cpu_ops、suspend_ops、arm_idle_driver以及machine_restart/machine_power_off到底层PSCI Firmware分析

在内核中针对的cpu的操作，比如arm_cpuidle_init、arm_cpuidle_suspend、boot_secondary、secondary_start_kernel、op_cpu_disable、op_cpu_kill、cpu_die、smp_cpu_setup、smp_prepare_cpus的都会回落到对cpu_ops的调用。

cpu_ops将针对底层cpu的操作抽象为一系列回调函数，以统一的形式向上层提供API。

cpu_psci_ops作为cpu_ops的一个特殊实现，将cpu_ops关联到PSCI的psci_ops。

psci_ops的函数在PSCI Firmware中实现，提供一系列基于Function ID的调用。

这种分层思想将内核通用cpu_operations和硬件相关部分分隔开。

从上图可知，cpu_ops和suspend_ops将内核通用API和底层arch-specific代码区隔开；cpu_ops和suspend_ops分别调用cpu_psci_ops和psci_suspend_ops，这些回调函数最终都会回落到PSCI Firmware提供的接口。machine_restart/machine_power_off直接调用PSCI提供的接口。

cpu_operations及应用场景

首先分析一些cpu_operations这个结构体：

struct cpu_operations {
    const char    *name;
    int        (*cpu_init)(unsigned int);  读取必要的数据准备初始化。
    int        (*cpu_prepare)(unsigned int);  启动前准备工作
    int        (*cpu_boot)(unsigned int);  启动一个CPU
    void        (*cpu_postboot)(void);  执行boot后的清理工作
#ifdef CONFIG_HOTPLUG_CPU
    int        (*cpu_disable)(unsigned int cpu);  关闭CPU之前的准备工作
    void        (*cpu_die)(unsigned int cpu);  关闭CPU
    int        (*cpu_kill)(unsigned int cpu);  确认是否关闭
#endif
#ifdef CONFIG_CPU_IDLE
    int        (*cpu_init_idle)(unsigned int);  读取CPU idle状态的参数
    int        (*cpu_suspend)(unsigned long);  suspend一个CPU，并且保存上下文
#endif
};

cpu_init

static int __init smp_cpu_setup(int cpu)
{
    if (cpu_read_ops(cpu))
        return -ENODEV;

if (cpu_ops[cpu]->cpu_init(cpu))
        return -ENODEV;

set_cpu_possible(cpu, true);

return 0;
}

获取指定cpu的cpu_ops，执行cpu_init回调函数进行初始化。并将此cpu设置为possible。

cpu_prepare

void __init smp_prepare_cpus(unsigned int max_cpus)
{
    int err;
    unsigned int cpu, ncores = num_possible_cpus();

init_cpu_topology();  填充cpu_topology结构体数组

smp_store_cpu_info(smp_processor_id());

/*
     * are we trying to boot more cores than exist?
     */
    if (max_cpus > ncores)  不能超过possible cpu数目
        max_cpus = ncores;

/* Don't bother if we're effectively UP */
    if (max_cpus <= 1)
        return;

/*
     * Initialise the present map (which describes the set of CPUs
     * actually populated at the present time) and release the
     * secondaries from the bootloader.
     *
     * Make sure we online at most (max_cpus - 1) additional CPUs.
     */
    max_cpus--;
    for_each_possible_cpu(cpu) {
        if (max_cpus == 0)
            break;

if (cpu == smp_processor_id())
            continue;

if (!cpu_ops[cpu])
            continue;

err = cpu_ops[cpu]->cpu_prepare(cpu);  执行.cpu_prepare回调函数，将指定cpu设置为present。
        if (err)
            continue;

set_cpu_present(cpu, true);
        max_cpus--;
    }
}

cpu_boot

static int boot_secondary(unsigned int cpu, struct task_struct *idle)
{
    if (cpu_ops[cpu]->cpu_boot)
        return cpu_ops[cpu]->cpu_boot(cpu);

return -EOPNOTSUPP;
}

cpu_postboot

asmlinkage void secondary_start_kernel(void)  被汇编调用，作为secondary CPU的启动入口
{
    struct mm_struct *mm = &init_mm;
    unsigned int cpu = smp_processor_id();

/*
     * All kernel threads share the same mm context; grab a
     * reference and switch to it.
     */
    atomic_inc(&mm->mm_count);
    current->active_mm = mm;

set_my_cpu_offset(per_cpu_offset(smp_processor_id()));

/*
     * TTBR0 is only used for the identity mapping at this stage. Make it
     * point to zero page to avoid speculatively fetching new entries.
     */
    cpu_set_reserved_ttbr0();
    local_flush_tlb_all();
    cpu_set_default_tcr_t0sz();

preempt_disable();
    trace_hardirqs_off();

/*
     * If the system has established the capabilities, make sure
     * this CPU ticks all of those. If it doesn't, the CPU will
     * fail to come online.
     */
    verify_local_cpu_capabilities();

if (cpu_ops[cpu]->cpu_postboot)
        cpu_ops[cpu]->cpu_postboot();

/*
     * Log the CPU info before it is marked online and might get read.
     */
    cpuinfo_store_cpu();

/*
     * Enable GIC and timers.
     */
    notify_cpu_starting(cpu);

smp_store_cpu_info(cpu);

/*
     * OK, now it's safe to let the boot CPU continue.  Wait for
     * the CPU migration code to notice that the CPU is online
     * before we continue.
     */
    pr_info("CPU%u: Booted secondary processor [%08x]\n",
                     cpu, read_cpuid_id());
    set_cpu_online(cpu, true); 至此CPU可以设置为online状态
    complete(&cpu_running);

local_dbg_enable();
    local_irq_enable();
    local_async_enable();

/*
     * OK, it's off to the idle thread for us
     */
    cpu_startup_entry(CPUHP_ONLINE);
}

cpu_disable

static int op_cpu_disable(unsigned int cpu)
{
    /*
     * If we don't have a cpu_die method, abort before we reach the point
     * of no return. CPU0 may not have an cpu_ops, so test for it.
     */
    if (!cpu_ops[cpu] || !cpu_ops[cpu]->cpu_die)
        return -EOPNOTSUPP;

/*
     * We may need to abort a hot unplug for some other mechanism-specific
     * reason.
     */
    if (cpu_ops[cpu]->cpu_disable)
        return cpu_ops[cpu]->cpu_disable(cpu);

return 0;
}

cpu_die

void cpu_die(void)
{
    unsigned int cpu = smp_processor_id();

idle_task_exit();

local_irq_disable();

/* Tell __cpu_die() that this CPU is now safe to dispose of */
    (void)cpu_report_death();

/*
     * Actually shutdown the CPU. This must never fail. The specific hotplug
     * mechanism must perform all required cache maintenance to ensure that
     * no dirty lines are lost in the process of shutting down the CPU.
     */
    cpu_ops[cpu]->cpu_die(cpu);

BUG();
}

cpu_kill

static int op_cpu_kill(unsigned int cpu)
{
    /*
     * If we have no means of synchronising with the dying CPU, then assume
     * that it is really dead. We can only wait for an arbitrary length of
     * time and hope that it's dead, so let's skip the wait and just hope.
     */
    if (!cpu_ops[cpu]->cpu_kill)
        return 0;

return cpu_ops[cpu]->cpu_kill(cpu);
}

cpu_init_idle和cpu_suspend

这两个回调函数主要用于idle初始化和进入idle状态。

arm_idle_init解析DeviceTree的"arm,idle-state"，注册ARM的cpuidle驱动arm_idle_driver，

static int __init arm_idle_init(void)
{
    int cpu, ret;
    struct cpuidle_driver *drv = &arm_idle_driver;
    struct cpuidle_device *dev;

/*
     * Initialize idle states data, starting at index 1.
     * This driver is DT only, if no DT idle states are detected (ret == 0)
     * let the driver initialization fail accordingly since there is no
     * reason to initialize the idle driver if only wfi is supported.
     */
    ret = dt_init_idle_driver(drv, arm_idle_state_match, 1);
    if (ret <= 0)
        return ret ? : -ENODEV;

ret = cpuidle_register_driver(drv);  注册arm_idle_driver驱动函数
    if (ret) {
        pr_err("Failed to register cpuidle driver\n");
        return ret;
    }

/*
     * Call arch CPU operations in order to initialize
     * idle states suspend back-end specific data
     */
    for_each_possible_cpu(cpu) {
        ret = arm_cpuidle_init(cpu);  获取arch-specific的idle处理参数，这里对应cpu_psci_cpu_init_idle。

/*
         * Skip the cpuidle device initialization if the reported
         * failure is a HW misconfiguration/breakage (-ENXIO).
         */
        if (ret == -ENXIO)
            continue;

if (ret) {
            pr_err("CPU %d failed to init idle CPU ops\n", cpu);
            goto out_fail;
        }

dev = kzalloc(sizeof(*dev), GFP_KERNEL);
        if (!dev) {
            pr_err("Failed to allocate cpuidle device\n");
            goto out_fail;
        }
        dev->cpu = cpu;

ret = cpuidle_register_device(dev);
        if (ret) {
            pr_err("Failed to register cpuidle device for CPU %d\n",
                   cpu);
            kfree(dev);
            goto out_fail;
        }
    }

return 0;
out_fail:
    while (--cpu >= 0) {
        dev = per_cpu(cpuidle_devices, cpu);
        cpuidle_unregister_device(dev);
        kfree(dev);
    }

cpuidle_unregister_driver(drv);

return ret;
}

arm_cpuidle_init调用.cpu_init_idle回调函数。

int __init arm_cpuidle_init(unsigned int cpu)
{
    int ret = -EOPNOTSUPP;

if (cpu_ops[cpu] && cpu_ops[cpu]->cpu_init_idle)
        ret = cpu_ops[cpu]->cpu_init_idle(cpu);

return ret;
}

arm_enter_idle_state根据参数idx使CPU进入特定的idle状态，

static int arm_enter_idle_state(struct cpuidle_device *dev,
                struct cpuidle_driver *drv, int idx)
{
    int ret;

if (!idx) {
        cpu_do_idle();  如果idx为0，则cpu_do_idle。
        return idx;
    }

ret = cpu_pm_enter();
    if (!ret) {
        /*
         * Pass idle state index to cpu_suspend which in turn will
         * call the CPU ops suspend protocol with idle index as a
         * parameter.
         */
        ret = arm_cpuidle_suspend(idx);  调用底层arch-specific处理函数。

cpu_pm_exit();
    }

return ret ? -1 : idx;
}

cpu_do_idle使CPU进入WFI状态。

ENTRY(cpu_do_idle)
    dsb    sy                // WFI may enter a low-power mode
    wfi
    ret
ENDPROC(cpu_do_idle)

int arm_cpuidle_suspend(int index)
{
    int cpu = smp_processor_id();

/*
     * If cpu_ops have not been registered or suspend
     * has not been initialized, cpu_suspend call fails early.
     */
    if (!cpu_ops[cpu] || !cpu_ops[cpu]->cpu_suspend)
        return -EOPNOTSUPP;
    return cpu_ops[cpu]->cpu_suspend(index);
}

cpu_ops到arch-dependent的关联

以start_kernel为起点，查看从内核开始到获取cpu_ops的路径如下：

start_kernel
    -->setup_arch
        -->cpu_read_bootcpu_ops  只获取bootcpu的cpu_ops
            -->cpu_read_bootcpu_ops
                -->cpu_read_ops(0)
        -->smp_init_cpus  获取nonboot cpu的cpu_ops
            -->smp_cpu_setup
                -->cpu_read_ops

cpu_read_ops是获取cpu_ops的关键，参数是cpu的序列号，输出是cpu_ops[cpu]。

int __init cpu_read_ops(int cpu)
{
    const char *enable_method = cpu_read_enable_method(cpu);  从DeviceTree获取enable_method字符串

if (!enable_method)
        return -ENODEV;

cpu_ops[cpu] = cpu_get_ops(enable_method);  根据enable_method字符串在supported_cpu_ops获取指针
    if (!cpu_ops[cpu]) {
        pr_warn("Unsupported enable-method: %s\n", enable_method);
        return -EOPNOTSUPP;
    }

return 0;
}

通过cpu0的DeviceTree可以看出enable-method为pcsi。

支持的cpu_operations有：

static const struct cpu_operations *supported_cpu_ops[] __initconst = {
    &smp_spin_table_ops,
    &cpu_psci_ops,
    NULL,
};

所以cpu_ops=&cpu_psci_ops。

suspend_ops

在enter_state—>suspend_devices_and_enter—>suspend_enter有针对suspend_ops->enter的调用，suspend_ops的赋值在psci_init_system_suspend中：

static void __init psci_init_system_suspend(void)
{
    int ret;

if (!IS_ENABLED(CONFIG_SUSPEND))
        return;

ret = psci_features(PSCI_FN_NATIVE(1_0, SYSTEM_SUSPEND));

if (ret != PSCI_RET_NOT_SUPPORTED)
        suspend_set_ops(&psci_suspend_ops);  suspend_ops指向psci_suspend_ops
}

psci_suspend_ops是platform_suspend_ops类型的函数结构体，这里只有两个成员。

static const struct platform_suspend_ops psci_suspend_ops = {
    .valid          = suspend_valid_only_mem,  返回系统支持的suspend类型
    .enter          = psci_system_suspend_enter,  进入suspend状态，参数是状态值，这里只能有mem
};

psci_system_suspend_enter调用cpu_suspend并且给出结束回调函数。

cpu_suspend调用__cpu_suspend_enter，并进行TTBR0、TLB、TCR、MM等的操作，这些都涉及到汇编处理。

__cpu_suspend_enter保存当前CPU状态，其中x0保存结束回调函数的参数，x1是结束回调函数指针地址。

ENTRY(__cpu_suspend_enter)
    stp    x29, lr, [sp, #-96]!
    stp    x19, x20, [sp,#16]
    stp    x21, x22, [sp,#32]
    stp    x23, x24, [sp,#48]
    stp    x25, x26, [sp,#64]
    stp    x27, x28, [sp,#80]
    /*
     * Stash suspend finisher and its argument in x20 and x19
     */
    mov    x19, x0
    mov    x20, x1
    mov    x2, sp
    sub    sp, sp, #CPU_SUSPEND_SZ    // allocate cpu_suspend_ctx
    mov    x0, sp
    /*
     * x0 now points to struct cpu_suspend_ctx allocated on the stack
     */
    str    x2, [x0, #CPU_CTX_SP]
    ldr    x1, =sleep_save_sp
    ldr    x1, [x1, #SLEEP_SAVE_SP_VIRT]
    mrs    x7, mpidr_el1
    ldr    x9, =mpidr_hash
    ldr    x10, [x9, #MPIDR_HASH_MASK]
    /*
     * Following code relies on the struct mpidr_hash
     * members size.
     */
    ldp    w3, w4, [x9, #MPIDR_HASH_SHIFTS]
    ldp    w5, w6, [x9, #(MPIDR_HASH_SHIFTS + 8)]
    compute_mpidr_hash x8, x3, x4, x5, x6, x7, x10
    add    x1, x1, x8, lsl #3
    bl    __cpu_suspend_save
    /*
     * Grab suspend finisher in x20 and its argument in x19
     */
    mov    x0, x19  将备份的arg和fn恢复到x0, x1
    mov    x1, x20
    /*
     * We are ready for power down, fire off the suspend finisher
     * in x1, with argument in x0
     */
    blr    x1  执行suspend结束函数回调，这里指的是psci_system_suspend。
        /*
     * Never gets here, unless suspend finisher fails.
     * Successful cpu_suspend should return from cpu_resume, returning
     * through this code path is considered an error
     * If the return value is set to 0 force x0 = -EOPNOTSUPP
     * to make sure a proper error condition is propagated
     */
    cmp    x0, #0
    mov    x3, #-EOPNOTSUPP
    csel    x0, x3, x0, eq
    add    sp, sp, #CPU_SUSPEND_SZ    // rewind stack pointer
    ldp    x19, x20, [sp, #16]
    ldp    x21, x22, [sp, #32]
    ldp    x23, x24, [sp, #48]
    ldp    x25, x26, [sp, #64]
    ldp    x27, x28, [sp, #80]
    ldp    x29, lr, [sp], #96
    ret
ENDPROC(__cpu_suspend_enter)

psci_system_suspend作为cpu_suspend收尾函数，调用psci的suspend函数，让CPU进入suspend。

static int psci_system_suspend(unsigned long unused)
{
    return invoke_psci_fn(PSCI_FN_NATIVE(1_0, SYSTEM_SUSPEND),
                  virt_to_phys(cpu_resume), 0, 0);
}

psci的SYSTEM_SUSPEND entry_point参数是cpu_resume，这个函数会在CPU唤醒之后执行的入口点。

cpu_resume在arc/arm64/kernel/sleep.S中定义，和__cpu_suspend_enter是相反的过程，恢复sp指针、pc指针、MMU等。

ENTRY(cpu_resume)
    bl    el2_setup        // if in EL2 drop to EL1 cleanly
    mrs    x1, mpidr_el1
    adrp    x8, mpidr_hash
    add x8, x8, #:lo12:mpidr_hash // x8 = struct mpidr_hash phys address
        /* retrieve mpidr_hash members to compute the hash */
    ldr    x2, [x8, #MPIDR_HASH_MASK]
    ldp    w3, w4, [x8, #MPIDR_HASH_SHIFTS]
    ldp    w5, w6, [x8, #(MPIDR_HASH_SHIFTS + 8)]
    compute_mpidr_hash x7, x3, x4, x5, x6, x1, x2
        /* x7 contains hash index, let's use it to grab context pointer */
    ldr_l    x0, sleep_save_sp + SLEEP_SAVE_SP_PHYS
    ldr    x0, [x0, x7, lsl #3]
    /* load sp from context */
    ldr    x2, [x0, #CPU_CTX_SP]
    /* load physical address of identity map page table in x1 */
    adrp    x1, idmap_pg_dir
    mov    sp, x2
    /*
     * cpu_do_resume expects x0 to contain context physical address
     * pointer and x1 to contain physical address of 1:1 page tables
     */
    bl    cpu_do_resume        // PC relative jump, MMU off
    b    cpu_resume_mmu        // Resume MMU, never returns
ENDPROC(cpu_resume)

cpu_psci_ops分析

cpu_psci_ops结构体可以说是cpu_operations和psci_operations的桥梁，他讲cpu_operations的一些列回调函数，映射到psci_operations。

const struct cpu_operations cpu_psci_ops = {
    .name        = "psci",
#ifdef CONFIG_CPU_IDLE
    .cpu_init_idle    = cpu_psci_cpu_init_idle,  从DeviceTree获取CPU idle状态数据
    .cpu_suspend    = cpu_psci_cpu_suspend,  根据是否丢失上下文来选择是psci_ops.cpu_suspend还是cpu_suspend
#endif
    .cpu_init    = cpu_psci_cpu_init,  为空
    .cpu_prepare    = cpu_psci_cpu_prepare,  只是判断psci_ops.cpu_on是否存在，不存在则返回错误。
    .cpu_boot    = cpu_psci_cpu_boot,  调用psci_ops.cpu_on
#ifdef CONFIG_HOTPLUG_CPU
    .cpu_disable    = cpu_psci_cpu_disable,  检查是否支持psci_ops.cpu_off。
    .cpu_die    = cpu_psci_cpu_die,  调用psci_ops.cpu_off
    .cpu_kill    = cpu_psci_cpu_kill,  检查指定cpu是否已经被kill
#endif
}

cpu_psci_cpu_boot

static int cpu_psci_cpu_boot(unsigned int cpu)
{
    int err = psci_ops.cpu_on(cpu_logical_map(cpu), __pa(secondary_entry));
    if (err)
        pr_err("failed to boot CPU%d (%d)\n", cpu, err);

return err;
}

CPU_ON用于secondary boot、hotplug或者big.LITTLE迁移。如果需要从一个核启动另一个核，通过CPU_ON提供一个入口地址和上下文标识。

PCSI提供必要的操作启动一个核，并且在提供的入口地址开始执行，上下文标识必须存在R0或者W0中。这里的入口地址就对应secondary_entry。

在arch/arm64/kernel/head.S中：

secondary_entry—>secondary_startup—>__secondary_switched—>secondary_start_kernel

ENTRY(secondary_entry)
    bl    el2_setup            // Drop to EL1
    bl    set_cpu_boot_mode_flag
    b    secondary_startup
ENDPROC(secondary_entry)

ENTRY(secondary_startup)
    /*
     * Common entry point for secondary CPUs.
     */
    adrp    x25, idmap_pg_dir
    adrp    x26, swapper_pg_dir
    bl    __cpu_setup            // initialise processor

ldr    x21, =secondary_data
    ldr    x27, =__secondary_switched    // address to jump to after enabling the MMU
    b    __enable_mmu
ENDPROC(secondary_startup)

ENTRY(__secondary_switched)
    ldr    x0, [x21]            // get secondary_data.stack
    mov    sp, x0
    mov    x29, #0
    b    secondary_start_kernel
ENDPROC(__secondary_switched)

在secondary_start_kernel将CPU设置为online，并调用.cpu_postboot回调函数，进行boot后处理。然后cpu_startup_entry启动idle线程。

cpu_psci_cpu_init_idle

static int __maybe_unused cpu_psci_cpu_init_idle(unsigned int cpu)
{
    int i, ret, count = 0;
    u32 *psci_states;
    struct device_node *state_node, *cpu_node;

cpu_node = of_get_cpu_node(cpu, NULL);
    if (!cpu_node)
        return -ENODEV;

/*
     * If the PSCI cpu_suspend function hook has not been initialized
     * idle states must not be enabled, so bail out
     */
    if (!psci_ops.cpu_suspend)
        return -EOPNOTSUPP;

/* Count idle states */
    while ((state_node = of_parse_phandle(cpu_node, "cpu-idle-states",
                          count))) {
        count++;
        of_node_put(state_node);
    }

if (!count)
        return -ENODEV;

psci_states = kcalloc(count, sizeof(*psci_states), GFP_KERNEL);
    if (!psci_states)
        return -ENOMEM;

for (i = 0; i < count; i++) {
        u32 state;

state_node = of_parse_phandle(cpu_node, "cpu-idle-states", i);

ret = of_property_read_u32(state_node,
                       "arm,psci-suspend-param",
                       &state);
        if (ret) {
            pr_warn(" * %s missing arm,psci-suspend-param property\n",
                state_node->full_name);
            of_node_put(state_node);
            goto free_mem;
        }

of_node_put(state_node);
        pr_debug("psci-power-state %#x index %d\n", state, i);
        if (!psci_power_state_is_valid(state)) {
            pr_warn("Invalid PSCI power state %#x\n", state);
            ret = -EINVAL;
            goto free_mem;
        }
        psci_states[i] = state;
    }
    /* Idle states parsed correctly, initialize per-cpu pointer */
    per_cpu(psci_power_state, cpu) = psci_states;
    return 0;

free_mem:
    kfree(psci_states);
    return ret;
}

1.解析DeviceTree中cpu下的cpu-idle-states属性

2.从每个state中获取arm,psci-suspend-param的参数，并验证是否有效。

3.初始化per-CPU类型的指针psci_power_state。

cpu_psci_cpu_suspend

static int __maybe_unused cpu_psci_cpu_suspend(unsigned long index)
{
    int ret;
    u32 *state = __this_cpu_read(psci_power_state);  从psci_power_state中读取suspend的state参数。
    /*
     * idle state index 0 corresponds to wfi, should never be called
     * from the cpu_suspend operations
     */
    if (WARN_ON_ONCE(!index))
        return -EINVAL;

if (!psci_power_state_loses_context(state[index - 1]))
        ret = psci_ops.cpu_suspend(state[index - 1], 0);
    else
        ret = cpu_suspend(index, psci_suspend_finisher);

return ret;
}

psci_ops

由于acpi_disabled，所以psci通过DeviceTree获取相关参数。

start_kernel
    -->setup_arch
        -->psci_dt_init  这个函数在cpu_ops之前，因为cpu_ops依赖psci_ops

psci有不同版本，需要通过DeviceTree获取版本信息和使用的method（是smc还是）。

通过查看DeviceTree可以看到对应的是psci_0_2_init。

static const struct of_device_id const psci_of_match[] __initconst = {
    { .compatible = "arm,psci",    .data = psci_0_1_init},
    { .compatible = "arm,psci-0.2",    .data = psci_0_2_init},
    { .compatible = "arm,psci-1.0",    .data = psci_0_2_init},
    {},
};

psci_dt_init解析DeviceTree执行对应psci版本的初始化函数。

int __init psci_dt_init(void)
{
    struct device_node *np;
    const struct of_device_id *matched_np;
    psci_initcall_t init_fn;

np = of_find_matching_node_and_match(NULL, psci_of_match, &matched_np);

if (!np)
        return -ENODEV;

init_fn = (psci_initcall_t)matched_np->data;
    return init_fn(np);
}

psci_0_2_init设置method，然后调用psci_probe：

static int __init psci_0_2_init(struct device_node *np)
{
    int err;

err = get_set_conduit_method(np);  从DeviceTree可知invoke_psci_fn = __invoke_psci_fn_smc

if (err)
        goto out_put_node;
    /*
     * Starting with v0.2, the PSCI specification introduced a call
     * (PSCI_VERSION) that allows probing the firmware version, so
     * that PSCI function IDs and version specific initialization
     * can be carried out according to the specific version reported
     * by firmware
     */
    err = psci_probe();

out_put_node:
    of_node_put(np);
    return err;
}

psci_probe设置版本高于0.2的PSCI回调函数，以及arm_pm_restart和pm_power_off。

static void __init psci_0_2_set_functions(void)
{
    pr_info("Using standard PSCI v0.2 function IDs\n");
    psci_function_id[PSCI_FN_CPU_SUSPEND] =
                    PSCI_FN_NATIVE(0_2, CPU_SUSPEND);
    psci_ops.cpu_suspend = psci_cpu_suspend;

psci_function_id[PSCI_FN_CPU_OFF] = PSCI_0_2_FN_CPU_OFF;
    psci_ops.cpu_off = psci_cpu_off;

psci_function_id[PSCI_FN_CPU_ON] = PSCI_FN_NATIVE(0_2, CPU_ON);
    psci_ops.cpu_on = psci_cpu_on;

psci_function_id[PSCI_FN_MIGRATE] = PSCI_FN_NATIVE(0_2, MIGRATE);
    psci_ops.migrate = psci_migrate;

psci_ops.affinity_info = psci_affinity_info;

psci_ops.migrate_info_type = psci_migrate_info_type;

arm_pm_restart = psci_sys_reset;

pm_power_off = psci_sys_poweroff;
}

这些函数都有一个共性invoke_psci_fn(PSCI_0_2_FN_SYSTEM_OFF, 0, 0, 0)，着这里invoke_psci_fn指向__invoke_psci_fn_smc 。

__invoke_psci_fn_smc指向arch/arm64/kernel/psci-call.S定义的函数：

/* int __invoke_psci_fn_smc(u64 function_id, u64 arg0, u64 arg1, u64 arg2) */
ENTRY(__invoke_psci_fn_smc)
    smc    #0
    ret
ENDPROC(__invoke_psci_fn_smc)

http://infocenter.arm.com/help/topic/com.arm.doc.den0022c/DEN0022C_Power_State_Coordination_Interface.pdf Chapter5有PSCI函数圆形和相关参数返回值的介绍。

第一个参数是Function ID，后面三个参数作为Function ID的参数。如果使用的是32位的参数，后三个参数通过r0-r3传递给Function ID，r0存放返回值；如果使用64位的参数，后三个参数通过W0-W3传递，w0存放返回值。这些Function ID的实现，在对应的Firmware中，但是可以通过上述pdf查看输入输出细节。

PSCI除了提供psci_ops的回调函数之外，还提供以restart和power off的arch-dependent函数arm_pm_restart和pm_power_off

比如machine_power_off和machine_restart调用：

void machine_power_off(void)
{
    local_irq_disable();
    smp_send_stop();
    if (pm_power_off)
        pm_power_off();
}

void machine_restart(char *cmd)
{
    /* Disable interrupts first */
    local_irq_disable();
    smp_send_stop();

/*
     * UpdateCapsule() depends on the system being reset via
     * ResetSystem().
     */
    if (efi_enabled(EFI_RUNTIME_SERVICES))
        efi_reboot(reboot_mode, NULL);

/* Now call the architecture specific reboot code. */
    if (arm_pm_restart)
        arm_pm_restart(reboot_mode, cmd);
    else
        do_kernel_restart(cmd);

/*
     * Whoops - the architecture was unable to reboot.
     */
    printk("Reboot failed -- System halted\n");
    while (1);
}

参考文档

Linux CPU core的电源管理(3)_cpu ops：http://www.wowotech.net/pm_subsystem/cpu_ops.html