本文基於openjdk11及hotspot
java
在正式探討JVM對象的建立前,先簡單地介紹一下hotspot中實現的Java的對象模型。在JVM中,並無直接將Java對象映射成C++對象,而是採用了oop-klass模型,主要是不但願每一個對象中都包含有一份虛函數表,其中:數組
簡單地說,一個Java類在JVM中被拆分爲了兩個部分:數據和描述信息,分別對應OOP和Klass。bash
在具體的JVM源碼中,當加載一個Class時,會建立一個InstanceKlass對象,實例化的對象則對應InstanceOopDesc,其中InstanceKlass存放在元空間,InstanceOopDesc存放在堆中。數據結構
首先先來看InstanceOopDesc的數據結構,InstanceOopDesc繼承了OopDesc,數據結構以下併發
// 此處爲了方便閱讀,改寫了一下代碼
class instanceOopDesc : public oopDesc {
private:
volatile markOop _mark;
union _metadata {
Klass* _klass;
narrowKlass _compressed_klass;
} _metadata;
};
複製代碼
其中_metadata指向該對象的InstanceKlass,而_mark中則存儲了對象運行時的狀態數據,數據結構以下(圖中爲32位的狀況下的數據,64位也大同小異)oracle
32 bits:
--------
hash:25 ------------>| age:4 biased_lock:1 lock:2 (normal object)
JavaThread*:23 epoch:2 age:4 biased_lock:1 lock:2 (biased object)
size:32 ------------------------------------------>| (CMS free block)
PromotedObject*:29 ---------->| promo_bits:3 ----->| (CMS promoted object)
複製代碼
每一行都表明了一種狀況,描述了哈希碼、GC分代年齡、鎖等狀態信息,以下:jvm
hash: 哈希碼
age: 分代年齡
biased_lock: 偏向鎖標識位
lock: 鎖狀態標識位
JavaThread*: 持有偏向鎖的線程ID
epoch: 偏向時間戳
複製代碼
instanceOopDesc其實保存的是對象的頭部信息,除了頭部信息,對象還有數據,對象數據緊跟着頭部後面,圖示以下:ide
上圖截取了一段程序字節碼,紅線所框對應了Java中new操做的字節碼,Java中的new操做對應了字節碼的三個操做,本文主要講述第一個操做(new)。字節碼中new操做對應JVM中的InterpreterRuntime::_new,代碼以下,函數
// hotspot/share/interpreter/interpreterRuntime.cpp
IRT_ENTRY(void, InterpreterRuntime::_new(JavaThread* thread, ConstantPool* pool, int index))
Klass* k = pool->klass_at(index, CHECK);
InstanceKlass* klass = InstanceKlass::cast(k);
klass->check_valid_for_instantiation(true, CHECK); // 校驗:接口/抽象類/Class不能實例化
klass->initialize(CHECK); // 初始化klass
oop obj = klass->allocate_instance(CHECK); // 分配實例
thread->set_vm_result(obj);
IRT_END
複製代碼
裏面主要包含了兩個部分:初始化klass和分配實例oop
// hotspot/share/oops/instanceKlass.cpp
void InstanceKlass::initialize(TRAPS) {
if (this->should_be_initialized()) {
initialize_impl(CHECK);
} else {
assert(is_initialized(), "sanity check");
}
}
複製代碼
在這裏咱們繼續看initialize_impl()
方法
// hotspot/share/oops/instanceKlass.cpp
void InstanceKlass::initialize_impl(TRAPS) {
HandleMark hm(THREAD);
link_class(CHECK); // 連接class
bool wait = false;
// Step 1
{
Handle h_init_lock(THREAD, init_lock());
ObjectLocker ol(h_init_lock, THREAD, h_init_lock() != NULL);
Thread *self = THREAD;
// Step 2
while(is_being_initialized() && !is_reentrant_initialization(self)) {
wait = true;
ol.waitUninterruptibly(CHECK);
}
// Step 3
if (is_being_initialized() && is_reentrant_initialization(self)) {
DTRACE_CLASSINIT_PROBE_WAIT(recursive, -1, wait);
return;
}
// Step 4
if (is_initialized()) {
DTRACE_CLASSINIT_PROBE_WAIT(concurrent, -1, wait);
return;
}
// Step 5
if (is_in_error_state()) {
DTRACE_CLASSINIT_PROBE_WAIT(erroneous, -1, wait);
ResourceMark rm(THREAD);
const char* desc = "Could not initialize class ";
const char* className = external_name();
size_t msglen = strlen(desc) + strlen(className) + 1;
char* message = NEW_RESOURCE_ARRAY(char, msglen);
if (NULL == message) {
// Out of memory: can't create detailed error message
THROW_MSG(vmSymbols::java_lang_NoClassDefFoundError(), className);
} else {
jio_snprintf(message, msglen, "%s%s", desc, className);
THROW_MSG(vmSymbols::java_lang_NoClassDefFoundError(), message);
}
}
// Step 6
set_init_state(being_initialized);
set_init_thread(self);
}
// Step 7
if (!is_interface()) {
Klass* super_klass = super();
if (super_klass != NULL && super_klass->should_be_initialized()) {
super_klass->initialize(THREAD);
}
if (!HAS_PENDING_EXCEPTION && has_nonstatic_concrete_methods()) {
initialize_super_interfaces(THREAD);
}
if (HAS_PENDING_EXCEPTION) {
Handle e(THREAD, PENDING_EXCEPTION);
CLEAR_PENDING_EXCEPTION;
{
EXCEPTION_MARK;
// Locks object, set state, and notify all waiting threads
set_initialization_state_and_notify(initialization_error, THREAD);
CLEAR_PENDING_EXCEPTION;
}
DTRACE_CLASSINIT_PROBE_WAIT(super__failed, -1, wait);
THROW_OOP(e());
}
}
AOTLoader::load_for_klass(this, THREAD);
// Step 8
{
assert(THREAD->is_Java_thread(), "non-JavaThread in initialize_impl");
JavaThread* jt = (JavaThread*)THREAD;
DTRACE_CLASSINIT_PROBE_WAIT(clinit, -1, wait);
PerfClassTraceTime timer(ClassLoader::perf_class_init_time(),
ClassLoader::perf_class_init_selftime(),
ClassLoader::perf_classes_inited(),
jt->get_thread_stat()->perf_recursion_counts_addr(),
jt->get_thread_stat()->perf_timers_addr(),
PerfClassTraceTime::CLASS_CLINIT);
call_class_initializer(THREAD);
}
// Step 9
if (!HAS_PENDING_EXCEPTION) {
set_initialization_state_and_notify(fully_initialized, CHECK);
{
debug_only(vtable().verify(tty, true);)
}
}
else {
// Step 10 and 11
Handle e(THREAD, PENDING_EXCEPTION);
CLEAR_PENDING_EXCEPTION;
JvmtiExport::clear_detected_exception((JavaThread*)THREAD);
{
EXCEPTION_MARK;
set_initialization_state_and_notify(initialization_error, THREAD);
CLEAR_PENDING_EXCEPTION;
JvmtiExport::clear_detected_exception((JavaThread*)THREAD);
}
DTRACE_CLASSINIT_PROBE_WAIT(error, -1, wait);
if (e->is_a(SystemDictionary::Error_klass())) {
THROW_OOP(e());
} else {
JavaCallArguments args(e);
THROW_ARG(vmSymbols::java_lang_ExceptionInInitializerError(),
vmSymbols::throwable_void_signature(),
&args);
}
}
DTRACE_CLASSINIT_PROBE_WAIT(end, -1, wait);
}
複製代碼
// hotspot/share/oops/instanceKlass.cpp
bool InstanceKlass::link_class_impl(bool throw_verifyerror, TRAPS) {
if (is_linked()) {
return true;
}
assert(THREAD->is_Java_thread(), "non-JavaThread in link_class_impl");
JavaThread* jt = (JavaThread*)THREAD;
// 先連接父類
Klass* super_klass = super();
if (super_klass != NULL) {
if (super_klass->is_interface()) {
ResourceMark rm(THREAD);
Exceptions::fthrow(
THREAD_AND_LOCATION,
vmSymbols::java_lang_IncompatibleClassChangeError(),
"class %s has interface %s as super class",
external_name(),
super_klass->external_name()
);
return false;
}
InstanceKlass* ik_super = InstanceKlass::cast(super_klass);
ik_super->link_class_impl(throw_verifyerror, CHECK_false);
}
// 連接該類的全部藉口
Array<Klass*>* interfaces = local_interfaces();
int num_interfaces = interfaces->length();
for (int index = 0; index < num_interfaces; index++) {
InstanceKlass* interk = InstanceKlass::cast(interfaces->at(index));
interk->link_class_impl(throw_verifyerror, CHECK_false);
}
if (is_linked()) {
return true;
}
// 驗證 & 重寫
{
HandleMark hm(THREAD);
Handle h_init_lock(THREAD, init_lock());
ObjectLocker ol(h_init_lock, THREAD, h_init_lock() != NULL);
if (!is_linked()) {
if (!is_rewritten()) {
{
bool verify_ok = verify_code(throw_verifyerror, THREAD);
if (!verify_ok) {
return false;
}
}
if (is_linked()) {
return true;
}
// 重寫類
rewrite_class(CHECK_false);
} else if (is_shared()) {
SystemDictionaryShared::check_verification_constraints(this, CHECK_false);
}
// 重寫完成後連接方法
link_methods(CHECK_false);
// 初始化vtable和itable
ClassLoaderData * loader_data = class_loader_data();
if (!(is_shared() &&
loader_data->is_the_null_class_loader_data())) {
ResourceMark rm(THREAD);
vtable().initialize_vtable(true, CHECK_false);
itable().initialize_itable(true, CHECK_false);
}
// 將類的狀態標記爲已連接
set_init_state(linked);
if (JvmtiExport::should_post_class_prepare()) {
Thread *thread = THREAD;
assert(thread->is_Java_thread(), "thread->is_Java_thread()");
JvmtiExport::post_class_prepare((JavaThread *) thread, this);
}
}
}
return true;
}
複製代碼
class連接的過程就是這樣,主要步驟總結以下:
關於重寫類和初始化vtable、itable的內容有空新開一章,本文就不描述具體細節了。
這段初始化klass步驟在JVM規範中有詳細描述,假設當前類(接口)爲C,它持有一個獨有的初始化鎖LC
上文爲JVM11規範中的步驟,實際中能夠看到hotspot在實現時和規範所寫略有誤差,但基本差很少。
// hotspot/share/oops/instanceKlass.cpp
instanceOop InstanceKlass::allocate_instance(TRAPS) {
bool has_finalizer_flag = has_finalizer(); // 是否存在非空finalize()方法
int size = size_helper(); // 類的大小
instanceOop i;
i = (instanceOop)Universe::heap()->obj_allocate(this, size, CHECK_NULL); // 分配對象
if (has_finalizer_flag && !RegisterFinalizersAtInit) {
i = register_finalizer(i, CHECK_NULL);
}
return i;
}
複製代碼
在這裏咱們比較關注的是堆空間分配對象環節,
代碼以下:
// hotspot/share/gc/shared/memAllocator.cpp
oop MemAllocator::allocate() const {
oop obj = NULL;
{
Allocation allocation(*this, &obj);
HeapWord* mem = mem_allocate(allocation);
if (mem != NULL) {
obj = initialize(mem);
}
}
return obj;
}
複製代碼
很容易能夠看到,此處的主流程分爲兩個部分,內存分配和初始化。
直接打開代碼,以下:
// hotspot/share/gc/shared/memAllocator.cpp
HeapWord* MemAllocator::mem_allocate(Allocation& allocation) const {
if (UseTLAB) {
HeapWord* result = allocate_inside_tlab(allocation);
if (result != NULL) {
return result;
}
}
return allocate_outside_tlab(allocation);
}
複製代碼
在這段代碼中,咱們能夠看到一個很耳熟的東西——TLAB(ThreadLocalAllocBuffer),默認狀況下TLAB是打開狀態,並且其對Java性能提高很是顯著。首先,先簡單介紹一下TLAB的概念,
由於JVM堆空間是全部線程共享的,所以分配一個對象時會鎖住整個堆,這樣效率就會比較低下。所以JVM在eden區分配了一塊空間做爲線程的私有緩衝區,這個緩衝區稱爲TLAB。不一樣線程不共享TLAB,所以在TLAB中分配對象時是無需上鎖的,從而能夠快速分配。
在這段代碼中,內存分配劃分爲了兩個部分——TLAB內分配和TLAB外分配。
咱們先來看看TLAB內分配的過程,
// hotspot/share/gc/shared/memAllocator.cpp
HeapWord* MemAllocator::allocate_inside_tlab(Allocation& allocation) const {
HeapWord* mem = _thread->tlab().allocate(_word_size);
if (mem != NULL) {
return mem;
}
return allocate_inside_tlab_slow(allocation);
}
複製代碼
一樣的在TLAB的分配的過程當中,也被拆成了兩種狀況,一種是直接使用線程現有的TLAB來進行分配,代碼以下,在下面的這段代碼中,咱們能夠看到TLAB的分配就只是簡單地將top指針向上增長了size大小,而且將原先top的位置分配給了obj,所以分配效率能夠說是極速了。(事實上,TLAB就是經過start、top、end等指針標記了TLAB的存儲信息以及分配空間)
// hotspot/share/gc/shared/threadLocalAllocBuffer.inline.hpp
inline HeapWord* ThreadLocalAllocBuffer::allocate(size_t size) {
invariants(); // 校驗TLAB是否合法
HeapWord* obj = top();
if (pointer_delta(end(), obj) >= size) {
set_top(obj + size);
invariants();
return obj;
}
return NULL;
}
複製代碼
接下來咱們來看看TLAB內的慢分配,
// hotspot/share/gc/shared/memAllocator.cpp
HeapWord* MemAllocator::allocate_inside_tlab_slow(Allocation& allocation) const {
HeapWord* mem = NULL;
ThreadLocalAllocBuffer& tlab = _thread->tlab();
if (ThreadHeapSampler::enabled()) {
tlab.set_back_allocation_end();
mem = tlab.allocate(_word_size);
if (mem != NULL) {
allocation._tlab_end_reset_for_sample = true;
return mem;
}
}
// 若是TLAB的剩餘空間大於閾值,則保留TLAB,這樣就會進入TLAB外分配。在這裏,每次TLAB分配失敗,該TLAB都會調大該閾值,以防線程重複分配一樣大小的對象
if (tlab.free() > tlab.refill_waste_limit()) {
tlab.record_slow_allocation(_word_size);
return NULL;
}
// 計算一個新的TLAB的大小,公式=min{可用空間,期待空間+對象佔據空間,最大TLAB空間}
size_t new_tlab_size = tlab.compute_size(_word_size);
// 清理原先的TLAB。會將剩餘的未使用空間填充進一個假數組,創造EDEN連續的假象,而且將start、end、top等指針所有置爲空
tlab.clear_before_allocation();
if (new_tlab_size == 0) {
return NULL;
}
// 建立一個新的TLAB,空間可能在min_tlab_size到new_tlab_size之間
size_t min_tlab_size = ThreadLocalAllocBuffer::compute_min_size(_word_size);
mem = _heap->allocate_new_tlab(min_tlab_size, new_tlab_size, &allocation._allocated_tlab_size);
if (mem == NULL) {
return NULL;
}
// 將分配的空間數據所有清0
if (ZeroTLAB) {
Copy::zero_to_words(mem, allocation._allocated_tlab_size);
}
// 將mem位置分配word_size大小給obj
tlab.fill(mem, mem + _word_size, allocation._allocated_tlab_size);
return mem;
}
複製代碼
// hotspot/share/gc/shared/memAllocator.cpp
HeapWord* MemAllocator::allocate_outside_tlab(Allocation& allocation) const {
allocation._allocated_outside_tlab = true;
HeapWord* mem = _heap->mem_allocate(_word_size, &allocation._overhead_limit_exceeded);
if (mem == NULL) {
return mem;
}
NOT_PRODUCT(_heap->check_for_non_bad_heap_word_value(mem, _word_size));
size_t size_in_bytes = _word_size * HeapWordSize;
_thread->incr_allocated_bytes(size_in_bytes);
return mem;
}
複製代碼
這裏的核心關注點只有一個——堆內存分配,此處以openjdk11的默認GC——G1爲例,看一看分配的過程。
// hotspot/share/gc/g1/g1CollectedHeap.cpp
HeapWord* G1CollectedHeap::mem_allocate(size_t word_size,
bool* gc_overhead_limit_was_exceeded) {
assert_heap_not_locked_and_not_at_safepoint();
if (is_humongous(word_size)) {
return attempt_allocation_humongous(word_size);
}
size_t dummy = 0;
return attempt_allocation(word_size, word_size, &dummy);
}
複製代碼
在G1中,對象的分配分爲了兩種形式:大對象分配、普通分配。因爲代碼比較長,簡單描述大對象的分配過程以下:
接下來的普通分配過程較爲複雜,本文就再也不深刻探究了。
代碼以下
// hotspot/share/gc/shared/memAllocator.cpp
oop ObjAllocator::initialize(HeapWord* mem) const {
mem_clear(mem);
return finish(mem);
}
複製代碼
其中mem_clear()
方法比較簡單,就是將對象除頭部之外的數據所有置爲0,代碼以下,
// hotspot/share/gc/shared/memAllocator.cpp
void MemAllocator::mem_clear(HeapWord* mem) const {
const size_t hs = oopDesc::header_size();
oopDesc::set_klass_gap(mem, 0);
Copy::fill_to_aligned_words(mem + hs, _word_size - hs);
}
複製代碼
接下來看看finish()
函數,
// hotspot/share/gc/shared/memAllocator.cpp
oop MemAllocator::finish(HeapWord* mem) const {
assert(mem != NULL, "NULL object pointer");
if (UseBiasedLocking) {
oopDesc::set_mark_raw(mem, _klass->prototype_header());
} else {
oopDesc::set_mark_raw(mem, markOopDesc::prototype());
}
oopDesc::release_set_klass(mem, _klass);
return oop(mem);
}
複製代碼
還記得對象頭中有兩個屬性mark和metadata嗎?finish()
方法就是設置對象的頭部數據。
因爲平時幾乎不多用到finalize(),且內部邏輯比較複雜,所以本文暫時不探究finalize的註冊機制。
整個JVM對象分配的總體流程大體以下,
[1] The Java® Virtual Machine Specification Java SE 11 Edition