神秘的subsys_initcall【转】

转自：http://blog.chinaunix.net/uid-12567959-id-161015.html

在内核代码里到处都能看到这个subsys_initcall()，而它到底是干什么的呢？让我们来揭开它的神秘面纱。

先来看一段代码：

---------------------------------------------------------------------

include/linux/init.h

174 /*

175  * Early initcalls run before initializing SMP.

176  *

177  * Only for built-in code, not modules.

178  */

179 #define early_initcall(fn)    __define_initcall("early",fn,early)

180

181 /*

182  * A "pure" initcall has no dependencies on anything else, and purely

183  * initializes variables that couldn't be statically initialized.

184  *

185  * This only exists for built-in code, not for modules.

186  */

187 #define pure_initcall(fn)              __define_initcall("",fn,0)

188

189 #define core_initcall(fn)             __define_initcall("1",fn,1)

190 #define core_initcall_sync(fn)      __define_initcall("1s",fn,1s)

191 #define postcore_initcall(fn)         __define_initcall("2",fn,2)

192 #define postcore_initcall_sync(fn)  __define_initcall("2s",fn,2s)

193 #define arch_initcall(fn)             __define_initcall("3",fn,3)

194 #define arch_initcall_sync(fn)      __define_initcall("3s",fn,3s)

195 #define subsys_initcall(fn)           __define_initcall("4",fn,4)

196 #define subsys_initcall_sync(fn)    __define_initcall("4s",fn,4s)

197 #define fs_initcall(fn)               __define_initcall("5",fn,5)

198 #define fs_initcall_sync(fn)        __define_initcall("5s",fn,5s)

199 #define rootfs_initcall(fn) __define_initcall("rootfs",fn,rootfs)

200 #define device_initcall(fn)           __define_initcall("6",fn,6)

201 #define device_initcall_sync(fn)    __define_initcall("6s",fn,6s)

202 #define late_initcall(fn)             __define_initcall("7",fn,7)

203 #define late_initcall_sync(fn)      __define_initcall("7s",fn,7s)

---------------------------------------------------------------------

类似于subsys_initcall()还有很多，但它们都依赖于__define_initcall()，再来看__define_initcall()的定义：

---------------------------------------------------------------------

include/linux/init.h

131 typedef int (*initcall_t)(void);

165  *

166  * The `id' arg to __define_initcall() is needed so that multiple

167  * initcalls can point at the same handler without causing duplicate-symbol build errors.

168  */

169

170 #define __define_initcall(level,fn,id) \

171       static initcall_t __initcall_##fn##id __used \

172       __attribute__((__section__(".initcall" level ".init"))) = fn

173

---------------------------------------------------------------------

__define_initcall()宏只是定义一个initcall_t类型的静态变量，并且声明要把这个静态变量放在特定的段里而已。上面我们看到initcall_t即是指向一个无参数有int返回值的函数的指针。

许多的子系统都有自己的初始化函数，而这些初始化的函数又根据功能不同被分开在不同的子段里，子段的排列顺序则由链接决定。为了向后兼容，initcall()把调用，也就是一个个指向初始化函数的函数指针放进设备初始化子段里。

在各个平台的链接脚本文件arch/xxx/kernel/vmlinux.lds.S中，我们总能看到下面的语句：

INIT_CALLS

这个宏有如下的定义：

---------------------------------------------------------------------

include/asm-generic/vmlinux.lds.h

606 #define INIT_CALLS                                           \

607                 VMLINUX_SYMBOL(__initcall_start) = .;        \

608                 INITCALLS                                    \

609                 VMLINUX_SYMBOL(__initcall_end) = .;

---------------------------------------------------------------------

INIT_CALLS即是定义一个新的段，而定义段的字段的任务则由宏INITCALLS完成：

---------------------------------------------------------------------

include/asm-generic/vmlinux.lds.h

585 #define INITCALLS                                            \

586         *(.initcallearly.init)                               \

587         VMLINUX_SYMBOL(__early_initcall_end) = .;            \

588         *(.initcall0.init)                                   \

589         *(.initcall0s.init)                                  \

590         *(.initcall1.init)                                   \

591         *(.initcall1s.init)                                  \

592         *(.initcall2.init)                                   \

593         *(.initcall2s.init)                                  \

594         *(.initcall3.init)                                   \

595         *(.initcall3s.init)                                  \

596         *(.initcall4.init)                                   \

597         *(.initcall4s.init)                                  \

598         *(.initcall5.init)                                   \

599         *(.initcall5s.init)                                  \

600         *(.initcallrootfs.init)                              \

601         *(.initcall6.init)                                   \

602         *(.initcall6s.init)                                  \

603         *(.initcall7.init)                                   \

604         *(.initcall7s.init)

---------------------------------------------------------------------

而这些初始化函数又是在何时调用的呢？我们看到start_kernel()-> rest_init()-> kernel_init()-> do_basic_setup(void)-> do_initcalls()，而正是do_initcalls()处理了这些初始化函数，其定义为：

---------------------------------------------------------------------

init/main.c

765 extern initcall_t __initcall_start[], __initcall_end[], __early_initcall_end[];

766

767 static void __init do_initcalls(void)

768 {

769         initcall_t *fn;

770

771         for (fn = __early_initcall_end; fn < __initcall_end; fn++)

772                 do_one_initcall(*fn);

773

774         /* Make sure there is no pending stuff from the initcall sequence */

775         flush_scheduled_work();

776 }

---------------------------------------------------------------------

do_initcalls()又调用do_one_initcall()函数类处理这些调用。

---------------------------------------------------------------------

init/main.c

715 static char msgbuf[64];

716 static struct boot_trace_call call;

717 static struct boot_trace_ret ret;

718

719 int do_one_initcall(initcall_t fn)

720 {

721      int count = preempt_count();

722      ktime_t calltime, delta, rettime;

723

724      if (initcall_debug) {

725         call.caller = task_pid_nr(current);

726         printk("calling  %pF @ %i\n", fn, call.caller);

727         calltime = ktime_get();

728         trace_boot_call(&call, fn);

729         enable_boot_trace();

730     }

731

732     ret.result = fn();

733

734     if (initcall_debug) {

735        disable_boot_trace();

736        rettime = ktime_get();

737        delta = ktime_sub(rettime, calltime);

738        ret.duration = (unsigned long long) ktime_to_ns(delta) >> 10;

739        trace_boot_ret(&ret, fn);

740        printk("initcall %pF returned %d after %Ld usecs\n", fn,

741        ret.result, ret.duration);

742     }

743

744     msgbuf[0] = 0;

745

746     if (ret.result && ret.result != -ENODEV && initcall_debug)

747        sprintf(msgbuf, "error code %d ", ret.result);

748

749     if (preempt_count() != count) {

750        strlcat(msgbuf, "preemption imbalance ", sizeof(msgbuf));

751        preempt_count() = count;

752     }

753     if (irqs_disabled()) {

754        strlcat(msgbuf, "disabled interrupts ", sizeof(msgbuf));

755        local_irq_enable();

756     }

757     if (msgbuf[0]) {

758        printk("initcall %pF returned with %s\n", fn, msgbuf);

759     }

760

761     return ret.result;

762 }

---------------------------------------------------------------------