Re: [RFC]confusion about syscall

Peter Teoh <htmldeveloper@xxxxxxxxx> · Sun, 15 Jul 2012 23:33:04 +0800

I think this is useful:
http://stackoverflow.com/questions/9355097/looking-for-system-calls-implementation-on-linux-kernel

On Sun, Jul 15, 2012 at 11:24 PM, Peter Teoh <htmldeveloper@xxxxxxxxx> wrote:

just sharing my analysis, correct me if wrong:

On Sun, Jul 15, 2012 at 8:36 PM, 王哲 <wangzhe5004@xxxxxxxxx> wrote:

2012/7/15 Peter Teoh <htmldeveloper@xxxxxxxxx>

Hi Mulyadi and WangZhe,
Nice to write to you again....:-).

On Sun, Jul 15, 2012 at 1:49 PM, Mulyadi Santosa <mulyadi.santosa@xxxxxxxxx> wrote:

Hi...

On Sun, Jul 15, 2012 at 9:28 AM, 王哲 <wangzhe5004@xxxxxxxxx> wrote:

> and the second program:

>

> #include <stdio.h>

> #include <unistd.h>

>

> int main(void)

> {

>     unsigned long value = 0;

>     value = getpid();

>     return 0;

> }

>

> and disassembling it:( objdump -d a.out)

> ...

> 08048300 <getpid@plt>:

>  8048300:    ff 25 00 a0 04 08        jmp    *0x804a000

>  8048306:    68 00 00 00 00           push   $0x0

>  804830b:    e9 e0 ff ff ff           jmp    80482f0 <_init+0x3c>

Looks like jumping into vsyscall page to me...

after I start the process, and doing a gdb -p <pid>:

(gdb) disassemble main 
Dump of assembler code for function main:

   0x0000000000400564 <+0>:	push   %rbp
   0x0000000000400565 <+1>:	mov    %rsp,%rbp
   0x0000000000400568 <+4>:	sub    $0x10,%rsp
   0x000000000040056c <+8>:	movq   $0x0,-0x8(%rbp)
   0x0000000000400574 <+16>:	mov    $0x0,%eax
   0x0000000000400579 <+21>:	callq  0x400460 <getpid@plt>

   0x000000000040057e <+26>:	cltq   
   0x0000000000400580 <+28>:	mov    %rax,-0x8(%rbp)
   0x0000000000400584 <+32>:	movabs $0x9184e72a000,%rdi
   0x000000000040058e <+42>:	mov    $0x0,%eax
   0x0000000000400593 <+47>:	callq  0x400470 <sleep@plt>
   0x0000000000400598 <+52>:	mov    $0x0,%eax

   0x000000000040059d <+57>:	leaveq 
   0x000000000040059e <+58>:	retq   
End of assembler dump.
(gdb) disassemble getpid
Dump of assembler code for function getpid:
   0x00007f19ae558530 <+0>:	mov    %fs:0x2d4,%edx

   0x00007f19ae558538 <+8>:	cmp    $0x0,%edx
   0x00007f19ae55853b <+11>:	jle    0x7f19ae558540 <getpid+16>

   0x00007f19ae55853d <+13>:	mov    %edx,%eax
   0x00007f19ae55853f <+15>:	retq   
   0x00007f19ae558540 <+16>:	jne    0x7f19ae558554 <getpid+36>
   0x00007f19ae558542 <+18>:	mov    %fs:0x2d0,%eax

   0x00007f19ae55854a <+26>:	test   %eax,%eax
   0x00007f19ae55854c <+28>:	nopl   0x0(%rax)
   0x00007f19ae558550 <+32>:	je     0x7f19ae558554 <getpid+36>
   0x00007f19ae558552 <+34>:	repz retq 

   0x00007f19ae558554 <+36>:	mov    $0x27,%eax
   0x00007f19ae558559 <+41>:	syscall 
   0x00007f19ae55855b <+43>:	test   %edx,%edx
   0x7f19ae55855d <getpid+45>:	jne    0x7f19ae558552 <getpid+34>

   0x7f19ae55855f <getpid+47>:	mov    %eax,%fs:0x2d0
   0x7f19ae558567 <getpid+55>:	retq   

   Hi peter:
       question1: why your system is "0x00007f19ae558554 <+36>:	mov    $0x27,%eax",

getpid syscall  number is 0x14

yes u are right - for 32-bit kernel:

In arch/x86/kernel>
grep getpid *.S
syscall_table_32.S:	.long sys_getpid	/* 20 */

but my linux kernel is 64-bit. 

       question2: i use gdb disassemble getpid just like you and the result:

    (gdb) disassemble getpid
 Dump of assembler code for function getpid:
   0xb7771a40 <+0>:    mov    %gs:0x6c,%edx

   0xb7771a47 <+7>:    cmp    $0x0,%edx
   0xb7771a4a <+10>:    jle    0xb7771a50 <getpid+16>
   0xb7771a4c <+12>:    mov    %edx,%eax
   0xb7771a4e <+14>:    repz ret 
   0xb7771a50 <+16>:    jne    0xb7771a62 <getpid+34>

   0xb7771a52 <+18>:    mov    %gs:0x68,%eax
   0xb7771a58 <+24>:    test   %eax,%eax
   0xb7771a5a <+26>:    lea    0x0(%esi),%esi
   0xb7771a60 <+32>:    jne    0xb7771a4e <getpid+14>

   0xb7771a62 <+34>:    mov    $0x14,%eax
   0xb7771a67 <+39>:    call   *%gs:0x10

See the comment for gs in entry_32.S:

/*
 * User gs save/restore
 *
 * %gs is used for userland TLS and kernel only uses it for stack
 * canary which is required to be at %gs:20 by gcc.  Read the comment

 * at the top of stackprotector.h for more info.
 *
 * Local labels 98 and 99 are used.
 */
#ifdef CONFIG_X86_32_LAZY_GS

And inside stackprotector.h, content of which is still beyond my completely understanding at the moment, I copied it here:

/*
 * GCC stack protector support.
 *
 * Stack protector works by putting predefined pattern at the start of
 * the stack frame and verifying that it hasn't been overwritten when

 * returning from the function.  The pattern is called stack canary
 * and unfortunately gcc requires it to be at a fixed offset from %gs.
 * On x86_64, the offset is 40 bytes and on x86_32 20 bytes.  x86_64

 * and x86_32 use segment registers differently and thus handles this
 * requirement differently.
 *
 * On x86_64, %gs is shared by percpu area and stack canary.  All
 * percpu symbols are zero based and %gs points to the base of percpu

 * area.  The first occupant of the percpu area is always
 * irq_stack_union which contains stack_canary at offset 40.  Userland
 * %gs is always saved and restored on kernel entry and exit using

 * swapgs, so stack protector doesn't add any complexity there.
 *
 * On x86_32, it's slightly more complicated.  As in x86_64, %gs is
 * used for userland TLS.  Unfortunately, some processors are much

 * slower at loading segment registers with different value when
 * entering and leaving the kernel, so the kernel uses %fs for percpu
 * area and manages %gs lazily so that %gs is switched only when

 * necessary, usually during task switch.
 *
 * As gcc requires the stack canary at %gs:20, %gs can't be managed
 * lazily if stack protector is enabled, so the kernel saves and

 * restores userland %gs on kernel entry and exit.  This behavior is
* controlled by CONFIG_X86_32_LAZY_GS and accessors are defined in
 * system.h to hide the details.
 */

Yes, gs register is valid for userspace TLS and thus is per-process, and for more info:

http://www.akkadia.org/drepper/tls.pdf

http://www.ibm.com/developerworks/linux/library/l-user-space-apps/index.html

http://stackoverflow.com/questions/6021273/how-to-allocate-thread-local-storage

(and lots of relevant links besides it).

  can you explain the meaning of "call   *%gs:0x10"?

  Thanks! 

And to check the address space:

(gdb) info sharedlibrary 
From                To                  Syms Read   Shared Object Library

0x00007f19ae4cb8c0  0x00007f19ae5dec60  Yes (*)     /lib/libc.so.6

0x00007f19ae830af0  0x00007f19ae849704  Yes (*)     /lib64/ld-linux-x86-64.so.2
(*): Shared library is missing debugging information.

and if u want:

cat /proc/2282/maps 

7f19ae82a000-7f19ae82b000 rw-p 0017d000 08:05 9922                       /lib/libc-2.11.1.so
7f19ae830000-7f19ae850000 r-xp 00000000 08:05 8824                       /lib/ld-2.11.1.so

7ffff2031000-7ffff2052000 rw-p 00000000 00:00 0                          [stack]
7ffff21af000-7ffff21b0000 r-xp 00000000 00:00 0                          [vdso]
ffffffffff600000-ffffffffff601000 r-xp 00000000 00:00 0                  [vsyscall]

noticed also that static analysis tools like "objdump -d" is generally avoided, if u want to understand dynamic addresses.   From above, we can conclude that the "sysenter" (this is intel syntax, or "syscall", in AMD syntax as used by gdb disassembly above) is used for the transition to the kernel - as embedded inside the libc.so.6.

--

regards,

Mulyadi Santosa

Freelance Linux trainer and consultant

blog: the-hydra.blogspot.com

training: mulyaditraining.blogspot.com

_______________________________________________

Kernelnewbies mailing list

Kernelnewbies@xxxxxxxxxxxxxxxxx

http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies

-- 
Regards,
Peter Teoh

-- 
Regards,
Peter Teoh

-- 
Regards,
Peter Teoh

_______________________________________________
Kernelnewbies mailing list
Kernelnewbies@xxxxxxxxxxxxxxxxx
http://lists.kernelnewbies.org/mailman/listinfo/kernelnewbies