Speed up your kernel

development cycle with QEMU

Stefan Hajnoczi <[email protected]>

Kernel Recipes 2015
1 KERNEL RECIPES 2015 | STEFAN HAJNOCZI
Agenda

● Kernel development cycle

● Introduction to QEMU
● Basics
● Testing kernels inside virtual machines
● Debugging virtual machines
● Advanced topics
● Cross-architecture testing
● Device bring-up
● Error injection

2 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

About me
QEMU contributor since 2010
● Subsystem maintainer
● Google Summer of Code & Outreachy
mentor/admin
● http://qemu-advent-calendar.org/

Occassional kernel patch contributor

● vsock, tcm_vhost, virtio_scsi, line6 staging driver

Work in Red Hat's Virtualization team

3 KERNEL RECIPES 2015 | STEFAN HAJNOCZI
Kernel development cycle

Write code

Test Build kernel/modules

Deploy
This
presentation

4 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

If you are doing kernel development...
USB
PCI Tracing ftrace
Device drivers
etc ebpf

LIO SCSI target .

Storage File systems .
device-mapper targets .

Network protocols
Networking Netfilter
OpenVSwitch

Resource Cgroups
management & Linux Security Modules
security Namespaces

5 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

...you might have these challenges

In situ debugging mechanisms like kgdb or kdump

● Not 100% reliable since they share the environment
● Crashes interrupt your browser/text editor session

Web Text
test.ko
R.I.P. browser editor
All My Work

Gone CRASH!
2015/09/01 Development machine

6 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Dedicated test machines

Ex situ debugging requires an additional machine

● More cumbersome to deploy code and run tests
● May require special hardware (JTAG, etc)
● Less mobile, hard to travel with multiple machines

PXE boot kernel/initramfs

Test Run tests Dev

Machine Machine
Debug or collect crash dump

7 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Virtual machines: best of both worlds!
● Easy to start/stop
● Full access to memory & CPU state
● Cross-architecture support using emulation
● Programmable hardware (e.g. error injection)

test.ko
Web Text
CRASH!
browser editor

Virtual machine

Development machine
8 KERNEL RECIPES 2015 | STEFAN HAJNOCZI
QEMU emulator and virtualizer

EMU Logo by Benoît Canet

Website: http://qemu-project.org/
Runs on Linux, Mac, BSD, and Windows
Emulates 17 CPU architectures (x86, arm, ppc, ...)
Supports fast hardware virtualization using KVM
Open source GPLv2 license

9 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

QEMU overview

Guest code runs in a virtual

machine
Hardware devices are Guest
emulated
QEMU performs I/O on behalf
of guest QEMU
QEMU appears as a normal
userspace process on the
host
Host kernel

10 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Cross-architecture emulation

Run another type of machine on your laptop

● qemu­system­arm
● qemu­system­ppc
● ...

Uses just-in-time compilation to achieve reasonable

speed
● Overhead can still be noticable

11 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Launching a virtual machine

Example with 1024 MB RAM and 2 CPUs:

qemu­system­x86_64 ­m 1024 \
                   ­smp 2 \
                   ­enable­kvm

Drop ­enable­kvm for emulation (e.g. ARM on x86)

Boots up to BIOS but there are no bootable drives...

12 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

QEMU virtual machine in BIOS/PXE

13 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

How to boot a development kernel

qemu­system­x86_64 ­enable­kvm ­m 1024 \
    ­kernel /boot/vmlinuz \
    ­initrd /boot/initramfs.img \
    ­append param1=value1

These options are similar to bootloader (GRUB, etc)

options.

14 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Small tests can run from initramfs

Initramfs can be customized to contain test programs

No need for full root file system

● Kick off tests from /init executable

Rebuild initramfs when kernel or test code changes

Result: Fast deployment & test

15 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Deploying kernel build products

arch/x86_64/boot/bzImage busybox

Custom init script & tests

initramfs
Kernel modules

qemu­system­x86_64 … \
    ­kernel vmlinuz \
    ­initrd initramfs.img \
    ­append param1=value1
16 KERNEL RECIPES 2015 | STEFAN HAJNOCZI
Building initramfs with gen_init_cpio

gen_init_cpio takes description file as input:

file /init my­init.sh 0755 0 0
dir /bin 0755 0 0
nod /dev/zero 0666 0 0 c 1 5
file /sbin/busybox /sbin/busybox 0755 0 0
slink /bin/sh /sbin/busybox 0755 0 0
Produces cpio archive as output:
$ usr/gen_init_cpio input | gzip >initramfs.img
Included in Linux source tree (usr/gen_init_cpio)

17 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Build process
Compile your kernel modules:
$ make M=drivers/virtio \
       CONFIG_VIRTIO_PCI=m modules
Build initramfs:
$ usr/gen_init_cpio input | gzip >initramfs.img
Run virtual machine:
$ qemu­system­x86_64 ­m 1024 ­enable­kvm \
          ­kernel arch/x86_64/boot/bzImage \
          ­initrd initramfs.img \
          ­append 'console=ttyS0' \
          ­nographic

18 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Using QEMU serial port for testing

I snuck in the QEMU ­nographic option

● Disables GUI
● Puts serial port onto stdin/stdout
● Perfect for running tests from terminal
● Easy to copy-paste error messages from output

Tell kernel to use console=ttyS0

19 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Challenges with manually built initramfs

Shared library dependencies must be found with

ldd(1) and added
Paths on the host may change across package
upgrades, breaking your initramfs build process
Rebuilding large initramfs every time is wasteful

Maybe it's time for a real root file system?

20 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Persistent root file system

Two options:
1)Share directory with host using virtfs or NFS
Pro: Easy to manipulate and inspect on host
2)Use disk image file with partition table and file
systems
Pro: Easy to install full Linux distro

Kernel can still be provided by ­kernel option.

Kernel modules need to be in initramfs and/or root file
system.

21 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Debugging a virtual machine

How do I inspect CPU registers and memory?

How do I set breakpoints on kernel code inside the

virtual machine?

QEMU supports GDB remote debugging to attach to

the virtual machine.

kgdb is not required inside virtual machine.

22 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Remote debugging != debugging QEMU

Often causes confusion:

If you want to debug what the virtual machine sees,

use remote debugging (gdbstub).

If you want to debug device emulation or QEMU

internals, use gdb -p $QEMU_PID.

23 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

GDB remote debugging
Protocol for remote debugging:
● Get/set CPU registers
● Load/store memory
Guest state
● Add/remove breakpoints
(CPU + RAM)
● Single-step and run

GDB stub
GDB
(client)
QEMU
24 KERNEL RECIPES 2015 | STEFAN HAJNOCZI
QEMU gdbstub example

qemu­system­x86_64 ­s ­enable­kvm ­m 1024 \
           ­drive if=virtio,file=test.img

(gdb) set architecture i386:x86­64
(gdb) file vmlinux
(gdb) target remote 127.0.0.1:1234
(gdb) backtrace
#0  native_safe_halt () at 
./arch/x86/include/asm/irqflags.h:50
#1  0xffffffff8101efae in arch_safe_halt ()
...

25 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Things to remember with remote debugging

Tell GDB which (sub-)architecture to use

● x86: 16-bit vs 32-bit vs 64-bit mode, check RFLAGS
register
● Careful with guest programs that switch modes!

Memory addresses are generally virtual addresses

(i.e. memory translation applies)

GDB doesn't know much about current userspace

process or swapped out pages!

26 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Device bring-up
Challenges for driver developers:
● Real hardware is not available yet
● Hardware is expensive
● Hardware/software co-development

How to develop & test drivers under these conditions?

1)Implement device emulation in QEMU

2)Develop driver against emulated device
3)Verify against real hardware when available

27 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

QEMU for device bring-up
Write C code in QEMU to emulate your device
● Out-of-tree solutions for hw simulation exist too!

QEMU device emulation covers common busses:

● PCI, USB, I2C

Examples where this approach was used:

● Rocker OpenFlow network switch
● NVDIMM persistent memory
● NVMe PCI flash storage controller
28 KERNEL RECIPES 2015 | STEFAN HAJNOCZI
QEMU device model
Object-oriented device model:
device

pci-device

e1000-base

e1000-82540em

Allows you to focus on unique device functionality

instead of common behavior.

29 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Memory API
Device register memory regions for PIO and MMIO
hardware register access:
static const MemoryRegionOps vmport_ops = {
    .read = vmport_ioport_read,
    .write = vmport_ioport_write,
    .impl = {
        .min_access_size = 4,
        .max_access_size = 4,
    },
    .endianness = DEVICE_LITTLE_ENDIAN,
};

memory_region_init_io(&s­>io, OBJECT(s),
         &vmport_ops, s, "vmport", 1);
isa_register_ioport(isadev, &s­>io, 0x5658);
30 KERNEL RECIPES 2015 | STEFAN HAJNOCZI
Interrupts

Devices use bus-specific methods to raise interrupts:

void pci_set_irq(PCIDevice *pci_dev, int level)

QEMU emulates interrupt controllers and injecting

interrupts
● Interrupt controller state is updated
● Guest CPU interrupt vector is taken

31 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

More information on device emulation

Plenty of examples in QEMU hw/ directory

● Learn from existing devices
● Documentation is sparse

Post patches to [email protected] for

feedback
● Guidelines for submitting patches:
http://qemu-project.org/Contribute/SubmitAPatch

32 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Error injection

How do I exercise rare error code paths in kernel?

QEMU can simulate error conditions

● Without overheating or damaging real hardware
● Without reaching into a box to pull cables

Simple scenarios:
● Test hot unplug while device is in use
(qemu) device_del e1000.0

33 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Advanced error injection

QEMU's block layer has an error injection engine:

[set­state]
state = "1"
event = "write_aio"
new_state = "2"
[inject­error]
state = "2"
event = "read_aio"
errno = "5"
This script fails disk reads after the first write.
Documentation: docs/blkdebug.txt

34 KERNEL RECIPES 2015 | STEFAN HAJNOCZI

Questions?

Email: [email protected]
IRC: stefanha on #qemu irc.oftc.net
Blog: http://blog.vmsplice.net/

QEMU: http://qemu-project.org/

Slides available on my website: http://vmsplice.net/

35 KERNEL RECIPES 2015 | STEFAN HAJNOCZI