Framework - Extension/ directory¶
This directory contains sources (C) to implement Python bindings to
internal subroutines of crash and GDB.
In addition, its Makefile is used to build the binary distribution of
PyKdump - a single self-sufficient file mpykdump.so that can
loaded as an extension in crash environment.
Directory Structure¶
configure - configuration script to generate extra Makefiles
epython.c - initilization and invoking programs
functions.c - generic functions and bindings to crash internals
gdbspec.c - bindings to GDB internals
Makefile - main Makefile
makestdlib.py - create and package a subset of Python Standard Library
minpylib-3.8.lst - list of files from Python Standard Library to include
pyconf.py - getting info about Python used for builds
pykdump.h - main header
pyparsing.py - 3rd party Python module to implement parsers
Setup.local-3.8 - used when you build Python from sources
testmod/ - a test DLKM, used for framework testing
writeREADME.py - generates README with the contents of mpykdump.so
Dependency on Python and Crash Versions¶
Usually there is no need to change anything when a new minor version
of crash or Python is released. But when there is a new
major release of either crash or Python, it is quite
possible that C-module will need some changes.
Dependency on crash Version¶
There are four types of changes in crash per se that might need changes
in C-module (in addition, we depend on GDB emebedded in crash)
Generating Makefiles during Configuration¶
Before running make command to build the extension, you need to
run configure script to generates Makefiles.
In particular, we need to get crash version (to be used in reports
and checks), GDB version and target. crash version used to be
specified directly in its Makefile until recently:
VERSION=7.2.8
but after crash project migration to Github it is not there anymore:
VERSION=
As a result, pyconf.py script that was used to extract
version/target from this Makefile needed to be modified, now we rely
on build_data.c:
char *build_command = "crash";
char *build_data = "Tue Sep 1 08:12:55 EDT 2020 by uid=1000(alexs) on zbook";
char *build_target = "X86_64";
char *build_version = "7.2.8";
char *compiler_version = "gcc (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0";
But it is possible that in the future it changes again.
Internal subroutines/variables/macros¶
C-module depends on several internal crash subroutines, variables
and macros, providing wrappers (usually called Python bindings) for
them so that these subroutines can be accessed from Python. Examples of such
subroutines/macros:
symbol_exists(symbol)
MEMBER_SIZE(name, member)
Some constants:
KVADDR/UVADDR/PHYSADDR
PAGESIZE/PAGE_CACHE_SHIFT
When a new version of crash is released, it is possible that some
of these subroutines change or become unavailable (happened at least once)
An example: recent versions of crash-7 have a set_error() subroutine that can be used for to redirect/suppress internal crash/GDB errors messages.
This subroutine is used in PyKdump C-module. As a result, if you try to load mpykdump.so with crash as shipped e.g. on RHEL7, you will see:
extend: /usr/local/lib/mpykdump64.so: undefined symbol: set_error
To workaround this we can use weak symbols (a feature of GCC), like that:
// Weak symbols as needed for compatibility with older versions of crash
extern FILE * set_error(char *target) __attribute__ ((weak));
<snip>
// Python bindings to crash internal subroutine set_error()
// You provide a single argument - a string with target name, and
// subroutine returns old target name
static PyObject *
py_crash_set_error(PyObject *self, PyObject *pyargs) {
char *target;
PyObject *rc;
// If set_error() is unavailable, do nothing and return None
if (!set_error) {
Py_INCREF(Py_None);
return Py_None;
}
Logic is as follows: if ‘set_error’ cannot be resolved when loading .so, it will be NULL.
After that, we check whether it is NULL and if yes, return None without doing anything (in more complex cases, we’ll might use our own subroutine instead, or print warning).
Signal Handlers and Executing crash Commands¶
crash has its own signal handlers for several signals,
e.g. SIGINT. When we execute Python code, Python has its own signal
handlers. To make everything work properly, we need to store/restore
signal handlers when executing Python.
PyKdump provides several commands to execute crash built-in and
return result as a string. The logic is rather complicated:
we need to feed a string as a command-line for
crashto execute, modify file descriptors to get output and after command completion do some cleanupwhile executing
crashbuiltin, we need to use its own signal handler and after that install Python signal handler again
Dependency on GDB¶
crash is built on top of GDB and to access symbolic
information (such as struct/union definitions) we need to execute
internal GDB subroutines. New major releases of crash are
usually rebased on newer major GDB version.
As a result, some enumeration definitions (used by GDB) can change
(happened twice), subroutine signature can change, and GDB
cleanup/error processing can change.
So in case PyKdump built on top of a new major crash does not work
properly, be ready to look not only on crash sources but GDB
sources as well (provided with patches in crash tarfile).
Dependency on Python Version¶
When there is a new major release of Python, three things might need change:
The contents of Setup.local used to build Python from sources
The list of Python Standard Library subroutines to be included
the way to initalize the Python environment and execute Python code
Setup.local and minpylib-3.N.lst¶
These two files are often updated together. If you decide to include another module from Python Standard Library, this often (but not always) needs linking statically another C-module included in Python sources distribution. This has nothing to do with Python major version change.
But it is not unusual that a new major release of Python rearranges library, so that you will need to change the contents of minpylib-3.N.lst to make things work. This happened e.g. while migrating from Python-3.6 to Python-3.7.
Python Initalization and Code Execution¶
Python environment is initialized only once, while loading the extension. Initialization subroutines - part of Python C-API - are regularly improved and older ones are somtimes obsoleted. This means that we might need to modify the login of _init_python() subroutine (defined in epython.c).
There are two sources of PyKdump Python code to execute:
from real files (either user-developed programs or local GIT-repo)
from ZIP-file - part of binary mpykdump.so module
Once again, new major releases of Python regularly improve the existing C-API subroutines and some old ones are being obsoleted.