectSynchronizer::enter(_obj, &_lock, _thread); 780 } 781 } 782 783 ObjectLocker::~ObjectLocker() { 784 if (_dolock) { 785 ObjectSynchronizer::exit(_obj(), &_lock, _thread); 786 } 787 } 788 789 790 // ----------------------------------------------------------------------------- 791 // Wait/Notify/NotifyAll 792 // NOTE: must use heavy weight monitor to handle wait() 793 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { 794 if (UseBiasedLocking) { 795 BiasedLocking::revoke(obj, THREAD); 796 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 797 } 798 if (millis < 0) { 799 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 800 } 801 ObjectMonitorHandle omh; 802 inflate(&omh, THREAD, obj(), inflate_cause_wait); 803 ObjectMonitor* monitor = omh.om_ptr(); 804 805 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); 806 monitor->wait(millis, true, THREAD); 807 808 // This dummy call is in place to get around dtrace bug 6254741. Once 809 // that's fixed we can uncomment the following line, remove the call 810 // and change this function back into a "void" func. 811 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); 812 int ret_code = dtrace_waited_probe(monitor, obj, THREAD); 813 return ret_code; 814 } 815 816 void ObjectSynchronizer::wait_uninterruptibly(Handle obj, jlong millis, TRAPS) { 817 if (UseBiasedLocking) { 818 BiasedLocking::revoke(obj, THREAD); 819 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 820 } 821 if (millis < 0) { 822 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); 823 } 824 ObjectMonitorHandle omh; 825 inflate(&omh, THREAD, obj(), inflate_cause_wait); 826 omh.om_ptr()->wait(millis, false, THREAD); 827 } 828 829 void ObjectSynchronizer::notify(Handle obj, TRAPS) { 830 if (UseBiasedLocking) { 831 BiasedLocking::revoke(obj, THREAD); 832 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 833 } 834 835 markWord mark = obj->mark(); 836 if (mark.has_locker() && THREAD->is_lock_owned((address)mark.locker())) { 837 return; 838 } 839 ObjectMonitorHandle omh; 840 inflate(&omh, THREAD, obj(), inflate_cause_notify); 841 omh.om_ptr()->notify(THREAD); 842 } 843 844 // NOTE: see comment of notify() 845 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { 846 if (UseBiasedLocking) { 847 BiasedLocking::revoke(obj, THREAD); 848 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 849 } 850 851 markWord mark = obj->mark(); 852 if (mark.has_locker() && THREAD->is_lock_owned((address)mark.locker())) { 853 return; 854 } 855 ObjectMonitorHandle omh; 856 inflate(&omh, THREAD, obj(), inflate_cause_notify); 857 omh.om_ptr()->notifyAll(THREAD); 858 } 859 860 // ----------------------------------------------------------------------------- 861 // Hash Code handling 862 // 863 // Performance concern: 864 // OrderAccess::storestore() calls release() which at one time stored 0 865 // into the global volatile OrderAccess::dummy variable. This store was 866 // unnecessary for correctness. Many threads storing into a common location 867 // causes considerable cache migration or "sloshing" on large SMP systems. 868 // As such, I avoided using OrderAccess::storestore(). In some cases 869 // OrderAccess::fence() -- which incurs local latency on the executing 870 // processor -- is a better choice as it scales on SMP systems. 871 // 872 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for 873 // a discussion of coherency costs. Note that all our current reference 874 // platforms provide strong ST-ST order, so the issue is moot on IA32, 875 // x64, and SPARC. 876 // 877 // As a general policy we use "volatile" to control compiler-based reordering 878 // and explicit fences (barriers) to control for architectural reordering 879 // performed by the CPU(s) or platform. 880 881 struct SharedGlobals { 882 char _pad_prefix[OM_CACHE_LINE_SIZE]; 883 // These are highly shared mostly-read variables. 884 // To avoid false-sharing they need to be the sole occupants of a cache line. 885 volatile int stw_random; 886 volatile int stw_cycle; 887 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int) * 2); 888 // Hot RW variable -- Sequester to avoid false-sharing 889 volatile int hc_sequence; 890 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int)); 891 }; 892 893 static SharedGlobals GVars; 894 static int MonitorScavengeThreshold = 1000000; 895 static volatile int ForceMonitorScavenge = 0; // Scavenge required and pending 896 897 static markWord read_stable_mark(oop obj) { 898 markWord mark = obj->mark(); 899 if (!mark.is_being_inflated()) { 900 return mark; // normal fast-path return 901 } 902 903 int its = 0; 904 for (;;) { 905 markWord mark = obj->mark(); 906 if (!mark.is_being_inflated()) { 907 return mark; // normal fast-path return 908 } 909 910 // The object is being inflated by some other thread. 911 // The caller of read_stable_mark() must wait for inflation to complete. 912 // Avoid live-lock 913 // TODO: consider calling SafepointSynchronize::do_call_back() while 914 // spinning to see if there's a safepoint pending. If so, immediately 915 // yielding or blocking would be appropriate. Avoid spinning while 916 // there is a safepoint pending. 917 // TODO: add inflation contention performance counters. 918 // TODO: restrict the aggregate number of spinners. 919 920 ++its; 921 if (its > 10000 || !os::is_MP()) { 922 if (its & 1) { 923 os::naked_yield(); 924 } else { 925 // Note that the following code attenuates the livelock problem but is not 926 // a complete remedy. A more complete solution would require that the inflating 927 // thread hold the associated inflation lock. The following code simply restricts 928 // the number of spinners to at most one. We'll have N-2 threads blocked 929 // on the inflationlock, 1 thread holding the inflation lock and using 930 // a yield/park strategy, and 1 thread in the midst of inflation. 931 // A more refined approach would be to change the encoding of INFLATING 932 // to allow encapsulation of a native thread pointer. Threads waiting for 933 // inflation to complete would use CAS to push themselves onto a singly linked 934 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag 935 // and calling park(). When inflation was complete the thread that accomplished inflation 936 // would detach the list and set the markword to inflated with a single CAS and 937 // then for each thread on the list, set the flag and unpark() the thread. 938 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease 939 // wakes at most one thread whereas we need to wake the entire list. 940 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1); 941 int YieldThenBlock = 0; 942 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant"); 943 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant"); 944 Thread::muxAcquire(gInflationLocks + ix, "gInflationLock"); 945 while (obj->mark() == markWord::INFLATING()) { 946 // Beware: NakedYield() is advisory and has almost no effect on some platforms 947 // so we periodically call self->_ParkEvent->park(1). 948 // We use a mixed spin/yield/block mechanism. 949 if ((YieldThenBlock++) >= 16) { 950 Thread::current()->_ParkEvent->park(1); 951 } else { 952 os::naked_yield(); 953 } 954 } 955 Thread::muxRelease(gInflationLocks + ix); 956 } 957 } else { 958 SpinPause(); // SMP-polite spinning 959 } 960 } 961 } 962 963 // hashCode() generation : 964 // 965 // Possibilities: 966 // * MD5Digest of {obj,stw_random} 967 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function. 968 // * A DES- or AES-style SBox[] mechanism 969 // * One of the Phi-based schemes, such as: 970 // 2654435761 = 2^32 * Phi (golden ratio) 971 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ; 972 // * A variation of Marsaglia's shift-xor RNG scheme. 973 // * (obj ^ stw_random) is appealing, but can result 974 // in undesirable regularity in the hashCode values of adjacent objects 975 // (objects allocated back-to-back, in particular). This could potentially 976 // result in hashtable collisions and reduced hashtable efficiency. 977 // There are simple ways to "diffuse" the middle address bits over the 978 // generated hashCode values: 979 980 static inline intptr_t get_next_hash(Thread* self, oop obj) { 981 intptr_t value = 0; 982 if (hashCode == 0) { 983 // This form uses global Park-Miller RNG. 984 // On MP system we'll have lots of RW access to a global, so the 985 // mechanism induces lots of coherency traffic. 986 value = os::random(); 987 } else if (hashCode == 1) { 988 // This variation has the property of being stable (idempotent) 989 // between STW operations. This can be useful in some of the 1-0 990 // synchronization schemes. 991 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3; 992 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random; 993 } else if (hashCode == 2) { 994 value = 1; // for sensitivity testing 995 } else if (hashCode == 3) { 996 value = ++GVars.hc_sequence; 997 } else if (hashCode == 4) { 998 value = cast_from_oop<intptr_t>(obj); 999 } else { 1000 // Marsaglia's xor-shift scheme with thread-specific state 1001 // This is probably the best overall implementation -- we'll 1002 // likely make this the default in future releases. 1003 unsigned t = self->_hashStateX; 1004 t ^= (t << 11); 1005 self->_hashStateX = self->_hashStateY; 1006 self->_hashStateY = self->_hashStateZ; 1007 self->_hashStateZ = self->_hashStateW; 1008 unsigned v = self->_hashStateW; 1009 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)); 1010 self->_hashStateW = v; 1011 value = v; 1012 } 1013 1014 value &= markWord::hash_mask; 1015 if (value == 0) value = 0xBAD; 1016 assert(value != markWord::no_hash, "invariant"); 1017 return value; 1018 } 1019 1020 intptr_t ObjectSynchronizer::FastHashCode(Thread* self, oop obj) { 1021 if (UseBiasedLocking) { 1022 // NOTE: many places throughout the JVM do not expect a safepoint 1023 // to be taken here, in particular most operations on perm gen 1024 // objects. However, we only ever bias Java instances and all of 1025 // the call sites of identity_hash that might revoke biases have 1026 // been checked to make sure they can handle a safepoint. The 1027 // added check of the bias pattern is to avoid useless calls to 1028 // thread-local storage. 1029 if (obj->mark().has_bias_pattern()) { 1030 // Handle for oop obj in case of STW safepoint 1031 Handle hobj(self, obj); 1032 // Relaxing assertion for bug 6320749. 1033 assert(Universe::verify_in_progress() || 1034 !SafepointSynchronize::is_at_safepoint(), 1035 "biases should not be seen by VM thread here"); 1036 BiasedLocking::revoke(hobj, JavaThread::current()); 1037 obj = hobj(); 1038 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now"); 1039 } 1040 } 1041 1042 // hashCode() is a heap mutator ... 1043 // Relaxing assertion for bug 6320749. 1044 assert(Universe::verify_in_progress() || DumpSharedSpaces || 1045 !SafepointSynchronize::is_at_safepoint(), "invariant"); 1046 assert(Universe::verify_in_progress() || DumpSharedSpaces || 1047 self->is_Java_thread() , "invariant"); 1048 assert(Universe::verify_in_progress() || DumpSharedSpaces || 1049 ((JavaThread *)self)->thread_state() != _thread_blocked, "invariant"); 1050 1051 while (true) { 1052 ObjectMonitor* monitor = NULL; 1053 markWord temp, test; 1054 intptr_t hash; 1055 markWord mark = read_stable_mark(obj); 1056 1057 // object should remain ineligible for biased locking 1058 assert(!mark.has_bias_pattern(), "invariant"); 1059 1060 if (mark.is_neutral()) { 1061 hash = mark.hash(); // this is a normal header 1062 if (hash != 0) { // if it has hash, just return it 1063 return hash; 1064 } 1065 hash = get_next_hash(self, obj); // allocate a new hash code 1066 temp = mark.copy_set_hash(hash); // merge the hash code into header 1067 // use (machine word version) atomic operation to install the hash 1068 test = obj->cas_set_mark(temp, mark); 1069 if (test == mark) { 1070 return hash; 1071 } 1072 // If atomic operation failed, we must inflate the header 1073 // into heavy weight monitor. We could add more code here 1074 // for fast path, but it does not worth the complexity. 1075 } else if (mark.has_monitor()) { 1076 ObjectMonitorHandle omh; 1077 if (!omh.save_om_ptr(obj, mark)) { 1078 // Lost a race with async deflation so try again. 1079 assert(AsyncDeflateIdleMonitors, "sanity check"); 1080 continue; 1081 } 1082 monitor = omh.om_ptr(); 1083 temp = monitor->header(); 1084 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1085 hash = temp.hash(); 1086 if (hash != 0) { 1087 return hash; 1088 } 1089 // Skip to the following code to reduce code size 1090 } else if (self->is_lock_owned((address)mark.locker())) { 1091 temp = mark.displaced_mark_helper(); // this is a lightweight monitor owned 1092 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1093 hash = temp.hash(); // by current thread, check if the displaced 1094 if (hash != 0) { // header contains hash code 1095 return hash; 1096 } 1097 // WARNING: 1098 // The displaced header in the BasicLock on a thread's stack 1099 // is strictly immutable. It CANNOT be changed in ANY cases. 1100 // So we have to inflate the stack lock into an ObjectMonitor 1101 // even if the current thread owns the lock. The BasicLock on 1102 // a thread's stack can be asynchronously read by other threads 1103 // during an inflate() call so any change to that stack memory 1104 // may not propagate to other threads correctly. 1105 } 1106 1107 // Inflate the monitor to set hash code 1108 ObjectMonitorHandle omh; 1109 inflate(&omh, self, obj, inflate_cause_hash_code); 1110 monitor = omh.om_ptr(); 1111 // Load displaced header and check it has hash code 1112 mark = monitor->header(); 1113 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value()); 1114 hash = mark.hash(); 1115 if (hash == 0) { 1116 hash = get_next_hash(self, obj); 1117 temp = mark.copy_set_hash(hash); // merge hash code into header 1118 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value()); 1119 uintptr_t v = Atomic::cmpxchg(temp.value(), (volatile uintptr_t*)monitor->header_addr(), mark.value()); 1120 test = markWord(v); 1121 if (test != mark) { 1122 // The only non-deflation update to the ObjectMonitor's 1123 // header/dmw field is to merge in the hash code. If someone 1124 // adds a new usage of the header/dmw field, please update 1125 // this code. 1126 // ObjectMonitor::install_displaced_markword_in_object() 1127 // does mark the header/dmw field as part of async deflation, 1128 // but that protocol cannot happen now due to the 1129 // ObjectMonitorHandle above. 1130 hash = test.hash(); 1131 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value()); 1132 assert(hash != 0, "Trivial unexpected object/monitor header usage."); 1133 } 1134 } 1135 // We finally get the hash 1136 return hash; 1137 } 1138 } 1139 1140 // Deprecated -- use FastHashCode() instead. 1141 1142 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { 1143 return FastHashCode(Thread::current(), obj()); 1144 } 1145 1146 1147 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, 1148 Handle h_obj) { 1149 if (UseBiasedLocking) { 1150 BiasedLocking::revoke(h_obj, thread); 1151 assert(!h_obj->mark().has_bias_pattern(), "biases should be revoked by now"); 1152 } 1153 1154 assert(thread == JavaThread::current(), "Can only be called on current thread"); 1155 oop obj = h_obj(); 1156 1157 while (true) { 1158 markWord mark = read_stable_mark(obj); 1159 1160 // Uncontended case, header points to stack 1161 if (mark.has_locker()) { 1162 return thread->is_lock_owned((address)mark.locker()); 1163 } 1164 // Contended case, header points to ObjectMonitor (tagged pointer) 1165 if (mark.has_monitor()) { 1166 ObjectMonitorHandle omh; 1167 if (!omh.save_om_ptr(obj, mark)) { 1168 // Lost a race with async deflation so try again. 1169 assert(AsyncDeflateIdleMonitors, "sanity check"); 1170 continue; 1171 } 1172 bool ret_code = omh.om_ptr()->is_entered(thread) != 0; 1173 return ret_code; 1174 } 1175 // Unlocked case, header in place 1176 assert(mark.is_neutral(), "sanity check"); 1177 return false; 1178 } 1179 } 1180 1181 // Be aware of this method could revoke bias of the lock object. 1182 // This method queries the ownership of the lock handle specified by 'h_obj'. 1183 // If the current thread owns the lock, it returns owner_self. If no 1184 // thread owns the lock, it returns owner_none. Otherwise, it will return 1185 // owner_other. 1186 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership 1187 (JavaThread *self, Handle h_obj) { 1188 // The caller must beware this method can revoke bias, and 1189 // revocation can result in a safepoint. 1190 assert(!SafepointSynchronize::is_at_safepoint(), "invariant"); 1191 assert(self->thread_state() != _thread_blocked, "invariant"); 1192 1193 // Possible mark states: neutral, biased, stack-locked, inflated 1194 1195 if (UseBiasedLocking && h_obj()->mark().has_bias_pattern()) { 1196 // CASE: biased 1197 BiasedLocking::revoke(h_obj, self); 1198 assert(!h_obj->mark().has_bias_pattern(), 1199 "biases should be revoked by now"); 1200 } 1201 1202 assert(self == JavaThread::current(), "Can only be called on current thread"); 1203 oop obj = h_obj(); 1204 1205 while (true) { 1206 markWord mark = read_stable_mark(obj); 1207 1208 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. 1209 if (mark.has_locker()) { 1210 return self->is_lock_owned((address)mark.locker()) ? 1211 owner_self : owner_other; 1212 } 1213 1214 // CASE: inflated. Mark (tagged pointer) points to an ObjectMonitor. 1215 // The Object:ObjectMonitor relationship is stable as long as we're 1216 // not at a safepoint and AsyncDeflateIdleMonitors is false. 1217 if (mark.has_monitor()) { 1218 ObjectMonitorHandle omh; 1219 if (!omh.save_om_ptr(obj, mark)) { 1220 // Lost a race with async deflation so try again. 1221 assert(AsyncDeflateIdleMonitors, "sanity check"); 1222 continue; 1223 } 1224 ObjectMonitor* monitor = omh.om_ptr(); 1225 void* owner = monitor->_owner; 1226 if (owner == NULL) return owner_none; 1227 return (owner == self || 1228 self->is_lock_owned((address)owner)) ? owner_self : owner_other; 1229 } 1230 1231 // CASE: neutral 1232 assert(mark.is_neutral(), "sanity check"); 1233 return owner_none; // it's unlocked 1234 } 1235 } 1236 1237 // FIXME: jvmti should call this 1238 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) { 1239 if (UseBiasedLocking) { 1240 if (SafepointSynchronize::is_at_safepoint()) { 1241 BiasedLocking::revoke_at_safepoint(h_obj); 1242 } else { 1243 BiasedLocking::revoke(h_obj, JavaThread::current()); 1244 } 1245 assert(!h_obj->mark().has_bias_pattern(), "biases should be revoked by now"); 1246 } 1247 1248 oop obj = h_obj(); 1249 1250 while (true) { 1251 address owner = NULL; 1252 markWord mark = read_stable_mark(obj); 1253 1254 // Uncontended case, header points to stack 1255 if (mark.has_locker()) { 1256 owner = (address) mark.locker(); 1257 } 1258 1259 // Contended case, header points to ObjectMonitor (tagged pointer) 1260 else if (mark.has_monitor()) { 1261 ObjectMonitorHandle omh; 1262 if (!omh.save_om_ptr(obj, mark)) { 1263 // Lost a race with async deflation so try again. 1264 assert(AsyncDeflateIdleMonitors, "sanity check"); 1265 continue; 1266 } 1267 ObjectMonitor* monitor = omh.om_ptr(); 1268 assert(monitor != NULL, "monitor should be non-null"); 1269 owner = (address) monitor->owner(); 1270 } 1271 1272 if (owner != NULL) { 1273 // owning_thread_from_monitor_owner() may also return NULL here 1274 return Threads::owning_thread_from_monitor_owner(t_list, owner); 1275 } 1276 1277 // Unlocked case, header in place 1278 // Cannot have assertion since this object may have been 1279 // locked by another thread when reaching here. 1280 // assert(mark.is_neutral(), "sanity check"); 1281 1282 return NULL; 1283 } 1284 } 1285 1286 // Visitors ... 1287 1288 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { 1289 PaddedObjectMonitor* block = OrderAccess::load_acquire(&g_block_list); 1290 while (block != NULL) { 1291 assert(block->object() == CHAINMARKER, "must be a block header"); 1292 for (int i = _BLOCKSIZE - 1; i > 0; i--) { 1293 ObjectMonitor* mid = (ObjectMonitor *)(block + i); 1294 if (mid->is_active()) { 1295 ObjectMonitorHandle omh(mid); 1296 1297 if (mid->object() == NULL || 1298 (AsyncDeflateIdleMonitors && mid->ref_count() < 0)) { 1299 // Only process with closure if the object is set. 1300 // For async deflation, race here if monitor is not owned! 1301 // The above ref_count bump (in ObjectMonitorHandle ctr) 1302 // will cause subsequent async deflation to skip it. 1303 // However, previous or concurrent async deflation is a race 1304 // so skip this ObjectMonitor if it is being async deflated. 1305 continue; 1306 } 1307 closure->do_monitor(mid); 1308 } 1309 } 1310 // unmarked_next() is not needed with g_block_list (no next field marking). 1311 block = (PaddedObjectMonitor*)OrderAccess::load_acquire(&block->_next_om); 1312 } 1313 } 1314 1315 static bool monitors_used_above_threshold() { 1316 if (OrderAccess::load_acquire(&g_om_population) == 0) { 1317 return false; 1318 } 1319 if (MonitorUsedDeflationThreshold > 0) { 1320 int monitors_used = OrderAccess::load_acquire(&g_om_population) - 1321 OrderAccess::load_acquire(&g_om_free_count); 1322 int monitor_usage = (monitors_used * 100LL) / 1323 OrderAccess::load_acquire(&g_om_population); 1324 return monitor_usage > MonitorUsedDeflationThreshold; 1325 } 1326 return false; 1327 } 1328 1329 // Returns true if MonitorBound is set (> 0) and if the specified 1330 // cnt is > MonitorBound. Otherwise returns false. 1331 static bool is_MonitorBound_exceeded(const int cnt) { 1332 const int mx = MonitorBound; 1333 return mx > 0 && cnt > mx; 1334 } 1335 1336 bool ObjectSynchronizer::is_async_deflation_needed() { 1337 if (!AsyncDeflateIdleMonitors) { 1338 return false; 1339 } 1340 if (is_async_deflation_requested()) { 1341 // Async deflation request. 1342 return true; 1343 } 1344 if (AsyncDeflationInterval > 0 && 1345 time_since_last_async_deflation_ms() > AsyncDeflationInterval && 1346 monitors_used_above_threshold()) { 1347 // It's been longer than our specified deflate interval and there 1348 // are too many monitors in use. We don't deflate more frequently 1349 // than AsyncDeflationInterval (unless is_async_deflation_requested) 1350 // in order to not swamp the ServiceThread. 1351 _last_async_deflation_time_ns = os::javaTimeNanos(); 1352 return true; 1353 } 1354 if (is_MonitorBound_exceeded(OrderAccess::load_acquire(&g_om_population) - 1355 OrderAccess::load_acquire(&g_om_free_count))) { 1356 // Not enough ObjectMonitors on the global free list. 1357 return true; 1358 } 1359 return false; 1360 } 1361 1362 bool ObjectSynchronizer::is_safepoint_deflation_needed() { 1363 if (!AsyncDeflateIdleMonitors) { 1364 if (monitors_used_above_threshold()) { 1365 // Too many monitors in use. 1366 return true; 1367 } 1368 return false; 1369 } 1370 if (is_special_deflation_requested()) { 1371 // For AsyncDeflateIdleMonitors only do a safepoint deflation 1372 // if there is a special deflation request. 1373 return true; 1374 } 1375 return false; 1376 } 1377 1378 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() { 1379 return (os::javaTimeNanos() - _last_async_deflation_time_ns) / (NANOUNITS / MILLIUNITS); 1380 } 1381 1382 void ObjectSynchronizer::oops_do(OopClosure* f) { 1383 // We only scan the global used list here (for moribund threads), and 1384 // the thread-local monitors in Thread::oops_do(). 1385 global_used_oops_do(f); 1386 } 1387 1388 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) { 1389 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1390 list_oops_do(OrderAccess::load_acquire(&g_om_in_use_list), OrderAccess::load_acquire(&g_om_in_use_count), f); 1391 } 1392 1393 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) { 1394 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1395 list_oops_do(OrderAccess::load_acquire(&thread->om_in_use_list), OrderAccess::load_acquire(&thread->om_in_use_count), f); 1396 } 1397 1398 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, int count, OopClosure* f) { 1399 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 1400 chk_for_list_loop(list, count); 1401 // The oops_do() phase does not overlap with monitor deflation 1402 // so no need to update the ObjectMonitor's ref_count for this 1403 // ObjectMonitor* use. 1404 for (ObjectMonitor* mid = list; mid != NULL; mid = unmarked_next(mid)) { 1405 if (mid->object() != NULL) { 1406 f->do_oop((oop*)mid->object_addr()); 1407 } 1408 } 1409 } 1410 1411 1412 // ----------------------------------------------------------------------------- 1413 // ObjectMonitor Lifecycle 1414 // ----------------------- 1415 // Inflation unlinks monitors from the global g_free_list and 1416 // associates them with objects. Deflation -- which occurs at 1417 // STW-time -- disassociates idle monitors from objects. Such 1418 // scavenged monitors are returned to the g_free_list. 1419 // 1420 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. 1421 // 1422 // Lifecycle: 1423 // -- unassigned and on the global free list 1424 // -- unassigned and on a thread's private om_free_list 1425 // -- assigned to an object. The object is inflated and the mark refers 1426 // to the objectmonitor. 1427 1428 1429 // Constraining monitor pool growth via MonitorBound ... 1430 // 1431 // If MonitorBound is not set (<= 0), MonitorBound checks are disabled. 1432 // 1433 // When safepoint deflation is being used (!AsyncDeflateIdleMonitors): 1434 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the 1435 // the rate of scavenging is driven primarily by GC. As such, we can find 1436 // an inordinate number of monitors in circulation. 1437 // To avoid that scenario we can artificially induce a STW safepoint 1438 // if the pool appears to be growing past some reasonable bound. 1439 // Generally we favor time in space-time tradeoffs, but as there's no 1440 // natural back-pressure on the # of extant monitors we need to impose some 1441 // type of limit. Beware that if MonitorBound is set to too low a value 1442 // we could just loop. In addition, if MonitorBound is set to a low value 1443 // we'll incur more safepoints, which are harmful to performance. 1444 // See also: GuaranteedSafepointInterval 1445 // 1446 // The current implementation uses asynchronous VM operations. 1447 // 1448 // When safepoint deflation is being used and MonitorBound is set, the 1449 // boundry applies to 1450 // (g_om_population - g_om_free_count) 1451 // i.e., if there are not enough ObjectMonitors on the global free list, 1452 // then a safepoint deflation is induced. Picking a good MonitorBound value 1453 // is non-trivial. 1454 // 1455 // When async deflation is being used: 1456 // The monitor pool is still grow-only. Async deflation is requested 1457 // by a safepoint's cleanup phase or by the ServiceThread at periodic 1458 // intervals when is_async_deflation_needed() returns true. In 1459 // addition to other policies that are checked, if there are not 1460 // enough ObjectMonitors on the global free list, then 1461 // is_async_deflation_needed() will return true. The ServiceThread 1462 // calls deflate_global_idle_monitors_using_JT() and also calls 1463 // deflate_per_thread_idle_monitors_using_JT() as needed. 1464 1465 static void InduceScavenge(Thread* self, const char * Whence) { 1466 assert(!AsyncDeflateIdleMonitors, "is not used by async deflation"); 1467 1468 // Induce STW safepoint to trim monitors 1469 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. 1470 // More precisely, trigger an asynchronous STW safepoint as the number 1471 // of active monitors passes the specified threshold. 1472 // TODO: assert thread state is reasonable 1473 1474 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { 1475 // Induce a 'null' safepoint to scavenge monitors 1476 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted 1477 // to the VMthread and have a lifespan longer than that of this activation record. 1478 // The VMThread will delete the op when completed. 1479 VMThread::execute(new VM_ScavengeMonitors()); 1480 } 1481 } 1482 1483 ObjectMonitor* ObjectSynchronizer::om_alloc(Thread* self, 1484 const InflateCause cause) { 1485 // A large MAXPRIVATE value reduces both list lock contention 1486 // and list coherency traffic, but also tends to increase the 1487 // number of ObjectMonitors in circulation as well as the STW 1488 // scavenge costs. As usual, we lean toward time in space-time 1489 // tradeoffs. 1490 const int MAXPRIVATE = 1024; 1491 1492 stringStream ss; 1493 for (;;) { 1494 ObjectMonitor* m; 1495 1496 // 1: try to allocate from the thread's local om_free_list. 1497 // Threads will attempt to allocate first from their local list, then 1498 // from the global list, and only after those attempts fail will the 1499 // thread attempt to instantiate new monitors. Thread-local free lists 1500 // improve allocation latency, as well as reducing coherency traffic 1501 // on the shared global list. 1502 m = take_from_start_of_om_free_list(self); 1503 if (m != NULL) { 1504 guarantee(m->object() == NULL, "invariant"); 1505 m->set_allocation_state(ObjectMonitor::New); 1506 prepend_to_om_in_use_list(self, m); 1507 return m; 1508 } 1509 1510 // 2: try to allocate from the global g_free_list 1511 // CONSIDER: use muxTry() instead of muxAcquire(). 1512 // If the muxTry() fails then drop immediately into case 3. 1513 // If we're using thread-local free lists then try 1514 // to reprovision the caller's free list. 1515 if (OrderAccess::load_acquire(&g_free_list) != NULL) { 1516 // Reprovision the thread's om_free_list. 1517 // Use bulk transfers to reduce the allocation rate and heat 1518 // on various locks. 1519 for (int i = self->om_free_provision; --i >= 0;) { 1520 ObjectMonitor* take = take_from_start_of_g_free_list(); 1521 if (take == NULL) { 1522 break; // No more are available. 1523 } 1524 guarantee(take->object() == NULL, "invariant"); 1525 if (AsyncDeflateIdleMonitors) { 1526 // We allowed 3 field values to linger during async deflation. 1527 // We clear header and restore ref_count here, but we leave 1528 // owner == DEFLATER_MARKER so the simple C2 ObjectMonitor 1529 // enter optimization can no longer race with async deflation 1530 // and reuse. 1531 take->set_header(markWord::zero()); 1532 if (take->ref_count() < 0) { 1533 // Add back max_jint to restore the ref_count field to its 1534 // proper value. 1535 Atomic::add(max_jint, &take->_ref_count); 1536 1537 assert(take->ref_count() >= 0, "must not be negative: ref_count=%d", 1538 take->ref_count()); 1539 } 1540 } 1541 take->Recycle(); 1542 assert(take->is_free(), "invariant"); 1543 om_release(self, take, false); 1544 } 1545 self->om_free_provision += 1 + (self->om_free_provision/2); 1546 if (self->om_free_provision > MAXPRIVATE) self->om_free_provision = MAXPRIVATE; 1547 1548 if (!AsyncDeflateIdleMonitors && 1549 is_MonitorBound_exceeded(OrderAccess::load_acquire(&g_om_population) - 1550 OrderAccess::load_acquire(&g_om_free_count))) { 1551 // Not enough ObjectMonitors on the global free list. 1552 // We can't safely induce a STW safepoint from om_alloc() as our thread 1553 // state may not be appropriate for such activities and callers may hold 1554 // naked oops, so instead we defer the action. 1555 InduceScavenge(self, "om_alloc"); 1556 } 1557 continue; 1558 } 1559 1560 // 3: allocate a block of new ObjectMonitors 1561 // Both the local and global free lists are empty -- resort to malloc(). 1562 // In the current implementation ObjectMonitors are TSM - immortal. 1563 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want 1564 // each ObjectMonitor to start at the beginning of a cache line, 1565 // so we use align_up(). 1566 // A better solution would be to use C++ placement-new. 1567 // BEWARE: As it stands currently, we don't run the ctors! 1568 assert(_BLOCKSIZE > 1, "invariant"); 1569 size_t neededsize = sizeof(PaddedObjectMonitor) * _BLOCKSIZE; 1570 PaddedObjectMonitor* temp; 1571 size_t aligned_size = neededsize + (OM_CACHE_LINE_SIZE - 1); 1572 void* real_malloc_addr = NEW_C_HEAP_ARRAY(char, aligned_size, mtInternal); 1573 temp = (PaddedObjectMonitor*)align_up(real_malloc_addr, OM_CACHE_LINE_SIZE); 1574 (void)memset((void *) temp, 0, neededsize); 1575 1576 // Format the block. 1577 // initialize the linked list, each monitor points to its next 1578 // forming the single linked free list, the very first monitor 1579 // will points to next block, which forms the block list. 1580 // The trick of using the 1st element in the block as g_block_list 1581 // linkage should be reconsidered. A better implementation would 1582 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } 1583 1584 for (int i = 1; i < _BLOCKSIZE; i++) { 1585 OrderAccess::release_store(&temp[i]._next_om, (ObjectMonitor*)&temp[i+1]); 1586 assert(temp[i].is_free(), "invariant"); 1587 } 1588 1589 // terminate the last monitor as the end of list 1590 OrderAccess::release_store(&temp[_BLOCKSIZE - 1]._next_om, (ObjectMonitor*)NULL); 1591 1592 // Element [0] is reserved for global list linkage 1593 temp[0].set_object(CHAINMARKER); 1594 1595 // Consider carving out this thread's current request from the 1596 // block in hand. This avoids some lock traffic and redundant 1597 // list activity. 1598 1599 prepend_block_to_lists(temp); 1600 } 1601 } 1602 1603 // Place "m" on the caller's private per-thread om_free_list. 1604 // In practice there's no need to clamp or limit the number of 1605 // monitors on a thread's om_free_list as the only non-allocation time 1606 // we'll call om_release() is to return a monitor to the free list after 1607 // a CAS attempt failed. This doesn't allow unbounded #s of monitors to 1608 // accumulate on a thread's free list. 1609 // 1610 // Key constraint: all ObjectMonitors on a thread's free list and the global 1611 // free list must have their object field set to null. This prevents the 1612 // scavenger -- deflate_monitor_list() or deflate_monitor_list_using_JT() 1613 // -- from reclaiming them while we are trying to release them. 1614 1615 void ObjectSynchronizer::om_release(Thread* self, ObjectMonitor* m, 1616 bool from_per_thread_alloc) { 1617 guarantee(m->header().value() == 0, "invariant"); 1618 guarantee(m->object() == NULL, "invariant"); 1619 stringStream ss; 1620 guarantee((m->is_busy() | m->_recursions) == 0, "freeing in-use monitor: " 1621 "%s, recursions=" INTX_FORMAT, m->is_busy_to_string(&ss), 1622 m->_recursions); 1623 m->set_allocation_state(ObjectMonitor::Free); 1624 // _next_om is used for both per-thread in-use and free lists so 1625 // we have to remove 'm' from the in-use list first (as needed). 1626 if (from_per_thread_alloc) { 1627 // Need to remove 'm' from om_in_use_list. 1628 // We use the more complicated mark-cur_mid_in_use-and-mid-as-we-go 1629 // protocol because async deflation can do list deletions in parallel. 1630 ObjectMonitor* cur_mid_in_use = NULL; 1631 ObjectMonitor* mid = NULL; 1632 ObjectMonitor* next = NULL; 1633 bool extracted = false; 1634 1635 if (!mark_list_head(&self->om_in_use_list, &mid, &next)) { 1636 fatal("thread=" INTPTR_FORMAT " in-use list must not be empty.", p2i(self)); 1637 } 1638 while (true) { 1639 if (m == mid) { 1640 // We found 'm' on the per-thread in-use list so try to extract it. 1641 // First try the list head: 1642 if (Atomic::cmpxchg(next, &self->om_in_use_list, mid) != mid) { 1643 // We could not switch the list head to next. 1644 ObjectMonitor* marked_mid = mark_om_ptr(mid); 1645 // Switch cur_mid_in_use's next field to next (which also 1646 // unmarks cur_mid_in_use): 1647 ADIM_guarantee(cur_mid_in_use != NULL, "must not be NULL"); 1648 if (Atomic::cmpxchg(next, &cur_mid_in_use->_next_om, marked_mid) 1649 != marked_mid) { 1650 // We could not switch cur_mid_in_use's next field. This 1651 // should not be possible since it was marked so we: 1652 fatal("mid=" INTPTR_FORMAT " must be referred to by the list " 1653 "head: &om_in_use_list=" INTPTR_FORMAT " or by " 1654 "cur_mid_in_use's next field: cur_mid_in_use=" INTPTR_FORMAT 1655 ", next_om=" INTPTR_FORMAT, p2i(mid), 1656 p2i((ObjectMonitor**)&self->om_in_use_list), 1657 p2i(cur_mid_in_use), p2i(cur_mid_in_use->_next_om)); 1658 } 1659 } 1660 extracted = true; 1661 Atomic::dec(&self->om_in_use_count); 1662 // Unmark mid, but leave the next value for any lagging list 1663 // walkers. It will get cleaned up when mid is prepended to 1664 // the thread's free list: 1665 set_next(mid, next); 1666 break; 1667 } 1668 if (cur_mid_in_use != NULL) { 1669 set_next(cur_mid_in_use, mid); // umark cur_mid_in_use 1670 } 1671 // The next cur_mid_in_use keeps mid's marked next field so 1672 // that it is stable for a possible next field change. It 1673 // cannot be deflated while it is marked. 1674 cur_mid_in_use = mid; 1675 mid = next; 1676 if (mid == NULL) { 1677 // Reached end of the list and didn't find m so: 1678 fatal("must find m=" INTPTR_FORMAT "on om_in_use_list=" INTPTR_FORMAT, 1679 p2i(m), p2i(self->om_in_use_list)); 1680 } 1681 // Mark mid's next field so we can possibly extract it: 1682 next = mark_next_loop(mid); 1683 } 1684 } 1685 1686 prepend_to_om_free_list(self, m); 1687 guarantee(m->is_free(), "invariant"); 1688 } 1689 1690 // Return ObjectMonitors on a moribund thread's free and in-use 1691 // lists to the appropriate global lists. The ObjectMonitors on the 1692 // per-thread in-use list may still be in use by other threads. 1693 // 1694 // We currently call om_flush() from Threads::remove() before the 1695 // thread has been excised from the thread list and is no longer a 1696 // mutator. This means that om_flush() cannot run concurrently with 1697 // a safepoint and interleave with deflate_idle_monitors(). In 1698 // particular, this ensures that the thread's in-use monitors are 1699 // scanned by a GC safepoint, either via Thread::oops_do() (before 1700 // om_flush() is called) or via ObjectSynchronizer::oops_do() (after 1701 // om_flush() is called). 1702 // 1703 // With AsyncDeflateIdleMonitors, deflate_global_idle_monitors_using_JT() 1704 // and deflate_per_thread_idle_monitors_using_JT() (in another thread) can 1705 // run at the same time as om_flush() so we have to follow a careful 1706 // protocol to prevent list corruption. 1707 1708 void ObjectSynchronizer::om_flush(Thread* self) { 1709 // This function can race with an async deflater thread. Since 1710 // deflation has to process the per-thread in-use list before 1711 // prepending the deflated ObjectMonitors to the global free list, 1712 // we process the per-thread lists in the same order to prevent 1713 // ordering races. 1714 int in_use_count = 0; 1715 ObjectMonitor* in_use_list = NULL; 1716 ObjectMonitor* in_use_tail = NULL; 1717 ObjectMonitor* next = NULL; 1718 1719 // An async deflation thread checks to see if the target thread 1720 // is exiting, but if it has made it past that check before we 1721 // started exiting, then it is racing to get to the in-use list. 1722 if (mark_list_head(&self->om_in_use_list, &in_use_list, &next)) { 1723 chk_for_list_loop(in_use_list, OrderAccess::load_acquire(&self->om_in_use_count)); 1724 // At this point, we have marked the in-use list head so an 1725 // async deflation thread cannot come in after us. If an async 1726 // deflation thread is ahead of us, then we'll detect that and 1727 // wait for it to finish its work. 1728 // 1729 // The thread is going away, however the ObjectMonitors on the 1730 // om_in_use_list may still be in-use by other threads. Link 1731 // them to in_use_tail, which will be linked into the global 1732 // in-use list g_om_in_use_list below. 1733 // 1734 // Account for the in-use list head before the loop since it is 1735 // already marked (by this thread): 1736 in_use_tail = in_use_list; 1737 in_use_count++; 1738 for (ObjectMonitor* cur_om = unmarked_next(in_use_list); cur_om != NULL;) { 1739 if (is_next_marked(cur_om)) { 1740 // This next field is marked so there must be an async deflater 1741 // thread ahead of us so we'll give it a chance to finish. 1742 while (is_next_marked(cur_om)) { 1743 os::naked_short_sleep(1); 1744 } 1745 // Refetch the possibly changed next field and try again. 1746 cur_om = unmarked_next(in_use_tail); 1747 continue; 1748 } 1749 if (!cur_om->is_active()) { 1750 // cur_om was deflated and the allocation state was changed 1751 // to Free while it was marked. We happened to see it just 1752 // after it was unmarked (and added to the free list). 1753 // Refetch the possibly changed next field and try again. 1754 cur_om = unmarked_next(in_use_tail); 1755 continue; 1756 } 1757 in_use_tail = cur_om; 1758 in_use_count++; 1759 cur_om = unmarked_next(cur_om); 1760 } 1761 guarantee(in_use_tail != NULL, "invariant"); 1762 int l_om_in_use_count = OrderAccess::load_acquire(&self->om_in_use_count); 1763 ADIM_guarantee(l_om_in_use_count == in_use_count, "in-use counts don't " 1764 "match: l_om_in_use_count=%d, in_use_count=%d", 1765 l_om_in_use_count, in_use_count); 1766 // Clear the in-use count before unmarking the in-use list head 1767 // to avoid races: 1768 OrderAccess::release_store(&self->om_in_use_count, 0); 1769 // Clear the in-use list head (which also unmarks it): 1770 OrderAccess::release_store(&self->om_in_use_list, (ObjectMonitor*)NULL); 1771 // Unmark the disconnected list head: 1772 set_next(in_use_list, next); 1773 } 1774 1775 int free_count = 0; 1776 ObjectMonitor* free_list = OrderAccess::load_acquire(&self->om_free_list); 1777 ObjectMonitor* free_tail = NULL; 1778 if (free_list != NULL) { 1779 chk_for_list_loop(free_list, OrderAccess::load_acquire(&self->om_free_count)); 1780 // The thread is going away. Set 'free_tail' to the last per-thread free 1781 // monitor which will be linked to g_free_list below. 1782 stringStream ss; 1783 for (ObjectMonitor* s = free_list; s != NULL; s = unmarked_next(s)) { 1784 free_count++; 1785 free_tail = s; 1786 guarantee(s->object() == NULL, "invariant"); 1787 guarantee(!s->is_busy(), "must be !is_busy: %s", s->is_busy_to_string(&ss)); 1788 } 1789 guarantee(free_tail != NULL, "invariant"); 1790 int l_om_free_count = OrderAccess::load_acquire(&self->om_free_count); 1791 ADIM_guarantee(l_om_free_count == free_count, "free counts don't match: " 1792 "l_om_free_count=%d, free_count=%d", l_om_free_count, 1793 free_count); 1794 OrderAccess::release_store(&self->om_free_list, (ObjectMonitor*)NULL); 1795 OrderAccess::release_store(&self->om_free_count, 0); 1796 } 1797 1798 if (free_tail != NULL) { 1799 prepend_list_to_g_free_list(free_list, free_tail, free_count); 1800 } 1801 1802 if (in_use_tail != NULL) { 1803 prepend_list_to_g_om_in_use_list(in_use_list, in_use_tail, in_use_count); 1804 } 1805 1806 LogStreamHandle(Debug, monitorinflation) lsh_debug; 1807 LogStreamHandle(Info, monitorinflation) lsh_info; 1808 LogStream* ls = NULL; 1809 if (log_is_enabled(Debug, monitorinflation)) { 1810 ls = &lsh_debug; 1811 } else if ((free_count != 0 || in_use_count != 0) && 1812 log_is_enabled(Info, monitorinflation)) { 1813 ls = &lsh_info; 1814 } 1815 if (ls != NULL) { 1816 ls->print_cr("om_flush: jt=" INTPTR_FORMAT ", free_count=%d" 1817 ", in_use_count=%d" ", om_free_provision=%d", 1818 p2i(self), free_count, in_use_count, self->om_free_provision); 1819 } 1820 } 1821 1822 static void post_monitor_inflate_event(EventJavaMonitorInflate* event, 1823 const oop obj, 1824 ObjectSynchronizer::InflateCause cause) { 1825 assert(event != NULL, "invariant"); 1826 assert(event->should_commit(), "invariant"); 1827 event->set_monitorClass(obj->klass()); 1828 event->set_address((uintptr_t)(void*)obj); 1829 event->set_cause((u1)cause); 1830 event->commit(); 1831 } 1832 1833 // Fast path code shared by multiple functions 1834 void ObjectSynchronizer::inflate_helper(ObjectMonitorHandle* omh_p, oop obj) { 1835 while (true) { 1836 markWord mark = obj->mark(); 1837 if (mark.has_monitor()) { 1838 if (!omh_p->save_om_ptr(obj, mark)) { 1839 // Lost a race with async deflation so try again. 1840 assert(AsyncDeflateIdleMonitors, "sanity check"); 1841 continue; 1842 } 1843 ObjectMonitor* monitor = omh_p->om_ptr(); 1844 assert(ObjectSynchronizer::verify_objmon_isinpool(monitor), "monitor is invalid"); 1845 markWord dmw = monitor->header(); 1846 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value()); 1847 return; 1848 } 1849 inflate(omh_p, Thread::current(), obj, inflate_cause_vm_internal); 1850 return; 1851 } 1852 } 1853 1854 void ObjectSynchronizer::inflate(ObjectMonitorHandle* omh_p, Thread* self, 1855 oop object, const InflateCause cause) { 1856 // Inflate mutates the heap ... 1857 // Relaxing assertion for bug 6320749. 1858 assert(Universe::verify_in_progress() || 1859 !SafepointSynchronize::is_at_safepoint(), "invariant"); 1860 1861 EventJavaMonitorInflate event; 1862 1863 for (;;) { 1864 const markWord mark = object->mark(); 1865 assert(!mark.has_bias_pattern(), "invariant"); 1866 1867 // The mark can be in one of the following states: 1868 // * Inflated - just return 1869 // * Stack-locked - coerce it to inflated 1870 // * INFLATING - busy wait for conversion to complete 1871 // * Neutral - aggressively inflate the object. 1872 // * BIASED - Illegal. We should never see this 1873 1874 // CASE: inflated 1875 if (mark.has_monitor()) { 1876 if (!omh_p->save_om_ptr(object, mark)) { 1877 // Lost a race with async deflation so try again. 1878 assert(AsyncDeflateIdleMonitors, "sanity check"); 1879 continue; 1880 } 1881 ObjectMonitor* inf = omh_p->om_ptr(); 1882 markWord dmw = inf->header(); 1883 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1884 assert(inf->object() == object, "invariant"); 1885 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); 1886 return; 1887 } 1888 1889 // CASE: inflation in progress - inflating over a stack-lock. 1890 // Some other thread is converting from stack-locked to inflated. 1891 // Only that thread can complete inflation -- other threads must wait. 1892 // The INFLATING value is transient. 1893 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. 1894 // We could always eliminate polling by parking the thread on some auxiliary list. 1895 if (mark == markWord::INFLATING()) { 1896 read_stable_mark(object); 1897 continue; 1898 } 1899 1900 // CASE: stack-locked 1901 // Could be stack-locked either by this thread or by some other thread. 1902 // 1903 // Note that we allocate the objectmonitor speculatively, _before_ attempting 1904 // to install INFLATING into the mark word. We originally installed INFLATING, 1905 // allocated the objectmonitor, and then finally STed the address of the 1906 // objectmonitor into the mark. This was correct, but artificially lengthened 1907 // the interval in which INFLATED appeared in the mark, thus increasing 1908 // the odds of inflation contention. 1909 // 1910 // We now use per-thread private objectmonitor free lists. 1911 // These list are reprovisioned from the global free list outside the 1912 // critical INFLATING...ST interval. A thread can transfer 1913 // multiple objectmonitors en-mass from the global free list to its local free list. 1914 // This reduces coherency traffic and lock contention on the global free list. 1915 // Using such local free lists, it doesn't matter if the om_alloc() call appears 1916 // before or after the CAS(INFLATING) operation. 1917 // See the comments in om_alloc(). 1918 1919 LogStreamHandle(Trace, monitorinflation) lsh; 1920 1921 if (mark.has_locker()) { 1922 ObjectMonitor* m = om_alloc(self, cause); 1923 // Optimistically prepare the objectmonitor - anticipate successful CAS 1924 // We do this before the CAS in order to minimize the length of time 1925 // in which INFLATING appears in the mark. 1926 m->Recycle(); 1927 m->_Responsible = NULL; 1928 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class 1929 1930 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark); 1931 if (cmp != mark) { 1932 om_release(self, m, true); 1933 continue; // Interference -- just retry 1934 } 1935 1936 // We've successfully installed INFLATING (0) into the mark-word. 1937 // This is the only case where 0 will appear in a mark-word. 1938 // Only the singular thread that successfully swings the mark-word 1939 // to 0 can perform (or more precisely, complete) inflation. 1940 // 1941 // Why do we CAS a 0 into the mark-word instead of just CASing the 1942 // mark-word from the stack-locked value directly to the new inflated state? 1943 // Consider what happens when a thread unlocks a stack-locked object. 1944 // It attempts to use CAS to swing the displaced header value from the 1945 // on-stack BasicLock back into the object header. Recall also that the 1946 // header value (hash code, etc) can reside in (a) the object header, or 1947 // (b) a displaced header associated with the stack-lock, or (c) a displaced 1948 // header in an ObjectMonitor. The inflate() routine must copy the header 1949 // value from the BasicLock on the owner's stack to the ObjectMonitor, all 1950 // the while preserving the hashCode stability invariants. If the owner 1951 // decides to release the lock while the value is 0, the unlock will fail 1952 // and control will eventually pass from slow_exit() to inflate. The owner 1953 // will then spin, waiting for the 0 value to disappear. Put another way, 1954 // the 0 causes the owner to stall if the owner happens to try to 1955 // drop the lock (restoring the header from the BasicLock to the object) 1956 // while inflation is in-progress. This protocol avoids races that might 1957 // would otherwise permit hashCode values to change or "flicker" for an object. 1958 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable. 1959 // 0 serves as a "BUSY" inflate-in-progress indicator. 1960 1961 1962 // fetch the displaced mark from the owner's stack. 1963 // The owner can't die or unwind past the lock while our INFLATING 1964 // object is in the mark. Furthermore the owner can't complete 1965 // an unlock on the object, either. 1966 markWord dmw = mark.displaced_mark_helper(); 1967 // Catch if the object's header is not neutral (not locked and 1968 // not marked is what we care about here). 1969 ADIM_guarantee(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 1970 1971 // Setup monitor fields to proper values -- prepare the monitor 1972 m->set_header(dmw); 1973 1974 // Optimization: if the mark.locker stack address is associated 1975 // with this thread we could simply set m->_owner = self. 1976 // Note that a thread can inflate an object 1977 // that it has stack-locked -- as might happen in wait() -- directly 1978 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. 1979 m->set_owner(mark.locker()); 1980 m->set_object(object); 1981 // TODO-FIXME: assert BasicLock->dhw != 0. 1982 1983 omh_p->set_om_ptr(m); 1984 assert(m->is_new(), "freshly allocated monitor must be new"); 1985 m->set_allocation_state(ObjectMonitor::Old); 1986 1987 // Must preserve store ordering. The monitor state must 1988 // be stable at the time of publishing the monitor address. 1989 guarantee(object->mark() == markWord::INFLATING(), "invariant"); 1990 object->release_set_mark(markWord::encode(m)); 1991 1992 // Hopefully the performance counters are allocated on distinct cache lines 1993 // to avoid false sharing on MP systems ... 1994 OM_PERFDATA_OP(Inflations, inc()); 1995 if (log_is_enabled(Trace, monitorinflation)) { 1996 ResourceMark rm(self); 1997 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark=" 1998 INTPTR_FORMAT ", type='%s'", p2i(object), 1999 object->mark().value(), object->klass()->external_name()); 2000 } 2001 if (event.should_commit()) { 2002 post_monitor_inflate_event(&event, object, cause); 2003 } 2004 ADIM_guarantee(!m->is_free(), "inflated monitor to be returned cannot be free"); 2005 return; 2006 } 2007 2008 // CASE: neutral 2009 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. 2010 // If we know we're inflating for entry it's better to inflate by swinging a 2011 // pre-locked ObjectMonitor pointer into the object header. A successful 2012 // CAS inflates the object *and* confers ownership to the inflating thread. 2013 // In the current implementation we use a 2-step mechanism where we CAS() 2014 // to inflate and then CAS() again to try to swing _owner from NULL to self. 2015 // An inflateTry() method that we could call from enter() would be useful. 2016 2017 // Catch if the object's header is not neutral (not locked and 2018 // not marked is what we care about here). 2019 ADIM_guarantee(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT,mark.value()); 2020 ObjectMonitor* m = om_alloc(self, cause); 2021 // prepare m for installation - set monitor to initial state 2022 m->Recycle(); 2023 m->set_header(mark); 2024 // If we leave _owner == DEFLATER_MARKER here, then the simple C2 2025 // ObjectMonitor enter optimization can no longer race with async 2026 // deflation and reuse. 2027 m->set_object(object); 2028 m->_Responsible = NULL; 2029 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class 2030 2031 omh_p->set_om_ptr(m); 2032 assert(m->is_new(), "freshly allocated monitor must be new"); 2033 m->set_allocation_state(ObjectMonitor::Old); 2034 2035 if (object->cas_set_mark(markWord::encode(m), mark) != mark) { 2036 guarantee(!m->owner_is_DEFLATER_MARKER() || m->ref_count() >= 0, 2037 "race between deflation and om_release() with m=" INTPTR_FORMAT 2038 ", _owner=" INTPTR_FORMAT ", ref_count=%d", p2i(m), 2039 p2i(m->_owner), m->ref_count()); 2040 m->set_header(markWord::zero()); 2041 m->set_object(NULL); 2042 m->Recycle(); 2043 omh_p->set_om_ptr(NULL); 2044 // om_release() will reset the allocation state 2045 om_release(self, m, true); 2046 m = NULL; 2047 continue; 2048 // interference - the markword changed - just retry. 2049 // The state-transitions are one-way, so there's no chance of 2050 // live-lock -- "Inflated" is an absorbing state. 2051 } 2052 2053 // Hopefully the performance counters are allocated on distinct 2054 // cache lines to avoid false sharing on MP systems ... 2055 OM_PERFDATA_OP(Inflations, inc()); 2056 if (log_is_enabled(Trace, monitorinflation)) { 2057 ResourceMark rm(self); 2058 lsh.print_cr("inflate(neutral): object=" INTPTR_FORMAT ", mark=" 2059 INTPTR_FORMAT ", type='%s'", p2i(object), 2060 object->mark().value(), object->klass()->external_name()); 2061 } 2062 if (event.should_commit()) { 2063 post_monitor_inflate_event(&event, object, cause); 2064 } 2065 ADIM_guarantee(!m->is_free(), "inflated monitor to be returned cannot be free"); 2066 return; 2067 } 2068 } 2069 2070 2071 // We maintain a list of in-use monitors for each thread. 2072 // 2073 // For safepoint based deflation: 2074 // deflate_thread_local_monitors() scans a single thread's in-use list, while 2075 // deflate_idle_monitors() scans only a global list of in-use monitors which 2076 // is populated only as a thread dies (see om_flush()). 2077 // 2078 // These operations are called at all safepoints, immediately after mutators 2079 // are stopped, but before any objects have moved. Collectively they traverse 2080 // the population of in-use monitors, deflating where possible. The scavenged 2081 // monitors are returned to the global monitor free list. 2082 // 2083 // Beware that we scavenge at *every* stop-the-world point. Having a large 2084 // number of monitors in-use could negatively impact performance. We also want 2085 // to minimize the total # of monitors in circulation, as they incur a small 2086 // footprint penalty. 2087 // 2088 // Perversely, the heap size -- and thus the STW safepoint rate -- 2089 // typically drives the scavenge rate. Large heaps can mean infrequent GC, 2090 // which in turn can mean large(r) numbers of ObjectMonitors in circulation. 2091 // This is an unfortunate aspect of this design. 2092 // 2093 // For async deflation: 2094 // If a special deflation request is made, then the safepoint based 2095 // deflation mechanism is used. Otherwise, an async deflation request 2096 // is registered with the ServiceThread and it is notified. 2097 2098 void ObjectSynchronizer::do_safepoint_work(DeflateMonitorCounters* counters) { 2099 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 2100 2101 // The per-thread in-use lists are handled in 2102 // ParallelSPCleanupThreadClosure::do_thread(). 2103 2104 if (!AsyncDeflateIdleMonitors || is_special_deflation_requested()) { 2105 // Use the older mechanism for the global in-use list or if a 2106 // special deflation has been requested before the safepoint. 2107 ObjectSynchronizer::deflate_idle_monitors(counters); 2108 return; 2109 } 2110 2111 log_debug(monitorinflation)("requesting async deflation of idle monitors."); 2112 // Request deflation of idle monitors by the ServiceThread: 2113 set_is_async_deflation_requested(true); 2114 MonitorLocker ml(Service_lock, Mutex::_no_safepoint_check_flag); 2115 ml.notify_all(); 2116 } 2117 2118 // Deflate a single monitor if not in-use 2119 // Return true if deflated, false if in-use 2120 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, 2121 ObjectMonitor** free_head_p, 2122 ObjectMonitor** free_tail_p) { 2123 bool deflated; 2124 // Normal case ... The monitor is associated with obj. 2125 const markWord mark = obj->mark(); 2126 guarantee(mark == markWord::encode(mid), "should match: mark=" 2127 INTPTR_FORMAT ", encoded mid=" INTPTR_FORMAT, mark.value(), 2128 markWord::encode(mid).value()); 2129 // Make sure that mark.monitor() and markWord::encode() agree: 2130 guarantee(mark.monitor() == mid, "should match: monitor()=" INTPTR_FORMAT 2131 ", mid=" INTPTR_FORMAT, p2i(mark.monitor()), p2i(mid)); 2132 const markWord dmw = mid->header(); 2133 guarantee(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value()); 2134 2135 if (mid->is_busy() || mid->ref_count() != 0) { 2136 // Easy checks are first - the ObjectMonitor is busy or ObjectMonitor* 2137 // is in use so no deflation. 2138 deflated = false; 2139 } else { 2140 // Deflate the monitor if it is no longer being used 2141 // It's idle - scavenge and return to the global free list 2142 // plain old deflation ... 2143 if (log_is_enabled(Trace, monitorinflation)) { 2144 ResourceMark rm; 2145 log_trace(monitorinflation)("deflate_monitor: " 2146 "object=" INTPTR_FORMAT ", mark=" 2147 INTPTR_FORMAT ", type='%s'", p2i(obj), 2148 mark.value(), obj->klass()->external_name()); 2149 } 2150 2151 // Restore the header back to obj 2152 obj->release_set_mark(dmw); 2153 if (AsyncDeflateIdleMonitors) { 2154 // clear() expects the owner field to be NULL and we won't race 2155 // with the simple C2 ObjectMonitor enter optimization since 2156 // we're at a safepoint. 2157 mid->set_owner(NULL); 2158 } 2159 mid->clear(); 2160 2161 assert(mid->object() == NULL, "invariant: object=" INTPTR_FORMAT, 2162 p2i(mid->object())); 2163 assert(mid->is_free(), "invariant"); 2164 2165 // Move the deflated ObjectMonitor to the working free list 2166 // defined by free_head_p and free_tail_p. No races on this list 2167 // so no need for load_acquire() or store_release(). 2168 if (*free_head_p == NULL) *free_head_p = mid; 2169 if (*free_tail_p != NULL) { 2170 // We append to the list so the caller can use mid->_next_om 2171 // to fix the linkages in its context. 2172 ObjectMonitor* prevtail = *free_tail_p; 2173 // Should have been cleaned up by the caller: 2174 // Note: Should not have to mark prevtail here since we're at a 2175 // safepoint and ObjectMonitors on the local free list should 2176 // not be accessed in parallel. 2177 assert(prevtail->_next_om == NULL, "must be NULL: _next_om=" 2178 INTPTR_FORMAT, p2i(prevtail->_next_om)); 2179 set_next(prevtail, mid); 2180 } 2181 *free_tail_p = mid; 2182 // At this point, mid->_next_om still refers to its current 2183 // value and another ObjectMonitor's _next_om field still 2184 // refers to this ObjectMonitor. Those linkages have to be 2185 // cleaned up by the caller who has the complete context. 2186 deflated = true; 2187 } 2188 return deflated; 2189 } 2190 2191 // Deflate the specified ObjectMonitor if not in-use using a JavaThread. 2192 // Returns true if it was deflated and false otherwise. 2193 // 2194 // The async deflation protocol sets owner to DEFLATER_MARKER and 2195 // makes ref_count negative as signals to contending threads that 2196 // an async deflation is in progress. There are a number of checks 2197 // as part of the protocol to make sure that the calling thread has 2198 // not lost the race to a contending thread or to a thread that just 2199 // wants to use the ObjectMonitor*. 2200 // 2201 // The ObjectMonitor has been successfully async deflated when: 2202 // (owner == DEFLATER_MARKER && ref_count < 0) 2203 // Contending threads or ObjectMonitor* using threads that see those 2204 // values know to retry their operation. 2205 // 2206 bool ObjectSynchronizer::deflate_monitor_using_JT(ObjectMonitor* mid, 2207 ObjectMonitor** free_head_p, 2208 ObjectMonitor** free_tail_p) { 2209 assert(AsyncDeflateIdleMonitors, "sanity check"); 2210 assert(Thread::current()->is_Java_thread(), "precondition"); 2211 // A newly allocated ObjectMonitor should not be seen here so we 2212 // avoid an endless inflate/deflate cycle. 2213 assert(mid->is_old(), "must be old: allocation_state=%d", 2214 (int) mid->allocation_state()); 2215 2216 if (mid->is_busy() || mid->ref_count() != 0) { 2217 // Easy checks are first - the ObjectMonitor is busy or ObjectMonitor* 2218 // is in use so no deflation. 2219 return false; 2220 } 2221 2222 if (Atomic::replace_if_null(DEFLATER_MARKER, &(mid->_owner))) { 2223 // ObjectMonitor is not owned by another thread. Our setting 2224 // owner to DEFLATER_MARKER forces any contending thread through 2225 // the slow path. This is just the first part of the async 2226 // deflation dance. 2227 2228 if (mid->_contentions != 0 || mid->_waiters != 0) { 2229 // Another thread has raced to enter the ObjectMonitor after 2230 // mid->is_busy() above or has already entered and waited on 2231 // it which makes it busy so no deflation. Restore owner to 2232 // NULL if it is still DEFLATER_MARKER. 2233 Atomic::cmpxchg((void*)NULL, &mid->_owner, DEFLATER_MARKER); 2234 return false; 2235 } 2236 2237 if (Atomic::cmpxchg(-max_jint, &mid->_ref_count, (jint)0) == 0) { 2238 // Make ref_count negative to force any contending threads or 2239 // ObjectMonitor* using threads to retry. This is the second 2240 // part of the async deflation dance. 2241 2242 if (mid->owner_is_DEFLATER_MARKER()) { 2243 // If owner is still DEFLATER_MARKER, then we have successfully 2244 // signaled any contending threads to retry. If it is not, then we 2245 // have lost the race to an entering thread and the ObjectMonitor 2246 // is now busy. This is the third and final part of the async 2247 // deflation dance. 2248 // Note: This owner check solves the ABA problem with ref_count 2249 // where another thread acquired the ObjectMonitor, finished 2250 // using it and restored the ref_count to zero. 2251 2252 // Sanity checks for the races: 2253 guarantee(mid->_contentions == 0, "must be 0: contentions=%d", 2254 mid->_contentions); 2255 guarantee(mid->_waiters == 0, "must be 0: waiters=%d", mid->_waiters); 2256 guarantee(mid->_cxq == NULL, "must be no contending threads: cxq=" 2257 INTPTR_FORMAT, p2i(mid->_cxq)); 2258 guarantee(mid->_EntryList == NULL, 2259 "must be no entering threads: EntryList=" INTPTR_FORMAT, 2260 p2i(mid->_EntryList)); 2261 2262 const oop obj = (oop) mid->object(); 2263 if (log_is_enabled(Trace, monitorinflation)) { 2264 ResourceMark rm; 2265 log_trace(monitorinflation)("deflate_monitor_using_JT: " 2266 "object=" INTPTR_FORMAT ", mark=" 2267 INTPTR_FORMAT ", type='%s'", 2268 p2i(obj), obj->mark().value(), 2269 obj->klass()->external_name()); 2270 } 2271 2272 // Install the old mark word if nobody else has already done it. 2273 mid->install_displaced_markword_in_object(obj); 2274 mid->clear_using_JT(); 2275 2276 assert(mid->object() == NULL, "must be NULL: object=" INTPTR_FORMAT, 2277 p2i(mid->object())); 2278 assert(mid->is_free(), "must be free: allocation_state=%d", 2279 (int) mid->allocation_state()); 2280 2281 // Move the deflated ObjectMonitor to the working free list 2282 // defined by free_head_p and free_tail_p. No races on this list 2283 // so no need for load_acquire() or store_release(). 2284 if (*free_head_p == NULL) { 2285 // First one on the list. 2286 *free_head_p = mid; 2287 } 2288 if (*free_tail_p != NULL) { 2289 // We append to the list so the caller can use mid->_next_om 2290 // to fix the linkages in its context. 2291 ObjectMonitor* prevtail = *free_tail_p; 2292 // Should have been cleaned up by the caller: 2293 ObjectMonitor* next = mark_next_loop(prevtail); 2294 assert(unmarked_next(prevtail) == NULL, "must be NULL: _next_om=" 2295 INTPTR_FORMAT, p2i(unmarked_next(prevtail))); 2296 set_next(prevtail, mid); // prevtail now points to mid (and is unmarked) 2297 } 2298 *free_tail_p = mid; 2299 2300 // At this point, mid->_next_om still refers to its current 2301 // value and another ObjectMonitor's _next_om field still 2302 // refers to this ObjectMonitor. Those linkages have to be 2303 // cleaned up by the caller who has the complete context. 2304 2305 // We leave owner == DEFLATER_MARKER and ref_count < 0 2306 // to force any racing threads to retry. 2307 return true; // Success, ObjectMonitor has been deflated. 2308 } 2309 2310 // The owner was changed from DEFLATER_MARKER so we lost the 2311 // race since the ObjectMonitor is now busy. 2312 2313 // Add back max_jint to restore the ref_count field to its 2314 // proper value (which may not be what we saw above): 2315 Atomic::add(max_jint, &mid->_ref_count); 2316 2317 assert(mid->ref_count() >= 0, "must not be negative: ref_count=%d", 2318 mid->ref_count()); 2319 return false; 2320 } 2321 2322 // The ref_count was no longer 0 so we lost the race since the 2323 // ObjectMonitor is now busy or the ObjectMonitor* is now is use. 2324 // Restore owner to NULL if it is still DEFLATER_MARKER: 2325 Atomic::cmpxchg((void*)NULL, &mid->_owner, DEFLATER_MARKER); 2326 } 2327 2328 // The owner field is no longer NULL so we lost the race since the 2329 // ObjectMonitor is now busy. 2330 return false; 2331 } 2332 2333 // Walk a given monitor list, and deflate idle monitors. 2334 // The given list could be a per-thread list or a global list. 2335 // 2336 // In the case of parallel processing of thread local monitor lists, 2337 // work is done by Threads::parallel_threads_do() which ensures that 2338 // each Java thread is processed by exactly one worker thread, and 2339 // thus avoid conflicts that would arise when worker threads would 2340 // process the same monitor lists concurrently. 2341 // 2342 // See also ParallelSPCleanupTask and 2343 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and 2344 // Threads::parallel_java_threads_do() in thread.cpp. 2345 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor* volatile * list_p, 2346 int volatile * count_p, 2347 ObjectMonitor** free_head_p, 2348 ObjectMonitor** free_tail_p) { 2349 ObjectMonitor* cur_mid_in_use = NULL; 2350 ObjectMonitor* mid = NULL; 2351 ObjectMonitor* next = NULL; 2352 int deflated_count = 0; 2353 2354 // We use the simpler mark-mid-as-we-go protocol since there are no 2355 // parallel list deletions since we are at a safepoint. 2356 if (!mark_list_head(list_p, &mid, &next)) { 2357 return 0; // The list is empty so nothing to deflate. 2358 } 2359 2360 while (true) { 2361 oop obj = (oop) mid->object(); 2362 if (obj != NULL && deflate_monitor(mid, obj, free_head_p, free_tail_p)) { 2363 // Deflation succeeded and already updated free_head_p and 2364 // free_tail_p as needed. Finish the move to the local free list 2365 // by unlinking mid from the global or per-thread in-use list. 2366 if (Atomic::cmpxchg(next, list_p, mid) != mid) { 2367 // We could not switch the list head to next. 2368 ADIM_guarantee(cur_mid_in_use != NULL, "must not be NULL"); 2369 if (Atomic::cmpxchg(next, &cur_mid_in_use->_next_om, mid) != mid) { 2370 // deflate_monitor_list() is called at a safepoint so the 2371 // global or per-thread in-use list should not be modified 2372 // in parallel so we: 2373 fatal("mid=" INTPTR_FORMAT " must be referred to by the list head: " 2374 "list_p=" INTPTR_FORMAT " or by cur_mid_in_use's next field: " 2375 "cur_mid_in_use=" INTPTR_FORMAT ", next_om=" INTPTR_FORMAT, 2376 p2i(mid), p2i((ObjectMonitor**)list_p), p2i(cur_mid_in_use), 2377 p2i(cur_mid_in_use->_next_om)); 2378 } 2379 } 2380 // At this point mid is disconnected from the in-use list so 2381 // its marked next field no longer has any effects. 2382 deflated_count++; 2383 Atomic::dec(count_p); 2384 chk_for_list_loop(OrderAccess::load_acquire(list_p), 2385 OrderAccess::load_acquire(count_p)); 2386 chk_om_not_on_list(mid, OrderAccess::load_acquire(list_p), 2387 OrderAccess::load_acquire(count_p)); 2388 // mid is current tail in the free_head_p list so NULL terminate it 2389 // (which also unmarks it): 2390 set_next(mid, NULL); 2391 2392 // All the list management is done so move on to the next one: 2393 mid = next; 2394 } else { 2395 set_next(mid, next); // unmark next field 2396 2397 // All the list management is done so move on to the next one: 2398 cur_mid_in_use = mid; 2399 mid = next; 2400 } 2401 if (mid == NULL) { 2402 break; // Reached end of the list so nothing more to deflate. 2403 } 2404 // Mark mid's next field so we can possibly deflate it: 2405 next = mark_next_loop(mid); 2406 } 2407 return deflated_count; 2408 } 2409 2410 // Walk a given ObjectMonitor list and deflate idle ObjectMonitors using 2411 // a JavaThread. Returns the number of deflated ObjectMonitors. The given 2412 // list could be a per-thread in-use list or the global in-use list. 2413 // If a safepoint has started, then we save state via saved_mid_in_use_p 2414 // and return to the caller to honor the safepoint. 2415 // 2416 int ObjectSynchronizer::deflate_monitor_list_using_JT(ObjectMonitor* volatile * list_p, 2417 int volatile * count_p, 2418 ObjectMonitor** free_head_p, 2419 ObjectMonitor** free_tail_p, 2420 ObjectMonitor** saved_mid_in_use_p) { 2421 assert(AsyncDeflateIdleMonitors, "sanity check"); 2422 assert(Thread::current()->is_Java_thread(), "precondition"); 2423 2424 ObjectMonitor* cur_mid_in_use = NULL; 2425 ObjectMonitor* mid = NULL; 2426 ObjectMonitor* next = NULL; 2427 ObjectMonitor* next_next = NULL; 2428 int deflated_count = 0; 2429 2430 // We use the more complicated mark-cur_mid_in_use-and-mid-as-we-go 2431 // protocol because om_release() can do list deletions in parallel. 2432 // We also mark-next-next-as-we-go to prevent an om_flush() that is 2433 // behind this thread from passing us. 2434 if (*saved_mid_in_use_p == NULL) { 2435 // No saved state so start at the beginning. 2436 // Mark the list head's next field so we can possibly deflate it: 2437 if (!mark_list_head(list_p, &mid, &next)) { 2438 return 0; // The list is empty so nothing to deflate. 2439 } 2440 } else { 2441 // We're restarting after a safepoint so restore the necessary state 2442 // before we resume. 2443 cur_mid_in_use = *saved_mid_in_use_p; 2444 // Mark cur_mid_in_use's next field so we can possibly update its 2445 // next field to extract a deflated ObjectMonitor. 2446 mid = mark_next_loop(cur_mid_in_use); 2447 if (mid == NULL) { 2448 set_next(cur_mid_in_use, NULL); // unmark next field 2449 *saved_mid_in_use_p = NULL; 2450 return 0; // The remainder is empty so nothing more to deflate. 2451 } 2452 // Mark mid's next field so we can possibly deflate it: 2453 next = mark_next_loop(mid); 2454 } 2455 2456 while (true) { 2457 // The current mid's next field is marked at this point. If we have 2458 // a cur_mid_in_use, then its next field is also marked at this point. 2459 2460 if (next != NULL) { 2461 // We mark the next -> next field so that an om_flush() 2462 // thread that is behind us cannot pass us when we 2463 // unmark the current mid's next field. 2464 next_next = mark_next_loop(next); 2465 } 2466 2467 // Only try to deflate if there is an associated Java object and if 2468 // mid is old (is not newly allocated and is not newly freed). 2469 if (mid->object() != NULL && mid->is_old() && 2470 deflate_monitor_using_JT(mid, free_head_p, free_tail_p)) { 2471 // Deflation succeeded and already updated free_head_p and 2472 // free_tail_p as needed. Finish the move to the local free list 2473 // by unlinking mid from the global or per-thread in-use list. 2474 if (Atomic::cmpxchg(next, list_p, mid) != mid) { 2475 // We could not switch the list head to next. 2476 ObjectMonitor* marked_mid = mark_om_ptr(mid); 2477 ObjectMonitor* marked_next = mark_om_ptr(next); 2478 // Switch cur_mid_in_use's next field to marked next: 2479 ADIM_guarantee(cur_mid_in_use != NULL, "must not be NULL"); 2480 if (Atomic::cmpxchg(marked_next, &cur_mid_in_use->_next_om, 2481 marked_mid) != marked_mid) { 2482 // We could not switch cur_mid_in_use's next field. This 2483 // should not be possible since it was marked so we: 2484 fatal("mid=" INTPTR_FORMAT " must be referred to by the list head: " 2485 "&list_p=" INTPTR_FORMAT " or by cur_mid_in_use's next field: " 2486 "cur_mid_in_use=" INTPTR_FORMAT ", next_om=" INTPTR_FORMAT, 2487 p2i(mid), p2i((ObjectMonitor**)list_p), p2i(cur_mid_in_use), 2488 p2i(cur_mid_in_use->_next_om)); 2489 } 2490 } 2491 // At this point mid is disconnected from the in-use list so 2492 // its marked next field no longer has any effects. 2493 deflated_count++; 2494 Atomic::dec(count_p); 2495 chk_for_list_loop(OrderAccess::load_acquire(list_p), 2496 OrderAccess::load_acquire(count_p)); 2497 chk_om_not_on_list(mid, OrderAccess::load_acquire(list_p), 2498 OrderAccess::load_acquire(count_p)); 2499 // mid is current tail in the free_head_p list so NULL terminate it 2500 // (which also unmarks it): 2501 set_next(mid, NULL); 2502 2503 // All the list management is done so move on to the next one: 2504 mid = next; // mid keeps non-NULL next's marked next field 2505 next = next_next; 2506 } else { 2507 // mid is considered in-use if it does not have an associated 2508 // Java object or mid is not old or deflation did not succeed. 2509 // A mid->is_new() node can be seen here when it is freshly 2510 // returned by om_alloc() (and skips the deflation code path). 2511 // A mid->is_old() node can be seen here when deflation failed. 2512 // A mid->is_free() node can be seen here when a fresh node from 2513 // om_alloc() is released by om_release() due to losing the race 2514 // in inflate(). 2515 2516 // All the list management is done so move on to the next one: 2517 if (cur_mid_in_use != NULL) { 2518 set_next(cur_mid_in_use, mid); // umark cur_mid_in_use 2519 } 2520 // The next cur_mid_in_use keeps mid's marked next field so 2521 // that it is stable for a possible next field change. It 2522 // cannot be modified by om_release() while it is marked. 2523 cur_mid_in_use = mid; 2524 mid = next; // mid keeps non-NULL next's marked next field 2525 next = next_next; 2526 2527 if (SafepointSynchronize::is_synchronizing() && 2528 cur_mid_in_use != OrderAccess::load_acquire(list_p) && 2529 cur_mid_in_use->is_old()) { 2530 // If a safepoint has started and cur_mid_in_use is not the list 2531 // head and is old, then it is safe to use as saved state. Return 2532 // to the caller before blocking. 2533 *saved_mid_in_use_p = cur_mid_in_use; 2534 set_next(cur_mid_in_use, mid); // umark cur_mid_in_use 2535 if (mid != NULL) { 2536 set_next(mid, next); // umark mid 2537 } 2538 return deflated_count; 2539 } 2540 } 2541 if (mid == NULL) { 2542 if (cur_mid_in_use != NULL) { 2543 set_next(cur_mid_in_use, mid); // umark cur_mid_in_use 2544 } 2545 break; // Reached end of the list so nothing more to deflate. 2546 } 2547 2548 // The current mid's next field is marked at this point. If we have 2549 // a cur_mid_in_use, then its next field is also marked at this point. 2550 } 2551 // We finished the list without a safepoint starting so there's 2552 // no need to save state. 2553 *saved_mid_in_use_p = NULL; 2554 return deflated_count; 2555 } 2556 2557 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) { 2558 OrderAccess::release_store(&counters->n_in_use, 0); // currently associated with objects 2559 OrderAccess::release_store(&counters->n_in_circulation, 0); // extant 2560 OrderAccess::release_store(&counters->n_scavenged, 0); // reclaimed (global and per-thread) 2561 OrderAccess::release_store(&counters->per_thread_scavenged, 0); // per-thread scavenge total 2562 counters->per_thread_times = 0.0; // per-thread scavenge times 2563 } 2564 2565 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) { 2566 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 2567 2568 if (AsyncDeflateIdleMonitors) { 2569 // Nothing to do when global idle ObjectMonitors are deflated using 2570 // a JavaThread unless a special deflation has been requested. 2571 if (!is_special_deflation_requested()) { 2572 return; 2573 } 2574 } 2575 2576 bool deflated = false; 2577 2578 ObjectMonitor* free_head_p = NULL; // Local SLL of scavenged monitors 2579 ObjectMonitor* free_tail_p = NULL; 2580 elapsedTimer timer; 2581 2582 if (log_is_enabled(Info, monitorinflation)) { 2583 timer.start(); 2584 } 2585 2586 // Note: the thread-local monitors lists get deflated in 2587 // a separate pass. See deflate_thread_local_monitors(). 2588 2589 // For moribund threads, scan g_om_in_use_list 2590 int deflated_count = 0; 2591 if (OrderAccess::load_acquire(&g_om_in_use_list) != NULL) { 2592 // Update n_in_circulation before g_om_in_use_count is updated by deflation. 2593 Atomic::add(OrderAccess::load_acquire(&g_om_in_use_count), &counters->n_in_circulation); 2594 2595 deflated_count = deflate_monitor_list(&g_om_in_use_list, &g_om_in_use_count, &free_head_p, &free_tail_p); 2596 Atomic::add(OrderAccess::load_acquire(&g_om_in_use_count), &counters->n_in_use); 2597 } 2598 2599 if (free_head_p != NULL) { 2600 // Move the deflated ObjectMonitors back to the global free list. 2601 // No races on the working free list so no need for load_acquire(). 2602 guarantee(free_tail_p != NULL && deflated_count > 0, "invariant"); 2603 assert(free_tail_p->_next_om == NULL, "must be NULL: _next_om=" 2604 INTPTR_FORMAT, p2i(free_tail_p->_next_om)); 2605 prepend_list_to_g_free_list(free_head_p, free_tail_p, deflated_count); 2606 Atomic::add(deflated_count, &counters->n_scavenged); 2607 } 2608 timer.stop(); 2609 2610 LogStreamHandle(Debug, monitorinflation) lsh_debug; 2611 LogStreamHandle(Info, monitorinflation) lsh_info; 2612 LogStream* ls = NULL; 2613 if (log_is_enabled(Debug, monitorinflation)) { 2614 ls = &lsh_debug; 2615 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) { 2616 ls = &lsh_info; 2617 } 2618 if (ls != NULL) { 2619 ls->print_cr("deflating global idle monitors, %3.7f secs, %d monitors", timer.seconds(), deflated_count); 2620 } 2621 } 2622 2623 // Deflate global idle ObjectMonitors using a JavaThread. 2624 // 2625 void ObjectSynchronizer::deflate_global_idle_monitors_using_JT() { 2626 assert(AsyncDeflateIdleMonitors, "sanity check"); 2627 assert(Thread::current()->is_Java_thread(), "precondition"); 2628 JavaThread* self = JavaThread::current(); 2629 2630 deflate_common_idle_monitors_using_JT(true /* is_global */, self); 2631 } 2632 2633 // Deflate the specified JavaThread's idle ObjectMonitors using a JavaThread. 2634 // 2635 void ObjectSynchronizer::deflate_per_thread_idle_monitors_using_JT(JavaThread* target) { 2636 assert(AsyncDeflateIdleMonitors, "sanity check"); 2637 assert(Thread::current()->is_Java_thread(), "precondition"); 2638 2639 deflate_common_idle_monitors_using_JT(false /* !is_global */, target); 2640 } 2641 2642 // Deflate global or per-thread idle ObjectMonitors using a JavaThread. 2643 // 2644 void ObjectSynchronizer::deflate_common_idle_monitors_using_JT(bool is_global, JavaThread* target) { 2645 JavaThread* self = JavaThread::current(); 2646 2647 int deflated_count = 0; 2648 ObjectMonitor* free_head_p = NULL; // Local SLL of scavenged ObjectMonitors 2649 ObjectMonitor* free_tail_p = NULL; 2650 ObjectMonitor* saved_mid_in_use_p = NULL; 2651 elapsedTimer timer; 2652 2653 if (log_is_enabled(Info, monitorinflation)) { 2654 timer.start(); 2655 } 2656 2657 if (is_global) { 2658 OM_PERFDATA_OP(MonExtant, set_value(OrderAccess::load_acquire(&g_om_in_use_count))); 2659 } else { 2660 OM_PERFDATA_OP(MonExtant, inc(OrderAccess::load_acquire(&target->om_in_use_count))); 2661 } 2662 2663 do { 2664 int local_deflated_count; 2665 if (is_global) { 2666 local_deflated_count = deflate_monitor_list_using_JT(&g_om_in_use_list, &g_om_in_use_count, &free_head_p, &free_tail_p, &saved_mid_in_use_p); 2667 } else { 2668 local_deflated_count = deflate_monitor_list_using_JT(&target->om_in_use_list, &target->om_in_use_count, &free_head_p, &free_tail_p, &saved_mid_in_use_p); 2669 } 2670 deflated_count += local_deflated_count; 2671 2672 if (free_head_p != NULL) { 2673 // Move the deflated ObjectMonitors to the global free list. 2674 // No races on the working list so no need for load_acquire(). 2675 guarantee(free_tail_p != NULL && local_deflated_count > 0, "free_tail_p=" INTPTR_FORMAT ", local_deflated_count=%d", p2i(free_tail_p), local_deflated_count); 2676 // Note: The target thread can be doing an om_alloc() that 2677 // is trying to prepend an ObjectMonitor on its in-use list 2678 // at the same time that we have deflated the current in-use 2679 // list head and put it on the local free list. prepend_to_common() 2680 // will detect the race and retry which avoids list corruption, 2681 // but the next field in free_tail_p can flicker to marked 2682 // and then unmarked while prepend_to_common() is sorting it 2683 // all out. 2684 assert(unmarked_next(free_tail_p) == NULL, "must be NULL: _next_om=" 2685 INTPTR_FORMAT, p2i(unmarked_next(free_tail_p))); 2686 2687 prepend_list_to_g_free_list(free_head_p, free_tail_p, local_deflated_count); 2688 2689 OM_PERFDATA_OP(Deflations, inc(local_deflated_count)); 2690 } 2691 2692 if (saved_mid_in_use_p != NULL) { 2693 // deflate_monitor_list_using_JT() detected a safepoint starting. 2694 timer.stop(); 2695 { 2696 if (is_global) { 2697 log_debug(monitorinflation)("pausing deflation of global idle monitors for a safepoint."); 2698 } else { 2699 log_debug(monitorinflation)("jt=" INTPTR_FORMAT ": pausing deflation of per-thread idle monitors for a safepoint.", p2i(target)); 2700 } 2701 assert(SafepointSynchronize::is_synchronizing(), "sanity check"); 2702 ThreadBlockInVM blocker(self); 2703 } 2704 // Prepare for another loop after the safepoint. 2705 free_head_p = NULL; 2706 free_tail_p = NULL; 2707 if (log_is_enabled(Info, monitorinflation)) { 2708 timer.start(); 2709 } 2710 } 2711 } while (saved_mid_in_use_p != NULL); 2712 timer.stop(); 2713 2714 LogStreamHandle(Debug, monitorinflation) lsh_debug; 2715 LogStreamHandle(Info, monitorinflation) lsh_info; 2716 LogStream* ls = NULL; 2717 if (log_is_enabled(Debug, monitorinflation)) { 2718 ls = &lsh_debug; 2719 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) { 2720 ls = &lsh_info; 2721 } 2722 if (ls != NULL) { 2723 if (is_global) { 2724 ls->print_cr("async-deflating global idle monitors, %3.7f secs, %d monitors", timer.seconds(), deflated_count); 2725 } else { 2726 ls->print_cr("jt=" INTPTR_FORMAT ": async-deflating per-thread idle monitors, %3.7f secs, %d monitors", p2i(target), timer.seconds(), deflated_count); 2727 } 2728 } 2729 } 2730 2731 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) { 2732 // Report the cumulative time for deflating each thread's idle 2733 // monitors. Note: if the work is split among more than one 2734 // worker thread, then the reported time will likely be more 2735 // than a beginning to end measurement of the phase. 2736 // Note: AsyncDeflateIdleMonitors only deflates per-thread idle 2737 // monitors at a safepoint when a special deflation has been requested. 2738 log_info(safepoint, cleanup)("deflating per-thread idle monitors, %3.7f secs, monitors=%d", 2739 counters->per_thread_times, 2740 OrderAccess::load_acquire(&counters->per_thread_scavenged)); 2741 2742 bool needs_special_deflation = is_special_deflation_requested(); 2743 if (!AsyncDeflateIdleMonitors || needs_special_deflation) { 2744 // AsyncDeflateIdleMonitors does not use these counters unless 2745 // there is a special deflation request. 2746 2747 OM_PERFDATA_OP(Deflations, inc(counters->n_scavenged)); 2748 OM_PERFDATA_OP(MonExtant, set_value(counters->n_in_circulation)); 2749 } 2750 2751 if (log_is_enabled(Debug, monitorinflation)) { 2752 // exit_globals()'s call to audit_and_print_stats() is done 2753 // at the Info level. 2754 ObjectSynchronizer::audit_and_print_stats(false /* on_exit */); 2755 } else if (log_is_enabled(Info, monitorinflation)) { 2756 log_info(monitorinflation)("g_om_population=%d, g_om_in_use_count=%d, " 2757 "g_om_free_count=%d", 2758 OrderAccess::load_acquire(&g_om_population), 2759 OrderAccess::load_acquire(&g_om_in_use_count), 2760 OrderAccess::load_acquire(&g_om_free_count)); 2761 } 2762 2763 ForceMonitorScavenge = 0; // Reset 2764 GVars.stw_random = os::random(); 2765 GVars.stw_cycle++; 2766 if (needs_special_deflation) { 2767 set_is_special_deflation_requested(false); // special deflation is done 2768 } 2769 } 2770 2771 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) { 2772 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); 2773 2774 if (AsyncDeflateIdleMonitors && !is_special_deflation_requested()) { 2775 // Nothing to do if a special deflation has NOT been requested. 2776 return; 2777 } 2778 2779 ObjectMonitor* free_head_p = NULL; // Local SLL of scavenged monitors 2780 ObjectMonitor* free_tail_p = NULL; 2781 elapsedTimer timer; 2782 2783 if (log_is_enabled(Info, safepoint, cleanup) || 2784 log_is_enabled(Info, monitorinflation)) { 2785 timer.start(); 2786 } 2787 2788 // Update n_in_circulation before om_in_use_count is updated by deflation. 2789 Atomic::add(OrderAccess::load_acquire(&thread->om_in_use_count), &counters->n_in_circulation); 2790 2791 int deflated_count = deflate_monitor_list(&thread->om_in_use_list, &thread->om_in_use_count, &free_head_p, &free_tail_p); 2792 Atomic::add(OrderAccess::load_acquire(&thread->om_in_use_count), &counters->n_in_use); 2793 2794 if (free_head_p != NULL) { 2795 // Move the deflated ObjectMonitors back to the global free list. 2796 // No races on the working list so no need for load_acquire(). 2797 guarantee(free_tail_p != NULL && deflated_count > 0, "invariant"); 2798 assert(free_tail_p->_next_om == NULL, "must be NULL: _next_om=" 2799 INTPTR_FORMAT, p2i(free_tail_p->_next_om)); 2800 prepend_list_to_g_free_list(free_head_p, free_tail_p, deflated_count); 2801 Atomic::add(deflated_count, &counters->n_scavenged); 2802 Atomic::add(deflated_count, &counters->per_thread_scavenged); 2803 } 2804 2805 timer.stop(); 2806 // Safepoint logging cares about cumulative per_thread_times and 2807 // we'll capture most of the cost, but not the muxRelease() which 2808 // should be cheap. 2809 counters->per_thread_times += timer.seconds(); 2810 2811 LogStreamHandle(Debug, monitorinflation) lsh_debug; 2812 LogStreamHandle(Info, monitorinflation) lsh_info; 2813 LogStream* ls = NULL; 2814 if (log_is_enabled(Debug, monitorinflation)) { 2815 ls = &lsh_debug; 2816 } else if (deflated_count != 0 && log_is_enabled(Info, monitorinflation)) { 2817 ls = &lsh_info; 2818 } 2819 if (ls != NULL) { 2820 ls->print_cr("jt=" INTPTR_FORMAT ": deflating per-thread idle monitors, %3.7f secs, %d monitors", p2i(thread), timer.seconds(), deflated_count); 2821 } 2822 } 2823 2824 // Monitor cleanup on JavaThread::exit 2825 2826 // Iterate through monitor cache and attempt to release thread's monitors 2827 // Gives up on a particular monitor if an exception occurs, but continues 2828 // the overall iteration, swallowing the exception. 2829 class ReleaseJavaMonitorsClosure: public MonitorClosure { 2830 private: 2831 TRAPS; 2832 2833 public: 2834 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} 2835 void do_monitor(ObjectMonitor* mid) { 2836 if (mid->owner() == THREAD) { 2837 (void)mid->complete_exit(CHECK); 2838 } 2839 } 2840 }; 2841 2842 // Release all inflated monitors owned by THREAD. Lightweight monitors are 2843 // ignored. This is meant to be called during JNI thread detach which assumes 2844 // all remaining monitors are heavyweight. All exceptions are swallowed. 2845 // Scanning the extant monitor list can be time consuming. 2846 // A simple optimization is to add a per-thread flag that indicates a thread 2847 // called jni_monitorenter() during its lifetime. 2848 // 2849 // Instead of No_Savepoint_Verifier it might be cheaper to 2850 // use an idiom of the form: 2851 // auto int tmp = SafepointSynchronize::_safepoint_counter ; 2852 // <code that must not run at safepoint> 2853 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; 2854 // Since the tests are extremely cheap we could leave them enabled 2855 // for normal product builds. 2856 2857 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { 2858 assert(THREAD == JavaThread::current(), "must be current Java thread"); 2859 NoSafepointVerifier nsv; 2860 ReleaseJavaMonitorsClosure rjmc(THREAD); 2861 ObjectSynchronizer::monitors_iterate(&rjmc); 2862 THREAD->clear_pending_exception(); 2863 } 2864 2865 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) { 2866 switch (cause) { 2867 case inflate_cause_vm_internal: return "VM Internal"; 2868 case inflate_cause_monitor_enter: return "Monitor Enter"; 2869 case inflate_cause_wait: return "Monitor Wait"; 2870 case inflate_cause_notify: return "Monitor Notify"; 2871 case inflate_cause_hash_code: return "Monitor Hash Code"; 2872 case inflate_cause_jni_enter: return "JNI Monitor Enter"; 2873 case inflate_cause_jni_exit: return "JNI Monitor Exit"; 2874 default: 2875 ShouldNotReachHere(); 2876 } 2877 return "Unknown"; 2878 } 2879 2880 //------------------------------------------------------------------------------ 2881 // Debugging code 2882 2883 u_char* ObjectSynchronizer::get_gvars_addr() { 2884 return (u_char*)&GVars; 2885 } 2886 2887 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() { 2888 return (u_char*)&GVars.hc_sequence; 2889 } 2890 2891 size_t ObjectSynchronizer::get_gvars_size() { 2892 return sizeof(SharedGlobals); 2893 } 2894 2895 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() { 2896 return (u_char*)&GVars.stw_random; 2897 } 2898 2899 void ObjectSynchronizer::audit_and_print_stats(bool on_exit) { 2900 assert(on_exit || SafepointSynchronize::is_at_safepoint(), "invariant"); 2901 2902 LogStreamHandle(Debug, monitorinflation) lsh_debug; 2903 LogStreamHandle(Info, monitorinflation) lsh_info; 2904 LogStreamHandle(Trace, monitorinflation) lsh_trace; 2905 LogStream* ls = NULL; 2906 if (log_is_enabled(Trace, monitorinflation)) { 2907 ls = &lsh_trace; 2908 } else if (log_is_enabled(Debug, monitorinflation)) { 2909 ls = &lsh_debug; 2910 } else if (log_is_enabled(Info, monitorinflation)) { 2911 ls = &lsh_info; 2912 } 2913 assert(ls != NULL, "sanity check"); 2914 2915 // Log counts for the global and per-thread monitor lists: 2916 int chk_om_population = log_monitor_list_counts(ls); 2917 int error_cnt = 0; 2918 2919 ls->print_cr("Checking global lists:"); 2920 2921 // Check g_om_population: 2922 if (OrderAccess::load_acquire(&g_om_population) == chk_om_population) { 2923 ls->print_cr("g_om_population=%d equals chk_om_population=%d", 2924 OrderAccess::load_acquire(&g_om_population), 2925 chk_om_population); 2926 } else { 2927 ls->print_cr("ERROR: g_om_population=%d is not equal to " 2928 "chk_om_population=%d", 2929 OrderAccess::load_acquire(&g_om_population), 2930 chk_om_population); 2931 error_cnt++; 2932 } 2933 2934 // Check g_om_in_use_list and g_om_in_use_count: 2935 chk_global_in_use_list_and_count(ls, &error_cnt); 2936 2937 // Check g_free_list and g_om_free_count: 2938 chk_global_free_list_and_count(ls, &error_cnt); 2939 2940 ls->print_cr("Checking per-thread lists:"); 2941 2942 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 2943 // Check om_in_use_list and om_in_use_count: 2944 chk_per_thread_in_use_list_and_count(jt, ls, &error_cnt); 2945 2946 // Check om_free_list and om_free_count: 2947 chk_per_thread_free_list_and_count(jt, ls, &error_cnt); 2948 } 2949 2950 if (error_cnt == 0) { 2951 ls->print_cr("No errors found in monitor list checks."); 2952 } else { 2953 log_error(monitorinflation)("found monitor list errors: error_cnt=%d", error_cnt); 2954 } 2955 2956 if ((on_exit && log_is_enabled(Info, monitorinflation)) || 2957 (!on_exit && log_is_enabled(Trace, monitorinflation))) { 2958 // When exiting this log output is at the Info level. When called 2959 // at a safepoint, this log output is at the Trace level since 2960 // there can be a lot of it. 2961 log_in_use_monitor_details(ls); 2962 } 2963 2964 ls->flush(); 2965 2966 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt); 2967 } 2968 2969 // Check a free monitor entry; log any errors. 2970 void ObjectSynchronizer::chk_free_entry(JavaThread* jt, ObjectMonitor* n, 2971 outputStream * out, int *error_cnt_p) { 2972 stringStream ss; 2973 if (n->is_busy()) { 2974 if (jt != NULL) { 2975 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2976 ": free per-thread monitor must not be busy: %s", p2i(jt), 2977 p2i(n), n->is_busy_to_string(&ss)); 2978 } else { 2979 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor " 2980 "must not be busy: %s", p2i(n), n->is_busy_to_string(&ss)); 2981 } 2982 *error_cnt_p = *error_cnt_p + 1; 2983 } 2984 if (n->header().value() != 0) { 2985 if (jt != NULL) { 2986 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 2987 ": free per-thread monitor must have NULL _header " 2988 "field: _header=" INTPTR_FORMAT, p2i(jt), p2i(n), 2989 n->header().value()); 2990 *error_cnt_p = *error_cnt_p + 1; 2991 } else if (!AsyncDeflateIdleMonitors) { 2992 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor " 2993 "must have NULL _header field: _header=" INTPTR_FORMAT, 2994 p2i(n), n->header().value()); 2995 *error_cnt_p = *error_cnt_p + 1; 2996 } 2997 } 2998 if (n->object() != NULL) { 2999 if (jt != NULL) { 3000 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 3001 ": free per-thread monitor must have NULL _object " 3002 "field: _object=" INTPTR_FORMAT, p2i(jt), p2i(n), 3003 p2i(n->object())); 3004 } else { 3005 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": free global monitor " 3006 "must have NULL _object field: _object=" INTPTR_FORMAT, 3007 p2i(n), p2i(n->object())); 3008 } 3009 *error_cnt_p = *error_cnt_p + 1; 3010 } 3011 } 3012 3013 // Check the global free list and count; log the results of the checks. 3014 void ObjectSynchronizer::chk_global_free_list_and_count(outputStream * out, 3015 int *error_cnt_p) { 3016 int chk_om_free_count = 0; 3017 for (ObjectMonitor* n = OrderAccess::load_acquire(&g_free_list); n != NULL; n = unmarked_next(n)) { 3018 chk_free_entry(NULL /* jt */, n, out, error_cnt_p); 3019 chk_om_free_count++; 3020 } 3021 if (OrderAccess::load_acquire(&g_om_free_count) == chk_om_free_count) { 3022 out->print_cr("g_om_free_count=%d equals chk_om_free_count=%d", 3023 OrderAccess::load_acquire(&g_om_free_count), 3024 chk_om_free_count); 3025 } else { 3026 // With lock free access to g_free_list, it is possible for an 3027 // ObjectMonitor to be prepended to g_free_list after we started 3028 // calculating chk_om_free_count so g_om_free_count may not 3029 // match anymore. 3030 out->print_cr("WARNING: g_om_free_count=%d is not equal to " 3031 "chk_om_free_count=%d", 3032 OrderAccess::load_acquire(&g_om_free_count), 3033 chk_om_free_count); 3034 } 3035 } 3036 3037 // Check the global in-use list and count; log the results of the checks. 3038 void ObjectSynchronizer::chk_global_in_use_list_and_count(outputStream * out, 3039 int *error_cnt_p) { 3040 int chk_om_in_use_count = 0; 3041 for (ObjectMonitor* n = OrderAccess::load_acquire(&g_om_in_use_list); n != NULL; n = unmarked_next(n)) { 3042 chk_in_use_entry(NULL /* jt */, n, out, error_cnt_p); 3043 chk_om_in_use_count++; 3044 } 3045 if (OrderAccess::load_acquire(&g_om_in_use_count) == chk_om_in_use_count) { 3046 out->print_cr("g_om_in_use_count=%d equals chk_om_in_use_count=%d", 3047 OrderAccess::load_acquire(&g_om_in_use_count), 3048 chk_om_in_use_count); 3049 } else { 3050 out->print_cr("ERROR: g_om_in_use_count=%d is not equal to chk_om_in_use_count=%d", 3051 OrderAccess::load_acquire(&g_om_in_use_count), 3052 chk_om_in_use_count); 3053 *error_cnt_p = *error_cnt_p + 1; 3054 } 3055 } 3056 3057 // Check an in-use monitor entry; log any errors. 3058 void ObjectSynchronizer::chk_in_use_entry(JavaThread* jt, ObjectMonitor* n, 3059 outputStream * out, int *error_cnt_p) { 3060 if (n->header().value() == 0) { 3061 if (jt != NULL) { 3062 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 3063 ": in-use per-thread monitor must have non-NULL _header " 3064 "field.", p2i(jt), p2i(n)); 3065 } else { 3066 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use global monitor " 3067 "must have non-NULL _header field.", p2i(n)); 3068 } 3069 *error_cnt_p = *error_cnt_p + 1; 3070 } 3071 if (n->object() == NULL) { 3072 if (jt != NULL) { 3073 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 3074 ": in-use per-thread monitor must have non-NULL _object " 3075 "field.", p2i(jt), p2i(n)); 3076 } else { 3077 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use global monitor " 3078 "must have non-NULL _object field.", p2i(n)); 3079 } 3080 *error_cnt_p = *error_cnt_p + 1; 3081 } 3082 const oop obj = (oop)n->object(); 3083 const markWord mark = obj->mark(); 3084 if (!mark.has_monitor()) { 3085 if (jt != NULL) { 3086 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 3087 ": in-use per-thread monitor's object does not think " 3088 "it has a monitor: obj=" INTPTR_FORMAT ", mark=" 3089 INTPTR_FORMAT, p2i(jt), p2i(n), p2i(obj), mark.value()); 3090 } else { 3091 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use global " 3092 "monitor's object does not think it has a monitor: obj=" 3093 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n), 3094 p2i(obj), mark.value()); 3095 } 3096 *error_cnt_p = *error_cnt_p + 1; 3097 } 3098 ObjectMonitor* const obj_mon = mark.monitor(); 3099 if (n != obj_mon) { 3100 if (jt != NULL) { 3101 out->print_cr("ERROR: jt=" INTPTR_FORMAT ", monitor=" INTPTR_FORMAT 3102 ": in-use per-thread monitor's object does not refer " 3103 "to the same monitor: obj=" INTPTR_FORMAT ", mark=" 3104 INTPTR_FORMAT ", obj_mon=" INTPTR_FORMAT, p2i(jt), 3105 p2i(n), p2i(obj), mark.value(), p2i(obj_mon)); 3106 } else { 3107 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use global " 3108 "monitor's object does not refer to the same monitor: obj=" 3109 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon=" 3110 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon)); 3111 } 3112 *error_cnt_p = *error_cnt_p + 1; 3113 } 3114 } 3115 3116 // Check the thread's free list and count; log the results of the checks. 3117 void ObjectSynchronizer::chk_per_thread_free_list_and_count(JavaThread *jt, 3118 outputStream * out, 3119 int *error_cnt_p) { 3120 int chk_om_free_count = 0; 3121 for (ObjectMonitor* n = OrderAccess::load_acquire(&jt->om_free_list); n != NULL; n = unmarked_next(n)) { 3122 chk_free_entry(jt, n, out, error_cnt_p); 3123 chk_om_free_count++; 3124 } 3125 if (OrderAccess::load_acquire(&jt->om_free_count) == chk_om_free_count) { 3126 out->print_cr("jt=" INTPTR_FORMAT ": om_free_count=%d equals " 3127 "chk_om_free_count=%d", p2i(jt), 3128 OrderAccess::load_acquire(&jt->om_free_count), 3129 chk_om_free_count); 3130 } else { 3131 out->print_cr("ERROR: jt=" INTPTR_FORMAT ": om_free_count=%d is not " 3132 "equal to chk_om_free_count=%d", p2i(jt), 3133 OrderAccess::load_acquire(&jt->om_free_count), 3134 chk_om_free_count); 3135 *error_cnt_p = *error_cnt_p + 1; 3136 } 3137 } 3138 3139 // Check the thread's in-use list and count; log the results of the checks. 3140 void ObjectSynchronizer::chk_per_thread_in_use_list_and_count(JavaThread *jt, 3141 outputStream * out, 3142 int *error_cnt_p) { 3143 int chk_om_in_use_count = 0; 3144 for (ObjectMonitor* n = OrderAccess::load_acquire(&jt->om_in_use_list); n != NULL; n = unmarked_next(n)) { 3145 chk_in_use_entry(jt, n, out, error_cnt_p); 3146 chk_om_in_use_count++; 3147 } 3148 if (OrderAccess::load_acquire(&jt->om_in_use_count) == chk_om_in_use_count) { 3149 out->print_cr("jt=" INTPTR_FORMAT ": om_in_use_count=%d equals " 3150 "chk_om_in_use_count=%d", p2i(jt), 3151 OrderAccess::load_acquire(&jt->om_in_use_count), 3152 chk_om_in_use_count); 3153 } else { 3154 out->print_cr("ERROR: jt=" INTPTR_FORMAT ": om_in_use_count=%d is not " 3155 "equal to chk_om_in_use_count=%d", p2i(jt), 3156 OrderAccess::load_acquire(&jt->om_in_use_count), 3157 chk_om_in_use_count); 3158 *error_cnt_p = *error_cnt_p + 1; 3159 } 3160 } 3161 3162 // Log details about ObjectMonitors on the in-use lists. The 'BHL' 3163 // flags indicate why the entry is in-use, 'object' and 'object type' 3164 // indicate the associated object and its type. 3165 void ObjectSynchronizer::log_in_use_monitor_details(outputStream * out) { 3166 stringStream ss; 3167 if (OrderAccess::load_acquire(&g_om_in_use_count) > 0) { 3168 out->print_cr("In-use global monitor info:"); 3169 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)"); 3170 out->print_cr("%18s %s %7s %18s %18s", 3171 "monitor", "BHL", "ref_cnt", "object", "object type"); 3172 out->print_cr("================== === ======= ================== =================="); 3173 for (ObjectMonitor* n = OrderAccess::load_acquire(&g_om_in_use_list); n != NULL; n = unmarked_next(n)) { 3174 const oop obj = (oop) n->object(); 3175 const markWord mark = n->header(); 3176 ResourceMark rm; 3177 out->print(INTPTR_FORMAT " %d%d%d %7d " INTPTR_FORMAT " %s", 3178 p2i(n), n->is_busy() != 0, mark.hash() != 0, 3179 n->owner() != NULL, (int)n->ref_count(), p2i(obj), 3180 obj->klass()->external_name()); 3181 if (n->is_busy() != 0) { 3182 out->print(" (%s)", n->is_busy_to_string(&ss)); 3183 ss.reset(); 3184 } 3185 out->cr(); 3186 } 3187 } 3188 3189 out->print_cr("In-use per-thread monitor info:"); 3190 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)"); 3191 out->print_cr("%18s %18s %s %7s %18s %18s", 3192 "jt", "monitor", "BHL", "ref_cnt", "object", "object type"); 3193 out->print_cr("================== ================== === ======= ================== =================="); 3194 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 3195 for (ObjectMonitor* n = OrderAccess::load_acquire(&jt->om_in_use_list); n != NULL; n = unmarked_next(n)) { 3196 const oop obj = (oop) n->object(); 3197 const markWord mark = n->header(); 3198 ResourceMark rm; 3199 out->print(INTPTR_FORMAT " " INTPTR_FORMAT " %d%d%d %7d " 3200 INTPTR_FORMAT " %s", p2i(jt), p2i(n), n->is_busy() != 0, 3201 mark.hash() != 0, n->owner() != NULL, (int)n->ref_count(), 3202 p2i(obj), obj->klass()->external_name()); 3203 if (n->is_busy() != 0) { 3204 out->print(" (%s)", n->is_busy_to_string(&ss)); 3205 ss.reset(); 3206 } 3207 out->cr(); 3208 } 3209 } 3210 3211 out->flush(); 3212 } 3213 3214 // Log counts for the global and per-thread monitor lists and return 3215 // the population count. 3216 int ObjectSynchronizer::log_monitor_list_counts(outputStream * out) { 3217 int pop_count = 0; 3218 out->print_cr("%18s %10s %10s %10s", 3219 "Global Lists:", "InUse", "Free", "Total"); 3220 out->print_cr("================== ========== ========== =========="); 3221 out->print_cr("%18s %10d %10d %10d", "", 3222 OrderAccess::load_acquire(&g_om_in_use_count), 3223 OrderAccess::load_acquire(&g_om_free_count), 3224 OrderAccess::load_acquire(&g_om_population)); 3225 pop_count += OrderAccess::load_acquire(&g_om_in_use_count) + 3226 OrderAccess::load_acquire(&g_om_free_count); 3227 3228 out->print_cr("%18s %10s %10s %10s", 3229 "Per-Thread Lists:", "InUse", "Free", "Provision"); 3230 out->print_cr("================== ========== ========== =========="); 3231 3232 for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) { 3233 out->print_cr(INTPTR_FORMAT " %10d %10d %10d", p2i(jt), 3234 OrderAccess::load_acquire(&jt->om_in_use_count), 3235 OrderAccess::load_acquire(&jt->om_free_count), 3236 jt->om_free_provision); 3237 pop_count += OrderAccess::load_acquire(&jt->om_in_use_count) + 3238 OrderAccess::load_acquire(&jt->om_free_count); 3239 } 3240 return pop_count; 3241 } 3242 3243 #ifndef PRODUCT 3244 3245 // Check if monitor belongs to the monitor cache 3246 // The list is grow-only so it's *relatively* safe to traverse 3247 // the list of extant blocks without taking a lock. 3248 3249 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { 3250 PaddedObjectMonitor* block = OrderAccess::load_acquire(&g_block_list); 3251 while (block != NULL) { 3252 assert(block->object() == CHAINMARKER, "must be a block header"); 3253 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) { 3254 address mon = (address)monitor; 3255 address blk = (address)block; 3256 size_t diff = mon - blk; 3257 assert((diff % sizeof(PaddedObjectMonitor)) == 0, "must be aligned"); 3258 return 1; 3259 } 3260 // unmarked_next() is not needed with g_block_list (no next field marking). 3261 block = (PaddedObjectMonitor*)OrderAccess::load_acquire(&block->_next_om); 3262 } 3263 return 0; 3264 } 3265 3266 #endif