By now, the PRi OS is already a fairly complicated bare metal program, but to be honest, we still can't call it an operating system. The reason is that it can't do any of the core tasks that any OS should do. One of such core tasks is called process scheduling. By scheduling I mean that an operating system should be able to share CPU time between different processes. The hard part of it is that a process should be unaware of the scheduling happening: it should view itself as the only one occupying the CPU. In this lesson, we are going to add this functionality to the RPi OS.
If we want to manage processes, the first thing we should do is to create a struct that describes a process. Linux has such a struct and it is called task_struct (in Linux both thread and processes are just different types of tasks). As we are mostly mimicking Linux implementation, we are going to do the same. RPi OS task_struct looks like the following.
struct cpu_context {
unsigned long x19;
unsigned long x20;
unsigned long x21;
unsigned long x22;
unsigned long x23;
unsigned long x24;
unsigned long x25;
unsigned long x26;
unsigned long x27;
unsigned long x28;
unsigned long fp;
unsigned long sp;
unsigned long pc;
};
struct task_struct {
struct cpu_context cpu_context;
long state;
long counter;
long priority;
long preempt_count;
};
This struct has the following members:
cpu_contextThis is an embedded structure that contains values of all registers that might be different between the tasks, that are being switched. A reasonable question to ask is why do we save not all registers, but only registersx19 - x30andsp? (fpisx29andpcisx30) The answer is that actual context switch happens only when a task calls cpu_switch_to function. So, from the point of view of the task that is being switched, it just callscpu_switch_tofunction and it returns after some (potentially long) time. The task doesn't notice that another task happens to runs during this period. Accordingly to ARM calling conventions registersx0 - x18can be overwritten by the called function, so the caller must not assume that the values of those registers will survive after a function call. That's why it doesn't make sense to savex0 - x18registers.stateThis is the state of the currently running task. For tasks that are just doing some work on the CPU the state will always be TASK_RUNNING. Actually, this is the only state that the RPi OS is going to support for now. However, later we will have to add a few additional states. For example, a task that is waiting for an interrupt should be moved to a different state, because it doesn't make sense to awake the task while the required interrupt hasn't yet happened.counterThis field is used to determine how long the current task has been running.counterdecreases by 1 each timer tick and when it reaches 0 another task is scheduled.priorityWhen a new task is scheduled itspriorityis copied tocounter. By setting tasks priority, we can regulate the amount of processor time that the task gets relative to other tasks.preempt_countIf this field has a non-zero value it is an indicator that right now the current task is executing some critical function that must not be interrupted (for example, it runs the scheduling function) If timer tick occurs at such time it is ignored and rescheduling is not triggered.
After the kernel startup, there is only one task running: the one that runs kernel_main function. It is called "init task". Before the scheduler functionality is enabled, we must fill task_struct corresponding to the init task. This is done here.
All tasks are stored in task array. This array has only 64 slots - that is the maximum number of simultaneous tasks that we can have in the RPi OS. It is definitely not the best solution for the production-ready OS, but it is ok for our goals.
There is also a very important variable called current that always points to the currently executing task. Both current and task array are initially set to hold a pointer to the init task. There is also a global variable called nr_tasks - it contains the number of currently running tasks in the system.
Those are all structures and global variables that we are going to use to implement the scheduler functionality. In the description of the task_struct I already briefly mentioned some aspects of how scheduling works, because it is impossible to understand the meaning of a particular task_struct field without understanding how this field is used. Now we are going to examine the scheduling algorithm in much more details and we will start with the kernel_main function.
Before we dig into the scheduler implementation, I want to quickly show you how we are going to prove that the scheduler actually works. To understand it, you need to take a look at the kernel.c file. Let me copy the relevant content here.
void kernel_main(void)
{
uart_init();
init_printf(0, putc);
irq_vector_init();
timer_init();
enable_interrupt_controller();
enable_irq();
int res = copy_process((unsigned long)&process, (unsigned long)"12345");
if (res != 0) {
printf("error while starting process 1");
return;
}
res = copy_process((unsigned long)&process, (unsigned long)"abcde");
if (res != 0) {
printf("error while starting process 2");
return;
}
while (1){
schedule();
}
}
There are a few important things about this code.
- New function copy_process is introduced.
copy_processtakes 2 arguments: a function to execute in a new thread and an argument that need to be passed to this function.copy_processallocates a newtask_structand makes it available for the scheduler. - Another new function for us is called schedule. This is the core scheduler function: it checks whether there is a new task that needs to preemt the current one. A task can voluntarily call
scheduleif it doesn't have any work to do at the moment.scheduleis also called from the timer interrupt handler.
We are calling copy_process 2 times, each time passing a pointer to the process function as the first argument. process function is very simple.
void process(char *array)
{
while (1){
for (int i = 0; i < 5; i++){
uart_send(array[i]);
delay(100000);
}
}
}
It just keeps printing on the screen characters from the array, that is passed as an argument The first time it is called with the argument "12345" and second time the argument is "abcde". If our scheduler implementation is correct, we should see on the screen mixed output from both threads.
Each task in the system should have its dedicated stack. That's why when creating a new task we must have a way to allocate memory. For now, our memory allocator is extremely primitive. (The implementation can be found in mm.c file)
static unsigned short mem_map [ PAGING_PAGES ] = {0,};
unsigned long get_free_page()
{
for (int i = 0; i < PAGING_PAGES; i++){
if (mem_map[i] == 0){
mem_map[i] = 1;
return LOW_MEMORY + i*PAGE_SIZE;
}
}
return 0;
}
void free_page(unsigned long p){
mem_map[p / PAGE_SIZE] = 0;
}
The allocator can work only with memory pages (each page is 4 KB in size). There is an array called mem_map that for each page in the system holds its status: whether it is allocated or free. Whenever we need to allocate a new page, we just loop through this array and return the first free page. This implementation is based on 2 assumptions:
- We know the total amount of memory in the system. It is
1 GB - 1 MB(the last megabyte of memory is reserved for device registers) This value is stored in the HIGH_MEMORY constant. - First 2 MB of memory are reserved for the kernel image and init task stack. This value is stored in LOW_MEMORY constant. All memory allocations start right after this point.
New task allocation is implemented in copy_process function.
int copy_process(unsigned long fn, unsigned long arg)
{
preempt_disable();
struct task_struct *p;
p = (struct task_struct *) get_free_page();
if (!p)
return 1;
p->priority = current->priority;
p->state = TASK_RUNNING;
p->counter = p->priority;
p->preempt_count = 1; //disable preemtion until schedule_tail
p->cpu_context.x19 = fn;
p->cpu_context.x20 = arg;
p->cpu_context.pc = (unsigned long)ret_from_fork;
p->cpu_context.sp = (unsigned long)p + THREAD_SIZE;
int pid = nr_tasks++;
task[pid] = p;
preempt_enable();
return 0;
}
Now, we are going to examine it in details.
preempt_disable();
struct task_struct *p;
The function starts with disabling preemption and allocating a pointer for the new task. Preemption is disabled because we don't want to be rescheduled to a different task in the middle of the copy_process function.
p = (struct task_struct *) get_free_page();
if (!p)
return 1;
Next, a new page is allocated. At the bottom of this page, we are putting the task_struct for the newly created task. The rest of this page will be used as the task stack.
p->priority = current->priority;
p->state = TASK_RUNNING;
p->counter = p->priority;
p->preempt_count = 1; //disable preemtion until schedule_tail
After the task_struct is allocated, we can initialize its properties. Priority and initial counter are set based on the current task priority. The state is set to TASK_RUNNING, indicating that the new task is ready to be started. preempt_count is set to 1, meaning that after the task is executed it should not be rescheduled until it completes some initialization work.
p->cpu_context.x19 = fn;
p->cpu_context.x20 = arg;
p->cpu_context.pc = (unsigned long)ret_from_fork;
p->cpu_context.sp = (unsigned long)p + THREAD_SIZE;
This is the most important part of the function. Here cpu_context is initialized. The stack pointer is set to the top of the newly allocated memory page. pc is set to the ret_from_fork function, and we need to look at this function now in order to understand why the rest of the cpu_context registers are initialized in the way they are.
.globl ret_from_fork
ret_from_fork:
bl schedule_tail
mov x0, x20
blr x19 //should never return
As you can see ret_from_fork first call schedule_tail, which just enabled preemption, and then it calls the function stored in x19 register with the argument stored in x20. x19 and x20 are restored from the cpu_context just before ret_from_fork function is called.
Now, let's go back to copy_process.
int pid = nr_tasks++;
task[pid] = p;
preempt_enable();
return 0;
Finally, copy_process adds the newly created task to the task array and enables preemption for the current task.
An important thing to understand about the copy_process function is that after it finishes execution, no context switch happens. The function only prepares new task_struct and adds it to the task array — this task will be executed only after schedule function is called.
Before we get into the details of the schedule function, lets first figure out how schedule is called. There are 2 scenarios.
- When some task doesn't have anything to do right now, but it nevertheless can't be terminated, it can call
schedulevoluntarily. That is somethingkernel_mainfunction does. scheduleis also called on a regular basis from the timer interrupt handler.
Now let's take a look at timer_tick function, which is called from the timer interrupt.
void timer_tick()
{
--current->counter;
if (current->counter>0 || current->preempt_count >0) {
return;
}
current->counter=0;
enable_irq();
_schedule();
disable_irq();
First of all, it decreases current task's counter. If the counter is greater then 0 or preemption is currently disabled the function returns, but otherwiseschedule is called with interrupts enabled. (We are inside an interrupt handler, and interrupts are disabled by default) We will see why interrupts must be enabled during scheduler execution in the next section.
Finally, we are ready to look at the scheduler algorithm. I almost precisely copied this algorithm from the first release of the Linux kernel. You can find the original version here
void _schedule(void)
{
preempt_disable();
int next,c;
struct task_struct * p;
while (1) {
c = -1;
next = 0;
for (int i = 0; i < NR_TASKS; i++){
p = task[i];
if (p && p->state == TASK_RUNNING && p->counter > c) {
c = p->counter;
next = i;
}
}
if (c) {
break;
}
for (int i = 0; i < NR_TASKS; i++) {
p = task[i];
if (p) {
p->counter = (p->counter >> 1) + p->priority;
}
}
}
switch_to(task[next]);
preempt_enable();
}
The algorithm works like the following:
-
The first inner
forloop iterates over all tasks and tries to find a task inTASK_RUNNINGstate with the maximum counter. If such task is found and its counter is greater then 0, we immediately break from the outerwhileloop and switch to this task. If no such task is found this means that no tasks inTASK_RUNNINGstate currently exist or all such tasks have 0 counters. In a real OS, the first case might happen, for example, when all tasks are waiting for an interrupt. In this case, the second nestedforloop is executed. For each task (no matter what state it is in) this loop increases its counter. The counter increase is done in a very smart way:- The more iterations of the second
forloop a task passes, the more its counter will be increased. - A task counter can never get larger than
2 * priority
- The more iterations of the second
-
Then the process is repeated. If there are at least one task in
TASK_RUNNIGstate, the second iteration of the outerwhileloop will be the last one because after the first iteration all counters are already non-zero. However, if noTASK_RUNNINGtasks are there, the process is repeated over and over again until some of the tasks will move toTASK_RUNNINGstate. But if we are running on a single CPU, how then a task state can change while this loop is running? The answer is that if some task is waiting for an interrupt, this interrupt can happen whileschedulefunction is executed and interrupt handler can change the state of the task. This actually explains why interrupts must be enabled duringscheduleexecution. This also demonstrates an important distinction between disabling interrupts and disabling preemption.scheduledisables preemption for the duration of the whole function. This ensures that nesterschedulewill not be called while we are in the middle of the original function execution. However, interrupts can legally happen duringschedulefunction execution.
I paid a lot of attention to the situation were some task is waiting for an interrupt, though this functionality isn't implemented in the RPi OS yet. But I still consider it necessary to understand this case because it is a part of the core scheduler algorithm and similar functionality will be added later.
After the task in TASK_RUNNING state with non-zero counter is found, switch_to function is called. It looks like this.
void switch_to(struct task_struct * next)
{
if (current == next)
return;
struct task_struct * prev = current;
current = next;
cpu_switch_to(prev, next);
}
Here we check that next process is not the same as the current, and if not, current variable is updated. The actual work is redirected to cpu_switch_to function.
.globl cpu_switch_to
cpu_switch_to:
mov x10, #THREAD_CPU_CONTEXT
add x8, x0, x10
mov x9, sp
stp x19, x20, [x8], #16 // store callee-saved registers
stp x21, x22, [x8], #16
stp x23, x24, [x8], #16
stp x25, x26, [x8], #16
stp x27, x28, [x8], #16
stp x29, x9, [x8], #16
str x30, [x8]
add x8, x1, x10
ldp x19, x20, [x8], #16 // restore callee-saved registers
ldp x21, x22, [x8], #16
ldp x23, x24, [x8], #16
ldp x25, x26, [x8], #16
ldp x27, x28, [x8], #16
ldp x29, x9, [x8], #16
ldr x30, [x8]
mov sp, x9
ret
This is the place where the real context switch happens. Let's examine it line by line.
mov x10, #THREAD_CPU_CONTEXT
add x8, x0, x10
THREAD_CPU_CONTEXT constant contains offset of the cpu_context structure in the task_struct. x0 contains a pointer to the first argument, which is the current task_struct (by current here I mean the one we are switching from). After the copied 2 lines are executed, x8 will contain a pointer to the current cpu_context
mov x9, sp
stp x19, x20, [x8], #16 // store callee-saved registers
stp x21, x22, [x8], #16
stp x23, x24, [x8], #16
stp x25, x26, [x8], #16
stp x27, x28, [x8], #16
stp x29, x9, [x8], #16
str x30, [x8]
Next all calle-saved registers are stored in the order, in wich they are defined in cpu_context struture. x30, which is the link register and contains function return address, is stored as pc, current stack pointer is saved as sp and x29 is saved as fp (frame pointer).
add x8, x1, x10
Now x10 contains an offset of the cpu_context structure inside task_struct, x1 is a pointer to the next task_struct, so x8 will contain a pointer to the next cpu_context.
ldp x19, x20, [x8], #16 // restore callee-saved registers
ldp x21, x22, [x8], #16
ldp x23, x24, [x8], #16
ldp x25, x26, [x8], #16
ldp x27, x28, [x8], #16
ldp x29, x9, [x8], #16
ldr x30, [x8]
mov sp, x9
Callee saved registers are restored from the next cpu_context.
ret
Function returns to the location pointed to by the link register (x30) If we are switching to some task for the first time, this will be ret_from_fork function. In all other cases this will be the location, previously saved in the cpu_context by the cpu_switch_to function.
In the previous lesson, we have seen how kernel_entry and kernel_exit macros are used to save and restore the processor state. After the scheduler has been introduced, a new problem arrives: now it becomes fully legal to enter an interrupt as one task and leave it as a different one. This is a problem, because eret instruction, which we are using to return from an interrupts, relies on the fact that return address should be stored in elr_el1 and processor state in spsr_el1 registers. So, if we want to switch tasks while processing an interrupt, we must save and restore those 2 registers alongside with all other general purpose registers. The code that does this is very straightforward, you can find the save part here and restore here
We have already examined all source code related to the context switch. However, that code contains a lot of asynchronous interactions that make it difficult to fully understand how the state of the whole system changes over time. In this section I want to make this task easier for you: I want to describe the sequence of events that happen from system startup to the time of the second context switch. For each such event, I will also include a diagram representing the state of the memory at the time of the event. I hope that such representation will help you to get a deep understanding of how the scheduler works. So, let's start!
- The kernel is initialized and kernel_main function is executed. The initial stack is configured to start at LOW_MEMORY, which is 4 MB.
0 +------------------+ | kernel image | |------------------| | | |------------------| | init task stack | 0x00400000 +------------------+ | | | | 0x3F000000 +------------------+ | device registers | 0x40000000 +------------------+ kernel_maincalls copy_process for the first time. New 4 KB memory page is allocated, andtask_structis placed at the bottom of this page. (Later I will refer to the task created at this point as "task 1")0 +------------------+ | kernel image | |------------------| | | |------------------| | init task stack | 0x00400000 +------------------+ | task_struct 1 | |------------------| | | 0x00401000 +------------------+ | | | | 0x3F000000 +------------------+ | device registers | 0x40000000 +------------------+kernel_maincallscopy_processfor the second time and the same process repeats. Task 2 is created and added to the task list.0 +------------------+ | kernel image | |------------------| | | |------------------| | init task stack | 0x00400000 +------------------+ | task_struct 1 | |------------------| | | 0x00401000 +------------------+ | task_struct 2 | |------------------| | | 0x00402000 +------------------+ | | | | 0x3F000000 +------------------+ | device registers | 0x40000000 +------------------+kernel_mainvolantirely calls schedule function and it desides to run task 1.- cpu_switch_to saves calee-saved registers in the init task
cpu_context, wich is located inside the kernel image. cpu_switch_torestores calee-saved registers from task 1cpu_context.spnow points to0x00401000, link register to ret_from_fork function,x19contains a pointer to process function andx20a pointer to string "12345", wich is located somewhere in the kernel image.cpu_switch_tocallsretinstruction, which jums to theret_from_forkfunction.ret_from_forkreadsx19andx20registers and callsprocessfunction with the argument "12345". Afterprocessfunction starts to execute its stack begins to grow.0 +------------------+ | kernel image | |------------------| | | |------------------| | init task stack | 0x00400000 +------------------+ | task_struct 1 | |------------------| | | |------------------| | task 1 stack | 0x00401000 +------------------+ | task_struct 2 | |------------------| | | 0x00402000 +------------------+ | | | | 0x3F000000 +------------------+ | device registers | 0x40000000 +------------------+- A timer interrupt occured. kernel_entry macro saves all general purpose registers +
elr_el1andspsr_el1to the bottom of task 1 stack.0 +------------------------+ | kernel image | |------------------------| | | |------------------------| | init task stack | 0x00400000 +------------------------+ | task_struct 1 | |------------------------| | | |------------------------| | task 1 saved registers | |------------------------| | task 1 stack | 0x00401000 +------------------------+ | task_struct 2 | |------------------------| | | 0x00402000 +------------------------+ | | | | 0x3F000000 +------------------------+ | device registers | 0x40000000 +------------------------+ scheduleis called and it decides to run task 2. But we still run task 1 and its stack continues to grow below task 1 saved registers region. On the diagram, this part of the stack is marked as (int), which means "interrupt stack"0 +------------------------+ | kernel image | |------------------------| | | |------------------------| | init task stack | 0x00400000 +------------------------+ | task_struct 1 | |------------------------| | | |------------------------| | task 1 stack (int) | |------------------------| | task 1 saved registers | |------------------------| | task 1 stack | 0x00401000 +------------------------+ | task_struct 2 | |------------------------| | | 0x00402000 +------------------------+ | | | | 0x3F000000 +------------------------+ | device registers | 0x40000000 +------------------------+cpu_switch_toruns task 2. In order to do this, it executes exactly the same sequence of steps that it does for task 1. Task 2 started to execute and it stack grows. Note, that we didn't return from an interrupt at this point, but this is ok because interrupts now are enabled (interrupts have been enabled previously in timer_tick beforeschedulewas called)0 +------------------------+ | kernel image | |------------------------| | | |------------------------| | init task stack | 0x00400000 +------------------------+ | task_struct 1 | |------------------------| | | |------------------------| | task 1 stack (int) | |------------------------| | task 1 saved registers | |------------------------| | task 1 stack | 0x00401000 +------------------------+ | task_struct 2 | |------------------------| | | |------------------------| | task 2 stack | 0x00402000 +------------------------+ | | | | 0x3F000000 +------------------------+ | device registers | 0x40000000 +------------------------+- Another timer interrupt happens and
kernel_entrysaves all general purpose registers +elr_el1andspsr_el1at the bottom of task 2 stack. Task 2 interrupt stack begins to grow.0 +------------------------+ | kernel image | |------------------------| | | |------------------------| | init task stack | 0x00400000 +------------------------+ | task_struct 1 | |------------------------| | | |------------------------| | task 1 stack (int) | |------------------------| | task 1 saved registers | |------------------------| | task 1 stack | 0x00401000 +------------------------+ | task_struct 2 | |------------------------| | | |------------------------| | task 2 stack (int) | |------------------------| | task 2 saved registers | |------------------------| | task 2 stack | 0x00402000 +------------------------+ | | | | 0x3F000000 +------------------------+ | device registers | 0x40000000 +------------------------+ scheduleis called. It observes that all tasks have their counters set to 0 and set counters to their tasks priorities.scheduleselects init task to run. (This is because all tasks now have their counters set to 1 and init task is the first in the list). But actually, it would be fully legal forscheduleto select task 1 or task 2 at this point, because their counters has equal values. We are more interested in the case when task 1 is selected so let's now assume that this is what had happened.cpu_switch_tois called and it restores previously saved callee-saved registers from task 1cpu_context. Link register now points here because this is the place from whichcpu_switch_towas called last time when task 1 was executed.sppoints to the bottom of task 1 interrupt stack.timer_tickfunction resumes execution, starting from this line. It disables interrupts and finally kernel_exit is executed. By the timekernel_exitis called, task 1 interrupt stack is collapsed to 0.0 +------------------------+ | kernel image | |------------------------| | | |------------------------| | init task stack | 0x00400000 +------------------------+ | task_struct 1 | |------------------------| | | |------------------------| | task 1 saved registers | |------------------------| | task 1 stack | 0x00401000 +------------------------+ | task_struct 2 | |------------------------| | | |------------------------| | task 2 stack (int) | |------------------------| | task 2 saved registers | |------------------------| | task 2 stack | 0x00402000 +------------------------+ | | | | 0x3F000000 +------------------------+ | device registers | 0x40000000 +------------------------+kernel_exitrestores all general purpose registers as well aselr_el1andspsr_el1.elr_el1now points somewhere in the middle of theprocessfunction.sppoints to the bottom of task 1 stack.0 +------------------------+ | kernel image | |------------------------| | | |------------------------| | init task stack | 0x00400000 +------------------------+ | task_struct 1 | |------------------------| | | |------------------------| | task 1 stack | 0x00401000 +------------------------+ | task_struct 2 | |------------------------| | | |------------------------| | task 2 stack (int) | |------------------------| | task 2 saved registers | |------------------------| | task 2 stack | 0x00402000 +------------------------+ | | | | 0x3F000000 +------------------------+ | device registers | 0x40000000 +------------------------+kernel_exitexecuteseretinstruction witch useselr_el1register to jump back toprocessfunction. Task 1 resumes it normal execution.
The described above sequence of steps is very important — I personally consider it one of the most important things in the whole tutorial. If you have difficulties with understanding it, I can advise you to work on the exercise number 1 from this lesson.
We are done with scheduling, but right now our kernel can manage only kernel threads: they are executed at EL1 and can directly access any kernel functions or data. In the next 2 lessons we are going fix this and introduce system calls and virtual memory.