The Exec Microkernel

This document describes the inner workings of exec.library. It's not intended for application programmers but for those who are interested the systems inner workings. Most of the things described herein are system private - they may or may not be true for future system versions.

Exec is a shared library giving the ability for multitasking, including communication between tasks, memory management, shared libraries, and device I/O. Large parts of this library are formed by a set of functions that are called on the user's schedule just like any normal C function would be (though they do things you'd usually expect in a kernel). These functions make up the library part of exec, with this part being built on a smaller set of functions which form the core of exec (you may call them the real kernel if you like). This machine-dependent core serves things like supervisor mode, traps, interrupts, task exceptions, and context-switching routines.

The core of exec in short:

ULONG Supervisor(ULONG (*)());

Calls a user-function in supervisor mode.

APTR tc_TrapData; and APTR tc_TrapCode;

User settable trap vector.

IntVects[i]

The different interrupt vectors.

void Disable(void);

Disable interrupts.

void Enable(void);

Enable interrupts.

BYTE IDNestCnt;

-1 if interrupts are allowed.

APTR PrepareContext(APTR);

Prepare a new stack to hold a context.

void Switch(void);

Do a task switch.

AttnFlags & 0x8000

Delayed dispatch flag.

AttnFlags & 0x80

Delayed switch flag.

tc_Flags & TF_SWITCH

tc_Switch is active.

void (*tc_Switch)();

Vector to be called before task loses CPU.

tc_Flags & TF_LAUNCH

tc_Launch is active.

void (*tc_Launch)();

Vector to be called before task gets CPU.

UBYTE tc_State;

Where the actual task structure is sorted in.

struct Task *ThisTask;

Pointer to actual task structure.

struct List TaskReady;

Sorted list of tasks waiting for CPU.

tc_Flags & TF_EXCEPT

Raise a task exception for this task at the next context switch.

void Exception(void);

Is called to handle the task exception.

The most important part are the context switching routines. The context of every task is stored on the task's user stack. The stack pointer is stored in the task structure. Context switches are raised by moving the current task to one of the two task lists (must be done with interrupts Disable()d) and then calling Switch(). However, if this happens inside an interrupt (which is allowed and quite useful) the user context is unavailable. Therefore Switch() checks if it was called from user mode and sets the delayed dispatch flag if not. This flag tells the interrupt primitives to do a context switch (by calling Dispatch()) before falling down to user mode. When called from user mode Switch() calls Dispatch() directly. Dispatch() is the function which does the real work, i.e. it replaces the current task with the first task in the ready list. If the TF_EXCEPT flag for the new task is set, Dispatch() raises a task exception by calling the Exception() vector of exec with interrupts Disable()d. Also, tc_Switch and tc_Launch are both handled by Dispatch().

The usual way to do a context switch is as follows:

if(I_need_a_task_switch)
{
  /* The task lists must be protected with a Disable() */
  Disable();
  /* check if task switches are allowed */
  if(SysBase->TDNestCnt<0&&SysBase->IDNestCnt<=0)
  {
    /* Check if there is already a pending task switch (this check can
       be skipped if I can guarantee to be in user mode). */
    if(SysBase->ThisTask->tc_State==TS_RUN)
    {
      /* No. Move the current task away */
      Enqueue(new_list,SysBase->ThisTask);
      SysBase->ThisTask->tc_State=new_state;
      /* And do the switch */
      Switch();
    }
  }else
    /* Memorize task switch for a later Enable() or Permit(). */
    SysBase->AttnFlags|=0x80;
  Enable();
}

Interrupts and traps are delivered with direct function calls. The different interrupt levels call different interrupt vectors (stored in IntVects[i]). Traps are raised by calling the tc_TrapCode vector of the current task (usually this ends up in an Alert() - the exec equivalent of a kernel panic). Interrupts may be disabled completely (by calling Disable(), Enable() enables them again, Disable() calls may be nested) but only for a short time.