Android binder from Top to Bottom
本文分六部分介绍Android binder机制和结构; 理解和行文难免有误, 一切以代码为准.
转载请注明出处.
Android Binder机制之RefBase & smart pointer
在组件技术中必然需要引用计数的机制，用于控制指针所指向的本端代理对象和所对应的远端进程内的组件对象的生命周期。
LPC通用行为和smart pointer
如同Windows COM二进制兼容跨进程接口调用机制，LPC（Local Procedure Call）三要素：
QueryInterface
AddRef/ReleaseRef
Method Invoke
另外，实现上对于本进程内的组件，要可以直接引用，这关乎运行效率。这也是LPC实现普遍要考虑的一个点（特别是用socket实现时）。
对于Android Binder Framework，Interface Query由ServiceManager负责，AddRef/ReleaseRef由RefBase计数负责；Method Invoke由对应Service的Interface和Binder负责。
正是因为组件有计数功能，smart pointer的“smart”才有了着落，即通过指针对象自身的生命周期来增减引用的组件的计数，从而控制组件的生命周期。
在Android中，android::RefBase及其派生类负责计数，而smart pointer（sp
RefBase和sp/wp的静态关系如下图(此图从网络引来)：
计数的祖宗---基类RefBase和LightRefBase
实现计数的对象的类是RefBase和LightRefBase。绝大多数情况下的类是继承自RefBase。
RefBase虽然借助关联的嵌套类对象实现了强引用计数和弱引用计数，但是对外只表露出了增减强引用计数的方法(如incStrong、decStrong)。
LightRefBase的Light表现在只实现强引用计数，没有弱引用计数！LightRefBase主要用在OpenGL ES需要的FrameBuffer NativeWindow实现中。
RefBase的设计原则
1. 因为调用incStrong/decStrong的时候，需要相应地调用incWeak/decWeak，如果表露出incWeak/decWeak就需要外部遵循不同情景的调用规则，可能会导致计数的混乱，所以incWeak/decWeak尽量不对外使用，内部实现类维护。
2. incStrong/decStrong表露出来是因为有些情况，直接由普通指针调用，否则，都不应该表露出，而是应该sp
3. RefBase既然需要对象的强引用计数和弱引用计数，由于强引用指针和弱引用指针的相互引用和相互赋值及弱引用提升为强引用，所以，为了设计的方便和运行的效率，强/弱引用计数及其操作的内部实现又不能过于独立。
4. 弱引用计数接口（android::RefBase::weakref_type）供弱引用计数实现使用，需要分离出来。
综上设计原则，于是有了Android RefBase的略显怪胎式的实现：
In frameworks/base/include/utils/RefBase.h
class RefBase
{
public:
void incStrong(const void* id) const;
void decStrong(const void* id) const;
void forceIncStrong(const void* id) const;
//! DEBUGGING ONLY: Get current strong ref count.
int32_t getStrongCount() const;
class weakref_type
{
public:
RefBase* refBase() const;
void incWeak(const void* id);
void decWeak(const void* id);
// acquires a strong reference if there is already one.
bool attemptIncStrong(const void* id);
// acquires a weak reference if there is already one.
// This is not always safe. see ProcessState.cpp and BpBinder.cpp
// for proper use.
bool attemptIncWeak(const void* id);
//! DEBUGGING ONLY: Get current weak ref count.
int32_t getWeakCount() const;
//! DEBUGGING ONLY: Print references held on object.
void printRefs() const;
//! DEBUGGING ONLY: Enable tracking for this object.
// enable -- enable/disable tracking
// retain -- when tracking is enable, if true, then we save a stack trace
// for each reference and dereference; when retain == false, we
// match up references and dereferences and keep only the
// outstanding ones.
void trackMe(bool enable, bool retain);
};
weakref_type* createWeak(const void* id) const;
weakref_type* getWeakRefs() const;
//! DEBUGGING ONLY: Print references held on object.
inline void printRefs() const { getWeakRefs()->printRefs(); }
//! DEBUGGING ONLY: Enable tracking of object.
inline void trackMe(bool enable, bool retain)
{
getWeakRefs()->trackMe(enable, retain);
}
typedef RefBase basetype;
protected:
RefBase();
virtual ~RefBase();
//! Flags for extendObjectLifetime()
enum {
OBJECT_LIFETIME_STRONG = 0x0000,
OBJECT_LIFETIME_WEAK = 0x0001,
OBJECT_LIFETIME_MASK = 0x0001
};
void extendObjectLifetime(int32_t mode);
//! Flags for onIncStrongAttempted()
enum {
FIRST_INC_STRONG = 0x0001
};
virtual void onFirstRef();
virtual void onLastStrongRef(const void* id);
virtual bool onIncStrongAttempted(uint32_t flags, const void* id);
virtual void onLastWeakRef(const void* id);
private:
friend class ReferenceMover;
static void moveReferences(void* d, void const* s, size_t n,
const ReferenceConverterBase& caster);
private:
friend class weakref_type;
class weakref_impl;
RefBase(const RefBase& o); //copy constructor
RefBase& operator=(const RefBase& o); //assignment operator overload
weakref_impl* const mRefs;
};
In frameworks/base/libs/utils/RefBase.cpp
class RefBase::weakref_impl : public RefBase::weakref_type
{
public:
volatile int32_t mStrong;
volatile int32_t mWeak;
RefBase* const mBase; //de-refer to RefBase
volatile int32_t mFlags;
#if !DEBUG_REFS
weakref_impl(RefBase* base)
: mStrong(INITIAL_STRONG_VALUE)
, mWeak(0)
, mBase(base)
, mFlags(0)
{
}
void addStrongRef(const void* /*id*/) { }
void removeStrongRef(const void* /*id*/) { }
void renameStrongRefId(const void* /*old_id*/, const void* /*new_id*/) { }
void addWeakRef(const void* /*id*/) { }
void removeWeakRef(const void* /*id*/) { }
void renameWeakRefId(const void* /*old_id*/, const void* /*new_id*/) { }
void printRefs() const { }
void trackMe(bool, bool) { }
#else
………
mutable Mutex mMutex;
ref_entry* mStrongRefs;
ref_entry* mWeakRefs;
bool mTrackEnabled;
// Collect stack traces on addref and removeref, instead of deleting the stack references
// on removeref that match the address ones.
bool mRetain;
#endif
};
可以看出，RefBase::weakref_type采用内部类分离出来，做到弱引用接口的独立。class RefBase::weakref_impl继承自RefBase::weakref_type，实现了弱计数需要的接口incWeak/decWeak。在实现RefBase的弱计数功能时，调用weakref_impl弱引用计数相关的接口。RefBase::weakref_impl对象和被计数的 DerivedRefbase对象分离，是弱引用计数实现的需要，因为RefBase对象生命周期仅受强引用控制时，wp
RefBase::weakref_impl的函数addStrongRef/ removeStrongRef，addWeakRef/ removeWeakRef等用来追踪对象的被哪些指针强弱引用的，可用于调试。
RefBase::incStrong/RefBase::decStrong则由RefBase类自己实现，会分别调用RefBase::weakref_type::incWeak/ RefBase::weakref_type::decWeak函数，即强引用计数时一定会伴随着弱引用计数发生。
总之，虽然可读性不怎么样，但是却体现了上述设计原则。
RefBase::weakref_impl己经不仅仅是个weakref实现，除了没有直接实现incStrong/decStrong函数，已经是一个强弱计数RefBase::ref_impl的实现。但是，反过来说，mStrong却一定不能放在RefBase基类中。
下面是强弱计数实现的两个示例函数。
void RefBase::incStrong(const void* id) const
{
weakref_impl* const refs = mRefs;
refs->incWeak(id);
refs->addStrongRef(id);
const int32_t c =android_atomic_inc(&refs->mStrong);
LOG_ASSERT(c > 0, "incStrong() called on %p after last strong ref", refs);
#if PRINT_REFS
LOGD("incStrong of %p from %p: cnt=%d\n", this, id, c);
#endif
if (c != INITIAL_STRONG_VALUE) {
return;
}
android_atomic_add(-INITIAL_STRONG_VALUE, &refs->mStrong);
refs->mBase->onFirstRef();
}
void RefBase::weakref_type::incWeak(const void* id)
{
weakref_impl* const impl = static_cast
impl->addWeakRef(id);
const int32_t c = android_atomic_inc(&impl->mWeak);
LOG_ASSERT(c >= 0, "incWeak called on %p after last weak ref", this);
}
从weaktype_impl代码中可以看到mStrong在其中，weaktype_impl对于强引用也有追踪，所以weaktype_impl可以直接实现强计数的几个接口，做为强弱引用计数的通用实现ref_impl。只在weaktype_impl中访问其成员如mStrong，也符合数据封装的原则。然后RefBase直接调用其对应的强引用计数接口实现即可。如：
inline Refbase::incStrong(const void* id) const
{
ref_impl* const refs = mRefs;
refs->incStrong(id)；
}
另外，mStrong的初始值（INITIAL_STRONG_VALUE）不为0的原因是将初始值和递减计数到0做区别。系统中不可能有INITIAL_STRONG_VALUE这么多个引用到对象。
RefBase的生命周期
使用函数extendObjectLifetime改变默认的对象LIFETIME MODE，默认是LIFETIME_STRONG，即只受强计数控制；改为LIFETIME_WEAK，只有弱计数也为零时，才释放m_ptr。另，弱指针提升时用到attemptIncStrong。
BpBinder::BpBinder(int32_t handle)
: mHandle(handle)
, mAlive(1)
, mObitsSent(0)
, mObituaries(NULL)
{
LOGV("Creating BpBinder %p handle %d\n", this, mHandle);
extendObjectLifetime(OBJECT_LIFETIME_WEAK);
IPCThreadState::self()->incWeakHandle(handle);
}
BpRefBase::BpRefBase(const sp
: mRemote(o.get()), mRefs(NULL), mState(0)
{
extendObjectLifetime(OBJECT_LIFETIME_WEAK);
if (mRemote) {
mRemote->incStrong(this); // Removed on first IncStrong().
mRefs = mRemote->createWeak(this); // Held for our entire lifetime.
}
}
void BpRefBase::onFirstRef()
{
android_atomic_or(kRemoteAcquired, &mState);
}
在此适度解释BpBinder和BpRefBase调用extendObjectLifetime的用途。
BpBinder and base class of BpSubInterface, BpRefBase are OBJECT_LIFETIME_WEAK life mode controlled; BpBinder和BpSubInterface两个对象本身是弱计数控制的。
BpSubInterface级别的BpRefBase增减BpBinder的强弱计数，而BpBinder增减远端组件的计数，准确的说是增减binder驱动中binder_ref的强弱计数通过控制binder_node的计数，间接控制BBinder子类的强弱计数。析构时，反之亦反。BpBinder从Binder级别控制；而BpRefbase从Interface级别控制。
如sp
当getService返回binder handle构造出BpBinder时，使用inWeakHandle增加binder_ref的弱计数；构造BpSubInterface时，表明BpBinder被引用，增加BpBinder的强计数引发其onFirstRef通过incStrongHandle增加binder_ref强计数；赋值给sp
此处增加flow图。
考察分离对于生命周期控制的作用[超前]！！！可以考察以下语句的执行来探索（不考虑编译器优化，需要结合binder driver）。
sp
sp
sp
sp
{
sp
AutoMutex _l(mLock);
handle_entry* e = lookupHandleLocked(handle);
if (e != NULL) {
// We need to create a new BpBinder if there isn't currently one, OR we
// are unable to acquire a weak reference on this current one. See comment
// in getWeakProxyForHandle() for more info about this.
IBinder* b = e->binder;
if (b == NULL || !e->refs->attemptIncWeak(this)) {
b = new BpBinder(handle);
e->binder = b;
if (b) e->refs = b->getWeakRefs();
result = b;
} else {
// This little bit of nastyness is to allow us to add a primary
// reference to the remote proxy when this team doesn't have one
// but another team is sending the handle to us.
result.force_set(b);
e->refs->decWeak(this);
}
}
return result;
}
wp
{
wp
AutoMutex _l(mLock);
handle_entry* e = lookupHandleLocked(handle);
if (e != NULL) {
// We need to create a new BpBinder if there isn't currently one, OR we
// are unable to acquire a weak reference on this current one. The
// attemptIncWeak() is safe because we know the BpBinder destructor will always
// call expungeHandle(), which acquires the same lock we are holding now.
// We need to do this because there is a race condition between someone
// releasing a reference on this BpBinder, and a new reference on its handle
// arriving from the driver.
IBinder* b = e->binder;
if (b == NULL || !e->refs->attemptIncWeak(this)) {
b = new BpBinder(handle);
result = b;
e->binder = b;
if (b) e->refs = b->getWeakRefs();
} else {
result = b;
e->refs->decWeak(this);
}
}
return result;
}
与binder object向service manager注册时有什么关系？？
RefBase类的三个重要虚函数：
virtual void RefBase::onFirstRef() {}
virtual void RefBase::onLastStrongRef(const void* /*id*/) {}
virtual void RefBase::onLastWeakRef(const void* /*id*/) {}
此三个虚函数，默认实现为空。分别用于当开始强计数、强计数减为0和弱计数减为0时，提供一个处理时机；一个重要用途是提供代理类对组件类计数的调用点，其大致作用如下。
子类BpBinder和BBinder重载了这三个函数。BpBinder的重载如下：
void BpBinder::onFirstRef()
{
LOGV("onFirstRef BpBinder %p handle %d\n", this, mHandle);
IPCThreadState* ipc = IPCThreadState::self();
if (ipc) ipc->incStrongHandle(mHandle);
}
void BpBinder::onLastStrongRef(const void* id)
{
LOGV("onLastStrongRef BpBinder %p handle %d\n", this, mHandle);
IF_LOGV() {
printRefs();
}
IPCThreadState* ipc = IPCThreadState::self();
if (ipc) ipc->decStrongHandle(mHandle);
}
BBinder的会留给具体的服务类实现自己重载。onFirstRef主要用于完成服务的一些初始化工作。Binder组件对象是继承自BnSubInterface（BBinder的子类），收到BR_ACQUIRE和BR_RELEASE后，会使用RefBase的引用计数操作incStrong/decStrong等函数实现来响应，不同的是在使用android_atomic_inc/android_atomic_dec完成引用计数操作后，调用到onFirstRef和onLastStrongRef虚函数时，调用到的是继承自RefBase的两个空实现，什么也不做。这是因为本对象的计数操作是直接增减，已经操作完毕。
当客户和服务处于同一进程空间时，getService返回的IBinder指针直接指向BBinder类型对象，所以当incStrong、decStrong、incWeak、decWeak时也是直接调用BBinder的相应函数，调用到RefBase的相应实现，过程同客户和服务在不同进程时远端组件的响应过程。
总结：RefBase通过IncStrong/decStrong/inWeak/decWeak实现了指针直接所指的本地代理对象的引用计数的操作，代理子类BpBinder重载RefBase的onFirstRef和onLastStrongRef的默认空实现。
为什么只有一个onFirstRef而没有分成onFirstStrongRef和onFirstWeakRef？onFirstRef的准确意义就是onFirstStrongRef()。具体在sp
smart pointer ---sp
Android下C++中的smart pointer分为强指针(Strong Pointer)sp和弱指针(Weak Pointer)wp。没有类似虚拟机层面的垃圾回收模块，而仅是利用c++的语言特性来解决内存使用和自动回收的一种手段。
wp in frameworks/base/include/utils/RefBase.h
template
class wp
{
public:
typedef typename RefBase::weakref_type weakref_type;
inline wp() : m_ptr(0) { }
wp(T* other);
wp(const wp
wp(const sp
template
template
template
~wp();
// Assignment
wp& operator = (T* other);
wp& operator = (const wp
wp& operator = (const sp
template
template
template
void set_object_and_refs(T* other, weakref_type* refs);
// promotion to sp
sp
// Reset
void clear();
// Accessors
inline weakref_type* get_refs() const { return m_refs; }
inline T* unsafe_get() const { return m_ptr; }
// Operators
COMPARE_WEAK(==)
COMPARE_WEAK(!=)
COMPARE_WEAK(>)
COMPARE_WEAK(<)
COMPARE_WEAK(<=)
COMPARE_WEAK(>=)
inline bool operator == (const wp
return (m_ptr == o.m_ptr) && (m_refs == o.m_refs);
}
template
inline bool operator == (const wp& o) const {
return m_ptr == o.m_ptr;
}
inline bool operator > (const wp
return (m_ptr == o.m_ptr) ? (m_refs > o.m_refs) : (m_ptr > o.m_ptr);
}
template
inline bool operator > (const wp& o) const {
return (m_ptr == o.m_ptr) ? (m_refs > o.m_refs) : (m_ptr > o.m_ptr);
}
inline bool operator < (const wp
return (m_ptr == o.m_ptr) ? (m_refs < o.m_refs) : (m_ptr < o.m_ptr);
}
template
inline bool operator < (const wp& o) const {
return (m_ptr == o.m_ptr) ? (m_refs < o.m_refs) : (m_ptr < o.m_ptr);
}
inline bool operator != (const wp
template
inline bool operator <= (const wp
template
inline bool operator >= (const wp
template
private:
template
template
T* m_ptr;
weakref_type* m_refs;
};
其中promote用来尝试提升为sp
wp
: m_ptr(other)
{
if (other) m_refs = other->createWeak(this);
}
template
sp
{
sp
if (m_ptr && m_refs->attemptIncStrong(&result)) {
result.set_pointer(m_ptr);
}
return result;
}
promote()不成功时，会直接返回空sp
attemptIncStrong在由弱指针提升到强指针时用到；而attemptIncWeak在OBJECT_LIFETIME_FOREVER时试图增加弱计数用到。
bool RefBase::weakref_type::attemptIncWeak(const void* id)
{
weakref_impl* const impl = static_cast
int32_t curCount = impl->mWeak;
LOG_ASSERT(curCount >= 0, "attemptIncWeak called on %p after underflow",
this);
while (curCount > 0) {
if (android_atomic_cmpxchg(curCount, curCount+1, &impl->mWeak) == 0) {
break;
}
curCount = impl->mWeak;
}
if (curCount > 0) {
impl->addWeakRef(id);
}
return curCount > 0;
}
bool RefBase::weakref_type::attemptIncStrong(const void* id)
{
incWeak(id);
weakref_impl* const impl = static_cast
int32_t curCount = impl->mStrong;
LOG_ASSERT(curCount >= 0, "attemptIncStrong called on %p after underflow",
this);
while (curCount > 0 && curCount != INITIAL_STRONG_VALUE) {
if (android_atomic_cmpxchg(curCount, curCount+1, &impl->mStrong) == 0) {
break;
}
curCount = impl->mStrong;
}
if (curCount <= 0 || curCount == INITIAL_STRONG_VALUE) {
bool allow;
if (curCount == INITIAL_STRONG_VALUE) {
// Attempting to acquire first strong reference... this is allowed
// if the object does NOT have a longer lifetime (meaning the
// implementation doesn't need to see this), or if the implementation
// allows it to happen.
allow = (impl->mFlags&OBJECT_LIFETIME_WEAK) != OBJECT_LIFETIME_WEAK
|| impl->mBase->onIncStrongAttempted(FIRST_INC_STRONG, id);
} else {
// Attempting to revive the object... this is allowed
// if the object DOES have a longer lifetime (so we can safely
// call the object with only a weak ref) and the implementation
// allows it to happen.
allow = (impl->mFlags&OBJECT_LIFETIME_WEAK) == OBJECT_LIFETIME_WEAK
&& impl->mBase->onIncStrongAttempted(FIRST_INC_STRONG, id);
}
if (!allow) {
decWeak(id);
return false;
}
curCount = android_atomic_inc(&impl->mStrong);
// If the strong reference count has already been incremented by
// someone else, the implementor of onIncStrongAttempted() is holding
// an unneeded reference. So call onLastStrongRef() here to remove it.
// (No, this is not pretty.) Note that we MUST NOT do this if we
// are in fact acquiring the first reference.
if (curCount > 0 && curCount < INITIAL_STRONG_VALUE) {
impl->mBase->onLastStrongRef(id);
}
}
impl->addStrongRef(id);
#if PRINT_REFS
LOGD("attemptIncStrong of %p from %p: cnt=%d\n", this, id, curCount);
#endif
if (curCount == INITIAL_STRONG_VALUE) {
android_atomic_add(-INITIAL_STRONG_VALUE, &impl->mStrong);
impl->mBase->onFirstRef();
}
return true;
}
Note:
android_atomic_cmpxchg return 0 for success, not failure.所以attemptInc时，是只要没有lock xchg成功就一直尝试增加，直至成功。
forceIncStrong在sp
void RefBase::forceIncStrong(const void* id) const
{
weakref_impl* const refs = mRefs;
refs->incWeak(id);
refs->addStrongRef(id);
const int32_t c = android_atomic_inc(&refs->mStrong);
LOG_ASSERT(c >= 0, "forceIncStrong called on %p after ref count underflow",
refs);
#if PRINT_REFS
LOGD("forceIncStrong of %p from %p: cnt=%d\n", this, id, c);
#endif
switch (c) {
case INITIAL_STRONG_VALUE:
android_atomic_add(-INITIAL_STRONG_VALUE, &refs->mStrong);
// fall through...
case 0:
refs->mBase->onFirstRef();
}
}
sp in frameworks/base/include/utils/StrongPointer.h
template
class sp
{
public:
inline sp() : m_ptr(0) { }
sp(T* other);
sp(const sp
template
template
~sp();
// Assignment
sp& operator = (T* other);
sp& operator = (const sp
template
template
//! Special optimization for use by ProcessState (and nobody else).
void force_set(T* other);
// Reset
void clear();
// Accessors
inline T& operator* () const { return *m_ptr; }
inline T* operator-> () const { return m_ptr; }
inline T* get() const { return m_ptr; }
// Operators
COMPARE(==)
COMPARE(!=)
COMPARE(>)
COMPARE(<)
COMPARE(<=)
COMPARE(>=)
private:
template
template
void set_pointer(T* ptr);
T* m_ptr;
};
sp
可以看出wp
sp
T& operator*(){return *m_ptr}
T* operator->(){return m_ptr}
这不仅仅是让智能指针的操作更像是指针；而是增加’->’左值结合，返回T*，从而能且仅能以指针形式访问T的成员或方法；wp
其余，sp
sp
template
void wp
{
if (m_ptr) {
m_refs->decWeak(this);
m_ptr = 0;
}
}
template
void sp
{
if (m_ptr) {
m_ptr->decStrong(this);
m_ptr = 0;
}
}
对象内存布局引用关系如：
+-------------------------------------------------+
+---------------+---------------->+------------------+ |-->+--------------------+
| wp | +---->|===========|---------------->| |
+---------------+ | | | | weakref_impl |
+---------------+ | | RefBase | | |
| sp |--------+ | | +--------------------+
+---------------+ +------------------+
现在来描述上节的问题为什么RefBase的onFirstRef()就是onFirstStrongRef而不可能有onFirstWeakRef？因为只有弱计数即只通过wp
int * mem = (int*)malloc(sizeof(DerivedRefBase));
DerivedRefBase*embryo = (DerivedRefBase*)mem;
mem[OFF_REFS] = new RefBase::weakref_impl(embyro); // to access privateRefBase::mRefs
//对象还未构造;不会出错,对象不被使用wp
wp
// 在内存上构造对象；注，DerivedRefBase使构造函数为public；会新建并覆盖mRefs
DerivedRefBase * p = new (mem) DerivedRefBase;
// 替换新创建的没有引用计数的mRefs，为wobj中记录的
delete((RefBase::weakref_impl*)mem[OFF_REFS]);
mem[OFF_REFS] = wobj. get_refs();
// 在AttemptIncStrong中如果成功,会调用onFirstRef
sp
/* 参考代码
template
wp
{
if (other) m_refs = other->createWeak(this);
}
RefBase::weakref_type* RefBase::createWeak(const void* id) const
{
mRefs->incWeak(id);
return mRefs;
}
RefBase::RefBase() : mRefs(new weakref_impl(this))
{
}
*/
Android Binder机制之IBinder & IInterface
Binder的工作机制示意图如下图（摘自网络，出处如标）：
IBinder接口需要为IInterface提供基础通信能力，而IInterface接口则负责定义具体的功能。对于通用的跨进程调用机制的三个要素，Query[Remote]Interface由ServiceManager管理，AddRef/ReleaseRef由基类Refbase实现并通过onFirstRef/onLastXxxxRef形成引用菊花链；Method Invoke分别由IBinder::transact(…)函数实现。
下面将通过查看代码的方式分析binder和interface。
IBinder & BpBinder & BBinder
IBinder接口的定义摘录如下：
/**
* Base class and low-level protocol for a remotable object.
* You can derive from this class to create an object for which other
* processes can hold references to it. Communication between processes
* (method calls, property get and set) is down through a low-level
* protocol implemented on top of the transact() API.
*/
class IBinder : publicvirtual RefBase
{
public:
enum {
FIRST_CALL_TRANSACTION = 0x00000001,
LAST_CALL_TRANSACTION = 0x00ffffff,
PING_TRANSACTION = B_PACK_CHARS('_','P','N','G'),
DUMP_TRANSACTION = B_PACK_CHARS('_','D','M','P'),
INTERFACE_TRANSACTION = B_PACK_CHARS('_', 'N', 'T', 'F'),
// Corresponds to TF_ONE_WAY -- an asynchronous call.
FLAG_ONEWAY = 0x00000001
};
IBinder();
/**
* Check if this IBinder implements the interface named by
* @a descriptor. If it does, the base pointer to it is returned,
* which you can safely static_cast<> to the concrete C++ interface.
*/
virtual sp
/**
* Return thecanonical name of the interface provided by this IBinder
* object.
*/
virtual const String16&getInterfaceDescriptor() const = 0;
virtual bool isBinderAlive() const = 0;
virtual status_t pingBinder() = 0;
virtual status_t dump(int fd, const Vector
virtual status_t transact( uint32_t code, const Parcel& data, Parcel* reply, uint32_tflags = 0) = 0;
/**
* This method allows you to add data that is transported through
* IPC along with your IBinder pointer. When implementing a Binder
* object, override it to write your desired data in to @a outData.
* You can then call getConstantData() on your IBinder to retrieve
* that data, from any process. You MUST return the number of bytes
* written in to the parcel (including padding).
*/
class DeathRecipient : publicvirtual RefBase
{
public:
virtual voidbinderDied(const wp
};
/**
* Register the @a recipient for a notification if this binder
* goes away. If this binder object unexpectedly goes away
* (typically because its hosting process has been killed),
* then DeathRecipient::binderDied() will be called with a reference
* to this.
*
* The @a cookie is optional -- if non-NULL, it should be a
* memory address that you own (that is, you know it is unique).
*
* @note You will only receive death notifications for remote binders,
* as local binders by definition can't die without you dying as well.
* Trying to use this function on a local binder will result in an
* INVALID_OPERATION code being returned and nothing happening.
*
* @note This link always holds a weak reference to its recipient.
*
* @note You will only receive a weak reference to the dead
* binder. You should not try to promote this to a strong reference. (the binder nodehas died. lzz)
* (Nor should you need to, as there is nothing useful you can
* directly do with it now that it has passed on.)
*/
virtual status_t linkToDeath(const sp
void* cookie = NULL,
uint32_t flags = 0) = 0;
/**
* Remove a previously registered death notification.
* The @a recipient will no longer be called if this object
* dies. The @a cookie is optional. If non-NULL, you can
* supply a NULL @a recipient, and the recipient previously
* added with that cookie will be unlinked.
*/
virtual status_t unlinkToDeath( const wp
void* cookie = NULL,
uint32_t flags = 0,
wp
virtual bool checkSubclass(const void* subclassID) const;
typedef void (*object_cleanup_func)(const void* id, void* obj, void* cleanupCookie);
virtual void attachObject( const void* objectID,
void* object,
void* cleanupCookie,
object_cleanup_func func) = 0;
virtual void* findObject(const void* objectID) const = 0;
virtual void detachObject(const void* objectID) = 0;
virtual BBinder* localBinder();
virtual BpBinder* remoteBinder();
protected:
virtual ~IBinder();
private:
};
可以看到这几个函数非虚即纯虚，在设计上，这说明子类BpBinder和BBinder会重载这几个函数。QueryLocalInterface、localBinder、remoteBinder三个虚函数的默认实现都是返回NULL，基本是空实现，子类根据需要重载特定的即可；而transact是要子类实现特定跨进程调用Method，基类（接口）transact当然是纯虚的。
其中，localBinder用于判断Binder是否是和客户同进程的服务，若是，则直接返回指向BBinder对象的指针；remoteBinder用于判断服务是否是远端服务，若是，则返回本端代理对象BpBinder的指针。BpBinder重载了remoteBinder {return this;}，而BBinder重载了localBinder {return this;}，这就为IBinder对象提供了一种非常初级的能力，用于判断实际组件是本端还是远端。在flatten_binder_object(...)/unflatten_binder_object(...)中会使用。
queryLocalInterface()虚函数默认实现返回NULL，从Binder到Interface的转换会用到， BBinder和BpBinder类都没有重载，BBinder子类BnInterface
getInterfaceDescriptor()纯虚函数，用于Check Interface Token。BBinder和BpBinder分别做了实现，BpBinder通过Binder通信获得BBinder的InterfaceDescriptor。BBinder的实现基本是个空实现。后面会看到子类BnInterface
const String16& BBinder::getInterfaceDescriptor() const @ Binder.cpp
{
// This is a local static rather than a global static,
// to avoid static initializer ordering issues.
static String16 sEmptyDescriptor;
LOGW("reached BBinder::getInterfaceDescriptor (this=%p)", this);
return sEmptyDescriptor;
}
const String16& BpBinder::getInterfaceDescriptor() const @ BpBinder.cpp
{
if (isDescriptorCached() == false) {
Parcel send, reply;
// do the IPC without a lock held.
status_t err = const_cast
INTERFACE_TRANSACTION, send, &reply);
if (err == NO_ERROR) {
String16 res(reply.readString16());
Mutex::Autolock _l(mLock);
// mDescriptorCache could have been assigned while the lock was
// released.
if (mDescriptorCache.size() == 0)
mDescriptorCache = res;
}
}
// we're returning a reference to a non-static object here. Usually this
// is not something smart to do, however, with binder objects it is
// (usually) safe because they are reference-counted.
return mDescriptorCache;
}
status_t BBinder::onTransact(uint32_t code, const Parcel& data, Parcel* reply, uint32_t flags)
{
switch (code) {
case INTERFACE_TRANSACTION:
reply->writeString16(getInterfaceDescriptor());
return NO_ERROR;
case DUMP_TRANSACTION: {
.......
}
default:
return UNKNOWN_TRANSACTION;
}
}
linkToDeath/unlinkToDeath/ DeathRecipient::binderDied，用于注册binder死亡通知。因为binder服务组件是否还存在，其进程是否还存在，这是客户端要关心的。
attachObject/findObject/detachObject从名字就可以看出是用于client side将BpBinder绑定到已存在的对象上。主要是用于java binder object。
BpBinder用于client端的通讯处理，类声明如下：
In frameworks/base/include/binder/BpBinder.h
class BpBinder : public IBinder
{
public:
BpBinder(int32_t handle);
inline int32_t handle() const { return mHandle; }
virtual const String16& getInterfaceDescriptor() const;
virtual bool isBinderAlive() const;
virtual status_t pingBinder();
virtual status_t dump(int fd, const Vector
virtual status_t transact( uint32_t code,
const Parcel& data,
Parcel* reply,
uint32_t flags = 0);
virtual status_t linkToDeath(const sp
void* cookie = NULL,
uint32_t flags = 0);
virtual status_t unlinkToDeath( const wp
void* cookie = NULL,
uint32_t flags = 0,
wp
virtual void attachObject( const void* objectID,
void* object,
void* cleanupCookie,
object_cleanup_func func);
virtual void* findObject(const void* objectID) const;
virtual void detachObject(const void* objectID);
virtual BpBinder* remoteBinder();
status_t setConstantData(const void* data, size_t size);
void sendObituary();
class ObjectManager
{
public:
ObjectManager();
~ObjectManager();
void attach( const void* objectID,
void* object,
void* cleanupCookie,
IBinder::object_cleanup_func func);
void* find(const void* objectID) const;
void detach(const void* objectID);
void kill();
private:
ObjectManager(const ObjectManager&);
ObjectManager& operator=(const ObjectManager&);
struct entry_t
{
void* object;
void* cleanupCookie;
IBinder::object_cleanup_func func;
};
KeyedVector
};
protected:
virtual ~BpBinder();
virtual void onFirstRef();
virtual void onLastStrongRef(const void* id);
virtual bool onIncStrongAttempted(uint32_t flags, const void* id);
private:
const int32_t mHandle;
struct Obituary {
wp
void* cookie;
uint32_t flags;
};
void reportOneDeath(const Obituary& obit);
bool isDescriptorCached() const;
mutable Mutex mLock; // lock used to operate Obituaries
volatile int32_t mAlive;
volatile int32_t mObitsSent; // if obituary is sent
Vector
ObjectManager mObjects; // Java objects attached to
Parcel* mConstantData;
mutable String16 mDescriptorCache;
};
BpBinder的最实质性的内容就是包含了mHandle to binder ref。其余成员见注释，函数功能同IBinder接口。
BBinder用于service端的通讯处理，类声明如下：
In frameworks/base/include/binder/Binder.h
class BBinder : public IBinder
{
public:
BBinder();
virtual const String16& getInterfaceDescriptor() const;
virtual bool isBinderAlive() const;
virtual status_t pingBinder();
virtual status_t dump(int fd, const Vector
virtual status_t transact( uint32_t code,
const Parcel& data,
Parcel* reply,
uint32_t flags = 0);
virtual status_t linkToDeath(const sp
void* cookie = NULL,
uint32_t flags = 0);
virtual status_t unlinkToDeath( const wp
void* cookie = NULL,
uint32_t flags = 0,
wp
virtual void attachObject( const void* objectID,
void* object,
void* cleanupCookie,
object_cleanup_func func);
virtual void* findObject(const void* objectID) const;
virtual void detachObject(const void* objectID);
virtual BBinder* localBinder();
protected:
virtual ~BBinder();
virtual status_t onTransact( uint32_t code,
const Parcel& data,
Parcel* reply,
uint32_t flags = 0);
private:
BBinder(const BBinder& o);
BBinder& operator=(const BBinder& o);
class Extras;
Extras* mExtras;
void* mReserved0;
};
class BBinder::Extras
{
public:
Mutex mLock;
BpBinder::ObjectManagermObjects;
};
mExtras作用同BpBinder的mObjects，用于service side的java binder node object绑定。
IInterface & BpInterface & BnInterface
IInterface用来定义接口的跨进程调用方法（Interface的特指），是所有自定义ISubInterface接口（IBinder当然是例外）的基类，在frameworks/base/includes/binder/Iinterface.h中，声明如下：
class IInterface : publicvirtual RefBase
{
public:
sp
sp
protected:
virtual IBinder* onAsBinder() = 0;
};
从类声明可以看出这个基类中的函数主要是由IInterface到IBinder的转换，得到对应的binder。使用代码示例：
sp
sp
还是在IInterface.h中，定义了一个interface_cast模板函数，用于从IBinder转换得到IInterface。
template
inline sp
{
return INTERFACE::asInterface(obj);
}
使用代码示例：
sp
BnInterface和BpInterface模板类继承于自定义的ISubInterface（模板中记为即为INTERFACE）。声明在IInterface.h中：
template
class BnInterface : public INTERFACE, public BBinder
{
public:
virtual sp
virtual const String16& getInterfaceDescriptor() const;
protected:
virtual IBinder* onAsBinder();
};
template
class BpInterface : public INTERFACE, public BpRefBase
{
public:
BpInterface(const sp
protected:
virtual IBinder* onAsBinder();
};
class BpRefBase : public virtual RefBase
{
protected:
BpRefBase(const sp
virtual ~BpRefBase();
virtual void onFirstRef();
virtual void onLastStrongRef(const void* id);
virtual bool onIncStrongAttempted(uint32_t flags, const void* id);
inline IBinder* remote() { return mRemote; }
inline IBinder* remote() const { return mRemote; }
private:
BpRefBase(const BpRefBase& o);
BpRefBase& operator=(const BpRefBase& o);
IBinder* const mRemote; //注意是普通指针
RefBase::weakref_type* mRefs; // BpRefBase ‘s mRefs, not private mRefs of base class.
volatile int32_t mState;
};
继承关系图如下：
RefBase RefBase
/ \ / \
IBinder IInterface BpBinder<--BpRefBase IInterface
| | | |
BBinder ISubInterface | ISubInterface
\ / \ /
BnInterface
| |
BnSubInterface BpSubInterface
思考对象内存布局（注意virtual继承）。
两张更专业和漂亮的图如下（引用自网络）：
BnInterface和BpInterface模板类继承都采用多重继承，BnInterface
在设计上，为什么BpInterface
答案是BpBinder的生命周期与BpInterface
BpRefBase的两个字段mRemote和mRefs的作用：跨进程binder意义上的强和弱引用，角色就像wp时，对m_ptr和m_refs的作用。内存布局如下。
在由IInterface得到关联的Binder时，IInterface::asBinder() calls IInterface::onAsBinder()=0，显然BpInterface
In frameworks/base/include/binder/IInterface.h
template
IBinder* BnInterface
{
return this; // the BBinder itself.
}
template
inline IBinder* BpInterface
{
return remote(); // the mRemote whcih is a BpBinder encapsuling the binder handle.
}
对于BpInterface和BpBinder，结合对象内存布局可知转换是得到对应的对象，而不是同一个对象的upcast或downcast。
引入IInterface后，涉及的静态类结构图（引用自网络）：
跨进程方法调用流程图如下图（引用自网络）：
下面，从客户和服务同进程和不同进程两种情况来详细审视一下interface_cast Binder to Interface背后的逻辑。
首先，ISubInterface的类声明引用DECLARE_META_INTERFACE(INTERFACE)宏，在实现文件中要引用宏IMPLEMENT_META_INTERFACE(INTERFACE, NAME)实现定义。
#define DECLARE_META_INTERFACE(INTERFACE) \
static const android::String16 descriptor; \
static android::sp
const android::sp
virtual const android::String16& getInterfaceDescriptor() const; \
I##INTERFACE(); \
virtual ~I##INTERFACE(); \
#define IMPLEMENT_META_INTERFACE(INTERFACE, NAME) \
const android::String16 I##INTERFACE::descriptor(NAME); \
const android::String16& \
I##INTERFACE::getInterfaceDescriptor() const { \
return I##INTERFACE::descriptor; \
} \
android::sp
const android::sp
{ \
android::sp
if (obj != NULL) { \
intr = static_cast
obj->queryLocalInterface( \
I##INTERFACE::descriptor).get()); \
if (intr == NULL) { \
intr =new Bp##INTERFACE(obj); \
} \
} \
return intr; \
} \
I##INTERFACE::I##INTERFACE() { } \
I##INTERFACE::~I##INTERFACE() { } \
#define CHECK_INTERFACE(interface, data, reply) \
if (!data.checkInterface(this)) { return PERMISSION_DENIED; } \
INTERFACE倒把自己搞的像个MFC。这样ISubInterface就有静态成员descriptor，静态方法asInterface()和虚方法queryInterfaceDescriptor()，暂且不管queryInterfaceDescriptor()。
由interface_cast实现调到ISubInterface::asInterface(sp
BnInterface
template
inline sp
const String16& _descriptor)
{
if (_descriptor == INTERFACE::descriptor) return this;
return NULL;
}
从而，如果参数sp
预为熟悉，BpSubIinterface构造过程中将BpBinder指针赋值给BpRefBase的IBinder* mRemote 变量。incStrong会增加远端Binder的引用计数，通过transact(BC_ACQUIRE)通信完成。
template
inline BpInterface
下面介绍getInterfaceDescriptor()：
template
inline const String16& BnInterface
{
return INTERFACE::getInterfaceDescriptor();
}
IBinder和ISubInterface均有各自的getInterfaceDescriptor()虚方法，所以调用时，要加上namespace，如BnSunbInterface可以用this->BBinder::getInterfaceDescriptor()，this->ISubInterface::getInterfaceDescriptor()；而BpSubInterface可以直接用getInterfaceDescriptor()，等同于this->ISubInterface::getInterfaceDescriptor()，用使用BpBinder的getInterfaceDescriptor()可以使用this->remote()->getInterfaceDescriptor()。
CHECK_INTERFACE宏，用于接口检查，目前代码中用于BnSubInterface实现中。从Parcel中读取InterfaceDescriptor与自己this（downcast to IBinder *）->getInterfaceDescriptor()比较，如果失败，终止处理，返回。
下面接着考察BBinder和BpBinder对IBinder的重要的纯虚函数transact的实现，即LPC功能的实现：
status_t BpBinder::transact(uint32_t code, const Parcel& data, Parcel* reply, uint32_t flags)
{
// Once a binder has died, it will never come back to life.
if (mAlive) {
status_t status =IPCThreadState::self()->transact(mHandle, code, data, reply, flags);
if (status == DEAD_OBJECT) mAlive = 0;
return status;
}
return DEAD_OBJECT;
}
status_t BBinder::transact(uint32_t code, const Parcel& data, Parcel* reply, uint32_t flags)
{
data.setDataPosition(0);
status_t err = NO_ERROR;
switch (code) {
case PING_TRANSACTION:
reply->writeInt32(pingBinder());
break;
default:
err = onTransact(code, data, reply, flags); // virtual, overload it in derived class!
break;
}
if (reply != NULL) {
reply->setDataPosition(0);
}
return err;
}
flag 参数同步调用是0，异步调用时用IBinder.FLAG_ONEWAY。FLAG_ONEWAY是class IBinder的匿名enum，所以不是IBinder::FLAG_ONEWAY，对应驱动的FLAG_ONE_WAY。
最后，对于ISubInterfaceClient/BpSubInterfaceClient/BnSubInterfaceClient等带”Client”字样的类，一是用于service端的client管理，BnSubInterfaceClient实现由client端使用，BpSubInterfaceClient实现由service端使用；更重要的，是实现一种异步的event sink机制。
Android Binder机制之defaultServiceManager
在继续探讨binder通信之前，首先大致分析defaultServicemanager，会预引入binder driver里面一些结构及BC/BR。泛泛描述一节，目的是从servicemanager视角对binder通信数据包结构有所了解。
defaultServiceManager提供服务的添加、获取和枚举功能，先于system_init启动各种服务之前启动，因为system_init中启动的服务需要向defaultServiceManager进行注册。binder component分为命名binder和不具名binder，角色类似于socket中的知名监听端口和accept后的普通端口。命名binder向ServiceManager注册用于寻找服务。
在Phone和Emulator平台(TARGET_SIMULATOR=false)上和Simulator平台(TARGET_SIMULATOR=true)上，defaultServiceManager使用不同的实现。但是，注意，TARGET_SIMULATOR = true几乎已经绝迹!!!
Should TARGET_SIMULATOR be set to true when using QEMU?
no, the simulator is a special target forruning the dalivk vm on Linux and it is not maintained for a long time. It has nothing to do with qemu.
在手机和Qemu上运行时，编译选项TARGET_SIMULATOR = false，ServiceManager的启动在Init.rc中以如下服务启动：
service servicemanager/system/bin/servicemanager
class core
user system
group system
critical
onrestart restart zygote
onrestart restart media
onrestart restart surfaceflinger
onrestart restart drm
编译控制frameworks/base/cmds/servicemanager/Android.mk:
1LOCAL_PATH:= $(call my-dir)
2
3#include $(CLEAR_VARS)
4#LOCAL_SRC_FILES := bctest.c binder.c
5#LOCAL_MODULE := bctest
6#include $(BUILD_EXECUTABLE)
7
8include $(CLEAR_VARS)
9LOCAL_SHARED_LIBRARIES := liblog
10LOCAL_SRC_FILES := servicemanager.c binder.c
11LOCAL_MODULE := servicemanager
12ifeq ($(BOARD_USE_LVMX),true)
13 LOCAL_CFLAGS += -DLVMX
14endif
15include $(BUILD_EXECUTABLE)
作为servicemanager。servicemanager在service_manager.c文件中没有使用IServiceManager接口，直接按照binder和serviceemanager的通信协议使用c实现，启动方式如下：
int main(int argc, char **argv) // service_manager.c forservicemanager
{
struct binder_state *bs;
void *svcmgr = BINDER_SERVICE_MANAGER;
bs = binder_open(128*1024);
// 打开/dev/binder，并mmap
/* the only binder_become_context_manager call*/
if (binder_become_context_manager(bs))
// 给使用ioctl给binder driver发送BINDER_SET_CONTEXT_MGR命令，在驱动里面会处理
{
LOGE("cannot become context manager (%s)\n", strerror(errno));
return -1;
}
svcmgr_handle = svcmgr;
binder_loop(bs, svcmgr_handler);
// BC_ENTER_LOOPER通知driver本线程进入线程池；进入主服务循环，阻塞读在fd上
return 0;
}
ServiceManager的功能比较简单，阻塞读在binder fd上，接收其他进程的请求处理即可，因为只有GET_SERVICE_TRANSACTION、CHECK_SERVICE_TRANSACTION、ADD_SERVICE_TRANSACTION和LIST_SERVICES_TRANSACTION几个请求且处理很简单，所以根本不需要多线程并行处理之类的复杂模型。代码就在service_manager.c和binder.c两个文件中。在列出几个重要的数据结构后，再详细考察svcmgr_handler。
In frameworks/cmds/servicemanager/binder.h (except struct binder_write_read)
struct binder_object
{
uint32_t type;
uint32_t flags;
void *pointer;
void *cookie;
};
binder对象。
struct binder_write_read {
signed long write_size;
signed long write_consumed;
unsigned long write_buffer;
signed long read_size;
signed long read_consumed;
unsigned long read_buffer;
};
传给ioctl输入输出缓冲区参数，用于数据的IO。当传送给binder driver的IOCTL CODE是BINDER_WRITE_READ时，传送此结构的参数。可以简单看成IOCTL的IO数据缓冲区。
struct binder_txn
{
void *target;
void *cookie;
uint32_t code;
uint32_t flags;
uint32_t sender_pid;
uint32_t sender_euid;
uint32_t data_size;
uint32_t offs_size;
void *data;
void *offs;
};
同驱动中的binder_transaction_data结构，Binder route和权限控制，序列化数据buffer地址，是binder通信命令BC_TRANSANCTION和BC_REPLY的负载数据。
struct binder_io
{
char *data; /* pointer to read/write from */
uint32_t *offs; /* array of offsets */
uint32_t data_avail; /* bytes available in data buffer */
uint32_t offs_avail; /* entries available in offsets array */
char *data0; /* start of data buffer */
uint32_t *offs0; /* start of offsets buffer */
uint32_t flags;
uint32_t unused;
};
主要用作binder_txn的offs和data字段指向的缓冲区。是通过binder_txn /binder_transaction_data route到特定binder，给该binder的LPC方法ID和参数序列化数据，作用同Parcel类和驱动中的binder_buffer。
binder_io.offs/binder_txn.offs中记录的是跨进程通信中的struct flat_binder_object对象，驱动根据offs中的偏移量找到每个struct flat_binder_object对象在.data中的位置，对binder对象进行转换。如果target binder所在的进程是target process，那么binder.type会被驱动转成BINDER_TYPE_BINDER/BINDER_TYPE_WEAK_BINDER，并可用于BnSubInterface/BBinder的直接寻址；不同时（从服务进程传到客户端进程或从ServiceManager进程传到客户端进程），则binder.type会是BINDER_TYPE_HANDLE/BINDER_TYPE_WEAK_HANDLE。
struct binder_death {
void (*func)(struct binder_state *bs, void *ptr);
void *ptr;
};
/* the one magic object */
#define BINDER_SERVICE_MANAGER ((void*) 0)
#define SVC_MGR_NAME "android.os.IServiceManager"
enum {
SVC_MGR_GET_SERVICE = 1,
SVC_MGR_CHECK_SERVICE,
SVC_MGR_ADD_SERVICE,
SVC_MGR_LIST_SERVICES,
};
虽未多描述，但此处，希望完全记清楚这几个数据结构的层次关系；因附下图(x)。
int binder_write(struct binder_state *bs, void *data, unsigned len)
{
struct binder_write_read bwr;
int res;
bwr.write_size = len;
bwr.write_consumed = 0;
bwr.write_buffer = (unsigned) data;
bwr.read_size = 0;
bwr.read_consumed = 0;
bwr.read_buffer = 0;
res = ioctl(bs->fd, BINDER_WRITE_READ, &bwr);
if (res < 0) {
fprintf(stderr,"binder_write: ioctl failed (%s)\n",
strerror(errno));
}
return res;
}
int binder_read(struct binder_state *bs, void *data, unsigned len)
{
struct binder_write_read bwr;
int res;
bwr.write_size = 0;
bwr.write_consumed = 0;
bwr.write_buffer =0;
bwr.read_size = len;
bwr.read_consumed = 0;
bwr.read_buffer = (unsigned) data;
res = ioctl(bs->fd, BINDER_WRITE_READ, &bwr);
if (res < 0) {
fprintf(stderr,"binder_write: ioctl failed (%s)\n",
strerror(errno));
}
return bwr.read_consumed;
}
void binder_loop(struct binder_state *bs, binder_handler func)
{
int res; int len;
unsigned readbuf[32]; // unsigned int default.
readbuf[0] = BC_ENTER_LOOPER;
binder_write(bs, readbuf, sizeof(unsigned)); // Enter thread pool as main work thread
for (;;) {
res = binder_read(bs->fd, readbuf, sizeof(readbuf)); // block read the request
if (res < 0) {
LOGE("binder_loop: ioctl failed (%s)\n", strerror(errno));
break;
}
len = res;
res =binder_parse(bs, 0, readbuf, len, func); // parse request and handle, callback func if BR_TRANSACTION
if (res == 0) {
LOGE("binder_loop: unexpected reply?!\n");
break;
}
if (res < 0) {
LOGE("binder_loop: io error %d %s\n", res, strerror(errno));
break;
}
}
}
svcmgr_handler是ServiceManager实现自己功能的回调函数，角色同BBinder::onTransact()。对于BR_TRANSACTION，binder_parse会调用回调函数func进行处理。
int svcmgr_handler(struct binder_state *bs,
struct binder_txn *txn,
struct binder_io *msg,
struct binder_io *reply)
{
struct svcinfo *si;
uint16_t *s;
unsigned len;
void *ptr;
uint32_t strict_policy;
// LOGI("target=%p code=%d pid=%d uid=%d\n",
// txn->target, txn->code, txn->sender_pid, txn->sender_euid);
if (txn->target != svcmgr_handle)
return -1;
// Equivalent to Parcel::enforceInterface(), reading the RPC
// header with the strict mode policy mask and the interface name.
// Note that we ignore the strict_policy and don't propagate it
// further (since we do no outbound RPCs anyway).
strict_policy = bio_get_uint32(msg);
s = bio_get_string16(msg, &len);
if ((len != (sizeof(svcmgr_id) / 2)) ||
memcmp(svcmgr_id, s, sizeof(svcmgr_id))) {
fprintf(stderr,"invalid id %s\n", str8(s));
return -1;
}
switch(txn->code) {
case SVC_MGR_GET_SERVICE:
case SVC_MGR_CHECK_SERVICE:
s = bio_get_string16(msg, &len);
ptr = do_find_service(bs, s, len);
if (!ptr)
break;
bio_put_ref(reply, ptr);
return 0;
case SVC_MGR_ADD_SERVICE:
s = bio_get_string16(msg, &len);
ptr = bio_get_ref(msg);
if (do_add_service(bs, s, len, ptr, txn->sender_euid))
return -1;
break;
case SVC_MGR_LIST_SERVICES: {
unsigned n = bio_get_uint32(msg);
si = svclist;
while ((n-- > 0) && si)
si = si->next;
if (si) {
bio_put_string16(reply, si->name);
return 0;
}
return -1;
}
default:
LOGE("unknown code %d\n", txn->code);
return -1;
}
bio_put_uint32(reply, 0);
return 0;
}
可以看出，回调完成SVC_MGR_[GET/ CHECK/ ADD/ LIST] _SERVICE的处理，就是用个链表的结点插入、查找、删除操作。
对于SVC_MGR_ADD_SERVICE，ServiceManager最终会使用下列函数处理，注意binder_io中的binder_object已经是被binder driver转换过的binder ref，是handle形式。
int do_add_service(struct binder_state *bs,
uint16_t *s, unsigned len,
void *ptr, unsigned uid)
{
struct svcinfo *si;
// LOGI("add_service('%s',%p) uid=%d\n", str8(s), ptr, uid);
if (!ptr || (len == 0) || (len > 127))
return -1;
if (!svc_can_register(uid, s)) {
LOGE("add_service('%s',%p) uid=%d - PERMISSION DENIED\n",
str8(s), ptr, uid);
return -1;
}
si = find_svc(s, len);
if (si) {
if (si->ptr) {
LOGE("add_service('%s',%p) uid=%d - ALREADY REGISTERED, OVERRIDE\n",
str8(s), ptr, uid);
svcinfo_death(bs, si);
}
si->ptr = ptr;
} else {
si = malloc(sizeof(*si) + (len + 1) * sizeof(uint16_t));
if (!si) {
LOGE("add_service('%s',%p) uid=%d - OUT OF MEMORY\n",
str8(s), ptr, uid);
return -1;
}
si->ptr = ptr;
si->len = len;
memcpy(si->name, s, (len + 1) * sizeof(uint16_t));
si->name[len] = '\0';
si->death.func = svcinfo_death;
si->death.ptr = si;
si->next = svclist;
svclist = si;
}
binder_acquire(bs, ptr);
binder_link_to_death(bs, ptr, &si->death);
return 0;
}
service_manager会立刻发送BC_ACQUIRE命令给binder driver增加对service binder_node的binder_ref的强计数(binder_ref.strong)；而没有发送BC_INCREFS增加binder_ref.weak；这是与service的普通客户端不同的地方。从servicemanger的debugfs/binder/proc中的binder_ref可以看到这种差异。具体后面协议分析时详述。
注意service实现端调用sm->addService(name, new BnSubInterface)返回前会收到BR_INCREFS和BR_ACQUIRE增加强弱计数（BR_ACQUIRE不是因为此处的BC_ACQUIRE而产生），从而将组件钳住，避免临时对象sp
对于工作线程返回给请求者不具名binder node时，同样有此需求和逻辑。
问题[超前]：调用sm->addService(name, new BnSubInterface)时，通常用户态main线程还没有joiThreadPool，Service所在进程的worker thread还没有spawn出来，内核如何调度在用户态实现BR_INCREFS和BR_ACQUIRE的处理？答案：在driver中需要进行BINDER_LOOPER_STATE_ENTERED判断的是binder工作线程。此时虽然是在Service进程中，但是是作为对sm的请求者线程，而请求者线程是不需要进行判断BINDER_LOOPER_STATE_ENTERED的，所以可以实现在addService返回前增加组件计数。问题中是混淆了请求者线程和工作线程处理BR_INCREFS和BR_ACQUIRE的情景。详细后述。
然后BC_REQUEST_DEATH_NOTIFICATION注册binder node death notification。
BR_TRANSACTION调用func后，会BC_REPLY，函数如下（注意同时发送了BC_FREE_BUFFER）：
void binder_send_reply(struct binder_state *bs,
struct binder_io *reply,
void *buffer_to_free,
int status)
{
struct {
uint32_t cmd_free;
void *buffer;
uint32_t cmd_reply;
struct binder_txn txn;
} __attribute__((packed)) data;
data.cmd_free = BC_FREE_BUFFER;
data.buffer = buffer_to_free;
data.cmd_reply =BC_REPLY;
data.txn.target = 0;
data.txn.cookie = 0;
data.txn.code = 0;
if (status) {
data.txn.flags = TF_STATUS_CODE;
data.txn.data_size = sizeof(int);
data.txn.offs_size = 0;
data.txn.data = &status;
data.txn.offs = 0;
} else {
data.txn.flags = 0;
data.txn.data_size = reply->data - reply->data0;
data.txn.offs_size = ((char*) reply->offs) - ((char*) reply->offs0);
data.txn.data = reply->data0;
data.txn.offs = reply->offs0;
}
binder_write(bs, &data, sizeof(data));
}
在向ServiceManager中注册服务时，会进行注册进程权限的检查，函数是svc_can_register，可以看出只有来自0号进程（init）、system_server、在allowed数组中项（AID_MEDIA, AID_RADIO, etc）才能注册。
int svc_can_register(unsigned uid, uint16_t *name)
{
unsigned n;
if ((uid == 0) || (uid == AID_SYSTEM))
return 1;
for (n = 0; n < sizeof(allowed) / sizeof(allowed[0]); n++)
if ((uid == allowed[n].uid) && str16eq(name, allowed[n].name))
return 1;
return 0;
}
static struct {
unsigned uid;
const char *name;
} allowed[] = {
#ifdef LVMX
{ AID_MEDIA, "com.lifevibes.mx.ipc" },
#endif
{ AID_MEDIA, "media.audio_flinger" },
{ AID_MEDIA, "media.player" },
{ AID_MEDIA, "media.camera" },
{ AID_MEDIA, "media.audio_policy" },
{ AID_DRM, "drm.drmManager" },
{ AID_NFC, "nfc" },
{ AID_RADIO, "radio.phone" },
{ AID_RADIO, "radio.sms" },
{ AID_RADIO, "radio.phonesubinfo" },
{ AID_RADIO, "radio.simphonebook" },
/* TODO: remove after phone services are updated: */
{ AID_RADIO, "phone" },
{ AID_RADIO, "sip" },
{ AID_RADIO, "isms" },
{ AID_RADIO, "iphonesubinfo" },
{ AID_RADIO, "simphonebook" },
{ AID_RADIO, "phone_msim" },
{ AID_RADIO, "simphonebook_msim" },
{ AID_RADIO, "iphonesubinfo_msim" },
{ AID_RADIO, "isms_msim" },
};
system_server中的进程可以参阅《Android的系统服务一览》
http://blog.csdn.net/freshui/article/details/5993195
更详细地，要知道init和system_server会注册哪些service，可以参阅Android启动过程。
在SIMULATOR上，编译选项TARGET_SIMULATOR = true时，runtime service。在init.rc中如下：
service runtime /system/bin/runtime
user system
group system
源码在frameworks/base/cmds/runtime/目录下，编译出模块runtime。在该服务中进行一些设置，启动其中内置实现的servicemanager，然后启动Zygote，启动system_server。在硬件平台和Qemu上，这些是分开的，在frameworks/base/cmds/目录下分别启动，如system_server是在system_server/目录下。
defaultServiceManager使用和普通service相同的实现方式，继承关系BServiceManager->BnServiceManager->BnInterface
static void boot_init() ------>main() ---Main_Runtime.cpp runtime
{
LOGI("Entered boot_init()!\n");
sp
// proc(ProcessState::self())单例，其构造函数会打开/dev/binder设备，
// 并使用mmap将binder fd映射到一段内存空间，就是用于获取一段用户态的内存空间。
LOGD("ProcessState: %p\n", proc.get());
proc->becomeContextManager(contextChecker, NULL);
// 直接调用ioctl(mDriverFD, BINDER_SET_CONTEXT_MGR, &dummy)，
// 通过BINDER_SET_CONTEXT_MGR命令，driver会建立一个binder_node，
// 并作为CONTEXT_MGR。
if (proc->supportsProcesses()) {
LOGI("Binder driver opened. Multiprocess enabled.\n"); //taken
} else {
LOGI("Binder driver not found. Processes not supported.\n");
}
sp
// BServiceManager继承自BnServiceManager，真正实现Manager功能。
// 通过维护一张服务向量表，实现getService/checkService/addService/listService操作。
// BnServiceManager的onTransact将GET_SERVICE_TRANSACTION/
// CHECK_SERVICE_TRANSACTION、ADD_SERVICE_TRANSACTION和
// LIST_SERVICES_TRANSACTION分别用上述四个函数处理。
proc->setContextObject(sm);
// 将sm用” default”做名字首先将其注册到KeyedVector向量表中
}
boot_init()之后，
static void finish_system_init(sp
{
// If we are running multiprocess, we now need to have the
// thread pool started here. We don't do this in boot_init()
// because when running single process we need to start the
// thread pool after the Android runtime has been started (so
// the pool uses Dalvik threads).
if (proc->supportsProcesses()) {
proc->startThreadPool(); //--> spawnPooledThread()新spawn一个用于处理服务的线程
}
}
SIMULATOR用的很少，通讯协议相同，不再叙述。代码使用BServiceManager : BnServiceManager和ProcessState，需要可以查阅。
Android Binder机制之IPCThreadState & Parcel
本小节描述应用层远端方式下，client端的transant实现和service的onTransact如何被调用到；同时描述binder在Framework层的线程模型。试图描述大致结构及设计理念，不涉及协议处理细节；但仍然将重要代码列出。
这个实现主要涉及三个重要的类，ProcessState、IPCThreadState和Parcel，其中ProcessState主要用于打开/dev/binder，进行mmap，和维护进程状态。我们主要讨论后面的两个，IPCThreadState用于实现binder通信的线程模型和Binder通信协议，Parcel则负责数据的序列化（serialization）。
数据序列化
Parcel是通信数据的封装类，和servicemanger中的binder_io角色对等。由于binder对象需要在内核中进行管理，每进程内的binder需要一个唯一个key。详细的需求是对于BINDER_TYPE_BINDER，binder_driver中会有一个binder_node结构与之对应，对于BINDER_TYPE_HANDLE，binder driver中会有一个binder_ref与之对应。Binder_node需要使用一个唯一的key挂接到内核中binder_proc维护的一棵rb_tree中，flatten就是要使BINDER_TYPE_BINDER/BINDER_TYPE_HANDLE时，pointer字段唯一。这是通过在用户空间中使用struct flat_binder_object结构，完成每进程空间中binder comp的唯一标记。
Parcel.cpp中的几个辅助函数列出如下。
status_t flatten_binder(const sp
{
flat_binder_object obj;
obj.flags = 0x7f | FLAT_BINDER_FLAG_ACCEPTS_FDS;
if (binder != NULL) {
IBinder *local = binder->localBinder();
if (!local) {
BpBinder *proxy = binder->remoteBinder();
if (proxy == NULL) {
LOGE("null proxy");
}
const int32_t handle = proxy ? proxy->handle() : 0;
obj.type = BINDER_TYPE_HANDLE;
obj.handle = handle;
obj.cookie = NULL;
} else {
obj.type = BINDER_TYPE_BINDER;
obj.binder = local->getWeakRefs();
obj.cookie = local;
}
} else {
obj.type = BINDER_TYPE_BINDER;
obj.binder = NULL;
obj.cookie = NULL;
}
return finish_flatten_binder(binder, obj, out);
}
status_t flatten_binder(const sp
{
flat_binder_object obj;
obj.flags = 0x7f | FLAT_BINDER_FLAG_ACCEPTS_FDS;
if (binder != NULL) {
sp
if (real != NULL) {
IBinder *local = real->localBinder();
if (!local) {
BpBinder *proxy = real->remoteBinder();
if (proxy == NULL) {
LOGE("null proxy");
}
const int32_t handle = proxy ? proxy->handle() : 0;
obj.type = BINDER_TYPE_WEAK_HANDLE;
obj.handle = handle;
obj.cookie = NULL;
} else {
obj.type = BINDER_TYPE_WEAK_BINDER;
obj.binder = binder.get_refs();
obj.cookie = binder.unsafe_get();
}
return finish_flatten_binder(real, obj, out);
}
// XXX How to deal? In order to flatten the given binder,
// we need to probe it for information, which requires a primary
// reference... but we don't have one.
//
// The OpenBinder implementation uses a dynamic_cast<> here,
// but we can't do that with the different reference counting
// implementation we are using.
LOGE("Unable to unflatten Binder weak reference!");
obj.type = BINDER_TYPE_BINDER;
obj.binder = NULL;
obj.cookie = NULL;
return finish_flatten_binder(NULL, obj, out);
} else {
obj.type = BINDER_TYPE_BINDER;
obj.binder = NULL;
obj.cookie = NULL;
return finish_flatten_binder(NULL, obj, out);
}
}
status_t unflatten_binder(const sp
{
const flat_binder_object* flat = in.readObject(false);
if (flat) {
switch (flat->type) {
case BINDER_TYPE_BINDER:
*out = static_cast
return finish_unflatten_binder(NULL, *flat, in);
case BINDER_TYPE_HANDLE:
*out = proc->getStrongProxyForHandle(flat->handle);
return finish_unflatten_binder(
static_cast
}
}
return BAD_TYPE;
}
status_t unflatten_binder(const sp
{
const flat_binder_object* flat = in.readObject(false);
if (flat) {
switch (flat->type) {
case BINDER_TYPE_BINDER:
*out = static_cast
return finish_unflatten_binder(NULL, *flat, in);
case BINDER_TYPE_WEAK_BINDER:
if (flat->binder != NULL) {
out->set_object_and_refs(
static_cast
static_cast
} else {
*out = NULL;
}
return finish_unflatten_binder(NULL, *flat, in);
case BINDER_TYPE_HANDLE:
case BINDER_TYPE_WEAK_HANDLE:
*out = proc->getWeakProxyForHandle(flat->handle);
return finish_unflatten_binder(
static_cast
}
}
return BAD_TYPE;
}
void acquire_object(const sp
{
switch (obj.type) {
case BINDER_TYPE_BINDER:
if (obj.binder) {
LOG_REFS("Parcel %p acquiring reference on local %p", who, obj.cookie);
static_cast
}
return;
case BINDER_TYPE_WEAK_BINDER:
if (obj.binder)
static_cast
return;
case BINDER_TYPE_HANDLE: {
const sp
if (b != NULL) {
LOG_REFS("Parcel %p acquiring reference on remote %p", who, b.get());
b->incStrong(who);
}
return;
}
case BINDER_TYPE_WEAK_HANDLE: {
const wp
if (b != NULL) b.get_refs()->incWeak(who);
return;
}
case BINDER_TYPE_FD: {
// intentionally blank -- nothing to do to acquire this, but we do
// recognize it as a legitimate object type.
return;
}
}
LOGD("Invalid object type 0x%08lx", obj.type);
}
void release_object(const sp
{
switch (obj.type) {
case BINDER_TYPE_BINDER:
if (obj.binder) {
LOG_REFS("Parcel %p releasing reference on local %p", who, obj.cookie);
static_cast
}
return;
case BINDER_TYPE_WEAK_BINDER:
if (obj.binder)
static_cast
return;
case BINDER_TYPE_HANDLE: {
const sp
if (b != NULL) {
LOG_REFS("Parcel %p releasing reference on remote %p", who, b.get());
b->decStrong(who);
}
return;
}
case BINDER_TYPE_WEAK_HANDLE: {
const wp
if (b != NULL) b.get_refs()->decWeak(who);
return;
}
case BINDER_TYPE_FD: {
if (obj.cookie != (void*)0) close(obj.handle);
return;
}
}
LOGE("Invalid object type 0x%08lx", obj.type);
}
struct flat_binder_object结构如下：
struct flat_binder_object{
unsigned long type;
unsigned long flags;
union {
void *binder; /* local object */
signed long handle; /* remote object */
};
void *cookie; /* address of bbinder if local binder*/
};
| .type\ |
TYPE_BINDER |
TYPE_HANDLE |
TYPE_WEAK_BINDER |
TYPE_WEAK_HANDLE |
| union {binder; handle}.pointer |
mRefs |
handle |
mRefs |
handle |
| .cookie |
BBinder |
NULL |
BBinder |
NULL |
binder.type为BINDER_TYPE_BINDER/ BINDER_TYPE_WEAK_BINDER时，维护mRefs和BBinder两个字段指针的用途是显然的，回看RefBase和RefBase::weakref_impl对象在内存中关系。
线程模型
下表格图是service端的线程模型。
| interface |
| binder component pool |
| binder worker thread pool |
| binder driver |
线程池和对象池
因为所有binder对象包括BBinder和BpBinder都是可路由的，所以binder worker thread并不与特定binder component绑定，而是寻找到目标binder，执行其代码逻辑；执行transaction时，根据接口函数代码派送execution到特定目标函数。
线程池可以设定最少和最大工作线程数，可以主动或接受dirvier指示spawn出的最大线程数默认是0x10，这些线程的名字是Binder Thread #num(ICS)或Binder_num(JB)。为了不拖延程序启动速度，创建线程池时，只创建一个isMain线程；然后main线程如果没有更多工作，也会加入线程池，main主线程不占用0x10这个数目。当工作线程不够用时，binder drvier会指示用户层创建non-isMain工作线程并加入到线程池中。用户态主动创建的线程，其属性isMian为true，通过BC_ENTER_LOOPER通知driver进入工作循环；而根据driver指示spawn出来的线程，其isMain属性为false，通过BC_REGISTER_LOOPER通知driver进入工作循环。当non-isMian工作线程长时间无事可做时，可以退出，但是android binder并没有完全实现这一点；isMain线程不退出。
工作线程优先级模型
binder driver会根据请求者线程的优先级设置工作线程的优先级，高优先级的请求者可以以高优先级调度，保证用户体验。binder node可以设置执行自己逻辑的工作线程的最低优先级，binder driver也会根据这个优先级设置nice运行其逻辑的工作线程的优先级。当工作线程服务完毕后，恢复其默认优先级，一般是其主线程的优先级。
此处加FOREGROUD BKGROUD和线程优先级部分。
IO效率
线程通过系统调用ioctl完成数据IO，考虑到频繁的系统调用的效率开销，会将多个命令打包进行一个IO，通常除transaction外，其余命令都不是紧急的。talkWithDriver还通过其它技巧尽量减少IO次数，降低系统调用开销，提高效率，具体详见代码。
同步异步模型
支持一发一收的双向同步方式，发送-阻塞-接收；也支持只发送不接收的ONE_WAY方式，发送后非阻塞返回，可以使用event sink通知机制。
waitForResponse和executeCommand
executeCommand处理的命令集合S = { BR_ERROR/BR_OK、BR_INCREFS/BR_ACQUIRE/BR_DECREFS/BR_RELEASE、BR_TRANSACTION、BR_... }(注意不包含BR_REPLY)。因为工作线程joinThreadPool后，阻塞读等待命令，所以talkWithDriver返回后，接收到的一定是集合S中的命令，因此只需要调用executeCommand处理。
请求者线程发送命令和工作线程答复命令时，数据写后都需要检查数据写入结果，即在waitForResponse中检查明确列出的BR_CMDs，命令集合W= { BR_DEAD_REPLY, BR_FAILED_REPLY, BR_ACQUIRE_RESULT, BR_REPLY}。
BR_NOOP，可能用来做标记作用，区分什么东西；或者将来做hook用？
对于Transaction请求者线程等待的回应包括两部分内容，一部分是transaction_write是否成功（此处描述先这么意会吧），一部分是BR_REPLY数据包；但是有两种情况会更复杂一点：一是调用ServiceManager::addService注册binder node和向服务端传送event sink binder comp时，需要waitForResponse->executeCommand(BR_INCREFS/BR_ACQUIRE)；二是交叉式transaction时，需要waitForResponse->executeCommand(BR_TRANSACTION)。
而对于工作线程，当向请求者返回不具名binder comp时，显然也需要executeCommand(BR_INCREFS/BR_ACQUIRE)；而交叉式transaction时，也需要waitForResponse{ handle BR_REPLY }。BR_TRANSACTION和BR_REPLY普通情况下前者是在binder工作线程W1中，后者在请求者线程T1中。但交叉式嵌套binder调用时，则线程角色互换，即BR_TRANSACTION在T1中，BR_REPLY在W1中。当然这需要binder drvier的鼎力支持，找到正确的线程并唤醒，然后才是用户态transaction处理的嵌套。
针对这些情形，所以源码实现中有executeCommand和waitForResponse嵌套调用的地方。具体情形具体流程，参考源码。
重要函数的简单标注
本节为初始版本文档中的描述，未作增补。
IPCThreadState* IPCThreadState::self()
每线程对应的线程对象表达，由static代码寻找实体对象。实现方法：若是线程组中第一次调用此函数，在TLS中申请一个slot。每个线程将IPCThreadState* this保存其中。以后每次self()，取出返回即可。
void IPCThreadState::joinThreadPool(bool isMain)
工作线程使用该函数加入到线程池。进入循环后，首先处理延迟（deferred）的减引用计数，为什么有延迟减计数？这是因为在executeCommand中处理减计数命令时，并没有立刻进行减计数，而是延迟操作。
然后主体部分是talkWithDriver，若接收到数据，则调用executeCommand进行处理。
void setTheContextObject(sp
调用本函数成为全系统唯一的ServiceManager；the_context_object全局变量。但系统的ServiceManager却不是调用本函数设置的，而是调用ProcessState::setContextObject(const sp
进入run后会进入joinThreadLoop循环，如果执行完一次循环后
// Let this thread exit the thread pool if it is no longer
// needed and it is not the main process thread.
if(result == TIMED_OUT && !isMain)
break;
那么，循环退出，执行到proc->setContextObject(NULL);，需要设置新的ServiceManager。但是除非TIMED_OUT，否则这种情况不会发生，因为run()中IPCThreadState::self()->joinThreadPool();而IPCThreadState::self()->joinThreadPool(isMain=true);默认参数isMain为true。
status_t IPCThreadState::talkWithDriver(bool doReceive)
调用ioctl进行数据收发，当参数doReceive为true，且输入缓冲区中数据不够时，则读数据，并且捎带要发送的数据；参数doReceive为true，且输入缓冲区中数据足够时，不读，也不发送，直接返回；若是需要立刻发送数据，使用doReceive=false参数，如flushCommands即是如此。
status_t IPCThreadState::waitForResponse(Parcel *reply, status_t *acquireResult)
在transaction写入数据后调用本函数等待应答。Initiator在writeTransactionData后调用时，对于binder driver发出的响应要求直接进行处理；对于作为responser接收到initiator发来的命令时，转executeCommand进行处理。注意其可被executeCommand嵌套。
status_t IPCThreadState::executeCommand(int32_t cmd)
处理initiator发送的BC_TRANSACTION经过binder driver转换投递的BR_TRANSACTION命令和binder driver发出的BR_XXXXX命令，其命令集合如前所述。对于BR_TRANSACTION会调用BBinder::transact()，从而会调用到BBinder::onTransact()。
void IPCThreadState::freeBuffer(……)
用于释放mmap中的内核写入数据用户态读取数据的缓冲区(binder_transaction_data)，在BR_TRANSACTION和BR_REPLY命令处理中使用完该缓冲区后释放，或者设置freeBuffer为Parcel的mOwner回调（通过Parcel::ipcSetDataReference()将数据写入Parcel时设置），在Parcel清理数据时自动调用，从而保持Parcel中数据和mmap缓冲区中数据同步。还有一个重要功能，内核中释放缓冲区时，做binder的减计数。
status_t IPCThreadState::writeTransactionData(int32_t cmd, ……)
向对端发送数据，是交互的initiator时，使用BC_TRANSACTION cmd，是交互的responser回答（sendReply()）initiator时，使用BC_REPLY cmd。writeTransactionData仅仅是将要发送的数据写入了Parcel mOut输出缓冲区中，等下一轮talkWithDriver时发送，
对于控制生命周期的计数命令和无应答的单向命令（TF_ONE_WAY）并不需要立刻发送。对于需要应答的命令，则是writeTransactionData之后，立刻使用waitForResponse调用talkWithDriver()，由于需要接收，所以数据会立刻发送出去，并且挂起等待对端应答。
Client & Server通信的大致框图如下：
附代码带注释
void IPCThreadState::joinThreadPool(bool isMain) // PoolThread的主体joinThreadPool
{
LOG_THREADPOOL("**** THREAD %p (PID %d) IS JOINING THE THREAD POOL\n", (void*)pthread_self(), getpid());
mOut.writeInt32(isMain ? BC_ENTER_LOOPER : BC_REGISTER_LOOPER);
// 对于main pool thread和spawn pool thread使用不同命令通知driver。
// This thread may have been spawned by a thread that was in the background
// scheduling group, so first we will make sure it is in the default/foreground
// one to avoid performing an initial transaction in the background.
androidSetThreadSchedulingGroup(mMyThreadId, ANDROID_TGROUP_DEFAULT);
// 改变线程优先级，优先响应用户操作，用户体验为主
status_t result;
do {
int32_t cmd;
//When we've cleared the incoming command queue, process any pending derefs
if (mIn.dataPosition() >= mIn.dataSize()) {
size_t numPending = mPendingWeakDerefs.size();
if (numPending > 0) {
for (size_t i = 0; i < numPending; i++) {
RefBase::weakref_type* refs = mPendingWeakDerefs[i];
refs->decWeak(mProcess.get());
}
mPendingWeakDerefs.clear();
}
numPending = mPendingStrongDerefs.size();
if (numPending > 0) {
for (size_t i = 0; i < numPending; i++) {
BBinder* obj = mPendingStrongDerefs[i];
obj->decStrong(mProcess.get());
}
mPendingStrongDerefs.clear();
}
}
// now get the next command to be processed, waiting if necessary
result = talkWithDriver();
if (result >= NO_ERROR) {
size_t IN = mIn.dataAvail();
if (IN < sizeof(int32_t)) continue;
cmd = mIn.readInt32();
IF_LOG_COMMANDS() {
alog << "Processing top-level Command: "
<< getReturnString(cmd) << endl;
}
result = executeCommand(cmd);
}
// After executing the command, ensure that the thread is returned to the
// default cgroup before rejoining the pool. The driver takes care of
// restoring the priority, but doesn't do anything with cgroups so we
// need to take care of that here in userspace. Note that we do make
// sure to go in the foreground after executing a transaction, but
// there are other callbacks into user code that could have changed
// our group so we want to make absolutely sure it is put back.
androidSetThreadSchedulingGroup(mMyThreadId, ANDROID_TGROUP_DEFAULT);
// Let this thread exit the thread pool if it is no longer
// needed and it is not the main process thread.
if(result == TIMED_OUT && !isMain) {
break;
}
} while (result != -ECONNREFUSED && result != -EBADF);
LOG_THREADPOOL("**** THREAD %p (PID %d) IS LEAVING THE THREAD POOL err=%p\n",
(void*)pthread_self(), getpid(), (void*)result);
mOut.writeInt32(BC_EXIT_LOOPER);
talkWithDriver(false);
}
status_t IPCThreadState::talkWithDriver(bool doReceive = true) //IO，参数指定是IO方向。
{
LOG_ASSERT(mProcess->mDriverFD >= 0, "Binder driver is not opened");
binder_write_read bwr;
// Is the read buffer empty?
const bool needRead = mIn.dataPosition() >= mIn.dataSize();
// 如果当前读指针已超出缓冲区中数据，则读，提高效率。正常情况下，不会出现命令码和半个
// 包的情况，应该是整的命令数据包，足够一个命令处理循环。
// We don't want to write anything if we are still reading
// from data left in the input buffer and the caller
// has requested to read the next data.
const size_t outAvail = (!doReceive || needRead) ? mOut.dataSize() : 0;
// 发送数据，两个OR条件：
// 1 doReceive为false，说明是写操作
// 2 needRead需要读数据时，向外写命令，可以认为是延迟写稍带写，减少ioctl系统调用，
// 提高效率；但是反过来说，读响应，必然需要将以前的命令全部发送出去
bwr.write_size = outAvail;
bwr.write_buffer = (long unsigned int)mOut.data();
// This is what we'll read.
// 指示读操作，则结合是否真有必要读取
if (doReceive && needRead) {
bwr.read_size = mIn.dataCapacity();
bwr.read_buffer = (long unsigned int)mIn.data();
} else {
bwr.read_size = 0;
bwr.read_buffer = 0;
}
IF_LOG_COMMANDS() {
TextOutput::Bundle _b(alog);
if (outAvail != 0) {
alog << "Sending commands to driver: " << indent;
const void* cmds = (const void*)bwr.write_buffer;
const void* end = ((const uint8_t*)cmds)+bwr.write_size;
alog << HexDump(cmds, bwr.write_size) << endl;
while (cmds < end) cmds = printCommand(alog, cmds);
alog << dedent;
}
alog << "Size of receive buffer: " << bwr.read_size
<< ", needRead: " << needRead << ", doReceive: " << doReceive << endl;
}
// Return immediately if there is nothing to do.
if ((bwr.write_size == 0) && (bwr.read_size == 0)) return NO_ERROR;
bwr.write_consumed = 0;
bwr.read_consumed = 0;
status_t err;
do {
IF_LOG_COMMANDS() {
alog << "About to read/write, write size = " << mOut.dataSize() << endl;
}
#if defined(HAVE_ANDROID_OS)
if (ioctl(mProcess->mDriverFD, BINDER_WRITE_READ, &bwr) >= 0)
err = NO_ERROR;
else
err = -errno;
#else
err = INVALID_OPERATION;
#endif
IF_LOG_COMMANDS() {
alog << "Finished read/write, write size = " << mOut.dataSize() << endl;
}
} while (err == -EINTR);
IF_LOG_COMMANDS() {
alog << "Our err: " << (void*)err << ", write consumed: "
<< bwr.write_consumed << " (of " << mOut.dataSize()
<< "), read consumed: " << bwr.read_consumed << endl;
}
if (err >= NO_ERROR) {
if (bwr.write_consumed > 0) {
if (bwr.write_consumed < (ssize_t)mOut.dataSize())
mOut.remove(0, bwr.write_consumed);
else
mOut.setDataSize(0);
}
if (bwr.read_consumed > 0) {
mIn.setDataSize(bwr.read_consumed);
mIn.setDataPosition(0);
}
IF_LOG_COMMANDS() {
TextOutput::Bundle _b(alog);
alog << "Remaining data size: " << mOut.dataSize() << endl;
alog << "Received commands from driver: " << indent;
const void* cmds = mIn.data();
const void* end = mIn.data() + mIn.dataSize();
alog << HexDump(cmds, mIn.dataSize()) << endl;
while (cmds < end) cmds = printReturnCommand(alog, cmds);
alog << dedent;
}
return NO_ERROR;
}
return err;
}
Android Binder机制之binder driver
因为binder driver提供了binder通信的机制，所以在对Binder Process Framework有了初步熟悉之后分析binder driver。Binder driver是个知识密集型的driver，涉及内存分配、进程唤醒、文件系统、debugfs、工作队列等。
先描述几个重要的数据结构表达的对象，并列出其定义。注意，结构定义仍是cupecake版本的，注释是当时学习binder driver时做的，数据结构可能与ICS以后版本有所差异，不再做对比，仅对注释根据理解做少量更正。
binder_node:
用于描述的BBinder子类实例对应的内核对象，binder_node只存在用户态BBinder所在进程对应的内核对象binder_proc中，其余的进程都使用下面的binder_ref做引用。每进程拥有的所有binder_node都挂载到该进程对应的binder_proc的nodes rb_tree(binder.pointer as key)上。含有强弱引用计数。
内嵌的binder_work用于binder_node计数控制时，链入binder_work todo list。
binder_node活着时，挂到其binder_proc的rb_tree上，死掉后就从树上摘下来，放到hlist上。
TODO：加下列重要字段的说明
internal_strong_refs表示有多少binder_ref引用到该binder_node并且实际增加了binder_node的强计数；没有internal_weak_refs这个字段，直接从binder_ref的个数得到。
binder_node必然首先是在传输缓冲区中检测到，在驱动中为目标进程新建binder_ref。
local_strong_refs / has_strong_refs / pending_strong_refs;
local_weak_refs / has_weak_refs / pending_weak_refs;
async_todo has_async_transaction;异步transaction使用。
/**
** component. binder node agent in ownner thread, ownner proc.
** use mountpoint/entry to express the node of list, hlist, tree, etc.
**/
struct binder_node {
int debug_id;
struct binder_work work; /* work entry used to implenment reference count */
union {
struct rb_node rb_node; /* mount to rb_tree of binder_proc */
struct hlist_node dead_node; /* mount to hlist_nodd if binder component is dead */
};
struct binder_proc *proc; /* user space ownner process agent. */
struct hlist_head refs; /* all the binder_refs refering the binder_node */
int internal_strong_refs; /* 'driver internally' reference count. reference count of this node refered by binder_refs; always inc when the 'binder_node' is in binder_buffer at first time. */
int local_weak_refs; /* 'system locally' weak reference count, inc/dec when the binder_node is target node or binder_node is in binder_buffer */
int local_strong_refs; /* 'system locally' strong reference count, inc/dec when the binder_node is target node or binder_node is in binder_buffer */
void __user *ptr; /* unique id in user process, it is the address of mRefs */
void __user *cookie; /* unique id in user process, it is the address of BBinder */
unsigned has_strong_ref : 1; /* used as a bool cooperating with local_strong_refs to indicate first strong reference inc or last strong reference dec */
unsigned pending_strong_ref : 1;/* if set, it indicates that BR_ACQUIRE is send to userspace to do BBinder::incStrong, but userspace has not responsed BC_ACQUIRED_DONE. When userspace sent back BC_ACQUIRED_DONE, clear the bit and dec local_strong_refs. The bit is used for foreincreasing local_strong_refs to protect binder_node from being released. */
unsigned has_weak_ref : 1; /* used as a bool cooperating with local_weak_refs to indicate first weak reference inc or last weak reference dec */
unsigned pending_weak_ref : 1; /* if set, it indicates that BR_INCREFS is send to userspace to do BBinder::incWeak, but userspace has not responsed BC_INCREFS_DONE. When userspace sent back BC_INCREFS_DONE, clear the bit and dec local_weak_refs. The bit is used for foreincreasing local_weak_refs to protect binder_node from being released. */
unsigned has_async_transaction : 1; /* if has asynchronized transaction, ONE_WAY */
unsigned accept_fds : 1; /* if accepte fd, accept mostly */
int min_priority : 8; /* bit filed – part of flags. nice of thread to access this node */
struct list_head async_todo; /* async task */
};
binder_ref:
表示一个binder_proc内到另外进程binder_proc的binder_node的引用，包含binder_node的指针，含有本身强弱计数。
注意结构开头的注释，binder_ref查找情景和需要的元素，分别对应debug_id后面三个字段的用途。为了提高不同情景的查找或使用效率，使用不同的数据结构和key。根据desc/handle查找binder_ref时，使用desc-keyed rb_tree，根据给定的binder_node查找对应的本进程的binder_ref时，使用binder_node keyed rb_tree，而node_entry则是用来链入该binder_ref引用的binder_node的的binder_ref hlist，当binder_node所在进程死掉，binder_node死掉之后，遍历该hlist给所有注册该binder_node死亡通知的binder_ref发送binder_node死亡讣告。
struct binder_ref {
/* Lookups needed: */
/* node + proc => ref (transaction) */
/* desc + proc => ref (transaction, inc/dec ref) */
/* node => refs + procs (proc exit) */
int debug_id;
struct rb_node rb_node_desc; /* entry to desc/handle-keyed rb_tree */
struct rb_node rb_node_node; /* entry to binder_node-keyed rb_tree */
struct hlist_node node_entry; /* entry to the refered binder_node's refer hlist */
struct binder_proc *proc; /* user space ownner proc agent */
struct binder_node *node; /* binder_node refered by this binder_ref */
uint32_t desc; /* from userspace viewpoint, it is the handle of BpBinder*/
int strong; /* reference count of binder_ref itself*/
int weak; /* reference count of binder_ref itself */
struct binder_ref_death *death; /* binder_node death notification structure list */
};
binder_ref_death:
注册的binder_node death notification保存在此结构中，binder_node死亡后，本结构作为work挂到binder_ref所属线程的todo上，用户层执行时，收到死亡通知；因为binder_node已经死亡，所以回调binderDied中参数是wp
struct binder_ref_death {
struct binder_work work; /* work entry for binder_ref when its referencing binder component is dead */
void __user *cookie; /* weak reference of the dead binder component */
};
binder_proc:
该结构用于binder driver中对用户态进程的内核描述，比如binder通信的initiator和responser所在的进程，binder serivce使用addSerice注册时，对应的binder节点也是添加在addService的调用进程对应的这个binder_proc。这个结构维护了属于该binder_proc的所有binder节点，以及向其他进程的binder节点的所有引用，还维护了其所有线程对应的binder_thread。维护了进程mmap到的用户空间和对应的内核空间，用于transaction buffer的管理。
TODO：加重要字段的说明
struct binder_proc {
struct hlist_node proc_node; /* global binder_procs hash list mountpoint */
struct rb_root threads; /* ownning threads rb_tree */
struct rb_root nodes; /* ownning binder_nodes rb_tree keyed by user space unique(flattened) id */
struct rb_root refs_by_desc; /* ownning binder_nodes rb_tree keyed by desc(handle: from user space viewpoint) */
struct rb_root refs_by_node; /* ownning binder_nodes rb_tree keyed by binder_node(address) */
int pid;
struct vm_area_struct *vma; /* as type, used for mmap for user space */
struct task_struct *tsk; /* as type, get from current */
void *buffer; /* buffer start address from viewpoint of kernel linear address */
size_t user_buffer_offset; /* difference between user_vma and kernel lma */
struct list_head buffers; /* buffers list for allocating binder_buffer */
struct rb_root free_buffers; /* free binder_buffer(little binder_buffer kids) rb_tree used for fast buffer ops */
struct rb_root allocated_buffers; /* bufferes have used */
size_t free_async_space; /* async used buffer limit size */
struct page **pages; /* kernel page array pointer table ptr*/
size_t buffer_size; /* the whole size */
uint32_t buffer_free; /* free, can use to allocate binder_buffer */
struct list_head todo; /* if any proc-scope binder_work? mount to it, the binder worker thread waken will fetch it and do the work */
wait_queue_head_t wait; /* binder worker thread wait queue */
struct binder_stats stats; /* statistics */
struct list_head delivered_death; /* mountpoint of dead binder_node which has sent BR_DEAD_BINDER to userspace */
int max_threads; /* max threads of the thread pool */
int requested_threads; /* thread management and schdule used */
int requested_threads_started; /* thread management and schdule used */
int ready_threads; /* thread management and schdule used */
long default_priority; /* proc priority saved as binder worker thread default priority */
};
binder_thread:
维护其所拥有的binder_node；有进行线程等待和唤醒队列；维护transaction描述结构
struct binder_thread {
struct binder_proc *proc; /* ownned proc */
struct rb_node rb_node; /* ownning binder_node rb_tree */
int pid; /* saved current->pid, leader pid */
int looper; /* thread action control flag*/
struct binder_transaction *transaction_stack; /* all transactions are here as a stack to support cross-call transaction */
struct list_head todo; /* binder_work this thread should do sunch as transaction_completion, binder_ref_cnt_control */
uint32_t return_error; /* Write failed, return error code in read buf */
uint32_t return_error2; /* Write failed, return error code in read */
/* buffer. Used when sending a reply to a dead process that */
/* we are also waiting on */
wait_queue_head_t wait; /* wait queue this thread waiting on */
struct binder_stats stats; /* statistics */
};
looper_state:
线程状态描述。
enum {
BINDER_LOOPER_STATE_REGISTERED = 0x01, /* spawn binder worker thread joining looper */
BINDER_LOOPER_STATE_ENTERED = 0x02, /* isMain binder worker thread joining looper */
BINDER_LOOPER_STATE_EXITED = 0x04,
BINDER_LOOPER_STATE_INVALID = 0x08,
BINDER_LOOPER_STATE_WAITING = 0x10,
BINDER_LOOPER_STATE_NEED_RETURN = 0x20
};
binder_work:
描述工作的结构，工作队列模型实现串行化。
/**
** often used embedded in binder_transaction to classify work type.
** can also used in self binder_thread_read wait.
**/
struct binder_work {
struct list_head entry; /* entry mounted to proc->todo or thread->todo task queue */
enum {
BINDER_WORK_TRANSACTION = 1, /* transaction work */
BINDER_WORK_TRANSACTION_COMPLETE, /* transaction deliver ok result work */
BINDER_WORK_NODE, /* binder_node reference count operation work */
BINDER_WORK_DEAD_BINDER, /* binder_node is dead notification work */
BINDER_WORK_DEAD_BINDER_AND_CLEAR, /* clear death n10n for dead node work */
BINDER_WORK_CLEAR_DEATH_NOTIFICATION, /* clear death n10 work */
} type; /* work task type */
};
binder_transaction:
首先作为一种binder_work with type BINDER_WORK_TRANSACTION。
用于描述一个交互，有交互的源和目的用于binder路由，会同时挂接到源和目线程的binder_transaction栈中。带有给目的binder的功能代码和数据负载。
/**
** used by transaction to route and transfer control and payload.
**/
struct binder_transaction {
int debug_id;
struct binder_work work; /* embedded work type in transaction container, entry to the todo list */
struct binder_thread *from; /* from which thread this transaction is requested */
struct binder_transaction *from_parent; /* transaction_stack link->next, indicating the binder_transaction this binder_transaction is nested in */
struct binder_proc *to_proc; /* to which proc this transaction is deliverred */
struct binder_thread *to_thread; /* to which thread, set when the worker thread fetch the binder_work rather than when it is mount to source transaction_stack */
struct binder_transaction *to_parent; /* parent binder_transaction when embedded for cross-call */
unsigned need_reply : 1;
/*unsigned is_dead : 1;*/ /* not used at the moment */
struct binder_buffer *buffer; /* transaction payload */
unsigned int code; /* interface function code */
unsigned int flags; /* extra flags. refer to user space, such as if_async, accept_fds, src_priority */
long priority; /* priority of requester thread, nice vaule precisely */
long saved_priority; /* saved priority of the binder worker thread before it is niced, used to restore the binder worker thread's priority */
uid_t sender_euid; /* priviledge control */
};
binder_buffer:
transaction数据，负载数据紧随其后，类同binder_io和Parcel。
struct binder_buffer {
struct list_head entry; /* free and allocated entries by addesss */
struct rb_node rb_node; /* free entry by size or allocated entry by address */
unsigned free : 1;
unsigned allow_user_free : 1;
unsigned async_transaction : 1;
unsigned debug_id : 29;
struct binder_transaction *transaction; /* de-refer to the ownner binder_transaction */
struct binder_node *target_node; /* target binder_node for transaction. NULL to indicate it is for reply */
size_t data_size; /* as userspace same name concept */
size_t offsets_size; /* as userspace same name concept */
uint8_t data[0]; /* data_size+offsets_size payload of size_t */
};
下面介绍binder driver的几个关键部分（仍是cupecake时的描述）。
1. 设备打开和内存映射
通过binder_open实现设备打开，新分配一个binder_proc，通过cuurent和current->gourp_leader得到pid，将proc指针保存到file->private_data中，实现进程到对应binder_proc的映射。
不同进程通过binder_mmap实现虚拟内存映射到相同物理页面，binder_buffer从该内存范围中分配。驱动程序直接操作物理页面，为不同进程建立页表进行映射到线性地址空间。一个原则：物理页面最少分配，需要时分配，空闲时释放。
TODO：适当增加物理页面、线性空间、进程虚拟空间映射的描述
驱动程序使用该段内存首地址的内核线性空间地址和用户线性空间地址的差值为不同进程调整地址即可访问该段内存中的数据，减少数据拷贝，提高效率。如下图：
2. 缓冲区管理
管理和分配binder_buffer。已经分配的即正在被使用的binder_buffer挂到address keyed rb-tree上，便于binder_buffer按地址查找；释放binder_buffer时将相邻的空闲buffer合并成一个大的binder_buffer后挂到size keyed rb-tree上，以后分配时可以按照size快速查找best-fit的binder_buffer。
注意其中的几个数据结构 list和rb_tree.
3. binder_node/ binder_ref管理的数据结构
一个binder_proc的binder_node都挂载在一棵rb_tree上，键值是pointer。其实是BBinder的mRefs地址；BBinder本身的地址在cookie字段中。用来由flatten_binder_node.pointer查找对应的binder_node。
一个进程内的所有binder_ref也分别按照desc(handle)、node挂载到两颗rb_tree上，前者用于按照handle查找binder_ref，为新binder_ref分配handle；后者用于查找binder_ref指向的binder_node（target_proc的binder_node，binder_node是内核自己分配的空间，不同proc不会冲突）。
4. 文件共享
文件系统主要有struct FILE、file descriptor、struct file、struct inode几个层次。进程组内打开的同一个文件在共享同一个struct file，而不同进程内打开同一个文件是共享同一个struct inode。binder为了在不同进程间使用文件句柄实现文件共享，就要在下一个层次就struct file级别共享。
5. binder_node/ binder_ref计数的字段
首先service端收到的BR_INCREFS/BR_ACQUIRE并不是client端BC_INCREFS/BC_ACQUIRE的因果，而是在客户端发送BC_INCREFS/BC_ACQUIRE之前就已经发送给BBinder，具体后面内容陆续叙述。RELEASE/DECREFS同理。
binder_ref和binder_node自身都有强弱计数，生命周期受强计数和弱计数共同控制，总体上可以认为是LIFETIME_MODE_WEAK。
binder_ref初始计数时，必然是binder_transaction的binder_buffer中含有BBinder时，这时驱动初次检测到该BBinder，为其在所属binder_proc中创建 binder_node，并在binder_transaction的目标binder_proc中创建binder_ref；binder_ref初次计数并增加binder_node的内部计数，并形成计数增加小状态机，增加用户空间binder component的计数。
对于每次交互，都会对target binder_node使用binder_inc_node增加计数，当交互完成后，调用binder_dec_node对原target binder_node减计数；
每次交互中，对于binder_buffer中的binder object对应的binder_ref和binder_node都会增加计数；当该binder_buffer被freed时，对所有binder object减少计数。
binder_release时，会发出DEATH_NOTIFICATION，回调所有death函数，所有binder_ref减计数，会导致binder_node的work重新挂接到thread->todo上，类似小状态机，转两次，从而会发出BR_RELEASE/BR_DECREFS，组件减计数，销毁。
6工作线程优先级调整
参见Framework部分工作线程优先级模型部分。
7 Async Transaction -- ONE_WAY
对于异步操作，其处理优先级低于同步操作。可以想象，异步操作的大部分情景是实时周期性或突发性的大批量数据。为了实现异步操作的优先级低于同步操作，除了用于同步的binder_proc::todo和binder_thread::todo队列，binder driver还准备了一个异步队列binder_node::async_todo。若当前的work queue中没有异步transaction，那么binder driver会将异步transaction work挂到todo上，若是当前transaction中已有异步请求，那么会将异步transaction work挂到目标binder_node的async_todo上；当前一个异步请求处理完毕BC_FREE_BUFFER的时候，再摘一个同target_node的async_todo上的异步请求挂到todo上，这样将异步请求分散开处理。这样甚至可以发送取消异步请求的同步请求，唯一的一个缺点是，实现取消异步队列中所有的异步请求时，需要应用程序层做些协议配合以区别哪些异步请求是在取消同步请求之前和之后发送的。
8 LOOPER SPAWN
当目标进程中的binder worker thread不够用时，驱动会发送BR_SPAWN_LOOPER指示用户层新建工作线程。
case BR_SPAWN_LOOPER:
mProcess->spawnPooledThread(false);
break;
void ProcessState::spawnPooledThread(bool isMain)
{
if (mThreadPoolStarted) {
int32_t s = android_atomic_add(1, &mThreadPoolSeq);
char buf[16];
snprintf(buf, sizeof(buf), "Binder_%X", s);
ALOGV("Spawning new pooled thread, name=%s\n", buf);
sp
t->run(buf);
}
}
class PoolThread : public Thread
{
public:
PoolThread(bool isMain)
: mIsMain(isMain)
{
}
protected:
virtual bool threadLoop()
{
IPCThreadState::self()->joinThreadPool(mIsMain);
return false;
}
const bool mIsMain;
};
void IPCThreadState::joinThreadPool(bool isMain)
{
mOut.writeInt32(isMain ? BC_ENTER_LOOPER : BC_REGISTER_LOOPER);
// This thread may have been spawned by a thread that was in the background
// scheduling group, so first we will make sure it is in the foreground
// one to avoid performing an initial transaction in the background.
set_sched_policy(mMyThreadId, SP_FOREGROUND);
……..
}
isMain工作线程通过BC_ENTER_LOOPER通知驱动进入工作线程池，非isMain线程通过BC_REGISTER_LOOPER通知驱动进入工作线程池。
BINDER_LOOPER_STATE_NEED_RETURN用于在驱动中控制thread退出工作循环。用户层可以通过发送BC_EXIT_LOOPER来使线程退出线程池。
Android Binder机制之协议分析
将协议的命令分散到几个具体的常用情景中加以考察。各个情景重点考察的命令如下表格所列。
| 情景1. |
BC_INCREFS BC_ACQUIRE BC_RELEASE BC_DECREFS |
| 情景2. |
BC_TRANSACTION BC_REPLY BR_TRANSACTION BR_REPLY BC_FREE_BUFFER BR_INCREFS BR_ACQUIRE BC_INCREFS_DONE BC_ACQUIRE_DONE |
| 情景5 |
BC_REQUEST_DEATH_NOTIFICATION BC_CLEAR_DEATH_NOTIFICATION BC_DEAD_BINDER_DONE BR_DEAD_BINDER BR_CLEAR_DEATH_NOTIFICATION_DONE |
情景1.获取defaultServiceManager
主要侧重用户态BpBinder转BpServiceManager的过程。
In frameworks/base/libs/binder/IServiceManager.cpp
sp
{
if (gDefaultServiceManager != NULL) return gDefaultServiceManager;
{
AutoMutex _l(gDefaultServiceManagerLock);
if (gDefaultServiceManager == NULL) {
gDefaultServiceManager =interface_cast
ProcessState::self()->getContextObject(NULL));
}
}
return gDefaultServiceManager;
}
首先看ProcessState::self()->getContextObject(NULL)构造BpBinder(0)的过程。
In frameworks/base/libs/binder/ProcessState.cpp
sp
{
return getStrongProxyForHandle(0);
}
sp
{
sp
AutoMutex _l(mLock);
handle_entry* e = lookupHandleLocked(handle);
if (e != NULL) {
// We need to create a new BpBinder if there isn't currently one, OR we
// are unable to acquire a weak reference on this current one. See comment
// in getWeakProxyForHandle() for more info about this.
IBinder* b = e->binder;
if (b == NULL || !e->refs->attemptIncWeak(this)) {
b = new BpBinder(handle);
e->binder = b;
if (b) e->refs = b->getWeakRefs();
result = b;
} else {
// This little bit of nastyness is to allow us to add a primary
// reference to the remote proxy when this team doesn't have one
// but another team is sending the handle to us.
result.force_set(b);
e->refs->decWeak(this);
}
}
return result;
}
BpBinder构造时，会extendObjectLifetime(OBJECT_LIFETIME_WEAK)，从而使BpBinder对象本身是弱计数控制；
BpBinder::BpBinder(int32_t handle)
: mHandle(handle)
, mAlive(1)
, mObitsSent(0)
, mObituaries(NULL)
{
LOGV("Creating BpBinder %p handle %d\n", this, mHandle);
extendObjectLifetime(OBJECT_LIFETIME_WEAK);
IPCThreadState::self()->incWeakHandle(handle);
}
然后增加驱动中其所对应的binder_ref的弱计数，此处是增加对应于binder_context_mgr_node的binder_ref的弱计数；
void IPCThreadState::incWeakHandle(int32_t handle)
{
LOG_REMOTEREFS("IPCThreadState::incWeakHandle(%d)\n", handle);
mOut.writeInt32(BC_INCREFS);
mOut.writeInt32(handle);
}
此时可以认为后面图中弱引用1建立起来；
出于效率原因，BC_INCREFS并没有立刻发送到驱动。但为了便于描述，继续沿此条线索考察。
static int binder_inc_ref(struct binder_ref *ref, weak, NULL)
{
int ret;
{
if (ref->weak == 0) {
ret = binder_inc_node(ref->node, 0, 1, target_list);
if (ret)
return ret;
}
ref->weak++;
}
return 0;
}
static int binder_inc_node(struct binder_node *node, weak, TRUE, NULL)
{
return 0;
}
此时可以认为后面图中弱引用2建立起来。
然后asInterface引发BpRefbase构造时，会extendObjectLifetime(OBJECT_LIFETIME_WEAK)使 BpRefbase对象本身是弱计数控制；
BpRefBase::BpRefBase(const sp
: mRemote(o.get()), mRefs(NULL), mState(0)
{
extendObjectLifetime(OBJECT_LIFETIME_WEAK);
if (mRemote) {
mRemote->incStrong(this); // Removed on first IncStrong().
mRefs = mRemote->createWeak(this); // Held for our entire lifetime.
}
}
mRemote->incStrong(this)增加BpBinder对象的强引用计数（注意mRemote是IBinder *普通指针），此时建立了一条到BpBinder的强引用3.
这会引发BpBinder::onFirstRef()向驱动发送BC_ACQUIRE增加binder_ref的强计数。这时可以认为强引用4建立起来。
void BpBinder::onFirstRef()
{
LOGV("onFirstRef BpBinder %p handle %d\n", this, mHandle);
IPCThreadState* ipc = IPCThreadState::self();
if (ipc) ipc->incStrongHandle(mHandle);
}
void IPCThreadState::incStrongHandle(int32_t handle)
{
LOG_REMOTEREFS("IPCThreadState::incStrongHandle(%d)\n", handle);
mOut.writeInt32(BC_ACQUIRE);
mOut.writeInt32(handle);
}
BC_ACQUIRE在驱动中的处理如下
static int binder_inc_ref(struct binder_ref *ref, strong, NULL)
{
int ret;
if (strong) {
if (ref->strong == 0) {
ret = binder_inc_node(ref->node, 1, 1, target_list);
if (ret)
return ret;
}
ref->strong++;
}
return 0;
}
static int binder_inc_node(struct binder_node *node, strong, true, NULL)
{
if (strong) {
if (internal) {
node->internal_strong_refs++;
}
}
return 0;
}
可以认为强引用计数5建立起来。
mRefs = mRemote->createWeak(this)建立一个到BpBinder（弱生命周期）的弱引用。
RefBase::weakref_type* RefBase::createWeak(const void* id) const
{
mRefs->incWeak(id);
return mRefs;
}
void RefBase::weakref_type::incWeak(const void* id)
{
weakref_impl* const impl = static_cast
impl->addWeakRef(id);
const int32_t c = android_atomic_inc(&impl->mWeak);
LOG_ASSERT(c >= 0, "incWeak called on %p after last weak ref", this);
}
此时建立弱引用6。
最后sp
template
sp
{
if (other) other->incStrong(this);
}
BpRefBase::onFirstRef()会被调用，设置状态为kRemoteAcquired，表征真正有了外部sp
void BpRefBase::onFirstRef()
{
android_atomic_or(kRemoteAcquired, &mState);
}
此时强引用7建立起来。
扩展地，如果有wp
需要说明的是：
图中序号并不表示真正的引用时序。引用1、4、2、5是在发送BC_INCREFS和BC_ACQUIRE到驱动时建立的，即是在发送BC_TRANSACTION即调用ServiceManger的方法时一起发送给驱动，真正增加引用的。binder driver部分对BC_INCREFS和BC_ACQUIRE的处理，二者最后归结到binder_inc_ref和binder_inc_node上，由于binder drvier在处理BINDER_SET_CONTEXT_MGR已经有如下处理，
case BINDER_SET_CONTEXT_MGR:
binder_context_mgr_node->local_weak_refs++;
binder_context_mgr_node->local_strong_refs++;
binder_context_mgr_node->has_strong_ref = 1;
binder_context_mgr_node->has_weak_ref = 1;
所以target== binder_context_mgr_node时， BC_INCREFS命令处理流程是没有操作；BC_ACQUIRE命令处理流程是node->internal_strong_refs++，表示有了binder_ref到binder_node的驱动内部维护的引用计数。
下面考察一般binder component的减计数过程，还是主要侧重用户态。
首先sp
template
sp
{
if (m_ptr) m_ptr->decStrong(this);
}
BpRefBase减少自身强计数，如果计数为零，则引发BpRefBase::onLastStrongRef，引起强引用3的释放。
void RefBase::decStrong(const void* id) const
{
weakref_impl* const refs = mRefs;
refs->removeStrongRef(id);
const int32_t c = android_atomic_dec(&refs->mStrong);
LOG_ASSERT(c >= 1, "decStrong() called on %p too many times", refs);
if (c == 1) {
refs->mBase->onLastStrongRef(id); //why this form? why not this->onLastStrongRef? vtbl.
if ((refs->mFlags&OBJECT_LIFETIME_MASK) == OBJECT_LIFETIME_STRONG) {
delete this;
}
}
refs->decWeak(id);
}
void BpRefBase::onLastStrongRef(const void* id)
{
if (mRemote) {
mRemote->decStrong(this);
}
}
BpBinder::decStrong减少BpBinder自身强计数，如果强计数减到0则发送BC_RELEASE减少binder_ref的强计数。此时可以认为强引用4释放。
void BpBinder::onLastStrongRef(const void* id)
{
LOGV("onLastStrongRef BpBinder %p handle %d\n", this, mHandle);
IF_LOGV() {
printRefs();
}
IPCThreadState* ipc = IPCThreadState::self();
if (ipc) ipc->decStrongHandle(mHandle);
}
void IPCThreadState::decStrongHandle(int32_t handle)
{
LOG_REMOTEREFS("IPCThreadState::decStrongHandle(%d)\n", handle);
mOut.writeInt32(BC_RELEASE);
mOut.writeInt32(handle);
}
在驱动中，可以认为会引起强引用5的释放。
static int binder_dec_ref(struct binder_ref *ref, true)
{
if (strong) {
if (ref->strong == 0) {
binder_user_error("binder: %d invalid dec strong, "
"ref %d desc %d s %d w %d\n",
ref->proc->pid, ref->debug_id,
ref->desc, ref->strong, ref->weak);
return -EINVAL;
}
ref->strong--;
if (ref->strong == 0) {
int ret;
ret = binder_dec_node(ref->node, strong, 1);
if (ret)
return ret;
}
if (ref->strong == 0 && ref->weak == 0)
binder_delete_ref(ref);
return 0;
}
返回到BpBinder::decStrong继续执行会减自身弱计数BpBinder::RefBase::mRefs，因为BpRefBase的mRefs增加了其一次弱计数，所以BpBinder此时不会销毁。执行到此时，上图中所有强引用都已经释放，只剩下弱引用。
返回到BpRefBase::decStrong继续执行会减自身弱计数BpRefBase::RefBase::mRefs，因为没有wp
void RefBase::weakref_type::decWeak(const void* id)
{
weakref_impl* const impl = static_cast
impl->removeWeakRef(id);
const int32_t c = android_atomic_dec(&impl->mWeak);
LOG_ASSERT(c >= 1, "decWeak called on %p too many times", this);
if (c != 1) return;
if ((impl->mFlags&OBJECT_LIFETIME_WEAK) == OBJECT_LIFETIME_STRONG) {
// This is the regular lifetime case. The object is destroyed
// when the last strong reference goes away. Since weakref_impl
// outlive the object, it is not destroyed in the dtor, and
// we'll have to do it here.
if (impl->mStrong == INITIAL_STRONG_VALUE) {
// Special case: we never had a strong reference, so we need to
// destroy the object now.
delete impl->mBase;
} else {
// LOGV("Freeing refs %p of old RefBase %p\n", this, impl->mBase);
delete impl;
}
} else {
// less common case: lifetime is OBJECT_LIFETIME_{WEAK|FOREVER}
impl->mBase->onLastWeakRef(id);
if ((impl->mFlags&OBJECT_LIFETIME_MASK) == OBJECT_LIFETIME_WEAK) {
// this is the OBJECT_LIFETIME_WEAK case. The last weak-reference
// is gone, we can destroy the object.
delete impl->mBase;
}
}
}
BpRefBase::~BpRefBase()
{
if (mRemote) {
if (!(mState&kRemoteAcquired)) {
mRemote->decStrong(this);
}
mRefs->decWeak(this); // BpBinder::mRefs that is BpBinder::RefBase::mRefs.
}
}
这时会减BpBinder::RefBase::mRefs弱计数，这次减弱计数对应于BpRefBase构造函数中createWeak的那次增加弱计数，此时上图中弱引用6释放。这时BpBinder的弱计数减为0，delete impl->mBase销毁BpBinder对象。
BpBinder::~BpBinder()
{
LOGV("Destroying BpBinder %p handle %d\n", this, mHandle);
IPCThreadState* ipc = IPCThreadState::self();
mLock.lock();
Vector
if(obits != NULL) {
if (ipc) ipc->clearDeathNotification(mHandle, this);
mObituaries = NULL;
}
mLock.unlock();
if (obits != NULL) {
// XXX Should we tell any remaining DeathRecipient
// objects that the last strong ref has gone away, so they
// are no longer linked?
delete obits;
}
if (ipc) {
ipc->expungeHandle(mHandle, this);
ipc->decWeakHandle(mHandle);
}
}
void IPCThreadState::decWeakHandle(int32_t handle)
{
LOG_REMOTEREFS("IPCThreadState::decWeakHandle(%d)\n", handle);
mOut.writeInt32(BC_DECREFS);
mOut.writeInt32(handle);
}
ipc->decWeakHandle(mHandle)会向驱动发送BC_DECREFS，可以认为弱引用1释放。
在驱动中，引起binder_ref的销毁，弱引用2释放。
static int binder_dec_ref(struct binder_ref *ref, weak)
{
{
if (ref->weak == 0) {
binder_user_error("binder: %d invalid dec weak, "
"ref %d desc %d s %d w %d\n",
ref->proc->pid, ref->debug_id,
ref->desc, ref->strong, ref->weak);
return -EINVAL;
}
ref->weak--;
}
if (ref->strong == 0 && ref->weak == 0)
binder_delete_ref(ref);
return 0;
}
如果当只剩下弱计数时有wp
此时，可以看出extendObjectLifetime(LIFETIME_MODE_WEAK)的作用，weak说明生命力更强！只要有wp
情景2. 向defaultServiceManager注册NamedService
defaultServiceManager()->addService(…)即
virtual status_t BpServiceManager::addService(const String16& name, const sp
{
Parcel data, reply;
data.writeInterfaceToken(IServiceManager::getInterfaceDescriptor());
data.writeString16(name);
data.writeStrongBinder(service);
status_t err = remote()->transact(ADD_SERVICE_TRANSACTION, data, &reply);
return err == NO_ERROR ? reply.readExceptionCode() : err;
}
目标binder component是ServiceManager。要注册的命名binder放在传输的数据中，NamedService binder component在writeStrongBinder=>flatten_binder中做如下serialize，
obj.type = BINDER_TYPE_BINDER;
obj.binder = local->getWeakRefs();
obj.cookie = local;
然后remote()->transact(…)即BpBinder::transact(…)调用IPCThreadState::transact(…)=>talkWithDriver()=>ioctl()。
在binder driver中，binder_ioctl() => binder_thread_write()，对于BC_TRANSACTION处理的是 binder_transaction()。
binder_transaction的工作是将TRANSACTION请求封装成一个binder_work，交给目标binder_node所在的进程，由binder工作线程执行目标binder的方法。非交叉式调用时，binder_transaction请求时总是提交binder_work给目标binder_node所在的进程binder_proc。在Service进程中，由空闲的binder_thread从binder_proc::todo list上取下工作执行。需要提个为什么吗？不太需要，因为Service端并一定是哪个工作线程来执行请求。binder_transation答复时则是将binder_work挂接到请求者binder_thread::todo上，即将transaction_reply返回给等待结果的请求者线程。
此情景中，target_node是 binder_context_mgr_node。target_proc是service_manager进程。service_manager的工作比较简单，因此，设计上只有一个进程按部就班地顺序处理请求，没有多余的work thread。
然后，binder_transaction用请求者信息填充binder_transaction结构，请求者线程的euid、priority信息将会带到并被设置到工作线程，使工作线程像请求者线程附体一样。
在将binder_work挂到服务进程待做队列(todo)之前，需要对目标binder_node做加计数防止在请求处理之前计数被减到0引起binder_node的销毁，否则，情况就不妙了。在Service端处理完请求后，Service端释放binder_buffer并做binder_node的减计数。
1576 t->buffer->target_node = target_node;
1577 if (target_node)
1578 binder_inc_node(target_node, 1, 0, NULL);
此时增加的计数是binder_node::local_strong_refs。需要区别的是，binder_node::internal_strong_refs是新创建了引用该binder_node的binder_ref时增加。
另外的line #1576设置binder_buffer::target_node为目标node，binder_buffer中保存target_node的作用不在于node route，而是用于binder_node的对应地减计数；当该binder_buffer释放时，做target_node和缓冲区内部binder的减计数，这是一个时机。t->buffer->target_node只有在binder transaction request的过程设置；binder transaction reply过程，t->buffer->target_node是NULL，因为reply时没有binder_node，只需要返回请求者线程就可以。
然后binder_transaction做binder_buffer中binder的转换，本情景中，是将命令服务对应的BINDER_TYPE_BINDER binder node转为service_manager使用的BINDER_TYPE_HANDLE，转换规则不再赘述，binder_node计数后面详述。
然后，会将该binder_transaction挂到线程自己的transaction stack(栈—功能在交叉式调用时叙述，栈一般保持空)上，并将其通过内嵌的binder_work挂到target_proc::todo list上；将BINDER_WORK_TRANSACTION_COMPLETE类型的binder_work挂到线程自己的todo list上，表示transaction已经成功投递给目标进程(请求处理完恢复执行时处理)；唤醒等待在target_proc->wait queue上的binder_thread。在Serivce中，binder工作线程空闲后就等待在进程的wait queue上，总之，进程。本情景中，就是发送BC_ENTER_LOOPER进入线程池后的service_manager进程本身。
下面详述binder_buffer中binder_node的计数，这是缓冲区中作为数据负载的binder_node，不是目标binder_node。
switch (fp->type) {
case BINDER_TYPE_BINDER:
case BINDER_TYPE_WEAK_BINDER: {
struct binder_ref *ref;
struct binder_node *node = binder_get_node(proc, fp->binder);
if (node == NULL) {
node = binder_new_node(proc, fp->binder, fp->cookie);
if (node == NULL) {
return_error = BR_FAILED_REPLY;
goto err_binder_new_node_failed;
}
node->min_priority = fp->flags & FLAT_BINDER_FLAG_PRIORITY_MASK;
node->accept_fds = !!(fp->flags & FLAT_BINDER_FLAG_ACCEPTS_FDS);
}
ref = binder_get_ref_for_node(target_proc, node);
if (fp->type == BINDER_TYPE_BINDER)
fp->type = BINDER_TYPE_HANDLE;
else
fp->type = BINDER_TYPE_WEAK_HANDLE;
fp->handle = ref->desc;
binder_inc_ref(ref, fp->type == BINDER_TYPE_HANDLE, &thread->todo);
} break;
当binder_buffer中出现BINDER_TYPE_BINDER的binder时，如果其所属的源进程中还没有对应binder_node，则创建binder_node；如果目标进程中还不存在引用到该binder_node的binder_ref，那么为其创建binder_ref，然后binder_inc_ref及以后设置完成至关重要的计数工作。下面仅考察强计数的情况。
static int binder_inc_ref(struct binder_ref *ref, int strong, struct list_head *target_list)
{
int ret;
if (strong) {
if (ref->strong == 0) {
ret = binder_inc_node(ref->node, 1, 1, target_list);
if (ret)
return ret;
}
ref->strong++;
}
return 0;
}
static int binder_inc_node(struct binder_node *node, int strong, int internal, struct list_head *target_list)
{
if (strong) {
if (internal) {
node->internal_strong_refs++;
} else
node->local_strong_refs++;
if (!node->has_strong_ref && target_list) {
list_del_init(&node->work.entry);
list_add_tail(&node->work.entry, target_list);
}
return 0;
}
首先参见前面数据结构介绍中binder_node各字段的涵义。
binder_inc_node先做internal_strong_refs加计数，表示有新的binder_ref引用到此binder_node。接着会将BINDER_NODE_WORK binder_node::binder_work挂到源线程的待办队列上。
source_thread->todo上已经有一个TRANSACTION_COMPLETE binder_work，此时又加上一个用于binder_node生命周期控制的binder_work，当service_manager处理完毕后，还会加一个返回transaction reply的TRANSACTION binder_work，这时源线程从source_thread->wait queue上被唤醒，依次处理待办binder_work。
此处提前介绍BINDER_NODE_WORK binder_work的处理。
对于BINDER_WORK_TRANSACTION_COMPLETE，向源线程输入缓冲区写BR_TRANSACTION_COMPLETE，摘除binder_work；
对于BINDER_WORK_NODE，则是类似一个小状态机对该binder_work转两轮。第一轮向BBinder发送BR_INCREFS，第二轮向BBinder发送BR_ACQUIRE，分别增加BBinder的弱计数和强计数。第一轮时不摘除BINDER_WORK_NODE binder_work，第二轮时才把它从链表摘除，实现状态转换。
int strong = node->internal_strong_refs || node->local_strong_refs;
int weak = !hlist_empty(&node->refs) || node->local_weak_refs || strong;
if (weak && !node->has_weak_ref) {
cmd = BR_INCREFS;
cmd_name = "BR_INCREFS";
node->has_weak_ref = 1;
node->pending_weak_ref = 1;
node->local_weak_refs++;
} else if (strong && !node->has_strong_ref) {
cmd = BR_ACQUIRE;
cmd_name = "BR_ACQUIRE";
node->has_strong_ref = 1;
node->pending_strong_ref = 1;
node->local_strong_refs++;
}
strong标识已经有了对binder_node的强引用，包含两部分强引用，或者是源自binder_ref的internal_strong_ref，或者是源自向对应binder component发送命令过程中的local_strong_ref；weak标识已经有了对binder_node的弱引用，当实际有了强引用时，是一定需要弱引用的，所以weak上逻辑或strong。而has_strong_ref和has_weak_ref则是用来判断初次计数和末次计数的标志，用于向BBinder发送命令进行生命周期控制。pending_strong_ref和pending_weak_ref表示已经将命令写进用户态读缓冲区，但是由于源线程要等待到transaction reply后才能被唤醒并一同处理命令，所以在用户态进行下一次ioctl返回给驱动BC_INCREFS_DONE和BC_ACQUIRE_DONE之前，保持pending。
加上以后的transaction reply的BR_REPLY，源线程读缓冲区中的命令（命令的数据从略）如下：
源线程在waitForResponse(…)=>executeCommand(…)中对BR_INCREFS和BR_ACQUIRE两个命令处理如下：
case BR_ACQUIRE:
refs = (RefBase::weakref_type*)mIn.readInt32();
obj = (BBinder*)mIn.readInt32();
obj->incStrong(mProcess.get());
mOut.writeInt32(BC_ACQUIRE_DONE);
mOut.writeInt32((int32_t)refs);
mOut.writeInt32((int32_t)obj);
break;
case BR_INCREFS:
refs = (RefBase::weakref_type*)mIn.readInt32();
obj = (BBinder*)mIn.readInt32();
refs->incWeak(mProcess.get());
mOut.writeInt32(BC_INCREFS_DONE);
mOut.writeInt32((int32_t)refs);
mOut.writeInt32((int32_t)obj);
break;
当再次ioctl将命令写到驱动中时，将pending清除，命令交互结束，做减计数，local_weak_refs和local_strong_refs清零。
case BC_INCREFS_DONE:
case BC_ACQUIRE_DONE: {
void __user *node_ptr;
void *cookie;
struct binder_node *node;
if (get_user(node_ptr, (void * __user *)ptr))
return -EFAULT;
ptr += sizeof(void *);
if (get_user(cookie, (void * __user *)ptr))
return -EFAULT;
ptr += sizeof(void *);
node = binder_get_node(proc, node_ptr);
if (cmd == BC_ACQUIRE_DONE) {
node->pending_strong_ref = 0;
} else {
node->pending_weak_ref = 0;
}
binder_dec_node(node, cmd == BC_ACQUIRE_DONE, 0);
break;
}
因为BR_INCREFS和BR_ACQUIRE在waitForResponse中在得到处理，从而BBinder的计数在BpServiceManager::addService(const String16& name, const sp
在一系列命令过程中，binder_node各字段变化如下：
|
|
hs |
ls |
is |
ps |
hw |
lw |
iw |
pw |
| after kalloced |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
| inc ref for new node |
0 |
0 |
1 |
0 |
0 |
0 |
1 |
0 |
| BINDER_WORK_NODE (r1) |
0 |
0 |
1 |
0 |
1 |
1 |
1 |
1 |
| BINDER_WORK_NODE (r2) |
1 |
1 |
1 |
1 |
1 |
1 |
1 |
1 |
| BC_INCREFS_DONE |
1 |
1 |
1 |
1 |
1 |
0 |
1 |
0 |
| BC_ACQUIRE_DONE |
1 |
0 |
1 |
0 |
1 |
0 |
1 |
0 |
最后的字段值也就是从debugfs/binder/proc/中dump出来的绝大多数binder_node的字段值。
下面返回service_manager进程考察其对BINDER_TRANSACTION_WORK的处理。
binder工作线程都是睡眠在主进程的等待队列上等待请求者线程挂在目标主线程待办队列上的工作，当遇到BINDER_TRANSACTION_WORK类型的binder_work时，就准备返回到用户态，由用户层继续处理。对于其它类型的的binder_work，则延迟到transaction时一并返回到用户态处理。在binder_thread_read中继续根据请求者传递过来的binder_transaction和binder_buffer中的信息调整工作线程的有效用户权限和线程优先级，根据内存映射调整binder_buffer的地址供工作线程在用户态访问，把该binder_work摘除，设置binder_buffer在transaction处理完毕之后允许释放标记，并且在此时设置目标线程binder_transaction::to_thread，然后就返回切换到用户态。
service_manager进程从ioctl返回后，如defaultServiceManager节中所述，
binder_acquire(bs, ptr); => BC_ACQUIRE
binder_link_to_death(bs, ptr, &si->death); => BC_REQUEST_DEATH_NOTIFICATION
binder_send_reply(bs, reply, buffer_to_free, status); => BC_FREE_BUFFER + BC_REPLY
这些命令序列又通过ioctl返回到驱动，进入binder_thread_write transaction reply过程。
对于BC_ACQUIRE，仅仅是增加binder_ref的强计数，因为不是初始计数，所以这次不会有binder_inc_node。BC_ACQUIRE的作用在于建立BpBinder到binder_ref的强引用，这样在后续BC_FREE_BUFFER释放binder_buffer做binder_dec_ref时，不至于binder_ref强计数减到0，从而引起binder_node减计数的错误行为。
对于BC_REQUEST_DEATH_NOTIFICATION，创建binder_ref_death结构用于binder_node death notification，具体后面详述。
对于BC_FREE_BUFFER，使用binder_transaction_buffer_release(…)处理，binder_dec_node(target_node)，对binder_buffer中的binder根据类型分别做binder_dec_node(…), binder_dec_ref(…), filp_close(…)。
对于BC_REPLY，使用binder_transaction reply处理。首先恢复工作线程的默认优先级，target_thread为binder_transaction中保存的请求者线程，经过与binder_transaction request时类似的处理后，唤醒睡眠在target_thread->wait queue上的线程即请求者线程。然后执行到binder_thread_read中，睡眠在binder_proc::wait queue上或是执行binder_proc::todo list上的binder_work。
请求者线程在binder_thread_read中被唤醒并被调度执行后，即执行前面介绍的binder_work todo list，依次处理WORK_ TRANSACTION_COMPLETE、WORK_BINDER_NODE、WORK_TRANSACTION。最后的WORK_TRANSACTION binder work是内嵌在replyed binder_transaction中的，此时t->buffer->target_node==NULL，所以写入请求者线程读缓冲的命令是BR_REPLY；然后返回用户态。BR_TRANSACTION_COMPLETE、BR_INCREFS、BR_ACQUIRE在用户层的处理前面已经描述，用户层对BR_REPLY的处理是在IPCThreadState::transact(…)=>IPCThreadState::waitForResponse(…)中。
case BR_REPLY:
{
binder_transaction_data tr;
err = mIn.read(&tr, sizeof(tr));
ALOG_ASSERT(err == NO_ERROR, "Not enough command data for brREPLY");
if (err != NO_ERROR) goto finish;
if (reply) {
if ((tr.flags & TF_STATUS_CODE) == 0) {
reply->ipcSetDataReference(
reinterpret_cast
tr.data_size,
reinterpret_cast
tr.offsets_size/sizeof(size_t),
freeBuffer, this);
} else {
err = *static_cast
freeBuffer(NULL,
reinterpret_cast
tr.data_size,
reinterpret_cast
tr.offsets_size/sizeof(size_t), this);
}
} else {
freeBuffer(NULL,
reinterpret_cast
tr.data_size,
reinterpret_cast
tr.offsets_size/sizeof(size_t), this);
continue;
}
}
goto finish;
将输出缓冲区中的数据读出，放入Parcel reply中供IPCThreadState::transact(…)返回给BpServiceManager::remote()::transact(…)。Parcel中数据使用完毕，对象销毁时，自动freeBuffer=>BC_FREE_BUFFER释放内核中的binder_buffer。
情景3. Get NamedService from defaultServiceManager
可以综合情景1和情景2得到本情景。考察请求者和原service注册者不在同一进程的情况。
请求者将SVC_MGR_GET_SERVICE发送到service_manager后，service_manager将BINDER_TYPE_HANDLE binder作为结果写到缓冲区。在驱动binder_transaction reply中，为BINDER_TYPE_HANDLE binder在请求者进程中创建binder_ref并binder_inc_ref，这会增加binder_node::internal_strong_refs计数。然后和情景1类似返回到请求者进程的用户空间，构造BpBinder和BpSubInterface。
普通进程获取NamedService后像情景1会立刻向驱动发送BC_INCREFS和BC_ACQUIRE，这一点和service_manager作为NamedService的一个客户端只发送BC_ACQUIRE是稍微有点区别的。
这从debugfs/binder dump出来的binder_ref::weak的值可以看出来，servive_manager所在进程的绝大多数binder_ref::weak值为0，binder_ref::strong值为1；而其它进程的绝大多数binder_ref::weak和binder_ref::strong值相同，不在transaction时，绝大多数是1。与service_manager在处理SVC_MGR_ADD_SERVICE时没有发送BC_INCREFS的事实符合。
情景4. Transfer Non-named binder component
不具名binder component的在缓冲区中初次传输主要有两种情况，一种是从NameService获取一个工作binder component或者用于接入管理的表征客户端的binder component，一种是客户端向Service端注册一个异步Event Sink。
这个情景与情景2类似，不同点仍是使用binder的客户端会向binder component所在端发送BC_INCREFS和BC_ACQUIRE同时增加binder_ref的弱计数和强计数。
扩展地，思考上图中各结构的自身计数、引用链使用的各结构的字段链（概念上都有mPtr和mRefs，只是具体体现不一样）和某些界面之间的命令链。
情景5.binder death notification registering and notification
由于是跨进程调用，因此客户端会关心Service端进程是否还活着，这通过register/notify机制来实现。由于Service进程死掉后，是不可能再和客户端交互的，所以后续的交互都是客户端和driver在交互。
不管进程是正常退出还是异常死亡，不管用户态有没有显式close(fd)，最后内核都会清理资源，其中对文件会调用filp_close，回调到file_operations::flush和file_operations::release。
当调用到/dev/binder对应的filp_close时，binder driver使用workqueue机制完成后续的binder death notification和binder占用资源清理。
当release时，在deferred work中会逐个对binder_node::refs上每个binder_ref设置ref->node->proc == NULL表示该binder_ref引用的binder_node已经是个DEAD_NODE。并且DEAD binder会被从rb-tree上摘下来，放到binder_dead_nodes hlist上。binder_deferred_release(…)会对binder_threads, incoming_refs, outgoing_refs, binder_transaction, binder_buffer, memory page, task_struct, binder_proc进行清理，在此情景中，只关心incoming_refs即别的进程中引用到退出进程的binder_nodes的binder_refs.
以处理过程中binder_ref的death->work的binder_work类型为状态，得到以下状态转换图。
1是在binder node died之后另一进程发送BC_REQUEST_DEATH_NOTIFICATION，或者先发送BC_REQUEST_DEATH_NOTIFICATION却由于不同进程的异步引起在binder node died之后执行，这时也会引起BINDER_WORK_DEAD_BINDER，因为用户层认为自己已经注册了通知。
2是已经BC_REQUEST_DEATH_NOTIFICATION，然后binder node died，这是大多数正常情况下binder_node死亡通知，引起BINDER_WORK_DEAD_BINDER。
路径3、4、OVER是绝大多数情况下用户层不发送BC_CLEAR_DEATH_NOTIFICATION的binder_node death notification处理流程。
3在BINDER_WORK_DEAD_BINDER工作处理前，因为进程异步，可能会收到客户端的BC_CLEAR_DEATH_NOTIFICATION，这时设置binder_ref::death:: work状态为BINDER_WORK_DEAD_BINDER_AND_CLEAR，表示binder_node已死并且需要清death notification，执行路径5；则首先发出BR_DEAD_BINDER，在收到客户端用户层处理讣告完毕后沿路径8、9处理完毕。
3在发出BR_DEAD_BINDER后，在用户层的death notification处理中，用户层也会发送BC_CLEAR_DEATH_NOTIFICATION（见BpBinder::sendObituary()代码），这时执行路径7，在收到客户端用户层处理讣告完毕后沿路径8、9处理完毕。
以上是binder driver中的控制流程，用户层对BR_DEAD_BINDER的处理代码如下。
case BR_DEAD_BINDER:
{
BpBinder *proxy = (BpBinder*)mIn.readInt32();
proxy->sendObituary();
mOut.writeInt32(BC_DEAD_BINDER_DONE);
mOut.writeInt32((int32_t)proxy);
} break;
case BR_CLEAR_DEATH_NOTIFICATION_DONE:
{
BpBinder *proxy = (BpBinder*)mIn.readInt32();
proxy->getWeakRefs()->decWeak(proxy);
} break;
void BpBinder::sendObituary()
{
mAlive = 0;
if (mObitsSent) return;
mLock.lock();
Vector
if(obits != NULL) {
ALOGV("Clearing sent death notification: %p handle %d\n", this, mHandle);
IPCThreadState* self = IPCThreadState::self();
self->clearDeathNotification(mHandle, this);
self->flushCommands();
mObituaries = NULL;
}
mObitsSent = 1;
mLock.unlock();
ALOGV("Reporting death of proxy %p for %d recipients\n", this, obits ? obits->size() : 0);
if (obits != NULL) {
const size_t N = obits->size();
for (size_t i=0; i
reportOneDeath(obits->itemAt(i));
}
delete obits;
}
}
BpBinder::sendObituary()是线程安全的。
void BpBinder::reportOneDeath(const Obituary& obit)
{
sp
ALOGV("Reporting death to recipient: %p\n", recipient.get());
if (recipient == NULL) return;
recipient->binderDied(this);
}
binderDied()是接口IBinder::DeathRecipient的方法。
class IBinder::DeathRecipient : publicvirtual RefBase
{
public:
virtual voidbinderDied(const wp
};
参数为wp
注意：所谓的死亡通知对于client来说并非立刻异步被通知到。在driver中已经放到client进程的队列中，但是如果client进程此刻没有阻塞在ioctl上，那么是下次调用ioctl的时候处理死亡通知。
从上图中，可以考察service端先死掉的情形和client端先死掉的情形。
情景5. cross-call transaction
主要就是transaction_stack这个栈式结构，引用universus的<
在请求者把binder_work挂到target_proc->todo前，把binder_transaction挂到自己的transaction_stack上；需要注意的是，只有当工作线程从进程todo上取下该binder_work处理时，才把该binder_transaction挂到线程自己的transaction_stack上并设置to_thread，这个处理节点以后从请求者线程的transaction_stack上的binder_transaction::to_thread考察target_thread才有意义，否则其值是NULL，表示虽然挂到target_proc->todo上，但还没有得到工作线程处理。
(引用universus的原话)当进程P1的线程T1向进程P2发送请求时，驱动会先查看一下线程T1是否也正在处理来自P2某个线程请求但尚未完成（没有发送回复）。这种情况通常发生在两个进程都有Binder实体并互相对发时请求时。假如驱动在进程P2中发现了这样的线程，比如说T2，就会要求T2来处理T1的这次请求。因为T2既然向T1发送了请求尚未得到返回包，说明T2肯定（或将会）阻塞在读取返回包的状态。 这时候可以让T2顺便做点事情，总比等在那里闲着好。
这个场景之下，binder驱动查看T1是否正在处理来自P2某线程的请求但尚未完成的工作是在binder_transaction()函数中完成的，也就是前文2.3节中提到过，但是没有分析，现在又了前面这些分析，应该对binder_transaction、binder_buffer结构体，特别是binder_transaction.transaction_stack这个链表有所理解，这个时候来分析这个优化才是最佳时机。
如上面的情景：本来是T1向P2发送的同步请求(交互1)，ioctl运行到这里，而如果在这之前，T1也正在处理来自P2进程T2的请求还没结束(交互2)，那么交互2中接收方T1也就是交互1的发送方了。其实这里完全可以不用优化，直接将交互1的请求投进P2进程的全局todo队列进行处理，但是为了提高效率和节省资源，反正交互2的发送方T2也在等待接收方的回复，也没有做事，这个时候让其做点事情岂不是更好，所以出于这样的考虑，才有了如下的优化过程。
交互2中发送方T2和接收方T1的binder_thread.transaction_stack都应该是指向同一个binder_transaction结构体，而交互1中的发送方T1也就是交互2中的接收方。
if (!(tr->flags & TF_ONE_WAY) && thread->transaction_stack) {
struct binder_transaction *tmp;
tmp = thread->transaction_stack;
if (tmp->to_thread != thread) {
… // 需要当前task是前次同步传输中的接收方才能进行发送优化
return_error = BR_FAILED_REPLY;
goto err_bad_call_stack;
}
while (tmp) { /* 如果此时T1在这之前，不仅仅只是接收到T2的请求，而
且还接收到了其他很多线程的请求均没有完成的话，那么这个transaction_stack链表中就有多个binder_transaction结构体存在，所以需要查找。查找的时候第一个条件都满足，只是第二个条件就不好满足了，就要一个一个比较，需要交互1的目标binder_proc匹配。 */
if (tmp->from && tmp->from->proc == target_proc)
target_thread = tmp->from;
// 找到之后，直接将交互2的发送者设置成交互1的接收者。
tmp = tmp->from_parent; // 否则，继续查找链表
}
} // if (!(tr->flags & TF_ONE_WAY) && thread->transaction_stack)
} // else
if (target_thread) { // 如果这个优化有结果，找到了对应的线程
e->to_thread = target_thread->pid;
target_list = &target_thread->todo; // 目标任务列表
target_wait = &target_thread->wait; // 目标线程等待队列
} else {// 优化不成功的话，将任务送到进程对于的全局任务队列
target_list = &target_proc->todo;
target_wait = &target_proc->wait;
}
…
} // binder_transaction()