1.
home
2.
PDF

How system calls work in Linux

Nasser M. Abbasi

May 29, 2000   page compiled on June 28, 2015 at 4:55am

These are notes I wrote while learning how system calls work on a Linux system.

To help show this how system call works, I show flow of a typical system call such as fopen().

fopen() is a function call defined in the C standard library. I use glibc-2.1 as an implementation.

From the UNIX98 standard, fopen() is defined as

  #include <stdio.h>
  
      FILE *fopen(const char *filename, const char *mode);
  
   DESCRIPTION
  
      The fopen() function opens the file whose pathname is the string pointed
      to by filename, and associates a stream with it.
  
      The argument mode points to a string beginning with one of the following sequences:
  
      r or rb
          Open file for reading.
      w or wb
          Truncate to zero length or create file for writing.
      a or ab
          Append; open or create file for writing at end-of-file.
      r+ or rb+ or r+b
          Open file for update (reading and writing).
      w+ or wb+ or w+b
          Truncate to zero length or create file for update.
      a+ or ab+ or a+b
          Append; open or create file for update, writing at end-of-file.

Create the following t.c C program to use to test with:

  #include <stdio.h>
  
  int main(int argc, char *argv[])
  {
    FILE *f;
  
    f = fopen("test.txt","r");
  
    return 0;
  }

To step into fopen(), glibc 2.1 was build in debug and the new build libc.a was linked against instead of the default installed libc on my linux box.

To build glibc, the following are steps performed. A good reference is the glibc2 HOWTO, http://www.linux.ps.pl/doc/other/LDP/HOWTO/Glibc2-HOWTO.html

First, I downloaded the glibc tar file to /usr/src/packages/SOURCES. Extracted It and it created glibc-2.1/ directory. Then copied the crypt tar file into glibc-2.1/ and extracted that. It created crypt/ directory under glibc-2.1/. Next I did

    cd glibc-2.1
    ./configure --enable-add-ons
    make
    make check

Next, I installed the library into a direcory called INSTALL_LIB under glibc-2.1.

      make install install_root=/usr/src/packages/SOURCES/glibc-2.1/INSTALL_LIB

OK, now glibc-2.1 is compiled and ready to use. Back to the little C program we have above. Lets now compile it and link it to the above library.

   gcc -static -g -I /usr/src/packages/SOURCES/glibc-2.1/INSTALL_LIB/usr/local/include \
      -L/usr/src/packages/SOURCES/glibc-2.1/INSTALL_LIB/usr/local/lib t.c

Ok, now lets step through it.

  $gdb ./a.out
  GNU gdb 4.18
  (gdb) break main
  Breakpoint 1 at 0x80481b6: file t.c, line 7.
  
  (gdb) run
  Starting program: /export/g/nabbasi/data/my_misc_programs/my_c/./a.out
  
  Breakpoint 1, main (argc=1, argv=0xbffff434) at t.c:7
  7         f = fopen("test.txt","r");
  (gdb) list
  2
  3       int main(int argc, char *argv[])
  4       {
  5         FILE *f;
  6
  7         f = fopen("test.txt","r");
  8
  9         return 0;
  10      }
  
  (gdb) disassemble main
  Dump of assembler code for function main:
  0x80481b0 <main>:       push   %ebp
  0x80481b1 <main+1>:     mov    %esp,%ebp
  0x80481b3 <main+3>:     sub    $0x4,%esp
  0x80481b6 <main+6>:     push   $0x8071ba8
  0x80481bb <main+11>:    push   $0x8071baa
  0x80481c0 <main+16>:    call   0x8048710 <_IO_new_fopen>
  0x80481c5 <main+21>:    add    $0x8,%esp
  0x80481c8 <main+24>:    mov    %eax,%eax
  0x80481ca <main+26>:    mov    %eax,0xfffffffc(%ebp)
  0x80481cd <main+29>:    xor    %eax,%eax
  0x80481cf <main+31>:    jmp    0x80481e0 <main+48>
  0x80481d1 <main+33>:    jmp    0x80481e0 <main+48>
  0x80481e0 <main+48>:    mov    %ebp,%esp
  0x80481e2 <main+50>:    pop    %ebp
  0x80481e3 <main+51>:    ret
  End of assembler dump.

Humm... what happened to printf call? you will notice, it is now a call to _IO_new_fopen. But I was calling fopen, not _IO_new_fopen?.

Lets step into _IO_new_fopen and see what happened.

  (gdb) s
  _IO_new_fopen (filename=0x8071baa "test.txt", mode=0x8071ba8 "r") at iofopen.c:42
  42        } *new_f = (struct locked_FILE *) malloc (sizeof (struct locked_FILE));

So, _IO_new_fopen is an entry in iofopen.c. Where is this file?

  cd glibc-2.1
  find . -name iofopen.c

will show it as glibc-2.1/libio/iofopen.c Lets look at it

  #include "libioP.h"
  #ifdef __STDC__
  #include <stdlib.h>
  #endif
  _IO_FILE *
  _IO_new_fopen (filename, mode)
       const char *filename;
       const char *mode;
  {
    struct locked_FILE
    {
      struct _IO_FILE_plus fp;
  #ifdef _IO_MTSAFE_IO
      _IO_lock_t lock;
  #endif
    } *new_f = (struct locked_FILE *) malloc (sizeof (struct locked_FILE));
    if (new_f == NULL)
      return NULL;
  #ifdef _IO_MTSAFE_IO
    new_f->fp.file._lock = &new_f->lock;
  #endif
    _IO_init (&new_f->fp.file, 0);
    _IO_JUMPS (&new_f->fp) = &_IO_file_jumps;
    _IO_file_init (&new_f->fp.file);
  #if  !_IO_UNIFIED_JUMPTABLES
    new_f->fp.vtable = NULL;
  #endif
    if (_IO_file_fopen (&new_f->fp.file, filename, mode, 1) != NULL)
      return (_IO_FILE *) &new_f->fp;
    _IO_un_link (&new_f->fp.file);
    free (new_f);
    return NULL;
  }
  #if defined PIC && DO_VERSIONING
  strong_alias (_IO_new_fopen, __new_fopen)
  default_symbol_version (_IO_new_fopen, _IO_fopen, GLIBC_2.1);
  default_symbol_version (__new_fopen, fopen, GLIBC_2.1);
  #else
  # ifdef weak_alias
  weak_alias (_IO_new_fopen, _IO_fopen)
  weak_alias (_IO_new_fopen, fopen)
  # endif
  #endif

Notice at the end what it says, it says weak_alias (_IO_new_fopen, fopen). This tells gcc that _IO_new_fopen is an alias to fopen. (weak alias). Let me make sure. Looking at libc.a now

  cd /usr/src/packages/SOURCES/glibc-2.1/INSTALL_LIB/usr/local/lib
  nm libc.a
  ...
  iofopen.o:
           U _IO_file_fopen
           U _IO_file_init
           U _IO_file_jumps
  00000000 W _IO_fopen
           U _IO_init
  00000000 T _IO_new_fopen
           U _IO_un_link
           U __pthread_atfork
           U __pthread_getspecific
           U __pthread_initialize
           U __pthread_key_create
           U __pthread_mutex_destroy
           U __pthread_mutex_init
           U __pthread_mutex_lock
           U __pthread_mutex_trylock
           U __pthread_mutex_unlock
           U __pthread_mutexattr_destroy
           U __pthread_mutexattr_init
           U __pthread_mutexattr_settype
           U __pthread_once
           U __pthread_setspecific
           U _pthread_cleanup_pop_restore
           U _pthread_cleanup_push_defer
  00000000 W fopen
           U free
           U malloc
  ...

Notice fopen has W next to it, meaning a Weak symbol. So, the linker when it sees a call to fopen will bind the call to _IO_new_fopen.

It is just a different name for fopen. This way, library can create different implementations for calls without the user program having to change.

Ok, now, lets continue to see where we will end up. back to gdb.

  (gdb) disassemble fopen
  Dump of assembler code for function _IO_new_fopen:
  0x8048710 <_IO_new_fopen>:      push   %ebp
  0x8048711 <_IO_new_fopen+1>:    mov    %esp,%ebp
  0x8048713 <_IO_new_fopen+3>:    push   %ebx
  0x8048714 <_IO_new_fopen+4>:    push   $0xb0
  0x8048719 <_IO_new_fopen+9>:    call   0x804b020 <__libc_malloc>
  0x804871e <_IO_new_fopen+14>:   mov    %eax,%ebx
  0x8048720 <_IO_new_fopen+16>:   add    $0x4,%esp
  0x8048723 <_IO_new_fopen+19>:   test   %ebx,%ebx
  0x8048725 <_IO_new_fopen+21>:   jne    0x8048730 <_IO_new_fopen+32>
  0x8048727 <_IO_new_fopen+23>:   xor    %eax,%eax
  0x8048729 <_IO_new_fopen+25>:   jmp    0x8048782 <_IO_new_fopen+114>
  0x804872b <_IO_new_fopen+27>:   nop
  0x804872c <_IO_new_fopen+28>:   lea    0x0(%esi,1),%esi
  0x8048730 <_IO_new_fopen+32>:   lea    0x98(%ebx),%edx
  0x8048736 <_IO_new_fopen+38>:   mov    %edx,0x48(%ebx)
  0x8048739 <_IO_new_fopen+41>:   push   $0x0
  0x804873b <_IO_new_fopen+43>:   push   %ebx
  0x804873c <_IO_new_fopen+44>:   call   0x804a030 <_IO_init>
  0x8048741 <_IO_new_fopen+49>:   movl   $0x807a360,0x94(%ebx)
  0x804874b <_IO_new_fopen+59>:   push   %ebx
  0x804874c <_IO_new_fopen+60>:   call   0x80487a0 <_IO_new_file_init>
  0x8048751 <_IO_new_fopen+65>:   push   $0x1
  0x8048753 <_IO_new_fopen+67>:   mov    0xc(%ebp),%eax
  0x8048756 <_IO_new_fopen+70>:   push   %eax
  0x8048757 <_IO_new_fopen+71>:   mov    0x8(%ebp),%eax
  0x804875a <_IO_new_fopen+74>:   push   %eax
  0x804875b <_IO_new_fopen+75>:   push   %ebx
  0x804875c <_IO_new_fopen+76>:   call   0x80488e0 <_IO_new_file_fopen>
  0x8048761 <_IO_new_fopen+81>:   add    $0x1c,%esp
  0x8048764 <_IO_new_fopen+84>:   test   %eax,%eax
  0x8048766 <_IO_new_fopen+86>:   jne    0x8048780 <_IO_new_fopen+112>
  0x8048768 <_IO_new_fopen+88>:   push   %ebx
  0x8048769 <_IO_new_fopen+89>:   call   0x80497a0 <_IO_un_link>
  0x804876e <_IO_new_fopen+94>:   push   %ebx
  0x804876f <_IO_new_fopen+95>:   call   0x804b9f0 <__libc_free>
  0x8048774 <_IO_new_fopen+100>:  xor    %eax,%eax
  0x8048776 <_IO_new_fopen+102>:  jmp    0x8048782 <_IO_new_fopen+114>
  0x8048778 <_IO_new_fopen+104>:  nop
  0x8048779 <_IO_new_fopen+105>:  lea    0x0(%esi,1),%esi
  0x8048780 <_IO_new_fopen+112>:  mov    %ebx,%eax
  0x8048782 <_IO_new_fopen+114>:  mov    0xfffffffc(%ebp),%ebx
  0x8048785 <_IO_new_fopen+117>:  mov    %ebp,%esp
  0x8048787 <_IO_new_fopen+119>:  pop    %ebp
  0x8048788 <_IO_new_fopen+120>:  ret
  End of assembler dump.

The call I am interested in is _IO_new_file_fopen. The earlier calls were calls that create and initialize data structures. I am interested in finding the call that will result in interrupt 0x80. So, lets step to _IO_new_file_fopen.

  (gdb) break _IO_new_file_fopen
  Breakpoint 3 at 0x80488ec: file fileops.c, line 204.
  (gdb) continue
  Continuing.
  
  Breakpoint 3, _IO_new_file_fopen (fp=0x807c838, filename=0x8071baa "test.txt", mode=0x8071ba8 "r", is32not64=1) at fileops.c:204
  204       int oflags = 0, omode;
  (gdb)

The file fileops.c is located in glibc-2.1/libio/, lets look at the source code for _IO_file_fopen() in that file:

  _IO_FILE *
  _IO_new_file_fopen (fp, filename, mode, is32not64)
       _IO_FILE *fp;
       const char *filename;
       const char *mode;
       int is32not64;
  {
    int oflags = 0, omode;
    int read_write;
    int oprot = 0666;
    int i;
    if (_IO_file_is_open (fp))
      return 0;
    switch (*mode)
      {
      case 'r':
        omode = O_RDONLY;
        read_write = _IO_NO_WRITES;
        break;
      case 'w':
        omode = O_WRONLY;
        oflags = O_CREAT|O_TRUNC;
        read_write = _IO_NO_READS;
        break;
      case 'a':
        omode = O_WRONLY;
        oflags = O_CREAT|O_APPEND;
        read_write = _IO_NO_READS|_IO_IS_APPENDING;
        break;
      default:
        __set_errno (EINVAL);
        return NULL;
      }
    for (i = 1; i < 4; ++i)
      {
        switch (*++mode)
   {
   case '\0':
     break;
   case '+':
     omode = O_RDWR;
     read_write &= _IO_IS_APPENDING;
     continue;
   case 'x':
     oflags |= O_EXCL;
     continue;
   case 'b':
   default:
     /* Ignore.  */
     continue;
   }
        break;
      }
  
    return _IO_file_open (fp, filename, omode|oflags, oprot, read_write,  ----> step here
   is32not64);
  }

Let us assume the file is not allready open, the next call will be _IO_file_open()

Setting a break point there. But notice, looking at source code in fileops.c, the above call to _IO_file_open is inlined (for performance?)

  #if defined __GNUC__ && __GNUC__ >= 2
  __inline__
  #endif
  _IO_FILE *
  _IO_file_open (fp, filename, posix_mode, prot, read_write, is32not64)
       _IO_FILE *fp;
       const char *filename;
       int posix_mode;
       int prot;
       int read_write;
       int is32not64;
  {
    int fdesc;
  #ifdef _G_OPEN64
    fdesc = (is32not64
      ? open (filename, posix_mode, prot)
      : _G_OPEN64 (filename, posix_mode, prot));
  #else
    fdesc = open (filename, posix_mode, prot);
  #endif
    if (fdesc < 0)
      return NULL;
    fp->_fileno = fdesc;
    _IO_mask_flags (fp, read_write,_IO_NO_READS+_IO_NO_WRITES+_IO_IS_APPENDING);
    if (read_write & _IO_IS_APPENDING)
      if (_IO_SEEKOFF (fp, (_IO_off64_t)0, _IO_seek_end, _IOS_INPUT|_IOS_OUTPUT)
   == _IO_pos_BAD && errno != ESPIPE)
        return NULL;
    _IO_link_in (fp);
    return fp;
  }

Setting a break point at the call to open above.

  (gdb) where
  #0  _IO_new_file_fopen (fp=0x807c838, filename=0x8071baa "test.txt", mode=0x8071ba8 "r", is32not64=1) at fileops.c:179
  #1  0x8048761 in _IO_new_fopen (filename=0x8071baa "test.txt", mode=0x8071ba8 "r") at iofopen.c:55
  #2  0x80481c5 in main (argc=1, argv=0xbffff434) at t.c:7
  (gdb) list
  174          int read_write;
  175          int is32not64;
  176     {
  177       int fdesc;
  178     #ifdef _G_OPEN64
  179       fdesc = (is32not64
  180                ? open (filename, posix_mode, prot)   -------> This is call we need
  181                : _G_OPEN64 (filename, posix_mode, prot));
  182     #else
  183       fdesc = open (filename, posix_mode, prot);
  (gdb) break open
  Breakpoint 2 at 0x804df80
  (gdb)

Since _IO_file_fopen is inlined inside _IO_new_file_fopen, we can look at the assembler call to open above by disassembly of _IO_new_file_fopen().

I'll show only the part where the call to open is made

  (gdb) disassemble _IO_new_file_fopen
  Dump of assembler code for function _IO_new_file_fopen:
  ...
  0x80489e4 <_IO_new_file_fopen+260>:     push   $0x1b6
  0x80489e9 <_IO_new_file_fopen+265>:     push   %eax
  0x80489ea <_IO_new_file_fopen+266>:     mov    0xc(%ebp),%edi
  0x80489ed <_IO_new_file_fopen+269>:     push   %edi
  0x80489ee <_IO_new_file_fopen+270>:     call   0x804df80 <__libc_open>  ----> this is open
  ...

Ok, back to gdb, setting a breakpoint at open and stepping into it

  (gdb) where
  #0  _IO_new_file_fopen (fp=0x807c838, filename=0x8071baa "test.txt", mode=0x8071ba8 "r", is32not64=1) at fileops.c:179
  #1  0x8048761 in _IO_new_fopen (filename=0x8071baa "test.txt", mode=0x8071ba8 "r") at iofopen.c:55
  #2  0x80481c5 in main (argc=1, argv=0xbffff434) at t.c:7
  (gdb) s
  
  Breakpoint 2, 0x804df80 in __libc_open ()
  (gdb) disassemble
  Dump of assembler code for function __libc_open:
  0x804df80 <__libc_open>:        push   %ebx
  0x804df81 <__libc_open+1>:      mov    0x10(%esp,1),%edx
  0x804df85 <__libc_open+5>:      mov    0xc(%esp,1),%ecx
  0x804df89 <__libc_open+9>:      mov    0x8(%esp,1),%ebx
  0x804df8d <__libc_open+13>:     mov    $0x5,%eax
  0x804df92 <__libc_open+18>:     int    $0x80    -----> to kernel mode.
  0x804df94 <__libc_open+20>:     pop    %ebx
  0x804df95 <__libc_open+21>:     cmp    $0xfffff001,%eax
  0x804df9a <__libc_open+26>:     jae    0x804e450 <__syscall_error>
  0x804dfa0 <__libc_open+32>:     ret
  End of assembler dump.
  (gdb)

We are finally there. The open() call being made from _IO_file_open(), is translated to __libc_open() and __libc_open() will issue the interupt 0x80, which will turn the processor to run in kernel more, and the interrupt handler will locate the kernel system call to process open().

But before jumping into kernel mode, lets see how did the call to open() become a call to __libc_open() It turns out that when building glibc-2.1, there is a file called glibc-2.1/sysdeps/unix/syscalls.list

This file is used by the glibc build system to generate the wrapper for open() and call it __libc_open.

  >cat glibc-2.1/sysdeps/unix/syscalls.list
  # File name     Caller  Syscall name    # args  Strong name     Weak names
  
  access          -       access          2       __access        access
  acct            -       acct            1       acct
  chdir           -       chdir           1       __chdir         chdir
  chmod           -       chmod           2       __chmod         chmod
  chown           -       chown           3       __chown         chown
  chroot          -       chroot          1       chroot
  close           -       close           1       __libc_close    __close close
  dup             -       dup             1       __dup           dup
  dup2            -       dup2            2       __dup2          dup2
  fchdir          -       fchdir          1       __fchdir        fchdir
  fcntl           -       fcntl           3       __libc_fcntl    __fcntl fcntl
  fstatfs         -       fstatfs         2       __fstatfs       fstatfs
  fsync           -       fsync           1       __libc_fsync    fsync
  getdomain       -       getdomainname   2       getdomainname
  getgid          -       getgid          0       __getgid        getgid
  getgroups       -       getgroups       2       __getgroups     getgroups
  getitimer       -       getitimer       2       __getitimer     getitimer
  getpid          -       getpid          0       __getpid        getpid
  getpriority     -       getpriority     2       getpriority
  getrlimit       -       getrlimit       2       __getrlimit     getrlimit
  getuid          -       getuid          0       __getuid        getuid
  ioctl           -       ioctl           3       __ioctl         ioctl
  kill            -       kill            2       __kill          kill
  link            -       link            2       __link          link
  lseek           -       lseek           3       __libc_lseek    __lseek lseek
  mkdir           -       mkdir           2       __mkdir         mkdir
  open            -       open            3       __libc_open     __open open
  profil          -       profil          4       profil
  ptrace          -       ptrace          4       ptrace
  read            -       read            3       __libc_read     __read read
  readlink        -       readlink        3       __readlink      readlink
  readv           -       readv           3       __readv         readv
  reboot          -       reboot          1       reboot
  rename          -       rename          2       rename
  rmdir           -       rmdir           1       __rmdir         rmdir
  select          -       select          5       __select        select
  setdomain       -       setdomainname   2       setdomainname
  setegid         -       setegid         1       __setegid       setegid
  seteuid         -       seteuid         1       __seteuid       seteuid
  setgid          -       setgid          1       __setgid        setgid
  setgroups       -       setgroups       2       setgroups
  setitimer       -       setitimer       3       __setitimer     setitimer
  setpriority     -       setpriority     3       setpriority
  setrlimit       -       setrlimit       2       setrlimit
  setsid          -       setsid          0       __setsid        setsid
  settimeofday    -       settimeofday    2       __settimeofday  settimeofday
  setuid          -       setuid          1       __setuid        setuid
  sigsuspend      -       sigsuspend      1       sigsuspend
  sstk            -       sstk            1       sstk
  statfs          -       statfs          2       __statfs        statfs
  swapoff         -       swapoff         1       swapoff
  swapon          -       swapon          1       swapon
  symlink         -       symlink         2       __symlink       symlink
  sync            -       sync            0       sync
  sys_fstat       fxstat  fstat           2       __syscall_fstat
  sys_mknod       xmknod  mknod           3       __syscall_mknod
  sys_stat        xstat   stat            2       __syscall_stat
  umask           -       umask           1       __umask         umask
  uname           -       uname           1       uname
  unlink          -       unlink          1       __unlink        unlink
  utimes          -       utimes          2       __utimes        utimes
  write           -       write           3       __libc_write    __write write
  writev          -       writev          3       __writev        writev

I extraced open.o from libc.a and dumped the open.o

  
  use ar -x libc.a, in some temp dir.
  
  >objdump --show-raw-insn open.o
  open.o:     file format elf32-i386
  >objdump  --disassemble open.o
  open.o:     file format elf32-i386
  Disassembly of section .text:
  00000000 <__libc_open>:
     0:   53                      pushl  %ebx
     1:   8b 54 24 10             movl   0x10(%esp,1),%edx
     5:   8b 4c 24 0c             movl   0xc(%esp,1),%ecx
     9:   8b 5c 24 08             movl   0x8(%esp,1),%ebx
     d:   b8 05 00 00 00          movl   $0x5,%eax
    12:   cd 80                   int    $0x80
    14:   5b                      popl   %ebx
    15:   3d 01 f0 ff ff          cmpl   $0xfffff001,%eax
    1a:   0f 83 fc ff ff ff       jae    1c <__libc_open+0x1c>
    20:   c3                      ret

How does the glibc build system generate the wrapper call to open()? It happens when the glibc-2.1/io directory is build. This is the output where it happens:

  make[1]: Entering directory /export/g/src/packages/SOURCES/glibc-2.1/io'
  (echo '#include <sysdep.h>'; \
   echo 'PSEUDO (__libc_open, open, 3)'; \
   echo ' ret'; \
   echo 'PSEUDO_END(__libc_open)'; \
   echo 'weak_alias (__libc_open, __open)'; \
   echo 'weak_alias (__libc_open, open)'; \
  ) | gcc -c  -I../include -I.  -I.. -I../libio  -I../sysdeps/i386/elf
  -I../crypt/sysdeps/unix -I../linuxthreads/
  sysdeps/unix/sysv/linux -I../linuxthreads/sysdeps/pthread
  -I../linuxthreads/sysdeps/unix/sysv -I../linuxthreads
  /sysdeps/unix -I../linuxthreads/sysdeps/i386/i686 -I../linuxthreads/sysdeps/i386
  -I../sysdeps/unix/sysv/linux/i
  386/i686 -I../sysdeps/unix/sysv/linux/i386 -I../sysdeps/unix/sysv/linux
  -I../sysdeps/gnu -I../sysdeps/unix/comm
  on -I../sysdeps/unix/mman -I../sysdeps/unix/inet -I../sysdeps/unix/sysv/i386
  -I../sysdeps/unix/sysv -I../sysdeps/unix/i386 -I../sysdeps/unix -I../sysdeps/posix
  -I../sysdeps/i386/i686 -I../sysdeps/i386/i486 -I../sysdeps/lib
  m-i387/i686 -I../sysdeps/i386/fpu -I../sysdeps/libm-i387 -I../sysdeps/i386
  -I../sysdeps/wordsize-32 -I../sysdeps/ieee754 -I../sysdeps/libm-ieee754
  -I../sysdeps/generic/elf -I../sysdeps/generic   -D_LIBC_REENTRANT
  -include ../include/libc-symbols.h     -DASSEMBLER  -DGAS_SYNTAX  -x assembler-with-cpp
  -o open.o -echo 'io/utime.o io/mkfifo.o io/stat.o io/fstat.o io/lstat.o
  io/mknod.o io/stat64.o io/fstat64.o io/lstat64.o io/xstat.o io/fxstat.o
  io/lxstat.o io/xmknod.o io/xstat64.o io/fxstat64.o io/lxstat64.o io/statfs.o io/fstatfs.o
   io/statfs64.o io/fstatfs64.o io/statvfs.o io/fstatvfs.o io/statvfs64.o
  io/fstatvfs64.o io/umask.o io/chmod.o io/fchmod.o io/mkdir.o io/open.o
  io/open64.o io/close.o io/read.o io/write.o io/lseek.o io/lseek64.o io/access.o
   io/euidaccess.o io/fcntl.o io/flock.o io/lockf.o io/lockf64.o io/dup.o io/dup2.o
  io/pipe.o io/creat.o io/creat64.o io/chdir.o io/fchdir.o io/getcwd.o
  io/getwd.o io/getdirname.o io/chown.o io/fchown.o io/lchown.o io/ttyname.o
  io/ttyname_r.o io/isatty.o io/link.o io/symlink.o io/readlink.o io/unlink.o
  io/rmdir.o io/ftw.o io/ftw64.o io/fts.o io/poll.o' > stamp.oT
  mv -f stamp.oT stamp.o

I do not understand the above, as I do not see where is the C source code for the call wrapper. Maybe one day I will understand the above.

But as a result of the above, we get open.o in libc.a, with the __libc_open entry there as an alias for 'open'. OK, now let me look more at the code generated in __libc_open.

Here it is again

  >objdump  --disassemble open.o
  open.o:     file format elf32-i386
  Disassembly of section .text:
  00000000 <__libc_open>:
     0:   53                      pushl  %ebx
     1:   8b 54 24 10             movl   0x10(%esp,1),%edx
     5:   8b 4c 24 0c             movl   0xc(%esp,1),%ecx
     9:   8b 5c 24 08             movl   0x8(%esp,1),%ebx
     d:   b8 05 00 00 00          movl   $0x5,%eax
    12:   cd 80                   int    $0x80
    14:   5b                      popl   %ebx
    15:   3d 01 f0 ff ff          cmpl   $0xfffff001,%eax
    1a:   0f 83 fc ff ff ff       jae    1c <__libc_open+0x1c>
    20:   c3                      ret

Notice that open() takes 3 arguments

    open (filename, posix_mode, prot)

Notice the asembler shows using registers eds, ecx, and ebx to pass the data, then it moves 5 to eax. What is 5? This got to be the number that kernel uses to identify which system call it is.

Actually this will end up as an index used by the interrupt handler to locate the system call. Lets look around.

  cd glibc-2.1
  >find . -name '*.h' | grep syscal
  ./include/syscall.h
  ./misc/syscall.h
  ./misc/syscall-list.h
  ./sysdeps/generic/sys/syscall.h
  ./sysdeps/mach/sys/syscall.h
  ./sysdeps/unix/sysv/linux/mips/sys/syscall.h
  ./sysdeps/unix/sysv/linux/sys/syscall.h
  ./sysdeps/unix/sysv/sco3.2.4/sys/syscall.h
  ./sysdeps/unix/sysv/sysv4/solaris2/sys/syscall.h
  ./INSTALL_LIB/usr/local/include/sys/syscall.h
  ./INSTALL_LIB/usr/local/include/bits/syscall.h
  ./INSTALL_LIB/usr/local/include/syscall.h
  
  >more ./include/syscall.h
  #include <misc/syscall.h>
  
  >more ./misc/syscall.h
  #include <sys/syscall.h>
  >

Ok, getting closer, lets look at /usr/include/sys/syscall.h

  
  >more /usr/include/sys/syscall.h
  /* Copyright (C) 1995, 1996, 1997 Free Software Foundation, Inc.
     This file is part of the GNU C Library.
  
     The GNU C Library is free software; you can redistribute it and/or
     modify it under the terms of the GNU Library General Public License as
     published by the Free Software Foundation; either version 2 of the
     License, or (at your option) any later version.
  
     The GNU C Library is distributed in the hope that it will be useful,
     but WITHOUT ANY WARRANTY; without even the implied warranty of
     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
     Library General Public License for more details.
  
     You should have received a copy of the GNU Library General Public
     License along with the GNU C Library; see the file COPYING.LIB.  If not,
     write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330,
     Boston, MA 02111-1307, USA.  */
  
  #ifndef _SYSCALL_H
  #define _SYSCALL_H      1
  
  /* This file should list the numbers of the system the system knows.
     But instead of duplicating this we use the information available
     from the kernel sources.  */
  #include <asm/unistd.h>
  
  #ifndef _LIBC
  /* The Linux kernel header file defines macros __NR_<name>', but some
     programs expect the traditional form SYS_<name>'.  So in building libc
     we scan the kernel's list and produce <bits/syscall.h> with macros for
     all the SYS_' names.  */
  # include <bits/syscall.h>
  #endif
  
  #endif

Ok, I am getting really close now.

  
  >more /usr/include/asm/unistd.h
  #ifndef _ASM_I386_UNISTD_H_
  #define _ASM_I386_UNISTD_H_
  
  /*
   * This file contains the system call numbers.
   */
  
  #define __NR_exit                 1
  #define __NR_fork                 2
  #define __NR_read                 3
  #define __NR_write                4
  #define __NR_open                 5  ------> HERE IT IS !!!
  ....
  

yahoo! found it. So, 5 is moved to register eax, and interrupt 0x80 is invoked.

When interrupt returns, system call is complete. It does not seem that the syscall macros defined in /usr/inlcude/asm/unistd.h are used in glibc?

OK, so far so good, now I'll switch hats, and jump into kernel mode to see how the open() call is processed. I need to find the code for that processes the interrupt 0x80.

The interrupt routine that is bound to interrupt 0x80 is found in

/usr/src/linux/arch/i386/kernel/entry.S

the entry point is called ENTRY(system_call).

Lets look at the code for the interrupt routine:

  ENTRY(system_call)
   pushl %eax # save orig_eax
   SAVE_ALL
   GET_CURRENT(%ebx)
   cmpl $(NR_syscalls),%eax    -----------> Notice, eax is where the system call number is saved.
   jae badsys
   testb $0x20,flags(%ebx) # PF_TRACESYS
   jne tracesys
   call *SYMBOL_NAME(sys_call_table)(,%eax,4)  -----> Here we index into the sys_call_table using the above number.
   movl %eax,EAX(%esp) # save the return value
  ENTRY(ret_from_sys_call)
  #ifdef __SMP__
   movl processor(%ebx),%eax
   shll $5,%eax
   movl SYMBOL_NAME(softirq_state)(,%eax),%ecx
   testl SYMBOL_NAME(softirq_state)+4(,%eax),%ecx
  #else
   movl SYMBOL_NAME(softirq_state),%ecx
   testl SYMBOL_NAME(softirq_state)+4,%ecx
  #endif
   jne   handle_softirq
  
  ret_with_reschedule:
   cmpl $0,need_resched(%ebx)
   jne reschedule
   cmpl $0,sigpending(%ebx)
   jne signal_return
  restore_all:
   RESTORE_ALL
  
   ALIGN
  signal_return:
   sti # we can get here from an interrupt handler
   testl $(VM_MASK),EFLAGS(%esp)
   movl %esp,%eax
   jne v86_signal_return
   xorl %edx,%edx
   call SYMBOL_NAME(do_signal)
   jmp restore_all
  
   ALIGN
  v86_signal_return:
   call SYMBOL_NAME(save_v86_state)
   movl %eax,%esp
   xorl %edx,%edx
   call SYMBOL_NAME(do_signal)
   jmp restore_all
  
   ALIGN
  tracesys:
   movl $-ENOSYS,EAX(%esp)
   call SYMBOL_NAME(syscall_trace)
   movl ORIG_EAX(%esp),%eax
   cmpl $(NR_syscalls),%eax
   jae tracesys_exit
   call *SYMBOL_NAME(sys_call_table)(,%eax,4)
   movl %eax,EAX(%esp) # save the return value
  tracesys_exit:
   call SYMBOL_NAME(syscall_trace)
   jmp ret_from_sys_call
  badsys:
   movl $-ENOSYS,EAX(%esp)
   jmp ret_from_sys_call
  
   ALIGN
  ret_from_exception:
  #ifdef __SMP__
   GET_CURRENT(%ebx)
   movl processor(%ebx),%eax
   shll $5,%eax
   movl SYMBOL_NAME(softirq_state)(,%eax),%ecx
   testl SYMBOL_NAME(softirq_state)+4(,%eax),%ecx
  #else
   movl SYMBOL_NAME(softirq_state),%ecx
   testl SYMBOL_NAME(softirq_state)+4,%ecx
  #endif
   jne   handle_softirq
  
  ENTRY(ret_from_intr)
   GET_CURRENT(%ebx)
   movl EFLAGS(%esp),%eax # mix EFLAGS and CS
   movb CS(%esp),%al
   testl $(VM_MASK | 3),%eax # return to VM86 mode or non-supervisor?
   jne ret_with_reschedule
   jmp restore_all
  
   ALIGN
  handle_softirq:
   call SYMBOL_NAME(do_softirq)
   jmp ret_from_intr
  
   ALIGN
  reschedule:
   call SYMBOL_NAME(schedule)    # test
   jmp ret_from_sys_call
  
  ENTRY(divide_error)
   pushl $0 # no error code
   pushl $ SYMBOL_NAME(do_divide_error)
   ALIGN
  error_code:
   pushl %ds
   pushl %eax
   xorl %eax,%eax
   pushl %ebp
   pushl %edi
   pushl %esi
   pushl %edx
   decl %eax # eax = -1
   pushl %ecx
   pushl %ebx
   cld
   movl %es,%ecx
   xchgl %eax, ORIG_EAX(%esp) # orig_eax (get the error code. )
   movl %esp,%edx
   xchgl %ecx, ES(%esp) # get the address and save es.
   pushl %eax # push the error code
   pushl %edx
   movl $(__KERNEL_DS),%edx
   movl %edx,%ds
   movl %edx,%es
   GET_CURRENT(%ebx)
   call *%ecx
   addl $8,%esp
   jmp ret_from_exception

The sys_call_table itself is located in .data segment in entry.S, this is the start of the table

  
  .data
  ENTRY(sys_call_table)
   .long SYMBOL_NAME(sys_ni_syscall) /* 0  -  old "setup()" system call*/
   .long SYMBOL_NAME(sys_exit)
   .long SYMBOL_NAME(sys_fork)
   .long SYMBOL_NAME(sys_read)
   .long SYMBOL_NAME(sys_write)
   .long SYMBOL_NAME(sys_open) /* 5 */
   .long SYMBOL_NAME(sys_mincore)
   .long SYMBOL_NAME(sys_madvise)
  
  .....
  
   /*
    * NOTE!! This doesn't have to be exact - we just have
    * to make sure we have _enough_ of the "sys_ni_syscall"
    * entries. Don't panic if you notice that this hasn't
    * been shrunk every time we add a new system call.
    */
   .rept NR_syscalls-219
   .long SYMBOL_NAME(sys_ni_syscall)
   .endr

Ok, lets follow the system call. I see from the dispatch table above, that the open() call is implemented in kernel using sys_open.

Where is sys_open() ? All the sys calls related to IO are locatd in linux/fs/. Looking at linux/fs/open.c, this is the sys_open function.

  
  asmlinkage long sys_open(const char * filename, int flags, int mode)
  {
   char * tmp;
   int fd, error;
  
  #if BITS_PER_LONG != 32
   flags |= O_LARGEFILE;
  #endif
   tmp = getname(filename);
   fd = PTR_ERR(tmp);
   if (!IS_ERR(tmp)) {
   fd = get_unused_fd();
   if (fd >= 0) {
   struct file * f;
   lock_kernel();
   f = filp_open(tmp, flags, mode);
   unlock_kernel();
   error = PTR_ERR(f);
   if (IS_ERR(f))
   goto out_error;
   fd_install(fd, f);
   }
  out:
   putname(tmp);
   }
   return fd;
  
  out_error:
   put_unused_fd(fd);
   fd = error;
   goto out;
  }

The function filp_open() is in the same above file as sys_open(). Here is the function

  struct file *filp_open(const char * filename, int flags, int mode)
  {
   int namei_flags, error;
   struct nameidata nd;
  
   namei_flags = flags;
   if ((namei_flags+1) & O_ACCMODE)
   namei_flags++;
   if (namei_flags & O_TRUNC)
   namei_flags |= 2;
  
   error = open_namei(filename, namei_flags, mode, &nd);
   if (!error)
   return dentry_open(nd.dentry, nd.mnt, flags);
  
   return ERR_PTR(error);
  }

Notice the call to open_namei(), this is the interface to the virtual file system. calls into VFS are named _namei (verify?).

open_namei() is defined in linux/fs/namei.c.

After some access checking, and pathname checking, and possibly allocating an inode, a kernel internal struct file is allocated for the file. The file struct contains a pointer to file_operations struct, which contains the address of functions to process operations on this filesystem, that must have been initialized when the file system was mounted.

  
  struct file {
  456         struct list_head        f_list;
  457         struct dentry           *f_dentry;
  458         struct vfsmount         *f_vfsmnt;
  459         struct file_operations  *f_op;
  460         atomic_t                f_count;
  461         unsigned int            f_flags;
  462         mode_t                  f_mode;
  463         loff_t                  f_pos;
  464         unsigned long           f_reada, f_ramax, f_raend, f_ralen, f_rawin;
  465         struct fown_struct      f_owner;
  466         unsigned int            f_uid, f_gid;
  467         int                     f_error;
  468
  469         unsigned long           f_version;
  470
  471         /* needed for tty driver, and maybe others */
  472         void                    *private_data;
  473 };
  
  struct file_operations {
  693         loff_t (*llseek) (struct file *, loff_t, int);
  694         ssize_t (*read) (struct file *, char *, size_t, loff_t *);
  695         ssize_t (*write) (struct file *, const char *, size_t, loff_t *);
  696         int (*readdir) (struct file *, void *, filldir_t);
  697         unsigned int (*poll) (struct file *, struct poll_table_struct *);
  698         int (*ioctl) (struct inode *, struct file *, unsigned int, unsigned long);
  699         int (*mmap) (struct file *, struct vm_area_struct *);
  700         int (*open) (struct inode *, struct file *);
  701         int (*flush) (struct file *);
  702         int (*release) (struct inode *, struct file *);
  703         int (*fsync) (struct file *, struct dentry *);
  704         int (*fasync) (int, struct file *, int);
  705         int (*lock) (struct file *, int, struct file_lock *);
  706         ssize_t (*readv) (struct file *, const struct iovec *, unsigned long, loff_t *);
  707         ssize_t (*writev) (struct file *, const struct iovec *, unsigned long, loff_t *);
  708 };

Ok, time to go sleep now.