问题
I want to hook a function by modify elf file's .text binary, my mean is replace an instruction like 'bl xxxx' with 'bl yyyy', the 'yyyy' is point to an padding area in elf file. After jump, I save the registers and call dlopen&dlsym to get new function's addr of another lib, call it, then restore the registers and jump back to 'xxxx'.
It is not very hard, and I have almost successful except for a problem: I cannot use 64-bits var in my hook function. The int type is no problem, but when I printf int64_t var, it's always shows a wrong number.
1 This is the src code:
test_ori
// test_ori.c
#include <stdio.h>
#include <dlfcn.h>
// I will hook sub and jump to myfn
void sub() {
printf("sub called...\n");
}
// The purpose of sub2 is just for let me know the addr of dlopen&dlsym
void (*func)();
void sub2() {
void *p = dlopen("/system/lib/libyyy.so", RTLD_NOW);
func = (void (*)())dlsym(p,"myfn2");
func();
}
int main(){
sub();
sub2();
return 0;
}
libyyy.so
// yyy.c
#include <stdio.h>
void myfn() {
int x = 1;
uint32_t y = 2;
uint64_t z = 3;
printf("x=%d, y=%u, z=%llu\n", x, y, z);
}
void myfn2() {}
2 Use objump to found the addrs of dlopen&dlsym
// dlopen is 0x8440, dlsym is 0x844c
84a8: f7ff efca blx 8440 <dlopen@plt>
...
84b2: f7ff efcc blx 844c <dlsym@plt>
// sub is 0x84d4
84e0: 003c movs r4, r7
84e2: 0000 movs r0, r0
84e4: b510 push {r4, lr}
84e6: f7ff fff5 bl 84d4 <puts@plt+0x7c>
84ea: f7ff ffd9 bl 84a0 <puts@plt+0x48>
3 Find padding area and modify elf file
// I use the offset 0x550(it's padding area) as my jump destination, the addr is 0x8550
// by the way, I also modify the segment's size field(0x580->0x600) so my new code can be loaded
ori -> 84e6: f7ff fff5 bl 84d4
new -> 84e6: f000 f833 bl 8550
4 The hook process from 0x8550, by asm:
1. push {r0-r7} // save registers
2. push {lr} // save lr
3. mov r1, #0 // param2 of dlopen(RTLD_NOW)
4. mov r0, pc
5. add r0, #xx // param1 of dlopen(addr of "libyyy.so")
6. blx xxxx // call dlopen
7. mov r1, pc
8. add r1, #xx // param2 of dlsym(addr of "myfn")
9. blx xxxx // call dlsym
10. blx r0 // call myfn
11. pop {r3} //
12. mov lr, r3 // restore lr
13. pop {r0-r7} // restore registers
14. b xxxx // jump back
5 Modify elf file: write code to the padding area
// I convert the asm above to machine code and write it(and strings "libyyy.so" & "myfn") to file
// then I check it in gdb:
(gdb) x/20i 0x8550
0x8550: push {r0, r1, r2, r3, r4, r5, r6, r7}
0x8552: push {lr}
0x8554: movs r1, #0
0x8556: mov r0, pc
0x8558: adds r0, #24
0x855a: blx 0x8440
0x855e: mov r1, pc
0x8560: adds r1, #26
0x8562: blx 0x844c
0x8566: blx r0
0x8568: pop {r3}
0x856a: mov lr, r3
0x856c: pop {r0, r1, r2, r3, r4, r5, r6, r7}
0x856e: b.w 0x84d4
6 The result
# ./test_new
x=1, y=2, z=12884901888
sub called...
As you see, the x and y is normal, but z(uint64_t) is wrong. It shoule be 3, but I get 12884901888(0x300000000) here.
It seems the high/low register for z is incorrect, but can you tell me why and how to fix it?
Thanks for your attention!
回答1:
The answer to this question is in the ARM Procedure call standard.
printf("x=%d, y=%u, z=%llu\n", x, y, z);
This has four arguments. The format string, the 32bit x
and y
and the 3rdz
value. As the first three arguments land z
at an odd address, the compiler pads a space so that ldrd
and strd
instruction can work; these 64-bit loads do not work if the stack is not aligned.
It is not clear whether you are showing assembler for the C code you compiled or if you are trying to modify the generated code. Most likely you do not ensure that the stack is eight byte aligned as var args must be put on the stack.
Ps: There is a good stack-overflow question about ARM 8-byte alignment, but I can not find it now. Feel free to edit my question or leave a comment.
来源:https://stackoverflow.com/questions/18000149/a-puzzle-about-hook-arm-function-by-modify-elf-file