Browse WireGuard specs (11 of 38)

Proofs / WireGuard / Replay protection

rp_too_old

A nonce more than 8128 below the newest accepted nonce is rejected.

spec
26 lines of C
with its headers
66 lines
status
Proved
11 of 38
On this page

The spec

The spec is the C program harness/rp_too_old.c, 26 lines, shown whole as it is in the repository at 3c1fc13. It includes harness/wg_laws.h (40 lines). Its first comment states the property in full, with every bound.

C
// Law rp_too_old: For a new key's replay window (zero, as keypair_create makes it), after any sequence of any number
// of any nonces given to counter_validate, any nonce more than COUNTER_WINDOW_SIZE (8128) below the largest nonce
// accepted so far is rejected.
#include "wg_laws.h"

int main(void) {
	// Choose the inputs: any nonces; remember the largest one accepted
	struct noise_replay_counter window;
	new_counter(&window);
	u64 newest = 0;
	while (nondet_u8() != 0) {
		u64 nonce = any_u64();
		bool accepted = wg_counter_validate(&window, nonce);
		if (accepted && nonce > newest)
			newest = nonce;
	}
	// Any nonce n with newest - n > COUNTER_WINDOW_SIZE
	if (newest <= COUNTER_WINDOW_SIZE)
		return 0;
	u64 n = newest - COUNTER_WINDOW_SIZE - 1 - upto(newest - COUNTER_WINDOW_SIZE - 1);
	// Check. RFC 6479, section 2: "If S<WB, the packet is dropped."; receive.c: "This is RFC6479", with the window
	// [newest - COUNTER_WINDOW_SIZE, newest] (selftest/counter.c, tests 15, 19-21, 24, 34).
	bool accepted = wg_counter_validate(&window, n);
	gp_assert(accepted == 0);
	return 0;
}
harness/wg_laws.h, 40 lines
C
// wg_laws.h: trivial helpers shared by the WireGuard laws (each law includes it once). No models, no algorithms.
#include "quickproof.h"
#include "noise.h"
#include <crypto/blake2s.h>
#include <crypto/chacha.h>
#include <crypto/poly1305.h>
#include <crypto/chacha20poly1305.h>
bool wg_counter_validate(struct noise_replay_counter *counter, u64 nonce);   // receive.c's counter_validate (S19)
#define REJECT_AFTER (U64_MAX - COUNTER_WINDOW_SIZE - 1)   // messages.h: REJECT_AFTER_MESSAGES, 2^64 - 8130
static size_t upto(size_t max) {                      // any value from 0 to max
	size_t value = nondet_size();
	return value < max ? value : max;
}
static u64 any_u64(void) {                            // any 64-bit value
	u64 high = nondet_u32();
	return high << 32 | nondet_u32();
}
static u8 *any_bytes(size_t n) {                      // a new object of n arbitrary bytes (NULL if malloc fails)
	u8 *bytes = malloc(n);
	if (bytes != NULL && n != 0)
		nondet_bytes(bytes, n);
	return bytes;
}
static u8 *copy_of(const u8 *source, size_t n) {      // a new object holding source[0..n-1]
	u8 *copy = malloc(n);
	for (size_t i = 0; copy != NULL && i < n; i++)
		copy[i] = source[i];
	return copy;
}
static int bytes_eq(const u8 *a, const u8 *b, size_t n) {   // a[0..n-1] == b[0..n-1]
	for (size_t i = 0; i < n; i++)
		if (a[i] != b[i])
			return 0;
	return 1;
}
static void new_counter(struct noise_replay_counter *c) {    // a new key's replay counter: zero (keypair_create)
	c->counter = 0;
	for (size_t i = 0; i < COUNTER_BITS_TOTAL / BITS_PER_LONG; i++)
		c->backtrack[i] = 0;
}

The theorem

QuickProof generated this statement from the spec, in src/contracts/rp_too_old.v:

Rocq
Theorem rp_too_old : harness_safe whole_prog.

src/contracts/laws.v states it on the spec's own program, whole_prog (the spec linked with the code under test), with its proof:

Rocq
Theorem rp_too_old : harness_safe WgNoise.contracts.rp_too_old.whole_prog.
Proof. exact (proj1 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 (proj2 WgNoise.contracts.suite.suite))))))))))))))))))). Qed.

Read as a statement about runs, it says that no run of the program has undefined behaviour or a failing check, for every value the spec can draw. That reading rests on VST's adequacy theorem, which is not yet instantiated in Rocq (formerly Coq) for QuickProof's environment.

The C files (units) its program links
  • patch/counter_validate.c

What it leaves out

  • What the first comment of the spec excludes, and its bounds.
  • Like every spec of this proof: runs of more than one thread (the locks and RCU are those of one thread), termination (a run that never returns is not a counterexample), and binaries built by a compiler (the theorem is about the code as CompCert defines C for 64-bit x86).
  • The kernel around the code is a shim of plain C for one thread, and the code carries rewrites of idioms that VST cannot handle. Both are tested against the kernel's code, not proved.

The summary page says what the proof as a whole covers.

How it was proved

AI agents wrote the proof, and Rocq checks it. qp verify passes on this repository: this statement is the one generated from the spec at the contract commit, and its proof uses only Rocq's standard axioms and VST's abstract malloc library.