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.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/globalDefinitions.hpp" 44 #include "utilities/macros.hpp" 45 #include "utilities/ostream.hpp" 46 #include "utilities/spinYield.hpp" 47 48 OopStorage::AllocateEntry::AllocateEntry() : _prev(NULL), _next(NULL) {} 49 50 OopStorage::AllocateEntry::~AllocateEntry() { 51 assert(_prev == NULL, "deleting attached block"); 52 assert(_next == NULL, "deleting attached block"); 53 } 54 55 OopStorage::AllocateList::AllocateList(const AllocateEntry& (*get_entry)(const Block& block)) : 56 _head(NULL), _tail(NULL), _get_entry(get_entry) 57 {} 58 59 OopStorage::AllocateList::~AllocateList() { 60 // ~OopStorage() empties its lists before destroying them. 61 assert(_head == NULL, "deleting non-empty block list"); 62 assert(_tail == NULL, "deleting non-empty block list"); 63 } 64 65 void OopStorage::AllocateList::push_front(const Block& block) { 66 const Block* old = _head; 67 if (old == NULL) { 68 assert(_tail == NULL, "invariant"); 69 _head = _tail = █ 70 } else { 71 _get_entry(block)._next = old; 72 _get_entry(*old)._prev = █ 73 _head = █ 74 } 75 } 76 77 void OopStorage::AllocateList::push_back(const Block& block) { 78 const Block* old = _tail; 79 if (old == NULL) { 80 assert(_head == NULL, "invariant"); 81 _head = _tail = █ 82 } else { 83 _get_entry(*old)._next = █ 84 _get_entry(block)._prev = old; 85 _tail = █ 86 } 87 } 88 89 void OopStorage::AllocateList::unlink(const Block& block) { 90 const AllocateEntry& block_entry = _get_entry(block); 91 const Block* prev_blk = block_entry._prev; 92 const Block* next_blk = block_entry._next; 93 block_entry._prev = NULL; 94 block_entry._next = NULL; 95 if ((prev_blk == NULL) && (next_blk == NULL)) { 96 assert(_head == &block, "invariant"); 97 assert(_tail == &block, "invariant"); 98 _head = _tail = NULL; 99 } else if (prev_blk == NULL) { 100 assert(_head == &block, "invariant"); 101 _get_entry(*next_blk)._prev = NULL; 102 _head = next_blk; 103 } else if (next_blk == NULL) { 104 assert(_tail == &block, "invariant"); 105 _get_entry(*prev_blk)._next = NULL; 106 _tail = prev_blk; 107 } else { 108 _get_entry(*next_blk)._prev = prev_blk; 109 _get_entry(*prev_blk)._next = next_blk; 110 } 111 } 112 113 OopStorage::ActiveArray::ActiveArray(size_t size) : 114 _size(size), 115 _block_count(0), 116 _refcount(0) 117 {} 118 119 OopStorage::ActiveArray::~ActiveArray() { 120 assert(_refcount == 0, "precondition"); 121 } 122 123 OopStorage::ActiveArray* OopStorage::ActiveArray::create(size_t size, AllocFailType alloc_fail) { 124 size_t size_in_bytes = blocks_offset() + sizeof(Block*) * size; 125 void* mem = NEW_C_HEAP_ARRAY3(char, size_in_bytes, mtGC, CURRENT_PC, alloc_fail); 126 if (mem == NULL) return NULL; 127 return new (mem) ActiveArray(size); 128 } 129 130 void OopStorage::ActiveArray::destroy(ActiveArray* ba) { 131 ba->~ActiveArray(); 132 FREE_C_HEAP_ARRAY(char, ba); 133 } 134 135 size_t OopStorage::ActiveArray::size() const { 136 return _size; 137 } 138 139 size_t OopStorage::ActiveArray::block_count() const { 140 return _block_count; 141 } 142 143 size_t OopStorage::ActiveArray::block_count_acquire() const { 144 return OrderAccess::load_acquire(&_block_count); 145 } 146 147 void OopStorage::ActiveArray::increment_refcount() const { 148 int new_value = Atomic::add(1, &_refcount); 149 assert(new_value >= 1, "negative refcount %d", new_value - 1); 150 } 151 152 bool OopStorage::ActiveArray::decrement_refcount() const { 153 int new_value = Atomic::sub(1, &_refcount); 154 assert(new_value >= 0, "negative refcount %d", new_value); 155 return new_value == 0; 156 } 157 158 bool OopStorage::ActiveArray::push(Block* block) { 159 size_t index = _block_count; 160 if (index < _size) { 161 block->set_active_index(index); 162 *block_ptr(index) = block; 163 // Use a release_store to ensure all the setup is complete before 164 // making the block visible. 165 OrderAccess::release_store(&_block_count, index + 1); 166 return true; 167 } else { 168 return false; 169 } 170 } 171 172 void OopStorage::ActiveArray::remove(Block* block) { 173 assert(_block_count > 0, "array is empty"); 174 size_t index = block->active_index(); 175 assert(*block_ptr(index) == block, "block not present"); 176 size_t last_index = _block_count - 1; 177 Block* last_block = *block_ptr(last_index); 178 last_block->set_active_index(index); 179 *block_ptr(index) = last_block; 180 _block_count = last_index; 181 } 182 183 void OopStorage::ActiveArray::copy_from(const ActiveArray* from) { 184 assert(_block_count == 0, "array must be empty"); 185 size_t count = from->_block_count; 186 assert(count <= _size, "precondition"); 187 Block* const* from_ptr = from->block_ptr(0); 188 Block** to_ptr = block_ptr(0); 189 for (size_t i = 0; i < count; ++i) { 190 Block* block = *from_ptr++; 191 assert(block->active_index() == i, "invariant"); 192 *to_ptr++ = block; 193 } 194 _block_count = count; 195 } 196 197 // Blocks start with an array of BitsPerWord oop entries. That array 198 // is divided into conceptual BytesPerWord sections of BitsPerByte 199 // entries. Blocks are allocated aligned on section boundaries, for 200 // the convenience of mapping from an entry to the containing block; 201 // see block_for_ptr(). Aligning on section boundary rather than on 202 // the full _data wastes a lot less space, but makes for a bit more 203 // work in block_for_ptr(). 204 205 const unsigned section_size = BitsPerByte; 206 const unsigned section_count = BytesPerWord; 207 const unsigned block_alignment = sizeof(oop) * section_size; 208 209 OopStorage::Block::Block(const OopStorage* owner, void* memory) : 210 _data(), 211 _allocated_bitmask(0), 212 _owner(owner), 213 _memory(memory), 214 _active_index(0), 215 _allocate_entry(), 216 _deferred_updates_next(NULL), 217 _release_refcount(0) 218 { 219 STATIC_ASSERT(_data_pos == 0); 220 STATIC_ASSERT(section_size * section_count == ARRAY_SIZE(_data)); 221 assert(offset_of(Block, _data) == _data_pos, "invariant"); 222 assert(owner != NULL, "NULL owner"); 223 assert(is_aligned(this, block_alignment), "misaligned block"); 224 } 225 226 OopStorage::Block::~Block() { 227 assert(_release_refcount == 0, "deleting block while releasing"); 228 assert(_deferred_updates_next == NULL, "deleting block with deferred update"); 229 // Clear fields used by block_for_ptr and entry validation, which 230 // might help catch bugs. Volatile to prevent dead-store elimination. 231 const_cast<uintx volatile&>(_allocated_bitmask) = 0; 232 const_cast<OopStorage* volatile&>(_owner) = NULL; 233 } 234 235 const OopStorage::AllocateEntry& OopStorage::Block::get_allocate_entry(const Block& block) { 236 return block._allocate_entry; 237 } 238 239 size_t OopStorage::Block::allocation_size() { 240 // _data must be first member, so aligning Block aligns _data. 241 STATIC_ASSERT(_data_pos == 0); 242 return sizeof(Block) + block_alignment - sizeof(void*); 243 } 244 245 size_t OopStorage::Block::allocation_alignment_shift() { 246 return exact_log2(block_alignment); 247 } 248 249 inline bool is_full_bitmask(uintx bitmask) { return ~bitmask == 0; } 250 inline bool is_empty_bitmask(uintx bitmask) { return bitmask == 0; } 251 252 bool OopStorage::Block::is_full() const { 253 return is_full_bitmask(allocated_bitmask()); 254 } 255 256 bool OopStorage::Block::is_empty() const { 257 return is_empty_bitmask(allocated_bitmask()); 258 } 259 260 uintx OopStorage::Block::bitmask_for_entry(const oop* ptr) const { 261 return bitmask_for_index(get_index(ptr)); 262 } 263 264 // A block is deletable if 265 // (1) It is empty. 266 // (2) There is not a release() operation currently operating on it. 267 // (3) It is not in the deferred updates list. 268 // The order of tests is important for proper interaction between release() 269 // and concurrent deletion. 270 bool OopStorage::Block::is_deletable() const { 271 return (OrderAccess::load_acquire(&_allocated_bitmask) == 0) && 272 (OrderAccess::load_acquire(&_release_refcount) == 0) && 273 (OrderAccess::load_acquire(&_deferred_updates_next) == NULL); 274 } 275 276 OopStorage::Block* OopStorage::Block::deferred_updates_next() const { 277 return _deferred_updates_next; 278 } 279 280 void OopStorage::Block::set_deferred_updates_next(Block* block) { 281 _deferred_updates_next = block; 282 } 283 284 bool OopStorage::Block::contains(const oop* ptr) const { 285 const oop* base = get_pointer(0); 286 return (base <= ptr) && (ptr < (base + ARRAY_SIZE(_data))); 287 } 288 289 size_t OopStorage::Block::active_index() const { 290 return _active_index; 291 } 292 293 void OopStorage::Block::set_active_index(size_t index) { 294 _active_index = index; 295 } 296 297 size_t OopStorage::Block::active_index_safe(const Block* block) { 298 STATIC_ASSERT(sizeof(intptr_t) == sizeof(block->_active_index)); 299 assert(CanUseSafeFetchN(), "precondition"); 300 return SafeFetchN((intptr_t*)&block->_active_index, 0); 301 } 302 303 unsigned OopStorage::Block::get_index(const oop* ptr) const { 304 assert(contains(ptr), PTR_FORMAT " not in block " PTR_FORMAT, p2i(ptr), p2i(this)); 305 return static_cast<unsigned>(ptr - get_pointer(0)); 306 } 307 308 oop* OopStorage::Block::allocate() { 309 // Use CAS loop because release may change bitmask outside of lock. 310 uintx allocated = allocated_bitmask(); 311 while (true) { 312 assert(!is_full_bitmask(allocated), "attempt to allocate from full block"); 313 unsigned index = count_trailing_zeros(~allocated); 314 uintx new_value = allocated | bitmask_for_index(index); 315 uintx fetched = Atomic::cmpxchg(new_value, &_allocated_bitmask, allocated); 316 if (fetched == allocated) { 317 return get_pointer(index); // CAS succeeded; return entry for index. 318 } 319 allocated = fetched; // CAS failed; retry with latest value. 320 } 321 } 322 323 OopStorage::Block* OopStorage::Block::new_block(const OopStorage* owner) { 324 // _data must be first member: aligning block => aligning _data. 325 STATIC_ASSERT(_data_pos == 0); 326 size_t size_needed = allocation_size(); 327 void* memory = NEW_C_HEAP_ARRAY_RETURN_NULL(char, size_needed, mtGC); 328 if (memory == NULL) { 329 return NULL; 330 } 331 void* block_mem = align_up(memory, block_alignment); 332 assert(sizeof(Block) + pointer_delta(block_mem, memory, 1) <= size_needed, 333 "allocated insufficient space for aligned block"); 334 return ::new (block_mem) Block(owner, memory); 335 } 336 337 void OopStorage::Block::delete_block(const Block& block) { 338 void* memory = block._memory; 339 block.Block::~Block(); 340 FREE_C_HEAP_ARRAY(char, memory); 341 } 342 343 // This can return a false positive if ptr is not contained by some 344 // block. For some uses, it is a precondition that ptr is valid, 345 // e.g. contained in some block in owner's _active_array. Other uses 346 // require additional validation of the result. 347 OopStorage::Block* 348 OopStorage::Block::block_for_ptr(const OopStorage* owner, const oop* ptr) { 349 assert(CanUseSafeFetchN(), "precondition"); 350 STATIC_ASSERT(_data_pos == 0); 351 // Const-ness of ptr is not related to const-ness of containing block. 352 // Blocks are allocated section-aligned, so get the containing section. 353 oop* section_start = align_down(const_cast<oop*>(ptr), block_alignment); 354 // Start with a guess that the containing section is the last section, 355 // so the block starts section_count-1 sections earlier. 356 oop* section = section_start - (section_size * (section_count - 1)); 357 // Walk up through the potential block start positions, looking for 358 // the owner in the expected location. If we're below the actual block 359 // start position, the value at the owner position will be some oop 360 // (possibly NULL), which can never match the owner. 361 intptr_t owner_addr = reinterpret_cast<intptr_t>(owner); 362 for (unsigned i = 0; i < section_count; ++i, section += section_size) { 363 Block* candidate = reinterpret_cast<Block*>(section); 364 intptr_t* candidate_owner_addr 365 = reinterpret_cast<intptr_t*>(&candidate->_owner); 366 if (SafeFetchN(candidate_owner_addr, 0) == owner_addr) { 367 return candidate; 368 } 369 } 370 return NULL; 371 } 372 373 ////////////////////////////////////////////////////////////////////////////// 374 // Allocation 375 // 376 // Allocation involves the _allocate_list, which contains a subset of the 377 // blocks owned by a storage object. This is a doubly-linked list, linked 378 // through dedicated fields in the blocks. Full blocks are removed from this 379 // list, though they are still present in the _active_array. Empty blocks are 380 // kept at the end of the _allocate_list, to make it easy for empty block 381 // deletion to find them. 382 // 383 // allocate(), and delete_empty_blocks_concurrent() lock the 384 // _allocate_mutex while performing any list and array modifications. 385 // 386 // allocate() and release() update a block's _allocated_bitmask using CAS 387 // loops. This prevents loss of updates even though release() performs 388 // its updates without any locking. 389 // 390 // allocate() obtains the entry from the first block in the _allocate_list, 391 // and updates that block's _allocated_bitmask to indicate the entry is in 392 // use. If this makes the block full (all entries in use), the block is 393 // removed from the _allocate_list so it won't be considered by future 394 // allocations until some entries in it are released. 395 // 396 // release() is performed lock-free. release() first looks up the block for 397 // the entry, using address alignment to find the enclosing block (thereby 398 // avoiding iteration over the _active_array). Once the block has been 399 // determined, its _allocated_bitmask needs to be updated, and its position in 400 // the _allocate_list may need to be updated. There are two cases: 401 // 402 // (a) If the block is neither full nor would become empty with the release of 403 // the entry, only its _allocated_bitmask needs to be updated. But if the CAS 404 // update fails, the applicable case may change for the retry. 405 // 406 // (b) Otherwise, the _allocate_list also needs to be modified. This requires 407 // locking the _allocate_mutex. To keep the release() operation lock-free, 408 // rather than updating the _allocate_list itself, it instead performs a 409 // lock-free push of the block onto the _deferred_updates list. Entries on 410 // that list are processed by allocate() and delete_empty_blocks_XXX(), while 411 // they already hold the necessary lock. That processing makes the block's 412 // list state consistent with its current _allocated_bitmask. The block is 413 // added to the _allocate_list if not already present and the bitmask is not 414 // full. The block is moved to the end of the _allocated_list if the bitmask 415 // is empty, for ease of empty block deletion processing. 416 417 oop* OopStorage::allocate() { 418 MutexLockerEx ml(_allocate_mutex, Mutex::_no_safepoint_check_flag); 419 // Do some deferred update processing every time we allocate. 420 // Continue processing deferred updates if _allocate_list is empty, 421 // in the hope that we'll get a block from that, rather than 422 // allocating a new block. 423 while (reduce_deferred_updates() && (_allocate_list.head() == NULL)) {} 424 425 // Use the first block in _allocate_list for the allocation. 426 Block* block = _allocate_list.head(); 427 if (block == NULL) { 428 // No available blocks; make a new one, and add to storage. 429 { 430 MutexUnlockerEx mul(_allocate_mutex, Mutex::_no_safepoint_check_flag); 431 block = Block::new_block(this); 432 } 433 if (block == NULL) { 434 while (_allocate_list.head() == NULL) { 435 if (!reduce_deferred_updates()) { 436 // Failed to make new block, no other thread made a block 437 // available while the mutex was released, and didn't get 438 // one from a deferred update either, so return failure. 439 log_info(oopstorage, ref)("%s: failed block allocation", name()); 440 return NULL; 441 } 442 } 443 } else { 444 // Add new block to storage. 445 log_info(oopstorage, blocks)("%s: new block " PTR_FORMAT, name(), p2i(block)); 446 447 // Add new block to the _active_array, growing if needed. 448 if (!_active_array->push(block)) { 449 if (expand_active_array()) { 450 guarantee(_active_array->push(block), "push failed after expansion"); 451 } else { 452 log_info(oopstorage, blocks)("%s: failed active array expand", name()); 453 Block::delete_block(*block); 454 return NULL; 455 } 456 } 457 // Add to end of _allocate_list. The mutex release allowed 458 // other threads to add blocks to the _allocate_list. We prefer 459 // to allocate from non-empty blocks, to allow empty blocks to 460 // be deleted. 461 _allocate_list.push_back(*block); 462 } 463 block = _allocate_list.head(); 464 } 465 // Allocate from first block. 466 assert(block != NULL, "invariant"); 467 assert(!block->is_full(), "invariant"); 468 if (block->is_empty()) { 469 // Transitioning from empty to not empty. 470 log_debug(oopstorage, blocks)("%s: block not empty " PTR_FORMAT, name(), p2i(block)); 471 } 472 oop* result = block->allocate(); 473 assert(result != NULL, "allocation failed"); 474 assert(!block->is_empty(), "postcondition"); 475 Atomic::inc(&_allocation_count); // release updates outside lock. 476 if (block->is_full()) { 477 // Transitioning from not full to full. 478 // Remove full blocks from consideration by future allocates. 479 log_debug(oopstorage, blocks)("%s: block full " PTR_FORMAT, name(), p2i(block)); 480 _allocate_list.unlink(*block); 481 } 482 log_info(oopstorage, ref)("%s: allocated " PTR_FORMAT, name(), p2i(result)); 483 return result; 484 } 485 486 // Create a new, larger, active array with the same content as the 487 // current array, and then replace, relinquishing the old array. 488 // Return true if the array was successfully expanded, false to 489 // indicate allocation failure. 490 bool OopStorage::expand_active_array() { 491 assert_lock_strong(_allocate_mutex); 492 ActiveArray* old_array = _active_array; 493 size_t new_size = 2 * old_array->size(); 494 log_info(oopstorage, blocks)("%s: expand active array " SIZE_FORMAT, 495 name(), new_size); 496 ActiveArray* new_array = ActiveArray::create(new_size, AllocFailStrategy::RETURN_NULL); 497 if (new_array == NULL) return false; 498 new_array->copy_from(old_array); 499 replace_active_array(new_array); 500 relinquish_block_array(old_array); 501 return true; 502 } 503 504 OopStorage::ProtectActive::ProtectActive() : _enter(0), _exit() {} 505 506 // Begin read-side critical section. 507 uint OopStorage::ProtectActive::read_enter() { 508 return Atomic::add(2u, &_enter); 509 } 510 511 // End read-side critical section. 512 void OopStorage::ProtectActive::read_exit(uint enter_value) { 513 Atomic::add(2u, &_exit[enter_value & 1]); 514 } 515 516 // Wait until all readers that entered the critical section before 517 // synchronization have exited that critical section. 518 void OopStorage::ProtectActive::write_synchronize() { 519 SpinYield spinner; 520 // Determine old and new exit counters, based on bit0 of the 521 // on-entry _enter counter. 522 uint value = OrderAccess::load_acquire(&_enter); 523 volatile uint* new_ptr = &_exit[(value + 1) & 1]; 524 // Atomically change the in-use exit counter to the new counter, by 525 // adding 1 to the _enter counter (flipping bit0 between 0 and 1) 526 // and initializing the new exit counter to that enter value. Note: 527 // The new exit counter is not being used by read operations until 528 // this change succeeds. 529 uint old; 530 do { 531 old = value; 532 *new_ptr = ++value; 533 value = Atomic::cmpxchg(value, &_enter, old); 534 } while (old != value); 535 // Readers that entered the critical section before we changed the 536 // selected exit counter will use the old exit counter. Readers 537 // entering after the change will use the new exit counter. Wait 538 // for all the critical sections started before the change to 539 // complete, e.g. for the value of old_ptr to catch up with old. 540 volatile uint* old_ptr = &_exit[old & 1]; 541 while (old != OrderAccess::load_acquire(old_ptr)) { 542 spinner.wait(); 543 } 544 } 545 546 // Make new_array the _active_array. Increments new_array's refcount 547 // to account for the new reference. The assignment is atomic wrto 548 // obtain_active_array; once this function returns, it is safe for the 549 // caller to relinquish the old array. 550 void OopStorage::replace_active_array(ActiveArray* new_array) { 551 // Caller has the old array that is the current value of _active_array. 552 // Update new_array refcount to account for the new reference. 553 new_array->increment_refcount(); 554 // Install new_array, ensuring its initialization is complete first. 555 OrderAccess::release_store(&_active_array, new_array); 556 // Wait for any readers that could read the old array from _active_array. 557 _protect_active.write_synchronize(); 558 // All obtain critical sections that could see the old array have 559 // completed, having incremented the refcount of the old array. The 560 // caller can now safely relinquish the old array. 561 } 562 563 // Atomically (wrto replace_active_array) get the active array and 564 // increment its refcount. This provides safe access to the array, 565 // even if an allocate operation expands and replaces the value of 566 // _active_array. The caller must relinquish the array when done 567 // using it. 568 OopStorage::ActiveArray* OopStorage::obtain_active_array() const { 569 uint enter_value = _protect_active.read_enter(); 570 ActiveArray* result = OrderAccess::load_acquire(&_active_array); 571 result->increment_refcount(); 572 _protect_active.read_exit(enter_value); 573 return result; 574 } 575 576 // Decrement refcount of array and destroy if refcount is zero. 577 void OopStorage::relinquish_block_array(ActiveArray* array) const { 578 if (array->decrement_refcount()) { 579 assert(array != _active_array, "invariant"); 580 ActiveArray::destroy(array); 581 } 582 } 583 584 class OopStorage::WithActiveArray : public StackObj { 585 const OopStorage* _storage; 586 ActiveArray* _active_array; 587 588 public: 589 WithActiveArray(const OopStorage* storage) : 590 _storage(storage), 591 _active_array(storage->obtain_active_array()) 592 {} 593 594 ~WithActiveArray() { 595 _storage->relinquish_block_array(_active_array); 596 } 597 598 ActiveArray& active_array() const { 599 return *_active_array; 600 } 601 }; 602 603 OopStorage::Block* OopStorage::find_block_or_null(const oop* ptr) const { 604 assert(ptr != NULL, "precondition"); 605 return Block::block_for_ptr(this, ptr); 606 } 607 608 static void log_release_transitions(uintx releasing, 609 uintx old_allocated, 610 const OopStorage* owner, 611 const void* block) { 612 Log(oopstorage, blocks) log; 613 LogStream ls(log.debug()); 614 if (is_full_bitmask(old_allocated)) { 615 ls.print_cr("%s: block not full " PTR_FORMAT, owner->name(), p2i(block)); 616 } 617 if (releasing == old_allocated) { 618 ls.print_cr("%s: block empty " PTR_FORMAT, owner->name(), p2i(block)); 619 } 620 } 621 622 void OopStorage::Block::release_entries(uintx releasing, Block* volatile* deferred_list) { 623 assert(releasing != 0, "preconditon"); 624 // Prevent empty block deletion when transitioning to empty. 625 Atomic::inc(&_release_refcount); 626 627 // Atomically update allocated bitmask. 628 uintx old_allocated = _allocated_bitmask; 629 while (true) { 630 assert((releasing & ~old_allocated) == 0, "releasing unallocated entries"); 631 uintx new_value = old_allocated ^ releasing; 632 uintx fetched = Atomic::cmpxchg(new_value, &_allocated_bitmask, old_allocated); 633 if (fetched == old_allocated) break; // Successful update. 634 old_allocated = fetched; // Retry with updated bitmask. 635 } 636 637 // Now that the bitmask has been updated, if we have a state transition 638 // (updated bitmask is empty or old bitmask was full), atomically push 639 // this block onto the deferred updates list. Some future call to 640 // reduce_deferred_updates will make any needed changes related to this 641 // block and _allocate_list. This deferral avoids list updates and the 642 // associated locking here. 643 if ((releasing == old_allocated) || is_full_bitmask(old_allocated)) { 644 // Log transitions. Both transitions are possible in a single update. 645 if (log_is_enabled(Debug, oopstorage, blocks)) { 646 log_release_transitions(releasing, old_allocated, _owner, this); 647 } 648 // Attempt to claim responsibility for adding this block to the deferred 649 // list, by setting the link to non-NULL by self-looping. If this fails, 650 // then someone else has made such a claim and the deferred update has not 651 // yet been processed and will include our change, so we don't need to do 652 // anything further. 653 if (Atomic::replace_if_null(this, &_deferred_updates_next)) { 654 // Successfully claimed. Push, with self-loop for end-of-list. 655 Block* head = *deferred_list; 656 while (true) { 657 _deferred_updates_next = (head == NULL) ? this : head; 658 Block* fetched = Atomic::cmpxchg(this, deferred_list, head); 659 if (fetched == head) break; // Successful update. 660 head = fetched; // Retry with updated head. 661 } 662 log_debug(oopstorage, blocks)("%s: deferred update " PTR_FORMAT, 663 _owner->name(), p2i(this)); 664 } 665 } 666 // Release hold on empty block deletion. 667 Atomic::dec(&_release_refcount); 668 } 669 670 // Process one available deferred update. Returns true if one was processed. 671 bool OopStorage::reduce_deferred_updates() { 672 assert_locked_or_safepoint(_allocate_mutex); 673 // Atomically pop a block off the list, if any available. 674 // No ABA issue because this is only called by one thread at a time. 675 // The atomicity is wrto pushes by release(). 676 Block* block = OrderAccess::load_acquire(&_deferred_updates); 677 while (true) { 678 if (block == NULL) return false; 679 // Try atomic pop of block from list. 680 Block* tail = block->deferred_updates_next(); 681 if (block == tail) tail = NULL; // Handle self-loop end marker. 682 Block* fetched = Atomic::cmpxchg(tail, &_deferred_updates, block); 683 if (fetched == block) break; // Update successful. 684 block = fetched; // Retry with updated block. 685 } 686 block->set_deferred_updates_next(NULL); // Clear tail after updating head. 687 // Ensure bitmask read after pop is complete, including clearing tail, for 688 // ordering with release(). Without this, we may be processing a stale 689 // bitmask state here while blocking a release() operation from recording 690 // the deferred update needed for its bitmask change. 691 OrderAccess::storeload(); 692 // Process popped block. 693 uintx allocated = block->allocated_bitmask(); 694 695 // Make membership in list consistent with bitmask state. 696 if ((_allocate_list.ctail() != NULL) && 697 ((_allocate_list.ctail() == block) || 698 (_allocate_list.next(*block) != NULL))) { 699 // Block is in the allocate list. 700 assert(!is_full_bitmask(allocated), "invariant"); 701 } else if (!is_full_bitmask(allocated)) { 702 // Block is not in the allocate list, but now should be. 703 _allocate_list.push_front(*block); 704 } // Else block is full and not in list, which is correct. 705 706 // Move empty block to end of list, for possible deletion. 707 if (is_empty_bitmask(allocated)) { 708 _allocate_list.unlink(*block); 709 _allocate_list.push_back(*block); 710 } 711 712 log_debug(oopstorage, blocks)("%s: processed deferred update " PTR_FORMAT, 713 name(), p2i(block)); 714 return true; // Processed one pending update. 715 } 716 717 inline void check_release_entry(const oop* entry) { 718 assert(entry != NULL, "Releasing NULL"); 719 assert(*entry == NULL, "Releasing uncleared entry: " PTR_FORMAT, p2i(entry)); 720 } 721 722 void OopStorage::release(const oop* ptr) { 723 check_release_entry(ptr); 724 Block* block = find_block_or_null(ptr); 725 assert(block != NULL, "%s: invalid release " PTR_FORMAT, name(), p2i(ptr)); 726 log_info(oopstorage, ref)("%s: released " PTR_FORMAT, name(), p2i(ptr)); 727 block->release_entries(block->bitmask_for_entry(ptr), &_deferred_updates); 728 Atomic::dec(&_allocation_count); 729 } 730 731 void OopStorage::release(const oop* const* ptrs, size_t size) { 732 size_t i = 0; 733 while (i < size) { 734 check_release_entry(ptrs[i]); 735 Block* block = find_block_or_null(ptrs[i]); 736 assert(block != NULL, "%s: invalid release " PTR_FORMAT, name(), p2i(ptrs[i])); 737 log_info(oopstorage, ref)("%s: released " PTR_FORMAT, name(), p2i(ptrs[i])); 738 size_t count = 0; 739 uintx releasing = 0; 740 for ( ; i < size; ++i) { 741 const oop* entry = ptrs[i]; 742 check_release_entry(entry); 743 // If entry not in block, finish block and resume outer loop with entry. 744 if (!block->contains(entry)) break; 745 // Add entry to releasing bitmap. 746 log_info(oopstorage, ref)("%s: released " PTR_FORMAT, name(), p2i(entry)); 747 uintx entry_bitmask = block->bitmask_for_entry(entry); 748 assert((releasing & entry_bitmask) == 0, 749 "Duplicate entry: " PTR_FORMAT, p2i(entry)); 750 releasing |= entry_bitmask; 751 ++count; 752 } 753 // Release the contiguous entries that are in block. 754 block->release_entries(releasing, &_deferred_updates); 755 Atomic::sub(count, &_allocation_count); 756 } 757 } 758 759 const char* dup_name(const char* name) { 760 char* dup = NEW_C_HEAP_ARRAY(char, strlen(name) + 1, mtGC); 761 strcpy(dup, name); 762 return dup; 763 } 764 765 const size_t initial_active_array_size = 8; 766 767 OopStorage::OopStorage(const char* name, 768 Mutex* allocate_mutex, 769 Mutex* active_mutex) : 770 _name(dup_name(name)), 771 _active_array(ActiveArray::create(initial_active_array_size)), 772 _allocate_list(&Block::get_allocate_entry), 773 _deferred_updates(NULL), 774 _allocate_mutex(allocate_mutex), 775 _active_mutex(active_mutex), 776 _allocation_count(0), 777 _concurrent_iteration_active(false) 778 { 779 _active_array->increment_refcount(); 780 assert(_active_mutex->rank() < _allocate_mutex->rank(), 781 "%s: active_mutex must have lower rank than allocate_mutex", _name); 782 assert(_active_mutex->_safepoint_check_required != Mutex::_safepoint_check_always, 783 "%s: active mutex requires safepoint check", _name); 784 assert(_allocate_mutex->_safepoint_check_required != Mutex::_safepoint_check_always, 785 "%s: allocate mutex requires safepoint check", _name); 786 } 787 788 void OopStorage::delete_empty_block(const Block& block) { 789 assert(block.is_empty(), "discarding non-empty block"); 790 log_info(oopstorage, blocks)("%s: delete empty block " PTR_FORMAT, name(), p2i(&block)); 791 Block::delete_block(block); 792 } 793 794 OopStorage::~OopStorage() { 795 Block* block; 796 while ((block = _deferred_updates) != NULL) { 797 _deferred_updates = block->deferred_updates_next(); 798 block->set_deferred_updates_next(NULL); 799 } 800 while ((block = _allocate_list.head()) != NULL) { 801 _allocate_list.unlink(*block); 802 } 803 bool unreferenced = _active_array->decrement_refcount(); 804 assert(unreferenced, "deleting storage while _active_array is referenced"); 805 for (size_t i = _active_array->block_count(); 0 < i; ) { 806 block = _active_array->at(--i); 807 Block::delete_block(*block); 808 } 809 ActiveArray::destroy(_active_array); 810 FREE_C_HEAP_ARRAY(char, _name); 811 } 812 813 void OopStorage::delete_empty_blocks_safepoint() { 814 assert_at_safepoint(); 815 // Process any pending release updates, which may make more empty 816 // blocks available for deletion. 817 while (reduce_deferred_updates()) {} 818 // Don't interfere with a concurrent iteration. 819 if (_concurrent_iteration_active) return; 820 // Delete empty (and otherwise deletable) blocks from end of _allocate_list. 821 for (Block* block = _allocate_list.tail(); 822 (block != NULL) && block->is_deletable(); 823 block = _allocate_list.tail()) { 824 _active_array->remove(block); 825 _allocate_list.unlink(*block); 826 delete_empty_block(*block); 827 } 828 } 829 830 void OopStorage::delete_empty_blocks_concurrent() { 831 MutexLockerEx ml(_allocate_mutex, Mutex::_no_safepoint_check_flag); 832 // Other threads could be adding to the empty block count while we 833 // release the mutex across the block deletions. Set an upper bound 834 // on how many blocks we'll try to release, so other threads can't 835 // cause an unbounded stay in this function. 836 size_t limit = block_count(); 837 838 for (size_t i = 0; i < limit; ++i) { 839 // Additional updates might become available while we dropped the 840 // lock. But limit number processed to limit lock duration. 841 reduce_deferred_updates(); 842 843 Block* block = _allocate_list.tail(); 844 if ((block == NULL) || !block->is_deletable()) { 845 // No block to delete, so done. There could be more pending 846 // deferred updates that could give us more work to do; deal with 847 // that in some later call, to limit lock duration here. 848 return; 849 } 850 851 { 852 MutexLockerEx aml(_active_mutex, Mutex::_no_safepoint_check_flag); 853 // Don't interfere with a concurrent iteration. 854 if (_concurrent_iteration_active) return; 855 _active_array->remove(block); 856 } 857 // Remove block from _allocate_list and delete it. 858 _allocate_list.unlink(*block); 859 // Release mutex while deleting block. 860 MutexUnlockerEx ul(_allocate_mutex, Mutex::_no_safepoint_check_flag); 861 delete_empty_block(*block); 862 } 863 } 864 865 OopStorage::EntryStatus OopStorage::allocation_status(const oop* ptr) const { 866 const Block* block = find_block_or_null(ptr); 867 if (block != NULL) { 868 // Prevent block deletion and _active_array modification. 869 MutexLockerEx ml(_allocate_mutex, Mutex::_no_safepoint_check_flag); 870 // Block could be a false positive, so get index carefully. 871 size_t index = Block::active_index_safe(block); 872 if ((index < _active_array->block_count()) && 873 (block == _active_array->at(index)) && 874 block->contains(ptr)) { 875 if ((block->allocated_bitmask() & block->bitmask_for_entry(ptr)) != 0) { 876 return ALLOCATED_ENTRY; 877 } else { 878 return UNALLOCATED_ENTRY; 879 } 880 } 881 } 882 return INVALID_ENTRY; 883 } 884 885 size_t OopStorage::allocation_count() const { 886 return _allocation_count; 887 } 888 889 size_t OopStorage::block_count() const { 890 WithActiveArray wab(this); 891 // Count access is racy, but don't care. 892 return wab.active_array().block_count(); 893 } 894 895 size_t OopStorage::total_memory_usage() const { 896 size_t total_size = sizeof(OopStorage); 897 total_size += strlen(name()) + 1; 898 total_size += sizeof(ActiveArray); 899 WithActiveArray wab(this); 900 const ActiveArray& blocks = wab.active_array(); 901 // Count access is racy, but don't care. 902 total_size += blocks.block_count() * Block::allocation_size(); 903 total_size += blocks.size() * sizeof(Block*); 904 return total_size; 905 } 906 907 // Parallel iteration support 908 909 uint OopStorage::BasicParState::default_estimated_thread_count(bool concurrent) { 910 uint configured = concurrent ? ConcGCThreads : ParallelGCThreads; 911 return MAX2(1u, configured); // Never estimate zero threads. 912 } 913 914 OopStorage::BasicParState::BasicParState(const OopStorage* storage, 915 uint estimated_thread_count, 916 bool concurrent) : 917 _storage(storage), 918 _active_array(_storage->obtain_active_array()), 919 _block_count(0), // initialized properly below 920 _next_block(0), 921 _estimated_thread_count(estimated_thread_count), 922 _concurrent(concurrent) 923 { 924 assert(estimated_thread_count > 0, "estimated thread count must be positive"); 925 update_iteration_state(true); 926 // Get the block count *after* iteration state updated, so concurrent 927 // empty block deletion is suppressed and can't reduce the count. But 928 // ensure the count we use was written after the block with that count 929 // was fully initialized; see ActiveArray::push. 930 _block_count = _active_array->block_count_acquire(); 931 } 932 933 OopStorage::BasicParState::~BasicParState() { 934 _storage->relinquish_block_array(_active_array); 935 update_iteration_state(false); 936 } 937 938 void OopStorage::BasicParState::update_iteration_state(bool value) { 939 if (_concurrent) { 940 MutexLockerEx ml(_storage->_active_mutex, Mutex::_no_safepoint_check_flag); 941 assert(_storage->_concurrent_iteration_active != value, "precondition"); 942 _storage->_concurrent_iteration_active = value; 943 } 944 } 945 946 bool OopStorage::BasicParState::claim_next_segment(IterationData* data) { 947 data->_processed += data->_segment_end - data->_segment_start; 948 size_t start = OrderAccess::load_acquire(&_next_block); 949 if (start >= _block_count) { 950 return finish_iteration(data); // No more blocks available. 951 } 952 // Try to claim several at a time, but not *too* many. We want to 953 // avoid deciding there are many available and selecting a large 954 // quantity, get delayed, and then end up claiming most or all of 955 // the remaining largish amount of work, leaving nothing for other 956 // threads to do. But too small a step can lead to contention 957 // over _next_block, esp. when the work per block is small. 958 size_t max_step = 10; 959 size_t remaining = _block_count - start; 960 size_t step = MIN2(max_step, 1 + (remaining / _estimated_thread_count)); 961 // Atomic::add with possible overshoot. This can perform better 962 // than a CAS loop on some platforms when there is contention. 963 // We can cope with the uncertainty by recomputing start/end from 964 // the result of the add, and dealing with potential overshoot. 965 size_t end = Atomic::add(step, &_next_block); 966 // _next_block may have changed, so recompute start from result of add. 967 start = end - step; 968 // _next_block may have changed so much that end has overshot. 969 end = MIN2(end, _block_count); 970 // _next_block may have changed so much that even start has overshot. 971 if (start < _block_count) { 972 // Record claimed segment for iteration. 973 data->_segment_start = start; 974 data->_segment_end = end; 975 return true; // Success. 976 } else { 977 // No more blocks to claim. 978 return finish_iteration(data); 979 } 980 } 981 982 bool OopStorage::BasicParState::finish_iteration(const IterationData* data) const { 983 log_debug(oopstorage, blocks, stats) 984 ("Parallel iteration on %s: blocks = " SIZE_FORMAT 985 ", processed = " SIZE_FORMAT " (%2.f%%)", 986 _storage->name(), _block_count, data->_processed, 987 percent_of(data->_processed, _block_count)); 988 return false; 989 } 990 991 const char* OopStorage::name() const { return _name; } 992 993 #ifndef PRODUCT 994 995 void OopStorage::print_on(outputStream* st) const { 996 size_t allocations = _allocation_count; 997 size_t blocks = _active_array->block_count(); 998 999 double data_size = section_size * section_count; 1000 double alloc_percentage = percent_of((double)allocations, blocks * data_size); 1001 1002 st->print("%s: " SIZE_FORMAT " entries in " SIZE_FORMAT " blocks (%.F%%), " SIZE_FORMAT " bytes", 1003 name(), allocations, blocks, alloc_percentage, total_memory_usage()); 1004 if (_concurrent_iteration_active) { 1005 st->print(", concurrent iteration active"); 1006 } 1007 } 1008 1009 #endif // !PRODUCT