Unexpected valgrind "invalid write" when fstack-check is used

Question

I have the following very simple C program (test.c):

int f(int i)
{
  return i;
}

int main(int argC, char* argV[])
{
  int x = f(12);
  return 1;
}

I have done some tests with valgrind, playing with the -fstack-check compiler flag in two different systems.

Test done at Debian 8.6 with gcc 4.9.2 and valgrind 3.10.0.

Without -fstack-check:

$ gcc test.c -o test
$ valgrind ./test
==103703== Memcheck, a memory error detector
==103703== Copyright (C) 2002-2013, and GNU GPL'd, by Julian Seward et al.
==103703== Using Valgrind-3.10.0 and LibVEX; rerun with -h for copyright info
==103703== Command: ./test
==103703== 
==103703== 
==103703== HEAP SUMMARY:
==103703==     in use at exit: 0 bytes in 0 blocks
==103703==   total heap usage: 0 allocs, 0 frees, 0 bytes allocated
==103703== 
==103703== All heap blocks were freed -- no leaks are possible
==103703== 
==103703== For counts of detected and suppressed errors, rerun with: -v
==103703== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 1 from 1)

With -fstack-check:

$ gcc -fstack-check test.c -o test
$ valgrind ./test
==103726== Memcheck, a memory error detector
==103726== Copyright (C) 2002-2013, and GNU GPL'd, by Julian Seward et al.
==103726== Using Valgrind-3.10.0 and LibVEX; rerun with -h for copyright info
==103726== Command: ./test
==103726== 
==103726== 
==103726== HEAP SUMMARY:
==103726==     in use at exit: 0 bytes in 0 blocks
==103726==   total heap usage: 0 allocs, 0 frees, 0 bytes allocated
==103726== 
==103726== All heap blocks were freed -- no leaks are possible
==103726== 
==103726== For counts of detected and suppressed errors, rerun with: -v
==103726== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 1 from 1)

Test done in CentOS 6.6 with gcc 4.4.7 and valgrind 3.8.1:

Without -fstack-check:

$ gcc test.c -o test
$ valgrind ./test
==16390== Memcheck, a memory error detector
==16390== Copyright (C) 2002-2012, and GNU GPL'd, by Julian Seward et al.
==16390== Using Valgrind-3.8.1 and LibVEX; rerun with -h for copyright info
==16390== Command: ./test
==16390== 
==16390== 
==16390== HEAP SUMMARY:
==16390==     in use at exit: 0 bytes in 0 blocks
==16390==   total heap usage: 0 allocs, 0 frees, 0 bytes allocated
==16390== 
==16390== All heap blocks were freed -- no leaks are possible
==16390== 
==16390== For counts of detected and suppressed errors, rerun with: -v
==16390== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 6 from 6)

With -fstack-check:

$ gcc -fstack-check test.c -o test
$ valgrind ./test
==16441== Memcheck, a memory error detector
==16441== Copyright (C) 2002-2012, and GNU GPL'd, by Julian Seward et al.
==16441== Using Valgrind-3.8.1 and LibVEX; rerun with -h for copyright info
==16441== Command: ./test
==16441== 
==16441== Invalid write of size 8
==16441==    at 0x400497: main (in /home/fermin/valgrindtest/test)
==16441==  Address 0x7feffd058 is not stack'd, malloc'd or (recently) free'd
==16441== 
==16441== 
==16441== Process terminating with default action of signal 11 (SIGSEGV)
==16441==  Access not within mapped region at address 0x7FEFFD058
==16441==    at 0x400497: main (in /home/fermin/valgrindtest/test)
==16441==  If you believe this happened as a result of a stack
==16441==  overflow in your program's main thread (unlikely but
==16441==  possible), you can try to increase the size of the
==16441==  main thread stack using the --main-stacksize= flag.
==16441==  The main thread stack size used in this run was 10485760.
==16441== 
==16441== HEAP SUMMARY:
==16441==     in use at exit: 0 bytes in 0 blocks
==16441==   total heap usage: 0 allocs, 0 frees, 0 bytes allocated
==16441== 
==16441== All heap blocks were freed -- no leaks are possible
==16441== 
==16441== For counts of detected and suppressed errors, rerun with: -v
==16441== ERROR SUMMARY: 1 errors from 1 contexts (suppressed: 6 from 6)

In this case, in addition, the program ends with segmentation fault.

In summary, while in Debian all seems to be ok with -fstack-check, in CentOS I'm getting an "Invalid write" error that I'm not able to explain given the code of my program.

Reading about the -fstack-check flag maybe the error trace is due to the stack check mechanism is allocating 8 bytes in heap space for the x variable but valgrind (due to some reason) is not able to get aware of that, so it marks it as an invalid write. Maybe this is a limitation in valgrind 3.8.1 that has been solved in 3.10.0 (given the fact that in the Debian setup, using 3.10.0, I'm not getting that trace)?

In addition, I have seem at valgring FAQ that "-fomit-frame-pointer and -fstack-check can make stack traces worse" but they don't provide too much information about the reason for this so I cannot know if the problem is related...

Any help or hint which helps to explain this is highly appreciated.

EDIT: as suggested in one comment, I'm including the dissasembled code for the CentOS 6.x case, compiling with -g3 and using objdump --source test. I'm including here the part corresponding to main() and f() functions, however the full dump can be found here.

Binary compiled with -fstack-check:

0000000000400474 <f>:
int f(int i)
{
  400474:   55                      push   %rbp
  400475:   48 89 e5                mov    %rsp,%rbp
  400478:   89 7d fc                mov    %edi,-0x4(%rbp)
  return i;
  40047b:   8b 45 fc                mov    -0x4(%rbp),%eax
}
  40047e:   c9                      leaveq 
  40047f:   c3                      retq   

0000000000400480 <main>:

int main(int argC, char* argV[])
{
  400480:   55                      push   %rbp
  400481:   48 89 e5                mov    %rsp,%rbp
  400484:   48 83 ec 20             sub    $0x20,%rsp
  400488:   89 7d ec                mov    %edi,-0x14(%rbp)
  40048b:   48 89 75 e0             mov    %rsi,-0x20(%rbp)
  int x = f(12);
  return 0;
  40048f:   48 8d 84 24 08 d0 ff    lea    -0x2ff8(%rsp),%rax
  400496:   ff 
  400497:   48 c7 00 00 00 00 00    movq   $0x0,(%rax)
  return i;
}

int main(int argC, char* argV[])
{
  int x = f(12);
  40049e:   bf 0c 00 00 00          mov    $0xc,%edi
  4004a3:   e8 cc ff ff ff          callq  400474 <f>
  4004a8:   89 45 fc                mov    %eax,-0x4(%rbp)
  return 0;
  4004ab:   b8 00 00 00 00          mov    $0x0,%eax
}
  4004b0:   c9                      leaveq 
  4004b1:   c3                      retq 

Binary compiled without -fstack-check:

0000000000400474 <f>:
int f(int i)
{
  400474:   55                      push   %rbp
  400475:   48 89 e5                mov    %rsp,%rbp
  400478:   89 7d fc                mov    %edi,-0x4(%rbp)
  return i;
  40047b:   8b 45 fc                mov    -0x4(%rbp),%eax
}
  40047e:   c9                      leaveq 
  40047f:   c3                      retq   

0000000000400480 <main>:

int main(int argC, char* argV[])
{
  400480:   55                      push   %rbp
  400481:   48 89 e5                mov    %rsp,%rbp
  400484:   48 83 ec 20             sub    $0x20,%rsp
  400488:   89 7d ec                mov    %edi,-0x14(%rbp)
  40048b:   48 89 75 e0             mov    %rsi,-0x20(%rbp)
  int x = f(12);
  40048f:   bf 0c 00 00 00          mov    $0xc,%edi
  400494:   e8 db ff ff ff          callq  400474 <f>
  400499:   89 45 fc                mov    %eax,-0x4(%rbp)
  return 0;
  40049c:   b8 00 00 00 00          mov    $0x0,%eax
}
  4004a1:   c9                      leaveq 
  4004a2:   c3                      retq   

EDIT2: I have tested with newest valgrind version at the present moment (3.13.0) in CentOS 6.8 and I get the same problem.

Answer 1

I have installed gcc 4.7.2 in my CentOS 6.6 system (using this procedure) and re-done the test compiling with -fstack-check:

$ /opt/centos/devtoolset-1.1/root/usr/bin/gcc --version
gcc (GCC) 4.7.2 20121015 (Red Hat 4.7.2-5)
Copyright (C) 2012 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

$ /opt/centos/devtoolset-1.1/root/usr/bin/gcc -fstack-check test.c -o test
$ valgrind ./test
==19374== Memcheck, a memory error detector
==19374== Copyright (C) 2002-2012, and GNU GPL'd, by Julian Seward et al.
==19374== Using Valgrind-3.8.1 and LibVEX; rerun with -h for copyright info
==19374== Command: ./test
==19374== 
==19374== 
==19374== HEAP SUMMARY:
==19374==     in use at exit: 0 bytes in 0 blocks
==19374==   total heap usage: 0 allocs, 0 frees, 0 bytes allocated
==19374== 
==19374== All heap blocks were freed -- no leaks are possible
==19374== 
==19374== For counts of detected and suppressed errors, rerun with: -v
==19374== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 6 from 6)

Note that now the "invalid write" error is not shown. Thus, it seems that is a problem in the gcc compiler, some kind of bug fixed between 4.4.7 and 4.7.2, as @n.m. suggested in the comments.