H. Peter Anvin <hpa@zytor.com>
   Last update 2005-09-02

On the i386 platform, the Linux kernel uses a rather complicated boot
convention.  This has evolved partially due to historical aspects, as
well as the desire in the early days to have the kernel itself be a
bootable image, the complicated PC memory model and due to changed
expectations in the PC industry caused by the effective demise of
real-mode DOS as a mainstream operating system.

Currently, four versions of the Linux/i386 boot protocol exist.

Old kernels: zImage/Image support only.  Some very early kernels
  may not even support a command line.

Protocol 2.00: (Kernel 1.3.73) Added bzImage and initrd support, as
  well as a formalized way to communicate between the
  boot loader and the kernel.  setup.S made relocatable,
  although the traditional setup area still assumed
  writable.

Protocol 2.01: (Kernel 1.3.76) Added a heap overrun warning.

Protocol 2.02: (Kernel 2.4.0-test3-pre3) New command line protocol.
  Lower the conventional memory ceiling. No overwrite
  of the traditional setup area, thus making booting
  safe for systems which use the EBDA from SMM or 32-bit
  BIOS entry points.  zImage deprecated but still
  supported.

Protocol 2.03: (Kernel 2.4.18-pre1) Explicitly makes the highest possible
  initrd address available to the bootloader.

Protocol 2.04: (Kernel 2.6.14) Extend the syssize field to four bytes.

**** MEMORY LAYOUT

The traditional memory map for the kernel loader, used for Image or
zImage kernels, typically looks like:

 |    |
0A0000 +------------------------+
 |  Reserved for BIOS  | Do not use.  Reserved for BIOS EBDA.
09A000 +------------------------+
 |  Stack/heap/cmdline  | For use by the kernel real-mode code.
098000 +------------------------+ 
 |  Kernel setup   | The kernel real-mode code.
090200 +------------------------+
 |  Kernel boot sector  | The kernel legacy boot sector.
090000 +------------------------+
 |  Protected-mode kernel | The bulk of the kernel image.
010000 +------------------------+
 |  Boot loader   | <- Boot sector entry point 0000:7C00
001000 +------------------------+
 |  Reserved for MBR/BIOS |
000800 +------------------------+
 |  Typically used by MBR |
000600 +------------------------+
 |  BIOS use only  |
000000 +------------------------+

When using bzImage, the protected-mode kernel was relocated to
0x100000 ("high memory"), and the kernel real-mode block (boot sector,
setup, and stack/heap) was made relocatable to any address between
0x10000 and end of low memory. Unfortunately, in protocols 2.00 and
2.01 the command line is still required to live in the 0x9XXXX memory
range, and that memory range is still overwritten by the early kernel.
The 2.02 protocol resolves that problem.

It is desirable to keep the "memory ceiling" -- the highest point in
low memory touched by the boot loader -- as low as possible, since
some newer BIOSes have begun to allocate some rather large amounts of
memory, called the Extended BIOS Data Area, near the top of low
memory.  The boot loader should use the "INT 12h" BIOS call to verify
how much low memory is available.

Unfortunately, if INT 12h reports that the amount of memory is too
low, there is usually nothing the boot loader can do but to report an
error to the user.  The boot loader should therefore be designed to
take up as little space in low memory as it reasonably can.  For
zImage or old bzImage kernels, which need data written into the
0x90000 segment, the boot loader should make sure not to use memory
above the 0x9A000 point; too many BIOSes will break above that point.

**** THE REAL-MODE KERNEL HEADER

In the following text, and anywhere in the kernel boot sequence, "a
sector" refers to 512 bytes.  It is independent of the actual sector
size of the underlying medium.

The first step in loading a Linux kernel should be to load the
real-mode code (boot sector and setup code) and then examine the
following header at offset 0x01f1.  The real-mode code can total up to
32K, although the boot loader may choose to load only the first two
sectors (1K) and then examine the bootup sector size.

The header looks like:

Offset Proto Name  Meaning
/Size

01F1/1 ALL(1 setup_sects The size of the setup in sectors
01F2/2 ALL root_flags If set, the root is mounted readonly
01F4/4 2.04+(2 syssize  The size of the 32-bit code in 16-byte paras
01F8/2 ALL ram_size DO NOT USE - for bootsect.S use only
01FA/2 ALL vid_mode Video mode control
01FC/2 ALL root_dev Default root device number
01FE/2 ALL boot_flag 0xAA55 magic number
0200/2 2.00+ jump  Jump instruction
0202/4 2.00+ header  Magic signature "HdrS"
0206/2 2.00+ version  Boot protocol version supported
0208/4 2.00+ realmode_swtch Boot loader hook (see below)
020C/2 2.00+ start_sys The load-low segment (0x1000) (obsolete)
020E/2 2.00+ kernel_version Pointer to kernel version string
0210/1 2.00+ type_of_loader Boot loader identifier
0211/1 2.00+ loadflags Boot protocol option flags
0212/2 2.00+ setup_move_size Move to high memory size (used with hooks)
0214/4 2.00+ code32_start Boot loader hook (see below)
0218/4 2.00+ ramdisk_image initrd load address (set by boot loader)
021C/4 2.00+ ramdisk_size initrd size (set by boot loader)
0220/4 2.00+ bootsect_kludge DO NOT USE - for bootsect.S use only
0224/2 2.01+ heap_end_ptr Free memory after setup end
0226/2 N/A pad1  Unused
0228/4 2.02+ cmd_line_ptr 32-bit pointer to the kernel command line
022C/4 2.03+ initrd_addr_max Highest legal initrd address

(1) For backwards compatibility, if the setup_sects field contains 0, the
    real value is 4.

(2) For boot protocol prior to 2.04, the upper two bytes of the syssize
    field are unusable, which means the size of a bzImage kernel
    cannot be determined.

If the "HdrS" (0x53726448) magic number is not found at offset 0x202,
the boot protocol version is "old".  Loading an old kernel, the
following parameters should be assumed:

 Image type = zImage
 initrd not supported
 Real-mode kernel must be located at 0x90000.

Otherwise, the "version" field contains the protocol version,
e.g. protocol version 2.01 will contain 0x0201 in this field.  When
setting fields in the header, you must make sure only to set fields
supported by the protocol version in use.

The "kernel_version" field, if set to a nonzero value, contains a
pointer to a null-terminated human-readable kernel version number
string, less 0x200.  This can be used to display the kernel version to
the user.  This value should be less than (0x200*setup_sects).  For
example, if this value is set to 0x1c00, the kernel version number
string can be found at offset 0x1e00 in the kernel file.  This is a
valid value if and only if the "setup_sects" field contains the value
14 or higher.

Most boot loaders will simply load the kernel at its target address
directly.  Such boot loaders do not need to worry about filling in
most of the fields in the header.  The following fields should be
filled out, however:

  vid_mode:
 Please see the section on SPECIAL COMMAND LINE OPTIONS.

  type_of_loader:
 If your boot loader has an assigned id (see table below), enter
 0xTV here, where T is an identifier for the boot loader and V is
 a version number.  Otherwise, enter 0xFF here.

 Assigned boot loader ids:
 0  LILO
 1  Loadlin
 2  bootsect-loader
 3  SYSLINUX
 4  EtherBoot
 5  ELILO
 7  GRuB
 8  U-BOOT

 Please contact <hpa@zytor.com> if you need a bootloader ID
 value assigned.

  loadflags, heap_end_ptr:
 If the protocol version is 2.01 or higher, enter the
 offset limit of the setup heap into heap_end_ptr and set the
 0x80 bit (CAN_USE_HEAP) of loadflags.  heap_end_ptr appears to
 be relative to the start of setup (offset 0x0200).

  setup_move_size:
 When using protocol 2.00 or 2.01, if the real mode
 kernel is not loaded at 0x90000, it gets moved there later in
 the loading sequence.  Fill in this field if you want
 additional data (such as the kernel command line) moved in
 addition to the real-mode kernel itself.

  ramdisk_image, ramdisk_size:
 If your boot loader has loaded an initial ramdisk (initrd),
 set ramdisk_image to the 32-bit pointer to the ramdisk data
 and the ramdisk_size to the size of the ramdisk data.

 The initrd should typically be located as high in memory as
 possible, as it may otherwise get overwritten by the early
 kernel initialization sequence.  However, it must never be
 located above the address specified in the initrd_addr_max
 field. The initrd should be at least 4K page aligned.

  cmd_line_ptr:
 If the protocol version is 2.02 or higher, this is a 32-bit
 pointer to the kernel command line.  The kernel command line
 can be located anywhere between the end of setup and 0xA0000.
 Fill in this field even if your boot loader does not support a
 command line, in which case you can point this to an empty
 string (or better yet, to the string "auto".)  If this field
 is left at zero, the kernel will assume that your boot loader
 does not support the 2.02+ protocol.

  ramdisk_max:
 The maximum address that may be occupied by the initrd
 contents.  For boot protocols 2.02 or earlier, this field is
 not present, and the maximum address is 0x37FFFFFF.  (This
 address is defined as the address of the highest safe byte, so
 if your ramdisk is exactly 131072 bytes long and this field is
 0x37FFFFFF, you can start your ramdisk at 0x37FE0000.)

**** THE KERNEL COMMAND LINE

The kernel command line has become an important way for the boot
loader to communicate with the kernel.  Some of its options are also
relevant to the boot loader itself, see "special command line options"
below.

The kernel command line is a null-terminated string currently up to
255 characters long, plus the final null.  A string that is too long
will be automatically truncated by the kernel, a boot loader may allow
a longer command line to be passed to permit future kernels to extend
this limit.

If the boot protocol version is 2.02 or later, the address of the
kernel command line is given by the header field cmd_line_ptr (see
above.)  This address can be anywhere between the end of the setup
heap and 0xA0000.

If the protocol version is *not* 2.02 or higher, the kernel
command line is entered using the following protocol:

 At offset 0x0020 (word), "cmd_line_magic", enter the magic
 number 0xA33F.

 At offset 0x0022 (word), "cmd_line_offset", enter the offset
 of the kernel command line (relative to the start of the
 real-mode kernel).
 
 The kernel command line *must* be within the memory region
 covered by setup_move_size, so you may need to adjust this
 field.

**** SAMPLE BOOT CONFIGURATION

As a sample configuration, assume the following layout of the real
mode segment (this is a typical, and recommended layout):

 0x0000-0x7FFF Real mode kernel
 0x8000-0x8FFF Stack and heap
 0x9000-0x90FF Kernel command line

Such a boot loader should enter the following fields in the header:

 unsigned long base_ptr; /* base address for real-mode segment */

 if ( setup_sects == 0 ) {
  setup_sects = 4;
 }

 if ( protocol >= 0x0200 ) {
  type_of_loader = <type code>;
  if ( loading_initrd ) {
   ramdisk_image = <initrd_address>;
   ramdisk_size = <initrd_size>;
  }
  if ( protocol >= 0x0201 ) {
   heap_end_ptr = 0x9000 - 0x200;
   loadflags |= 0x80; /* CAN_USE_HEAP */
  }
  if ( protocol >= 0x0202 ) {
   cmd_line_ptr = base_ptr + 0x9000;
  } else {
   cmd_line_magic = 0xA33F;
   cmd_line_offset = 0x9000;
   setup_move_size = 0x9100;
  }
 } else {
  /* Very old kernel */

  cmd_line_magic = 0xA33F;
  cmd_line_offset = 0x9000;

  /* A very old kernel MUST have its real-mode code
     loaded at 0x90000 */

  if ( base_ptr != 0x90000 ) {
   /* Copy the real-mode kernel */
   memcpy(0x90000, base_ptr, (setup_sects+1)*512);
   /* Copy the command line */
   memcpy(0x99000, base_ptr+0x9000, 256);

   base_ptr = 0x90000;   /* Relocated */
  }

  /* It is recommended to clear memory up to the 32K mark */
  memset(0x90000 + (setup_sects+1)*512, 0,
         (64-(setup_sects+1))*512);
 }

**** LOADING THE REST OF THE KERNEL

The 32-bit (non-real-mode) kernel starts at offset (setup_sects+1)*512
in the kernel file (again, if setup_sects == 0 the real value is 4.)
It should be loaded at address 0x10000 for Image/zImage kernels and
0x100000 for bzImage kernels.

The kernel is a bzImage kernel if the protocol >= 2.00 and the 0x01
bit (LOAD_HIGH) in the loadflags field is set:

 is_bzImage = (protocol >= 0x0200) && (loadflags & 0x01);
 load_address = is_bzImage ? 0x100000 : 0x10000;

Note that Image/zImage kernels can be up to 512K in size, and thus use
the entire 0x10000-0x90000 range of memory.  This means it is pretty
much a requirement for these kernels to load the real-mode part at
0x90000.  bzImage kernels allow much more flexibility.

**** SPECIAL COMMAND LINE OPTIONS

If the command line provided by the boot loader is entered by the
user, the user may expect the following command line options to work.
They should normally not be deleted from the kernel command line even
though not all of them are actually meaningful to the kernel.  Boot
loader authors who need additional command line options for the boot
loader itself should get them registered in
Documentation/kernel-parameters.txt to make sure they will not
conflict with actual kernel options now or in the future.

  vga=<mode>
 <mode> here is either an integer (in C notation, either
 decimal, octal, or hexadecimal) or one of the strings
 "normal" (meaning 0xFFFF), "ext" (meaning 0xFFFE) or "ask"
 (meaning 0xFFFD).  This value should be entered into the
 vid_mode field, as it is used by the kernel before the command
 line is parsed.

  mem=<size>
 <size> is an integer in C notation optionally followed by K, M
 or G (meaning << 10, << 20 or << 30).  This specifies the end
 of memory to the kernel. This affects the possible placement
 of an initrd, since an initrd should be placed near end of
 memory.  Note that this is an option to *both* the kernel and
 the bootloader!

  initrd=<file>
 An initrd should be loaded.  The meaning of <file> is
 obviously bootloader-dependent, and some boot loaders
 (e.g. LILO) do not have such a command.

In addition, some boot loaders add the following options to the
user-specified command line:

  BOOT_IMAGE=<file>
 The boot image which was loaded.  Again, the meaning of <file>
 is obviously bootloader-dependent.

  auto
 The kernel was booted without explicit user intervention.

If these options are added by the boot loader, it is highly
recommended that they are located *first*, before the user-specified
or configuration-specified command line.  Otherwise, "init=/bin/sh"
gets confused by the "auto" option.

**** RUNNING THE KERNEL

The kernel is started by jumping to the kernel entry point, which is
located at *segment* offset 0x20 from the start of the real mode
kernel.  This means that if you loaded your real-mode kernel code at
0x90000, the kernel entry point is 9020:0000.

At entry, ds = es = ss should point to the start of the real-mode
kernel code (0x9000 if the code is loaded at 0x90000), sp should be
set up properly, normally pointing to the top of the heap, and
interrupts should be disabled.  Furthermore, to guard against bugs in
the kernel, it is recommended that the boot loader sets fs = gs = ds =
es = ss.

In our example from above, we would do:

 /* Note: in the case of the "old" kernel protocol, base_ptr must
    be == 0x90000 at this point; see the previous sample code */

 seg = base_ptr >> 4;

 cli(); /* Enter with interrupts disabled! */

 /* Set up the real-mode kernel stack */
 _SS = seg;
 _SP = 0x9000; /* Load SP immediately after loading SS! */

 _DS = _ES = _FS = _GS = seg;
 jmp_far(seg+0x20, 0); /* Run the kernel */

If your boot sector accesses a floppy drive, it is recommended to
switch off the floppy motor before running the kernel, since the
kernel boot leaves interrupts off and thus the motor will not be
switched off, especially if the loaded kernel has the floppy driver as
a demand-loaded module!

**** ADVANCED BOOT TIME HOOKS

If the boot loader runs in a particularly hostile environment (such as
LOADLIN, which runs under DOS) it may be impossible to follow the
standard memory location requirements.  Such a boot loader may use the
following hooks that, if set, are invoked by the kernel at the
appropriate time.  The use of these hooks should probably be
considered an absolutely last resort!

IMPORTANT: All the hooks are required to preserve %esp, %ebp, %esi and
%edi across invocation.

  realmode_swtch:
 A 16-bit real mode far subroutine invoked immediately before
 entering protected mode.  The default routine disables NMI, so
 your routine should probably do so, too.

  code32_start:
 A 32-bit flat-mode routine *jumped* to immediately after the
 transition to protected mode, but before the kernel is
 uncompressed.  No segments, except CS, are set up; you should
 set them up to KERNEL_DS (0x18) yourself.

 After completing your hook, you should jump to the address
 that was in this field before your boot loader overwrote it.