Skip to content

fork()

Memory Manager

We need to duplicate the address space in the memory manager side. Follow the traditional fork() semantic, both the existing and newly created address space will be write-protected.

Since we have the flexibility to implement any VM organization, we should be careful while duplicating the address space. Currently, we are using page-based VM, thus the duplicating is basically creating a new pgd and copy existing pgtables, and further downgrade permission to read-only. This is now performed by lego_copy_page_range().

The final write-protect is performed by lego_copy_one_pte():

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
static inline int lego_copy_one_pte(..)
{
    ..
    /*
     * If it's a COW mapping, write protect it both
     * in the parent and the child
     */
    if (is_cow_mapping(vm_flags)) {
        ptep_set_wrprotect(src_pte);   
        pte = pte_wrprotect(pte);      
    }
    ...
}

Duplicate VM Free Pool

TODO Yutong

Processor Manager

Boring implementation details in the processor manager side.

Entry Points

  • fork()
  • vfork()
  • clone()
  • kernel_thread()

All of them land on do_fork(), which is Lego’s main fork function.

do_fork()

There are mainly three parts within do_fork(): 1) copy_process(), which duplicates a new task based on current, including allocate new kernel stack, new task_struct, increase mm reference counter, etc. 2) If we are creating a new process, then tell global monitor or memory manager to let them update bookkeeping and create corresponding data structures. 3) wake_up_new_task(), which gives away the newly created task to local scheduler.

copy_process()

The routine is kind of boring. It do a lot dirty work to copy information from calling thread to new thread. The most important data structures of course are task_struct, mm_sturct, sighand, and so on. This section only talks about few of them, and leave others to readers who are interested.

Sanity Checking

Mainly check if clone_flags are passed properly. For example, if user is creating a new thread, that implies certain data structures are shared, cause new thread belongs to the same process with the calling thread. If CLONE_THREAD is passed, then CLONE_SIGHAND, CLONE_VM, and so on must be set as well.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
    /*
     * Thread groups must share signals as well, and detached threads
     * can only be started up within the thread group.
     */
    if ((clone_flags & CLONE_THREAD) && !(clone_flags & CLONE_SIGHAND))
        return ERR_PTR(-EINVAL);

    /*
     * Shared signal handlers imply shared VM. By way of the above,
     * thread groups also imply shared VM. Blocking this case allows
     * for various simplifications in other code.
     */
    if ((clone_flags & CLONE_SIGHAND) && !(clone_flags & CLONE_VM))
        return ERR_PTR(-EINVAL);

dup_task_struct()

Two main things: 1) duplicate a new task_struct, 2) duplicate a new kernel stack. x86 is just a weird architecture, the size of task_struct depends on the size of fpu. So the allocation and duplication need to callback to x86-specific code to duplicate the task_struct and fpu info.

1
2
3
4
5
6
int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
{
    memcpy(dst, src, arch_task_struct_size);

    return fpu__copy(&dst->thread.fpu, &src->thread.fpu);
}
The stack duplication is fairly simple, just copy everything from the old stack to new stack. Of course, it needs to setup the thread_info to points to this new thread, so the current macro will work.
1
2
3
4
5
6
7
8
static void setup_thread_stack(struct task_struct *p, struct task_struct *org)
{
        /* Duplicate whole stack! */
        *task_thread_info(p) = *task_thread_info(org);

        /* Make the `current' macro work */
        task_thread_info(p)->task = p;
}

copy_mm()

This is where threads within a process will share the virtual address space happens. If we are creating a new process, then this function will create a new mm_struct, and also a new pgd:

1
2
3
4
5
6
7
8
9
/*
 * pgd_alloc() will duplicate the identity kernel mapping
 * but leaves other entries empty:
 */
mm->pgd = pgd_alloc(mm);
if (unlikely(!mm->pgd)) {
        kfree(mm);
        return NULL;
}

Duplicate pcache data

TODO

TODO: hook with pcache

We need to duplicate the pcache vm_range array, once Yutong finished the code.

setup_sched_fork()

Callback to scheduler to setup this new task. It may reset all scheduler related information. Here we also have a chance to change this task’s scheduler class:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
int setup_sched_fork(unsigned long clone_flags, struct task_struct *p)
{
        int cpu = get_cpu();

        __sched_fork(clone_flags, p);

        p->state = TASK_NEW;
        ...
        if (unlikely(p->sched_reset_on_fork)) {
                if (task_has_rt_policy(p)) {
                        p->policy = SCHED_NORMAL;
                        p->static_prio = NICE_TO_PRIO(0);
                        p->rt_priority = 0;
                } else if (PRIO_TO_NICE(p->static_prio) < 0)
                        p->static_prio = NICE_TO_PRIO(0);

                p->prio = p->normal_prio = __normal_prio(p);
                set_load_weight(p);
                ...
        }    

        if (rt_prio(p->prio))
                p->sched_class = &rt_sched_class;
        else {
                p->sched_class = &fair_sched_class;
                set_load_weight(p);
        }    

        __set_task_cpu(p, cpu);
        if (p->sched_class->task_fork)
                p->sched_class->task_fork(p);

        ...
}
Allocate new pid

In both Lego and Linux, we don’t allocate new pid for a new thread, if that thread is an idle thread. So callers of do_fork needs to pass something to let do_fork know. In Linux, they use struct pid, init_struct_pid to check. In Lego, we introduce an new clone_flag CLONE_IDLE_THREAD. If that flag is set, do_fork() will try to allocate a new pid for the new thread. Otherwise, it will be 0:

1
2
3
4
5
6
/* clone idle thread, whose pid is 0 */
if (!(clone_flags & CLONE_IDLE_THREAD)) {
        pid = alloc_pid(p);
        if (!pid)
                goto out_cleanup_thread;
}

So, only the init_idle() function can pass this CLONE_IDLE_THREAD down. All other usages are wrong and should be reported.

In order to avoid conflict with Linux clone_flag, we define it as:

1
#define CLONE_IDLE_THREAD       0x100000000

SETTID/CLEARTID

These are some futex related stuff. I will cover these stuff in futex document:

1
2
3
4
5
6
7
8
9
p->set_child_tid = (clone_flags & CLONE_CHILD_SETTID) ? child_tidptr : NULL;
/*  
 * Clear TID on mm_release()?
 */
p->clear_child_tid = (clone_flags & CLONE_CHILD_CLEARTID) ? child_tidptr : NULL;

#ifdef CONFIG_FUTEX
p->robust_list = NULL;
#endif

copy_thread_tls()

This is the most interesting function. Cover later.

p2m_fork()

In order to track user activities, we need to know when user are going to create new process. Fork is the best time and the only time we kernel know. So, Lego adds this special hook to tell remote global monitor or memory manager that there is a new process going to be created. Upon receiving this message, remote monitor will update its bookkeeping for this specific user/vNode.

1
2
3
4
5
6
7
8
/* Tell remote memory component */
#ifdef CONFIG_COMP_PROCESSOR
if (clone_flags & CLONE_GLOBAL_THREAD) {
        ...
        p2m_fork(p, clone_flags);
        ...
}   
#endif

The CLONE_GLOBAL_THREAD should only be set, if the following cases happen:

  • fork()
  • vfork()
  • clone(), without CLONE_THREAD being set

In order to avoid conflict with Linux clone_flag, we define it as:

1
#define CLONE_GLOBAL_THREAD     0x200000000

wake_up_new_task()

The last step of do_fork is waking up the new thread or process, which is performed by wake_up_new_task() function. The first question this function will ask is: which cpu to land? The answer comes from select_task_rq():

1
2
3
4
5
6
7
8
9
static inline
int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
{
        if (p->nr_cpus_allowed > 1)
                cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
        else
                cpu = cpumask_any(&p->cpus_allowed);
        ...
}

Clearly, this is determined by cpus_allowed, which is the same with its parent at this point. That being said, if the parent is only able to run on one specific CPU, then all its children will end up running on the same CPU when they wake up (they could change their affinity later). This is also the default on Linux: A child created via fork(2) inherits its parent's CPU affinity mask. The affinity mask is preserved across an execve(2).

After landing CPU is selected, following operation is simple: just enqueue this task into landing CPU’s runqueue, and we are done:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
void wake_up_new_task(struct task_struct *p)
{
        ...
/* Select a CPU for new thread to run */
#ifdef CONFIG_SMP
        /*   
         * Fork balancing, do it here and not earlier because:
         *  - cpus_allowed can change in the fork path
         *  - any previously selected cpu might disappear through hotplug
         */
        set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
#endif

        rq = __task_rq_lock(p);
        activate_task(rq, p, 0);
        p->on_rq = TASK_ON_RQ_QUEUED;
        ...
}


Yizhou Shan
Created: Feb 11, 2018
Last Updated: Feb 27, 2018