hack.cpp

// Compile with:
// g++ -c -fpic -o hack.o hack.cpp && g++ -shared -o libhack.so hack.o
// Then load with the Python
// import ctypes; ctypes.CDLL('libhack.so')

#include <stdio.h>
#include <dlfcn.h>
#include <stddef.h>
#include <stdint.h>
#include <sys/mman.h>
#include <sys/user.h>


static int cudaGetDeviceFake(int* device) {
    *device = 0;
    return 0;
}

static int cudaSetDeviceFake(int device) {
    return 0;
}

// Taken from http://jmkeyes.github.io/post/hot-patching-unconditional-jumps-x86-x86-64/
static void write_redirection_medium(unsigned char *source, unsigned char *destination) {
    ptrdiff_t delta = (destination - (source + 5));
    if (delta < -2147483648 || delta > 2147483647) {
        perror("unable to redirect function: jmp to large");
        return;
    }

    /* 0xe9 is the unconditional jump opcode. */
    source[0] = 0xe9; 

    /* Write each of the octets of the delta. */
    source[1] = (delta >>  0) & 0xff;
    source[2] = (delta >>  8) & 0xff;
    source[3] = (delta >> 16) & 0xff;
    source[4] = (delta >> 24) & 0xff;

    return;
}

static void alter_page_attributes(const unsigned char *address, unsigned long length, int mask) {
    /* mprotect() needs an address that is page-aligned. */
    unsigned char *addr = (unsigned char *) ((unsigned long) address & PAGE_MASK);
    length += (address - addr);

   if (mprotect(addr, length, mask) == -1) {
       perror("mprotect");
    }

    return;
}

static void redirect(const char* sym, void* dst) {
    unsigned char* src = (unsigned char*)dlsym(NULL, sym);
    if (!src) {
       perror("src symbol not found");
       return;
    }
    alter_page_attributes(src, 5, PROT_READ|PROT_WRITE);
    write_redirection_medium(src, (unsigned char*)dst);
    alter_page_attributes(src, 5, PROT_READ|PROT_EXEC);
    printf("patched %s\n", sym);
}

__attribute__((constructor)) static void init() {
    redirect("cudaGetDevice", (void*)&cudaGetDeviceFake);
    redirect("cudaSetDevice", (void*)&cudaSetDeviceFake);
}

awesome hack, thanks.