1/*
2 * Copyright (c) 2018, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25#include "precompiled.hpp"
26#include "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/interfaceSupport.inline.hpp"
35#include "runtime/mutex.hpp"
36#include "runtime/mutexLocker.hpp"
37#include "runtime/orderAccess.hpp"
38#include "runtime/os.hpp"
39#include "runtime/safepoint.hpp"
40#include "runtime/stubRoutines.hpp"
41#include "runtime/thread.hpp"
42#include "utilities/align.hpp"
43#include "utilities/count_trailing_zeros.hpp"
44#include "utilities/debug.hpp"
45#include "utilities/globalDefinitions.hpp"
46#include "utilities/macros.hpp"
47#include "utilities/ostream.hpp"
48
49OopStorage::AllocationListEntry::AllocationListEntry() : _prev(NULL), _next(NULL) {}
50
51OopStorage::AllocationListEntry::~AllocationListEntry() {
52 assert(_prev == NULL, "deleting attached block");
53 assert(_next == NULL, "deleting attached block");
54}
55
56OopStorage::AllocationList::AllocationList() : _head(NULL), _tail(NULL) {}
57
58OopStorage::AllocationList::~AllocationList() {
59 // ~OopStorage() empties its lists before destroying them.
60 assert(_head == NULL, "deleting non-empty block list");
61 assert(_tail == NULL, "deleting non-empty block list");
62}
63
64void OopStorage::AllocationList::push_front(const Block& block) {
65 const Block* old = _head;
66 if (old == NULL) {
67 assert(_tail == NULL, "invariant");
68 _head = _tail = █
69 } else {
70 block.allocation_list_entry()._next = old;
71 old->allocation_list_entry()._prev = █
72 _head = █
73 }
74}
75
76void OopStorage::AllocationList::push_back(const Block& block) {
77 const Block* old = _tail;
78 if (old == NULL) {
79 assert(_head == NULL, "invariant");
80 _head = _tail = █
81 } else {
82 old->allocation_list_entry()._next = █
83 block.allocation_list_entry()._prev = old;
84 _tail = █
85 }
86}
87
88void OopStorage::AllocationList::unlink(const Block& block) {
89 const AllocationListEntry& block_entry = block.allocation_list_entry();
90 const Block* prev_blk = block_entry._prev;
91 const Block* next_blk = block_entry._next;
92 block_entry._prev = NULL;
93 block_entry._next = NULL;
94 if ((prev_blk == NULL) && (next_blk == NULL)) {
95 assert(_head == &block, "invariant");
96 assert(_tail == &block, "invariant");
97 _head = _tail = NULL;
98 } else if (prev_blk == NULL) {
99 assert(_head == &block, "invariant");
100 next_blk->allocation_list_entry()._prev = NULL;
101 _head = next_blk;
102 } else if (next_blk == NULL) {
103 assert(_tail == &block, "invariant");
104 prev_blk->allocation_list_entry()._next = NULL;
105 _tail = prev_blk;
106 } else {
107 next_blk->allocation_list_entry()._prev = prev_blk;
108 prev_blk->allocation_list_entry()._next = next_blk;
109 }
110}
111
112OopStorage::ActiveArray::ActiveArray(size_t size) :
113 _size(size),
114 _block_count(0),
115 _refcount(0)
116{}
117
118OopStorage::ActiveArray::~ActiveArray() {
119 assert(_refcount == 0, "precondition");
120}
121
122OopStorage::ActiveArray* OopStorage::ActiveArray::create(size_t size, AllocFailType alloc_fail) {
123 size_t size_in_bytes = blocks_offset() + sizeof(Block*) * size;
124 void* mem = NEW_C_HEAP_ARRAY3(char, size_in_bytes, mtGC, CURRENT_PC, alloc_fail);
125 if (mem == NULL) return NULL;
126 return new (mem) ActiveArray(size);
127}
128
129void OopStorage::ActiveArray::destroy(ActiveArray* ba) {
130 ba->~ActiveArray();
131 FREE_C_HEAP_ARRAY(char, ba);
132}
133
134size_t OopStorage::ActiveArray::size() const {
135 return _size;
136}
137
138size_t OopStorage::ActiveArray::block_count() const {
139 return _block_count;
140}
141
142size_t OopStorage::ActiveArray::block_count_acquire() const {
143 return OrderAccess::load_acquire(&_block_count);
144}
145
146void OopStorage::ActiveArray::increment_refcount() const {
147 int new_value = Atomic::add(1, &_refcount);
148 assert(new_value >= 1, "negative refcount %d", new_value - 1);
149}
150
151bool OopStorage::ActiveArray::decrement_refcount() const {
152 int new_value = Atomic::sub(1, &_refcount);
153 assert(new_value >= 0, "negative refcount %d", new_value);
154 return new_value == 0;
155}
156
157bool OopStorage::ActiveArray::push(Block* block) {
158 size_t index = _block_count;
159 if (index < _size) {
160 block->set_active_index(index);
161 *block_ptr(index) = block;
162 // Use a release_store to ensure all the setup is complete before
163 // making the block visible.
164 OrderAccess::release_store(&_block_count, index + 1);
165 return true;
166 } else {
167 return false;
168 }
169}
170
171void OopStorage::ActiveArray::remove(Block* block) {
172 assert(_block_count > 0, "array is empty");
173 size_t index = block->active_index();
174 assert(*block_ptr(index) == block, "block not present");
175 size_t last_index = _block_count - 1;
176 Block* last_block = *block_ptr(last_index);
177 last_block->set_active_index(index);
178 *block_ptr(index) = last_block;
179 _block_count = last_index;
180}
181
182void OopStorage::ActiveArray::copy_from(const ActiveArray* from) {
183 assert(_block_count == 0, "array must be empty");
184 size_t count = from->_block_count;
185 assert(count <= _size, "precondition");
186 Block* const* from_ptr = from->block_ptr(0);
187 Block** to_ptr = block_ptr(0);
188 for (size_t i = 0; i < count; ++i) {
189 Block* block = *from_ptr++;
190 assert(block->active_index() == i, "invariant");
191 *to_ptr++ = block;
192 }
193 _block_count = count;
194}
195
196// Blocks start with an array of BitsPerWord oop entries. That array
197// is divided into conceptual BytesPerWord sections of BitsPerByte
198// entries. Blocks are allocated aligned on section boundaries, for
199// the convenience of mapping from an entry to the containing block;
200// see block_for_ptr(). Aligning on section boundary rather than on
201// the full _data wastes a lot less space, but makes for a bit more
202// work in block_for_ptr().
203
204const unsigned section_size = BitsPerByte;
205const unsigned section_count = BytesPerWord;
206const unsigned block_alignment = sizeof(oop) * section_size;
207
208OopStorage::Block::Block(const OopStorage* owner, void* memory) :
209 _data(),
210 _allocated_bitmask(0),
211 _owner(owner),
212 _memory(memory),
213 _active_index(0),
214 _allocation_list_entry(),
215 _deferred_updates_next(NULL),
216 _release_refcount(0)
217{
218 STATIC_ASSERT(_data_pos == 0);
219 STATIC_ASSERT(section_size * section_count == ARRAY_SIZE(_data));
220 assert(offset_of(Block, _data) == _data_pos, "invariant");
221 assert(owner != NULL, "NULL owner");
222 assert(is_aligned(this, block_alignment), "misaligned block");
223}
224
225OopStorage::Block::~Block() {
226 assert(_release_refcount == 0, "deleting block while releasing");
227 assert(_deferred_updates_next == NULL, "deleting block with deferred update");
228 // Clear fields used by block_for_ptr and entry validation, which
229 // might help catch bugs. Volatile to prevent dead-store elimination.
230 const_cast<uintx volatile&>(_allocated_bitmask) = 0;
231 const_cast<OopStorage* volatile&>(_owner) = NULL;
232}
233
234size_t OopStorage::Block::allocation_size() {
235 // _data must be first member, so aligning Block aligns _data.
236 STATIC_ASSERT(_data_pos == 0);
237 return sizeof(Block) + block_alignment - sizeof(void*);
238}
239
240size_t OopStorage::Block::allocation_alignment_shift() {
241 return exact_log2(block_alignment);
242}
243
244inline bool is_full_bitmask(uintx bitmask) { return ~bitmask == 0; }
245inline bool is_empty_bitmask(uintx bitmask) { return bitmask == 0; }
246
247bool OopStorage::Block::is_full() const {
248 return is_full_bitmask(allocated_bitmask());
249}
250
251bool OopStorage::Block::is_empty() const {
252 return is_empty_bitmask(allocated_bitmask());
253}
254
255uintx OopStorage::Block::bitmask_for_entry(const oop* ptr) const {
256 return bitmask_for_index(get_index(ptr));
257}
258
259// An empty block is not yet deletable if either:
260// (1) There is a release() operation currently operating on it.
261// (2) It is in the deferred updates list.
262// For interaction with release(), these must follow the empty check,
263// and the order of these checks is important.
264bool OopStorage::Block::is_safe_to_delete() const {
265 assert(is_empty(), "precondition");
266 OrderAccess::loadload();
267 return (OrderAccess::load_acquire(&_release_refcount) == 0) &&
268 (OrderAccess::load_acquire(&_deferred_updates_next) == NULL);
269}
270
271OopStorage::Block* OopStorage::Block::deferred_updates_next() const {
272 return _deferred_updates_next;
273}
274
275void OopStorage::Block::set_deferred_updates_next(Block* block) {
276 _deferred_updates_next = block;
277}
278
279bool OopStorage::Block::contains(const oop* ptr) const {
280 const oop* base = get_pointer(0);
281 return (base <= ptr) && (ptr < (base + ARRAY_SIZE(_data)));
282}
283
284size_t OopStorage::Block::active_index() const {
285 return _active_index;
286}
287
288void OopStorage::Block::set_active_index(size_t index) {
289 _active_index = index;
290}
291
292size_t OopStorage::Block::active_index_safe(const Block* block) {
293 STATIC_ASSERT(sizeof(intptr_t) == sizeof(block->_active_index));
294 assert(CanUseSafeFetchN(), "precondition");
295 return SafeFetchN((intptr_t*)&block->_active_index, 0);
296}
297
298unsigned OopStorage::Block::get_index(const oop* ptr) const {
299 assert(contains(ptr), PTR_FORMAT " not in block " PTR_FORMAT, p2i(ptr), p2i(this));
300 return static_cast<unsigned>(ptr - get_pointer(0));
301}
302
303oop* OopStorage::Block::allocate() {
304 // Use CAS loop because release may change bitmask outside of lock.
305 uintx allocated = allocated_bitmask();
306 while (true) {
307 assert(!is_full_bitmask(allocated), "attempt to allocate from full block");
308 unsigned index = count_trailing_zeros(~allocated);
309 uintx new_value = allocated | bitmask_for_index(index);
310 uintx fetched = Atomic::cmpxchg(new_value, &_allocated_bitmask, allocated);
311 if (fetched == allocated) {
312 return get_pointer(index); // CAS succeeded; return entry for index.
313 }
314 allocated = fetched; // CAS failed; retry with latest value.
315 }
316}
317
318OopStorage::Block* OopStorage::Block::new_block(const OopStorage* owner) {
319 // _data must be first member: aligning block => aligning _data.
320 STATIC_ASSERT(_data_pos == 0);
321 size_t size_needed = allocation_size();
322 void* memory = NEW_C_HEAP_ARRAY_RETURN_NULL(char, size_needed, mtGC);
323 if (memory == NULL) {
324 return NULL;
325 }
326 void* block_mem = align_up(memory, block_alignment);
327 assert(sizeof(Block) + pointer_delta(block_mem, memory, 1) <= size_needed,
328 "allocated insufficient space for aligned block");
329 return ::new (block_mem) Block(owner, memory);
330}
331
332void OopStorage::Block::delete_block(const Block& block) {
333 void* memory = block._memory;
334 block.Block::~Block();
335 FREE_C_HEAP_ARRAY(char, memory);
336}
337
338// This can return a false positive if ptr is not contained by some
339// block. For some uses, it is a precondition that ptr is valid,
340// e.g. contained in some block in owner's _active_array. Other uses
341// require additional validation of the result.
342OopStorage::Block*
343OopStorage::Block::block_for_ptr(const OopStorage* owner, const oop* ptr) {
344 assert(CanUseSafeFetchN(), "precondition");
345 STATIC_ASSERT(_data_pos == 0);
346 // Const-ness of ptr is not related to const-ness of containing block.
347 // Blocks are allocated section-aligned, so get the containing section.
348 oop* section_start = align_down(const_cast<oop*>(ptr), block_alignment);
349 // Start with a guess that the containing section is the last section,
350 // so the block starts section_count-1 sections earlier.
351 oop* section = section_start - (section_size * (section_count - 1));
352 // Walk up through the potential block start positions, looking for
353 // the owner in the expected location. If we're below the actual block
354 // start position, the value at the owner position will be some oop
355 // (possibly NULL), which can never match the owner.
356 intptr_t owner_addr = reinterpret_cast<intptr_t>(owner);
357 for (unsigned i = 0; i < section_count; ++i, section += section_size) {
358 Block* candidate = reinterpret_cast<Block*>(section);
359 intptr_t* candidate_owner_addr
360 = reinterpret_cast<intptr_t*>(&candidate->_owner);
361 if (SafeFetchN(candidate_owner_addr, 0) == owner_addr) {
362 return candidate;
363 }
364 }
365 return NULL;
366}
367
368//////////////////////////////////////////////////////////////////////////////
369// Allocation
370//
371// Allocation involves the _allocation_list, which contains a subset of the
372// blocks owned by a storage object. This is a doubly-linked list, linked
373// through dedicated fields in the blocks. Full blocks are removed from this
374// list, though they are still present in the _active_array. Empty blocks are
375// kept at the end of the _allocation_list, to make it easy for empty block
376// deletion to find them.
377//
378// allocate(), and delete_empty_blocks() lock the
379// _allocation_mutex while performing any list and array modifications.
380//
381// allocate() and release() update a block's _allocated_bitmask using CAS
382// loops. This prevents loss of updates even though release() performs
383// its updates without any locking.
384//
385// allocate() obtains the entry from the first block in the _allocation_list,
386// and updates that block's _allocated_bitmask to indicate the entry is in
387// use. If this makes the block full (all entries in use), the block is
388// removed from the _allocation_list so it won't be considered by future
389// allocations until some entries in it are released.
390//
391// release() is performed lock-free. (Note: This means it can't notify the
392// service thread of pending cleanup work. It must be lock-free because
393// it is called in all kinds of contexts where even quite low ranked locks
394// may be held.) release() first looks up the block for
395// the entry, using address alignment to find the enclosing block (thereby
396// avoiding iteration over the _active_array). Once the block has been
397// determined, its _allocated_bitmask needs to be updated, and its position in
398// the _allocation_list may need to be updated. There are two cases:
399//
400// (a) If the block is neither full nor would become empty with the release of
401// the entry, only its _allocated_bitmask needs to be updated. But if the CAS
402// update fails, the applicable case may change for the retry.
403//
404// (b) Otherwise, the _allocation_list also needs to be modified. This requires
405// locking the _allocation_mutex. To keep the release() operation lock-free,
406// rather than updating the _allocation_list itself, it instead performs a
407// lock-free push of the block onto the _deferred_updates list. Entries on
408// that list are processed by allocate() and delete_empty_blocks(), while
409// they already hold the necessary lock. That processing makes the block's
410// list state consistent with its current _allocated_bitmask. The block is
411// added to the _allocation_list if not already present and the bitmask is not
412// full. The block is moved to the end of the _allocation_list if the bitmask
413// is empty, for ease of empty block deletion processing.
414
415oop* OopStorage::allocate() {
416 MutexLocker ml(_allocation_mutex, Mutex::_no_safepoint_check_flag);
417
418 Block* block = block_for_allocation();
419 if (block == NULL) return NULL; // Block allocation failed.
420 assert(!block->is_full(), "invariant");
421 if (block->is_empty()) {
422 // Transitioning from empty to not empty.
423 log_debug(oopstorage, blocks)("%s: block not empty " PTR_FORMAT, name(), p2i(block));
424 }
425 oop* result = block->allocate();
426 assert(result != NULL, "allocation failed");
427 assert(!block->is_empty(), "postcondition");
428 Atomic::inc(&_allocation_count); // release updates outside lock.
429 if (block->is_full()) {
430 // Transitioning from not full to full.
431 // Remove full blocks from consideration by future allocates.
432 log_debug(oopstorage, blocks)("%s: block full " PTR_FORMAT, name(), p2i(block));
433 _allocation_list.unlink(*block);
434 }
435 log_trace(oopstorage, ref)("%s: allocated " PTR_FORMAT, name(), p2i(result));
436 return result;
437}
438
439bool OopStorage::try_add_block() {
440 assert_lock_strong(_allocation_mutex);
441 Block* block;
442 {
443 MutexUnlocker ul(_allocation_mutex, Mutex::_no_safepoint_check_flag);
444 block = Block::new_block(this);
445 }
446 if (block == NULL) return false;
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_debug(oopstorage, blocks)("%s: failed active array expand", name());
454 Block::delete_block(*block);
455 return false;
456 }
457 }
458 // Add to end of _allocation_list. The mutex release allowed other
459 // threads to add blocks to the _allocation_list. We prefer to
460 // allocate from non-empty blocks, to allow empty blocks to be
461 // deleted. But we don't bother notifying about the empty block
462 // because we're (probably) about to allocate an entry from it.
463 _allocation_list.push_back(*block);
464 log_debug(oopstorage, blocks)("%s: new block " PTR_FORMAT, name(), p2i(block));
465 return true;
466}
467
468OopStorage::Block* OopStorage::block_for_allocation() {
469 assert_lock_strong(_allocation_mutex);
470 while (true) {
471 // Use the first block in _allocation_list for the allocation.
472 Block* block = _allocation_list.head();
473 if (block != NULL) {
474 return block;
475 } else if (reduce_deferred_updates()) {
476 // Might have added a block to the _allocation_list, so retry.
477 } else if (try_add_block()) {
478 // Successfully added a new block to the list, so retry.
479 assert(_allocation_list.chead() != NULL, "invariant");
480 } else if (_allocation_list.chead() != NULL) {
481 // Trying to add a block failed, but some other thread added to the
482 // list while we'd dropped the lock over the new block allocation.
483 } else if (!reduce_deferred_updates()) { // Once more before failure.
484 // Attempt to add a block failed, no other thread added a block,
485 // and no deferred updated added a block, then allocation failed.
486 log_debug(oopstorage, blocks)("%s: failed block allocation", name());
487 return NULL;
488 }
489 }
490}
491
492// Create a new, larger, active array with the same content as the
493// current array, and then replace, relinquishing the old array.
494// Return true if the array was successfully expanded, false to
495// indicate allocation failure.
496bool OopStorage::expand_active_array() {
497 assert_lock_strong(_allocation_mutex);
498 ActiveArray* old_array = _active_array;
499 size_t new_size = 2 * old_array->size();
500 log_debug(oopstorage, blocks)("%s: expand active array " SIZE_FORMAT,
501 name(), new_size);
502 ActiveArray* new_array = ActiveArray::create(new_size, AllocFailStrategy::RETURN_NULL);
503 if (new_array == NULL) return false;
504 new_array->copy_from(old_array);
505 replace_active_array(new_array);
506 relinquish_block_array(old_array);
507 return true;
508}
509
510// Make new_array the _active_array. Increments new_array's refcount
511// to account for the new reference. The assignment is atomic wrto
512// obtain_active_array; once this function returns, it is safe for the
513// caller to relinquish the old array.
514void OopStorage::replace_active_array(ActiveArray* new_array) {
515 // Caller has the old array that is the current value of _active_array.
516 // Update new_array refcount to account for the new reference.
517 new_array->increment_refcount();
518 // Install new_array, ensuring its initialization is complete first.
519 OrderAccess::release_store(&_active_array, new_array);
520 // Wait for any readers that could read the old array from _active_array.
521 // Can't use GlobalCounter here, because this is called from allocate(),
522 // which may be called in the scope of a GlobalCounter critical section
523 // when inserting a StringTable entry.
524 _protect_active.synchronize();
525 // All obtain critical sections that could see the old array have
526 // completed, having incremented the refcount of the old array. The
527 // caller can now safely relinquish the old array.
528}
529
530// Atomically (wrto replace_active_array) get the active array and
531// increment its refcount. This provides safe access to the array,
532// even if an allocate operation expands and replaces the value of
533// _active_array. The caller must relinquish the array when done
534// using it.
535OopStorage::ActiveArray* OopStorage::obtain_active_array() const {
536 SingleWriterSynchronizer::CriticalSection cs(&_protect_active);
537 ActiveArray* result = OrderAccess::load_acquire(&_active_array);
538 result->increment_refcount();
539 return result;
540}
541
542// Decrement refcount of array and destroy if refcount is zero.
543void OopStorage::relinquish_block_array(ActiveArray* array) const {
544 if (array->decrement_refcount()) {
545 assert(array != _active_array, "invariant");
546 ActiveArray::destroy(array);
547 }
548}
549
550class OopStorage::WithActiveArray : public StackObj {
551 const OopStorage* _storage;
552 ActiveArray* _active_array;
553
554public:
555 WithActiveArray(const OopStorage* storage) :
556 _storage(storage),
557 _active_array(storage->obtain_active_array())
558 {}
559
560 ~WithActiveArray() {
561 _storage->relinquish_block_array(_active_array);
562 }
563
564 ActiveArray& active_array() const {
565 return *_active_array;
566 }
567};
568
569OopStorage::Block* OopStorage::find_block_or_null(const oop* ptr) const {
570 assert(ptr != NULL, "precondition");
571 return Block::block_for_ptr(this, ptr);
572}
573
574static void log_release_transitions(uintx releasing,
575 uintx old_allocated,
576 const OopStorage* owner,
577 const void* block) {
578 Log(oopstorage, blocks) log;
579 LogStream ls(log.debug());
580 if (is_full_bitmask(old_allocated)) {
581 ls.print_cr("%s: block not full " PTR_FORMAT, owner->name(), p2i(block));
582 }
583 if (releasing == old_allocated) {
584 ls.print_cr("%s: block empty " PTR_FORMAT, owner->name(), p2i(block));
585 }
586}
587
588void OopStorage::Block::release_entries(uintx releasing, OopStorage* owner) {
589 assert(releasing != 0, "preconditon");
590 // Prevent empty block deletion when transitioning to empty.
591 Atomic::inc(&_release_refcount);
592
593 // Atomically update allocated bitmask.
594 uintx old_allocated = _allocated_bitmask;
595 while (true) {
596 assert((releasing & ~old_allocated) == 0, "releasing unallocated entries");
597 uintx new_value = old_allocated ^ releasing;
598 uintx fetched = Atomic::cmpxchg(new_value, &_allocated_bitmask, old_allocated);
599 if (fetched == old_allocated) break; // Successful update.
600 old_allocated = fetched; // Retry with updated bitmask.
601 }
602
603 // Now that the bitmask has been updated, if we have a state transition
604 // (updated bitmask is empty or old bitmask was full), atomically push
605 // this block onto the deferred updates list. Some future call to
606 // reduce_deferred_updates will make any needed changes related to this
607 // block and _allocation_list. This deferral avoids _allocation_list
608 // updates and the associated locking here.
609 if ((releasing == old_allocated) || is_full_bitmask(old_allocated)) {
610 // Log transitions. Both transitions are possible in a single update.
611 if (log_is_enabled(Debug, oopstorage, blocks)) {
612 log_release_transitions(releasing, old_allocated, _owner, this);
613 }
614 // Attempt to claim responsibility for adding this block to the deferred
615 // list, by setting the link to non-NULL by self-looping. If this fails,
616 // then someone else has made such a claim and the deferred update has not
617 // yet been processed and will include our change, so we don't need to do
618 // anything further.
619 if (Atomic::replace_if_null(this, &_deferred_updates_next)) {
620 // Successfully claimed. Push, with self-loop for end-of-list.
621 Block* head = owner->_deferred_updates;
622 while (true) {
623 _deferred_updates_next = (head == NULL) ? this : head;
624 Block* fetched = Atomic::cmpxchg(this, &owner->_deferred_updates, head);
625 if (fetched == head) break; // Successful update.
626 head = fetched; // Retry with updated head.
627 }
628 // Only request cleanup for to-empty transitions, not for from-full.
629 // There isn't any rush to process from-full transitions. Allocation
630 // will reduce deferrals before allocating new blocks, so may process
631 // some. And the service thread will drain the entire deferred list
632 // if there are any pending to-empty transitions.
633 if (releasing == old_allocated) {
634 owner->record_needs_cleanup();
635 }
636 log_debug(oopstorage, blocks)("%s: deferred update " PTR_FORMAT,
637 _owner->name(), p2i(this));
638 }
639 }
640 // Release hold on empty block deletion.
641 Atomic::dec(&_release_refcount);
642}
643
644// Process one available deferred update. Returns true if one was processed.
645bool OopStorage::reduce_deferred_updates() {
646 assert_lock_strong(_allocation_mutex);
647 // Atomically pop a block off the list, if any available.
648 // No ABA issue because this is only called by one thread at a time.
649 // The atomicity is wrto pushes by release().
650 Block* block = OrderAccess::load_acquire(&_deferred_updates);
651 while (true) {
652 if (block == NULL) return false;
653 // Try atomic pop of block from list.
654 Block* tail = block->deferred_updates_next();
655 if (block == tail) tail = NULL; // Handle self-loop end marker.
656 Block* fetched = Atomic::cmpxchg(tail, &_deferred_updates, block);
657 if (fetched == block) break; // Update successful.
658 block = fetched; // Retry with updated block.
659 }
660 block->set_deferred_updates_next(NULL); // Clear tail after updating head.
661 // Ensure bitmask read after pop is complete, including clearing tail, for
662 // ordering with release(). Without this, we may be processing a stale
663 // bitmask state here while blocking a release() operation from recording
664 // the deferred update needed for its bitmask change.
665 OrderAccess::fence();
666 // Process popped block.
667 uintx allocated = block->allocated_bitmask();
668
669 // Make membership in list consistent with bitmask state.
670 if ((_allocation_list.ctail() != NULL) &&
671 ((_allocation_list.ctail() == block) ||
672 (_allocation_list.next(*block) != NULL))) {
673 // Block is in the _allocation_list.
674 assert(!is_full_bitmask(allocated), "invariant");
675 } else if (!is_full_bitmask(allocated)) {
676 // Block is not in the _allocation_list, but now should be.
677 _allocation_list.push_front(*block);
678 } // Else block is full and not in list, which is correct.
679
680 // Move empty block to end of list, for possible deletion.
681 if (is_empty_bitmask(allocated)) {
682 _allocation_list.unlink(*block);
683 _allocation_list.push_back(*block);
684 }
685
686 log_debug(oopstorage, blocks)("%s: processed deferred update " PTR_FORMAT,
687 name(), p2i(block));
688 return true; // Processed one pending update.
689}
690
691inline void check_release_entry(const oop* entry) {
692 assert(entry != NULL, "Releasing NULL");
693 assert(*entry == NULL, "Releasing uncleared entry: " PTR_FORMAT, p2i(entry));
694}
695
696void OopStorage::release(const oop* ptr) {
697 check_release_entry(ptr);
698 Block* block = find_block_or_null(ptr);
699 assert(block != NULL, "%s: invalid release " PTR_FORMAT, name(), p2i(ptr));
700 log_trace(oopstorage, ref)("%s: released " PTR_FORMAT, name(), p2i(ptr));
701 block->release_entries(block->bitmask_for_entry(ptr), this);
702 Atomic::dec(&_allocation_count);
703}
704
705void OopStorage::release(const oop* const* ptrs, size_t size) {
706 size_t i = 0;
707 while (i < size) {
708 check_release_entry(ptrs[i]);
709 Block* block = find_block_or_null(ptrs[i]);
710 assert(block != NULL, "%s: invalid release " PTR_FORMAT, name(), p2i(ptrs[i]));
711 log_trace(oopstorage, ref)("%s: released " PTR_FORMAT, name(), p2i(ptrs[i]));
712 size_t count = 0;
713 uintx releasing = 0;
714 for ( ; i < size; ++i) {
715 const oop* entry = ptrs[i];
716 check_release_entry(entry);
717 // If entry not in block, finish block and resume outer loop with entry.
718 if (!block->contains(entry)) break;
719 // Add entry to releasing bitmap.
720 log_trace(oopstorage, ref)("%s: released " PTR_FORMAT, name(), p2i(entry));
721 uintx entry_bitmask = block->bitmask_for_entry(entry);
722 assert((releasing & entry_bitmask) == 0,
723 "Duplicate entry: " PTR_FORMAT, p2i(entry));
724 releasing |= entry_bitmask;
725 ++count;
726 }
727 // Release the contiguous entries that are in block.
728 block->release_entries(releasing, this);
729 Atomic::sub(count, &_allocation_count);
730 }
731}
732
733const char* dup_name(const char* name) {
734 char* dup = NEW_C_HEAP_ARRAY(char, strlen(name) + 1, mtGC);
735 strcpy(dup, name);
736 return dup;
737}
738
739const size_t initial_active_array_size = 8;
740
741OopStorage::OopStorage(const char* name,
742 Mutex* allocation_mutex,
743 Mutex* active_mutex) :
744 _name(dup_name(name)),
745 _active_array(ActiveArray::create(initial_active_array_size)),
746 _allocation_list(),
747 _deferred_updates(NULL),
748 _allocation_mutex(allocation_mutex),
749 _active_mutex(active_mutex),
750 _allocation_count(0),
751 _concurrent_iteration_count(0),
752 _needs_cleanup(false)
753{
754 _active_array->increment_refcount();
755 assert(_active_mutex->rank() < _allocation_mutex->rank(),
756 "%s: active_mutex must have lower rank than allocation_mutex", _name);
757 assert(Service_lock->rank() < _active_mutex->rank(),
758 "%s: active_mutex must have higher rank than Service_lock", _name);
759 assert(_active_mutex->_safepoint_check_required == Mutex::_safepoint_check_never,
760 "%s: active mutex requires never safepoint check", _name);
761 assert(_allocation_mutex->_safepoint_check_required == Mutex::_safepoint_check_never,
762 "%s: allocation mutex requires never safepoint check", _name);
763}
764
765void OopStorage::delete_empty_block(const Block& block) {
766 assert(block.is_empty(), "discarding non-empty block");
767 log_debug(oopstorage, blocks)("%s: delete empty block " PTR_FORMAT, name(), p2i(&block));
768 Block::delete_block(block);
769}
770
771OopStorage::~OopStorage() {
772 Block* block;
773 while ((block = _deferred_updates) != NULL) {
774 _deferred_updates = block->deferred_updates_next();
775 block->set_deferred_updates_next(NULL);
776 }
777 while ((block = _allocation_list.head()) != NULL) {
778 _allocation_list.unlink(*block);
779 }
780 bool unreferenced = _active_array->decrement_refcount();
781 assert(unreferenced, "deleting storage while _active_array is referenced");
782 for (size_t i = _active_array->block_count(); 0 < i; ) {
783 block = _active_array->at(--i);
784 Block::delete_block(*block);
785 }
786 ActiveArray::destroy(_active_array);
787 FREE_C_HEAP_ARRAY(char, _name);
788}
789
790// Managing service thread notifications.
791//
792// We don't want cleanup work to linger indefinitely, but we also don't want
793// to run the service thread too often. We're also very limited in what we
794// can do in a release operation, where cleanup work is created.
795//
796// When a release operation changes a block's state to empty, it records the
797// need for cleanup in both the associated storage object and in the global
798// request state. A safepoint cleanup task notifies the service thread when
799// there may be cleanup work for any storage object, based on the global
800// request state. But that notification is deferred if the service thread
801// has run recently, and we also avoid duplicate notifications. The service
802// thread updates the timestamp and resets the state flags on every iteration.
803
804// Global cleanup request state.
805static volatile bool needs_cleanup_requested = false;
806
807// Flag for avoiding duplicate notifications.
808static bool needs_cleanup_triggered = false;
809
810// Time after which a notification can be made.
811static jlong cleanup_trigger_permit_time = 0;
812
813// Minimum time since last service thread check before notification is
814// permitted. The value of 500ms was an arbitrary choice; frequent, but not
815// too frequent.
816const jlong cleanup_trigger_defer_period = 500 * NANOSECS_PER_MILLISEC;
817
818void OopStorage::trigger_cleanup_if_needed() {
819 MonitorLocker ml(Service_lock, Monitor::_no_safepoint_check_flag);
820 if (Atomic::load(&needs_cleanup_requested) &&
821 !needs_cleanup_triggered &&
822 (os::javaTimeNanos() > cleanup_trigger_permit_time)) {
823 needs_cleanup_triggered = true;
824 ml.notify_all();
825 }
826}
827
828bool OopStorage::has_cleanup_work_and_reset() {
829 assert_lock_strong(Service_lock);
830 cleanup_trigger_permit_time =
831 os::javaTimeNanos() + cleanup_trigger_defer_period;
832 needs_cleanup_triggered = false;
833 // Set the request flag false and return its old value.
834 // Needs to be atomic to avoid dropping a concurrent request.
835 // Can't use Atomic::xchg, which may not support bool.
836 return Atomic::cmpxchg(false, &needs_cleanup_requested, true);
837}
838
839// Record that cleanup is needed, without notifying the Service thread.
840// Used by release(), where we can't lock even Service_lock.
841void OopStorage::record_needs_cleanup() {
842 // Set local flag first, else service thread could wake up and miss
843 // the request. This order may instead (rarely) unnecessarily notify.
844 OrderAccess::release_store(&_needs_cleanup, true);
845 OrderAccess::release_store_fence(&needs_cleanup_requested, true);
846}
847
848bool OopStorage::delete_empty_blocks() {
849 // Service thread might have oopstorage work, but not for this object.
850 // Check for deferred updates even though that's not a service thread
851 // trigger; since we're here, we might as well process them.
852 if (!OrderAccess::load_acquire(&_needs_cleanup) &&
853 (OrderAccess::load_acquire(&_deferred_updates) == NULL)) {
854 return false;
855 }
856
857 MutexLocker ml(_allocation_mutex, Mutex::_no_safepoint_check_flag);
858
859 // Clear the request before processing.
860 OrderAccess::release_store_fence(&_needs_cleanup, false);
861
862 // Other threads could be adding to the empty block count or the
863 // deferred update list while we're working. Set an upper bound on
864 // how many updates we'll process and blocks we'll try to release,
865 // so other threads can't cause an unbounded stay in this function.
866 // We add a bit of slop because the reduce_deferred_updates clause
867 // can cause blocks to be double counted. If there are few blocks
868 // and many of them are deferred and empty, we might hit the limit
869 // and spin the caller without doing very much work. Otherwise,
870 // we don't normally hit the limit anyway, instead running out of
871 // work to do.
872 size_t limit = block_count() + 10;
873
874 for (size_t i = 0; i < limit; ++i) {
875 // Process deferred updates, which might make empty blocks available.
876 // Continue checking once deletion starts, since additional updates
877 // might become available while we're working.
878 if (reduce_deferred_updates()) {
879 // Be safepoint-polite while looping.
880 MutexUnlocker ul(_allocation_mutex, Mutex::_no_safepoint_check_flag);
881 ThreadBlockInVM tbiv(JavaThread::current());
882 } else {
883 Block* block = _allocation_list.tail();
884 if ((block == NULL) || !block->is_empty()) {
885 return false;
886 } else if (!block->is_safe_to_delete()) {
887 // Look for other work while waiting for block to be deletable.
888 break;
889 }
890
891 // Try to delete the block. First, try to remove from _active_array.
892 {
893 MutexLocker aml(_active_mutex, Mutex::_no_safepoint_check_flag);
894 // Don't interfere with an active concurrent iteration.
895 // Instead, give up immediately. There is more work to do,
896 // but don't re-notify, to avoid useless spinning of the
897 // service thread. Instead, iteration completion notifies.
898 if (_concurrent_iteration_count > 0) return true;
899 _active_array->remove(block);
900 }
901 // Remove block from _allocation_list and delete it.
902 _allocation_list.unlink(*block);
903 // Be safepoint-polite while deleting and looping.
904 MutexUnlocker ul(_allocation_mutex, Mutex::_no_safepoint_check_flag);
905 delete_empty_block(*block);
906 ThreadBlockInVM tbiv(JavaThread::current());
907 }
908 }
909 // Exceeded work limit or can't delete last block. This will
910 // cause the service thread to loop, giving other subtasks an
911 // opportunity to run too. There's no need for a notification,
912 // because we are part of the service thread (unless gtesting).
913 record_needs_cleanup();
914 return true;
915}
916
917OopStorage::EntryStatus OopStorage::allocation_status(const oop* ptr) const {
918 const Block* block = find_block_or_null(ptr);
919 if (block != NULL) {
920 // Prevent block deletion and _active_array modification.
921 MutexLocker ml(_allocation_mutex, Mutex::_no_safepoint_check_flag);
922 // Block could be a false positive, so get index carefully.
923 size_t index = Block::active_index_safe(block);
924 if ((index < _active_array->block_count()) &&
925 (block == _active_array->at(index)) &&
926 block->contains(ptr)) {
927 if ((block->allocated_bitmask() & block->bitmask_for_entry(ptr)) != 0) {
928 return ALLOCATED_ENTRY;
929 } else {
930 return UNALLOCATED_ENTRY;
931 }
932 }
933 }
934 return INVALID_ENTRY;
935}
936
937size_t OopStorage::allocation_count() const {
938 return _allocation_count;
939}
940
941size_t OopStorage::block_count() const {
942 WithActiveArray wab(this);
943 // Count access is racy, but don't care.
944 return wab.active_array().block_count();
945}
946
947size_t OopStorage::total_memory_usage() const {
948 size_t total_size = sizeof(OopStorage);
949 total_size += strlen(name()) + 1;
950 total_size += sizeof(ActiveArray);
951 WithActiveArray wab(this);
952 const ActiveArray& blocks = wab.active_array();
953 // Count access is racy, but don't care.
954 total_size += blocks.block_count() * Block::allocation_size();
955 total_size += blocks.size() * sizeof(Block*);
956 return total_size;
957}
958
959// Parallel iteration support
960
961uint OopStorage::BasicParState::default_estimated_thread_count(bool concurrent) {
962 uint configured = concurrent ? ConcGCThreads : ParallelGCThreads;
963 return MAX2(1u, configured); // Never estimate zero threads.
964}
965
966OopStorage::BasicParState::BasicParState(const OopStorage* storage,
967 uint estimated_thread_count,
968 bool concurrent) :
969 _storage(storage),
970 _active_array(_storage->obtain_active_array()),
971 _block_count(0), // initialized properly below
972 _next_block(0),
973 _estimated_thread_count(estimated_thread_count),
974 _concurrent(concurrent)
975{
976 assert(estimated_thread_count > 0, "estimated thread count must be positive");
977 update_concurrent_iteration_count(1);
978 // Get the block count *after* iteration state updated, so concurrent
979 // empty block deletion is suppressed and can't reduce the count. But
980 // ensure the count we use was written after the block with that count
981 // was fully initialized; see ActiveArray::push.
982 _block_count = _active_array->block_count_acquire();
983}
984
985OopStorage::BasicParState::~BasicParState() {
986 _storage->relinquish_block_array(_active_array);
987 update_concurrent_iteration_count(-1);
988 if (_concurrent) {
989 // We may have deferred some cleanup work.
990 const_cast<OopStorage*>(_storage)->record_needs_cleanup();
991 }
992}
993
994void OopStorage::BasicParState::update_concurrent_iteration_count(int value) {
995 if (_concurrent) {
996 MutexLocker ml(_storage->_active_mutex, Mutex::_no_safepoint_check_flag);
997 _storage->_concurrent_iteration_count += value;
998 assert(_storage->_concurrent_iteration_count >= 0, "invariant");
999 }
1000}
1001
1002bool OopStorage::BasicParState::claim_next_segment(IterationData* data) {
1003 data->_processed += data->_segment_end - data->_segment_start;
1004 size_t start = OrderAccess::load_acquire(&_next_block);
1005 if (start >= _block_count) {
1006 return finish_iteration(data); // No more blocks available.
1007 }
1008 // Try to claim several at a time, but not *too* many. We want to
1009 // avoid deciding there are many available and selecting a large
1010 // quantity, get delayed, and then end up claiming most or all of
1011 // the remaining largish amount of work, leaving nothing for other
1012 // threads to do. But too small a step can lead to contention
1013 // over _next_block, esp. when the work per block is small.
1014 size_t max_step = 10;
1015 size_t remaining = _block_count - start;
1016 size_t step = MIN2(max_step, 1 + (remaining / _estimated_thread_count));
1017 // Atomic::add with possible overshoot. This can perform better
1018 // than a CAS loop on some platforms when there is contention.
1019 // We can cope with the uncertainty by recomputing start/end from
1020 // the result of the add, and dealing with potential overshoot.
1021 size_t end = Atomic::add(step, &_next_block);
1022 // _next_block may have changed, so recompute start from result of add.
1023 start = end - step;
1024 // _next_block may have changed so much that end has overshot.
1025 end = MIN2(end, _block_count);
1026 // _next_block may have changed so much that even start has overshot.
1027 if (start < _block_count) {
1028 // Record claimed segment for iteration.
1029 data->_segment_start = start;
1030 data->_segment_end = end;
1031 return true; // Success.
1032 } else {
1033 // No more blocks to claim.
1034 return finish_iteration(data);
1035 }
1036}
1037
1038bool OopStorage::BasicParState::finish_iteration(const IterationData* data) const {
1039 log_info(oopstorage, blocks, stats)
1040 ("Parallel iteration on %s: blocks = " SIZE_FORMAT
1041 ", processed = " SIZE_FORMAT " (%2.f%%)",
1042 _storage->name(), _block_count, data->_processed,
1043 percent_of(data->_processed, _block_count));
1044 return false;
1045}
1046
1047const char* OopStorage::name() const { return _name; }
1048
1049#ifndef PRODUCT
1050
1051void OopStorage::print_on(outputStream* st) const {
1052 size_t allocations = _allocation_count;
1053 size_t blocks = _active_array->block_count();
1054
1055 double data_size = section_size * section_count;
1056 double alloc_percentage = percent_of((double)allocations, blocks * data_size);
1057
1058 st->print("%s: " SIZE_FORMAT " entries in " SIZE_FORMAT " blocks (%.F%%), " SIZE_FORMAT " bytes",
1059 name(), allocations, blocks, alloc_percentage, total_memory_usage());
1060 if (_concurrent_iteration_count > 0) {
1061 st->print(", concurrent iteration active");
1062 }
1063}
1064
1065#endif // !PRODUCT
1066