TF_NewTensor Segmentation Fault: Possible Bug?

Question

I'm using Tensorflow 2.1 git master branch (commit id:db8a74a737cc735bb2a4800731d21f2de6d04961) and compile it locally. Playing around with the C API to call TF_LoadSessionFromSavedModel but seems to get segmentation fault. I've managed to drill down the error in the sample code below.

TF_NewTensor call is crashing and causing a segmentation fault.

int main()
{
    TF_Tensor** InputValues = (TF_Tensor**)malloc(sizeof(TF_Tensor*)*1);
    int ndims = 1;
    int64_t* dims = malloc(sizeof(int64_t));
    int ndata = sizeof(int32_t);
    int32_t* data = malloc(sizeof(int32_t));
    dims[0] = 1;
    data[0] = 10;

    // Crash on the next line
    TF_Tensor* int_tensor = TF_NewTensor(TF_INT32, dims, ndims, data, ndata, NULL, NULL);
    if(int_tensor == NULL)
    {
        printf("ERROR");
    }
    else
    {
        printf("OK");
    }
    return 0;
}

However, when i move the TF_Tensor** InputValues = (TF_Tensor**)malloc(sizeof(TF_Tensor*)*1); after the TF_NewTensor call, it doesn't crash. Like below:

int main()
{

    int ndims = 1;
    int64_t* dims = malloc(sizeof(int64_t));
    int ndata = sizeof(int32_t);
    int32_t* data = malloc(sizeof(int32_t));
    dims[0] = 1;
    data[0] = 10;

    // NO more crash
    TF_Tensor* int_tensor = TF_NewTensor(TF_INT32, dims, ndims, data, ndata, NULL, NULL);
    if(int_tensor == NULL)
    {
        printf("ERROR");
    }
    else
    {
        printf("OK");
    }

    TF_Tensor** InputValues = (TF_Tensor**)malloc(sizeof(TF_Tensor*)*1);


    return 0;
}

Is it a possible bug or I'm using it wrong? I don't understand how mallocq an independent variable could cause a segmentation fault.

can anybody reproduce?

I'm using gcc (Ubuntu 9.2.1-9ubuntu2) 9.2.1 20191008 to compile.

UPDATE:

can be further simplified the error as below. This is even without the InputValues being allocated.

#include <stdlib.h>
#include <stdio.h>
#include "tensorflow/c/c_api.h"

int main()
{
    int ndims = 1;
    int ndata = 1;

    int64_t dims[] = { 1 };
    int32_t data[] = { 10 };

    TF_Tensor* int_tensor = TF_NewTensor(TF_INT32, dims, ndims, data, ndata, NULL, NULL);
    if(int_tensor == NULL)
    {
        printf("ERROR Tensor");
    }
    else
    {
        printf("OK");
    }

    return 0;
}

compile with

gcc -I<tensorflow_path>/include/ -L<tensorflow_path>/lib test.c -ltensorflow -o test2.out

Solution

As point up by Raz, pass empty deallocater instead of NULL, and ndata should be size in terms of byte.

#include "tensorflow/c/c_api.h"
void NoOpDeallocator(void* data, size_t a, void* b) {}

int main(){
    int ndims = 2;
    int64_t dims[] = {1,1};
    int64_t data[] = {20};
    int ndata = sizeof(int64_t); // This is tricky, it number of bytes not number of element

    TF_Tensor* int_tensor = TF_NewTensor(TF_INT64, dims, ndims, data, ndata, &NoOpDeallocator, 0);
    if (int_tensor != NULL)\
        printf("TF_NewTensor is OK\n");
    else
        printf("ERROR: Failed TF_NewTensor\n");
}

checkout my Github on full code of running/compile TensorFlow's C API here

Answer 1

You set ndata to be sizeof(int32_t) which is 4. Your ndata is passed as len argument to TF_NewTensor() which represents the number of elements in data (can be seen in GitHub). Therefore, it should be set to 1 in your example, as you have a single element.

By the way, you can avoid using malloc() here (as you don't check for return values, and this may be error-pront and less elegant in general) and just use local variables instead.

UPDATE

In addition, you pass NULL both for deallocator and deallocator_arg. I'm pretty sure this is the issue as the comment states "Clients must provide a custom deallocator function..." (can be seen here). The deallocator is called by the TF_NewTensor() (can be seen here) and this may be the cause for the segmentation fault.

So, summing it all up, try the next code:

void my_deallocator(void * data, size_t len, void * arg)
{
    printf("Deallocator called with data %p\n", data);
}

void main()
{
    int64_t dims[] = { 1 };
    int32_t data[] = { 10 };

    ... = TF_NewTensor(TF_INT32, dims, /*num_dims=*/ 1, data, /*len=*/ 1, my_deallocator, NULL);
}