I am currently debugging a Kernel module and to this purpose, I built the whole kernel with debug information (produces kallsyms, etc ...).
When I try nm my_module.ko, I get the list of symbols included by my module. All is allright except that some symbols are kind of missing as they do not appear in the symbol list. My feeling about this is that the related functions are being automatically inlined.
Anyway, when running the kernel with qemu-kgdb/gdb, I am able to see that the "missing" function is called. This means the compiler did not wipe it out because it was never used in any code path (hence my "feeling").
Since the symbol does not appear, I can't set a breakpoint on it and gdb won't unroll it so that I can see the running code path - understand I don't know how to tell gdb to unroll it.
Unfortunately, I want to see this part of the code path ... How can I do so ?
EDIT : As suggested in Tom's answer, I tried using the file:line syntax as below :
My code file looks like this :
int foo(int arg) // The function that I suspect to be inlined - here is line 1
{
/* Blabla */
return 42;
}
void foo2(void)
{
foo(0); // Line 9
}
I tried b file.c:1, and the breakpoint was hit but the foo() function is not unrolled.
Of course, I am producing debug symbols, since I also set a breakpoint to foo2 to check what happened (which worked well).
You don't say what version of gdb you are using.
Very old versions of gdb don't have any support for inline functions. This was true for 6.8 and maybe even 7.0 -- I don't recall. You can look at the NEWS file for your gdb to see.
Then there were some versions of gdb that supported breakpoints on inline functions, but only using the "file:line" syntax. So what you would do is look up the function in your editor, and find its line number and enter, e.g.:
(gdb) break myfile.c:777
Even more recent versions of gdb, starting with 7.4 or 7.5 (I forget) will handle "break function" just fine if "function" was inlined.
All of this only works if you have debuginfo available. So if you tried this, and it failed, either you have an older gdb, or you forgot to use -g.
There's no good way inside gdb to see what objects in a compilation were missing -g. You can see it pretty easily from the shell, though, by running "readelf -WS" on the .o files, and looking for files that don't have a .debug_info section.
Setting a breakpoint to the signature line of the function did not work. But setting one to the line of an instruction of the inlined function solved the problem for me. For instance, considering the following function inline_foo, found in myfile.c:
inline int inline_foo(int arg) // l.1
{
int a_var = 0;
do_smth(&a_var);
do_some_other_thing(); // l.5
if (a_var) {
a_var = blob();
} else {
a_var = blub();
return a_var; // l.10
}
I was trying b myfile.c:1, which did not appear to work. But if I tried b myfile.c:3 instead, the breakpoint was well handled by GDB.
Since the technique is the same as the one described previously by Tom, I'll accept his answer.
Related
I'm working on an application which defines it's own printf() to get around differences between the different CRTs out there or because some other platforms don't have it.
When building the application with gcc this automatically seems to work and the custom printf is used instead of libc's one; if I understand it correctly this is because of the order in which object files/libraries appear in the link command or maybe because object files are always searched before CRT libs, correct?
I'd like to do the same using msvc. Just building the project gives the expected 'LNK2005: _printf already defined in printf.obj' because printf is also in msvcrtd.lib. Fair enough. I know about /NODEFAULTLIB but that excludes everything resulting in unresolved references for everything but printf. I scanned through the other linker settings but couldn't find anything which allows this (apart from /FORCE maybe, but the 'might produce an invalid executable' comment doesn't make it sound like a good idea). Also nothing in the module definition file docs; the latter got me thinking it might be possible to create a stub library which has all exports from msvcrt.lib except printf but that seems a brittle solution even if it works.
In the end the question is simple: how do I tell msvc's linker it should skip msvcrt's printf definition and use the one from my printf.obj instead. Basically /NODEFAULTFUNCTION:printf or so. Just an answer for one single executable is ok, though I'd also be interested to know if and how it can be done when building a dll instead where the custom printf is exported: how to tell the linker it should use the export from my .lib instead of msvcrt.lib?
edit simplest repo I could find: create a file main.c:
#include <stdio.h>
int main(int argc, char** argv)
{
printf("Hello");
return 0;
}
and a file printf.c:
int printf(const char *fmt, ...)
{
write(1, "ok\n", 3);
return 3;
}
For VS2013 (though the other versions might work as well): create a new empty C++ project and add both files then build. (For gcc: just gcc main.c printf.c and the resulting a.out prints 'ok')
The culptrit for VS is #include : without that it works ok but I have yet to find out if the original code allows getting rid of it in some way. But even if it does I'd still want to know if this can be solved at the link level.
I have some C++ code like this that I'm stepping through with GDB:
void foo(int num) { ... }
void main() {
Baz baz;
foo (baz.get());
}
When I'm in main(), I want to step into foo(), but I want to step over baz.get().
The GDB docs say that "the step command only enters a function if there is line number information for the function", so I'd be happy if I could remove the line number information for baz.get() from my executable. But ideally, I'd be able to tell GDB "never step into any function in the Baz class".
Does anyone know how to do this?
Starting with GDB 7.4, skip can be used.
Run info skip, or check out the manual for details: https://sourceware.org/gdb/onlinedocs/gdb/Skipping-Over-Functions-and-Files.html
Instead of choosing to "step", you can use the "until" command to usually behave in the way that you desire:
(gdb) until foo
I don't know of any way to permanently configure gdb to skip certain symbols (aside from eliding their debugging information).
Edit: actually, the GDB documentation states that you can't use until to jump to locations that aren't in the same frame. I don't think this is true, but in the event that it is, you can use advance for the same purpose:
(gdb) advance foo
Page 85 of the GDB manual defines what can be used as "location" arguments for commands that take them. Just putting "foo" will make it look for a function named foo, so as long as it can find it, you should be fine. Alternatively you're stuck typing things like the filename:linenum for foo, in which case you might just be better off setting a breakpoint on foo and using continue to advance to it.
(I think this might be better suited as a comment rather than an answer, but I don't have enough reputation to add a comment yet.)
So I've also been wanting to ignore STL, Boost, et al (collectively '3rd Party') files when debugging for a while. Yesterday I finally decided to look for a solution and it seems the nearest capability is the 'skip' command in GDB.
I found the 'skip' ability in GDB to be helpful, but it's still a nuisance for me because my program uses a lot of STL and other "3rd Party" template code. In this case I have to mark a bunch of files as skip. After the 2nd time doing so I realized it would be more helpful to be able to skip an entire directory--and most helpful to skip a directory and all subdirectories. That way I can skip, for example, /usr since none of my code lives there and I typically have no interest in debugging through 3rd party code. So I extended the 'skip' command in gdb to support a new type 'dir'. I can now do this in gdb:
skip dir /usr
and then I'm never stopped in any of my 3rd party headers.
Here's a webpage w/ this info + the patch if it helps anyone: info & patch to skip directories in GDB
It appears that this isn't possible in GDB. I've filed a bug.
Meanwhile, gdb has the skip function command. Just execute it when you are inside the uninteresting function and it will not bother you again.
skip file is also very useful to get rid of the STL internals.
As Justin has said, it has been added in gdb 7.4. For more details, take a look at the documentation.
I am trying to measure my code coverage utilization on a C project consist of several libraries, and main program.
Is there a utility that can help me find which function I dont use from both libraries and main program.
I want to build list of functions (public functions) that are not used by my main program, in order to ignore them in my code coverage report.
If you are using gcc you compile your code with option:
-Wunused-function
Warn whenever a static function is declared but not defined or a non-inline static function is unused. This warning is enabled by -Wall.
cflow can create a call graph for the program, but it doesn't work well with pointers to functions in some cases.
for eaxample:
#include <stdio.h>
static int f1(){
return 1;
}
int (*p_f1)() = f1;
int main() {
p_f1();
return 0;
}
There are coverage tools available for free - for example "gcov" that runs on go with the gcc tool suite. However, Code Coverage only tells you which functions get hit by your testing (or whatever you do to excercise the code), so for example
ptr = malloc(...);
if (!ptr)
{
allocation_failed(__FILE__, __LINE__);
}
would only show that allocation_failed is called if you are also using some tool that makes your allocations fail from time to time.
I'm not aware of a tool that will show you what functions are not used across larger systems (with multiple libraries, etc). I expect you could make something by using the output of "nm" and a bit of "pulling things in". It won't cover foo and bar as unusued in this case:
unit1.c:
extern int foo(void);
int bar()
{
return foo();
}
unit2.c:
int foo(void)
{
return 42;
}
int baz(void)
{
return bar();
}
and then baz isn't used anywhere. But if you remove baz, it will show that bar is not called, and then you can remove foo after that...
Edit: Crazy idea time - how about taking every C file in the project and just concatenating the whole thing into a a single .C file, then add static at the beginning of every function, and compiling with -Wunused-functon - I'm sure there will be some "interesting" effects from this if your code isn't extremely well written, but it may be worth a try [it would be fairly easy to do this in a Linux system, something like find . -name "*.c" -print | xargs cat {} > giantsource.c - you then need a little bit of sed or something to label all functions static, which I'm not quite sure how you'd go about doing - it depends very much on the formatting of your code.
You may want to have a look at this:
http://www.gedanken.demon.co.uk/cxref/
I haven't used it, but any decent cross referencing tool should be able to identify anything that is "not used" as not having any references. Of course, you'll probably still have to run over the code severa times to weed out the functions that are used by functions that aren't being called, etc.
cflow has an option to build a cross-reference table: --xref
The format of the output is described by GNU cflow: Cross-Reference
GNU cflow is also able to produce cross-reference listings. This mode is enabled by --xref (-x) command line option. Cross-reference output lists each symbol occurrence on a separate line. Each line shows the identifier and the source location where it appears. If this location is where the symbol is defined, it is additionally marked with an asterisk and followed by the definition. For example, here is a fragment of a cross-reference output for d.c program:
printdir * d.c:42 void printdir (int level,char *name)
printdir d.c:74
printdir d.c:102
It shows that the function printdir is defined in line 42 and referenced twice, in lines 74 and 102.
To detect unused functions, search the line with a star not followed by a line with the same prefix. The following GNU Awk code print the unused functions:
{
if( $2 == "*" ) {
if( f ) {
print f
}
f = $1
}
else {
f = ""
}
}
The command may be:
cflow -x src/*.c src-gen/*.c | awk -f find-unused-functions.awk
While debugging a C program in gdb I have a breakpoint in a for loop. I cannot print the value of "i" ( I get : No symbol "i" in current context.). I can print the value of all the other variables. Is it normal?
Here is the loop:
for (i=0; i < datasize; i++){
if ( feature_mask[i] > 0 ){
k = feature_mask[i] - 1;
if (neighbors[k][nmax-1] != 0){
neighbors[k][nmax-1] = bvalue;
feature_mask[i] = -feature_mask[i];
}
}
}
It has probably been optimised out of your compiled code as you only use feature_mask[i] within the loop.
Did you specify an optimization level when you called your compiler? If you were using gcc, then just omit any -O options and try again.
I encountered this issue recently. I compiled GCC 5.1 and then used it to compile a C++11 codebase. And, although I could step through the program's code in gdb, I couldn't print the value of any variable, I kept getting “No symbol "xyz" in current context” errors, for every variable.
I was using gdb 7.4, but the latest version available at the time was 7.9. I downloaded the latest version of gdb and compiled it (using GCC 5.1) and when using gdb 7.9 I was able to successfully inspect variable values again.
I guess the debug information of GCC 5.1 is incompatible with gdb 7.4.
Make sure the program is compiled without optimization, and with debugging information enabled. It's quite likely that the loop counter ends up in a register.
Check your optimization options. It's possible the GCC could replace the variable with a pointer into feature_mask.
You can try declaring i as volatile. That will prevent some compiler optimizations (and hopefully make i visible inside the debugger).
In case anyone else is using Google's Bazel build system for your project, I'd like to add that if you cannot print any variables from gdb, it may be because you need to properly add the -ggdb and -O0 (update: use -Og instead of -O0 use -O0 over -Og) C build flags using the --copt= option, INSTEAD OF using the --per_file_copt= option. In my case, although they both built and ran just fine, only the --copt= technique allowed me to fully use gdb and print variables, whereas the --per_file_copt= one also allowed me to use gdb but would NOT allow me to print variables.
Note: in the below examples, just replace test with build if you do NOT need to run the unit tests as well.
UDPATE: it turns out, you should prefer -Og over -O0 when doing debugging, so I'm updating these examples accordingly. See here: What's the difference between a compiler's `-O0` option and `-Og` option?.
So, do this:
time bazel test --copt=-ggdb --copt=-O0 \
//my/build/folder1/... //my/build/folder2/...
INSTEAD OF this:
time bazel test --per_file_copt=//my/build/folder1/...,//my/build/folder2/...#-ggdb,-O0 \
//my/build/folder1/... //my/build/folder2/...
...in order to be able to print variables from within gdb.
Again, both of the above techniques build and run just fine, and both allow me to run and use gdb, but only the first one actually allows me to use gdb to its full extent.
Lastly, if the first command above still doesn't work, try adding the --strip=never Bazel flag described here to prevent Bazel from ever stripping debugging information. Now the command will look like this:
time bazel test --copt=-ggdb --copt=-O0 --strip=never \
//my/build/folder1/... //my/build/folder2/...
Reference documentation:
--copt=:
https://docs.bazel.build/versions/master/command-line-reference.html#flag--copt
[better, with examples] https://docs.bazel.build/versions/master/user-manual.html#flag--copt
--per_file_copt:
https://docs.bazel.build/versions/master/command-line-reference.html#flag--per_file_copt
[better, with examples] https://docs.bazel.build/versions/master/user-manual.html#flag--per_file_copt
--strip=never:
https://docs.bazel.build/versions/master/user-manual.html#flag--strip
[my own Q&A] Prefer -Og over -O0 -O0 over -Og for debugging: What's the difference between a compiler's `-O0` option and `-Og` option?
I'm having some trouble stepping over in GDB. I've built an example program from the ffmpeg library with debug symbols on and stripping off. Although I configured the ffmpeg library to static and explicitly disabled shared, it looks like the program I'm debugging is linking dynamically, since its file size is only 99kB. I don't know that this is the issue but thought to mention it.
After I set and hit a breakpoint in av_seek_frame, I use the 'next' command to step over. However, this steps into the first function within av_seek_frame(), as you can see below. Furthermore if a do a second 'next', the backtrace looses track of where it is. Am I set up wrong? How can I step over? I should note I double checked that 'set step-mode off' is off as the default (as I believe this will break at the first piece of code without debug info.)
Breakpoint 1, av_seek_frame (s=0x16429000, stream_index=0, timestamp=29727438, flags=0) at l
(gdb) list
1648
1649 return 0;
1650 }
1651
1652 int av_seek_frame(AVFormatContext *s, int stream_index, int64_t timestamp, int flags
1653 {
1654 int ret;
1655 AVStream *st;
1656
1657 ff_read_frame_flush(s);
(gdb) next
ff_read_frame_flush (s=0x16429000) at libavformat/utils.c:1248
(gdb) list
1243
1244 /**
1245 * Flush the frame reader.
1246 **/
1247 void ff_read_frame_flush(AVFormatContext *s)
1248 {
1249 AVStream *st;
1250 int i, j;
1251
1252 flush_packet_queue(s);
(gdb) next
ff_read_frame_flush (s=0x16429000) at libavformat/utils.c:1252
(gdb) where
#0 ff_read_frame_flush (s=0x16429000) at libavformat/utils.c:1252
#1 0x00000000 in ?? ()
If you're not sure whether or not your binary is statically linked, you can check it with ldd and see a message like this:
% ldd ffmpeg
not a dynamic executable
Next, make sure that you give gdb the full path to the executable so that you're not accidentally picking up a binary installed elsewhere on the system that happens to be in your PATH.
Most likely you're loading the wrong binary. Even without using --disable-stripping and --disable-optimizations I can use gdb fine using both step and next commands. You don't need to use --disable-stripping because inside gdb you can use the ffmpeg_g binary (or if you happen to run the ffmpeg binary you can load the symbols from it using file ffmpeg_g).
For debugging purposes it is nice to use the --disable-optimizations so that you don't get value optimized out when inspecting variables, but strictly speaking you don't need to use the option to get emacs/gdb to behave... I have no problems stepping through the code when optimizations are used.
There is one things to keep in mind, though, when setting breakpoints with gud/gdb inside of Emacs that might lead to confusion: the gud-break command uses only the base part of the filename for setting breakpoints, not the absolute path to it, which in the case of ffmpeg means that if, for example, you set a breakpoint in utils.c it might not work correctly depending on the value of the source code search paths that you have set in gdb because ffmpeg has multiple files named utils.c in different paths (in fact, there are a total of 5 utils.c files, one in each of the lib* subdirectories) . By default the search path is set to $cdir:$cwd, but if you have it set to something like /path/to/ffmpeg:$cdir:$cwd and you try to set a breakpoint in the utils.c of libavformat, it might find the one in libavutil-- in which case if you're lucky it will complain that the line where you want to set a breakpoint does not exist (because the one in libavutil is shorter), or it might set a breakpoint on the line you want, but in the wrong utils.c.
This issue with gud/gdb should be considered a bug. When I get a moment I'll submit a patch for gud-break/gud-format-command to fix the problem.