Introduction

Lab objectives

• presenting the rules and objectives of the Operating Systems 2 lab
• introducing the lab documentation
• introducing the Linux kernel and related resources

Keywords

• kernel, kernel programming
• Linux, vanilla, http://www.kernel.org
• cscope, LXR
• gdb, /proc/kcore, addr2line, dump_stack

About this laboratory

The Operating Systems 2 lab is a kernel programming and driver development lab. The objectives of the laboratory are:

• deepening the notions presented in the course
• presentation of kernel programming interfaces (kernel API)
• gaining documenting, development and debugging skills on a freestanding environment
• acquiring knowledge and skills for drivers development

A laboratory will present a set of concepts, applications and commands specific to a given problem. The lab will start with a presentation (each lab will have a set of slides) (15 minutes) and the remaining time will be allocated to the lab exercises (80 minutes).

For best laboratory performance, we recommend that you read the related slides. To fully understand a laboratory, we recommend going through the lab support. For in-depth study, use the supporting documentation.

References

• Linux
  • Linux Kernel Development, 3rd Edition
  • Linux Device Drivers, 3rd Edition
  • Essential Linux Device Drivers
• General
  • mailing list (searching the mailing list)

Source code navigation

cscope

Cscope is a tool for efficient navigation of C sources. To use it, a cscope database must be generated from the existing sources. In a Linux tree, the command make ARCH=x86 cscope is sufficient. Specification of the architecture through the ARCH variable is optional but recommended; otherwise, some architecture dependent functions will appear multiple times in the database.

You can build the cscope database with the command make ARCH=x86 COMPILED_SOURCE=1 cscope. This way, the cscope database will only contain symbols that have already been used in the compile process before, thus resulting in better performance when searching for symbols.

Cscope can also be used as stand-alone, but it is more useful when combined with an editor. To use cscope with vim, it is necessary to install both packages and add the following lines to the file .vimrc (the machine in the lab already has the settings):

if has("cscope")
        " Look for a 'cscope.out' file starting from the current directory,
        " going up to the root directory.
        let s:dirs = split(getcwd(), "/")
        while s:dirs != []
                let s:path = "/" . join(s:dirs, "/")
                if (filereadable(s:path . "/cscope.out"))
                        execute "cs add " . s:path . "/cscope.out " . s:path . " -v"
                        break
                endif
                let s:dirs = s:dirs[:-2]
        endwhile

        set csto=0  " Use cscope first, then ctags
        set cst     " Only search cscope
        set csverb  " Make cs verbose

        nmap `<C-\>`s :cs find s `<C-R>`=expand("`<cword>`")`<CR>``<CR>`
        nmap `<C-\>`g :cs find g `<C-R>`=expand("`<cword>`")`<CR>``<CR>`
        nmap `<C-\>`c :cs find c `<C-R>`=expand("`<cword>`")`<CR>``<CR>`
        nmap `<C-\>`t :cs find t `<C-R>`=expand("`<cword>`")`<CR>``<CR>`
        nmap `<C-\>`e :cs find e `<C-R>`=expand("`<cword>`")`<CR>``<CR>`
        nmap `<C-\>`f :cs find f `<C-R>`=expand("`<cfile>`")`<CR>``<CR>`
        nmap `<C-\>`i :cs find i ^`<C-R>`=expand("`<cfile>`")`<CR>`$`<CR>`
        nmap `<C-\>`d :cs find d `<C-R>`=expand("`<cword>`")`<CR>``<CR>`
        nmap <F6> :cnext <CR>
        nmap <F5> :cprev <CR>

        " Open a quickfix window for the following queries.
        set cscopequickfix=s-,c-,d-,i-,t-,e-,g-
endif

The script searches for a file called cscope.out in the current directory, or in parent directories. If vim finds this file, you can use the shortcut Ctrl +] or Ctrl+\ g (the combination control-\ followed by g) to jump directly to the definition of the word under the cursor (function, variable, structure, etc.). Similarly, you can use Ctrl+\ s to go where the word under the cursor is used.

You can take a cscope-enabled .vimrc file (also contains other goodies) from https://github.com/ddvlad/cfg/blob/master/_vimrc. The following guidelines are based on this file, but also show basic vim commands that have the same effect.

If there are more than one results (usually there are) you can move between them using F6 and F5 (:ccnext and :cprev). You can also open a new panel showing the results using :copen. To close the panel, use the :cclose command.

To return to the previous location, use Ctrl+o (o, not zero). The command can be used multiple times and works even if cscope changed the file you are currently editing.

To go to a symbol definition directly when vim starts, use vim -t <symbol_name> (for example vim -t task_struct). Otherwise, if you started vim and want to search for a symbol by name, use cs find g <symbol_name> (for example cs find g task_struct).

If you found more than one results and opened a panel showing all the matches (using :copen) and you want to find a symbol of type structure, it is recommended to search in the results panel (using / -- slash) the character { (opening brace).

Important

You can get a summary of all the cscope commands using :cs help.

For more info, use the vim built-in help command: :h cscope or :h copen.

If you use emacs, install the xcscope-el package and add the following lines in ~/.emacs.

(require ‘xcscope)
(cscope-setup)

These commands will activate cscope for the C and C++ modes automatically. C-s s is the key bindings prefix and C-s s s is used to search for a symbol (if you call it when the cursor is over a word, it will use that). For more details, check https://github.com/dkogan/xcscope.el

clangd

Clangd is a language server that provides tools for navigating C and C++ code. Language Server Protocol facilitates features like go-to-definition, find-references, hover, completion, etc., using semantic whole project analysis.

Clangd requires a compilation database to understand the kernel source code. It can be generated with:

make defconfig
make
scripts/clang-tools/gen_compile_commands.py

LSP clients:

• Vim/Neovim (coc.nvim, nvim-lsp, vim-lsc, vim-lsp)
• Emacs (lsp-mode)
• VSCode (clangd extension)

Kscope

For a simpler interface, Kscope is a cscope frontend which uses QT. It is lightweight, very fast and very easy to use. It allows searching using regular expressions, call graphs, etc. Kscope is no longer mantained.

There is also a port of version 1.6 for Qt4 and KDE 4 which keeps the integration of the text editor Kate and is easier to use than the last version on SourceForge.

LXR Cross-Reference

LXR (LXR Cross-Reference) is a tool that allows indexing and referencing the symbols in the source code of a program using a web interface. The web interface shows links to locations in files where a symbol is defined or used. Development website for LXR is http://sourceforge.net/projects/lxr. Similar tools are OpenGrok and Gonzui.

Although LXR was originally intended for the Linux kernel sources, it is also used in the sources of Mozilla, Apache HTTP Server and FreeBSD.

There are a number of sites that use LXR for cross-referencing the the sources of the Linux kernel, the main site being the original site of development which does not work anymore. You can use https://elixir.bootlin.com/.

LXR allows searching for an identifier (symbol), after a free text or after a file name. The main feature and, at the same time, the main advantage provided is the ease of finding the declaration of any global identifier. This way, it facilitates quick access to function declarations, variables, macro definitions and the code can be easily navigated. Also, the fact that it can detect what code areas are affected when a variable or function is changed is a real advantage in the development and debugging phase.

SourceWeb

SourceWeb is a source code indexer for C and C++. It uses the framework provided by the Clang compiler to index the code.

The main difference between cscope and SourceWeb is the fact that SourceWeb is, in a way, a compiler pass. SourceWeb doesn't index all the code, but only the code that was efectively compiled by the compiler. This way, some problems are eliminated, such as ambiguities about which variant of a function defined in multiple places is used. This also means that the indexing takes more time, because the compiled files must pass one more time through the indexer to generate the references.

Usage example:

make oldconfig
sw-btrace make -j4
sw-btrace-to-compile-db
sw-clang-indexer --index-project
sourceweb index

sw-btrace is a script that adds the libsw-btrace.so library to LD_PRELOAD. This way, the library is loaded by every process started by make (basically, the compiler), registers the commands used to start the processes and generates a filed called btrace.log. This file is then used by sw-btrace-to-compile-db which converts it to a format defined by clang: JSON Compilation Database. This JSON Compilation Database resulted from the above steps is then used by the indexer, which makes one more pass through the compiled source files and generates the index used by the GUI.

Word of advice: don't index the sources you are working with, but use a copy, because SourceWeb doesn't have, at this moment, the capability to regenerate the index for a single file and you will have to regenerate the complete index.

Kernel Debugging

Debugging a kernel is a much more difficult process than the debugging of a program, because there is no support from the operating system. This is why this process is usually done using two computers, connected on serial interfaces.

gdb (Linux)

A simpler debug method on Linux, but with many disadvantages, is local debugging, using gdb, the uncompressed kernel image (vmlinux) and /proc/kcore (the real-time kernel image). This method is usually used to inspect the kernel and detect certain inconsistencies while it runs. The method is useful especially if the kernel was compiled using the -g option, which keeps debug information. Some well-known debug techniques can't be used by this method, such as breakpoints of data modification.

Note

Because /proc is a virtual filesystem, /proc/kcore does not physically exist on the disk. It is generated on-the-fly by the kernel when a program tries to access proc/kcore.

It is used for debugging purposes.

From man proc, we have:

/proc/kcore
This file represents the physical memory of the system and is stored in the ELF core file format.  With this pseudo-file, and
an unstripped kernel (/usr/src/linux/vmlinux) binary, GDB can be used to examine the current state of any kernel data struc‐
tures.

The uncompressed kernel image offers information about the data structures and symbols it contains.

student@eg106$ cd ~/src/linux
student@eg106$ file vmlinux
vmlinux: ELF 32-bit LSB executable, Intel 80386, ...
student@eg106$ nm vmlinux | grep sys_call_table
c02e535c R sys_call_table
student@eg106$ cat System.map | grep sys_call_table
c02e535c R sys_call_table

The nm utility is used to show the symbols in an object or executable file. In our case, vmlinux is an ELF file. Alternately, we can use the file System.map to view information about the symbols in kernel.

Then we use gdb to inspect the symbols using the uncompressed kernel image. A simple gdb session is the following:

student@eg106$ cd ~/src/linux
stduent@eg106$ gdb --quiet vmlinux
Using host libthread_db library "/lib/tls/libthread_db.so.1".
(gdb) x/x 0xc02e535c
0xc02e535c `<sys_call_table>`:    0xc011bc58
(gdb) x/16 0xc02e535c
0xc02e535c `<sys_call_table>`:    0xc011bc58      0xc011482a      0xc01013d3     0xc014363d
0xc02e536c `<sys_call_table+16>`: 0xc014369f      0xc0142d4e      0xc0142de5     0xc011548b
0xc02e537c `<sys_call_table+32>`: 0xc0142d7d      0xc01507a1      0xc015042c     0xc0101431
0xc02e538c `<sys_call_table+48>`: 0xc014249e      0xc0115c6c      0xc014fee7     0xc0142725
(gdb) x/x sys_call_table
0xc011bc58 `<sys_restart_syscall>`:       0xffe000ba
(gdb) x/x &sys_call_table
0xc02e535c `<sys_call_table>`:    0xc011bc58
(gdb) x/16 &sys_call_table
0xc02e535c `<sys_call_table>`:    0xc011bc58      0xc011482a      0xc01013d3     0xc014363d
0xc02e536c `<sys_call_table+16>`: 0xc014369f      0xc0142d4e      0xc0142de5     0xc011548b
0xc02e537c `<sys_call_table+32>`: 0xc0142d7d      0xc01507a1      0xc015042c     0xc0101431
0xc02e538c `<sys_call_table+48>`: 0xc014249e      0xc0115c6c      0xc014fee7     0xc0142725
(gdb) x/x sys_fork
0xc01013d3 `<sys_fork>`:  0x3824548b
(gdb) disass sys_fork
Dump of assembler code for function sys_fork:
0xc01013d3 `<sys_fork+0>`:        mov    0x38(%esp),%edx
0xc01013d7 `<sys_fork+4>`:        mov    $0x11,%eax
0xc01013dc `<sys_fork+9>`:        push   $0x0
0xc01013de `<sys_fork+11>`:       push   $0x0
0xc01013e0 `<sys_fork+13>`:       push   $0x0
0xc01013e2 `<sys_fork+15>`:       lea    0x10(%esp),%ecx
0xc01013e6 `<sys_fork+19>`:       call   0xc0111aab `<do_fork>`
0xc01013eb `<sys_fork+24>`:       add    $0xc,%esp
0xc01013ee `<sys_fork+27>`:       ret
End of assembler dump.

It can be noticed that the uncompressed kernel image was used as an argument for gdb. The image can be found in the root of the kernel sources after compilation.

A few commands used for debugging using gdb are:

• x (examine) - Used to show the contents of the memory area whose address is specified as an argument to the command (this address can be the value of a physical address, a symbol or the address of a symbol). It can take as arguments (preceded by /): the format to display the data in (x for hexadecimal, d for decimal, etc.), how many memory units to display and the size of a memory unit.
• disassemble - Used to disassemble a function.
• p (print) - Used to evaluate and show the value of an expression. The format to show the data in can be specified as an argument (/x for hexadecimal, /d for decimal, etc.).

The analysis of the kernel image is a method of static analysis. If we want to perform dynamic analysis (analyzing how the kernel runs, not only its static image) we can use /proc/kcore; this is a dynamic image (in memory) of the kernel.

student@eg106$ gdb ~/src/linux/vmlinux /proc/kcore
Core was generated by `root=/dev/hda3 ro'.
#0  0x00000000 in ?? ()
(gdb) p sys_call_table
$1 = -1072579496
(gdb) p /x sys_call_table
$2 = 0xc011bc58
(gdb) p /x &sys_call_table
$3 = 0xc02e535c
(gdb) x/16 &sys_call_table
0xc02e535c `<sys_call_table>`:    0xc011bc58      0xc011482a      0xc01013d3     0xc014363d
0xc02e536c `<sys_call_table+16>`: 0xc014369f      0xc0142d4e      0xc0142de5     0xc011548b
0xc02e537c `<sys_call_table+32>`: 0xc0142d7d      0xc01507a1      0xc015042c     0xc0101431
0xc02e538c `<sys_call_table+48>`: 0xc014249e      0xc0115c6c      0xc014fee7     0xc0142725

Using the dynamic image of the kernel is useful for detecting rootkits.

• Linux Device Drivers 3rd Edition - Debuggers and Related Tools
• Detecting Rootkits and Kernel-level Compromises in Linux
• User-Mode Linux

Getting a stack trace

Sometimes, you will want information about the trace the execution reaches a certain point. You can determine this information using cscope or LXR, but some function are called from many execution paths, which makes this method difficult.

In these situations, it is useful to get a stack trace, which can be simply done using the function dump_stack().

Documentation

Kernel development is a difficult process, compared to user space programming. The API is different and the complexity of the subsystems in kernel requires additional preparation. The associated documentation is heterogeneous, sometimes requiring the inspection of multiple sources to have a more complete understanding of a certain aspect.

The main advantages of the Linux kernel are the access to sources and the open development system. Because of this, the Internet offers a larger number of documentation for the kernel.

A few links related to the Linux kernel are shown bellow:

• KernelNewbies
• KernelNewbies - Kernel Hacking
• Kernel Analysis - HOWTO
• Linux Kernel Programming
• Linux kernel - Wikibooks

The links are not comprehensive. Using The Internet and kernel source code is essential.

Exercises

Remarks

Note

• Usually, the steps used to develop a kernel module are the following:
  • editing the module source code (on the physical machine);
  • module compilation (on the physical machine);
  • generation of the minimal image for the virtual machine; this image contains the kernel, your module, busybox and eventually test programs;
  • starting the virtual machine using QEMU;
  • running the tests in the virtual machine.
• When using cscope, use ~/src/linux. If there is no cscope.out file, you can generate it using the command make ARCH=x86 cscope.
• You can find more details about the virtual machine at Recommended Setup.

Important

Before solving an exercice, carefully read all its bullets.

Booting the virtual machine

A summary of the virtual machine infrastructure:

• ~/src/linux - Linux kernel sources, needed to compile modules. The directory contains the file cscope.out, used for navigation in the source tree.
• ~/src/linux/tools/labs/qemu- scripts and auxiliary files used to generate and run the QEMU VM.

To start the VM, run make boot in the directory ~/src/linux/tools/labs:

student@eg106:~$ cd ~/src/linux/tools/labs
student@eg106:~/src/linux/tools/labs$ make boot

By default, you will not get a prompt or any graphical interface, but you can connect to a console exposed by the virtual machine using minicom or screen.

student@eg106:~/src/linux/tools/labs$ minicom -D serial.pts

<press enter>

qemux86 login:
Poky (Yocto Project Reference Distro) 2.3 qemux86 /dev/hvc0

Alternatively, you can start the virtual machine with graphical interface support, using the QEMU_DISPLAY=gtk make boot.

Note

To access the virtual machine, at the login prompt, enter the username root; there is no need to enter a password. The virtual machine will start with the permissions of the root account.

Adding and using a virtual disk

Note

If you don't have the file mydisk.img, you can download it from the address http://elf.cs.pub.ro/so2/res/laboratoare/mydisk.img. The file must be placed in tools/labs.

In the ~/src/linux/tools/labs directory, you have a new virtual machine disk, in the file mydisk.img. We want to add the disk to the virtual machine and use it within the virtual machine.

Edit qemu/Makefile and add -drive file=mydisk.img,if=virtio,format=raw to the QEMU_OPTS variable.

Note

There are already two disks added to qemu (disk1.img and disk2.img). You will need to add the new one after them. In this case, the new disk can be accessed as /dev/vdd (vda is the root partition, vdb is disk1 and vdc is disk2).

Hint

You do not need to manually create the entry for the new disk in /dev because the virtual machine uses devtmpfs.

Run make in tools/labs to boot the virtual machine. Create /test directory and try to mount the new disk:

mkdir /test
mount /dev/vdd /test

The reason why we can not mount the virtual disk is because we do not have support in the kernel for the filesystem with which the mydisk.img is formatted. You will need to identify the filesystem for mydisk.img and compile kernel support for that filesystem.

Close the virtual machine (close the QEMU window, you do not need to use another command). Use the file command on the physical machine to find out with which filesystem the mydisk.img file is formatted. You will identify the btrfs file system.

You will need to enable btrfs support in the kernel and recompile the kernel image.

Warning

If you receive an error while executing the make menuconfig command, you probably do not have the libncurses5-dev package installed. Install it using the command:

sudo apt-get install libncurses5-dev

Hint

Enter the ~/src/linux/ subdirectory. Run make menuconfig and go to the File systems section. Enable Btrfs filesystem support. You will need to use the builtin option (not the module), i.e. <*> must appear next to the option (not <M>).

Save the configuration you have made. Use the default configuration file (config).

In the kernel source subdirectory (~/src/linux/) recompile using the command:

make

To wait less, you can use the -j option run multiple jobs in parallel. Generally, it is recommended to use number of CPUs+1:

make -j5

After the kernel recompilation finishes, restart the QEMU virtual machine: that is, launch the make command in the subdirectory. You do not need to copy anything, because the bzImage file is a symlink to the kernel image you just recompiled.

Inside the QEMU virtual machine, repeat the mkdir and mount operations. With support for the btrfs filesystem, now mount will finish successfully.

Note

When doing your homework, there is no need to recompile the kernel because you will only use kernel modules. However, it is important to be familiar with configuring and recompiling a kernel.

If you still plan to recompile the kernel, make a backup of the bzImage file (follow the link in ~/src/linux for the full path). This will allow you to return to the initial setup in order to have an environment identical to the one used by vmchecker.

GDB and QEMU

We can investigate and troubleshoot the QEMU virtual machine in real time.

Note

You can also use the GDB Dashboard plugin for a user-friendly interface. gdb must be compiled with Python support.

In order to install it, you can just run:

wget -P ~ git.io/.gdbinit

To do this, we start the QEMU virtual machine first. Then, we can connect with gdb to a running QEMU virtual machine using the command

make gdb

We used the QEMU command with the -s parameter, which means listening to port 1234 from gdb. We can do debugging using a remote target for gdb. The existing Makefile takes care of the details.

When you attach a debugger to a process, the process is suspended. You can add breakpoints and inspect the current status of the process.

Attach to the QEMU virtual machine (using the make gdb command) and place a breakpoint in the sys_access function using the following command in the gdb console:

break sys_access

At this time, the virtual machine is suspended. To continue executing it (up to the possible call of the sys_access function), use the command:

continue

in the gdb console.

At this time, the virtual machine is active and has a usable console. To make a sys_access call, issue a ls command. Note that the virtual machine was again suspended by gdb and the corresponding sys_access callback message appeared within the gdb console.

Trace code execution using step instruction, continue or next instruction. You probably do not understand everything that happens, so use commands such as list and backtrace to trace the execution.

Hint

At the gdb prompt, you can press Enter (without anything else) to rerun the last command.

4. GDB spelunking

Use gdb to display the source code of the function that creates kernel threads (kernel_thread).

Note

You can use GDB for static kernel analysis using, in the kernel source directory, a command such as:

gdb vmlinux

Go over the gdb (Linux) section of the lab.

Use gdb to find the address of the jiffies variable in memory and its contents. The jiffies variable holds the number of ticks (clock beats) since the system started.

Hint

To track the value of the jiffies variable, use dynamic analysis in gdb by running the command:

make gdb

as in the previous exercise.

Go over the gdb (Linux) section of the lab.

Hint

The jiffies is a 64-bit variable. You can see that its address is the same as the jiffies_64 variable.

To explore the contents of a 64-bit variable, use in the gdb console the command:

x/gx & jiffies

If you wanted to display the contents of the 32-bit variable, you would use in the gdb console the command:

x/wx & jiffies

5. Cscope spelunking

Use LXR or cscope in the ~/src/linux/ directory to discover the location of certain structures or functions.

Cscope index files are already generated. Use vim and other related commands to scroll through the source code. For example, use the command:

vim

for opening the vim editor. Afterwards, inside the editor, use commands such as:

:cs find g task_struct.

Find the file in which the following data types are defined:

• struct task_struct
• struct semaphore
• struct list_head
• spinlock_t
• struct file_system_type

Hint

For a certain structure, only its name needs to be searched.

For instance, in the case of struct task_struct, search for the task_struct string.

Usually, you will get more matches. To locate the one you are interested in, do the following:

1. List all matches by using, in vim, :copen command.
2. Look for the right match (where the structure is defined) by looking for an open character ({), a single character on the structure definition line. To search for the open braid you use in vim the construction /{.
3. On the respective line, press Enter to get into the source code where the variable is defined.
4. Close the secondary window using the command: :cclose command.

Find the file in which the following global kernel variables are declared:

• sys_call_table
• file_systems
• current
• chrdevs

Hint

To do this, use a vim command with the syntax:

:cs f g <symbol>

where <symbol> is the name of the symbol being searched.

Find the file in which the following functions are declared:

• copy_from_user
• vmalloc
• schedule_timeout
• add_timer

Hint

To do this, use a vim command with the syntax:

:cs f g <symbol>

where <symbol> is the name of the symbol being searched.

Scroll through the following sequence of structures:

• struct task_struct
• struct mm_struct
• struct vm_area_struct
• struct vm_operations_struct

That is, you access a structure and then you find fields with the data type of the next structure, access the respective fields and so on. Note in which files these structures are defined; this will be useful to the following labs.

Hint

In order to search for a symbol in vim (with cscope support) when the cursor is placed on it, use the Ctrl+] keyboard shortcut.

To return to the previous match (the one before search/jump), use the Ctrl+o keyboard shortcut.

To move forward with the search (to return to matches before Ctrl+o), use the Ctrl+i keyboard shortcut.

Following the above instructions, find and go through the function call sequence:

• bio_alloc
• bio_alloc_bioset
• bvec_alloc
• kmem_cache_alloc
• slab_alloc

Note

Read cscope or LXR Cross-Reference sections of the lab.