Debugging C/C++ code

Creation Date: 8 Feb 2012; Last Edit Date: 27 Jun 2014

The following applies to non-Windows operating systems. This is not for the faint of heart, and requires some C-level familiarity.

For those who learn better visually, see the video by a Bioconductor alum about using gdb to debug R packages with native code.

The benefit of the diagnose-a-crash and Case study examples is all the steps and logic are written out; one does not need to rewind a video to review the steps.

Set-up

The first, essential, step is to write a short script that reliably and quickly reproduces the error. Call this script buggy.R.

For debugging package code at the C/C++ level, one usually starts by installing the package without any compiler optimizations, e.g., by following

RShowDoc("R-admin")

section 6.3.3. setting for instance

CFLAGS=-ggdb -O0

in .R/Makevars. See the relevant Package Guidelines section for more examples and information.

Detecting memory errors (Valgrind)

Valgrind is a mature suite of tools for low-level program analysis. Valgrind's memory error checker (Memcheck) is the premier tool for diagnosing C/C++ memory errors.

Valgrind can be used to spot memory access problems, which are a common source of segfaults in C/C++ code. When the bug is isolated and easily produced in buggy.R, start R with:

R -d valgrind -f buggy.R

This runs quite slowly, and will flag invalid memory read and write locations. The former typically contribute to bad data, the latter to memory corruption and spectacular failures. The output will require C familiarity to interpret. It is helpful to run the buggy code with a package that has been installed without compiler optimizations. See section 4.3 of RShowDoc("R-exts") and the relevant Package Guidelines section.

Interactive debugging (gdb or lldb)

If you have never used a command-line debugger, there are many fine quick-start guides on the web; it is not as daunting as it seems.

On Linux the preferred debugger is gdb, but lldb is default for the Mavericks platform. The interfaces are similar, but if you are accustomed to gdb, see the gdb to lldb command map.

Start R with a C-level debugger such as gdb.

R -d gdb -f buggy.R

You'll end up at the gdb prompt

(gdb)

and a typical operation is to (r)un or (c)ontinue execution

(gdb) r

to run buggy.R. You'll end up back in C when there is a segfault, or you press cntrl-C (^C, below), or when you've inserted a (b)reakpoint at some C-level function that you suspect is buggy, e.g.,

> ^C
(gdb) b some_buggy_fun
(gdb) c

When you do end up back in the debugger, you can print C variables or the C representation of R variables (provided R isn't too confused by this point)

(gdb) p c_var
(gdb) call Rf_PrintValue(some_R_variable)

You can also view a (b)ack(t)race of the call stack, navigate (u)p and (d)own the call stack, etc. See

(gdb) help

and our mutual friend Google for additional information.

Finding clues for causes of a program crash

Perhaps the most useful function of debuggers is to provide a breadcrumb trail ("backtrace") of the routines that led to the crash of a buggy program. With this knowledge we can narrow our inquiry to code that affects parts of the program state relevant at the time of the crash.

It is worth reiterating that it is essential optimizations be turned off and the compiler be instructed to include debugging symbols if one hopes to have a fruitful debugging session. See the relevant Package Guidelines section.

Though the debugger output in the example may differ slightly compared to output from other computing environments, the underlying techniques apply for diagnosing program crashes on any platform. See the Case study for a real-world example that uses Valgrind and gdb in conjunction.

We will use a contrived example to demonstrate how to identify potential places in our code that cause a crash. You should be able to use the example files exactly as they appear. Some extraneous output has been omitted for brevity.

C++ file buggy.cpp:

#include <map>
#include <utility>
#ifdef __cplusplus
#define R_NO_REMAP
#endif
#include <Rdefines.h>

extern "C" SEXP buggy_function();

SEXP buggy_function() {
    std::map<int, int> m;
    m.insert(std::make_pair(5, 7));
    m.insert(std::make_pair(9, 42));
    std::map<int,int>::const_iterator it = m.begin();
    ++it;
    ++it;
    ++it;

    return R_NilValue;
}

Compile with R CMD SHLIB buggy.cpp -o buggy.so.

source()ing this file (buggy.R) within an R session (or entering the commands in an R session) will lead to a program crash:

dyn.load("buggy.so")
.Call("buggy_function")

Unfortunately R's diagnostics are not very illuminating:

> source("buggy.R")

 *** caught segfault ***
address 0x2, cause 'memory not mapped'

Traceback:
 1: .Call("buggy_function")
 2: eval(expr, envir, enclos)
 3: eval(ei, envir)
 4: withVisible(eval(ei, envir))
 5: source("buggy.R")

Now we turn to the debugger. Start R with the lldb debugger (or equivalent for your platform):

R -d lldb
(lldb) run
## R startup messages elided
## now in R session
> source("buggy.R")

At this point R crashes, lldb produces some output, and we are back at the lldb prompt. The lldb output looks like this (showing us the frame (#0) in the call stack where the crash occurred):

Process 21657 stopped
* thread #1: tid = 0xbcb4ab, 0x00000001028fcbb0 buggy.so`buggy_function [inlined] std::__1::__tree_node_base<void*>* std::__1::__tree_min<std::__1::__tree_node_base<void*>*>(std::__1::__tree_node_base<void*>*) at __tree:134, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x2)
    frame #0: 0x00000001028fcbb0 buggy.so`buggy_function [inlined] std::__1::__tree_node_base<void*>* std::__1::__tree_min<std::__1::__tree_node_base<void*>*>(std::__1::__tree_node_base<void*>*) at __tree:134
   131  _NodePtr
   132  __tree_min(_NodePtr __x) _NOEXCEPT
   133  {
-> 134      while (__x->__left_ != nullptr)
   135          __x = __x->__left_;
   136      return __x;
   137  }

It looks like the debugger is telling us there was a memory access error when fetching a tree node. (Trees are a common underlying data structure for the standard library map). The output is voluminous and looks confusing, but only the gist is important right now.

Still in the same lldb session, enter the bt command (for "backtrace") at the lldb prompt and we see all the stack frames (and function calls) that preceded the crash. Frames are listed in ascending order, starting with the frame in which crash happened. (Note frame #0 here is the same as frame #0 given above.) This means when diagnosing a crash, it usually makes sense to start with lower-numbered frames and proceed upwards.

(lldb) bt
* thread #1: tid = 0xbcb4ab, 0x00000001028fcbb0 buggy.so`buggy_function [inlined] std::__1::__tree_node_base<void*>* std::__1::__tree_min<std::__1::__tree_node_base<void*>*>(std::__1::__tree_node_base<void*>*) at __tree:134, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x2)
  * frame #0: 0x00000001028fcbb0 buggy.so`buggy_function [inlined] std::__1::__tree_node_base<void*>* std::__1::__tree_min<std::__1::__tree_node_base<void*>*>(std::__1::__tree_node_base<void*>*) at __tree:134
    frame #1: 0x00000001028fcbb0 buggy.so`buggy_function [inlined] std::__1::__tree_node_base<void*>* std::__1::__tree_next<std::__1::__tree_node_base<void*>*>(std::__1::__tree_node_base<void*>*) + 20 at __tree:158
    frame #2: 0x00000001028fcb9c buggy.so`buggy_function [inlined] std::__1::__tree_const_iterator<std::__1::__value_type<int, int>, std::__1::__tree_node<std::__1::__value_type<int, int>, void*>*, long>::operator++() at __tree:747
    frame #3: 0x00000001028fcb9c buggy.so`buggy_function [inlined] std::__1::__map_const_iterator<std::__1::__tree_const_iterator<std::__1::__value_type<int, int>, std::__1::__tree_node<std::__1::__value_type<int, int>, void*>*, long> >::operator++() at map:750
    frame #4: 0x00000001028fcb9c buggy.so`buggy_function + 188 at buggy.cpp:17
    frame #5: 0x0000000100073a13 libR.dylib`do_dotcall(call=<unavailable>, op=<unavailable>, args=<unavailable>, env=<unavailable>) + 323 at dotcode.c:578

frame #5 mentions do_dotcall, which is the native function (in the R library) that corresponds to the .Call("buggy_function") line in buggy.R where we call our C entry point. We can reasonably conclude the useful information for our bug is probably in frames #0-4.

Here is a possible chain of thought that leads to the right conclusion:

  1. frames #0-2 look like they are dealing with tree/map internals; ignore for the moment.

  2. frame #3 indicates we are probably talking about our map const_iterator variable declared at line #14 in buggy.cpp (std::map<int,int>::const_iterator it = m.begin();).

  3. frame #4 is the key: it tells us line (#17) in the buggy.cpp file (++it;) is where execution went from C++ code we wrote into the map iterator internals that produced the error.

  4. Eureka! By carefully reading the code in buggy.cpp we realize that after the inserts the size of map m is 2. That means after incrementing iterator it at line #16 (++it;), the value of it is the special past-the-end value. Incrementing an iterator beyond past-the-end (the third ++it; at line #17) is undefined behavior!

If we revise buggy.cpp to not increment it beyond past-the-end by removing the third ++it; the program runs without complaint. Problem solved!

As you can see, the debugger was not able to immediately tell us why the program crashed, just where the program crashed. We used the information about where the crash happened to home in on the parts of our code that affected program state relevant at the time of the crash. Obviously this example is contrived; in a real-world scenario the extra help afforded by the insight about relevant program state is invaluable.

Case study

As a case study, a colleague reported that their complicated program would, on one particular computer, produce a segmentation fault or just stop responding. The same series of actions wouldn't cause problems on other computers. This sounds like a classic memory problem, with the segfault and difficulty of reproduction.

The first advice was to develop a simple script that reproduced the problem: the original report had too many moving parts. A big insight was that the bug could be produced by running part of the code that uses RCurl, followed by a call to the garbage collector, gc(). The role of the garbage collector suggests again memory corruption of some sort, and in particular that perhaps RCurl is allocating (at the C level) an R object but not properly PROTECT'ing it from garbage collection. We suspect RCurl rather than R or libcurl (other possible players) because it is the least tested of the code. We could be wrong, of course... After many iterations, my colleague arrived at buggy24.R:

library(RCurl)

foo <- function() {
    url <- "https://google.com"
    curl <- getCurlHandle()
    opts <- list(followlocation=NULL, ssl.verifypeer=TRUE)
    d <- debugGatherer()
    getURL(url,customrequest="GET",curl=curl,debugfunction=d$update,.opts=opts)
}

execute <- function() {
    foo()
    gc()
}

execute()

This is pretty simple, and doesn't require access to any special resources (like the server that was originally being queried). This script doesn't cause a segfault when run on all systems, but running valgrind (having installed RCurl without any optimizations) shows...

> R -d valgrind -f buggy24.R
...
==10859== Conditional jump or move depends on uninitialised value(s)
==10859==    at 0x11BF00F6: getCurlPointerForData (curl.c:798)
==10859==    by 0x11BF0E80: R_curl_easy_setopt (curl.c:164)
==10859==    by 0x11BF17AD: R_curl_easy_perform (curl.c:89)
==10859==    by 0x4ED5499: do_dotcall (dotcode.c:588)
==10859==    by 0x4F1CAA4: Rf_eval (eval.c:593)
==10859==    by 0x4F2BD5C: do_set (eval.c:1828)
==10859==    by 0x4F1C8B7: Rf_eval (eval.c:567)
==10859==    by 0x4F2B957: do_begin (eval.c:1514)
==10859==    by 0x4F1C8B7: Rf_eval (eval.c:567)
==10859==    by 0x4F297E9: Rf_applyClosure (eval.c:960)
==10859==    by 0x4F1CBA5: Rf_eval (eval.c:611)
==10859==    by 0x4F2BD5C: do_set (eval.c:1828)

Look around the C source code in RCurl's curl.c, as suggested by the backtrace, just to get oriented. Then do

R -d gdb -f buggy24.R

to run the script under gdb. Run our test script

(gdb) r

No error. Don't give up, set a break point

(gdb) b curl.c:798

and run again

(gdb) r
Breakpoint 1, getCurlPointerForData (el=0x79e038,
    option=CURLOPT_WRITEFUNCTION, isProtected=FALSE, curl=0x1d9bdc0)
    at curl.c:798
798    curl.c: No such file or directory.
(gdb)

That 'no such file' means that gdb doesn't know where to find the RCurl package src/ directory, so tell it and (l)ist the context, and (p)rint the value of the C variable isProtected, which seems to be the source of the valgrind warning

(gdb) dir ~/tmp/RCurl/src
(gdb) l
793                        }
794                    }
795                }
796                break;
797              case CLOSXP:
798              (gdb) l
793                        }
794                    }
795                }
796                break;
797              case CLOSXP:
798                  if(!isProtected) {
799                R_PreserveObject(el);
800                }
801                ptr = (void *) el;
802                break;
(gdb) p isProtected
$5 = FALSE

isProtected has a value (it has to!), and furthermore the value of FALSE results in PROTECT'ing the object el across C calls (this is what R_PreserveObject does). This is pretty interesting, because we're aware that garbage collection triggers the segfault. valgrind is telling us that the value of isProtected isn't actually the result of an assignment, it could be the result of accessing an array out of bounds. Let's head up the call stack and see where this value is coming from

(gdb) up
#1  0x00007ffff426e273 in R_curl_easy_setopt (handle=0x15d9600,
    values=0x1445788, opts=0xf3d418, isProtected=0xb7d308, encoding=0x776db0)
    at curl.c:164
164            val = getCurlPointerForData(el, opt, LOGICAL(isProtected)[ i % n ], obj);
(gdb) l
159        /* Loop over all the options we are setting. */
160        for(i = 0; i < n; i++) {
161            opt = INTEGER(opts)[i];
162            el = VECTOR_ELT(values, i);
163                 /* Turn the R value into something we can use in libcurl. */
164            val = getCurlPointerForData(el, opt, LOGICAL(isProtected)[ i % n ], obj);
165   
166                    if(opt == CURLOPT_WRITEFUNCTION && TYPEOF(el) == CLOSXP) {
167                data->fun = val; useData++;
168                status =  curl_easy_setopt(obj, CURLOPT_WRITEFUNCTION, &R_curl_write_data);
(gdb)

We're entering the function getCurlPointerForData with the value LOGICAL(isProtected)[ i % n ]. Here, isProtected is now an R object, not a C variable. Looking at the surrounding code, that i % n doesn't look right -- it's probably meant to recycle isProtected in the case where a shorter logical variable is provided than the vector of elements requiring protection, but the value of n is not necessarily the length of isProtected. Let's have a look at what we've got, using a C-level R function Rf_PrintValue to print R values (SEXP's) in an R fashion

(gdb) p isProtected
$1 = (SEXP) 0xaad8a0
(gdb) call Rf_PrintValue(isProtected )
[1] FALSE

isProtected is a logical vector of length 1.

(gdb) p i
$7 = 1
(gdb) p n
$8 = 6
(gdb) p i % n
$9 = 1

...and we're trying to access element 1 of it. But the C representation of R vectors is zero-based, so the only valid value of the index is 0 -- we're out of bounds! This could well be our bug, and it's time to try fixing it (naively, LOGICAL(isProtected)[ i % LENGTH(isProtected) ]) to confirm our diagnosis, or report to the packageDescription("RCurl")$Maintainer who might have a better sense of the overall structure and intention of the code.

Fred Hutchinson Cancer Research Center