Grub2 Booting Process

Grub2 Booting Process
Mike Wang
May 2nd, 2016

Agenda
 Scope
 Classical booting process (MBR booting)
 Build grub2 image
 UEFI booting
 Booting failure analysis

Scope
 Grub2 supports many target, but we will only focus on:
 i386-pc
 X86 CPU, BIOSbased, Local disk boot
 Linux distributors use it for both 32-bit system and 64-bit system
 x86_64-efi
 X86 64-bit CPU, UEFI based, Local disk boot
 Not include XEN, not include PXE
 Not include multiboot
 Reference
 grub source code 2.02 beta3
 Linux distributions
 RHEL 7
 SLES 12
 Ubuntu 16.04 LTS

BIOS
 Intel x86 CPU always start (in real mode) by running the instructions at
0xFFFFFFF0 (mapping to ROM)
 That’s the entry point of BIOS code
 BIOS goes through a list of pre-configuredboot devices, until it finds a
bootable device
 Bootable device is the one which last two bytes of first sector contains boot
signature 0xAA55
 BIOS read the first sector to memory address 0x7C00
 Error message depends on BIOS vendor
 No operating system
 Operating system not found
 Booting failure ...

MBR
 Grub image file boot.img
 Code is at offset 0x65
 Load 1st sector of core image (diskboot.img) at memory address 0x8000
 INT13H is used to read disk (try LBA, and then CHS)
 Contains boot information
offset Length (bytes) description
0x5a 2 Memory address to execute 1st sector of core
image
0x5c 8 Where to load 1st sector of core image
0x64 1 The disk to load core image. 0xff means
current boot disk.
0x1be 64 Partition table
Check MBR: hexdump –C /dev/sda | more

MBR - message
message type description
GRUB information Print this message at the beginning of
grub start
(Ubuntu: shown when key “shift” is
pressed)
(RedHat: not show anything)
Hard Disk Error error Not be able to get geometry
information from hard disk. CHS only.
Geom Error error Invalid valuefor start sector address.
CHS only.
Read Error error Failedto read hard disk. CHS only.
Note 1: INT13 CHS mode may only read first 8GB of disk (or 137GB)
Note 2: INT13 needs buffer must be in one memory segment.
Note 3: It may run into CHS mode because start sector exceeds disk size which
causes LBA mode fail.

Core image
 Grub image file core.img
Core image maximum size
relies on post-MBR gap size:
• Traditional: 63 sectors
• Today: 2048 sectors

Boot sector of core image
 Grub image file diskboot.img
 Start at memory address 0x8000
 Load core image based on block list to memory address 0x8200
 Grub-bios-setup (grub-install) will update block list for the area which
core.img is actually installed
 Continuous area is not a must

Boot sector of core image – disk layout

Boot sector of core image - message
message type description
Loading… information Print "loading" at the beginning of the
code
print one dot after load one block
print new line after load all blocks
(Ubuntu: shown when key “shift” is
pressed)
(RedHat: not print "Loading")
Geom Error error The value of block list is invalid. CHS only.
Read Error error Failedto read hard disk

Decompressor of Core Image
 Grub image file:lzma_decompress.img
 Start at memory address 0x8200
 Not really only for decompressing
 Transition to 32-bit protected mode
 Error correction using Reed-Solomon
 Decompress
 Decompress data to 0x100000, and then copy kernel.img code(not include
modules and its later) to 0x9000

Decompressor of Core Image – disk
layout
Note: Post MBR gap is 2048 sectors.

Core Image – kernel.img
 Load embedded modules
 Set “prefix” and “root”
 They are used when load external modules
 “prefix” can be derived from boot drive, if it’s not set explicitly
 “root” is the device part of “prefix”
 Register core commands – commands which do not rely on external modules
 Only four: set, unset, ls, insmod
 Parse early configure file
 Load “normal” module (and its dependencies) and run it
 When command “normal” is executed:
 Load grub external script file grub.cfg
 Show menu interface
 If something goes wrong,run into “rescue” mode

Grub2 modules
 Folder /boot/grub2/<target_name>/
 For example, /boot/grub2/i386-pc
 External modules are not recommended for UEFI boot
 Much like Linux kernel module
 ELF file format
 There is “init” and “fini” function
 Check license type during loading
 Register commands in “init” function
 Contains dependency information
 So its dependent modules can be loaded automatically
 Important files
 moddep.lst: dependency relationship, used by grub-install, grub-mkimage
 command.lst: command-module relationship. So you can use a command without
need “insmod”
 modinfo.sh: grub compile information, include version number

Core Image – boot Linux kernel
 Linux kernel file contains three parts
 Boot sector
 Setup code (16-bit real mode)
 Protected mode code (vmlinux)
 Use command “linux” to load Linux kernel
 32-bit boot protocol is used
 Grub will prepare environment for Linux kernel, and load protected mode code
 Jump to Linux kernel protected mode code when boots
 Use command “linux16” to load Linux kernel
 Load both setup code and protected mode code
 Jump to setup code (means go back to real mode) when boots
 Kernel setup code will prepare environment, and then run protected mode code
 For 64-bit kernel, it's kernel's code enters into long mode

SUSE is different
 MBR code is from syslinux
 Relocate itself to 0x0000:0x0600
 Scan active partition to load its first sector to 0x0000:0x7c00
 Jump to 0x7c00 to continue execute
 The 1st sector of active partition (Volume Boot Record, VBR)
 It’s grub’s MBR code
 Error message
Multiple active partitions.rn When multiple partitions are in active state
Operating system load error.rn Failed to read disk
Missing operating system.rn Not found an active partition to boot

Create and install grub
 Grub-mkimage
 Create core image based on input
 Compress algorithm
 For i386-pc-*, LZMA implemented by grub
 For x86_64-efi, no compress is used
 Grub-bios-setup
 Install MBR
 Avoid overwrite BPB area (used by Windows NT)
 If it’s hard disk, replace offset 0x66 with two 0x90 (workaround for buggy BIOS firmware)
 Avoid overwriting Windows NT magic code and partition table
 Install core image
 Update block list in boot sector
 Update MBR for where boot sector can be accessed
 Update Reed-Solomon redundancy field if it’s used
 Grub-install
 Higher level command includes functionsof grub-mkimage and grub-bios-setup
 Decides which modules should be embedded based on configuration of current machine

Store core image in file system?
 Not support cases
 Software RAID/LVM
 Certain file systems (e.g. BtrFS, ZFS)
 Cross disk install(/boot being on one disk but MBR on another)
 Show warning to user that this is not reliable and is discouraged
 Only continue if user has specified "--force" option
 Get block list of file core.img and store it in boot sector
 Modify MBR to point to boot sector

Core Image – which modules should
be embedded?
 Include:
 Access disk
 Parse partition
 Parse volume manager(if applied)
 Parse file system
 And the dependencies of above
 For example (for target i386-pc, BIOS based, local disk boot):
 Biosdisk (INT13H is used)
 Part_msdos
 ext2

Example prefix
 For simple volume
 (hd0,1)/boot/grub
 For LVM
 (lvm/lvm_group_name-lvm_logical_boot_partition_name)/boot/grub
 (lv/system-root)/boot/grub
 For Software RAID
 (md/md0)/boot/grub
 (mduuid/123456789abcdef0123456789abcdef0)/boot/grub
 Or a better way
 search.fs_uuid 01234567-89ab-cdef-0123-456789abcdef root
 set prefix=($root)/boot/grub

UEFI Booting
For target “x86_64-efi”

UEFI boot
 Firmware enables protected mode
 Firmware understands partition table and FAT file system
 Try to find GPT partition with GUID “C12A7328-F81F-11D2-BA4B-
00A0C93EC93B” or MBR type 0xEF.
 The partition should contains FAT file system
 Try to find boot loader from
 The default file: /EFI/BOOT/BOOTX64.EFI
 But it can be changed using command “efibootmgr”
 Two stages booting is used because of Secure Boot requirement
 Different vendor uses different boot loader
 Red Hat, SUSE 12.x: grub2
 SUSE 11.x: elilo

UEFI boot – boot entry
Use “blkid” to get PartUUID for your partitions.
You can also get some information from /sys/firmware/efi/

UEFI Boot – boot files
 Shim.efi is signed by Microsoft’s KEK, and can be verify by the certificate in UEFI firmware. It
contains OS vendor’s public key
 Try to load 2nd stage boot loader (grubx64.efi) by using function provided by UEFI
firmware
 If it’s failed, it will verify signature of 2nd stage boot loader by itself, and then load and
run it
 grubx64.efi (name cannot be changed because it’s hardcoded in shim.efi) contains:
 1KB PE header
 Signature
 Core image data without compressing
 Additional module “linuxefi” (not in upstream)
 Think about what modules should be embedded?
 Answer is: all (because grub2 modules has no signature, and cannot be
verified)
 Call back to shim to verifylinux kernel signature

UEFI Boot – boot files
 MokManager.efi
 Machine Owner Key Manager
 Add trust to secure boot
 Can add as root key
 Command “mokutil” can be used to register key
 The key import request is recognized by shim.efi, and it will call MokManager.efi
 Need reboot and a physical user’s interaction
 Use command “pesign -S –I <file>” to view signature
 You will find there is bug in CentOS 7.1 where shim.efi is signed by Red Hat Inc.

Secure Boot
 Platform Key
 Control key update
 Installed in firmware by hardware
vendor
 Key Exchanged Key
 Used to verify boot loader
 Microsoft's key is generally
installed by default
 Machine Owner Key
 No need trust relationship with PK
or KEK
Ref: https://www.suse.com/documentation/sles-12/book_sle_admin/data/sec_uefi_secboot.html

Debug MBR/Boot Sector
 Find clue from error message
 Consult source code
 Reverseopcode to ASM
 Try to skip data area, otherwise you may get wrong result sometimes
objdump -D -b binary -mi386 -Maddr16,data16 mbr.img
 Modify opcode for debugging purpose
ASM code Opcode Description
nop 0x90 Nooperation
jmpN 0xeb0x(N-2) JumpoffsetN tocurrent address
ljmp<segment><ADDR> 0xea0x(ADDR_LOW) 0x(ADDR_HIGH) 0x(SEG_LOW) 0x(SEG_HIGH) Long jumptospecifiedaddress
call <ADDR> 0xe8 0x(OFFSET_LOW) 0x(OFFSET_HIGH) Call a functionatADDR. OFFSET=(ADD-
CURRENT_ADDR) - 3
ret 0xc3 Returnfromfunctioncall
echo -ne "x90x90x90" | dd of=/dev/sda bs=1 seek=6 conv=notrunc

Debug MBR/Boot Sector – qemu + gdb
 Don’t enable KVM
 Skip interrupt call (e.g. INT13H, break at its next instruction)
 ASM may not correct
 Use “nexti” or “stepi”
 If you use partial image file, be care of its size is valid for MBR code
 Qemu’s BIOS supports LBA mode (is it the same as your source machine?)
Qemu:
qemu-system-x86_64 -m 16M -drive file=/dev/sdb,format=raw,if=scsi,readonly -
gdb tcp::1234 -nographic -vnc :0
Gdb:
set architecture i8086
target remote localhost:1234
break *0x7c00
layout asm
x /16bx 0x7df0 #print memory content
set *0x7df0 = 10 #modify memory content
info registers eax

Figure out where core image is
 Where is the first sector of core image?
 In MBR (mentioned previously)
 Get partition boundary
parted /dev/sda "unit s print"
parted /dev/sda "unit chs print"
sfdisk -u S -l /dev/sda
 File system boundary
File System SuperBlock Offset
(to device)
Magic NumberOffset
(to superblock)
Magic NumberContent
Ext4 0x400 (1KB) 0x38 0xEF53
XFS 0x00 0x00 0x58465342("XFSB")
BTRFS 0x10000 (64KB) 0x28 "_BHRfS_M"
ReiserFS 0x10000 (64KB) 0x34 "ReIsEr2Fs"
filefrag -v /file
 File blocks mapping

Debug in grub environment
 Show debug message
set pager=1 pauseoutput after each screenfuland wait for keyboard input
set debug=all Candidate value can also be specific module, e.g. disk,linux,partition,modules
 Test case: recover from grub boot failure
 Ubuntu 16.04 LTS
 /boot is a separated volume
 Rename folder i386-pc to i386-pc.org.

Grub menu will show up. But
we cannot use the menu to
boot linux kernel because
many modules are missing.
Press "c" to enter command
line mode.
Linux kernel is booted now.

Grub2 Booting Process

More Related Content

What's hot

Viewers also liked

Similar to Grub2 Booting Process

Recently uploaded

In this document

Grub2 Booting Process