6. linux/arch/i386/kernel/head.S

Resident kernel image linux/vmlinux is in place finally! It requires two inputs:

ESI points to the parameter area from the 16-bit real mode code, which will be copied to empty_zero_page later. ESI is only valid for BSP.

BSP (BootStrap Processor) and APs (Application Processors) are Intel terminologies. Check IA-32 Manual (Vol.3. Ch.7.5. Multiple-Processor (MP) Initialization) and MultiProcessor Specification for MP intialization issue.

From a software point of view, in a multiprocessor system, BSP and APs share the physical memory but use their own register sets. BSP runs the kernel code first, setups OS execution enviornment and triggers APs to run over it too. AP will be sleeping until BSP kicks it.

6.1. Enable Paging

.text
///////////////////////////////////////////////////////////////////////////////
startup_32()
{
        /* set segments to known values */
        cld;
        DS = ES = FS = GS = __KERNEL_DS;

#ifdef CONFIG_SMP
#define cr4_bits mmu_cr4_features-__PAGE_OFFSET
        /* long mmu_cr4_features defined in linux/arch/i386/kernel/setup.c
         * __PAGE_OFFSET = 0xC0000000, i.e. 3G */

        // AP with CR4 support (> Intel 486) will copy CR4 from BSP
        if (BX && cr4_bits) {
                // turn on paging options (PSE, PAE, ...)
                CR4 |= cr4_bits;
        } else
#endif
        {
                /* only BSP initializes page tables (pg0..empty_zero_page-1)
                 *   pg0 at .org 0x2000
                 *   empty_zero_page at .org 0x4000
                 *   total (0x4000-0x2000)/4 = 0x0800 entries */
                pg0 = {
                        0x00000007,             // 7 = PRESENT + RW + USER
                        0x00001007,             // 0x1000 = 4096 = 4K
                        0x00002007,
                        ...
                pg1:    0x00400007,
                        ...
                        0x007FF007              // total 8M
                empty_zero_page:
                };
        }
Why do we have to add "-__PAGE_OFFSET" when referring a kernel symbol, for example, like pg0?

In linux/arch/i386/vmlinux.lds, we have:
  . = 0xC0000000 + 0x100000;
  _text = .;                    /* Text and read-only data */
  .text : {
        *(.text)
...
As pg0 is at offset 0x2000 of section .text in linux/arch/i386/kernel/head.o, which is the first file to be linked for linux/vmlinux, it will be at offset 0x2000 in output section .text. Thus it will be located at address 0xC0000000+0x100000+0x2000 after linking.
[root@localhost boot]# nm --defined /boot/vmlinux-2.4.20-28.9 | grep 'startup_32
\|mmu_cr4_features\|pg0\|\<empty_zero_page\>' | sort
c0100000 t startup_32
c0102000 T pg0
c0104000 T empty_zero_page
c0376404 B mmu_cr4_features
In protected mode without paging enabled, linear address will be mapped directly to physical address. "movl $pg0-__PAGE_OFFSET,%edi" will set EDI=0x102000, which is equal to the physical address of pg0 (as linux/vmlinux is relocated to 0x100000). Without this "-PAGE_OFFSET" scheme, it will access physical address 0xC0102000, which will be wrong and probably beyond RAM space.

mmu_cr4_features is in .bss section and is located at physical address 0x376404 in the above example.

After page tables are initialized, paging can be enabled.
        // set page directory base pointer, physical address
        CR3 = swapper_pg_dir - __PAGE_OFFSET;
        // paging enabled!
        CR0 |= 0x80000000;      // set PG bit
        goto 1f;                // flush prefetch-queue
1:
        EAX = &1f;              // address following the next instruction
        goto *(EAX);            // relocate EIP
1:
        SS:ESP = *stack_start;
Page directory swapper_pg_dir (see definition in Section 6.5), together with page tables pg0 and pg1, defines that both linear address 0..8M-1 and 3G..3G+8M-1 are mapped to physical address 0..8M-1. We can access kernel symbols without "-__PAGE_OFFSET" from now on, because kernel space (resides in linear address >=3G) will be correctly mapped to its physical addresss after paging is enabled.

"lss stack_start,%esp" (SS:ESP = *stack_start) is the first example to reference a symbol without "-PAGE_OFFSET", which sets up a new stack. For BSP, the stack is at the end of init_task_union. For AP, stack_start.esp has been redefined by linux/arch/i386/kernel/smpboot.c:do_boot_cpu() to be "(void *) (1024 + PAGE_SIZE + (char *)idle)" in Section 8.2.

For paging mechanism and data structures, refer to IA-32 Manual Vol.3. (Ch.3.7. Page Translation Using 32-Bit Physical Addressing, Ch.9.8.3. Initializing Paging, Ch.9.9.1. Switching to Protected Mode, and Ch.18.26.3. Enabling and Disabling Paging).

6.2. Get Kernel Parameters

#define OLD_CL_MAGIC_ADDR       0x90020
#define OLD_CL_MAGIC            0xA33F
#define OLD_CL_BASE_ADDR        0x90000
#define OLD_CL_OFFSET           0x90022
#define NEW_CL_POINTER          0x228   /* Relative to real mode data */

#ifdef CONFIG_SMP
        if (BX) {
                EFLAGS = 0;             // AP clears EFLAGS
        } else
#endif
        {
                // Initial CPU cleans BSS
                clear BSS;              // i.e. __bss_start .. _end
                setup_idt() {
                        /* idt_table[256] defined in arch/i386/kernel/traps.c
                         *   located in section .data.idt
                        EAX = __KERNEL_CS << 16 + ignore_int;
                        DX = 0x8E00;    // interrupt gate, dpl = 0, present
                        idt_table[0..255] = {EAX, EDX};
                }
                EFLAGS = 0;
                /*
                 * Copy bootup parameters out of the way. First 2kB of
                 * _empty_zero_page is for boot parameters, second 2kB
                 * is for the command line.
                 */
                move *ESI (real-mode header) to empty_zero_page, 2KB;
                clear empty_zero_page+2K, 2KB;
                ESI = empty_zero_page[NEW_CL_POINTER];
                if (!ESI) {             // 32-bit command line pointer
                        if (OLD_CL_MAGIC==(uint16)[OLD_CL_MAGIC_ADDR]) {
                                ESI = [OLD_CL_BASE_ADDR]
                                      + (uint16)[OLD_CL_OFFSET];
                                move *ESI to empty_zero_page+2K, 2KB;
                        }
                } else {                // valid in 2.02+
                        move *ESI to empty_zero_page+2K, 2KB;
                }
        }
}
For BSP, kernel parameters are copied from memory pointed by ESI to empty_zero_page. Kernel command line will be copied to empty_zero_page+2K if applicable.

6.3. Check CPU Type

Refer to IA-32 Manual Vol.1. (Ch.13. Processor Identification and Feature Determination) on how to identify processor type and processor features.

struct cpuinfo_x86;     // see include/asm-i386/processor.h
struct cpuinfo_x86 boot_cpu_data;       // see arch/i386/kernel/setup.c

#define CPU_PARAMS      SYMBOL_NAME(boot_cpu_data)
#define X86             CPU_PARAMS+0
#define X86_VENDOR      CPU_PARAMS+1
#define X86_MODEL       CPU_PARAMS+2
#define X86_MASK        CPU_PARAMS+3
#define X86_HARD_MATH   CPU_PARAMS+6
#define X86_CPUID       CPU_PARAMS+8
#define X86_CAPABILITY  CPU_PARAMS+12
#define X86_VENDOR_ID   CPU_PARAMS+28

checkCPUtype:
{
        X86_CPUID = -1;                 // no CPUID

        X86 = 3;                        // at least 386
        save original EFLAGS to ECX;
        flip AC bit (0x40000) in EFLAGS;
        if (AC bit not changed) goto is386;

        X86 = 4;                        // at least 486
        flip ID bit (0X200000) in EFLAGS;
        restore original EFLAGS;        // for AC & ID flags
        if (ID bit can not be changed) goto is486;

        // get CPU info
        CPUID(EAX=0);
        X86_CPUID = EAX;
        X86_VENDOR_ID = {EBX, EDX, ECX};
        if (!EAX) goto is486;

        CPUID(EAX=1);
        CL = AL;
        X86 = AH & 0x0f;                // family
        X86_MODEL = (AL & 0xf0) >> 4;   // model
        X86_MASK = CL & 0x0f;           // stepping id
        X86_CAPABILITY = EDX;           // feature

Refer to IA-32 Manual Vol.3. (Ch.9.2. x87 FPU Initialization, and Ch.18.14. x87 FPU) on how to setup x87 FPU.

is486:
        // save PG, PE, ET and set AM, WP, NE, MP
        EAX = (CR0 & 0x80000011) | 0x50022;
        goto 2f;                        // skip "is386:" processing
is386:
        restore original EFLAGS from ECX;
        // save PG, PE, ET and set MP
        EAX = (CR0 & 0x80000011) | 0x02;

        /* ET: Extension Type (bit 4 of CR0).
         * In the Intel 386 and Intel 486 processors, this flag indicates
         * support of Intel 387 DX math coprocessor instructions when set.
         * In the Pentium 4, Intel Xeon, and P6 family processors,
         * this flag is hardcoded to 1.
         *     -- IA-32 Manual Vol.3. Ch.2.5. Control Registers (p.2-14) */

2:      CR0 = EAX;
        check_x87() {
                /* We depend on ET to be correct.
                 * This checks for 287/387. */
                X86_HARD_MATH = 0;
                clts;                   // CR0.TS = 0;
                fninit;                 // Init FPU;
                fstsw AX;               // AX = ST(0);
                if (AL) {
                        CR0 ^= 0x04;    // no coprocessor, set EM
                } else {
                        ALIGN
1:                      X86_HARD_MATH = 1;
                        /* IA-32 Manual Vol.3. Ch.18.14.7.14. FSETPM Instruction
                         * inform 287 that processor is in protected mode
                         * 287 only, ignored by 387 */
                        fsetpm;
                }
        }
}
Macro ALIGN, defined in linux/include/linux/linkage.h, specifies 16-bytes alignment and fill value 0x90 (opcode for NOP). See also Using as: Assembler Directives for the meaning of directive .align.

6.4. Go Start Kernel

        ready:  .byte 0;        // global variable
{
        ready++;                // how many CPUs are ready
        lgdt gdt_descr;         // use new descriptor table in safe place
        lidt idt_descr;
        goto __KERNEL_CS:$1f;   // reload segment registers after "lgdt"
1:      DS = ES = FS = GS = __KERNEL_DS;
#ifdef CONFIG_SMP
        SS = __KERNEL_DS;       // reload segment only
#else
        SS:ESP = *stack_start;  /* end of init_task_union, defined
                                 *   in linux/arch/i386/kernel/init_task.c */
#endif
        EAX = 0;
        lldt AX;
        cld;

#ifdef CONFIG_SMP
        if (1!=ready) {         // not first CPU
                initialize_secondary();
                // see linux/arch/i386/kernel/smpboot.c
        } else
#endif
        {
                start_kernel(); // see linux/init/main.c
        }
L6:     goto L6;
}
The first CPU (BSP) will call linux/init/main.c:start_kernel() and the others (AP) will call linux/arch/i386/kernel/smpboot.c:initialize_secondary(). See start_kernel() in Section 7 and initialize_secondary() in Section 8.4.

init_task_union happens to be the task struct for the first process, "idle" process (pid=0), whose stack grows from the tail of init_task_union. The following is the code related to init_task_union:
ENTRY(stack_start)
        .long init_task_union+8192;
        .long __KERNEL_DS;

#ifndef INIT_TASK_SIZE
# define INIT_TASK_SIZE 2048*sizeof(long)
#endif

union task_union {
        struct task_struct task;
        unsigned long stack[INIT_TASK_SIZE/sizeof(long)];
};

/* INIT_TASK is used to set up the first task table, touch at
 * your own risk! Base=0, limit=0x1fffff (=2MB) */
union task_union init_task_union
        __attribute__((__section__(".data.init_task"))) =
                { INIT_TASK(init_task_union.task) };

init_task_union is for BSP "idle" process. Don't confuse it with "init" process, which will be mentioned in Section 7.2.

6.5. Miscellaneous

///////////////////////////////////////////////////////////////////////////////
// default interrupt "handler"
ignore_int() { printk("Unknown interrupt\n"); iret; }

/*
 * The interrupt descriptor table has room for 256 idt's,
 * the global descriptor table is dependent on the number
 * of tasks we can have..
 */
#define IDT_ENTRIES     256
#define GDT_ENTRIES     (__TSS(NR_CPUS))

.globl SYMBOL_NAME(idt)
.globl SYMBOL_NAME(gdt)

        ALIGN
        .word 0
idt_descr:
        .word IDT_ENTRIES*8-1           # idt contains 256 entries
SYMBOL_NAME(idt):
        .long SYMBOL_NAME(idt_table)

        .word 0
gdt_descr:
        .word GDT_ENTRIES*8-1
SYMBOL_NAME(gdt):
        .long SYMBOL_NAME(gdt_table)

/*
 * This is initialized to create an identity-mapping at 0-8M (for bootup
 * purposes) and another mapping of the 0-8M area at virtual address
 * PAGE_OFFSET.
 */
.org 0x1000
ENTRY(swapper_pg_dir)   // "ENTRY" defined in linux/include/linux/linkage.h
        .long 0x00102007
        .long 0x00103007
        .fill BOOT_USER_PGD_PTRS-2,4,0
        /* default: 766 entries */
        .long 0x00102007
        .long 0x00103007
        /* default: 254 entries */
        .fill BOOT_KERNEL_PGD_PTRS-2,4,0

/*
 * The page tables are initialized to only 8MB here - the final page
 * tables are set up later depending on memory size.
 */
.org 0x2000
ENTRY(pg0)

.org 0x3000
ENTRY(pg1)

/*
 * empty_zero_page must immediately follow the page tables ! (The
 * initialization loop counts until empty_zero_page)
 */
.org 0x4000
ENTRY(empty_zero_page)

/*
 * Real beginning of normal "text" segment
 */
.org 0x5000
ENTRY(stext)
ENTRY(_stext)

///////////////////////////////////////////////////////////////////////////////
/*
 * This starts the data section. Note that the above is all
 * in the text section because it has alignment requirements
 * that we cannot fulfill any other way.
 */
.data

ALIGN
/*
 * This contains typically 140 quadwords, depending on NR_CPUS.
 *
 * NOTE! Make sure the gdt descriptor in head.S matches this if you
 * change anything.
 */
ENTRY(gdt_table)
        .quad 0x0000000000000000        /* NULL descriptor */
        .quad 0x0000000000000000        /* not used */
        .quad 0x00cf9a000000ffff        /* 0x10 kernel 4GB code at 0x00000000 */
        .quad 0x00cf92000000ffff        /* 0x18 kernel 4GB data at 0x00000000 */
        .quad 0x00cffa000000ffff        /* 0x23 user   4GB code at 0x00000000 */
        .quad 0x00cff2000000ffff        /* 0x2b user   4GB data at 0x00000000 */
        .quad 0x0000000000000000        /* not used */
        .quad 0x0000000000000000        /* not used */
        /*
         * The APM segments have byte granularity and their bases
         * and limits are set at run time.
         */
        .quad 0x0040920000000000        /* 0x40 APM set up for bad BIOS's */
        .quad 0x00409a0000000000        /* 0x48 APM CS    code */
        .quad 0x00009a0000000000        /* 0x50 APM CS 16 code (16 bit) */
        .quad 0x0040920000000000        /* 0x58 APM DS    data */
        .fill NR_CPUS*4,8,0             /* space for TSS's and LDT's */
Macro ALIGN, before idt_descr and gdt_table, is for performance consideration.

6.6. Reference