1 /* 2 * Copyright (c) 2018, 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 "gc/shared/oopStorage.inline.hpp" 27 #include "gc/shared/oopStorageParState.inline.hpp" 28 #include "logging/log.hpp" 29 #include "logging/logStream.hpp" 30 #include "memory/allocation.inline.hpp" 31 #include "runtime/atomic.hpp" 32 #include "runtime/globals.hpp" 33 #include "runtime/handles.inline.hpp" 34 #include "runtime/mutex.hpp" 35 #include "runtime/mutexLocker.hpp" 36 #include "runtime/orderAccess.inline.hpp" 37 #include "runtime/safepoint.hpp" 38 #include "runtime/stubRoutines.hpp" 39 #include "runtime/thread.hpp" 40 #include "utilities/align.hpp" 41 #include "utilities/count_trailing_zeros.hpp" 42 #include "utilities/debug.hpp" 43 #include "utilities/globalCounter.inline.hpp" 44 #include "utilities/globalDefinitions.hpp" 45 #include "utilities/macros.hpp" 46 #include "utilities/ostream.hpp" 47 #include "utilities/spinYield.hpp" 48 49 OopStorage::AllocateEntry::AllocateEntry() : _prev(NULL), _next(NULL) {} 50 51 OopStorage::AllocateEntry::~AllocateEntry() { 52 assert(_prev == NULL, "deleting attached block"); 53 assert(_next == NULL, "deleting attached block"); 54 } 55 56 OopStorage::AllocateList::AllocateList(const AllocateEntry& (*get_entry)(const Block& block)) : 57 _head(NULL), _tail(NULL), _get_entry(get_entry) 58 {} 59 60 OopStorage::AllocateList::~AllocateList() { 61 // ~OopStorage() empties its lists before destroying them. 62 assert(_head == NULL, "deleting non-empty block list"); 63 assert(_tail == NULL, "deleting non-empty block list"); 64 } 65 66 void OopStorage::AllocateList::push_front(const Block& block) { 67 const Block* old = _head; 68 if (old == NULL) { 69 assert(_tail == NULL, "invariant"); 70 _head = _tail = █ 71 } else { 72 _get_entry(block)._next = old; 73 _get_entry(*old)._prev = █ 74 _head = █ 75 } 76 } 77 78 void OopStorage::AllocateList::push_back(const Block& block) { 79 const Block* old = _tail; 80 if (old == NULL) { 81 assert(_head == NULL, "invariant"); 82 _head = _tail = █ 83 } else { 84 _get_entry(*old)._next = █ 85 _get_entry(block)._prev = old; 86 _tail = █ 87 } 88 } 89 90 void OopStorage::AllocateList::unlink(const Block& block) { 91 const AllocateEntry& block_entry = _get_entry(block); 92 const Block* prev_blk = block_entry._prev; 93 const Block* next_blk = block_entry._next; 94 block_entry._prev = NULL; 95 block_entry._next = NULL; 96 if ((prev_blk == NULL) && (next_blk == NULL)) { 97 assert(_head == &block, "invariant"); 98 assert(_tail == &block, "invariant"); 99 _head = _tail = NULL; 100 } else if (prev_blk == NULL) { 101 assert(_head == &block, "invariant"); 102 _get_entry(*next_blk)._prev = NULL; 103 _head = next_blk; 104 } else if (next_blk == NULL) { 105 assert(_tail == &block, "invariant"); 106 _get_entry(*prev_blk)._next = NULL; 107 _tail = prev_blk; 108 } else { 109 _get_entry(*next_blk)._prev = prev_blk; 110 _get_entry(*prev_blk)._next = next_blk; 111 } 112 } 113 114 OopStorage::ActiveArray::ActiveArray(size_t size) : 115 _size(size), 116 _block_count(0), 117 _refcount(0) 118 {} 119 120 OopStorage::ActiveArray::~ActiveArray() { 121 assert(_refcount == 0, "precondition"); 122 } 123 124 OopStorage::ActiveArray* OopStorage::ActiveArray::create(size_t size, AllocFailType alloc_fail) { 125 size_t size_in_bytes = blocks_offset() + sizeof(Block*) * size; 126 void* mem = NEW_C_HEAP_ARRAY3(char, size_in_bytes, mtGC, CURRENT_PC, alloc_fail); 127 if (mem == NULL) return NULL; 128 return new (mem) ActiveArray(size); 129 } 130 131 void OopStorage::ActiveArray::destroy(ActiveArray* ba) { 132 ba->~ActiveArray(); 133 FREE_C_HEAP_ARRAY(char, ba); 134 } 135 136 size_t OopStorage::ActiveArray::size() const { 137 return _size; 138 } 139 140 size_t OopStorage::ActiveArray::block_count() const { 141 return _block_count; 142 } 143 144 size_t OopStorage::ActiveArray::block_count_acquire() const { 145 return OrderAccess::load_acquire(&_block_count); 146 } 147 148 void OopStorage::ActiveArray::increment_refcount() const { 149 int new_value = Atomic::add(1, &_refcount); 150 assert(new_value >= 1, "negative refcount %d", new_value - 1); 151 } 152 153 bool OopStorage::ActiveArray::decrement_refcount() const { 154 int new_value = Atomic::sub(1, &_refcount); 155 assert(new_value >= 0, "negative refcount %d", new_value); 156 return new_value == 0; 157 } 158 159 bool OopStorage::ActiveArray::push(Block* block) { 160 size_t index = _block_count; 161 if (index < _size) { 162 block->set_active_index(index); 163 *block_ptr(index) = block; 164 // Use a release_store to ensure all the setup is complete before 165 // making the block visible. 166 OrderAccess::release_store(&_block_count, index + 1); 167 return true; 168 } else { 169 return false; 170 } 171 } 172 173 void OopStorage::ActiveArray::remove(Block* block) { 174 assert(_block_count > 0, "array is empty"); 175 size_t index = block->active_index(); 176 assert(*block_ptr(index) == block, "block not present"); 177 size_t last_index = _block_count - 1; 178 Block* last_block = *block_ptr(last_index); 179 last_block->set_active_index(index); 180 *block_ptr(index) = last_block; 181 _block_count = last_index; 182 } 183 184 void OopStorage::ActiveArray::copy_from(const ActiveArray* from) { 185 assert(_block_count == 0, "array must be empty"); 186 size_t count = from->_block_count; 187 assert(count <= _size, "precondition"); 188 Block* const* from_ptr = from->block_ptr(0); 189 Block** to_ptr = block_ptr(0); 190 for (size_t i = 0; i < count; ++i) { 191 Block* block = *from_ptr++; 192 assert(block->active_index() == i, "invariant"); 193 *to_ptr++ = block; 194 } 195 _block_count = count; 196 } 197 198 // Blocks start with an array of BitsPerWord oop entries. That array 199 // is divided into conceptual BytesPerWord sections of BitsPerByte 200 // entries. Blocks are allocated aligned on section boundaries, for 201 // the convenience of mapping from an entry to the containing block; 202 // see block_for_ptr(). Aligning on section boundary rather than on 203 // the full _data wastes a lot less space, but makes for a bit more 204 // work in block_for_ptr(). 205 206 const unsigned section_size = BitsPerByte; 207 const unsigned section_count = BytesPerWord; 208 const unsigned block_alignment = sizeof(oop) * section_size; 209 210 OopStorage::Block::Block(const OopStorage* owner, void* memory) : 211 _data(), 212 _allocated_bitmask(0), 213 _owner(owner), 214 _memory(memory), 215 _active_index(0), 216 _allocate_entry(), 217 _deferred_updates_next(NULL), 218 _release_refcount(0) 219 { 220 STATIC_ASSERT(_data_pos == 0); 221 STATIC_ASSERT(section_size * section_count == ARRAY_SIZE(_data)); 222 assert(offset_of(Block, _data) == _data_pos, "invariant"); 223 assert(owner != NULL, "NULL owner"); 224 assert(is_aligned(this, block_alignment), "misaligned block"); 225 } 226 227 OopStorage::Block::~Block() { 228 assert(_release_refcount == 0, "deleting block while releasing"); 229 assert(_deferred_updates_next == NULL, "deleting block with deferred update"); 230 // Clear fields used by block_for_ptr and entry validation, which 231 // might help catch bugs. Volatile to prevent dead-store elimination. 232 const_cast<uintx volatile&>(_allocated_bitmask) = 0; 233 const_cast<OopStorage* volatile&>(_owner) = NULL; 234 } 235 236 const OopStorage::AllocateEntry& OopStorage::Block::get_allocate_entry(const Block& block) { 237 return block._allocate_entry; 238 } 239 240 size_t OopStorage::Block::allocation_size() { 241 // _data must be first member, so aligning Block aligns _data. 242 STATIC_ASSERT(_data_pos == 0); 243 return sizeof(Block) + block_alignment - sizeof(void*); 244 } 245 246 size_t OopStorage::Block::allocation_alignment_shift() { 247 return exact_log2(block_alignment); 248 } 249 250 inline bool is_full_bitmask(uintx bitmask) { return ~bitmask == 0; } 251 inline bool is_empty_bitmask(uintx bitmask) { return bitmask == 0; } 252 253 bool OopStorage::Block::is_full() const { 254 return is_full_bitmask(allocated_bitmask()); 255 } 256 257 bool OopStorage::Block::is_empty() const { 258 return is_empty_bitmask(allocated_bitmask()); 259 } 260 261 uintx OopStorage::Block::bitmask_for_entry(const oop* ptr) const { 262 return bitmask_for_index(get_index(ptr)); 263 } 264 265 // A block is deletable if 266 // (1) It is empty. 267 // (2) There is not a release() operation currently operating on it. 268 // (3) It is not in the deferred updates list. 269 // The order of tests is important for proper interaction between release() 270 // and concurrent deletion. 271 bool OopStorage::Block::is_deletable() const { 272 return (OrderAccess::load_acquire(&_allocated_bitmask) == 0) && 273 (OrderAccess::load_acquire(&_release_refcount) == 0) && 274 (OrderAccess::load_acquire(&_deferred_updates_next) == NULL); 275 } 276 277 OopStorage::Block* OopStorage::Block::deferred_updates_next() const { 278 return _deferred_updates_next; 279 } 280 281 void OopStorage::Block::set_deferred_updates_next(Block* block) { 282 _deferred_updates_next = block; 283 } 284 285 bool OopStorage::Block::contains(const oop* ptr) const { 286 const oop* base = get_pointer(0); 287 return (base <= ptr) && (ptr < (base + ARRAY_SIZE(_data))); 288 } 289 290 size_t OopStorage::Block::active_index() const { 291 return _active_index; 292 } 293 294 void OopStorage::Block::set_active_index(size_t index) { 295 _active_index = index; 296 } 297 298 size_t OopStorage::Block::active_index_safe(const Block* block) { 299 STATIC_ASSERT(sizeof(intptr_t) == sizeof(block->_active_index)); 300 assert(CanUseSafeFetchN(), "precondition"); 301 return SafeFetchN((intptr_t*)&block->_active_index, 0); 302 } 303 304 unsigned OopStorage::Block::get_index(const oop* ptr) const { 305 assert(contains(ptr), PTR_FORMAT " not in block " PTR_FORMAT, p2i(ptr), p2i(this)); 306 return static_cast<unsigned>(ptr - get_pointer(0)); 307 } 308 309 oop* OopStorage::Block::allocate() { 310 // Use CAS loop because release may change bitmask outside of lock. 311 uintx allocated = allocated_bitmask(); 312 while (true) { 313 assert(!is_full_bitmask(allocated), "attempt to allocate from full block"); 314 unsigned index = count_trailing_zeros(~allocated); 315 uintx new_value = allocated | bitmask_for_index(index); 316 uintx fetched = Atomic::cmpxchg(new_value, &_allocated_bitmask, allocated); 317 if (fetched == allocated) { 318 return get_pointer(index); // CAS succeeded; return entry for index. 319 } 320 allocated = fetched; // CAS failed; retry with latest value. 321 } 322 } 323 324 OopStorage::Block* OopStorage::Block::new_block(const OopStorage* owner) { 325 // _data must be first member: aligning block => aligning _data. 326 STATIC_ASSERT(_data_pos == 0); 327 size_t size_needed = allocation_size(); 328 void* memory = NEW_C_HEAP_ARRAY_RETURN_NULL(char, size_needed, mtGC); 329 if (memory == NULL) { 330 return NULL; 331 } 332 void* block_mem = align_up(memory, block_alignment); 333 assert(sizeof(Block) + pointer_delta(block_mem, memory, 1) <= size_needed, 334 "allocated insufficient space for aligned block"); 335 return ::new (block_mem) Block(owner, memory); 336 } 337 338 void OopStorage::Block::delete_block(const Block& block) { 339 void* memory = block._memory; 340 block.Block::~Block(); 341 FREE_C_HEAP_ARRAY(char, memory); 342 } 343 344 // This can return a false positive if ptr is not contained by some 345 // block. For some uses, it is a precondition that ptr is valid, 346 // e.g. contained in some block in owner's _active_array. Other uses 347 // require additional validation of the result. 348 OopStorage::Block* 349 OopStorage::Block::block_for_ptr(const OopStorage* owner, const oop* ptr) { 350 assert(CanUseSafeFetchN(), "precondition"); 351 STATIC_ASSERT(_data_pos == 0); 352 // Const-ness of ptr is not related to const-ness of containing block. 353 // Blocks are allocated section-aligned, so get the containing section. 354 oop* section_start = align_down(const_cast<oop*>(ptr), block_alignment); 355 // Start with a guess that the containing section is the last section, 356 // so the block starts section_count-1 sections earlier. 357 oop* section = section_start - (section_size * (section_count - 1)); 358 // Walk up through the potential block start positions, looking for 359 // the owner in the expected location. If we're below the actual block 360 // start position, the value at the owner position will be some oop 361 // (possibly NULL), which can never match the owner. 362 intptr_t owner_addr = reinterpret_cast<intptr_t>(owner); 363 for (unsigned i = 0; i < section_count; ++i, section += section_size) { 364 Block* candidate = reinterpret_cast<Block*>(section); 365 intptr_t* candidate_owner_addr 366 = reinterpret_cast<intptr_t*>(&candidate->_owner); 367 if (SafeFetchN(candidate_owner_addr, 0) == owner_addr) { 368 return candidate; 369 } 370 } 371 return NULL; 372 } 373 374 ////////////////////////////////////////////////////////////////////////////// 375 // Allocation 376 // 377 // Allocation involves the _allocate_list, which contains a subset of the 378 // blocks owned by a storage object. This is a doubly-linked list, linked 379 // through dedicated fields in the blocks. Full blocks are removed from this 380 // list, though they are still present in the _active_array. Empty blocks are 381 // kept at the end of the _allocate_list, to make it easy for empty block 382 // deletion to find them. 383 // 384 // allocate(), and delete_empty_blocks_concurrent() lock the 385 // _allocate_mutex while performing any list and array modifications. 386 // 387 // allocate() and release() update a block's _allocated_bitmask using CAS 388 // loops. This prevents loss of updates even though release() performs 389 // its updates without any locking. 390 // 391 // allocate() obtains the entry from the first block in the _allocate_list, 392 // and updates that block's _allocated_bitmask to indicate the entry is in 393 // use. If this makes the block full (all entries in use), the block is 394 // removed from the _allocate_list so it won't be considered by future 395 // allocations until some entries in it are released. 396 // 397 // release() is performed lock-free. release() first looks up the block for 398 // the entry, using address alignment to find the enclosing block (thereby 399 // avoiding iteration over the _active_array). Once the block has been 400 // determined, its _allocated_bitmask needs to be updated, and its position in 401 // the _allocate_list may need to be updated. There are two cases: 402 // 403 // (a) If the block is neither full nor would become empty with the release of 404 // the entry, only its _allocated_bitmask needs to be updated. But if the CAS 405 // update fails, the applicable case may change for the retry. 406 // 407 // (b) Otherwise, the _allocate_list also needs to be modified. This requires 408 // locking the _allocate_mutex. To keep the release() operation lock-free, 409 // rather than updating the _allocate_list itself, it instead performs a 410 // lock-free push of the block onto the _deferred_updates list. Entries on 411 // that list are processed by allocate() and delete_empty_blocks_XXX(), while 412 // they already hold the necessary lock. That processing makes the block's 413 // list state consistent with its current _allocated_bitmask. The block is 414 // added to the _allocate_list if not already present and the bitmask is not 415 // full. The block is moved to the end of the _allocated_list if the bitmask 416 // is empty, for ease of empty block deletion processing. 417 418 oop* OopStorage::allocate() { 419 MutexLockerEx ml(_allocate_mutex, Mutex::_no_safepoint_check_flag); 420 // Do some deferred update processing every time we allocate. 421 // Continue processing deferred updates if _allocate_list is empty, 422 // in the hope that we'll get a block from that, rather than 423 // allocating a new block. 424 while (reduce_deferred_updates() && (_allocate_list.head() == NULL)) {} 425 426 // Use the first block in _allocate_list for the allocation. 427 Block* block = _allocate_list.head(); 428 if (block == NULL) { 429 // No available blocks; make a new one, and add to storage. 430 { 431 MutexUnlockerEx mul(_allocate_mutex, Mutex::_no_safepoint_check_flag); 432 block = Block::new_block(this); 433 } 434 if (block == NULL) { 435 while (_allocate_list.head() == NULL) { 436 if (!reduce_deferred_updates()) { 437 // Failed to make new block, no other thread made a block 438 // available while the mutex was released, and didn't get 439 // one from a deferred update either, so return failure. 440 log_info(oopstorage, ref)("%s: failed block allocation", name()); 441 return NULL; 442 } 443 } 444 } else { 445 // Add new block to storage. 446 log_info(oopstorage, blocks)("%s: new block " PTR_FORMAT, name(), p2i(block)); 447 448 // Add new block to the _active_array, growing if needed. 449 if (!_active_array->push(block)) { 450 if (expand_active_array()) { 451 guarantee(_active_array->push(block), "push failed after expansion"); 452 } else { 453 log_info(oopstorage, blocks)("%s: failed active array expand", name()); 454 Block::delete_block(*block); 455 return NULL; 456 } 457 } 458 // Add to end of _allocate_list. The mutex release allowed 459 // other threads to add blocks to the _allocate_list. We prefer 460 // to allocate from non-empty blocks, to allow empty blocks to 461 // be deleted. 462 _allocate_list.push_back(*block); 463 } 464 block = _allocate_list.head(); 465 } 466 // Allocate from first block. 467 assert(block != NULL, "invariant"); 468 assert(!block->is_full(), "invariant"); 469 if (block->is_empty()) { 470 // Transitioning from empty to not empty. 471 log_debug(oopstorage, blocks)("%s: block not empty " PTR_FORMAT, name(), p2i(block)); 472 } 473 oop* result = block->allocate(); 474 assert(result != NULL, "allocation failed"); 475 assert(!block->is_empty(), "postcondition"); 476 Atomic::inc(&_allocation_count); // release updates outside lock. 477 if (block->is_full()) { 478 // Transitioning from not full to full. 479 // Remove full blocks from consideration by future allocates. 480 log_debug(oopstorage, blocks)("%s: block full " PTR_FORMAT, name(), p2i(block)); 481 _allocate_list.unlink(*block); 482 } 483 log_info(oopstorage, ref)("%s: allocated " PTR_FORMAT, name(), p2i(result)); 484 return result; 485 } 486 487 // Create a new, larger, active array with the same content as the 488 // current array, and then replace, relinquishing the old array. 489 // Return true if the array was successfully expanded, false to 490 // indicate allocation failure. 491 bool OopStorage::expand_active_array() { 492 assert_lock_strong(_allocate_mutex); 493 ActiveArray* old_array = _active_array; 494 size_t new_size = 2 * old_array->size(); 495 log_info(oopstorage, blocks)("%s: expand active array " SIZE_FORMAT, 496 name(), new_size); 497 ActiveArray* new_array = ActiveArray::create(new_size, AllocFailStrategy::RETURN_NULL); 498 if (new_array == NULL) return false; 499 new_array->copy_from(old_array); 500 replace_active_array(new_array); 501 relinquish_block_array(old_array); 502 return true; 503 } 504 505 // Make new_array the _active_array. Increments new_array's refcount 506 // to account for the new reference. The assignment is atomic wrto 507 // obtain_active_array; once this function returns, it is safe for the 508 // caller to relinquish the old array. 509 void OopStorage::replace_active_array(ActiveArray* new_array) { 510 // Caller has the old array that is the current value of _active_array. 511 // Update new_array refcount to account for the new reference. 512 new_array->increment_refcount(); 513 // Install new_array, ensuring its initialization is complete first. 514 OrderAccess::release_store(&_active_array, new_array); 515 // Wait for any readers that could read the old array from _active_array. 516 GlobalCounter::write_synchronize(); 517 // All obtain_actuve_array critical sections that could see the old array 518 // have completed, having incremented the refcount of the old array. The 519 // caller can now safely relinquish the old array. 520 } 521 522 // Atomically (wrto replace_active_array) get the active array and 523 // increment its refcount. This provides safe access to the array, 524 // even if an allocate operation expands and replaces the value of 525 // _active_array. The caller must relinquish the array when done 526 // using it. 527 OopStorage::ActiveArray* OopStorage::obtain_active_array() const { 528 GlobalCounter::CriticalSection cs(Thread::current()); 529 ActiveArray* result = OrderAccess::load_acquire(&_active_array); 530 result->increment_refcount(); 531 return result; 532 } 533 534 // Decrement refcount of array and destroy if refcount is zero. 535 void OopStorage::relinquish_block_array(ActiveArray* array) const { 536 if (array->decrement_refcount()) { 537 assert(array != _active_array, "invariant"); 538 ActiveArray::destroy(array); 539 } 540 } 541 542 class OopStorage::WithActiveArray : public StackObj { 543 const OopStorage* _storage; 544 ActiveArray* _active_array; 545 546 public: 547 WithActiveArray(const OopStorage* storage) : 548 _storage(storage), 549 _active_array(storage->obtain_active_array()) 550 {} 551 552 ~WithActiveArray() { 553 _storage->relinquish_block_array(_active_array); 554 } 555 556 ActiveArray& active_array() const { 557 return *_active_array; 558 } 559 }; 560 561 OopStorage::Block* OopStorage::find_block_or_null(const oop* ptr) const { 562 assert(ptr != NULL, "precondition"); 563 return Block::block_for_ptr(this, ptr); 564 } 565 566 static void log_release_transitions(uintx releasing, 567 uintx old_allocated, 568 const OopStorage* owner, 569 const void* block) { 570 Log(oopstorage, blocks) log; 571 LogStream ls(log.debug()); 572 if (is_full_bitmask(old_allocated)) { 573 ls.print_cr("%s: block not full " PTR_FORMAT, owner->name(), p2i(block)); 574 } 575 if (releasing == old_allocated) { 576 ls.print_cr("%s: block empty " PTR_FORMAT, owner->name(), p2i(block)); 577 } 578 } 579 580 void OopStorage::Block::release_entries(uintx releasing, Block* volatile* deferred_list) { 581 assert(releasing != 0, "preconditon"); 582 // Prevent empty block deletion when transitioning to empty. 583 Atomic::inc(&_release_refcount); 584 585 // Atomically update allocated bitmask. 586 uintx old_allocated = _allocated_bitmask; 587 while (true) { 588 assert((releasing & ~old_allocated) == 0, "releasing unallocated entries"); 589 uintx new_value = old_allocated ^ releasing; 590 uintx fetched = Atomic::cmpxchg(new_value, &_allocated_bitmask, old_allocated); 591 if (fetched == old_allocated) break; // Successful update. 592 old_allocated = fetched; // Retry with updated bitmask. 593 } 594 595 // Now that the bitmask has been updated, if we have a state transition 596 // (updated bitmask is empty or old bitmask was full), atomically push 597 // this block onto the deferred updates list. Some future call to 598 // reduce_deferred_updates will make any needed changes related to this 599 // block and _allocate_list. This deferral avoids list updates and the 600 // associated locking here. 601 if ((releasing == old_allocated) || is_full_bitmask(old_allocated)) { 602 // Log transitions. Both transitions are possible in a single update. 603 if (log_is_enabled(Debug, oopstorage, blocks)) { 604 log_release_transitions(releasing, old_allocated, _owner, this); 605 } 606 // Attempt to claim responsibility for adding this block to the deferred 607 // list, by setting the link to non-NULL by self-looping. If this fails, 608 // then someone else has made such a claim and the deferred update has not 609 // yet been processed and will include our change, so we don't need to do 610 // anything further. 611 if (Atomic::replace_if_null(this, &_deferred_updates_next)) { 612 // Successfully claimed. Push, with self-loop for end-of-list. 613 Block* head = *deferred_list; 614 while (true) { 615 _deferred_updates_next = (head == NULL) ? this : head; 616 Block* fetched = Atomic::cmpxchg(this, deferred_list, head); 617 if (fetched == head) break; // Successful update. 618 head = fetched; // Retry with updated head. 619 } 620 log_debug(oopstorage, blocks)("%s: deferred update " PTR_FORMAT, 621 _owner->name(), p2i(this)); 622 } 623 } 624 // Release hold on empty block deletion. 625 Atomic::dec(&_release_refcount); 626 } 627 628 // Process one available deferred update. Returns true if one was processed. 629 bool OopStorage::reduce_deferred_updates() { 630 assert_locked_or_safepoint(_allocate_mutex); 631 // Atomically pop a block off the list, if any available. 632 // No ABA issue because this is only called by one thread at a time. 633 // The atomicity is wrto pushes by release(). 634 Block* block = OrderAccess::load_acquire(&_deferred_updates); 635 while (true) { 636 if (block == NULL) return false; 637 // Try atomic pop of block from list. 638 Block* tail = block->deferred_updates_next(); 639 if (block == tail) tail = NULL; // Handle self-loop end marker. 640 Block* fetched = Atomic::cmpxchg(tail, &_deferred_updates, block); 641 if (fetched == block) break; // Update successful. 642 block = fetched; // Retry with updated block. 643 } 644 block->set_deferred_updates_next(NULL); // Clear tail after updating head. 645 // Ensure bitmask read after pop is complete, including clearing tail, for 646 // ordering with release(). Without this, we may be processing a stale 647 // bitmask state here while blocking a release() operation from recording 648 // the deferred update needed for its bitmask change. 649 OrderAccess::storeload(); 650 // Process popped block. 651 uintx allocated = block->allocated_bitmask(); 652 653 // Make membership in list consistent with bitmask state. 654 if ((_allocate_list.ctail() != NULL) && 655 ((_allocate_list.ctail() == block) || 656 (_allocate_list.next(*block) != NULL))) { 657 // Block is in the allocate list. 658 assert(!is_full_bitmask(allocated), "invariant"); 659 } else if (!is_full_bitmask(allocated)) { 660 // Block is not in the allocate list, but now should be. 661 _allocate_list.push_front(*block); 662 } // Else block is full and not in list, which is correct. 663 664 // Move empty block to end of list, for possible deletion. 665 if (is_empty_bitmask(allocated)) { 666 _allocate_list.unlink(*block); 667 _allocate_list.push_back(*block); 668 } 669 670 log_debug(oopstorage, blocks)("%s: processed deferred update " PTR_FORMAT, 671 name(), p2i(block)); 672 return true; // Processed one pending update. 673 } 674 675 inline void check_release_entry(const oop* entry) { 676 assert(entry != NULL, "Releasing NULL"); 677 assert(*entry == NULL, "Releasing uncleared entry: " PTR_FORMAT, p2i(entry)); 678 } 679 680 void OopStorage::release(const oop* ptr) { 681 check_release_entry(ptr); 682 Block* block = find_block_or_null(ptr); 683 assert(block != NULL, "%s: invalid release " PTR_FORMAT, name(), p2i(ptr)); 684 log_info(oopstorage, ref)("%s: released " PTR_FORMAT, name(), p2i(ptr)); 685 block->release_entries(block->bitmask_for_entry(ptr), &_deferred_updates); 686 Atomic::dec(&_allocation_count); 687 } 688 689 void OopStorage::release(const oop* const* ptrs, size_t size) { 690 size_t i = 0; 691 while (i < size) { 692 check_release_entry(ptrs[i]); 693 Block* block = find_block_or_null(ptrs[i]); 694 assert(block != NULL, "%s: invalid release " PTR_FORMAT, name(), p2i(ptrs[i])); 695 log_info(oopstorage, ref)("%s: released " PTR_FORMAT, name(), p2i(ptrs[i])); 696 size_t count = 0; 697 uintx releasing = 0; 698 for ( ; i < size; ++i) { 699 const oop* entry = ptrs[i]; 700 check_release_entry(entry); 701 // If entry not in block, finish block and resume outer loop with entry. 702 if (!block->contains(entry)) break; 703 // Add entry to releasing bitmap. 704 log_info(oopstorage, ref)("%s: released " PTR_FORMAT, name(), p2i(entry)); 705 uintx entry_bitmask = block->bitmask_for_entry(entry); 706 assert((releasing & entry_bitmask) == 0, 707 "Duplicate entry: " PTR_FORMAT, p2i(entry)); 708 releasing |= entry_bitmask; 709 ++count; 710 } 711 // Release the contiguous entries that are in block. 712 block->release_entries(releasing, &_deferred_updates); 713 Atomic::sub(count, &_allocation_count); 714 } 715 } 716 717 const char* dup_name(const char* name) { 718 char* dup = NEW_C_HEAP_ARRAY(char, strlen(name) + 1, mtGC); 719 strcpy(dup, name); 720 return dup; 721 } 722 723 const size_t initial_active_array_size = 8; 724 725 OopStorage::OopStorage(const char* name, 726 Mutex* allocate_mutex, 727 Mutex* active_mutex) : 728 _name(dup_name(name)), 729 _active_array(ActiveArray::create(initial_active_array_size)), 730 _allocate_list(&Block::get_allocate_entry), 731 _deferred_updates(NULL), 732 _allocate_mutex(allocate_mutex), 733 _active_mutex(active_mutex), 734 _allocation_count(0), 735 _concurrent_iteration_active(false) 736 { 737 _active_array->increment_refcount(); 738 assert(_active_mutex->rank() < _allocate_mutex->rank(), 739 "%s: active_mutex must have lower rank than allocate_mutex", _name); 740 assert(_active_mutex->_safepoint_check_required != Mutex::_safepoint_check_always, 741 "%s: active mutex requires safepoint check", _name); 742 assert(_allocate_mutex->_safepoint_check_required != Mutex::_safepoint_check_always, 743 "%s: allocate mutex requires safepoint check", _name); 744 } 745 746 void OopStorage::delete_empty_block(const Block& block) { 747 assert(block.is_empty(), "discarding non-empty block"); 748 log_info(oopstorage, blocks)("%s: delete empty block " PTR_FORMAT, name(), p2i(&block)); 749 Block::delete_block(block); 750 } 751 752 OopStorage::~OopStorage() { 753 Block* block; 754 while ((block = _deferred_updates) != NULL) { 755 _deferred_updates = block->deferred_updates_next(); 756 block->set_deferred_updates_next(NULL); 757 } 758 while ((block = _allocate_list.head()) != NULL) { 759 _allocate_list.unlink(*block); 760 } 761 bool unreferenced = _active_array->decrement_refcount(); 762 assert(unreferenced, "deleting storage while _active_array is referenced"); 763 for (size_t i = _active_array->block_count(); 0 < i; ) { 764 block = _active_array->at(--i); 765 Block::delete_block(*block); 766 } 767 ActiveArray::destroy(_active_array); 768 FREE_C_HEAP_ARRAY(char, _name); 769 } 770 771 void OopStorage::delete_empty_blocks_safepoint() { 772 assert_at_safepoint(); 773 // Process any pending release updates, which may make more empty 774 // blocks available for deletion. 775 while (reduce_deferred_updates()) {} 776 // Don't interfere with a concurrent iteration. 777 if (_concurrent_iteration_active) return; 778 // Delete empty (and otherwise deletable) blocks from end of _allocate_list. 779 for (Block* block = _allocate_list.tail(); 780 (block != NULL) && block->is_deletable(); 781 block = _allocate_list.tail()) { 782 _active_array->remove(block); 783 _allocate_list.unlink(*block); 784 delete_empty_block(*block); 785 } 786 } 787 788 void OopStorage::delete_empty_blocks_concurrent() { 789 MutexLockerEx ml(_allocate_mutex, Mutex::_no_safepoint_check_flag); 790 // Other threads could be adding to the empty block count while we 791 // release the mutex across the block deletions. Set an upper bound 792 // on how many blocks we'll try to release, so other threads can't 793 // cause an unbounded stay in this function. 794 size_t limit = block_count(); 795 796 for (size_t i = 0; i < limit; ++i) { 797 // Additional updates might become available while we dropped the 798 // lock. But limit number processed to limit lock duration. 799 reduce_deferred_updates(); 800 801 Block* block = _allocate_list.tail(); 802 if ((block == NULL) || !block->is_deletable()) { 803 // No block to delete, so done. There could be more pending 804 // deferred updates that could give us more work to do; deal with 805 // that in some later call, to limit lock duration here. 806 return; 807 } 808 809 { 810 MutexLockerEx aml(_active_mutex, Mutex::_no_safepoint_check_flag); 811 // Don't interfere with a concurrent iteration. 812 if (_concurrent_iteration_active) return; 813 _active_array->remove(block); 814 } 815 // Remove block from _allocate_list and delete it. 816 _allocate_list.unlink(*block); 817 // Release mutex while deleting block. 818 MutexUnlockerEx ul(_allocate_mutex, Mutex::_no_safepoint_check_flag); 819 delete_empty_block(*block); 820 } 821 } 822 823 OopStorage::EntryStatus OopStorage::allocation_status(const oop* ptr) const { 824 const Block* block = find_block_or_null(ptr); 825 if (block != NULL) { 826 // Prevent block deletion and _active_array modification. 827 MutexLockerEx ml(_allocate_mutex, Mutex::_no_safepoint_check_flag); 828 // Block could be a false positive, so get index carefully. 829 size_t index = Block::active_index_safe(block); 830 if ((index < _active_array->block_count()) && 831 (block == _active_array->at(index)) && 832 block->contains(ptr)) { 833 if ((block->allocated_bitmask() & block->bitmask_for_entry(ptr)) != 0) { 834 return ALLOCATED_ENTRY; 835 } else { 836 return UNALLOCATED_ENTRY; 837 } 838 } 839 } 840 return INVALID_ENTRY; 841 } 842 843 size_t OopStorage::allocation_count() const { 844 return _allocation_count; 845 } 846 847 size_t OopStorage::block_count() const { 848 WithActiveArray wab(this); 849 // Count access is racy, but don't care. 850 return wab.active_array().block_count(); 851 } 852 853 size_t OopStorage::total_memory_usage() const { 854 size_t total_size = sizeof(OopStorage); 855 total_size += strlen(name()) + 1; 856 total_size += sizeof(ActiveArray); 857 WithActiveArray wab(this); 858 const ActiveArray& blocks = wab.active_array(); 859 // Count access is racy, but don't care. 860 total_size += blocks.block_count() * Block::allocation_size(); 861 total_size += blocks.size() * sizeof(Block*); 862 return total_size; 863 } 864 865 // Parallel iteration support 866 867 uint OopStorage::BasicParState::default_estimated_thread_count(bool concurrent) { 868 return concurrent ? ConcGCThreads : ParallelGCThreads; 869 } 870 871 OopStorage::BasicParState::BasicParState(const OopStorage* storage, 872 uint estimated_thread_count, 873 bool concurrent) : 874 _storage(storage), 875 _active_array(_storage->obtain_active_array()), 876 _block_count(0), // initialized properly below 877 _next_block(0), 878 _estimated_thread_count(estimated_thread_count), 879 _concurrent(concurrent) 880 { 881 assert(estimated_thread_count > 0, "estimated thread count must be positive"); 882 update_iteration_state(true); 883 // Get the block count *after* iteration state updated, so concurrent 884 // empty block deletion is suppressed and can't reduce the count. But 885 // ensure the count we use was written after the block with that count 886 // was fully initialized; see ActiveArray::push. 887 _block_count = _active_array->block_count_acquire(); 888 } 889 890 OopStorage::BasicParState::~BasicParState() { 891 _storage->relinquish_block_array(_active_array); 892 update_iteration_state(false); 893 } 894 895 void OopStorage::BasicParState::update_iteration_state(bool value) { 896 if (_concurrent) { 897 MutexLockerEx ml(_storage->_active_mutex, Mutex::_no_safepoint_check_flag); 898 assert(_storage->_concurrent_iteration_active != value, "precondition"); 899 _storage->_concurrent_iteration_active = value; 900 } 901 } 902 903 bool OopStorage::BasicParState::claim_next_segment(IterationData* data) { 904 data->_processed += data->_segment_end - data->_segment_start; 905 size_t start = OrderAccess::load_acquire(&_next_block); 906 if (start >= _block_count) { 907 return finish_iteration(data); // No more blocks available. 908 } 909 // Try to claim several at a time, but not *too* many. We want to 910 // avoid deciding there are many available and selecting a large 911 // quantity, get delayed, and then end up claiming most or all of 912 // the remaining largish amount of work, leaving nothing for other 913 // threads to do. But too small a step can lead to contention 914 // over _next_block, esp. when the work per block is small. 915 size_t max_step = 10; 916 size_t remaining = _block_count - start; 917 size_t step = MIN2(max_step, 1 + (remaining / _estimated_thread_count)); 918 // Atomic::add with possible overshoot. This can perform better 919 // than a CAS loop on some platforms when there is contention. 920 // We can cope with the uncertainty by recomputing start/end from 921 // the result of the add, and dealing with potential overshoot. 922 size_t end = Atomic::add(step, &_next_block); 923 // _next_block may have changed, so recompute start from result of add. 924 start = end - step; 925 // _next_block may have changed so much that end has overshot. 926 end = MIN2(end, _block_count); 927 // _next_block may have changed so much that even start has overshot. 928 if (start < _block_count) { 929 // Record claimed segment for iteration. 930 data->_segment_start = start; 931 data->_segment_end = end; 932 return true; // Success. 933 } else { 934 // No more blocks to claim. 935 return finish_iteration(data); 936 } 937 } 938 939 bool OopStorage::BasicParState::finish_iteration(const IterationData* data) const { 940 log_debug(oopstorage, blocks, stats) 941 ("Parallel iteration on %s: blocks = " SIZE_FORMAT 942 ", processed = " SIZE_FORMAT " (%2.f%%)", 943 _storage->name(), _block_count, data->_processed, 944 percent_of(data->_processed, _block_count)); 945 return false; 946 } 947 948 const char* OopStorage::name() const { return _name; } 949 950 #ifndef PRODUCT 951 952 void OopStorage::print_on(outputStream* st) const { 953 size_t allocations = _allocation_count; 954 size_t blocks = _active_array->block_count(); 955 956 double data_size = section_size * section_count; 957 double alloc_percentage = percent_of((double)allocations, blocks * data_size); 958 959 st->print("%s: " SIZE_FORMAT " entries in " SIZE_FORMAT " blocks (%.F%%), " SIZE_FORMAT " bytes", 960 name(), allocations, blocks, alloc_percentage, total_memory_usage()); 961 if (_concurrent_iteration_active) { 962 st->print(", concurrent iteration active"); 963 } 964 } 965 966 #endif // !PRODUCT