Debugging Crashes in Proprietary Binaries – Case Study for an ARM Library

So there I was. I did have to use a proprietary library, for which I had no sources and no real hope of support from the creators. I built my program against it, I ran it, and I got a segmentation fault. An exception that seemed to happen inside that insidious library, which was of course stripped of all debugging information. I scratched my head, changed my code, checked traces, tried valgrind, strace, and other debugging tools, but found no obvious error. Finally, I assumed that I had to dig deeper and do some serious debugging of the library’s assembly code with gdb. The rest of the post is dedicated to the steps I followed to find out what was happening inside the wily proprietary library that we will call libProprietary. Prerequisites for this article are some knowledge of gdb and ARM architecture.

Some background on the task I was doing: I am a Canonical employee that works as developer for Ubuntu for Phones. In most, if not all, phones, the BSP code is not 100% open and we have to use proprietary libraries built for Android. Therefore, these libraries use bionic, Android’s libc implementation. As we want to call them inside binaries compiled with glibc, we resort to libhybris, an ingenious library that is able to load and call libraries compiled against bionic while the rest of the process uses glibc. This will turn out to be critical in this debugging. Note also that we are debugging ARM 32-bits binaries here.

The Debugging Session

To start, I made sure I had installed glibc and other libraries symbols and started to debug by using gdb in the usual way:

$ gdb myprogram
GNU gdb (Ubuntu 7.9-1ubuntu1) 7.9
Starting program: myprogram
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/arm-linux-gnueabihf/".
[New Thread 0xf49de460 (LWP 7101)]
[New Thread 0xf31de460 (LWP 7104)]
[New Thread 0xf39de460 (LWP 7103)]
[New Thread 0xf41de460 (LWP 7102)]
[New Thread 0xf51de460 (LWP 7100)]

Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0xf49de460 (LWP 7101)]
0x00000000 in ?? ()
(gdb) bt
#0  0x00000000 in ?? ()
#1  0xf520bd06 in ?? ()
Backtrace stopped: previous frame identical to this frame (corrupt stack?)
(gdb) info proc mappings
process 7097
Mapped address spaces:

	Start Addr   End Addr       Size     Offset objfile
	   0x10000    0x17000     0x7000        0x0 /usr/bin/myprogram
	0xf41e0000 0xf49df000   0x7ff000        0x0 [stack:7101]
	0xf51f6000 0xf5221000    0x2b000        0x0 /android/system/lib/
	0xf5221000 0xf5222000     0x1000        0x0 
	0xf5222000 0xf5224000     0x2000    0x2b000 /android/system/lib/
	0xf5224000 0xf5225000     0x1000    0x2d000 /android/system/lib/

We can see here that we get the promised crash. I execute a couple of gdb commands after that to see the backtrace and part of the process address space that will be of interest in the following discussion. The backtrace shows that a segment violation happened when the CPU tried to execute instructions in address zero, and we can see by checking the process mappings that the previous frame lives inside the text segment of There is no backtrace beyond that point, but that should come as no surprise as there is no DWARF information in libProprietary, and also noting that usage of frame pointer is optimized away quite commonly these days.

After this I tried to get a bit more information on the CPU state when the crash happened:

(gdb) info reg
r0             0x0	0
r1             0x0	0
r2             0x0	0
r3             0x9	9
r4             0x0	0
r5             0x0	0
r6             0x0	0
r7             0x0	0
r8             0x0	0
r9             0x0	0
r10            0x0	0
r11            0x0	0
r12            0xffffffff	4294967295
sp             0xf49dde70	0xf49dde70
lr             0xf520bd07	-182403833
pc             0x0	0x0
cpsr           0x60000010	1610612752
(gdb) disassemble 0xf520bd02,+10
Dump of assembler code from 0xf520bd02 to 0xf520bd0c:
   0xf520bd02:	b	0xf49c9cd6
   0xf520bd06:	movwpl	pc, #18628	; 0x48c4	<UNPREDICTABLE>
   0xf520bd0a:	andlt	r4, r11, r8, lsr #12
End of assembler dump.

Hmm, we are starting to see weird things here. First, in 0xf520bd02 (which probably has been executed little before the crash) we get an unconditional branch to some point in the thread stack (see mappings in previous figure). Second, the instruction in 0xf520bd06 (which should be executed after returning from the procedure that provokes the crash) would load into the pc (program counter) an address that is not mapped: we saw that the first mapped address is 0x10000 in the previous figure. The movw instruction has also a “pl” suffix that makes the instruction execute only when the operand is positive or zero… which is obviously unnecessary as 0x48c4 is encoded in the instruction.

I resorted to doing objdump -d to disassemble the library and compare with gdb output. objdump shows, in that part of the file (subtracting the library load address gives us the offset inside the file: 0xf520bd02-0xf51f6000=0x15d02):

   15d02:	f7f3 eade 	blx	92c0 <__android_log_print@plt>;
   15d06:	f8c4 5304 	str.w	r5, [r4, #772]	; 0x304
   15d0a:	4628      	mov	r0, r5
   15d0c:	b00b      	add	sp, #44	; 0x2c
   15d0e:	e8bd 8ff0 	ldmia.w	sp!, {r4, r5, r6, r7, r8, r9, sl, fp, pc}

which is completely different from what gdb shows! What is happening here? Taking a look at addresses for both code chunks, we see that instructions are always 4 bytes in gdb output, while they are 2 or 4 in objdump‘s. Well, you have guessed, don’t you? We are seeing “normal” ARM instructions in gdb, while objdump is decoding THUMB-2 instructions. Certainly objdump seems to be right here as the output is more sensible: we have a call to an executable part of the process space in 0x15d02 (it is resolved to a known function, __android_log_print), and the following instructions seems like a normal function epilogue in ARM: a return value is stored in r0, the sp (stack pointer) is incremented (we are freeing space in the stack), and we restore registers.

If we get back to the register values, we see that cpsr (current program status register [1]) does not have the T bit set, so gdb thinks we are using ARM instructions. We can change this by doing

(gdb) set $cpsr=0x60000030
(gdb) disass 0xf520bd02,+15
Dump of assembler code from 0xf520bd02 to 0xf520bd11:
   0xf520bd02:	blx	0xf51ff2c0
   0xf520bd06:	str.w	r5, [r4, #772]	; 0x304
   0xf520bd0a:	mov	r0, r5
   0xf520bd0c:	add	sp, #44	; 0x2c
   0xf520bd0e:	ldmia.w	sp!, {r4, r5, r6, r7, r8, r9, r10, r11, pc}
End of assembler dump.

Ok, much better now [2]. The thumb bit in cpsr is determined by last bx/blx call: if the address is odd, the procedure to which we are calling contains THUMB instructions, otherwise they are ARM (a good reference for these instructions is [3]). In this case, after an exception the CPU moves to arm mode, and gdb is unable to know which is the right mode when disassembling. We can search for hints on which parts of the code are arm/thumb by looking at the values in registers used by bx/blx, or by looking at the lr (link register): we can see above that the value after the crash was 0xf520bd07, which is odd and indicates that 0xf520bd06 contains a thumb instruction. However, for some reason gdb is not able to take advantage of this information.

Of course this problem does not happen if we have debugging information: in that case we have special symbols that let gdb know if the section where the code is contains thumb instructions or not [4]. As those are not found, gdb uses the cpsr value. Here objdump seems to have better heuristics though.

After solving this issue with instruction decoding, I started to debug __android_log_print to check what was happening there, as it looked like the crash was happening in that call. I spent quite a lot of time there, but found nothing. All looked fine, and I started to despair. Until I inserted a breakpoint in address 0xf520bd06, right after the call to __android_log_print, run the program… and it stopped at that address, no crash happened. I started to execute the program instruction by instruction after that:

(gdb) b *0xf520bd06
(gdb) run
Breakpoint 1, 0xf520bd06 in ?? ()
(gdb) si
0xf520bd0a in ?? ()
(gdb) si
0xf520bd0c in ?? ()
(gdb) si
0xf520bd0e in ?? ()
Cannot insert breakpoint 0.
Cannot access memory at address 0x0

Something was apparently wrong with instruction ldmia, which restores registers, including the pc, from the stack. I took a look at the stack in that moment (taking into account that ldmia had already modified the sp after restoring 9 registers == 36 bytes):

(gdb) x/16xw $sp-36
0xf49dde4c:	0x00000000	0x00000000	0x00000000	0x00000000
0xf49dde5c:	0x00000000	0x00000000	0x00000000	0x00000000
0xf49dde6c:	0x00000000	0x00000000	0x00000000	0x00000000
0xf49dde7c:	0x00000000	0x00000000	0x00000000	0x00000000

All zeros! At this point it is clear that this is the real point where the crash is happening, as we are loading 0 into the pc. This looked clearly like a stack corruption issue.

But, before moving forward, why are we getting a wrong backtrace from gdb? Well, gdb is seeing a corrupted stack, so it is not able to unwind it. It would not be able to unwind it even if having full debug information. The only hint it has is the lr. This register contains the return address after execution of a bl/blx instruction [3]. If the called procedure is non-leaf, it is saved in the prologue, and restored in the epilogue, because it gets overwritten when branching to other procedures. In this case, it is restored on the pc and sometimes it is also saved back in the lr, depending on whether we have arm-thumb interworking built in the procedure or not [5]. It is not overwritten if we have a leaf procedure (as there are no procedure calls inside these).

As gdb has no additional information, it uses the lr to build the backtrace, assuming we are in a leaf procedure. However this is not true and the backtrace turns out to be wrong. Nonetheless, this information was not completely useless: lr was pointing to the instruction right after the last bl/blx instruction that was executed, which was not that far away from the real point where the program was crashing. This happened because fortunately __android_log_print has interworking code and restores the lr, otherwise the value of lr could have been from a point much far away from the point where the real crash happens. Believe or not, but it could have been even worse!

Having now a clear idea of where and why the crash was happening, things accelerated. The procedure where the crash happened, as disassembled by objdump, was (I include here only the more relevant parts of the code)

00015b1c <ProprietaryProcedure@@Base>:
   15b1c:	e92d 4ff0 	stmdb	sp!, {r4, r5, r6, r7, r8, r9, sl, fp, lr}
   15b20:	b08b      	sub	sp, #44	; 0x2c
   15b22:	497c      	ldr	r1, [pc, #496]	; (15d14 <ProprietaryProcedure@@Base+0x1f8>)
   15b24:	2500      	movs	r5, #0
   15b26:	9500      	str	r5, [sp, #0]
   15b28:	4604      	mov	r4, r0
   15b2a:	4479      	add	r1, pc
   15b2c:	462b      	mov	r3, r5
   15b2e:	f8df 81e8 	ldr.w	r8, [pc, #488]	; 15d18 <ProprietaryProcedure@@Base+0x1fc>
   15b32:	462a      	mov	r2, r5
   15b34:	f8df 91e4 	ldr.w	r9, [pc, #484]	; 15d1c <ProprietaryProcedure@@Base+0x200>
   15b38:	ae06      	add	r6, sp, #24
   15b3a:	f8df a1e4 	ldr.w	sl, [pc, #484]	; 15d20 <ProprietaryProcedure@@Base+0x204>
   15b3e:	200f      	movs	r0, #15
   15b40:	f8df b1e0 	ldr.w	fp, [pc, #480]	; 15d24 <ProprietaryProcedure@@Base+0x208>
   15b44:	f7f3 ef76 	blx	9a34 <prctl@plt>
   15b48:	44f8      	add	r8, pc
   15b4a:	4629      	mov	r1, r5
   15b4c:	44f9      	add	r9, pc
   15b4e:	2210      	movs	r2, #16
   15b50:	44fa      	add	sl, pc
   15b52:	4630      	mov	r0, r6
   15b54:	44fb      	add	fp, pc
   15b56:	f7f3 ea40 	blx	8fd8 <memset@plt>
   15b5a:	a807      	add	r0, sp, #28
   15b5c:	f7f3 ef70 	blx	9a40 <sigemptyset@plt>
   15b60:	4b71      	ldr	r3, [pc, #452]	; (15d28 <ProprietaryProcedure@@Base+0x20c>)
   15b62:	462a      	mov	r2, r5
   15b64:	9508      	str	r5, [sp, #32]
   15b66:	4631      	mov	r1, r6
   15b68:	447b      	add	r3, pc
   15b6a:	681b      	ldr	r3, [r3, #0]
   15b6c:	200a      	movs	r0, #10
   15b6e:	9306      	str	r3, [sp, #24]
   15b70:	f7f3 ef6c 	blx	9a4c <sigaction@plt>
   15d02:	f7f3 eade 	blx	92c0 <__android_log_print@plt>
   15d06:	f8c4 5304 	str.w	r5, [r4, #772]	; 0x304
   15d0a:	4628      	mov	r0, r5
   15d0c:	b00b      	add	sp, #44	; 0x2c
   15d0e:	e8bd 8ff0 	ldmia.w	sp!, {r4, r5, r6, r7, r8, r9, sl, fp, pc}

The addresses where this code is loaded can be easily computed by adding 0xf51f6000 to the file offsets shown in the first column. We see that a few calls to different external functions [6] are performed by ProprietaryProcedure, which is itself an exported symbol.

I restarted the debug session, added a breakpoint at the start of ProprietaryProcedure, right after stmdb saves the state, and checked the stack values:

(gdb) b *0xf520bb20
Breakpoint 1 at 0xf520bb20
(gdb) cont
Breakpoint 1, 0xf520bb20 in ?? ()
(gdb) p $sp
$1 = (void *) 0xf49dde4c
(gdb) x/16xw $sp
0xf49dde4c:	0xf49de460	0x0007df00	0x00000000	0xf49dde70
0xf49dde5c:	0xf49de694	0x00000000	0xf77e9000	0x00000000
0xf49dde6c:	0xf75b4491	0x00000000	0xf49de460	0x00000000
0xf49dde7c:	0x00000000	0xfd5b4eba	0xfe9dd4a3	0xf49de460

We can see that the stack contains something, including a return address that looks valid (0xf75b4491). Note also that the procedure must never touch this part of the stack, as it belongs to the caller of ProprietaryProcedure.

Now it is a simply a matter of bisecting the code between the beginning and the end of ProprietaryProcedure to find out where we are clobbering the stack. I will save you of developing here this tedious process. Instead, I will just show, that, in the end, it turned out that the call to sigemptyset() is the culprit [7]:

(gdb) b *0xf520bb5c
Breakpoint 1 at 0xf520bb5c
(gdb) b *0xf520bb60
Breakpoint 2 at 0xf520bb60
(gdb) run
Breakpoint 1, 0xf520bb5c in ?? ()
(gdb) x/16xw 0xf49dde4c
0xf49dde4c:	0xf49de460	0x0007df00	0x00000000	0xf49dde70
0xf49dde5c:	0xf49de694	0x00000000	0xf77e9000	0x00000000
0xf49dde6c:	0xf75b4491	0x00000000	0xf49de460	0x00000000
0xf49dde7c:	0x00000000	0xfd5b4eba	0xfe9dd4a3	0xf49de460
(gdb) cont
Breakpoint 2, 0xf520bb60 in ?? ()
(gdb) x/16xw 0xf49dde4c
0xf49dde4c:	0x00000000	0x00000000	0x00000000	0x00000000
0xf49dde5c:	0x00000000	0x00000000	0x00000000	0x00000000
0xf49dde6c:	0x00000000	0x00000000	0x00000000	0x00000000
0xf49dde7c:	0x00000000	0x00000000	0x00000000	0x00000000

Note here that I am printing the part of the stack not reserved by the function (0xf49dde4c is the value of the sp before execution of the line at offset 0x15b20, see the code).

What is going wrong here? Now, remember that at the beginning of the article I mentioned that we were using libhybris. libProprietary assumes a bionic environment, and the libc functions it calls are from bionic’s libc. However, libhybris has hooks for some bionic functions: for them bionic is not called, instead the hook is invoked. libhybris does this to avoid conflicts between bionic and glibc: for instance having two allocators fighting for process address space is a recipe for disaster, so malloc() and related functions are hooked and the hooks call in the end the glibc implementation. Signals related functions were hooked too, including sigemptyset(), and in this case the hook simply called glibc implementation.

I looked at glibc and bionic implementations, in both cases sigemptyset() is a very simple utility function that clears with memset() a sigset_t variable. All pointed to different definitions of sigset_t depending on the library. Definition turned out to be a bit messy when looking at the code as it depended on build time definitions, so I resorted to gdb to print the type. For a executable compiled for glibc, I saw

(gdb) ptype sigset_t
type = struct {
    unsigned long __val[32];

and for one using bionic

(gdb) ptype sigset_t
type = unsigned long

This finally confirms where the bug is, and explains it: we are overwriting the stack because libProprietary reserves in the stack memory for bionic’s sigset_t, while we are using glibc’s sigemptyset(), which uses a different definition for it. As this definition is much bigger, the stack gets overwritten after the call to memset(). And we get the crash later when trying to restore registers when the function returns.

After knowing this, the solution was simple: I removed the libhybris hooks for signal functions, recompiled it, and… all worked just fine, no crashes anymore!

However, this is not the final solution: as signals are shared resources, it makes sense to hook them in libhybris. But, to do it properly, the hooks have to translate types between bionic in glibc, thing that we were not doing (we were simply calling glibc implementation). That, however, is “just work”.

Of course I wondered why the heck a library that is kind of generic needs to mess around with signals, but hey, that is not my fault ;-).


I can say I learned several things while debugging this:

  1. Not having the sources is terrible for debugging (well, I already knew this). Unfortunately not open sourcing the code is still a standard practice in part of the industry.
  2. The most interesting technical bit here is IMHO that we need to be very cautious with the backtrace that debuggers shows after a crash. If you start to see things that do not make sense it is possible that registers or stack have been messed up and the real crash happens elsewhere. Bear in mind that the very first thing to do when a program crashes is to make sure that we know the exact point where that happens.
  3. We have to be careful in ARM when disassembling, because if there is no debug information we could be seeing the wrong instruction set. We can check evenness of addresses used by bx/blx and of the lr to make sure we are in the right mode.
  4. Some times taking a look at assembly code can help us when debugging, even when we have the sources. Note that if I had had the C sources I would have seen the crash happening right when returning from a function, and it might not have been that immediate to find out that the stack was messed up. The assembly clearly pointed to an overwritten stack.
  5. Finally, I personally learned some bits of ARM architecture that I did not know, which was great.

Well, this is it. I hope you enjoyed the (lengthy, I know) article. Thanks for your reading!

[2] We can get the same result by executing in gdb set arm fallback-mode thumb, but changing the register seemed more pedagogical here.
[6] In fact the calls are to the PLT section, which is inside the library. The PLT calls in turn, by using addresses in the GOT data section, either directly the function or the dynamic loader, as we are doing lazy loading. See, for instance.
[7] I had to use two breakpoints between consecutive instructions because the “ni” gdb command was not working well here.