Consider this simple python program - 

  1 #!/usr/bin/python
  2
  3 import pdb
  4 import sys
  5 import traceback
  6
  7 def trace_exception( type, value, tb ):
  8     traceback.print_tb( tb )
  9     pdb.post_mortem( tb )
 10
 11 sys.excepthook = trace_exception
 12
 13 def func():
 14     print ( "Am here in func" )
 15     pdb.set_trace()
 16     print ( "Am here after pdb" )
 17     print ( "Am going to assert now" )
 18     assert False
 19     print ( "Am here after exception" )
 20
 21 def main():
 22     func()
 23
 24 if __name__ == "__main__":
 25     main()


On running this program - 

% ./python /tmp/test.py
Am here in func
> /tmp/test.py(16)func()
-> print ( "Am here after pdb" )
(Pdb) c
Am here after pdb
Am going to assert now
  File "/tmp/test.py", line 25, in <module>
    main()
  File "/tmp/test.py", line 22, in main
    func()
  File "/tmp/test.py", line 16, in func
    print ( "Am here after pdb" )
> /tmp/test.py(16)func()
-> print ( "Am here after pdb" ) >>>> This should have been at the line corresponding to "Am going to assert now"
(Pdb)


This seems to be an bug ( due to a performance consideration ) with the way python bdb's set_continue() has been implemented -

https://hg.python.org/cpython/file/2.7/Lib/bdb.py#l227

    def set_continue(self):
        # Don't stop except at breakpoints or when finished
        self._set_stopinfo(self.botframe, None, -1)
        if not self.breaks:
            # no breakpoints; run without debugger overhead
            sys.settrace(None)
            frame = sys._getframe().f_back
            while frame and frame is not self.botframe:
                del frame.f_trace
                frame = frame.f_back

Basically what happens after "c" in a "(Pdb)" prompt is that bdb optimizes for the case where there are no more break points found by cleaning up the trace callback from all the frames.

However, all of this happens in the context of tracing itself and hence the trace_dispatch function in https://hg.python.org/cpython/file/2.7/Lib/bdb.py#l45 still returns back the trace_dispatch as the new system trace function. For more details on sys.settrace(), check https://docs.python.org/2/library/sys.html#sys.settrace

Check, the function trace_trampoline at https://hg.python.org/cpython/file/2.7/Python/sysmodule.c#l353
which sets f->f_trace back to result at https://hg.python.org/cpython/file/2.7/Python/sysmodule.c#l377

This seems to be an bug ( due to a performance consideration ) with the way python bdb's set_continue() has been implemented -

https://hg.python.org/cpython/file/2.7/Lib/bdb.py#l227

    def set_continue(self):
        # Don't stop except at breakpoints or when finished
        self._set_stopinfo(self.botframe, None, -1)
        if not self.breaks:
            # no breakpoints; run without debugger overhead
            sys.settrace(None)
            frame = sys._getframe().f_back
            while frame and frame is not self.botframe:
                del frame.f_trace
                frame = frame.f_back

Basically what happens after "c" in a "(Pdb)" prompt is that bdb optimizes for the case where there are no more break points found by cleaning up the trace callback from all the frames.

However, all of this happens in the context of tracing itself and hence the trace_dispatch function in https://hg.python.org/cpython/file/2.7/Lib/bdb.py#l45 still returns back the trace_dispatch as the new system trace function. For more details on sys.settrace(), check https://docs.python.org/2/library/sys.html#sys.settrace

Check, the function trace_trampoline at https://hg.python.org/cpython/file/2.7/Python/sysmodule.c#l353
which sets f->f_trace back to result at https://hg.python.org/cpython/file/2.7/Python/sysmodule.c#l377

Now, check the function PyFrame_GetLineNumber() which is used by the traceback to get the frame line number https://hg.python.org/cpython/file/2.7/Objects/frameobject.c#l63

int
PyFrame_GetLineNumber(PyFrameObject *f)
{
    if (f->f_trace)
        return f->f_lineno;
    else
        return PyCode_Addr2Line(f->f_code, f->f_lasti);
}

Basically this function returns back the stored f->f_lineno in case the f->f_trace is enabled.

The fix is fortunately simple - 

Just set self.trace_dispatch to None if pdb set_continue decides to run without debugger overhead.

ping

ping.

Gentle reminder to review the patch

See also issue 24565.

Ran into a similar issue today

This issue (and mine is fixed) in python 3.10. Most probaly related to #24565 fix

I meant: this issue is fixed in 3.10