PyFile_WriteString() calls PyObject_Str() which calls PyErr_CheckSignals(). If a signal was catched, the signal handler is called. If the signal handler raises an error, PyObject_Str() and then PyFile_WriteString() return NULL. mywrite() ignores all PyFile_WriteString() errors. It should maybe only ignores errors from the file (except IOError: ...) and not any error. Another problem: mywrite() is called from PySys_WriteStdout() and PySys_WriteStderr() which are procedures. PySys_WriteStdout()/PySys_WriteStderr() caller cannot detect the error. There are 65 calls to PySys_WriteStd...
Attached patch fixes this issue: PyFile_WriteObject() doesn't call PyObject_Str() to avoid PyErr_CheckSignals(). I'm not sure that it's the right approch because it may change the behaviour of existing code when getting a signal.
Ok, forget my pyfile_writeobject_nosignal.patch, it's not the right approach. New patch: mywrite() uses its own implementation PyFile_WriteString(), sys_pyfile_write(), which doesn't call PyErr_CheckSignals(): /* Implementation of PyFile_WriteString() no calling PyErr_CheckSignals(): * mywrite() should not execute any Python signal handler to avoid raising an * error because mywrite() ignores all errors */
The goal is this issue is also to catch SIGINT when starting Python. It now works in Python trunk and py3k, but not in verbose mode because mywrite() eats errors (especially the KeyboardInterrupt raised by the default SIGINT handler) and calls indirectly Python signal handlers.
