Give the attached file execute permission and run it. At the first breakpoint, the backtrace will be correct. Continue. At the second breakpoint, a backtrace will show the foo('first call') on the stack when, in fact, the call came from foo('second call'), as verified by the printed message. I am running this on an up-to-date 64-bit Arch Linux system. Python 3.3.1.
On python-list, Robert Kern used "where" to demonstrate the problem: $ python pdbbug.py first call --Return-- > /Users/rkern/scratch/pdbbug.py(4)foo()->None -> pdb.set_trace() (Pdb) where /Users/rkern/scratch/pdbbug.py(5)() -> foo('first call') > /Users/rkern/scratch/pdbbug.py(4)foo()->None -> pdb.set_trace() (Pdb) c second call --Return-- > /Users/rkern/scratch/pdbbug.py(4)foo()->None -> pdb.set_trace() (Pdb) where /Users/rkern/scratch/pdbbug.py(5)() -> foo('first call') > /Users/rkern/scratch/pdbbug.py(4)foo()->None -> pdb.set_trace() (Pdb) Searching other open pdb issues, applying the suggested fix from Issue16482 seems to correct the problem here.
The call to set_trace() installs a local trace function on all the frames of the stack, including on the oldest frame, i.e. the module level frame. This causes the invocation of frame_settrace() in frameobject.c and the module frame f_lineno is evaluated by PyCode_Addr2Line() (this happens only on the first call to set_trace() since the module frame local trace is never removed). Next, the 'continue' command sets the global trace function to None and removes the local trace function of all the frames of the stack, *except* the module frame (see set_continue() in bdb.py). On printing the stack with the 'where' command, the bdb get_stack() function invokes frame_getlineno() that calls PyFrame_GetLineNumber() that assumes that the module frame f_lineno is correct since it has a local trace function. But as the global trace function has been set to None by the 'continue' command, f_lineno is not updated by the interpreter tracing code when running the module level frame. Hence the problem. The root cause is that PyFrame_GetLineNumber assumes incorrectly that when a frame has a local trace function, its f_lineno is correct. Issue 7238 and issue 17277 are also related to this issue.
