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