Number of Outputs
The number of functions that a function calls.
Interpretation
The attribute that number of outputs is expected to quantify is testability. A file that is difficult to exhaustively test is likely to have a high value for the number of outputs metric.
Evidence
Number of Outputs has been empirically-validated to be associated with historical vulnerabilities in software in the following peer-reviewed research studies:
- To Fear or Not to Fear That is the Question: Code Characteristics of a Vulnerable Function with an Existing Exploit [2]
The empirical evidence overwhelming supports the notion that a source code file with high number of outputs is more likely to contain a security vulnerability.
Implications
The security implication(s) of a file having high number of outputs could be one or more of the following:
- Difficulty in exhaustively testing a file may increase the potential for latent vulnerabilities.
Mitigations
The theoretical mitigation to lowering the number of outputs of a file is to have functions in source code files call, or be dependent on, fewer other functions. However, the theoretical mitigation is not practical because modern software is inherently complex requiring functions to call other functions to implement features. Therefore, the risk of latent vulnerabilities in a file with high number of outputs could be mitigated using one or more of the following suggestions:
- Refactor functions in the file to call fewer other functions when possible leveraging common design patterns as appropriate.
- Leverage automated testing to ensure all functions are appropriately tested to a satisfactory level of exhaustiveness.
Implementation
In our implementation of the metric, we use SciTools Understand™ to collect the number of outputs metric from functions. The metric is aggregated at the file level by computing the sum of the number of outputs of all functions in a file.
The source code of the implementation of the metric will be made available on GitHub. If you need to collect the metric from your project, the implementation will also be made available as a container image on Docker Hub.
Languages
The metric implementation is limited to projects written in C/C++, C#, Fortran, Java.
Example(s)
In this section, we present examples of the metric collected from popular open-source software projects.
Chromium
In this subsection, we present examples of the metric collected from the Chromium, the open-source project behind the Google Chrome web browser.
The metric examples presented here were collected at
6b9bf768231f commit to the master branch of the Chromium source code repository.
Summary
Shown in Figure 1.1 is the distribution of the metric collected from source code files in the Chromium project. Shown in Figure 1.2 is the comparison of the distribution of the metric collected from source code files in the Chromium project that were not historically vulnerable and those that were.
Thresholds
The thresholds of the metric in the Chromium project determined using the approach prescribed by Alves et al. [1] is shown in the table below.
| Metric Range | value < 319 | 319 ≤ value < 554 | 554 ≤ value < 1,222 | 1,222 ≤ value |
|---|---|---|---|---|
| Risk Level | Low | Medium | High | Critical |
Risky Files
The thresholds are used to classify source code files into appropriate risk levels. Shown below are the top and bottom three source code files from the Chromium project in each of the three non-trivial risk levels.
| Path | Number of Outputs | Percentile |
|---|---|---|
chrome/browser/ui/views/page_info/page_info_bubble_view_browsertest.cc |
319 | 70.0488 |
third_party/tcmalloc/chromium/src/profiledata.h |
319 | 70.0488 |
components/sync_sessions/session_store_unittest.cc |
319 | 70.0488 | ... |
third_party/sqlite/sqlite-src-3280000/src/where.c |
552 | 79.9817 |
chrome/browser/profiles/profile_manager_unittest.cc |
552 | 79.9817 |
chrome/browser/browsing_data/cookies_tree_model.h |
553 | 79.9912 |
| Path | Number of Outputs | Percentile |
|---|---|---|
third_party/sqlite/patched/src/where.c |
554 | 80.0878 |
chrome/browser/page_load_metrics/observers/from_gws_page_load_metrics_observer_unittest.cc |
557 | 80.1114 |
third_party/abseil-cpp/absl/container/internal/raw_hash_set_test.cc |
558 | 80.1619 | ... |
third_party/sqlite/patched/ext/fts5/fts5_index.c |
1,161 | 89.8892 |
chrome/browser/ui/views/tabs/tab_strip.h |
1,171 | 89.8981 |
chrome/renderer/net/net_error_helper_core_unittest.cc |
1,195 | 89.9609 |
| Path | Number of Outputs | Percentile |
|---|---|---|
chrome/browser/sync/test/integration/two_client_bookmarks_sync_test.cc |
1,222 | 90.0207 |
content/browser/webauth/authenticator_impl_unittest.cc |
1,238 | 90.2554 |
third_party/sqlite/sqlite-src-3280000/ext/fts2/fts2.c |
1,238 | 90.2554 | ... |
third_party/wtl/include/atlwince.h |
8,567 | 96.6272 |
third_party/libxml/src/testapi.c |
11,115 | 97.6838 |
third_party/sqlite/amalgamation/sqlite3.c |
14,612 | 100 |
OpenBSD
In this subsection, we present examples of the metric collected from the UNIX-like operating system developed by the OpenBSD project.
The metric examples presented here were collected at dbdab68da3b
commit to the master branch of the OpenBSD source code repository.
Summary
Shown in Figure 2.1 is the distribution of the metric collected from source code files in the OpenBSD project. Shown in Figure 2.2 is the comparison of the distribution of the metric collected from source code files in the OpenBSD project that were not historically vulnerable and those that were.
Thresholds
The thresholds of the metric in the OpenBSD project determined using the approach prescribed by Alves et al. [1] is shown in the table below.
| Metric Range | value < 415 | 415 ≤ value < 636 | 636 ≤ value < 1,103 | 1,103 ≤ value |
|---|---|---|---|---|
| Risk Level | Low | Medium | High | Critical |
Risky Files
The thresholds are used to classify source code files into appropriate risk levels. Shown below are the top and bottom three source code files from the OpenBSD project in each of the three non-trivial risk levels.
| Path | Number of Outputs | Percentile |
|---|---|---|
gnu/llvm/lib/Transforms/Scalar/LoopRerollPass.cpp |
415 | 70.0482 |
gnu/gcc/gcc/config/h8300/h8300.c |
415 | 70.0482 |
gnu/usr.bin/binutils/gdb/s390-tdep.c |
415 | 70.0482 | ... |
usr.sbin/unbound/util/netevent.c |
632 | 79.9835 |
gnu/gcc/gcc/gcc.c |
632 | 79.9835 |
gnu/llvm/tools/clang/lib/CodeGen/CGExprConstant.cpp |
632 | 79.9835 |
| Path | Number of Outputs | Percentile |
|---|---|---|
sys/dev/pci/drm/i915/intel_sdvo.c |
636 | 80.0126 |
usr.sbin/ldapd/btree.c |
639 | 80.0446 |
gnu/llvm/lib/Analysis/ScalarEvolution.cpp |
641 | 80.2240 | ... |
gnu/llvm/tools/clang/lib/Sema/SemaDecl.cpp |
1,099 | 89.8299 |
sys/dev/softraid.c |
1,102 | 89.9442 |
gnu/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp |
1,102 | 89.9442 |
| Path | Number of Outputs | Percentile |
|---|---|---|
gnu/gcc/gcc/config/s390/s390.c |
1,103 | 90.0289 |
gnu/gcc/gcc/tree-cfg.c |
1,110 | 90.0798 |
gnu/usr.bin/binutils/gdb/breakpoint.c |
1,115 | 90.1510 | ... |
gnu/llvm/tools/clang/lib/AST/ExprConstant.cpp |
3,621 | 99.4902 |
gnu/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp |
5,081 | 99.6455 |
gnu/llvm/lib/Target/X86/X86ISelLowering.cpp |
7,128 | 100 |
Reference(s)
[1] Tiago L. Alves, Christiaan Ypma, and Joost Visser. 2010. Deriving Metric Thresholds From Benchmark Data. In Proceedings of the 26th International Conference on Software Maintenance (ICSM '10). 1-10. https://doi.org/10.1109/ICSM.2010.5609747
[2] Awad Younis, Yashwant Malaiya, Charles Anderson, and Indrajit Ray. 2016. To Fear or Not to Fear That is the Question: Code Characteristics of a Vulnerable Function with an Existing Exploit. In Proceedings of the 6th ACM Conference on Data and Application Security and Privacy (CODASPY '16). New York, NY, USA, 97–104. https://doi.org/10.1145/2857705.2857750