Android系统进程间通信（IPC）机制Binder中的Server启动过程源代码分析

Android系统进程间通信（IPC）机制Binder中的Server启动过程源代码分析
文章转载至CSDN社区罗升阳的安卓之旅，原文地址：http://blog.csdn.net/luoshengyang/article/details/6629298

在前面一篇文章浅谈Android系统进程间通信（IPC）机制Binder中的Server和Client获得Service Manager接口之路中，介绍了在Android系统中Binder进程间通信机制中的Server角色是如何获得Service Manager远程接口的，即defaultServiceManager函数的实现。Server获得了Service Manager远程接口之后，就要把自己的Service添加到Service Manager中去，然后把自己启动起来，等待Client的请求。本文将通过分析源代码了解Server的启动过程是怎么样的。

   本文通过一个具体的例子来说明Binder机制中Server的启动过程。我们知道，在Android系统中，提供了多媒体播放的功能，这个功能是以服务的形式来提供的。这里，我们就通过分析MediaPlayerService的实现来了解Media Server的启动过程。

   首先，看一下MediaPlayerService的类图，以便我们理解下面要描述的内容。

   我们将要介绍的主角MediaPlayerService继承于BnMediaPlayerService类，熟悉Binder机制的同学应该知道 BnMediaPlayerService是一个Binder Native类，用来处理Client请求的。BnMediaPlayerService继承于 BnInterface<IMediaPlayerService>类，BnInterface是一个模板类，它定义在 frameworks/base/include/binder/IInterface.h文件中：
[cpp] view plain copy
1. template<typename INTERFACE>
2. class BnInterface : public INTERFACE, public BBinder
3. {
4. public:
5. virtual sp<IInterface> queryLocalInterface(const String16& _descriptor);
6. virtual const String16& getInterfaceDescriptor() const;
8. protected:
9. virtual IBinder* onAsBinder();
10. };
   这里可以看出，BnMediaPlayerService实际是继承了IMediaPlayerService和BBinder类。 IMediaPlayerService和BBinder类又分别继承了IInterface和IBinder类，IInterface和IBinder 类又同时继承了RefBase类。

   实际上，BnMediaPlayerService并不是直接接收到Client处发送过来的请求，而是使用了IPCThreadState接收 Client处发送过来的请求，而IPCThreadState又借助了ProcessState类来与Binder驱动程序交互。有关 IPCThreadState和ProcessState的关系，可以参考上一篇文章浅谈Android系统进程间通信（IPC）机制Binder中的Server和Client获得Service Manager接口之路，接下来也会有相应的描述。IPCThreadState接收到了Client处的请求后，就会调用BBinder类的transact函数，并传入相关参数，BBinder类的transact函数最终调用BnMediaPlayerService类的onTransact函数，于是，就开始真正地处理 Client的请求了。

   了解了MediaPlayerService类结构之后，就要开始进入到本文的主题了。

   首先，看看MediaPlayerService是如何启动的。启动MediaPlayerService的代码位于frameworks/base/media/mediaserver/main_mediaserver.cpp文件中：
[cpp] view plain copy
1. int main(int argc, char** argv)
2. {
3. sp<ProcessState> proc(ProcessState::self());
4. sp<IServiceManager> sm = defaultServiceManager();
5. LOGI("ServiceManager: %p", sm.get());
6. AudioFlinger::instantiate();
7. MediaPlayerService::instantiate();
8. CameraService::instantiate();
9. AudioPolicyService::instantiate();
10. ProcessState::self()->startThreadPool();
11. IPCThreadState::self()->joinThreadPool();
12. }
这里我们不关注AudioFlinger和CameraService相关的代码。

先看下面这句代码：
[cpp] view plain copy
1. sp<ProcessState> proc(ProcessState::self());
这句代码的作用是通过ProcessState::self()调用创建一个ProcessState实例。ProcessState::self()是 ProcessState类的一个静态成员变量，定义在frameworks/base/libs/binder/ProcessState.cpp文件中：
[cpp] view plain copy
1. sp<ProcessState> ProcessState::self()
2. {
3. if (gProcess != NULL) return gProcess;
5. AutoMutex _l(gProcessMutex);
6. if (gProcess == NULL) gProcess = new ProcessState;
7. return gProcess;
8. }
这里可以看出，这个函数作用是返回一个全局唯一的ProcessState实例gProcess。全局唯一实例变量gProcess定义在frameworks/base/libs/binder/Static.cpp文件中：
[cpp] view plain copy
1. Mutex gProcessMutex;
2. sp<ProcessState> gProcess;
再来看ProcessState的构造函数：
[cpp] view plain copy
1. ProcessState::ProcessState()
2. : mDriverFD(open_driver())
3. , mVMStart(MAP_FAILED)
4. , mManagesContexts(false)
5. , mBinderContextCheckFunc(NULL)
6. , mBinderContextUserData(NULL)
7. , mThreadPoolStarted(false)
8. , mThreadPoolSeq(1)
9. {
10. if (mDriverFD >= 0) {
11. // XXX Ideally, there should be a specific define for whether we
12. // have mmap (or whether we could possibly have the kernel module
13. // availabla).
14. #if !defined(HAVE_WIN32_IPC)
15. // mmap the binder, providing a chunk of virtual address space to receive transactions.
16. mVMStart = mmap(0, BINDER_VM_SIZE, PROT_READ, MAP_PRIVATE | MAP_NORESERVE, mDriverFD, 0);
17. if (mVMStart == MAP_FAILED) {
18. // *sigh*
19. LOGE("Using /dev/binder failed: unable to mmap transaction memory. ");
20. close(mDriverFD);
21. mDriverFD = -1;
22. }
23. #else
24. mDriverFD = -1;
25. #endif
26. }
27. if (mDriverFD < 0) {
28. // Need to run without the driver, starting our own thread pool.
29. }
30. }
这个函数有两个关键地方，一是通过open_driver函数打开Binder设备文件/dev/binder，并将打开设备文件描述符保存在成员变量mDriverFD中；二是通过mmap来把设备文件/dev/binder映射到内存中。

先看open_driver函数的实现，这个函数同样位于frameworks/base/libs/binder/ProcessState.cpp文件中：
[cpp] view plain copy
1. static int open_driver()
2. {
3. if (gSingleProcess) {
4. return -1;
5. }
7. int fd = open("/dev/binder", O_RDWR);
8. if (fd >= 0) {
9. fcntl(fd, F_SETFD, FD_CLOEXEC);
10. int vers;
11. #if defined(HAVE_ANDROID_OS)
12. status_t result = ioctl(fd, BINDER_VERSION, &vers);
13. #else
14. status_t result = -1;
15. errno = EPERM;
16. #endif
17. if (result == -1) {
18. LOGE("Binder ioctl to obtain version failed: %s", strerror(errno));
19. close(fd);
20. fd = -1;
21. }
22. if (result != 0 || vers != BINDER_CURRENT_PROTOCOL_VERSION) {
23. LOGE("Binder driver protocol does not match user space protocol!");
24. close(fd);
25. fd = -1;
26. }
27. #if defined(HAVE_ANDROID_OS)
28. size_t maxThreads = 15;
29. result = ioctl(fd, BINDER_SET_MAX_THREADS, &maxThreads);
30. if (result == -1) {
31. LOGE("Binder ioctl to set max threads failed: %s", strerror(errno));
32. }
33. #endif
35. } else {
36. LOGW("Opening '/dev/binder' failed: %s ", strerror(errno));
37. }
38. return fd;
39. }
   这个函数的作用主要是通过open文件操作函数来打开/dev/binder设备文件，然后再调用ioctl文件控制函数来分别执行 BINDER_VERSION和BINDER_SET_MAX_THREADS两个命令来和Binder驱动程序进行交互，前者用于获得当前Binder 驱动程序的版本号，后者用于通知Binder驱动程序，MediaPlayerService最多可同时启动15个线程来处理Client端的请求。

   open在Binder驱动程序中的具体实现，请参考前面一篇文章浅谈Service Manager成为Android进程间通信（IPC）机制Binder守护进程之路，这里不再重复描述。打开/dev/binder设备文件后，Binder驱动程序就为MediaPlayerService进程创建了一个struct binder_proc结构体实例来维护MediaPlayerService进程上下文相关信息。

   我们来看一下ioctl文件操作函数执行BINDER_VERSION命令的过程：
[cpp] view plain copy
1. status_t result = ioctl(fd, BINDER_VERSION, &vers);
这个函数调用最终进入到Binder驱动程序的binder_ioctl函数中，我们只关注BINDER_VERSION相关的部分逻辑：
[cpp] view plain copy
1. static long binder_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
2. {
3. int ret;
4. struct binder_proc *proc = filp->private_data;
5. struct binder_thread *thread;
6. unsigned int size = _IOC_SIZE(cmd);
7. void __user *ubuf = (void __user *)arg;
9. /*printk(KERN_INFO "binder_ioctl: %d:%d %x %lx ", proc->pid, current->pid, cmd, arg);*/
11. ret = wait_event_interruptible(binder_user_error_wait, binder_stop_on_user_error < 2);
12. if (ret)
13. return ret;
15. mutex_lock(&binder_lock);
16. thread = binder_get_thread(proc);
17. if (thread == NULL) {
18. ret = -ENOMEM;
19. goto err;
20. }
22. switch (cmd) {
23. ......
24. case BINDER_VERSION:
25. if (size != sizeof(struct binder_version)) {
26. ret = -EINVAL;
27. goto err;
28. }
29. if (put_user(BINDER_CURRENT_PROTOCOL_VERSION, &((struct binder_version *)ubuf)->protocol_version)) {
30. ret = -EINVAL;
31. goto err;
32. }
33. break;
34. ......
35. }
36. ret = 0;
37. err:
38. ......
39. return ret;
40. }
很简单，只是将BINDER_CURRENT_PROTOCOL_VERSION写入到传入的参数arg指向的用户缓冲区中去就返回了。 BINDER_CURRENT_PROTOCOL_VERSION是一个宏，定义在kernel/common/drivers/staging /android/binder.h文件中：
[cpp] view plain copy
1. /* This is the current protocol version. */
2. #define BINDER_CURRENT_PROTOCOL_VERSION 7
这里为什么要把ubuf转换成struct binder_version之后，再通过其protocol_version成员变量再来写入呢，转了一圈，最终内容还是写入到ubuf中。我们看一下 struct binder_version的定义就会明白，同样是在kernel/common/drivers/staging/android/binder.h 文件中：
[cpp] view plain copy
1. /* Use with BINDER_VERSION, driver fills in fields. */
2. struct binder_version {
3. /* driver protocol version -- increment with incompatible change */
4. signed long protocol_version;
5. };
   从注释中可以看出来，这里是考虑到兼容性，因为以后很有可能不是用signed long来表示版本号。

   这里有一个重要的地方要注意的是，由于这里是打开设备文件/dev/binder之后，第一次进入到binder_ioctl函数，因此，这里调用 binder_get_thread的时候，就会为当前线程创建一个struct binder_thread结构体变量来维护线程上下文信息，具体可以参考浅谈Service Manager成为Android进程间通信（IPC）机制Binder守护进程之路一文。

   接着我们再来看一下ioctl文件操作函数执行BINDER_SET_MAX_THREADS命令的过程：
[cpp] view plain copy
1. result = ioctl(fd, BINDER_SET_MAX_THREADS, &maxThreads);
这个函数调用最终进入到Binder驱动程序的binder_ioctl函数中，我们只关注BINDER_SET_MAX_THREADS相关的部分逻辑：
[cpp] view plain copy
1. static long binder_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
2. {
3. int ret;
4. struct binder_proc *proc = filp->private_data;
5. struct binder_thread *thread;
6. unsigned int size = _IOC_SIZE(cmd);
7. void __user *ubuf = (void __user *)arg;
9. /*printk(KERN_INFO "binder_ioctl: %d:%d %x %lx ", proc->pid, current->pid, cmd, arg);*/
11. ret = wait_event_interruptible(binder_user_error_wait, binder_stop_on_user_error < 2);
12. if (ret)
13. return ret;
15. mutex_lock(&binder_lock);
16. thread = binder_get_thread(proc);
17. if (thread == NULL) {
18. ret = -ENOMEM;
19. goto err;
20. }
22. switch (cmd) {
23. ......
24. case BINDER_SET_MAX_THREADS:
25. if (copy_from_user(&proc->max_threads, ubuf, sizeof(proc->max_threads))) {
26. ret = -EINVAL;
27. goto err;
28. }
29. break;
30. ......
31. }
32. ret = 0;
33. err:
34. ......
35. return ret;
36. }
这里实现也是非常简单，只是简单地把用户传进来的参数保存在proc->max_threads中就完毕了。注意，这里再调用 binder_get_thread函数的时候，就可以在proc->threads中找到当前线程对应的struct binder_thread结构了，因为前面已经创建好并保存在proc->threads红黑树中。

回到ProcessState的构造函数中，这里还通过mmap函数来把设备文件/dev/binder映射到内存中，这个函数在浅谈Service Manager成为Android进程间通信（IPC）机制Binder守护进程之路一文也已经有详细介绍，这里不再重复描述。宏BINDER_VM_SIZE就定义在ProcessState.cpp文件中：
[cpp] view plain copy
1. #define BINDER_VM_SIZE ((1*1024*1024) - (4096 *2))
   mmap函数调用完成之后，Binder驱动程序就为当前进程预留了BINDER_VM_SIZE大小的内存空间了。

   这样，ProcessState全局唯一变量gProcess就创建完毕了，回到frameworks/base/media/mediaserver /main_mediaserver.cpp文件中的main函数，下一步是调用defaultServiceManager函数来获得Service Manager的远程接口，这个已经在上一篇文章浅谈Android系统进程间通信（IPC）机制Binder中的Server和Client获得Service Manager接口之路有详细描述，读者可以回过头去参考一下。

   再接下来，就进入到MediaPlayerService::instantiate函数把MediaPlayerService添加到Service Manger中去了。这个函数定义在frameworks/base/media/libmediaplayerservice /MediaPlayerService.cpp文件中：
[cpp] view plain copy
1. void MediaPlayerService::instantiate() {
2. defaultServiceManager()->addService(
3. String16("media.player"), new MediaPlayerService());
4. }
我们重点看一下IServiceManger::addService的过程，这有助于我们加深对Binder机制的理解。

在上一篇文章浅谈Android系统进程间通信（IPC）机制Binder中的Server和Client获得Service Manager接口之路中说到，defaultServiceManager返回的实际是一个BpServiceManger类实例，因此，我们看一下 BpServiceManger::addService的实现，这个函数实现在frameworks/base/libs/binder /IServiceManager.cpp文件中：
[cpp] view plain copy
1. class BpServiceManager : public BpInterface<IServiceManager>
2. {
3. public:
4. BpServiceManager(const sp<IBinder>& impl)
5. : BpInterface<IServiceManager>(impl)
6. {
7. }
9. ......
11. virtual status_t addService(const String16& name, const sp<IBinder>& service)
12. {
13. Parcel data, reply;
14. data.writeInterfaceToken(IServiceManager::getInterfaceDescriptor());
15. data.writeString16(name);
16. data.writeStrongBinder(service);
17. status_t err = remote()->transact(ADD_SERVICE_TRANSACTION, data, &reply);
18. return err == NO_ERROR ? reply.readExceptionCode()
19. }
21. ......
23. };
这里的Parcel类是用来于序列化进程间通信数据用的。

先来看这一句的调用：
[cpp] view plain copy
1. data.writeInterfaceToken(IServiceManager::getInterfaceDescriptor());
IServiceManager::getInterfaceDescriptor()返回来的是一个字符串，即"android.os.IServiceManager"，具体可以参考IServiceManger的实现。我们看一下 Parcel::writeInterfaceToken的实现，位于frameworks/base/libs/binder/Parcel.cpp文件中：
[cpp] view plain copy
1. // Write RPC headers. (previously just the interface token)
2. status_t Parcel::writeInterfaceToken(const String16& interface)
3. {
4. writeInt32(IPCThreadState::self()->getStrictModePolicy() |
5. STRICT_MODE_PENALTY_GATHER);
6. // currently the interface identification token is just its name as a string
7. return writeString16(interface);
8. }
它的作用是写入一个整数和一个字符串到Parcel中去。

再来看下面的调用：
[cpp] view plain copy
1. data.writeString16(name);
这里又是写入一个字符串到Parcel中去，这里的name即是上面传进来的“media.player”字符串。

往下看：
[cpp] view plain copy
1. data.writeStrongBinder(service);
这里定入一个Binder对象到Parcel去。我们重点看一下这个函数的实现，因为它涉及到进程间传输Binder实体的问题，比较复杂，需要重点关注，同时，也是理解Binder机制的一个重点所在。注意，这里的service参数是一个MediaPlayerService对象。
[cpp] view plain copy
1. status_t Parcel::writeStrongBinder(const sp<IBinder>& val)
2. {
3. return flatten_binder(ProcessState::self(), val, this);
4. }
看到flatten_binder函数，是不是似曾相识的感觉？我们在前面一篇文章浅谈Service Manager成为Android进程间通信（IPC）机制Binder守护进程之路中，曾经提到在Binder驱动程序中，使用struct flat_binder_object来表示传输中的一个binder对象，它的定义如下所示：
[cpp] view plain copy
1. /*
2. * This is the flattened representation of a Binder object for transfer
3. * between processes. The 'offsets' supplied as part of a binder transaction
4. * contains offsets into the data where these structures occur. The Binder
5. * driver takes care of re-writing the structure type and data as it moves
6. * between processes.
7. */
8. struct flat_binder_object {
9. /* 8 bytes for large_flat_header. */
10. unsigned long type;
11. unsigned long flags;
13. /* 8 bytes of data. */
14. union {
15. void *binder; /* local object */
16. signed long handle; /* remote object */
17. };
19. /* extra data associated with local object */
20. void *cookie;
21. };
各个成员变量的含义请参考资料Android Binder设计与实现。

我们进入到flatten_binder函数看看：
[cpp] view plain copy
1. status_t flatten_binder(const sp<ProcessState>& proc,
2. const sp<IBinder>& binder, Parcel* out)
3. {
4. flat_binder_object obj;
6. obj.flags = 0x7f | FLAT_BINDER_FLAG_ACCEPTS_FDS;
7. if (binder != NULL) {
8. IBinder *local = binder->localBinder();
9. if (!local) {
10. BpBinder *proxy = binder->remoteBinder();
11. if (proxy == NULL) {
12. LOGE("null proxy");
13. }
14. const int32_t handle = proxy ? proxy->handle() : 0;
15. obj.type = BINDER_TYPE_HANDLE;
16. obj.handle = handle;
17. obj.cookie = NULL;
18. } else {
19. obj.type = BINDER_TYPE_BINDER;
20. obj.binder = local->getWeakRefs();
21. obj.cookie = local;
22. }
23. } else {
24. obj.type = BINDER_TYPE_BINDER;
25. obj.binder = NULL;
26. obj.cookie = NULL;
27. }
29. return finish_flatten_binder(binder, obj, out);
30. }
首先是初始化flat_binder_object的flags域：
[cpp] view plain copy
1. obj.flags = 0x7f | FLAT_BINDER_FLAG_ACCEPTS_FDS;
0x7f表示处理本Binder实体请求数据包的线程的最低优先级，FLAT_BINDER_FLAG_ACCEPTS_FDS表示这个Binder实体可以接受文件描述符，Binder实体在收到文件描述符时，就会在本进程中打开这个文件。

传进来的binder即为MediaPlayerService::instantiate函数中new出来的MediaPlayerService实例，因此，不为空。又由于MediaPlayerService继承自BBinder类，它是一个本地Binder实体，因此 binder->localBinder返回一个BBinder指针，而且肯定不为空，于是执行下面语句：
[cpp] view plain copy
1. obj.type = BINDER_TYPE_BINDER;
2. obj.binder = local->getWeakRefs();
3. obj.cookie = local;
设置了flat_binder_obj的其他成员变量，注意，指向这个Binder实体地址的指针local保存在flat_binder_obj的成员变量cookie中。

函数调用finish_flatten_binder来将这个flat_binder_obj写入到Parcel中去：
[cpp] view plain copy
1. inline static status_t finish_flatten_binder(
2. const sp<IBinder>& binder, const flat_binder_object& flat, Parcel* out)
3. {
4. return out->writeObject(flat, false);
5. }
Parcel::writeObject的实现如下：
[cpp] view plain copy
1. status_t Parcel::writeObject(const flat_binder_object& val, bool nullMetaData)
2. {
3. const bool enoughData = (mDataPos+sizeof(val)) <= mDataCapacity;
4. const bool enoughObjects = mObjectsSize < mObjectsCapacity;
5. if (enoughData && enoughObjects) {
6. restart_write:
7. *reinterpret_cast<flat_binder_object*>(mData+mDataPos) = val;
9. // Need to write meta-data?
10. if (nullMetaData || val.binder != NULL) {
11. mObjects[mObjectsSize] = mDataPos;
12. acquire_object(ProcessState::self(), val, this);
13. mObjectsSize++;
14. }
16. // remember if it's a file descriptor
17. if (val.type == BINDER_TYPE_FD) {
18. mHasFds = mFdsKnown = true;
19. }
21. return finishWrite(sizeof(flat_binder_object));
22. }
24. if (!enoughData) {
25. const status_t err = growData(sizeof(val));
26. if (err != NO_ERROR) return err;
27. }
28. if (!enoughObjects) {
29. size_t newSize = ((mObjectsSize+2)*3)/2;
30. size_t* objects = (size_t*)realloc(mObjects, newSize*sizeof(size_t));
31. if (objects == NULL) return NO_MEMORY;
32. mObjects = objects;
33. mObjectsCapacity = newSize;
34. }
36. goto restart_write;
37. }
这里除了把flat_binder_obj写到Parcel里面之内，还要记录这个flat_binder_obj在Parcel里面的偏移位置：
[cpp] view plain copy
1. mObjects[mObjectsSize] = mDataPos;
这里因为，如果进程间传输的数据间带有Binder对象的时候，Binder驱动程序需要作进一步的处理，以维护各个Binder实体的一致性，下面我们将会看到Binder驱动程序是怎么处理这些Binder对象的。

再回到BpServiceManager::addService函数中，调用下面语句：
[cpp] view plain copy
1. status_t err = remote()->transact(ADD_SERVICE_TRANSACTION, data, &reply);
回到浅谈Android系统进程间通信（IPC）机制Binder中的Server和Client获得Service Manager接口之路一文中的类图中去看一下，这里的remote成员函数来自于BpRefBase类，它返回一个BpBinder指针。因此，我们继续进入到BpBinder::transact函数中去看看：
[cpp] view plain copy
1. status_t BpBinder::transact(
2. uint32_t code, const Parcel& data, Parcel* reply, uint32_t flags)
3. {
4. // Once a binder has died, it will never come back to life.
5. if (mAlive) {
6. status_t status = IPCThreadState::self()->transact(
7. mHandle, code, data, reply, flags);
8. if (status == DEAD_OBJECT) mAlive = 0;
9. return status;
10. }
12. return DEAD_OBJECT;
13. }
这里又调用了IPCThreadState::transact进执行实际的操作。注意，这里的mHandle为0，code为 ADD_SERVICE_TRANSACTION。ADD_SERVICE_TRANSACTION是上面以参数形式传进来的，那mHandle为什么是 0呢？因为这里表示的是Service Manager远程接口，它的句柄值一定是0，具体请参考浅谈Android系统进程间通信（IPC）机制Binder中的Server和Client获得Service Manager接口之路一文。
再进入到IPCThreadState::transact函数，看看做了些什么事情：
[cpp] view plain copy
1. status_t IPCThreadState::transact(int32_t handle,
2. uint32_t code, const Parcel& data,
3. Parcel* reply, uint32_t flags)
4. {
5. status_t err = data.errorCheck();
7. flags |= TF_ACCEPT_FDS;
9. IF_LOG_TRANSACTIONS() {
10. TextOutput::Bundle _b(alog);
11. alog << "BC_TRANSACTION thr " << (void*)pthread_self() << " / hand "
12. << handle << " / code " << TypeCode(code) << ": "
13. << indent << data << dedent << endl;
14. }
16. if (err == NO_ERROR) {
17. LOG_ONEWAY(">>>> SEND from pid %d uid %d %s", getpid(), getuid(),
18. (flags & TF_ONE_WAY) == 0 ? "READ REPLY" : "ONE WAY");
19. err = writeTransactionData(BC_TRANSACTION, flags, handle, code, data, NULL);
20. }
22. if (err != NO_ERROR) {
23. if (reply) reply->setError(err);
24. return (mLastError = err);
25. }
27. if ((flags & TF_ONE_WAY) == 0) {
28. #if 0
29. if (code == 4) { // relayout
30. LOGI(">>>>>> CALLING transaction 4");
31. } else {
32. LOGI(">>>>>> CALLING transaction %d", code);
33. }
34. #endif
35. if (reply) {
36. err = waitForResponse(reply);
37. } else {
38. Parcel fakeReply;
39. err = waitForResponse(&fakeReply);
40. }
41. #if 0
42. if (code == 4) { // relayout
43. LOGI("<<<<<< RETURNING transaction 4");
44. } else {
45. LOGI("<<<<<< RETURNING transaction %d", code);
46. }
47. #endif
49. IF_LOG_TRANSACTIONS() {
50. TextOutput::Bundle _b(alog);
51. alog << "BR_REPLY thr " << (void*)pthread_self() << " / hand "
52. << handle << ": ";
53. if (reply) alog << indent << *reply << dedent << endl;
54. else alog << "(none requested)" << endl;
55. }
56. } else {
57. err = waitForResponse(NULL, NULL);
58. }
60. return err;
61. }
IPCThreadState::transact函数的参数flags是一个默认值为0的参数，上面没有传相应的实参进来，因此，这里就为0。

函数首先调用writeTransactionData函数准备好一个struct binder_transaction_data结构体变量，这个是等一下要传输给Binder驱动程序的。struct binder_transaction_data的定义我们在浅谈Service Manager成为Android进程间通信（IPC）机制Binder守护进程之路一文中有详细描述，读者不妨回过去读一下。这里为了方便描述，将struct binder_transaction_data的定义再次列出来：
[cpp] view plain copy
1. struct binder_transaction_data {
2. /* The first two are only used for bcTRANSACTION and brTRANSACTION,
3. * identifying the target and contents of the transaction.
4. */
5. union {
6. size_t handle; /* target descriptor of command transaction */
7. void *ptr; /* target descriptor of return transaction */
8. } target;
9. void *cookie; /* target object cookie */
10. unsigned int code; /* transaction command */
12. /* General information about the transaction. */
13. unsigned int flags;
14. pid_t sender_pid;
15. uid_t sender_euid;
16. size_t data_size; /* number of bytes of data */
17. size_t offsets_size; /* number of bytes of offsets */
19. /* If this transaction is inline, the data immediately
20. * follows here; otherwise, it ends with a pointer to
21. * the data buffer.
22. */
23. union {
24. struct {
25. /* transaction data */
26. const void *buffer;
27. /* offsets from buffer to flat_binder_object structs */
28. const void *offsets;
29. } ptr;
30. uint8_t buf[8];
31. } data;
32. };
writeTransactionData函数的实现如下：
[cpp] view plain copy
1. status_t IPCThreadState::writeTransactionData(int32_t cmd, uint32_t binderFlags,
2. int32_t handle, uint32_t code, const Parcel& data, status_t* statusBuffer)
3. {
4. binder_transaction_data tr;
6. tr.target.handle = handle;
7. tr.code = code;
8. tr.flags = binderFlags;
10. const status_t err = data.errorCheck();
11. if (err == NO_ERROR) {
12. tr.data_size = data.ipcDataSize();
13. tr.data.ptr.buffer = data.ipcData();
14. tr.offsets_size = data.ipcObjectsCount()*sizeof(size_t);
15. tr.data.ptr.offsets = data.ipcObjects();
16. } else if (statusBuffer) {
17. tr.flags |= TF_STATUS_CODE;
18. *statusBuffer = err;
19. tr.data_size = sizeof(status_t);
20. tr.data.ptr.buffer = statusBuffer;
21. tr.offsets_size = 0;
22. tr.data.ptr.offsets = NULL;
23. } else {
24. return (mLastError = err);
25. }
27. mOut.writeInt32(cmd);
28. mOut.write(&tr, sizeof(tr));
30. return NO_ERROR;
31. }
注意，这里的cmd为BC_TRANSACTION。这个函数很简单，在这个场景下，就是执行下面语句来初始化本地变量tr：
[cpp] view plain copy
1. tr.data_size = data.ipcDataSize();
2. tr.data.ptr.buffer = data.ipcData();
3. tr.offsets_size = data.ipcObjectsCount()*sizeof(size_t);
4. tr.data.ptr.offsets = data.ipcObjects();
回忆一下上面的内容，写入到tr.data.ptr.buffer的内容相当于下面的内容：
[cpp] view plain copy
1. writeInt32(IPCThreadState::self()->getStrictModePolicy() |
2. STRICT_MODE_PENALTY_GATHER);
3. writeString16("android.os.IServiceManager");
4. writeString16("media.player");
5. writeStrongBinder(new MediaPlayerService());
其中包含了一个Binder实体MediaPlayerService，因此需要设置tr.offsets_size就为 1，tr.data.ptr.offsets就指向了这个MediaPlayerService的地址在tr.data.ptr.buffer中的偏移量。最后，将tr的内容保存在IPCThreadState的成员变量mOut中。
回到IPCThreadState::transact函数中，接下去看，(flags & TF_ONE_WAY) == 0为true，并且reply不为空，所以最终进入到waitForResponse(reply)这条路径来。我们看一下 waitForResponse函数的实现：
[cpp] view plain copy
1. status_t IPCThreadState::waitForResponse(Parcel *reply, status_t *acquireResult)
2. {
3. int32_t cmd;
4. int32_t err;
6. while (1) {
7. if ((err=talkWithDriver()) < NO_ERROR) break;
8. err = mIn.errorCheck();
9. if (err < NO_ERROR) break;
10. if (mIn.dataAvail() == 0) continue;
12. cmd = mIn.readInt32();
14. IF_LOG_COMMANDS() {
15. alog << "Processing waitForResponse Command: "
16. << getReturnString(cmd) << endl;
17. }
19. switch (cmd) {
20. case BR_TRANSACTION_COMPLETE:
21. if (!reply && !acquireResult) goto finish;
22. break;
24. case BR_DEAD_REPLY:
25. err = DEAD_OBJECT;
26. goto finish;
28. case BR_FAILED_REPLY:
29. err = FAILED_TRANSACTION;
30. goto finish;
32. case BR_ACQUIRE_RESULT:
33. {
34. LOG_ASSERT(acquireResult != NULL, "Unexpected brACQUIRE_RESULT");
35. const int32_t result = mIn.readInt32();
36. if (!acquireResult) continue;
37. *acquireResult = result ? NO_ERROR : INVALID_OPERATION;
38. }
39. goto finish;
41. case BR_REPLY:
42. {
43. binder_transaction_data tr;
44. err = mIn.read(&tr, sizeof(tr));
45. LOG_ASSERT(err == NO_ERROR, "Not enough command data for brREPLY");
46. if (err != NO_ERROR) goto finish;
48. if (reply) {
49. if ((tr.flags & TF_STATUS_CODE) == 0) {
50. reply->ipcSetDataReference(
51. reinterpret_cast<const uint8_t*>(tr.data.ptr.buffer),
52. tr.data_size,
53. reinterpret_cast<const size_t*>(tr.data.ptr.offsets),
54. tr.offsets_size/sizeof(size_t),
55. freeBuffer, this);
56. } else {
57. err = *static_cast<const status_t*>(tr.data.ptr.buffer);
58. freeBuffer(NULL,
59. reinterpret_cast<const uint8_t*>(tr.data.ptr.buffer),
60. tr.data_size,
61. reinterpret_cast<const size_t*>(tr.data.ptr.offsets),
62. tr.offsets_size/sizeof(size_t), this);
63. }
64. } else {
65. freeBuffer(NULL,
66. reinterpret_cast<const uint8_t*>(tr.data.ptr.buffer),
67. tr.data_size,
68. reinterpret_cast<const size_t*>(tr.data.ptr.offsets),
69. tr.offsets_size/sizeof(size_t), this);
70. continue;
71. }
72. }
73. goto finish;
75. default:
76. err = executeCommand(cmd);
77. if (err != NO_ERROR) goto finish;
78. break;
79. }
80. }
82. finish:
83. if (err != NO_ERROR) {
84. if (acquireResult) *acquireResult = err;
85. if (reply) reply->setError(err);
86. mLastError = err;
87. }
89. return err;
90. }
这个函数虽然很长，但是主要调用了talkWithDriver函数来与Binder驱动程序进行交互：
[cpp] view plain copy
1. status_t IPCThreadState::talkWithDriver(bool doReceive)
2. {
3. LOG_ASSERT(mProcess->mDriverFD >= 0, "Binder driver is not opened");
5. binder_write_read bwr;
7. // Is the read buffer empty?
8. const bool needRead = mIn.dataPosition() >= mIn.dataSize();
10. // We don't want to write anything if we are still reading
11. // from data left in the input buffer and the caller
12. // has requested to read the next data.
13. const size_t outAvail = (!doReceive || needRead) ? mOut.dataSize() : 0;
15. bwr.write_size = outAvail;
16. bwr.write_buffer = (long unsigned int)mOut.data();
18. // This is what we'll read.
19. if (doReceive && needRead) {
20. bwr.read_size = mIn.dataCapacity();
21. bwr.read_buffer = (long unsigned int)mIn.data();
22. } else {
23. bwr.read_size = 0;
24. }
26. IF_LOG_COMMANDS() {
27. TextOutput::Bundle _b(alog);
28. if (outAvail != 0) {
29. alog << "Sending commands to driver: " << indent;
30. const void* cmds = (const void*)bwr.write_buffer;
31. const void* end = ((const uint8_t*)cmds)+bwr.write_size;
32. alog << HexDump(cmds, bwr.write_size) << endl;
33. while (cmds < end) cmds = printCommand(alog, cmds);
34. alog << dedent;
35. }
36. alog << "Size of receive buffer: " << bwr.read_size
37. << ", needRead: " << needRead << ", doReceive: " << doReceive << endl;
38. }
40. // Return immediately if there is nothing to do.
41. if ((bwr.write_size == 0) && (bwr.read_size == 0)) return NO_ERROR;
43. bwr.write_consumed = 0;
44. bwr.read_consumed = 0;
45. status_t err;
46. do {
47. IF_LOG_COMMANDS() {
48. alog << "About to read/write, write size = " << mOut.dataSize() << endl;
49. }
50. #if defined(HAVE_ANDROID_OS)
51. if (ioctl(mProcess->mDriverFD, BINDER_WRITE_READ, &bwr) >= 0)
52. err = NO_ERROR;
53. else
54. err = -errno;
55. #else
56. err = INVALID_OPERATION;
57. #endif
58. IF_LOG_COMMANDS() {
59. alog << "Finished read/write, write size = " << mOut.dataSize() << endl;
60. }
61. } while (err == -EINTR);
63. IF_LOG_COMMANDS() {
64. alog << "Our err: " << (void*)err << ", write consumed: "
65. << bwr.write_consumed << " (of " << mOut.dataSize()
66. << "), read consumed: " << bwr.read_consumed << endl;
67. }
69. if (err >= NO_ERROR) {
70. if (bwr.write_consumed > 0) {
71. if (bwr.write_consumed < (ssize_t)mOut.dataSize())
72. mOut.remove(0, bwr.write_consumed);
73. else
74. mOut.setDataSize(0);
75. }
76. if (bwr.read_consumed > 0) {
77. mIn.setDataSize(bwr.read_consumed);
78. mIn.setDataPosition(0);
79. }
80. IF_LOG_COMMANDS() {
81. TextOutput::Bundle _b(alog);
82. alog << "Remaining data size: " << mOut.dataSize() << endl;
83. alog << "Received commands from driver: " << indent;
84. const void* cmds = mIn.data();
85. const void* end = mIn.data() + mIn.dataSize();
86. alog << HexDump(cmds, mIn.dataSize()) << endl;
87. while (cmds < end) cmds = printReturnCommand(alog, cmds);
88. alog << dedent;
89. }
90. return NO_ERROR;
91. }
93. return err;
94. }
这里doReceive和needRead均为1，有兴趣的读者可以自已分析一下。因此，这里告诉Binder驱动程序，先执行write操作，再执行read操作，下面我们将会看到。

最后，通过ioctl(mProcess->mDriverFD, BINDER_WRITE_READ, &bwr)进行到Binder驱动程序的binder_ioctl函数，我们只关注cmd为BINDER_WRITE_READ的逻辑：
[cpp] view plain copy
1. static long binder_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
2. {
3. int ret;
4. struct binder_proc *proc = filp->private_data;
5. struct binder_thread *thread;
6. unsigned int size = _IOC_SIZE(cmd);
7. void __user *ubuf = (void __user *)arg;
9. /*printk(KERN_INFO "binder_ioctl: %d:%d %x %lx ", proc->pid, current->pid, cmd, arg);*/
11. ret = wait_event_interruptible(binder_user_error_wait, binder_stop_on_user_error < 2);
12. if (ret)
13. return ret;
15. mutex_lock(&binder_lock);
16. thread = binder_get_thread(proc);
17. if (thread == NULL) {
18. ret = -ENOMEM;
19. goto err;
20. }
22. switch (cmd) {
23. case BINDER_WRITE_READ: {
24. struct binder_write_read bwr;
25. if (size != sizeof(struct binder_write_read)) {
26. ret = -EINVAL;
27. goto err;
28. }
29. if (copy_from_user(&bwr, ubuf, sizeof(bwr))) {
30. ret = -EFAULT;
31. goto err;
32. }
33. if (binder_debug_mask & BINDER_DEBUG_READ_WRITE)
34. printk(KERN_INFO "binder: %d:%d write %ld at %08lx, read %ld at %08lx ",
35. proc->pid, thread->pid, bwr.write_size, bwr.write_buffer, bwr.read_size, bwr.read_buffer);
36. if (bwr.write_size > 0) {
37. ret = binder_thread_write(proc, thread, (void __user *)bwr.write_buffer, bwr.write_size, &bwr.write_consumed);
38. if (ret < 0) {
39. bwr.read_consumed = 0;
40. if (copy_to_user(ubuf, &bwr, sizeof(bwr)))
41. ret = -EFAULT;
42. goto err;
43. }
44. }
45. if (bwr.read_size > 0) {
46. ret = binder_thread_read(proc, thread, (void __user *)bwr.read_buffer, bwr.read_size, &bwr.read_consumed, filp->f_flags & O_NONBLOCK);
47. if (!list_empty(&proc->todo))
48. wake_up_interruptible(&proc->wait);
49. if (ret < 0) {
50. if (copy_to_user(ubuf, &bwr, sizeof(bwr)))
51. ret = -EFAULT;
52. goto err;
53. }
54. }
55. if (binder_debug_mask & BINDER_DEBUG_READ_WRITE)
56. printk(KERN_INFO "binder: %d:%d wrote %ld of %ld, read return %ld of %ld ",
57. proc->pid, thread->pid, bwr.write_consumed, bwr.write_size, bwr.read_consumed, bwr.read_size);
58. if (copy_to_user(ubuf, &bwr, sizeof(bwr))) {
59. ret = -EFAULT;
60. goto err;
61. }
62. break;
63. }
64. ......
65. }
66. ret = 0;
67. err:
68. ......
69. return ret;
70. }
函数首先是将用户传进来的参数拷贝到本地变量struct binder_write_read bwr中去。这里bwr.write_size > 0为true，因此，进入到binder_thread_write函数中，我们只关注BC_TRANSACTION部分的逻辑：
[cpp] view plain copy
1. binder_thread_write(struct binder_proc *proc, struct binder_thread *thread,
2. void __user *buffer, int size, signed long *consumed)
3. {
4. uint32_t cmd;
5. void __user *ptr = buffer + *consumed;
6. void __user *end = buffer + size;
8. while (ptr < end && thread->return_error == BR_OK) {
9. if (get_user(cmd, (uint32_t __user *)ptr))
10. return -EFAULT;
11. ptr += sizeof(uint32_t);
12. if (_IOC_NR(cmd) < ARRAY_SIZE(binder_stats.bc)) {
13. binder_stats.bc[_IOC_NR(cmd)]++;
14. proc->stats.bc[_IOC_NR(cmd)]++;
15. thread->stats.bc[_IOC_NR(cmd)]++;
16. }
17. switch (cmd) {
18. .....
19. case BC_TRANSACTION:
20. case BC_REPLY: {
21. struct binder_transaction_data tr;
23. if (copy_from_user(&tr, ptr, sizeof(tr)))
24. return -EFAULT;
25. ptr += sizeof(tr);
26. binder_transaction(proc, thread, &tr, cmd == BC_REPLY);
27. break;
28. }
29. ......
30. }
31. *consumed = ptr - buffer;
32. }
33. return 0;
34. }
首先将用户传进来的transact参数拷贝在本地变量struct binder_transaction_data tr中去，接着调用binder_transaction函数进一步处理，这里我们忽略掉无关代码：
[cpp] view plain copy
1. static void
2. binder_transaction(struct binder_proc *proc, struct binder_thread *thread,
3. struct binder_transaction_data *tr, int reply)
4. {
5. struct binder_transaction *t;
6. struct binder_work *tcomplete;
7. size_t *offp, *off_end;
8. struct binder_proc *target_proc;
9. struct binder_thread *target_thread = NULL;
10. struct binder_node *target_node = NULL;
11. struct list_head *target_list;
12. wait_queue_head_t *target_wait;
13. struct binder_transaction *in_reply_to = NULL;
14. struct binder_transaction_log_entry *e;
15. uint32_t return_error;
17. ......
19. if (reply) {
20. ......
21. } else {
22. if (tr->target.handle) {
23. ......
24. } else {
25. target_node = binder_context_mgr_node;
26. if (target_node == NULL) {
27. return_error = BR_DEAD_REPLY;
28. goto err_no_context_mgr_node;
29. }
30. }
31. ......
32. target_proc = target_node->proc;
33. if (target_proc == NULL) {
34. return_error = BR_DEAD_REPLY;
35. goto err_dead_binder;
36. }
37. ......
38. }
39. if (target_thread) {
40. ......
41. } else {
42. target_list = &target_proc->todo;
43. target_wait = &target_proc->wait;
44. }
46. ......
48. /* TODO: reuse incoming transaction for reply */
49. t = kzalloc(sizeof(*t), GFP_KERNEL);
50. if (t == NULL) {
51. return_error = BR_FAILED_REPLY;
52. goto err_alloc_t_failed;
53. }
54. ......
56. tcomplete = kzalloc(sizeof(*tcomplete), GFP_KERNEL);
57. if (tcomplete == NULL) {
58. return_error = BR_FAILED_REPLY;
59. goto err_alloc_tcomplete_failed;
60. }
62. ......
64. if (!reply && !(tr->flags & TF_ONE_WAY))
65. t->from = thread;
66. else
67. t->from = NULL;
68. t->sender_euid = proc->tsk->cred->euid;
69. t->to_proc = target_proc;
70. t->to_thread = target_thread;
71. t->code = tr->code;
72. t->flags = tr->flags;
73. t->priority = task_nice(current);
74. t->buffer = binder_alloc_buf(target_proc, tr->data_size,
75. tr->offsets_size, !reply && (t->flags & TF_ONE_WAY));
76. if (t->buffer == NULL) {
77. return_error = BR_FAILED_REPLY;
78. goto err_binder_alloc_buf_failed;
79. }
80. t->buffer->allow_user_free = 0;
81. t->buffer->debug_id = t->debug_id;
82. t->buffer->transaction = t;
83. t->buffer->target_node = target_node;
84. if (target_node)
85. binder_inc_node(target_node, 1, 0, NULL);
87. offp = (size_t *)(t->buffer->data + ALIGN(tr->data_size, sizeof(void *)));
89. if (copy_from_user(t->buffer->data, tr->data.ptr.buffer, tr->data_size)) {
90. ......
91. return_error = BR_FAILED_REPLY;
92. goto err_copy_data_failed;
93. }
94. if (copy_from_user(offp, tr->data.ptr.offsets, tr->offsets_size)) {
95. ......
96. return_error = BR_FAILED_REPLY;
97. goto err_copy_data_failed;
98. }
99. ......
101. off_end = (void *)offp + tr->offsets_size;
102. for (; offp < off_end; offp++) {
103. struct flat_binder_object *fp;
104. ......
105. fp = (struct flat_binder_object *)(t->buffer->data + *offp);
106. switch (fp->type) {
107. case BINDER_TYPE_BINDER:
108. case BINDER_TYPE_WEAK_BINDER: {
109. struct binder_ref *ref;
110. struct binder_node *node = binder_get_node(proc, fp->binder);
111. if (node == NULL) {
112. node = binder_new_node(proc, fp->binder, fp->cookie);
113. if (node == NULL) {
114. return_error = BR_FAILED_REPLY;
115. goto err_binder_new_node_failed;
116. }
117. node->min_priority = fp->flags & FLAT_BINDER_FLAG_PRIORITY_MASK;
118. node->accept_fds = !!(fp->flags & FLAT_BINDER_FLAG_ACCEPTS_FDS);
119. }
120. if (fp->cookie != node->cookie) {
121. ......
122. goto err_binder_get_ref_for_node_failed;
123. }
124. ref = binder_get_ref_for_node(target_proc, node);
125. if (ref == NULL) {
126. return_error = BR_FAILED_REPLY;
127. goto err_binder_get_ref_for_node_failed;
128. }
129. if (fp->type == BINDER_TYPE_BINDER)
130. fp->type = BINDER_TYPE_HANDLE;
131. else
132. fp->type = BINDER_TYPE_WEAK_HANDLE;
133. fp->handle = ref->desc;
134. binder_inc_ref(ref, fp->type == BINDER_TYPE_HANDLE, &thread->todo);
135. ......
137. } break;
138. ......
139. }
140. }
142. if (reply) {
143. ......
144. } else if (!(t->flags & TF_ONE_WAY)) {
145. BUG_ON(t->buffer->async_transaction != 0);
146. t->need_reply = 1;
147. t->from_parent = thread->transaction_stack;
148. thread->transaction_stack = t;
149. } else {
150. ......
151. }
152. t->work.type = BINDER_WORK_TRANSACTION;
153. list_add_tail(&t->work.entry, target_list);
154. tcomplete->type = BINDER_WORK_TRANSACTION_COMPLETE;
155. list_add_tail(&tcomplete->entry, &thread->todo);
156. if (target_wait)
157. wake_up_interruptible(target_wait);
158. return;
159. ......
160. }
注意，这里传进来的参数reply为0，tr->target.handle也为0。因此，target_proc、target_thread、target_node、target_list和target_wait的值分别为：
[cpp] view plain copy
1. target_node = binder_context_mgr_node;
2. target_proc = target_node->proc;
3. target_list = &target_proc->todo;
4. target_wait = &target_proc->wait;
接着，分配了一个待处理事务t和一个待完成工作项tcomplete，并执行初始化工作：
[cpp] view plain copy
1. /* TODO: reuse incoming transaction for reply */
2. t = kzalloc(sizeof(*t), GFP_KERNEL);
3. if (t == NULL) {
4. return_error = BR_FAILED_REPLY;
5. goto err_alloc_t_failed;
6. }
7. ......
9. tcomplete = kzalloc(sizeof(*tcomplete), GFP_KERNEL);
10. if (tcomplete == NULL) {
11. return_error = BR_FAILED_REPLY;
12. goto err_alloc_tcomplete_failed;
13. }
15. ......
17. if (!reply && !(tr->flags & TF_ONE_WAY))
18. t->from = thread;
19. else
20. t->from = NULL;
21. t->sender_euid = proc->tsk->cred->euid;
22. t->to_proc = target_proc;
23. t->to_thread = target_thread;
24. t->code = tr->code;
25. t->flags = tr->flags;
26. t->priority = task_nice(current);
27. t->buffer = binder_alloc_buf(target_proc, tr->data_size,
28. tr->offsets_size, !reply && (t->flags & TF_ONE_WAY));
29. if (t->buffer == NULL) {
30. return_error = BR_FAILED_REPLY;
31. goto err_binder_alloc_buf_failed;
32. }
33. t->buffer->allow_user_free = 0;
34. t->buffer->debug_id = t->debug_id;
35. t->buffer->transaction = t;
36. t->buffer->target_node = target_node;
37. if (target_node)
38. binder_inc_node(target_node, 1, 0, NULL);
40. offp = (size_t *)(t->buffer->data + ALIGN(tr->data_size, sizeof(void *)));
42. if (copy_from_user(t->buffer->data, tr->data.ptr.buffer, tr->data_size)) {
43. ......
44. return_error = BR_FAILED_REPLY;
45. goto err_copy_data_failed;
46. }
47. if (copy_from_user(offp, tr->data.ptr.offsets, tr->offsets_size)) {
48. ......
49. return_error = BR_FAILED_REPLY;
50. goto err_copy_data_failed;
51. }
注意，这里的事务t是要交给target_proc处理的，在这个场景之下，就是Service Manager了。因此，下面的语句：
[cpp] view plain copy
1. t->buffer = binder_alloc_buf(target_proc, tr->data_size,
2. tr->offsets_size, !reply && (t->flags & TF_ONE_WAY));
就是在Service Manager的进程空间中分配一块内存来保存用户传进入的参数了：
[cpp] view plain copy
1. if (copy_from_user(t->buffer->data, tr->data.ptr.buffer, tr->data_size)) {
2. ......
3. return_error = BR_FAILED_REPLY;
4. goto err_copy_data_failed;
5. }
6. if (copy_from_user(offp, tr->data.ptr.offsets, tr->offsets_size)) {
7. ......
8. return_error = BR_FAILED_REPLY;
9. goto err_copy_data_failed;
10. }
由于现在target_node要被使用了，增加它的引用计数：
[cpp] view plain copy
1. if (target_node)
2. binder_inc_node(target_node, 1, 0, NULL);
接下去的for循环，就是用来处理传输数据中的Binder对象了。在我们的场景中，有一个类型为BINDER_TYPE_BINDER的Binder实体MediaPlayerService：
[cpp] view plain copy
1. switch (fp->type) {
2. case BINDER_TYPE_BINDER:
3. case BINDER_TYPE_WEAK_BINDER: {
4. struct binder_ref *ref;
5. struct binder_node *node = binder_get_node(proc, fp->binder);
6. if (node == NULL) {
7. node = binder_new_node(proc, fp->binder, fp->cookie);
8. if (node == NULL) {
9. return_error = BR_FAILED_REPLY;
10. goto err_binder_new_node_failed;
11. }
12. node->min_priority = fp->flags & FLAT_BINDER_FLAG_PRIORITY_MASK;
13. node->accept_fds = !!(fp->flags & FLAT_BINDER_FLAG_ACCEPTS_FDS);
14. }
15. if (fp->cookie != node->cookie) {
16. ......
17. goto err_binder_get_ref_for_node_failed;
18. }
19. ref = binder_get_ref_for_node(target_proc, node);
20. if (ref == NULL) {
21. return_error = BR_FAILED_REPLY;
22. goto err_binder_get_ref_for_node_failed;
23. }
24. if (fp->type == BINDER_TYPE_BINDER)
25. fp->type = BINDER_TYPE_HANDLE;
26. else
27. fp->type = BINDER_TYPE_WEAK_HANDLE;
28. fp->handle = ref->desc;
29. binder_inc_ref(ref, fp->type == BINDER_TYPE_HANDLE, &thread->todo);
30. ......
32. } break;
   由于是第一次在Binder驱动程序中传输这个MediaPlayerService，调用binder_get_node函数查询这个Binder实体时，会返回空，于是binder_new_node在proc中新建一个，下次就可以直接使用了。

   现在，由于要把这个Binder实体MediaPlayerService交给target_proc，也就是Service Manager来管理，也就是说Service Manager要引用这个MediaPlayerService了，于是通过binder_get_ref_for_node为 MediaPlayerService创建一个引用，并且通过binder_inc_ref来增加这个引用计数，防止这个引用还在使用过程当中就被销毁。注意，到了这里的时候，t->buffer中的flat_binder_obj的type已经改为 BINDER_TYPE_HANDLE，handle已经改为ref->desc，跟原来不一样了，因为这个flat_binder_obj是最终是要传给Service Manager的，而Service Manager只能够通过句柄值来引用这个Binder实体。

   最后，把待处理事务加入到target_list列表中去：
[cpp] view plain copy
1. list_add_tail(&t->work.entry, target_list);
并且把待完成工作项加入到本线程的todo等待执行列表中去：
[cpp] view plain copy
1. list_add_tail(&tcomplete->entry, &thread->todo);
现在目标进程有事情可做了，于是唤醒它：
[cpp] view plain copy
1. if (target_wait)
2. wake_up_interruptible(target_wait);
   这里就是要唤醒Service Manager进程了。回忆一下前面浅谈Service Manager成为Android进程间通信（IPC）机制Binder守护进程之路这篇文章，此时， Service Manager正在binder_thread_read函数中调用wait_event_interruptible进入休眠状态。

   这里我们先忽略一下Service Manager被唤醒之后的场景，继续MedaPlayerService的启动过程，然后再回来。

   回到binder_ioctl函数，bwr.read_size > 0为true，于是进入binder_thread_read函数：
[cpp] view plain copy
1. static int
2. binder_thread_read(struct binder_proc *proc, struct binder_thread *thread,
3. void __user *buffer, int size, signed long *consumed, int non_block)
4. {
5. void __user *ptr = buffer + *consumed;
6. void __user *end = buffer + size;
8. int ret = 0;
9. int wait_for_proc_work;
11. if (*consumed == 0) {
12. if (put_user(BR_NOOP, (uint32_t __user *)ptr))
13. return -EFAULT;
14. ptr += sizeof(uint32_t);
15. }
17. retry:
18. wait_for_proc_work = thread->transaction_stack == NULL && list_empty(&thread->todo);
20. .......
22. if (wait_for_proc_work) {
23. .......
24. } else {
25. if (non_block) {
26. if (!binder_has_thread_work(thread))
27. ret = -EAGAIN;
28. } else
29. ret = wait_event_interruptible(thread->wait, binder_has_thread_work(thread));
30. }
32. ......
34. while (1) {
35. uint32_t cmd;
36. struct binder_transaction_data tr;
37. struct binder_work *w;
38. struct binder_transaction *t = NULL;
40. if (!list_empty(&thread->todo))
41. w = list_first_entry(&thread->todo, struct binder_work, entry);
42. else if (!list_empty(&proc->todo) && wait_for_proc_work)
43. w = list_first_entry(&proc->todo, struct binder_work, entry);
44. else {
45. if (ptr - buffer == 4 && !(thread->looper & BINDER_LOOPER_STATE_NEED_RETURN)) /* no data added */
46. goto retry;
47. break;
48. }
50. if (end - ptr < sizeof(tr) + 4)
51. break;
53. switch (w->type) {
54. ......
55. case BINDER_WORK_TRANSACTION_COMPLETE: {
56. cmd = BR_TRANSACTION_COMPLETE;
57. if (put_user(cmd, (uint32_t __user *)ptr))
58. return -EFAULT;
59. ptr += sizeof(uint32_t);
61. binder_stat_br(proc, thread, cmd);
62. if (binder_debug_mask & BINDER_DEBUG_TRANSACTION_COMPLETE)
63. printk(KERN_INFO "binder: %d:%d BR_TRANSACTION_COMPLETE ",
64. proc->pid, thread->pid);
66. list_del(&w->entry);
67. kfree(w);
68. binder_stats.obj_deleted[BINDER_STAT_TRANSACTION_COMPLETE]++;
69. } break;
70. ......
71. }
73. if (!t)
74. continue;
76. ......
77. }
79. done:
80. ......
81. return 0;
82. }
这里，thread->transaction_stack和thread->todo均不为空，于是 wait_for_proc_work为false，由于binder_has_thread_work的时候，返回true，这里因为 thread->todo不为空，因此，线程虽然调用了wait_event_interruptible，但是不会睡眠，于是继续往下执行。

由于thread->todo不为空，执行下列语句：
[cpp] view plain copy
1. if (!list_empty(&thread->todo))
2. w = list_first_entry(&thread->todo, struct binder_work, entry);
w->type为BINDER_WORK_TRANSACTION_COMPLETE，这是在上面的binder_transaction函数设置的，于是执行：
[cpp] view plain copy
1. switch (w->type) {
2. ......
3. case BINDER_WORK_TRANSACTION_COMPLETE: {
4. cmd = BR_TRANSACTION_COMPLETE;
5. if (put_user(cmd, (uint32_t __user *)ptr))
6. return -EFAULT;
7. ptr += sizeof(uint32_t);
9. ......
10. list_del(&w->entry);
11. kfree(w);
13. } break;
14. ......
15. }
这里就将w从thread->todo删除了。由于这里t为空，重新执行while循环，这时由于已经没有事情可做了，最后就返回到 binder_ioctl函数中。注间，这里一共往用户传进来的缓冲区buffer写入了两个整数，分别是BR_NOOP和 BR_TRANSACTION_COMPLETE。

binder_ioctl函数返回到用户空间之前，把数据消耗情况拷贝回用户空间中：
[cpp] view plain copy
1. if (copy_to_user(ubuf, &bwr, sizeof(bwr))) {
2. ret = -EFAULT;
3. goto err;
4. }
最后返回到IPCThreadState::talkWithDriver函数中，执行下面语句：
[cpp] view plain copy
1. if (err >= NO_ERROR) {
2. if (bwr.write_consumed > 0) {
3. if (bwr.write_consumed < (ssize_t)mOut.dataSize())
4. mOut.remove(0, bwr.write_consumed);
5. else
6. mOut.setDataSize(0);
7. }
8. if (bwr.read_consumed > 0) {
9. <pre code_snippet_id="134056" snippet_file_name="blog_20131230_54_6706870" name="code" class="cpp"> mIn.setDataSize(bwr.read_consumed);
10. mIn.setDataPosition(0);</pre> } ...... return NO_ERROR; }
首先是把mOut的数据清空：
[cpp] view plain copy
1. mOut.setDataSize(0);
然后设置已经读取的内容的大小：
[cpp] view plain copy
1. mIn.setDataSize(bwr.read_consumed);
2. mIn.setDataPosition(0);
然后返回到IPCThreadState::waitForResponse函数中。在 IPCThreadState::waitForResponse函数，先是从mIn读出一个整数，这个便是BR_NOOP了，这是一个空操作，什么也不做。然后继续进入IPCThreadState::talkWithDriver函数中。
这时候，下面语句执行后：
[cpp] view plain copy
1. const bool needRead = mIn.dataPosition() >= mIn.dataSize();
needRead为false，因为在mIn中，尚有一个整数BR_TRANSACTION_COMPLETE未读出。

这时候，下面语句执行后：
[cpp] view plain copy
1. const size_t outAvail = (!doReceive || needRead) ? mOut.dataSize() : 0;
outAvail等于0。因此，最后bwr.write_size和bwr.read_size均为 0，IPCThreadState::talkWithDriver函数什么也不做，直接返回到 IPCThreadState::waitForResponse函数中。在IPCThreadState::waitForResponse函数，又继续从mIn读出一个整数，这个便是BR_TRANSACTION_COMPLETE：
[cpp] view plain copy
1. switch (cmd) {
2. case BR_TRANSACTION_COMPLETE:
3. if (!reply && !acquireResult) goto finish;
4. break;
5. ......
6. }
reply不为NULL，因此，IPCThreadState::waitForResponse的循环没有结束，继续执行，又进入到IPCThreadState::talkWithDrive中。

这次，needRead就为true了，而outAvail仍为0，所以bwr.read_size不为0，bwr.write_size为0。于是通过：
[cpp] view plain copy
1. ioctl(mProcess->mDriverFD, BINDER_WRITE_READ, &bwr)
进入到Binder驱动程序中的binder_ioctl函数中。由于bwr.write_size为0，bwr.read_size不为0，这次直接就进入到binder_thread_read函数中。这时候，thread->transaction_stack不等于0，但是 thread->todo为空，于是线程就通过：
[cpp] view plain copy
1. wait_event_interruptible(thread->wait, binder_has_thread_work(thread));
进入睡眠状态，等待Service Manager来唤醒了。

现在，我们可以回到Service Manager被唤醒的过程了。我们接着前面浅谈Service Manager成为Android进程间通信（IPC）机制Binder守护进程之路这篇文章的最后，继续描述。此时， Service Manager正在binder_thread_read函数中调用wait_event_interruptible_exclusive进入休眠状态。上面被MediaPlayerService启动后进程唤醒后，继续执行binder_thread_read函数：
[cpp] view plain copy
1. static int
2. binder_thread_read(struct binder_proc *proc, struct binder_thread *thread,
3. void __user *buffer, int size, signed long *consumed, int non_block)
4. {
5. void __user *ptr = buffer + *consumed;
6. void __user *end = buffer + size;
8. int ret = 0;
9. int wait_for_proc_work;
11. if (*consumed == 0) {
12. if (put_user(BR_NOOP, (uint32_t __user *)ptr))
13. return -EFAULT;
14. ptr += sizeof(uint32_t);
15. }
17. retry:
18. wait_for_proc_work = thread->transaction_stack == NULL && list_empty(&thread->todo);
20. ......
22. if (wait_for_proc_work) {
23. ......
24. if (non_block) {
25. if (!binder_has_proc_work(proc, thread))
26. ret = -EAGAIN;
27. } else
28. ret = wait_event_interruptible_exclusive(proc->wait, binder_has_proc_work(proc, thread));
29. } else {
30. ......
31. }
33. ......
35. while (1) {
36. uint32_t cmd;
37. struct binder_transaction_data tr;
38. struct binder_work *w;
39. struct binder_transaction *t = NULL;
41. if (!list_empty(&thread->todo))
42. w = list_first_entry(&thread->todo, struct binder_work, entry);
43. else if (!list_empty(&proc->todo) && wait_for_proc_work)
44. w = list_first_entry(&proc->todo, struct binder_work, entry);
45. else {
46. if (ptr - buffer == 4 && !(thread->looper & BINDER_LOOPER_STATE_NEED_RETURN)) /* no data added */
47. goto retry;
48. break;
49. }
51. if (end - ptr < sizeof(tr) + 4)
52. break;
54. switch (w->type) {
55. case BINDER_WORK_TRANSACTION: {
56. t = container_of(w, struct binder_transaction, work);
57. } break;
58. ......
59. }
61. if (!t)
62. continue;
64. BUG_ON(t->buffer == NULL);
65. if (t->buffer->target_node) {
66. struct binder_node *target_node = t->buffer->target_node;
67. tr.target.ptr = target_node->ptr;
68. tr.cookie = target_node->cookie;
69. ......
70. cmd = BR_TRANSACTION;
71. } else {
72. ......
73. }
74. tr.code = t->code;
75. tr.flags = t->flags;
76. tr.sender_euid = t->sender_euid;
78. if (t->from) {
79. struct task_struct *sender = t->from->proc->tsk;
80. tr.sender_pid = task_tgid_nr_ns(sender, current->nsproxy->pid_ns);
81. } else {
82. tr.sender_pid = 0;
83. }
85. tr.data_size = t->buffer->data_size;
86. tr.offsets_size = t->buffer->offsets_size;
87. tr.data.ptr.buffer = (void *)t->buffer->data + proc->user_buffer_offset;
88. tr.data.ptr.offsets = tr.data.ptr.buffer + ALIGN(t->buffer->data_size, sizeof(void *));
90. if (put_user(cmd, (uint32_t __user *)ptr))
91. return -EFAULT;
92. ptr += sizeof(uint32_t);
93. if (copy_to_user(ptr, &tr, sizeof(tr)))
94. return -EFAULT;
95. ptr += sizeof(tr);
97. ......
99. list_del(&t->work.entry);
100. t->buffer->allow_user_free = 1;
101. if (cmd == BR_TRANSACTION && !(t->flags & TF_ONE_WAY)) {
102. t->to_parent = thread->transaction_stack;
103. t->to_thread = thread;
104. thread->transaction_stack = t;
105. } else {
106. t->buffer->transaction = NULL;
107. kfree(t);
108. binder_stats.obj_deleted[BINDER_STAT_TRANSACTION]++;
109. }
110. break;
111. }
113. done:
115. ......
116. return 0;
117. }
Service Manager被唤醒之后，就进入while循环开始处理事务了。这里wait_for_proc_work等于1，并且proc->todo不为空，所以从proc->todo列表中得到第一个工作项：
[cpp] view plain copy
1. w = list_first_entry(&proc->todo, struct binder_work, entry);
从上面的描述中，我们知道，这个工作项的类型为BINDER_WORK_TRANSACTION，于是通过下面语句得到事务项：
[cpp] view plain copy
1. t = container_of(w, struct binder_transaction, work);
接着就是把事务项t中的数据拷贝到本地局部变量struct binder_transaction_data tr中去了：
[cpp] view plain copy
1. if (t->buffer->target_node) {
2. struct binder_node *target_node = t->buffer->target_node;
3. tr.target.ptr = target_node->ptr;
4. tr.cookie = target_node->cookie;
5. ......
6. cmd = BR_TRANSACTION;
7. } else {
8. ......
9. }
10. tr.code = t->code;
11. tr.flags = t->flags;
12. tr.sender_euid = t->sender_euid;
14. if (t->from) {
15. struct task_struct *sender = t->from->proc->tsk;
16. tr.sender_pid = task_tgid_nr_ns(sender, current->nsproxy->pid_ns);
17. } else {
18. tr.sender_pid = 0;
19. }
21. tr.data_size = t->buffer->data_size;
22. tr.offsets_size = t->buffer->offsets_size;
23. tr.data.ptr.buffer = (void *)t->buffer->data + proc->user_buffer_offset;
24. tr.data.ptr.offsets = tr.data.ptr.buffer + ALIGN(t->buffer->data_size, sizeof(void *));
这里有一个非常重要的地方，是Binder进程间通信机制的精髓所在：
[cpp] view plain copy
1. tr.data.ptr.buffer = (void *)t->buffer->data + proc->user_buffer_offset;
2. tr.data.ptr.offsets = tr.data.ptr.buffer + ALIGN(t->buffer->data_size, sizeof(void *));
t->buffer->data所指向的地址是内核空间的，现在要把数据返回给Service Manager进程的用户空间，而Service Manager进程的用户空间是不能访问内核空间的数据的，所以这里要作一下处理。怎么处理呢？我们在学面向对象语言的时候，对象的拷贝有深拷贝和浅拷贝之分，深拷贝是把另外分配一块新内存，然后把原始对象的内容搬过去，浅拷贝是并没有为新对象分配一块新空间，而只是分配一个引用，而个引用指向原始对象。 Binder机制用的是类似浅拷贝的方法，通过在用户空间分配一个虚拟地址，然后让这个用户空间虚拟地址与 t->buffer->data 这个内核空间虚拟地址指向同一个物理地址，这样就可以实现浅拷贝了。怎么样用户空间和内核空间的虚拟地址同时指向同一个物理地址呢？请参考前面一篇文章浅谈Service Manager成为Android进程间通信（IPC）机制Binder守护进程之路，那里有详细描述。这里只要将t->buffer->data加上一个偏移值proc->user_buffer_offset就可以得到t->buffer->data对应的用户空间虚拟地址了。调整了tr.data.ptr.buffer的值之后，不要忘记也要一起调整 tr.data.ptr.offsets的值。

接着就是把tr的内容拷贝到用户传进来的缓冲区去了，指针ptr指向这个用户缓冲区的地址：
[cpp] view plain copy
1. if (put_user(cmd, (uint32_t __user *)ptr))
2. return -EFAULT;
3. ptr += sizeof(uint32_t);
4. if (copy_to_user(ptr, &tr, sizeof(tr)))
5. return -EFAULT;
6. ptr += sizeof(tr);
这里可以看出，这里只是对作tr.data.ptr.bufferr和tr.data.ptr.offsets的内容作了浅拷贝。

最后，由于已经处理了这个事务，要把它从todo列表中删除：
[cpp] view plain copy
1. list_del(&t->work.entry);
2. t->buffer->allow_user_free = 1;
3. if (cmd == BR_TRANSACTION && !(t->flags & TF_ONE_WAY)) {
4. t->to_parent = thread->transaction_stack;
5. t->to_thread = thread;
6. thread->transaction_stack = t;
7. } else {
8. t->buffer->transaction = NULL;
9. kfree(t);
10. binder_stats.obj_deleted[BINDER_STAT_TRANSACTION]++;
11. }
注意，这里的cmd == BR_TRANSACTION && !(t->flags & TF_ONE_WAY)为true，表明这个事务虽然在驱动程序中已经处理完了，但是它仍然要等待Service Manager完成之后，给驱动程序一个确认，也就是需要等待回复，于是把当前事务t放在thread->transaction_stack队列的头部：
[cpp] view plain copy
1. t->to_parent = thread->transaction_stack;
2. t->to_thread = thread;
3. thread->transaction_stack = t;
如果cmd == BR_TRANSACTION && !(t->flags & TF_ONE_WAY)为false，那就不需要等待回复了，直接把事务t删掉。

这个while最后通过一个break跳了出来，最后返回到binder_ioctl函数中：
[cpp] view plain copy
1. static long binder_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
2. {
3. int ret;
4. struct binder_proc *proc = filp->private_data;
5. struct binder_thread *thread;
6. unsigned int size = _IOC_SIZE(cmd);
7. void __user *ubuf = (void __user *)arg;
9. ......
11. switch (cmd) {
12. case BINDER_WRITE_READ: {
13. struct binder_write_read bwr;
14. if (size != sizeof(struct binder_write_read)) {
15. ret = -EINVAL;
16. goto err;
17. }
18. if (copy_from_user(&bwr, ubuf, sizeof(bwr))) {
19. ret = -EFAULT;
20. goto err;
21. }
22. ......
23. if (bwr.read_size > 0) {
24. ret = binder_thread_read(proc, thread, (void __user *)bwr.read_buffer, bwr.read_size, &bwr.read_consumed, filp->f_flags & O_NONBLOCK);
25. if (!list_empty(&proc->todo))
26. wake_up_interruptible(&proc->wait);
27. if (ret < 0) {
28. if (copy_to_user(ubuf, &bwr, sizeof(bwr)))
29. ret = -EFAULT;
30. goto err;
31. }
32. }
33. ......
34. if (copy_to_user(ubuf, &bwr, sizeof(bwr))) {
35. ret = -EFAULT;
36. goto err;
37. }
38. break;
39. }
40. ......
41. default:
42. ret = -EINVAL;
43. goto err;
44. }
45. ret = 0;
46. err:
47. ......
48. return ret;
49. }
从binder_thread_read返回来后，再看看proc->todo是否还有事务等待处理，如果是，就把睡眠在 proc->wait队列的线程唤醒来处理。最后，把本地变量struct binder_write_read bwr的内容拷贝回到用户传进来的缓冲区中，就返回了。

这里就是返回到frameworks/base/cmds/servicemanager/binder.c文件中的binder_loop函数了：
[cpp] view plain copy
1. void binder_loop(struct binder_state *bs, binder_handler func)
2. {
3. int res;
4. struct binder_write_read bwr;
5. unsigned readbuf[32];
7. bwr.write_size = 0;
8. bwr.write_consumed = 0;
9. bwr.write_buffer = 0;
11. readbuf[0] = BC_ENTER_LOOPER;
12. binder_write(bs, readbuf, sizeof(unsigned));
14. for (;;) {
15. bwr.read_size = sizeof(readbuf);
16. bwr.read_consumed = 0;
17. bwr.read_buffer = (unsigned) readbuf;
19. res = ioctl(bs->fd, BINDER_WRITE_READ, &bwr);
21. if (res < 0) {
22. LOGE("binder_loop: ioctl failed (%s) ", strerror(errno));
23. break;
24. }
26. res = binder_parse(bs, 0, readbuf, bwr.read_consumed, func);
27. if (res == 0) {
28. LOGE("binder_loop: unexpected reply?! ");
29. break;
30. }
31. if (res < 0) {
32. LOGE("binder_loop: io error %d %s ", res, strerror(errno));
33. break;
34. }
35. }
36. }
返回来的数据都放在readbuf中，接着调用binder_parse进行解析：
[cpp] view plain copy
1. int binder_parse(struct binder_state *bs, struct binder_io *bio,
2. uint32_t *ptr, uint32_t size, binder_handler func)
3. {
4. int r = 1;
5. uint32_t *end = ptr + (size / 4);
7. while (ptr < end) {
8. uint32_t cmd = *ptr++;
9. ......
10. case BR_TRANSACTION: {
11. struct binder_txn *txn = (void *) ptr;
12. if ((end - ptr) * sizeof(uint32_t) < sizeof(struct binder_txn)) {
13. LOGE("parse: txn too small! ");
14. return -1;
15. }
16. binder_dump_txn(txn);
17. if (func) {
18. unsigned rdata[256/4];
19. struct binder_io msg;
20. struct binder_io reply;
21. int res;
23. bio_init(&reply, rdata, sizeof(rdata), 4);
24. bio_init_from_txn(&msg, txn);
25. res = func(bs, txn, &msg, &reply);
26. binder_send_reply(bs, &reply, txn->data, res);
27. }
28. ptr += sizeof(*txn) / sizeof(uint32_t);
29. break;
30. }
31. ......
32. default:
33. LOGE("parse: OOPS %d ", cmd);
34. return -1;
35. }
36. }
38. return r;
39. }
首先把从Binder驱动程序读出来的数据转换为一个struct binder_txn结构体，保存在txn本地变量中，struct binder_txn定义在frameworks/base/cmds/servicemanager/binder.h文件中：
[cpp] view plain copy
1. struct binder_txn
2. {
3. void *target;
4. void *cookie;
5. uint32_t code;
6. uint32_t flags;
8. uint32_t sender_pid;
9. uint32_t sender_euid;
11. uint32_t data_size;
12. uint32_t offs_size;
13. void *data;
14. void *offs;
15. };
函数中还用到了另外一个数据结构struct binder_io，也是定义在frameworks/base/cmds/servicemanager/binder.h文件中：
[cpp] view plain copy
1. struct binder_io
2. {
3. char *data; /* pointer to read/write from */
4. uint32_t *offs; /* array of offsets */
5. uint32_t data_avail; /* bytes available in data buffer */
6. uint32_t offs_avail; /* entries available in offsets array */
8. char *data0; /* start of data buffer */
9. uint32_t *offs0; /* start of offsets buffer */
10. uint32_t flags;
11. uint32_t unused;
12. };
接着往下看，函数调bio_init来初始化reply变量：
[cpp] view plain copy
1. void bio_init(struct binder_io *bio, void *data,
2. uint32_t maxdata, uint32_t maxoffs)
3. {
4. uint32_t n = maxoffs * sizeof(uint32_t);
6. if (n > maxdata) {
7. bio->flags = BIO_F_OVERFLOW;
8. bio->data_avail = 0;
9. bio->offs_avail = 0;
10. return;
11. }
13. bio->data = bio->data0 = data + n;
14. bio->offs = bio->offs0 = data;
15. bio->data_avail = maxdata - n;
16. bio->offs_avail = maxoffs;
17. bio->flags = 0;
18. }
接着又调用bio_init_from_txn来初始化msg变量：
[cpp] view plain copy
1. void bio_init_from_txn(struct binder_io *bio, struct binder_txn *txn)
2. {
3. bio->data = bio->data0 = txn->data;
4. bio->offs = bio->offs0 = txn->offs;
5. bio->data_avail = txn->data_size;
6. bio->offs_avail = txn->offs_size / 4;
7. bio->flags = BIO_F_SHARED;
8. }
最后，真正进行处理的函数是从参数中传进来的函数指针func，这里就是定义在frameworks/base/cmds/servicemanager/service_manager.c文件中的svcmgr_handler函数：
[cpp] view plain copy
1. int svcmgr_handler(struct binder_state *bs,
2. struct binder_txn *txn,
3. struct binder_io *msg,
4. struct binder_io *reply)
5. {
6. struct svcinfo *si;
7. uint16_t *s;
8. unsigned len;
9. void *ptr;
10. uint32_t strict_policy;
12. if (txn->target != svcmgr_handle)
13. return -1;
15. // Equivalent to Parcel::enforceInterface(), reading the RPC
16. // header with the strict mode policy mask and the interface name.
17. // Note that we ignore the strict_policy and don't propagate it
18. // further (since we do no outbound RPCs anyway).
19. strict_policy = bio_get_uint32(msg);
20. s = bio_get_string16(msg, &len);
21. if ((len != (sizeof(svcmgr_id) / 2)) ||
22. memcmp(svcmgr_id, s, sizeof(svcmgr_id))) {
23. fprintf(stderr,"invalid id %s ", str8(s));
24. return -1;
25. }
27. switch(txn->code) {
28. ......
29. case SVC_MGR_ADD_SERVICE:
30. s = bio_get_string16(msg, &len);
31. ptr = bio_get_ref(msg);
32. if (do_add_service(bs, s, len, ptr, txn->sender_euid))
33. return -1;
34. break;
35. ......
36. }
38. bio_put_uint32(reply, 0);
39. return 0;
40. }
回忆一下，在BpServiceManager::addService时，传给Binder驱动程序的参数为：
[cpp] view plain copy
1. writeInt32(IPCThreadState::self()->getStrictModePolicy() | STRICT_MODE_PENALTY_GATHER);
2. writeString16("android.os.IServiceManager");
3. writeString16("media.player");
4. writeStrongBinder(new MediaPlayerService());
这里的语句：
[cpp] view plain copy
1. strict_policy = bio_get_uint32(msg);
2. s = bio_get_string16(msg, &len);
3. s = bio_get_string16(msg, &len);
4. ptr = bio_get_ref(msg);
就是依次把它们读取出来了，这里，我们只要看一下bio_get_ref的实现。先看一个数据结构struct binder_obj的定义：
[cpp] view plain copy
1. struct binder_object
2. {
3. uint32_t type;
4. uint32_t flags;
5. void *pointer;
6. void *cookie;
7. };
这个结构体其实就是对应struct flat_binder_obj的。

接着看bio_get_ref实现：
[cpp] view plain copy
1. void *bio_get_ref(struct binder_io *bio)
2. {
3. struct binder_object *obj;
5. obj = _bio_get_obj(bio);
6. if (!obj)
7. return 0;
9. if (obj->type == BINDER_TYPE_HANDLE)
10. return obj->pointer;
12. return 0;
13. }
_bio_get_obj这个函数就不跟进去看了，它的作用就是从binder_io中取得第一个还没取获取过的binder_object。在这个场景下，就是我们最开始传过来代表MediaPlayerService的flat_binder_obj了，这个原始的flat_binder_obj的 type为BINDER_TYPE_BINDER，binder为指向MediaPlayerService的弱引用的地址。在前面我们说过，在 Binder驱动驱动程序里面，会把这个flat_binder_obj的type改为BINDER_TYPE_HANDLE，handle改为一个句柄值。这里的handle值就等于obj->pointer的值。

回到svcmgr_handler函数，调用do_add_service进一步处理：
[cpp] view plain copy
1. int do_add_service(struct binder_state *bs,
2. uint16_t *s, unsigned len,
3. void *ptr, unsigned uid)
4. {
5. struct svcinfo *si;
6. // LOGI("add_service('%s',%p) uid=%d ", str8(s), ptr, uid);
8. if (!ptr || (len == 0) || (len > 127))
9. return -1;
11. if (!svc_can_register(uid, s)) {
12. LOGE("add_service('%s',%p) uid=%d - PERMISSION DENIED ",
13. str8(s), ptr, uid);
14. return -1;
15. }
17. si = find_svc(s, len);
18. if (si) {
19. if (si->ptr) {
20. LOGE("add_service('%s',%p) uid=%d - ALREADY REGISTERED ",
21. str8(s), ptr, uid);
22. return -1;
23. }
24. si->ptr = ptr;
25. } else {
26. si = malloc(sizeof(*si) + (len + 1) * sizeof(uint16_t));
27. if (!si) {
28. LOGE("add_service('%s',%p) uid=%d - OUT OF MEMORY ",
29. str8(s), ptr, uid);
30. return -1;
31. }
32. si->ptr = ptr;
33. si->len = len;
34. memcpy(si->name, s, (len + 1) * sizeof(uint16_t));
35. si->name[len] = '';
36. si->death.func = svcinfo_death;
37. si->death.ptr = si;
38. si->next = svclist;
39. svclist = si;
40. }
42. binder_acquire(bs, ptr);
43. binder_link_to_death(bs, ptr, &si->death);
44. return 0;
45. }
这个函数的实现很简单，就是把MediaPlayerService这个Binder实体的引用写到一个struct svcinfo结构体中，主要是它的名称和句柄值，然后插入到链接svclist的头部去。这样，Client来向Service Manager查询服务接口时，只要给定服务名称，Service Manger就可以返回相应的句柄值了。

这个函数执行完成后，返回到svcmgr_handler函数，函数的最后，将一个错误码0写到reply变量中去，表示一切正常：
[cpp] view plain copy
1. bio_put_uint32(reply, 0);
svcmgr_handler函数执行完成后，返回到binder_parse函数，执行下面语句：
[cpp] view plain copy
1. binder_send_reply(bs, &reply, txn->data, res);
我们看一下binder_send_reply的实现，从函数名就可以猜到它要做什么了，告诉Binder驱动程序，它完成了Binder驱动程序交给它的任务了。
[cpp] view plain copy
1. void binder_send_reply(struct binder_state *bs,
2. struct binder_io *reply,
3. void *buffer_to_free,
4. int status)
5. {
6. struct {
7. uint32_t cmd_free;
8. void *buffer;
9. uint32_t cmd_reply;
10. struct binder_txn txn;
11. } __attribute__((packed)) data;
13. data.cmd_free = BC_FREE_BUFFER;
14. data.buffer = buffer_to_free;
15. data.cmd_reply = BC_REPLY;
16. data.txn.target = 0;
17. data.txn.cookie = 0;
18. data.txn.code = 0;
19. if (status) {
20. data.txn.flags = TF_STATUS_CODE;
21. data.txn.data_size = sizeof(int);
22. data.txn.offs_size = 0;
23. data.txn.data = &status;
24. data.txn.offs = 0;
25. } else {
26. data.txn.flags = 0;
27. data.txn.data_size = reply->data - reply->data0;
28. data.txn.offs_size = ((char*) reply->offs) - ((char*) reply->offs0);
29. data.txn.data = reply->data0;
30. data.txn.offs = reply->offs0;
31. }
32. binder_write(bs, &data, sizeof(data));
33. }
从这里可以看出，binder_send_reply告诉Binder驱动程序执行BC_FREE_BUFFER和BC_REPLY命令，前者释放之前在 binder_transaction分配的空间，地址为buffer_to_free，buffer_to_free这个地址是Binder驱动程序把自己在内核空间用的地址转换成用户空间地址再传给Service Manager的，所以Binder驱动程序拿到这个地址后，知道怎么样释放这个空间；后者告诉MediaPlayerService，它的 addService操作已经完成了，错误码是0，保存在data.txn.data中。

再来看binder_write函数：
[cpp] view plain copy
1. int binder_write(struct binder_state *bs, void *data, unsigned len)
2. {
3. struct binder_write_read bwr;
4. int res;
5. bwr.write_size = len;
6. bwr.write_consumed = 0;
7. bwr.write_buffer = (unsigned) data;
8. bwr.read_size = 0;
9. bwr.read_consumed = 0;
10. bwr.read_buffer = 0;
11. res = ioctl(bs->fd, BINDER_WRITE_READ, &bwr);
12. if (res < 0) {
13. fprintf(stderr,"binder_write: ioctl failed (%s) ",
14. strerror(errno));
15. }
16. return res;
17. }
   这里可以看出，只有写操作，没有读操作，即read_size为0。

   这里又是一个ioctl的BINDER_WRITE_READ操作。直入到驱动程序的binder_ioctl函数后，执行BINDER_WRITE_READ命令，这里就不累述了。

   最后，从binder_ioctl执行到binder_thread_write函数，我们首先看第一个命令BC_FREE_BUFFER：
[cpp] view plain copy
1. int
2. binder_thread_write(struct binder_proc *proc, struct binder_thread *thread,
3. void __user *buffer, int size, signed long *consumed)
4. {
5. uint32_t cmd;
6. void __user *ptr = buffer + *consumed;
7. void __user *end = buffer + size;
9. while (ptr < end && thread->return_error == BR_OK) {
10. if (get_user(cmd, (uint32_t __user *)ptr))
11. return -EFAULT;
12. ptr += sizeof(uint32_t);
13. if (_IOC_NR(cmd) < ARRAY_SIZE(binder_stats.bc)) {
14. binder_stats.bc[_IOC_NR(cmd)]++;
15. proc->stats.bc[_IOC_NR(cmd)]++;
16. thread->stats.bc[_IOC_NR(cmd)]++;
17. }
18. switch (cmd) {
19. ......
20. case BC_FREE_BUFFER: {
21. void __user *data_ptr;
22. struct binder_buffer *buffer;
24. if (get_user(data_ptr, (void * __user *)ptr))
25. return -EFAULT;
26. ptr += sizeof(void *);
28. buffer = binder_buffer_lookup(proc, data_ptr);
29. if (buffer == NULL) {
30. binder_user_error("binder: %d:%d "
31. "BC_FREE_BUFFER u%p no match ",
32. proc->pid, thread->pid, data_ptr);
33. break;
34. }
35. if (!buffer->allow_user_free) {
36. binder_user_error("binder: %d:%d "
37. "BC_FREE_BUFFER u%p matched "
38. "unreturned buffer ",
39. proc->pid, thread->pid, data_ptr);
40. break;
41. }
42. if (binder_debug_mask & BINDER_DEBUG_FREE_BUFFER)
43. printk(KERN_INFO "binder: %d:%d BC_FREE_BUFFER u%p found buffer %d for %s transaction ",
44. proc->pid, thread->pid, data_ptr, buffer->debug_id,
45. buffer->transaction ? "active" : "finished");
47. if (buffer->transaction) {
48. buffer->transaction->buffer = NULL;
49. buffer->transaction = NULL;
50. }
51. if (buffer->async_transaction && buffer->target_node) {
52. BUG_ON(!buffer->target_node->has_async_transaction);
53. if (list_empty(&buffer->target_node->async_todo))
54. buffer->target_node->has_async_transaction = 0;
55. else
56. list_move_tail(buffer->target_node->async_todo.next, &thread->todo);
57. }
58. binder_transaction_buffer_release(proc, buffer, NULL);
59. binder_free_buf(proc, buffer);
60. break;
61. }
63. ......
64. *consumed = ptr - buffer;
65. }
66. return 0;
67. }
首先通过看这个语句：
[cpp] view plain copy
1. get_user(data_ptr, (void * __user *)ptr)
这个是获得要删除的Buffer的用户空间地址，接着通过下面这个语句来找到这个地址对应的struct binder_buffer信息：
[cpp] view plain copy
1. buffer = binder_buffer_lookup(proc, data_ptr);
因为这个空间是前面在binder_transaction里面分配的，所以这里一定能找到。

最后，就可以释放这块内存了：
[cpp] view plain copy
1. binder_transaction_buffer_release(proc, buffer, NULL);
2. binder_free_buf(proc, buffer);
再来看另外一个命令BC_REPLY：
[cpp] view plain copy
1. int
2. binder_thread_write(struct binder_proc *proc, struct binder_thread *thread,
3. void __user *buffer, int size, signed long *consumed)
4. {
5. uint32_t cmd;
6. void __user *ptr = buffer + *consumed;
7. void __user *end = buffer + size;
9. while (ptr < end && thread->return_error == BR_OK) {
10. if (get_user(cmd, (uint32_t __user *)ptr))
11. return -EFAULT;
12. ptr += sizeof(uint32_t);
13. if (_IOC_NR(cmd) < ARRAY_SIZE(binder_stats.bc)) {
14. binder_stats.bc[_IOC_NR(cmd)]++;
15. proc->stats.bc[_IOC_NR(cmd)]++;
16. thread->stats.bc[_IOC_NR(cmd)]++;
17. }
18. switch (cmd) {
19. ......
20. case BC_TRANSACTION:
21. case BC_REPLY: {
22. struct binder_transaction_data tr;
24. if (copy_from_user(&tr, ptr, sizeof(tr)))
25. return -EFAULT;
26. ptr += sizeof(tr);
27. binder_transaction(proc, thread, &tr, cmd == BC_REPLY);
28. break;
29. }
31. ......
32. *consumed = ptr - buffer;
33. }
34. return 0;
35. }
又再次进入到binder_transaction函数：
[cpp] view plain copy
1. static void
2. binder_transaction(struct binder_proc *proc, struct binder_thread *thread,
3. struct binder_transaction_data *tr, int reply)
4. {
5. struct binder_transaction *t;
6. struct binder_work *tcomplete;
7. size_t *offp, *off_end;
8. struct binder_proc *target_proc;
9. struct binder_thread *target_thread = NULL;
10. struct binder_node *target_node = NULL;
11. struct list_head *target_list;
12. wait_queue_head_t *target_wait;
13. struct binder_transaction *in_reply_to = NULL;
14. struct binder_transaction_log_entry *e;
15. uint32_t return_error;
17. ......
19. if (reply) {
20. in_reply_to = thread->transaction_stack;
21. if (in_reply_to == NULL) {
22. ......
23. return_error = BR_FAILED_REPLY;
24. goto err_empty_call_stack;
25. }
26. binder_set_nice(in_reply_to->saved_priority);
27. if (in_reply_to->to_thread != thread) {
28. .......
29. goto err_bad_call_stack;
30. }
31. thread->transaction_stack = in_reply_to->to_parent;
32. target_thread = in_reply_to->from;
33. if (target_thread == NULL) {
34. return_error = BR_DEAD_REPLY;
35. goto err_dead_binder;
36. }
37. if (target_thread->transaction_stack != in_reply_to) {
38. ......
39. return_error = BR_FAILED_REPLY;
40. in_reply_to = NULL;
41. target_thread = NULL;
42. goto err_dead_binder;
43. }
44. target_proc = target_thread->proc;
45. } else {
46. ......
47. }
48. if (target_thread) {
49. e->to_thread = target_thread->pid;
50. target_list = &target_thread->todo;
51. target_wait = &target_thread->wait;
52. } else {
53. ......
54. }
57. /* TODO: reuse incoming transaction for reply */
58. t = kzalloc(sizeof(*t), GFP_KERNEL);
59. if (t == NULL) {
60. return_error = BR_FAILED_REPLY;
61. goto err_alloc_t_failed;
62. }
65. tcomplete = kzalloc(sizeof(*tcomplete), GFP_KERNEL);
66. if (tcomplete == NULL) {
67. return_error = BR_FAILED_REPLY;
68. goto err_alloc_tcomplete_failed;
69. }
71. if (!reply && !(tr->flags & TF_ONE_WAY))
72. t->from = thread;
73. else
74. t->from = NULL;
75. t->sender_euid = proc->tsk->cred->euid;
76. t->to_proc = target_proc;
77. t->to_thread = target_thread;
78. t->code = tr->code;
79. t->flags = tr->flags;
80. t->priority = task_nice(current);
81. t->buffer = binder_alloc_buf(target_proc, tr->data_size,
82. tr->offsets_size, !reply && (t->flags & TF_ONE_WAY));
83. if (t->buffer == NULL) {
84. return_error = BR_FAILED_REPLY;
85. goto err_binder_alloc_buf_failed;
86. }
87. t->buffer->allow_user_free = 0;
88. t->buffer->debug_id = t->debug_id;
89. t->buffer->transaction = t;
90. t->buffer->target_node = target_node;
91. if (target_node)
92. binder_inc_node(target_node, 1, 0, NULL);
94. offp = (size_t *)(t->buffer->data + ALIGN(tr->data_size, sizeof(void *)));
96. if (copy_from_user(t->buffer->data, tr->data.ptr.buffer, tr->data_size)) {
97. binder_user_error("binder: %d:%d got transaction with invalid "
98. "data ptr ", proc->pid, thread->pid);
99. return_error = BR_FAILED_REPLY;
100. goto err_copy_data_failed;
101. }
102. if (copy_from_user(offp, tr->data.ptr.offsets, tr->offsets_size)) {
103. binder_user_error("binder: %d:%d got transaction with invalid "
104. "offsets ptr ", proc->pid, thread->pid);
105. return_error = BR_FAILED_REPLY;
106. goto err_copy_data_failed;
107. }
109. ......
111. if (reply) {
112. BUG_ON(t->buffer->async_transaction != 0);
113. binder_pop_transaction(target_thread, in_reply_to);
114. } else if (!(t->flags & TF_ONE_WAY)) {
115. ......
116. } else {
117. ......
118. }
119. t->work.type = BINDER_WORK_TRANSACTION;
120. list_add_tail(&t->work.entry, target_list);
121. tcomplete->type = BINDER_WORK_TRANSACTION_COMPLETE;
122. list_add_tail(&tcomplete->entry, &thread->todo);
123. if (target_wait)
124. wake_up_interruptible(target_wait);
125. return;
126. ......
127. }
注意，这里的reply为1，我们忽略掉其它无关代码。

前面Service Manager正在binder_thread_read函数中被MediaPlayerService启动后进程唤醒后，在最后会把当前处理完的事务放在thread->transaction_stack中：
[cpp] view plain copy
1. if (cmd == BR_TRANSACTION && !(t->flags & TF_ONE_WAY)) {
2. t->to_parent = thread->transaction_stack;
3. t->to_thread = thread;
4. thread->transaction_stack = t;
5. }
所以，这里，首先是把它这个binder_transaction取回来，并且放在本地变量in_reply_to中：
[cpp] view plain copy
1. in_reply_to = thread->transaction_stack;
接着就可以通过in_reply_to得到最终发出这个事务请求的线程和进程：
[cpp] view plain copy
1. target_thread = in_reply_to->from;
2. target_proc = target_thread->proc;
然后得到target_list和target_wait：
[cpp] view plain copy
1. target_list = &target_thread->todo;
2. target_wait = &target_thread->wait;
下面这一段代码：
[cpp] view plain copy
1. /* TODO: reuse incoming transaction for reply */
2. t = kzalloc(sizeof(*t), GFP_KERNEL);
3. if (t == NULL) {
4. return_error = BR_FAILED_REPLY;
5. goto err_alloc_t_failed;
6. }
9. tcomplete = kzalloc(sizeof(*tcomplete), GFP_KERNEL);
10. if (tcomplete == NULL) {
11. return_error = BR_FAILED_REPLY;
12. goto err_alloc_tcomplete_failed;
13. }
15. if (!reply && !(tr->flags & TF_ONE_WAY))
16. t->from = thread;
17. else
18. t->from = NULL;
19. t->sender_euid = proc->tsk->cred->euid;
20. t->to_proc = target_proc;
21. t->to_thread = target_thread;
22. t->code = tr->code;
23. t->flags = tr->flags;
24. t->priority = task_nice(current);
25. t->buffer = binder_alloc_buf(target_proc, tr->data_size,
26. tr->offsets_size, !reply && (t->flags & TF_ONE_WAY));
27. if (t->buffer == NULL) {
28. return_error = BR_FAILED_REPLY;
29. goto err_binder_alloc_buf_failed;
30. }
31. t->buffer->allow_user_free = 0;
32. t->buffer->debug_id = t->debug_id;
33. t->buffer->transaction = t;
34. t->buffer->target_node = target_node;
35. if (target_node)
36. binder_inc_node(target_node, 1, 0, NULL);
38. offp = (size_t *)(t->buffer->data + ALIGN(tr->data_size, sizeof(void *)));
40. if (copy_from_user(t->buffer->data, tr->data.ptr.buffer, tr->data_size)) {
41. binder_user_error("binder: %d:%d got transaction with invalid "
42. "data ptr ", proc->pid, thread->pid);
43. return_error = BR_FAILED_REPLY;
44. goto err_copy_data_failed;
45. }
46. if (copy_from_user(offp, tr->data.ptr.offsets, tr->offsets_size)) {
47. binder_user_error("binder: %d:%d got transaction with invalid "
48. "offsets ptr ", proc->pid, thread->pid);
49. return_error = BR_FAILED_REPLY;
50. goto err_copy_data_failed;
51. }
我们在前面已经分析过了，这里不再重复。但是有一点要注意的是，这里target_node为NULL，因此，t->buffer->target_node也为NULL。

函数本来有一个for循环，用来处理数据中的Binder对象，这里由于没有Binder对象，所以就略过了。到了下面这句代码：
[cpp] view plain copy
1. binder_pop_transaction(target_thread, in_reply_to);
我们看看做了什么事情：
[cpp] view plain copy
1. static void
2. binder_pop_transaction(
3. struct binder_thread *target_thread, struct binder_transaction *t)
4. {
5. if (target_thread) {
6. BUG_ON(target_thread->transaction_stack != t);
7. BUG_ON(target_thread->transaction_stack->from != target_thread);
8. target_thread->transaction_stack =
9. target_thread->transaction_stack->from_parent;
10. t->from = NULL;
11. }
12. t->need_reply = 0;
13. if (t->buffer)
14. t->buffer->transaction = NULL;
15. kfree(t);
16. binder_stats.obj_deleted[BINDER_STAT_TRANSACTION]++;
17. }
由于到了这里，已经不需要in_reply_to这个transaction了，就把它删掉。

回到binder_transaction函数：
[cpp] view plain copy
1. t->work.type = BINDER_WORK_TRANSACTION;
2. list_add_tail(&t->work.entry, target_list);
3. tcomplete->type = BINDER_WORK_TRANSACTION_COMPLETE;
4. list_add_tail(&tcomplete->entry, &thread->todo);
和前面一样，分别把t和tcomplete分别放在target_list和thread->todo队列中，这里的target_list指的就是最初调用IServiceManager::addService的MediaPlayerService的Server主线程的的 thread->todo队列了，而thread->todo指的是Service Manager中用来回复IServiceManager::addService请求的线程。

最后，唤醒等待在target_wait队列上的线程了，就是最初调用IServiceManager::addService的 MediaPlayerService的Server主线程了，它最后在binder_thread_read函数中睡眠在 thread->wait上，就是这里的target_wait了：
[cpp] view plain copy
1. if (target_wait)
2. wake_up_interruptible(target_wait);
这样，Service Manger回复调用IServiceManager::addService请求就算完成了，重新回到frameworks/base/cmds /servicemanager/binder.c文件中的binder_loop函数等待下一个Client请求的到来。事实上，Service Manger回到binder_loop函数再次执行ioctl函数时候，又会再次进入到binder_thread_read函数。这时个会发现 thread->todo不为空，这是因为刚才我们调用了：
[cpp] view plain copy
1. list_add_tail(&tcomplete->entry, &thread->todo);
把一个工作项tcompelete放在了在thread->todo中，这个tcompelete的type为BINDER_WORK_TRANSACTION_COMPLETE，因此，Binder驱动程序会执行下面操作：
[cpp] view plain copy
1. switch (w->type) {
2. case BINDER_WORK_TRANSACTION_COMPLETE: {
3. cmd = BR_TRANSACTION_COMPLETE;
4. if (put_user(cmd, (uint32_t __user *)ptr))
5. return -EFAULT;
6. ptr += sizeof(uint32_t);
8. list_del(&w->entry);
9. kfree(w);
11. } break;
12. ......
13. }
binder_loop函数执行完这个ioctl调用后，才会在下一次调用ioctl进入到Binder驱动程序进入休眠状态，等待下一次Client的请求。

上面讲到调用IServiceManager::addService的MediaPlayerService的Server主线程被唤醒了，于是，重新执行binder_thread_read函数：
[cpp] view plain copy
1. static int
2. binder_thread_read(struct binder_proc *proc, struct binder_thread *thread,
3. void __user *buffer, int size, signed long *consumed, int non_block)
4. {
5. void __user *ptr = buffer + *consumed;
6. void __user *end = buffer + size;
8. int ret = 0;
9. int wait_for_proc_work;
11. if (*consumed == 0) {
12. if (put_user(BR_NOOP, (uint32_t __user *)ptr))
13. return -EFAULT;
14. ptr += sizeof(uint32_t);
15. }
17. retry:
18. wait_for_proc_work = thread->transaction_stack == NULL && list_empty(&thread->todo);
20. ......
22. if (wait_for_proc_work) {
23. ......
24. } else {
25. if (non_block) {
26. if (!binder_has_thread_work(thread))
27. ret = -EAGAIN;
28. } else
29. ret = wait_event_interruptible(thread->wait, binder_has_thread_work(thread));
30. }
32. ......
34. while (1) {
35. uint32_t cmd;
36. struct binder_transaction_data tr;
37. struct binder_work *w;
38. struct binder_transaction *t = NULL;
40. if (!list_empty(&thread->todo))
41. w = list_first_entry(&thread->todo, struct binder_work, entry);
42. else if (!list_empty(&proc->todo) && wait_for_proc_work)
43. w = list_first_entry(&proc->todo, struct binder_work, entry);
44. else {
45. if (ptr - buffer == 4 && !(thread->looper & BINDER_LOOPER_STATE_NEED_RETURN)) /* no data added */
46. goto retry;
47. break;
48. }
50. ......
52. switch (w->type) {
53. case BINDER_WORK_TRANSACTION: {
54. t = container_of(w, struct binder_transaction, work);
55. } break;
56. ......
57. }
59. if (!t)
60. continue;
62. BUG_ON(t->buffer == NULL);
63. if (t->buffer->target_node) {
64. ......
65. } else {
66. tr.target.ptr = NULL;
67. tr.cookie = NULL;
68. cmd = BR_REPLY;
69. }
70. tr.code = t->code;
71. tr.flags = t->flags;
72. tr.sender_euid = t->sender_euid;
74. if (t->from) {
75. ......
76. } else {
77. tr.sender_pid = 0;
78. }
80. tr.data_size = t->buffer->data_size;
81. tr.offsets_size = t->buffer->offsets_size;
82. tr.data.ptr.buffer = (void *)t->buffer->data + proc->user_buffer_offset;
83. tr.data.ptr.offsets = tr.data.ptr.buffer + ALIGN(t->buffer->data_size, sizeof(void *));
85. if (put_user(cmd, (uint32_t __user *)ptr))
86. return -EFAULT;
87. ptr += sizeof(uint32_t);
88. if (copy_to_user(ptr, &tr, sizeof(tr)))
89. return -EFAULT;
90. ptr += sizeof(tr);
92. ......
94. list_del(&t->work.entry);
95. t->buffer->allow_user_free = 1;
96. if (cmd == BR_TRANSACTION && !(t->flags & TF_ONE_WAY)) {
97. ......
98. } else {
99. t->buffer->transaction = NULL;
100. kfree(t);
101. binder_stats.obj_deleted[BINDER_STAT_TRANSACTION]++;
102. }
103. break;
104. }
106. done:
107. ......
108. return 0;
109. }
在while循环中，从thread->todo得到w，w->type为BINDER_WORK_TRANSACTION，于是，得到t。从上面可以知道，Service Manager反回了一个0回来，写在t->buffer->data里面，现在把t->buffer->data加上 proc->user_buffer_offset，得到用户空间地址，保存在tr.data.ptr.buffer里面，这样用户空间就可以访问这个返回码了。由于cmd不等于BR_TRANSACTION，这时就可以把t删除掉了，因为以后都不需要用了。

执行完这个函数后，就返回到binder_ioctl函数，执行下面语句，把数据返回给用户空间：
[cpp] view plain copy
1. if (copy_to_user(ubuf, &bwr, sizeof(bwr))) {
2. ret = -EFAULT;
3. goto err;
4. }
接着返回到用户空间IPCThreadState::talkWithDriver函数，最后返回到IPCThreadState::waitForResponse函数，最终执行到下面语句：
[cpp] view plain copy
1. status_t IPCThreadState::waitForResponse(Parcel *reply, status_t *acquireResult)
2. {
3. int32_t cmd;
4. int32_t err;
6. while (1) {
7. if ((err=talkWithDriver()) < NO_ERROR) break;
9. ......
11. cmd = mIn.readInt32();
13. ......
15. switch (cmd) {
16. ......
17. case BR_REPLY:
18. {
19. binder_transaction_data tr;
20. err = mIn.read(&tr, sizeof(tr));
21. LOG_ASSERT(err == NO_ERROR, "Not enough command data for brREPLY");
22. if (err != NO_ERROR) goto finish;
24. if (reply) {
25. if ((tr.flags & TF_STATUS_CODE) == 0) {
26. reply->ipcSetDataReference(
27. reinterpret_cast<const uint8_t*>(tr.data.ptr.buffer),
28. tr.data_size,
29. reinterpret_cast<const size_t*>(tr.data.ptr.offsets),
30. tr.offsets_size/sizeof(size_t),
31. freeBuffer, this);
32. } else {
33. ......
34. }
35. } else {
36. ......
37. }
38. }
39. goto finish;
41. ......
42. }
43. }
45. finish:
46. ......
47. return err;
48. }
注意，这里的tr.flags等于0，这个是在上面的binder_send_reply函数里设置的。最终把结果保存在reply了：
[cpp] view plain copy
1. reply->ipcSetDataReference(
2. reinterpret_cast<const uint8_t*>(tr.data.ptr.buffer),
3. tr.data_size,
4. reinterpret_cast<const size_t*>(tr.data.ptr.offsets),
5. tr.offsets_size/sizeof(size_t),
6. freeBuffer, this);
   这个函数我们就不看了，有兴趣的读者可以研究一下。

   从这里层层返回，最后回到MediaPlayerService::instantiate函数中。

   至此，IServiceManager::addService终于执行完毕了。这个过程非常复杂，但是如果我们能够深刻地理解这一过程，将能很好地理解 Binder机制的设计思想和实现过程。这里，对IServiceManager::addService过程中MediaPlayerService、 ServiceManager和BinderDriver之间的交互作一个小结：

   回到frameworks/base/media/mediaserver/main_mediaserver.cpp文件中的main函数，接下去还要执行下面两个函数：
[cpp] view plain copy
1. ProcessState::self()->startThreadPool();
2. IPCThreadState::self()->joinThreadPool();
首先看ProcessState::startThreadPool函数的实现：
[cpp] view plain copy
1. void ProcessState::startThreadPool()
2. {
3. AutoMutex _l(mLock);
4. if (!mThreadPoolStarted) {
5. mThreadPoolStarted = true;
6. spawnPooledThread(true);
7. }
8. }
这里调用spwanPooledThread：
[cpp] view plain copy
1. void ProcessState::spawnPooledThread(bool isMain)
2. {
3. if (mThreadPoolStarted) {
4. int32_t s = android_atomic_add(1, &mThreadPoolSeq);
5. char buf[32];
6. sprintf(buf, "Binder Thread #%d", s);
7. LOGV("Spawning new pooled thread, name=%s ", buf);
8. sp<Thread> t = new PoolThread(isMain);
9. t->run(buf);
10. }
11. }
这里主要是创建一个线程，PoolThread继续Thread类，Thread类定义在frameworks/base/libs/utils /Threads.cpp文件中，其run函数最终调用子类的threadLoop函数，这里即为PoolThread::threadLoop函数：
[cpp] view plain copy
1. virtual bool threadLoop()
2. {
3. IPCThreadState::self()->joinThreadPool(mIsMain);
4. return false;
5. }
这里和frameworks/base/media/mediaserver/main_mediaserver.cpp文件中的main函数一样，最终都是调用了IPCThreadState::joinThreadPool函数，它们的区别是，一个参数是true，一个是默认值false。我们来看一下这个函数的实现：
[cpp] view plain copy
1. void IPCThreadState::joinThreadPool(bool isMain)
2. {
3. LOG_THREADPOOL("**** THREAD %p (PID %d) IS JOINING THE THREAD POOL ", (void*)pthread_self(), getpid());
5. mOut.writeInt32(isMain ? BC_ENTER_LOOPER : BC_REGISTER_LOOPER);
7. ......
9. status_t result;
10. do {
11. int32_t cmd;
13. .......
15. // now get the next command to be processed, waiting if necessary
16. result = talkWithDriver();
17. if (result >= NO_ERROR) {
18. size_t IN = mIn.dataAvail();
19. if (IN < sizeof(int32_t)) continue;
20. cmd = mIn.readInt32();
21. ......
22. }
24. result = executeCommand(cmd);
25. }
27. ......
28. } while (result != -ECONNREFUSED && result != -EBADF);
30. .......
32. mOut.writeInt32(BC_EXIT_LOOPER);
33. talkWithDriver(false);
34. }
这个函数最终是在一个无穷循环中，通过调用talkWithDriver函数来和Binder驱动程序进行交互，实际上就是调用 talkWithDriver来等待Client的请求，然后再调用executeCommand来处理请求，而在executeCommand函数中，最终会调用BBinder::transact来真正处理Client的请求：
[cpp] view plain copy
1. status_t IPCThreadState::executeCommand(int32_t cmd)
2. {
3. BBinder* obj;
4. RefBase::weakref_type* refs;
5. status_t result = NO_ERROR;
7. switch (cmd) {
8. ......
10. case BR_TRANSACTION:
11. {
12. binder_transaction_data tr;
13. result = mIn.read(&tr, sizeof(tr));
15. ......
17. Parcel reply;
19. ......
21. if (tr.target.ptr) {
22. sp<BBinder> b((BBinder*)tr.cookie);
23. const status_t error = b->transact(tr.code, buffer, &reply, tr.flags);
24. if (error < NO_ERROR) reply.setError(error);
26. } else {
27. const status_t error = the_context_object->transact(tr.code, buffer, &reply, tr.flags);
28. if (error < NO_ERROR) reply.setError(error);
29. }
31. ......
32. }
33. break;
35. .......
36. }
38. if (result != NO_ERROR) {
39. mLastError = result;
40. }
42. return result;
43. }
接下来再看一下BBinder::transact的实现：
[cpp] view plain copy
1. status_t BBinder::transact(
2. uint32_t code, const Parcel& data, Parcel* reply, uint32_t flags)
3. {
4. data.setDataPosition(0);
6. status_t err = NO_ERROR;
7. switch (code) {
8. case PING_TRANSACTION:
9. reply->writeInt32(pingBinder());
10. break;
11. default:
12. err = onTransact(code, data, reply, flags);
13. break;
14. }
16. if (reply != NULL) {
17. reply->setDataPosition(0);
18. }
20. return err;
21. }
最终会调用onTransact函数来处理。在这个场景中，BnMediaPlayerService继承了BBinder类，并且重载了 onTransact函数，因此，这里实际上是调用了BnMediaPlayerService::onTransact函数，这个函数定义在 frameworks/base/libs/media/libmedia/IMediaPlayerService.cpp文件中：
[cpp] view plain copy
1. status_t BnMediaPlayerService::onTransact(
2. uint32_t code, const Parcel& data, Parcel* reply, uint32_t flags)
3. {
4. switch(code) {
5. case CREATE_URL: {
6. ......
7. } break;
8. case CREATE_FD: {
9. ......
10. } break;
11. case DECODE_URL: {
12. ......
13. } break;
14. case DECODE_FD: {
15. ......
16. } break;
17. case CREATE_MEDIA_RECORDER: {
18. ......
19. } break;
20. case CREATE_METADATA_RETRIEVER: {
21. ......
22. } break;
23. case GET_OMX: {
24. ......
25. } break;
26. default:
27. return BBinder::onTransact(code, data, reply, flags);
28. }
29. }
至此，我们就以MediaPlayerService为例，完整地介绍了Android系统进程间通信Binder机制中的Server启动过程。 Server启动起来之后，就会在一个无穷循环中等待Client的请求了。在下一篇文章中，我们将介绍Client如何通过Service Manager远程接口来获得Server远程接口，进而调用Server远程接口来使用Server提供的服务，敬请关注。

老罗的新浪微博：http://weibo.com/shengyangluo，欢迎关注！
相关阅读:
silverlight 会越来越好
 如何在C#里实现端口监视呢？
我就这么活着
 无意间看到的两句话
 唉，心情
 有一种美丽的缘分，叫错过！
HTML5之Canvas绘图——阴影效果呈现方法
 jQuery——动态给表格添加序号
 PHP代码——Curl实现网页代理proxy
HTML5之Canvas绘图——Canvas画布调整之移动、缩放、旋转
原文地址：https://www.cnblogs.com/Free-Thinker/p/4142093.html