Linux syscall过程分析（万字长文）
文章有点长，慢慢看。
一、背景
为了安全，Linux 中分为用户态和内核态两种运行状态。对于普通进程，平时都是运行在用户态下，仅拥有基本的运行能力。当进行一些敏感操作，比如说要打开文件(open)然后进行写入(write)、分配内存(malloc)时，就会切换到内核态。内核态进行相应的检查，如果通过了，则按照进程的要求执行相应的操作，分配相应的资源。这种机制被称为系统调用，用户态进程发起调用，切换到内核态，内核态完成，返回用户态继续执行，是用户态唯一主动切换到内核态的合法手段(exception 和 interrupt 是被动切换)。
关于系统调用的详细定义可以通过 man syscalls 查看，它列出了目前 Linux Kernel 提供的系统调用 ABI 。我们熟悉的调用比如 open， read ，close 之类的都属于系统调用，但它们都经过了 C 库 (glibc)的封装。实际上，只要符合 ABI 规范，我们可以自己用汇编代码来进行调用。
历史上，x86 的系统调用实现经历了 int / iret 到 sysenter / sysexit 再到 syscall / sysret 的演变。
以下的分析基于 Linux kernel 4.9.76 ，glibc 为 2.25.90。
二、int / iret
很久很久以前，我们通过 int 0x80 进行系统调用(open)：
mov 0x05 ,eax /* 设置系统调用号 */
int 0x80
在 arch/x86/kernel/traps.c 的 trap_init 中，定义了各种 set_intr_gate / set_intr_gate_ist / set_system_intr_gate 。其中 set_system_intr_gate 用于在中断描述符表(IDT)上设置系统调用门：
#ifdef CONFIG_X86_32
set_system_intr_gate(IA32_SYSCALL_VECTOR, entry_INT80_32);
set_bit(IA32_SYSCALL_VECTOR, used_vectors);
#endif根据 arch/x86/include/asm/irq_vectors.h， IA32_SYSCALL_VECTOR 值为 0x80。
于是在调用 int 0x80 后，硬件根据向量号在 IDT 中找到对应的表项，即中断描述符，进行特权级检查，发现 DPL = CPL = 3 ，允许调用。然后硬件将切换到内核栈 (tss.ss0 : tss.esp0)。接着根据中断描述符的 segment selector 在 GDT / LDT 中找到对应的段描述符，从段描述符拿到段的基址，加载到 cs 。将 offset 加载到 eip。最后硬件将 ss / sp / eflags / cs / ip / error code 依次压到内核栈。
于是从 entry_INT80_32 开始执行，其定义在 arch/x86/entry/entry_32.S ：
ENTRY(entry_INT80_32)
ASM_CLAC
pushl %eax /* pt_regs->orig_ax */
SAVE_ALL pt_regs_ax=$-ENOSYS /* save rest */
/*
* User mode is traced as though IRQs are on, and the interrupt gate
* turned them off.
*/
TRACE_IRQS_OFF
movl %esp, %eax
call do_int80_syscall_32
...它将存在 eax 中的系统调用号压入栈中，然后调用 SAVE_ALL 将其他寄存器的值压入栈中进行保存：
.macro SAVE_ALL pt_regs_ax=%eax
cld
PUSH_GS
pushl %fs
pushl %es
pushl %ds
pushl \pt_regs_ax
pushl %ebp
pushl %edi
pushl %esi
pushl %edx
pushl %ecx
pushl %ebx
movl $(__USER_DS), %edx
movl %edx, %ds
movl %edx, %es
movl $(__KERNEL_PERCPU), %edx
movl %edx, %fs
SET_KERNEL_GS %edx
.endm保存完毕后，关闭中断，将当前栈指针保存到 eax ，调用 do_int80_syscall_32 => do_syscall_32_irqs_on ，该函数在 arch/x86/entry/common.c 中定义：
static __always_inline void do_syscall_32_irqs_on(struct pt_regs *regs)
{
struct thread_info *ti = current_thread_info();
unsigned int nr = (unsigned int)regs->orig_ax;
#ifdef CONFIG_IA32_EMULATION
current->thread.status |= TS_COMPAT;
#endif
if (READ_ONCE(ti->flags) & _TIF_WORK_SYSCALL_ENTRY) {
/*
* Subtlety here: if ptrace pokes something larger than
* 2^32-1 into orig_ax, this truncates it. This may or
* may not be necessary, but it matches the old asm
* behavior.
*/
nr = syscall_trace_enter(regs);
}
if (likely(nr < IA32_NR_syscalls)) {
/*
* It's possible that a 32-bit syscall implementation
* takes a 64-bit parameter but nonetheless assumes that
* the high bits are zero. Make sure we zero-extend all
* of the args.
*/
regs->ax = ia32_sys_call_table[nr](
(unsigned int)regs->bx, (unsigned int)regs->cx,
(unsigned int)regs->dx, (unsigned int)regs->si,
(unsigned int)regs->di, (unsigned int)regs->bp);
}
syscall_return_slowpath(regs);
}这个函数的参数 regs(struct pt_regs 定义见 arch/x86/include/asm/ptrace.h )就是先前在 entry_INT80_32 依次被压入栈的寄存器值。这里先取出系统调用号，从系统调用表(ia32_sys_call_table) 中取出对应的处理函数，然后通过先前寄存器中的参数调用之。
系统调用表 ia32_sys_call_table 在 arch/x86/entry/syscall_32.c 中定义，但内容有点奇怪，看上去表的内容是 include 进来的：
/* System call table for i386. */
#include <linux/linkage.h>
#include <linux/sys.h>
#include <linux/cache.h>
#include <asm/asm-offsets.h>
#include <asm/syscall.h>
#define __SYSCALL_I386(nr, sym, qual) extern asmlinkage long sym(unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long) ;
#include <asm/syscalls_32.h>
#undef __SYSCALL_I386
#define __SYSCALL_I386(nr, sym, qual) [nr] = sym,
extern asmlinkage long sys_ni_syscall(unsigned long, unsigned long, unsigned long, unsigned long, unsigned long, unsigned long);
__visible const sys_call_ptr_t ia32_sys_call_table[__NR_syscall_compat_max+1] = {
/*
* Smells like a compiler bug -- it doesn't work
* when the & below is removed.
*/
[0 ... __NR_syscall_compat_max] = &sys_ni_syscall,
#include <asm/syscalls_32.h>
};然而我们到源码的 arch/x86/include/asm 目录下却找不到 syscalls_32.h 的，但在编译 kernel 后的 arch/x86/include/generated/asm 里面发现了它：
__SYSCALL_I386(0, sys_restart_syscall, )
__SYSCALL_I386(1, sys_exit, )
#ifdef CONFIG_X86_32
__SYSCALL_I386(2, sys_fork, )
#else
__SYSCALL_I386(2, sys_fork, )
#endif
__SYSCALL_I386(3, sys_read, )
__SYSCALL_I386(4, sys_write, )
#ifdef CONFIG_X86_32
__SYSCALL_I386(5, sys_open, )
#else
__SYSCALL_I386(5, compat_sys_open, )
...这说明 syscalls_32.h 是在编译过程中动态生成的，请看脚本 arch/x86/entry/syscalls/syscalltbl.sh，它读取了同目录下的 syscall_32.tbl ，为每一有效行都生成了 __SYSCALL_${abi}($nr, $real_entry, $qualifier) 结构。然后在宏 __SYSCALL_I386 的作用下形成了这样的定义：
__visible const sys_call_ptr_t ia32_sys_call_table[__NR_syscall_compat_max+1] = {
[0 ... __NR_syscall_compat_max] = &sys_ni_syscall,
[0] = sys_restart_syscall,
[1] = sys_exit,
[2] = sys_fork,
[3] = sys_read,
[4] = sys_write,
[5] = sys_open,
...
};根据 GCC文档 ，这样的初始化方法在 ISO C99 中定义，个人称之为数组的乱序初始化。
因为我们的调用号是 0x05 ，所以这里调用了 sys_open ，定义在 fs/open.c 中定义：
SYSCALL_DEFINE3(open, const char __user *, filename, int, flags, umode_t, mode)
{
if (force_o_largefile())
flags |= O_LARGEFILE;
return do_sys_open(AT_FDCWD, filename, flags, mode);
}宏 SYSCALL_DEFINE3 及相关定义如下：
#define SYSCALL_DEFINE3(name, ...) SYSCALL_DEFINEx(3, _##name, __VA_ARGS__)
#define SYSCALL_DEFINEx(x, sname, ...) \
SYSCALL_METADATA(sname, x, __VA_ARGS__) \
__SYSCALL_DEFINEx(x, sname, __VA_ARGS__)
#define __SYSCALL_DEFINEx(x, name, ...) \
asmlinkage long sys##name(__MAP(x,__SC_DECL,__VA_ARGS__)) \
__attribute__((alias(__stringify(SyS##name)))); \
\
static inline long SYSC##name(__MAP(x,__SC_DECL,__VA_ARGS__)); \
\
asmlinkage long SyS##name(__MAP(x,__SC_LONG,__VA_ARGS__)); \
\
asmlinkage long SyS##name(__MAP(x,__SC_LONG,__VA_ARGS__)) \
{ \
long ret = SYSC##name(__MAP(x,__SC_CAST,__VA_ARGS__)); \
__MAP(x,__SC_TEST,__VA_ARGS__); \
__PROTECT(x, ret,__MAP(x,__SC_ARGS,__VA_ARGS__)); \
return ret; \
} \
\
static inline long SYSC##name(__MAP(x,__SC_DECL,__VA_ARGS__))SYSCALL_METADATA 保存了调用的基本信息，供调试程序跟踪使用( kernel 需开启 CONFIG_FTRACE_SYSCALLS )。
而 __SYSCALL_DEFINEx 用于拼接函数，函数名被拼接为 sys##_##open，参数也通过 __SC_DECL 拼接，最终得到展开后的定义：
asmlinkage long sys_open(const char __user * filename, int flags, umode_t mode)
{
if (force_o_largefile())
flags |= O_LARGEFILE;
return do_sys_open(AT_FDCWD, filename, flags, mode);
}sys_open 是对 do_sys_open 的封装：
long do_sys_open(int dfd, const char __user *filename, int flags, umode_t mode)
{
struct open_flags op;
int fd = build_open_flags(flags, mode, &op);
struct filename *tmp;
if (fd)
return fd;
tmp = getname(filename);
if (IS_ERR(tmp))
return PTR_ERR(tmp);
fd = get_unused_fd_flags(flags);
if (fd >= 0) {
struct file *f = do_filp_open(dfd, tmp, &op);
if (IS_ERR(f)) {
put_unused_fd(fd);
fd = PTR_ERR(f);
} else {
fsnotify_open(f);
fd_install(fd, f);
}
}
putname(tmp);
return fd;
}getname 将处于用户态的文件名拷到内核态，然后通过 get_unused_fd_flags 获取一个没用过的文件描述符，然后 do_filp_open 创建 struct file ， fd_install 将 fd 和 struct file 绑定(task_struct->files->fdt[fd] = file)，然后返回 fd。
fd一直返回到 do_syscall_32_irqs_on ，被设置到 regs->ax (eax) 中。接着返回 entry_INT80_32 继续执行，最后执行 INTERRUPT_RETURN 。INTERRUPT_RETURN 在 arch/x86/include/asm/irqflags.h 中定义为 iret ，负责恢复先前压栈的寄存器，返回用户态。系统调用执行完毕。
在目前主流的系统调用库(glibc) 中，int 0x80 只有在硬件不支持快速系统调用(sysenter / syscall)的时候才会调用，但目前的硬件都支持快速系统调用，所以为了能够看看 int 0x80 的效果，我们手撸汇编：
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main(){
char * filename = "/tmp/test";
char * buffer = malloc(80);
memset(buffer, 0, 80);
int count;
__asm__ __volatile__("movl $0x5, %%eax\n\t"
"movl %1, %%ebx\n\t"
"movl $0, %%ecx\n\t"
"movl $0664, %%edx\n\t"
"int $0x80\n\t"
"movl %%eax, %%ebx\n\t"
"movl $0x3, %%eax\n\t"
"movl %2, %%ecx\n\t"
"movl $80, %%edx\n\t"
"int $0x80\n\t"
"movl %%eax, %0\n\t"
:"=m"(count)
:"g"(filename), "g"(buffer)
:"%eax", "%ebx", "%ecx", "%edx");
printf("%d\n", count);
printf("%s\n", buffer);
free(buffer);
}这段代码首先通过 int 0x80 调用系统调用 open 得到 fd (由 eax 返回)，再作为 read 的参数传入，从而读出了文件中的内容。但比较奇怪的是如果 buffer 存储在栈中 (buffer[80])，则调用 read 失败。只有将 buffer 作为全局变量或存储在堆中，才能调用成功。希望有知道的大大指点一下。
三、sysenter / sysexit
接下来介绍的是 32位下 Intel 提出的快速系统调用 sysenter/sysexit，它和同期AMD的 syscall/sysret 机制类似。
之所以提出新指令，是因为通过软中断来实现系统调用实在太慢了。于是 Intel x86 CPU 自 Pentium II（Family 6, Model 3, Stepping 3）之后，开始支持新的系统调用指令 sysenter/sysexit。前者用于从低特权级切换到 ring 0，后者用于 从ring 0 切换到低特权级。没有特权级别检查(CPL, DPL)，也没有压栈的操作，快最重要！
在 Intel SDM 中阐述了sysenter指令。首先 CPU 有一堆特殊的寄存器，名为 Model-Specific Register(MSR)，这些寄存器在操作系统运行过程中起着重要作用。对于这些寄存器，需要采用专门的指令 RDMSR 和 WRMSR 进行读写。
sysenter 用到了以下 MSR (定义在 arch/x86/include/asm/msr-index.h)：
- IA32_SYSENTER_CS(174H)：存放内核态处理代码的段选择符
- IA32_SYSENTER_EIP(175H)：存放内核态栈顶偏移量
- IA32_SYSENTER_ESP(176H)：存放内核态处理代码偏移量
当执行 sysenter 时，执行以下操作：
- 清除 FLAGS 的 VM 标志，确保在保护模式下运行
- 清除 FLAGS 的 IF 标志，屏蔽中断
- 加载 IA32_SYSENTER_ESP 的值到 esp
- 加载 IA32_SYSENTER_EIP 的值到 eip
- 加载 SYSENTER_CS_MSR 的值到 CS
- 将 SYSENTER_CS_MSR + 8 的值加载到 ss 。因为在GDT中， ss 就跟在 cs 后面
- 开始执行(cs:eip)指向的代码
这些 MSR 在 arch/x86/kernel/cpu/common.c 的 enable_sep_cpu 中初始化：
void enable_sep_cpu(void)
{
struct tss_struct *tss;
int cpu;
if (!boot_cpu_has(X86_FEATURE_SEP))
return;
cpu = get_cpu();
tss = &per_cpu(cpu_tss, cpu);
/*
* We cache MSR_IA32_SYSENTER_CS's value in the TSS's ss1 field --
* see the big comment in struct x86_hw_tss's definition.
*/
tss->x86_tss.ss1 = __KERNEL_CS;
wrmsr(MSR_IA32_SYSENTER_CS, tss->x86_tss.ss1, 0);
wrmsr(MSR_IA32_SYSENTER_ESP,
(unsigned long)tss + offsetofend(struct tss_struct, SYSENTER_stack),
0);
wrmsr(MSR_IA32_SYSENTER_EIP, (unsigned long)entry_SYSENTER_32, 0);
put_cpu();
}这里将 __KERNEL_CS 设置到 MSR_IA32_SYSENTER_CS 中，将 tss.SYSENTER_stack 地址设置到 MSR_IA32_SYSENTER_ESP 中，最后将内核入口点 entry_SYSENTER_32 的地址设置到 MSR_IA32_SYSENTER_EIP 中。
当用户程序进行系统调用时，实际上在用户态中最终会调用到 VDSO 中映射的 __kernel_vsyscall ，其定义位于 arch/x86/entry/vdso/vdso32/system_call.S：
__kernel_vsyscall:
CFI_STARTPROC
pushl %ecx
CFI_ADJUST_CFA_OFFSET 4
CFI_REL_OFFSET ecx, 0
pushl %edx
CFI_ADJUST_CFA_OFFSET 4
CFI_REL_OFFSET edx, 0
pushl %ebp
CFI_ADJUST_CFA_OFFSET 4
CFI_REL_OFFSET ebp, 0
#define SYSENTER_SEQUENCE "movl %esp, %ebp; sysenter"
#define SYSCALL_SEQUENCE "movl %ecx, %ebp; syscall"
#ifdef CONFIG_X86_64
/* If SYSENTER (Intel) or SYSCALL32 (AMD) is available, use it. */
ALTERNATIVE_2 "", SYSENTER_SEQUENCE, X86_FEATURE_SYSENTER32, \
SYSCALL_SEQUENCE, X86_FEATURE_SYSCALL32
#else
ALTERNATIVE "", SYSENTER_SEQUENCE, X86_FEATURE_SEP
#endif
/* Enter using int $0x80 */
int $0x80
GLOBAL(int80_landing_pad)
/*
* Restore EDX and ECX in case they were clobbered. EBP is not
* clobbered (the kernel restores it), but it's cleaner and
* probably faster to pop it than to adjust ESP using addl.
*/
popl %ebp
CFI_RESTORE ebp
CFI_ADJUST_CFA_OFFSET -4
popl %edx
CFI_RESTORE edx
CFI_ADJUST_CFA_OFFSET -4
popl %ecx
CFI_RESTORE ecx
CFI_ADJUST_CFA_OFFSET -4
ret
CFI_ENDPROC
.size __kernel_vsyscall,.-__kernel_vsyscall
.previous__kernel_vsyscall 首先将寄存器当前值压栈保存，因为这些寄存器以后要用作系统调用传参。然后填入参数，调用 sysenter
ALTERNATIVE_2 宏实际上是在做选择，如果支持 X86_FEATURE_SYSENTER32(Intel CPU) ，则执行 SYSENTER_SEQUENCE ，如果支持 X86_FEATURE_SYSCALL32(AMD CPU)，则执行 SYSCALL_SEQUENCE 。如果都不支持，那么啥都不干(???)。如果啥都没干，那么接着往下执行，即执行 int $0x80，退化到传统(legacy)方式进行系统调用。
注意 sysenter 指令会覆盖掉 esp ，因此 SYSENTER_SEQUENCE 中会将当前 esp 保存到 ebp 中。sysenter 同样会覆盖 eip ，但由于返回地址是固定的(__kernel_vsyscall 函数结尾)，因此无需保存。
前文提到过，执行了 sysenter 指令之后直接切换到内核态，同时寄存器也都设置好了：eip 被设置为 IA32_SYSENTER_EIP 即 entry_SYSENTER_32 的地址，其定义在arch/x86/entry/entry_32.S中：
ENTRY(entry_SYSENTER_32)
movl TSS_sysenter_sp0(%esp), %esp
sysenter_past_esp:
pushl $__USER_DS /* pt_regs->ss */
pushl %ebp /* pt_regs->sp (stashed in bp) */
pushfl /* pt_regs->flags (except IF = 0) */
orl $X86_EFLAGS_IF, (%esp) /* Fix IF */
pushl $__USER_CS /* pt_regs->cs */
pushl $0 /* pt_regs->ip = 0 (placeholder) */
pushl %eax /* pt_regs->orig_ax */
SAVE_ALL pt_regs_ax=$-ENOSYS /* save rest */
testl $X86_EFLAGS_NT|X86_EFLAGS_AC|X86_EFLAGS_TF, PT_EFLAGS(%esp)
jnz .Lsysenter_fix_flags
.Lsysenter_flags_fixed:
/*
* User mode is traced as though IRQs are on, and SYSENTER
* turned them off.
*/
TRACE_IRQS_OFF
movl %esp, %eax
call do_fast_syscall_32
...
/* arch/x86/kernel/asm-offsets_32.c */
/* Offset from the sysenter stack to tss.sp0 */
DEFINE(TSS_sysenter_sp0, offsetof(struct cpu_entry_area, tss.x86_tss.sp0) -
offsetofend(struct cpu_entry_area, entry_stack_page.stack));前文提到过，sysenter 会将 IA32_SYSENTER_ESP 加载到 esp 中，但 IA32_SYSENTER_ESP 保存的是 SYSENTER_stack 的地址，需要通过 TSS_sysenter_sp0 进行修正，指向进程的内核栈。
然后开始按照 pt_regs 的结构将相关寄存器中的值压入栈中，包括在 sysenter 前保存到 ebp 的用户态栈顶指针。由于 eip 无需保存，于是压入 0 用于占位。
最后调用 do_fast_syscall_32 ，该函数在 arch/x86/entry/common.c 中定义：
/* Returns 0 to return using IRET or 1 to return using SYSEXIT/SYSRETL. */
__visible long do_fast_syscall_32(struct pt_regs *regs)
{
/*
* Called using the internal vDSO SYSENTER/SYSCALL32 calling
* convention. Adjust regs so it looks like we entered using int80.
*/
unsigned long landing_pad = (unsigned long)current->mm->context.vdso +
vdso_image_32.sym_int80_landing_pad;
/*
* SYSENTER loses EIP, and even SYSCALL32 needs us to skip forward
* so that 'regs->ip -= 2' lands back on an int $0x80 instruction.
* Fix it up.
*/
regs->ip = landing_pad;
enter_from_user_mode();
local_irq_enable();
/* Fetch EBP from where the vDSO stashed it. */
if (
#ifdef CONFIG_X86_64
/*
* Micro-optimization: the pointer we're following is explicitly
* 32 bits, so it can't be out of range.
*/
__get_user(*(u32 *)®s->bp,
(u32 __user __force *)(unsigned long)(u32)regs->sp)
#else
get_user(*(u32 *)®s->bp,
(u32 __user __force *)(unsigned long)(u32)regs->sp)
#endif
) {
/* User code screwed up. */
local_irq_disable();
regs->ax = -EFAULT;
prepare_exit_to_usermode(regs);
return 0; /* Keep it simple: use IRET. */
}
/* Now this is just like a normal syscall. */
do_syscall_32_irqs_on(regs);
#ifdef CONFIG_X86_64
/*
* Opportunistic SYSRETL: if possible, try to return using SYSRETL.
* SYSRETL is available on all 64-bit CPUs, so we don't need to
* bother with SYSEXIT.
*
* Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP,
* because the ECX fixup above will ensure that this is essentially
* never the case.
*/
return regs->cs == __USER32_CS && regs->ss == __USER_DS &&
regs->ip == landing_pad &&
(regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF)) == 0;
#else
/*
* Opportunistic SYSEXIT: if possible, try to return using SYSEXIT.
*
* Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP,
* because the ECX fixup above will ensure that this is essentially
* never the case.
*
* We don't allow syscalls at all from VM86 mode, but we still
* need to check VM, because we might be returning from sys_vm86.
*/
return static_cpu_has(X86_FEATURE_SEP) &&
regs->cs == __USER_CS && regs->ss == __USER_DS &&
regs->ip == landing_pad &&
(regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF | X86_EFLAGS_VM)) == 0;
#endif
}由于没有保存 eip，我们需要计算系统调用完毕后返回到用户态的地址：current->mm->context.vdso + vdso_image_32.sym_int80_landing_pad (即跳过 sym_int80_landing_pad 来到 __kernel_vsyscall 的结尾) 覆盖掉先前压栈的 0 。
接下来就和 int 0x80 的流程一样，通过 do_syscall_32_irqs_on 从系统调用表中找到相应的处理函数进行调用。完成后，如果都符合 sysexit 的要求，返回 1，否则返回 0 。
...
call do_fast_syscall_32
/* XEN PV guests always use IRET path */
ALTERNATIVE "testl %eax, %eax; jz .Lsyscall_32_done", \
"jmp .Lsyscall_32_done", X86_FEATURE_XENPV
/* Opportunistic SYSEXIT */
TRACE_IRQS_ON /* User mode traces as IRQs on. */
movl PT_EIP(%esp), %edx /* pt_regs->ip */
movl PT_OLDESP(%esp), %ecx /* pt_regs->sp */
1: mov PT_FS(%esp), %fs
PTGS_TO_GS
popl %ebx /* pt_regs->bx */
addl $2*4, %esp /* skip pt_regs->cx and pt_regs->dx */
popl %esi /* pt_regs->si */
popl %edi /* pt_regs->di */
popl %ebp /* pt_regs->bp */
popl %eax /* pt_regs->ax */
/*
* Restore all flags except IF. (We restore IF separately because
* STI gives a one-instruction window in which we won't be interrupted,
* whereas POPF does not.)
*/
addl $PT_EFLAGS-PT_DS, %esp /* point esp at pt_regs->flags */
btr $X86_EFLAGS_IF_BIT, (%esp)
popfl
/*
* Return back to the vDSO, which will pop ecx and edx.
* Don't bother with DS and ES (they already contain __USER_DS).
*/
sti
sysexit根据 testl %eax, %eax; jz .Lsyscall_32_done ，如果 do_fast_syscall_32 的返回值(eax)为 0 ，表示不支持快速返回，于是跳转到 Lsyscall_32_done ，通过 iret 返回。否则继续执行下面代码，将内核栈中保存的值保存到相应寄存器中，然后通过 sysexit 返回。
注意这里将原有的 eip 设置到 edx、 esp 设置到 ecx ，这是因为根据 Intel SDM，sysexit 会用 edx 来设置 eip，用 ecx 来设置 esp ，从而指向先前用户空间的代码偏移和栈偏移。并加载 SYSENTER_CS_MSR+16 到 cs，加载 SYSENTER_CS_MSR+24 到 ss 。如此一来就回到了用户态的 __kernel_vsyscall 尾端。
四、实验
我们通过 gdb 一个 C 程序来检验一下：
#include <unistd.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
int main(int argc, char *argv[]){
char buffer[80] = "/tmp/test";
int fd = open(buffer, O_RDONLY);
int size = read(fd, buffer, sizeof(buffer));
close(fd);
}
$ gcc -m32 -g -static -o read read.c
$ file read
read: ELF 32-bit LSB executable, Intel 80386, version 1 (GNU/Linux), statically linked, for GNU/Linux 2.6.32, BuildID[sha1]=8a7f3d69d3e4c9582551934b0617ad78e492e48c, not stripped
[txt]
(gdb) disas
0x0804888a <+14>: push %ecx
0x0804888b <+15>: sub $0x70,%esp
0x0804888e <+18>: mov %ecx,%eax
0x08048890 <+20>: mov 0x4(%eax),%eax
0x08048893 <+23>: mov %eax,-0x6c(%ebp)
0x08048896 <+26>: mov %gs:0x14,%eax
0x0804889c <+32>: mov %eax,-0xc(%ebp)
0x0804889f <+35>: xor %eax,%eax
0x080488a1 <+37>: movl $0x706d742f,-0x5c(%ebp)
0x080488a8 <+44>: movl $0x7365742f,-0x58(%ebp)
0x080488af <+51>: movl $0x74,-0x54(%ebp)
0x080488b6 <+58>: lea -0x50(%ebp),%edx
0x080488b9 <+61>: mov $0x0,%eax
0x080488be <+66>: mov $0x11,%ecx
0x080488c3 <+71>: mov %edx,%edi
0x080488c5 <+73>: rep stos %eax,%es:(%edi)
0x080488c7 <+75>: sub $0x8,%esp
0x080488ca <+78>: push $0x0
0x080488cc <+80>: lea -0x5c(%ebp),%eax
0x080488cf <+83>: push %eax
0x080488d0 <+84>: call 0x806cf30 <open>
0x080488d5 <+89>: add $0x10,%esp
0x080488d8 <+92>: mov %eax,-0x64(%ebp)
0x080488db <+95>: sub $0x4,%esp
0x080488de <+98>: push $0x50
0x080488e0 <+100>: lea -0x5c(%ebp),%eax
0x080488e3 <+103>: push %eax
0x080488e4 <+104>: pushl -0x64(%ebp)
0x080488e7 <+107>: call 0x806cfa0 <read>
0x080488ec <+112>: add $0x10,%esp
0x080488ef <+115>: mov %eax,-0x60(%ebp)
=> 0x080488f2 <+118>: sub $0xc,%esp
0x080488f5 <+121>: pushl -0x64(%ebp)
0x080488f8 <+124>: call 0x806d150 <close>
0x080488fd <+129>: add $0x10,%esp
0x08048900 <+132>: mov $0x0,%eax
0x08048905 <+137>: mov -0xc(%ebp),%edx
0x08048908 <+140>: xor %gs:0x14,%edx
0x0804890f <+147>: je 0x8048916 <main+154>
0x08048911 <+149>: call 0x806ef90 <__stack_chk_fail>
0x08048916 <+154>: lea -0x8(%ebp),%esp
0x08048919 <+157>: pop %ecx
0x0804891a <+158>: pop %edi
0x0804891b <+159>: pop %ebp
0x0804891c <+160>: lea -0x4(%ecx),%esp
0x0804891f <+163>: ret
End of assembler dump.首先是 open ，将将参数 O_RDONLY (根据 #define O_RDONLY 0，值为 0x0 )，将 buffer 地址(eax) 压栈后调用系统调用 glibc 的 open 函数，disas 之：
(gdb) disas 0x806cf30
Dump of assembler code for function open:
0x0806cf30 <+0>: cmpl $0x0,%gs:0xc
0x0806cf38 <+8>: jne 0x806cf5f
0x0806cf3a <+0>: push %ebx
0x0806cf3b <+1>: mov 0x10(%esp),%edx
0x0806cf3f <+5>: mov 0xc(%esp),%ecx
0x0806cf43 <+9>: mov 0x8(%esp),%ebx
0x0806cf47 <+13>: mov $0x5,%eax
0x0806cf4c <+18>: call *0x80ea9f0
0x0806cf52 <+24>: pop %ebx
0x0806cf53 <+25>: cmp $0xfffff001,%eax
0x0806cf58 <+30>: jae 0x8070590 <__syscall_error>
0x0806cf5e <+36>: ret
0x0806cf5f <+47>: call 0x806ea80 <__libc_enable_asynccancel>
0x0806cf64 <+52>: push %eax
0x0806cf65 <+53>: push %ebx
0x0806cf66 <+54>: mov 0x14(%esp),%edx
0x0806cf6a <+58>: mov 0x10(%esp),%ecx
0x0806cf6e <+62>: mov 0xc(%esp),%ebx
0x0806cf72 <+66>: mov $0x5,%eax
0x0806cf77 <+71>: call *0x80ea9f0
0x0806cf7d <+77>: pop %ebx
0x0806cf7e <+78>: xchg %eax,(%esp)
0x0806cf81 <+81>: call 0x806eaf0 <__libc_disable_asynccancel>
0x0806cf86 <+86>: pop %eax
0x0806cf87 <+87>: cmp $0xfffff001,%eax
0x0806cf8c <+92>: jae 0x8070590 <__syscall_error>
0x0806cf92 <+98>: ret
End of assembler dump.将压入栈中的参数保存到寄存器中，然后调用了 0x80ea9f0，用 x 查看该地址的值：
(gdb) x 0x80ea9f0
0x80ea9f0 <_dl_sysinfo>: 0xf7ffcc80disas 之，发现来到了 __kernel_vsyscall ，并执行了sysenter指令：
(gdb) disas 0xf7ffcc80
Dump of assembler code for function __kernel_vsyscall:
0xf7ffcc80 <+0>: push %ecx
0xf7ffcc81 <+1>: push %edx
0xf7ffcc82 <+2>: push %ebp
0xf7ffcc83 <+3>: mov %esp,%ebp
0xf7ffcc85 <+5>: sysenter
0xf7ffcc87 <+7>: int $0x80
0xf7ffcc89 <+9>: pop %ebp
0xf7ffcc8a <+10>: pop %edx
0xf7ffcc8b <+11>: pop %ecx
0xf7ffcc8c <+12>: ret
End of assembler dump.read 同理，只是有三个参数，需要 push 三次而已。
五、syscall / sysret
前文提到过，在32位下 Intel 和 AMD 对快速系统调用指令的定义有分歧，一个使用 sysenter ，另一个使用 syscall 。但到了64位下，为啥都统一成 syscall 了呢？
关于这个我在网上也没有找到权威的答案，只是一些道途听说：为什么IA-64指令集架构失败了？
在 64 位架构的开发上，Intel 和 AMD 选择了不同的道路：Intel搞出了一套全新的架构，名为安腾(IA-64)，这套架构性能完爆x86，这样用户为了更好的性能需要进行硬件换代，岂不是喜滋滋？然而这种做法在商业上取得了失败。因为 IA-64 架构虽然提高了性能，却不能向后兼容，即原来能在 x86 下跑的程序到新架构下就跑不了了，用户非常 angry 。AMD 就比较厚道，老老实实地做出了兼容 x86 的 x86_64 ，能够运行 32 位下的程序。于是农企日常翻身，逼得 Intel 反过来兼容 x86_64 架构，于是只能支持 AMD 标准中定义的 syscall 了。
这次我们直接从gdb出发，同样是之前的代码，只是这次编译成 64 位：
(gdb) disas
Dump of assembler code for function main:
0x00000000004009ae <+0>: push %rbp
0x00000000004009af <+1>: mov %rsp,%rbp
0x00000000004009b2 <+4>: add $0xffffffffffffff80,%rsp
0x00000000004009b6 <+8>: mov %edi,-0x74(%rbp)
0x00000000004009b9 <+11>: mov %rsi,-0x80(%rbp)
0x00000000004009bd <+15>: mov %fs:0x28,%rax
0x00000000004009c6 <+24>: mov %rax,-0x8(%rbp)
0x00000000004009ca <+28>: xor %eax,%eax
0x00000000004009cc <+30>: movabs $0x7365742f706d742f,%rax
0x00000000004009d6 <+40>: mov %rax,-0x60(%rbp)
0x00000000004009da <+44>: movq $0x74,-0x58(%rbp)
0x00000000004009e2 <+52>: lea -0x50(%rbp),%rdx
0x00000000004009e6 <+56>: mov $0x0,%eax
0x00000000004009eb <+61>: mov $0x8,%ecx
0x00000000004009f0 <+66>: mov %rdx,%rdi
0x00000000004009f3 <+69>: rep stos %rax,%es:(%rdi)
0x00000000004009f6 <+72>: lea -0x60(%rbp),%rax
0x00000000004009fa <+76>: mov $0x0,%esi
0x00000000004009ff <+81>: mov %rax,%rdi
0x0000000000400a02 <+84>: mov $0x0,%eax
0x0000000000400a07 <+89>: callq 0x43e650 <open64>
0x0000000000400a0c <+94>: mov %eax,-0x68(%rbp)
0x0000000000400a0f <+97>: lea -0x60(%rbp),%rcx
0x0000000000400a13 <+101>: mov -0x68(%rbp),%eax
0x0000000000400a16 <+104>: mov $0x50,%edx
0x0000000000400a1b <+109>: mov %rcx,%rsi
0x0000000000400a1e <+112>: mov %eax,%edi
0x0000000000400a20 <+114>: callq 0x43e6b0 <read>
0x0000000000400a25 <+119>: mov %eax,-0x64(%rbp)
=> 0x0000000000400a28 <+122>: mov -0x68(%rbp),%eax
0x0000000000400a2b <+125>: mov %eax,%edi
0x0000000000400a2d <+127>: callq 0x43e900 <close>
0x0000000000400a32 <+132>: mov $0x0,%eax
0x0000000000400a37 <+137>: mov -0x8(%rbp),%rdx
0x0000000000400a3b <+141>: xor %fs:0x28,%rdx
0x0000000000400a44 <+150>: je 0x400a4b <main+157>
0x0000000000400a46 <+152>: callq 0x442010 <__stack_chk_fail>
0x0000000000400a4b <+157>: leaveq
0x0000000000400a4c <+158>: retq
End of assembler dump.
(gdb) disas 0x43e650
Dump of assembler code for function open64:
0x000000000043e650 <+0>: cmpl $0x0,0x28db65(%rip) # 0x6cc1bc <__libc_multiple_threads>
0x000000000043e657 <+7>: jne 0x43e66d <open64+29>
0x000000000043e659 <+0>: mov $0x2,%eax
0x000000000043e65e <+5>: syscall
0x000000000043e660 <+7>: cmp $0xfffffffffffff001,%rax
0x000000000043e666 <+13>: jae 0x4436b0 <__syscall_error>
0x000000000043e66c <+19>: retq
0x000000000043e66d <+29>: sub $0x8,%rsp
0x000000000043e671 <+33>: callq 0x441b70 <__libc_enable_asynccancel>
0x000000000043e676 <+38>: mov %rax,(%rsp)
0x000000000043e67a <+42>: mov $0x2,%eax
0x000000000043e67f <+47>: syscall
0x000000000043e681 <+49>: mov (%rsp),%rdi
0x000000000043e685 <+53>: mov %rax,%rdx
0x000000000043e688 <+56>: callq 0x441bd0 <__libc_disable_asynccancel>
0x000000000043e68d <+61>: mov %rdx,%rax
0x000000000043e690 <+64>: add $0x8,%rsp
0x000000000043e694 <+68>: cmp $0xfffffffffffff001,%rax
0x000000000043e69a <+74>: jae 0x4436b0 <__syscall_error>
0x000000000043e6a0 <+80>: retq
End of assembler dump.open64 定义在 glibc 的 sysdeps/posix/open64.c中：
#include <fcntl.h>
#include <stdarg.h>
#include <sysdep-cancel.h>
/* Open FILE with access OFLAG. If O_CREAT or O_TMPFILE is in OFLAG,
a third argument is the file protection. */
int
__libc_open64 (const char *file, int oflag, ...)
{
int mode = 0;
if (__OPEN_NEEDS_MODE (oflag))
{
va_list arg;
va_start (arg, oflag);
mode = va_arg (arg, int);
va_end (arg);
}
if (SINGLE_THREAD_P)
return __libc_open (file, oflag | O_LARGEFILE, mode);
int oldtype = LIBC_CANCEL_ASYNC ();
int result = __libc_open (file, oflag | O_LARGEFILE, mode);
LIBC_CANCEL_RESET (oldtype);
return result;
}
weak_alias (__libc_open64, __open64)
libc_hidden_weak (__open64)
weak_alias (__libc_open64, open64)
再看 __libc_open ，定义在 unix/sysv/linux/generic/open.c ：
#include <errno.h>
#include <fcntl.h>
#include <stdarg.h>
#include <stdio.h>
#include <sysdep-cancel.h>
/* Open FILE with access OFLAG. If O_CREAT or O_TMPFILE is in OFLAG,
a third argument is the file protection. */
int
__libc_open (const char *file, int oflag, ...)
{
int mode = 0;
if (__OPEN_NEEDS_MODE (oflag))
{
va_list arg;
va_start (arg, oflag);
mode = va_arg (arg, int);
va_end (arg);
}
return SYSCALL_CANCEL (openat, AT_FDCWD, file, oflag, mode);
}我们将宏展开：
SYSCALL_CANCEL(openat, AT_FDCWD, file, oflag, mode)
=> __SYSCALL_CALL(openat, AT_FDCWD, file, oflag, mode)
=> __SYSCALL_DISP(__SYSCALL, openat, AT_FDCWD, file, oflag, mode)
=> __SYSCALL_CONCAT(__SYSCALL, 4)(openat, AT_FDCWD, file, oflag, mode)
=> __SYSCALL_CONCAT_X(__SYSCALL, 4)(openat, AT_FDCWD, file, oflag, mode)
=> __SYSCALL5(openat, AT_FDCWD, file, oflag, mode)
=> INLINE_SYSCALL (openat, 4, AT_FDCWD, file, oflag, mode)
=> INTERNAL_SYSCALL (openat, _, 4, AT_FDCWD, file, oflag, mode)
=> INTERNAL_SYSCALL_NCS (__NR_openat, _, 4, AT_FDCWD, file, oflag, mode)最终到达 INTERNAL_SYSCALL_NCS ：
# define INTERNAL_SYSCALL_NCS(name, err, nr, args...) \
({ \
unsigned long int resultvar; \
LOAD_ARGS_##nr (args) \
LOAD_REGS_##nr \
asm volatile ( \
"syscall\n\t" \
: "=a" (resultvar) \
: "0" (name) ASM_ARGS_##nr : "memory", REGISTERS_CLOBBERED_BY_SYSCALL); \
(long int) resultvar; })LOAD_ARGS_##nr 负责把参数 args 展开，然后由 LOAD_REGS_##nr 设置到相应的寄存器中，因为 syscall 通过寄存器传参。最终调用 syscall 。
根据 Intel SDM，syscall 会将当前 rip 存到 rcx ，然后将 IA32_LSTAR 加载到 rip 。同时将 IA32_STAR[47:32] 加载到cs，IA32_STAR[47:32] + 8 加载到 ss (在 GDT 中，ss 就跟在 cs 后面)。
MSR IA32_LSTAR (MSR_LSTAR) 和 IA32_STAR (MSR_STAR) 在 arch/x86/kernel/cpu/common.c 的 syscall_init 中初始化：
void syscall_init(void)
{
wrmsr(MSR_STAR, 0, (__USER32_CS << 16) | __KERNEL_CS);
wrmsrl(MSR_LSTAR, (unsigned long)entry_SYSCALL_64);
#ifdef CONFIG_IA32_EMULATION
wrmsrl(MSR_CSTAR, (unsigned long)entry_SYSCALL_compat);
/*
* This only works on Intel CPUs.
* On AMD CPUs these MSRs are 32-bit, CPU truncates MSR_IA32_SYSENTER_EIP.
* This does not cause SYSENTER to jump to the wrong location, because
* AMD doesn't allow SYSENTER in long mode (either 32- or 64-bit).
*/
wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)__KERNEL_CS);
wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
wrmsrl_safe(MSR_IA32_SYSENTER_EIP, (u64)entry_SYSENTER_compat);
#else
wrmsrl(MSR_CSTAR, (unsigned long)ignore_sysret);
wrmsrl_safe(MSR_IA32_SYSENTER_CS, (u64)GDT_ENTRY_INVALID_SEG);
wrmsrl_safe(MSR_IA32_SYSENTER_ESP, 0ULL);
wrmsrl_safe(MSR_IA32_SYSENTER_EIP, 0ULL);
#endif
/* Flags to clear on syscall */
wrmsrl(MSR_SYSCALL_MASK,
X86_EFLAGS_TF|X86_EFLAGS_DF|X86_EFLAGS_IF|
X86_EFLAGS_IOPL|X86_EFLAGS_AC|X86_EFLAGS_NT);
}可以看到 MSR_STAR 的第 32-47 位设置为 kernel mode 的 cs，48-63位设置为 user mode 的 cs。而 IA32_LSTAR 被设置为函数 entry_SYSCALL_64 的起始地址。
于是 syscall 时，跳转到 entry_SYSCALL_64 开始执行，其定义在 arch/x86/entry/entry_64.S：
ENTRY(entry_SYSCALL_64)
/*
* Interrupts are off on entry.
* We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
* it is too small to ever cause noticeable irq latency.
*/
SWAPGS_UNSAFE_STACK
// KAISER 进内核态需要切到内核页表
SWITCH_KERNEL_CR3_NO_STACK
/*
* A hypervisor implementation might want to use a label
* after the swapgs, so that it can do the swapgs
* for the guest and jump here on syscall.
*/
GLOBAL(entry_SYSCALL_64_after_swapgs)
// 将用户栈偏移保存到 per-cpu 变量 rsp_scratch 中
movq %rsp, PER_CPU_VAR(rsp_scratch)
// 加载内核栈偏移
movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp
TRACE_IRQS_OFF
/* Construct struct pt_regs on stack */
pushq $__USER_DS /* pt_regs->ss */
pushq PER_CPU_VAR(rsp_scratch) /* pt_regs->sp */
pushq %r11 /* pt_regs->flags */
pushq $__USER_CS /* pt_regs->cs */
pushq %rcx /* pt_regs->ip */
pushq %rax /* pt_regs->orig_ax */
pushq %rdi /* pt_regs->di */
pushq %rsi /* pt_regs->si */
pushq %rdx /* pt_regs->dx */
pushq %rcx /* pt_regs->cx */
pushq $-ENOSYS /* pt_regs->ax */
pushq %r8 /* pt_regs->r8 */
pushq %r9 /* pt_regs->r9 */
pushq %r10 /* pt_regs->r10 */
pushq %r11 /* pt_regs->r11 */
// 为r12-r15, rbp, rbx保留位置
sub $(6*8), %rsp /* pt_regs->bp, bx, r12-15 not saved */
/*
* If we need to do entry work or if we guess we'll need to do
* exit work, go straight to the slow path.
*/
movq PER_CPU_VAR(current_task), %r11
testl $_TIF_WORK_SYSCALL_ENTRY|_TIF_ALLWORK_MASK, TASK_TI_flags(%r11)
jnz entry_SYSCALL64_slow_path
entry_SYSCALL_64_fastpath:
/*
* Easy case: enable interrupts and issue the syscall. If the syscall
* needs pt_regs, we'll call a stub that disables interrupts again
* and jumps to the slow path.
*/
TRACE_IRQS_ON
ENABLE_INTERRUPTS(CLBR_NONE)
#if __SYSCALL_MASK == ~0
// 确保系统调用号没超过最大值，超过了则跳转到后面的符号 1 处进行返回
cmpq $__NR_syscall_max, %rax
#else
andl $__SYSCALL_MASK, %eax
cmpl $__NR_syscall_max, %eax
#endif
ja 1f /* return -ENOSYS (already in pt_regs->ax) */
// 除系统调用外的其他调用都通过 rcx 来传第四个参数，因此将 r10 的内容设置到 rcx
movq %r10, %rcx
/*
* This call instruction is handled specially in stub_ptregs_64.
* It might end up jumping to the slow path. If it jumps, RAX
* and all argument registers are clobbered.
*/
// 调用系统调用表中对应的函数
call *sys_call_table(, %rax, 8)
.Lentry_SYSCALL_64_after_fastpath_call:
// 将函数返回值压到栈中，返回时弹出
movq %rax, RAX(%rsp)
1:
/*
* If we get here, then we know that pt_regs is clean for SYSRET64.
* If we see that no exit work is required (which we are required
* to check with IRQs off), then we can go straight to SYSRET64.
*/
DISABLE_INTERRUPTS(CLBR_NONE)
TRACE_IRQS_OFF
movq PER_CPU_VAR(current_task), %r11
testl $_TIF_ALLWORK_MASK, TASK_TI_flags(%r11)
jnz 1f
LOCKDEP_SYS_EXIT
TRACE_IRQS_ON /* user mode is traced as IRQs on */
movq RIP(%rsp), %rcx
movq EFLAGS(%rsp), %r11
RESTORE_C_REGS_EXCEPT_RCX_R11
/*
* This opens a window where we have a user CR3, but are
* running in the kernel. This makes using the CS
* register useless for telling whether or not we need to
* switch CR3 in NMIs. Normal interrupts are OK because
* they are off here.
*/
SWITCH_USER_CR3
movq RSP(%rsp), %rsp
USERGS_SYSRET64
1:
/*
* The fast path looked good when we started, but something changed
* along the way and we need to switch to the slow path. Calling
* raise(3) will trigger this, for example. IRQs are off.
*/
TRACE_IRQS_ON
ENABLE_INTERRUPTS(CLBR_NONE)
SAVE_EXTRA_REGS
movq %rsp, %rdi
call syscall_return_slowpath /* returns with IRQs disabled */
jmp return_from_SYSCALL_64
entry_SYSCALL64_slow_path:
/* IRQs are off. */
SAVE_EXTRA_REGS
movq %rsp, %rdi
call do_syscall_64 /* returns with IRQs disabled */
return_from_SYSCALL_64:
RESTORE_EXTRA_REGS
TRACE_IRQS_IRETQ /* we're about to change IF */
/*
* Try to use SYSRET instead of IRET if we're returning to
* a completely clean 64-bit userspace context.
*/
movq RCX(%rsp), %rcx
movq RIP(%rsp), %r11
cmpq %rcx, %r11 /* RCX == RIP */
jne opportunistic_sysret_failed
/*
* On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
* in kernel space. This essentially lets the user take over
* the kernel, since userspace controls RSP.
*
* If width of "canonical tail" ever becomes variable, this will need
* to be updated to remain correct on both old and new CPUs.
*/
.ifne __VIRTUAL_MASK_SHIFT - 47
.error "virtual address width changed -- SYSRET checks need update"
.endif
/* Change top 16 bits to be the sign-extension of 47th bit */
shl $(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
sar $(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
/* If this changed %rcx, it was not canonical */
cmpq %rcx, %r11
jne opportunistic_sysret_failed
cmpq $__USER_CS, CS(%rsp) /* CS must match SYSRET */
jne opportunistic_sysret_failed
movq R11(%rsp), %r11
cmpq %r11, EFLAGS(%rsp) /* R11 == RFLAGS */
jne opportunistic_sysret_failed
/*
* SYSCALL clears RF when it saves RFLAGS in R11 and SYSRET cannot
* restore RF properly. If the slowpath sets it for whatever reason, we
* need to restore it correctly.
*
* SYSRET can restore TF, but unlike IRET, restoring TF results in a
* trap from userspace immediately after SYSRET. This would cause an
* infinite loop whenever #DB happens with register state that satisfies
* the opportunistic SYSRET conditions. For example, single-stepping
* this user code:
*
* movq $stuck_here, %rcx
* pushfq
* popq %r11
* stuck_here:
*
* would never get past 'stuck_here'.
*/
testq $(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
jnz opportunistic_sysret_failed
/* nothing to check for RSP */
cmpq $__USER_DS, SS(%rsp) /* SS must match SYSRET */
jne opportunistic_sysret_failed
/*
* We win! This label is here just for ease of understanding
* perf profiles. Nothing jumps here.
*/
syscall_return_via_sysret:
/* rcx and r11 are already restored (see code above) */
RESTORE_C_REGS_EXCEPT_RCX_R11
/*
* This opens a window where we have a user CR3, but are
* running in the kernel. This makes using the CS
* register useless for telling whether or not we need to
* switch CR3 in NMIs. Normal interrupts are OK because
* they are off here.
*/
// KAISER 返回用户态需要切回用户页表
SWITCH_USER_CR3
/* 根据压栈的内容，恢复 rsp 为用户态的栈顶 */
movq RSP(%rsp), %rsp
USERGS_SYSRET64
// 无法快速返回，只能退化到 iret
opportunistic_sysret_failed:
/*
* This opens a window where we have a user CR3, but are
* running in the kernel. This makes using the CS
* register useless for telling whether or not we need to
* switch CR3 in NMIs. Normal interrupts are OK because
* they are off here.
*/
SWITCH_USER_CR3
SWAPGS
jmp restore_c_regs_and_iret
END(entry_SYSCALL_64)注意 syscall 不会保存栈指针，因此 handler 首先将当前用户态栈偏移 rsp 存到 per-cpu 变量 rsp_scratch 中，然后将 per-cpu 变量 cpu_current_top_of_stack ，即内核态的栈偏移加载到 rsp。
随后将各寄存器中的值压入内核态的栈中，包括：
- rax system call number
- rcx return address
- r11 saved rflags (note: r11 is callee-clobbered register in C ABI)
- rdi arg0
- rsi arg1
- rdx arg2
- r10 arg3 (needs to be moved to rcx to conform to C ABI)
- r8 arg4
- r9 arg5
接着根据系统调用号从系统调用表(sys_call_table) 中找到相应的处理函数，如 sys_open ，进行调用。64位下系统调用定义在 arch/x86/entry/syscalls/syscall_64.tbl中，ABI 和 32 位不同。
如果一切顺利的话，最终通过 USERGS_SYSRET64 ，即 sysretq 返回。
六、总结
本文主要分析了Linux下的三种系统调用方式：int 0x80 ，sysenter 和 syscall 。
传统系统调用(int 0x80) 通过中断/异常实现，在执行 int 指令时，发生 trap。硬件找到在中断描述符表中的表项，在自动切换到内核栈 (tss.ss0 : tss.esp0) 后根据中断描述符的 segment selector 在 GDT / LDT 中找到对应的段描述符，从段描述符拿到段的基址，加载到 cs ，将 offset 加载到 eip。最后硬件将 ss / sp / eflags / cs / ip / error code 依次压到内核栈。返回时，iret 将先前压栈的 ss / sp / eflags / cs / ip 弹出，恢复用户态调用时的寄存器上下文。
sysenter 和 syscall 是为了加速系统调用所引入的新指令，通过引入新的 MSR 来存放内核态的代码和栈的段号和偏移量，从而实现快速跳转：
在调用 sysenter 时将 SYSENTER_CS_MSR 加载到 cs，将 SYSENTER_CS_MSR + 8 加载到 ss，将 IA32_SYSENTER_EIP 加载到 eip ，将 IA32_SYSENTER_ESP 加载到 esp ，整套切换到内核态。返回时，sysexit 将 IA32_SYSENTER_CS + 16 加载到 cs ，将 IA32_SYSENTER_CS + 24 加载到 cs ，而 eip 和 esp 分别从 edx 和 ecx 中加载，因此返回前应该将压栈的用户态 eip(计算出来的) 和 esp(调用前用户态保存到 ebp 进行传递) 设置到这两个寄存器中。
在调用 syscall 时，会自动将 rip 保存到 rcx ，然后将 IA32_LSTAR 加载到 rip 。同时将 IA32_STAR[47:32] 加载到 cs ，IA32_STAR[47:32] + 8 加载到 ss 。栈顶指针的切换会延迟到内核态系统调用入口点 entry_SYSCALL_64 后进行处理，将用户态栈偏移 rsp 存到 per-cpu 变量 rsp_scratch 中，然后将 per-cpu 变量 cpu_current_top_of_stack ，即内核态的栈偏移加载到 rsp。返回时，sysret 将 IA32_STAR[63:48] 加载到 cs ，IA32_STAR[63:48] + 8 加载到 ss ，而 rip 从 rcx 中加载，因此返回前应该将压栈的用户态 rip 设置到 rcx 中。对于 rsp ，返回前根据先前压栈内容先设置为用户态 rsp。
文章中肯定有遗漏或理解错误的地方，欢迎留言指正，不胜感激。
参考：
https://0xax.gitbooks.io/linux-insides/content/SysCall/
https://blog.packagecloud.io/eng/2016/04/05/the-definitive-guide-to-linux-system-calls/
http://www.ibm.com/developerworks/cn/linux/kernel/l-k26ncpu/index.html
https://lwn.net/Articles/604287/
https://lwn.net/Articles/604515/腾讯云开发者
