Booting/linux-bootstrap-3.md
In the previous part, we have seen first pieces of C code that run in the Linux kernel. One of the main goal of this stage is to switch into the protected mode, but before this, we have seen some early setup code which executes early initialization procedures, such as:
In this part, we continue to explore the next steps before transitioning to protected mode.
Previously, we stopped right at the point where the kernel setup code was about to initialize the video mode.
The setup code is located in arch/x86/boot/video.c and implemented by the set_video function. Now let's take a look at the implementation of the set_video function:
void set_video(void)
{
u16 mode = boot_params.hdr.vid_mode;
RESET_HEAP();
store_mode_params();
save_screen();
probe_cards(0);
for (;;) {
if (mode == ASK_VGA)
mode = mode_menu();
if (!set_mode(mode))
break;
printf("Undefined video mode number: %x\n", mode);
mode = ASK_VGA;
}
boot_params.hdr.vid_mode = mode;
vesa_store_edid();
store_mode_params();
if (do_restore)
restore_screen();
}
In the next section, let's try to understand what a video mode is and how this function initializes it.
A video mode is a predefined configuration of a screen that tells the video hardware information about:
The next goal of the kernel is to collect this information and initialize a suitable video mode. This allows the kernel to use a special API to print messages on the screen.
The implementation of the set_video function starts by getting the video mode from the boot_params.hdr structure:
u16 mode = boot_params.hdr.vid_mode;
[!NOTE] Instead of old good standard C data types like
int,short,unsigned short, Linux kernel provides own data types for numeric values. Here is the table that will help you to remember them:
Type char short int long u8 u16 u32 u64 Size 1 2 4 8 1 2 4 8
The initial value of the video mode can be filled by the bootloader. This header field defined in the Linux kernel boot protocol:
Offset Proto Name Meaning
/Size
01FA/2 ALL vid_mode Video mode control
Information about potential values for this field can be also found in the Linux kernel boot protocol document:
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.
This tells us that we can add the vga option to the kernel's command line. As mentioned in the description above, this option can have different values. For example, it can be an integer number 0xFFFD or ask. If you pass ask to vga, you see a menu with the possible video modes. We can test it using QEMU virtual machine as we did in the previous chapters:
sudo qemu-system-x86_64 -kernel ./linux/arch/x86/boot/bzImage \
-nographic \
-append "console=ttyS0 nokaslr vga=ask" \
-initrd /boot/initramfs-6.17.0-rc3-g1b237f190eb3.img
If you did everything correctly, after the kernel is loaded it will ask you to press the ENTER. By pressing on it you should see something like this:
Booting from ROM...
Probing EDD (edd=off to disable)... ok
Press <ENTER> to see video modes available, <SPACE> to continue, or wait 30 sec
Mode: Resolution: Type: Mode: Resolution: Type: Mode: Resolution: Type:
0 F00 80x25 VGA 1 F01 80x50 VGA 2 F02 80x43 VGA
3 F03 80x28 VGA 4 F05 80x30 VGA 5 F06 80x34 VGA
6 F07 80x60 VGA 7 340 320x200x32 VESA 8 341 640x400x32 VESA
9 342 640x480x32 VESA a 343 800x600x32 VESA b 344 1024x768x32 VESA
c 345 1280x1024x32 VESA d 347 1600x1200x32 VESA e 34C 1152x864x32 VESA
f 377 1280x768x32 VESA g 37A 1280x800x32 VESA h 37D 1280x960x32 VESA
i 380 1440x900x32 VESA j 383 1400x1050x32 VESA k 386 1680x1050x32 VESA
l 389 1920x1200x32 VESA m 38C 2560x1600x32 VESA n 38F 1280x720x32 VESA
o 392 1920x1080x32 VESA p 300 640x400x8 VESA q 301 640x480x8 VESA
r 303 800x600x8 VESA s 305 1024x768x8 VESA t 307 1280x1024x8 VESA
u 30D 320x200x15 VESA v 30E 320x200x16 VESA w 30F 320x200x24 VESA
x 310 640x480x15 VESA y 311 640x480x16 VESA z 312 640x480x24 VESA
313 800x600x15 VESA 314 800x600x16 VESA 315 800x600x24 VESA
316 1024x768x15 VESA 317 1024x768x16 VESA 318 1024x768x24 VESA
319 1280x1024x15 VESA 31A 1280x1024x16 VESA 31B 1280x1024x24 VESA
31C 1600x1200x8 VESA 31D 1600x1200x15 VESA 31E 1600x1200x16 VESA
31F 1600x1200x24 VESA 346 320x200x8 VESA 348 1152x864x8 VESA
349 1152x864x15 VESA 34A 1152x864x16 VESA 34B 1152x864x24 VESA
375 1280x768x16 VESA 376 1280x768x24 VESA 378 1280x800x16 VESA
379 1280x800x24 VESA 37B 1280x960x16 VESA 37C 1280x960x24 VESA
37E 1440x900x16 VESA 37F 1440x900x24 VESA 381 1400x1050x16 VESA
382 1400x1050x24 VESA 384 1680x1050x16 VESA 385 1680x1050x24 VESA
387 1920x1200x16 VESA 388 1920x1200x24 VESA 38A 2560x1600x16 VESA
38B 2560x1600x24 VESA 38D 1280x720x16 VESA 38E 1280x720x24 VESA
390 1920x1080x16 VESA 391 1920x1080x24 VESA 393 1600x900x16 VESA
394 1600x900x24 VESA 395 1600x900x32 VESA 396 2560x1440x16 VESA
397 2560x1440x24 VESA 398 2560x1440x32 VESA 399 3840x2160x16 VESA
200 40x25 VESA 201 40x25 VESA 202 80x25 VESA
203 80x25 VESA 207 80x25 VESA 213 320x200x8 VESA
Enter a video mode or "scan" to scan for additional modes:
Before proceeding further to investigate what the set_video function does, it will be useful to take a look at the API for the management of the kernel's early heap.
After getting the video mode set by the bootloader, we can see resetting the heap value by the RESET_HEAP macro. The definition of this macro is in the arch/x86/boot/boot.h:
#define RESET_HEAP() ((void *)( HEAP = _end ))
If you have read part 2, you should remember the initialization of the heap memory area. This memory area starts right after the end of BSS and lasts till the stack.
The kernel setup code provides a couple of utility macros and functions for managing the early heap. Let's take a look at some of them, especially at those relevant for this chapter.
The RESET_HEAP macro resets the heap by setting the HEAP variable to _end, which represents the end of the early setup kernel's image, including the early code, data, and BSS memory areas. By doing this, we set the heap pointer to the very beginning of the heap.
The next useful macro is:
<!-- https://raw.githubusercontent.com/torvalds/linux/refs/heads/master/arch/x86/boot/boot.h#L184-L185 -->#define GET_HEAP(type, n) \
((type *)__get_heap(sizeof(type),__alignof__(type),(n)))
The goal of this macro is to allocate memory on the early heap. This macro calls the __get_heap function from the same header file with the following parameters:
The implementation of __get_heap is:
static inline char *__get_heap(size_t s, size_t a, size_t n)
{
char *tmp;
HEAP = (char *)(((size_t)HEAP+(a-1)) & ~(a-1));
tmp = HEAP;
HEAP += s*n;
return tmp;
}
Let's try to understand how the __get_heap function works. First of all we can see here that HEAP pointer is assigned to the aligned address of the memory. The address is aligned based on the size of data type for which we want to allocate memory. After we have got the initial aligned address, we just move the HEAP pointer by the requested size.
The last but not least API of the early heap that we will see is the heap_free function which checks the availability of the given size of memory on the heap:
static inline bool heap_free(size_t n)
{
return (int)(heap_end-HEAP) >= (int)n;
}
As you may see, the implementation of this function is pretty trivial. It just subtracts the current value of the heap pointer from the address which represents the end of heap memory area. The function returns true if there is enough memory for n or false otherwise.
Since the kernel initialized the heap and the heap pointer is in the right place, we can move directly to video mode initialization.
The first step during the process of a video mode initialization is the store_mode_params function, which stores currently available video mode parameters in boot_params.screen_info. This structure is defined in include/uapi/linux/screen_info.h header file and provides basic information about the screen and video mode:
The store_mode_params function asks the BIOS services about this information and stores it in this structure for later usage.
The next step is saving the current contents of the screen to the heap by calling the save_screen function. This function collects all the data that we got in the previous functions (like rows and columns) and stores it in the saved_screen structure, which is defined as:
static struct saved_screen {
int x, y;
int curx, cury;
u16 *data;
} saved;
After the contents of the screen is saved, the next step is to collect currently available video modes in the system. This job is done by the probe_cards function defined in the arch/x86/boot/video-mode.c. It goes over all video_cards and collects the information about them:
for (card = video_cards; card < video_cards_end; card++) {
/* collecting the number of video modes */
}
The video_cards is an array defined as:
#define __videocard struct card_info __section(".videocards") __attribute__((used))
extern struct card_info video_cards[], video_cards_end[];
The __videocard macro allows to define structures which describe video cards and the linker will put them into the video_cards array. Example of such structure can be found in the arch/x86/boot/video-vga.c:
static __videocard video_vga = {
.card_name = "VGA",
.probe = vga_probe,
.set_mode = vga_set_mode,
};
After the probe_cards function is executed, we have a set of structures in our video_cards array, along with the known number of video modes they support. At the next step, the kernel setup code prints a menu with available video modes if the vid_mode=ask option was passed to the kernel command line, and sets up the video mode with all the parameters that we collected in the previous steps.
The video mode is set by the set_mode function which is defined in video-mode.c. This function expects one parameter - the video mode identifier. This identifier is set by the bootloader or based on the choice of the video modes menu. The set_mode function goes over all available video cards defined in the video_cards array, and if the given mode belongs to the given card, the card->set_mode() callback is called to set up the video mode.
Let's take a look at the example of setting up the VGA video mode:
<!-- https://raw.githubusercontent.com/torvalds/linux/refs/heads/master/arch/x86/boot/video-vga.c#L191-L224 -->static int vga_set_mode(struct mode_info *mode)
{
/* Set the basic mode */
vga_set_basic_mode();
/* Override a possibly broken BIOS */
force_x = mode->x;
force_y = mode->y;
switch (mode->mode) {
case VIDEO_80x25:
break;
case VIDEO_8POINT:
vga_set_8font();
break;
case VIDEO_80x43:
vga_set_80x43();
break;
case VIDEO_80x28:
vga_set_14font();
break;
case VIDEO_80x30:
vga_set_80x30();
break;
case VIDEO_80x34:
vga_set_80x34();
break;
case VIDEO_80x60:
vga_set_80x60();
break;
}
return 0;
}
The vga_set_mode function is responsible for configuring the VGA display to a specific text mode, based on the settings which we collected in the previous steps. The vga_set_basic_mode function resets the VGA hardware into a standard text mode. The next statement sets up the video mode based on the video mode that was selected. Most of these functions have very similar implementation based on the 0x10 BIOS interrupt.
After this step, the video mode is configured and we save all the information about it again for later use. Having done this, the video mode setup is complete and now we can take a look at the last preparation before we will see the switch into the protected mode.
Returning to the main function of the early kernel setup code, we finally can see:
/* Do the last things and invoke protected mode */
go_to_protected_mode();
As the comment says: Do the last things and invoke protected mode, so let's see what these last things are and switch into protected mode.
The go_to_protected_mode function is defined in arch/x86/boot/pm.c. It contains routines that make the final preparations before we jump into protected mode, so let's look at it and try to understand what it does and how it works.
The very first function that we can see in go_to_protected_mode is the realmode_switch_hook function. This function invokes the real mode switch hook if it is present, or disables NMI otherwise. The hooks are used if the bootloader runs in a hostile environment. You can read more about hooks in the boot protocol (see ADVANCED BOOT LOADER HOOKS). Interrupts must be disabled before switching to protected mode because otherwise the CPU could receive an interrupt when there is no valid interrupt table or handlers. Once the kernel sets up the protected-mode interrupt infrastructure, interrupts are enabled again.
We will consider only a standard use case, when the bootloader does not provide any hooks. In this case, we just disable non-maskable interrupts:
<!-- https://raw.githubusercontent.com/torvalds/linux/refs/heads/master/arch/x86/boot/pm.c#L28-L30 --> asm volatile("cli");
outb(0x80, 0x70); /* Disable NMI */
io_delay();
An interrupt is a signal to the CPU that is emitted by hardware or software. After getting such a signal, the CPU suspends the current instruction sequence, saves its state, and transfers control to the interrupt handler. After the interrupt handler has finished its work, it transfers control back to the interrupted instruction. Non-maskable interrupts (NMI) are interrupts that are always processed, independently of permission. They cannot be ignored and are typically used to signal non-recoverable hardware errors. We will not dive into the details of interrupts now, but we will discuss them in the next parts.
At the first line, there is an inline assembly statement with the cli instruction, which clears the interrupt flag. After this, external interrupts are disabled. The next line disables NMI (non-maskable interrupt).
Let's get back to the code. In the second line, we set the byte 0x0 to the port 0x80. After that, a call to the io_delay function occurs. io_delay causes a little delay and looks like this:
static inline void io_delay(void)
{
const u16 DELAY_PORT = 0x80;
outb(0, DELAY_PORT);
}
Writing any byte to port 0x80 introduces a delay of 1 microsecond. This delay ensures that the change to the NMI mask has fully taken effect. After this delay, all interrupts are disabled.
The next step is the enable_a20 function, which enables the A20 line. Enabling this line allows the kernel to have access to more than 1 megabyte of memory.
The enable_a20 function is defined in arch/x86/boot/a20.c. It enables the A20 gate using the different approaches. The first is the a20_test_short function, which checks if A20 is already enabled using the a20_test function:
static int a20_test(int loops)
{
int ok = 0;
int saved, ctr;
set_fs(0x0000);
set_gs(0xffff);
saved = ctr = rdfs32(A20_TEST_ADDR);
while (loops--) {
wrfs32(++ctr, A20_TEST_ADDR);
io_delay(); /* Serialize and make delay constant */
ok = rdgs32(A20_TEST_ADDR+0x10) ^ ctr;
if (ok)
break;
}
wrfs32(saved, A20_TEST_ADDR);
return ok;
}
To verify whether the A20 line is already enabled or not, the kernel performs a simple memory test. It begins by setting the FS register to 0x0000 and the GS register to 0xffff values. By doing this, an access to FS:0x200 (A20_TEST_ADDR) points into the very beginning of memory, while an access to GS:0x2010 refers to a location just past the one-megabyte boundary. If the A20 line is disabled, the latter will wrap around and point to the same physical address.
If the A20 gate is disabled, the kernel will try to enable it using different methods which you can find in enable_a20 function. For example, it can be done with a call to the 0x15 BIOS interrupt with AH register set to 0x2041. If this function finished with a failure, print an error message and call the function die which will stop the process of the kernel setup.
After the A20 gate is successfully enabled, the reset_coprocessor function is called:
static void reset_coprocessor(void)
{
outb(0, 0xf0);
io_delay();
outb(0, 0xf1);
io_delay();
}
This function resets the math coprocessor to ensure it is in a clean state before switching to protected mode. The reset is performed by writing 0 to port 0xF0, followed by writing 0 to port 0xF1.
The next step is the mask_all_interrupts function:
static void mask_all_interrupts(void)
{
outb(0xff, 0xa1); /* Mask all interrupts on the secondary PIC */
io_delay();
outb(0xfb, 0x21); /* Mask all but cascade on the primary PIC */
io_delay();
}
This function masks or in other words forbids all interrupts on the primary and secondary PICs. This is needed for safeness, we forbid all the interrupts from the PIC so nothing can interrupt the CPU while the kernel is doing transition into protected mode.
All the operations before this point, were executed for safe transition to the protected mode. The next operations will prepare the transition to the protected mode. Let's take a look at them.
At this point, we are very close to see the switching into protected mode of the Linux kernel.
Only two last steps remain:
And that’s all! Once these two structures will be configured, the Linux kernel can make the jump into protected mode.
Before the CPU can safely enter protected mode, it needs to know where to find the handlers that are triggered in the case of interrupts and exceptions. In real mode, the CPU relies on the Interrupt Vector Table. In protected mode, this mechanism changes to the Interrupt Descriptor Table.
The Interrupt Descriptor Table is a special structure located in memory that contains descriptors. This structure describes where the CPU can find handlers for interrupts and exceptions. We will see the full description of the Interrupt Description Table and its entries later, because for now, we have disabled all interrupts in the previous steps. Let's take a look at the function that sets up a zero-filled Interrupt Descriptor Table:
<!-- https://raw.githubusercontent.com/torvalds/linux/refs/heads/master/arch/x86/boot/pm.c#L94-L98 -->static void setup_idt(void)
{
static const struct gdt_ptr null_idt = {0, 0};
asm volatile("lidtl %0" : : "m" (null_idt));
}
As we can see, it just loads the IDT (which is filled with zeros) using the lidtl instruction. The null_idt has type gdt_ptr, which is a structure defined in the same arch/x86/boot/pm.c file:
struct gdt_ptr {
u16 len;
u32 ptr;
} __attribute__((packed));
This structure provides information about the pointer to the Interrupt Descriptor Table.
Next, we set up the Global Descriptor Table. As you may remember, the memory access is based on the segment:offset addressing in real mode. The protected mode introduces a different model based on the Global Descriptor Table. If you forgot the details about the Global Description Table structure, you can find more information in the previous chapter.
Instead of fixed segment bases and limits, the CPU now looks for memory regions defined by descriptors located in the Global Descriptor Table. The goal of the kernel is to set up these descriptors.
All the job will be done by the setup_gdt function, which is defined in the same source code file. Let's take a look at the definition of this function:
static void setup_gdt(void)
{
/* There are machines which are known to not boot with the GDT
being 8-byte unaligned. Intel recommends 16 byte alignment. */
static const u64 boot_gdt[] __attribute__((aligned(16))) = {
/* CS: code, read/execute, 4 GB, base 0 */
[GDT_ENTRY_BOOT_CS] = GDT_ENTRY(DESC_CODE32, 0, 0xfffff),
/* DS: data, read/write, 4 GB, base 0 */
[GDT_ENTRY_BOOT_DS] = GDT_ENTRY(DESC_DATA32, 0, 0xfffff),
/* TSS: 32-bit tss, 104 bytes, base 4096 */
/* We only have a TSS here to keep Intel VT happy;
we don't actually use it for anything. */
[GDT_ENTRY_BOOT_TSS] = GDT_ENTRY(DESC_TSS32, 4096, 103),
};
/* Xen HVM incorrectly stores a pointer to the gdt_ptr, instead
of the gdt_ptr contents. Thus, make it static so it will
stay in memory, at least long enough that we switch to the
proper kernel GDT. */
static struct gdt_ptr gdt;
gdt.len = sizeof(boot_gdt)-1;
gdt.ptr = (u32)&boot_gdt + (ds() << 4);
asm volatile("lgdtl %0" : : "m" (gdt));
}
The initial memory descriptors specified by the items of the boot_gdt array. The setup_gdt function just loads the pointer to the Global Descriptor Table filled with these items using the lgdtl instruction. Let's take a closer look at the memory descriptors definition.
Initially, the 3 memory descriptors specified:
We will skip the description of the task state segment for now, as it was added there (according to the comment) to make Intel VT happy.
The other two segments correspond to the memory regions used by the kernel code and data sections. Both memory descriptors are defined using the GDT_ENTRY macro. This macro itself is defined in arch/x86/include/asm/segment.h and expects three arguments:
flagsbaselimitLet's take a look at the definition of the code memory segment:
[GDT_ENTRY_BOOT_CS] = GDT_ENTRY(DESC_CODE32, 0, 0xfffff),
The base address of this memory segment is defined as 0 and the limit as 0xFFFFF. The DESC_CODE32 value describes the flags of this segment. If we take a look at the flags, we can see that the granularity (bit G) of this segment is set to 4 KB units. This means that the segment covers addresses 0x00000000–0xFFFFFFFF, which is the entire 4 GB linear address space. The same base address and limit are defined for the data segment. This is because the Linux kernel uses the so-called flat memory model.
Besides the granularity bit, the DESC_CODE32 specifies other flags. Among them, you can find a 32-bit segment that is readable, executable, and present in memory. The privilege level is set to the highest value as the kernel needs.
Looking at the documentation of the Global Descriptor Table and its entries, you can check all the initial segments by yourself. It is not so hard.
Finally, we are standing right before it – Interrupts are disabled, and the Interrupt Descriptor Table and Global Descriptor Table are initialized. Now the kernel can execute a jump into protected mode! But despite the good news, we need to return to the assembly again 😅
The transition to protected mode is defined in arch/x86/boot/pmjump.S. Let's take a look at it:
<!-- https://raw.githubusercontent.com/torvalds/linux/refs/heads/master/arch/x86/boot/pmjump.S#L24-L45 -->SYM_FUNC_START_NOALIGN(protected_mode_jump)
movl %edx, %esi # Pointer to boot_params table
xorl %ebx, %ebx
movw %cs, %bx
shll $4, %ebx
addl %ebx, 2f
jmp 1f # Short jump to serialize on 386/486
1:
movw $__BOOT_DS, %cx
movw $__BOOT_TSS, %di
movl %cr0, %edx
orb $X86_CR0_PE, %dl # Protected mode
movl %edx, %cr0
# Transition to 32-bit mode
.byte 0x66, 0xea # ljmpl opcode
2: .long .Lin_pm32 # offset
.word __BOOT_CS # segment
SYM_FUNC_END(protected_mode_jump)
First of all, we preserve the address of the boot_params structure in the esi register since we continue to use parameters that the kernel got during boot in later stages.
After this, we compute the physical base address of the current code segment and store it in the ebx register. Having it, we add it to the value stored at memory location 2f so that the jump instruction to the first protected mode code will contain the proper offset.
The next jump to the label 1 may look quite unexpected. Why does the kernel even need this jump? Right now, the CPU works in real mode. While it is executing the current instruction, it may have already fetched several subsequent instruction bytes into its internal prefetch queue. At this moment, all prefetched instructions were fetched under the assumption that the processor is still operating in real mode. If we were to continue executing instructions that were prefetched before the jump to the protected mode, the processor could continue decoding and executing them without fully synchronizing its internal state with the new mode. The jump instruction prevents this.
At the next steps, we save the segment addresses of the data and task state in general-purpose registers cx and di and set the PE bit in the control register cr0. From this point, the protected mode is turned on, and we just need to jump into it to set the proper value of the code segment:
# Transition to 32-bit mode
.byte 0x66, 0xea # ljmpl opcode
2: .long .Lin_pm32 # offset
.word __BOOT_CS # segment
The kernel is in protected mode now 🥳🥳🥳
<!-- https://raw.githubusercontent.com/torvalds/linux/refs/heads/master/arch/x86/boot/pmjump.S#L47-L49 --> .code32
.section ".text32","ax"
SYM_FUNC_START_LOCAL_NOALIGN(.Lin_pm32)
Let's look at the first steps taken in the protected mode. First of all we set up the data segment with the data segment address that we preserved in the cx register at the previous step:
# Set up data segments for flat 32-bit mode
movl %ecx, %ds
movl %ecx, %es
movl %ecx, %fs
movl %ecx, %gs
movl %ecx, %ss
Since we are in protected mode, our segment bases point to zero. Because of this, the stack pointer will point somewhere below the kernel code, so we need to adjust it to at least its previous state. Before the jump, we stored the base address of the code segment in the ebx register, so now we can use this value to adjust the stack pointer:
addl %ebx, %esp
The last step before the jump into actual 32-bit entry point is to clear the general purpose registers:
<!-- https://raw.githubusercontent.com/torvalds/linux/refs/heads/master/arch/x86/boot/pmjump.S#L65-L69 --> xorl %ecx, %ecx
xorl %edx, %edx
xorl %ebx, %ebx
xorl %ebp, %ebp
xorl %edi, %edi
Now everything is ready. The kernel is in the protected mode and we can jump to the next code, address of which was passed in the eax register:
jmpl *%eax # Jump to the 32-bit entrypoint
This is the end of the third part about Linux kernel insides. If you have questions or suggestions, feel free ping me on X - 0xAX, drop me an email, or just create an issue.
Here is the list of the links that you may find useful during reading of this chapter: