A crash found on tensorflow_jni.so when create interpreter using byteBufferMode

[zhangto...@gmail.com]

Dear tensorflow developers：

       A crash is found when using java nio to create interpreter of tensorflowLite, which make us puzzled for a long time. I hope you can help us to solve the issue, thanks & best regard. The issues will be described in detail as follows:

       Naturally, we use the function provided by tensorflow below to create model & interpreter.      

 

       NativeInterpreterWrapper(ByteBuffer modelByteBuffer) {

     errorHandle = createErrorReporter(ERROR_BUFFER_SIZE);
     modelHandle = createModelWithBuffer(modelByteBuffer, errorHandle);
     interpreterHandle = createInterpreter(modelHandle);
   }

      step 1.  We create the Input parameter modelByteBuffer. The ByteBuffer as follows:

          fileChannel.map(FileChannel.MapMode.READ_ONLY, startOffset, declaredLength);

          

          The following conditions should be satisfied at the same time, 

          1)  startOffset is not zero,  for example startOffset = 200

               If the startOffset is zero, the issue will not occur.

          2)  the byteBuffer size is very large, for example size = 80MB

               if the size is small, the issue will not occur.

      Step 2. We call the funtion NativeInterpreterWrapper with byteBuffer.

      Unfortunately, after running  step 1 & 2, the crash is occured as follows:

01-02 18:58:54.544 21135-21135/? A/DEBUG: *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** ***

01-02 18:58:54.545 21135-21135/? A/DEBUG: Build fingerprint: 'google/marlin/marlin:8.0.0/OPR3.170623.013/4397526:user/release-keys'

01-02 18:58:54.545 21135-21135/? A/DEBUG: Revision: '0'

01-02 18:58:54.545 21135-21135/? A/DEBUG: ABI: 'arm'

01-02 18:58:54.545 21135-21135/? A/DEBUG: pid: 20837, tid: 20837, name: fish.xxxxxx  >>> com.taobao.idlexxxx.xxxxxx <<<

01-02 18:58:54.545 21135-21135/? A/DEBUG: signal 11 (SIGSEGV), code 1 (SEGV_MAPERR), fault addr 0xa733f5a6

01-02 18:58:54.545 21135-21135/? A/DEBUG:     r0 a733f5a6  r1 00000004  r2 00000006  r3 00000000

01-02 18:58:54.545 21135-21135/? A/DEBUG:     r4 e0f59c58  r5 de9734b0  r6 a73f5160  r7 ffc5de20

01-02 18:58:54.545 21135-21135/? A/DEBUG:     r8 00000004  r9 00000000  sl ffc5de90  fp a733f57a

01-02 18:58:54.545 21135-21135/? A/DEBUG:     ip edee063c  sp ffc5dd90  lr edea7741  pc a73aa208  cpsr 200f1830

01-02 18:58:54.546 21135-21135/? A/DEBUG: backtrace:

01-02 18:58:54.546 21135-21135/? A/DEBUG:     #00 pc 00062208  /data/app/com.taobao.idlexxxx.xxxxxx-qw202S8jC-x2xWFXznSPlw==/lib/arm/libtensorflowlite_jni.so

01-02 18:58:54.546 21135-21135/? A/DEBUG:     #01 pc 00062d53  /data/app/com.taobao.idlexxxx.xxxxxx-qw202S8jC-x2xWFXznSPlw==/lib/arm/libtensorflowlite_jni.so

01-02 18:58:54.546 21135-21135/? A/DEBUG:     #02 pc 00006d93  /data/app/com.taobao.idlexxxx.xxxxxx-qw202S8jC-x2xWFXznSPlw==/lib/arm/libtensorflowlite_jni.so (Java_org_tensorflow_lite_NativeInterpreterWrapper_createInterpreter+50)

01-02 18:58:54.546 21135-21135/? A/DEBUG:     #03 pc 000281ef  /data/app/com.taobao.idlexxxx.xxxxxx-qw202S8jC-x2xWFXznSPlw==/oat/arm/base.odex (offset 0x1c000)

01-02 18:58:55.403 742-742/? E//system/bin/tombstoned: Tombstone written to: /data/tombstones//tombstone_03

While,  If we reallocate memory,  the issue will also not occur .The relative code is as follows:

…….

rf = new RandomAccessFile(modelPath, "r");
declaredLength = rf.length() - startOffset;
rf.seek(startOffset);
ByteBuffer byteBuffer = ByteBuffer.allocateDirect((int) declaredLength);

…….

byteBuffer.put(tfbyte);

…….

Is there a memory allocation bug? And could you tell me how to avoid the crash if not reallocate memory？ Thanks very much.