This is the most native setup, and therefore the best one if you are on one of the supported Ubuntu: 16.04 or 17.10. We will try to keep up with both the latest LTS and release.

Reserve 12Gb of disk and run:

The first configure will take a while (30 minutes to 2 hours) to clone and build, see Benchmark initial build for more details.

If you don’t want to wait, you could also try to compile the examples and run them on your host computer as explained on at Run on host, but as explained on that section, that is dangerous, limited, and will likely not work.

After QEMU opens up, you can start playing with the kernel modules:

This should print to the screen:

which are printk messages from init and cleanup methods of those modules.

Once you use GDB step debug and tmux, your terminal will look a bit like this:

All available modules can be found in the kernel_module directory.

This is a good option if you are on a Linux host, but the native build failed due to your weird host distribution.

Note however that most things in this repository are highly Linux-portable, should just work once you have found the corresponding apt-get package manager commands in the configure for your distro. In theory :-)

Before anything, you must get ride of any host build files on out/ if you have any. A simple way to do this it to:

A cleaner option is to make a separate clone of this repository just for Docker, although this will require another submodule update.

Then install Docker, e.g. on Ubuntu:

The very first time you launch Docker, create the container with:

You are now left inside a shell in the Docker guest.

From there, run the exact same commands that you would on a native install: Getting started

The host git top level directory is mounted inside the guest, which means for example that you can use your host’s GUI text editor directly on the files.

Just don’t forget that if you nuke that directory on the guest, then it gets nuked on the host as well!

Trying to run the output from Docker from host won’t however, I think the main reason is that the absolute paths inside Docker are will be different than the host ones, but even if we fix that there will likely be other problems.

TODO make files created inside Docker be owned by the current user in host instead of root: https://stackoverflow.com/questions/23544282/what-is-the-best-way-to-manage-permissions-for-docker-shared-volumes

Quit and stop the container:

Restart the container:

Open a second shell in a running container:

You will need this for example to use GDB step debug.

Start a second shell, and run a command in it at the same time:

Docker stops if and only if you quit the initial shell, you can quit this one without consequences.

If you mistakenly run ./rundocker twice, it opens two mirrored terminals. To quit one of them do https://stackoverflow.com/questions/19688314/how-do-you-attach-and-detach-from-dockers-process:

To use Graphic mode from Docker:

and then on host:

Destroy the docker container:

Since we mount the guest’s working directory on the host git top-level, you will likely not lose data from doing this, just the apt-get installs.

To get back to a host build, don’t forget to clean up out/ again:

After this, to start using Docker again will you need another:

By default, we show the serial console directly on the current terminal, without opening a QEMU window.

Quit QEMU immediately:

https://superuser.com/questions/1087859/how-to-quit-the-qemu-monitor-when-not-using-a-gui

Alternative methods:

• /poweroff.out. Not immediate, and sometimes you can’t do it, e.g. kernel panic. Furthermore, it does not work for ARM: arm shutdown

• echo quit | ./qemumonitor

• Ctrl-A C then quit

• pkill qemu

Toggle between QEMU monitor and the shell:

• http://stackoverflow.com/questions/14165158/how-to-switch-to-qemu-monitor-console-when-running-with-curses

• https://superuser.com/questions/488263/how-to-switch-to-the-qemu-control-panel-with-nographics

But doing:

is generally more practical since you can use host shell functionality like shell history.

Getting everything to work requires careful choice of QEMU command line options:

• https://stackoverflow.com/questions/49716931/how-to-run-qemu-with-nographic-and-monitor-but-still-be-able-to-send-ctrlc-to/49751144#49751144

• https://unix.stackexchange.com/questions/167165/how-to-pass-ctrl-c-to-the-guest-when-running-qemu-with-nographic/436321#436321

TODO: if you hit Ctrl-C several times while arm or aarch64 are booting, after boot the userland shell does not show any updates when you type. Why?

Enable graphic mode:

Text mode is the default due to the following considerable advantages:

• copy and paste commands and stdout output to / from host

• get full panic traces when you start making the kernel crash :-) See also: https://unix.stackexchange.com/questions/208260/how-to-scroll-up-after-a-kernel-panic

• have a large scroll buffer, and be able to search it, e.g. by using tmux on host

• one less window floating around to think about in addition to your shell :-)

• graphics mode has only been properly tested on x86_64.

Text mode has the following limitations over graphics mode:

• you can’t see graphics such as those produced by X11

• very early kernel messages such as early console in extract_kernel only show on the GUI, since at such early stages, not even the serial has been setup.

When debugging a module, it becomes tedious to wait for build and re-type:

every time.

To automate that, use the methods described at: init

We use printk a lot, and it shows on the terminal by default, along with stdout and what you type.

Hide all printk messages:

or equivalently:

See also: https://superuser.com/questions/351387/how-to-stop-kernel-messages-from-flooding-my-console

Do it with a Kernel command line parameters to affect the boot itself:

and now only boot warning messages or worse show, which is useful to identify problems.

Our default printk format is:

e.g.:

where:

• LEVEL: higher means less serious

• TIMESTAMP: seconds since boot

This format is selected by the following boot options:

• console_msg_format=syslog: add the <LEVEL> part

• printk.time=y: add the [TIMESTAMP] part

Scroll up in Graphic mode:

but I never managed to increase that buffer:

• https://askubuntu.com/questions/709697/how-to-increase-scrollback-lines-in-ubuntu14-04-2-server-edition

• https://unix.stackexchange.com/questions/346018/how-to-increase-the-scrollback-buffer-size-for-tty

The superior alternative is to use text mode and GNU screen or tmux.

Many of the modules have userland test scripts / executables with the same name as the module, e.g. form inside the guest:

The sources of those tests will further clarify what the corresponding kernel modules does. To find them on the host, do a quick:

After making changes to a package, you must explicitly request it to be rebuilt.

For example, you you modify the kernel modules, you must rebuild with:

which is just an alias for:

where kernel_module is the name of out Buildroot package that contains the kernel modules.

Other important targets are:

which rebuild the Linux kernel, and QEMU and gem5 respectively. They are essentially aliases for:

However, some of our aliases such as -l also have some magic convenience properties. So generally just use the aliases instead.

We don’t rebuild by default because, even with make incremental rebuilds, the timestamp check takes a few annoying seconds.

Not all packages have an alias, when they don’t, just use the form:

It gets annoying to retype -a aarch64 for every single command, or to remember ./build -B setups.

So simplify that, do:

and then edit the data/cli file to your needs.

That file is used to pass extra command line arguments to most of our utilities.

Of course, you could get by with the shell history, or your own aliases, but we’ve felt that it was worth introducing a common mechanism for that.

You did something crazy, and nothing seems to work anymore?

All builds are stored under buildroot/,

The most coarse thing you can do is:

To only nuke one architecture, do:

Only nuke one one package:

This is sometimes necessary when changing the version of the submodules, and then builds fail. We should try to understand why and report bugs.

The root filesystem is persistent across:

then:

The command:

also saves the disk.

This is particularly useful to re-run shell commands from the history of a previous session with Ctrl-R.

However, when you do:

the disk image gets overwritten by a fresh filesystem and you lose all changes.

Remember that if you forcibly turn QEMU off without sync or poweroff from inside the VM, e.g. by closing the QEMU window, disk changes may not be saved.

When booting from initrd however without a disk, persistency is lost.

Bootloaders can pass a string as input to the Linux kernel when it is booting to control its behaviour, much like the execve system call does to userland processes.

This allows us to control the behaviour of the kernel without rebuilding anything.

With QEMU, QEMU itself acts as the bootloader, and provides the -append option and we expose it through ./run -e, e.g.:

Then inside the host, you can check which options were given with:

They are also printed at the beginning of the boot message:

See also:

• https://unix.stackexchange.com/questions/48601/how-to-display-the-linux-kernel-command-line-parameters-given-for-the-current-bo

• https://askubuntu.com/questions/32654/how-do-i-find-the-boot-parameters-used-by-the-running-kernel

The arguments are documented in the kernel documentation: https://www.kernel.org/doc/html/v4.14/admin-guide/kernel-parameters.html

When dealing with real boards, extra command line options are provided on some magic bootloader configuration file, e.g.:

• GRUB configuration files: https://askubuntu.com/questions/19486/how-do-i-add-a-kernel-boot-parameter

• Raspberry pi /boot/cmdline.txt on a magic partition: https://raspberrypi.stackexchange.com/questions/14839/how-to-change-the-kernel-commandline-for-archlinuxarm-on-raspberry-pi-effectly

Double quotes can be used to escape spaces as in opt="a b", but double quotes themselves cannot be escaped, e.g. opt"a\"b"

This even lead us to use base64 encoding with -E!

If you are feeling fancy, you can also insert modules with:

This method also deals with module dependencies, which we almost don’t use to make examples simpler:

• https://askubuntu.com/questions/20070/whats-the-difference-between-insmod-and-modprobe

• https://stackoverflow.com/questions/22891705/whats-the-difference-between-insmod-and-modprobe

Removal also removes required modules that have zero usage count:

but it can’t know if you actually insmodded them separately or not:

so it is a bit risky.

modprobe searches for modules under:

Kernel modules built from the Linux mainline tree with CONFIG_SOME_MOD=m, are automatically available with modprobe, e.g.:

https://stackoverflow.com/questions/5947286/how-to-load-linux-kernel-modules-from-c-code

If you are feeling raw, you can insert and remove modules with our own minimal module inserter and remover!

which teaches you how it is done from C code.

The Linux kernel offers two system calls for module insertion:

• init_module

• finit_module

and:

documents that:

The finit_module() system call is like init_module(), but reads the module to be loaded from the file descriptor fd. It is useful when the authenticity of a kernel module can be determined from its location in the filesystem; in cases where that is possible, the overhead of using cryptographically signed modules to determine the authenticity of a module can be avoided. The param_values argument is as for init_module().

finit is newer and was added only in v3.8. More rationale: https://lwn.net/Articles/519010/

-d makes QEMU wait for a GDB connection, otherwise we could accidentally go past the point we want to break at:

Say you want to break at start_kernel. So on another shell:

or at a given line:

Now QEMU will stop there, and you can use the normal GDB commands:

See also:

• http://stackoverflow.com/questions/11408041/how-to-debug-the-linux-kernel-with-gdb-and-qemu/33203642#33203642

• http://stackoverflow.com/questions/4943857/linux-kernel-live-debugging-how-its-done-and-what-tools-are-used/42316607#42316607

Let’s observe the kernel as it reacts to some userland actions.

Start QEMU with just:

and after boot inside a shell run:

which counts to infinity to stdout. Then in another shell, run:

and then hit:

And you now control the counting on the first shell from GDB!

When you hit Ctrl-C, if we happen to be inside kernel code at that point, which is very likely if there are no heavy background tasks waiting, and we are just waiting on a sleep type system call of the command prompt, we can already see the source for the random place inside the kernel where we stopped.

https://unix.stackexchange.com/questions/152738/how-to-split-a-new-window-and-run-a-command-in-this-new-window-using-tmux/432111#432111

tmux just makes things even more fun by allowing us to see both terminals at once without dragging windows around!

Gives two panes:

• left: usual QEMU

• right: gdb

and focuses on the GDB pane.

To start again, switch back to the QEMU (<prefix>-o) pane, and re-run:

This automatically kills the GDB pane.

To quit for good, exit GDB, and quit the right shell with Ctrl-D.

Pass extra GDB arguments with:

If you are using gem5 instead of QEMU, -u has a different effect: it opens the gem5 terminal instead of the debugger:

If you also want to use the debugger with gem5, you will need to create your own panes or windows.

http://stackoverflow.com/questions/28607538/how-to-debug-linux-kernel-modules-with-qemu/44095831#44095831

Loadable kernel modules are a bit trickier since the kernel can place them at different memory locations depending on load order.

So we cannot set the breakpoints before insmod.

However, the Linux kernel GDB scripts offer the lx-symbols command, which takes care of that beautifully for us.

Shell 1:

Wait for the boot to end and run:

This prints a message to dmesg every second.

Shell 2:

In GDB, hit Ctrl-C, and note how it says:

That’s lx-symbols working! Now simply:

and we now control the callback from GDB!

Just don’t forget to remove your breakpoints after rmmod, or they will point to stale memory locations.

TODO: why does break work_func for insmod kthread.ko not break the first time I insmod, but breaks the second time?

TODO: why can’t we break at early startup stuff such as:

Maybe it is because they are being copied around at specific locations instead of being run directly from inside the main image, which is where the debug information points to?

See also: https://stackoverflow.com/questions/2589845/what-are-the-first-operations-that-the-linux-kernel-executes-on-boot

QEMU’s -gdb GDB breakpoints are set on virtual addresses, so you can in theory debug userland processes as well.

• https://stackoverflow.com/questions/26271901/is-it-possible-to-use-gdb-and-qemu-to-debug-linux-user-space-programs-and-kernel

• https://stackoverflow.com/questions/16273614/debug-init-on-qemu-using-gdb

You will generally want to use gdbserver for this as it is more reliable, but this method can overcome the following limitations of gdbserver:

• the emulator does not support host to guest networking. This seems to be the case for gem5: gem5 host to guest networking

• cannot see the start of the init process easily

• gdbserver alters the working of the kernel, and makes your run less representative

Known limitations of direct userland debugging:

• the kernel might switch context to another process or to the kernel itself e.g. on a system call, and then TODO confirm the PIC would go to weird places and source code would be missing.

• TODO step into shared libraries. If I attempt to load them explicitly:

  (gdb) sharedlibrary ../../staging/lib/libc.so.0
  No loaded shared libraries match the pattern `../../staging/lib/libc.so.0'.

  since GDB does not know that libc is loaded.

GDB can call functions as explained at: https://stackoverflow.com/questions/1354731/how-to-evaluate-functions-in-gdb

However this is failing for us:

• some symbols are not visible to call even though b sees them

• for those that are, call fails with an E14 error

E.g.: if we break on sys_write on /count.sh:

even though fdget_pos is the first thing sys_write does:

I also noticed that I get the same error:

when trying to use:

on many (all?) functions.

See also: cirosantilli#19

./run -k
./rungdb -k

c

/count.sh &
/kgdb.sh

b sys_write
c
c
c
c

GDB not connecting to KGDB in ARM. Possibly linked to -serial stdio. See also: https://stackoverflow.com/questions/14155577/how-to-use-kgdb-on-arm

Main shell just falls on:

and GDB shell gives:

In QEMU:

In GDB:

and you now control the count.

TODO: if I -ex lx-symbols to the gdb command, just like done for QEMU -gdb, the kernel oops. How to automate this step?

If you modify runqemu to use:

instead of kgdboc=ttyS0,115200, you enter a different debugging mode called KDB.

Usage: in QEMU:

Boot finishes, then:

And you are back in KDB. Now you can:

And you will break whenever sys_write is hit.

The other KDB commands allow you to instruction steps, view memory, registers and some higher level kernel runtime data.

But TODO I don’t think you can see where you are in the kernel source code and line step as from GDB, since the kernel source is not available on guest (ah, if only debugging information supported full source).

/gdbserver.sh /myinsmod.out /hello.ko

./rungdbserver kernel_module-1.0/user/myinsmod.out

find out/x86_64/buildroot/build -name myinsmod.out

As usual, different archs work with:

BusyBox executables are all symlinks, so if you do on guest:

on host you need:

Our setup gives you the rare opportunity to step debug libc and other system libraries e.g. with:

Or simply by stepping into calls:

This is made possible by the GDB command:

which automatically finds unstripped shared libraries on the host for us.

See also: https://stackoverflow.com/questions/8611194/debugging-shared-libraries-with-gdbserver/45252113#45252113

./build -a arm
./run -a arm

./run -a arm -d
# On another terminal.
./rungdb -a arm

As usual, we use Buildroot’s recommended QEMU setup QEMU aarch64 setup:

• https://github.com/buildroot/buildroot/blob/2017.08/board/qemu/aarch64-virt/readme.txt

• https://github.com/buildroot/buildroot/blob/2017.08/configs/qemu_aarch64_virt_defconfig

This makes aarch64 a bit different from arm:

• uses -M virt. https://wiki.qemu.org/Documentation/Platforms/ARM explains:

  Most of the machines QEMU supports have annoying limitations (small amount of RAM, no PCI or other hard disk, etc) which are there because that’s what the real hardware is like. If you don’t care about reproducing the idiosyncrasies of a particular bit of hardware, the best choice today is the "virt" machine.

  -M virt has some limitations, e.g. I could not pass -drive if=scsi as for arm, and so Snapshot fails.

Keep in mind that MIPS has the worst support compared to our other architectures due to the smaller community. Patches welcome as usual.

TODOs:

• networking is not working. See also:

  • https://stackoverflow.com/questions/21496449/networking-is-not-working-on-qemu-guest-malta-mips

  • https://unix.stackexchange.com/questions/208266/setting-up-qemu-and-mipsel-networking-trouble

  • https://unix.stackexchange.com/questions/354127/qemu-mips-and-debian

• GDB step debug does not work properly, does not find start_kernel

To have more control over the system, you can replace BusyBox’s init with your own.

The -E option replaces init and evals a command from the Kernel command line parameters:

It is basically a shortcut for:

although -E is smarter:

• allows quoting and newlines by using base64 encoding, see: Kernel command line parameters escaping

• automatically chooses between init= and rcinit= for you, see: Path to init

so you should almost always use it, unless you are really counting each cycle ;-)

This method replaces BusyBox' init completely, which makes things more minimal, but also has has the following consequences:

• /etc/fstab mounts are not done, notably /proc and /sys, test it out with:

  ./run -E 'echo asdf;ls /proc;ls /sys;echo qwer'

• no shell is launched at the end of boot for you to interact with the system. You could explicitly add a sh at the end of your commands however:

  ./run -E 'echo hello;sh'

The best way to overcome those limitations is to use: Run command at the end of BusyBox init

If the script is large, you can add it to a gitignored file and pass that to -E as in:

or add it to a file to the root filesystem guest and rebuild:

Remember that if your init returns, the kernel will panic, there are just two non-panic possibilities:

• run forever in a loop or long sleep

• poweroff the machine

Use the -F option is for you rely on something that BusyBox' init set up for you like /etc/fstab:

After the commands run, you are left on an interactive shell.

The above command is basically equivalent to:

where the lkmc_eval option gets evaled by our default S98 startup script if present.

However, -F is smarter and uses base64 encoding, much like -E vs -e, so you will just use -F most of the time.

Alternatively, add them to a new init.d entry to run at the end o the BusyBox init:

and they will be run automatically before the login prompt.

Scripts under /etc/init.d are run by /etc/init.d/rcS, which gets called by the line ::sysinit:/etc/init.d/rcS in /etc/inittab.

The init is selected at:

• initrd or initramfs system: /init, a custom one can be set with the rdinit= kernel command line parameter

• otherwise: default is /sbin/init, followed by some other paths, a custom one can be set with init=

More details: https://unix.stackexchange.com/questions/30414/what-can-make-passing-init-path-to-program-to-the-kernel-not-start-program-as-i/430614#430614

Documented at https://www.kernel.org/doc/html/v4.14/admin-guide/kernel-parameters.html:

The kernel parses parameters from the kernel command line up to "-"; if it doesn’t recognize a parameter and it doesn’t contain a '.', the parameter gets passed to init: parameters with '=' go into init’s environment, others are passed as command line arguments to init. Everything after "-" is passed as an argument to init.

And you can try it out with:

Also note how the annoying dash - also gets passed as a parameter to init, which makes it impossible to use this method for most executables.

Finally, the docs are lying, arguments with dots that come after - are still treated specially (of the form subsystem.somevalue) and disappear:

We disable networking by default because it starts an userland process, and we want to keep the number of userland processes to a minimum to make the system more understandable.

Enable:

Disable:

Test:

BusyBox' ping does not work with hostnames even when networking is working fine:

TODO why: https://unix.stackexchange.com/questions/124283/busybox-ping-ip-works-but-hostname-nslookup-fails-with-bad-address

To enable networking by default, use the methods documented at Automatic startup commands

./run -K

panic: KVM: Failed to enter virtualized mode (hw reason: 0x80000021)

./build -b br2_x11
./run -x

startx

xcalc
xeyes

TODO 9076c1d9bcc13b6efdb8ef502274f846d8d4e6a1 I’m 100% sure that it was working before, but I didn’t run it forever, and it stopped working at some point. Needs bisection, on whatever commit last touched x11 stuff.

• https://askubuntu.com/questions/730891/how-can-i-get-a-mouse-cursor-in-qemu

• https://stackoverflow.com/questions/19665412/mouse-and-keyboard-not-working-in-qemu-emulator

-show-cursor did not help, I just get to see the host cursor, but the guest cursor still does not move.

Doing:

shows that interrupts do happen when mouse and keyboard presses are done, so I expect that it is some wrong either with:

• QEMU. Same behaviour if I try the host’s QEMU 2.10.1 however.

• X11 configuration. We do have BR2_PACKAGE_XDRIVER_XF86_INPUT_MOUSE=y.

/var/log/Xorg.0.log contains the following interesting lines:

The file /dev/inputs/mice does not exist.

On ARM, startx hangs at a message:

and nothing shows on the screen, and:

says:

A friend told me this but I haven’t tried it yet:

• xf86-video-modesetting is likely the missing ingredient, but it does not seem possible to activate it from Buildroot currently without patching things.

• xf86-video-fbdev should work as well, but we need to make sure fbdev is enabled, and maybe add some line to the Xorg.conf

Haven’t tried it, doubt it will work out of the box! :-)

Maybe: https://stackoverflow.com/questions/47857023/booting-a-graphical-mips-qemu-machine

./build -i
./run -i

cat ./out/x86_64/run.sh

date >f
poweroff
cat f
# can't open 'f': No such file or directory

end Kernel panic - not syncing: Out of memory and no killable processes...

./run -im 256M

Most modern desktop distributions have an initrd in their root disk to do early setup.

The rationale for this is described at: https://en.wikipedia.org/wiki/Initial_ramdisk

One obvious use case is having an encrypted root filesystem: you keep the initrd in an unencrypted partition, and then setup decryption from there.

I think GRUB then knows read common disk formats, and then loads that initrd to memory with a /boot/grub/grub.cfg directive of type:

Related: https://stackoverflow.com/questions/6405083/initrd-and-booting-the-linux-kernel

initramfs is just like initrd, but you also glue the image directly to the kernel image itself.

So the only argument that QEMU needs is the -kernel, no -drive not even -initrd! Pretty cool.

Try it out with:

The -l (ell) should only be used the first time you move to / from a different root filesystem method (ext2 or cpio) to initramfs to overcome: https://stackoverflow.com/questions/49260466/why-when-i-change-br2-linux-kernel-custom-config-file-and-run-make-linux-reconfi

It is interesting to see how this increases the size of the kernel image if you do a:

before and after using initramfs, since the .cpio is now glued to the kernel image.

In the background, it uses BR2_TARGET_ROOTFS_INITRAMFS, and this makes the kernel config option CONFIG_INITRAMFS_SOURCE point to the CPIO that will be embedded in the kernel image.

http://nairobi-embedded.org/initramfs_tutorial.html shows a full manual setup.

TODO we were not able to get it working yet: https://stackoverflow.com/questions/49261801/how-to-boot-the-linux-kernel-with-initrd-or-initramfs-with-gem5

We try to use the latest possible kernel major release version.

In QEMU:

or in the source:

During update all you kernel modules may break since the kernel API is not stable.

They are usually trivial breaks of things moving around headers or to sub-structs.

The userland, however, should simply not break, as Linus enforces strict backwards compatibility of userland interfaces.

This backwards compatibility is just awesome, it makes getting and running the latest master painless.

This also makes this repo the perfect setup to develop the Linux kernel.

The kernel is not forward compatible, however, so downgrading the Linux kernel requires downgrading the userland too to the latest Buildroot branch that supports it.

The default Linux kernel version is bumped in Buildroot with commit messages of type:

So you can try:

Those commits change BR2_LINUX_KERNEL_LATEST_VERSION in /linux/Config.in.

You should then look up if there is a branch that supports that kernel. Staying on branches is a good idea as they will get backports, in particular ones that fix the build as newer host versions come out.

To test out kernel panics and oops in controlled circumstances, try out the modules:

A panic can also be generated with:

Panic vs oops: https://unix.stackexchange.com/questions/91854/whats-the-difference-between-a-kernel-oops-and-a-kernel-panic

How to generate them:

• https://unix.stackexchange.com/questions/66197/how-to-cause-kernel-panic-with-a-single-command

• https://stackoverflow.com/questions/23484147/generate-kernel-oops-or-crash-in-the-code

When a panic happens, Shift-PgUp does not work as it normally does, and it is hard to get the logs if on are on Graphic mode:

• https://superuser.com/questions/848412/scrolling-up-the-failed-screen-with-kernel-panic

• https://superuser.com/questions/269228/write-qemu-booting-virtual-machine-output-to-a-file

• http://www.reactos.org/wiki/QEMU#Redirect_to_a_file

I once got UML running on a minimal Buildroot setup at: https://unix.stackexchange.com/questions/73203/how-to-create-rootfs-for-user-mode-linux-on-fedora-18/372207#372207

But in part because it is dying, I didn’t spend much effort to integrate it into this repo, although it would be a good fit in principle, since it is essentially a virtualization method.

Maybe some brave should will send a pull request one day.

https://stackoverflow.com/questions/3177338/how-is-the-linux-kernel-tested

https://stackoverflow.com/questions/40227651/does-qemu-emulator-have-checkpoint-function/48724371#48724371

QEMU allows us to take snapshots at any time through the monitor.

You can then restore CPU, memory and disk state back at any time.

qcow2 filesystems must be used for that to work.

To test it out, login into the VM with and run:

On another shell, take a snapshot:

The counting continues.

Restore the snapshot:

and the counting goes back to where we saved. This shows that CPU and memory states were reverted.

We can also verify that the disk state is also reversed. Guest:

Monitor:

Guest:

Monitor:

Guest:

And the output is 0.

Our setup does not allow for snapshotting while using initrd.

We have added and interacted with a few educational hardware models in QEMU.

This is useful to learn:

• how to create new hardware models for QEMU. Overview: https://stackoverflow.com/questions/28315265/how-to-add-a-new-device-in-qemu-source-code

• how the Linux kernel interacts with hardware

To get started, have a look at the "Hardware device drivers" section under kernel_module/README.adoc, and try to run those modules, and then grep the QEMU source code.

This protocol allows sharing a mountable filesystem between guest and host.

With networking, it’s boring, we can just use any of the old tools like sshfs and NFS.

https://superuser.com/questions/628169/how-to-share-a-directory-with-the-host-without-networking-in-qemu

One advantage of this method over NFS is that can run without sudo on host, or having to pass host credentials on guest for sshfs.

TODO performance compared to NFS.

As usual, we have already set everything up for you. On host:

Guest:

Host:

The main ingredients for this are:

• 9P settings in our kernel_config_fragment

• 9p entry on our rootfs_overlay/etc/fstab

  Alternatively, you could also mount your own with:

  mkdir /mnt/my9p
  mount -t 9p -o trans=virtio,version=9p2000.L host0 /mnt/my9p

• Launch QEMU with -virtfs as in your run script

  When we tried:

  security_model=mapped

  writes from guest failed due to user mismatch problems: https://serverfault.com/questions/342801/read-write-access-for-passthrough-9p-filesystems-with-libvirt-qemu

The feature is documented at: https://wiki.qemu.org/Documentation/9psetup

First ensure that networking is enabled before trying out anything in this section: Networking

This has nothing to do with the Linux kernel, but it is cool:

This uses QEMU’s user-mode emulation mode that allows us to run cross-compiled userland programs directly on the host.

The reason this is cool, is that ls is not statically compiled, but since we have the Buildroot image, we are still able to find the shared linker and the shared library at the given path.

In other words, much cooler than:

It is also possible to compile QEMU user mode from source with BR2_PACKAGE_HOST_QEMU_LINUX_USER_MODE=y, but then your compilation will likely fail with:

since we are using a bleeding edge kernel, which is a sanity check in the Buildroot QEMU package.

Anyways, this warns us that the userland emulation will likely not be reliable, which is good to know. TODO: where is it documented the host kernel must be as new as the target one?

GDB step debugging is also possible with:

TODO: find source. Lazy now.

When you start interacting with QEMU hardware, it is useful to see what is going on inside of QEMU itself.

This is of course trivial since QEMU is just an userland program on the host, but we make it a bit easier with:

Then you could:

And in QEMU:

Just make sure that you never click inside the QEMU window when doing that, otherwise you mouse gets captured forever, and the only solution I can find is to go to a TTY with Ctrl-Alt-F1 and kill QEMU.

You can still send key presses to QEMU however even without the mouse capture, just either click on the title bar, or alt tab to give it focus.

QEMU has a mechanism to log all instructions executed to a file.

To do it for the Linux kernel boot we have a helper:

You can then inspect the instructions with:

This functionality relies on the following setup:

• ./configure --enable-trace-backends=simple. This logs in a binary format to the trace file.

  It makes 3x execution faster than the default trace backend which logs human readable data to stdout.

  Logging with the default backend log greatly slows down the CPU, and in particular leads to this boot message:

  All QSes seen, last rcu_sched kthread activity 5252 (4294901421-4294896169), jiffies_till_next_fqs=1, root ->qsmask 0x0
  swapper/0       R  running task        0     1      0 0x00000008
   ffff880007c03ef8 ffffffff8107aa5d ffff880007c16b40 ffffffff81a3b100
   ffff880007c03f60 ffffffff810a41d1 0000000000000000 0000000007c03f20
   fffffffffffffedc 0000000000000004 fffffffffffffedc ffffffff00000000
  Call Trace:
   <IRQ>  [<ffffffff8107aa5d>] sched_show_task+0xcd/0x130
   [<ffffffff810a41d1>] rcu_check_callbacks+0x871/0x880
   [<ffffffff810a799f>] update_process_times+0x2f/0x60

  in which the boot appears to hang for a considerable time.

• patch QEMU source to remove the disable from exec_tb in the trace-events file. See also: https://rwmj.wordpress.com/2016/03/17/tracing-qemu-guest-execution/

Sometimes in Ubuntu 14.04, after the QEMU SDL GUI starts, it does not get updated after keyboard strokes, and there are artifacts like disappearing text.

We have not managed to track this problem down yet, but the following workaround always works:

This started happening when we switched to building QEMU through Buildroot, and has not been observed on later Ubuntu.

Using text mode is another workaround if you don’t need GUI features.

gem5 is a system simulator, much like QEMU: http://gem5.org/

For the most part, just add the -g option to all commands and everything should magically work:

On another shell:

A full rebuild is currently needed even if you already have QEMU working unfortunately, see: gem5 and QEMU with the same kernel configuration

Tested architectures:

• arm

• aarch64

• x86_64

Like QEMU, gem5 also has a syscall emulation mode (SE), but in this tutorial we focus on the full system emulation mode (FS). For a comparison see: https://stackoverflow.com/questions/48986597/when-should-you-use-full-system-fs-vs-syscall-emulation-se-with-userland-program

• advantages of gem5:

  • simulates a generic more realistic pipelined and optionally out of order CPU cycle by cycle, including a realistic DRAM memory access model with latencies, caches and page table manipulations. This allows us to:

    • do much more realistic performance benchmarking with it, which makes absolutely no sense in QEMU, which is purely functional

    • make certain functional observations that are not possible in QEMU, e.g.:

      • use Linux kernel APIs that flush cache memory like DMA, which are crucial for driver development. In QEMU, the driver would still work even if we forget to flush caches.

      • spectre / meltdown:

        • https://www.mail-archive.com/[email protected]/msg15319.html

        • https://github.com/jlpresearch/gem5/tree/spectre-test

          It is not of course truly cycle accurate, as that

  • would require exposing proprietary information of the CPU designs: https://stackoverflow.com/questions/17454955/can-you-check-performance-of-a-program-running-with-qemu-simulator/33580850#33580850

  • would make the simulation even slower TODO confirm, by how much

    but the approximation is reasonable.

    It is used mostly for microarchitecture research purposes: when you are making a new chip technology, you don’t really need to specialize enormously to an existing microarchitecture, but rather develop something that will work with a wide range of future architectures.

  • runs are deterministic by default, unlike QEMU which has a special QEMU record and replay mode, that requires first playing the content once and then replaying

• disadvantage of gem5: slower than QEMU, see: Benchmark Linux kernel boot

  This implies that the user base is much smaller, since no Android devs.

  Instead, we have only chip makers, who keep everything that really works closed, and researchers, who can’t version track or document code properly >:-) And this implies that:

  • the documentation is more scarce

  • it takes longer to support new hardware features

  Well, not that AOSP is that much better anyways.

• not sure: gem5 has BSD license while QEMU has GPL

  This suits chip makers that want to distribute forks with secret IP to their customers.

  On the other hand, the chip makers tend to upstream less, and the project becomes more crappy in average :-)

OK, this is why we used gem5 in the first place, performance measurements!

Let’s benchmark Dhrystone which Buildroot provides.

The most flexible way is to do:

These commands output the approximate number of CPU cycles it took Dhrystone to run.

For more serious tests, you will likely want to automate logging the commands ran and results to files, a good example is: gem5-bench-cache.

A more naive and simpler to understand approach would be a direct:

but the problem is that this method does not allow to easily run a different script without running the boot again, see: gem5 checkpoint restore and run a different script

A few imperfections of our benchmarking method are:

• when we do m5 resetstats and m5 exit, there is some time passed before the exec system call returns and the actual benchmark starts and ends

• the benchmark outputs to stdout, which means so extra cycles in addition to the actual computation. But TODO: how to get the output to check that it is correct without such IO cycles?

Solutions to these problems include:

• modify benchmark code with instrumentation directly, as PARSEC and ARM employees have been doing: https://github.com/arm-university/arm-gem5-rsk/blob/aa3b51b175a0f3b6e75c9c856092ae0c8f2a7cdc/parsec_patches/xcompile-patch.diff#L230

• monitor known addresses

Discussion at: https://stackoverflow.com/questions/48944587/how-to-count-the-number-of-cpu-clock-cycles-between-the-start-and-end-of-a-bench/48944588#48944588

Those problems should be insignificant if the benchmark runs for long enough however.

Now you can play a fun little game with your friends:

• pick a computational problem

• make a program that solves the computation problem, and outputs output to stdout

• write the code that runs the correct computation in the smallest number of cycles possible

To find out why your program is slow, a good first step is to have a look at the statistics for the run:

Whenever we run m5 dumpstats or m5 exit, a section with the following format is added to that file:

Analogous to QEMU:

Internals: when we give --command-line= to gem5, it overrides default command lines, including some mandatory ones which are required to boot properly.

Our run script hardcodes the require options in the default --command-line and appends extra options given by -e.

To find the default options in the first place, we removed --command-line and ran:

and then looked at the line of the Linux kernel that starts with:

Analogous to QEMU’s Snapshot, but better since it can be started from inside the guest, so we can easily checkpoint after a specific guest event, e.g. just before init is done.

Documentation: http://gem5.org/Checkpoints

In the guest, wait for the boot to end and run:

To restore the checkpoint, kill the VM and run:

Let’s create a second checkpoint to see how it works, in guest:

Kill the VM, and try it out:

and now in the guest:

contains the date. The file f wouldn’t exist had we used the first checkpoint with -r 1.

If you automate things with Kernel command line parameters as in:

Then there is no need to pass the kernel command line again to gem5 for replay:

since boot has already happened, and the parameters are already in the RAM of the snapshot.

Our scripts "namespace" with the checkpoint by architecture with --checkpoint-dir, so if you make two checkpoints:

• one in x86

• the other in arm

Then you would still restore both of them with -- -r 1.

This makes it easier to remember which checkpoint is which, especially since there appears to be no runtime way to set the checkpoint names.

Internals:

• the checkpoints are stored under out/$arch/gem5/m5out/cpt.$todo_whatisthis

• m5 is a guest utility present inside the gem5 tree which we cross-compiled and installed into the guest

Pass options to the fs.py script:

• get help:

  ./run -g -- -h

• boot with the more detailed and slow HPI CPU model:

  ./run -a arm -g -- --caches --cpu-type=HPI

Pass options to the gem5 executable itself:

• get help:

  ./run -G '-h' -g

gem5 just assigns new ports if some ports are occupied, so we can do:

And a second instance:

TODO Now we just need to network them up to have some more fun! See dist-gem5: http://publish.illinois.edu/icsl-pdgem5/

We would like to be able to run both gem5 and QEMU with the same minimal kernel build to:

• do a single Buildroot build for both. Otherwise, we have to create two full out/.*/buildroot/ directories, which takes up a lot of time.

  Alternatively, we could try to be brave and switch between two kernel builds inside out/.*/buildroot/, but that would be too hackish.

• be able to compare behaviour between QEMU and gem5 when one is doing something weird.

  Note however that there are also variations which need to be controlled, e.g. kernel command line, DTB and QEMU’s non-determinism.

Unfortunately, we have only managed to find a working config for aarch64, which just works transparently.

The others use the Buildroot config for QEMU, and magic huge post-olddefconfig config files floating around the web for GEM5.

Subsections of this section document our failed attempts so far.

This is the strategy that we used to make it work for aarch64:

• make savedefconfig on the working gem5 kernel tree

• paste the result on kernel_config_fragment

• bisect it up

but this strategy failed for the other archs for some reason.

m5 is a guest command line utility that is installed and run on the guest.

It generates magic instructions, which lead gem5 to do magic things, like dumpstats or exit.

It is however under-documented, so let’s document some of its capabilities here.

Lets try to understand some stats better.

• networking not working. We currently just disable it from inittab by default to prevent waiting at startup

This method runs the kernel modules directly on your host computer without a VM, and saves you the compilation time and disk usage of the virtual machine method.

It has however severe limitations, and you will soon see that the compilation time and disk usage are well worth it:

• can’t control which kernel version and build options to use. So some of the modules will likely not compile because of kernel API changes, since the Linux kernel does not have a stable kernel module API.

• bugs can easily break you system. E.g.:

  • segfaults can trivially lead to a kernel crash, and require a reboot

  • your disk could get erased. Yes, this can also happen with sudo from userland. But you should not use sudo when developing newbie programs. And for the kernel you don’t have the choice not to use sudo

  • even more subtle system corruption such as not being able to rmmod

• can’t control which hardware is used, notably the CPU architecture

• can’t step debug it with GDB easily

Still interested?

If the compilation of any of the C files fails because of kernel or toolchain differences that we don’t control on the host, just rename it to remove the .c extension and try again:

Once you manage to compile, and have come to terms with the fact that this may blow up your host, try it out with:

Once you are done with this method, you must clean up the in-tree build objects before you decide to do the right thing and move on to the superior ./build Buildroot method:

otherwise they will cause problems.

Minimal host build system sanity check example.

We provide the following mechanisms:

• ./build -b data/br2: append the Buildroot configuration file data/br2 to a single build. Must be passed every time you run ./build. The format is the same as br2.

• ./build -B 'BR2_SOME_OPTION="myval"': append a single option to a single build.

You will then likely want to make those more permanent with: Don’t retype arguments all the time

make menuconfig is a convenient way to find Buildroot configurations:

Hit / and search for the settings.

Save and quit.

Then copy and paste the diff additions to br2 to make them permanent.

At startup, we login automatically as the root user as explained at: https://unix.stackexchange.com/questions/299408/how-to-login-automatically-without-typing-root-in-buildroot-x86-64-qemu/300152#300152

If you want to switch to another user to test some permissions, we have already created an user0 user through the user_table file, and you can just login as that user with:

and password:

Then test that the user changed with:

which gives:

We have enabled ccached builds by default.

BR2_CCACHE_USE_BASEDIR=n is used, which means that:

• absolute paths are used and GDB can find source files

• but builds are not reused across separated LKMC directories

ccache can considerably speed up builds when you:

• are switching between multiple configurations for a given package to bisect something out, as mentioned at: Use your own kernel config

• clean the build because things stopped working. We store the cache outside of this repository, so you can nuke away without fear

The default ccache environment variables are honored if you have them set, which we recommend you do. E.g., in your .bashrc:

The choice basically comes down to:

• should I store my cache on my HD or SSD?

• how big is my build, and how many build configurations do I need to keep around at a time?

If you don’t set it, the default is to use ~/.buildroot-ccache with 5G, which is a bit small for us.

I find it very relaxing to watch ccache at work with:

or if you have it installed on host and the environment variables exported simply with:

while a build is going on in another terminal and my cooler is humming. Especially when the hit count goes up ;-) The joys of system programming.

First, see if you can’t get away without actually adding a new package, for example:

• if you have a standalone C file with no dependencies besides the C standard library to be compiled with GCC, just add a new file under kernel_module/user and you are done

• if you have a dependency on a library, first check if Buildroot doesn’t have a package for it already with ls buildroot/package. If yes, just enable that package as explained at: Custom Buildroot options

If none of those methods are flexible enough for you, create a new package as follows:

• use sample_package as a starting point

• fork this repository, and modify that package to do what you want

• read the comments on that package to get an idea of how to start

• check the main manual for more complicated things: https://buildroot.org/downloads/manual/manual.html

• don’t forget to rebuild with:

  ./build -- sample_package-reconfigure

  if you make any changes to that package after the initial build: Rebuild

When adding new large package to the Buildroot root filesystem, it may fail with the message:

The solution is to simply add:

where 512Mb is "large enough".

Note that dots cannot be used as in 1.5G, so just use Megs as in 1500M instead.

Unfortunately, TODO we don’t have a perfect way to find the right value for BR2_TARGET_ROOTFS_EXT2_SIZE. One good heuristic is:

https://stackoverflow.com/questions/49211241/is-there-a-way-to-automatically-detect-the-minimum-required-br2-target-rootfs-ex

One way to overcome this problem is to mount benchmarks from host instead of adding them to the root filesystem, e.g. with: 9P

Buildroot is not designed for large root filesystem images, and the rebuild becomes very slow when we add a large package to it.

This is due mainly to the pkg-generic GLOBAL_INSTRUMENTATION_HOOKS sanitation which go over the entire tree doing complex operations…​ I no like, in particular check_bin_arch and check_host_rpath, which get stuck for a long time on the message:

The pause is followed by:

so which shows that the whole delay is inside our install itself.

I put an echo f in check_bin_arch, and it just loops forever, does not stop on a particular package.

When asking for help on upstream repositories outside of this repository, you will need to provide the commands that you are running in detail without referencing our scripts.

For example, QEMU developers will only want to see the final QEMU command that you are running.

For the configure and build, search for the Building and Configuring parts of the build log, then try to strip down all Buildroot related paths, to keep only options that seem to matter.

We make that easy by building commands as strings, and then echoing them before evaling.

So for example when you run:

Stdout shows a line with the full command of type:

and this line is also saved to a file for convenience:

Next, you will also want to give the relevant images to save them time. Zip the images with:

and then upload the out/images-*.zip file somewhere, e.g. GitHub release assets as in https://github.com/cirosantilli/linux-kernel-module-cheat/releases/tag/test-replay-arm

Finally, do a clone of the relevant repository out of tree and reproduce the bug there, to be 100% sure that it is an actual upstream bug, and to provide developers with the cleanest possible commands. For example as was done in this QEMU bug report: https://bugs.launchpad.net/qemu/+bug/1762179

Our philosophy is:

• if something adds little to the build time, build it in by default

• otherwise, make it optional

• try to keep the toolchain (GCC, Binutils) unchanged, otherwise a full rebuild is required.

  So we generally just enable all toolchain options by defaut, even though this adds a bit of time to the build.

  The biggest build time hog is always GCC, and it does not look like we can use a precompiled one: https://stackoverflow.com/questions/10833672/buildroot-environment-with-host-toolchain

• if something is very vaulable, we just add it by default even if it increases the Build time, notably GDB and QEMU

• runtime is sacred.

  We do our best to reduce the instruction and feature count to the bare minimum needed, to make the system:

  • easier to understand

  • run faster, specially for gem5

  One possibility we could play with is to build loadable modules instead of built-in modules to reduce runtime, but make it easier to get started with the modules.