Stack Backtracing Inside Your Program
At this point, we have in hand a tool that is able to print the list of function calls up to the current execution point. This can be a useful tool in many different contexts. Think of having a complex program and needing to know who's calling a given function with the wrong parameters. With a simple check and a call to our show_stackframe() function, the faulty caller can be spotted easily.
An even more useful application for this technique is putting a stack backtrace inside a signal handler and having the latter catch all the "bad" signals your program can receive (SIGSEGV, SIGBUS, SIGILL, SIGFPE and the like). This way, if your program unfortunately crashes and you were not running it with a debugger, you can get a stack trace and know where the fault happened. This technique also can be used to understand where your program is looping in case it stops responding. All you need to do is set up a SIGUSR1/2 handler and send such a signal when needed. Before presenting an example, we need to open a parenthesis on signal handling.
Backtracing from within a signal handler requires some interesting intricacies that take us on a little detour through signal delivery to processes. Going into deep detail on this matter is outside the scope of this article, but we briefly can summarize it this way:
When the kernel needs to notify a signal of a given process, it prepares some data structures attached to the process' task struct and sets a signal-pending bit.
Later on, when the signalee process is scheduled for execution, its stack frame is altered by the kernel in order to have EIP point to the process' signal handler. This way, when the process runs it behaves as if it had called its own signal handler by itself before being suspended.
The initial steps of user space signal management are taken care of inside libc, which eventually calls the real process' signal handling routines which, in turn, execute our stack backtrace function.
As a consequence of this mechanism, the first two entries in the stack frame chain when you get into the signal handler contain, respectively, a return address inside your signal handler and one inside sigaction() in libc. The stack frame of the last function called before the signal (which, in case of fault signals, also is the one that supposedly caused the problem) is lost. Thus, if function B called function A, which in turn caused a SIGSEGV, a plain backtrace would list these entry points:
your_sig_handler() sigaction() in libc.so func_B() main()
and no trace of the call to function A would be found. For more details, have a look at the manuals for signal() and sigaction().
In order to get a meaningful backtrace, we need a workaround. Luckily, when you have the sources of both the kernel and libc, you can find a workaround for nearly anything. In Listing 2 we exploit an undocumented parameter of type sigcontext that is passed to the signal handler (see the UNDOCUMENTED section in man sigaction) and contains, among other things, the value of EIP when the signal was raised. After the call to backtrace(), we use this value to overwrite the useless entry corresponding to the sigaction() return address in the trace array. When we later call backtrace_symbols(), the address we inserted is resolved the same as any other entry in the array. Finally, when we print the backtrace, we start from the second entry (i=1 in the loop), because the first one always would be inside our signal handler.
Since kernel version 2.2 the undocumented parameter to the signal handler has been declared obsolete in adherence with POSIX.1b. A more correct way to retrieve additional information is to use the SA_SIGINFO option when setting the handler, as shown in Listing 3 and documented in the man page. Unfortunately, the siginfo_t structure provided to the handler does not contain the EIP value we need, so we are forced to resort again to an undocumented feature: the third parameter to the signal handler. No man page is going to tell you that such a parameter points to an ucontext_t structure that contains the values of the CPU registers when the signal was raised. From this structure, we are able to extract the value of EIP and proceed as in the previous case.
- Download "Linux Management with Red Hat Satellite: Measuring Business Impact and ROI"
- Use Linux as a SAN Provider
- Numerical Python
- diff -u: What's New in Kernel Development
- NSA: Linux Journal is an "extremist forum" and its readers get flagged for extra surveillance
- RSS Feeds
- Linux Systems Administrator
- Tech Tip: Really Simple HTTP Server with Python
- Senior Perl Developer
- Technical Support Rep