StackMirror tool for fixing of stack corruption problems

19 July 2011

Debugging buffer overruns is hard. Usually you have to solve these bugs with logs + apply divide and conquer techniques. Even valgrind can’t help here ! GCC has the compiler options -fstack-protector-all , but it does not give accurate stack traces, it is more designed to shut down an application so as to avoid remote exploit of a buffer overflow error.

So a new open source tool is presented here that can help with most such cases; StackMirror to the rescue! It detects even the nastiest stack smashers and reports an accurate stack trace of the function that caused it.

First of all how to use it

  • The tool works with gcc (versions or upward)

  • The application must be re compiled with gcc compiler option -finstrument-functions

  • During runtime the application has to load library, so either LD_PRELOAD environment has to be set to the shared library, or application has to link with it.

  • Application should not be compiled with Optimization (-O1 -O2 -O3 ) and frame pointer omission (Please use both -fno-omit-frame-pointer and -On in order to prevent frame pointer omission).
  • I have tested the tool on both i686 and x86_64; please look if it works on other platforms too.

Download instructions

Get my stuff chdir stack make (for specific platform builds do make ARCH=x86_64 or make ARCH=i686 ) make install the preceeding line will install to /usr/local make install INSTALL_PREFIX=/opt/stack will install stuff to prefix /opt/stack

Step by step usage instruction

Prior to running your program, please set the following environment variable export LD_PRELOAD=/usr/local/lib/ Please note that value of LD_PRELOAD must be the full path name of

Run you program, if it smashes the stack and we are ok then an error message of the following form is displayed. Don’t worry; it will be formatted in the next step.

ERROR: Stack corruption detected
**** start stack ***
frame 0 ip: 0x80484e4
frame 0 ip: 0x80485c3
frame 1 ip: 0x804856f
frame 1 ip: 0x80485e7
frame 2 ip: 0x804856f
frame 2 ip: 0x80485e7
frame 3 ip: 0x804856f
frame 3 ip: 0x80485e7
frame 4 ip: 0x804856f
frame 4 ip: 0x80485e7
frame 5 ip: 0x804856f
frame 5 ip: 0x80485e7
frame 6 ip: 0x804856f
frame 6 ip: 0x80485e7
frame 7 ip: 0x804856f
frame 7 ip: 0x80485e7
frame 8 ip: 0x804856f
frame 8 ip: 0x80485e7
frame 9 ip: 0x804856f
frame 9 ip: 0x80485e7
frame 10 ip: 0x804856f
frame 10 ip: 0x804864e
frame 11 ip: 0x8048627
frame 11 ip: 0x80486c7
frame 12 ip: 0x8048670
frame 12 ip: 0x40049e37

08048000-08049000 00000000 /home/mmoser/archive/cstuff/GNU-Linux-2.6.38-8-generic-i686/test/tst
40000000-4001c000 00000000 /lib/i386-linux-gnu/
4001e000-4001f000 00000000 [vdso]
40021000-40023000 00000000 /home/mmoser/archive/cstuff/GNU-Linux-2.6.38-8-generic-i686/lib/
40033000-4018d000 00000000 /lib/i386-linux-gnu/
*** eof stack ***

Save this error message to a file crash.log and run /usr/local/bin/ crash.log

It should create the following message:

ERROR: Stack corruption detected
ip 0x80484e4 from /home/mmoser/archive/cstuff/GNU-Linux-2.6.38-8-generic-i686/test/tst

ip 0x80485c3 from /home/mmoser/archive/cstuff/GNU-Linux-2.6.38-8-generic-i686/test/tst

ip 0x804856f from /home/mmoser/archive/cstuff/GNU-Linux-2.6.38-8-generic-i686/test/tst

ip 0x80485e7 from /home/mmoser/archive/cstuff/GNU-Linux-2.6.38-8-generic-i686/test/tst

ip 0x804856f from /home/mmoser/archive/cstuff/GNU-Linux-2.6.38-8-generic-i686/test/tst

ip 0x80485e7 from /home/mmoser/archive/cstuff/GNU-Linux-2.6.38-8-generic-i686/test/tst

ip 0x804856f from /home/mmoser/archive/cstuff/GNU-Linux-2.6.38-8-generic-i686/test/tst

ip 0x80485e7 from /home/mmoser/archive/cstuff/GNU-Linux-2.6.38-8-generic-i686/test/tst

How does it work?

Compiler option -finstrument-function is usually used for building profilers; for each function the compiler inserts two function calls; on entering the function it calls _cyg_profile_func_enter on leaving the function it calls_cyg_profile_func_exit . The tool implements these two functions; on entering an application function the tool records the function return address, and when it is time to leave this function it checks if that address has been modified; if that is the case then we know that the stack was smashed here right by this function.

In more detail: For each application thread a stack data structure is maintained and referenced by a thread local storage entry. For each function call this stack keeps an entry; this entry contains the function return address, the function frame pointer address and the value received from dereferencing the frame pointer address. Typically the value obtained from dereferencing of frame pointer address is the return address from the called function. The stack entry is created when _cyg_profile_func_enter is called. When the function is left and_cyg_profile_func_exit is called, we check that the value obtained from dereferencing of the frame pointer did not change. If it actually did change, then we know for certain that the stack was overwritten, just right now here by this function. We also have the accurate application stack trace, just by looking at the entries in the stack data structures. So when an error occurs, the return address of all functions is printed to standard error, in addition to that, we print the addresses of the text segment for all shared libraries + executable. This information is then converted (offline after the crash) to symbolic names by the script ; we do not want to do this at crash time.

The tool keeps entries that shadow the application stack - hence the name StackMirror. Of course this shadow stack can also be overwritten by some pointer gone wild, but this is much less likely than occurrence of a buffer overflow error. Another limitation is that we only return address of the current function is checked. In some complex example an application function call could smash the return address of a calling function; we do not check for that here, (though an even more paranoid version of the tool could do that). We will still get there eventually, as the stack unwinds.

.h3 Other limitations

This library makes an effort to recover from situations with setjmp / longjmp trick; this function is a non local form of goto ; it undoes the stack up until some remembered levy. However there is one trick that we can’t cope with makecontext / setcontext If the application decides to do its own cooperative multitasking / user level threading or if it needs co procedures then all our efforts here are toast. Note that it’s becoming something of a habit here - for every trick find when it breaks ;-)