matos/arch/x86/cpu_context.c

/* Copyright (C) 2005  David Decotigny
   Copyright (C) 2000-2004, The KOS team

   Initially taken from SOS
*/

#include "assert.h"
#include "klibc.h"
#include "segment.h"

#include "cpu_context.h"

/**
 * Here is the definition of a CPU context for IA32 processors. This
 * is a Matos/SOS convention, not a specification given by the IA32
 * spec. However there is a strong constraint related to the x86
 * interrupt handling specification: the top of the stack MUST be
 * compatible with the 'iret' instruction, ie there must be the
 * err_code (might be 0), eip, cs and eflags of the destination
 * context in that order (see Intel x86 specs vol 3, figure 5-4).
 *
 * @note IMPORTANT: This definition MUST be consistent with the way
 * the registers are stored on the stack in
 * irq_wrappers.S/exception_wrappers.S !!! Hence the constraint above.
 */
struct cpu_state {
    /* (Lower addresses) */

    /* These are Matos/SOS convention */
    uint16_t gs;
    uint16_t fs;
    uint16_t es;
    uint16_t ds;
    uint16_t cpl0_ss;           /* This is ALWAYS the Stack Segment of the
                   Kernel context (CPL0) of the interrupted
                   thread, even for a user thread */
    uint16_t alignment_padding; /* unused */
    uint32_t edi;
    uint32_t esi;
    uint32_t esp;
    uint32_t ebp;
    uint32_t ebx;
    uint32_t edx;
    uint32_t ecx;
    uint32_t eax;

    /* MUST NEVER CHANGE (dependent on the IA32 iret instruction) */
    uint32_t error_code;
    vaddr_t eip;
    uint32_t cs; /* 32bits according to the specs ! However, the CS
          register is really 16bits long */
    uint32_t eflags;

    /* (Higher addresses) */
} __attribute__((packed));

/**
 * The CS value pushed on the stack by the CPU upon interrupt, and
 * needed by the iret instruction, is 32bits long while the real CPU
 * CS register is 16bits only: this macro simply retrieves the CPU
 * "CS" register value from the CS value pushed on the stack by the
 * CPU upon interrupt.
 *
 * The remaining 16bits pushed by the CPU should be considered
 * "reserved" and architecture dependent. IMHO, the specs don't say
 * anything about them. Considering that some architectures generate
 * non-zero values for these 16bits (at least Cyrix), we'd better
 * ignore them.
 */
#define GET_CPU_CS_REGISTER_VALUE(pushed_ui32_cs_value) ((pushed_ui32_cs_value)&0xffff)

/**
 * Structure of an interrupted Kernel thread's context
 */
struct cpu_kstate {
    struct cpu_state regs;
} __attribute__((packed));

/**
 * THE main operation of a kernel thread. This routine calls the
 * kernel thread function start_func and calls exit_func when
 * start_func returns.
 */
static void core_routine(cpu_kstate_function_arg1_t *start_func, void *start_arg,
                         cpu_kstate_function_arg1_t *exit_func, void *exit_arg)
    __attribute__((noreturn));

static void core_routine(cpu_kstate_function_arg1_t *start_func, void *start_arg,
                         cpu_kstate_function_arg1_t *exit_func, void *exit_arg)
{
    start_func(start_arg);
    exit_func(exit_arg);

    assert(!"The exit function of the thread should NOT return !");
    for (;;)
        ;
}

int cpu_kstate_init(struct cpu_state **ctxt, cpu_kstate_function_arg1_t *start_func,
                    vaddr_t start_arg, vaddr_t stack_bottom, size_t stack_size,
                    cpu_kstate_function_arg1_t *exit_func, vaddr_t exit_arg)
{
    /* We are initializing a Kernel thread's context */
    struct cpu_kstate *kctxt;

    /* This is a critical internal function, so that it is assumed that
       the caller knows what he does: we legitimally assume that values
       for ctxt, start_func, stack_* and exit_func are allways VALID ! */

    /* Setup the stack.
     *
     * On x86, the stack goes downward. Each frame is configured this
     * way (higher addresses first):
     *
     *  - (optional unused space. As of gcc 3.3, this space is 24 bytes)
     *  - arg n
     *  - arg n-1
     *  - ...
     *  - arg 1
     *  - return instruction address: The address the function returns to
     *    once finished
     *  - local variables
     *
     * The remaining of the code should be read from the end upward to
     * understand how the processor will handle it.
     */

    vaddr_t tmp_vaddr = stack_bottom + stack_size;
    uint32_t *stack   = (uint32_t *)tmp_vaddr;

    /* If needed, poison the stack */
#ifdef CPU_STATE_DETECT_UNINIT_KERNEL_VARS
    memset((void *)stack_bottom, CPU_STATE_STACK_POISON, stack_size);
#elif defined(CPU_STATE_DETECT_KERNEL_STACK_OVERFLOW)
    cpu_state_prepare_detect_kernel_stack_overflow(stack_bottom, stack_size);
#endif

    /* Simulate a call to the core_routine() function: prepare its
       arguments */
    *(--stack) = exit_arg;
    *(--stack) = (uint32_t)exit_func;
    *(--stack) = start_arg;
    *(--stack) = (uint32_t)start_func;
    *(--stack) = 0; /* Return address of core_routine => force page fault */

    /*
     * Setup the initial context structure, so that the CPU will execute
     * the function core_routine() once this new context has been
     * restored on CPU
     */

    /* Compute the base address of the structure, which must be located
       below the previous elements */
    tmp_vaddr = ((vaddr_t)stack) - sizeof(struct cpu_kstate);
    kctxt     = (struct cpu_kstate *)tmp_vaddr;

    /* Initialize the CPU context structure */
    memset(kctxt, 0x0, sizeof(struct cpu_kstate));

    /* Tell the CPU context structure that the first instruction to
       execute will be that of the core_routine() function */
    kctxt->regs.eip = (uint32_t)core_routine;

    /* Setup the segment registers */
    kctxt->regs.cs      = BUILD_SEGMENT_REG_VALUE(0, FALSE, SEG_KCODE); /* Code */
    kctxt->regs.ds      = BUILD_SEGMENT_REG_VALUE(0, FALSE, SEG_KDATA); /* Data */
    kctxt->regs.es      = BUILD_SEGMENT_REG_VALUE(0, FALSE, SEG_KDATA); /* Data */
    kctxt->regs.cpl0_ss = BUILD_SEGMENT_REG_VALUE(0, FALSE, SEG_KDATA); /* Stack */
    /* fs and gs unused for the moment. */

    /* The newly created context is initially interruptible */
    kctxt->regs.eflags = (1 << 9); /* set IF bit */

    /* Finally, update the generic kernel/user thread context */
    *ctxt = (struct cpu_state *)kctxt;

    return 0;
}

#if defined(CPU_STATE_DETECT_KERNEL_STACK_OVERFLOW)
void cpu_state_prepare_detect_kernel_stack_overflow(const struct cpu_state *ctxt,
                                                    vaddr_t stack_bottom, size_t stack_size)
{
    (void)ctxt;
    size_t poison_size = CPU_STATE_DETECT_KERNEL_STACK_OVERFLOW;
    if (poison_size > stack_size)
        poison_size = stack_size;

    memset((void *)stack_bottom, CPU_STATE_STACK_POISON, poison_size);
}

void cpu_state_detect_kernel_stack_overflow(const struct cpu_state *ctxt, vaddr_t stack_bottom,
                                            size_t stack_size)
{
    unsigned char *c;
    size_t i;

    /* On Matos/SOS, "ctxt" corresponds to the address of the esp register of
       the saved context in Kernel mode (always, even for the interrupted
       context of a user thread). Here we make sure that this stack
       pointer is within the allowed stack area */
    assert(((vaddr_t)ctxt) >= stack_bottom);
    assert(((vaddr_t)ctxt) + sizeof(struct cpu_kstate) <= stack_bottom + stack_size);

    /* Check that the bottom of the stack has not been altered */
    for (c = (unsigned char *)stack_bottom, i = 0;
         (i < CPU_STATE_DETECT_KERNEL_STACK_OVERFLOW) && (i < stack_size); c++, i++) {
        assert(CPU_STATE_STACK_POISON == *c);
    }
}
#endif

/* =======================================================================
 * Public Accessor functions
 */

vaddr_t cpu_context_get_PC(const struct cpu_state *ctxt)
{
    assert(NULL != ctxt);

    /* This is the PC of the interrupted context (ie kernel or user
       context). */
    return ctxt->eip;
}

vaddr_t cpu_context_get_SP(const struct cpu_state *ctxt)
{
    assert(NULL != ctxt);

    /* On Matos/SOS, "ctxt" corresponds to the address of the esp register of
       the saved context in Kernel mode (always, even for the interrupted
       context of a user thread). */
    return (vaddr_t)ctxt;
}

void cpu_context_dump(const struct cpu_state *ctxt)
{
    printf("CPU: eip=%x esp=%x eflags=%x cs=%x ds=%x ss=%x err=%x", (unsigned)ctxt->eip,
           (unsigned)ctxt, (unsigned)ctxt->eflags,
           (unsigned)GET_CPU_CS_REGISTER_VALUE(ctxt->cs), (unsigned)ctxt->ds,
           (unsigned)ctxt->cpl0_ss, (unsigned)ctxt->error_code);
}