宋宝华 Barry Song 21cnbao@gmail.com
新浪微博： @宋宝华Barry
在Linux系统中，对于多核的ARM芯片而言，Bootrom代码中，CPU0会率先起来，引导Bootloader和Linux内核执行，而其他的核则在上电时Bootrom一般将自身置于WFI或者WFE状态，并等待CPU0给其发CPU核间中断（IPI）或事件（一般透过SEV指令）唤醒之。一个典型的启动过程如下图：
被CPU0唤醒的CPUn可以在运行过程中进行热插拔。譬如运行如下命令即可卸载CPU1并且将CPU1上的任务全部迁移到其他CPU：
# echo 0 /sys/devices/system/cpu/cpu1/online
同样地，运行如下命令可以再次启动CPU1:
# echo 1 /sys/devices/system/cpu/cpu1/online
之后CPU1会主动参与系统中各个CPU之间要运行任务的负载均衡工作。
CPU0唤醒其他 CPU的动作在内核中被封装为一个smp_operations的结构体，该结构体的成员如下：
83struct smp_operations {
84#ifdef CONFIG_SMP
85 /*
86 * Setup the set of possible CPUs (via set_cpu_possible)
87 */
88 void (*smp_init_cpus)(void);
89 /*
90 * Initialize cpu_possible map, and enable coherency
91 */
92 void (*smp_prepare_cpus)(unsigned int max_cpus);
93
94 /*
95 * Perform platform specific initialisation of the specified CPU.
96 */
97 void (*smp_secondary_init)(unsigned int cpu);
98 /*
99 * Boot a secondary CPU, and assign it the specified idle task.
100 * This also gives us the initial stack to use for this CPU.
101 */
102 int (*smp_boot_secondary)(unsigned int cpu, struct task_struct *idle);
103#ifdef CONFIG_HOTPLUG_CPU
104 int (*cpu_kill)(unsigned int cpu);
105 void (*cpu_die)(unsigned int cpu);
106 int (*cpu_disable)(unsigned int cpu);
107#endif
108#endif
109};
我们从arch/arm/mach-vexpress/v2m.c看到VEXPRESS电路板用到的smp_ops为vexpress_smp_ops：
666DT_MACHINE_START(VEXPRESS_DT, ARM-Versatile Express )
667 .dt_compat = v2m_dt_match,
668 .smp = smp_ops(vexpress_smp_ops),
669 .map_io = v2m_dt_map_io,
670 .init_early = v2m_dt_init_early,
671 .init_irq = v2m_dt_init_irq,
672 .timer = v2m_dt_timer,
673 .init_machine = v2m_dt_init,
674 .handle_irq = gic_handle_irq,
675 .restart = v2m_restart,
676MACHINE_END
透过arch/arm/mach-vexpress/platsmp.c的实现代码可以看出，smp_operations的成员函数smp_init_cpus() 即vexpress_smp_init_cpus()会探测SoC内CPU核的个数，并设置了核间通信的方式为gic_raise_softirq()。可见于vexpress_smp_init_cpus()中调用的vexpress_dt_smp_init_cpus()：
103static void __init vexpress_dt_smp_init_cpus(void)
104{
128 for (i = 0; i ncores; ++i)
129 set_cpu_possible(i, true);
130
131 set_smp_cross_call(gic_raise_softirq);
132}
而smp_operations的成员函数smp_prepare_cpus()即vexpress_smp_prepare_cpus()则会透过v2m_flags_set(virt_to_phys(versatile_secondary_startup))设置其他CPU的启动地址为versatile_secondary_startup：
179static void __init vexpress_smp_prepare_cpus(unsigned int max_cpus)
180{
181
189
190 /*
191 * Write the address of secondary startup into the
192 * system-wide flags register. The boot monitor waits
193 * until it receives a soft interrupt, and then the
194 * secondary CPU branches to this address.
195 */
196 v2m_flags_set(virt_to_phys(versatile_secondary_startup));
197}
注意这部分的具体实现方法是SoC相关的，由芯片的设计以及芯片内部的Bootrom决定。对于VEXPRESS来讲，设置方法如下：
139void __init v2m_flags_set(u32 data)
140{
141 writel(~0, v2m_sysreg_base + V2M_SYS_FLAGSCLR);
142 writel(data, v2m_sysreg_base + V2M_SYS_FLAGSSET);
143}
即填充v2m_sysreg_base + V2M_SYS_FLAGSCLR地址为0xFFFFFFFF,将其他CPU初始启动执行的指令地址填入v2m_sysreg_base + V2M_SYS_FLAGSSET。这2个地址属于芯片实现时候设定的。填入的CPUn的起始地址都透过virt_to_phys()转化为物理地址，因为此时CPUn的MMU尚未开启。
比较关键的是smp_operations的成员函数smp_boot_secondary()，它完成最终的CPUn的唤醒工作：
27static void __cpuinit write_pen_release(int val)
28{
29 pen_release = val;
30 smp_wmb();
31 __cpuc_flush_dcache_area((void *) pen_release, sizeof(pen_release));
32 outer_clean_range(__pa( pen_release), __pa( pen_release + 1));
33}
59int __cpuinit versatile_boot_secondary(unsigned int cpu, struct task_struct *idle)
60{
61 unsigned long timeout;
62
63 /*
64 * Set synchronisation state between this boot processor
65 * and the secondary one
66 */
67 spin_lock( boot_lock);
68
69 /*
70 * This is really belt and braces; we hold unintended secondary
71 * CPUs in the holding pen until we’re ready for them. However,
72 * since we haven’t sent them a soft interrupt, they shouldn’t
73 * be there.
74 */
75 write_pen_release(cpu_logical_map(cpu));
76
77 /*
78 * Send the secondary CPU a soft interrupt, thereby causing
79 * the boot monitor to read the system wide flags register,
80 * and branch to the address found there.
81 */
82 gic_raise_softirq(cpumask_of(cpu), 0);
83
84 timeout = jiffies + (1 * HZ);
85 while (time_before(jiffies, timeout)) {
86 smp_rmb();
87 if (pen_release == -1)
88 break;
89
90 udelay(10);
91 }
92
93 /*
94 * now the secondary core is starting up let it run its
95 * calibrations, then wait for it to finish
96 */
97 spin_unlock( boot_lock);
98
99 return pen_release != -1 ? -ENOSYS : 0;
100}
上述代码中高亮的部分首先会将pen_release变量设置为要唤醒的CPU核的CPU号cpu_logical_map(cpu)，而后透过gic_raise_softirq(cpumask_of(cpu), 0)给CPUcpu发起0号IPI，这个时候，CPUcpu核会从前面smp_operations中的smp_prepare_cpus()成员函数即vexpress_smp_prepare_cpus()透过v2m_flags_set()设置的其他CPU核的起始地址versatile_secondary_startup开始执行，如果顺利的话，该CPU会将原先为正数的pen_release写为-1，以便CPU0从等待pen_release成为-1的循环中跳出。
versatile_secondary_startup实现于arch/arm/plat-versatile/headsmp.S，是一段汇编：
21ENTRY(versatile_secondary_startup)
22 mrc p15, 0, r0, c0, c0, 5
23 and r0, r0, #15
24 adr r4, 1f
25 ldmia r4, {r5, r6}
26 sub r4, r4, r5
27 add r6, r6, r4
28pen: ldr r7, [r6]
29 cmp r7, r0
30 bne pen
31
32 /*
33 * we’ve been released from the holding pen: secondary_stack
34 * should now contain the SVC stack for this core
35 */
36 b secondary_startup
37
38 .align
391: .long .
40 .long pen_release
41ENDPROC(versatile_secondary_startup)
第1段高亮的部分实际上是等待pen_release成为CPU0设置的cpu_logical_map(cpu)，一般直接就成立了。第2段高亮的部分则调用到内核通用的secondary_startup()函数，经过一系列的初始化如MMU等，最终新的被唤醒的CPU将调用到smp_operations的smp_secondary_init()成员函数，对于本例为versatile_secondary_init()：
37void __cpuinit versatile_secondary_init(unsigned int cpu)
38{
39 /*
40 * if any interrupts are already enabled for the primary
41 * core (e.g. timer irq), then they will not have been enabled
42 * for us: do so
43 */
44 gic_secondary_init(0);
45
46 /*
47 * let the primary processor know we’re out of the
48 * pen, then head off into the C entry point
49 */
50 write_pen_release(-1);
51
52 /*
53 * Synchronise with the boot thread.
54 */
55 spin_lock( boot_lock);
56 spin_unlock( boot_lock);
57}
上述代码中高亮的那1行会将pen_release写为-1，于是CPU0还在执行的 versatile_boot_secondary()函数中的如下循环就退出了：
85 while (time_before(jiffies, timeout)) {
86 smp_rmb();
87 if (pen_release == -1)
88 break;
89
90 udelay(10);
91 }
此后CPU0和新唤醒的其他CPU各自狂奔。整个系统在运行过程中会进行实时进程和正常进程的动态负载均衡。
CPU hotplug的实现也是芯片相关的，对于VEXPRESS而言，实现了smp_operations的cpu_die()成员函数即vexpress_cpu_die()。它会在进行CPUn的拔除操作时将CPUn投入低功耗的WFI状态，相关代码位于arch/arm/mach-vexpress/hotplug.c：
90void __ref vexpress_cpu_die(unsigned int cpu)
91{
92 int spurious = 0;
93
94 /*
95 * we’re ready for shutdown now, so do it
96 */
97 cpu_enter_lowpower();
98 platform_do_lowpower(cpu, spurious);
99
100 /*
101 * bring this CPU back into the world of cache
102 * coherency, and then restore interrupts
103 */
104 cpu_leave_lowpower();
105
106 if (spurious)
107 pr_warn( CPU%u: %u spurious wakeup calls\n , cpu, spurious);
108}
57static inline void platform_do_lowpower(unsigned int cpu, int *spurious)
58{
59 /*
60 * there is no power-control hardware on this platform, so all
61 * we can do is put the core into WFI; this is safe as the calling
62 * code will have already disabled interrupts
63 */
64 for (;;) {
65 wfi();
66
67 if (pen_release == cpu_logical_map(cpu)) {
68 /*
69 * OK, proper wakeup, we’re done
70 */
71 break;
72 }
73
74 /*
75 * Getting here, means that we have come out of WFI without
76 * having been woken up – this shouldn’t happen
77 *
78 * Just note it happening – when we’re woken, we can report
79 * its occurrence.
80 */
81 (*spurious)++;
82 }
83}
CPUn睡眠于wfi()，之后再次online的时候，又会因为CPU0给它发出的IPI而从wfi()函数返回继续执行，醒来时CPUn也判决了是否pen_release == cpu_logical_map(cpu)成立，以确定该次醒来确确实实是由CPU0唤醒的一次正常醒来。
可偏偏。多么温柔，一出口便是相互指责和嘲讽。
