cudaMemcpy segmentation fault

I've been haunted by this error for quite a while so I decided to post it here.

This segmentation fault happened when a cudaMemcpy is called:

CurrentGrid->cdata[i] = new float[size];
cudaMemcpy(CurrentGrid->cdata[i], Grid_dev->cdata[i], size*sizeof(float),\
                cudaMemcpyDeviceToHost);

CurrentGrid and Grid_dev are pointer to a grid class object on host and device respectively and i=0 in this context. Class member cdata is a float type pointer array. For debugging, right before this cudaMemcpy call I printed out the value of each element of Grid_Dev->cdata[i], the address of CurrentGrid->cdata[i] and Grid_dev->cdata[i] and the value of size, which all looks good. But it still ends up with "Segmentation fault (core dumped)", which is the only error message. cuda-memcheck only gave "process didn't terminate successfully". I'm not able to use cuda-gdb at the moment. Any suggestion about where to go?

UPDATE: It seems now I have solved this problem by cudaMalloc another float pointer A on device and cudaMemcpy the value of Grid_dev->cdata[i] to A, and then cudaMemcpy A to host. So the segment of code written above becomes:

float * A;
cudaMalloc((void**)&A, sizeof(float));
...
...
cudaMemcpy(&A, &(Grid_dev->cdata[i]), sizeof(float *), cudaMemcpyDeviceToHost);    
CurrentGrid->cdata[i] = new float[size];
cudaMemcpy(CurrentGrid->cdata[i], A, size*sizeof(float), cudaMemcpyDeviceToHost);            

I did this because valgrind popped up "invalid read of size 8", which I thought referring to Grid_dev->cdata[i]. I checked it again with gdb, printing out the value of Grid_dev->cdata[i] being NULL. So I guess I cannot directly dereference the device pointer even in this cudaMemcpy call. But why ? According to the comment at the bottom of this thread , we should be able to dereference device pointer in cudaMemcpy function.

Also, I don't know the the underlying mechanism of how cudaMalloc and cudaMemcpy work but I think by cudaMalloc a pointer, say A here, we actually assign this pointer to point to a certain address on the device. And by cudaMemcpy the Grid_dev->cdata[i] to A as in the modified code above, we re-assign the pointer A to point to the array. Then don't we lose the track of the previous address that A pointed to when it is cudaMalloced? Could this cause memory leak or something? If yes, how should I work around this situation properly? Thanks!

For reference I put the code of the complete function in which this error happened below.

Many thanks!

__global__ void Print(grid *, int);
__global__ void Printcell(grid *, int);
void CopyDataToHost(param_t p, grid * CurrentGrid, grid * Grid_dev){

    cudaMemcpy(CurrentGrid, Grid_dev, sizeof(grid), cudaMemcpyDeviceToHost);
#if DEBUG_DEV
    cudaCheckErrors("cudaMemcpy1 error");
#endif
    printf("\nBefore copy cell data\n");
    Print<<<1,1>>>(Grid_dev, 0);            //Print out some Grid_dev information for 
    cudaDeviceSynchronize();                //debug 
    int NumberOfBaryonFields = CurrentGrid->ReturnNumberOfBaryonFields();
    int size = CurrentGrid->ReturnSize();
    int vsize = CurrentGrid->ReturnVSize();
    CurrentGrid->FieldType = NULL;
    CurrentGrid->FieldType = new int[NumberOfBaryonFields];
    printf("CurrentGrid size is %d\n", size);
    for( int i = 0; i < p.NumberOfFields; i++){
        CurrentGrid->cdata[i] = NULL;
        CurrentGrid->vdata[i] = NULL;
        CurrentGrid->cdata[i] = new float[size];
        CurrentGrid->vdata[i] = new float[vsize];

        Printcell<<<1,1>>>(Grid_dev, i);//Print out element value of Grid_dev->cdata[i]
        cudaDeviceSynchronize();        

        cudaMemcpy(CurrentGrid->cdata[i], Grid_dev->cdata[i], size*sizeof(float),\
                cudaMemcpyDeviceToHost);               //where error occurs
#if DEBUG_DEV
        cudaCheckErrors("cudaMemcpy2 error");
#endif
        printf("\nAfter copy cell data\n");
        Print<<<1,1>>>(Grid_dev, i);
        cudaDeviceSynchronize();
        cudaMemcpy(CurrentGrid->vdata[i], Grid_dev->vdata[i], vsize*sizeof(float),\
                cudaMemcpyDeviceToHost);
#if DEBUG_DEV
        cudaCheckErrors("cudaMemcpy3 error");
#endif
    }
    cudaMemcpy(CurrentGrid->FieldType, Grid_dev->FieldType,\
            NumberOfBaryonFields*sizeof(int), cudaMemcpyDeviceToHost);
#if DEBUG_DEV
    cudaCheckErrors("cudaMemcpy4 error");
#endif
}

EDIT: here is the information from valgrind, from which I'm trying to track down where the memory leak happened.

==19340== Warning: set address range perms: large range [0x800000000, 0xd00000000) (noaccess)
==19340== Warning: set address range perms: large range [0x200000000, 0x400000000) (noaccess)
==19340== Invalid read of size 8
==19340==    at 0x402C79: CopyDataToHost(param_t, grid*, grid*) (CheckDevice.cu:48)
==19340==    by 0x403646: CheckDevice(param_t, grid*, grid*) (CheckDevice.cu:186)
==19340==    by 0x40A6CD: main (Transport.cu:81)
==19340==  Address 0x2003000c0 is not stack'd, malloc'd or (recently) free'd
==19340== 
==19340== 
==19340== Process terminating with default action of signal 11 (SIGSEGV)
==19340==  Bad permissions for mapped region at address 0x2003000C0
==19340==    at 0x402C79: CopyDataToHost(param_t, grid*, grid*) (CheckDevice.cu:48)
==19340==    by 0x403646: CheckDevice(param_t, grid*, grid*) (CheckDevice.cu:186)
==19340==    by 0x40A6CD: main (Transport.cu:81)
==19340== 
==19340== HEAP SUMMARY:
==19340==     in use at exit: 2,611,365 bytes in 5,017 blocks
==19340==   total heap usage: 5,879 allocs, 862 frees, 4,332,278 bytes allocated
==19340== 
==19340== LEAK SUMMARY:
==19340==    definitely lost: 0 bytes in 0 blocks
==19340==    indirectly lost: 0 bytes in 0 blocks
==19340==      possibly lost: 37,416 bytes in 274 blocks
==19340==    still reachable: 2,573,949 bytes in 4,743 blocks
==19340==         suppressed: 0 bytes in 0 blocks
==19340== Rerun with --leak-check=full to see details of leaked memory
==19340== 
==19340== For counts of detected and suppressed errors, rerun with: -v
==19340== ERROR SUMMARY: 1 errors from 1 contexts (suppressed: 2 from 2)

Solution 1:

I believe I know what the problem is, but to confirm it, it would be useful to see the code that you are using to set up the Grid_dev classes on the device.

When a class or other data structure is to be used on the device, and that class has pointers in it which refer to other objects or buffers in memory (presumably in device memory, for a class that will be used on the device), then the process of making this top-level class usable on the device becomes more complicated.

Suppose I have a class like this:

class myclass{
  int myval;
  int *myptr;
  }

I could instantiate the above class on the host, and then malloc an array of int and assign that pointer to myptr, and everything would be fine. To make this class usable on the device and the device only, the process could be similar. I could:

1. cudaMalloc a pointer to device memory that will hold myclass
2. (optionally) copy an instantiated object of myclass on the host to the device pointer from step 1 using cudaMemcpy
3. on the device, use malloc or new to allocate device storage for myptr

The above sequence is fine if I never want to access the storage allocated for myptr on the host. But if I do want that storage to be visible from the host, I need a different sequence:

1. cudaMalloc a pointer to device memory that will hold myclass, let's call this mydevobj
2. (optionally) copy an instantiated object of myclass on the host to the device pointer mydevobj from step 1 using cudaMemcpy
3. Create a separate int pointer on the host, let's call it myhostptr
4. cudaMalloc int storage on the device for myhostptr
5. cudaMemcpy the pointer value of myhostptr from the host to the device pointer &(mydevobj->myptr)

After that, you can cudaMemcpy the data pointed to by the embedded pointer myptr to the region allocated (via cudaMalloc) on myhostptr

Note that in step 5, because I am taking the address of this pointer location, this cudaMemcpy operation only requires the mydevobj pointer on the host, which is valid in a cudaMemcpy operation (only).

The value of the device pointer myint will then be properly set up to do the operations you are trying to do. If you then want to cudaMemcpy data to and from myint to the host, you use the pointer myhostptr in any cudaMemcpy calls, not mydevobj->myptr. If we tried to use mydevobj->myptr, it would require dereferencing mydevobj and then using it to retrieve the pointer that is stored in myptr, and then using that pointer as the copy to/from location. This is not acceptable in host code. If you try to do it, you will get a seg fault. (Note that by way of analogy, my mydevobj is like your Grid_dev and my myptr is like your cdata)

Overall it is a concept that requires some careful thought the first time you run into it, and so questions like this come up with some frequency on SO. You may want to study some of these questions to see code examples (since you haven't provided your code that sets up Grid_dev):

1. example 1
2. example 2
3. example 3