/* * Copyright (c) 2018, 2022, Oracle and/or its affiliates. All rights reserved. * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * * This code is free software; you can redistribute it and/or modify it * under the terms of the GNU General Public License version 2 only, as * published by the Free Software Foundation. * * This code is distributed in the hope that it will be useful, but WITHOUT * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License * version 2 for more details (a copy is included in the LICENSE file that * accompanied this code). * * You should have received a copy of the GNU General Public License version * 2 along with this work; if not, write to the Free Software Foundation, * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. * * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA * or visit www.oracle.com if you need additional information or have any * questions. *
*/
bool OopStorage::ActiveArray::push(Block* block) {
size_t index = _block_count; if (index < _size) {
block->set_active_index(index);
*block_ptr(index) = block; // Use a release_store to ensure all the setup is complete before // making the block visible.
Atomic::release_store(&_block_count, index + 1); returntrue;
} else { returnfalse;
}
}
void OopStorage::ActiveArray::copy_from(const ActiveArray* from) {
assert(_block_count == 0, "array must be empty");
size_t count = from->_block_count;
assert(count <= _size, "precondition");
Block* const* from_ptr = from->block_ptr(0);
Block** to_ptr = block_ptr(0); for (size_t i = 0; i < count; ++i) {
Block* block = *from_ptr++;
assert(block->active_index() == i, "invariant");
*to_ptr++ = block;
}
_block_count = count;
}
// Blocks start with an array of BitsPerWord oop entries. That array // is divided into conceptual BytesPerWord sections of BitsPerByte // entries. Blocks are allocated aligned on section boundaries, for // the convenience of mapping from an entry to the containing block; // see block_for_ptr(). Aligning on section boundary rather than on // the full _data wastes a lot less space, but makes for a bit more // work in block_for_ptr().
// An empty block is not yet deletable if either: // (1) There is a release() operation currently operating on it. // (2) It is in the deferred updates list. // For interaction with release(), these must follow the empty check, // and the order of these checks is important. bool OopStorage::Block::is_safe_to_delete() const {
assert(is_empty(), "precondition");
OrderAccess::loadload(); return (Atomic::load_acquire(&_release_refcount) == 0) &&
(Atomic::load_acquire(&_deferred_updates_next) == NULL);
}
unsigned OopStorage::Block::get_index(const oop* ptr) const {
assert(contains(ptr), PTR_FORMAT " not in block " PTR_FORMAT, p2i(ptr), p2i(this)); returnstatic_cast<unsigned>(ptr - get_pointer(0));
}
// Merge new allocation bits into _allocated_bitmask. Only one thread at a // time is ever allocating from a block, but other threads may concurrently // release entries and clear bits in _allocated_bitmask. // precondition: _allocated_bitmask & add == 0 void OopStorage::Block::atomic_add_allocated(uintx add) { // Since the current allocated bitmask has no set bits in common with add, // we can use an atomic add to implement the operation. The assert post // facto verifies the precondition held; if there were any set bits in // common, then after the add at least one of them will be zero.
uintx sum = Atomic::add(&_allocated_bitmask, add);
assert((sum & add) == add, "some already present: " UINTX_FORMAT ":" UINTX_FORMAT,
sum, add);
}
oop* OopStorage::Block::allocate() {
uintx allocated = allocated_bitmask();
assert(!is_full_bitmask(allocated), "attempt to allocate from full block"); unsigned index = count_trailing_zeros(~allocated); // Use atomic update because release may change bitmask.
atomic_add_allocated(bitmask_for_index(index)); return get_pointer(index);
}
uintx OopStorage::Block::allocate_all() {
uintx new_allocated = ~allocated_bitmask();
assert(new_allocated != 0, "attempt to allocate from full block"); // Use atomic update because release may change bitmask.
atomic_add_allocated(new_allocated); return new_allocated;
}
// This can return a false positive if ptr is not contained by some // block. For some uses, it is a precondition that ptr is valid, // e.g. contained in some block in owner's _active_array. Other uses // require additional validation of the result.
OopStorage::Block*
OopStorage::Block::block_for_ptr(const OopStorage* owner, const oop* ptr) {
STATIC_ASSERT(_data_pos == 0); // Const-ness of ptr is not related to const-ness of containing block. // Blocks are allocated section-aligned, so get the containing section.
oop* section_start = align_down(const_cast<oop*>(ptr), block_alignment); // Start with a guess that the containing section is the last section, // so the block starts section_count-1 sections earlier.
oop* section = section_start - (section_size * (section_count - 1)); // Walk up through the potential block start positions, looking for // the owner in the expected location. If we're below the actual block // start position, the value at the owner position will be some oop // (possibly NULL), which can never match the owner.
intptr_t owner_addr = reinterpret_cast<intptr_t>(owner); for (unsigned i = 0; i < section_count; ++i, section += section_size) {
Block* candidate = reinterpret_cast<Block*>(section); if (SafeFetchN(&candidate->_owner_address, 0) == owner_addr) { return candidate;
}
} return NULL;
}
////////////////////////////////////////////////////////////////////////////// // Allocation // // Allocation involves the _allocation_list, which contains a subset of the // blocks owned by a storage object. This is a doubly-linked list, linked // through dedicated fields in the blocks. Full blocks are removed from this // list, though they are still present in the _active_array. Empty blocks are // kept at the end of the _allocation_list, to make it easy for empty block // deletion to find them. // // allocate(), and delete_empty_blocks() lock the // _allocation_mutex while performing any list and array modifications. // // allocate() and release() update a block's _allocated_bitmask using CAS // loops. This prevents loss of updates even though release() performs // its updates without any locking. // // allocate() obtains the entry from the first block in the _allocation_list, // and updates that block's _allocated_bitmask to indicate the entry is in // use. If this makes the block full (all entries in use), the block is // removed from the _allocation_list so it won't be considered by future // allocations until some entries in it are released. // // release() is performed lock-free. (Note: This means it can't notify the // service thread of pending cleanup work. It must be lock-free because // it is called in all kinds of contexts where even quite low ranked locks // may be held.) release() first looks up the block for // the entry, using address alignment to find the enclosing block (thereby // avoiding iteration over the _active_array). Once the block has been // determined, its _allocated_bitmask needs to be updated, and its position in // the _allocation_list may need to be updated. There are two cases: // // (a) If the block is neither full nor would become empty with the release of // the entry, only its _allocated_bitmask needs to be updated. But if the CAS // update fails, the applicable case may change for the retry. // // (b) Otherwise, the _allocation_list also needs to be modified. This requires // locking the _allocation_mutex. To keep the release() operation lock-free, // rather than updating the _allocation_list itself, it instead performs a // lock-free push of the block onto the _deferred_updates list. Entries on // that list are processed by allocate() and delete_empty_blocks(), while // they already hold the necessary lock. That processing makes the block's // list state consistent with its current _allocated_bitmask. The block is // added to the _allocation_list if not already present and the bitmask is not // full. The block is moved to the end of the _allocation_list if the bitmask // is empty, for ease of empty block deletion processing.
Block* block = block_for_allocation(); if (block == NULL) return NULL; // Block allocation failed.
assert(!block->is_full(), "invariant"); if (block->is_empty()) { // Transitioning from empty to not empty.
log_block_transition(block, "not empty");
}
oop* result = block->allocate();
assert(result != NULL, "allocation failed");
assert(!block->is_empty(), "postcondition");
Atomic::inc(&_allocation_count); // release updates outside lock. if (block->is_full()) { // Transitioning from not full to full. // Remove full blocks from consideration by future allocates.
log_block_transition(block, "full");
_allocation_list.unlink(*block);
}
log_trace(oopstorage, ref)("%s: allocated " PTR_FORMAT, name(), p2i(result)); return result;
}
// Bulk allocation takes the first block off the _allocation_list, and marks // all remaining entries in that block as allocated. It then drops the lock // and fills buffer with those newly allocated entries. If more entries // were obtained than requested, the remaining entries are released back // (which is a lock-free operation). Finally, the number actually added to // the buffer is returned. It's best to request at least as many entries as // a single block can provide, to avoid the release case. That number is // available as bulk_allocate_limit.
size_t OopStorage::allocate(oop** ptrs, size_t size) {
assert(size > 0, "precondition");
Block* block;
uintx taken;
{
MutexLocker ml(_allocation_mutex, Mutex::_no_safepoint_check_flag);
block = block_for_allocation(); if (block == NULL) return 0; // Block allocation failed. // Taking all remaining entries, so remove from list.
_allocation_list.unlink(*block); // Transitioning from empty to not empty. if (block->is_empty()) {
log_block_transition(block, "not empty");
}
taken = block->allocate_all(); // Safe to drop the lock, since we have claimed our entries.
assert(!is_empty_bitmask(taken), "invariant");
} // Drop lock, now that we've taken all available entries from block.
size_t num_taken = population_count(taken);
Atomic::add(&_allocation_count, num_taken); // Fill ptrs from those taken entries.
size_t limit = MIN2(num_taken, size); for (size_t i = 0; i < limit; ++i) {
assert(taken != 0, "invariant"); unsigned index = count_trailing_zeros(taken);
taken ^= block->bitmask_for_index(index);
ptrs[i] = block->get_pointer(index);
} // If more entries taken than requested, release remainder. if (taken == 0) {
assert(num_taken == limit, "invariant");
} else {
assert(size == limit, "invariant");
assert(num_taken == (limit + population_count(taken)), "invariant");
block->release_entries(taken, this);
Atomic::sub(&_allocation_count, num_taken - limit);
}
log_trace(oopstorage, ref)("%s: bulk allocate %zu, returned %zu",
name(), limit, num_taken - limit); return limit; // Return number allocated.
}
// Add new block to the _active_array, growing if needed. if (!_active_array->push(block)) { if (expand_active_array()) {
guarantee(_active_array->push(block), "push failed after expansion");
} else {
log_debug(oopstorage, blocks)("%s: failed active array expand", name());
Block::delete_block(*block); returnfalse;
}
} // Add to end of _allocation_list. The mutex release allowed other // threads to add blocks to the _allocation_list. We prefer to // allocate from non-empty blocks, to allow empty blocks to be // deleted. But we don't bother notifying about the empty block // because we're (probably) about to allocate an entry from it.
_allocation_list.push_back(*block);
log_debug(oopstorage, blocks)("%s: new block " PTR_FORMAT, name(), p2i(block)); returntrue;
}
OopStorage::Block* OopStorage::block_for_allocation() {
assert_lock_strong(_allocation_mutex); while (true) { // Use the first block in _allocation_list for the allocation.
Block* block = _allocation_list.head(); if (block != NULL) { return block;
} elseif (reduce_deferred_updates()) { // Might have added a block to the _allocation_list, so retry.
} elseif (try_add_block()) { // Successfully added a new block to the list, so retry.
assert(_allocation_list.chead() != NULL, "invariant");
} elseif (_allocation_list.chead() != NULL) { // Trying to add a block failed, but some other thread added to the // list while we'd dropped the lock over the new block allocation.
} elseif (!reduce_deferred_updates()) { // Once more before failure. // Attempt to add a block failed, no other thread added a block, // and no deferred updated added a block, then allocation failed.
log_info(oopstorage, blocks)("%s: failed block allocation", name()); return NULL;
}
}
}
// Create a new, larger, active array with the same content as the // current array, and then replace, relinquishing the old array. // Return true if the array was successfully expanded, false to // indicate allocation failure. bool OopStorage::expand_active_array() {
assert_lock_strong(_allocation_mutex);
ActiveArray* old_array = _active_array;
size_t new_size = 2 * old_array->size();
log_debug(oopstorage, blocks)("%s: expand active array " SIZE_FORMAT,
name(), new_size);
ActiveArray* new_array = ActiveArray::create(new_size,
memflags(),
AllocFailStrategy::RETURN_NULL); if (new_array == NULL) returnfalse;
new_array->copy_from(old_array);
replace_active_array(new_array);
relinquish_block_array(old_array); returntrue;
}
// Make new_array the _active_array. Increments new_array's refcount // to account for the new reference. The assignment is atomic wrto // obtain_active_array; once this function returns, it is safe for the // caller to relinquish the old array. void OopStorage::replace_active_array(ActiveArray* new_array) { // Caller has the old array that is the current value of _active_array. // Update new_array refcount to account for the new reference.
new_array->increment_refcount(); // Install new_array, ensuring its initialization is complete first.
Atomic::release_store(&_active_array, new_array); // Wait for any readers that could read the old array from _active_array. // Can't use GlobalCounter here, because this is called from allocate(), // which may be called in the scope of a GlobalCounter critical section // when inserting a StringTable entry.
_protect_active.synchronize(); // All obtain critical sections that could see the old array have // completed, having incremented the refcount of the old array. The // caller can now safely relinquish the old array.
}
// Atomically (wrto replace_active_array) get the active array and // increment its refcount. This provides safe access to the array, // even if an allocate operation expands and replaces the value of // _active_array. The caller must relinquish the array when done // using it.
OopStorage::ActiveArray* OopStorage::obtain_active_array() const {
SingleWriterSynchronizer::CriticalSection cs(&_protect_active);
ActiveArray* result = Atomic::load_acquire(&_active_array);
result->increment_refcount(); return result;
}
// Decrement refcount of array and destroy if refcount is zero. void OopStorage::relinquish_block_array(ActiveArray* array) const { if (array->decrement_refcount()) {
assert(array != _active_array, "invariant");
ActiveArray::destroy(array);
}
}
class OopStorage::WithActiveArray : public StackObj { const OopStorage* _storage;
ActiveArray* _active_array;
// Now that the bitmask has been updated, if we have a state transition // (updated bitmask is empty or old bitmask was full), atomically push // this block onto the deferred updates list. Some future call to // reduce_deferred_updates will make any needed changes related to this // block and _allocation_list. This deferral avoids _allocation_list // updates and the associated locking here. if ((releasing == old_allocated) || is_full_bitmask(old_allocated)) { // Log transitions. Both transitions are possible in a single update.
log_release_transitions(releasing, old_allocated, owner, this); // Attempt to claim responsibility for adding this block to the deferred // list, by setting the link to non-NULL by self-looping. If this fails, // then someone else has made such a claim and the deferred update has not // yet been processed and will include our change, so we don't need to do // anything further. if (Atomic::replace_if_null(&_deferred_updates_next, this)) { // Successfully claimed. Push, with self-loop for end-of-list.
Block* head = owner->_deferred_updates; while (true) {
_deferred_updates_next = (head == NULL) ? this : head;
Block* fetched = Atomic::cmpxchg(&owner->_deferred_updates, head, this); if (fetched == head) break; // Successful update.
head = fetched; // Retry with updated head.
} // Only request cleanup for to-empty transitions, not for from-full. // There isn't any rush to process from-full transitions. Allocation // will reduce deferrals before allocating new blocks, so may process // some. And the service thread will drain the entire deferred list // if there are any pending to-empty transitions. if (releasing == old_allocated) {
owner->record_needs_cleanup();
}
log_trace(oopstorage, blocks)("%s: deferred update " PTR_FORMAT,
owner->name(), p2i(this));
}
} // Release hold on empty block deletion.
Atomic::dec(&_release_refcount);
}
// Process one available deferred update. Returns true if one was processed. bool OopStorage::reduce_deferred_updates() {
assert_lock_strong(_allocation_mutex); // Atomically pop a block off the list, if any available. // No ABA issue because this is only called by one thread at a time. // The atomicity is wrto pushes by release().
Block* block = Atomic::load_acquire(&_deferred_updates); while (true) { if (block == NULL) returnfalse; // Try atomic pop of block from list.
Block* tail = block->deferred_updates_next(); if (block == tail) tail = NULL; // Handle self-loop end marker.
Block* fetched = Atomic::cmpxchg(&_deferred_updates, block, tail); if (fetched == block) break; // Update successful.
block = fetched; // Retry with updated block.
}
block->set_deferred_updates_next(NULL); // Clear tail after updating head. // Ensure bitmask read after pop is complete, including clearing tail, for // ordering with release(). Without this, we may be processing a stale // bitmask state here while blocking a release() operation from recording // the deferred update needed for its bitmask change.
OrderAccess::fence(); // Make list state consistent with bitmask state.
uintx allocated = block->allocated_bitmask(); if (is_full_bitmask(allocated)) { // If full then it shouldn't be in the list, and should stay that way.
assert(!_allocation_list.contains(*block), "invariant");
} elseif (_allocation_list.contains(*block)) { // Block is in list. If empty, move to the end for possible deletion. if (is_empty_bitmask(allocated)) {
_allocation_list.unlink(*block);
_allocation_list.push_back(*block);
}
} elseif (is_empty_bitmask(allocated)) { // Block is empty and not in list. Add to back for possible deletion.
_allocation_list.push_back(*block);
} else { // Block is neither full nor empty, and not in list. Add to front.
_allocation_list.push_front(*block);
}
// Managing service thread notifications. // // We don't want cleanup work to linger indefinitely, but we also don't want // to run the service thread too often. We're also very limited in what we // can do in a release operation, where cleanup work is created. // // When a release operation changes a block's state to empty, it records the // need for cleanup in both the associated storage object and in the global // request state. A safepoint cleanup task notifies the service thread when // there may be cleanup work for any storage object, based on the global // request state. But that notification is deferred if the service thread // has run recently, and we also avoid duplicate notifications. The service // thread updates the timestamp and resets the state flags on every iteration.
// Global cleanup request state. staticvolatilebool needs_cleanup_requested = false;
// Flag for avoiding duplicate notifications. staticbool needs_cleanup_triggered = false;
// Time after which a notification can be made. static jlong cleanup_trigger_permit_time = 0;
// Minimum time since last service thread check before notification is // permitted. The value of 500ms was an arbitrary choice; frequent, but not // too frequent. const jlong cleanup_trigger_defer_period = 500 * NANOSECS_PER_MILLISEC;
bool OopStorage::has_cleanup_work_and_reset() {
assert_lock_strong(Service_lock);
cleanup_trigger_permit_time =
os::javaTimeNanos() + cleanup_trigger_defer_period;
needs_cleanup_triggered = false; // Set the request flag false and return its old value. // Needs to be atomic to avoid dropping a concurrent request. // Can't use Atomic::xchg, which may not support bool. return Atomic::cmpxchg(&needs_cleanup_requested, true, false);
}
// Record that cleanup is needed, without notifying the Service thread. // Used by release(), where we can't lock even Service_lock. void OopStorage::record_needs_cleanup() { // Set local flag first, else service thread could wake up and miss // the request. This order may instead (rarely) unnecessarily notify.
Atomic::release_store(&_needs_cleanup, true);
Atomic::release_store_fence(&needs_cleanup_requested, true);
}
bool OopStorage::delete_empty_blocks() { // Service thread might have oopstorage work, but not for this object. // Check for deferred updates even though that's not a service thread // trigger; since we're here, we might as well process them. if (!Atomic::load_acquire(&_needs_cleanup) &&
(Atomic::load_acquire(&_deferred_updates) == NULL)) { returnfalse;
}
// Clear the request before processing.
Atomic::release_store_fence(&_needs_cleanup, false);
// Other threads could be adding to the empty block count or the // deferred update list while we're working. Set an upper bound on // how many updates we'll process and blocks we'll try to release, // so other threads can't cause an unbounded stay in this function. // We add a bit of slop because the reduce_deferred_updates clause // can cause blocks to be double counted. If there are few blocks // and many of them are deferred and empty, we might hit the limit // and spin the caller without doing very much work. Otherwise, // we don't normally hit the limit anyway, instead running out of // work to do.
size_t limit = block_count() + 10;
for (size_t i = 0; i < limit; ++i) { // Process deferred updates, which might make empty blocks available. // Continue checking once deletion starts, since additional updates // might become available while we're working. if (reduce_deferred_updates()) { // Be safepoint-polite while looping.
MutexUnlocker ul(_allocation_mutex, Mutex::_no_safepoint_check_flag);
ThreadBlockInVM tbiv(JavaThread::current());
} else {
Block* block = _allocation_list.tail(); if ((block == NULL) || !block->is_empty()) { returnfalse;
} elseif (!block->is_safe_to_delete()) { // Look for other work while waiting for block to be deletable. break;
}
// Try to delete the block. First, try to remove from _active_array.
{
MutexLocker aml(_active_mutex, Mutex::_no_safepoint_check_flag); // Don't interfere with an active concurrent iteration. // Instead, give up immediately. There is more work to do, // but don't re-notify, to avoid useless spinning of the // service thread. Instead, iteration completion notifies. if (_concurrent_iteration_count > 0) returntrue;
_active_array->remove(block);
} // Remove block from _allocation_list and delete it.
_allocation_list.unlink(*block); // Be safepoint-polite while deleting and looping.
MutexUnlocker ul(_allocation_mutex, Mutex::_no_safepoint_check_flag);
delete_empty_block(*block);
ThreadBlockInVM tbiv(JavaThread::current());
}
} // Exceeded work limit or can't delete last block. This will // cause the service thread to loop, giving other subtasks an // opportunity to run too. There's no need for a notification, // because we are part of the service thread (unless gtesting).
record_needs_cleanup(); returntrue;
}
OopStorage::EntryStatus OopStorage::allocation_status(const oop* ptr) const { const Block* block = find_block_or_null(ptr); if (block != NULL) { // Prevent block deletion and _active_array modification.
MutexLocker ml(_allocation_mutex, Mutex::_no_safepoint_check_flag); // Block could be a false positive, so get index carefully.
size_t index = Block::active_index_safe(block); if ((index < _active_array->block_count()) &&
(block == _active_array->at(index)) &&
block->contains(ptr)) { if ((block->allocated_bitmask() & block->bitmask_for_entry(ptr)) != 0) { return ALLOCATED_ENTRY;
} else { return UNALLOCATED_ENTRY;
}
}
} return INVALID_ENTRY;
}
OopStorage::BasicParState::BasicParState(const OopStorage* storage,
uint estimated_thread_count, bool concurrent) :
_storage(storage),
_active_array(_storage->obtain_active_array()),
_block_count(0), // initialized properly below
_next_block(0),
_estimated_thread_count(estimated_thread_count),
_concurrent(concurrent),
_num_dead(0)
{
assert(estimated_thread_count > 0, "estimated thread count must be positive");
update_concurrent_iteration_count(1); // Get the block count *after* iteration state updated, so concurrent // empty block deletion is suppressed and can't reduce the count. But // ensure the count we use was written after the block with that count // was fully initialized; see ActiveArray::push.
_block_count = _active_array->block_count_acquire();
}
OopStorage::BasicParState::~BasicParState() {
_storage->relinquish_block_array(_active_array);
update_concurrent_iteration_count(-1); if (_concurrent) { // We may have deferred some cleanup work. const_cast<OopStorage*>(_storage)->record_needs_cleanup();
}
}
bool OopStorage::BasicParState::claim_next_segment(IterationData* data) {
data->_processed += data->_segment_end - data->_segment_start;
size_t start = Atomic::load_acquire(&_next_block); if (start >= _block_count) { return finish_iteration(data); // No more blocks available.
} // Try to claim several at a time, but not *too* many. We want to // avoid deciding there are many available and selecting a large // quantity, get delayed, and then end up claiming most or all of // the remaining largish amount of work, leaving nothing for other // threads to do. But too small a step can lead to contention // over _next_block, esp. when the work per block is small.
size_t max_step = 10;
size_t remaining = _block_count - start;
size_t step = MIN2(max_step, 1 + (remaining / _estimated_thread_count)); // Atomic::add with possible overshoot. This can perform better // than a CAS loop on some platforms when there is contention. // We can cope with the uncertainty by recomputing start/end from // the result of the add, and dealing with potential overshoot.
size_t end = Atomic::add(&_next_block, step); // _next_block may have changed, so recompute start from result of add.
start = end - step; // _next_block may have changed so much that end has overshot.
end = MIN2(end, _block_count); // _next_block may have changed so much that even start has overshot. if (start < _block_count) { // Record claimed segment for iteration.
data->_segment_start = start;
data->_segment_end = end; returntrue; // Success.
} else { // No more blocks to claim. return finish_iteration(data);
}
}
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung ist noch experimentell.
