For EIP-4844, Ethereum purchasers want the flexibility to compute and confirm KZG commitments. Quite than every shopper rolling their very own crypto, researchers and builders got here collectively to write down c-kzg-4844, a comparatively small C library with bindings for higher-level languages. The thought was to create a sturdy and environment friendly cryptographic library that each one purchasers might use. The Protocol Safety Analysis staff on the Ethereum Basis had the chance to overview and enhance this library. This weblog submit will talk about some issues we do to make C tasks safer.
Fuzz
Fuzzing is a dynamic code testing approach that includes offering random inputs to find bugs in a program. LibFuzzer and afl++ are two standard fuzzing frameworks for C tasks. They’re each in-process, coverage-guided, evolutionary fuzzing engines. For c-kzg-4844, we used LibFuzzer since we have been already well-integrated with LLVM challenge’s different choices.
Here is the fuzzer for verify_kzg_proof, one in every of c-kzg-4844’s features:
#embody "../base_fuzz.h" static const size_t COMMITMENT_OFFSET = 0; static const size_t Z_OFFSET = COMMITMENT_OFFSET + BYTES_PER_COMMITMENT; static const size_t Y_OFFSET = Z_OFFSET + BYTES_PER_FIELD_ELEMENT; static const size_t PROOF_OFFSET = Y_OFFSET + BYTES_PER_FIELD_ELEMENT; static const size_t INPUT_SIZE = PROOF_OFFSET + BYTES_PER_PROOF; int LLVMFuzzerTestOneInput(const uint8_t* information, size_t measurement) { initialize(); if (measurement == INPUT_SIZE) { bool okay; verify_kzg_proof( &okay, (const Bytes48 *)(information + COMMITMENT_OFFSET), (const Bytes32 *)(information + Z_OFFSET), (const Bytes32 *)(information + Y_OFFSET), (const Bytes48 *)(information + PROOF_OFFSET), &s ); } return 0; }
When executed, that is what the output seems to be like. If there have been an issue, it will write the enter to disk and cease executing. Ideally, you must be capable to reproduce the issue.
There’s additionally differential fuzzing, which is a method which fuzzes two or extra implementations of the identical interface and compares the outputs. For a given enter, if the output is completely different, and also you anticipated them to be the identical, one thing is fallacious. This method could be very standard in Ethereum as a result of we prefer to have a number of implementations of the identical factor. This diversification gives an additional degree of security, realizing that if one implementation have been flawed the others might not have the identical difficulty.
For KZG libraries, we developed kzg-fuzz which differentially fuzzes c-kzg-4844 (via its Golang bindings) and go-kzg-4844. To date, there have not been any variations.
Protection
Subsequent, we used llvm-profdata and llvm-cov to generate a protection report from working the exams. This can be a nice approach to confirm code is executed (“lined”) and examined. See the protection goal in c-kzg-4844’s Makefile for an instance of the best way to generate this report.
When this goal is run (i.e., make protection) it produces a desk that serves as a high-level overview of how a lot of every perform is executed. The exported features are on the prime and the non-exported (static) features are on the underside.
There’s a variety of inexperienced within the desk above, however there may be some yellow and purple too. To find out what’s and is not being executed, seek advice from the HTML file (protection.html) that was generated. This webpage exhibits all the supply file and highlights non-executed code in purple. On this challenge’s case, many of the non-executed code offers with hard-to-test error instances resembling reminiscence allocation failures. For instance, here is some non-executed code:
Initially of this perform, it checks that the trusted setup is sufficiently big to carry out a pairing test. There is not a check case which gives an invalid trusted setup, so this does not get executed. Additionally, as a result of we solely check with the proper trusted setup, the results of is_monomial_form is at all times the identical and does not return the error worth.
Profile
We do not suggest this for all tasks, however since c-kzg-4844 is a efficiency essential library we expect it is essential to profile its exported features and measure how lengthy they take to execute. This may also help determine inefficiencies which might doubtlessly DoS nodes. For this, we used gperftools (Google Efficiency Instruments) as an alternative of llvm-xray as a result of we discovered it to be extra feature-rich and simpler to make use of.
The next is an easy instance which profiles my_function. Profiling works by checking which instruction is being executed once in a while. If a perform is quick sufficient, it is probably not seen by the profiler. To cut back the possibility of this, it’s possible you’ll must name your perform a number of occasions. On this instance, we name my_function 1000 occasions.
#embody <gperftools/profiler.h> int task_a(int n) { if (n <= 1) return 1; return task_a(n - 1) * n; } int task_b(int n) { if (n <= 1) return 1; return task_b(n - 2) + n; } void my_function(void) { for (int i = 0; i < 500; i++) { if (i % 2 == 0) { task_a(i); } else { task_b(i); } } } int foremost(void) { ProfilerStart("instance.prof"); for (int i = 0; i < 1000; i++) { my_function(); } ProfilerStop(); return 0; }
Use ProfilerStart(“<filename>”) and ProfilerStop() to mark which elements of your program to profile. When re-compiled and executed, it is going to write a file to disk with profiling information. You may then use pprof to visualise this information.
Right here is the graph generated from the command above:
Here is a much bigger instance from one in every of c-kzg-4844’s features. The next picture is the profiling graph for compute_blob_kzg_proof. As you’ll be able to see, 80% of this perform’s time is spent performing Montgomery multiplications. That is anticipated.
Reverse
Subsequent, view your binary in a software program reverse engineering (SRE) instrument resembling Ghidra or IDA. These instruments may also help you perceive how high-level constructs are translated into low-level machine code. We expect it helps to overview your code this fashion; like how studying a paper in a special font will power your mind to interpret sentences in another way. It is also helpful to see what kind of optimizations your compiler makes. It is uncommon, however typically the compiler will optimize out one thing which it deemed pointless. Preserve a watch out for this, one thing like this truly occurred in c-kzg-4844, a few of the exams have been being optimized out.
Whenever you view a decompiled perform, it is not going to have variable names, advanced varieties, or feedback. When compiled, this info is not included within the binary. It is going to be as much as you to reverse engineer this. You will usually see features are inlined right into a single perform, a number of variables declared in code are optimized right into a single buffer, and the order of checks are completely different. These are simply compiler optimizations and are typically nice. It might assist to construct your binary with DWARF debugging info; most SREs can analyze this part to offer higher outcomes.
For instance, that is what blob_to_kzg_commitment initially seems to be like in Ghidra:
With a bit of work, you’ll be able to rename variables and add feedback to make it simpler to learn. Here is what it might seem like after a couple of minutes:
Static Evaluation
Clang comes built-in with the Clang Static Analyzer, which is a wonderful static evaluation instrument that may determine many issues that the compiler will miss. Because the title “static” suggests, it examines code with out executing it. That is slower than the compiler, however quite a bit quicker than “dynamic” evaluation instruments which execute code.
Here is a easy instance which forgets to free arr (and has one other drawback however we’ll discuss extra about that later). The compiler is not going to determine this, even with all warnings enabled as a result of technically that is utterly legitimate code.
#embody <stdlib.h> int foremost(void) { int* arr = malloc(5 * sizeof(int)); arr[5] = 42; return 0; }
The unix.Malloc checker will determine that arr wasn’t freed. The road within the warning message is a bit deceptive, nevertheless it is sensible if you concentrate on it; the analyzer reached the return assertion and seen that the reminiscence hadn’t been freed.
Not the entire findings are that easy although. Here is a discovering that Clang Static Analyzer present in c-kzg-4844 when initially launched to the challenge:
Given an surprising enter, it was doable to shift this worth by 32 bits which is undefined conduct. The answer was to limit the enter with CHECK(log2_pow2(n) != 0) in order that this was not possible. Good job, Clang Static Analyzer!
Sanitize
Santizers are dynamic evaluation instruments which instrument (add directions) to packages which may level out points throughout execution. These are significantly helpful at discovering widespread errors related to reminiscence dealing with. Clang comes built-in with a number of sanitizers; listed below are the 4 we discover most helpful and simple to make use of.
Tackle
AddressSanitizer (ASan) is a quick reminiscence error detector which may determine out-of-bounds accesses, use-after-free, use-after-return, use-after-scope, double-free, and reminiscence leaks.
Right here is similar instance from earlier. It forgets to free arr and it’ll set the sixth aspect in a 5 aspect array. This can be a easy instance of a heap-buffer-overflow:
#embody <stdlib.h> int foremost(void) { int* arr = malloc(5 * sizeof(int)); arr[5] = 42; return 0; }
When compiled with -fsanitize=handle and executed, it is going to output the next error message. This factors you in route (a 4-byte write in foremost). This binary may very well be considered in a disassembler to determine precisely which instruction (at foremost+0x84) is inflicting the issue.
Equally, here is an instance the place it finds a heap-use-after-free:
#embody <stdlib.h> int foremost(void) { int *arr = malloc(5 * sizeof(int)); free(arr); return arr[2]; }
It tells you that there is a 4-byte learn of freed reminiscence at foremost+0x8c.
Reminiscence
MemorySanitizer (MSan) is a detector of uninitialized reads. Here is a easy instance which reads (and returns) an uninitialized worth:
int foremost(void) { int information[2]; return information[0]; }
When compiled with -fsanitize=reminiscence and executed, it is going to output the next error message:
Undefined Conduct
UndefinedBehaviorSanitizer (UBSan) detects undefined conduct, which refers back to the state of affairs the place a program’s conduct is unpredictable and never specified by the langauge normal. Some widespread examples of this are accessing out-of-bounds reminiscence, dereferencing an invalid pointer, studying uninitialized variables, and overflow of a signed integer. For instance, right here we increment INT_MAX which is undefined conduct.
#embody <limits.h> int foremost(void) { int a = INT_MAX; return a + 1; }
When compiled with -fsanitize=undefined and executed, it is going to output the next error message which tells us precisely the place the issue is and what the situations are:
Thread
ThreadSanitizer (TSan) detects information races, which may happen in multi-threaded packages when two or extra threads entry a shared reminiscence location on the similar time. This case introduces unpredictability and may result in undefined conduct. Here is an instance during which two threads increment a world counter variable. There are no locks or semaphores, so it is solely doable that these two threads will increment the variable on the similar time.
#embody <pthread.h> int counter = 0; void *increment(void *arg) { (void)arg; for (int i = 0; i < 1000000; i++) counter++; return NULL; } int foremost(void) { pthread_t thread1, thread2; pthread_create(&thread1, NULL, increment, NULL); pthread_create(&thread2, NULL, increment, NULL); pthread_join(thread1, NULL); pthread_join(thread2, NULL); return 0; }
When compiled with -fsanitize=thread and executed, it is going to output the next error message:
This error message tells us that there is a information race. In two threads, the increment perform is writing to the identical 4 bytes on the similar time. It even tells us that the reminiscence is counter.
Valgrind
Valgrind is a robust instrumentation framework for constructing dynamic evaluation instruments, however its finest recognized for figuring out reminiscence errors and leaks with its built-in Memcheck instrument.
The next picture exhibits the output from working c-kzg-4844’s exams with Valgrind. Within the purple field is a legitimate discovering for a “conditional bounce or transfer [that] will depend on uninitialized worth(s).”
This recognized an edge case in expand_root_of_unity. If the fallacious root of unity or width have been supplied, it was doable that the loop will break earlier than out[width] was initialized. On this state of affairs, the ultimate test would depend upon an uninitialized worth.
static C_KZG_RET expand_root_of_unity( fr_t *out, const fr_t *root, uint64_t width ) { out[0] = FR_ONE; out[1] = *root; for (uint64_t i = 2; !fr_is_one(&out[i - 1]); i++) { CHECK(i <= width); blst_fr_mul(&out[i], &out[i - 1], root); } CHECK(fr_is_one(&out[width])); return C_KZG_OK; }
Safety Evaluation
After growth stabilizes, it has been completely examined, and your staff has manually reviewed the codebase themselves a number of occasions, it is time to get a safety overview by a good safety group. This may not be a stamp of approval, nevertheless it exhibits that your challenge is a minimum of considerably safe. Have in mind there isn’t any such factor as good safety. There’ll at all times be the danger of vulnerabilities.
For c-kzg-4844 and go-kzg-4844, the Ethereum Basis contracted Sigma Prime to conduct a safety overview. They produced this report with 8 findings. It accommodates one essential vulnerability in go-kzg-4844 that was a extremely good discover. The BLS12-381 library that go-kzg-4844 makes use of, gnark-crypto, had a bug which allowed invalid G1 and G2 factors to be sucessfully decoded. Had this not been fastened, this might have resulted in a consensus bug (a disagreement between implementations) in Ethereum.
Bug Bounty
If a vulnerability in your challenge may very well be exploited for positive factors, like it’s for Ethereum, think about organising a bug bounty program. This enables safety researchers, or anybody actually, to submit vulnerability studies in alternate for cash. Typically, that is particularly for findings which may show that an exploit is feasible. If the bug bounty payouts are affordable, bug finders will notify you of the bug reasonably than exploiting it or promoting it to a different celebration. We suggest beginning your bug bounty program after the findings from the primary safety overview are resolved; ideally, the safety overview would value lower than the bug bounty payouts.
Conclusion
The event of sturdy C tasks, particularly within the essential area of blockchain and cryptocurrencies, requires a multi-faceted strategy. Given the inherent vulnerabilities related to the C language, a mixture of finest practices and instruments is crucial for producing resilient software program. We hope our experiences and findings from our work with c-kzg-4844 present helpful insights and finest practices for others embarking on comparable tasks.
