Scroll to navigation

DRGN(1) drgn DRGN(1)

NAME

drgn - drgn 0.0.26

drgn (pronounced "dragon") is a debugger with an emphasis on programmability. drgn exposes the types and variables in a program for easy, expressive scripting in Python. For example, you can debug the Linux kernel:

>>> from drgn.helpers.linux import list_for_each_entry
>>> for mod in list_for_each_entry('struct module',
...                                prog['modules'].address_of_(),
...                                'list'):
...    if mod.refcnt.counter > 10:
...        print(mod.name)
...
(char [56])"snd"
(char [56])"evdev"
(char [56])"i915"


Although other debuggers like GDB have scripting support, drgn aims to make scripting as natural as possible so that debugging feels like coding. This makes it well-suited for introspecting the complex, inter-connected state in large programs.

Additionally, drgn is designed as a library that can be used to build debugging and introspection tools; see the official tools.

drgn was developed at Meta for debugging the Linux kernel (as an alternative to the crash utility), but it can also debug userspace programs written in C. C++ support is in progress.

In addition to the main Python API, an experimental C library, libdrgn, is also available.

See the Installation instructions. Then, start with the User Guide.

LICENSE

Copyright (c) Meta Platforms, Inc. and affiliates.

drgn is licensed under the LGPLv2.1 or later.

ACKNOWLEDGEMENTS

drgn is named after this because dragons eat dwarves.

TABLE OF CONTENTS

Installation

There are several options for installing drgn.

Dependencies

drgn depends on:

  • Python 3.6 or newer
  • elfutils 0.165 or newer

It optionally depends on:

•
libkdumpfile for makedumpfile compressed kernel core dump format support

The build requires:

  • GCC
  • GNU Make
  • pkgconf
  • setuptools

Building from the Git repository (rather than a release tarball) additionally requires:

  • autoconf
  • automake
  • libtool

Installation

Package Manager

drgn can be installed using the package manager on some Linux distributions.

•
Fedora >= 32

$ sudo dnf install drgn


•
RHEL/CentOS >= 8

Enable EPEL. Then:

$ sudo dnf install drgn


  • Arch Linux

    Install the drgn package from the AUR.

  • Debian >= 12 (Bookworm)

$ sudo apt install python3-drgn


•
openSUSE

$ sudo zypper install python3-drgn


•
Ubuntu

Enable the michel-slm/kernel-utils PPA. Then:

$ sudo apt install python3-drgn



pip

If your Linux distribution doesn't package the latest release of drgn, you can install it with pip.

First, install pip. Then, run:

$ sudo pip3 install drgn


This will install a binary wheel by default. If you get a build error, then pip wasn't able to use the binary wheel. Install the dependencies listed below and try again.

Note that RHEL/CentOS 6, Debian Stretch, Ubuntu Trusty, and Ubuntu Xenial (and older) ship Python versions which are too old. Python 3.6 or newer must be installed.

From Source

To get the development version of drgn, you will need to build it from source. First, install dependencies:

•
Fedora

$ sudo dnf install autoconf automake elfutils-devel gcc git libkdumpfile-devel libtool make pkgconf python3 python3-devel python3-pip python3-setuptools


•
RHEL/CentOS

$ sudo dnf install autoconf automake elfutils-devel gcc git libtool make pkgconf python3 python3-devel python3-pip python3-setuptools


Optionally, install libkdumpfile-devel from EPEL on RHEL/CentOS >= 8 or install libkdumpfile from source if you want support for the makedumpfile format.

Replace dnf with yum for RHEL/CentOS < 8.

•
Debian/Ubuntu

$ sudo apt-get install autoconf automake gcc git liblzma-dev libelf-dev libdw-dev libtool make pkgconf python3 python3-dev python3-pip python3-setuptools zlib1g-dev


Optionally, install libkdumpfile from source if you want support for the makedumpfile format.

•
Arch Linux

$ sudo pacman -S --needed autoconf automake gcc git libelf libtool make pkgconf python python-pip python-setuptools


Optionally, install libkdumpfile from the AUR or from source if you want support for the makedumpfile format.

•
openSUSE

$ sudo zypper install autoconf automake gcc git libdw-devel libelf-devel libkdumpfile-devel libtool make pkgconf python3 python3-devel python3-pip python3-setuptools



Then, run:

$ git clone https://github.com/osandov/drgn.git
$ cd drgn
$ python3 setup.py build
$ sudo python3 setup.py install


Virtual Environment

The above options all install drgn globally. You can also install drgn in a virtual environment, either with pip:

$ python3 -m venv drgnenv
$ source drgnenv/bin/activate
(drgnenv) $ pip3 install drgn
(drgnenv) $ drgn --help


Or from source:

$ python3 -m venv drgnenv
$ source drgnenv/bin/activate
(drgnenv) $ python3 setup.py install
(drgnenv) $ drgn --help


Running Locally

If you build drgn from source, you can also run it without installing it:

$ python3 setup.py build_ext -i
$ python3 -m drgn --help


User Guide

Quick Start

drgn debugs the running kernel by default; run sudo drgn. To debug a running program, run sudo drgn -p $PID. To debug a core dump (either a kernel vmcore or a userspace core dump), run drgn -c $PATH. Make sure to install debugging symbols for whatever you are debugging.

Then, you can access variables in the program with prog['name'] and access structure members with .:

$ sudo drgn
>>> prog['init_task'].comm
(char [16])"swapper/0"


You can use various predefined helpers:

>>> len(list(bpf_prog_for_each()))
11
>>> task = find_task(115)
>>> cmdline(task)
[b'findmnt', b'-p']


You can get stack traces with stack_trace() and access parameters or local variables with trace['name']:

>>> trace = stack_trace(task)
>>> trace[5]
#5 at 0xffffffff8a5a32d0 (do_sys_poll+0x400/0x578) in do_poll at ./fs/select.c:961:8 (inlined)
>>> poll_list = trace[5]['list']
>>> file = fget(task, poll_list.entries[0].fd)
>>> d_path(file.f_path.address_of_())
b'/proc/115/mountinfo'


Core Concepts

The most important interfaces in drgn are programs, objects, and helpers.

Programs

A program being debugged is represented by an instance of the drgn.Program class. The drgn CLI is initialized with a Program named prog; unless you are using the drgn library directly, this is usually the only Program you will need.

A Program is used to look up type definitions, access variables, and read arbitrary memory:

>>> prog.type('unsigned long')
prog.int_type(name='unsigned long', size=8, is_signed=False)
>>> prog['jiffies']
Object(prog, 'volatile unsigned long', address=0xffffffffbe405000)
>>> prog.read(0xffffffffbe411e10, 16)
b'swapper/0\x00\x00\x00\x00\x00\x00\x00'


The drgn.Program.type(), drgn.Program.variable(), drgn.Program.constant(), and drgn.Program.function() methods look up those various things in a program. drgn.Program.read() reads memory from the program's address space. The [] operator looks up a variable, constant, or function:

>>> prog['jiffies'] == prog.variable('jiffies')
True


It is usually more convenient to use the [] operator rather than the variable(), constant(), or function() methods unless the program has multiple objects with the same name, in which case the methods provide more control.

Objects

Variables, constants, functions, and computed values are all called objects in drgn. Objects are represented by the drgn.Object class. An object may exist in the memory of the program (a reference):

>>> Object(prog, 'int', address=0xffffffffc09031a0)


Or, an object may be a constant or temporary computed value (a value):

>>> Object(prog, 'int', value=4)


What makes drgn scripts expressive is that objects can be used almost exactly like they would be in the program's own source code. For example, structure members can be accessed with the dot (.) operator, arrays can be subscripted with [], arithmetic can be performed, and objects can be compared:

>>> print(prog['init_task'].comm[0])
(char)115
>>> print(repr(prog['init_task'].nsproxy.mnt_ns.mounts + 1))
Object(prog, 'unsigned int', value=34)
>>> prog['init_task'].nsproxy.mnt_ns.pending_mounts > 0
False


Python doesn't have all of the operators that C or C++ do, so some substitutions are necessary:

  • Instead of *ptr, dereference a pointer with ptr[0].
  • Instead of ptr->member, access a member through a pointer with ptr.member.
  • Instead of &var, get the address of a variable with var.address_of_().

A common use case is converting a drgn.Object to a Python value so it can be used by a standard Python library. There are a few ways to do this:

  • The drgn.Object.value_() method gets the value of the object with the directly corresponding Python type (i.e., integers and pointers become int, floating-point types become float, booleans become bool, arrays become list, structures and unions become dict).
  • The drgn.Object.string_() method gets a null-terminated string as bytes from an array or pointer.
  • The int(), float(), and bool() functions do an explicit conversion to that Python type.

Objects have several attributes; the most important are drgn.Object.prog_ and drgn.Object.type_. The former is the drgn.Program that the object is from, and the latter is the drgn.Type of the object.

Note that all attributes and methods of the Object class end with an underscore (_) in order to avoid conflicting with structure or union members. The Object attributes and methods always take precedence; use drgn.Object.member_() if there is a conflict.

References vs. Values

The main difference between reference objects and value objects is how they are evaluated. References are read from the program's memory every time they are evaluated; values simply return the stored value (drgn.Object.read_() reads a reference object and returns it as a value object):

>>> import time
>>> jiffies = prog['jiffies']
>>> jiffies.value_()
4391639989
>>> time.sleep(1)
>>> jiffies.value_()
4391640290
>>> jiffies2 = jiffies.read_()
>>> jiffies2.value_()
4391640291
>>> time.sleep(1)
>>> jiffies2.value_()
4391640291
>>> jiffies.value_()
4391640593


References have a drgn.Object.address_ attribute, which is the object's address as a Python int. This is slightly different from the drgn.Object.address_of_() method, which returns the address as a drgn.Object. Of course, both references and values can have a pointer type; address_ refers to the address of the pointer object itself, and drgn.Object.value_() refers to the value of the pointer (i.e., the address it points to):

>>> address = prog['jiffies'].address_
>>> type(address)
<class 'int'>
>>> print(hex(address))
0xffffffffbe405000
>>> jiffiesp = prog['jiffies'].address_of_()
>>> jiffiesp
Object(prog, 'volatile unsigned long *', value=0xffffffffbe405000)
>>> print(hex(jiffiesp.value_()))
0xffffffffbe405000


Absent Objects

In addition to reference objects and value objects, objects may also be absent.

>>> Object(prog, "int").value_()
Traceback (most recent call last):


  File "<console>", line 1, in <module>
_drgn.ObjectAbsentError: object absent

This represents an object whose value or address is not known. For example, this can happen if the object was optimized out of the program by the compiler.

Any attempt to operate on an absent object results in a drgn.ObjectAbsentError exception, although basic information including its type may still be accessed.

Helpers

Some programs have common data structures that you may want to examine. For example, consider linked lists in the Linux kernel:

struct list_head {


    struct list_head *next, *prev;
};
#define list_for_each(pos, head) \


    for (pos = (head)->next; pos != (head); pos = pos->next)


When working with these lists, you'd probably want to define a function:

def list_for_each(head):


    pos = head.next


    while pos != head:


        yield pos


        pos = pos.next


Then, you could use it like so for any list you need to look at:

>>> for pos in list_for_each(head):
...     do_something_with(pos)


Of course, it would be a waste of time and effort for everyone to have to define these helpers for themselves, so drgn includes a collection of helpers for many use cases. See Helpers.

Validators

Validators are a special category of helpers that check the consistency of a data structure. In general, helpers assume that the data structures that they examine are valid. Validators do not make this assumption and do additional (potentially expensive) checks to detect broken invariants, corruption, etc.

Validators raise drgn.helpers.ValidationError if the data structure is not valid or drgn.FaultError if the data structure is invalid in a way that causes a bad memory access. They have names prefixed with validate_.

For example, drgn.helpers.linux.list.validate_list() checks the consistency of a linked list in the Linux kernel (in particular, the consistency of the next and prev pointers):

>>> validate_list(prog["my_list"].address_of_())
drgn.helpers.ValidationError: (struct list_head *)0xffffffffc029e460 next 0xffffffffc029e000 has prev 0xffffffffc029e450


drgn.helpers.linux.list.validate_list_for_each_entry() does the same checks while also returning the entries in the list for further validation:

def validate_my_list(prog):


     for entry in validate_list_for_each_entry(


         "struct my_entry",


         prog["my_list"].address_of_(),


         "list",


     ):


         if entry.value < 0:


             raise ValidationError("list contains negative entry")


Other Concepts

In addition to the core concepts above, drgn provides a few additional abstractions.

Threads

The drgn.Thread class represents a thread. drgn.Program.threads(), drgn.Program.thread(), drgn.Program.main_thread(), and drgn.Program.crashed_thread() can be used to find threads:

>>> for thread in prog.threads():
...     print(thread.tid)
...
39143
39144
>>> print(prog.main_thread().tid)
39143
>>> print(prog.crashed_thread().tid)
39144


Stack Traces

drgn represents stack traces with the drgn.StackTrace and drgn.StackFrame classes. drgn.stack_trace(), drgn.Program.stack_trace(), and drgn.Thread.stack_trace() return the call stack for a thread. The [] operator looks up an object in the scope of a StackFrame:

>>> trace = stack_trace(115)
>>> trace
#0  context_switch (./kernel/sched/core.c:4683:2)
#1  __schedule (./kernel/sched/core.c:5940:8)
#2  schedule (./kernel/sched/core.c:6019:3)
#3  schedule_hrtimeout_range_clock (./kernel/time/hrtimer.c:2148:3)
#4  poll_schedule_timeout (./fs/select.c:243:8)
#5  do_poll (./fs/select.c:961:8)
#6  do_sys_poll (./fs/select.c:1011:12)
#7  __do_sys_poll (./fs/select.c:1076:8)
#8  __se_sys_poll (./fs/select.c:1064:1)
#9  __x64_sys_poll (./fs/select.c:1064:1)
#10 do_syscall_x64 (./arch/x86/entry/common.c:50:14)
#11 do_syscall_64 (./arch/x86/entry/common.c:80:7)
#12 entry_SYSCALL_64+0x7c/0x15b (./arch/x86/entry/entry_64.S:113)
#13 0x7f3344072af7
>>> trace[5]
#5 at 0xffffffff8a5a32d0 (do_sys_poll+0x400/0x578) in do_poll at ./fs/select.c:961:8 (inlined)
>>> prog['do_poll']
(int (struct poll_list *list, struct poll_wqueues *wait, struct timespec64 *end_time))<absent>
>>> trace[5]['list']
*(struct poll_list *)0xffffacca402e3b50 = {


        .next = (struct poll_list *)0x0,


        .len = (int)1,


        .entries = (struct pollfd []){},
}


Symbols

The symbol table of a program is a list of identifiers along with their address and size. drgn represents symbols with the drgn.Symbol class, which is returned by drgn.Program.symbol().

Types

drgn automatically obtains type definitions from the program. Types are represented by the drgn.Type class and created by various factory functions like drgn.Program.int_type():

>>> prog.type('int')
prog.int_type(name='int', size=4, is_signed=True)


You won't usually need to work with types directly, but see Types if you do.

Platforms

Certain operations and objects in a program are platform-dependent; drgn allows accessing the platform that a program runs with the drgn.Platform class.

Command Line Interface

The drgn CLI is basically a wrapper around the drgn library which automatically creates a drgn.Program. The CLI can be run in interactive mode or script mode.

Script Mode

Script mode is useful for reusable scripts. Simply pass the path to the script along with any arguments:

$ cat script.py
import sys
from drgn.helpers.linux import find_task
pid = int(sys.argv[1])
uid = find_task(pid).cred.uid.val.value_()
print(f'PID {pid} is being run by UID {uid}')
$ sudo drgn script.py 601
PID 601 is being run by UID 1000


It's even possible to run drgn scripts directly with the proper shebang:

$ cat script2.py
#!/usr/bin/env drgn
mounts = prog['init_task'].nsproxy.mnt_ns.mounts.value_()
print(f'You have {mounts} filesystems mounted')
$ sudo ./script2.py
You have 36 filesystems mounted


Interactive Mode

Interactive mode uses the Python interpreter's interactive mode and adds a few nice features, including:

  • History
  • Tab completion
  • Automatic import of relevant helpers
  • Pretty printing of objects and types

The default behavior of the Python REPL is to print the output of repr(). For drgn.Object and drgn.Type, this is a raw representation:

>>> print(repr(prog['jiffies']))
Object(prog, 'volatile unsigned long', address=0xffffffffbe405000)
>>> print(repr(prog.type('atomic_t')))
prog.typedef_type(name='atomic_t', type=prog.struct_type(tag=None, size=4, members=(TypeMember(prog.type('int'), name='counter', bit_offset=0),)))


The standard print() function uses the output of str(). For drgn objects and types, this is a representation in programming language syntax:

>>> print(prog['jiffies'])
(volatile unsigned long)4395387628
>>> print(prog.type('atomic_t'))
typedef struct {


        int counter;
} atomic_t


In interactive mode, the drgn CLI automatically uses str() instead of repr() for objects and types, so you don't need to call print() explicitly:

$ sudo drgn
>>> prog['jiffies']
(volatile unsigned long)4395387628
>>> prog.type('atomic_t')
typedef struct {


        int counter;
} atomic_t


Next Steps

Refer to the API Reference. Look through the Helpers. Read some Case Studies. Browse through the tools. Check out the community contributions.

Advanced Usage

The User Guide covers basic usage of drgn, but drgn also supports more advanced use cases which are covered here.

Loading Debugging Symbols

drgn will automatically load debugging information based on the debugged program (e.g., from loaded kernel modules or loaded shared libraries). drgn.Program.load_debug_info() can be used to load additional debugging information:

>>> prog.load_debug_info(['./libfoo.so', '/usr/lib/libbar.so'])


Library

In addition to the CLI, drgn is also available as a library. drgn.program_from_core_dump(), drgn.program_from_kernel(), and drgn.program_from_pid() correspond to the -c, -k, and -p command line options, respectively; they return a drgn.Program that can be used just like the one initialized by the CLI:

>>> import drgn
>>> prog = drgn.program_from_kernel()


C Library

The core functionality of drgn is implemented in C and is available as a C library, libdrgn. See drgn.h.

Full documentation can be generated by running doxygen in the libdrgn directory of the source code. Note that the API and ABI are not yet stable.

Custom Programs

The main components of a drgn.Program are the program memory, types, and symbols. The CLI and equivalent library interfaces automatically determine these. However, it is also possible to create a "blank" Program and plug in the main components. The drgn.cli.run_interactive() function allows you to run the same drgn CLI once you've created a drgn.Program, so it's easy to make a custom program which allows interactive debugging.

drgn.Program.add_memory_segment() defines a range of memory and how to read that memory. The following example uses a Btrfs filesystem image as the program "memory":

import drgn
import os
import sys
from drgn.cli import run_interactive
def btrfs_debugger(dev):


    file = open(dev, 'rb')


    size = file.seek(0, 2)


    def read_file(address, count, offset, physical):


        file.seek(offset)


        return file.read(count)


    platform = drgn.Platform(drgn.Architecture.UNKNOWN,


                             drgn.PlatformFlags.IS_LITTLE_ENDIAN)


    prog = drgn.Program(platform)


    prog.add_memory_segment(0, size, read_file)


    prog.load_debug_info([f'/lib/modules/{os.uname().release}/kernel/fs/btrfs/btrfs.ko'])


    return prog
prog = btrfs_debugger(sys.argv[1] if len(sys.argv) >= 2 else '/dev/sda')
print(drgn.Object(prog, 'struct btrfs_super_block', address=65536))
run_interactive(prog, banner_func=lambda _: "BTRFS debugger")


drgn.Program.add_type_finder() and drgn.Program.add_object_finder() are the equivalent methods for plugging in types and objects.

Environment Variables

Some of drgn's behavior can be modified through environment variables:

The maximum number of individual errors to report in a drgn.MissingDebugInfoError. Any additional errors are truncated. The default is 5; -1 is unlimited.
Whether to prefer using ORC over DWARF for stack unwinding (0 or 1). The default is 0. Note that drgn will always fall back to ORC for functions lacking DWARF call frame information and vice versa. This environment variable is mainly intended for testing and may be ignored in the future.
Whether drgn should use libdwfl to find debugging information for core dumps instead of its own implementation (0 or 1). The default is 0. This environment variable is mainly intended as an escape hatch in case of bugs in drgn's implementation and will be ignored in the future.
Whether drgn should use libkdumpfile for ELF vmcores (0 or 1). The default is 0. This functionality will be removed in the future.
Whether drgn should use /sys/module to find information about loaded kernel modules for the running kernel instead of getting them from the core dump (0 or 1). The default is 1. This environment variable is mainly intended for testing and may be ignored in the future.

Linux Kernel Special Objects

When debugging the Linux kernel, there are some special drgn.Objects accessible with drgn.Program.object() and drgn.Program[]. Some of these are available even without debugging information, thanks to metadata called "vmcoreinfo" which is present in kernel core dumps. These special objects include:

Object type: const char []

This corresponds to the UTS_RELEASE macro in the Linux kernel source code. This is the exact kernel release (i.e., the output of uname -r).

To use this as a Python string, you must convert it:

>>> release = prog["UTS_RELEASE"].string_().decode("ascii")


This is available without debugging information.

Object type: unsigned long
Object type: unsigned int
Object type: unsigned long

These correspond to the macros of the same name in the Linux kernel source code. The page size is the smallest contiguous unit of physical memory which can be allocated or mapped by the kernel.

>>> prog['PAGE_SIZE']
(unsigned long)4096
>>> prog['PAGE_SHIFT']
(int)12
>>> prog['PAGE_MASK']
(unsigned long)18446744073709547520
>>> 1 << prog['PAGE_SHIFT'] == prog['PAGE_SIZE']
True
>>> ~(prog['PAGE_SIZE'] - 1) == prog['PAGE_MASK']
True

These are available without debugging information.

Object type: volatile unsigned long

This is a counter of timer ticks. It is actually an alias of jiffies_64 on 64-bit architectures, or the least significant 32 bits of jiffies_64 on 32-bit architectures. Since this alias is defined via the linker, drgn handles it specially.

This is not available without debugging information.

Object type: struct page *

This is a pointer to the "virtual memory map", an array of struct page for each physical page of memory. While the purpose and implementation details of this array are beyond the scope of this documentation, it is enough to say that it is represented in the kernel source in an architecture-dependent way, frequently as a macro or constant. The definition provided by drgn ensures that users can access it without resorting to architecture-specific logic.

This is not available without debugging information.

Object type: const char []

This is the data contained in the vmcoreinfo note, which is present either as an ELF note in /proc/kcore or ELF vmcores, or as a special data section in kdump-formatted vmcores. The vmcoreinfo note contains critical data necessary for interpreting the kernel image, such as KASLR offsets and data structure locations.

In the Linux kernel, this data is normally stored in a variable called vmcoreinfo_data. However, drgn reads this information from ELF note or from the diskdump header. It is possible (in rare cases, usually with vmcores created by hypervisors) for a vmcore to contain vmcoreinfo which differs from the data in vmcoreinfo_data, so it is important to distinguish the contents. For that reason, we use the name VMCOREINFO to distinguish it from the kernel variable vmcoreinfo_data.

This is available without debugging information.


API Reference

Programs

A Program represents a crashed or running program. It can be used to lookup type definitions, access variables, and read arbitrary memory.

The main functionality of a Program is looking up objects (i.e., variables, constants, or functions). This is usually done with the [] operator.

Create a Program with no target program. It is usually more convenient to use one of the Program Constructors.
platform -- The platform of the program, or None if it should be determined automatically when a core dump or symbol file is added.


Flags which apply to this program.

Platform that this program runs on, or None if it has not been determined yet.

Default programming language of the program.

This is used for interpreting the type name given to type() and when creating an Object without an explicit type.

For the Linux kernel, this defaults to Language.C. For userspace programs, this defaults to the language of main in the program, falling back to Language.C. This heuristic may change in the future.

This can be explicitly set to a different language (e.g., if the heuristic was incorrect).


__getitem__(name: str) -> Object
Implement self[name]. Get the object (variable, constant, or function) with the given name.

This is equivalent to prog.object(name) except that this raises KeyError instead of LookupError if no objects with the given name are found.

If there are multiple objects with the same name, one is returned arbitrarily. In this case, the variable(), constant(), function(), or object() methods can be used instead.

>>> prog['jiffies']
Object(prog, 'volatile unsigned long', address=0xffffffff94c05000)
name -- Object name.


__contains__(name: str) -> bool
Implement name in self. Return whether an object (variable, constant, or function) with the given name exists in the program.
name -- Object name.


Get the variable with the given name. 

>>> prog.variable('jiffies')
Object(prog, 'volatile unsigned long', address=0xffffffff94c05000)

This is equivalent to prog.object(name, FindObjectFlags.VARIABLE, filename).

  • name -- The variable name.
  • filename -- The source code file that contains the definition. See Filenames.

LookupError -- if no variables with the given name are found in the given file


Get the constant (e.g., enumeration constant) with the given name.

Note that support for macro constants is not yet implemented for DWARF files, and most compilers don't generate macro debugging information by default anyways.

>>> prog.constant('PIDTYPE_MAX')
Object(prog, 'enum pid_type', value=4)

This is equivalent to prog.object(name, FindObjectFlags.CONSTANT, filename).

  • name -- The constant name.
  • filename -- The source code file that contains the definition. See Filenames.

LookupError -- if no constants with the given name are found in the given file


Get the function with the given name. 

>>> prog.function('schedule')
Object(prog, 'void (void)', address=0xffffffff94392370)

This is equivalent to prog.object(name, FindObjectFlags.FUNCTION, filename).

  • name -- The function name.
  • filename -- The source code file that contains the definition. See Filenames.

LookupError -- if no functions with the given name are found in the given file


Get the object (variable, constant, or function) with the given name.

When debugging the Linux kernel, this can look up certain special objects documented in Linux Kernel Special Objects, sometimes without any debugging information loaded.

  • name -- The object name.
  • flags -- Flags indicating what kind of object to look for.
  • filename -- The source code file that contains the definition. See Filenames.

LookupError -- if no objects with the given name are found in the given file


Get a symbol containing the given address, or a symbol with the given name.

Global symbols are preferred over weak symbols, and weak symbols are preferred over other symbols. In other words: if a matching SymbolBinding.GLOBAL or SymbolBinding.UNIQUE symbol is found, it is returned. Otherwise, if a matching SymbolBinding.WEAK symbol is found, it is returned. Otherwise, any matching symbol (e.g., SymbolBinding.LOCAL) is returned. If there are multiple matching symbols with the same binding, one is returned arbitrarily. To retrieve all matching symbols, use symbols().

address_or_name -- Address or name to search for. This parameter is positional-only.
LookupError -- if no symbol contains the given address or matches the given name


Get a list of global and local symbols, optionally matching a name or address.

If a string argument is given, this returns all symbols matching that name. If an integer-like argument given, this returns a list of all symbols containing that address. If no argument is given, all symbols in the program are returned. In all cases, the symbols are returned in an unspecified order.

address_or_name -- Address or name to search for. This parameter is positional-only.


Get the stack trace for the given thread in the program.

thread may be a thread ID (as defined by gettid(2)), in which case this will unwind the stack for the thread with that ID. The ID may be a Python int or an integer Object

thread may also be a struct pt_regs or struct pt_regs * object, in which case the initial register values will be fetched from that object.

Finally, if debugging the Linux kernel, thread may be a struct task_struct * object, in which case this will unwind the stack for that task. See drgn.helpers.linux.pid.find_task().

This is implemented for the Linux kernel (both live and core dumps) as well as userspace core dumps; it is not yet implemented for live userspace processes.

thread -- Thread ID, struct pt_regs object, or struct task_struct * object.


Get a stack trace with the supplied list of program counters.
pcs -- List of program counters.


Get the type with the given name. 

>>> prog.type('long')
prog.int_type(name='long', size=8, is_signed=True)
  • name -- The type name.
  • filename -- The source code file that contains the definition. See Filenames.

LookupError -- if no types with the given name are found in the given file


Return the given type.

This is mainly useful so that helpers can use prog.type() to get a Type regardless of whether they were given a str or a Type. For example:

def my_helper(obj: Object, type: Union[str, Type]) -> bool:


    # type may be str or Type.


    type = obj.prog_.type(type)


    # type is now always Type.


    return sizeof(obj) > sizeof(type)


type -- Type.
The exact same type.


Get an iterator over all of the threads in the program.

Get the thread with the given thread ID.
tid -- Thread ID (as defined by gettid(2)).
LookupError -- if no thread has the given thread ID


Get the main thread of the program.

This is only defined for userspace programs.

ValueError -- if the program is the Linux kernel


Get the thread that caused the program to crash.

For userspace programs, this is the thread that received the fatal signal (e.g., SIGSEGV or SIGQUIT).

For the kernel, this is the thread that panicked (either directly or as a result of an oops, BUG_ON(), etc.).

ValueError -- if the program is live (i.e., not a core dump)


Read size bytes of memory starting at address in the program. The address may be virtual (the default) or physical if the program supports it. 

>>> prog.read(0xffffffffbe012b40, 16)
b'swapper/0'
  • address -- The starting address.
  • size -- The number of bytes to read.
  • physical -- Whether address is a physical memory address. If False, then it is a virtual memory address. Physical memory can usually only be read when the program is an operating system kernel.

  • FaultError -- if the address range is invalid or the type of address (physical or virtual) is not supported by the program
  • ValueError -- if size is negative







Read an unsigned integer from the program's memory in the program's byte order.

read_u8(), read_u16(), read_u32(), and read_u64() read an 8-, 16-, 32-, or 64-bit unsigned integer, respectively. read_word() reads a program word-sized unsigned integer.

For signed integers, alternate byte order, or other formats, you can use read() and int.from_bytes() or the struct module.

  • address -- Address of the integer.
  • physical -- Whether address is a physical memory address; see read().

FaultError -- if the address is invalid; see read()


Define a region of memory in the program.

If it overlaps a previously registered segment, the new segment takes precedence.

  • address -- Address of the segment.
  • size -- Size of the segment in bytes.
  • physical -- Whether to add a physical memory segment. If False, then this adds a virtual memory segment.
  • read_fn -- Callable to call to read memory from the segment. It is passed the address being read from, the number of bytes to read, the offset in bytes from the beginning of the segment, and whether the address is physical: (address, count, offset, physical). It should return the requested number of bytes as bytes or another buffer type.



Register a callback for finding types in the program.

Callbacks are called in reverse order of the order they were added until the type is found. So, more recently added callbacks take precedence.

fn -- Callable taking a TypeKind, name, and filename: (kind, name, filename). The filename should be matched with filename_matches(). This should return a Type or None if not found.


Register a callback for finding objects in the program.

Callbacks are called in reverse order of the order they were added until the object is found. So, more recently added callbacks take precedence.

fn -- Callable taking a program, name, FindObjectFlags, and filename: (prog, name, flags, filename). The filename should be matched with filename_matches(). This should return an Object or None if not found.


Set the program to a core dump.

This loads the memory segments from the core dump and determines the mapped executable and libraries. It does not load any debugging symbols; see load_default_debug_info().

path -- Core dump file path or open file descriptor.


Set the program to the running operating system kernel.

This loads the memory of the running kernel and thus requires root privileges. It does not load any debugging symbols; see load_default_debug_info().


Set the program to a running process.

This loads the memory of the process and determines the mapped executable and libraries. It does not load any debugging symbols; see load_default_debug_info().

pid -- Process ID.


Load debugging information for a list of executable or library files.

Note that this is parallelized, so it is usually faster to load multiple files at once rather than one by one.

  • paths -- Paths of binary files.
  • default --

    Also load debugging information which can automatically be determined from the program.

    For the Linux kernel, this tries to load vmlinux and any loaded kernel modules from a few standard locations.

    For userspace programs, this tries to load the executable and any loaded libraries.

    This implies main=True.

  • main --

    Also load debugging information for the main executable.

    For the Linux kernel, this tries to load vmlinux.

    This is currently ignored for userspace programs.


MissingDebugInfoError -- if debugging information was not available for some files; other files with debugging information are still loaded


Load debugging information which can automatically be determined from the program.

This is equivalent to load_debug_info(None, True).


Dictionary for caching program metadata.

This isn't used by drgn itself. It is intended to be used by helpers to cache metadata about the program. For example, if a helper for a program depends on the program version or an optional feature, the helper can detect it and cache it for subsequent invocations:

def my_helper(prog):


    try:


        have_foo = prog.cache['have_foo']


    except KeyError:


        have_foo = detect_foo_feature(prog)


        prog.cache['have_foo'] = have_foo


    if have_foo:


        return prog['foo']


    else:


        return prog['bar']




Bases: enum.Flag

ProgramFlags are flags that can apply to a Program (e.g., about what kind of program it is).

The program is the Linux kernel.

The program is currently running (e.g., it is the running operating system kernel or a running process).

The program is running on the local machine.


Bases: enum.Flag

FindObjectFlags are flags for Program.object(). These can be combined to search for multiple kinds of objects at once.






A thread in a program.
Thread ID (as defined by gettid(2)).

If the program is the Linux kernel, the struct task_struct * object for this thread. Otherwise, not defined.

Get the stack trace for this thread.

This is equivalent to prog.stack_trace(thread.tid). See Program.stack_trace().



Filenames

The Program.type(), Program.object(), Program.variable(), Program.constant(), and Program.function() methods all take a filename parameter to distinguish between multiple definitions with the same name. The filename refers to the source code file that contains the definition. It is matched with filename_matches(). If multiple definitions match, one is returned arbitrarily.

Return whether a filename containing a definition (haystack) matches a filename being searched for (needle).

The filename is matched from right to left, so 'stdio.h', 'include/stdio.h', 'usr/include/stdio.h', and '/usr/include/stdio.h' would all match a definition in /usr/include/stdio.h. If needle is None or empty, it matches any definition. If haystack is None or empty, it only matches if needle is also None or empty.

  • haystack -- Path of file containing definition.
  • needle -- Filename to match.



Program Constructors

The drgn command line interface automatically creates a Program named prog. However, drgn may also be used as a library without the CLI, in which case a Program must be created manually.

Create a Program from a core dump file. The type of program (e.g., userspace or kernel) is determined automatically.
path -- Core dump file path or open file descriptor.


Create a Program from the running operating system kernel. This requires root privileges.

Create a Program from a running program with the given PID. This requires appropriate permissions (on Linux, ptrace(2) attach permissions).
pid -- Process ID of the program to debug.


Default Program

Most functions that take a Program can be called without the prog argument. In that case, the default program argument is used, which is determined by the rules below.

NOTE:

In the drgn CLI, you probably don't need to care about these details. Simply omit prog: 

# Equivalent in the CLI.
find_task(pid)
find_task(prog, pid)
find_task(prog["init_pid_ns"].address_of_(), pid)




1.
If prog is given explicitly, either as a positional or keyword argument, then it is used.
2.
Otherwise, if the first argument is an Object, then Object.prog_ is used.
3.
Otherwise, the default program is used.

The default program is set automatically in the CLI. Library users can get and set it manually. The default program is a per-thread setting. See Thread Safety.

Get the default program for the current thread.
NoDefaultProgramError -- if the default program is not set


Set the default program for the current thread.
prog -- Program to set as the default, or None to unset it.


Bases: Exception

Error raised when trying to use the default program when it is not set.


For helpers, it is recommended to use the decorators from the drgn.helpers.common.prog module instead.

Platforms

A Platform represents the environment (i.e., architecture and ABI) that a program runs on.
Create a Platform.
  • arch -- Platform.arch
  • flags -- Platform.flags; if None, default flags for the architecture are used.



Instruction set architecture of this platform.

Flags which apply to this platform.

Processor registers on this platform.


Bases: enum.Enum

An Architecture represents an instruction set architecture.

The x86-64 architecture, a.k.a. AMD64 or Intel 64.

The 32-bit x86 architecture, a.k.a. i386 or IA-32.

The AArch64 architecture, a.k.a. ARM64.

The 32-bit Arm architecture.

The 64-bit PowerPC architecture.

The 64-bit RISC-V architecture.

The 32-bit RISC-V architecture.

The s390x architecture, a.k.a. IBM Z or z/Architecture.

The 32-bit s390 architecture, a.k.a. System/390.

An architecture which is not known to drgn. Certain features are not available when the architecture is unknown, but most of drgn will still work.


Bases: enum.Flag

PlatformFlags are flags describing a Platform.

Platform is 64-bit.

Platform is little-endian.


A Register represents information about a processor register.
Names of this register.


The platform of the host which is running drgn.

Languages

A Language represents a programming language supported by drgn.

This class cannot be constructed; there are singletons for the supported languages.

Name of the programming language.

The C programming language.

The C++ programming language.


Objects

An Object represents a symbol or value in a program. An object may exist in the memory of the program (a reference), it may be a constant or temporary computed value (a value), or it may be absent entirely (an absent object).

All instances of this class have two attributes: prog_, the program that the object is from; and type_, the type of the object. Reference objects also have an address_ and a bit_offset_. Objects may also have a bit_field_size_.

repr() of an object returns a Python representation of the object:

>>> print(repr(prog['jiffies']))
Object(prog, 'volatile unsigned long', address=0xffffffffbf005000)

str() returns a "pretty" representation of the object in programming language syntax:

>>> print(prog['jiffies'])
(volatile unsigned long)4326237045

The output format of str() can be modified by using the format_() method instead:

>>> sysname = prog['init_uts_ns'].name.sysname
>>> print(sysname)
(char [65])"Linux"
>>> print(sysname.format_(type_name=False))
"Linux"
>>> print(sysname.format_(string=False))
(char [65]){ 76, 105, 110, 117, 120 }

NOTE:

The drgn CLI is set up so that objects are displayed in the "pretty" format instead of with repr() (the latter is the default behavior of Python's interactive mode). Therefore, it's usually not necessary to call print() in the drgn CLI.


Objects support the following operators:

  • Arithmetic operators: +, -, *, /, %
  • Bitwise operators: <<, >>, &, |, ^, ~
  • Relational operators: ==, !=, <, >, <=, >=
  • Subscripting: [] (Python does not have a unary * operator, so pointers are dereferenced with ptr[0])
  • Member access: . (Python does not have a -> operator, so . is also used to access members of pointers to structures)
  • The address-of operator: drgn.Object.address_of_() (this is a method because Python does not have a & operator)
  • Array length: len()

These operators all have the semantics of the program's programming language. For example, adding two objects from a program written in C results in an object with a type and value according to the rules of C:

>>> Object(prog, 'unsigned long', 2**64 - 1) + Object(prog, 'int', 1)
Object(prog, 'unsigned long', value=0)

If only one operand to a binary operator is an object, the other operand will be converted to an object according to the language's rules for literals:

>>> Object(prog, 'char', 0) - 1
Object(prog, 'int', value=-1)

The standard int(), float(), and bool() functions convert an object to that Python type. Conversion to bool uses the programming language's notion of "truthiness". Additionally, certain Python functions will automatically coerce an object to the appropriate Python type (e.g., hex(), round(), and list subscripting).

Object attributes and methods are named with a trailing underscore to avoid conflicting with structure, union, or class members. The attributes and methods always take precedence; use member_() if there is a conflict.

Objects are usually obtained directly from a Program, but they can be constructed manually, as well (for example, if you got a variable address from a log file).

Create a value object given its type and value.
  • prog -- Program to create the object in.
  • type -- Type of the object.
  • value -- Value of the object. See value_().
  • bit_field_size -- Size in bits of the object if it is a bit field. The default is None, which means the object is not a bit field.



Create a value object from a "literal".

This is used to emulate a literal number in the source code of the program. The type is deduced from value according to the language's rules for literals.

value -- Value of the literal.


Create a reference object.
  • address -- Address of the object in the program.
  • bit_offset -- Offset in bits from address to the beginning of the object.



Create an absent object.

Program that this object is from.

Type of this object.

Whether this object is absent.

This is False for all values and references (even if the reference has an invalid address).


Address of this object if it is a reference, None if it is a value or absent.

Offset in bits from this object's address to the beginning of the object if it is a reference, None otherwise. This can only be non-zero for scalars.

Size in bits of this object if it is a bit field, None if it is not.

__getattr__(name: str) -> Object
Implement self.name.

This corresponds to both the member access (.) and member access through pointer (->) operators in C.

Note that if name is an attribute or method of the Object class, then that takes precedence. Otherwise, this is equivalent to member_().

>>> print(prog['init_task'].pid)
(pid_t)0
name -- Attribute name.


__getitem__(idx: IntegerLike) -> Object
Implement self[idx]. Get the array element at the given index. 

>>> print(prog['init_task'].comm[1])
(char)119

[0] is also the equivalent of the pointer dereference (*) operator in C:

>>> ptr_to_ptr
*(void **)0xffff9b86801e2968 = 0xffff9b86801e2460
>>> print(ptr_to_ptr[0])
(void *)0xffff9b86801e2460

This is only valid for pointers and arrays.

NOTE:

Negative indices behave as they would in the object's language (as opposed to the Python semantics of indexing from the end of the array).


idx -- The array index.
TypeError -- if this object is not a pointer or array


__len__() -> int
Implement len(self). Get the number of elements in this object. 

>>> len(prog['init_task'].comm)
16

This is only valid for arrays.

TypeError -- if this object is not an array with complete type


Get the value of this object as a Python object.

For basic types (integer, floating-point, boolean), this returns an object of the directly corresponding Python type (int, float, bool). For pointers, this returns the address value of the pointer. For enums, this returns an int. For structures and unions, this returns a dict of members. For arrays, this returns a list of values.

  • FaultError -- if reading the object causes a bad memory access
  • TypeError -- if this object has an unreadable type (e.g., void)



Read a null-terminated string pointed to by this object.

This is only valid for pointers and arrays. The element type is ignored; this operates byte-by-byte.

For pointers and flexible arrays, this stops at the first null byte.

For complete arrays, this stops at the first null byte or at the end of the array.

  • FaultError -- if reading the string causes a bad memory access
  • TypeError -- if this object is not a pointer or array



Get a member of this object.

This is valid for structures, unions, classes, and pointers to any of those. If the object is a pointer, it is automatically dereferenced first.

Normally the dot operator (.) can be used to accomplish the same thing, but this method can be used if there is a name conflict with an Object attribute or method.

name -- Name of the member.
  • TypeError -- if this object is not a structure, union, class, or a pointer to one of those
  • LookupError -- if this object does not have a member with the given name



Get a pointer to this object.

This corresponds to the address-of (&) operator in C. It is only possible for reference objects, as value objects don't have an address in the program.

As opposed to address_, this returns an Object, not an int.

ValueError -- if this object is a value


Read this object (which may be a reference or a value) and return it as a value object.

This is useful if the object can change in the running program (but of course nothing stops the program from modifying the object while it is being read).

As opposed to value_(), this returns an Object, not a standard Python type.

  • FaultError -- if reading this object causes a bad memory access
  • TypeError -- if this object has an unreadable type (e.g., void)



Return the binary representation of this object's value.

Return a value object from its binary representation.
  • prog -- Program to create the object in.
  • type -- Type of the object.
  • bytes -- Buffer containing value of the object.
  • bit_offset -- Offset in bits from the beginning of bytes to the beginning of the object.
  • bit_field_size -- Size in bits of the object if it is a bit field. The default is None, which means the object is not a bit field.



Format this object in programming language syntax.

Various format options can be passed (as keyword arguments) to control the output. Options that aren't passed or are passed as None fall back to a default. Specifically, obj.format_() (i.e., with no passed options) is equivalent to str(obj).

>>> workqueues = prog['workqueues']
>>> print(workqueues)
(struct list_head){


        .next = (struct list_head *)0xffff932ecfc0ae10,


        .prev = (struct list_head *)0xffff932e3818fc10,
}
>>> print(workqueues.format_(type_name=False,
...                          member_type_names=False,
...                          member_names=False,
...                          members_same_line=True))
{ 0xffff932ecfc0ae10, 0xffff932e3818fc10 }
  • columns -- Number of columns to limit output to when the expression can be reasonably wrapped. Defaults to no limit.
  • dereference -- If this object is a pointer, include the dereferenced value. This does not apply to structure, union, or class members, or array elements, as dereferencing those could lead to an infinite loop. Defaults to True.
  • symbolize -- Include a symbol name and offset for pointer objects. Defaults to True.
  • string -- Format the values of objects with string type as strings. For C, this applies to pointers to and arrays of char, signed char, and unsigned char. Defaults to True.
  • char -- Format objects with character type as character literals. For C, this applies to char, signed char, and unsigned char. Defaults to False.
  • type_name -- Include the type name of this object. Defaults to True.
  • member_type_names -- Include the type names of structure, union, and class members. Defaults to True.
  • element_type_names -- Include the type names of array elements. Defaults to False.
  • members_same_line -- Place multiple structure, union, and class members on the same line if they fit within the specified number of columns. Defaults to False.
  • elements_same_line -- Place multiple array elements on the same line if they fit within the specified number of columns. Defaults to True.
  • member_names -- Include the names of structure, union, and class members. Defaults to True.
  • element_indices -- Include the indices of array elements. Defaults to False.
  • implicit_members -- Include structure, union, and class members which have an implicit value (i.e., for C, zero-initialized). Defaults to True.
  • implicit_elements -- Include array elements which have an implicit value (i.e., for C, zero-initialized). Defaults to False.




Get an object representing NULL casted to the given type.

This is equivalent to Object(prog, type, 0).

  • prog -- The program.
  • type -- The type.



Get the value of the given object casted to another type.

Objects with a scalar type (integer, boolean, enumerated, floating-point, or pointer) can be casted to a different scalar type. Other objects can only be casted to the same type. This always results in a value object. See also drgn.reinterpret().

>>> cast("unsigned int", Object(prog, "float", 2.0))
(unsigned int)2
  • type -- Type to cast to.
  • obj -- Object to cast.



Get the representation of an object reinterpreted as another type.

This reinterprets the raw memory of the object, so an object can be reinterpreted as any other type. Reinterpreting a reference results in a reference, and reinterpreting a value results in a value. See also drgn.cast().

>>> reinterpret("unsigned int", Object(prog, "float", 2.0))
(unsigned int)1073741824
  • type -- Type to reinterpret as.
  • obj -- Object to reinterpret.



Get the containing object of a pointer object.

This corresponds to the container_of() macro in C.

  • ptr -- Pointer to member in containing object.
  • type -- Type of containing object.
  • member -- Name of member in containing object. May include one or more member references and zero or more array subscripts.

Pointer to containing object.
  • TypeError -- if ptr is not a pointer or type is not a structure, union, or class type
  • ValueError -- if the member is not byte-aligned (e.g., because it is a bit field)
  • LookupError -- if type does not have a member with the given name



Symbols

A Symbol represents an entry in the symbol table of a program, i.e., an identifier along with its corresponding address range in the program.
Name of this symbol.

Start address of this symbol.

Size of this symbol in bytes.

Linkage behavior and visibility of this symbol.

Kind of entity represented by this symbol.


Bases: enum.Enum

A SymbolBinding describes the linkage behavior and visibility of a symbol.

Unknown.

Not visible outside of the object file containing its definition.

Globally visible.

Globally visible but may be overridden by a non-weak global symbol.

Globally visible even if dynamic shared object is loaded locally. See GCC's -fno-gnu-unique option.


Bases: enum.Enum

A SymbolKind describes the kind of entity that a symbol represents.

Unknown or not defined.

Data object (e.g., variable or array).

Function or other executable code.

Object file section.

Source file.

Data object in common block.

Thread-local storage entity.

Indirect function.


Stack Traces

Stack traces are retrieved with stack_trace(), Program.stack_trace(), or Thread.stack_trace().

Get the stack trace for the given thread using the default program argument.

See Program.stack_trace() for more details.

thread -- Thread ID, struct pt_regs object, or struct task_struct * object.


A StackTrace is a sequence of StackFrame.

len(trace) is the number of stack frames in the trace. trace[0] is the innermost stack frame, trace[1] is its caller, and trace[len(trace) - 1] is the outermost frame. Negative indexing also works: trace[-1] is the outermost frame and trace[-len(trace)] is the innermost frame. It is also iterable:

for frame in trace:


    if frame.name == 'io_schedule':


        print('Thread is doing I/O')


str() returns a pretty-printed stack trace:

>>> prog.stack_trace(1)
#0  context_switch (kernel/sched/core.c:4339:2)
#1  __schedule (kernel/sched/core.c:5147:8)
#2  schedule (kernel/sched/core.c:5226:3)
#3  do_wait (kernel/exit.c:1534:4)
#4  kernel_wait4 (kernel/exit.c:1678:8)
#5  __do_sys_wait4 (kernel/exit.c:1706:13)
#6  do_syscall_64 (arch/x86/entry/common.c:47:14)
#7  entry_SYSCALL_64+0x7c/0x15b (arch/x86/entry/entry_64.S:112)
#8  0x4d49dd

The format is subject to change. The drgn CLI is set up so that stack traces are displayed with str() by default.

Program that this stack trace is from.


A StackFrame represents a single frame in a thread's call stack.

str() returns a pretty-printed stack frame:

>>> prog.stack_trace(1)[0]
#0 at 0xffffffffb64ac287 (__schedule+0x227/0x606) in context_switch at kernel/sched/core.c:4339:2 (inlined)

This includes more information than when printing the full stack trace. The format is subject to change. The drgn CLI is set up so that stack frames are displayed with str() by default.

The [] operator can look up function parameters, local variables, and global variables in the scope of the stack frame:

>>> prog.stack_trace(1)[0]['prev'].pid
(pid_t)1
>>> prog.stack_trace(1)[0]['scheduler_running']
(int)1
Name of the function at this frame, or None if it could not be determined.

The name cannot be determined if debugging information is not available for the function, e.g., because it is implemented in assembly. It may be desirable to use the symbol name or program counter as a fallback:

name = frame.name
if name is None:


    try:


        name = frame.symbol().name


    except LookupError:


        name = hex(frame.pc)



Whether this frame is for an inlined call.

An inline frame shares the same stack frame in memory as its caller. Therefore, it has the same registers (including program counter and thus symbol).


Whether this stack frame was interrupted (for example, by a hardware interrupt, signal, trap, etc.).

If this is True, then the register values in this frame are the values at the time that the frame was interrupted.

This is False if the frame is for a function call, in which case the register values are the values when control returns to this frame. In particular, the program counter is the return address, which is typically the instruction after the call instruction.


Program counter at this stack frame.

Stack pointer at this stack frame.

__getitem__(name: str) -> Object
Implement self[name]. Get the object (variable, function parameter, constant, or function) with the given name in the scope of this frame.

If the object exists but has been optimized out, this returns an absent object.

name -- Object name.


__contains__(name: str) -> bool
Implement name in self. Return whether an object with the given name exists in the scope of this frame.
name -- Object name.


Get a list of the names of all local objects (local variables, function parameters, local constants, and nested functions) in the scope of this frame.

Not all names may have present values, but they can be used with the [] operator to check.


Get the source code location of this frame.
Location as a (filename, line, column) triple.
LookupError -- if the source code location is not available


Get the function symbol at this stack frame.

This is equivalent to:

prog.symbol(frame.pc - (0 if frame.interrupted else 1))



Get the value of the given register at this stack frame.
reg -- Register name.
  • ValueError -- if the register name is not recognized
  • LookupError -- if the register value is not known



Get the values of all available registers at this stack frame as a dictionary with the register names as keys.


Types

A Type object describes a type in a program. Each kind of type (e.g., integer, structure) has different attributes (e.g., name, size). Types can also have qualifiers (e.g., constant, atomic). Accessing an attribute which does not apply to a type raises an AttributeError.

repr() of a Type returns a Python representation of the type:

>>> print(repr(prog.type('sector_t')))
prog.typedef_type(name='sector_t', type=prog.int_type(name='unsigned long', size=8, is_signed=False))

str() returns a representation of the type in programming language syntax:

>>> print(prog.type('sector_t'))
typedef unsigned long sector_t

The drgn CLI is set up so that types are displayed with str() instead of repr() by default.

This class cannot be constructed directly. Instead, use one of the Type Constructors.

Program that this type is from.

Kind of this type.

If this is a primitive type (e.g., int or double), the kind of primitive type. Otherwise, None.

Bitmask of this type's qualifier.

Programming language of this type.

Name of this type. This is present for integer, boolean, floating-point, and typedef types.

Tag of this type, or None if this is an anonymous type. This is present for structure, union, class, and enumerated types.

Size of this type in bytes, or None if this is an incomplete type. This is present for integer, boolean, floating-point, structure, union, class, and pointer types.

Number of elements in this type, or None if this is an incomplete type. This is only present for array types.

Whether this type is signed. This is only present for integer types.

Byte order of this type: 'little' if it is little-endian, or 'big' if it is big-endian. This is present for integer, boolean, floating-point, and pointer types.

Type underlying this type, defined as follows:
  • For typedef types, the aliased type.
  • For enumerated types, the compatible integer type, which is None if this is an incomplete type.
  • For pointer types, the referenced type.
  • For array types, the element type.
  • For function types, the return type.

For other types, this attribute is not present.


List of members of this type, or None if this is an incomplete type. This is present for structure, union, and class types.

List of enumeration constants of this type, or None if this is an incomplete type. This is only present for enumerated types.

List of parameters of this type. This is only present for function types.

Whether this type takes a variable number of arguments. This is only present for function types.

List of template parameters of this type. This is present for structure, union, class, and function types.

Get a descriptive full name of this type.

Get whether this type is complete (i.e., the type definition is known). This is always False for void types. It may be False for structure, union, class, enumerated, and array types, as well as typedef types where the underlying type is one of those. Otherwise, it is always True.

Get a copy of this type with different qualifiers.

Note that the original qualifiers are replaced, not added to.

qualifiers -- New type qualifiers.


Get a copy of this type with no qualifiers.

Look up a member in this type by name.

If this type has any unnamed members, this also matches members of those unnamed members, recursively. If the member is found in an unnamed member, TypeMember.bit_offset and TypeMember.offset are adjusted accordingly.

name -- Name of the member.
  • TypeError -- if this type is not a structure, union, or class type
  • LookupError -- if this type does not have a member with the given name



Return whether this type has a member with the given name.

If this type has any unnamed members, this also matches members of those unnamed members, recursively.

name -- Name of the member.
TypeError -- if this type is not a structure, union, or class type



A TypeMember represents a member of a structure, union, or class type.
Create a TypeMember.
•
object_or_type --

One of:

1.
TypeMember.object as an Object.
2.
TypeMember.type as a Type. In this case, object is set to an absent object with that type.
3.
A callable that takes no arguments and returns one of the above. It is called when object or type is first accessed, and the result is cached.

  • name -- TypeMember.name
  • bit_offset -- TypeMember.bit_offset



Member as an Object.

This is the default initializer for the member, or an absent object if the member has no default initializer. (However, the DWARF specification as of version 5 does not actually support default member initializers, so this is usually absent.)


Member type.

This is a shortcut for TypeMember.object.type.


Member name, or None if the member is unnamed.

Offset of the member from the beginning of the type in bits.

Offset of the member from the beginning of the type in bytes. If the offset is not byte-aligned, accessing this attribute raises ValueError.

Size in bits of this member if it is a bit field, None if it is not.

This is a shortcut for TypeMember.object.bit_field_size_.



A TypeEnumerator represents a constant in an enumerated type.

Its name and value may be accessed as attributes or unpacked:

>>> prog.type('enum pid_type').enumerators[0].name
'PIDTYPE_PID'
>>> name, value = prog.type('enum pid_type').enumerators[0]
>>> value
0
Create a TypeEnumerator.
  • name -- TypeEnumerator.name
  • value -- TypeEnumerator.value



Enumerator name.

Enumerator value.


A TypeParameter represents a parameter of a function type.
Create a TypeParameter.
•
default_argument_or_type --

One of:

1.
TypeParameter.default_argument as an Object.
2.
TypeParameter.type as a Type. In this case, default_argument is set to an absent object with that type.
3.
A callable that takes no arguments and returns one of the above. It is called when default_argument or type is first accessed, and the result is cached.

•
name -- TypeParameter.name



Default argument for parameter.

If the parameter does not have a default argument, then this is an absent object.

NOTE:

Neither GCC nor Clang emits debugging information for default arguments (as of GCC 10 and Clang 11), and drgn does not yet parse it, so this is usually absent.



Parameter type.

This is the same as TypeParameter.default_argument.type_.


Parameter name, or None if the parameter is unnamed.


A TypeTemplateParameter represents a template parameter of a structure, union, class, or function type.
Create a TypeTemplateParameter.
•
argument --

One of:

1.
TypeTemplateParameter.argument as a Type if the parameter is a type template parameter.
2.
TypeTemplateParameter.argument as a non-absent Object if the parameter is a non-type template parameter.
3.
A callable that takes no arguments and returns one of the above. It is called when argument is first accessed, and the result is cached.

  • name -- TypeTemplateParameter.name
  • is_default -- TypeTemplateParameter.is_default



Template argument.

If this is a type template parameter, then this is a Type. If this is a non-type template parameter, then this is an Object.


Template parameter name, or None if the parameter is unnamed.

Whether argument is the default for the template parameter.

NOTE:

There are two ways to interpret this:
1.
The argument was omitted entirely and thus defaulted to the default argument.
2.
The (specified or defaulted) argument is the same as the default argument.



Compilers are inconsistent about which interpretation they use.

GCC added this information in version 4.9. Clang added it in version 11 (and only when emitting DWARF version 5). If the program was compiled by an older version, this is always false.





Bases: enum.Enum

A TypeKind represents a kind of type.

Void type.

Integer type.

Boolean type.

Floating-point type.

Complex type.

Structure type.

Union type.

Class type.

Enumerated type.

Type definition (a.k.a. alias) type.

Pointer type.

Array type.

Function type.


Bases: enum.Enum

A PrimitiveType represents a primitive type known to drgn.




















Bases: enum.Flag

Qualifiers are modifiers on types.

No qualifiers.

Constant type.

Volatile type.

Restrict type.

Atomic type.


Get the offset (in bytes) of a member in a Type.

This corresponds to offsetof() in C.

  • type -- Structure, union, or class type.
  • member -- Name of member. May include one or more member references and zero or more array subscripts.

  • TypeError -- if type is not a structure, union, or class type
  • ValueError -- if the member is not byte-aligned (e.g., because it is a bit field)
  • LookupError -- if type does not have a member with the given name



Type Constructors

Custom drgn types can be created with the following factory functions. These can be used just like types obtained from Program.type().

Create a new void type. It has kind TypeKind.VOID.
  • qualifiers -- Type.qualifiers
  • lang -- Type.language



Create a new integer type. It has kind TypeKind.INT.
  • name -- Type.name
  • size -- Type.size
  • is_signed -- Type.is_signed
  • byteorder -- Type.byteorder, or None to use the program's default byte order.
  • qualifiers -- Type.qualifiers
  • lang -- Type.language



Create a new boolean type. It has kind TypeKind.BOOL.
  • name -- Type.name
  • size -- Type.size
  • byteorder -- Type.byteorder, or None to use the program's default byte order.
  • qualifiers -- Type.qualifiers
  • lang -- Type.language



Create a new floating-point type. It has kind TypeKind.FLOAT.
  • name -- Type.name
  • size -- Type.size
  • byteorder -- Type.byteorder, or None to use the program's default byte order.
  • qualifiers -- Type.qualifiers
  • lang -- Type.language



Create a new structure type. It has kind TypeKind.STRUCT.
  • tag -- Type.tag
  • size -- Type.size
  • members -- Type.members
  • template_parameters -- Type.template_parameters
  • qualifiers -- Type.qualifiers
  • lang -- Type.language



Create a new incomplete structure type.

Create a new union type. It has kind TypeKind.UNION. Otherwise, this is the same as as struct_type().

Create a new incomplete union type.

Create a new class type. It has kind TypeKind.CLASS. Otherwise, this is the same as as struct_type().

Create a new incomplete class type.

Create a new enumerated type. It has kind TypeKind.ENUM.
  • tag -- Type.tag
  • type -- The compatible integer type (Type.type)
  • enumerators -- Type.enumerators
  • qualifiers -- Type.qualifiers
  • lang -- Type.language



Create a new incomplete enumerated type.

Create a new typedef type. It has kind TypeKind.TYPEDEF.
  • name -- Type.name
  • type -- The aliased type (Type.type)
  • qualifiers -- Type.qualifiers
  • lang -- Type.language



Create a new pointer type. It has kind TypeKind.POINTER,
  • type -- The referenced type (Type.type)
  • size -- Type.size, or None to use the program's default pointer size.
  • byteorder -- Type.byteorder, or None to use the program's default byte order.
  • qualifiers -- Type.qualifiers
  • lang -- Type.language



Create a new array type. It has kind TypeKind.ARRAY.
  • type -- The element type (Type.type)
  • length -- Type.length
  • qualifiers -- Type.qualifiers
  • lang -- Type.language



Create a new function type. It has kind TypeKind.FUNCTION.
  • type -- The return type (Type.type)
  • parameters -- Type.parameters
  • is_variadic -- Type.is_variadic
  • template_parameters -- Type.template_parameters
  • qualifiers -- Type.qualifiers
  • lang -- Type.language



Miscellaneous

Get the size of a Type or Object in bytes.
type_or_obj -- Entity to get the size of.
TypeError -- if the type does not have a size (e.g., because it is incomplete or void)


Execute a script.

The script is executed in the same context as the caller: currently defined globals are available to the script, and globals defined by the script are added back to the calling context.

This is most useful for executing scripts from interactive mode. For example, you could have a script named exe.py:

"""Get all tasks executing a given file."""
import sys
from drgn.helpers.linux.fs import d_path
from drgn.helpers.linux.pid import find_task
def task_exe_path(task):


    if task.mm:


        return d_path(task.mm.exe_file.f_path).decode()


    else:


        return None
tasks = [


    task for task in for_each_task()


    if task_exe_path(task) == sys.argv[1]
]


Then, you could execute it and use the defined variables and functions:

>>> execscript('exe.py', '/usr/bin/bash')
>>> tasks[0].pid
(pid_t)358442
>>> task_exe_path(find_task(357954))
'/usr/bin/vim'
  • path -- File path of the script.
  • args -- Zero or more additional arguments to pass to the script. This is a variable argument list.



Bases: Protocol

An int or integer-like object.

Parameters annotated with this type expect an integer which may be given as a Python int or an Object with integer type.


Filesystem path.

Parameters annotated with this type accept a filesystem path as str, bytes, or os.PathLike.


Exceptions

Bases: Exception

This error is raised when a bad memory access is attempted (i.e., when accessing a memory address which is not valid in a program).

  • message -- FaultError.message
  • address -- FaultError.address



Error message.

Address that couldn't be accessed.


Bases: Exception

This error is raised when one or more files in a program do not have debug information.


Bases: Exception

This error is raised when attempting to use an absent object.


Bases: Exception

This error is raised when attempting to access beyond the bounds of a value object.


CLI

Functions for embedding the drgn CLI.

Return the version header printed at the beginning of a drgn session.

The run_interactive() function does not include this banner at the beginning of an interactive session. Use this function to retrieve one line of text to add to the beginning of the drgn banner, or print it before calling run_interactive().


Run drgn's Interactive Mode until the user exits.

This function allows your application to embed the same REPL that drgn provides when it is run on the command line in interactive mode.

  • prog -- Pre-configured program to run against. Available as a global named prog in the CLI.
  • banner_func -- Optional function to modify the printed banner. Called with the default banner, and must return a string to use as the new banner. The default banner does not include the drgn version, which can be retrieved via version_header().
  • globals_func -- Optional function to modify globals provided to the session. Called with a dictionary of default globals, and must return a dictionary to use instead.
  • quiet -- Ignored. Will be removed in the future.


NOTE:

This function uses readline and modifies some settings. Unfortunately, it is not possible for it to restore all settings. In particular, it clears the readline history and resets the TAB keybinding to the default.

Applications using readline should save their history and clear any custom settings before calling this function. After calling this function, applications should restore their history and settings before using readline.




Logging

drgn logs using the standard logging module to a logger named "drgn".

Thread Safety

Only one thread at a time should access the same Program (including Object, Type, StackTrace, etc. from that program). It is safe to use different Programs from concurrent threads.

Helpers

The drgn.helpers package contains subpackages which provide helpers for working with particular types of programs. Currently, there are common helpers and helpers for the Linux kernel. In the future, there may be helpers for, e.g., glibc and libstdc++.

Bases: Exception

Error raised by a validator when an inconsistent or invalid state is detected.


Common

The drgn.helpers.common package provides helpers that can be used with any program. The helpers are available from the individual modules in which they are defined and from this top-level package. E.g., the following are both valid:

>>> from drgn.helpers.common.memory import identify_address
>>> from drgn.helpers.common import identify_address

Some of these helpers may have additional program-specific behavior but are otherwise generic.

Formatting

The drgn.helpers.common.format module provides generic helpers for formatting different things as text.

Format an ASCII byte value as a character, possibly escaping it. Non-printable characters are always escaped. Non-printable characters other than \0, \a, \b, \t, \n, \v, \f, and \r are escaped in hexadecimal format (e.g., \x7f). By default, printable characters are never escaped.
  • c -- Character to escape.
  • escape_single_quote -- Whether to escape single quotes to \'.
  • escape_double_quote -- Whether to escape double quotes to \".
  • escape_backslash -- Whether to escape backslashes to \\.



Escape an iterable of ASCII byte values (e.g., bytes or bytearray). See escape_ascii_character().
buffer -- Byte array to escape.


Get a human-readable representation of a bitmask of flags.

By default, flags are specified by their bit number:

>>> decode_flags(2, [("BOLD", 0), ("ITALIC", 1), ("UNDERLINE", 2)])
'ITALIC'

They can also be specified by their value:

>>> decode_flags(2, [("BOLD", 1), ("ITALIC", 2), ("UNDERLINE", 4)],
...              bit_numbers=False)
'ITALIC'

Multiple flags are combined with "|":

>>> decode_flags(5, [("BOLD", 0), ("ITALIC", 1), ("UNDERLINE", 2)])
'BOLD|UNDERLINE'

If there are multiple names for the same bit, they are all included:

>>> decode_flags(2, [("SMALL", 0), ("BIG", 1), ("LARGE", 1)])
'BIG|LARGE'

If there are any unknown bits, their raw value is included:

>>> decode_flags(27, [("BOLD", 0), ("ITALIC", 1), ("UNDERLINE", 2)])
'BOLD|ITALIC|0x18'

Zero is returned verbatim:

>>> decode_flags(0, [("BOLD", 0), ("ITALIC", 1), ("UNDERLINE", 2)])
'0'
  • value -- Bitmask to decode.
  • flags -- List of flag names and their bit numbers or values.
  • bit_numbers -- Whether flags specifies the bit numbers (where 0 is the least significant bit) or values of the flags.



Get a human-readable representation of a bitmask of flags where the flags are specified by an enumerated drgn.Type.

This supports enums where the values are bit numbers:

>>> print(bits_enum)
enum style_bits {


        BOLD = 0,


        ITALIC = 1,


        UNDERLINE = 2,
}
>>> decode_enum_type_flags(5, bits_enum)
'BOLD|UNDERLINE'

Or the values of the flags:

>>> print(flags_enum)
enum style_flags {


        BOLD = 1,


        ITALIC = 2,


        UNDERLINE = 4,
}
>>> decode_enum_type_flags(5, flags_enum, bit_numbers=False)
'BOLD|UNDERLINE'

See decode_flags().

  • value -- Bitmask to decode.
  • type -- Enumerated type with bit numbers for enumerators.
  • bit_numbers -- Whether the enumerator values specify the bit numbers or values of the flags.



Format a number in binary units (i.e., "K" is 1024, "M" is 10242, etc.). 

>>> number_in_binary_units(1280)
'1.2K'

A precision can be specified:

>>> number_in_binary_units(1280, precision=2)
'1.25K'

Exact numbers are printed without a fractional part:

>>> number_in_binary_units(1024 * 1024)
'1M'

Numbers less than 1024 are not scaled:

>>> number_in_binary_units(10)
'10'
  • n -- Number to format.
  • precision -- Number of digits to include in fractional part.



Memory

The drgn.helpers.common.memory module provides helpers for working with memory and addresses.

Try to identify what an address refers to.

For all programs, this will identify addresses as follows:

  • Object symbols (e.g., addresses in global variables): object symbol: {symbol_name}+{hex_offset} (where hex_offset is the offset from the beginning of the symbol in hexadecimal).
  • Function symbols (i.e., addresses in functions): function symbol: {symbol_name}+{hex_offset}.
  • Other symbols: symbol: {symbol_name}+{hex_offset}.

Additionally, for the Linux kernel, this will identify:

  • Allocated slab objects: slab object: {slab_cache_name}+{hex_offset} (where hex_offset is the offset from the beginning of the object in hexadecimal).
  • Free slab objects: free slab object: {slab_cache_name}+{hex_offset}.

This may recognize other types of addresses in the future.

  • prog -- Program, which may be omitted to use the default program argument.
  • addr -- void *

Identity as string, or None if the address is unrecognized.


Program Decorators

The drgn.helpers.common.prog module provides decorators to transparently use the default program argument.

Wrap a function taking a Program so that it uses the default program argument if omitted.

@takes_program_or_default
def my_helper(prog: Program, n: IntegerLike) -> Foo:


    ...
my_helper(1)
# is equivalent to
my_helper(get_default_prog(), 1)
obj = Object(...)
my_helper(obj)
# is equivalent to
my_helper(obj.prog_, obj)



Wrap a function taking a Program and an optional Object so that it accepts a Program or an Object or neither, in which case the default program argument is used.

@takes_object_or_program_or_default
def my_helper(prog: Program, obj: Optional[Object], n: IntegerLike) -> Foo:


    ...
my_helper(prog, 1)
# is equivalent to
my_helper.__wrapped__(prog, None, 1)
obj = Object(...)
my_helper(obj, 1)
# is equivalent to
my_helper.__wrapped__(obj.prog_, obj, 1)
my_helper(1)
# is equivalent to
my_helper.__wrapped__(get_default_prog(), None, 1)
one_obj = Object(..., 1)
my_helper(one_obj)
# is equivalent to
my_helper.__wrapped__(one_obj.prog_, None, one_obj)


WARNING:

This cannot be used with positional parameters with a default value, as that would create ambiguity. Keyword-only parameters with a default value are OK. 

# NOT ALLOWED
@takes_object_or_program_or_default
def my_helper(prog: Program, obj: Optional[Object], foo: str = ""): ...
# OK
@takes_object_or_program_or_default
def my_helper(prog: Program, obj: Optional[Object], *, foo: str = ""): ...




NOTE:

The object parameter can be passed as a keyword, but because of limitations of the Python type system, type checkers do not recognize this.



Stack

The drgn.helpers.common.stack module provides helpers for working with stack traces.

Print the contents of stack memory in a stack trace, annotating values that can be identified.

Currently, this will identify any addresses on the stack with identify_address().

>>> print_annotated_stack(stack_trace(1))
STACK POINTER     VALUE
[stack frame #0 at 0xffffffff8dc93c41 (__schedule+0x429/0x488) in context_switch at ./kernel/sched/core.c:5209:2 (inlined)]
[stack frame #1 at 0xffffffff8dc93c41 (__schedule+0x429/0x488) in __schedule at ./kernel/sched/core.c:6521:8]
ffffa903c0013d28: ffffffff8d8497bf [function symbol: __flush_tlb_one_user+0x5]
ffffa903c0013d30: 000000008d849eb5
ffffa903c0013d38: 0000000000000001
ffffa903c0013d40: 0000000000000004
ffffa903c0013d48: efdea37bb7cb1f00
ffffa903c0013d50: ffff926641178000 [slab object: task_struct+0x0]
ffffa903c0013d58: ffff926641178000 [slab object: task_struct+0x0]
ffffa903c0013d60: ffffa903c0013e10
ffffa903c0013d68: ffff926641177ff0 [slab object: mm_struct+0x70]
ffffa903c0013d70: ffff926641178000 [slab object: task_struct+0x0]
ffffa903c0013d78: ffff926641178000 [slab object: task_struct+0x0]
ffffa903c0013d80: ffffffff8dc93d29 [function symbol: schedule+0x89]
...
trace -- Stack trace to print.


Types

The drgn.helpers.common.type module provides generic helpers for working with types in ways that aren't provided by the core drgn library.

Get an enum.IntEnum class from an enumerated drgn.Type.
  • type -- Enumerated type to convert.
  • name -- Name of the IntEnum type to create.
  • exclude -- Container (e.g., list or set) of enumerator names to exclude from the created IntEnum.
  • prefix -- Prefix to strip from the beginning of enumerator names.



Linux Kernel

The drgn.helpers.linux package contains several modules for working with data structures and subsystems in the Linux kernel. The helpers are available from the individual modules in which they are defined and from this top-level package. E.g., the following are both valid:

>>> from drgn.helpers.linux.list import list_for_each_entry
>>> from drgn.helpers.linux import list_for_each_entry

Iterator macros (for_each_foo) are a common idiom in the Linux kernel. The equivalent drgn helpers are implemented as Python generators. For example, the following code in C:

list_for_each(pos, head)


        do_something_with(pos);


Translates to the following code in Python:

for pos in list_for_each(head):


    do_something_with(pos)


Bit Operations

The drgn.helpers.linux.bitops module provides helpers for common bit operations in the Linux kernel.

Iterate over all set (one) bits in a bitmap.
  • bitmap -- unsigned long *
  • size -- Size of bitmap in bits.



Iterate over all clear (zero) bits in a bitmap.
  • bitmap -- unsigned long *
  • size -- Size of bitmap in bits.



Return whether a bit in a bitmap is set.
  • nr -- Bit number.
  • bitmap -- unsigned long *



Block Layer

The drgn.helpers.linux.block module provides helpers for working with the Linux block layer, including disks (struct gendisk) and partitions.

Since Linux v5.11, partitions are represented by struct block_device. Before that, they were represented by struct hd_struct.

Get a disk's device number.
disk -- struct gendisk *
dev_t


Get the name of a disk (e.g., sda).
disk -- struct gendisk *


Iterate over all disks in the system.
prog -- Program, which may be omitted to use the default program argument.
Iterator of struct gendisk * objects.


Print all of the disks in the system.
prog -- Program, which may be omitted to use the default program argument.


Get a partition's device number.
part -- struct block_device * or struct hd_struct * depending on the kernel version.
dev_t


Get the name of a partition (e.g., sda1).
part -- struct block_device * or struct hd_struct * depending on the kernel version.


Iterate over all partitions in the system.
prog -- Program, which may be omitted to use the default program argument.
Iterator of struct block_device * or struct hd_struct * objects depending on the kernel version.


Print all of the partitions in the system.
prog -- Program, which may be omitted to use the default program argument.


Boot

The drgn.helpers.linux.boot module provides helpers for inspecting the Linux kernel boot configuration.

Get the kernel address space layout randomization offset (zero if it is disabled).
prog -- Program, which may be omitted to use the default program argument.


Return whether 5-level paging is enabled.
prog -- Program, which may be omitted to use the default program argument.


BPF

The drgn.helpers.linux.bpf module provides helpers for working with BPF interface in include/linux/bpf.h, include/linux/bpf-cgroup.h, etc.

Iterate over all BTF objects.

This is only supported since Linux v4.18.

prog -- Program, which may be omitted to use the default program argument.
Iterator of struct btf * objects.


Iterate over all BPF links.

This is only supported since Linux v5.8.

prog -- Program, which may be omitted to use the default program argument.
Iterator of struct bpf_link * objects.


Iterate over all BPF maps.

This is only supported since Linux v4.13.

prog -- Program, which may be omitted to use the default program argument.
Iterator of struct bpf_map * objects.


Iterate over all BPF programs.

This is only supported since Linux v4.13.

prog -- Program, which may be omitted to use the default program argument.
Iterator of struct bpf_prog * objects.


Iterate over all cgroup BPF programs of the given attach type attached to the given cgroup.
  • cgrp -- struct cgroup *
  • bpf_attach_type -- enum bpf_attach_type

Iterator of struct bpf_prog * objects.


Iterate over all effective cgroup BPF programs of the given attach type for the given cgroup.
  • cgrp -- struct cgroup *
  • bpf_attach_type -- enum bpf_attach_type

Iterator of struct bpf_prog * objects.


Cgroup

The drgn.helpers.linux.cgroup module provides helpers for working with the cgroup interface in include/linux/cgroup.h. Only cgroup v2 is supported.

Get the cgroup for a socket from the given struct sock_cgroup_data * (usually from struct sock::sk_cgrp_data).
skcd -- struct sock_cgroup_data *
struct cgroup *


Return the parent cgroup of the given cgroup if it exists, NULL otherwise.
cgrp -- struct cgroup *
struct cgroup *


Get the name of the given cgroup.
cgrp -- struct cgroup *


Get the full path of the given cgroup.
cgrp -- struct cgroup *


Look up a cgroup from its default hierarchy path .
  • prog -- Program, which may be omitted to use the default program argument.
  • path -- Path name.



Get the next child (or NULL if there is none) of the given parent starting from the given position (NULL to initiate traversal).
  • pos -- struct cgroup_subsys_state *
  • parent -- struct cgroup_subsys_state *

struct cgroup_subsys_state *


Get the next pre-order descendant (or NULL if there is none) of the given css root starting from the given position (NULL to initiate traversal).
  • pos -- struct cgroup_subsys_state *
  • root -- struct cgroup_subsys_state *

struct cgroup_subsys_state *


Iterate through children of the given css.
css -- struct cgroup_subsys_state *
Iterator of struct cgroup_subsys_state * objects.


Iterate through the given css's descendants in pre-order.
css -- struct cgroup_subsys_state *
Iterator of struct cgroup_subsys_state * objects.


CPU Masks

The drgn.helpers.linux.cpumask module provides helpers for working with CPU masks from include/linux/cpumask.h.

Return the mask of online CPUs.
prog -- Program, which may be omitted to use the default program argument.
struct cpumask *


Return the mask of possible CPUs.
prog -- Program, which may be omitted to use the default program argument.
struct cpumask *


Return the mask of present CPUs.
prog -- Program, which may be omitted to use the default program argument.
struct cpumask *


Iterate over all of the CPUs in the given mask.
mask -- struct cpumask *


Iterate over all online CPUs.
prog -- Program, which may be omitted to use the default program argument.


Iterate over all possible CPUs.
prog -- Program, which may be omitted to use the default program argument.


Iterate over all present CPUs.
prog -- Program, which may be omitted to use the default program argument.


Return a CPU mask as a CPU list string. 

>>> cpumask_to_cpulist(mask)
0-3,8-11
mask -- struct cpumask *
String in the CPU list format.


Devices

The drgn.helpers.linux.device module provides helpers for working with Linux devices, including the kernel encoding of dev_t.

Return the major ID of a kernel dev_t.
dev -- dev_t object or int.


Return the minor ID of a kernel dev_t.
dev -- dev_t object or int.


Return a kernel dev_t from the major and minor IDs.
  • major -- Device major ID.
  • minor -- Device minor ID.



Virtual Filesystem Layer

The drgn.helpers.linux.fs module provides helpers for working with the Linux virtual filesystem (VFS) layer, including mounts, dentries, and inodes.

Look up the given path name.
  • root -- struct path * to use as the root directory. Defaults to the initial root filesystem if given a Program or omitted.
  • path -- Path to lookup.
  • allow_negative -- Whether to allow returning a negative dentry (i.e., a dentry for a non-existent path).

struct path
Exception -- if the dentry is negative and allow_negative is False, or if the path is not present in the dcache. The latter does not necessarily mean that the path does not exist; it may be uncached. On a live system, you can make the kernel cache the path by accessing it (e.g., with open() or os.stat()):

>>> path_lookup('/usr/include/stdlib.h')
...
Exception: could not find '/usr/include/stdlib.h' in dcache
>>> open('/usr/include/stdlib.h').close()
>>> path_lookup('/usr/include/stdlib.h')
(struct path){


        .mnt = (struct vfsmount *)0xffff8b70413cdca0,


        .dentry = (struct dentry *)0xffff8b702ac2c480,
}

Return the full path of a dentry given a struct path.
path -- struct path or struct path *


Return the full path of a dentry given a mount and dentry.
  • vfsmnt -- struct vfsmount *
  • dentry -- struct dentry *



Return the path of a dentry from the root of its filesystem.
dentry -- struct dentry *


Return any path of an inode from the root of its filesystem.
inode -- struct inode *
Path, or None if the inode has no aliases.


Return an iterator over all of the paths of an inode from the root of its filesystem.
inode -- struct inode *


Get the source device name for a mount.
mnt -- struct mount *


Get the path of a mount point.
mnt -- struct mount *


Get the filesystem type of a mount.
mnt -- struct mount *


Iterate over all of the mounts in a given namespace.
  • ns -- struct mnt_namespace *. Defaults to the initial mount namespace if given a Program or omitted.
  • src -- Only include mounts with this source device name.
  • dst -- Only include mounts with this destination path.
  • fstype -- Only include mounts with this filesystem type.

Iterator of struct mount * objects.


Print the mount table of a given namespace. The arguments are the same as for_each_mount(). The output format is similar to /proc/mounts but prints the value of each struct mount *.

Return the kernel file descriptor of the fd of a given task.
  • task -- struct task_struct *
  • fd -- File descriptor.

struct file *


Iterate over all of the files open in a given task.
task -- struct task_struct *
Iterator of (fd, struct file *) tuples.


Print the open files of a given task.
task -- struct task_struct *


IDR

The drgn.helpers.linux.idr module provides helpers for working with the IDR data structure in include/linux/idr.h. An IDR provides a mapping from an ID to a pointer.

Look up the entry with the given ID in an IDR.
  • idr -- struct idr *
  • id -- Entry ID.

void * found entry, or NULL if not found.


Iterate over all of the pointers in an IDR.
idr -- struct idr *
Iterator of (index, void *) tuples.


Iterate over all of the entries with the given type in an IDR.
  • idr -- struct idr *
  • type -- Entry type.

Iterator of (index, type *) tuples.


Kconfig

The drgn.helpers.linux.kconfig module provides helpers for reading the Linux kernel build configuration.

Get the kernel build configuration as a mapping from the option name to the value. 

>>> get_kconfig()['CONFIG_SMP']
'y'
>>> get_kconfig()['CONFIG_HZ']
'300'

This is only supported if the kernel was compiled with CONFIG_IKCONFIG. Note that most Linux distributions do not enable this option.

prog -- Program, which may be omitted to use the default program argument.


Kernfs

The drgn.helpers.linux.kernfs module provides helpers for working with the kernfs pseudo filesystem interface in include/linux/kernfs.h.

Get the name of the given kernfs node.
kn -- struct kernfs_node *


Get full path of the given kernfs node.
kn -- struct kernfs_node *


Find the kernfs node with the given path from the given parent kernfs node.
  • parent -- struct kernfs_node *
  • path -- Path name.

struct kernfs_node * (NULL if not found)


Linked Lists

The drgn.helpers.linux.list module provides helpers for working with the doubly-linked list implementations (struct list_head and struct hlist_head) in include/linux/list.h.

Return whether a list is empty.
head -- struct list_head *


Return whether a list has only one element.
head -- struct list_head *


Return the number of nodes in a list.
head -- struct list_head *


Return the first entry in a list.

The list is assumed to be non-empty.

See also list_first_entry_or_null().

  • head -- struct list_head *
  • type -- Entry type.
  • member -- Name of list node member in entry type.

type *


Return the first entry in a list or NULL if the list is empty.

See also list_first_entry().

  • head -- struct list_head *
  • type -- Entry type.
  • member -- Name of list node member in entry type.

type *


Return the last entry in a list.

The list is assumed to be non-empty.

  • head -- struct list_head *
  • type -- Entry type.
  • member -- Name of list node member in entry type.

type *


Return the next entry in a list.
  • pos -- type*
  • member -- Name of list node member in entry type.

type *


Return the previous entry in a list.
  • pos -- type*
  • member -- Name of list node member in entry type.

type *


Iterate over all of the nodes in a list.
head -- struct list_head *
Iterator of struct list_head * objects.


Iterate over all of the nodes in a list in reverse order.
head -- struct list_head *
Iterator of struct list_head * objects.


Iterate over all of the entries in a list.
  • type -- Entry type.
  • head -- struct list_head *
  • member -- Name of list node member in entry type.

Iterator of type * objects.


Iterate over all of the entries in a list in reverse order.
  • type -- Entry type.
  • head -- struct list_head *
  • member -- Name of list node member in entry type.

Iterator of type * objects.


Validate that the next and prev pointers in a list are consistent.
head -- struct list_head *
ValidationError -- if the list is invalid


Like list_for_each(), but validates the list like validate_list() while iterating.
head -- struct list_head *
ValidationError -- if the list is invalid


Like list_for_each_entry(), but validates the list like validate_list() while iterating.
  • type -- Entry type.
  • head -- struct list_head *
  • member -- Name of list node member in entry type.

ValidationError -- if the list is invalid


Return whether a hash list is empty.
head -- struct hlist_head *


Iterate over all of the nodes in a hash list.
head -- struct hlist_head *
Iterator of struct hlist_node * objects.


Iterate over all of the entries in a hash list.
  • type -- Entry type.
  • head -- struct hlist_head *
  • member -- Name of list node member in entry type.

Iterator of type * objects.


Nulls Lists

The drgn.helpers.linux.list_nulls module provides helpers for working with the special version of lists (struct hlist_nulls_head and struct hlist_nulls_node) in include/linux/list_nulls.h where the end of list is not a NULL pointer, but a "nulls" marker.

Return whether a a pointer is a nulls marker.
pos -- struct hlist_nulls_node *


Return whether a nulls hash list is empty.
head -- struct hlist_nulls_head *


Iterate over all the entries in a nulls hash list.
  • type -- Entry type.
  • head -- struct hlist_nulls_head *
  • member -- Name of list node member in entry type.

Iterator of type * objects.


Lockless Lists

The drgn.helpers.linux.llist module provides helpers for working with the lockless, NULL-terminated, singly-linked list implementation in include/linux/llist.h (struct llist_head and struct llist_node).

Return whether an llist is empty.
head -- struct llist_head *


Return whether an llist has only one element.
head -- struct llist_head *


Return the first entry in an llist.

The list is assumed to be non-empty.

See also llist_first_entry_or_null().

  • head -- struct llist_head *
  • type -- Entry type.
  • member -- Name of struct llist_node member in entry type.

type *


Return the first entry in an llist or NULL if the llist is empty.

See also llist_first_entry().

  • head -- struct llist_head *
  • type -- Entry type.
  • member -- Name of struct llist_node member in entry type.

type *


Return the next entry in an llist.
  • pos -- type*
  • member -- Name of struct llist_node member in entry type.

type *


Iterate over all of the nodes in an llist starting from a given node.
node -- struct llist_node *
Iterator of struct llist_node * objects.


Iterate over all of the entries in an llist starting from a given node.
  • type -- Entry type.
  • node -- struct llist_node *
  • member -- Name of struct llist_node member in entry type.

Iterator of type * objects.


Maple Trees

The drgn.helpers.linux.mapletree module provides helpers for working with maple trees from include/linux/maple_tree.h.

Maple trees were introduced in Linux 6.1.

Look up the entry at a given index in a maple tree. 

>>> entry = mtree_load(task.mm.mm_mt.address_of_(), 0x55d65cfaa000)
>>> cast("struct vm_area_struct *", entry)
*(struct vm_area_struct *)0xffff97ad82bfc930 = {


    ...
}
  • mt -- struct maple_tree *
  • index -- Entry index.
  • advanced -- Whether to return nodes only visible to the maple tree advanced API. If False, zero entries (see xa_is_zero()) will be returned as NULL.

void * found entry, or NULL if not found.


Iterate over all of the entries and their ranges in a maple tree. 

>>> for first_index, last_index, entry in mt_for_each(task.mm.mm_mt.address_of_()):
...     print(hex(first_index), hex(last_index), entry)
...
0x55d65cfaa000 0x55d65cfaafff (void *)0xffff97ad82bfc930
0x55d65cfab000 0x55d65cfabfff (void *)0xffff97ad82bfc0a8
0x55d65cfac000 0x55d65cfacfff (void *)0xffff97ad82bfc000
0x55d65cfad000 0x55d65cfadfff (void *)0xffff97ad82bfcb28
...
  • mt -- struct maple_tree *
  • advanced -- Whether to return nodes only visible to the maple tree advanced API. If False, zero entries (see xa_is_zero()) will be skipped.

Iterator of (first_index, last_index, void *) tuples. Both indices are inclusive.


Memory Management

The drgn.helpers.linux.mm module provides helpers for working with the Linux memory management (MM) subsystem. Only AArch64, ppc64, s390x, and x86-64 are currently supported.

Return whether the PG_active flag is set on a page.
page -- struct page *


Return whether the PG_checked flag is set on a page.
page -- struct page *


Return whether the PG_dirty flag is set on a page.
page -- struct page *


Return whether the PG_double_map flag is set on a page.
page -- struct page *


Return whether the PG_error flag is set on a page.
page -- struct page *


Return whether the PG_foreign flag is set on a page.
page -- struct page *


Return whether the PG_hwpoison flag is set on a page.
page -- struct page *


Return whether the PG_has_hwpoisoned flag is set on a page.
page -- struct page *


Return whether the PG_idle flag is set on a page.
page -- struct page *


Return whether the PG_isolated flag is set on a page.
page -- struct page *


Return whether the PG_lru flag is set on a page.
page -- struct page *


Return whether the PG_locked flag is set on a page.
page -- struct page *


Return whether the PG_mappedtodisk flag is set on a page.
page -- struct page *


Return whether the PG_mlocked flag is set on a page.
page -- struct page *


Return whether the PG_owner_priv_1 flag is set on a page.
page -- struct page *


Return whether the PG_pinned flag is set on a page.
page -- struct page *


Return whether the PG_private flag is set on a page.
page -- struct page *


Return whether the PG_private_2 flag is set on a page.
page -- struct page *


Return whether the PG_readahead flag is set on a page.
page -- struct page *


Return whether the PG_reclaim flag is set on a page.
page -- struct page *


Return whether the PG_referenced flag is set on a page.
page -- struct page *


Return whether the PG_reported flag is set on a page.
page -- struct page *


Return whether the PG_reserved flag is set on a page.
page -- struct page *


Return whether the PG_savepinned flag is set on a page.
page -- struct page *


Return whether the PG_skip_kasan_poison flag is set on a page.
page -- struct page *


Return whether the PG_slab flag is set on a page.
page -- struct page *


Return whether the PG_slob_free flag is set on a page.
page -- struct page *


Return whether the PG_swapbacked flag is set on a page.
page -- struct page *


Return whether the PG_uncached flag is set on a page.
page -- struct page *


Return whether the PG_unevictable flag is set on a page.
page -- struct page *


Return whether the PG_uptodate flag is set on a page.
page -- struct page *


Return whether the PG_vmemmap_self_hosted flag is set on a page.
page -- struct page *


Return whether the PG_waiters flag is set on a page.
page -- struct page *


Return whether the PG_workingset flag is set on a page.
page -- struct page *


Return whether the PG_writeback flag is set on a page.
page -- struct page *


Return whether the PG_xen_remapped flag is set on a page.
page -- struct page *


Return whether the PG_young flag is set on a page.
page -- struct page *


Return whether a page is part of a compound page.
page -- struct page *


Return whether a page is a head page in a compound page.
page -- struct page *


Return whether a page is a tail page in a compound page.
page -- struct page *


Get the head page associated with a page.

If page is a tail page, this returns the head page of the compound page it belongs to. Otherwise, it returns page.

page -- struct page *
struct page *


Return the allocation order of a potentially compound page.
page -- struct page *
unsigned int


Return the number of pages in a potentially compound page.
page -- struct page *
unsigned long


Return the number of bytes in a potentially compound page.
page -- struct page *
unsigned long


Get a human-readable representation of the flags set on a page. 

>>> decode_page_flags(page)
'PG_uptodate|PG_dirty|PG_lru|PG_reclaim|PG_swapbacked|PG_readahead|PG_savepinned|PG_isolated|PG_reported'
page -- struct page *


Iterate over every struct page * from the minimum to the maximum page.

NOTE:

This may include offline pages which don't have a valid struct page. Wrap accesses in a try ... except drgn.FaultError: 

>>> for page in for_each_page():
...     try:
...         if PageLRU(page):
...             print(hex(page))
...     except drgn.FaultError:
...         continue
0xfffffb4a000c0000
0xfffffb4a000c0040
...

This may be fixed in the future.



prog -- Program, which may be omitted to use the default program argument.
Iterator of struct page * objects.


Get the physical address of a page frame number (PFN).
  • prog -- Program, which may be omitted to use the default program argument.
  • pfn -- unsigned long

phys_addr_t


Get the page frame number (PFN) of a physical address.
  • prog -- Program, which may be omitted to use the default program argument.
  • addr -- phys_addr_t

unsigned long


Get the page frame number (PFN) of a page.
page -- struct page *
unsigned long


Get the physical address of a page.
page -- struct page *
phys_addr_t


Get the directly mapped virtual address of a page.
page -- struct page *
void *


Get the page with a page frame number (PFN).
  • prog -- Program, which may be omitted to use the default program argument.
  • pfn -- unsigned long

struct page *


Get the directly mapped virtual address of a page frame number (PFN).
  • prog -- Program, which may be omitted to use the default program argument.
  • pfn -- unsigned long

void *


Get the page containing a physical address.
  • prog -- Program, which may be omitted to use the default program argument.
  • addr -- phys_addr_t

struct page *


Get the directly mapped virtual address of a physical address.
  • prog -- Program, which may be omitted to use the default program argument.
  • addr -- phys_addr_t

void *


Get the page containing a directly mapped virtual address.

NOTE:

This only works for virtual addresses from the "direct map". This includes address from:
  • kmalloc
  • Slab allocator
  • Page allocator

But not:

  • vmalloc
  • vmap
  • ioremap
  • Symbols (function pointers, global variables)

For vmalloc or vmap addresses, use vmalloc_to_page(addr). For arbitrary kernel addresses, use follow_page(prog["init_mm"].address_of_(), addr).



  • prog -- Program, which may be omitted to use the default program argument.
  • addr -- void *

struct page *


Get the page frame number (PFN) of a directly mapped virtual address.

NOTE:

This only works for virtual addresses from the "direct map". For vmalloc or vmap addresses, use vmalloc_to_pfn(addr). For arbitrary kernel addresses, use follow_pfn(prog["init_mm"].address_of_(), addr).


  • prog -- Program, which may be omitted to use the default program argument.
  • addr -- void *

unsigned long


Get the physical address of a directly mapped virtual address.

NOTE:

This only works for virtual addresses from the "direct map". For arbitrary kernel addresses, use follow_phys(prog["init_mm"].address_of_(), addr).


  • prog -- Program, which may be omitted to use the default program argument.
  • addr -- void *

phys_addr_t


Get the page that a virtual address maps to in a virtual address space. 

>>> task = find_task(113)
>>> follow_page(task.mm, 0x7fffbbb6d4d0)
*(struct page *)0xffffbe4bc0337b80 = {


    ...
}
  • mm -- struct mm_struct *
  • addr -- void *

struct page *


Get the page frame number (PFN) that a virtual address maps to in a virtual address space. 

>>> task = find_task(113)
>>> follow_pfn(task.mm, 0x7fffbbb6d4d0)
(unsigned long)52718
  • mm -- struct mm_struct *
  • addr -- void *

unsigned long


Get the physical address that a virtual address maps to in a virtual address space. 

>>> task = find_task(113)
>>> follow_phys(task.mm, 0x7fffbbb6d4d0)
(phys_addr_t)215934160
  • mm -- struct mm_struct *
  • addr -- void *

phys_addr_t


Get the page containing a vmalloc or vmap address. 

>>> task = find_task(113)
>>> vmalloc_to_page(task.stack)
*(struct page *)0xffffbe4bc00a2200 = {


    ...
}
  • prog -- Program, which may be omitted to use the default program argument.
  • addr -- void *

struct page *


Get the page frame number (PFN) containing a vmalloc or vmap address. 

>>> task = find_task(113)
>>> vmalloc_to_pfn(task.stack)
(unsigned long)10376
  • prog -- Program, which may be omitted to use the default program argument.
  • addr -- void *

unsigned long


Read memory from a task's virtual address space. 

>>> task = find_task(1490152)
>>> access_process_vm(task, 0x7f8a62b56da0, 12)
b'hello, world'
  • task -- struct task_struct *
  • address -- Starting address.
  • size -- Number of bytes to read.



Read memory from a virtual address space. This is similar to access_process_vm(), but it takes a struct mm_struct * instead of a struct task_struct *. 

>>> task = find_task(1490152)
>>> access_remote_vm(task.mm, 0x7f8a62b56da0, 12)
b'hello, world'
  • mm -- struct mm_struct *
  • address -- Starting address.
  • size -- Number of bytes to read.



Get the list of command line arguments of a task, or None for kernel tasks. 

>>> cmdline(find_task(1495216))
[b'vim', b'drgn/helpers/linux/mm.py']
    

$ tr '\0' ' ' < /proc/1495216/cmdline
vim drgn/helpers/linux/mm.py


task -- struct task_struct *


Get the list of environment variables of a task, or None for kernel tasks. 

>>> environ(find_task(1497797))
[b'HOME=/root', b'PATH=/usr/local/sbin:/usr/local/bin:/usr/bin', b'LOGNAME=root']
    

$ tr '\0' '\n' < /proc/1497797/environ
HOME=/root
PATH=/usr/local/sbin:/usr/local/bin:/usr/bin
LOGNAME=root


task -- struct task_struct *


Return the virtual memory area (VMA) containing an address.
  • mm -- struct mm_struct *
  • addr -- Address to look up.

struct vm_area_struct * (NULL if not found)


Iterate over every virtual memory area (VMA) in a virtual address space. 

>>> for vma in for_each_vma(task.mm):
...     print(vma)
...
*(struct vm_area_struct *)0xffff97ad82bfc930 = {


    ...
}
*(struct vm_area_struct *)0xffff97ad82bfc0a8 = {


    ...
}
...
mm -- struct mm_struct *
Iterator of struct vm_area_struct * objects.


Return the total number of RAM pages.
prog -- Program, which may be omitted to use the default program argument.


Networking

The drgn.helpers.linux.net module provides helpers for working with the Linux kernel networking subsystem.

Get a socket from an inode referring to the socket.
inode -- struct inode *
struct socket *
ValueError -- If inode does not refer to a socket


Get the inode of a socket.
sock -- struct socket *
struct inode *


Iterate over all network namespaces in the system.
prog -- Program, which may be omitted to use the default program argument.
Iterator of struct net * objects.


Get a network namespace from a network namespace NSFS inode, e.g. /proc/$PID/ns/net or /var/run/netns/$NAME.
inode -- struct inode *
struct net *
ValueError -- if inode is not a network namespace inode


Get a network namespace from a task and a file descriptor referring to a network namespace NSFS inode, e.g. /proc/$PID/ns/net or /var/run/netns/$NAME.
  • task -- struct task_struct *
  • fd -- File descriptor.

struct net *
ValueError -- If fd does not refer to a network namespace inode


Iterate over all TX queues for a network device.
dev -- struct net_device *
Iterator of struct netdev_queue * objects.


Get the network device with the given interface index number.
  • net -- struct net *. Defaults to the initial network namespace if given a Program or omitted.
  • ifindex -- Network interface index number.

struct net_device * (NULL if not found)


Get the network device with the given interface name.
  • net -- struct net *. Defaults to the initial network namespace if given a Program or omitted.
  • name -- Network interface name.

struct net_device * (NULL if not found)


Return the private data of a network device. 

>>> dev = netdev_get_by_name("wlp0s20f3")
>>> netdev_priv(dev)
(void *)0xffff9419c9dec9c0
>>> netdev_priv(dev, "struct ieee80211_sub_if_data")
*(struct ieee80211_sub_if_data *)0xffff9419c9dec9c0 = {


    ...
}
  • dev -- struct net_device *
  • type -- Type of private data.

type *


Check whether a socket is a full socket, i.e., not a time-wait or request socket.
sk -- struct sock *


Iterate over all the entries in a nulls hash list of sockets specified by struct hlist_nulls_head head.
head -- struct hlist_nulls_head *
Iterator of struct sock * objects.


Get the shared info for a socket buffer.
skb -- struct sk_buff *
struct skb_shared_info *


NUMA Node Masks

The drgn.helpers.linux.nodemask module provides helpers for working with NUMA node masks from include/linux/nodemask.h.

Iterate over all of the NUMA nodes in the given mask.
mask -- nodemask_t


Iterate over all NUMA nodes in the given state.
  • prog -- Program, which may be omitted to use the default program argument.
  • state -- enum node_states (e.g., N_NORMAL_MEMORY)



Iterate over all possible NUMA nodes.
prog -- Program, which may be omitted to use the default program argument.


Iterate over all online NUMA nodes.
prog -- Program, which may be omitted to use the default program argument.


Return whether the given NUMA node has the given state.
  • node -- NUMA node number.
  • state -- enum node_states (e.g., N_NORMAL_MEMORY)



Per-CPU

The drgn.helpers.linux.percpu module provides helpers for working with per-CPU allocations from include/linux/percpu.h and per-CPU counters from include/linux/percpu_counter.h.

Return the per-CPU pointer for a given CPU. 

>>> prog["init_net"].loopback_dev.pcpu_refcnt
(int *)0x2c980
>>> per_cpu_ptr(prog["init_net"].loopback_dev.pcpu_refcnt, 7)
*(int *)0xffff925e3ddec980 = 4
  • ptr -- Per-CPU pointer, i.e., type __percpu *. For global variables, it's usually easier to use per_cpu().
  • cpu -- CPU number.

type * object.


Return the per-CPU variable for a given CPU. 

>>> print(repr(prog["runqueues"]))
Object(prog, 'struct rq', address=0x278c0)
>>> per_cpu(prog["runqueues"], 6).curr.comm
(char [16])"python3"
  • var -- Per-CPU variable, i.e., type __percpu (not a pointer; use per_cpu_ptr() for that).
  • cpu -- CPU number.

type object.


Return the sum of a per-CPU counter.
fbc -- struct percpu_counter *


Process IDS

The drgn.helpers.linux.pid module provides helpers for looking up process IDs and processes.

Return the struct task_struct * containing the given struct pid * of the given type.
  • pid -- struct pid *
  • pid_type -- enum pid_type

struct task_struct *


Return the struct pid * for the given PID number.
ns -- struct pid_namespace *. Defaults to the initial PID namespace if given a Program or omitted.
struct pid *


Iterate over all PIDs in a namespace.
ns -- struct pid_namespace *. Defaults to the initial PID namespace if given a Program or omitted.
Iterator of struct pid * objects.


Return the task with the given PID.
ns -- struct pid_namespace *. Defaults to the initial PID namespace if given a Program or omitted.
struct task_struct *


Iterate over all of the tasks visible in a namespace.
ns -- struct pid_namespace *. Defaults to the initial PID namespace if given a Program or omitted.
Iterator of struct task_struct * objects.


Priority-Sorted Lists

The drgn.helpers.linux.plist module provides helpers for working with descending-priority-sorted doubly-linked lists (struct plist_head and struct plist_node) from include/linux/plist.h.

Return whether a plist is empty.
head -- struct plist_head *


Return whether a plist node is empty (i.e., not on a list).
node -- struct plist_node *


Return the first (highest priority) entry in a plist.

The list is assumed to be non-empty.

  • head -- struct plist_head *
  • type -- Entry type.
  • member -- Name of list node member in entry type.

type *


Return the last (lowest priority) entry in a plist.

The list is assumed to be non-empty.

  • head -- struct plist_head *
  • type -- Entry type.
  • member -- Name of list node member in entry type.

type *


Iterate over all of the nodes in a plist.
head -- struct plist_head *
Iterator of struct plist_node * objects.


Iterate over all of the entries in a plist.
  • type -- Entry type.
  • head -- struct plist_head *
  • member -- Name of plist node member in entry type.

Iterator of type * objects.


Log Buffer

The drgn.helpers.linux.printk module provides helpers for reading the Linux kernel log buffer.

Bases: NamedTuple

Kernel log record.

Message text.

syslog(3) facility.

Log level.

Sequence number.

Timestamp in nanoseconds.

Thread ID of thread that logged this record, if available.

This is available if the message was logged from task context and if the kernel saves the printk() caller ID.

As of Linux 5.10, the kernel always saves the caller ID. From Linux 5.1 through 5.9, it is saved only if the kernel was compiled with CONFIG_PRINTK_CALLER. Before that, it is never saved.


Processor ID of CPU that logged this record, if available.

This is available only if the message was logged when not in task context (e.g., in an interrupt handler) and if the kernel saves the printk() caller ID.

See caller_tid for when the kernel saves the caller ID.


Whether this record is a continuation of a previous record.

Additional metadata for the message.

See the /dev/kmsg documentation for an explanation of the keys and values.



Get a list of records in the kernel log buffer.
prog -- Program, which may be omitted to use the default program argument.


Get the contents of the kernel log buffer formatted like dmesg(1).

If you just want to print the log buffer, use print_dmesg().

The format of each line is:

[   timestamp] message


If you need to format the log buffer differently, use get_printk_records() and format it yourself.

prog -- Program, which may be omitted to use the default program argument.


Print the contents of the kernel log buffer. 

>>> print_dmesg()
[    0.000000] Linux version 6.8.0-vmtest28.1default (drgn@drgn) (x86_64-linux-gcc (GCC) 12.2.0, GNU ld (GNU Binutils) 2.39) #1 SMP PREEMPT_DYNAMIC Mon Mar 11 06:38:45 UTC 2024
[    0.000000] Command line: rootfstype=9p rootflags=trans=virtio,cache=loose,msize=1048576 ro console=ttyS0,115200 panic=-1 crashkernel=256M init=/tmp/drgn-vmtest-rudzppeo/init
[    0.000000] BIOS-provided physical RAM map:
...
  • prog -- Program, which may be omitted to use the default program argument.
  • file -- File to print to. Defaults to sys.stdout.



Radix Trees

The drgn.helpers.linux.radixtree module provides helpers for working with radix trees from include/linux/radix-tree.h.

SEE ALSO:

XArrays, which were introduced in Linux 4.20 as a replacement for radix trees.


Look up the entry at a given index in a radix tree.
  • root -- struct radix_tree_root *
  • index -- Entry index.

void * found entry, or NULL if not found.


Iterate over all of the entries in a radix tree.
root -- struct radix_tree_root *
Iterator of (index, void *) tuples.


Red-Black Trees

The drgn.helpers.linux.rbtree module provides helpers for working with red-black trees from include/linux/rbtree.h.

Return whether a red-black tree is empty.
node -- struct rb_root *


Return whether a red-black tree node is empty, i.e., not inserted in a tree.
node -- struct rb_node *


Return the parent node of a red-black tree node.
node -- struct rb_node *
struct rb_node *


Return the first node (in sort order) in a red-black tree, or NULL if the tree is empty.
root -- struct rb_root *
struct rb_node *


Return the last node (in sort order) in a red-black tree, or NULL if the tree is empty.
root -- struct rb_root *
struct rb_node *


Return the next node (in sort order) after a red-black node, or NULL if the node is the last node in the tree or is empty.
node -- struct rb_node *
struct rb_node *


Return the previous node (in sort order) before a red-black node, or NULL if the node is the first node in the tree or is empty.
node -- struct rb_node *
struct rb_node *


Iterate over all of the nodes in a red-black tree, in sort order.
root -- struct rb_root *
Iterator of struct rb_node * objects.


Iterate over all of the entries in a red-black tree in sorted order.
  • type -- Entry type.
  • root -- struct rb_root *
  • member -- Name of struct rb_node member in entry type.

Iterator of type * objects.


Find an entry in a red-black tree given a key and a comparator function.

Note that this function does not have an analogue in the Linux kernel source code, as tree searches are all open-coded.

  • type -- Entry type.
  • root -- struct rb_root *
  • member -- Name of struct rb_node member in entry type.
  • key -- Key to find.
  • cmp -- Callback taking key and entry that returns < 0 if the key is less than the entry, > 0 if the key is greater than the entry, and 0 if the key matches the entry.

type * found entry, or NULL if not found.


Validate a red-black tree.

This checks that:

1.
The tree is a valid binary search tree ordered according to cmp.
2.
If allow_equal is False, there are no nodes that compare equal according to cmp.
3.
The rb_parent pointers are consistent.
4.
The red-black tree requirements are satisfied: the root node is black, no red node has a red child, and every path from any node to any of its descendant leaf nodes goes through the same number of black nodes.

  • type -- Entry type.
  • root -- struct rb_root *
  • member -- Name of struct rb_node member in entry type.
  • cmp -- Callback taking two type * entry objects that returns < 0 if the first entry is less than the second entry, > 0 if the first entry is greater than the second entry, and 0 if they are equal.
  • allow_equal -- Whether the tree may contain entries that compare equal to each other.

ValidationError -- if the tree is invalid


Like rbtree_inorder_for_each_entry(), but validates the red-black tree like validate_rbtree() while iterating.
  • type -- Entry type.
  • root -- struct rb_root *
  • member -- Name of struct rb_node member in entry type.
  • cmp -- Callback taking two type * entry objects that returns < 0 if the first entry is less than the second entry, > 0 if the first entry is greater than the second entry, and 0 if they are equal.
  • allow_equal -- Whether the tree may contain entries that compare equal to each other.

ValidationError -- if the tree is invalid


CPU Scheduler

The drgn.helpers.linux.sched module provides helpers for working with the Linux CPU scheduler.

Return the CPU number that the given task last ran on.
task -- struct task_struct *


Return the task running on the given CPU. 

>>> cpu_curr(7).comm
(char [16])"python3"
  • prog -- Program, which may be omitted to use the default program argument.
  • cpu -- CPU number.

struct task_struct *


Return the idle thread (PID 0, a.k.a swapper) for the given CPU. 

>>> idle_task(1).comm
(char [16])"swapper/1"
  • prog -- Program, which may be omitted to use the default program argument.
  • cpu -- CPU number.

struct task_struct *


Get the state of the task as a character (e.g., 'R' for running). See ps(1) for a description of the process state codes.
task -- struct task_struct *


Return system load averaged over 1, 5 and 15 minutes as tuple of three float values. 

>>> loadavg()
(2.34, 0.442, 1.33)
prog -- Program, which may be omitted to use the default program argument.


Slab Allocator

The drgn.helpers.linux.slab module provides helpers for working with the Linux slab allocator.

WARNING:

Beware of slab merging when using these helpers. See slab_cache_is_merged().


Return whether a slab cache has been merged with any other slab caches.

Unless configured otherwise, the kernel may merge slab caches of similar sizes together. See the SLUB users guide and slab_merge/slab_nomerge in the kernel parameters documentation.

This can cause confusion, as only the name of the first cache will be found, and objects of different types will be mixed in the same slab cache.

For example, suppose that we have two types, struct foo and struct bar, which have the same size but are otherwise unrelated. If the kernel creates a slab cache named foo for struct foo, then another slab cache named bar for struct bar, then slab cache foo will be reused instead of creating another cache for bar. So the following will fail:

find_slab_cache("bar")


And the following will also return struct bar * objects errantly casted to struct foo *:

slab_cache_for_each_allocated_object(find_slab_cache("foo"), "struct foo")


Unfortunately, these issues are difficult to work around generally, so one must be prepared to handle them on a case-by-case basis (e.g., by looking up the slab cache by its variable name and by checking that members of the structure make sense for the expected type).

slab_cache -- struct kmem_cache *


Return a dict mapping slab cache name to the cache it was merged with.

The SLAB and SLUB subsystems can merge caches with similar settings and object sizes, as described in the documentation of slab_cache_is_merged(). In some cases, the information about which caches were merged is lost, but in other cases, we can reconstruct the info. This function reconstructs the mapping, but requires that the kernel is configured with CONFIG_SLUB and CONFIG_SYSFS.

The returned dict maps from original cache name, to merged cache name. You can use this mapping to discover the correct cache to lookup via find_slab_cache(). The dict contains an entry only for caches which were merged into a cache of a different name.

>>> cache_to_merged = get_slab_cache_aliases()
>>> cache_to_merged["dnotify_struct"]
'avc_xperms_data'
>>> "avc_xperms_data" in cache_to_merged
False
>>> find_slab_cache("dnotify_struct") is None
True
>>> find_slab_cache("avc_xperms_data") is None
False
This function will only work on kernels which are built with CONFIG_SLUB and CONFIG_SYSFS enabled.
prog -- Program, which may be omitted to use the default program argument.
Mapping of slab cache name to final merged name
LookupError -- If the helper fails because the debugged kernel doesn't have the required configuration


Iterate over all slab caches.
prog -- Program, which may be omitted to use the default program argument.
Iterator of struct kmem_cache * objects.


Return the slab cache with the given name.
  • prog -- Program, which may be omitted to use the default program argument.
  • name -- Slab cache name.

struct kmem_cache *


Print the name and struct kmem_cache * value of all slab caches.
prog -- Program, which may be omitted to use the default program argument.


Iterate over all allocated objects in a given slab cache.

Only the SLUB and SLAB allocators are supported; SLOB does not store enough information to identify objects in a slab cache.

>>> dentry_cache = find_slab_cache("dentry")
>>> next(slab_cache_for_each_allocated_object(dentry_cache, "struct dentry"))
*(struct dentry *)0xffff905e41404000 = {


    ...
}
  • slab_cache -- struct kmem_cache *
  • type -- Type of object in the slab cache.

Iterator of type * objects.


Get information about an address if it is in a slab object. 

>>> ptr = find_task(1).comm.address_of_()
>>> info = slab_object_info(ptr)
>>> info
SlabObjectInfo(slab_cache=Object(prog, 'struct kmem_cache *', address=0xffffdb93c0045e18), slab=Object(prog, 'struct slab *', value=0xffffdb93c0045e00), address=0xffffa2bf81178000, allocated=True)

Note that SlabObjectInfo.address is the start address of the object, which may be less than addr if addr points to a member inside of the object:

>>> ptr.value_() - info.address
1496
>>> offsetof(prog.type("struct task_struct"), "comm")
1496

Note that SLOB does not store enough information to identify slab objects, so if the kernel is configured to use SLOB, this will always return None.

  • prog -- Program, which may be omitted to use the default program argument.
  • addr -- void *

SlabObjectInfo if addr is in a slab object, or None if not.


Information about an object in the slab allocator.
struct kmem_cache * that the slab object is from.

Slab containing the slab object.

Since Linux v5.17, this is a struct slab *. Before that, it is a struct page *.


Address of the slab object.

True if the object is allocated, False if it is free.


Get the slab cache that an address was allocated from, if any.

Note that SLOB does not store enough information to identify objects in a slab cache, so if the kernel is configured to use SLOB, this will always return NULL.

  • prog -- Program, which may be omitted to use the default program argument.
  • addr -- void *

struct kmem_cache * containing addr, or NULL if addr is not from a slab cache.


Stack Depot

The drgn.helpers.linux.stackdepot module provides helpers for working with the stack trace storage from include/linux/stackdepot.h used by KASAN and other kernel debugging tools.

Returns a stack trace for the given stack handle.
handle -- depot_stack_handle_t
The stack trace, or None if not available.


Traffic Control (TC)

The drgn.helpers.linux.tc module provides helpers for working with the Linux kernel Traffic Control (TC) subsystem.

Get a Qdisc from a device and a major handle number. It is worth noting that conventionally handles are hexadecimal, e.g. 10: in a tc command means major handle 0x10.
  • dev -- struct net_device *
  • major -- Qdisc major handle number.

struct Qdisc * (NULL if not found)


TCP

The drgn.helpers.linux.tcp module provides helpers for working with the TCP protocol in the Linux kernel.

Return the TCP protocol state of a socket.
sk -- struct sock *
TCP state enum value.


Users

The drgn.helpers.linux.user module provides helpers for working with users in the Linux kernel.

Return the user structure with the given UID.
  • prog -- Program, which may be omitted to use the default program argument.
  • uid -- kuid_t object or integer.

struct user_struct * (NULL if not found)


Iterate over all users in the system.
prog -- Program, which may be omitted to use the default program argument.
Iterator of struct user_struct * objects.


Wait Queues

The drgn.helpers.linux.wait module provides helpers for working with wait queues (wait_queue_head_t and wait_queue_entry_t) from include/linux/wait.h.

NOTE:

Since Linux 4.13, entries in a wait queue have type wait_queue_entry_t. Before that, the type was named wait_queue_t.


Return whether a wait queue has any waiters.
wq -- wait_queue_head_t *


Iterate over all entries in a wait queue.
wq -- wait_queue_head_t *
Iterator of wait_queue_entry_t * or wait_queue_t * objects depending on the kernel version.


Iterate over all tasks waiting on a wait queue.

WARNING:

This comes from wait_queue_entry_t::private, which usually stores a task. However, some wait queue entries store a different pointer type, in which case this will return garbage.


wq -- wait_queue_head_t *
Iterator of struct task_struct * objects.


XArrays

The drgn.helpers.linux.xarray module provides helpers for working with the XArray data structure from include/linux/xarray.h.

NOTE:

XArrays were introduced in Linux 4.20 as a replacement for radix trees. To make it easier to work with data structures that were changed from a radix tree to an XArray (like struct address_space::i_pages), drgn treats XArrays and radix trees interchangeably in some cases.

Specifically, xa_load() is equivalent to radix_tree_lookup(), and xa_for_each() is equivalent to radix_tree_for_each(), except that the radix tree helpers assume advanced=False. (Therefore, xa_load() and xa_for_each() also accept a struct radix_tree_root *, and radix_tree_lookup() and radix_tree_for_each() also accept a struct xarray *.)



Look up the entry at a given index in an XArray. 

>>> entry = xa_load(inode.i_mapping.i_pages.address_of_(), 2)
>>> cast("struct page *", entry)
*(struct page *)0xffffed6980306f40 = {


    ...
}
  • xa -- struct xarray *
  • index -- Entry index.
  • advanced -- Whether to return nodes only visible to the XArray advanced API. If False, zero entries (see xa_is_zero()) will be returned as NULL.

void * found entry, or NULL if not found.


Iterate over all of the entries in an XArray. 

>>> for index, entry in xa_for_each(inode.i_mapping.i_pages.address_of_()):
...     print(index, entry)
...
0 (void *)0xffffed6980356140
1 (void *)0xffffed6980306f80
2 (void *)0xffffed6980306f40
3 (void *)0xffffed6980355b40
  • xa -- struct xarray *
  • advanced -- Whether to return nodes only visible to the XArray advanced API. If False, zero entries (see xa_is_zero()) will be skipped.

Iterator of (index, void *) tuples.


Return whether an XArray entry is a value.

See xa_to_value().

entry -- void *


Return the value in an XArray entry.

In addition to pointers, XArrays can store integers between 0 and LONG_MAX. If xa_is_value() returns True, use this to get the stored integer.

>>> entry = xa_load(xa, 9)
>>> entry
(void *)0xc9
>>> xa_is_value(entry)
True
>>> xa_to_value(entry)
(unsigned long)100
entry -- void *
unsigned long


Return whether an XArray entry is a "zero" entry.

A zero entry is an entry that was reserved but is not present. These are only visible to the XArray advanced API, so they are only returned by xa_load() and xa_for_each() when advanced = True.

>>> entry = xa_load(xa, 10, advanced=True)
>>> entry
(void *)0x406
>>> xa_is_zero(entry)
True
>>> xa_load(xa, 10)
(void *)0
entry -- void *


Support Matrix

Architectures

Some features in drgn require architecture-specific support. The current status of this support is:

Architecture Linux Kernel Modules [1] Stack Traces [2] Virtual Address Translation [3]
x86-64 ✓ ✓ ✓
AArch64 ✓ ✓ ✓
s390x ✓ ✓ ✓
ppc64 ✓ ✓ ✓
i386 ✓
Arm ✓
RISC-V ✓

Key

[1]
Support for loading debugging symbols for Linux kernel modules.
[2]
Support for capturing stack traces (drgn.Program.stack_trace(), drgn.Thread.stack_trace()).
[3]
Support for translating virtual addresses, which is required for reading from vmalloc/vmap and module memory in Linux kernel vmcores and for various helpers in drgn.helpers.linux.mm.

The listed architectures are recognized in drgn.Architecture. Other architectures are represented by drgn.Architecture.UNKNOWN. Features not mentioned above should work on any architecture, listed or not.

Cross-Debugging

drgn can debug architectures different from the host. For example, you can debug an AArch64 (kernel or userspace) core dump from an x86-64 machine.

Linux Kernel Versions

drgn officially supports the current mainline, stable, and longterm kernel releases from kernel.org. (There may be some delay before a new mainline version is fully supported.) End-of-life versions are supported until it becomes too difficult to do so. The kernel versions currently fully supported are:

  • 6.0-6.8
  • 5.10-5.19
  • 5.4
  • 4.19
  • 4.14
  • 4.9

Other versions are not tested. They'll probably mostly work, but support is best-effort.

Kernel Configuration

drgn supports debugging kernels with various configurations:

  • SMP and !SMP.
  • Preemptible and non-preemptible.
  • SLUB, SLAB, and SLOB allocators.

drgn requires a kernel configured with CONFIG_PROC_KCORE=y for live kernel debugging.

Case Studies

These are writeups of real-world problems solved with drgn.

Recovering a dm-crypt Encryption Key

Author: Omar Sandoval
Date: January 11th, 2024

dm-crypt is the Linux kernel's transparent disk encryption subsystem. I recently had to recover the master key for an encrypted disk where the passphrase was no longer known, but the dm-crypt device was still open. Normally, the key is stored in kernel space and cannot be accessed by user space. However, with drgn, we can traverse kernel data structures to recover the key. This is a great example of how to jump between kernel code and drgn to navigate a subsystem.

WARNING:

The dm-crypt master key is obviously very sensitive information that shouldn't be exposed carelessly.

As a disclaimer for anyone concerned about the security implications: everything is working as intended here. Debugging the live kernel with drgn requires root, and root has many other ways to access sensitive information (loading kernel modules, triggering a kernel core dump, etc.). Solutions like inline encryption and kernel_lockdown(7) can be used for defense in depth if necessary.



Setup

For this writeup, I'm going to set up dm-crypt in a virtual machine running Linux 6.7.

# uname -r
6.7.0
# cryptsetup luksFormat /dev/vdb
WARNING!
========
This will overwrite data on /dev/vdb irrevocably.
Are you sure? (Type 'yes' in capital letters): YES
Enter passphrase for /dev/vdb: hello
Verify passphrase: hello
# cryptsetup open /dev/vdb mycrypt
Enter passphrase for /dev/vdb: hello


The default configuration is AES in XTS mode with a 512-bit key:

# cryptsetup status mycrypt
/dev/mapper/mycrypt is active.


  type:    LUKS2


  cipher:  aes-xts-plain64


  keysize: 512 bits


  key location: keyring


  device:  /dev/vdb


  sector size:  512


  offset:  32768 sectors


  size:    33521664 sectors


  mode:    read/write


The new device is dm-0:

# realpath /dev/mapper/mycrypt
/dev/dm-0


Getting from Device Mapper to the Crypto API

The dm-crypt documentation tells us that "Device-mapper is infrastructure in the Linux kernel that provides a generic way to create virtual layers of block devices. Device-mapper crypt target provides transparent encryption of block devices using the kernel crypto API."

Our first goal is therefore to get to whatever context is used by the crypto API, which likely includes the encryption key. To do that, we're going to have to navigate through the device mapper code.

To start, let's find the virtual disk for our dm-crypt target in drgn using the for_each_disk() and disk_name() helpers:

>>> for disk in for_each_disk():
...     if disk_name(disk) == b"dm-0":
...             print(disk)
...             break
...
*(struct gendisk *)0xffffa3b9421b2c00 = {


        ...
}

struct gendisk has a function table, fops, with callbacks to the disk driver. Specifically, the submit_bio callback intercepts disk reads and writes:

>>> disk.fops.submit_bio
(void (*)(struct bio *))dm_submit_bio+0x0 = 0xffffffffc05761e0


Let's take a look at dm_submit_bio():

static void dm_submit_bio(struct bio *bio)
{


        struct mapped_device *md = bio->bi_bdev->bd_disk->private_data;


        int srcu_idx;


        struct dm_table *map;


        map = dm_get_live_table(md, &srcu_idx);


        ...


        dm_split_and_process_bio(md, map, bio);


        ...
}


So the disk's private data is a struct mapped_device. Let's get it in drgn:

>>> md = cast("struct mapped_device *", disk.private_data)


dm_get_live_table() gets the device mapper table:

struct dm_table *dm_get_live_table(struct mapped_device *md,


                                   int *srcu_idx) __acquires(md->io_barrier)
{


        *srcu_idx = srcu_read_lock(&md->io_barrier);


        return srcu_dereference(md->map, &md->io_barrier);
}


SRCU is a synchronization mechanism which we can blithely ignore:

>>> map = cast("struct dm_table *", md.map)


dm_submit_bio() then calls dm_split_and_process_bio(), which calls __split_and_process_bio():

static blk_status_t __split_and_process_bio(struct clone_info *ci)
{


        struct bio *clone;


        struct dm_target *ti;


        unsigned int len;


        ti = dm_table_find_target(ci->map, ci->sector);


        ...


        __map_bio(clone);
}


dm_table_find_target() finds the appropriate device mapper target in a table:

struct dm_target *dm_table_find_target(struct dm_table *t, sector_t sector)
{


        ...


        return &t->targets[(KEYS_PER_NODE * n) + k];
}


Our simple case only has one target:

>>> map.num_targets
(unsigned int)1
>>> ti = map.targets


__split_and_process_bio() then calls __map_bio():

static void __map_bio(struct bio *clone)
{


        struct dm_target_io *tio = clone_to_tio(clone);


        struct dm_target *ti = tio->ti;


        struct dm_io *io = tio->io;


        struct mapped_device *md = io->md;


        int r;


        ...


                if (likely(ti->type->map == linear_map))


                        r = linear_map(ti, clone);


                else if (ti->type->map == stripe_map)


                        r = stripe_map(ti, clone);


                else


                        r = ti->type->map(ti, clone);


        ...
}


So we need to look at another callback:

>>> ti.type.map
(dm_map_fn)crypt_map+0x0 = 0xffffffffc08a03f0


crypt_map() is part of dm-crypt, so we've finally made it out of generic device mapper:

static int crypt_map(struct dm_target *ti, struct bio *bio)
{


        struct dm_crypt_io *io;


        struct crypt_config *cc = ti->private;


        ...


And we have the dm-crypt configuration:

>>> cc = cast("struct crypt_config *", ti.private)


Dumping it out reveals some crypto API context!

>>> cc
*(struct crypt_config *)0xffffa3b9421b2400 = {


        ...


        .cipher_tfm = (union <anonymous>){


                .tfms = (struct crypto_skcipher **)0xffffa3b9438667c0,


                ...


        },


        .tfms_count = (unsigned int)1,


        ...
}
>>> tfm = cc.cipher_tfm.tfms[0]


Descending Down the Crypto API

The Linux kernel crypto API is very generic and is implemented with a lot of runtime polymorphism. Our next goal is to traverse through the crypto API data structures to find the key.

The crypto API refers to cryptographic ciphers as "transformations". Transformations can be combined and nested in various ways. The tfm variable we found is a "transformation object", which is an instance of a transformation:

>>> tfm
*(struct crypto_skcipher *)0xffffa3b948218c00 = {


        .reqsize = (unsigned int)160,


        .base = (struct crypto_tfm){


                .refcnt = (refcount_t){


                        .refs = (atomic_t){


                                .counter = (int)1,


                        },


                },


                .crt_flags = (u32)0,


                .node = (int)-1,


                .exit = (void (*)(struct crypto_tfm *))crypto_skcipher_exit_tfm+0x0 = 0xffffffffb77d2600,


                .__crt_alg = (struct crypto_alg *)0xffffa3b943dab448,


                .__crt_ctx = (void *[]){},


        },
}
>>> tfm.base.__crt_alg
*(struct crypto_alg *)0xffffa3b943dab448 = {


        ...


        .cra_name = (char [128])"xts(aes)",


        ...
}


This is an skcipher, or a symmetric key cipher. It is using the xts(aes) algorithm as expected. __crt_ctx is an opaque context, which is promising if we can figure out how to interpret it. The exit callback looks like a cleanup function. That seems like a good way for us to figure out how __crt_ctx is used. Here are crypto_skcipher_exit_tfm() and the crypto_skcipher_alg() and crypto_skcipher_tfm() getters it uses:

static void crypto_skcipher_exit_tfm(struct crypto_tfm *tfm)
{


        struct crypto_skcipher *skcipher = __crypto_skcipher_cast(tfm);


        struct skcipher_alg *alg = crypto_skcipher_alg(skcipher);


        alg->exit(skcipher);
}
static inline struct skcipher_alg *crypto_skcipher_alg(


        struct crypto_skcipher *tfm)
{


        return container_of(crypto_skcipher_tfm(tfm)->__crt_alg,


                            struct skcipher_alg, base);
}
static inline struct crypto_tfm *crypto_skcipher_tfm(


        struct crypto_skcipher *tfm)
{


        return &tfm->base;
}


We can emulate the getters in drgn to find the underlying implementation:

>>> def crypto_skcipher_alg(tfm):
...     return container_of(tfm.base.__crt_alg, "struct skcipher_alg", "base")
...
>>> crypto_skcipher_alg(tfm).exit
(void (*)(struct crypto_skcipher *))simd_skcipher_exit+0x0 = 0xffffffffc058b1f0


My machine supports the AES-NI x86 extension. The kernel cannot use SIMD instructions like AES-NI in some contexts, so it has an extra layer of indirection to go through an asynchronous daemon when necessary. This involves a couple of wrapper transformation objects. simd_skcipher_exit() shows us how to unwrap the first one:

static void simd_skcipher_exit(struct crypto_skcipher *tfm)
{


        struct simd_skcipher_ctx *ctx = crypto_skcipher_ctx(tfm);


        cryptd_free_skcipher(ctx->cryptd_tfm);
}


We just need one more getter in drgn, crypto_skcipher_ctx():

>>> def crypto_skcipher_ctx(tfm):
...     return cast("void *", tfm.base.__crt_ctx)
...
>>> simd_ctx = cast("struct simd_skcipher_ctx *", crypto_skcipher_ctx(tfm))
>>> cryptd_tfm = simd_ctx.cryptd_tfm
>>> cryptd_tfm
*(struct cryptd_skcipher *)0xffffa3b94b5e4cc0 = {


        .base = (struct crypto_skcipher){


                .reqsize = (unsigned int)80,


                .base = (struct crypto_tfm){


                        .refcnt = (refcount_t){


                                .refs = (atomic_t){


                                        .counter = (int)1,


                                },


                        },


                        .crt_flags = (u32)0,


                        .node = (int)-1,


                        .exit = (void (*)(struct crypto_tfm *))crypto_skcipher_exit_tfm+0x0 = 0xffffffffb77d2600,


                        .__crt_alg = (struct crypto_alg *)0xffffa3b9421b2848,


                        .__crt_ctx = (void *[]){},


                },


        },
}


We saw crypto_skcipher_exit_tfm() earlier, so we know where to look next:

>>> crypto_skcipher_alg(cryptd_tfm.base).exit
(void (*)(struct crypto_skcipher *))cryptd_skcipher_exit_tfm+0x0 = 0xffffffffc04d6210


cryptd_skcipher_exit_tfm() shows us how to unwrap this transformation object:

static void cryptd_skcipher_exit_tfm(struct crypto_skcipher *tfm)
{


        struct cryptd_skcipher_ctx *ctx = crypto_skcipher_ctx(tfm);


        crypto_free_skcipher(ctx->child);
}


Now we can get the actual cipher transformation object:

>>> cryptd_ctx = cast("struct cryptd_skcipher_ctx *", crypto_skcipher_ctx(cryptd_tfm.base))
>>> child_tfm = cryptd_ctx.child
>>> child_tfm
*(struct crypto_skcipher *)0xffffa3b945dc4000 = {


        .reqsize = (unsigned int)0,


        .base = (struct crypto_tfm){


                .refcnt = (refcount_t){


                        .refs = (atomic_t){


                                .counter = (int)1,


                        },


                },


                .crt_flags = (u32)0,


                .node = (int)-1,


                .exit = (void (*)(struct crypto_tfm *))0x0,


                .__crt_alg = (struct crypto_alg *)0xffffffffc05e7d80,


                .__crt_ctx = (void *[]){},


        },
}


This one doesn't have an exit callback, so let's look at the algorithm:

>>> crypto_skcipher_alg(child_tfm)
*(struct skcipher_alg *)0xffffffffc05e7d40 = {


        .setkey = (int (*)(struct crypto_skcipher *, const u8 *, unsigned int))xts_aesni_setkey+0x0 = 0xffffffffc059efb0,


        ...
}


xts_aesni_setkey() is very enlightening:

static int xts_aesni_setkey(struct crypto_skcipher *tfm, const u8 *key,


                            unsigned int keylen)
{


        struct aesni_xts_ctx *ctx = aes_xts_ctx(tfm);


        int err;


        err = xts_verify_key(tfm, key, keylen);


        if (err)


                return err;


        keylen /= 2;


        /* first half of xts-key is for crypt */


        err = aes_set_key_common(&ctx->crypt_ctx, key, keylen);


        if (err)


                return err;


        /* second half of xts-key is for tweak */


        return aes_set_key_common(&ctx->tweak_ctx, key + keylen, keylen);
}


XTS splits the provided key into two keys: one for data and one for a "tweak". They are stored in ctx->crypt_ctx and ctx->tweak_ctx, respectively.

To reach ctx, we need one more getter, aes_xts_ctx():

static inline struct aesni_xts_ctx *aes_xts_ctx(struct crypto_skcipher *tfm)
{


        return aes_align_addr(crypto_skcipher_ctx(tfm));
}


Which uses aes_align_addr():

#define AESNI_ALIGN     16
static inline void *aes_align_addr(void *addr)
{


        if (crypto_tfm_ctx_alignment() >= AESNI_ALIGN)


                return addr;


        return PTR_ALIGN(addr, AESNI_ALIGN);
}


Implementing that in drgn gets us the key material!

>>> def aes_xts_ctx(tfm):
...     AESNI_ALIGN = 16
...     mask = AESNI_ALIGN - 1
...     ctx = cast("unsigned long", crypto_skcipher_ctx(tfm))
...     return cast("struct aesni_xts_ctx *", (ctx + mask) & ~mask)
...
>>> xts_ctx = aes_xts_ctx(cryptd_ctx.child)
>>> xts_ctx
*(struct aesni_xts_ctx *)0xffffa3b945dc4030 = {


        .tweak_ctx = (struct crypto_aes_ctx){


                .key_enc = (u32 [60]){


                        4053857025, 2535432618, 3497512106, 429624542,


                        190965574, 620881567, 2728140233, 1574816406,


                        1642869364, 4143158238, 646209396, 1059050410,


                        2124513770, 1537238901, 4181490364, 2766254122,


                        2225457809, 1918261583, 1423050299, 1808651665,


                        18645611, 1522328862, 2743115682, 123809672, 1080042880,


                        842431695, 1726249716, 220835685, 3602512678,


                        2349145656, 797278618, 686075410, 2304003180,


                        3143774371, 3716565591, 3501188402, 2797609477,


                        717569085, 88128935, 765727669, 1552680193, 3891148194,


                        979927029, 3938949831, 554080963, 197371646, 243473241,


                        589760748, 2460666129, 1967455411, 1328317254,


                        2783648129, 669994703, 741140529, 581956456, 25754500,


                        3453357406, 3096637933, 4156453547, 1381329706,


                },


                .key_dec = (u32 [60]){


                        3453357406, 3096637933, 4156453547, 1381329706,


                        1691590497, 1611861415, 2033812690, 3535200077,


                        1503779265, 1400120959, 2713205381, 402136101,


                        2278736107, 79729350, 422218101, 2878299039, 3072023845,


                        181796798, 4073463034, 3057657504, 2722800653,


                        2199015981, 501881779, 2997211882, 893456792,


                        3184435867, 4162446148, 1150040666, 3430456984,


                        559478304, 2667071902, 2941241689, 2504843709,


                        2291118851, 1171735007, 3163937054, 4210330224,


                        3978324152, 3214983102, 834109639, 179351664, 499339966,


                        3445158620, 4181891265, 4283462504, 399827656,


                        1384175366, 2383888249, 3581021031, 393470670,


                        3499860066, 874146333, 3319833674, 3901002144,


                        1163146702, 3700942975, 4053857025, 2535432618,


                        3497512106, 429624542,


                },


                .key_length = (u32)32,


        },


        .crypt_ctx = (struct crypto_aes_ctx){


                .key_enc = (u32 [60]){


                        91118336, 1683438947, 280915620, 1674463119, 3416529787,


                        95371281, 156839573, 539041733, 2748950209, 3348011938,


                        3610309894, 3036590729, 1176448220, 1135635661,


                        1256800856, 1791516061, 4259008143, 978703661,


                        3982827563, 1503367842, 2366333926, 3468365611,


                        2219986291, 4003074286, 3589535297, 4020642668,


                        46334791, 1532531173, 3026313791, 2061167892,


                        4270366823, 269660297, 1916354478, 2644450498,


                        2673614725, 3288632928, 2828270575, 3528005371,


                        750892700, 1020462613, 735205841, 3058517267, 689003158,


                        3977630966, 4257919917, 797156694, 54662090, 1066472927,


                        3047676072, 65707451, 721143597, 3354268635, 1004719636,


                        341928770, 388200584, 682782039, 4002672596, 3984159343,


                        3347232066, 7120537,


                },


                .key_dec = (u32 [60]){


                        4002672596, 3984159343, 3347232066, 7120537, 2275767381,


                        3582792214, 728749911, 250810445, 2145441323,


                        3415330885, 1171250799, 717236012, 72947820, 1378379331,


                        4276274497, 631031578, 3286455042, 3027306094,


                        2388528682, 1863317827, 1027747936, 1450278447,


                        2898961154, 3682468443, 2929020077, 2006078828,


                        976160836, 3780245353, 3002856629, 1798524495,


                        4206615853, 2008326489, 523503039, 3641121217,


                        1304255784, 3682533165, 3583917429, 3653810938,


                        2441646946, 2366602356, 2101484483, 3325238398,


                        2495235305, 2529403397, 1276800912, 206997391,


                        1212164504, 478670614, 2260253082, 3144746941,


                        1384732823, 41543404, 2858181789, 1078781983,


                        1142337047, 1422378638, 91118336, 1683438947, 280915620,


                        1674463119,


                },


                .key_length = (u32)32,


        },
}


Extracting the AES Key

Since we have a 512-bit key, XTS uses two 256-bit AES keys. You'll notice that the key_enc fields above are much larger than that. This is because AES expands the key into a number of "round keys" using a "key schedule". Luckily, the first few round keys are copied directly from the original key.

With that information, we can finally recover the original key:

>>> def aes_key_from_ctx(ctx):
...     words = ctx.key_enc.value_()[:ctx.key_length / 4]
...     return b"".join(word.to_bytes(4, "little") for word in words)
...
>>> aes_key_from_ctx(xts_ctx.crypt_ctx).hex()
'005b6e05633d5764a46ebe108f47ce637b1ba4cb1140af05952e5909c51f2120'
>>> aes_key_from_ctx(xts_ctx.tweak_ctx).hex()
'01f3a0f1aaa11f97aacc77d0de8c9b1946e7610b9fe60125c91d9ca296cadd5d'


Which we can double check with cryptsetup:



 # cryptsetup luksDump --dump-master-key /dev/vdb


 WARNING!


 ========


 The header dump with volume key is sensitive information


 that allows access to encrypted partition without a passphrase.


 This dump should be stored encrypted in a safe place.


 Are you sure? (Type 'yes' in capital letters): YES


 Enter passphrase for /dev/vdb: hello


 LUKS header information for /dev/vdb


 Cipher name:    aes


 Cipher mode:    xts-plain64


 Payload offset: 32768


 UUID:           b43cba2c-532b-4491-bbb9-763b55bd7f03


 MK bits:        512


 MK dump:        00 5b 6e 05 63 3d 57 64 a4 6e be 10 8f 47 ce 63


                 7b 1b a4 cb 11 40 af 05 95 2e 59 09 c5 1f 21 20


                 01 f3 a0 f1 aa a1 1f 97 aa cc 77 d0 de 8c 9b 19


                 46 e7 61 0b 9f e6 01 25 c9 1d 9c a2 96 ca dd 5d


Conclusion

Before this, I had almost no knowledge of device mapper or crypto API internals. drgn makes it easy to explore the kernel and learn how it works.

Note that different system configurations will have different representations in the crypto API. For example, different ciphers modes will obviously have different transformations. Even the lack of AES-NI with the same cipher mode results in different transformation objects.

I converted this case study to the dm_crypt_key.py script in drgn's contrib directory. It could be extended to cover other ciphers in the future.

Using Stack Trace Variables to Find a Kyber Bug

Author: Omar Sandoval
Date: June 9th, 2021

Jakub Kicinski reported a crash in the Kyber I/O scheduler when he was testing Linux 5.12. He captured a core dump and asked me to debug it. This is a quick writeup of that investigation.

First, we can get the task that crashed:

>>> task = per_cpu(prog["runqueues"], prog["crashing_cpu"]).curr


Then, we can get its stack trace:

>>> trace = prog.stack_trace(task)
>>> trace
#0  queued_spin_lock_slowpath (../kernel/locking/qspinlock.c:471:3)
#1  queued_spin_lock (../include/asm-generic/qspinlock.h:85:2)
#2  do_raw_spin_lock (../kernel/locking/spinlock_debug.c:113:2)
#3  spin_lock (../include/linux/spinlock.h:354:2)
#4  kyber_bio_merge (../block/kyber-iosched.c:573:2)
#5  blk_mq_sched_bio_merge (../block/blk-mq-sched.h:37:9)
#6  blk_mq_submit_bio (../block/blk-mq.c:2182:6)
#7  __submit_bio_noacct_mq (../block/blk-core.c:1015:9)
#8  submit_bio_noacct (../block/blk-core.c:1048:10)
#9  submit_bio (../block/blk-core.c:1125:9)
#10 submit_stripe_bio (../fs/btrfs/volumes.c:6553:2)
#11 btrfs_map_bio (../fs/btrfs/volumes.c:6642:3)
#12 btrfs_submit_data_bio (../fs/btrfs/inode.c:2440:8)
#13 submit_one_bio (../fs/btrfs/extent_io.c:175:9)
#14 submit_extent_page (../fs/btrfs/extent_io.c:3229:10)
#15 __extent_writepage_io (../fs/btrfs/extent_io.c:3793:9)
#16 __extent_writepage (../fs/btrfs/extent_io.c:3872:8)
#17 extent_write_cache_pages (../fs/btrfs/extent_io.c:4514:10)
#18 extent_writepages (../fs/btrfs/extent_io.c:4635:8)
#19 do_writepages (../mm/page-writeback.c:2352:10)
#20 __writeback_single_inode (../fs/fs-writeback.c:1467:8)
#21 writeback_sb_inodes (../fs/fs-writeback.c:1732:3)
#22 __writeback_inodes_wb (../fs/fs-writeback.c:1801:12)
#23 wb_writeback (../fs/fs-writeback.c:1907:15)
#24 wb_check_background_flush (../fs/fs-writeback.c:1975:10)
#25 wb_do_writeback (../fs/fs-writeback.c:2063:11)
#26 wb_workfn (../fs/fs-writeback.c:2091:20)
#27 process_one_work (../kernel/workqueue.c:2275:2)
#28 worker_thread (../kernel/workqueue.c:2421:4)
#29 kthread (../kernel/kthread.c:292:9)
#30 ret_from_fork+0x1f/0x2a (../arch/x86/entry/entry_64.S:294)


It looks like kyber_bio_merge() tried to lock an invalid spinlock. For reference, this is the source code of kyber_bio_merge():

static bool kyber_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio,


                            unsigned int nr_segs)
{


        struct kyber_hctx_data *khd = hctx->sched_data;


        struct blk_mq_ctx *ctx = blk_mq_get_ctx(hctx->queue);


        struct kyber_ctx_queue *kcq = &khd->kcqs[ctx->index_hw[hctx->type]];


        unsigned int sched_domain = kyber_sched_domain(bio->bi_opf);


        struct list_head *rq_list = &kcq->rq_list[sched_domain];


        bool merged;


        spin_lock(&kcq->lock);


        merged = blk_bio_list_merge(hctx->queue, rq_list, bio, nr_segs);


        spin_unlock(&kcq->lock);


        return merged;
}


When printed, the kcq structure containing the spinlock indeed looks like garbage (omitted for brevity).

A crash course on the Linux kernel block layer: for each block device, there is a "software queue" (struct blk_mq_ctx *ctx) for each CPU and a "hardware queue" (struct blk_mq_hw_ctx *hctx) for each I/O queue provided by the device. Each hardware queue has one or more software queues assigned to it. Kyber keeps additional data per hardware queue (struct kyber_hctx_data *khd) and per software queue (struct kyber_ctx_queue *kcq).

Let's try to figure out where the bad kcq came from. It should be an element of the khd->kcqs array (khd is optimized out, but we can recover it from hctx->sched_data):

>>> trace[4]["khd"]
(struct kyber_hctx_data *)<absent>
>>> hctx = trace[4]["hctx"]
>>> khd = cast("struct kyber_hctx_data *", hctx.sched_data)
>>> trace[4]["kcq"] - khd.kcqs
(ptrdiff_t)1
>>> hctx.nr_ctx
(unsigned short)1


So the kcq is for the second software queue, but the hardware queue is only supposed to have one software queue. Let's see which CPU was assigned to the hardware queue:

>>> hctx.ctxs[0].cpu
(unsigned int)6


Here's the problem: we're not running on CPU 6, we're running on CPU 19:

>>> prog["crashing_cpu"]
(int)19


And CPU 19 is assigned to a different hardware queue that actually does have two software queues:

>>> ctx = per_cpu_ptr(hctx.queue.queue_ctx, 19)
>>> other_hctx = ctx.hctxs[hctx.type]
>>> other_hctx == hctx
False
>>> other_hctx.nr_ctx
(unsigned short)2


The bug is that the caller gets the hctx for the current CPU, then kyber_bio_merge() gets the ctx for the current CPU, and if the thread is migrated to another CPU in between, they won't match. The fix is to get a consistent view of the hctx and ctx. The commit that fixes this is here.

Getting Debugging Symbols

Most Linux distributions don't install debugging symbols for installed packages by default. This page documents how to install debugging symbols on common distributions. If drgn prints an error like:

$ sudo drgn
could not get debugging information for:
kernel (could not find vmlinux for 5.14.14-200.fc34.x86_64)
...


Then you need to install debugging symbols.

Fedora

Fedora makes it very easy to install debugging symbols with the DNF debuginfo-install plugin, which is installed by default. Simply run sudo dnf debuginfo-install $package:

$ sudo dnf debuginfo-install python3


To find out what package owns a binary, use rpm -qf:

$ rpm -qf $(which python3)
python3-3.9.7-1.fc34.x86_64


To install symbols for the running kernel:

$ sudo dnf debuginfo-install kernel-$(uname -r)


Also see the Fedora documentation.

Debian

Debian requires you to manually add the debugging symbol repositories:

$ sudo tee /etc/apt/sources.list.d/debug.list << EOF
deb http://deb.debian.org/debian-debug/ $(lsb_release -cs)-debug main
deb http://deb.debian.org/debian-debug/ $(lsb_release -cs)-proposed-updates-debug main
EOF
$ sudo apt update


Then, debugging symbol packages can be installed with sudo apt install. Some debugging symbol packages are named with a -dbg suffix:

$ sudo apt install python3-dbg


And some are named with a -dbgsym suffix:

$ sudo apt install coreutils-dbgsym


You can use the find-dbgsym-packages command from the debian-goodies package to find the correct name:

$ sudo apt install debian-goodies
$ find-dbgsym-packages $(which python3)
libc6-dbg libexpat1-dbgsym python3.9-dbg zlib1g-dbgsym
$ find-dbgsym-packages $(which cat)
coreutils-dbgsym libc6-dbg


To install symbols for the running kernel:

$ sudo apt install linux-image-$(uname -r)-dbg


Also see the Debian documentation.

Ubuntu

On Ubuntu, you must install the debugging symbol archive signing key and manually add the debugging symbol repositories:

$ sudo apt update
$ sudo apt install ubuntu-dbgsym-keyring
$ sudo tee /etc/apt/sources.list.d/debug.list << EOF
deb http://ddebs.ubuntu.com $(lsb_release -cs) main restricted universe multiverse
deb http://ddebs.ubuntu.com $(lsb_release -cs)-updates main restricted universe multiverse
deb http://ddebs.ubuntu.com $(lsb_release -cs)-proposed main restricted universe multiverse
EOF
$ sudo apt update


Like Debian, some debugging symbol packages are named with a -dbg suffix and some are named with a -dbgsym suffix:

$ sudo apt install python3-dbg
$ sudo apt install coreutils-dbgsym


You can use the find-dbgsym-packages command from the debian-goodies package to find the correct name:

$ sudo apt install debian-goodies
$ find-dbgsym-packages $(which python3)
libc6-dbg libexpat1-dbgsym python3.9-dbg zlib1g-dbgsym
$ find-dbgsym-packages $(which cat)
coreutils-dbgsym libc6-dbg


To install symbols for the running kernel:

$ sudo apt install linux-image-$(uname -r)-dbgsym


Also see the Ubuntu documentation.

Arch Linux

Arch Linux unfortunately does not make debugging symbols available. Packages must be manually rebuilt with debugging symbols enabled. See the ArchWiki and the feature request.

Release Highlights

These are highlights of each release of drgn focusing on a few exciting items from the full release notes.

0.0.26 (Released March 11th, 2024)

These are some of the highlights of drgn 0.0.26. See the GitHub release for the full release notes, including more improvements and bug fixes.

Miscellaneous Helpers

This release added several new Linux kernel helpers with no particular theme:

  • print_dmesg(), a shortcut for printing the kernel log buffer.
  • idr_for_each_entry(), a shortcut for iterating over an IDR and casting its entries to a specific type.
  • stack_depot_fetch() for getting stack traces from the storage used by KASAN and other kernel debugging tools. This was contributed by Peter Collingbourne.
  • plist_head_empty(), plist_node_empty(), plist_first_entry(), plist_last_entry(), plist_for_each(), and plist_for_each_entry(), helpers for working with the kernel's priority-sorted lists.

fsrefs.py Tool

The fsrefs.py tool was added to the tools directory. It prints information about everything that is referencing a file or filesystem. This is similar to fuser(1) and lsof(8), but it can find more since it has access to kernel internals.

$ ./tools/fsrefs.py --inode /dev/urandom
pid 1349 (bluetoothd) fd 16 (struct file *)0xffff8881458cf000
pid 1368 (udisksd) fd 15 (struct file *)0xffff888145c13100
...
$ ./tools/fsrefs.py --super-block /run
mount /run (struct mount *)0xffff8881015cc140
pid 1 (systemd) fd 256 (struct file *)0xffff8881012f3d00 /run/initctl
pid 1 (systemd) fd 380 (struct file *)0xffff88810bf88800 /run/dmeventd-server
pid 1 (systemd) fd 385 (struct file *)0xffff88810bf88f00 /run/dmeventd-client
mount /run (mount namespace 4026532545) (struct mount *)0xffff8881474028c0
pid 2135770 (systemd-journal) vma 0x7f7d94f2a000-0x7f7d94f2b000 (struct file *)0xffff88813925bf00 /run/systemd/journal/kernel-seqnum
pid 2135770 (systemd-journal) vma 0x7f7d94f2b000-0x7f7d94f2c000 (struct file *)0xffff88813925a100 /run/systemd/journal/seqnum
...


fsrefs.py currently checks:

  • File descriptors
  • Task working directories
  • Task root directories
  • Memory mappings
  • Filesystem mounts
  • binfmt_misc
  • loop(4) devices
  • Swap files
  • uprobes

It will be extended to check more as the need arises, so feel free to report anything it missed.

(Note that as opposed to the contrib directory, scripts in the tools directory are regularly maintained and tested.)

DWARF Package Files

drgn now supports split DWARF package (.dwp) files. These are generated by the dwp and llvm-dwp tools.

Linux 6.8 Support

Linux 6.8 changed some filesystem internals in a way that broke a couple of drgn helpers. Here are some errors you might see with older versions of drgn that are fixed in this release.

From path_lookup() or for_each_mount() (fixed by Johannes Thumshirn):

AttributeError: 'struct mnt_namespace' has no member 'list'


From path_lookup():

AttributeError: 'struct dentry' has no member 'd_subdirs'


Python 3.13 Support

Python 3.13, currently in alpha, removed or changed some private APIs (_PyDict_GetItemIdWithError(), _PyDict_SetItemId(), and _PyLong_AsByteArray()) that drgn depended on, which caused build failures. This was fixed by using public APIs instead.

0.0.25 (Released December 1st, 2023)

These are some of the highlights of drgn 0.0.25. See the GitHub release for the full release notes, including more improvements and bug fixes.

Omitting the prog Argument

As a usability improvement, prog can now be omitted from most function calls. For example, instead of find_task(prog, 1234), you can now simply write find_task(1234). Additionally, instead of prog.stack_trace(1234), you can now write stack_trace(1234). (The old way will continue to be supported.)

Most CLI users don't need to worry about how this works, but library users may want to understand the Default Program.

It's tricky balancing interactive convenience and sensible APIs for scripting, but we think this is a nice improvement overall!

Running Without root

drgn debugs the live Linux kernel via /proc/kcore, which can only be accessed by the root user (or a user with the CAP_SYS_RAWIO capability, to be precise). However, it's not necessary (or ideal) for the rest of drgn to run as root.

Now when drgn is run against the live kernel as an unprivileged user, it will attempt to open /proc/kcore via sudo(8). The rest of drgn will then run without extra privileges.

In other words, in order to debug the live kernel, all you need to do is install debugging symbols and run:

$ drgn


This feature was contributed by Stephen Brennan.

Maple Tree Helpers

Maple trees were introduced in Linux 6.1, initially to store virtual memory areas (VMAs). This release adds a couple of helpers for working with them.

mtree_load() looks up an entry in a maple tree:

>>> mtree_load(task.mm.mm_mt.address_of_(), 0x55d65cfaa000)
(void *)0xffff97ad82bfc930


mt_for_each() iterates over a maple tree:

>>> for first_index, last_index, entry in mt_for_each(task.mm.mm_mt.address_of_()):
...     print(hex(first_index), hex(last_index), entry)
...
0x55d65cfaa000 0x55d65cfaafff (void *)0xffff97ad82bfc930
0x55d65cfab000 0x55d65cfabfff (void *)0xffff97ad82bfc0a8
0x55d65cfac000 0x55d65cfacfff (void *)0xffff97ad82bfc000
0x55d65cfad000 0x55d65cfadfff (void *)0xffff97ad82bfcb28
...


VMA Helpers

This release also adds higher-level helpers specifically for VMAs.

vma_find() looks up a VMA by address:

>>> vma_find(task.mm, 0x55d65cfaa000)
*(struct vm_area_struct *)0xffff97ad82bfc930 = {


    ...
}
>>> vma_find(task.mm, 0x55d65cfa9fff)
(struct vm_area_struct *)0


for_each_vma() iterates over every VMA in an address space:

>>> for vma in for_each_vma(task.mm):
...     print(vma)
...
*(struct vm_area_struct *)0xffff97ad82bfc930 = {


    ...
}
*(struct vm_area_struct *)0xffff97ad82bfc0a8 = {


    ...
}
...


These helpers also handle older kernels without maple trees.

Wait Queue Helpers

Wait queues are a fundamental data structure and synchronization mechanism in the Linux kernel. Imran Khan contributed a few helpers for working with them.

waitqueue_active() returns whether a wait queue has any waiters:

>>> wq
*(wait_queue_head_t *)0xffff8da80d618e18 = {


        .lock = (spinlock_t){


                .rlock = (struct raw_spinlock){


                        .raw_lock = (arch_spinlock_t){


                                .val = (atomic_t){


                                        .counter = (int)0,


                                },


                                .locked = (u8)0,


                                .pending = (u8)0,


                                .locked_pending = (u16)0,


                                .tail = (u16)0,


                        },


                },


        },


        .head = (struct list_head){


                .next = (struct list_head *)0xffffae44e3007ce8,


                .prev = (struct list_head *)0xffffae44e3007ce8,


        },
}
>>> waitqueue_active(wq)
True


waitqueue_for_each_entry() iterates over each entry in a wait queue:

>>> for entry in waitqueue_for_each_entry(wq):
...     print(entry)
...
*(wait_queue_entry_t *)0xffffae44e3007cd0 = {


        .flags = (unsigned int)0,


        .private = (void *)0xffff8da7863ec000,


        .func = (wait_queue_func_t)woken_wake_function+0x0 = 0xffffffffa8181010,


        .entry = (struct list_head){


                .next = (struct list_head *)0xffff8da80d618e20,


                .prev = (struct list_head *)0xffff8da80d618e20,


        },
}


waitqueue_for_each_task() iterates over each task waiting on a wait queue (although note that this does not work for some special wait queues that don't store tasks):

>>> for task in waitqueue_for_each_task(wq):
...     print(task.pid, task.comm)
...
(pid_t)294708 (char [16])"zsh"


ppc64 Radix MMU Support

Sourabh Jain contributed ppc64 radix MMU virtual address translation support. This is the state of architecture support in this release:

drgn 0.0.25 Architecture Support

Architecture Linux Kernel Modules Stack Traces Virtual Address Translation
x86-64 ✓ ✓ ✓
AArch64 ✓ ✓ ✓
s390x ✓ ✓ ✓
ppc64 ✓ ✓ ✓
i386 ✓
Arm ✓
RISC-V ✓

0.0.24 (Released September 8th, 2023)

These are some of the highlights of drgn 0.0.24. See the GitHub release for the full release notes, including more improvements and bug fixes.

Linked List Length Helper

This release added list_count_nodes(), which returns the length of a Linux kernel linked list:

>>> list_count_nodes(prog["workqueues"].address_of_())
29


Networking Helpers

This release added a couple of Linux kernel networking helpers requested by Jakub Kicinski.

netdev_priv() returns the private data of a network device:

>>> dev = netdev_get_by_name(prog, "wlp0s20f3")
>>> netdev_priv(dev)
(void *)0xffff9419c9dec9c0
>>> netdev_priv(dev, "struct ieee80211_sub_if_data")
*(struct ieee80211_sub_if_data *)0xffff9419c9dec9c0 = {


    ...
}


skb_shinfo() returns the shared info for a socket buffer.

C++ Lookups

This release added support for a few C++ features.

Simple Type Specifiers

Unlike C, C++ allows referring to class, struct, union, and enum types without their respective keywords. For example:

class Foo { ... };
Foo foo; // Equivalent to class Foo foo;


Previously, drgn always required the keyword, so prog.type("class Foo") would succeed but prog.type("Foo") would fail with a LookupError. This requirement was surprising to C++ developers, so it was removed. For C++ programs, prog.type("Foo") will now find a class, struct, union, or enum type named Foo (for C programs, the keyword is still required).

Nested Classes

Again unlike C, C++ allows class, struct, and union types to be defined inside of other class, struct, and union types. For example:

class Foo {
public:


  class Bar { ... };


  ...
};
Foo::Bar bar;


drgn can now find such types with prog.type("Foo::Bar").

Member Functions

C++ supports member functions (a.k.a. methods). For example:

class Foo {


  int method() { ... }
};


drgn can now find member functions with drgn.Program.function(), drgn.Program.object(), or drgn.Program[] (e.g., prog.function("Foo::method") or prog["Foo::method"]).

Split DWARF

drgn now supports split DWARF object (.dwo) files. This is enabled by the -gsplit-dwarf option in GCC and Clang or for the Linux kernel with CONFIG_DEBUG_INFO_SPLIT=y.

Split DWARF package (.dwp) file support is still in progress.

Performance Improvements

Thierry Treyer found a bug that made us search through much more debugging information than necessary when getting a stack trace. Fixing this made stack traces almost twice as fast.

The C++ lookup and split DWARF support mentioned above require processing more information in drgn's debugging information indexing step, which it does on startup and whenever debugging information is manually loaded. This could've been a performance regression, but instead, indexing was reworked from the ground up in a way that's usually faster despite the added features.

0.0.23 (Released June 28th, 2023)

These are some of the highlights of drgn 0.0.23. See the GitHub release for the full release notes, including more improvements and bug fixes.

Virtual Address Translation Helpers

This release added several Linux kernel helpers for translating virtual addresses.

follow_phys() translates a virtual address to a physical address in a given address space. For example, to get the physical address that virtual address 0x7f7fe46a4270 maps to in process 115:

>>> task = find_task(prog, 115)
>>> address = 0x7f7fe46a4270
>>> print(hex(follow_phys(task.mm, address)))
0x4090270


follow_page() translates a virtual address to the struct page * that it maps to:

>>> follow_page(task.mm, address)
*(struct page *)0xffffd20ac0102400 = {


    ...
}


follow_pfn() translates a virtual address to the page frame number (PFN) of the page that it maps to:

>>> follow_pfn(task.mm, address)
(unsigned long)16528


These can be used to translate arbitrary kernel virtual addresses by passing prog["init_mm"].address_of_():

>>> print(hex(follow_phys(prog["init_mm"].address_of_(), 0xffffffffc0483000)))
0x2e4b000


Vmalloc/Vmap Address Translation Helpers

vmalloc_to_page() is a special case of follow_page() for vmalloc and vmap addresses:

>>> vmalloc_to_page(prog, 0xffffffffc0477000)
*(struct page *)0xffffc902400b8980 = {


    ...
}


Likewise, vmalloc_to_pfn() is a special case of follow_pfn() for vmalloc and vmap addresses:

>>> vmalloc_to_pfn(prog, 0xffffffffc0477000)
(unsigned long)11814


contrib Directory

Martin Liška, Boris Burkov, and Johannes Thumshirn added lots of new scripts to the contrib directory:

  • btrfs_tree.py: work-in-progress helpers for Btrfs B-trees
  • btrfs_tree_mod_log.py: simulator for the Btrfs tree modification log
  • dump_btrfs_bgs.py: print block groups in a Btrfs filesystem
  • kcore_list.py: print memory regions from /proc/kcore
  • kernel_sys.py: print system information (similar to crash's sys command)
  • mount.py: print a filesystem mount table
  • platform_drivers.py: print registered platform drivers
  • vmmap.py: print memory mappings in a process (similar to /proc/$pid/maps)
  • vmstat.py: print information about kernel memory usage

Embedding Interactive Mode

drgn.cli.run_interactive() runs drgn's interactive mode. It can be used to embed drgn in another application. For example, you could use it for a custom drgn.Program that the standard drgn CLI can't set up:

import drgn
import drgn.cli
prog = drgn.Program()
prog.add_type_finder(...)
prog.add_object_finder(...)
prog.add_memory_segment(...)
drgn.cli.run_interactive(prog)


Full s390x Support

Sven Schnelle contributed s390x virtual address translation support. This is the state of architecture support in this release:

drgn 0.0.23 Architecture Support

Architecture Linux Kernel Modules Stack Traces Virtual Address Translation
x86-64 ✓ ✓ ✓
AArch64 ✓ ✓ ✓
ppc64 ✓ ✓
s390x ✓ ✓ ✓
i386 ✓
Arm ✓
RISC-V ✓

Linux 6.3 & 6.4 Support

Linux 6.3 and 6.4 had an unusual number of breaking changes for drgn. Here are some errors you might see with older versions of drgn that are fixed in this release.

On startup (fixed by Ido Schimmel):

warning: could not get debugging information for:
kernel modules (could not find loaded kernel modules: 'struct module' has no member 'core_size')


From drgn.Program.stack_trace() and drgn.Thread.stack_trace():

Exception: unknown ORC entry type 3


From compound_order() and compound_nr():

AttributeError: 'struct page' has no member 'compound_order'


From for_each_disk() and for_each_partition():

AttributeError: 'struct class' has no member 'p'


Python 3.12 Support

Python 3.12, currently in beta, changed an implementation detail that drgn depended on, which caused crashes like:

Py_SIZE: Assertion `ob->ob_type != &PyLong_Type' failed.


Stephen Brennan fixed this.

0.0.22 (Released January 5th, 2023)

These are some of the highlights of drgn 0.0.22. See the GitHub release for the full release notes, including more improvements and bug fixes.

Listing Stack Frame Locals

drgn.StackFrame.locals() returns the names of all arguments and local variables in the scope of a stack frame. This allows you to get a quick idea of what's going on in a function without needing to read the source code right away.

Let's use the __schedule stack frame from the following trace as an example:

>>> trace = prog.stack_trace(1)
>>> trace
#0  context_switch (./kernel/sched/core.c:5209:2)
#1  __schedule (./kernel/sched/core.c:6521:8)
#2  schedule (./kernel/sched/core.c:6597:3)
#3  do_wait (./kernel/exit.c:1562:4)
#4  kernel_wait4 (./kernel/exit.c:1706:8)
#5  __do_sys_wait4 (./kernel/exit.c:1734:13)
#6  do_syscall_x64 (./arch/x86/entry/common.c:50:14)
#7  do_syscall_64 (./arch/x86/entry/common.c:80:7)
#8  entry_SYSCALL_64+0x9b/0x197 (./arch/x86/entry/entry_64.S:120)
#9  0x7f6a34a00057
>>> trace[1].locals()
['sched_mode', 'prev', 'next', 'switch_count', 'prev_state', 'rf', 'rq', 'cpu']
>>> for name in trace[1].locals():
...     print(name, trace[1][name].format_(dereference=False))
...
sched_mode (unsigned int)0
prev (struct task_struct *)0xffffa3b601178000
next (struct task_struct *)0xffffa3b6026db800
switch_count (unsigned long *)0xffffa3b601178528
prev_state (unsigned long)<absent>
rf (struct rq_flags){


        .flags = (unsigned long)1,


        .cookie = (struct pin_cookie){},


        .clock_update_flags = (unsigned int)4,
}
rq (struct rq *)0xffffa3b67fda9640
cpu (int)<absent>


Compare this to the kernel source code. Note that some of the variables have been optimized out by the compiler.

This feature was contributed by Stephen Brennan.

Merged Slab Caches

The Linux kernel slab allocator merges "similar" slab caches as an optimization, which often causes confusion. slab_cache_is_merged() (added back in 0.0.20) returns whether or not a slab cache has been merged, but not what it was merged with. In this release, Stephen Brennan added get_slab_cache_aliases(), which provides a mapping from a slab cache name to the name of the cache it was merged into:

>>> get_slab_cache_aliases(prog)
{'io_kiocb': 'maple_node', 'ip_dst_cache': 'uid_cache', 'aio_kiocb': 'uid_cache', 'ip_fib_alias': 'Acpi-Parse', 'pid_namespace': 'pid', 'iommu_iova': 'vmap_area', 'fasync_cache': 'ftrace_event_field', 'dnotify_mark': 'Acpi-State', 'tcp_bind2_bucket': 'vmap_area', 'nsproxy': 'Acpi-Operand', 'shared_policy_node': 'ftrace_event_field', 'eventpoll_epi': 'pid', 'fib6_nodes': 'vmap_area', 'Acpi-Namespace': 'ftrace_event_field', 'posix_timers_cache': 'maple_node', 'inotify_inode_mark': 'Acpi-State', 'kernfs_iattrs_cache': 'trace_event_file', 'fs_cache': 'vmap_area', 'UDP-Lite': 'UDP', 'anon_vma_chain': 'vmap_area', 'ip6_dst_cache': 'maple_node', 'eventpoll_pwq': 'vmap_area', 'inet_peer_cache': 'uid_cache', 'fsnotify_mark_connector': 'numa_policy', 'ip_fib_trie': 'ftrace_event_field', 'filp': 'maple_node', 'dnotify_struct': 'numa_policy', 'UDPLITEv6': 'UDPv6', 'biovec-16': 'maple_node', 'PING': 'signal_cache', 'ep_head': 'blkdev_ioc', 'tcp_bind_bucket': 'pid', 'Acpi-ParseExt': 'Acpi-State', 'cred_jar': 'pid', 'ovl_aio_req': 'pid', 'pool_workqueue': 'maple_node', 'sigqueue': 'Acpi-State', 'file_lock_ctx': 'Acpi-Parse', 'kernfs_node_cache': 'pid'}


This means that if you're looking for io_kiocb allocations, you actually need to look at the maple_node slab cache. Conversely, if you're looking at the maple_node slab cache, you need to be aware that it also contains allocations from all of the following slab caches:

>>> [merged for merged, canonical in get_slab_cache_aliases(prog).items() if canonical == "maple_node"]
['io_kiocb', 'posix_timers_cache', 'ip6_dst_cache', 'filp', 'biovec-16', 'pool_workqueue']


Slab Address Information

This release extended identify_address() to show additional information about slab allocations:

>>> ptr1 = 0xffffa3b601178438
>>> ptr2 = 0xffffa3b601176cc0
>>> identify_address(prog, ptr1)
'slab object: task_struct+0x438'
>>> identify_address(prog, ptr2)
'free slab object: mm_struct+0x0'


This means that ptr1 is an address 0x438 bytes into an allocated object from the task_struct slab cache, and ptr2 is a free object from the mm_struct slab cache.

slab_object_info() provides the same information programmatically:

>>> slab_object_info(prog, ptr1)
SlabObjectInfo(slab_cache=Object(prog, 'struct kmem_cache *', value=0xffffa3b601045500), slab=Object(prog, 'struct slab *', value=0xffffe80840045e00), address=0xffffa3b601178000, allocated=True)
>>> slab_object_info(prog, ptr2)
SlabObjectInfo(slab_cache=Object(prog, 'struct kmem_cache *', value=0xffffa3b601045900), slab=Object(prog, 'struct slab *', value=0xffffe80840045c00), address=0xffffa3b601176cc0, allocated=False)


Annotated Stack Memory

print_annotated_stack() prints a stack trace and all of its memory, identifying anything that it can:

>>> print_annotated_stack(prog.stack_trace(1))
STACK POINTER     VALUE
[stack frame #0 at 0xffffffffaf8a68e9 (__schedule+0x429/0x488) in context_switch at ./kernel/sched/core.c:5209:2 (inlined)]
[stack frame #1 at 0xffffffffaf8a68e9 (__schedule+0x429/0x488) in __schedule at ./kernel/sched/core.c:6521:8]
ffffbb1ac0013d28: ffffffffaf4498f5 [function symbol: __flush_tlb_one_user+0x5]
ffffbb1ac0013d30: 00000000af449feb
ffffbb1ac0013d38: 0000000000000001
ffffbb1ac0013d40: 0000000000000004
ffffbb1ac0013d48: 25c5ff9539edc200
ffffbb1ac0013d50: ffffa3b601178000 [slab object: task_struct+0x0]
ffffbb1ac0013d58: ffffa3b601178000 [slab object: task_struct+0x0]
ffffbb1ac0013d60: ffffbb1ac0013e10
ffffbb1ac0013d68: ffffa3b601177ff0 [slab object: mm_struct+0x70]
ffffbb1ac0013d70: ffffa3b601178000 [slab object: task_struct+0x0]
ffffbb1ac0013d78: ffffa3b601178000 [slab object: task_struct+0x0]
ffffbb1ac0013d80: ffffffffaf8a69d1 [function symbol: schedule+0x89]
[stack frame #2 at 0xffffffffaf8a69d1 (schedule+0x89/0xc7) in schedule at ./kernel/sched/core.c:6597:3]
ffffbb1ac0013d88: ffffbb1ac0013de8
ffffbb1ac0013d90: 0000000000000000
ffffbb1ac0013d98: ffffffffaf4595ee [function symbol: do_wait+0x231]
[stack frame #3 at 0xffffffffaf4595ee (do_wait+0x231/0x2e3) in do_wait at ./kernel/exit.c:1562:4]
ffffbb1ac0013da0: ffffa3b601178450 [slab object: task_struct+0x450]
ffffbb1ac0013da8: ffffa3b601178000 [slab object: task_struct+0x0]
ffffbb1ac0013db0: 0000000000000004
ffffbb1ac0013db8: 0000000000000000
ffffbb1ac0013dc0: 00007ffe0984a170
ffffbb1ac0013dc8: 0000000000000000
ffffbb1ac0013dd0: fffffffffffffffd
ffffbb1ac0013dd8: 0000000000000004
ffffbb1ac0013de0: ffffffffaf45a42f [function symbol: kernel_wait4+0xc2]
[stack frame #4 at 0xffffffffaf45a42f (kernel_wait4+0xc2/0x11b) in kernel_wait4 at ./kernel/exit.c:1706:8]
ffffbb1ac0013de8: 0000000400000004
ffffbb1ac0013df0: 0000000000000000
ffffbb1ac0013df8: 0000000000000000
ffffbb1ac0013e00: 0000000000000000
ffffbb1ac0013e08: 0000000000000000
ffffbb1ac0013e10: ffffffff00000000
ffffbb1ac0013e18: ffffa3b601178000 [slab object: task_struct+0x0]
ffffbb1ac0013e20: ffffffffaf45890c [function symbol: child_wait_callback+0x0]
ffffbb1ac0013e28: ffffa3b601188028 [slab object: signal_cache+0x28]
ffffbb1ac0013e30: ffffa3b601188028 [slab object: signal_cache+0x28]
ffffbb1ac0013e38: 000055d500000000
ffffbb1ac0013e40: 25c5ff9539edc200
ffffbb1ac0013e48: 0000000000000000
ffffbb1ac0013e50: ffffbb1ac0013f30
ffffbb1ac0013e58: ffffbb1ac0013f58
ffffbb1ac0013e60: 0000000000000000
ffffbb1ac0013e68: 0000000000000000
ffffbb1ac0013e70: 0000000000000000
ffffbb1ac0013e78: ffffffffaf45a4c0 [function symbol: __do_sys_wait4+0x38]
[stack frame #5 at 0xffffffffaf45a4c0 (__do_sys_wait4+0x38/0x8c) in __do_sys_wait4 at ./kernel/exit.c:1734:13]
ffffbb1ac0013e80: ffffffffaf8aaa21 [function symbol: _raw_spin_unlock_irq+0x10]
ffffbb1ac0013e88: ffffffffaf46460c [function symbol: do_sigaction+0xf8]
ffffbb1ac0013e90: ffffa3b601180020 [slab object: sighand_cache+0x20]
ffffbb1ac0013e98: ffffa3b6028d02d0 [slab object: vm_area_struct+0x0]
ffffbb1ac0013ea0: 25c5ff9539edc200
ffffbb1ac0013ea8: 0000000000000002
ffffbb1ac0013eb0: 00007ffe09849fb0
ffffbb1ac0013eb8: ffffbb1ac0013f58
ffffbb1ac0013ec0: 0000000000000000
ffffbb1ac0013ec8: 0000000000000000
ffffbb1ac0013ed0: 0000000000000046
ffffbb1ac0013ed8: ffffa3b601178000 [slab object: task_struct+0x0]
ffffbb1ac0013ee0: ffffa3b601178000 [slab object: task_struct+0x0]
ffffbb1ac0013ee8: ffffbb1ac0013f58
ffffbb1ac0013ef0: 0000000000000000
ffffbb1ac0013ef8: ffffffffaf426def [function symbol: fpregs_assert_state_consistent+0x1b]
ffffbb1ac0013f00: 0000000000000000
ffffbb1ac0013f08: ffffffffaf4b2f53 [function symbol: exit_to_user_mode_prepare+0xa6]
ffffbb1ac0013f10: 0000000000000000
ffffbb1ac0013f18: 25c5ff9539edc200
ffffbb1ac0013f20: ffffbb1ac0013f58
ffffbb1ac0013f28: 0000000000000000
ffffbb1ac0013f30: ffffbb1ac0013f48
ffffbb1ac0013f38: ffffffffaf8a1573 [function symbol: do_syscall_64+0x70]
[stack frame #6 at 0xffffffffaf8a1573 (do_syscall_64+0x70/0x8a) in do_syscall_x64 at ./arch/x86/entry/common.c:50:14 (inlined)]
[stack frame #7 at 0xffffffffaf8a1573 (do_syscall_64+0x70/0x8a) in do_syscall_64 at ./arch/x86/entry/common.c:80:7]
ffffbb1ac0013f40: 0000000000000000
ffffbb1ac0013f48: 0000000000000000
ffffbb1ac0013f50: ffffffffafa0009b [symbol: entry_SYSCALL_64+0x9b]
[stack frame #8 at 0xffffffffafa0009b (entry_SYSCALL_64+0x9b/0x197) at ./arch/x86/entry/entry_64.S:120]
ffffbb1ac0013f58: 0000000000000000
[stack frame #9 at 0x7f6a34a00057]


Like drgn.StackFrame.locals(), this provides a nice overview of everything happening in a function, which might include useful hints. Keep in mind that it may identify "stale" addresses for anything that a function hasn't reinitialized yet, and as always, be careful of slab cache merging.

This was inspired by the crash bt -FF command. It was contributed by Nhat Pham.

XArray Helpers

XArrays were introduced in Linux 4.20 as a replacement for radix trees. drgn's radix tree helpers also support XArrays in some cases, but this is awkward, not obvious, and doesn't work for new, XArray-only functionality.

This release added dedicated XArray helpers like xa_load() and xa_for_each().

s390x Support

Sven Schnelle contributed s390x support for Linux kernel modules and stack traces. This is the state of architecture support in this release:

drgn 0.0.22 Architecture Support

Architecture Linux Kernel Modules Stack Traces Virtual Address Translation
x86-64 ✓ ✓ ✓
AArch64 ✓ ✓ ✓
ppc64 ✓ ✓
s390x ✓ ✓
i386 ✓
Arm ✓
RISC-V ✓

Relicensing to LGPL

drgn was originally licensed as GPLv3+. In this release, it was changed to LGPLv2.1+. The motivation for this change was to enable the long term vision for drgn that more projects can use it as a library providing programmatic interfaces for debugger functionality. For example, Object Introspection, a userspace memory profiler recently open sourced by Meta, uses drgn to parse debugging information.

AUTHOR

Omar Sandoval

COPYRIGHT

Omar Sandoval

March 12, 2024 0.0.26