A Guide to Improving Software Security

This document provides guidelines for improving software security through secure coding. It covers common programming languages and libraries, and focuses on concrete recommendations.

title logo

Copyright 2012-2018 Red Hat, Inc.

The text of and illustrations in this document are licensed by Red Hat under a Creative Commons Attribution–Share Alike 3.0 Unported license ("CC-BY-SA"). An explanation of CC-BY-SA is available at http://creativecommons.org/licenses/by-sa/3.0/. The original authors of this document, and Red Hat, designate the Fedora Project as the "Attribution Party" for purposes of CC-BY-SA. In accordance with CC-BY-SA, if you distribute this document or an adaptation of it, you must provide the URL for the original version.

Red Hat, as the licensor of this document, waives the right to enforce, and agrees not to assert, Section 4d of CC-BY-SA to the fullest extent permitted by applicable law.

Red Hat, Red Hat Enterprise Linux, the Shadowman logo, JBoss, MetaMatrix, Fedora, the Infinity Logo, and RHCE are trademarks of Red Hat, Inc., registered in the United States and other countries.

For guidelines on the permitted uses of the Fedora trademarks, refer to https://fedoraproject.org/wiki/Legal:Trademark_guidelines.

Linux is the registered trademark of Linus Torvalds in the United States and other countries.

Java is a registered trademark of Oracle and/or its affiliates.

XFS is a trademark of Silicon Graphics International Corp. or its subsidiaries in the United States and/or other countries.

MySQL is a registered trademark of MySQL AB in the United States, the European Union and other countries.

All other trademarks are the property of their respective owners.

Florian Weimer

Red Hat Product Security Team

Nikos Mavrogiannopoulos

Red Hat Crypto Team

Robert Relyea

Red Hat Crypto Team

1. Programming Languages

1.1. The C Programming Language

1.1.1. The Core Language

C provides no memory safety. Most recommendations in this section deal with this aspect of the language.

Undefined Behavior

Some C constructs are defined to be undefined by the C standard. This does not only mean that the standard does not describe what happens when the construct is executed. It also allows optimizing compilers such as GCC to assume that this particular construct is never reached. In some cases, this has caused GCC to optimize security checks away. (This is not a flaw in GCC or the C language. But C certainly has some areas which are more difficult to use than others.)

Common sources of undefined behavior are:

  • out-of-bounds array accesses

  • null pointer dereferences

  • overflow in signed integer arithmetic

Recommendations for Pointers and Array Handling

Always keep track of the size of the array you are working with. Often, code is more obviously correct when you keep a pointer past the last element of the array, and calculate the number of remaining elements by substracting the current position from that pointer. The alternative, updating a separate variable every time when the position is advanced, is usually less obviously correct.

Array processing in C shows how to extract Pascal-style strings from a character buffer. The two pointers kept for length checks are inend and outend. inp and outp are the respective positions. The number of input bytes is checked using the expression len > (size_t)(inend - inp). The cast silences a compiler warning; inend is always larger than inp.

Example 1. Array processing in C

extract_strings(const char *in, size_t inlen, char **out, size_t outlen)
  const char *inp = in;
  const char *inend = in + inlen;
  char **outp = out;
  char **outend = out + outlen;

  while (inp != inend) {
    size_t len;
    char *s;
    if (outp == outend) {
      errno = ENOSPC;
      goto err;
    len = (unsigned char)*inp;
    if (len > (size_t)(inend - inp)) {
      errno = EINVAL;
      goto err;
    s = malloc(len + 1);
    if (s == NULL) {
      goto err;
    memcpy(s, inp, len);
    inp += len;
    s[len] = '\0';
    *outp = s;
  return outp - out;
    int errno_old = errno;
    while (out != outp) {
    errno = errno_old;
  return -1;

It is important that the length checks always have the form len > (size_t)(inend - inp), where len is a variable of type size_t which denotes the total number of bytes which are about to be read or written next. In general, it is not safe to fold multiple such checks into one, as in len1 + len2 > (size_t)(inend - inp), because the expression on the left can overflow or wrap around (see Recommendations for Integer Arithmetic), and it no longer reflects the number of bytes to be processed.

Recommendations for Integer Arithmetic

Overflow in signed integer arithmetic is undefined. This means that it is not possible to check for overflow after it happened, see Incorrect overflow detection in C.

Example 2. Incorrect overflow detection in C

void report_overflow(void);

add(int a, int b)
  int result = a + b;
  if (a < 0 || b < 0) {
    return -1;
  // The compiler can optimize away the following if statement.
  if (result < 0) {
  return result;

The following approaches can be used to check for overflow, without actually causing it.

  • Use a wider type to perform the calculation, check that the result is within bounds, and convert the result to the original type. All intermediate results must be checked in this way.

  • Perform the calculation in the corresponding unsigned type and use bit fiddling to detect the overflow. Overflow checking for unsigned addition shows how to perform an overflow check for unsigned integer addition. For three or more terms, all the intermediate additions have to be checked in this way.

Example 3. Overflow checking for unsigned addition

void report_overflow(void);

add_unsigned(unsigned a, unsigned b)
  unsigned sum = a + b;
  if (sum < a) { // or sum < b
  return sum;
Example 4. Overflow checking for unsigned multiplication

mul(unsigned a, unsigned b)
  if (b && a > ((unsigned)-1) / b) {
  return a * b;

Basic arithmetic operations are commutative, so for bounds checks, there are two different but mathematically equivalent expressions. Sometimes, one of the expressions results in better code because parts of it can be reduced to a constant. This applies to overflow checks for multiplication a * b involving a constant a, where the expression is reduced to b > C for some constant C determined at compile time. The other expression, b && a > ((unsigned)-1) / b, is more difficult to optimize at compile time.

When a value is converted to a signed integer, GCC always chooses the result based on 2’s complement arithmetic. This GCC extension (which is also implemented by other compilers) helps a lot when implementing overflow checks.

Sometimes, it is necessary to compare unsigned and signed integer variables. This results in a compiler warning, comparison between signed and unsigned integer expressions, because the comparison often gives unexpected results for negative values. When adding a cast, make sure that negative values are covered properly. If the bound is unsigned and the checked quantity is signed, you should cast the checked quantity to an unsigned type as least as wide as either operand type. As a result, negative values will fail the bounds check. (You can still check for negative values separately for clarity, and the compiler will optimize away this redundant check.)

Legacy code should be compiled with the -fwrapv GCC option. As a result, GCC will provide 2’s complement semantics for integer arithmetic, including defined behavior on integer overflow.

Global Variables

Global variables should be avoided because they usually lead to thread safety hazards. In any case, they should be declared static, so that access is restricted to a single translation unit.

Global constants are not a problem, but declaring them can be tricky. Declaring a constant array of constant strings shows how to declare a constant array of constant strings. The second const is needed to make the array constant, and not just the strings. It must be placed after the *, and not before it.

Example 5. Declaring a constant array of constant strings

static const char *const string_list[] = {

Sometimes, static variables local to functions are used as a replacement for proper memory management. Unlike non-static local variables, it is possible to return a pointer to static local variables to the caller. But such variables are well-hidden, but effectively global (just as static variables at file scope). It is difficult to add thread safety afterwards if such interfaces are used. Merely dropping the static keyword in such cases leads to undefined behavior.

Another source for static local variables is a desire to reduce stack space usage on embedded platforms, where the stack may span only a few hundred bytes. If this is the only reason why the static keyword is used, it can just be dropped, unless the object is very large (larger than 128 kilobytes on 32-bit platforms). In the latter case, it is recommended to allocate the object using malloc, to obtain proper array checking, for the same reasons outlined in alloca and Other Forms of Stack-based Allocation.

1.1.2. The C Standard Library

Parts of the C standard library (and the UNIX and GNU extensions) are difficult to use, so you should avoid them.

Please check the applicable documentation before using the recommended replacements. Many of these functions allocate buffers using malloc which your code must deallocate explicitly using free.

Absolutely Banned Interfaces

The functions listed below must not be used because they are almost always unsafe. Use the indicated replacements instead.

  • getsfgets

  • getwdgetcwd or get_current_dir_name

  • readdir_rreaddir

  • realpath (with a non-NULL second parameter) ⟶ realpath with NULL as the second parameter, or canonicalize_file_name

The constants listed below must not be used, either. Instead, code must allocate memory dynamically and use interfaces with length checking.

  • NAME_MAX (limit not actually enforced by the kernel)

  • PATH_MAX (limit not actually enforced by the kernel)

  • _PC_NAME_MAX (This limit, returned by the pathconf function, is not enforced by the kernel.)

  • _PC_PATH_MAX (This limit, returned by the pathconf function, is not enforced by the kernel.)

The following structure members must not be used.

  • f_namemax in struct statvfs (limit not actually enforced by the kernel, see _PC_NAME_MAX above)

Functions to Avoid

The following string manipulation functions can be used securely in principle, but their use should be avoided because they are difficult to use correctly. Calls to these functions can be replaced with asprintf or vasprintf. (For non-GNU targets, these functions are available from Gnulib.) In some cases, the snprintf function might be a suitable replacement, see String Functions with Explicit Length Arguments.

  • sprintf

  • strcat

  • strcpy

  • vsprintf

Use the indicated replacements for the functions below.

String Functions with Explicit Length Arguments

The C run-time library provides string manipulation functions which not just look for NUL characters for string termination, but also honor explicit lengths provided by the caller. However, these functions evolved over a long period of time, and the lengths mean different things depending on the function.


The snprintf function provides a way to construct a string in a statically-sized buffer. (If the buffer size is allocated on the heap, consider use asprintf instead.)

char fraction[30];
snprintf(fraction, sizeof(fraction), "%d/%d", numerator, denominator);

The second argument to the snprintf call should always be the size of the buffer in the first argument (which should be a character array). Elaborate pointer and length arithmetic can introduce errors and nullify the security benefits of snprintf.

In particular, snprintf is not well-suited to constructing a string iteratively, by appending to an existing buffer. snprintf returns one of two values, -1 on errors, or the number of characters which would have been written to the buffer if the buffer were large enough. This means that adding the result of snprintf to the buffer pointer to skip over the characters just written is incorrect and risky. However, as long as the length argument is not zero, the buffer will remain null-terminated. Repeatedly writing to a buffer using snprintf works because end -current > 0 is a loop invariant. After the loop, the result string is in the buf variable.

Example 6. Repeatedly writing to a buffer using snprintf

char buf[512];
char *current = buf;
const char *const end = buf + sizeof(buf);
for (struct item *it = data; it->key; ++it) {
  snprintf(current, end - current, "%s%s=%d",
	       current == buf ? "" : ", ", it->key, it->value);
  current += strlen(current);

If you want to avoid the call to strlen for performance reasons, you have to check for a negative return value from snprintf and also check if the return value is equal to the specified buffer length or larger. Only if neither condition applies, you may advance the pointer to the start of the write buffer by the number return by snprintf. However, this optimization is rarely worthwhile.

Note that it is not permitted to use the same buffer both as the destination and as a source argument.

vsnprintf and Format Strings

If you use vsnprintf (or vasprintf or even snprintf) with a format string which is not a constant, but a function argument, it is important to annotate the function with a format function attribute, so that GCC can warn about misuse of your function (see The format function attribute).

Example 7. The format function attribute

void log_format(const char *format, ...) __attribute__((format(printf, 1, 2)));

log_format(const char *format, ...)
  char buf[1000];
  va_list ap;
  va_start(ap, format);
  vsnprintf(buf, sizeof(buf), format, ap);

The strncpy function does not ensure that the target buffer is null-terminated. A common idiom for ensuring NUL termination is:

char buf[10];
strncpy(buf, data, sizeof(buf));
buf[sizeof(buf) - 1] = '\0';

Another approach uses the strncat function for this purpose:

buf[0] = '\0';
strncat(buf, data, sizeof(buf) - 1);

The length argument of the strncat function specifies the maximum number of characters copied from the source buffer, excluding the terminating NUL character. This means that the required number of bytes in the destination buffer is the length of the original string, plus the length argument in the strncat call, plus one. Consequently, this function is rarely appropriate for performing a length-checked string operation, with the notable exception of the strcpy emulation described in strncpy.

To implement a length-checked string append, you can use an approach similar to Repeatedly writing to a buffer using snprintf:

char buf[10];
snprintf(buf, sizeof(buf), "%s", prefix);
snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), "%s", data);

In many cases, including this one, the string concatenation can be avoided by combining everything into a single format string:

snprintf(buf, sizeof(buf), "%s%s", prefix, data);

But you should must not dynamically construct format strings to avoid concatenation because this would prevent GCC from type-checking the argument lists.

It is not possible to use format strings like "%s%s" to implement concatenation, unless you use separate buffers. snprintf does not support overlapping source and target strings.

strlcpy and strlcat

Some systems support strlcpy and strlcat functions which behave this way, but these functions are not part of GNU libc. strlcpy is often replaced with snprintf with a "%s" format string. See strncpy for a caveat related to the snprintf return value.

To emulate strlcat, use the approach described in strncat.

ISO C11 Annex K *_s functions

ISO C11 adds another set of length-checking functions, but GNU libc currently does not implement them.

Other strn* and stpn* functions

GNU libc contains additional functions with different variants of length checking. Consult the documentation before using them to find out what the length actually means.

1.1.3. Memory Allocators

The C library interfaces for memory allocation are provided by malloc, free and realloc, and the calloc function. In addition to these generic functions, there are derived functions such as strdup which perform allocation using malloc internally, but do not return untyped heap memory (which could be used for any object).

The C compiler knows about these functions and can use their expected behavior for optimizations. For instance, the compiler assumes that an existing pointer (or a pointer derived from an existing pointer by arithmetic) will not point into the memory area returned by malloc.

If the allocation fails, realloc does not free the old pointer. Therefore, the idiom ptr = realloc(ptr, size); is wrong because the memory pointed to by ptr leaks in case of an error.

Use-after-free errors

After free, the pointer is invalid. Further pointer dereferences are not allowed (and are usually detected by valgrind). Less obvious is that any use of the old pointer value is not allowed, either. In particular, comparisons with any other pointer (or the null pointer) are undefined according to the C standard.

The same rules apply to realloc if the memory area cannot be enlarged in-place. For instance, the compiler may assume that a comparison between the old and new pointer will always return false, so it is impossible to detect movement this way.

Handling Memory Allocation Errors

Recovering from out-of-memory errors is often difficult or even impossible. In these cases, malloc and other allocation functions return a null pointer. Dereferencing this pointer lead to a crash. Such dereferences can even be exploitable for code execution if the dereference is combined with an array subscript.

In general, if you cannot check all allocation calls and handle failure, you should abort the program on allocation failure, and not rely on the null pointer dereference to terminate the process. See Recommendations for Manually-written Decoders for related memory allocation concerns.

alloca and Other Forms of Stack-based Allocation

Allocation on the stack is risky because stack overflow checking is implicit. There is a guard page at the end of the memory area reserved for the stack. If the program attempts to read from or write to this guard page, a SIGSEGV signal is generated and the program typically terminates.

This is sufficient for detecting typical stack overflow situations such as unbounded recursion, but it fails when the stack grows in increments larger than the size of the guard page. In this case, it is possible that the stack pointer ends up pointing into a memory area which has been allocated for a different purposes. Such misbehavior can be exploitable.

A common source for large stack growth are calls to alloca and related functions such as strdupa. These functions should be avoided because of the lack of error checking. (They can be used safely if the allocated size is less than the page size (typically, 4096 bytes), but this case is relatively rare.) Additionally, relying on alloca makes it more difficult to reorganize the code because it is not allowed to use the pointer after the function calling alloca has returned, even if this function has been inlined into its caller.

Similar concerns apply to variable-length arrays (VLAs), a feature of the C99 standard which started as a GNU extension. For large objects exceeding the page size, there is no error checking, either.

In both cases, negative or very large sizes can trigger a stack-pointer wraparound, and the stack pointer and end up pointing into caller stack frames, which is fatal and can be exploitable.

If you want to use alloca or VLAs for performance reasons, consider using a small on-stack array (less than the page size, large enough to fulfill most requests). If the requested size is small enough, use the on-stack array. Otherwise, call malloc. When exiting the function, check if malloc had been called, and free the buffer as needed.

Array Allocation

When allocating arrays, it is important to check for overflows. The calloc function performs such checks.

If malloc or realloc is used, the size check must be written manually. For instance, to allocate an array of n elements of type T, check that the requested size is not greater than ((size_t) -1) / sizeof(T). See Recommendations for Integer Arithmetic.

Custom Memory Allocators

Custom memory allocates come in two forms: replacements for malloc, and completely different interfaces for memory management. Both approaches can reduce the effectiveness of valgrind and similar tools, and the heap corruption detection provided by GNU libc, so they should be avoided.

Memory allocators are difficult to write and contain many performance and security pitfalls.

  • When computing array sizes or rounding up allocation requests (to the next allocation granularity, or for alignment purposes), checks for arithmetic overflow are required.

  • Size computations for array allocations need overflow checking. See Array Allocation.

  • It can be difficult to beat well-tuned general-purpose allocators. In micro benchmarks, pool allocators can show huge wins, and size-specific pools can reduce internal fragmentation. But often, utilization of individual pools is poor, and external fragmentation increases the overall memory usage.

Conservative Garbage Collection

Garbage collection can be an alternative to explicit memory management using malloc and free. The Boehm-Dehmers-Weiser allocator can be used from C programs, with minimal type annotations. Performance is competitive with malloc on 64-bit architectures, especially for multi-threaded programs. The stop-the-world pauses may be problematic for some real-time applications, though.

However, using a conservative garbage collector may reduce opportunities for code reduce because once one library in a program uses garbage collection, the whole process memory needs to be subject to it, so that no pointers are missed. The Boehm-Dehmers-Weiser collector also reserves certain signals for internal use, so it is not fully transparent to the rest of the program.

1.1.4. Other C-related Topics

Wrapper Functions

Some libraries provide wrappers for standard library functions. Common cases include allocation functions such as xmalloc which abort the process on allocation failure (instead of returning a NULL pointer), or alternatives to relatively recent library additions such as snprintf (along with implementations for systems which lack them).

In general, such wrappers are a bad idea, particularly if they are not implemented as inline functions or preprocessor macros. The compiler lacks knowledge of such wrappers outside the translation unit which defines them, which means that some optimizations and security checks are not performed. Adding attribute annotations to function declarations can remedy this to some extent, but these annotations have to be maintained carefully for feature parity with the standard implementation.

At the minimum, you should apply these attributes:

  • If you wrap function which accepts are GCC-recognized format string (for example, a printf-style function used for logging), you should add a suitable format attribute, as in The format function attribute.

  • If you wrap a function which carries a warn_unused_result attribute and you propagate its return value, your wrapper should be declared with warn_unused_result as well.

  • Duplicating the buffer length checks based on the __builtin_object_size GCC builtin is desirable if the wrapper processes arrays. (This functionality is used by the -D_FORTIFY_SOURCE=2 checks to guard against static buffer overflows.) However, designing appropriate interfaces and implementing the checks may not be entirely straightforward.

For other attributes (such as malloc), careful analysis and comparison with the compiler documentation is required to check if propagating the attribute is appropriate. Incorrectly applied attributes can result in undesired behavioral changes in the compiled code.

1.2. The C++ Programming Language

1.2.1. The Core Language

C++ includes a large subset of the C language. As far as the C subset is used, the recommendations in The C Programming Language apply.

Array Allocation with operator new[]

For very large values of n, an expression like new T[n] can return a pointer to a heap region which is too small. In other words, not all array elements are actually backed with heap memory reserved to the array. Current GCC versions generate code that performs a computation of the form sizeof(T) * size_t(n)
, where cookie_size is currently at most 8. This computation can overflow, and GCC versions prior to 4.8 generated code which did not detect this. (Fedora 18 was the first release which fixed this in GCC.)

The std::vector template can be used instead an explicit array allocation. (The GCC implementation detects overflow internally.)

If there is no alternative to operator new[] and the sources will be compiled with older GCC versions, code which allocates arrays with a variable length must check for overflow manually. For the new T[n] example, the size check could be n || (n > 0 && n > (size_t(-1) - 8) / sizeof(T)). (See Recommendations for Integer Arithmetic.) If there are additional dimensions (which must be constants according to the C++ standard), these should be included as factors in the divisor.

These countermeasures prevent out-of-bounds writes and potential code execution. Very large memory allocations can still lead to a denial of service. Recommendations for Manually-written Decoders contains suggestions for mitigating this problem when processing untrusted data.

See Array Allocation for array allocation advice for C-style memory allocation.


Do not overload functions with versions that have different security characteristics. For instance, do not implement a function strcat which works on std::string arguments. Similarly, do not name methods after such functions.

ABI compatibility and preparing for security updates

A stable binary interface (ABI) is vastly preferred for security updates. Without a stable ABI, all reverse dependencies need recompiling, which can be a lot of work and could even be impossible in some cases. Ideally, a security update only updates a single dynamic shared object, and is picked up automatically after restarting affected processes.

Outside of extremely performance-critical code, you should ensure that a wide range of changes is possible without breaking ABI. Some very basic guidelines are:

  • Avoid inline functions.

  • Use the pointer-to-implementation idiom.

  • Try to avoid templates. Use them if the increased type safety provides a benefit to the programmer.

  • Move security-critical code out of templated code, so that it can be patched in a central place if necessary.

The KDE project publishes a document with more extensive guidelines on ABI-preserving changes to C++ code, Policies/Binary Compatibility Issues With C++ (d-pointer refers to the pointer-to-implementation idiom).

C++0X and C++11 Support

GCC offers different language compatibility modes:

  • -std=c++98 for the original 1998 C++ standard

  • -std=c++03 for the 1998 standard with the changes from the TR1 technical report

  • -std=c++11 for the 2011 C++ standard. This option should not be used.

  • -std=c++0x for several different versions of C++11 support in development, depending on the GCC version. This option should not be used.

For each of these flags, there are variants which also enable GNU extensions (mostly language features also found in C99 or C11):

  • -std=gnu++98

  • -std=gnu++03

  • -std=gnu++11

Again, -std=gnu++11 should not be used.

If you enable C++11 support, the ABI of the standard C++ library libstdc++ will change in subtle ways. Currently, no C++ libraries are compiled in C++11 mode, so if you compile your code in C++11 mode, it will be incompatible with the rest of the system. Unfortunately, this is also the case if you do not use any C++11 features. Currently, there is no safe way to enable C++11 mode (except for freestanding applications).

The meaning of C++0X mode changed from GCC release to GCC release. Earlier versions were still ABI-compatible with C++98 mode, but in the most recent versions, switching to C++0X mode activates C++11 support, with its compatibility problems.

Some C++11 features (or approximations thereof) are available with TR1 support, that is, with -std=c03` or [option]`-std=gnu03 and in the <tr1/*> header files. This includes std::tr1::shared_ptr (from <tr1/memory>) and std::tr1::function (from <tr1/functional>). For other C++11 features, the Boost C++ library contains replacements.

1.2.2. The C++ Standard Library

The C++ standard library includes most of its C counterpart by reference, see The C Standard Library.

Functions That Are Difficult to Use

This section collects functions and function templates which are part of the standard library and are difficult to use.

Unpaired Iterators

Functions which use output operators or iterators which do not come in pairs (denoting ranges) cannot perform iterator range checking. (See Iterators) Function templates which involve output iterators are particularly dangerous:

  • std::copy

  • std::copy_backward

  • std::copy_if

  • std::move (three-argument variant)

  • std::move_backward

  • std::partition_copy_if

  • std::remove_copy

  • std::remove_copy_if

  • std::replace_copy

  • std::replace_copy_if

  • std::swap_ranges

  • std::transform

In addition, std::copy_n, std::fill_n and std::generate_n do not perform iterator checking, either, but there is an explicit count which has to be supplied by the caller, as opposed to an implicit length indicator in the form of a pair of forward iterators.

These output-iterator-expecting functions should only be used with unlimited-range output iterators, such as iterators obtained with the std::back_inserter function.

Other functions use single input or forward iterators, which can read beyond the end of the input range if the caller is not careful:

  • std::equal

  • std::is_permutation

  • std::mismatch

String Handling with std::string

The std::string class provides a convenient way to handle strings. Unlike C strings, std::string objects have an explicit length (and can contain embedded NUL characters), and storage for its characters is managed automatically. This section discusses std::string, but these observations also apply to other instances of the std::basic_string template.

The pointer returned by the data() member function does not necessarily point to a NUL-terminated string. To obtain a C-compatible string pointer, use c_str() instead, which adds the NUL terminator.

The pointers returned by the data() and c_str() functions and iterators are only valid until certain events happen. It is required that the exact std::string object still exists (even if it was initially created as a copy of another string object). Pointers and iterators are also invalidated when non-const member functions are called, or functions with a non-const reference parameter. The behavior of the GCC implementation deviates from that required by the C++ standard if multiple threads are present. In general, only the first call to a non-const member function after a structural modification of the string (such as appending a character) is invalidating, but this also applies to member function such as the non-const version of begin(), in violation of the C++ standard.

Particular care is necessary when invoking the c_str() member function on a temporary object. This is convenient for calling C functions, but the pointer will turn invalid as soon as the temporary object is destroyed, which generally happens when the outermost expression enclosing the expression on which c_str() is called completes evaluation. Passing the result of c_str() to a function which does not store or otherwise leak that pointer is safe, though.

Like with std::vector and std::array, subscribing with operator[] does not perform bounds checks. Use the at(size_type) member function instead. See Containers and operator[]. Furthermore, accessing the terminating NUL character using operator[] is not possible. (In some implementations, the c_str() member function writes the NUL character on demand.)

Never write to the pointers returned by data() or c_str() after casting away const. If you need a C-style writable string, use a std::vector<char> object and its data() member function. In this case, you have to explicitly add the terminating NUL character.

GCC’s implementation of std::string is currently based on reference counting. It is expected that a future version will remove the reference counting, due to performance and conformance issues. As a result, code that implicitly assumes sharing by holding to pointers or iterators for too long will break, resulting in run-time crashes or worse. On the other hand, non-const iterator-returning functions will no longer give other threads an opportunity for invalidating existing iterators and pointers because iterator invalidation does not depend on sharing of the internal character array object anymore.

Containers and operator[]

Many sequence containers similar to std::vector provide both operator[](size_type) and a member function at(size_type). This applies to std::vector itself, std::array, std::string and other instances of std::basic_string.

operator[](size_type) is not required by the standard to perform bounds checking (and the implementation in GCC does not). In contrast, at(size_type) must perform such a check. Therefore, in code which is not performance-critical, you should prefer at(size_type) over operator[](size_type), even though it is slightly more verbose.

The front() and back() member functions are undefined if a vector object is empty. You can use vec.at(0) and vec.at(vec.size() - 1) as checked replacements. For an empty vector, data() is defined; it returns an arbitrary pointer, but not necessarily the NULL pointer.


Iterators do not perform any bounds checking. Therefore, all functions that work on iterators should accept them in pairs, denoting a range, and make sure that iterators are not moved outside that range. For forward iterators and bidirectional iterators, you need to check for equality before moving the first or last iterator in the range. For random-access iterators, you need to compute the difference before adding or subtracting an offset. It is not possible to perform the operation and check for an invalid operator afterwards.

Output iterators cannot be compared for equality. Therefore, it is impossible to write code that detects that it has been supplied an output area that is too small, and their use should be avoided.

These issues make some of the standard library functions difficult to use correctly, see Unpaired Iterators.

1.3. The Java Programming Language

1.3.1. The Core Language

Implementations of the Java programming language provide strong memory safety, even in the presence of data races in concurrent code. This prevents a large range of security vulnerabilities from occurring, unless certain low-level features are used; see Low-level Features of the Virtual Machine.

Increasing Robustness when Reading Arrays

External data formats often include arrays, and the data is stored as an integer indicating the number of array elements, followed by this number of elements in the file or protocol data unit. This length specified can be much larger than what is actually available in the data source.

To avoid allocating extremely large amounts of data, you can allocate a small array initially and grow it as you read more data, implementing an exponential growth policy. See the readBytes(InputStream, int) function in Incrementally reading a byte array.

Example 8. Incrementally reading a byte array

static byte[] readBytes(InputStream in, int length) throws IOException {
	final int startSize = 65536;
    byte[] b = new byte[Math.min(length, startSize)];
    int filled = 0;
    while (true) {
        int remaining = b.length - filled;
	    readFully(in, b, filled, remaining);
        if (b.length == length) {
        filled = b.length;
        if (length - b.length <= b.length) {
            // Allocate final length.  Condition avoids overflow.
            b = Arrays.copyOf(b, length);
        } else {
            b = Arrays.copyOf(b, b.length * 2);
    return b;

static void readFully(InputStream in,byte[] b, int off, int len)
	    throws IOException {
	int startlen = len;
    while (len > 0) {
        int count = in.read(b, off, len);
        if (count < 0) {
            throw new EOFException();
        off += count;
        len -= count;

When reading data into arrays, hash maps or hash sets, use the default constructor and do not specify a size hint. You can simply add the elements to the collection as you read them.

Resource Management

Unlike C++, Java does not offer destructors which can deallocate resources in a predictable fashion. All resource management has to be manual, at the usage site. (Finalizers are generally not usable for resource management, especially in high-performance code; see Finalizers.)

The first option is the try-finally construct, as shown in Resource management with a try-finally block. The code in the finally block should be as short as possible and should not throw any exceptions.

Example 9. Resource management with a try-finally block

InputStream in = new BufferedInputStream(new FileInputStream(path));
try {
} finally {

Note that the resource allocation happens outside the try block, and that there is no null check in the finally block. (Both are common artifacts stemming from IDE code templates.)

If the resource object is created freshly and implements the java.lang.AutoCloseable interface, the code in Resource management using the try-with-resource construct can be used instead. The Java compiler will automatically insert the close() method call in a synthetic finally block.

Example 10. Resource management using the try-with-resource construct

try (InputStream in = new BufferedInputStream(new FileInputStream(path))) {

To be compatible with the try-with-resource construct, new classes should name the resource deallocation method close(), and implement the AutoCloseable interface (the latter breaking backwards compatibility with Java 6). However, using the try-with-resource construct with objects that are not freshly allocated is at best awkward, and an explicit finally block is usually the better approach.

In general, it is best to design the programming interface in such a way that resource deallocation methods like close() cannot throw any (checked or unchecked) exceptions, but this should not be a reason to ignore any actual error conditions.


Finalizers can be used a last-resort approach to free resources which would otherwise leak. Finalization is unpredictable, costly, and there can be a considerable delay between the last reference to an object going away and the execution of the finalizer. Generally, manual resource management is required; see Resource Management.

Finalizers should be very short and should only deallocate native or other external resources held directly by the object being finalized. In general, they must use synchronization: Finalization necessarily happens on a separate thread because it is inherently concurrent. There can be multiple finalization threads, and despite each object being finalized at most once, the finalizer must not assume that it has exclusive access to the object being finalized (in the this pointer).

Finalizers should not deallocate resources held by other objects, especially if those objects have finalizers on their own. In particular, it is a very bad idea to define a finalizer just to invoke the resource deallocation method of another object, or overwrite some pointer fields.

Finalizers are not guaranteed to run at all. For instance, the virtual machine (or the machine underneath) might crash, preventing their execution.

Objects with finalizers are garbage-collected much later than objects without them, so using finalizers to zero out key material (to reduce its undecrypted lifetime in memory) may have the opposite effect, keeping objects around for much longer and prevent them from being overwritten in the normal course of program execution.

For the same reason, code which allocates objects with finalizers at a high rate will eventually fail (likely with a java.lang.OutOfMemoryError exception) because the virtual machine has finite resources for keeping track of objects pending finalization. To deal with that, it may be necessary to recycle objects with finalizers.

The remarks in this section apply to finalizers which are implemented by overriding the finalize() method, and to custom finalization using reference queues.

Recovering from Exceptions and Errors

Java exceptions come in three kinds, all ultimately deriving from java.lang.Throwable:

  • Run-time exceptions do not have to be declared explicitly and can be explicitly thrown from any code, by calling code which throws them, or by triggering an error condition at run time, like division by zero, or an attempt at an out-of-bounds array access. These exceptions derive from from the java.lang.RuntimeException class (perhaps indirectly).

  • Checked exceptions have to be declared explicitly by functions that throw or propagate them. They are similar to run-time exceptions in other regards, except that there is no language construct to throw them (except the throw statement itself). Checked exceptions are only present at the Java language level and are only enforced at compile time. At run time, the virtual machine does not know about them and permits throwing exceptions from any code. Checked exceptions must derive (perhaps indirectly) from the java.lang.Exception class, but not from java.lang.RuntimeException.

  • Errors are exceptions which typically reflect serious error conditions. They can be thrown at any point in the program, and do not have to be declared (unlike checked exceptions). In general, it is not possible to recover from such errors; more on that below, in The Difficulty of Catching Errors. Error classes derive (perhaps indirectly) from java.lang.Error, or from java.lang.Throwable, but not from java.lang.Exception.

The general expection is that run-time errors are avoided by careful programming (e.g., not dividing by zero). Checked exception are expected to be caught as they happen (e.g., when an input file is unexpectedly missing). Errors are impossible to predict and can happen at any point and reflect that something went wrong beyond all expectations.

The Difficulty of Catching Errors

Errors (that is, exceptions which do not (indirectly) derive from java.lang.Exception), have the peculiar property that catching them is problematic. There are several reasons for this:

  • The error reflects a failed consistenty check, for example, java.lang.AssertionError.

  • The error can happen at any point, resulting in inconsistencies due to half-updated objects. Examples are java.lang.ThreadDeath, java.lang.OutOfMemoryError and java.lang.StackOverflowError.

  • The error indicates that virtual machine failed to provide some semantic guarantees by the Java programming language. java.lang.ExceptionInInitializerError is an example—it can leave behind a half-initialized class.

In general, if an error is thrown, the virtual machine should be restarted as soon as possible because it is in an inconsistent state. Continuing running as before can have unexpected consequences. However, there are legitimate reasons for catching errors because not doing so leads to even greater problems.

Code should be written in a way that avoids triggering errors. See Increasing Robustness when Reading Arrays for an example.

It is usually necessary to log errors. Otherwise, no trace of the problem might be left anywhere, making it very difficult to diagnose realted failures. Consequently, if you catch java.lang.Exception to log and suppress all unexpected exceptions (for example, in a request dispatching loop), you should consider switching to java.lang.Throwable instead, to also cover errors.

The other reason mainly applies to such request dispatching loops: If you do not catch errors, the loop stops looping, resulting in a denial of service.

However, if possible, catching errors should be coupled with a way to signal the requirement of a virtual machine restart.

1.3.2. Low-level Features of the Virtual Machine

Reflection and Private Parts

The setAccessible(boolean) method of the java.lang.reflect.AccessibleObject class allows a program to disable language-defined access rules for specific constructors, methods, or fields. Once the access checks are disabled, any code can use the java.lang.reflect.Constructor, java.lang.reflect.Method, or java.lang.reflect.Field object to access the underlying Java entity, without further permission checks. This breaks encapsulation and can undermine the stability of the virtual machine. (In contrast, without using the setAccessible(boolean) method, this should not happen because all the language-defined checks still apply.)

This feature should be avoided if possible.

Java Native Interface (JNI)

The Java Native Interface allows calling from Java code functions specifically written for this purpose, usually in C or C++.

The transition between the Java world and the C world is not fully type-checked, and the C code can easily break the Java virtual machine semantics. Therefore, extra care is needed when using this functionality.

To provide a moderate amount of type safety, it is recommended to recreate the class-specific header file using javah during the build process, include it in the implementation, and use the -Wmissing-declarations option.

Ideally, the required data is directly passed to static JNI methods and returned from them, and the code and the C side does not have to deal with accessing Java fields (or even methods).

When using GetPrimitiveArrayCritical or GetStringCritical, make sure that you only perform very little processing between the get and release operations. Do not access the file system or the network, and not perform locking, because that might introduce blocking. When processing large strings or arrays, consider splitting the computation into multiple sub-chunks, so that you do not prevent the JVM from reaching a safepoint for extended periods of time.

If necessary, you can use the Java long type to store a C pointer in a field of a Java class. On the C side, when casting between the jlong value and the pointer on the C side,

You should not try to perform pointer arithmetic on the Java side (that is, you should treat pointer-carrying long values as opaque). When passing a slice of an array to the native code, follow the Java convention and pass it as the base array, the integer offset of the start of the slice, and the integer length of the slice. On the native side, check the offset/length combination against the actual array length, and use the offset to compute the pointer to the beginning of the array.

Example 11. Array length checking in JNI code

  (JNIEnv *jEnv, jclass clazz, jbyteArray buffer, jint offset, jint length)
  assert(sizeof(jint) == sizeof(unsigned));
  if (offset < 0 || length < 0) {
    (*jEnv)->ThrowNew(jEnv, arrayIndexOutOfBoundsExceptionClass,
		      "negative offset/length");
    return 0;
  unsigned uoffset = offset;
  unsigned ulength = length;
  // This cannot overflow because of the check above.
  unsigned totallength = uoffset + ulength;
  unsigned actuallength = (*jEnv)->GetArrayLength(jEnv, buffer);
  if (totallength > actuallength) {
    (*jEnv)->ThrowNew(jEnv, arrayIndexOutOfBoundsExceptionClass,
		      "offset + length too large");
    return 0;
  unsigned char *ptr = (*jEnv)->GetPrimitiveArrayCritical(jEnv, buffer, 0);
  if (ptr == NULL) {
    return 0;
  unsigned long long sum = 0;
  for (unsigned char *p = ptr + uoffset, *end = p + ulength; p != end; ++p) {
    sum += *p;
  (*jEnv)->ReleasePrimitiveArrayCritical(jEnv, buffer, ptr, 0);
  return sum;

In any case, classes referring to native resources must be declared final, and must not be serializeable or cloneable. Initialization and mutation of the state used by the native side must be controlled carefully. Otherwise, it might be possible to create an object with inconsistent native state which results in a crash (or worse) when used (or perhaps only finalized) later. If you need both Java inheritance and native resources, you should consider moving the native state to a separate class, and only keep a reference to objects of that class. This way, cloning and serialization issues can be avoided in most cases.

If there are native resources associated with an object, the class should have an explicit resource deallocation method (Resource Management) and a finalizer (Finalizers) as a last resort. The need for finalization means that a minimum amount of synchronization is needed. Code on the native side should check that the object is not in a closed/freed state.

Many JNI functions create local references. By default, these persist until the JNI-implemented method returns. If you create many such references (e.g., in a loop), you may have to free them using DeleteLocalRef, or start using PushLocalFrame and PopLocalFrame. Global references must be deallocated with DeleteGlobalRef, otherwise there will be a memory leak, just as with malloc and free.

When throwing exceptions using Throw or ThrowNew, be aware that these functions return regularly. You have to return control manually to the JVM.

Technically, the JNIEnv pointer is not necessarily constant during the lifetime of your JNI module. Storing it in a global variable is therefore incorrect. Particularly if you are dealing with callbacks, you may have to store the pointer in a thread-local variable (defined with __thread). It is, however, best to avoid the complexity of calling back into Java code.

Keep in mind that C/C and Java are different languages, despite very similar syntax for expressions. The Java memory model is much more strict than the C or C memory models, and native code needs more synchronization, usually using JVM facilities or POSIX threads mutexes. Integer overflow in Java is defined, but in C/C++ it is not (for the jint and jlong types).


The sun.misc.Unsafe class is unportable and contains many functions explicitly designed to break Java memory safety (for performance and debugging). If possible, avoid using this class.

1.3.3. Interacting with the Security Manager

The Java platform is largely implemented in the Java language itself. Therefore, within the same JVM, code runs which is part of the Java installation and which is trusted, but there might also be code which comes from untrusted sources and is restricted by the Java sandbox (to varying degrees). The security manager draws a line between fully trusted, partially trusted and untrusted code.

The type safety and accessibility checks provided by the Java language and JVM would be sufficient to implement a sandbox. However, only some Java APIs employ such a capabilities-based approach. (The Java SE library contains many public classes with public constructors which can break any security policy, such as java.io.FileOutputStream.) Instead, critical functionality is protected by stack inspection: At a security check, the stack is walked from top (most-nested) to bottom. The security check fails if a stack frame for a method is encountered whose class lacks the permission which the security check requires.

This simple approach would not allow untrusted code (which lacks certain permissions) to call into trusted code while the latter retains trust. Such trust transitions are desirable because they enable Java as an implementation language for most parts of the Java platform, including security-relevant code. Therefore, there is a mechanism to mark certain stack frames as trusted (Re-gaining Privileges).

In theory, it is possible to run a Java virtual machine with a security manager that acts very differently from this approach, but a lot of code expects behavior very close to the platform default (including many classes which are part of the OpenJDK implementation).

Security Manager Compatibility

A lot of code can run without any additional permissions at all, with little changes. The following guidelines should help to increase compatibility with a restrictive security manager.

  • When retrieving system properties using System.getProperty(String) or similar methods, catch SecurityException exceptions and treat the property as unset.

  • Avoid unnecessary file system or network access.

  • Avoid explicit class loading. Access to a suitable class loader might not be available when executing as untrusted code.

If the functionality you are implementing absolutely requires privileged access and this functionality has to be used from untrusted code (hopefully in a restricted and secure manner), see Re-gaining Privileges.

Activating the Security Manager

The usual command to launch a Java application, java, does not activate the security manager. Therefore, the virtual machine does not enforce any sandboxing restrictions, even if explicitly requested by the code (for example, as described in Reducing Trust in Code).

The -Djava.security.manager option activates the security manager, with the fairly restrictive default policy. With a very permissive policy, most Java code will run unchanged. Assuming the policy in Most permissve OpenJDK policy file has been saved in a file grant-all.policy, this policy can be activated using the option -Djava.security.policy=grant-all.policy (in addition to the -Djava.security.manager option).

Example 12. Most permissve OpenJDK policy file

grant {
      permission java.security.AllPermission;

With this most permissive policy, the security manager is still active, and explicit requests to drop privileges will be honored.

Reducing Trust in Code

The Using the security manager to run code with reduced privileges example shows how to run a piece code of with reduced privileges.

Example 13. Using the security manager to run code with reduced privileges

Permissions permissions = new Permissions();
        ProtectionDomain protectionDomain =
    new ProtectionDomain(null, permissions);
        AccessControlContext context = new AccessControlContext(
            new ProtectionDomain[] { protectionDomain });

// This is expected to succeed.
try (FileInputStream in = new FileInputStream(path)) {
    System.out.format("FileInputStream: %s%n", in);

AccessController.doPrivileged(new PrivilegedExceptionAction<Void>() {
        public Void run() throws Exception {
	    // This code runs with reduced privileges and is
	    // expected to fail.
	    try (FileInputStream in = new FileInputStream(path)) {
		System.out.format("FileInputStream: %s%n", in);
	    return null;
    }, context);

The example above does not add any additional permissions to the permissions object. If such permissions are necessary, code like the following (which grants read permission on all files in the current directory) can be used:

permissions.add(new FilePermission(
            System.getProperty("user.dir") + "/-", "read"));

Calls to the java.security.AccessController.doPrivileged() methods do not enforce any additional restriction if no security manager has been set. Except for a few special exceptions, the restrictions no longer apply if the doPrivileged() has returned, even to objects created by the code which ran with reduced privileges. (This applies to object finalization in particular.)

The example code above does not prevent the called code from calling the java.security.AccessController.doPrivileged() methods. This mechanism should be considered an additional safety net, but it still can be used to prevent unexpected behavior of trusted code. As long as the executed code is not dynamic and came with the original application or library, the sandbox is fairly effective.

The context argument in Using the security manager to run code with reduced privileges is extremely important—otherwise, this code would increase privileges instead of reducing them.

For activating the security manager, see Activating the Security Manager. Unfortunately, this affects the virtual machine as a whole, so it is not possible to do this from a library.

Re-gaining Privileges

Ordinarily, when trusted code is called from untrusted code, it loses its privileges (because of the untrusted stack frames visible to stack inspection). The java.security.AccessController.doPrivileged() family of methods provides a controlled backdoor from untrusted to trusted code.


By design, this feature can undermine the Java security model and the sandbox. It has to be used very carefully. Most sandbox vulnerabilities can be traced back to its misuse.

In essence, the doPrivileged() methods cause the stack inspection to end at their call site. Untrusted code further down the call stack becomes invisible to security checks.

The following operations are common and safe to perform with elevated privileges.

  • Reading custom system properties with fixed names, especially if the value is not propagated to untrusted code. (File system paths including installation paths, host names and user names are sometimes considered private information and need to be protected.)

  • Reading from the file system at fixed paths, either determined at compile time or by a system property. Again, leaking the file contents to the caller can be problematic.

  • Accessing network resources under a fixed address, name or URL, derived from a system property or configuration file, information leaks not withstanding.

The Using the security manager to run code with increased privileges example shows how to request additional privileges.

Example 14. Using the security manager to run code with increased privileges

// This is expected to fail.
try {
} catch (SecurityException e) {
AccessController.doPrivileged(new PrivilegedAction<Void>() {
        public Void run() {
            // This should work.
            return null;

Obviously, this only works if the class containing the call to doPrivileged() is marked trusted (usually because it is loaded from a trusted class loader).

When writing code that runs with elevated privileges, make sure that you follow the rules below.

  • Make the privileged code as small as possible. Perform as many computations as possible before and after the privileged code section, even if it means that you have to define a new class to pass the data around.

  • Make sure that you either control the inputs to the privileged code, or that the inputs are harmless and cannot affect security properties of the privileged code.

  • Data that is returned from or written by the privileged code must either be restricted (that is, it cannot be accessed by untrusted code), or must be harmless. Otherwise, privacy leaks or information disclosures which affect security properties can be the result.

If the code calls back into untrusted code at a later stage (or performs other actions under control from the untrusted caller), you must obtain the original security context and restore it before performing the callback, as in Restoring privileges when invoking callbacks. (In this example, it would be much better to move the callback invocation out of the privileged code section, of course.)

Example 15. Restoring privileges when invoking callbacks

interface Callback<T> {
	T call(boolean flag);

class CallbackInvoker<T> {
	private final AccessControlContext context;
	Callback<T> callback;

	CallbackInvoker(Callback<T> callback) {
	    context = AccessController.getContext();
	    this.callback = callback;

	public T invoke() {
	    // Obtain increased privileges.
	    return AccessController.doPrivileged(new PrivilegedAction<T>() {
		    public T run() {
			// This operation would fail without
			// additional privileges.
			final boolean flag = Boolean.getBoolean("some.property");

			// Restore the original privileges.
			return AccessController.doPrivileged(
                        new PrivilegedAction<T>() {
				public T run() {
				    return callback.call(flag);
			    }, context);

1.4. The Python Programming Language

Python provides memory safety by default, so low-level security vulnerabilities are rare and typically needs fixing the Python interpreter or standard library itself.

Other sections with Python-specific advice include:

1.4.1. Dangerous Standard Library Features

Some areas of the standard library, notably the ctypes module, do not provide memory safety guarantees comparable to the rest of Python. If such functionality is used, the advice in The Core Language should be followed.

1.4.2. Run-time Compilation and Code Generation

The following Python functions and statements related to code execution should be avoided:

  • compile

  • eval

  • exec

  • execfile

If you need to parse integers or floating point values, use the int and float functions instead of eval. Sandboxing untrusted Python code does not work reliably.

1.4.3. Sandboxing

The rexec Python module cannot safely sandbox untrusted code and should not be used. The standard CPython implementation is not suitable for sandboxing.

1.5. Shell Programming and bash

This chapter contains advice about shell programming, specifically in bash. Most of the advice will apply to scripts written for other shells because extensions such as integer or array variables have been implemented there as well, with comparable syntax.

1.5.1. Consider Alternatives

Once a shell script is so complex that advice in this chapter applies, it is time to step back and consider the question: Is there a more suitable implementation language available?

For example, Python with its subprocess module can be used to write scripts which are almost as concise as shell scripts when it comes to invoking external programs, and Python offers richer data structures, with less arcane syntax and more consistent behavior.

1.5.2. Shell Language Features

The following sections cover subtleties concerning the shell programming languages. They have been written with the bash shell in mind, but some of these features apply to other shells as well.

Some of the features described may seem like implementation defects, but these features have been replicated across multiple independent implementations, so they now have to be considered part of the shell programming language.

Parameter Expansion

The mechanism by which named shell variables and parameters are expanded is called parameter expansion. The most basic syntax is “$variable” or “${variable}”.

In almost all cases, a parameter expansion should be enclosed in double quotation marks “…”.

external-program "$arg1" "$arg2"

If the double quotation marks are omitted, the value of the variable will be split according to the current value of the IFS variable. This may allow the injection of additional options which are then processed by external-program.

Parameter expansion can use special syntax for specific features, such as substituting defaults or performing string or array operations. These constructs should not be used because they can trigger arithmetic evaluation, which can result in code execution. See Arithmetic Evaluation.

Double Expansion

Double expansion occurs when, during the expansion of a shell variable, not just the variable is expanded, replacing it by its value, but the value of the variable is itself is expanded as well. This can trigger arbitrary code execution, unless the value of the variable is verified against a restrictive pattern.

The evaluation process is in fact recursive, so a self-referential expression can cause an out-of-memory condition and a shell crash.

Double expansion may seem like as a defect, but it is implemented by many shells, and has to be considered an integral part of the shell programming language. However, it does make writing robust shell scripts difficult.

Double expansion can be requested explicitly with the eval built-in command, or by invoking a subshell with “bash -c”. These constructs should not be used.

The following sections give examples of places where implicit double expansion occurs.

Arithmetic Evaluation

Arithmetic evaluation is a process by which the shell computes the integer value of an expression specified as a string. It is highly problematic for two reasons: It triggers double expansion (see Double Expansion), and the language of arithmetic expressions is not self-contained. Some constructs in arithmetic expressions (notably array subscripts) provide a trapdoor from the restricted language of arithmetic expressions to the full shell language, thus paving the way towards arbitrary code execution. Due to double expansion, input which is (indirectly) referenced from an arithmetic expression can trigger execution of arbitrary code, which is potentially harmful.

Arithmetic evaluation is triggered by the follow constructs:

  • The expression in “$expression” is evaluated. This construct is called arithmetic expansion.

  • $[expression]” is a deprecated syntax with the same effect.

  • The arguments to the let shell built-in are evaluated.

  • expression” is an alternative syntax for “let expression”.

  • Conditional expressions surrounded by “[[]]” can trigger arithmetic evaluation if certain operators such as -eq are used. (The test built-in does not perform arithmetic evaluation, even with integer operators such as -eq.)

    The conditional expression “[[ $variable =~ regexp ]]” can be used for input validation, assuming that regexp is a constant regular expression. See Performing Input Validation.

  • Certain parameter expansions, for example “${variable[expression]}” (array indexing) or “${variable:expression}” (string slicing), trigger arithmetic evaluation of expression.

  • Assignment to array elements using “array_variable[subscript]=expression” triggers evaluation of subscript, but not expression.

  • The expressions in the arithmetic for command, “for expression1; expression2; expression3; do commands; done” are evaluated. This does not apply to the regular for command, “for variable in list; do commands; done”.


Depending on the bash version, the above list may be incomplete.

If faced with a situation where using such shell features appears necessary, see Consider Alternatives.

If it is impossible to avoid shell arithmetic on untrusted inputs, refer to Performing Input Validation.

Type declarations

bash supports explicit type declarations for shell variables:

	declare -i integer_variable
	declare -a array_variable
	declare -A assoc_array_variable

	typeset -i integer_variable
	typeset -a array_variable
	typeset -A assoc_array_variable

	local -i integer_variable
	local -a array_variable
	local -A assoc_array_variable

	readonly -i integer_variable
	readonly -a array_variable
	readonly -A assoc_array_variable

Variables can also be declared as arrays by assigning them an array expression, as in:

array_variable=(1 2 3 4)

Some built-ins (such as mapfile) can implicitly create array variables.

Such type declarations should not be used because assignment to such variables (independent of the concrete syntax used for the assignment) triggers arithmetic expansion (and thus double expansion) of the right-hand side of the assignment operation. See Arithmetic Evaluation.

Shell scripts which use integer or array variables should be rewritten in another, more suitable language. Se Consider Alternatives.

Other Obscurities

Obscure shell language features should not be used. Examples are:

  • Exported functions (export -f or declare -f).

  • Function names which are not valid variable names, such as “module::function”.

  • The possibility to override built-ins or external commands with shell functions.

  • Changing the value of the IFS variable to tokenize strings.

1.5.3. Invoking External Commands

When passing shell variables as single command line arguments, they should always be surrounded by double quotes. See Parameter Expansion.

Care is required when passing untrusted values as positional parameters to external commands. If the value starts with a hyphen “-”, it may be interpreted by the external command as an option. Depending on the external program, a “--” argument stops option processing and treats all following arguments as positional parameters. (Double quotes are completely invisible to the command being invoked, so they do not prevent variable values from being interpreted as options.)

Cleaning the environment before invoking child processes is difficult to implement in script. bash keeps a hidden list of environment variables which do not correspond to shell variables, and unsetting them from within a bash script is not possible. To reset the environment, a script can re-run itself under the “env -i” command with an additional parameter which indicates the environment has been cleared and suppresses a further self-execution. Alternatively, individual commands can be executed with “env -i”.


Complete isolation from its original execution environment (which is required when the script is executed after a trust transition, e.g., triggered by the SUID mechanism) is impossible to achieve from within the shell script itself. Instead, the invoking process has to clear the process environment (except for few trusted variables) before running the shell script.

Checking for failures in executed external commands is recommended. If no elaborate error recovery is needed, invoking “set -e” may be sufficient. This causes the script to stop on the first failed command. However, failures in pipes (“command1 | command2”) are only detected for the last command in the pipe, errors in previous commands are ignored. This can be changed by invoking “set -o pipefail”. Due to architectural limitations, only the process that spawned the entire pipe can check for failures in individual commands; it is not possible for a process to tell if the process feeding data (or the process consuming data) exited normally or with an error.

See Creating Safe Processes for additional details on creating child processes.

1.5.4. Temporary Files

Temporary files should be created with the mktemp command, and temporary directories with “mktemp -d”.

To clean up temporary files and directories, write a clean-up shell function and register it as a trap handler, as shown in Creating and Cleaning up Temporary Files. Using a separate function avoids issues with proper quoting of variables.

Example 16. Creating and Cleaning up Temporary Files


cleanup () {
    rm -f -- "$tmpfile"

trap cleanup 0

1.5.5. Performing Input Validation

In some cases, input validation cannot be avoided. For example, if arithmetic evaluation is absolutely required, it is imperative to check that input values are, in fact, integers. See Arithmetic Evaluation.

Input validation in bash shows a construct which can be used to check if a string “$value” is an integer. This construct is specific to bash and not portable to POSIX shells.

Example 17. Input validation in bash

if [[ $value =~ ^-?[0-9]+$ ]] ; then
	echo value is an integer
	echo "value is not an integer" 1>&2
	exit 1

Using case statements for input validation is also possible and supported by other (POSIX) shells, but the pattern language is more restrictive, and it can be difficult to write suitable patterns.

The expr external command can give misleading results (e.g., if the value being checked contains operators itself) and should not be used.

1.5.6. Guarding Shell Scripts Against Changes

bash only reads a shell script up to the point it is needed for executed the next command. This means that if script is overwritten while it is running, execution can jump to a random part of the script, depending on what is modified in the script and how the file offsets change as a result. (This behavior is needed to support self-extracting shell archives whose script part is followed by a stream of bytes which does not follow the shell language syntax.)

Therefore, long-running scripts should be guarded against concurrent modification by putting as much of the program logic into a main function, and invoking the main function at the end of the script, using this syntax:

main "$@" ; exit $?

This construct ensures that bash will stop execution after the main function, instead of opening the script file and trying to read more commands.

1.6. The Go Programming Language

This chapter contains language-specific recommendations for Go.

1.6.1. Memory Safety

Go provides memory safety, but only if the program is not executed in parallel (that is, GOMAXPROCS is not larger than 1). The reason is that interface values and slices consist of multiple words are not updated atomically. Another thread of execution can observe an inconsistent pairing between type information and stored value (for interfaces) or pointer and length (for slices), and such inconsistency can lead to a memory safety violation.

Code which does not run in parallel and does not use the unsafe package (or other packages which expose unsafe constructs) is memory-safe. For example, invalid casts and out-of-range subscripting cause panics at run time.

Keep in mind that finalization can introduce parallelism because finalizers are executed concurrently, potentially interleaved with the rest of the program.

1.6.2. Error Handling

Only a few common operations (such as pointer dereference, integer division, array subscripting) trigger exceptions in Go, called panics. Most interfaces in the standard library use a separate return value of type error to signal error.

Not checking error return values can lead to incorrect operation and data loss (especially in the case of writes, using interfaces such as io.Writer).

The correct way to check error return values depends on the function or method being called. In the majority of cases, the first step after calling a function should be an error check against the nil value, handling any encountered error. See Regular error handling in Go for details.

Example 18. Regular error handling in Go

type Processor interface {
	Process(buf []byte) (message string, err error)

type ErrorHandler interface {
	Handle(err error)

func RegularError(buf []byte, processor Processor,
	handler ErrorHandler) (message string, err error) {
	message, err = processor.Process(buf)
	if err != nil {
		return "", err

However, with io.Reader, io.ReaderAt and related interfaces, it is necessary to check for a non-zero number of read bytes first, as shown in Read error handling in Go. If this pattern is not followed, data loss may occur. This is due to the fact that the io.Reader interface permits returning both data and an error at the same time.

Example 19. Read error handling in Go

func IOError(r io.Reader, buf []byte, processor Processor,
	handler ErrorHandler) (message string, err error) {
	n, err := r.Read(buf)
	// First check for available data.
	if n > 0 {
		message, err = processor.Process(buf[0:n])
		// Regular error handling.
		if err != nil {
			return "", err
	// Then handle any error.
	if err != nil {
		return "", err

1.6.3. Garbage Collector

Older Go releases (before Go 1.3) use a conservative garbage collector without blacklisting. This means that data blobs can cause retention of unrelated data structures because the data is conservatively interpreted as pointers. This phenomenon can be triggered accidentally on 32-bit architectures and is more likely to occur if the heap grows larger. On 64-bit architectures, it may be possible to trigger it deliberately—it is unlikely to occur spontaneously.

1.6.4. Marshaling and Unmarshaling

Several packages in the encoding hierarchy provide support for serialization and deserialization. The usual caveats apply (see Serialization and Deserialization).

As an additional precaution, the Unmarshal and Decode functions should only be used with fresh values in the interface{} argument. This is due to the way defaults for missing values are implemented: During deserialization, missing value do not result in an error, but the original value is preserved. Using a fresh value (with suitable default values if necessary) ensures that data from a previous deserialization operation does not leak into the current one. This is especially relevant when structs are deserialized.

1.7. The Vala Programming Language

Vala is a programming language mainly targeted at GNOME developers.

Its syntax is inspired by C# (and thus, indirectly, by Java). But unlike C# and Java, Vala does not attempt to provide memory safety: Vala is compiled to C, and the C code is compiled with GCC using typical compiler flags. Basic operations like integer arithmetic are directly mapped to C constructs. As a results, the recommendations in The C Programming Language apply.

In particular, the following Vala language constructs can result in undefined behavior at run time:

  • Integer arithmetic, as described in Recommendations for Integer Arithmetic.

  • Pointer arithmetic, string subscripting and the substring method on strings (the string class in the glib-2.0 package) are not range-checked. It is the responsibility of the calling code to ensure that the arguments being passed are valid. This applies even to cases (like substring) where the implementation would have range information to check the validity of indexes. See Recommendations for Pointers and Array Handling.

  • Similarly, Vala only performs garbage collection (through reference counting) for GObject values. For plain C pointers (such as strings), the programmer has to ensure that storage is deallocated once it is no longer needed (to avoid memory leaks), and that storage is not being deallocated while it is still being used (see Use-after-free errors).

2. Specific Programming Tasks

2.1. Library Design

Through this section, the term client code refers to applications and other libraries using the library.

2.1.1. State Management

Global State

Global state should be avoided.

If this is impossible, the global state must be protected with a lock. For C/C++, you can use the pthread_mutex_lock and pthread_mutex_unlock functions without linking against -lpthread because the system provides stubs for non-threaded processes.

For compatibility with fork, these locks should be acquired and released in helpers registered with pthread_atfork. This function is not available without -lpthread, so you need to use dlsym or a weak symbol to obtain its address.

If you need fork protection for other reasons, you should store the process ID and compare it to the value returned by getpid each time you access the global state. (getpid is not implemented as a system call and is fast.) If the value changes, you know that you have to re-create the state object. (This needs to be combined with locking, of course.)


Library state should be kept behind a curtain. Client code should receive only a handle. In C, the handle can be a pointer to an incomplete struct. In C++, the handle can be a pointer to an abstract base class, or it can be hidden using the pointer-to-implementation idiom.

The library should provide functions for creating and destroying handles. (In C++, it is possible to use virtual destructors for the latter.) Consistency between creation and destruction of handles is strongly recommended: If the client code created a handle, it is the responsibility of the client code to destroy it. (This is not always possible or convenient, so sometimes, a transfer of ownership has to happen.)

Using handles ensures that it is possible to change the way the library represents state in a way that is transparent to client code. This is important to facilitate security updates and many other code changes.

It is not always necessary to protect state behind a handle with a lock. This depends on the level of thread safety the library provides.

2.1.2. Object Orientation

Classes should be either designed as base classes, or it should be impossible to use them as base classes (like final classes in Java). Classes which are not designed for inheritance and are used as base classes nevertheless create potential maintenance hazards because it is difficult to predict how client code will react when calls to virtual methods are added, reordered or removed.

Virtual member functions can be used as callbacks. See Callbacks for some of the challenges involved.

2.1.3. Callbacks

Higher-order code is difficult to analyze for humans and computers alike, so it should be avoided. Often, an iterator-based interface (a library function which is called repeatedly by client code and returns a stream of events) leads to a better design which is easier to document and use.

If callbacks are unavoidable, some guidelines for them follow.

In modern C++ code, std::function objects should be used for callbacks.

In older C++ code and in C code, all callbacks must have an additional closure parameter of type void *, the value of which can be specified by client code. If possible, the value of the closure parameter should be provided by client code at the same time a specific callback is registered (or specified as a function argument). If a single closure parameter is shared by multiple callbacks, flexibility is greatly reduced, and conflicts between different pieces of client code using the same library object could be unresolvable. In some cases, it makes sense to provide a de-registration callback which can be used to destroy the closure parameter when the callback is no longer used.

Callbacks can throw exceptions or call longjmp. If possible, all library objects should remain in a valid state. (All further operations on them can fail, but it should be possible to deallocate them without causing resource leaks.)

The presence of callbacks raises the question if functions provided by the library are reentrant. Unless a library was designed for such use, bad things will happen if a callback function uses functions in the same library (particularly if they are invoked on the same objects and manipulate the same state). When the callback is invoked, the library can be in an inconsistent state. Reentrant functions are more difficult to write than thread-safe functions (by definition, simple locking would immediately lead to deadlocks). It is also difficult to decide what to do when destruction of an object which is currently processing a callback is requested.

2.1.4. Process Attributes

Several attributes are global and affect all code in the process, not just the library that manipulates them.

  • environment variables (see Accessing Environment Variables)

  • umask

  • user IDs, group IDs and capabilities

  • current working directory

  • signal handlers, signal masks and signal delivery

  • file locks (especially fcntl locks behave in surprising ways, not just in a multi-threaded environment)

Library code should avoid manipulating these global process attributes. It should not rely on environment variables, umask, the current working directory and signal masks because these attributes can be inherited from an untrusted source.

In addition, there are obvious process-wide aspects such as the virtual memory layout, the set of open files and dynamic shared objects, but with the exception of shared objects, these can be manipulated in a relatively isolated way.

2.2. File Descriptor Management

File descriptors underlie all input/output mechanisms offered by the system. They are used to implementation the FILE *-based functions found in <stdio.h>, and all the file and network communication facilities provided by the Python and Java environments are eventually implemented in them.

File descriptors are small, non-negative integers in userspace, and are backed on the kernel side with complicated data structures which can sometimes grow very large.

2.2.1. Closing Descriptors

If a descriptor is no longer used by a program and is not closed explicitly, its number cannot be reused (which is problematic in itself, see Dealing with the select Limit), and the kernel resources are not freed. Therefore, it is important to close all descriptors at the earliest point in time possible, but not earlier.

Error Handling during Descriptor Close

The close system call is always successful in the sense that the passed file descriptor is never valid after the function has been called. However, close still can return an error, for example if there was a file system failure. But this error is not very useful because the absence of an error does not mean that all caches have been emptied and previous writes have been made durable. Programs which need such guarantees must open files with O_SYNC or use fsync or fdatasync, and may also have to fsync the directory containing the file.

Closing Descriptors and Race Conditions

Unlike process IDs, which are recycle only gradually, the kernel always allocates the lowest unused file descriptor when a new descriptor is created. This means that in a multi-threaded program which constantly opens and closes file descriptors, descriptors are reused very quickly. Unless descriptor closing and other operations on the same file descriptor are synchronized (typically, using a mutex), there will be race conditons and I/O operations will be applied to the wrong file descriptor.

Sometimes, it is necessary to close a file descriptor concurrently, while another thread might be about to use it in a system call. In order to support this, a program needs to create a single special file descriptor, one on which all I/O operations fail. One way to achieve this is to use socketpair, close one of the descriptors, and call shutdown(fd, SHUTRDWR) on the other.

When a descriptor is closed concurrently, the program does not call close on the descriptor. Instead it program uses dup2 to replace the descriptor to be closed with the dummy descriptor created earlier. This way, the kernel will not reuse the descriptor, but it will carry out all other steps associated with calling a descriptor (for instance, if the descriptor refers to a stream socket, the peer will be notified).

This is just a sketch, and many details are missing. Additional data structures are needed to determine when it is safe to really close the descriptor, and proper locking is required for that.

Lingering State after Close

By default, closing a stream socket returns immediately, and the kernel will try to send the data in the background. This means that it is impossible to implement accurate accounting of network-related resource utilization from userspace.

The SO_LINGER socket option alters the behavior of close, so that it will return only after the lingering data has been processed, either by sending it to the peer successfully, or by discarding it after the configured timeout. However, there is no interface which could perform this operation in the background, so a separate userspace thread is needed for each close call, causing scalability issues.

Currently, there is no application-level countermeasure which applies universally. Mitigation is possible with iptables (the connlimit match type in particular) and specialized filtering devices for denial-of-service network traffic.

These problems are not related to the TIME_WAIT state commonly seen in netstat output. The kernel automatically expires such sockets if necessary.

2.2.2. Preventing File Descriptor Leaks to Child Processes

Child processes created with fork share the initial set of file descriptors with their parent process. By default, file descriptors are also preserved if a new process image is created with execve (or any of the other functions such as system or posix_spawn).

Usually, this behavior is not desirable. There are two ways to turn it off, that is, to prevent new process images from inheriting the file descriptors in the parent process:

  • Set the close-on-exec flag on all newly created file descriptors. Traditionally, this flag is controlled by the FD_CLOEXEC flag, using F_GETFD and F_SETFD operations of the fcntl function.

    However, in a multi-threaded process, there is a race condition: a subprocess could have been created between the time the descriptor was created and the FD_CLOEXEC was set. Therefore, many system calls which create descriptors (such as open and openat) now accept the O_CLOEXEC flag (SOCK_CLOEXEC for socket and socketpair), which cause the FD_CLOEXEC flag to be set for the file descriptor in an atomic fashion. In addition, a few new systems calls were introduced, such as pipe2 and dup3.

    The downside of this approach is that every descriptor needs to receive special treatment at the time of creation, otherwise it is not completely effective.

  • After calling fork, but before creating a new process image with execve, all file descriptors which the child process will not need are closed.

    Traditionally, this was implemented as a loop over file descriptors ranging from 3 to 255 and later 1023. But this is only an approximation because it is possible to create file descriptors outside this range easily (see Dealing with the select Limit). Another approach reads /proc/self/fd and closes the unexpected descriptors listed there, but this approach is much slower.

At present, environments which care about file descriptor leakage implement the second approach. OpenJDK 6 and 7 are among them.

2.2.3. Dealing with the select Limit

By default, a user is allowed to open only 1024 files in a single process, but the system administrator can easily change this limit (which is necessary for busy network servers). However, there is another restriction which is more difficult to overcome.

The select function only supports a maximum of FD_SETSIZE file descriptors (that is, the maximum permitted value for a file descriptor is FD_SETSIZE - 1, usually 1023.) If a process opens many files, descriptors may exceed such limits. It is impossible to query such descriptors using select.

If a library which creates many file descriptors is used in the same process as a library which uses select, at least one of them needs to be changed. Calls to select can be replaced with calls to poll or another event handling mechanism. Replacing the select function is the recommended approach.

Alternatively, the library with high descriptor usage can relocate descriptors above the FD_SETSIZE limit using the following procedure.

  • Create the file descriptor fd as usual, preferably with the O_CLOEXEC flag.

  • Before doing anything else with the descriptor fd, invoke:

	  int newfd = fcntl(fd, F_DUPFD_CLOEXEC, (long)FD_SETSIZE);
  • Check that newfd result is non-negative, otherwise close fd and report an error, and return.

  • Close fd and continue to use newfd.

The new descriptor has been allocated above the FD_SETSIZE. Even though this algorithm is racy in the sense that the FD_SETSIZE first descriptors could fill up, a very high degree of physical parallelism is required before this becomes a problem.

2.3. File System Manipulation

In this chapter, we discuss general file system manipulation, with a focus on access files and directories to which an other, potentially untrusted user has write access.

Temporary files are covered in their own chapter, Temporary Files.

2.3.1. Working with Files and Directories Owned by Other Users

Sometimes, it is necessary to operate on files and directories owned by other (potentially untrusted) users. For example, a system administrator could remove the home directory of a user, or a package manager could update a file in a directory which is owned by an application-specific user. This differs from accessing the file system as a specific user; see Accessing the File System as a Different User.

Accessing files across trust boundaries faces several challenges, particularly if an entire directory tree is being traversed:

  1. Another user might add file names to a writable directory at any time. This can interfere with file creation and the order of names returned by readdir.

  2. Merely opening and closing a file can have side effects. For instance, an automounter can be triggered, or a tape device rewound. Opening a file on a local file system can block indefinitely, due to mandatory file locking, unless the O_NONBLOCK flag is specified.

  3. Hard links and symbolic links can redirect the effect of file system operations in unexpected ways. The O_NOFOLLOW and AT_SYMLINK_NOFOLLOW variants of system calls only affected final path name component.

  4. The structure of a directory tree can change. For example, the parent directory of what used to be a subdirectory within the directory tree being processed could suddenly point outside that directory tree.

Files should always be created with the O_CREAT and O_EXCL flags, so that creating the file will fail if it already exists. This guards against the unexpected appearance of file names, either due to creation of a new file, or hard-linking of an existing file. In multi-threaded programs, rather than manipulating the umask, create the files with mode 000 if possible, and adjust it afterwards with fchmod.

To avoid issues related to symbolic links and directory tree restructuring, the “at” variants of system calls have to be used (that is, functions like openat, fchownat, fchmodat, and unlinkat, together with O_NOFOLLOW or AT_SYMLINK_NOFOLLOW). Path names passed to these functions must have just a single component (that is, without a slash). When descending, the descriptors of parent directories must be kept open. The missing opendirat function can be emulated with openat (with an O_DIRECTORY flag, to avoid opening special files with side effects), followed by fdopendir.

If the “at” functions are not available, it is possible to emulate them by changing the current directory. (Obviously, this only works if the process is not multi-threaded.) fchdir has to be used to change the current directory, and the descriptors of the parent directories have to be kept open, just as with the “at”-based approach. chdir("…​") is unsafe because it might ascend outside the intended directory tree.

This “at” function emulation is currently required when manipulating extended attributes. In this case, the lsetxattr function can be used, with a relative path name consisting of a single component. This also applies to SELinux contexts and the lsetfilecon function.

Currently, it is not possible to avoid opening special files and changes to files with hard links if the directory containing them is owned by an untrusted user. (Device nodes can be hard-linked, just as regular files.) fchmodat and fchownat affect files whose link count is greater than one. But opening the files, checking that the link count is one with fstat, and using fchmod and fchown on the file descriptor may have unwanted side effects, due to item 2 above. When creating directories, it is therefore important to change the ownership and permissions only after it has been fully created. Until that point, file names are stable, and no files with unexpected hard links can be introduced.

Similarly, when just reading a directory owned by an untrusted user, it is currently impossible to reliably avoid opening special files.

There is no workaround against the instability of the file list returned by readdir. Concurrent modification of the directory can result in a list of files being returned which never actually existed on disk.

Hard links and symbolic links can be safely deleted using unlinkat without further checks because deletion only affects the name within the directory tree being processed.

2.3.2. Accessing the File System as a Different User

This section deals with access to the file system as a specific user. This is different from accessing files and directories owned by a different, potentially untrusted user; see Accessing the File System as a Different User.

One approach is to spawn a child process which runs under the target user and group IDs (both effective and real IDs). Note that this child process can block indefinitely, even when processing regular files only. For example, a special FUSE file system could cause the process to hang in uninterruptible sleep inside a stat system call.

An existing process could change its user and group ID using setfsuid and setfsgid. (These functions are preferred over seteuid and setegid because they do not allow the impersonated user to send signals to the process.) These functions are not thread safe. In multi-threaded processes, these operations need to be performed in a single-threaded child process. Unexpected blocking may occur as well.

It is not recommended to try to reimplement the kernel permission checks in user space because the required checks are complex. It is also very difficult to avoid race conditions during path name resolution.

2.3.3. File System Limits

For historical reasons, there are preprocessor constants such as PATH_MAX, NAME_MAX. However, on most systems, the length of canonical path names (absolute path names with all symbolic links resolved, as returned by realpath or canonicalize_file_name) can exceed PATH_MAX bytes, and individual file name components can be longer than NAME_MAX. This is also true of the _PC_PATH_MAX and _PC_NAME_MAX values returned by pathconf, and the f_namemax member of struct statvfs. Therefore, these constants should not be used. This is also reason why the readdir_r should never be used (instead, use readdir).

You should not write code in a way that assumes that there is an upper limit on the number of subdirectories of a directory, the number of regular files in a directory, or the link count of an inode.

2.3.4. File system features

Not all file systems support all features. This makes it very difficult to write general-purpose tools for copying files. For example, a copy operation intending to preserve file permissions will generally fail when copying to a FAT file system.

  • Some file systems are case-insensitive. Most should be case-preserving, though.

  • Name length limits vary greatly, from eight to thousands of bytes. Path length limits differ as well. Most systems impose an upper bound on path names passed to the kernel, but using relative path names, it is possible to create and access files whose absolute path name is essentially of unbounded length.

  • Some file systems do not store names as fairly unrestricted byte sequences, as it has been traditionally the case on GNU systems. This means that some byte sequences (outside the POSIX safe character set) are not valid names. Conversely, names of existing files may not be representable as byte sequences, and the files are thus inaccessible on GNU systems. Some file systems perform Unicode canonicalization on file names. These file systems preserve case, but reading the name of a just-created file using readdir might still result in a different byte sequence.

  • Permissions and owners are not universally supported (and SUID/SGID bits may not be available). For example, FAT file systems assign ownership based on a mount option, and generally mark all files as executable. Any attempt to change permissions would result in an error.

  • Non-regular files (device nodes, FIFOs) are not generally available.

  • Only on some file systems, files can have holes, that is, not all of their contents is backed by disk storage.

  • ioctl support (even fairly generic functionality such as FIEMAP for discovering physical file layout and holes) is file-system-specific.

  • Not all file systems support extended attributes, ACLs and SELinux metadata. Size and naming restriction on extended attributes vary.

  • Hard links may not be supported at all (FAT) or only within the same directory (AFS). Symbolic links may not be available, either. Reflinks (hard links with copy-on-write semantics) are still very rare. Recent systems restrict creation of hard links to users which own the target file or have read/write access to it, but older systems do not.

  • Renaming (or moving) files using rename can fail (even when stat indicates that the source and target directories are located on the same file system). This system call should work if the old and new paths are located in the same directory, though.

  • Locking semantics vary among file systems. This affects advisory and mandatory locks. For example, some network file systems do not allow deleting files which are opened by any process.

  • Resolution of time stamps varies from two seconds to nanoseconds. Not all time stamps are available on all file systems. File creation time (birth time) is not exposed over the stat/fstat interface, even if stored by the file system.

2.3.5. Checking Free Space

The statvfs and fstatvfs functions allow programs to examine the number of available blocks and inodes, through the members f_bfree, f_bavail, f_ffree, and f_favail of struct statvfs. Some file systems return fictional values in the f_ffree and f_favail fields, so the only reliable way to discover if the file system still has space for a file is to try to create it. The f_bfree field should be reasonably accurate, though.

2.4. Temporary Files

In this chapter, we describe how to create temporary files and directories, how to remove them, and how to work with programs which do not create files in ways that are safe with a shared directory for temporary files. General file system manipulation is treated in a separate chapter, File System Manipulation.

Secure creation of temporary files has four different aspects.

  • The location of the directory for temporary files must be obtained in a secure manner (that is, untrusted environment variables must be ignored, see Accessing Environment Variables).

  • A new file must be created. Reusing an existing file must be avoided (the /tmp race condition). This is tricky because traditionally, system-wide temporary directories shared by all users are used.

  • The file must be created in a way that makes it impossible for other users to open it.

  • The descriptor for the temporary file should not leak to subprocesses.

All functions mentioned below will take care of these aspects.

Traditionally, temporary files are often used to reduce memory usage of programs. More and more systems use RAM-based file systems such as tmpfs for storing temporary files, to increase performance and decrease wear on Flash storage. As a result, spooling data to temporary files does not result in any memory savings, and the related complexity can be avoided if the data is kept in process memory.

2.4.1. Obtaining the Location of Temporary Directory

Some functions below need the location of a directory which stores temporary files. For C/C++ programs, use the following steps to obtain that directory:

  • Use secure_getenv to obtain the value of the TMPDIR environment variable. If it is set, convert the path to a fully-resolved absolute path, using realpath(path, NULL). Check if the new path refers to a directory and is writeable. In this case, use it as the temporary directory.

  • Fall back to /tmp.

In Python, you can use the tempfile.tempdir variable.

Java does not support SUID/SGID programs, so you can use the java.lang.System.getenv(String) method to obtain the value of the TMPDIR environment variable, and follow the two steps described above. (Java’s default directory selection does not honor TMPDIR.)

2.4.2. Named Temporary Files

The mkostemp function creates a named temporary file. You should specify the O_CLOEXEC flag to avoid file descriptor leaks to subprocesses. (Applications which do not use multiple threads can also use mkstemp, but libraries should use mkostemp.) For determining the directory part of the file name pattern, see Obtaining the Location of Temporary Directory.

The file is not removed automatically. It is not safe to rename or delete the file before processing, or transform the name in any way (for example, by adding a file extension). If you need multiple temporary files, call mkostemp multiple times. Do not create additional file names derived from the name provided by a previous mkostemp call. However, it is safe to close the descriptor returned by mkostemp and reopen the file using the generated name.

The Python class tempfile.NamedTemporaryFile provides similar functionality, except that the file is deleted automatically by default. Note that you may have to use the file attribute to obtain the actual file object because some programming interfaces cannot deal with file-like objects. The C function mkostemp is also available as tempfile.mkstemp.

In Java, you can use the java.io.File.createTempFile(String, String, File) function, using the temporary file location determined according to Obtaining the Location of Temporary Directory. Do not use java.io.File.deleteOnExit() to delete temporary files, and do not register a shutdown hook for each temporary file you create. In both cases, the deletion hint cannot be removed from the system if you delete the temporary file prior to termination of the VM, causing a memory leak.

2.4.3. Temporary Files without Names

The tmpfile function creates a temporary file and immediately deletes it, while keeping the file open. As a result, the file lacks a name and its space is deallocated as soon as the file descriptor is closed (including the implicit close when the process terminates). This avoids cluttering the temporary directory with orphaned files.

Alternatively, if the maximum size of the temporary file is known beforehand, the fmemopen function can be used to create a FILE * object which is backed by memory.

In Python, unnamed temporary files are provided by the tempfile.TemporaryFile class, and the tempfile.SpooledTemporaryFile class provides a way to avoid creation of small temporary files.

Java does not support unnamed temporary files.

2.4.4. Temporary Directories

The mkdtemp function can be used to create a temporary directory. (For determining the directory part of the file name pattern, see Obtaining the Location of Temporary Directory.) The directory is not automatically removed. In Python, this function is available as tempfile.mkdtemp. In Java 7, temporary directories can be created using the java.nio.file.Files.createTempDirectory(Path, String, FileAttribute…​) function.

When creating files in the temporary directory, use automatically generated names, e.g., derived from a sequential counter. Files with externally provided names could be picked up in unexpected contexts, and crafted names could actually point outside of the tempoary directory (due to directory traversal).

Removing a directory tree in a completely safe manner is complicated. Unless there are overriding performance concerns, the rm program should be used, with the -rf and -- options.

2.4.5. Compensating for Unsafe File Creation

There are two ways to make a function or program which excepts a file name safe for use with temporary files. See Creating Safe Processes, for details on subprocess creation.

  • Create a temporary directory and place the file there. If possible, run the program in a subprocess which uses the temporary directory as its current directory, with a restricted environment. Use generated names for all files in that temporary directory. (See Temporary Directories.)

  • Create the temporary file and pass the generated file name to the function or program. This only works if the function or program can cope with a zero-length existing file. It is safe only under additional assumptions:

    • The function or program must not create additional files whose name is derived from the specified file name or are otherwise predictable.

    • The function or program must not delete the file before processing it.

    • It must not access any existing files in the same directory.

      It is often difficult to check whether these additional assumptions are matched, therefore this approach is not recommended.

2.5. Processes

2.5.1. Creating Safe Processes

This section describes how to create new child processes in a safe manner. In addition to the concerns addressed below, there is the possibility of file descriptor leaks, see Preventing File Descriptor Leaks to Child Processes.

Obtaining the Program Path and the Command-line Template

The name and path to the program being invoked should be hard-coded or controlled by a static configuration file stored at a fixed location (at an file system absolute path). The same applies to the template for generating the command line.

The configured program name should be an absolute path. If it is a relative path, the contents of the PATH must be obtained in a secure manner (see Accessing Environment Variables). If the PATH variable is not set or untrusted, the safe default /bin:/usr/bin must be used.

If too much flexibility is provided here, it may allow invocation of arbitrary programs without proper authorization.

Bypassing the Shell

Child processes should be created without involving the system shell.

For C/C++, system should not be used. The posix_spawn function can be used instead, or a combination fork and execve. (In some cases, it may be preferable to use vfork or the Linux-specific clone system call instead of fork.)

In Python, the subprocess module bypasses the shell by default (when the shell keyword argument is not set to true). os.system should not be used.

The Java class java.lang.ProcessBuilder can be used to create subprocesses without interference from the system shell.

Portability notice

On Windows, there is no argument vector, only a single argument string. Each application is responsible for parsing this string into an argument vector. There is considerable variance among the quoting style recognized by applications. Some of them expand shell wildcards, others do not. Extensive application-specific testing is required to make this secure.

Note that some common applications (notably ssh) unconditionally introduce the use of a shell, even if invoked directly without a shell. It is difficult to use these applications in a secure manner. In this case, untrusted data should be supplied by other means. For example, standard input could be used, instead of the command line.

Specifying the Process Environment

Child processes should be created with a minimal set of environment variables. This is absolutely essential if there is a trust transition involved, either when the parent process was created, or during the creation of the child process.

In C/C++, the environment should be constructed as an array of strings and passed as the envp argument to posix_spawn or execve. The functions setenv, unsetenv and putenv should not be used. They are not thread-safe and suffer from memory leaks.

Python programs need to specify a dict for the the env argument of the subprocess.Popen constructor. The Java class java.lang.ProcessBuilder provides a environment() method, which returns a map that can be manipulated.

The following list provides guidelines for selecting the set of environment variables passed to the child process.

  • PATH should be initialized to /bin:/usr/bin.

  • USER and HOME can be inhereted from the parent process environment, or they can be initialized from the pwent structure for the user.

  • The DISPLAY and XAUTHORITY variables should be passed to the subprocess if it is an X program. Note that this will typically not work across trust boundaries because XAUTHORITY refers to a file with 0600 permissions.

  • The location-related environment variables LANG, LANGUAGE, LC_ADDRESS, LC_ALL, LC_COLLATE, LC_CTYPE, LC_IDENTIFICATION, LC_MEASUREMENT, LC_MESSAGES, LC_MONETARY, LC_NAME, LC_NUMERIC, LC_PAPER, LC_TELEPHONE and LC_TIME can be passed to the subprocess if present.

  • The called process may need application-specific environment variables, for example for passing passwords. (See Passing Secrets to Subprocesses.)

  • All other environment variables should be dropped. Names for new environment variables should not be accepted from untrusted sources.

Robust Argument List Processing

When invoking a program, it is sometimes necessary to include data from untrusted sources. Such data should be checked against embedded NUL characters because the system APIs will silently truncate argument strings at the first NUL character.

The following recommendations assume that the program being invoked uses GNU-style option processing using getopt_long. This convention is widely used, but it is just that, and individual programs might interpret a command line in a different way.

If the untrusted data has to go into an option, use the --option-name=VALUE syntax, placing the option and its value into the same command line argument. This avoids any potential confusion if the data starts with -.

For positional arguments, terminate the option list with a single -- marker after the last option, and include the data at the right position. The -- marker terminates option processing, and the data will not be treated as an option even if it starts with a dash.

Passing Secrets to Subprocesses

The command line (the name of the program and its argument) of a running process is traditionally available to all local users. The called program can overwrite this information, but only after it has run for a bit of time, during which the information may have been read by other processes. However, on Linux, the process environment is restricted to the user who runs the process. Therefore, if you need a convenient way to pass a password to a child process, use an environment variable, and not a command line argument. (See Specifying the Process Environment.)

Portability notice

On some UNIX-like systems (notably Solaris), environment variables can be read by any system user, just like command lines.

If the environment-based approach cannot be used due to portability concerns, the data can be passed on standard input. Some programs (notably gpg) use special file descriptors whose numbers are specified on the command line. Temporary files are an option as well, but they might give digital forensics access to sensitive data (such as passphrases) because it is difficult to safely delete them in all cases.

2.5.2. Handling Child Process Termination

When child processes terminate, the parent process is signalled. A stub of the terminated processes (a zombie, shown as <defunct> by ps) is kept around until the status information is collected (reaped) by the parent process. Over the years, several interfaces for this have been invented:

  • The parent process calls wait, waitpid, waitid, wait3 or wait4, without specifying a process ID. This will deliver any matching process ID. This approach is typically used from within event loops.

  • The parent process calls waitpid, waitid, or wait4, with a specific process ID. Only data for the specific process ID is returned. This is typically used in code which spawns a single subprocess in a synchronous manner.

  • The parent process installs a handler for the SIGCHLD signal, using sigaction, and specifies to the SA_NOCLDWAIT flag. This approach could be used by event loops as well.

None of these approaches can be used to wait for child process terminated in a completely thread-safe manner. The parent process might execute an event loop in another thread, which could pick up the termination signal. This means that libraries typically cannot make free use of child processes (for example, to run problematic code with reduced privileges in a separate address space).

At the moment, the parent process should explicitly wait for termination of the child process using waitpid or waitid, and hope that the status is not collected by an event loop first.

2.5.3. SUID/SGID processes

Programs can be marked in the file system to indicate to the kernel that a trust transition should happen if the program is run. The SUID file permission bit indicates that an executable should run with the effective user ID equal to the owner of the executable file. Similarly, with the SGID bit, the effective group ID is set to the group of the executable file.

Linux supports fscaps, which can grant additional capabilities to a process in a finer-grained manner. Additional mechanisms can be provided by loadable security modules.

When such a trust transition has happened, the process runs in a potentially hostile environment. Additional care is necessary not to rely on any untrusted information. These concerns also apply to libraries which can be linked into such processes.

Accessing Environment Variables

The following steps are required so that a program does not accidentally pick up untrusted data from environment variables.

  • Compile your C/C++ sources with -D_GNU_SOURCE. The Autoconf macro AC_GNU_SOURCE ensures this.

  • Check for the presence of the secure_getenv and secure_getenv function. The Autoconf directive AC_CHECK_FUNCS([secure_getenv secure_getenv]) performs these checks.

  • Arrange for a proper definition of the secure_getenv function. See Obtaining a definition for secure_getenv.

  • Use secure_getenv instead of getenv to obtain the value of critical environment variables. secure_getenv will pretend the variable has not bee set if the process environment is not trusted.

Critical environment variables are debugging flags, configuration file locations, plug-in and log file locations, and anything else that might be used to bypass security restrictions or cause a privileged process to behave in an unexpected way.

Either the secure_getenv function or the __secure_getenv is available from GNU libc.

Example 20. Obtaining a definition for secure_getenv
#include <stdlib.h>

#    define secure_getenv __secure_getenv
#  else
#    error neither secure_getenv nor __secure_getenv are available
#  endif

2.5.4. Daemons

Background processes providing system services (daemons) need to decouple themselves from the controlling terminal and the parent process environment:

  • Fork.

  • In the child process, call setsid. The parent process can simply exit (using _exit, to avoid running clean-up actions twice).

  • In the child process, fork again. Processing continues in the child process. Again, the parent process should just exit.

  • Replace the descriptors 0, 1, 2 with a descriptor for /dev/null. Logging should be redirected to syslog.

Older instructions for creating daemon processes recommended a call to umask(0). This is risky because it often leads to world-writable files and directories, resulting in security vulnerabilities such as arbitrary process termination by untrusted local users, or log file truncation. If the umask needs setting, a restrictive value such as 027 or 077 is recommended.

Other aspects of the process environment may have to changed as well (environment variables, signal handler disposition).

It is increasingly common that server processes do not run as background processes, but as regular foreground process under a supervising master process (such as systemd). Server processes should offer a command line option which disables forking and replacement of the standard output and standard error streams. Such an option is also useful for debugging.

2.5.5. Semantics of Command-line Arguments

After process creation and option processing, it is up to the child process to interpret the arguments. Arguments can be file names, host names, or URLs, and many other things. URLs can refer to the local network, some server on the Internet, or to the local file system. Some applications even accept arbitrary code in arguments (for example, python with the -c option).

Similar concerns apply to environment variables, the contents of the current directory and its subdirectories.

Consequently, careful analysis is required if it is safe to pass untrusted data to another program.

2.5.6. fork as a Primitive for Parallelism

A call to fork which is not immediately followed by a call to execve (perhaps after rearranging and closing file descriptors) is typically unsafe, especially from a library which does not control the state of the entire process. Such use of fork should be replaced with proper child processes or threads.

2.6. Serialization and Deserialization

Protocol decoders and file format parsers are often the most-exposed part of an application because they are exposed with little or no user interaction and before any authentication and security checks are made. They are also difficult to write robustly in languages which are not memory-safe.

2.6.1. Recommendations for Manually-written Decoders

For C and C++, the advice in Recommendations for Pointers and Array Handling applies. In addition, avoid non-character pointers directly into input buffers. Pointer misalignment causes crashes on some architectures.

When reading variable-sized objects, do not allocate large amounts of data solely based on the value of a size field. If possible, grow the data structure as more data is read from the source, and stop when no data is available. This helps to avoid denial-of-service attacks where little amounts of input data results in enormous memory allocations during decoding. Alternatively, you can impose reasonable bounds on memory allocations, but some protocols do not permit this.

2.6.2. Protocol Design

Binary formats with explicit length fields are more difficult to parse robustly than those where the length of dynamically-sized elements is derived from sentinel values. A protocol which does not use length fields and can be written in printable ASCII characters simplifies testing and debugging. However, binary protocols with length fields may be more efficient to parse.

In new datagram-oriented protocols, unique numbers such as sequence numbers or identifiers for fragment reassembly (see Fragmentation) should be at least 64 bits large, and really should not be smaller than 32 bits in size. Protocols should not permit fragments with overlapping contents.

2.6.3. Fragmentation

Some serialization formats use frames or protocol data units (PDUs) on lower levels which are smaller than the PDUs on higher levels. With such an architecture, higher-level PDUs may have to be fragmented into smaller frames during serialization, and frames may need reassembly into large PDUs during deserialization.

Serialization formats may use conceptually similar structures for completely different purposes, for example storing multiple layers and color channels in a single image file.

When fragmenting PDUs, establish a reasonable lower bound for the size of individual fragments (as large as possible—limits as low as one or even zero can add substantial overhead). Avoid fragmentation if at all possible, and try to obtain the maximum acceptable fragment length from a trusted data source.

When implementing reassembly, consider the following aspects.

  • Avoid allocating significant amount of resources without proper authentication. Allocate memory for the unfragmented PDU as more and more and fragments are encountered, and not based on the initially advertised unfragmented PDU size, unless there is a sufficiently low limit on the unfragmented PDU size, so that over-allocation cannot lead to performance problems.

  • Reassembly queues on top of datagram-oriented transports should be bounded, both in the combined size of the arrived partial PDUs waiting for reassembly, and the total number of partially reassembled fragments. The latter limit helps to reduce the risk of accidental reassembly of unrelated fragments, as it can happen with small fragment IDs (see Fragment IDs). It also guards to some extent against deliberate injection of fragments, by guessing fragment IDs.

  • Carefully keep track of which bytes in the unfragmented PDU have been covered by fragments so far. If message reordering is a concern, the most straightforward data structure for this is an array of bits, with one bit for every byte (or other atomic unit) in the unfragmented PDU. Complete reassembly can be determined by increasing a counter of set bits in the bit array as the bit array is updated, taking overlapping fragments into consideration.

  • Reject overlapping fragments (that is, multiple fragments which provide data at the same offset of the PDU being fragmented), unless the protocol explicitly requires accepting overlapping fragments. The bit array used for tracking already arrived bytes can be used for this purpose.

  • Check for conflicting values of unfragmented PDU lengths (if this length information is part of every fragment) and reject fragments which are inconsistent.

  • Validate fragment lengths and offsets of individual fragments against the unfragmented PDU length (if they are present). Check that the last byte in the fragment does not lie after the end of the unfragmented PDU. Avoid integer overflows in these computations (see Recommendations for Integer Arithmetic).

Fragment IDs

If the underlying transport is datagram-oriented (so that PDUs can be reordered, duplicated or be lost, like with UDP), fragment reassembly needs to take into account endpoint addresses of the communication channel, and there has to be some sort of fragment ID which identifies the individual fragments as part of a larger PDU. In addition, the fragmentation protocol will typically involve fragment offsets and fragment lengths, as mentioned above.

If the transport may be subject to blind PDU injection (again, like UDP), the fragment ID must be generated randomly. If the fragment ID is 64 bit or larger (strongly recommended), it can be generated in a completely random fashion for most traffic volumes. If it is less than 64 bits large (so that accidental collisions can happen if a lot of PDUs are transmitted), the fragment ID should be incremented sequentially from a starting value. The starting value should be derived using a HMAC-like construction from the endpoint addresses, using a long-lived random key. This construction ensures that despite the limited range of the ID, accidental collisions are as unlikely as possible. (This will not work reliable with really short fragment IDs, such as the 16 bit IDs used by the Internet Protocol.)

2.6.4. Library Support for Deserialization

For some languages, generic libraries are available which allow to serialize and deserialize user-defined objects. The deserialization part comes in one of two flavors, depending on the library. The first kind uses type information in the data stream to control which objects are instantiated. The second kind uses type definitions supplied by the programmer. The first one allows arbitrary object instantiation, the second one generally does not.

The following serialization frameworks are in the first category, are known to be unsafe, and must not be used for untrusted data:

  • Python’s pickle and cPickle modules, and wrappers such as shelve

  • Perl’s Storable package

  • Java serialization (java.io.ObjectInputStream), even if encoded in other formats (as with java.beans.XMLDecoder)

  • PHP serialization (unserialize)

  • Most implementations of YAML

When using a type-directed deserialization format where the types of the deserialized objects are specified by the programmer, make sure that the objects which can be instantiated cannot perform any destructive actions in their destructors, even when the data members have been manipulated.

In general, JSON decoders do not suffer from this problem. But you must not use the eval function to parse JSON objects in Javascript; even with the regular expression filter from RFC 4627, there are still information leaks remaining. JSON-based formats can still turn out risky if they serve as an encoding form for any if the serialization frameworks listed above.

For serialization in C and C++ projects, the Protocol Buffers serialization (protobuf) provides type safe automated serialization by relying on code generation. It is positioned as similar, but simpler and more efficient to XML serialization.

2.6.5. XML Serialization

External References

XML documents can contain external references. They can occur in various places.

  • In the DTD declaration in the header of an XML document:

      "-//W3C//DTD XHTML 1.0 Transitional//EN"
  • In a namespace declaration:

    <xsd:schema xmlns:xsd="http://www.w3.org/2001/XMLSchema">
  • In an entity defintion:

    <!ENTITY sys SYSTEM "http://www.example.com/ent.adoc[]>
    <!ENTITY pub PUBLIC "-//Example//Public Entity//EN"
  • In a notation:

    <!NOTATION not SYSTEM "../not.adoc[]>

Originally, these external references were intended as unique identifiers, but by many XML implementations, they are used for locating the data for the referenced element. This causes unwanted network traffic, and may disclose file system contents or otherwise unreachable network resources, so this functionality should be disabled.

Depending on the XML library, external referenced might be processed not just when parsing XML, but also when generating it.

Entity Expansion

When external DTD processing is disabled, an internal DTD subset can still contain entity definitions. Entity declarations can reference other entities. Some XML libraries expand entities automatically, and this processing cannot be switched off in some places (such as attribute values or content models). Without limits on the entity nesting level, this expansion results in data which can grow exponentially in length with size of the input. (If there is a limit on the nesting level, the growth is still polynomial, unless further limits are imposed.)

Consequently, the processing internal DTD subsets should be disabled if possible, and only trusted DTDs should be processed. If a particular XML application does not permit such restrictions, then application-specific limits are called for.

XInclude Processing

XInclude processing can reference file and network resources and include them into the document, much like external entity references. When parsing untrusted XML documents, XInclude processing should be turned off.

XInclude processing is also fairly complex and may pull in support for the XPointer and XPath specifications, considerably increasing the amount of code required for XML processing.

Algorithmic Complexity of XML Validation

DTD-based XML validation uses regular expressions for content models. The XML specification requires that content models are deterministic, which means that efficient validation is possible. However, some implementations do not enforce determinism, and require exponential (or just polynomial) amount of space or time for validating some DTD/document combinations.

XML schemas and RELAX NG (via the xsd: prefix) directly support textual regular expressions which are not required to be deterministic.

Using Expat for XML parsing

By default, Expat does not try to resolve external IDs, so no steps are required to block them. However, internal entity declarations are processed. Installing a callback which stops parsing as soon as such entities are encountered disables them, see Disabling XML entity processing with Expat. Expat does not perform any validation, so there are no problems related to that.

Example 21. Disabling XML entity processing with Expat

// Stop the parser when an entity declaration is encountered.
static void
EntityDeclHandler(void *userData,
		  const XML_Char *entityName, int is_parameter_entity,
		  const XML_Char *value, int value_length,
		  const XML_Char *base, const XML_Char *systemId,
		  const XML_Char *publicId, const XML_Char *notationName)
  XML_StopParser((XML_Parser)userData, XML_FALSE);

This handler must be installed when the XML_Parser object is created (Creating an Expat XML parser).

Example 22. Creating an Expat XML parser

XML_Parser parser = XML_ParserCreate("UTF-8");
if (parser == NULL) {
  fprintf(stderr, "XML_ParserCreate failed\n");
// EntityDeclHandler needs a reference to the parser to stop
// parsing.
XML_SetUserData(parser, parser);
// Disable entity processing, to inhibit entity expansion.
XML_SetEntityDeclHandler(parser, EntityDeclHandler);

It is also possible to reject internal DTD subsets altogether, using a suitable XML_StartDoctypeDeclHandler handler installed with XML_SetDoctypeDeclHandler.

Using Qt for XML Parsing

The XML component of Qt, QtXml, does not resolve external IDs by default, so it is not required to prevent such resolution. Internal entities are processed, though. To change that, a custom QXmlDeclHandler and QXmlSimpleReader subclasses are needed. It is not possible to use the QDomDocument::setContent(const QByteArray &) convenience methods.

A QtXml entity handler which blocks entity processing shows an entity handler which always returns errors, causing parsing to stop when encountering entity declarations.

Example 23. A QtXml entity handler which blocks entity processing

class NoEntityHandler : public QXmlDeclHandler {
  bool attributeDecl(const QString&, const QString&, const QString&,
		       const QString&, const QString&);
  bool internalEntityDecl(const QString&, const QString&);
  bool externalEntityDecl(const QString&, const QString&,
			    const QString&);
  QString errorString() const;

  (const QString&, const QString&, const QString&, const QString&,
   const QString&)
  return false;

NoEntityHandler::internalEntityDecl(const QString&, const QString&)
  return false;

NoEntityHandler::externalEntityDecl(const QString&, const QString&, const
  return false;

NoEntityHandler::errorString() const
  return "XML declaration not permitted";

This handler is used in the custom QXmlReader subclass in A QtXml XML reader which blocks entity processing. Some parts of QtXml will call the setDeclHandler(QXmlDeclHandler *) method. Consequently, we prevent overriding our custom handler by providing a definition of this method which does nothing. In the constructor, we activate namespace processing; this part may need adjusting.

Example 24. A QtXml XML reader which blocks entity processing

class NoEntityReader : public QXmlSimpleReader {
  NoEntityHandler handler;
  void setDeclHandler(QXmlDeclHandler *);

  setFeature("http://xml.org/sax/features/namespaces", true);
  setFeature("http://xml.org/sax/features/namespace-prefixes", false);

NoEntityReader::setDeclHandler(QXmlDeclHandler *)
  // Ignore the handler which was passed in.

Our NoEntityReader class can be used with one of the overloaded QDomDocument::setContent methods. Parsing an XML document with QDomDocument, without entity expansion shows how the buffer object (of type QByteArray) is wrapped as a QXmlInputSource. After calling the setContent method, you should check the return value and report any error.

Example 25. Parsing an XML document with QDomDocument, without entity expansion

NoEntityReader reader;
QBuffer buffer(&data);
QXmlInputSource source(&buffer);
QDomDocument doc;
QString errorMsg;
int errorLine;
int errorColumn;
bool okay = doc.setContent
  (&source, &reader, &errorMsg, &errorLine, &errorColumn);
Using OpenJDK for XML Parsing and Validation

OpenJDK contains facilities for DOM-based, SAX-based, and StAX-based document parsing. Documents can be validated against DTDs or XML schemas.

The approach taken to deal with entity expansion differs from the general recommendation in Entity Expansion. We enable the the feature flag javax.xml.XMLConstants.FEATURE_SECURE_PROCESSING, which enforces heuristic restrictions on the number of entity expansions. Note that this flag alone does not prevent resolution of external references (system IDs or public IDs), so it is slightly misnamed.

In the following sections, we use helper classes to prevent external ID resolution.

Example 26. Helper class to prevent DTD external entity resolution in OpenJDK

class NoEntityResolver implements EntityResolver {
    public InputSource resolveEntity(String publicId, String systemId)
            throws SAXException, IOException {
        // Throwing an exception stops validation.
        throw new IOException(String.format(
                "attempt to resolve \"%s\" \"%s\"", publicId, systemId));
Example 27. Helper class to prevent schema resolution in OpenJDK

class NoResourceResolver implements LSResourceResolver {
    public LSInput resolveResource(String type, String namespaceURI,
            String publicId, String systemId, String baseURI) {
        // Throwing an exception stops validation.
        throw new RuntimeException(String.format(
                "resolution attempt: type=%s namespace=%s " +
                "publicId=%s systemId=%s baseURI=%s",
                type, namespaceURI, publicId, systemId, baseURI));

Java imports for OpenJDK XML parsing shows the imports used by the examples.

Example 28. Java imports for OpenJDK XML parsing

import javax.xml.XMLConstants;
import javax.xml.parsers.DocumentBuilder;
import javax.xml.parsers.DocumentBuilderFactory;
import javax.xml.parsers.ParserConfigurationException;
import javax.xml.parsers.SAXParser;
import javax.xml.parsers.SAXParserFactory;
import javax.xml.transform.dom.DOMSource;
import javax.xml.transform.sax.SAXSource;
import javax.xml.validation.Schema;
import javax.xml.validation.SchemaFactory;
import javax.xml.validation.Validator;

import org.w3c.dom.Document;
import org.w3c.dom.ls.LSInput;
import org.w3c.dom.ls.LSResourceResolver;
import org.xml.sax.EntityResolver;
import org.xml.sax.ErrorHandler;
import org.xml.sax.InputSource;
import org.xml.sax.SAXException;
import org.xml.sax.SAXParseException;
import org.xml.sax.XMLReader;
DOM-based XML parsing and DTD validation in OpenJDK

This approach produces a org.w3c.dom.Document object from an input stream. DOM-based XML parsing in OpenJDK use the data from the java.io.InputStream instance in the inputStream variable.

Example 29. DOM-based XML parsing in OpenJDK

DocumentBuilderFactory factory = DocumentBuilderFactory.newInstance();
// Impose restrictions on the complexity of the DTD.
factory.setFeature(XMLConstants.FEATURE_SECURE_PROCESSING, true);

// Turn on validation.
// This step can be omitted if validation is not desired.

// Parse the document.
DocumentBuilder builder = factory.newDocumentBuilder();
builder.setEntityResolver(new NoEntityResolver());
builder.setErrorHandler(new Errors());
Document document = builder.parse(inputStream);

External entity references are prohibited using the NoEntityResolver class in Helper class to prevent DTD external entity resolution in OpenJDK. Because external DTD references are prohibited, DTD validation (if enabled) will only happen against the internal DTD subset embedded in the XML document.

To validate the document against an external DTD, use a javax.xml.transform.Transformer class to add the DTD reference to the document, and an entity resolver which whitelists this external reference.

XML Schema Validation in OpenJDK

SAX-based validation against an XML schema in OpenJDK shows how to validate a document against an XML Schema, using a SAX-based approach. The XML data is read from an java.io.InputStream in the inputStream variable.

Example 30. SAX-based validation against an XML schema in OpenJDK

SchemaFactory factory = SchemaFactory.newInstance(

// This enables restrictions on the schema and document
// complexity.
factory.setFeature(XMLConstants.FEATURE_SECURE_PROCESSING, true);

// This prevents resource resolution by the schema itself.
// If the schema is trusted and references additional files,
// this line must be omitted, otherwise loading these files
// will fail.
factory.setResourceResolver(new NoResourceResolver());

Schema schema = factory.newSchema(schemaFile);
Validator validator = schema.newValidator();

// This prevents external resource resolution.
validator.setResourceResolver(new NoResourceResolver());

validator.validate(new SAXSource(new InputSource(inputStream)));

The NoResourceResolver class is defined in Helper class to prevent schema resolution in OpenJDK.

If you need to validate a document against an XML schema, use the code in DOM-based XML parsing in OpenJDK to create the document, but do not enable validation at this point. Then use Validation of a DOM document against an XML schema in OpenJDK to perform the schema-based validation on the org.w3c.dom.Document instance document.

Example 31. Validation of a DOM document against an XML schema in OpenJDK

SchemaFactory factory = SchemaFactory.newInstance(

// This enables restrictions on schema complexity.
factory.setFeature(XMLConstants.FEATURE_SECURE_PROCESSING, true);

// The following line prevents resource resolution
// by the schema itself.
factory.setResourceResolver(new NoResourceResolver());

Schema schema = factory.newSchema(schemaFile);

Validator validator = schema.newValidator();

// This prevents external resource resolution.
validator.setResourceResolver(new NoResourceResolver());
validator.validate(new DOMSource(document));
Other XML Parsers in OpenJDK

OpenJDK contains additional XML parsing and processing facilities. Some of them are insecure.

The class java.beans.XMLDecoder acts as a bridge between the Java object serialization format and XML. It is close to impossible to securely deserialize Java objects in this format from untrusted inputs, so its use is not recommended, as with the Java object serialization format itself. See Library Support for Deserialization.

2.6.6. Protocol Encoders

For protocol encoders, you should write bytes to a buffer which grows as needed, using an exponential sizing policy. Explicit lengths can be patched in later, once they are known. Allocating the required number of bytes upfront typically requires separate code to compute the final size, which must be kept in sync with the actual encoding step, or vulnerabilities may result. In multi-threaded code, parts of the object being deserialized might change, so that the computed size is out of date.

You should avoid copying data directly from a received packet during encoding, disregarding the format. Propagating malformed data could enable attacks on other recipients of that data.

When using C or C++ and copying whole data structures directly into the output, make sure that you do not leak information in padding bytes between fields or at the end of the struct.

2.7. Cryptography

2.7.1. Primitives

Choosing from the following cryptographic primitives is recommended:

  • RSA with 2048-bit keys and OAEP or PSS padding

  • AES-128 in CBC mode

  • AES-128 in GCM mode

  • AES-256 in CBC mode

  • AES-256 in GCM mode

  • SHA-256

  • HMAC-SHA-256

  • HMAC-SHA-1

Other cryptographic algorithms can be used if they are required for interoperability with existing software:

  • RSA with key sizes larger than 1024 and legacy padding

  • AES-192

  • 3DES (triple DES, with two or three 56-bit keys), but strongly discouraged

  • RC4 (but very, very strongly discouraged)

  • SHA-1

  • HMAC-MD5


These primitives are difficult to use in a secure way. Custom implementation of security protocols should be avoided. For protecting confidentiality and integrity of network transmissions, TLS should be used (Transport Layer Security (TLS)).

In particular, when using AES in CBC mode, it is necessary to add integrity checking by other means, preferably using HMAC-SHA-256 and after encryption (that is, on the encrypted cipher text). For AES in GCM mode, correct construction of nonces is absolutely essential.

2.7.2. Randomness

The following facilities can be used to generate unpredictable and non-repeating values. When these functions are used without special safeguards, each individual random value should be at least 12 bytes long.

  • PK11_GenerateRandom in the NSS library (usable for high data rates)

  • RAND_bytes in the OpenSSL library (usable for high data rates)

  • gnutls_rnd in GNUTLS, with GNUTLS_RND_RANDOM as the first argument (usable for high data rates)

  • java.security.SecureRandom in Java (usable for high data rates)

  • os.urandom in Python

  • The getrandom system call since glibc 2.25

  • The getentropy call since glibc 2.25

  • Reading from the /dev/urandom character device

All these functions should be non-blocking, and they should not wait until physical randomness becomes available. (Some cryptography providers for Java can cause java.security.SecureRandom to block, however.) Those functions which do not obtain all bits directly from /dev/urandom are suitable for high data rates because they do not deplete the system-wide entropy pool.

Difficult to use API

Both RAND_bytes and PK11_GenerateRandom have three-state return values (with conflicting meanings). Careful error checking is required. Please review the documentation when using these functions.

Difficult to use API

The getrandom system call has three-state return values, hence requires careful error checking.

It was introduced in Linux kernel 3.17, but before glibc 2.25 no API wrappers were provided. As such one could only use it via the syscall interface as syscall(SYS_getrandom, (void*)dest, (size_t)size, (unsigned int)0). For portable code targetting multiple kernel versions one has to check for the function beingavailable on run-time, and switch to another facility if the running kernel does not support this call.

Other sources of randomness should be considered predictable.

Generating randomness for cryptographic keys in long-term use may need different steps and is best left to cryptographic libraries.

2.8. RPM Packaging

This chapter deals with security-related concerns around RPM packaging. It has to be read in conjunction with distribution-specific packaging guidelines.

2.8.1. Generating X.509 Self-signed Certificates during Installation

Some applications need X.509 certificates for authentication purposes. For example, a single private/public key pair could be used to define cluster membership, enabling authentication and encryption of all intra-cluster communication. (Lack of certification from a CA matters less in such a context.) For such use, generating the key pair at package installation time when preparing system images for use in the cluster is reasonable. For other use cases, it is necessary to generate the key pair before the service is started for the first time, see Generating X.509 Self-signed Certificates before Service Start, and Packaging:Initial Service Setup.


The way the key is generated may not be suitable for key material of critical value. (openssl genrsa uses, but does not require, entropy from a physical source of randomness, among other things.) Such keys should be stored in a hardware security module if possible, and generated from random bits reserved for this purpose derived from a non-deterministic physical source.

In the spec file, we define two RPM variables which contain the names of the files used to store the private and public key, and the user name for the service:

# Name of the user owning the file with the private key
%define tlsuser %{name}
# Name of the directory which contains the key and certificate files
%define tlsdir %{_sysconfdir}/%{name}
%define tlskey %{tlsdir}/%{name}.key
%define tlscert %{tlsdir}/%{name}.crt

These variables likely need adjustment based on the needs of the package.

Typically, the file with the private key needs to be owned by the system user which needs to read it, %{tlsuser} (not root). In order to avoid races, if the directory %{tlsdir} is owned by the services user, you should use the code in Creating a key pair in a user-owned directory. The invocation of su with the -s /bin/bash argument is necessary in case the login shell for the user has been disabled.

Example 32. Creating a key pair in a user-owned directory
if [ $1 -eq 1 ] ; then
  if ! test -e %{tlskey} ; then
    su -s /bin/bash \
      -c "umask 077 && openssl genrsa -out %{tlskey} 2048 2>/dev/null" \
  if ! test -e %{tlscert} ; then
    cn="Automatically generated certificate for the %{tlsuser} service"
    req_args="-key %{tlskey} -out %{tlscert} -days 7305 -subj \"/CN=$cn/\""
    su -s /bin/bash \
      -c "openssl req -new -x509 -extensions usr_cert $req_args" \

%dir %attr(0755,%{tlsuser},%{tlsuser]) %{tlsdir}
%ghost %attr(0600,%{tlsuser},%{tlsuser}) %config(noreplace) %{tlskey}
%ghost %attr(0644,%{tlsuser},%{tlsuser}) %config(noreplace) %{tlscert}

The files containing the key material are marked as ghost configuration files. This ensures that they are tracked in the RPM database as associated with the package, but RPM will not create them when the package is installed and not verify their contents (the %ghost), or delete the files when the package is uninstalled (the %config(noreplace) part).

If the directory %{tlsdir} is owned by root, use the code in Creating a key pair in a root-owned directory.

Example 33. Creating a key pair in a root-owned directory
if [ $1 -eq 1 ] ; then
  if ! test -e %{tlskey} ; then
    (umask 077 && openssl genrsa -out %{tlskey} 2048 2>/dev/null)
    chown %{tlsuser} %{tlskey}
  if ! test -e %{tlscert} ; then
    cn="Automatically generated certificate for the %{tlsuser} service"
    openssl req -new -x509 -extensions usr_cert \
      -key %{tlskey} -out %{tlscert} -days 7305 -subj "/CN=$cn/"

%dir %attr(0755,root,root]) %{tlsdir}
%ghost %attr(0600,%{tlsuser},%{tlsuser}) %config(noreplace) %{tlskey}
%ghost %attr(0644,root,root) %config(noreplace) %{tlscert}

In order for this to work, the package which generates the keys must require the openssl package. If the user which owns the key file is generated by a different package, the package generating the certificate must specify a Requires(pre): on the package which creates the user. This ensures that the user account will exist when it is needed for the su or chmod invocation.

2.8.2. Generating X.509 Self-signed Certificates before Service Start

An alternative way to automatically provide an X.509 key pair is to create it just before the service is started for the first time. This ensures that installation images which are created from installed RPM packages receive different key material. Creating the key pair at package installation time (see Generating X.509 Self-signed Certificates during Installation) would put the key into the image, which may or may not make sense.


The caveats about the way the key is generated in Generating X.509 Self-signed Certificates during Installation apply to this procedure as well.

Generating key material before service start may happen very early during boot, when the kernel randomness pool has not yet been initialized. Currently, the only way to check for the initialization is to look for the kernel message random: nonblocking pool is initialized, or ensure that the application used for generating the keys is utilizing the getrandom() system call.

In theory, it is also possible to use an application which reads from /dev/random while generating the key material (instead of /dev/urandom), but this can block not just during the boot process, but also much later at run time, and generally results in a poor user experience.

The requirements for generating such keys is documented at Packaging:Initial Service Setup.

3. Implementing Security Features

3.1. Authentication and Authorization

3.1.1. Authenticating Servers

When connecting to a server, a client has to make sure that it is actually talking to the server it expects. There are two different aspects, securing the network path, and making sure that the expected user runs the process on the target host. There are several ways to ensure that:

  • The server uses a TLS certificate which is valid according to the web browser public key infrastructure, and the client verifies the certificate and the host name.

  • The server uses a TLS certificate which is expected by the client (perhaps it is stored in a configuration file read by the client). In this case, no host name checking is required.

  • On Linux, UNIX domain sockets (of the PF_UNIX protocol family, sometimes called PF_LOCAL) are restricted by file system permissions. If the server socket path is not world-writable, the server identity cannot be spoofed by local users.

  • Port numbers less than 1024 (trusted ports) can only be used by root, so if a UDP or TCP server is running on the local host and it uses a trusted port, its identity is assured. (Not all operating systems enforce the trusted ports concept, and the network might not be trusted, so it is only useful on the local system.)

TLS (Transport Layer Security (TLS)) is the recommended way for securing connections over untrusted networks.

If the server port number is 1024 is higher, a local user can impersonate the process by binding to this socket, perhaps after crashing the real server by exploiting a denial-of-service vulnerability.

3.1.2. Host-based Authentication

Host-based authentication uses access control lists (ACLs) to accept or deny requests from clients. This authentication method comes in two flavors: IP-based (or, more generally, address-based) and name-based (with the name coming from DNS or /etc/hosts). IP-based ACLs often use prefix notation to extend access to entire subnets. Name-based ACLs sometimes use wildcards for adding groups of hosts (from entire DNS subtrees). (In the SSH context, host-based authentication means something completely different and is not covered in this section.)

Host-based authentication trust the network and may not offer sufficient granularity, so it has to be considered a weak form of authentication. On the other hand, IP-based authentication can be made extremely robust and can be applied very early in input processing, so it offers an opportunity for significantly reducing the number of potential attackers for many services.

The names returned by gethostbyaddr and getnameinfo functions cannot be trusted. (DNS PTR records can be set to arbitrary values, not just names belong to the address owner.) If these names are used for ACL matching, a forward lookup using gethostbyaddr or getaddrinfo has to be performed. The name is only valid if the original address is found among the results of the forward lookup (double-reverse lookup).

An empty ACL should deny all access (deny-by-default). If empty ACLs permits all access, configuring any access list must switch to deny-by-default for all unconfigured protocols, in both name-based and address-based variants.

Similarly, if an address or name is not matched by the list, it should be denied. However, many implementations behave differently, so the actual behavior must be documented properly.

IPv6 addresses can embed IPv4 addresses. There is no universally correct way to deal with this ambiguity. The behavior of the ACL implementation should be documented.

3.1.3. UNIX Domain Socket Authentication

UNIX domain sockets (with address family AF_UNIX or AF_LOCAL) are restricted to the local host and offer a special authentication mechanism: credentials passing.

Nowadays, most systems support the SO_PEERCRED (Linux) or LOCAL_PEERCRED (FreeBSD) socket options, or the getpeereid (other BSDs, OS X). These interfaces provide direct access to the (effective) user ID on the other end of a domain socket connect, without cooperation from the other end.

Historically, credentials passing was implemented using ancillary data in the sendmsg and recvmsg functions. On some systems, only credentials data that the peer has explicitly sent can be received, and the kernel checks the data for correctness on the sending side. This means that both peers need to deal with ancillary data. Compared to that, the modern interfaces are easier to use. Both sets of interfaces vary considerably among UNIX-like systems, unfortunately.

If you want to authenticate based on supplementary groups, you should obtain the user ID using one of these methods, and look up the list of supplementary groups using getpwuid (or getpwuid_r) and getgrouplist. Using the PID and information from /proc/PID/status is prone to race conditions and insecure.

Netlink messages are used as a high-performance data transfer mechanism between the kernel and the user space. Traditionally, they are used to exchange information related to the network stack, such as routing table entries.

When processing Netlink messages from the kernel, it is important to check that these messages actually originate from the kernel, by checking that the port ID (or PID) field nl_pid in the sockaddr_nl structure is 0. (This structure can be obtained using recvfrom or recvmsg, it is different from the nlmsghdr structure.) The kernel does not prevent other processes from sending unicast Netlink messages, but the nl_pid field in the sender’s socket address will be non-zero in such cases.

Applications should not use AF_NETLINK sockets as an IPC mechanism among processes, but prefer UNIX domain sockets for this tasks.

3.2. Transport Layer Security (TLS)

Transport Layer Security (TLS, formerly Secure Sockets Layer/SSL) is the recommended way to to protect integrity and confidentiality while data is transferred over an untrusted network connection, and to identify the endpoint. At this chapter we describe the available libraries in Fedora as well as known pitfalls, and safe ways to write applications with them.

When using any library, in addition to this guide, it is recommended to consult the library' documentation.

3.2.1. Common Pitfalls

TLS implementations are difficult to use, and most of them lack a clean API design. The following sections contain implementation-specific advice, and some generic pitfalls are mentioned below.

  • Most TLS implementations have questionable default TLS cipher suites. Most of them enable anonymous Diffie-Hellman key exchange (but we generally want servers to authenticate themselves). Many do not disable ciphers which are subject to brute-force attacks because of restricted key lengths. Some even disable all variants of AES in the default configuration.

    When overriding the cipher suite defaults, it is recommended to disable all cipher suites which are not present on a whitelist, instead of simply enabling a list of cipher suites. This way, if an algorithm is disabled by default in the TLS implementation in a future security update, the application will not re-enable it.

  • The name which is used in certificate validation must match the name provided by the user or configuration file. No host name canonicalization or IP address lookup must be performed.

  • The TLS handshake has very poor performance if the TCP Nagle algorithm is active. You should switch on the TCP_NODELAY socket option (at least for the duration of the handshake), or use the Linux-specific TCP_CORK option.

    Example 34. Deactivating the TCP Nagle algorithm
    const int val = 1;
    int ret = setsockopt(sockfd, IPPROTO_TCP, TCP_NODELAY, &val, sizeof(val));
    if (ret < 0) {
  • Implementing proper session resumption decreases handshake overhead considerably. This is important if the upper-layer protocol uses short-lived connections (like most application of HTTPS).

  • Both client and server should work towards an orderly connection shutdown, that is send close_notify alerts and respond to them. This is especially important if the upper-layer protocol does not provide means to detect connection truncation (like some uses of HTTP).

  • When implementing a server using event-driven programming, it is important to handle the TLS handshake properly because it includes multiple network round-trips which can block when an ordinary TCP accept would not. Otherwise, a client which fails to complete the TLS handshake for some reason will prevent the server from handling input from other clients.

  • Unlike regular file descriptors, TLS connections cannot be passed between processes. Some TLS implementations add additional restrictions, and TLS connections generally cannot be used across fork function calls (see fork as a Primitive for Parallelism).

OpenSSL Pitfalls

Some OpenSSL function use tri-state return values. Correct error checking is extremely important. Several functions return int values with the following meaning:

  • The value 1 indicates success (for example, a successful signature verification).

  • The value 0 indicates semantic failure (for example, a signature verification which was unsuccessful because the signing certificate was self-signed).

  • The value -1 indicates a low-level error in the system, such as failure to allocate memory using malloc.

Treating such tri-state return values as booleans can lead to security vulnerabilities. Note that some OpenSSL functions return boolean results or yet another set of status indicators. Each function needs to be checked individually.

Recovering precise error information is difficult. Obtaining OpenSSL error codes shows how to obtain a more precise error code after a function call on an SSL object has failed. However, there are still cases where no detailed error information is available (e.g., if SSL_shutdown fails due to a connection teardown by the other end).

Example 35. Obtaining OpenSSL error codes

static void __attribute__((noreturn))
ssl_print_error_and_exit(SSL *ssl, const char *op, int ret)
  int subcode = SSL_get_error(ssl, ret);
  switch (subcode) {
    fprintf(stderr, "error: %s: no error to report\n", op);
    fprintf(stderr, "error: %s: invalid blocking state %d\n", op, subcode);
    fprintf(stderr, "error: %s: TLS layer problem\n", op);
    fprintf(stderr, "error: %s: system call failed: %s\n", op, strerror(errno));
    fprintf(stderr, "error: %s: zero return\n", op);

The OPENSSL_config function is documented to never fail. In reality, it can terminate the entire process if there is a failure accessing the configuration file. An error message is written to standard error, but which might not be visible if the function is called from a daemon process.

OpenSSL contains two separate ASN.1 DER decoders. One set of decoders operate on BIO handles (the input/output stream abstraction provided by OpenSSL); their decoder function names start with d2i_ and end in _fp or _bio (e.g., d2i_X509_fp or d2i_X509_bio). These decoders must not be used for parsing data from untrusted sources; instead, the variants without the _fp and _bio (e.g., d2i_X509) shall be used. The BIO variants have received considerably less testing and are not very robust.

For the same reason, the OpenSSL command line tools (such as openssl x509) are generally generally less robust than the actual library code. They use the BIO functions internally, and not the more robust variants.

The command line tools do not always indicate failure in the exit status of the openssl process. For instance, a verification failure in openssl verify result in an exit status of zero.

OpenSSL command-line commands, such as openssl genrsa, do not ensure that physical entropy is used for key generation—they obtain entropy from /dev/urandom and other sources, but not from /dev/random. This can result in weak keys if the system lacks a proper entropy source (e.g., a virtual machine with solid state storage). Depending on local policies, keys generated by these OpenSSL tools should not be used in high-value, critical functions.

The OpenSSL server and client applications (openssl s_client and openssl s_server) are debugging tools and should never be used as generic clients. For instance, the s_client tool reacts in a surprising way to lines starting with R and Q.

OpenSSL allows application code to access private key material over documented interfaces. This can significantly increase the part of the code base which has to undergo security certification.

GnuTLS Pitfalls

Older versions of GnuTLS had several peculiarities described in previous versions of this guide; as of GnuTLS 3.3.10, these issues are no longer applicable.

OpenJDK Pitfalls

The Java cryptographic framework is highly modular. As a result, when you request an object implementing some cryptographic functionality, you cannot be completely sure that you end up with the well-tested, reviewed implementation in OpenJDK.

OpenJDK (in the source code as published by Oracle) and other implementations of the Java platform require that the system administrator has installed so-called unlimited strength jurisdiction policy files. Without this step, it is not possible to use the secure algorithms which offer sufficient cryptographic strength. Most downstream redistributors of OpenJDK remove this requirement.

Some versions of OpenJDK use /dev/random as the randomness source for nonces and other random data which is needed for TLS operation, but does not actually require physical randomness. As a result, TLS applications can block, waiting for more bits to become available in /dev/random.

NSS Pitfalls

NSS was not designed to be used by other libraries which can be linked into applications without modifying them. There is a lot of global state. There does not seem to be a way to perform required NSS initialization without race conditions.

If the NSPR descriptor is in an unexpected state, the SSL_ForceHandshake function can succeed, but no TLS handshake takes place, the peer is not authenticated, and subsequent data is exchanged in the clear.

NSS disables itself if it detects that the process underwent a fork after the library has been initialized. This behavior is required by the PKCS#11 API specification.

3.2.2. TLS Clients

Secure use of TLS in a client generally involves all of the following steps. (Individual instructions for specific TLS implementations follow in the next sections.)

  • The client must configure the TLS library to use a set of trusted root certificates. These certificates are provided by the system in various formats and files. These are documented in update-ca-trust man page in Fedora. Portable applications should not hard-code any paths; they should rely on APIs which set the default for the system trust store.

  • The client selects sufficiently strong cryptographic primitives and disables insecure ones (such as no-op encryption). Compression support and SSL version 3 or lower must be disabled (including the SSLv2-compatible handshake).

  • The client initiates the TLS connection. The Server Name Indication extension should be used if supported by the TLS implementation. Before switching to the encrypted connection state, the contents of all input and output buffers must be discarded.

  • The client needs to validate the peer certificate provided by the server, that is, the client must check that there is a cryptographically protected chain from a trusted root certificate to the peer certificate. (Depending on the TLS implementation, a TLS handshake can succeed even if the certificate cannot be validated.)

  • The client must check that the configured or user-provided server name matches the peer certificate provided by the server.

It is safe to provide users detailed diagnostics on certificate validation failures. Other causes of handshake failures and, generally speaking, any details on other errors reported by the TLS implementation (particularly exception tracebacks), must not be divulged in ways that make them accessible to potential attackers. Otherwise, it is possible to create decryption oracles.


Depending on the application, revocation checking (against certificate revocations lists or via OCSP) and session resumption are important aspects of production-quality client. These aspects are not yet covered.

Implementation TLS Clients With OpenSSL

In the following code, the error handling is only exploratory. Proper error handling is required for production use, especially in libraries.

The OpenSSL library needs explicit initialization (see OpenSSL library initialization).

Example 36. OpenSSL library initialization

// The following call prints an error message and calls exit() if
// the OpenSSL configuration file is unreadable.
// Provide human-readable error messages.
// Register ciphers.

After that, a context object has to be created, which acts as a factory for connection objects (OpenSSL client context creation). We use an explicit cipher list so that we do not pick up any strange ciphers when OpenSSL is upgraded. The actual version requested in the client hello depends on additional restrictions in the OpenSSL library. If possible, you should follow the example code and use the default list of trusted root certificate authorities provided by the system because you would have to maintain your own set otherwise, which can be cumbersome.

Example 37. OpenSSL client context creation

// Configure a client connection context.  Send a hendshake for the
// highest supported TLS version, and disable compression.
const SSL_METHOD *const req_method = SSLv23_client_method();
SSL_CTX *const ctx = SSL_CTX_new(req_method);
if (ctx == NULL) {

// Adjust the ciphers list based on a whitelist.  First enable all
// ciphers of at least medium strength, to get the list which is
// compiled into OpenSSL.
if (SSL_CTX_set_cipher_list(ctx, "HIGH:MEDIUM") != 1) {
  // Create a dummy SSL session to obtain the cipher list.
  SSL *ssl = SSL_new(ctx);
  if (ssl == NULL) {
  STACK_OF(SSL_CIPHER) *active_ciphers = SSL_get_ciphers(ssl);
  if (active_ciphers == NULL) {
  // Whitelist of candidate ciphers.
  static const char *const candidates[] =  {
    "AES128-GCM-SHA256", "AES128-SHA256", "AES256-SHA256", // strong ciphers
    "AES128-SHA", "AES256-SHA", // strong ciphers, also in older versions
    "RC4-SHA", "RC4-MD5", // backwards compatibility, supposed to be weak
    "DES-CBC3-SHA", "DES-CBC3-MD5", // more backwards compatibility
  // Actually selected ciphers.
  char ciphers[300];
  ciphers[0] = '\0';
  for (const char *const *c = candidates; *c; ++c) {
    for (int i = 0; i < sk_SSL_CIPHER_num(active_ciphers); ++i) {
	if (strcmp(SSL_CIPHER_get_name(sk_SSL_CIPHER_value(active_ciphers, i)),
		   *c) == 0) {
	  if (*ciphers) {
	    strcat(ciphers, ":");
	  strcat(ciphers, *c);
  // Apply final cipher list.
  if (SSL_CTX_set_cipher_list(ctx, ciphers) != 1) {

// Load the set of trusted root certificates.
if (!SSL_CTX_set_default_verify_paths(ctx)) {

A single context object can be used to create multiple connection objects. It is safe to use the same SSL_CTX object for creating connections concurrently from multiple threads, provided that the SSL_CTX object is not modified (e.g., callbacks must not be changed).

After creating the TCP socket and disabling the Nagle algorithm (per Deactivating the TCP Nagle algorithm), the actual connection object needs to be created, as show in OpenSSL client context creation. If the handshake started by SSL_connect fails, the ssl_print_error_and_exit function from Obtaining OpenSSL error codes is called.

The certificate_validity_override function provides an opportunity to override the validity of the certificate in case the OpenSSL check fails. If such functionality is not required, the call can be removed, otherwise, the application developer has to implement it.

The host name passed to the functions SSL_set_tlsext_host_name and X509_check_host must be the name that was passed to getaddrinfo or a similar name resolution function. No host name canonicalization must be performed. The X509_check_host function used in the final step for host name matching is currently only implemented in OpenSSL 1.1, which is not released yet. In case host name matching fails, the function certificate_host_name_override is called. This function should check user-specific certificate store, to allow a connection even if the host name does not match the certificate. This function has to be provided by the application developer. Note that the override must be keyed by both the certificate and the host name.

Example 38. Creating a client connection using OpenSSL

// Create the connection object.
SSL *ssl = SSL_new(ctx);
if (ssl == NULL) {
SSL_set_fd(ssl, sockfd);

// Enable the ServerNameIndication extension
if (!SSL_set_tlsext_host_name(ssl, host)) {

// Perform the TLS handshake with the server.
ret = SSL_connect(ssl);
if (ret != 1) {
  // Error status can be 0 or negative.
  ssl_print_error_and_exit(ssl, "SSL_connect", ret);

// Obtain the server certificate.
X509 *peercert = SSL_get_peer_certificate(ssl);
if (peercert == NULL) {
  fprintf(stderr, "peer certificate missing");

// Check the certificate verification result.  Allow an explicit
// certificate validation override in case verification fails.
int verifystatus = SSL_get_verify_result(ssl);
if (verifystatus != X509_V_OK && !certificate_validity_override(peercert)) {
  fprintf(stderr, "SSL_connect: verify result: %s\n",

// Check if the server certificate matches the host name used to
// establish the connection.
// FIXME: Currently needs OpenSSL 1.1.
if (X509_check_host(peercert, (const unsigned char *)host, strlen(host),
		      0) != 1
    && !certificate_host_name_override(peercert, host)) {
  fprintf(stderr, "SSL certificate does not match host name\n");


The connection object can be used for sending and receiving data, as in Using an OpenSSL connection to send and receive data. It is also possible to create a BIO object and use the SSL object as the underlying transport, using BIO_set_ssl.

Example 39. Using an OpenSSL connection to send and receive data

const char *const req = "GET / HTTP/1.0\r\n\r\n";
if (SSL_write(ssl, req, strlen(req)) < 0) {
  ssl_print_error_and_exit(ssl, "SSL_write", ret);
char buf[4096];
ret = SSL_read(ssl, buf, sizeof(buf));
if (ret < 0) {
  ssl_print_error_and_exit(ssl, "SSL_read", ret);

When it is time to close the connection, the SSL_shutdown function needs to be called twice for an orderly, synchronous connection termination (Closing an OpenSSL connection in an orderly fashion). This exchanges close_notify alerts with the server. The additional logic is required to deal with an unexpected close_notify from the server. Note that is necessary to explicitly close the underlying socket after the connection object has been freed.

Example 40. Closing an OpenSSL connection in an orderly fashion

// Send the close_notify alert.
ret = SSL_shutdown(ssl);
switch (ret) {
case 1:
  // A close_notify alert has already been received.
case 0:
  // Wait for the close_notify alert from the peer.
  ret = SSL_shutdown(ssl);
  switch (ret) {
  case 0:
    fprintf(stderr, "info: second SSL_shutdown returned zero\n");
  case 1:
    ssl_print_error_and_exit(ssl, "SSL_shutdown 2", ret);
  ssl_print_error_and_exit(ssl, "SSL_shutdown 1", ret);

Closing an OpenSSL connection in an orderly fashion shows how to deallocate the context object when it is no longer needed because no further TLS connections will be established.

Example 41. Closing an OpenSSL connection in an orderly fashion

Implementation TLS Clients With GnuTLS

This section describes how to implement a TLS client with full certificate validation (but without certificate revocation checking). Note that the error handling in is only exploratory and needs to be replaced before production use.

Before setting up TLS connections, a credentials objects has to be allocated and initialized with the set of trusted root CAs (Initializing a GnuTLS credentials structure).

Example 42. Initializing a GnuTLS credentials structure

// Load the trusted CA certificates.
gnutls_certificate_credentials_t cred = NULL;
int ret = gnutls_certificate_allocate_credentials (&cred);
if (ret != GNUTLS_E_SUCCESS) {
  fprintf(stderr, "error: gnutls_certificate_allocate_credentials: %s\n",

ret = gnutls_certificate_set_x509_system_trust(cred);
if (ret == 0) {
  fprintf(stderr, "error: no certificates found in system trust store\n");
if (ret < 0) {
  fprintf(stderr, "error: gnutls_certificate_set_x509_system_trust: %s\n",

After the last TLS connection has been closed, this credentials object should be freed:


During its lifetime, the credentials object can be used to initialize TLS session objects from multiple threads, provided that it is not changed.

Once the TCP connection has been established, the Nagle algorithm should be disabled (see Deactivating the TCP Nagle algorithm). After that, the socket can be associated with a new GnuTLS session object. The previously allocated credentials object provides the set of root CAs. Then the TLS handshake must be initiated. This is shown in Establishing a TLS client connection using GnuTLS.

Example 43. Establishing a TLS client connection using GnuTLS

// Create the session object.
gnutls_session_t session;
ret = gnutls_init(&session, GNUTLS_CLIENT);
if (ret != GNUTLS_E_SUCCESS) {
  fprintf(stderr, "error: gnutls_init: %s\n",

// Configure the cipher preferences.
const char *errptr = NULL;
ret = gnutls_priority_set_direct(session, "NORMAL", &errptr);
if (ret != GNUTLS_E_SUCCESS) {
  fprintf(stderr, "error: gnutls_priority_set_direct: %s\n"
	    "error: at: \"%s\"\n", gnutls_strerror(ret), errptr);

// Install the trusted certificates.
ret = gnutls_credentials_set(session, GNUTLS_CRD_CERTIFICATE, cred);
if (ret != GNUTLS_E_SUCCESS) {
  fprintf(stderr, "error: gnutls_credentials_set: %s\n",

// Associate the socket with the session object and set the server
// name.
gnutls_transport_set_int(session, sockfd);
ret = gnutls_server_name_set(session, GNUTLS_NAME_DNS,
			       host, strlen(host));
if (ret != GNUTLS_E_SUCCESS) {
  fprintf(stderr, "error: gnutls_server_name_set: %s\n",

// Establish the session.
ret = gnutls_handshake(session);
if (ret != GNUTLS_E_SUCCESS) {
  fprintf(stderr, "error: gnutls_handshake: %s\n",

After the handshake has been completed, the server certificate needs to be verified against the server’s hostname (Verifying a server certificate using GnuTLS). In the example, the user-defined certificate_validity_override function is called if the verification fails, so that a separate, user-specific trust store can be checked. This function call can be omitted if the functionality is not needed.

Example 44. Verifying a server certificate using GnuTLS

// Obtain the server certificate chain.  The server certificate
// itself is stored in the first element of the array.
unsigned certslen = 0;
const gnutls_datum_t *const certs =
  gnutls_certificate_get_peers(session, &certslen);
if (certs == NULL || certslen == 0) {
  fprintf(stderr, "error: could not obtain peer certificate\n");

// Validate the certificate chain.
unsigned status = (unsigned)-1;
ret = gnutls_certificate_verify_peers3(session, host, &status);
if (ret != GNUTLS_E_SUCCESS) {
  fprintf(stderr, "error: gnutls_certificate_verify_peers3: %s\n",
if (status != 0 && !certificate_validity_override(certs[0])) {
  gnutls_datum_t msg;
  int type = gnutls_certificate_type_get (session);
  ret = gnutls_certificate_verification_status_print(status, type, &out, 0);
  ret = -1;
  if (ret == 0) {
    fprintf(stderr, "error: %s\n", msg.data);
  } else {
    fprintf(stderr, "error: certificate validation failed with code 0x%x\n",

An established TLS session can be used for sending and receiving data, as in Using a GnuTLS session.

Example 45. Using a GnuTLS session

char buf[4096];
snprintf(buf, sizeof(buf), "GET / HTTP/1.0\r\nHost: %s\r\n\r\n", host);
ret = gnutls_record_send(session, buf, strlen(buf));
if (ret < 0) {
  fprintf(stderr, "error: gnutls_record_send: %s\n", gnutls_strerror(ret));
ret = gnutls_record_recv(session, buf, sizeof(buf));
if (ret < 0) {
  fprintf(stderr, "error: gnutls_record_recv: %s\n", gnutls_strerror(ret));

In order to shut down a connection in an orderly manner, you should call the gnutls_bye function. Finally, the session object can be deallocated using gnutls_deinit (see Closing a GnuTLS session in an orderly fashion).

Example 46. Closing a GnuTLS session in an orderly fashion

// Initiate an orderly connection shutdown.
ret = gnutls_bye(session, GNUTLS_SHUT_RDWR);
if (ret < 0) {
  fprintf(stderr, "error: gnutls_bye: %s\n", gnutls_strerror(ret));
// Free the session object.
Implementing TLS Clients With OpenJDK

The examples below use the following cryptographic-related classes:

import java.security.NoSuchAlgorithmException;
import java.security.NoSuchProviderException;
import java.security.cert.CertificateEncodingException;
import java.security.cert.CertificateException;
import java.security.cert.X509Certificate;
import javax.net.ssl.SSLContext;
import javax.net.ssl.SSLParameters;
import javax.net.ssl.SSLSocket;
import javax.net.ssl.TrustManager;
import javax.net.ssl.X509TrustManager;

import sun.security.util.HostnameChecker;

If compatibility with OpenJDK 6 is required, it is necessary to use the internal class sun.security.util.HostnameChecker. (The public OpenJDK API does not provide any support for dissecting the subject distinguished name of an X.509 certificate, so a custom-written DER parser is needed—or we have to use an internal class, which we do below.) In OpenJDK 7, the setEndpointIdentificationAlgorithm method was added to the javax.net.ssl.SSLParameters class, providing an official way to implement host name checking.

TLS connections are established using an SSLContext instance. With a properly configured OpenJDK installation, the SunJSSE provider uses the system-wide set of trusted root certificate authorities, so no further configuration is necessary. For backwards compatibility with OpenJDK6, the TLSv1 provider has to be supported as a fall-back option. This is shown in Setting up an SSLContext for OpenJDK TLS clients.

Example 47. Setting up an SSLContext for OpenJDK TLS clients

// Create the context.  Specify the SunJSSE provider to avoid
// picking up third-party providers.  Try the TLS 1.2 provider
// first, then fall back to TLS 1.0.
SSLContext ctx;
try {
    ctx = SSLContext.getInstance("TLSv1.2", "SunJSSE");
} catch (NoSuchAlgorithmException e) {
    try {
        ctx = SSLContext.getInstance("TLSv1", "SunJSSE");
    } catch (NoSuchAlgorithmException e1) {
        // The TLS 1.0 provider should always be available.
        throw new AssertionError(e1);
    } catch (NoSuchProviderException e1) {
        throw new AssertionError(e1);
} catch (NoSuchProviderException e) {
    // The SunJSSE provider should always be available.
    throw new AssertionError(e);
ctx.init(null, null, null);

In addition to the context, a TLS parameter object will be needed which adjusts the cipher suites and protocols (Setting up SSLParameters for TLS use with OpenJDK). Like the context, these parameters can be reused for multiple TLS connections.

Example 48. Setting up SSLParameters for TLS use with OpenJDK

// Prepare TLS parameters.  These have to applied to every TLS
// socket before the handshake is triggered.
SSLParameters params = ctx.getDefaultSSLParameters();
// Do not send an SSL-2.0-compatible Client Hello.
ArrayList<String> protocols = new ArrayList<String>(
params.setProtocols(protocols.toArray(new String[protocols.size()]));
// Adjust the supported ciphers.
ArrayList<String> ciphers = new ArrayList<String>(
params.setCipherSuites(ciphers.toArray(new String[ciphers.size()]));

As initialized above, the parameter object does not yet require host name checking. This has to be enabled separately, and this is only supported by OpenJDK 7 and later:


All application protocols can use the "HTTPS" algorithm. (The algorithms have minor differences with regard to wildcard handling, which should not matter in practice.)

Establishing a TLS connection with OpenJDK shows how to establish the connection. Before the handshake is initialized, the protocol and cipher configuration has to be performed, by applying the parameter object params. (After this point, changes to params will not affect this TLS socket.) As mentioned initially, host name checking requires using an internal API on OpenJDK 6.

Example 49. Establishing a TLS connection with OpenJDK

// Create the socket and connect it at the TCP layer.
SSLSocket socket = (SSLSocket) ctx.getSocketFactory()
    .createSocket(host, port);

// Disable the Nagle algorithm.

// Adjust ciphers and protocols.

// Perform the handshake.

// Validate the host name.  The match() method throws
// CertificateException on failure.
X509Certificate peer = (X509Certificate)
// This is the only way to perform host name checking on OpenJDK 6.
    host, peer);

Starting with OpenJDK 7, the last lines can be omitted, provided that host name verification has been enabled by calling the setEndpointIdentificationAlgorithm method on the params object (before it was applied to the socket).

The TLS socket can be used as a regular socket, as shown in Using a TLS client socket in OpenJDK.

Example 50. Using a TLS client socket in OpenJDK

socket.getOutputStream().write("GET / HTTP/1.0\r\n\r\n"
byte[] buffer = new byte[4096];
int count = socket.getInputStream().read(buffer);
System.out.write(buffer, 0, count);
Overriding server certificate validation with OpenJDK 6

Overriding certificate validation requires a custom trust manager. With OpenJDK 6, the trust manager lacks information about the TLS session, and to which server the connection is made. Certificate overrides have to be tied to specific servers (host names). Consequently, different TrustManager and SSLContext objects have to be used for different servers.

In the trust manager shown in A customer trust manager for OpenJDK TLS clients, the server certificate is identified by its SHA-256 hash.

Example 51. A customer trust manager for OpenJDK TLS clients

public class MyTrustManager implements X509TrustManager {
    private final byte[] certHash;

    public MyTrustManager(byte[] certHash) throws Exception {
        this.certHash = certHash;

    public void checkClientTrusted(X509Certificate[] chain, String authType)
            throws CertificateException {
        throw new UnsupportedOperationException();

    public void checkServerTrusted(X509Certificate[] chain,
            String authType) throws CertificateException {
        byte[] digest = getCertificateDigest(chain[0]);
        String digestHex = formatHex(digest);

        if (Arrays.equals(digest, certHash)) {
            System.err.println("info: accepting certificate: " + digestHex);
        } else {
            throw new CertificateException("certificate rejected: "  +

    public X509Certificate[] getAcceptedIssuers() {
        return new X509Certificate[0];

This trust manager has to be passed to the init method of the SSLContext object, as show in Using a custom TLS trust manager with OpenJDK.

Example 52. Using a custom TLS trust manager with OpenJDK

SSLContext ctx;
try {
    ctx = SSLContext.getInstance("TLSv1.2", "SunJSSE");
} catch (NoSuchAlgorithmException e) {
    try {
        ctx = SSLContext.getInstance("TLSv1", "SunJSSE");
    } catch (NoSuchAlgorithmException e1) {
        throw new AssertionError(e1);
    } catch (NoSuchProviderException e1) {
        throw new AssertionError(e1);
} catch (NoSuchProviderException e) {
    throw new AssertionError(e);
MyTrustManager tm = new MyTrustManager(certHash);
ctx.init(null, new TrustManager[] {tm}, null);

When certificate overrides are in place, host name verification should not be performed because there is no security requirement that the host name in the certificate matches the host name used to establish the connection (and it often will not). However, without host name verification, it is not possible to perform transparent fallback to certification validation using the system certificate store.

The approach described above works with OpenJDK 6 and later versions. Starting with OpenJDK 7, it is possible to use a custom subclass of the javax.net.ssl.X509ExtendedTrustManager class. The OpenJDK TLS implementation will call the new methods, passing along TLS session information. This can be used to implement certificate overrides as a fallback (if certificate or host name verification fails), and a trust manager object can be used for multiple servers because the server address is available to the trust manager.

Implementing TLS Clients With NSS

The following code shows how to implement a simple TLS client using NSS. These instructions apply to NSS version 3.14 and later. Versions before 3.14 need different initialization code.

Keep in mind that the error handling needs to be improved before the code can be used in production.

Using NSS needs several header files, as shown in Include files for NSS.

Example 53. Include files for NSS

// NSPR include files
#include <prerror.h>
#include <prinit.h>

// NSS include files
#include <nss.h>
#include <pk11pub.h>
#include <secmod.h>
#include <ssl.h>
#include <sslproto.h>

// Private API, no other way to turn a POSIX file descriptor into an
// NSPR handle.
NSPR_API(PRFileDesc*) PR_ImportTCPSocket(int);

Initializing the NSS library is shown in Initializing the NSS library. This initialization procedure overrides global state. We only call NSS_SetDomesticPolicy if there are no strong ciphers available, assuming that it has already been called otherwise. This avoids overriding the process-wide cipher suite policy unnecessarily.

The simplest way to configured the trusted root certificates involves loading the libnssckbi.so NSS module with a call to the SECMOD_LoadUserModule function. The root certificates are compiled into this module. (The PEM module for NSS, libnsspem.so, offers a way to load trusted CA certificates from a file.)

Example 54. Initializing the NSS library

NSSInitContext *const ctx =
  NSS_InitContext("sql:/etc/pki/nssdb", "", "", "", NULL,
if (ctx == NULL) {
  const PRErrorCode err = PR_GetError();
  fprintf(stderr, "error: NSPR error code %d: %s\n",
	    err, PR_ErrorToName(err));

// Ciphers to enable.
static const PRUint16 good_ciphers[] = {

// Check if the current policy allows any strong ciphers.  If it
// doesn't, set the cipher suite policy.  This is not thread-safe
// and has global impact.  Consequently, we only do it if absolutely
// necessary.
int found_good_cipher = 0;
for (const PRUint16 *p = good_ciphers; *p != SSL_NULL_WITH_NULL_NULL;
     ++p) {
  PRInt32 policy;
  if (SSL_CipherPolicyGet(*p, &policy) != SECSuccess) {
    const PRErrorCode err = PR_GetError();
    fprintf(stderr, "error: policy for cipher %u: error %d: %s\n",
	      (unsigned)*p, err, PR_ErrorToName(err));
  if (policy == SSL_ALLOWED) {
    fprintf(stderr, "info: found cipher %x\n", (unsigned)*p);
    found_good_cipher = 1;
if (!found_good_cipher) {
  if (NSS_SetDomesticPolicy() != SECSuccess) {
    const PRErrorCode err = PR_GetError();
    fprintf(stderr, "error: NSS_SetDomesticPolicy: error %d: %s\n",
	      err, PR_ErrorToName(err));

// Initialize the trusted certificate store.
char module_name[] = "library=libnssckbi.so name=\"Root Certs\"";
SECMODModule *module = SECMOD_LoadUserModule(module_name, NULL, PR_FALSE);
if (module == NULL || !module->loaded) {
  const PRErrorCode err = PR_GetError();
  fprintf(stderr, "error: NSPR error code %d: %s\n",
	    err, PR_ErrorToName(err));

Some of the effects of the initialization can be reverted with the following function calls:


After NSS has been initialized, the TLS connection can be created (Creating a TLS connection with NSS). The internal PR_ImportTCPSocket function is used to turn the POSIX file descriptor sockfd into an NSPR file descriptor. (This function is de-facto part of the NSS public ABI, so it will not go away.) Creating the TLS-capable file descriptor requires a model descriptor, which is configured with the desired set of protocols. The model descriptor is not needed anymore after TLS support has been activated for the existing connection descriptor.

The call to SSL_BadCertHook can be omitted if no mechanism to override certificate verification is needed. The bad_certificate function must check both the host name specified for the connection and the certificate before granting the override.

Triggering the actual handshake requires three function calls, SSL_ResetHandshake, SSL_SetURL, and SSL_ForceHandshake. (If SSL_ResetHandshake is omitted, SSL_ForceHandshake will succeed, but the data will not be encrypted.) During the handshake, the certificate is verified and matched against the host name.

Example 55. Creating a TLS connection with NSS

// Wrap the POSIX file descriptor.  This is an internal NSPR
// function, but it is very unlikely to change.
PRFileDesc* nspr = PR_ImportTCPSocket(sockfd);
sockfd = -1; // Has been taken over by NSPR.

// Add the SSL layer.
  PRFileDesc *model = PR_NewTCPSocket();
  PRFileDesc *newfd = SSL_ImportFD(NULL, model);
  if (newfd == NULL) {
    const PRErrorCode err = PR_GetError();
    fprintf(stderr, "error: NSPR error code %d: %s\n",
	      err, PR_ErrorToName(err));
  model = newfd;
  newfd = NULL;
  if (SSL_OptionSet(model, SSL_ENABLE_SSL2, PR_FALSE) != SECSuccess) {
    const PRErrorCode err = PR_GetError();
    fprintf(stderr, "error: set SSL_ENABLE_SSL2 error %d: %s\n",
	      err, PR_ErrorToName(err));
  if (SSL_OptionSet(model, SSL_V2_COMPATIBLE_HELLO, PR_FALSE) != SECSuccess) {
    const PRErrorCode err = PR_GetError();
    fprintf(stderr, "error: set SSL_V2_COMPATIBLE_HELLO error %d: %s\n",
	      err, PR_ErrorToName(err));
  if (SSL_OptionSet(model, SSL_ENABLE_DEFLATE, PR_FALSE) != SECSuccess) {
    const PRErrorCode err = PR_GetError();
    fprintf(stderr, "error: set SSL_ENABLE_DEFLATE error %d: %s\n",
	      err, PR_ErrorToName(err));

  // Allow overriding invalid certificate.
  if (SSL_BadCertHook(model, bad_certificate, (char *)host) != SECSuccess) {
    const PRErrorCode err = PR_GetError();
    fprintf(stderr, "error: SSL_BadCertHook error %d: %s\n",
	      err, PR_ErrorToName(err));

  newfd = SSL_ImportFD(model, nspr);
  if (newfd == NULL) {
    const PRErrorCode err = PR_GetError();
    fprintf(stderr, "error: SSL_ImportFD error %d: %s\n",
	      err, PR_ErrorToName(err));
  nspr = newfd;

// Perform the handshake.
if (SSL_ResetHandshake(nspr, PR_FALSE) != SECSuccess) {
  const PRErrorCode err = PR_GetError();
  fprintf(stderr, "error: SSL_ResetHandshake error %d: %s\n",
	    err, PR_ErrorToName(err));
if (SSL_SetURL(nspr, host) != SECSuccess) {
  const PRErrorCode err = PR_GetError();
  fprintf(stderr, "error: SSL_SetURL error %d: %s\n",
	    err, PR_ErrorToName(err));
if (SSL_ForceHandshake(nspr) != SECSuccess) {
  const PRErrorCode err = PR_GetError();
  fprintf(stderr, "error: SSL_ForceHandshake error %d: %s\n",
	    err, PR_ErrorToName(err));

After the connection has been established, Using NSS for sending and receiving data shows how to use the NSPR descriptor to communicate with the server.

Example 56. Using NSS for sending and receiving data

char buf[4096];
snprintf(buf, sizeof(buf), "GET / HTTP/1.0\r\nHost: %s\r\n\r\n", host);
PRInt32 ret = PR_Write(nspr, buf, strlen(buf));
if (ret < 0) {
  const PRErrorCode err = PR_GetError();
  fprintf(stderr, "error: PR_Write error %d: %s\n",
	    err, PR_ErrorToName(err));
ret = PR_Read(nspr, buf, sizeof(buf));
if (ret < 0) {
  const PRErrorCode err = PR_GetError();
  fprintf(stderr, "error: PR_Read error %d: %s\n",
	    err, PR_ErrorToName(err));

Closing NSS client connections shows how to close the connection.

Example 57. Closing NSS client connections

// Send close_notify alert.
if (PR_Shutdown(nspr, PR_SHUTDOWN_BOTH) != PR_SUCCESS) {
  const PRErrorCode err = PR_GetError();
  fprintf(stderr, "error: PR_Read error %d: %s\n",
	    err, PR_ErrorToName(err));
// Closes the underlying POSIX file descriptor, too.
Implementing TLS Clients With Python

The Python distribution provides a TLS implementation in the ssl module (actually a wrapper around OpenSSL). The exported interface is somewhat restricted, so that the client code shown below does not fully implement the recommendations in OpenSSL Pitfalls.


Currently, most Python function which accept https:// URLs or otherwise implement HTTPS support do not perform certificate validation at all. (For example, this is true for the httplib and xmlrpclib modules.) If you use HTTPS, you should not use the built-in HTTP clients. The Curl class in the curl module, as provided by the python-pycurl package implements proper certificate validation.

The ssl module currently does not perform host name checking on the server certificate. Implementing TLS host name checking Python (without wildcard support) shows how to implement certificate matching, using the parsed certificate returned by getpeercert.

Example 58. Implementing TLS host name checking Python (without wildcard support)

def check_host_name(peercert, name):
    """Simple certificate/host name checker.  Returns True if the
    certificate matches, False otherwise.  Does not support
    # Check that the peer has supplied a certificate.
    # None/{} is not acceptable.
    if not peercert:
        return False
    if peercert.has_key("subjectAltName"):
        for typ, val in peercert["subjectAltName"]:
            if typ == "DNS" and val == name:
                return True
        # Only check the subject DN if there is no subject alternative
        # name.
        cn = None
        for attr, val in peercert["subject"]:
            # Use most-specific (last) commonName attribute.
            if attr == "commonName":
                cn = val
        if cn is not None:
            return cn == name
    return False

To turn a regular, connected TCP socket into a TLS-enabled socket, use the ssl.wrap_socket function. The function call in Establishing a TLS client connection with Python provides additional arguments to override questionable defaults in OpenSSL and in the Python module.

  • ciphers="HIGH:-aNULL:-eNULL:-PSK:RC4-SHA:RC4-MD5" selects relatively strong cipher suites with certificate-based authentication. (The call to check_host_name function provides additional protection against anonymous cipher suites.)

  • ssl_version=ssl.PROTOCOL_TLSv1 disables SSL 2.0 support. By default, the ssl module sends an SSL 2.0 client hello, which is rejected by some servers. Ideally, we would request OpenSSL to negotiated the most recent TLS version supported by the server and the client, but the Python module does not allow this.

  • cert_reqs=ssl.CERT_REQUIRED turns on certificate validation.

  • ca_certs='/etc/ssl/certs/ca-bundle.crt' initializes the certificate store with a set of trusted root CAs. Unfortunately, it is necessary to hard-code this path into applications because the default path in OpenSSL is not available through the Python ssl module.

The ssl module (and OpenSSL) perform certificate validation, but the certificate must be compared manually against the host name, by calling the check_host_name defined above.

Example 59. Establishing a TLS client connection with Python

sock = ssl.wrap_socket(sock,
# getpeercert() triggers the handshake as a side effect.
if not check_host_name(sock.getpeercert(), host):
    raise IOError("peer certificate does not match host name")

After the connection has been established, the TLS socket can be used like a regular socket:

sock.write("GET / HTTP/1.1\r\nHost: " + host + "\r\n\r\n")
print sock.read()

Closing the TLS socket is straightforward as well:


3.3. Hardware Security Modules and Smart Cards

Hardware Security Modules (HSMs) are specialized hardware intended to protect private keys on server systems. They store internally the private keys (e.g., RSA keys), and provide access to operations with the keys without exposing the keys. That access, is provided using a standardized API, which across Fedora is PKCS#11.

Smart cards are small cards with a micro processor, often combined with a USB reader resembling a USB stick. They are very similar in nature with HSMs as they can also be used to protect private keys and are almost universally accessed via the PKCS#11 API. The main distinguishers from HSMs is their inferior performance and often, the available hardware protection mechanisms.

Typically a smart card or HSM relies on a shared library to provide functionality. This shared library follows the PKCS#11 API and thus is often referred to as a PKCS#11 module. In Fedora the opensc shared module (opensc-pkcs11.so) can be used for the majority of smart cards available in the market. By convention these modules are located at /usr/lib64/pkcs11. They can be used directly, or via a higher level library.

All the major crypto libraries (NSS, GnuTLS and OpenSSL in Fedora) support hardware security modules and smart cards, by providing wrappers over the PKCS#11 API. However, the level of support varies, as well as the ease of use of such modules and its integration to the overall library API.

  • The PKCS#11 API does provide an API to access HSMs or smart cards, but does not provide any method of discovering which HSMs or smart cards are available in the system. In Fedora and modules are registered via p11-kit configuration files, stored at /etc/pkcs11/modules/. For applications using engine_pkcs11 or GnuTLS the registered modules are available without further configuration. Other applications will have to load the p11-kit-proxy.so module.

  • Most crypto libraries support the PKCS#11 URLs scheme to identify objects stored in an HSM, however that support is not yet universal. Some support transparent usage of PKCS#11 objects, e.g., specifying a PKCS#11 object instead of a file, while others require to use specialized APIs for such objects.

  • Objects stored in an HSM or smart card can be protected with a PIN. As such, libraries typically require to set a PIN handling function for accessing private keys, or the PIN can be passed along with a PKCS#11 URL and the pin-value parameter.

  • Obtaining a Hardware Security Module, or including it on a continuous integration testing is not always feasible. For testing purposes smart cards supported by the OpenSC project can be used, as well as software modules like softhsm which provides a tool to setup a software HSM, and a PKCS#11 library.

  • The PKCS#11 API requires applications that use fork to reinitialize the used PKCS#11 modules. This is an uncommon requirement, which has led to several bugs across applications in Fedora which used PKCS#11 directly. To make things more complicated software PKCS#11 module like softhsm do not require this re-initialization leading to applications working against software modules but failing with hardware modules or smart cards. The wrapper PKCS#11 APIs provided by NSS, GnuTLS and engine_pkcs11 (OpenSSL) handle the reinitialization after fork requirement transparently.

3.3.1. OpenSSL HSM Support

OpenSSL does not have native support for PKCS#11. It can provide PKCS#11 support through the OpenSC’s project pkcs11 engine (formerly known as engine_pkcs11). As such software intended to use HSMs, must utilize that engine.

Engine pkcs11 supports loading stored objects via PKCS#11 URLs. If no PKCS#11 module is specified the engine will use the system-wide registered modules via p11-kit-proxy.so.

The following example demonstrates the initialization of the pkcs11 engine and its usage to sign data.

Example 60. Signing data with HSM and OpenSSL


e = ENGINE_by_id("pkcs11");
if (!e) {

if (module_path) {
  fprintf(stderr, "loading: %s\n", module_path);
  if (!ENGINE_ctrl_cmd_string(e, "MODULE_PATH", module_path, 0)) {

if (!ENGINE_init(e)) {

if (key_pass && !ENGINE_ctrl_cmd_string(e, "PIN", key_pass, 0)) {

private_key = ENGINE_load_private_key(e, private_key_name, NULL, NULL);
if (!private_key) {
  fprintf(stderr, "cannot load: %s\n", private_key_name);


digest_algo = EVP_get_digestbyname("sha256");

if (EVP_DigestInit(&ctx, digest_algo) <= 0) {

EVP_SignInit(&ctx, digest_algo);

#define TEST_DATA "test data"
if (EVP_SignUpdate(&ctx, TEST_DATA, sizeof(TEST_DATA) - 1) <= 0) {

n = sizeof(buf);
if (EVP_SignFinal(&ctx, buf, &n, private_key) <= 0) {


3.3.2. GnuTLS HSM Support

GnuTLS supports PKCS#11 natively. Most of the API functions accepting certificate files, can also accept PKCS#11 URLs, thus requiring minor or no modifications to applications in order to support HSMs. In most cases applications must be modified to install a PIN callback function.

The following example demonstrates the initialization of the pkcs11 engine and its usage to sign data.

Example 61. Signing data with HSM and GnuTLS

if (module_path) {
  ret = gnutls_pkcs11_init(GNUTLS_PKCS11_FLAG_MANUAL, NULL);
  if (ret < 0) {
    fprintf(stderr, "error in %d: %s\n", __LINE__, gnutls_strerror(ret));

  ret = gnutls_pkcs11_add_provider(module_path, NULL);
  if (ret < 0) {
    fprintf(stderr, "error in %d: %s\n", __LINE__, gnutls_strerror(ret));

if (key_pass)
  gnutls_pkcs11_set_pin_function(pin_function, key_pass);

ret = gnutls_privkey_init(&private_key);
if (ret < 0) {
  fprintf(stderr, "error in %d: %s\n", __LINE__, gnutls_strerror(ret));

ret = gnutls_privkey_import_url(private_key, private_key_name, 0);
if (ret < 0) {
  fprintf(stderr, "error in %d: %s\n", __LINE__, gnutls_strerror(ret));

ret = gnutls_privkey_sign_data(private_key, GNUTLS_DIG_SHA256, 0,
                               &testdata, &signature);
if (ret < 0) {
  fprintf(stderr, "error in %d: %s\n", __LINE__, gnutls_strerror(ret));


The PIN callback function can be either set globally as in the example above or locally by utilizing functions such as gnutls_privkey_set_pin_function. An example PIN callback function is shown below.

Example 62. An example PIN callback with GNUTLS

int pin_function(void *userdata, int attempt, const char *token_url,
                 const char *token_label, unsigned flags, char *pin, size_t pin_max)
  if (flags & GNUTLS_PIN_FINAL_TRY)
    printf("This is the final try before locking!\n");
  if (flags & GNUTLS_PIN_COUNT_LOW)
    printf("Only few tries left before locking!\n");
  if (flags & GNUTLS_PIN_WRONG)
    printf("Wrong PIN has been provided in the previous attempt\n");

  /* userdata is the second value passed to gnutls_pkcs11_set_pin_function()
   * in this example we passed the PIN as a null terminated value.
  snprintf(pin, pin_max, "%s", (char*)userdata);
  return 0;

3.3.3. NSS HSM Support

NSS supports PKCS#11 natively. In fact all NSS crypto operations, including built-in operations, go through PKCS #11 modules. NSS provides its own software PKCS #11 module called softoken. NSS automatically loads any PKCS #11 module specified in its module database, which can be manipulated with the modutil command. NSS uses the PKCS #11 module that contains the requested keys to do the crypto operations. As long as the application opens an NSS database and properly sets a pin callback. If it runs with native NSS, it should be able to use HSMs that provide PKCS #11 modules. Modules can also be loaded programatically, though this is less common.

The following example demonstrates a typical NSS application for signing.

Example 63. Signing data with HSM and NSS

SECStatus rv;
CERTCertificate *cert = NULL;
SECKEYPrivateKey *pvtkey = NULL;
SECItem signature = { siBuffer, NULL, 0 };
SECOidTag algTag;
int r = 1;
unsigned char buf[] = "test data to sign";
const char *cert_name;
unsigned i;

if (argc < 3) {
  fprintf(stderr, "usage: %s [cert name] [PIN]\n\n", argv[0]);

cert_name = argv[1];
pin = argv[2];

rv = NSS_Init(".");
if (rv != SECSuccess) {
  fprintf(stderr, "NSS initialization failed (err %d)\n", PR_GetError());
  goto cleanup;

cert = PK11_FindCertFromNickname(cert_name, NULL);
if (cert == NULL) {
  fprintf(stderr, "Couldn't find cert %s in NSS db (err %d: %s)\n",
	    cert_name, PR_GetError(), PORT_ErrorToString(PR_GetError()));
  goto cleanup;

fprintf(stderr, "Buffer being signed = \n%s\n", buf);

pvtkey = PK11_FindKeyByAnyCert(cert, NULL);
if (pvtkey == NULL) {
  fprintf(stderr, "Couldn't find private key for cert %s (err %d: %s)\n",
	    cert_name, PR_GetError(), PORT_ErrorToString(PR_GetError()));
  goto cleanup;

/* get the algtag. Pick the default hash algorithm */
algTag = SEC_GetSignatureAlgorithmOidTag(pvtkey->keyType, SEC_OID_UNKNOWN);

fprintf(stderr, "Signing with alg = %s (%d)\n",
	  SECOID_FindOIDTagDescription(algTag), algTag);

rv = SEC_SignData(&signature, buf, sizeof(buf)-1, pvtkey, algTag);
if (rv != SECSuccess) {
  fprintf(stderr, "sign with Private Key failed (err %d: %s)\n",
	    PR_GetError(), PORT_ErrorToString(PR_GetError()));
  goto cleanup;

To use the example above with an HSM or smart card you will need to do the following.

# add your HSM or token library to an NSS database (in the sample code the database is
# located in the current directory'.')
$ modutil -add "My HSM" -libfile ${path_to_pkcs11_file} -dbdir .
# Find the token name on your HSM
$ modutil -list -dbdir .
# find the cert on your token
$ certutil -L -h ${token_name} -d .
# pass the cert to your signing program
$ NSS_Sign_Example "${token_name}:${cert_name}"
Example 64. An example PIN callback with NSS

char *passwdcb(PK11SlotInfo * slot, PRBool retry, void *arg)
  if (!isatty(STDIN_FILENO) && retry) {
    /* we're just reading from a file, and the value is known to be wrong,
     * don't keep bounding the token with the wrong password. */
    return NULL;

  if (retry) {
    printf("Warning: Wrong PIN has been provided in the previous attempt\n");
    if (PK11_IsHW(slot)) {
	  ("  NOTE: multiple pin failures could result in locking your device\n");

  if (pin == NULL)
    return pin;
    return strdup(pin);

4. Revision History


Fri Dec 1 2017, Mirek Jahoda (mjahoda@redhat.com)

  • First release in AsciiDoc

  • Many updates in the crypto-related sections

  • Grammar and typography fixes


Mon Oct 13 2014, Florian Weimer (fweimer@redhat.com)

  • Go: Mention default value handling in deserialization

  • Shell: New chapter


Wed Jul 16 2014, Florian Weimer (fweimer@redhat.com)

  • C: Corrected the strncat example

  • C: Mention mixed signed/unsigned comparisons

  • C: Unsigned overflow checking example

  • C++: operator new[] has been fixed in GCC

  • C++: Additional material on std::string, iterators

  • OpenSSL: Mention openssl genrsa entropy issue

  • Packaging: X.509 key generation

  • Go, Vala: Add short chapters

  • Serialization: Notes on fragmentation and reassembly


Tue Aug 27 2013, Eric Christensen (sparks@redhat.com)

  • Add a chapter which covers some Java topics.

  • Deserialization: Warn about Java’s java.beans.XMLDecoder.

  • C: Correct the advice on array allocation (bug 995595).

  • C: Add material on global variables.


Thu May 09 2013, Eric Christensen (sparks@redhat.com)

  • Added more C and C++ examples.

  • TLS Client NSS: Rely on NSS 3.14 cipher suite defaults.


Thu Mar 7 2013, Eric Christensen (sparks@redhat.com)

  • Initial publication.