Core - ARM Cortex-M4
Compiler - GCC 5.3.0 ARM EABI
OS - Free RTOS
I am doing stack backtrace using gcc library function _Unwind_Reason_Code _Unwind_Backt
This is doable but needs access to internal details of how libgcc implements the _Unwind_Backtrace function. Fortunately the code is open-source, but depending on such internal details is brittle in that it may break in future versions of armgcc without any notice.
Generally, reading through the source of libgcc doing the backtrace, it creates an inmemory virtual representation of the CPU core registers, then uses this representation to walk up the stack, simulating exception throws. The first thing that _Unwind_Backtrace does is fill in this context from the current CPU registers, then call an internal implementation function.
Creating that context manually from the stacked exception structure is sufficient to fake the backtrace going from handler mode upwards through the call stack in most cases. Here is some example code (from https://github.com/bakerstu/openmrn/blob/62683863e8621cef35e94c9dcfe5abcaf996d7a2/src/freertos_drivers/common/cpu_profile.hxx#L162):
/// This struct definition mimics the internal structures of libgcc in
/// arm-none-eabi binary. It's not portable and might break in the future.
struct core_regs
{
unsigned r[16];
};
/// This struct definition mimics the internal structures of libgcc in
/// arm-none-eabi binary. It's not portable and might break in the future.
typedef struct
{
unsigned demand_save_flags;
struct core_regs core;
} phase2_vrs;
/// We store what we know about the external context at interrupt entry in this
/// structure.
phase2_vrs main_context;
/// Saved value of the lr register at the exception entry.
unsigned saved_lr;
/// Takes registers from the core state and the saved exception context and
/// fills in the structure necessary for the LIBGCC unwinder.
void fill_phase2_vrs(volatile unsigned *fault_args)
{
main_context.demand_save_flags = 0;
main_context.core.r[0] = fault_args[0];
main_context.core.r[1] = fault_args[1];
main_context.core.r[2] = fault_args[2];
main_context.core.r[3] = fault_args[3];
main_context.core.r[12] = fault_args[4];
// We add +2 here because first thing libgcc does with the lr value is
// subtract two, presuming that lr points to after a branch
// instruction. However, exception entry's saved PC can point to the first
// instruction of a function and we don't want to have the backtrace end up
// showing the previous function.
main_context.core.r[14] = fault_args[6] + 2;
main_context.core.r[15] = fault_args[6];
saved_lr = fault_args[5];
main_context.core.r[13] = (unsigned)(fault_args + 8); // stack pointer
}
extern "C"
{
_Unwind_Reason_Code __gnu_Unwind_Backtrace(
_Unwind_Trace_Fn trace, void *trace_argument, phase2_vrs *entry_vrs);
}
/// Static variable for trace_func.
void *last_ip;
/// Callback from the unwind backtrace function.
_Unwind_Reason_Code trace_func(struct _Unwind_Context *context, void *arg)
{
void *ip;
ip = (void *)_Unwind_GetIP(context);
if (strace_len == 0)
{
// stacktrace[strace_len++] = ip;
// By taking the beginning of the function for the immediate interrupt
// we will attempt to coalesce more traces.
// ip = (void *)_Unwind_GetRegionStart(context);
}
else if (last_ip == ip)
{
if (strace_len == 1 && saved_lr != _Unwind_GetGR(context, 14))
{
_Unwind_SetGR(context, 14, saved_lr);
allocator.singleLenHack++;
return _URC_NO_REASON;
}
return _URC_END_OF_STACK;
}
if (strace_len >= MAX_STRACE - 1)
{
++allocator.limitReached;
return _URC_END_OF_STACK;
}
// stacktrace[strace_len++] = ip;
last_ip = ip;
ip = (void *)_Unwind_GetRegionStart(context);
stacktrace[strace_len++] = ip;
return _URC_NO_REASON;
}
/// Called from the interrupt handler to take a CPU trace for the current
/// exception.
void take_cpu_trace()
{
memset(stacktrace, 0, sizeof(stacktrace));
strace_len = 0;
last_ip = nullptr;
phase2_vrs first_context = main_context;
__gnu_Unwind_Backtrace(&trace_func, 0, &first_context);
// This is a workaround for the case when the function in which we had the
// exception trigger does not have a stack saved LR. In this case the
// backtrace will fail after the first step. We manually append the second
// step to have at least some idea of what's going on.
if (strace_len == 1)
{
main_context.core.r[14] = saved_lr;
main_context.core.r[15] = saved_lr;
__gnu_Unwind_Backtrace(&trace_func, 0, &main_context);
}
unsigned h = hash_trace(strace_len, (unsigned *)stacktrace);
struct trace *t = find_current_trace(h);
if (!t)
{
t = add_new_trace(h);
}
if (t)
{
t->total_size += 1;
}
}
/// Change this value to runtime disable and enable the CPU profile gathering
/// code.
bool enable_profiling = 0;
/// Helper function to declare the CPU usage tick interrupt.
/// @param irq_handler_name is the name of the interrupt to declare, for example
/// timer4a_interrupt_handler.
/// @param CLEAR_IRQ_FLAG is a c++ statement or statements in { ... } that will
/// be executed before returning from the interrupt to clear the timer IRQ flag.
#define DEFINE_CPU_PROFILE_INTERRUPT_HANDLER(irq_handler_name, CLEAR_IRQ_FLAG) \
extern "C" \
{ \
void __attribute__((__noinline__)) load_monitor_interrupt_handler( \
volatile unsigned *exception_args, unsigned exception_return_code) \
{ \
if (enable_profiling) \
{ \
fill_phase2_vrs(exception_args); \
take_cpu_trace(); \
} \
cpuload_tick(exception_return_code & 4 ? 0 : 255); \
CLEAR_IRQ_FLAG; \
} \
void __attribute__((__naked__)) irq_handler_name(void) \
{ \
__asm volatile("mov r0, %0 \n" \
"str r4, [r0, 4*4] \n" \
"str r5, [r0, 5*4] \n" \
"str r6, [r0, 6*4] \n" \
"str r7, [r0, 7*4] \n" \
"str r8, [r0, 8*4] \n" \
"str r9, [r0, 9*4] \n" \
"str r10, [r0, 10*4] \n" \
"str r11, [r0, 11*4] \n" \
"str r12, [r0, 12*4] \n" \
"str r13, [r0, 13*4] \n" \
"str r14, [r0, 14*4] \n" \
: \
: "r"(main_context.core.r) \
: "r0"); \
__asm volatile(" tst lr, #4 \n" \
" ite eq \n" \
" mrseq r0, msp \n" \
" mrsne r0, psp \n" \
" mov r1, lr \n" \
" ldr r2, =load_monitor_interrupt_handler \n" \
" bx r2 \n" \
: \
: \
: "r0", "r1", "r2"); \
} \
}
This code is designed to take a CPU profile using a timer interrupt, but the backtrace unwinding can be reused from any handler including fault handlers. Read the code from the bottom to the top:
__naked__
, otherwise the function entry header of GCC will manipulate the state of the CPU in unpredictable way, modifying the stack pointer for example.