/*

 * Copyright (c) 2018, 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.

 *

 */

#include "precompiled.hpp"

#include "gc/shared/oopStorage.inline.hpp"

#include "gc/shared/oopStorageParState.inline.hpp"

#include "logging/log.hpp"

#include "logging/logStream.hpp"

#include "memory/allocation.inline.hpp"

#include "runtime/atomic.hpp"

#include "runtime/globals.hpp"

#include "runtime/handles.inline.hpp"

#include "runtime/mutex.hpp"

#include "runtime/mutexLocker.hpp"

#include "runtime/orderAccess.hpp"

#include "runtime/safepoint.hpp"

#include "runtime/stubRoutines.hpp"

#include "runtime/thread.hpp"

#include "utilities/align.hpp"

#include "utilities/count_trailing_zeros.hpp"

#include "utilities/debug.hpp"

#include "utilities/globalDefinitions.hpp"

#include "utilities/macros.hpp"

#include "utilities/ostream.hpp"

#include "utilities/spinYield.hpp"

OopStorage::AllocateEntry::AllocateEntry() : _prev(NULL), _next(NULL) {}

OopStorage::AllocateEntry::~AllocateEntry() {

  assert(_prev == NULL, "deleting attached block");

  assert(_next == NULL, "deleting attached block");

}

OopStorage::AllocateList::~AllocateList() {

  // ~OopStorage() empties its lists before destroying them.

  assert(_head == NULL, "deleting non-empty block list");

  assert(_tail == NULL, "deleting non-empty block list");

}

void OopStorage::AllocateList::push_front(const Block& block) {

  const Block* old = _head;

  if (old == NULL) {

    assert(_tail == NULL, "invariant");

    _head = _tail = █

  } else {

    _head = █

  }

}

void OopStorage::AllocateList::push_back(const Block& block) {

  const Block* old = _tail;

  if (old == NULL) {

    assert(_head == NULL, "invariant");

    _head = _tail = █

  } else {

    _tail = █

  }

}

void OopStorage::AllocateList::unlink(const Block& block) {

  const Block* prev_blk = block_entry._prev;

  const Block* next_blk = block_entry._next;

  block_entry._prev = NULL;

  block_entry._next = NULL;

  if ((prev_blk == NULL) && (next_blk == NULL)) {

    assert(_head == &block, "invariant");

    assert(_tail == &block, "invariant");

    _head = _tail = NULL;

  } else if (prev_blk == NULL) {

    assert(_head == &block, "invariant");

    _head = next_blk;

  } else if (next_blk == NULL) {

    assert(_tail == &block, "invariant");

    _tail = prev_blk;

  } else {

  }

}

OopStorage::ActiveArray::ActiveArray(size_t size) :

  _size(size),

  _block_count(0),

  _refcount(0)

{}

OopStorage::ActiveArray::~ActiveArray() {

  assert(_refcount == 0, "precondition");

}

OopStorage::ActiveArray* OopStorage::ActiveArray::create(size_t size, AllocFailType alloc_fail) {

  size_t size_in_bytes = blocks_offset() + sizeof(Block*) * size;

  void* mem = NEW_C_HEAP_ARRAY3(char, size_in_bytes, mtGC, CURRENT_PC, alloc_fail);

  if (mem == NULL) return NULL;

  return new (mem) ActiveArray(size);

}

void OopStorage::ActiveArray::destroy(ActiveArray* ba) {

  ba->~ActiveArray();

  FREE_C_HEAP_ARRAY(char, ba);

}

size_t OopStorage::ActiveArray::size() const {

  return _size;

}

size_t OopStorage::ActiveArray::block_count() const {

  return _block_count;

}

size_t OopStorage::ActiveArray::block_count_acquire() const {

  return OrderAccess::load_acquire(&_block_count);

}

void OopStorage::ActiveArray::increment_refcount() const {

  int new_value = Atomic::add(1, &_refcount);

  assert(new_value >= 1, "negative refcount %d", new_value - 1);

}

bool OopStorage::ActiveArray::decrement_refcount() const {

  int new_value = Atomic::sub(1, &_refcount);

  assert(new_value >= 0, "negative refcount %d", new_value);

  return new_value == 0;

}

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.

    OrderAccess::release_store(&_block_count, index + 1);

    return true;

  } else {

    return false;

  }

}

void OopStorage::ActiveArray::remove(Block* block) {

  assert(_block_count > 0, "array is empty");

  size_t index = block->active_index();

  assert(*block_ptr(index) == block, "block not present");

  size_t last_index = _block_count - 1;

  Block* last_block = *block_ptr(last_index);

  last_block->set_active_index(index);

  *block_ptr(index) = last_block;

  _block_count = last_index;

}

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().

const unsigned section_size = BitsPerByte;

const unsigned section_count = BytesPerWord;

const unsigned block_alignment = sizeof(oop) * section_size;

OopStorage::Block::Block(const OopStorage* owner, void* memory) :

  _data(),

  _allocated_bitmask(0),

  _owner(owner),

  _memory(memory),

  _active_index(0),

  _allocate_entry(),

  _deferred_updates_next(NULL),

  _release_refcount(0)

{

  STATIC_ASSERT(_data_pos == 0);

  STATIC_ASSERT(section_size * section_count == ARRAY_SIZE(_data));

  assert(offset_of(Block, _data) == _data_pos, "invariant");

  assert(owner != NULL, "NULL owner");

  assert(is_aligned(this, block_alignment), "misaligned block");

}

OopStorage::Block::~Block() {

  assert(_release_refcount == 0, "deleting block while releasing");

  assert(_deferred_updates_next == NULL, "deleting block with deferred update");

  // Clear fields used by block_for_ptr and entry validation, which

  // might help catch bugs.  Volatile to prevent dead-store elimination.

  const_cast<uintx volatile&>(_allocated_bitmask) = 0;

  const_cast<OopStorage* volatile&>(_owner) = NULL;

}

size_t OopStorage::Block::allocation_size() {

  // _data must be first member, so aligning Block aligns _data.

  STATIC_ASSERT(_data_pos == 0);

  return sizeof(Block) + block_alignment - sizeof(void*);

}

size_t OopStorage::Block::allocation_alignment_shift() {

  return exact_log2(block_alignment);

}

inline bool is_full_bitmask(uintx bitmask) { return ~bitmask == 0; }

inline bool is_empty_bitmask(uintx bitmask) { return bitmask == 0; }

bool OopStorage::Block::is_full() const {

  return is_full_bitmask(allocated_bitmask());

}

bool OopStorage::Block::is_empty() const {

  return is_empty_bitmask(allocated_bitmask());

}

uintx OopStorage::Block::bitmask_for_entry(const oop* ptr) const {

  return bitmask_for_index(get_index(ptr));

}

// A block is deletable if

// (1) It is empty.

// (2) There is not a release() operation currently operating on it.

// (3) It is not in the deferred updates list.

// The order of tests is important for proper interaction between release()

// and concurrent deletion.

bool OopStorage::Block::is_deletable() const {

  return (OrderAccess::load_acquire(&_allocated_bitmask) == 0) &&

         (OrderAccess::load_acquire(&_release_refcount) == 0) &&

         (OrderAccess::load_acquire(&_deferred_updates_next) == NULL);

}

OopStorage::Block* OopStorage::Block::deferred_updates_next() const {

  return _deferred_updates_next;

}

void OopStorage::Block::set_deferred_updates_next(Block* block) {

  _deferred_updates_next = block;

}

bool OopStorage::Block::contains(const oop* ptr) const {

  const oop* base = get_pointer(0);

  return (base <= ptr) && (ptr < (base + ARRAY_SIZE(_data)));

}

size_t OopStorage::Block::active_index() const {

  return _active_index;

}

void OopStorage::Block::set_active_index(size_t index) {

  _active_index = index;

}

size_t OopStorage::Block::active_index_safe(const Block* block) {

  STATIC_ASSERT(sizeof(intptr_t) == sizeof(block->_active_index));

  assert(CanUseSafeFetchN(), "precondition");

  return SafeFetchN((intptr_t*)&block->_active_index, 0);

}

unsigned OopStorage::Block::get_index(const oop* ptr) const {

  assert(contains(ptr), PTR_FORMAT " not in block " PTR_FORMAT, p2i(ptr), p2i(this));

  return static_cast<unsigned>(ptr - get_pointer(0));

}

oop* OopStorage::Block::allocate() {

  // Use CAS loop because release may change bitmask outside of lock.

  uintx allocated = allocated_bitmask();

  while (true) {

    assert(!is_full_bitmask(allocated), "attempt to allocate from full block");

    unsigned index = count_trailing_zeros(~allocated);

    uintx new_value = allocated | bitmask_for_index(index);

    uintx fetched = Atomic::cmpxchg(new_value, &_allocated_bitmask, allocated);

    if (fetched == allocated) {

      return get_pointer(index); // CAS succeeded; return entry for index.

    }

    allocated = fetched;       // CAS failed; retry with latest value.

  }

}

OopStorage::Block* OopStorage::Block::new_block(const OopStorage* owner) {

  // _data must be first member: aligning block => aligning _data.

  STATIC_ASSERT(_data_pos == 0);

  size_t size_needed = allocation_size();

  void* memory = NEW_C_HEAP_ARRAY_RETURN_NULL(char, size_needed, mtGC);

  if (memory == NULL) {

    return NULL;

  }

  void* block_mem = align_up(memory, block_alignment);

  assert(sizeof(Block) + pointer_delta(block_mem, memory, 1) <= size_needed,

         "allocated insufficient space for aligned block");

  return ::new (block_mem) Block(owner, memory);

}

void OopStorage::Block::delete_block(const Block& block) {

  void* memory = block._memory;

  block.Block::~Block();

  FREE_C_HEAP_ARRAY(char, memory);

}

// 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) {

  assert(CanUseSafeFetchN(), "precondition");

  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);

    intptr_t* candidate_owner_addr

      = reinterpret_cast<intptr_t*>(&candidate->_owner);

    if (SafeFetchN(candidate_owner_addr, 0) == owner_addr) {

      return candidate;

    }

  }

  return NULL;

}

//////////////////////////////////////////////////////////////////////////////

// Allocation

//

// Allocation involves the _allocate_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 _allocate_list, to make it easy for empty block

// deletion to find them.

//

// allocate(), and delete_empty_blocks_concurrent() lock the

// _allocate_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 _allocate_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 _allocate_list so it won't be considered by future

// allocations until some entries in it are released.

//

// release() is performed lock-free. 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 _allocate_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 _allocate_list also needs to be modified.  This requires

// locking the _allocate_mutex.  To keep the release() operation lock-free,

// rather than updating the _allocate_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_XXX(), 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 _allocate_list if not already present and the bitmask is not

// full.  The block is moved to the end of the _allocated_list if the bitmask

// is empty, for ease of empty block deletion processing.

oop* OopStorage::allocate() {

  MutexLockerEx ml(_allocate_mutex, Mutex::_no_safepoint_check_flag);

  // Do some deferred update processing every time we allocate.

  // Continue processing deferred updates if _allocate_list is empty,

  // in the hope that we'll get a block from that, rather than

  // allocating a new block.

  while (reduce_deferred_updates() && (_allocate_list.head() == NULL)) {}

  // Use the first block in _allocate_list for the allocation.

  Block* block = _allocate_list.head();

  if (block == NULL) {

    // No available blocks; make a new one, and add to storage.

    {

      MutexUnlockerEx mul(_allocate_mutex, Mutex::_no_safepoint_check_flag);

      block = Block::new_block(this);

    }

    if (block == NULL) {

      while (_allocate_list.head() == NULL) {

        if (!reduce_deferred_updates()) {

          // Failed to make new block, no other thread made a block

          // available while the mutex was released, and didn't get

          // one from a deferred update either, so return failure.

          log_info(oopstorage, ref)("%s: failed block allocation", name());

          return NULL;

        }

      }

    } else {

      // Add new block to storage.

      log_info(oopstorage, blocks)("%s: new block " PTR_FORMAT, name(), p2i(block));

      // 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_info(oopstorage, blocks)("%s: failed active array expand", name());

          Block::delete_block(*block);

          return NULL;

        }

      }

      // Add to end of _allocate_list.  The mutex release allowed

      // other threads to add blocks to the _allocate_list.  We prefer

      // to allocate from non-empty blocks, to allow empty blocks to

      // be deleted.

      _allocate_list.push_back(*block);

    }

    block = _allocate_list.head();

  }

  // Allocate from first block.

  assert(block != NULL, "invariant");

  assert(!block->is_full(), "invariant");

  if (block->is_empty()) {

    // Transitioning from empty to not empty.

    log_debug(oopstorage, blocks)("%s: block not empty " PTR_FORMAT, name(), p2i(block));

  }

  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_debug(oopstorage, blocks)("%s: block full " PTR_FORMAT, name(), p2i(block));

    _allocate_list.unlink(*block);

  }

  log_info(oopstorage, ref)("%s: allocated " PTR_FORMAT, name(), p2i(result));

  return result;

}

// 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(_allocate_mutex);

  ActiveArray* old_array = _active_array;

  size_t new_size = 2 * old_array->size();

  log_info(oopstorage, blocks)("%s: expand active array " SIZE_FORMAT,

                               name(), new_size);

  ActiveArray* new_array = ActiveArray::create(new_size, AllocFailStrategy::RETURN_NULL);

  if (new_array == NULL) return false;

  new_array->copy_from(old_array);

  replace_active_array(new_array);

  relinquish_block_array(old_array);

  return true;

}

OopStorage::ProtectActive::ProtectActive() : _enter(0), _exit() {}

// Begin read-side critical section.

uint OopStorage::ProtectActive::read_enter() {

  return Atomic::add(2u, &_enter);

}

// End read-side critical section.

void OopStorage::ProtectActive::read_exit(uint enter_value) {