What do you consider "best practice" when it comes to error handling errors in a consistent way in a C library.
There are two ways I've been thinking of:
Always return error code. A typical function would look like this:
MYAPI_ERROR getObjectSize(MYAPIHandle h, int* returnedSize);
The always provide an error pointer approach:
int getObjectSize(MYAPIHandle h, MYAPI_ERROR* returnedError);
When using the first approach it's possible to write code like this where the error handling check is directly placed on the function call:
int size;
if(getObjectSize(h, &size) != MYAPI_SUCCESS) {
// Error handling
}
Which looks better than the error handling code here.
MYAPIError error;
int size;
size = getObjectSize(h, &error);
if(error != MYAPI_SUCCESS) {
// Error handling
}
However, I think using the return value for returning data makes the code more readable, It's obvious that something was written to the size variable in the second example.
Do you have any ideas on why I should prefer any of those approaches or perhaps mix them or use something else? I'm not a fan of global error states since it tends to make multi threaded use of the library way more painful.
EDIT:
C++ specific ideas on this would also be interesting to hear about as long as they are not involving exceptions since it's not an option for me at the moment...
I've used both approaches, and they both worked fine for me. Whichever one I use, I always try to apply this principle:
If the only possible errors are programmer errors, don't return an error code, use asserts inside the function.
An assertion that validates the inputs clearly communicates what the function expects, while too much error checking can obscure the program logic. Deciding what to do for all the various error cases can really complicate the design. Why figure out how functionX should handle a null pointer if you can instead insist that the programmer never pass one?
I like the error as return-value way. If you're designing the api and you want to make use of your library as painless as possible think about these additions:
store all possible error-states in one typedef'ed enum and use it in your lib. Don't just return ints or even worse, mix ints or different enumerations with return-codes.
provide a function that converts errors into something human readable. Can be simple. Just error-enum in, const char* out.
I know this idea makes multithreaded use a bit difficult, but it would be nice if application programmer can set an global error-callback. That way they will be able to put a breakpoint into the callback during bug-hunt sessions.
There's a nice set of slides from CMU's CERT with recommendations for when to use each of the common C (and C++) error handling techniques. One of the best slides is this decision tree:
I would personally change two things about this flowcart.
First, I would clarify that sometimes objects should use return values to indicate errors. If a function only extracts data from an object but doesn't mutate the object, then the integrity of the object itself is not at risk and indicating errors using a return value is more appropriate.
Second, it's not always appropriate to use exceptions in C++. Exceptions are good because they can reduce the amount of source code devoted to error handling, they mostly don't affect function signatures, and they're very flexible in what data they can pass up the callstack. On the other hand, exceptions might not be the right choice for a few reasons:
C++ exceptions have very particular semantics. If you don't want those semantics, then C++ exceptions are a bad choice. An exception must be dealt with immediately after being thrown and the design favors the case where an error will need to unwind the callstack a few levels.
C++ functions that throw exceptions can't later be wrapped to not throw exceptions, at least not without paying the full cost of exceptions anyway. Functions that return error codes can be wrapped to throw C++ exceptions, making them more flexible. C++'s new gets this right by providing a non-throwing variant.
C++ exceptions are relatively expensive but this downside is mostly overblown for programs making sensible use of exceptions. A program simply shouldn't throw exceptions on a codepath where performance is a concern. It doesn't really matter how fast your program can report an error and exit.
Sometimes C++ exceptions are not available. Either they're literally not available in one's C++ implementation, or one's code guidelines ban them.
Since the original question was about a multithreaded context, I think the local error indicator technique (what's described in SirDarius's answer) was underappreciated in the original answers. It's threadsafe, doesn't force the error to be immediately dealt with by the caller, and can bundle arbitrary data describing the error. The downside is that it must be held by an object (or I suppose somehow associated externally) and is arguably easier to ignore than a return code.
I use the first approach whenever I create a library. There are several advantages of using a typedef'ed enum as a return code.
If the function returns a more complicated output such as an array and its length you do not need to create arbitrary structures to return.
rc = func(..., int **return_array, size_t *array_length);
It allows for simple, standardized error handling.
if ((rc = func(...)) != API_SUCCESS) {
/* Error Handling */
}
It allows for simple error handling in the library function.
/* Check for valid arguments */
if (NULL == return_array || NULL == array_length)
return API_INVALID_ARGS;
Using a typedef'ed enum also allows for the enum name to be visible in the debugger. This allows for easier debugging without the need to constantly consult a header file. Having a function to translate this enum into a string is helpful as well.
The most important issue regardless of approach used is to be consistent. This applies to function and argument naming, argument ordering and error handling.
Returning error code is the usual approach for error handling in C.
But recently we experimented with the outgoing error pointer approach as well.
It has some advantages over the return value approach:
You can use the return value for more meaningful purposes.
Having to write out that error parameter reminds you to handle the error or propagate it. (You never forget checking the return value of fclose, don't you?)
If you use an error pointer, you can pass it down as you call functions. If any of the functions set it, the value won't get lost.
By setting a data breakpoint on the error variable, you can catch where does the error occurred first. By setting a conditional breakpoint you can catch specific errors too.
It makes it easier to automatize the check whether you handle all errors. The code convention may force you to call your error pointer as err and it must be the last argument. So the script can match the string err); then check if it's followed by if (*err. Actually in practice we made a macro called CER (check err return) and CEG (check err goto). So you don't need to type it out always when we just want to return on error, and can reduce the visual clutter.
Not all functions in our code has this outgoing parameter though.
This outgoing parameter thing are used for cases where you would normally throw an exception.
Here's a simple program to demonstrate the first 2 bullets of Nils Pipenbrinck's answer here.
His first 2 bullets are:
store all possible error-states in one typedef'ed enum and use it in your lib. Don't just return ints or even worse, mix ints or different enumerations with return-codes.
provide a function that converts errors into something human readable. Can be simple. Just error-enum in, const char* out.
Assume you have written a module named mymodule. First, in mymodule.h, you define your enum-based error codes, and you write some error strings which correspond to these codes. Here I am using an array of C strings (char *), which only works well if your first enum-based error code has value 0, and you don't manipulate the numbers thereafter. If you do use error code numbers with gaps or other starting values, you'll simply have to change from using a mapped C-string array (as I do below) to using a function which uses a switch statement or if / else if statements to map from enum error codes to printable C strings (which I don't demonstrate). The choice is yours.
mymodule.h
/// #brief Error codes for library "mymodule"
typedef enum mymodule_error_e
{
/// No error
MYMODULE_ERROR_OK = 0,
/// Invalid arguments (ex: NULL pointer where a valid pointer is required)
MYMODULE_ERROR_INVARG,
/// Out of memory (RAM)
MYMODULE_ERROR_NOMEM,
/// Make up your error codes as you see fit
MYMODULE_ERROR_MYERROR,
// etc etc
/// Total # of errors in this list (NOT AN ACTUAL ERROR CODE);
/// NOTE: that for this to work, it assumes your first error code is value 0 and you let it naturally
/// increment from there, as is done above, without explicitly altering any error values above
MYMODULE_ERROR_COUNT,
} mymodule_error_t;
// Array of strings to map enum error types to printable strings
// - see important NOTE above!
const char* const MYMODULE_ERROR_STRS[] =
{
"MYMODULE_ERROR_OK",
"MYMODULE_ERROR_INVARG",
"MYMODULE_ERROR_NOMEM",
"MYMODULE_ERROR_MYERROR",
};
// To get a printable error string
const char* mymodule_error_str(mymodule_error_t err);
// Other functions in mymodule
mymodule_error_t mymodule_func1(void);
mymodule_error_t mymodule_func2(void);
mymodule_error_t mymodule_func3(void);
mymodule.c contains my mapping function to map from enum error codes to printable C strings:
mymodule.c
#include <stdio.h>
/// #brief Function to get a printable string from an enum error type
/// #param[in] err a valid error code for this module
/// #return A printable C string corresponding to the error code input above, or NULL if an invalid error code
/// was passed in
const char* mymodule_error_str(mymodule_error_t err)
{
const char* err_str = NULL;
// Ensure error codes are within the valid array index range
if (err >= MYMODULE_ERROR_COUNT)
{
goto done;
}
err_str = MYMODULE_ERROR_STRS[err];
done:
return err_str;
}
// Let's just make some empty dummy functions to return some errors; fill these in as appropriate for your
// library module
mymodule_error_t mymodule_func1(void)
{
return MYMODULE_ERROR_OK;
}
mymodule_error_t mymodule_func2(void)
{
return MYMODULE_ERROR_INVARG;
}
mymodule_error_t mymodule_func3(void)
{
return MYMODULE_ERROR_MYERROR;
}
main.c contains a test program to demonstrate calling some functions and printing some error codes from them:
main.c
#include <stdio.h>
int main()
{
printf("Demonstration of enum-based error codes in C (or C++)\n");
printf("err code from mymodule_func1() = %s\n", mymodule_error_str(mymodule_func1()));
printf("err code from mymodule_func2() = %s\n", mymodule_error_str(mymodule_func2()));
printf("err code from mymodule_func3() = %s\n", mymodule_error_str(mymodule_func3()));
return 0;
}
Output:
Demonstration of enum-based error codes in C (or C++)
err code from mymodule_func1() = MYMODULE_ERROR_OK
err code from mymodule_func2() = MYMODULE_ERROR_INVARG
err code from mymodule_func3() = MYMODULE_ERROR_MYERROR
References:
You can run this code yourself here: https://onlinegdb.com/ByEbKLupS.
My answer I frequently reference to see this type of error handling: STM32 how to get last reset status
I personally prefer the former approach (returning an error indicator).
Where necessary the return result should just indicate that an error occurred, with another function being used to find out the exact error.
In your getSize() example I'd consider that sizes must always be zero or positive, so returning a negative result can indicate an error, much like UNIX system calls do.
I can't think of any library that I've used that goes for the latter approach with an error object passed in as a pointer. stdio, etc all go with a return value.
The UNIX approach is most similar to your second suggestion. Return either the result or a single "it went wrong" value. For instance, open will return the file descriptor on success or -1 on failure. On failure it also sets errno, an external global integer to indicate which failure occurred.
For what it's worth, Cocoa has also been adopting a similar approach. A number of methods return BOOL, and take an NSError ** parameter, so that on failure they set the error and return NO. Then the error handling looks like:
NSError *error = nil;
if ([myThing doThingError: &error] == NO)
{
// error handling
}
which is somewhere between your two options :-).
Use setjmp.
http://en.wikipedia.org/wiki/Setjmp.h
http://aszt.inf.elte.hu/~gsd/halado_cpp/ch02s03.html
http://www.di.unipi.it/~nids/docs/longjump_try_trow_catch.html
#include <setjmp.h>
#include <stdio.h>
jmp_buf x;
void f()
{
longjmp(x,5); // throw 5;
}
int main()
{
// output of this program is 5.
int i = 0;
if ( (i = setjmp(x)) == 0 )// try{
{
f();
} // } --> end of try{
else // catch(i){
{
switch( i )
{
case 1:
case 2:
default: fprintf( stdout, "error code = %d\n", i); break;
}
} // } --> end of catch(i){
return 0;
}
#include <stdio.h>
#include <setjmp.h>
#define TRY do{ jmp_buf ex_buf__; if( !setjmp(ex_buf__) ){
#define CATCH } else {
#define ETRY } }while(0)
#define THROW longjmp(ex_buf__, 1)
int
main(int argc, char** argv)
{
TRY
{
printf("In Try Statement\n");
THROW;
printf("I do not appear\n");
}
CATCH
{
printf("Got Exception!\n");
}
ETRY;
return 0;
}
When I write programs, during initialization, I usually spin off a thread for error handling, and initialize a special structure for errors, including a lock. Then, when I detect an error, through return values, I enter in the info from the exception into the structure and send a SIGIO to the exception handling thread, then see if I can't continue execution. If I can't, I send a SIGURG to the exception thread, which stops the program gracefully.
I have done a lot of C programming in the past. And I really apreciated the error code return value. But is has several possible pitfalls:
Duplicate error numbers, this can be solved with a global errors.h file.
Forgetting to check the error code, this should be solved with a cluebat and long debugging hours. But in the end you will learn (or you will know that someone else will do the debugging).
I ran into this Q&A a number of times, and wanted to contribute a more comprehensive answer. I think the best way to think about this is how to return errors to the caller, and what you return.
How
There are 3 ways to return information from a function:
Return Value
Out Argument(s)
Out of Band, that includes non-local goto (setjmp/longjmp),
file or global scoped variables, file system etc.
Return Value
You can only return a single value (object); however, it can be an arbitrarily complex value. Here is an example of an error returning function:
enum error hold_my_beer(void);
One benefit of return values is that it allows chaining of calls for less intrusive error handling:
!hold_my_beer() &&
!hold_my_cigarette() &&
!hold_my_pants() ||
abort();
This not just about readability, but may also allow processing an array of such function pointers in a uniform way.
Out Argument(s)
You can return more via more than one object via arguments, but best practice does suggest to keep the total number of arguments low (say, <=4):
void look_ma(enum error *e, char *what_broke);
enum error e;
look_ma(e);
if(e == FURNITURE) {
reorder(what_broke);
} else if(e == SELF) {
tell_doctor(what_broke);
}
This forces caller to pass in object which may make it more likely that it's being checked. If you have a set of calls all returning errors, and you decide to allocate a new variable to each, then it add some clutter in the caller.
Out of Band
The best known example is probably the (thread-local) errno variable, which the called function sets. It's very easy for the callee to not check this variable, and you only get one which may be an issue if your function is complicated (for instance, two parts of the function returning the same error code).
With setjmp() you define a place and how you want to handle an int value, and you transfer control to that location via a longjmp(). See Practical usage of setjmp and longjmp in C.
What
Indicator
Code
Object
Callback
Indicator
An error indicator only tells you that there is a problem but nothing about the nature of said problem:
struct foo *f = foo_init();
if(!f) {
/// handle the absence of foo
}
This is the least powerful way for a function to communicate error state; however, it's perfect if the caller cannot respond to the error in a graduated manner anyways.
Code
An error code tells the caller about the nature of the problem, and may allow for a suitable response (from the above). It can be a return value, or like the look_ma() example above an error argument.
Object
With an error object, the caller can be informed about arbitrarily complicated issues. For example, an error code and a suitable human-readable message. It can also inform the caller that multiple things went wrong, or an error per item when processing a collection:
struct collection friends;
enum error *e = malloc(c.size * sizeof(enum error));
...
ask_for_favor(friends, reason);
for(int i = 0; i < c.size; i++) {
if(reason[i] == NOT_FOUND) find(friends[i]);
}
Instead of pre-allocating the error array, you can also (re)allocate it dynamically as needed of course.
Callback
Callback is the most powerful way to handle errors, as you can tell the function what behavior you would like to see happen when something goes wrong. A callback argument can be added to each function, or if customization uis only required per instance of a struct like this:
struct foo {
...
void (error_handler)(char *);
};
void default_error_handler(char *message) {
assert(f);
printf("%s", message);
}
void foo_set_error_handler(struct foo *f, void (*eh)(char *)) {
assert(f);
f->error_handler = eh;
}
struct foo *foo_init() {
struct foo *f = malloc(sizeof(struct foo));
foo_set_error_handler(f, default_error_handler);
return f;
}
struct foo *f = foo_init();
foo_something();
One interesting benefit of a callback is that it can be invoked multiple times, or none at all in the absence of errors in which there is no overhead on the happy path.
There is, however, an inversion of control. The calling code does not know if the callback was invoked. As such, it may make sense to use an indicator as well.
I was pondering this issue recently as well, and wrote up some macros for C that simulate try-catch-finally semantics using purely local return values. Hope you find it useful.
Here is an approach which I think is interesting, while requiring some discipline.
This assumes a handle-type variable is the instance on which operate all API functions.
The idea is that the struct behind the handle stores the previous error as a struct with necessary data (code, message...), and the user is provided with a function that returns a pointer to this error object. Each operation will update the pointed object so the user can check its status without even calling functions. As opposed to the errno pattern, the error code is not global, which make the approach thread-safe, as long as each handle is properly used.
Example:
MyHandle * h = MyApiCreateHandle();
/* first call checks for pointer nullity, since we cannot retrieve error code
on a NULL pointer */
if (h == NULL)
return 0;
/* from here h is a valid handle */
/* get a pointer to the error struct that will be updated with each call */
MyApiError * err = MyApiGetError(h);
MyApiFileDescriptor * fd = MyApiOpenFile("/path/to/file.ext");
/* we want to know what can go wrong */
if (err->code != MyApi_ERROR_OK) {
fprintf(stderr, "(%d) %s\n", err->code, err->message);
MyApiDestroy(h);
return 0;
}
MyApiRecord record;
/* here the API could refuse to execute the operation if the previous one
yielded an error, and eventually close the file descriptor itself if
the error is not recoverable */
MyApiReadFileRecord(h, &record, sizeof(record));
/* we want to know what can go wrong, here using a macro checking for failure */
if (MyApi_FAILED(err)) {
fprintf(stderr, "(%d) %s\n", err->code, err->message);
MyApiDestroy(h);
return 0;
}
First approach is better IMHO:
It's easier to write function that way. When you notice an error in the middle of the function you just return an error value. In second approach you need to assign error value to one of the parameters and then return something.... but what would you return - you don't have correct value and you don't return error value.
it's more popular so it will be easier to understand, maintain
I definitely prefer the first solution :
int size;
if(getObjectSize(h, &size) != MYAPI_SUCCESS) {
// Error handling
}
i would slightly modify it, to:
int size;
MYAPIError rc;
rc = getObjectSize(h, &size)
if ( rc != MYAPI_SUCCESS) {
// Error handling
}
In additional i will never mix legitimate return value with error even if currently the scope of function allowing you to do so, you never know which way function implementation will go in the future.
And if we already talking about error handling i would suggest goto Error; as error handling code, unless some undo function can be called to handle error handling correctly.
What you could do instead of returning your error, and thus forbidding you from returning data with your function, is using a wrapper for your return type:
typedef struct {
enum {SUCCESS, ERROR} status;
union {
int errCode;
MyType value;
} ret;
} MyTypeWrapper;
Then, in the called function:
MyTypeWrapper MYAPIFunction(MYAPIHandle h) {
MyTypeWrapper wrapper;
// [...]
// If there is an error somewhere:
wrapper.status = ERROR;
wrapper.ret.errCode = MY_ERROR_CODE;
// Everything went well:
wrapper.status = SUCCESS;
wrapper.ret.value = myProcessedData;
return wrapper;
}
Please note that with the following method, the wrapper will have the size of MyType plus one byte (on most compilers), which is quite profitable; and you won't have to push another argument on the stack when you call your function (returnedSize or returnedError in both of the methods you presented).
In addition to what has been said, prior to returning your error code, fire off an assert or similar diagnostic when an error is returned, as it will make tracing a lot easier. The way I do this is to have a customised assert that still gets compiled in at release but only gets fired when the software is in diagnostics mode, with an option to silently report to a log file or pause on screen.
I personally return error codes as negative integers with no_error as zero , but it does leave you with the possible following bug
if (MyFunc())
DoSomething();
An alternative is have a failure always returned as zero, and use a LastError() function to provide details of the actual error.
EDIT:If you need access only to the last error, and you don't work in multithreaded environment.
You can return only true/false (or some kind of #define if you work in C and don't support bool variables), and have a global Error buffer that will hold the last error:
int getObjectSize(MYAPIHandle h, int* returnedSize);
MYAPI_ERROR LastError;
MYAPI_ERROR* getLastError() {return LastError;};
#define FUNC_SUCCESS 1
#define FUNC_FAIL 0
if(getObjectSize(h, &size) != FUNC_SUCCESS ) {
MYAPI_ERROR* error = getLastError();
// error handling
}
Second approach lets the compiler produce more optimized code, because when address of a variable is passed to a function, the compiler cannot keep its value in register(s) during subsequent calls to other functions. The completion code usually is used only once, just after the call, whereas "real" data returned from the call may be used more often
I prefer error handling in C using the following technique:
struct lnode *insert(char *data, int len, struct lnode *list) {
struct lnode *p, *q;
uint8_t good;
struct {
uint8_t alloc_node : 1;
uint8_t alloc_str : 1;
} cleanup = { 0, 0 };
// allocate node.
p = (struct lnode *)malloc(sizeof(struct lnode));
good = cleanup.alloc_node = (p != NULL);
// good? then allocate str
if (good) {
p->str = (char *)malloc(sizeof(char)*len);
good = cleanup.alloc_str = (p->str != NULL);
}
// good? copy data
if(good) {
memcpy ( p->str, data, len );
}
// still good? insert in list
if(good) {
if(NULL == list) {
p->next = NULL;
list = p;
} else {
q = list;
while(q->next != NULL && good) {
// duplicate found--not good
good = (strcmp(q->str,p->str) != 0);
q = q->next;
}
if (good) {
p->next = q->next;
q->next = p;
}
}
}
// not-good? cleanup.
if(!good) {
if(cleanup.alloc_str) free(p->str);
if(cleanup.alloc_node) free(p);
}
// good? return list or else return NULL
return (good ? list : NULL);
}
Source: http://blog.staila.com/?p=114
In addition the other great answers, I suggest that you try to separate the error flag and the error code in order to save one line on each call, i.e.:
if( !doit(a, b, c, &errcode) )
{ (* handle *)
(* thine *)
(* error *)
}
When you have lots of error-checking, this little simplification really helps.
I have seen five main approaches used in error reporting by functions in C:
return value with no error code reporting or no return value
return value that is an error code only
return value that is a valid value or an error code value
return value indicating an error with some way of fetching an error code possibly with error context information
function argument that returns a value with an error code possibly with error context information
In addition to the choice of function error return mechanism there is also the consideration of error code mnemonics and ensuring that the error code mnemonics do not clash with any other error code mnemonics being used. Typically this requires the use of a Three Letter Prefix approach to the naming of mnemonics defining them with #define, enum, or const static int. See this discussion "static const" vs "#define" vs "enum"
There are a couple of different outcomes once an error is detected and that may be a consideration how functions provide error codes and error information. These outcomes are really divided into two camps, recoverable errors and unrecoverable errors:
document the system state and then abort
wait and retry the failed action
notify a human being and request assistance
continue execution in a degraded state
An error type may use more than one of these outcomes depending on the context of the error. For instance a file open that fails because the file doesn't exist may be retried with a different file name or notify a user and ask for assistance or continue execution in a degraded state.
Details on Five Main Approaches
Some functions do not provide an error code. The functions either can't fail or if they fail, they fail silently. An example of this type of function are the various is character test functions such as isdigit() which indicates if a character value is a digit or is not. A character value either is or is not a digit or an alphabetic character. Similarly with the strcmp() function, comparing two strings results in a value indicating which one is higher in the collating sequence than the other should they not be the same.
In some cases an error code is not necessary because a value indicating failure is a valid result. For example the strchr() function from the Standard Library returns a pointer to the searched for character if found in the string to be scanned or NULL if it is not found. In this case a failure to find the character is a valid and useful indicator. A function using strchr() may require the character searched for not be in the string to be successful and finding the character is an error condition.
Other functions do not return an error code but instead report an error through an external mechanism. This is used by most of the math library functions in the Standard Library which require the user to set errno to a value of zero, call the function, and then check that the value of errno is still zero. The range of output values from many of the math functions do not allow a special return value to be used to indicate an error and they do not have an error reporting argument in their interfaces.
Some functions perform an action and return an error code value with one of the possible error code values indicating success and the rest of the range of values indicating an error code. For example a function may return a value of 0 if successful or a positive or negative non-zero value indicating an error with the value returned being the error code.
Some functions may perform an action and return either a value from a range of valid values if successful or a value from a range of invalid values indicating an error code. A simple approach is to use a positive value (0, 1, 2, ...) for valid values and a negative value for error codes allowing a check such as if(status < 0) return error;.
Some functions return a valid value or an invalid value indicating an error requiring the additional step of fetching the error code by some means. For example the fopen() function returns either a pointer to a FILE object or it returns an invalid pointer value of NULL and sets errno to an error code indicating the reason for the failure. A number of Windows API functions that return a HANDLE value to reference a resource may also return a value of INVALID_HANDLE_VALUE and the function GetLastError() is used to obtain the error code. The OPOS Control Objects standard requires an OPOS Control Object to provide two functions, GetResultCode() and GetResultCodeExtended(), to allow for the retrieval of error status information in the event a COM object method call fails.
This same approach is used in other APIs that use a handle or reference to a resource in which there is a range of valid values with one or more values outside of that range used to indicate an error. A mechanism is then provided to fetch additional error information such as an error code.
A similar approach is used with functions that return a boolean value of true to indicate the function was successful or false to indicate an error. The programmer must then examine other data to determine an error code such as GetLastError() with the Windows API.
Some functions have a pointer argument containing the address of a memory area for the function called to provide an error code or error information. Where this approach really shines is when in addition to a simple error code there is additional, error context information that helps to pin point the error. For example a JSON string parsing function may not only return an error code but also a pointer to where in the JSON string the parsing failed.
I have also seen functions where the function returned an error indicator such as a boolean value with the argument used for error information. I recall that the error information argument could in some cases be NULL indicating the caller didn't want to know the specifics of a failure.
This approach to returning error code or error information seems to be uncommon in my experience though for some reason I think I've seen it used in the Windows API from time to time or perhaps with an XML parser.
Considerations for multi-threading
When using the approach of an additional error code access through a mechanism as in checking a global such as errno or using a function such as GetLastError() there is the problem of sharing the global across multiple threads.
Modern compilers and libraries deal with this by using thread local storage to ensure that each thread has its own storage that is not shared by other threads. However there is still the issue of multiple functions sharing the same thread local storage location for status information which may require some accomodation. For instance, a function that uses several files may need to work around the issue that all of the fopen() calls that may fail share a single errno in the same thread.
If the API uses some type of handle or reference then error code storage can be made handle specific. The fopen() function could be wrapped in another function which performs the fopen() and then sets an API control block with both the FILE * returned by the fopen() as well as the value of errno.
The approach I prefer
My preference is for an error code to be returned as a function return value so that I can either check it at the point of call or save it for later. In most cases, an error is something to be dealt with immediately which is why I prefer this approach.
An approach I have used with functions is to have the function return a simple struct which contains two members, a status code and the return value. For example:
struct FuncRet {
short sStatus; // status or error code
double dValue; // calculated value
};
struct FuncRet Func(double dInput)
{
struct FuncRet = {0, 0}; // sStatus == 0 indicates success
// calculate return value FuncRet.dValue and set
// status code FuncRet.sStatus in the event of an error.
return FuncRet;
}
// ... source code before using our function.
{
struct FuncRet s;
if ((s = Func(aDble)).sStatus == 0) {
// do things with the valid value s.dValue
} else {
// error so deal with the error reported in s.sStatus
}
}
This allows me to do an immediate check for an error. Many functions end up returning a status without returning an actual value as well because the data returned is complex. One or more arguments may be modified by the function but the function doesn't return a value other than a status code.
What is the intention to set handle to an object as pointer-to pointer but not pointer? Like following code:
FT_Library library;
FT_Error error = FT_Init_FreeType( &library );
where
typedef struct FT_LibraryRec_ *FT_Library
so &library is a FT_LIBraryRec_ handle of type FT_LIBraryRec_**
It's a way to emulate pass by reference in C, which otherwise only have pass by value.
The 'C' library function FT_Init_FreeType has two outputs, the error code and/or the library handle (which is a pointer).
In C++ we'd more naturally either:
return an object which encapsulated the success or failure of the call and the library handle, or
return one output - the library handle, and throw an exception on failure.
C APIs are generally not implemented this way.
It is not unusual for a C Library function to return a success code, and to be passed the addresses of in/out variables to be conditionally mutated, as per the case above.
The approach hides implementation. It speeds up compilation of your code. It allows to upgrade data structures used by the library without breaking existing code that uses them. Finally, it makes sure the address of that object never changes, and that you don’t copy these objects.
Here’s how the version with a single pointer might be implemented:
struct FT_Struct
{
// Some fields/properties go here, e.g.
int field1;
char* field2;
}
FT_Error Init( FT_Struct* p )
{
p->field1 = 11;
p->field2 = malloc( 100 );
if( nullptr == p->field2 )
return E_OUTOFMEMORY;
return S_OK;
}
Or C++ equivalent, without any pointers:
class FT_Struct
{
int field1;
std::vector<char> field2;
public:
FT_Struct() :
field1( 11 )
{
field2.resize( 100 );
}
};
As a user of the library, you have to include struct/class FT_Struct definition. Libraries can be very complex so this will slow down compilation of your code.
If the library is dynamic i.e. *.dll on windows, *.so on linux or *.dylib on osx, you upgrade the library and if the new version changes memory layout of the struct/class, old applications will crash.
Because of the way C++ works, objects are passed by value, i.e. you normally expect them to be movable and copiable, which is not necessarily what library author wants to support.
Now consider the following function instead:
FT_Error Init( FT_Struct** pp )
{
try
{
*pp = new FT_Struct();
return S_OK;
}
catch( std::exception& ex )
{
return E_FAIL;
}
}
As a user of the library, you no longer need to know what’s inside FT_Struct or even what size it is. You don’t need to #include the implementation details, i.e. compilation will be faster.
This plays nicely with dynamic libraries, library author can change memory layout however they please, as long as the C API is stable, old apps will continue to work.
The API guarantees you won’t copy or move the values, you can’t copy structures of unknown lengths.
I am writing a large C program for embedded use. Every module in this program has an init() function (like a constructor) to set up its static variables.
The problem is that I have to remember to call all of these init functions from main(). I also have to remember to put them back if I have commented them out for some reason.
Is there anything clever I do to make sure that all of these functions are getting called? Something along the lines of putting a macro in each init function that, when you call a check_inited() function later, sends a warning to STDOUT if not all the functions are called.
I could increment a counter, but I'd have to maintain the correct number of init functions somewhere and that is also prone to error.
Thoughts?
The following is the solution I decided on, with input from several people in this thread
My goal is to make sure that all my init functions are actually being called. I want to do
this without maintaining lists or counts of modules across several files. I can't call
them automatically as Nick D suggested because they need to be called in a certain order.
To accomplish this, a macro included in every module uses the gcc constructor attribute to
add the init function name to a global list.
Another macro included in the body of the init function updates the global list to make a
note that the function was actually called.
Finally, a check function is called in main() after all of the inits are done.
Notes:
I chose to copy the strings into an array. This not strictly necessary because the
function names passed will always be static strings in normal usage. If memory was short
you could just store a pointer to the string that was passed in.
My reusable library of utility functions is called "nx_lib". Thus all the 'nxl' designations.
This isn't the most efficient code in the world but it's only called a boot time so that
doesn't matter for me.
There are two lines of code that need to be added to each module. If either is omitted,
the check function will let you know.
you might be able to make the constructor function static, which would avoid the need to give it a name that is unique across the project.
this code is only lightly tested and it's really late so please check carefully before trusting it.
Thank you to:
pierr who introduced me to the constructor attribute.
Nick D for demonstrating the ## preprocessor trick and giving me the framework.
tod frye for a clever linker-based approach that will work with many compilers.
Everyone else for helping out and sharing useful tidbits.
nx_lib_public.h
This is the relevant fragment of my library header file
#define NX_FUNC_RUN_CHECK_NAME_SIZE 20
typedef struct _nxl_function_element{
char func[NX_FUNC_RUN_CHECK_NAME_SIZE];
BOOL called;
} nxl_function_element;
void nxl_func_run_check_add(char *func_name);
BOOL nxl_func_run_check(void);
void nxl_func_run_check_hit(char *func_name);
#define NXL_FUNC_RUN_CHECK_ADD(function_name) \
void cons_ ## function_name() __attribute__((constructor)); \
void cons_ ## function_name() { nxl_func_run_check_add(#function_name); }
nxl_func_run_check.c
This is the libary code that is called to add function names and check them later.
#define MAX_CHECKED_FUNCTIONS 100
static nxl_function_element m_functions[MAX_CHECKED_FUNCTIONS];
static int m_func_cnt = 0;
// call automatically before main runs to register a function name.
void nxl_func_run_check_add(char *func_name)
{
// fail and complain if no more room.
if (m_func_cnt >= MAX_CHECKED_FUNCTIONS) {
print ("nxl_func_run_check_add failed, out of space\r\n");
return;
}
strncpy (m_functions[m_func_cnt].func, func_name,
NX_FUNC_RUN_CHECK_NAME_SIZE);
m_functions[m_func_cnt].func[NX_FUNC_RUN_CHECK_NAME_SIZE-1] = 0;
m_functions[m_func_cnt++].called = FALSE;
}
// call from inside the init function
void nxl_func_run_check_hit(char *func_name)
{
int i;
for (i=0; i< m_func_cnt; i++) {
if (! strncmp(m_functions[i].func, func_name,
NX_FUNC_RUN_CHECK_NAME_SIZE)) {
m_functions[i].called = TRUE;
return;
}
}
print("nxl_func_run_check_hit(): error, unregistered function was hit\r\n");
}
// checks that all registered functions were called
BOOL nxl_func_run_check(void) {
int i;
BOOL success=TRUE;
for (i=0; i< m_func_cnt; i++) {
if (m_functions[i].called == FALSE) {
success = FALSE;
xil_printf("nxl_func_run_check error: %s() not called\r\n",
m_functions[i].func);
}
}
return success;
}
solo.c
This is an example of a module that needs initialization
#include "nx_lib_public.h"
NXL_FUNC_RUN_CHECK_ADD(solo_init)
void solo_init(void)
{
nxl_func_run_check_hit((char *) __func__);
/* do module initialization here */
}
You can use gcc's extension __attribute__((constructor)) if gcc is ok for your project.
#include <stdio.h>
void func1() __attribute__((constructor));
void func2() __attribute__((constructor));
void func1()
{
printf("%s\n",__func__);
}
void func2()
{
printf("%s\n",__func__);
}
int main()
{
printf("main\n");
return 0;
}
//the output
func2
func1
main
I don't know how ugly the following looks but I post it anyway :-)
(The basic idea is to register function pointers, like what atexit function does.
Of course atexit implementation is different)
In the main module we can have something like this:
typedef int (*function_t)(void);
static function_t vfunctions[100]; // we can store max 100 function pointers
static int vcnt = 0; // count the registered function pointers
int add2init(function_t f)
{
// todo: error checks
vfunctions[vcnt++] = f;
return 0;
}
...
int main(void) {
...
// iterate vfunctions[] and call the functions
...
}
... and in some other module:
typedef int (*function_t)(void);
extern int add2init(function_t f);
#define M_add2init(function_name) static int int_ ## function_name = add2init(function_name)
int foo(void)
{
printf("foo\n");
return 0;
}
M_add2init(foo); // <--- register foo function
Why not write a post processing script to do the checking for you. Then run that script as part of your build process... Or better yet, make it one of your tests. You are writing tests, right? :)
For example, if each of your modules has a header file, modX.c. And if the signature of your init() function is "void init()"...
Have your script grep through all your .h files, and create a list of module names that need to be init()ed. Then have the script check that init() is indeed called on each module in main().
If your single module represents "class" entity and has instance constructor, you can use following construction:
static inline void init(void) { ... }
static int initialized = 0;
#define INIT if (__predict_false(!initialized)) { init(); initialized = 1; }
struct Foo *
foo_create(void)
{
INIT;
...
}
where "__predict_false" is your compiler's branch prediction hint. When first object is created, module is auto-initialized (for once).
Splint (and probably other Lint variants) can give a warning about functions that are defined but not called.
It's interesting that most compilers will warn you about unused variables, but not unused functions.
Larger running time is not a problem
You can conceivably implement a kind of "state-machine" for each module, wherein the actions of a function depend on the state the module is in. This state can be set to BEFORE_INIT or INITIALIZED.
For example, let's say we have module A with functions foo and bar.
The actual logic of the functions (i.e., what they actually do) would be declared like so:
void foo_logic();
void bar_logic();
Or whatever the signature is.
Then, the actual functions of the module (i.e., the actual function declared foo()) will perform a run-time check of the condition of the module, and decide what to do:
void foo() {
if (module_state == BEFORE_INIT) {
handle_not_initialized_error();
}
foo_logic();
}
This logic is repeated for all functions.
A few things to note:
This will obviously incur a huge penalty performance-wise, so is
probably not a good idea (I posted
anyway because you said runtime is
not a problem).
This is not a real state-machine, since there are only two states which are checked using a basic if, without some kind of smart general logic.
This kind of "design-pattern" works great when you're using separate threads/tasks, and the functions you're calling are actually called using some kind of IPC.
A state machine can be nicely implemented in C++, might be worth reading up on it. The same kind of idea can conceivably be coded in C with arrays of function pointers, but it's almost certainly not worth your time.
you can do something along these lines with a linker section. whenever you define an init function, place a pointer to it in a linker section just for init function pointers. then you can at least find out how many init functions have been compiled.
and if it does not matter what order the init functions are called, and the all have the same prototype, you can just call them all in a loop from main.
the exact details elude my memory, but it works soemthing like this::
in the module file...
//this is the syntax in GCC..(or would be if the underscores came through in this text editor)
initFuncPtr thisInit __attribute((section(.myinits)))__= &moduleInit;
void moduleInit(void)
{
// so init here
}
this places a pointer to the module init function in the .myinits section, but leaves the code in the .code section. so the .myinits section is nothing but pointers. you can think of this as a variable length array that module files can add to.
then you can access the section start and end address from the main. and go from there.
if the init functions all have the same protoytpe, you can just iterate over this section, calling them all.
this, in effect, is creating your own static constructor system in C.
if you are doing a large project and your linker is not at least this fully featured, you may have a problem...
Can I put up an answer to my question?
My idea was to have each function add it's name to a global list of functions, like Nick D's solution.
Then I would run through the symbol table produced by -gstab, and look for any functions named init_* that had not been called.
This is an embedded app so I have the elf image handy in flash memory.
However I don't like this idea because it means I always have to include debugging info in the binary.
Im getting into kernel work for a bit of my summer research. We are looking to make modifications to the TCP, in specific RTT calculations. What I would like to do is replace the resolution of one of the functions in tcp_input.c to a function provided by a dynamically loaded kernel module. I think this would improve the pace at which we can develop and distribute the modification.
The function I'm interested in was declared as static, however I've recompiled the kernel with the function non-static and exported by EXPORT_SYMBOL. This means the function is now accessible to other modules/parts of the kernel. I have verified this by "cat /proc/kallsyms".
Now I'd like to be able to load a module that can rewrite the symbol address from the initial to my dynamically loaded function. Similarly, when the module is to be unloaded, it would restore the original address. Is this a feasible approach? Do you all have suggestions how this might be better implemented?
Thanks!
Same as Overriding functionality with modules in Linux kernel
Edit:
This was my eventual approach.
Given the following function (which I wanted to override, and is not exported):
static void internal_function(void)
{
// do something interesting
return;
}
modify like so:
static void internal_function_original(void)
{
// do something interesting
return;
}
static void (*internal_function)(void) = &internal_function_original;
EXPORT_SYMBOL(internal_function);
This redefines the expected function identifier instead as a function pointer (which can be called in a similar manner) pointing to the original implementation. EXPORT_SYMBOL() makes the address globally accessible, so we can modify it from a module (or other kernel location).
Now you can write a kernel module with the following form:
static void (*original_function_reference)(void);
extern void (*internal_function)(void);
static void new_function_implementation(void)
{
// do something new and interesting
// return
}
int init_module(void)
{
original_function_reference = internal_function;
internal_function = &new_function_implementation;
return 0;
}
void cleanup_module(void)
{
internal_function = original_function_reference;
}
This module replaces the original implementation with a dynamically loaded version. Upon unloading, the original reference (and implementation) is restored. In my specific case, I provided a new estimator for the RTT in TCP. By using a module, I am able to make small tweaks and restart testing, all without having to recompile and reboot the kernel.
I'm not sure that'll work - I believe the symbol resolution for the internal calls to the function you want to replace will have already been done by the time your module loads.
Instead, you could change the code by renaming the existing function, then creating a global function pointer with the original name of the function. Initialise the function pointer to the address of the internal function, so the existing code will work unmodified. Export the symbol of the global function pointer, then your module can just change its value by assignment at module load and unload time.
I once made a proof of concept of a hijack module that inserted it's own function in place of kernel function.
I just so happens that the new kernel tacing architecture uses a very similar system.
I injected my own function in the kernel by overwriting the first couple of bytes of code with a jump pointing to my custom function. As soon as the real function gets called, it jumps instead to my function that after it had done it's work called the original function.
#include <linux/module.h>
#include <linux/kernel.h>
#define CODESIZE 12
static unsigned char original_code[CODESIZE];
static unsigned char jump_code[CODESIZE] =
"\x48\xb8\x00\x00\x00\x00\x00\x00\x00\x00" /* movq $0, %rax */
"\xff\xe0" /* jump *%rax */
;
/* FILL THIS IN YOURSELF */
int (*real_printk)( char * fmt, ... ) = (int (*)(char *,...) )0xffffffff805e5f6e;
int hijack_start(void);
void hijack_stop(void);
void intercept_init(void);
void intercept_start(void);
void intercept_stop(void);
int fake_printk(char *, ... );
int hijack_start()
{
real_printk(KERN_INFO "I can haz hijack?\n" );
intercept_init();
intercept_start();
return 0;
}
void hijack_stop()
{
intercept_stop();
return;
}
void intercept_init()
{
*(long *)&jump_code[2] = (long)fake_printk;
memcpy( original_code, real_printk, CODESIZE );
return;
}
void intercept_start()
{
memcpy( real_printk, jump_code, CODESIZE );
}
void intercept_stop()
{
memcpy( real_printk, original_code, CODESIZE );
}
int fake_printk( char *fmt, ... )
{
int ret;
intercept_stop();
ret = real_printk(KERN_INFO "Someone called printk\n");
intercept_start();
return ret;
}
module_init( hijack_start );
module_exit( hijack_stop );
I'm warning you, when you're going to experiment with these kind of things, watch out for kernel panics and other disastrous events. I would advise you to do this in a virtualised environment. This is a proof-of-concept code I wrote a while ago, I'm not sure it still works.
It's a really easy principle, but very effective. Of course, a real solution would use locks to make sure nobody would call the function while you're overwriting it.
Have fun!
You can try using ksplice - you don't even need to make it non static.
I think what you want is Kprobe.
Another way that caf has mentioned is to add a hook to the original routine, and register/unregister hook in the module.