HELP : ptrace returns puzzling results
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 2 lines · Jun Sun
I am writing a gdbserver for linux/mips. The server can now talk with the gdb client, and can run to completion if you press 'c'.
Excellent.
Quoted 11 lines
However, the gdb client reads some confusing register values. I traced back to the gdbserver and found out that these values are returned from
ptrace(), which seem wrong. On the other hand, the native gdb must be using the sameptrace()to get register values. I wonder why it works there.Here is a sample output from gdbserver calling ptrace. Note that PC value does not corresponds to the executable image. sp also seems wrong.
reg #29 ($29,sp) = 2147483120 (
0x7ffffdf0)
The value of sp looks sane, it a value near the absolute top of the stack at 0x80000000.
Quoted 1 line
reg #37 ($64,pc) = 263607008 (
0x0fb652e0)
Also looks sane at first look, this value is in the typical address range where the dynamic linker gets mapped.
Many of the registers in the dump you gave have a value of zero and that is worrying me much more. All the caller saved registers are zero, that smells.
What kernel version are you using?
Ralf
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 5 lines · Jun Sun
However, step or next or setting another breakpoint do not work. It appears that gdbserver calls ptrace with
PTRACE_SINGLESTEPoption, which is not implemented in the kernel I am using (it is v2.3.99-pre3). Is this implemented in the latest version? If not, is it difficult to add one? Or can we get around without it?
PTRACE_SINGLESTEP is meant to be implemented using hardware single stepping. That's something which most MIPS CPUs don't support. It can be simulated in userspace, so there is no point in putting it into the kernel ergo we won't implement it.
There's actually a good number of CPUs which don't have hardware breakpoints so you should check what gdbserver does for those other architectures. The generic solution is to write a breakpoint into the logically next instruction(s). Usually that's at epc + 4 but there are branch instructions to make live more entertaining ...
Ralf