android 官方文档 JNI TIPS
文章地址 http://developer.android.com/training/articles/perf-jni.html
JNI Tips
JNI is the Java Native Interface. It defines a way for managed code (written in the Java programming language) to interact with native code (written in C/C++). It's vendor-neutral, has support for loading code from dynamic shared libraries, and while cumbersome at times is reasonably efficient.
If you're not already familiar with it, read through the Java Native Interface Specification to get a sense for how JNI works and what features are available.Some aspects of the interface aren't immediately obvious on first reading, so you may find the next few sections handy.
JavaVM and JNIEnv
JNI defines two key data structures, "JavaVM" and "JNIEnv". Both of these are essentially pointers to pointers to function tables. (In the C++ version, they're classes with a pointer to a function table and a member function for each JNI function that indirects through the table.) The JavaVM provides the "invocation interface" functions, which allow you to create and destroy a JavaVM. In theory you can have multiple JavaVMs per process, but Android only allows one.
The JNIEnv provides most of the JNI functions. Your native functions all receive a JNIEnv as the first argument.
The JNIEnv is used for thread-local storage. For this reason, you cannot share a JNIEnv between threads. If a piece of code has no other way to get its JNIEnv, you should share the JavaVM, and use GetEnv to discover the thread's JNIEnv.(Assuming it has one; see AttachCurrentThread below.)
The C declarations of JNIEnv and JavaVM are different from the C++ declarations. The "jni.h" include file provides different typedefs depending on whether it's included into C or C++. For this reason it's a bad idea to include JNIEnv arguments in header files included by both languages. (Put another way: if your header file requires #ifdef __cplusplus, you may have to do some extra work if anything in that header refers to JNIEnv.)
Threads
All threads are Linux threads, scheduled by the kernel. They're usually started from managed code (using Thread.start), but they can also be created elsewhere and then attached to the JavaVM. For example, a thread started withpthread_create can be attached with the JNI AttachCurrentThread orAttachCurrentThreadAsDaemon functions. Until a thread is attached, it has no JNIEnv, and cannot make JNI calls.
Attaching a natively-created thread causes a java.lang.Thread object to be constructed and added to the "main" ThreadGroup, making it visible to the debugger. Calling AttachCurrentThread on an already-attached thread is a no-op.
Android does not suspend threads executing native code. If garbage collection is in progress, or the debugger has issued a suspend request, Android will pause the thread the next time it makes a JNI call.
Threads attached through JNI must call DetachCurrentThread before they exit. If coding this directly is awkward, in Android 2.0 (Eclair) and higher you can usepthread_key_create to define a destructor function that will be called before the thread exits, and call DetachCurrentThread from there. (Use that key withpthread_setspecific to store the JNIEnv in thread-local-storage; that way it'll be passed into your destructor as the argument.)
jclass, jmethodID, and jfieldID
If you want to access an object's field from native code, you would do the following:
- Get the class object reference for the class with
FindClass - Get the field ID for the field with
GetFieldID - Get the contents of the field with something appropriate, such as
GetIntField
Similarly, to call a method, you'd first get a class object reference and then a method ID. The IDs are often just pointers to internal runtime data structures.Looking them up may require several string comparisons, but once you have them the actual call to get the field or invoke the method is very quick.
If performance is important, it's useful to look the values up once and cache the results in your native code. Because there is a limit of one JavaVM per process, it's reasonable to store this data in a static local structure.
The class references, field IDs, and method IDs are guaranteed valid until the class is unloaded. Classes are only unloaded if all classes associated with a ClassLoader can be garbage collected, which is rare but will not be impossible in Android. Note however that the jclass is a class reference and must be protected with a call to NewGlobalRef (see the next section).
If you would like to cache the IDs when a class is loaded, and automatically re-cache them if the class is ever unloaded and reloaded, the correct way to initialize the IDs is to add a piece of code that looks like this to the appropriate class:
/*
* We use a class initializer to allow the native code to cache some
* field offsets. This native function looks up and caches interesting
* class/field/method IDs. Throws on failure.
*/
private static native void nativeInit();
static {
nativeInit();
}
Create a nativeClassInit method in your C/C++ code that performs the ID lookups. The code will be executed once, when the class is initialized. If the class is ever unloaded and then reloaded, it will be executed again.
Local and Global References
Every argument passed to a native method, and almost every object returned by a JNI function is a "local reference". This means that it's valid for the duration of the current native method in the current thread. Even if the object itself continues to live on after the native method returns, the reference is not valid.
This applies to all sub-classes of jobject, including jclass, jstring, and jarray.(The runtime will warn you about most reference mis-uses when extended JNI checks are enabled.)
The only way to get non-local references is via the functions NewGlobalRef andNewWeakGlobalRef.
If you want to hold on to a reference for a longer period, you must use a "global" reference. The NewGlobalRef function takes the local reference as an argument and returns a global one. The global reference is guaranteed to be valid until you call DeleteGlobalRef.
This pattern is commonly used when caching a jclass returned from FindClass, e.g.:
jclass localClass = env->FindClass("MyClass");
jclass globalClass = reinterpret_cast<jclass>(env->NewGlobalRef(localClass));
All JNI methods accept both local and global references as arguments. It's possible for references to the same object to have different values. For example, the return values from consecutive calls to NewGlobalRef on the same object may be different. To see if two references refer to the same object, you must use the IsSameObject function. Never compare references with == in native code.
One consequence of this is that you must not assume object references are constant or unique in native code. The 32-bit value representing an object may be different from one invocation of a method to the next, and it's possible that two different objects could have the same 32-bit value on consecutive calls. Do not use jobject values as keys.
Programmers are required to "not excessively allocate" local references. In practical terms this means that if you're creating large numbers of local references, perhaps while running through an array of objects, you should free them manually with DeleteLocalRef instead of letting JNI do it for you. The implementation is only required to reserve slots for 16 local references, so if you need more than that you should either delete as you go or useEnsureLocalCapacity/PushLocalFrame to reserve more.
Note that jfieldIDs and jmethodIDs are opaque types, not object references, and should not be passed to NewGlobalRef. The raw data pointers returned by functions like GetStringUTFChars and GetByteArrayElements are also not objects.(They may be passed between threads, and are valid until the matching Release call.)
One unusual case deserves separate mention. If you attach a native thread withAttachCurrentThread, the code you are running will never automatically free local references until the thread detaches. Any local references you create will have to be deleted manually. In general, any native code that creates local references in a loop probably needs to do some manual deletion.
UTF-8 and UTF-16 Strings
The Java programming language uses UTF-16. For convenience, JNI provides methods that work with Modified UTF-8 as well. The modified encoding is useful for C code because it encodes \u0000 as 0xc0 0x80 instead of 0x00. The nice thing about this is that you can count on having C-style zero-terminated strings, suitable for use with standard libc string functions. The down side is that you cannot pass arbitrary UTF-8 data to JNI and expect it to work correctly.
If possible, it's usually faster to operate with UTF-16 strings. Android currently does not require a copy in GetStringChars, whereas GetStringUTFChars requires an allocation and a conversion to UTF-8. Note that UTF-16 strings are not zero-terminated, and \u0000 is allowed, so you need to hang on to the string length as well as the jchar pointer.
Don't forget to Release the strings you Get. The string functions return jchar* orjbyte*, which are C-style pointers to primitive data rather than local references.They are guaranteed valid until Release is called, which means they are not released when the native method returns.
Data passed to NewStringUTF must be in Modified UTF-8 format. A common mistake is reading character data from a file or network stream and handing it to NewStringUTF without filtering it. Unless you know the data is 7-bit ASCII, you need to strip out high-ASCII characters or convert them to proper Modified UTF-8 form. If you don't, the UTF-16 conversion will likely not be what you expect. The extended JNI checks will scan strings and warn you about invalid data, but they won't catch everything.
Primitive Arrays
JNI provides functions for accessing the contents of array objects. While arrays of objects must be accessed one entry at a time, arrays of primitives can be read and written directly as if they were declared in C.
To make the interface as efficient as possible without constraining the VM implementation, the Get<PrimitiveType>ArrayElements family of calls allows the runtime to either return a pointer to the actual elements, or allocate some memory and make a copy. Either way, the raw pointer returned is guaranteed to be valid until the corresponding Release call is issued (which implies that, if the data wasn't copied, the array object will be pinned down and can't be relocated as part of compacting the heap). You must Release every array you Get. Also, if the Get call fails, you must ensure that your code doesn't try to Release a NULL pointer later.
You can determine whether or not the data was copied by passing in a non-NULL pointer for the isCopy argument. This is rarely useful.
The Release call takes a mode argument that can have one of three values. The actions performed by the runtime depend upon whether it returned a pointer to the actual data or a copy of it:
0- Actual: the array object is un-pinned.
- Copy: data is copied back. The buffer with the copy is freed.
JNI_COMMIT- Actual: does nothing.
- Copy: data is copied back. The buffer with the copy is not freed.
JNI_ABORT- Actual: the array object is un-pinned. Earlier writes are not aborted.
- Copy: the buffer with the copy is freed; any changes to it are lost.
One reason for checking the isCopy flag is to know if you need to call Release withJNI_COMMIT after making changes to an array — if you're alternating between making changes and executing code that uses the contents of the array, you may be able to skip the no-op commit. Another possible reason for checking the flag is for efficient handling of JNI_ABORT. For example, you might want to get an array, modify it in place, pass pieces to other functions, and then discard the changes. If you know that JNI is making a new copy for you, there's no need to create another "editable" copy. If JNI is passing you the original, then you do need to make your own copy.
It is a common mistake (repeated in example code) to assume that you can skip the Release call if *isCopy is false. This is not the case. If no copy buffer was allocated, then the original memory must be pinned down and can't be moved by the garbage collector.
Also note that the JNI_COMMIT flag does not release the array, and you will need to call Release again with a different flag eventually.
Region Calls
There is an alternative to calls like Get<Type>ArrayElements and GetStringChars that may be very helpful when all you want to do is copy data in or out. Consider the following:
jbyte* data = env->GetByteArrayElements(array, NULL);
if (data != NULL) {
memcpy(buffer, data, len);
env->ReleaseByteArrayElements(array, data, JNI_ABORT);
}
This grabs the array, copies the first len byte elements out of it, and then releases the array. Depending upon the implementation, the Get call will either pin or copy the array contents. The code copies the data (for perhaps a second time), then calls Release; in this case JNI_ABORT ensures there's no chance of a third copy.
One can accomplish the same thing more simply:
env->GetByteArrayRegion(array, 0, len, buffer);
This has several advantages:
- Requires one JNI call instead of 2, reducing overhead.
- Doesn't require pinning or extra data copies.
- Reduces the risk of programmer error — no risk of forgetting to call
Releaseafter something fails.
Similarly, you can use the Set<Type>ArrayRegion call to copy data into an array, and GetStringRegion or GetStringUTFRegion to copy characters out of a String.
Exceptions
You must not call most JNI functions while an exception is pending. Your code is expected to notice the exception (via the function's return value,ExceptionCheck, or ExceptionOccurred) and return, or clear the exception and handle it.
The only JNI functions that you are allowed to call while an exception is pending are:
DeleteGlobalRefDeleteLocalRefDeleteWeakGlobalRefExceptionCheckExceptionClearExceptionDescribeExceptionOccurredMonitorExitPopLocalFramePushLocalFrameRelease<PrimitiveType>ArrayElementsReleasePrimitiveArrayCriticalReleaseStringCharsReleaseStringCriticalReleaseStringUTFChars
Many JNI calls can throw an exception, but often provide a simpler way of checking for failure. For example, if NewString returns a non-NULL value, you don't need to check for an exception. However, if you call a method (using a function like CallObjectMethod), you must always check for an exception, because the return value is not going to be valid if an exception was thrown.
Note that exceptions thrown by interpreted code do not unwind native stack frames, and Android does not yet support C++ exceptions. The JNI Throw andThrowNew instructions just set an exception pointer in the current thread. Upon returning to managed from native code, the exception will be noted and handled appropriately.
Native code can "catch" an exception by calling ExceptionCheck orExceptionOccurred, and clear it with ExceptionClear. As usual, discarding exceptions without handling them can lead to problems.
There are no built-in functions for manipulating the Throwable object itself, so if you want to (say) get the exception string you will need to find the Throwableclass, look up the method ID for getMessage "()Ljava/lang/String;", invoke it, and if the result is non-NULL use GetStringUTFChars to get something you can hand toprintf(3) or equivalent.
Extended Checking
JNI does very little error checking. Errors usually result in a crash. Android also offers a mode called CheckJNI, where the JavaVM and JNIEnv function table pointers are switched to tables of functions that perform an extended series of checks before calling the standard implementation.
The additional checks include:
- Arrays: attempting to allocate a negative-sized array.
- Bad pointers: passing a bad jarray/jclass/jobject/jstring to a JNI call, or passing a NULL pointer to a JNI call with a non-nullable argument.
- Class names: passing anything but the “java/lang/String” style of class name to a JNI call.
- Critical calls: making a JNI call between a “critical” get and its corresponding release.
- Direct ByteBuffers: passing bad arguments to
NewDirectByteBuffer. - Exceptions: making a JNI call while there’s an exception pending.
- JNIEnv*s: using a JNIEnv* from the wrong thread.
- jfieldIDs: using a NULL jfieldID, or using a jfieldID to set a field to a value of the wrong type (trying to assign a StringBuilder to a String field, say), or using a jfieldID for a static field to set an instance field or vice versa, or using a jfieldID from one class with instances of another class.
- jmethodIDs: using the wrong kind of jmethodID when making a
Call*MethodJNI call: incorrect return type, static/non-static mismatch, wrong type for ‘this’ (for non-static calls) or wrong class (for static calls). - References: using
DeleteGlobalRef/DeleteLocalRefon the wrong kind of reference. - Release modes: passing a bad release mode to a release call (something other than
0,JNI_ABORT, orJNI_COMMIT). - Type safety: returning an incompatible type from your native method (returning a StringBuilder from a method declared to return a String, say).
- UTF-8: passing an invalid Modified UTF-8 byte sequence to a JNI call.
(Accessibility of methods and fields is still not checked: access restrictions don't apply to native code.)
There are several ways to enable CheckJNI.
If you’re using the emulator, CheckJNI is on by default.
If you have a rooted device, you can use the following sequence of commands to restart the runtime with CheckJNI enabled:
adb shell stop
adb shell setprop dalvik.vm.checkjni true
adb shell start
In either of these cases, you’ll see something like this in your logcat output when the runtime starts:
D AndroidRuntime: CheckJNI is ON
If you have a regular device, you can use the following command:
adb shell setprop debug.checkjni 1
This won’t affect already-running apps, but any app launched from that point on will have CheckJNI enabled. (Change the property to any other value or simply rebooting will disable CheckJNI again.) In this case, you’ll see something like this in your logcat output the next time an app starts:
D Late-enabling CheckJNI
Native Libraries
You can load native code from shared libraries with the standardSystem.loadLibrary call. The preferred way to get at your native code is:
- Call
System.loadLibraryfrom a static class initializer. (See the earlier example, where one is used to callnativeClassInit.) The argument is the "undecorated" library name, so to load "libfubar.so" you would pass in "fubar". - Provide a native function:
jint JNI_OnLoad(JavaVM* vm, void* reserved) - In
JNI_OnLoad, register all of your native methods. You should declare the methods "static" so the names don't take up space in the symbol table on the device.
The JNI_OnLoad function should look something like this if written in C++:
jint JNI_OnLoad(JavaVM* vm, void* reserved)
{
JNIEnv* env;
if (vm->GetEnv(reinterpret_cast<void**>(&env), JNI_VERSION_1_6) != JNI_OK) {
return -1;
}
// Get jclass with env->FindClass.
// Register methods with env->RegisterNatives.
return JNI_VERSION_1_6;
}
You can also call System.load with the full path name of the shared library. For Android apps, you may find it useful to get the full path to the application's private data storage area from the context object.
This is the recommended approach, but not the only approach. Explicit registration is not required, nor is it necessary that you provide a JNI_OnLoadfunction. You can instead use "discovery" of native methods that are named in a specific way (see the JNI spec for details), though this is less desirable because if a method signature is wrong you won't know about it until the first time the method is actually used.
One other note about JNI_OnLoad: any FindClass calls you make from there will happen in the context of the class loader that was used to load the shared library. Normally FindClass uses the loader associated with the method at the top of the interpreted stack, or if there isn't one (because the thread was just attached) it uses the "system" class loader. This makes JNI_OnLoad a convenient place to look up and cache class object references.
64-bit Considerations
Android is currently expected to run on 32-bit platforms. In theory it could be built for a 64-bit system, but that is not a goal at this time. For the most part this isn't something that you will need to worry about when interacting with native code, but it becomes significant if you plan to store pointers to native structures in integer fields in an object. To support architectures that use 64-bit pointers, you need to stash your native pointers in a long field rather than an int.
Unsupported Features/Backwards Compatibility
All JNI 1.6 features are supported, with the following exception:
DefineClassis not implemented. Android does not use Java bytecodes or class files, so passing in binary class data doesn't work.
For backward compatibility with older Android releases, you may need to be aware of:
- Dynamic lookup of native functions
Until Android 2.0 (Eclair), the '$' character was not properly converted to "_00024" during searches for method names. Working around this requires using explicit registration or moving the native methods out of inner classes.
- Detaching threads
Until Android 2.0 (Eclair), it was not possible to use a
pthread_key_createdestructor function to avoid the "thread must be detached before exit" check.(The runtime also uses a pthread key destructor function, so it'd be a race to see which gets called first.) - Weak global references
Until Android 2.2 (Froyo), weak global references were not implemented.Older versions will vigorously reject attempts to use them. You can use the Android platform version constants to test for support.
Until Android 4.0 (Ice Cream Sandwich), weak global references could only be passed to
NewLocalRef,NewGlobalRef, andDeleteWeakGlobalRef. (The spec strongly encourages programmers to create hard references to weak globals before doing anything with them, so this should not be at all limiting.)From Android 4.0 (Ice Cream Sandwich) on, weak global references can be used like any other JNI references.
- Local references
Until Android 4.0 (Ice Cream Sandwich), local references were actually direct pointers. Ice Cream Sandwich added the indirection necessary to support better garbage collectors, but this means that lots of JNI bugs are undetectable on older releases. See JNI Local Reference Changes in ICS for more details.
- Determining reference type with
GetObjectRefTypeUntil Android 4.0 (Ice Cream Sandwich), as a consequence of the use of direct pointers (see above), it was impossible to implement
GetObjectRefTypecorrectly. Instead we used a heuristic that looked through the weak globals table, the arguments, the locals table, and the globals table in that order.The first time it found your direct pointer, it would report that your reference was of the type it happened to be examining. This meant, for example, that if you calledGetObjectRefTypeon a global jclass that happened to be the same as the jclass passed as an implicit argument to your static native method, you'd getJNILocalRefTyperather thanJNIGlobalRefType.
FAQ: Why do I get UnsatisfiedLinkError?
When working on native code it's not uncommon to see a failure like this:
java.lang.UnsatisfiedLinkError: Library foo not found
In some cases it means what it says — the library wasn't found. In other cases the library exists but couldn't be opened by dlopen(3), and the details of the failure can be found in the exception's detail message.
Common reasons why you might encounter "library not found" exceptions:
- The library doesn't exist or isn't accessible to the app. Use
adb shell ls -l <path>to check its presence and permissions. - The library wasn't built with the NDK. This can result in dependencies on functions or libraries that don't exist on the device.
Another class of UnsatisfiedLinkError failures looks like:
java.lang.UnsatisfiedLinkError: myfunc
at Foo.myfunc(Native Method)
at Foo.main(Foo.java:10)
In logcat, you'll see:
W/dalvikvm( 880): No implementation found for native LFoo;.myfunc ()V
This means that the runtime tried to find a matching method but was unsuccessful. Some common reasons for this are:
- The library isn't getting loaded. Check the logcat output for messages about library loading.
- The method isn't being found due to a name or signature mismatch. This is commonly caused by:
- For lazy method lookup, failing to declare C++ functions with
extern "C"and appropriate visibility (JNIEXPORT). Note that prior to Ice Cream Sandwich, the JNIEXPORT macro was incorrect, so using a new GCC with an oldjni.hwon't work. You can usearm-eabi-nmto see the symbols as they appear in the library; if they look mangled (something like_Z15Java_Foo_myfuncP7_JNIEnvP7_jclassrather thanJava_Foo_myfunc), or if the symbol type is a lowercase 't' rather than an uppercase 'T', then you need to adjust the declaration. - For explicit registration, minor errors when entering the method signature. Make sure that what you're passing to the registration call matches the signature in the log file. Remember that 'B' is
byteand 'Z' isboolean. Class name components in signatures start with 'L', end with ';', use '/' to separate package/class names, and use '$' to separate inner-class names (Ljava/util/Map$Entry;, say).
- For lazy method lookup, failing to declare C++ functions with
Using javah to automatically generate JNI headers may help avoid some problems.
FAQ: Why didn't FindClass find my class?
Make sure that the class name string has the correct format. JNI class names start with the package name and are separated with slashes, such asjava/lang/String. If you're looking up an array class, you need to start with the appropriate number of square brackets and must also wrap the class with 'L' and ';', so a one-dimensional array of String would be [Ljava/lang/String;.
If the class name looks right, you could be running into a class loader issue.FindClass wants to start the class search in the class loader associated with your code. It examines the call stack, which will look something like:
Foo.myfunc(Native Method)
Foo.main(Foo.java:10)
dalvik.system.NativeStart.main(Native Method)
The topmost method is Foo.myfunc. FindClass finds the ClassLoader object associated with the Foo class and uses that.
This usually does what you want. You can get into trouble if you create a thread yourself (perhaps by calling pthread_create and then attaching it withAttachCurrentThread). Now the stack trace looks like this:
dalvik.system.NativeStart.run(Native Method)
The topmost method is NativeStart.run, which isn't part of your application. If you call FindClass from this thread, the JavaVM will start in the "system" class loader instead of the one associated with your application, so attempts to find app-specific classes will fail.
There are a few ways to work around this:
- Do your
FindClasslookups once, inJNI_OnLoad, and cache the class references for later use. AnyFindClasscalls made as part of executingJNI_OnLoadwill use the class loader associated with the function that calledSystem.loadLibrary(this is a special rule, provided to make library initialization more convenient). If your app code is loading the library,FindClasswill use the correct class loader. - Pass an instance of the class into the functions that need it, by declaring your native method to take a Class argument and then passing
Foo.classin. - Cache a reference to the
ClassLoaderobject somewhere handy, and issueloadClasscalls directly. This requires some effort.
FAQ: How do I share raw data with native code?
You may find yourself in a situation where you need to access a large buffer of raw data from both managed and native code. Common examples include manipulation of bitmaps or sound samples. There are two basic approaches.
You can store the data in a byte[]. This allows very fast access from managed code. On the native side, however, you're not guaranteed to be able to access the data without having to copy it. In some implementations,GetByteArrayElements and GetPrimitiveArrayCritical will return actual pointers to the raw data in the managed heap, but in others it will allocate a buffer on the native heap and copy the data over.
The alternative is to store the data in a direct byte buffer. These can be created with java.nio.ByteBuffer.allocateDirect, or the JNI NewDirectByteBuffer function.Unlike regular byte buffers, the storage is not allocated on the managed heap, and can always be accessed directly from native code (get the address withGetDirectBufferAddress). Depending on how direct byte buffer access is implemented, accessing the data from managed code can be very slow.
The choice of which to use depends on two factors:
- Will most of the data accesses happen from code written in Java or in C/C++?
- If the data is eventually being passed to a system API, what form must it be in? (For example, if the data is eventually passed to a function that takes a byte[], doing processing in a direct
ByteBuffermight be unwise.)
If there's no clear winner, use a direct byte buffer. Support for them is built directly into JNI, and performance should improve in future releases.
