引言
__libc_open 是 glibc 中 openat 接口的封装，针对给定的 file 路径，打开文件，生成一个文件描述符 fd。可以创建文件/创建临时文件/打开已有文件。在很多 IO 接口函数中都有使用到，我们来一起看看它的实现原理。
源码分析
代码参考：glibc/sysdeps/unix/sysv/linux/open.c
28 /* Open FILE with access OFLAG. If O_CREAT or O_TMPFILE is in OFLAG, 29 a third argument is the file protection. */ 30 int 31 __libc_open (const char *file, int oflag, ...) 32 { 33 int mode = 0; 34 35 if (__OPEN_NEEDS_MODE (oflag)) 36 { 37 va_list arg; 38 va_start (arg, oflag); 39 mode = va_arg (arg, int); 40 va_end (arg); 41 } 42 43 return SYSCALL_CANCEL (openat, AT_FDCWD, file, oflag, mode); 44 }
复制代码
首先根据 oflag(里面是否包含 O_CREAT 或 O_TMPFILE)检查，是否需要获取第三个参数 mode(用来进行读写控制的)。
__OPEN_NEEDS_MODE 宏判定
具体实现里面就是通过位运算检测 oflag 中是否包含__O_TMPFILE 或 O_CREAT
//glibc/io/fcntl.h 37 /* Detect if open needs mode as a third argument (or for openat as a fourth 38 argument). */ 39 #ifdef __O_TMPFILE 40 # define __OPEN_NEEDS_MODE(oflag) \ 41 (((oflag) & O_CREAT) != 0 || ((oflag) & __O_TMPFILE) == __O_TMPFILE) 42 #else 43 # define __OPEN_NEEDS_MODE(oflag) (((oflag) & O_CREAT) != 0) 44 #endif
复制代码
SYSCALL_CANCEL
在这个宏中实际上最后还是调用 INLINE_SYSCALL_CALL 实现我们的调用，多余的部分实际上是 syscall 取消的检测，这里我们就不展开了。
//glibc/sysdeps/unix/sysdep.h111 #if IS_IN (rtld)112 /* All cancellation points are compiled out in the dynamic loader. */113 # define NO_SYSCALL_CANCEL_CHECKING 1114 #else115 # define NO_SYSCALL_CANCEL_CHECKING SINGLE_THREAD_P116 #endif117 118 #define SYSCALL_CANCEL(...) \ 119 ({ \120 long int sc_ret; \121 if (NO_SYSCALL_CANCEL_CHECKING) \122 sc_ret = INLINE_SYSCALL_CALL (__VA_ARGS__); \123 else \124 { \125 int sc_cancel_oldtype = LIBC_CANCEL_ASYNC (); \126 sc_ret = INLINE_SYSCALL_CALL (__VA_ARGS__); \127 LIBC_CANCEL_RESET (sc_cancel_oldtype); \128 } \129 sc_ret; \130 })
复制代码
INLINE_SYSCALL_CALL
这里就是对 syscall 的相关封装了，实际上就是要从我们之前传入的SYSCALL_CANCEL (openat, AT_FDCWD, file, oflag, mode);解析出具体需要执行的汇编指令
103 /* Issue a syscall defined by syscall number plus any other argument104 required. Any error will be handled using arch defined macros and errno105 will be set accordingly.106 It is similar to INLINE_SYSCALL macro, but without the need to pass the107 expected argument number as second parameter. */108 #define INLINE_SYSCALL_CALL(...) \109 __INLINE_SYSCALL_DISP (__INLINE_SYSCALL, __VA_ARGS__)
复制代码
宏展开之后__INLINE_SYSCALL_DISP (__INLINE_SYSCALL, openat, AT_FDCWD, file, oflag, mode)
100 #define __INLINE_SYSCALL_DISP(b,...) \101 __SYSCALL_CONCAT (b,__INLINE_SYSCALL_NARGS(__VA_ARGS__))(__VA_ARGS__)
复制代码
展开之后：__SYSCALL_CONCAT (__INLINE_SYSCALL, __INLINE_SYSCALL_NARGS(openat, AT_FDCWD, file, oflag, mode)) (openat, AT_FDCWD, file, oflag, mode)
__INLINE_SYSCALL_NARGS
先具体分析__INLINE_SYSCALL_NARGS 的宏展开方式：
98 #define __INLINE_SYSCALL_NARGS(...) \ 99 __INLINE_SYSCALL_NARGS_X (__VA_ARGS__,7,6,5,4,3,2,1,0,)
复制代码
展开如下，在原有参数后面增加了 8 个数字
__INLINE_SYSCALL_NARGS_X (openat, AT_FDCWD, file, oflag, mode,7,6,5,4,3,2,1,0,)
继续展开，可以看到这里是取第九个参数，对应上面这个表达式，也就是 4，可以推出，如果输入有 1 个参数，那么返回 0,输入有 2 个参数，返回 1,依次类推，输入有 8 个参数，返回 7。97 #define __INLINE_SYSCALL_NARGS_X(a,b,c,d,e,f,g,h,n,...) n
__SYSCALL_CONCAT
实际上就是把两个字符进行了连接，“##”在宏当中是进行字符连接的
27 #define __SYSCALL_CONCAT_X(a,b) a##b 28 #define __SYSCALL_CONCAT(a,b) __SYSCALL_CONCAT_X (a, b)
复制代码
至此，我们的宏展开就应该是这样的
__INLINE_SYSCALL4(openat, AT_FDCWD, file, oflag, mode)
__INLINE_SYSCALL4
对应的定义如下：就应该被展开为INLINE_SYSCALL (openat, 4, AT_FDCWD, file, oflag, mode)
80 #define __INLINE_SYSCALL0(name) \ 81 INLINE_SYSCALL (name, 0) 82 #define __INLINE_SYSCALL1(name, a1) \ 83 INLINE_SYSCALL (name, 1, a1) 84 #define __INLINE_SYSCALL2(name, a1, a2) \ 85 INLINE_SYSCALL (name, 2, a1, a2) 86 #define __INLINE_SYSCALL3(name, a1, a2, a3) \ 87 INLINE_SYSCALL (name, 3, a1, a2, a3) 88 #define __INLINE_SYSCALL4(name, a1, a2, a3, a4) \ 89 INLINE_SYSCALL (name, 4, a1, a2, a3, a4) 90 #define __INLINE_SYSCALL5(name, a1, a2, a3, a4, a5) \ 91 INLINE_SYSCALL (name, 5, a1, a2, a3, a4, a5) 92 #define __INLINE_SYSCALL6(name, a1, a2, a3, a4, a5, a6) \ 93 INLINE_SYSCALL (name, 6, a1, a2, a3, a4, a5, a6) 94 #define __INLINE_SYSCALL7(name, a1, a2, a3, a4, a5, a6, a7) \ 95 INLINE_SYSCALL (name, 7, a1, a2, a3, a4, a5, a6, a7)
复制代码
INLINE_SYSCALL
封装调用了 INTERNAL_SYSCALL 宏
INTERNAL_SYSCALL (openat, 4, AT_FDCWD, file, oflag, mode)
38 /* Define a macro which expands into the inline wrapper code for a system 39 call. It sets the errno and returns -1 on a failure, or the syscall 40 return value otherwise. */ 41 #undef INLINE_SYSCALL 42 #define INLINE_SYSCALL(name, nr, args...) \ 43 ({ \ 44 long int sc_ret = INTERNAL_SYSCALL (name, nr, args); \ 45 __glibc_unlikely (INTERNAL_SYSCALL_ERROR_P (sc_ret)) \ 46 ? SYSCALL_ERROR_LABEL (INTERNAL_SYSCALL_ERRNO (sc_ret)) \ 47 : sc_ret; \ 48 })
复制代码
INTERNAL_SYSCALL
这一步的实现就与每个平台不同了，这里以 x86_64 平台为例上面的被依次如下封装:
internal_syscall4 (SYS_ify (openat), AT_FDCWD, file, oflag, mode)
internal_syscall4 (__NR_openat, AT_FDCWD, file, oflag, mode)
//glibc/sysdeps/unix/sysv/linux/x86_64/sysdep.h
29 /* For Linux we can use the system call table in the header file 30 /usr/include/asm/unistd.h 31 of the kernel. But these symbols do not follow the SYS_* syntax 32 so we have to redefine the `SYS_ify' macro here. */ 33 #undef SYS_ify 34 #define SYS_ify(syscall_name) __NR_##syscall_name 233 #undef INTERNAL_SYSCALL234 #define INTERNAL_SYSCALL(name, nr, args...) \235 internal_syscall##nr (SYS_ify (name), args)
复制代码
最终我们的调用为internal_syscall4 (__NR_openat, AT_FDCWD, file, oflag, mode)
依次将 4 个参数装入寄存器中，然后调用 syscall 汇编代码实现调用，这里用到 C 语言的 asm 汇编语法
301 #undef internal_syscall4302 #define internal_syscall4(number, arg1, arg2, arg3, arg4) \303 ({ \304 unsigned long int resultvar; \305 TYPEFY (arg4, __arg4) = ARGIFY (arg4); \306 TYPEFY (arg3, __arg3) = ARGIFY (arg3); \307 TYPEFY (arg2, __arg2) = ARGIFY (arg2); \308 TYPEFY (arg1, __arg1) = ARGIFY (arg1); \309 register TYPEFY (arg4, _a4) asm ("r10") = __arg4; \310 register TYPEFY (arg3, _a3) asm ("rdx") = __arg3; \311 register TYPEFY (arg2, _a2) asm ("rsi") = __arg2; \312 register TYPEFY (arg1, _a1) asm ("rdi") = __arg1; \313 asm volatile ( \314 "syscall\n\t" \315 : "=a" (resultvar) \316 : "0" (number), "r" (_a1), "r" (_a2), "r" (_a3), "r" (_a4) \317 : "memory", REGISTERS_CLOBBERED_BY_SYSCALL); \318 (long int) resultvar; \319 })
复制代码
我们看一下该架构下使用最多参数的宏，（x86_64 架构最多只支持 6 个参数）
344 #undef internal_syscall6345 #define internal_syscall6(number, arg1, arg2, arg3, arg4, arg5, arg6) \346 ({ \347 unsigned long int resultvar; \348 TYPEFY (arg6, __arg6) = ARGIFY (arg6); \349 TYPEFY (arg5, __arg5) = ARGIFY (arg5); \350 TYPEFY (arg4, __arg4) = ARGIFY (arg4); \351 TYPEFY (arg3, __arg3) = ARGIFY (arg3); \352 TYPEFY (arg2, __arg2) = ARGIFY (arg2); \353 TYPEFY (arg1, __arg1) = ARGIFY (arg1); \354 register TYPEFY (arg6, _a6) asm ("r9") = __arg6; \355 register TYPEFY (arg5, _a5) asm ("r8") = __arg5; \356 register TYPEFY (arg4, _a4) asm ("r10") = __arg4; \357 register TYPEFY (arg3, _a3) asm ("rdx") = __arg3; \358 register TYPEFY (arg2, _a2) asm ("rsi") = __arg2; \359 register TYPEFY (arg1, _a1) asm ("rdi") = __arg1; \360 asm volatile ( \361 "syscall\n\t" \362 : "=a" (resultvar) \363 : "0" (number), "r" (_a1), "r" (_a2), "r" (_a3), "r" (_a4), \364 "r" (_a5), "r" (_a6) \365 : "memory", REGISTERS_CLOBBERED_BY_SYSCALL); \366 (long int) resultvar; \367 })
复制代码
寄存器的使用中参数依次装入 rdi,rsi,rdx,r10,r8,r9
支持最多 7 个参数的只有如下的架构：
riscv 架构：glibc/sysdeps/unix/sysv/linux/riscv/sysdep.h 分别使用 a0 到 a6 寄存器装参数
323 # define internal_syscall7(number, arg0, arg1, arg2, arg3, arg4, arg5, arg6) \ 324 ({ \325 long int _sys_result; \326 long int _arg0 = (long int) (arg0); \327 long int _arg1 = (long int) (arg1); \328 long int _arg2 = (long int) (arg2); \329 long int _arg3 = (long int) (arg3); \330 long int _arg4 = (long int) (arg4); \331 long int _arg5 = (long int) (arg5); \332 long int _arg6 = (long int) (arg6); \333 \334 { \335 register long int __a7 asm ("a7") = number; \336 register long int __a0 asm ("a0") = _arg0; \337 register long int __a1 asm ("a1") = _arg1; \338 register long int __a2 asm ("a2") = _arg2; \339 register long int __a3 asm ("a3") = _arg3; \340 register long int __a4 asm ("a4") = _arg4; \341 register long int __a5 asm ("a5") = _arg5; \342 register long int __a6 asm ("a6") = _arg6; \343 __asm__ volatile ( \344 "scall\n\t" \345 : "+r" (__a0) \346 : "r" (__a7), "r" (__a1), "r" (__a2), "r" (__a3), \347 "r" (__a4), "r" (__a5), "r" (__a6) \348 : __SYSCALL_CLOBBERS); \349 _sys_result = __a0; \350 } \351 _sys_result; \352 })
复制代码
mips 架构：glibc/sysdeps/unix/sysv/linux/mips/mips32/sysdep.h
315 #define internal_syscall7(v0_init, input, number, err, \316 arg1, arg2, arg3, arg4, arg5, arg6, arg7) \317 ({ \318 union __mips_syscall_return _sc_ret; \319 _sc_ret.val = __mips_syscall7 ((long int) (arg1), \320 (long int) (arg2), \321 (long int) (arg3), \322 (long int) (arg4), \323 (long int) (arg5), \ 324 (long int) (arg6), \325 (long int) (arg7), \326 (long int) (number)); \327 _sc_ret.reg.v1 != 0 ? -_sc_ret.reg.v0 : _sc_ret.reg.v0; \328 })
复制代码
总结
从最开始的SYSCALL_CANCEL (openat, AT_FDCWD, file, oflag, mode)宏，依次转换为
internal_syscall4 (__NR_openat, AT_FDCWD, file, oflag, mode)，然后调用汇编实现底层 syscall 操作。
评论