Android’s defense-in-depth technique applies not solely to the Android OS working on the Utility Processor (AP) but additionally the firmware that runs on units. We significantly prioritize hardening the mobile baseband given its distinctive mixture of working in an elevated privilege and parsing untrusted inputs which might be remotely delivered into the system.
This submit covers tips on how to use two high-value sanitizers which might forestall particular courses of vulnerabilities discovered inside the baseband. They’re structure agnostic, appropriate for bare-metal deployment, and needs to be enabled in present C/C++ code bases to mitigate unknown vulnerabilities. Past safety, addressing the problems uncovered by these sanitizers improves code well being and general stability, decreasing assets spent addressing bugs sooner or later.
As we outlined beforehand, safety analysis targeted on the baseband has highlighted a constant lack of exploit mitigations in firmware. Baseband Distant Code Execution (RCE) exploits have their very own categorization in well-known third-party marketplaces with a comparatively low payout. This means baseband bugs might probably be considerable and/or not too complicated to seek out and exploit, and their distinguished inclusion within the market demonstrates that they’re helpful.
Baseband safety and exploitation has been a recurring theme in safety conferences for the final decade. Researchers have additionally made a dent on this space in well-known exploitation contests. Most not too long ago, this space has change into distinguished sufficient that it’s frequent to seek out sensible baseband exploitation trainings in prime safety conferences.
Acknowledging this pattern, mixed with the severity and obvious abundance of those vulnerabilities, final yr we launched updates to the severity pointers of Android’s Vulnerability Rewards Program (VRP). For instance, we take into account vulnerabilities permitting Distant Code Execution (RCE) within the mobile baseband to be of CRITICAL severity.
Widespread courses of vulnerabilities will be mitigated via the usage of sanitizers offered by Clang-based toolchains. These sanitizers insert runtime checks towards frequent courses of vulnerabilities. GCC-based toolchains can also present some stage of assist for these flags as nicely, however is not going to be thought of additional on this submit. We encourage you to examine your toolchain’s documentation.
Two sanitizers included in Undefined Conduct Sanitizer (UBSan) will probably be our focus – Integer Overflow Sanitizer (IntSan) and BoundsSanitizer (BoundSan). These have been broadly deployed in Android userspace for years following a data-driven strategy. These two are nicely fitted to bare-metal environments such because the baseband since they don’t require assist from the OS or particular structure options, and so are typically supported for all Clang targets.
Integer Overflow Sanitizer (IntSan)
IntSan causes signed and unsigned integer overflows to abort execution except the overflow is made express. Whereas unsigned integer overflows are technically outlined conduct, it will probably typically result in unintentional conduct and vulnerabilities – particularly once they’re used to index into arrays.
As each intentional and unintentional overflows are probably current in most code bases, IntSan might require refactoring and annotating the code base to stop intentional or benign overflows from trapping (which we take into account a false constructive for our functions). Overflows which have to be addressed will be uncovered by way of testing (see the Deploying Sanitizers part)
BoundsSanitizer (BoundSan)
BoundSan inserts instrumentation to carry out bounds checks round some array accesses. These checks are solely added if the compiler can’t show at compile time that the entry will probably be protected and if the scale of the array will probably be recognized at runtime, in order that it may be checked towards. Observe that this is not going to cowl all array accesses as the scale of the array will not be recognized at runtime, similar to perform arguments that are arrays.
So long as the code is appropriately written C/C++, BoundSan ought to produce no false positives. Any violations found when first enabling BoundSan is a minimum of a bug, if not a vulnerability. Resolving even these which aren’t exploitable can significantly enhance stability and code high quality.
Modernize your toolchains
Adopting fashionable mitigations additionally means adopting (and sustaining) fashionable toolchains. The advantages of this transcend using sanitizers nevertheless. Sustaining an previous toolchain is just not free and entails hidden alternative prices. Toolchains include bugs that are addressed in subsequent releases. Newer toolchains deliver new efficiency optimizations, useful within the extremely constrained bare-metal setting that basebands function in. Safety points may even exist within the generated code of out-of-date compilers.
Sustaining a contemporary up-to-date toolchain for the baseband entails some prices by way of upkeep, particularly at first if the toolchain is especially previous, however over time the advantages, as outlined above, outweigh the prices.
Each BoundSan and IntSan have a measurable efficiency overhead. Though we had been capable of considerably scale back this overhead up to now (for instance to lower than 1% in media codecs), even very small will increase in CPU load can have a considerable influence in some environments.
Enabling sanitizers over the whole codebase gives probably the most profit, however enabling them in security-critical assault surfaces can function a primary step in an incremental deployment. For instance:
- Features parsing messages delivered over the air in 2G, 3G, 4G, and 5G (particularly capabilities dealing with pre-authentication messages that may be injected with a false/malicious base station)
- Libraries encoding/decoding complicated codecs (e.g. ASN.1, XML, DNS, and so forth…)
- IMS, TCP and IP stacks
- Messaging capabilities (SMS, MMS)
Within the specific case of 2G, the most effective technique is to disable the stack altogether by supporting Android’s “2G toggle”. Nonetheless, 2G continues to be a needed cellular entry expertise in sure elements of the world and a few customers may must have this legacy protocol enabled.
Having a transparent plan for deployment of sanitizers saves a variety of effort and time. We consider the deployment course of as having three phases:
- Detecting (and fixing) violations
- Measuring and decreasing overhead
- Soaking in pre-production
We additionally introduce two modes by which sanitizers needs to be run: diagnostics mode and trapping mode. These will probably be mentioned within the following sections, however briefly: diagnostics mode recovers from violations and gives useful debug data, whereas trapping mode actively mitigates vulnerabilities by trapping execution on violations.
Detecting (and Fixing) Violations
To efficiently ship these sanitizers, any benign integer overflows have to be made express and unintentional out-of-bounds accesses have to be addressed. These must be uncovered via testing. The upper the code protection your checks present, the extra points you possibly can uncover at this stage and the simpler deployment will probably be afterward.
To diagnose violations uncovered in testing, sanitizers can emit calls to runtime handlers with debug data such because the file, line quantity, and values resulting in the violation. Sanitizers can optionally proceed execution after a violation has occurred, permitting a number of violations to be found in a single check run. We confer with utilizing the sanitizers on this means as working them in “diagnostics mode”. Diagnostics mode is just not meant for manufacturing because it gives no safety advantages and provides excessive overhead.
Diagnostics mode for the sanitizers will be set utilizing the next flags:
-fsanitize=signed-integer-overflow,unsigned-integer-overflow,bounds -fsanitize-recover=all
Since Clang doesn’t present a UBSan runtime for bare-metal targets, a runtime will have to be outlined and offered at hyperlink time:
// integer overflow handlers __ubsan_handle_add_overflow(OverflowData *information, ValueHandle lhs, ValueHandle rhs) __ubsan_handle_sub_overflow(OverflowData *information, ValueHandle lhs, ValueHandle rhs) __ubsan_handle_mul_overflow(OverflowData *information, ValueHandle lhs, ValueHandle rhs) __ubsan_handle_divrem_overflow(OverflowData *information, ValueHandle lhs, ValueHandle rhs) __ubsan_handle_negate_overflow(OverflowData *information, ValueHandle old_val) // boundsan handler __ubsan_handle_out_of_bounds_overflow(OverflowData *information, ValueHandle old_val)
For instance, see the default Clang implementation; the Linux Kernels implementation can also be illustrative.
With the runtime outlined, allow the sanitizer over the whole baseband codebase and run all out there checks to uncover and tackle any violations. Vulnerabilities needs to be patched, and overflows ought to both be refactored or made express via the usage of checked arithmetic builtins which don’t set off sanitizer violations. Sure capabilities that are anticipated to have intentional overflows (similar to cryptographic capabilities) will be preemptively excluded from sanitization (see subsequent part).
Apart from uncovering safety vulnerabilities, this stage is extremely efficient at uncovering code high quality and stability bugs that would lead to instability on person units.
As soon as violations have been addressed and checks are now not uncovering new violations, the subsequent stage can start.
Measuring and Lowering Overhead
As soon as shallow violations have been addressed, benchmarks will be run and the overhead from the sanitizers (efficiency, code measurement, reminiscence footprint) will be measured.
Measuring overhead have to be performed utilizing manufacturing flags – particularly “trapping mode”, the place violations trigger execution to abort. The diagnostics runtime used within the first stage carries important overhead and isn’t indicative of the particular efficiency sanitizer overhead.
Trapping mode will be enabled utilizing the next flags:
-fsanitize=signed-integer-overflow,unsigned-integer-overflow,bounds -fsanitize-trap=all
Many of the overhead is probably going resulting from a small handful of “sizzling capabilities”, for instance these with tight long-running loops. Positive-grained per-function efficiency metrics (just like what Simpleperf gives for Android), permits evaluating metrics earlier than and after sanitizers and gives the best means to establish sizzling capabilities. These capabilities can both be refactored or, after guide inspection to confirm that they’re protected, have sanitization disabled.
Sanitizers will be disabled both inline within the supply or via the usage of ignorelists and the -fsanitize-ignorelist flag.
The bodily layer code, with its extraordinarily tight efficiency margins and decrease probability of exploitable vulnerabilities, could also be an excellent candidate to disable sanitization wholesale if preliminary efficiency appears prohibitive.
Soaking in Pre-production
With overhead minimized and shallow bugs resolved, the ultimate stage is enabling the sanitizers in trapping mode to mitigate vulnerabilities.
We strongly suggest a protracted interval of inner soak in pre-production with check populations to uncover any remaining violations not found in testing. The extra thorough the check protection and size of the soak interval, the much less danger there will probably be from undiscovered violations.
As above, the configuration for trapping mode is as follows:
-fsanitize=signed-integer-overflow,unsigned-integer-overflow,bounds -fsanitize-trap=all
Having infrastructure in place to gather bug studies which outcome from any undiscovered violations might help reduce the chance they current.
The advantages from deploying sanitizers in your present code base are tangible, nevertheless finally they tackle solely the bottom hanging fruit and won’t lead to a code base freed from vulnerabilities. Different courses of reminiscence security vulnerabilities stay unaddressed by these sanitizers. A long run answer is to start transitioning in the present day to memory-safe languages similar to Rust.
Rust is prepared for bare-metal environments such because the baseband, and we’re already utilizing it in different bare-metal parts in Android. There is no such thing as a must rewrite all the pieces in Rust, as Rust gives a powerful C FFI assist and simply interfaces with present C codebases. Simply writing new code in Rust can quickly scale back the variety of reminiscence security vulnerabilities. Rewrites needs to be restricted/prioritized just for probably the most important parts, similar to complicated parsers dealing with untrusted information.
The Android workforce has developed a Rust coaching meant to assist skilled builders shortly ramp up Rust fundamentals. A complete day for bare-metal Rust is included, and the course has been translated to numerous totally different languages.
Whereas the Rust compiler might not explicitly assist your bare-metal goal, as a result of it’s a front-end for LLVM, any goal supported by LLVM will be supported in Rust via customized goal definitions.
Because the high-level working system turns into a tougher goal for attackers to efficiently exploit, we count on that decrease stage parts such because the baseband will entice extra consideration. By utilizing fashionable toolchains and deploying exploit mitigation applied sciences, the bar for attacking the baseband will be raised as nicely. You probably have any questions, tell us – we’re right here to assist!
