Re: [lkcd-devel] issues during gdb-readable core dump conversion implementation

[Keith O. <ka...@sg...>]

On Wed, 31 Mar 2004 14:31:59 +0200, 
"Jan Beulich" <JBe...@no...> wrote:
>Following the code path from IPI_handler, I can't see where this
>creation of a switch_stack structure should happen.

Standard lkcd code does not create a switch_stack for active processes.
kdb does, and driving lkcd out of kdb gives lkcd the data that kdb has
already captured, for free.

>Similarly, I can't see how this would save the
>high floating point registers, since the only two places I can find
>where they are saved are in switch_to and the ptrace code.

IA64 kdb drives all cpus through kdba_main_loop() -> unw_init_running()
-> do_kdba_main_loop.  unw_init_running (a lovely chunk of ia64
assembler) creates a struct switch_stack that contains all the process
state that is not already in pt_regs.  kdb continues running on each
cpu after switch_stack is pushed, but the switch_stack makes it look
like the task is sleeping.  gdb needs the data in switch_stack to do a
proper unwind.

If you trace through unw_init_running, you find that much of the work
is done in save_switch_stack.  That saves f12-f31, amongst others.