Re: [libunwind] Unwinding using the context passed to a signal handler (doesn't work)

[Johan Walles]

Johan Walles wrote:
> I'm trying to create a stack dump from inside a signal handler.  The 
> context I'm using is the one passed as the third parameter to the signal 
> handler.  I'm doing this on Linux / ia64 using libunwind 0.96.
>
> This is the output from the program:
> "
> printStackTrace(): Got context at 0x60000fffffffae50
>   0: 0x4000000000001f30: main()
>   1: 0x200000000008c560: __libc_start_main()
>   2: 0x4000000000001500: _start()
> printStackTrace(): Got context at 0x60000fffffffa310
>   0: 0x4000000000001ce0: crashhandler()
>   1: 0xa0000000000040d0: Unknown
>   2: 0x4000000000001ec0: crash()
>   3: 0x4000000000001f30: main()
>   4: 0x200000000008c560: __libc_start_main()
>   5: 0x4000000000001500: _start()
> "
>
> Note that the function that crashes (crash()) isn't part of the first 
> stack trace (initialized using the context passed to the signal handler).
>
> Why isn't it?  And if I want that, what should I do?  Am I using 
> libunwind correctly?

To answer my own question, what happens here is the following (this has nothing 
to do with what registers are filled in or not):

1. Calling unw_getcontext() is the same as just calling libc's getcontext().

2. getcontext() on Linux / ia64 returns the context *inside the getcontext() 
function*, not the context at the place from where getcontext() is called (which 
is what I always thought it did).

3. To compensate for this, unw_init_local() starts by unwinding the context it 
receives by one frame.  Thus, unwinding the stack after doing 
unw_init_local(getcontext()) will get me a backtrace starting from the current 
frame, rather than from the inside of getcontext().

In the signal handler case, the signal handler receives as parameter the context 
as it looked at the time the signal was received.  If you pass that to 
unw_init_local(), it'll start by skipping the frame where the signal was 
received.  This (rather than all registers not being filled in) is why the 
attached program didn't work.

So some things that I think should be changed are:

1. The unw_getcontext() man page says:

"The  unw_getcontext()  routine  initializes  the  context  structure  pointed 
to  by ucp with the machine-state of the call-site."

This is how I think it should be working, but isn't in fact how it works in 
reality.  IMO, the man page is correct, and the implementation should be fixed.

2. The unw_init_local() man page says:

"The unw_init_local() routine initializes the unwind cursor pointed to by c with 
the machine-state in the context structure pointed to by ctxt.  As such, the 
machine-state pointed to by ctxt identifies the initial  stack  frame  at  which 
unwinding starts."

Again, this is how I think it should be working, but isn't in fact how it works 
in reality.  IMO, this man page is correct as well, and the implementation 
should be fixed.

  Regards //Johan