Linux: the first second

Linux: the first second
Alison Chaiken
alison@she-devel.com
January 25, 2018
All code for demos
Related blog post at opensource.com

Fast boot: US gov't requires reverse-video in 2s
Panic Concern ensues among automakers shipping Linux.

3
Slow boot: Linux boot on 8-bit AVR
“uARM is certainly no speed
demon. It takes about 2 hours
to boot to bash prompt”.
System:
8-bit micro,
external storage,
external RAM,
32-bit ARMv5 emulation.

How Linux starts
● What precisely does “off” mean?
● Fun with bootloaders
● ACPI vs DTB
● The kernel as PID 0
● How does PID 1 start?
● What is an initrd?
http://spartakirada.com/lawnmower-pullcord/

kernel/smp.c
smp_init()
Boot
ROM in
CPU
bootloader
(u-boot,
GRUB)
cpu_idle
head_64.S
“Kernel startup
entry point”
Decompress
start of
zImage
arch/arm/kernel/smp.c
secondary_start_kernel()
main.c start_kernel()
Boot
secondary
cores
init
kernel_init
do_initcalls()
rest_init()
device drivers
probe
devices
start
userspace
spawn
2nd
thread

6
Applying power
https://www.flickr.com/photos/nicknormal/11939558655/in/album-72157644792628117/

7
x86_64: Never genuinely off
Source: Intel
IPMI: run from Baseboard Management Controller
AMT: run from Platform Controller Hub

8
Source: Minnich et al., ELCE2017
on PCH
GRUB (x86)
or
u-boot (ARM)
{
on CPU

9
Purism, System76, Dell turn AMT off
Source: ExtremeTech, December 2017

11
Fun with u-boot's sandbox
(demo placeholder)
● How-to:
make ARCH=sandbox defconfig
make
./u-boot
● Even more fun:
make_test_disk.sh
file test.raw; gdisk -l test.raw
./u-boot
host bind 0 test.raw
printenv
gpt read host 0
fatls host 0:1
fdt addr $fdt_addr_
fdt header

Boot
ROM in
CPU
bootloader
(u-boot,
GRUB)
head_64.S
“Kernel startup
entry point”
Decompress
start of
zImage
How the system reaches the kernel initialization stage
Pass cmdline +
device-tree or ACPI

14
Kernel's “address book”: ACPI or Device-tree
● ACPI tables in SPI-NOR flash.
● At boot:
'dmesg | grep DT'
● Examine:
'acpidump | grep Windows'
● Get source: run iasl to extract
● Modify: boot-time 'BIOS' menu.
● device-tree in /boot.
● At boot:
each driver reads the DTB.
● Examine:
'strings /boot/foo.dtb'
● Get source: from kernel
● Modify: edit source, run dtc,
copy to /boot.

15
Starting up
the kernel
https://www.flickr.com/photos/nicknormal/11939856433/in/album-72157644792628117/

16
The kernel is an ELF binary
● Extract vmlinux from vmlinuz:
– <path-to-kernel-source>/scripts/extract-vmlinux
/boot/vmlinuz-$(uname -r) > vmlinux
● vmlinux is a regular ELF binary:
– file vmlinux; file /bin/ls
– readelf -e vmlinux; readelf -e /bin/ls
https://flic.kr/p/6xuhiK

17
Quiz:
How do ELF binaries
start?
?
?
?
?
?? ??
?

18
Quiz:
Where do argc and argv
come from?
?
?
?
?
?? ??
?

19
demo placeholder
Inspecting the start of ls with GDB

20
Examining ELF binary start with GDB
(results depend on toolchain and libc)
● Compile your C program with '-ggdb'.
● gdb <some-binary-executable>
● set backtrace past-main
on
● set backtrace past-entry
on
● Type 'run'
● frame 1; list
● Type 'info files'
● Find 'Entry point'.
● Type 'l *(hex address)'
● Type 'l 1,80'
● Type 'info functions' or
'info sources'
demo placeholder

21
The kernel as PID 0
● Userspace processes need to start need:
– stack,
– heap,
– STD* file descriptors
– environment
● glibc and libgcc allocate these resources.
– Source is in start.S (ARM) and libc-start.c.
● Corresponding kernel resources provided via inline ASM.
– Reads cmdline, device-tree or ACPI.

22
Examining ARM32 kernel start with GDB
(demo placeholder)
1 Type 'file vmlinux'. (If zImage, extract with linux/scripts/extract-vmlinux).
2 Type:
arm-linux-gnueabihf-gdb vmlinux
3 Type:
info files
4 Find 'Entry point'.
5 Type:
l *(hex address)
6 Type
l 1,80

23
What's in ARM's head.S?
● Type 'file vmlinux.o'
● Try 'arm-linux-gnueabihf-gdb vmlinux.o'
● Type 'info files'
● Type 'l *(0x0)' <---- actually works!
demo placeholder
Kernel starts in head.S,
not start.S.

24
Examining x86_64 kernel with GDB
(demo placeholder)
1 Type 'file vmlinux'. (If zImage, extract with linux/scripts/extract-vmlinux).
2 Type:
gdb vmlinux
3 Type:
info files
4 Find '.init.text'.
5 Type:
l *(hex address)
6 Type
l 200,290

25
What's in x86_64 head_64.S?
demo placeholder

26
The kernel's main() function
● start_kernel() {:
boot_cpu_init();
setup_arch(&command_line);
page_alloc_init();
pr_notice("Kernel command line: );
mm_init();
sched_init();
init_IRQ();
init_timers(); timekeeping_init();
console_init();
rest_init();
● }
“Activate the first processor.”
process the device-tree
All timestamps before
are [0.000000]
setup page tables
and start virtual memory
start
userspace
All on
one core!

cpu_idle
Boot
ROM in
CPU
bootloader
(u-boot,
GRUB)
head_64.S
“Kernel startup
entry point”
Decompress
start of
zImage
kernel_init
rest_init()
Finish core kernel bringup
spawn
2nd
thread

kernel/smp.c
smp_init()
Boot
ROM in
CPU
bootloader
(u-boot,
GRUB)
cpu_idle
head_64.S
“Kernel startup
entry point”
Decompress
start of
zImage
kernel_init
rest_init()
Boot
secondary
cores
Bring up symmetric
multiprocessing (SMP)
spawn
2nd
thread

29
Kernel boot via BCC
Stack for CPU0
Stack for 2nd
core
x86_64_start_kernel: head_64.S
demo placeholder

kernel/smp.c
smp_init()
Boot
ROM in
CPU
bootloader
(u-boot,
GRUB)
cpu_idle
head_64.S
“Kernel startup
entry point”
Decompress
spawn
2nd
thread
start of
zImage
kernel_init
do_initcalls()
rest_init()
device drivers
probe
devices
Boot
secondary
cores
Bring up
devices,
filesystems . . .

kernel/smp.c
smp_init()
Boot
ROM in
CPU
bootloader
(u-boot,
GRUB)
cpu_idle
head_64.S
“Kernel startup
entry point”
Decompress
init
start of
zImage
kernel_init
do_initcalls()
rest_init()
device drivers
probe
devices
start
userspace
Boot
secondary
cores
Finally,
userspace.
spawn
2nd
thread

kernel/smp.c
smp_init()
Boot
ROM in
CPU
bootloader
(u-boot,
GRUB)
cpu_idle
head_64.S
“Kernel startup
entry point”
Decompress
init
start of
zImage
kernel_init
do_initcalls()
rest_init()
device drivers
probe
devices
start
userspace
Boot
secondary
cores
spawn
2nd
thread

33
Summary
● Practicing with u-boot sandbox is comparatively
relaxing.
● Viewing the kernel as ELF helps to understand early
boot.
● Several processors and SW components participate in
boot.
● Until the scheduler and SMP start, the boot process is
relatively simple.

34
Acknowledgements
● Big thanks to Joel Fernandes and Akkana Peck for suggestions.
● Shout-out to Linaro for making ARM so much easier than x86.

35
Major References
● Embedded Linux Primer by Chris Hallinan and
Essential Linux Device Drivers by Sreekrishnan Venkateswaran (books)
● Booting ARM Linux by Russell King and THE LINUX/x86 BOOT PROTOCOL
(Documentation/)
● Program startup process in userspace at linux-insides blog, Michael Kerrisk's
TLPI (book)
● Matthew Garrett's comprehensive series on UEFI
● Status of Intel Management Engine on various laptops (Coreboot) and
servers (FSF)
● Nov, 2017 Intel Management Engine exploits and vulnerability detection tool
● All about ACPI talk by Darren Hart, ELCE 2013, Arch Wiki on
hacking ACPI tables
● 'apt-get install debian-kernel-handbook'; GDB docs chapter 8

36
Cold-boot may become rare
Source: Micron
● Non-volatile RAM  suspend even for brief
inactivity.
● Minimal diff between 'suspend' and 'hibernate'?
● Linux drivers: Matthew Wilcox, XIP DAX
AKA,
'Optane'
by Intel
Specs:
ArsTechnica

39
What is an initrd anyway?
● 'init ramdisk' = filesystem that is loaded into memory by the
kernel before the rootfs mounts.
● Why?
– To provide a 'rescue shell' in case rootfs doesn't mount.
– To provide modules that don't fit in zImage.
– To provide a safe environment to run agressive tests.
– To facilitate software updates on devices with limited
storage.

41
What's in an initrd and why?
● Boot into the rescue shell by providing a broken cmdline in
/boot/grub/grub.cfg
– Type 'ls'
● Or try 'lsinitramfs /boot/$(uname -r)'
● initrd is a gzipped cpio archive:
cp /boot/initrd-$(uname -r) /tmp/initrd.gz
gunzip /tmp/initrd.gz
cpio -t < /tmp/initrd

42
OMG! My life is over! (rescue shell tips)
Inhale on a 4-count, then exhale on a 10-count.
● Oh no! 'help' scrolls pages of unreadable crap!
Relax your jaw. Make circles with your neck.
● Read 'man busybox'.
● 'help | grep' works in busybox.
● Look in /bin and /sbin. There's fsck!!
● You have sed and vi (but not emacs ;-( )
● Type 'reboot -f' or 'exit' when you are bored.

43
How to create your own initrd
● Unpack one that already works with gunzip and 'cpio -i'
● Copy in your binary.
● Use gen_initramfs.h from kernel source tree:
– scripts/gen_initramfs_list.sh -o <archive> <path to source>
● Run 'lsinitramfs <archive>' to check the result.
● cp <archive> /boot; edit /boot/grub/grub.cfg
CAUTION: your system boots fine, right? You're crazy to mess
with the bootloader, you moron.
● Run grub-script-check.

45
The Lenovo laptop on which the slides were created has
known IME vulnerabilities described by unpatched CVEs. This
has nothing to do with Meltdown and Spectre.

46
Bootloaders according to Intel
Read
device-tree
(ARM)

47
Coming soon to a system near youSource:
Anandtech

48
Investigating your laptop's PCH
● Try:
lsmod | grep pch
● Try:
find /lib/modules/$(uname -r)/ -name “*pch*”
● Then (for example):
EG20T = Intel Topcliff PCH

49
Why bootloaders have two parts
● ARM: “SPL”, “XLoader” or “MLO” in addition to u-boot.img.
● Problem: DRAM controller must be initialized.
● Solution: load into SRAM ('OCRAM' in i.MX6, 'l2ram' for TI).
– Why this works: SRAM (and pNOR) are mapped memory.
● Problem: SRAM is little! (256K on i.MX6, 2 MB on DRA7x).
● Solution: start with a tiny SPL.

50
Warm vs. power-on reset
Clears memory?
Restarts clocks?
Pros Cons Examples
Power-on Reset Yes, then reads
boot-mode pins.
Won't fail. Slightly slower. Plug-in device
Warm Reset DDR set to 'self-
refresh', then
reset clocks and
jump to stored
address.
Faster; retains
'reset reason'
and RAM data.
Can fail. 'reboot';
watchdog;
JTAG

51
Advanced Configuration and Power Interface
Source: Intel

do_bootm_states = u-boot state machine
bootm.c<lib>
“Main Entry point for arm
bootm implementation”
BootROM
(internal EEPROM)
resume?
Read Reset
Controller registers
No
Yes
Jump to
stored image
bootm_start() bootm_find_os()
bootm_load_os() bootm_os_get_boot_func()
boot_selected_os
do_bootm_linux()
bootm_jump_linux()
bootm.c<common>
Das U-boot
***

53
Where do messages originate?
[ 54.590327] Starting kernel ...
[ 54.593459]
Uncompressing Linux... done, booting the kernel.
Linux version 3.0.35-2508-g54750ff (gcc version 4.6.3 #1 SMP PREEMPT
CPU: ARMv7 Processor [412fc09a] revision 10 (ARMv7), cr=10c53c7d
CPU: VIPT nonaliasing data cache, VIPT aliasing instruction cache
Machine: Freescale i.MX 6Quad/DualLite/Solo Sabre-SD Board
Memory policy: ECC disabled, Data cache writealloc
CPU identified as i.MX6Q, silicon rev 1.1
PERCPU: Embedded 7 pages/cpu @8c008000 s5440 r8192 d15040 u32768
Built 1 zonelists in Zone order, mobility grouping on. Total pages: 227328
Kernel command line: console=ttymxc1,115200 ip=dhcp rootwait root=/dev/nfs
nfsroot=172.17.0.1:/tftpboot/alison/mx6q/fsl-mx6,v3,tcp
u-boot misc.c
in
zImage
header
from kernel
proper
from CPU
from
kernel
passed from u-boot
}

54
Getting more detailed kernel messages at boot
● Remove 'quiet' from the kernel command line.
● How to keep 'quiet' from coming back:
– edit /etc/grub.d/10_linux and add:
export GRUB_DISABLE_SUBMENU=y
export GRUB_CMDLINE_LINUX_DEFAULT=""
CAUTION: your system boots fine, right? You're crazy to mess
with the bootloader, you moron.
● Always run 'grub-script-check /boot/grub/grub.cfg' afterwards.

55
Learning more with systemd-bootchart
● Make sure kernel is compiled with CONFIG_SCHEDSTATS=y.
● 'apt-get install systemd-bootchart'
● Interrupt grub by typing 'e'
● Append 'init=/lib/systemd/systemd-bootchart' to the line that
starts with 'linux'
● After boot, open the SVG image in /run/log/ with a browser.

A change in compiling your own kernel

57
Appendix: running QEMU
#!/bin/bash
ROOTDIR=/home/alison/ISOs
HDNAME=debian-testing
VERSION=4.9.5
# Load kernel via GRUB; console shows in QEMU window.
#qemu-system-x86_64 -machine accel=kvm -name ${HDNAME} -boot c -drive file=$
{ROOTDIR}/${HDNAME}.raw,format=raw -m 4096 -smp cpus=1 -net nic,model=e1000
-net user,hostfwd=tcp:127.0.0.1:6666-:22 -localtime -serial stdio
# Load kernel from external file; console shows in xterm; GRUB doesn't run.
qemu-system-x86_64 -machine accel=kvm -name ${HDNAME} -initrd
/home/alison/embedded/SCALE2017/kernel/initrd.img-${VERSION} -kernel
/home/alison/embedded/SCALE2017/kernel/vmlinuz-${VERSION} -boot c -drive file=$
{ROOTDIR}/${HDNAME}.raw,format=raw -m 4096 -smp cpus=1 -net nic,model=e1000
-net user,hostfwd=tcp:127.0.0.1:6666-:22 -localtime -serial stdio -append
"console=ttyAMA0 console=ttyS0 root=UUID=8e6a1c7e-b3c4-4a37-8e21-56a137c9dded
ro"

The ARM bootloader
● Read fundamental configuration from fuses, switches and
GPIOs.
● Then, for ARM:
1. Setup and initialise the RAM.
2. Initialise one serial port.
3. Detect the machine type.
4. Setup the kernel tagged list. device-tree
5. Load initramfs.
6. Call the kernel image.
Code in the SPL: board_init_f() and jump_to_image_linux()

60
Image, zImage, uImage, vmlinux, vmlinuz?
● Image is the raw executable.
● zImage is compressed version of Image with prepended
uncompression instructions in ASM.
● uImage is a zImage with a u-boot header.
● vmlinux is ELF executable containing Image in .text section.
● vmlinuz is a stripped version of vmlinux.

61Source: https://recon.cx/2014/slides/Recon%202014%20Skochinsky.pdf
Older x86 had ARC processor as
GMCH instead of PCH.

Linux: the first second

More Related Content

What's hot (20)

Similar to Linux: the first second (20)

More from Alison Chaiken (20)

Recently uploaded (20)

Linux: the first second