hash-test-template.h

  1/* SPDX-License-Identifier: GPL-2.0-or-later */
  2/*
  3 * Test cases for hash functions, including a benchmark.  This is included by
  4 * KUnit test suites that want to use it.  See sha512_kunit.c for an example.
  5 *
  6 * Copyright 2025 Google LLC
  7 */
  8#include <kunit/run-in-irq-context.h>
  9#include <kunit/test.h>
 10#include <linux/vmalloc.h>
 11
 12/* test_buf is a guarded buffer, i.e. &test_buf[TEST_BUF_LEN] is not mapped. */
 13#define TEST_BUF_LEN 16384
 14static u8 *test_buf;
 15
 16static u8 *orig_test_buf;
 17
 18static u64 random_seed;
 19
 20/*
 21 * This is a simple linear congruential generator.  It is used only for testing,
 22 * which does not require cryptographically secure random numbers.  A hard-coded
 23 * algorithm is used instead of <linux/prandom.h> so that it matches the
 24 * algorithm used by the test vector generation script.  This allows the input
 25 * data in random test vectors to be concisely stored as just the seed.
 26 */
 27static u32 rand32(void)
 28{
 29	random_seed = (random_seed * 25214903917 + 11) & ((1ULL << 48) - 1);
 30	return random_seed >> 16;
 31}
 32
 33static void rand_bytes(u8 *out, size_t len)
 34{
 35	for (size_t i = 0; i < len; i++)
 36		out[i] = rand32();
 37}
 38
 39static void rand_bytes_seeded_from_len(u8 *out, size_t len)
 40{
 41	random_seed = len;
 42	rand_bytes(out, len);
 43}
 44
 45static bool rand_bool(void)
 46{
 47	return rand32() % 2;
 48}
 49
 50/* Generate a random length, preferring small lengths. */
 51static size_t rand_length(size_t max_len)
 52{
 53	size_t len;
 54
 55	switch (rand32() % 3) {
 56	case 0:
 57		len = rand32() % 128;
 58		break;
 59	case 1:
 60		len = rand32() % 3072;
 61		break;
 62	default:
 63		len = rand32();
 64		break;
 65	}
 66	return len % (max_len + 1);
 67}
 68
 69static size_t rand_offset(size_t max_offset)
 70{
 71	return min(rand32() % 128, max_offset);
 72}
 73
 74static int hash_suite_init(struct kunit_suite *suite)
 75{
 76	/*
 77	 * Allocate the test buffer using vmalloc() with a page-aligned length
 78	 * so that it is immediately followed by a guard page.  This allows
 79	 * buffer overreads to be detected, even in assembly code.
 80	 */
 81	size_t alloc_len = round_up(TEST_BUF_LEN, PAGE_SIZE);
 82
 83	orig_test_buf = vmalloc(alloc_len);
 84	if (!orig_test_buf)
 85		return -ENOMEM;
 86
 87	test_buf = orig_test_buf + alloc_len - TEST_BUF_LEN;
 88	return 0;
 89}
 90
 91static void hash_suite_exit(struct kunit_suite *suite)
 92{
 93	vfree(orig_test_buf);
 94	orig_test_buf = NULL;
 95	test_buf = NULL;
 96}
 97
 98/*
 99 * Test the hash function against a list of test vectors.
100 *
101 * Note that it's only necessary to run each test vector in one way (e.g.,
102 * one-shot instead of incremental), since consistency between different ways of
103 * using the APIs is verified by other test cases.
104 */
105static void test_hash_test_vectors(struct kunit *test)
106{
107	for (size_t i = 0; i < ARRAY_SIZE(hash_testvecs); i++) {
108		size_t data_len = hash_testvecs[i].data_len;
109		u8 actual_hash[HASH_SIZE];
110
111		KUNIT_ASSERT_LE(test, data_len, TEST_BUF_LEN);
112		rand_bytes_seeded_from_len(test_buf, data_len);
113
114		HASH(test_buf, data_len, actual_hash);
115		KUNIT_ASSERT_MEMEQ_MSG(
116			test, actual_hash, hash_testvecs[i].digest, HASH_SIZE,
117			"Wrong result with test vector %zu; data_len=%zu", i,
118			data_len);
119	}
120}
121
122/*
123 * Test that the hash function produces correct results for *every* length up to
124 * 4096 bytes.  To do this, generate seeded random data, then calculate a hash
125 * value for each length 0..4096, then hash the hash values.  Verify just the
126 * final hash value, which should match only when all hash values were correct.
127 */
128static void test_hash_all_lens_up_to_4096(struct kunit *test)
129{
130	struct HASH_CTX ctx;
131	u8 hash[HASH_SIZE];
132
133	static_assert(TEST_BUF_LEN >= 4096);
134	rand_bytes_seeded_from_len(test_buf, 4096);
135	HASH_INIT(&ctx);
136	for (size_t len = 0; len <= 4096; len++) {
137		HASH(test_buf, len, hash);
138		HASH_UPDATE(&ctx, hash, HASH_SIZE);
139	}
140	HASH_FINAL(&ctx, hash);
141	KUNIT_ASSERT_MEMEQ(test, hash, hash_testvec_consolidated, HASH_SIZE);
142}
143
144/*
145 * Test that the hash function produces the same result with a one-shot
146 * computation as it does with an incremental computation.
147 */
148static void test_hash_incremental_updates(struct kunit *test)
149{
150	for (int i = 0; i < 1000; i++) {
151		size_t total_len, offset;
152		struct HASH_CTX ctx;
153		u8 hash1[HASH_SIZE];
154		u8 hash2[HASH_SIZE];
155		size_t num_parts = 0;
156		size_t remaining_len, cur_offset;
157
158		total_len = rand_length(TEST_BUF_LEN);
159		offset = rand_offset(TEST_BUF_LEN - total_len);
160		rand_bytes(&test_buf[offset], total_len);
161
162		/* Compute the hash value in one shot. */
163		HASH(&test_buf[offset], total_len, hash1);
164
165		/*
166		 * Compute the hash value incrementally, using a randomly
167		 * selected sequence of update lengths that sum to total_len.
168		 */
169		HASH_INIT(&ctx);
170		remaining_len = total_len;
171		cur_offset = offset;
172		while (rand_bool()) {
173			size_t part_len = rand_length(remaining_len);
174
175			HASH_UPDATE(&ctx, &test_buf[cur_offset], part_len);
176			num_parts++;
177			cur_offset += part_len;
178			remaining_len -= part_len;
179		}
180		if (remaining_len != 0 || rand_bool()) {
181			HASH_UPDATE(&ctx, &test_buf[cur_offset], remaining_len);
182			num_parts++;
183		}
184		HASH_FINAL(&ctx, hash2);
185
186		/* Verify that the two hash values are the same. */
187		KUNIT_ASSERT_MEMEQ_MSG(
188			test, hash1, hash2, HASH_SIZE,
189			"Incremental test failed with total_len=%zu num_parts=%zu offset=%zu",
190			total_len, num_parts, offset);
191	}
192}
193
194/*
195 * Test that the hash function does not overrun any buffers.  Uses a guard page
196 * to catch buffer overruns even if they occur in assembly code.
197 */
198static void test_hash_buffer_overruns(struct kunit *test)
199{
200	const size_t max_tested_len = TEST_BUF_LEN - sizeof(struct HASH_CTX);
201	void *const buf_end = &test_buf[TEST_BUF_LEN];
202	struct HASH_CTX *guarded_ctx = buf_end - sizeof(*guarded_ctx);
203
204	rand_bytes(test_buf, TEST_BUF_LEN);
205
206	for (int i = 0; i < 100; i++) {
207		size_t len = rand_length(max_tested_len);
208		struct HASH_CTX ctx;
209		u8 hash[HASH_SIZE];
210
211		/* Check for overruns of the data buffer. */
212		HASH(buf_end - len, len, hash);
213		HASH_INIT(&ctx);
214		HASH_UPDATE(&ctx, buf_end - len, len);
215		HASH_FINAL(&ctx, hash);
216
217		/* Check for overruns of the hash value buffer. */
218		HASH(test_buf, len, buf_end - HASH_SIZE);
219		HASH_INIT(&ctx);
220		HASH_UPDATE(&ctx, test_buf, len);
221		HASH_FINAL(&ctx, buf_end - HASH_SIZE);
222
223		/* Check for overuns of the hash context. */
224		HASH_INIT(guarded_ctx);
225		HASH_UPDATE(guarded_ctx, test_buf, len);
226		HASH_FINAL(guarded_ctx, hash);
227	}
228}
229
230/*
231 * Test that the caller is permitted to alias the output digest and source data
232 * buffer, and also modify the source data buffer after it has been used.
233 */
234static void test_hash_overlaps(struct kunit *test)
235{
236	const size_t max_tested_len = TEST_BUF_LEN - HASH_SIZE;
237	struct HASH_CTX ctx;
238	u8 hash[HASH_SIZE];
239
240	rand_bytes(test_buf, TEST_BUF_LEN);
241
242	for (int i = 0; i < 100; i++) {
243		size_t len = rand_length(max_tested_len);
244		size_t offset = HASH_SIZE + rand_offset(max_tested_len - len);
245		bool left_end = rand_bool();
246		u8 *ovl_hash = left_end ? &test_buf[offset] :
247					  &test_buf[offset + len - HASH_SIZE];
248
249		HASH(&test_buf[offset], len, hash);
250		HASH(&test_buf[offset], len, ovl_hash);
251		KUNIT_ASSERT_MEMEQ_MSG(
252			test, hash, ovl_hash, HASH_SIZE,
253			"Overlap test 1 failed with len=%zu offset=%zu left_end=%d",
254			len, offset, left_end);
255
256		/* Repeat the above test, but this time use init+update+final */
257		HASH(&test_buf[offset], len, hash);
258		HASH_INIT(&ctx);
259		HASH_UPDATE(&ctx, &test_buf[offset], len);
260		HASH_FINAL(&ctx, ovl_hash);
261		KUNIT_ASSERT_MEMEQ_MSG(
262			test, hash, ovl_hash, HASH_SIZE,
263			"Overlap test 2 failed with len=%zu offset=%zu left_end=%d",
264			len, offset, left_end);
265
266		/* Test modifying the source data after it was used. */
267		HASH(&test_buf[offset], len, hash);
268		HASH_INIT(&ctx);
269		HASH_UPDATE(&ctx, &test_buf[offset], len);
270		rand_bytes(&test_buf[offset], len);
271		HASH_FINAL(&ctx, ovl_hash);
272		KUNIT_ASSERT_MEMEQ_MSG(
273			test, hash, ovl_hash, HASH_SIZE,
274			"Overlap test 3 failed with len=%zu offset=%zu left_end=%d",
275			len, offset, left_end);
276	}
277}
278
279/*
280 * Test that if the same data is hashed at different alignments in memory, the
281 * results are the same.
282 */
283static void test_hash_alignment_consistency(struct kunit *test)
284{
285	u8 hash1[128 + HASH_SIZE];
286	u8 hash2[128 + HASH_SIZE];
287
288	for (int i = 0; i < 100; i++) {
289		size_t len = rand_length(TEST_BUF_LEN);
290		size_t data_offs1 = rand_offset(TEST_BUF_LEN - len);
291		size_t data_offs2 = rand_offset(TEST_BUF_LEN - len);
292		size_t hash_offs1 = rand_offset(128);
293		size_t hash_offs2 = rand_offset(128);
294
295		rand_bytes(&test_buf[data_offs1], len);
296		HASH(&test_buf[data_offs1], len, &hash1[hash_offs1]);
297		memmove(&test_buf[data_offs2], &test_buf[data_offs1], len);
298		HASH(&test_buf[data_offs2], len, &hash2[hash_offs2]);
299		KUNIT_ASSERT_MEMEQ_MSG(
300			test, &hash1[hash_offs1], &hash2[hash_offs2], HASH_SIZE,
301			"Alignment consistency test failed with len=%zu data_offs=(%zu,%zu) hash_offs=(%zu,%zu)",
302			len, data_offs1, data_offs2, hash_offs1, hash_offs2);
303	}
304}
305
306/* Test that HASH_FINAL zeroizes the context. */
307static void test_hash_ctx_zeroization(struct kunit *test)
308{
309	static const u8 zeroes[sizeof(struct HASH_CTX)];
310	struct HASH_CTX ctx;
311
312	rand_bytes(test_buf, 128);
313	HASH_INIT(&ctx);
314	HASH_UPDATE(&ctx, test_buf, 128);
315	HASH_FINAL(&ctx, test_buf);
316	KUNIT_ASSERT_MEMEQ_MSG(test, &ctx, zeroes, sizeof(ctx),
317			       "Hash context was not zeroized by finalization");
318}
319
320#define IRQ_TEST_DATA_LEN 256
321#define IRQ_TEST_NUM_BUFFERS 3 /* matches max concurrency level */
322
323struct hash_irq_test1_state {
324	u8 expected_hashes[IRQ_TEST_NUM_BUFFERS][HASH_SIZE];
325	atomic_t seqno;
326};
327
328/*
329 * Compute the hash of one of the test messages and verify that it matches the
330 * expected hash from @state->expected_hashes.  To increase the chance of
331 * detecting problems, cycle through multiple messages.
332 */
333static bool hash_irq_test1_func(void *state_)
334{
335	struct hash_irq_test1_state *state = state_;
336	u32 i = (u32)atomic_inc_return(&state->seqno) % IRQ_TEST_NUM_BUFFERS;
337	u8 actual_hash[HASH_SIZE];
338
339	HASH(&test_buf[i * IRQ_TEST_DATA_LEN], IRQ_TEST_DATA_LEN, actual_hash);
340	return memcmp(actual_hash, state->expected_hashes[i], HASH_SIZE) == 0;
341}
342
343/*
344 * Test that if hashes are computed in task, softirq, and hardirq context
345 * concurrently, then all results are as expected.
346 */
347static void test_hash_interrupt_context_1(struct kunit *test)
348{
349	struct hash_irq_test1_state state = {};
350
351	/* Prepare some test messages and compute the expected hash of each. */
352	rand_bytes(test_buf, IRQ_TEST_NUM_BUFFERS * IRQ_TEST_DATA_LEN);
353	for (int i = 0; i < IRQ_TEST_NUM_BUFFERS; i++)
354		HASH(&test_buf[i * IRQ_TEST_DATA_LEN], IRQ_TEST_DATA_LEN,
355		     state.expected_hashes[i]);
356
357	kunit_run_irq_test(test, hash_irq_test1_func, 100000, &state);
358}
359
360struct hash_irq_test2_hash_ctx {
361	struct HASH_CTX hash_ctx;
362	atomic_t in_use;
363	int offset;
364	int step;
365};
366
367struct hash_irq_test2_state {
368	struct hash_irq_test2_hash_ctx ctxs[IRQ_TEST_NUM_BUFFERS];
369	u8 expected_hash[HASH_SIZE];
370	u16 update_lens[32];
371	int num_steps;
372};
373
374static bool hash_irq_test2_func(void *state_)
375{
376	struct hash_irq_test2_state *state = state_;
377	struct hash_irq_test2_hash_ctx *ctx;
378	bool ret = true;
379
380	for (ctx = &state->ctxs[0]; ctx < &state->ctxs[ARRAY_SIZE(state->ctxs)];
381	     ctx++) {
382		if (atomic_cmpxchg(&ctx->in_use, 0, 1) == 0)
383			break;
384	}
385	if (WARN_ON_ONCE(ctx == &state->ctxs[ARRAY_SIZE(state->ctxs)])) {
386		/*
387		 * This should never happen, as the number of contexts is equal
388		 * to the maximum concurrency level of kunit_run_irq_test().
389		 */
390		return false;
391	}
392
393	if (ctx->step == 0) {
394		/* Init step */
395		HASH_INIT(&ctx->hash_ctx);
396		ctx->offset = 0;
397		ctx->step++;
398	} else if (ctx->step < state->num_steps - 1) {
399		/* Update step */
400		HASH_UPDATE(&ctx->hash_ctx, &test_buf[ctx->offset],
401			    state->update_lens[ctx->step - 1]);
402		ctx->offset += state->update_lens[ctx->step - 1];
403		ctx->step++;
404	} else {
405		/* Final step */
406		u8 actual_hash[HASH_SIZE];
407
408		if (WARN_ON_ONCE(ctx->offset != TEST_BUF_LEN))
409			ret = false;
410		HASH_FINAL(&ctx->hash_ctx, actual_hash);
411		if (memcmp(actual_hash, state->expected_hash, HASH_SIZE) != 0)
412			ret = false;
413		ctx->step = 0;
414	}
415	atomic_set_release(&ctx->in_use, 0);
416	return ret;
417}
418
419/*
420 * Test that if hashes are computed in task, softirq, and hardirq context
421 * concurrently, *including doing different parts of the same incremental
422 * computation in different contexts*, then all results are as expected.
423 * Besides detecting bugs similar to those that test_hash_interrupt_context_1
424 * can detect, this test case can also detect bugs where hash function
425 * implementations don't correctly handle these mixed incremental computations.
426 */
427static void test_hash_interrupt_context_2(struct kunit *test)
428{
429	struct hash_irq_test2_state *state;
430	int remaining = TEST_BUF_LEN;
431
432	state = kunit_kzalloc(test, sizeof(*state), GFP_KERNEL);
433	KUNIT_ASSERT_NOT_NULL(test, state);
434
435	rand_bytes(test_buf, TEST_BUF_LEN);
436	HASH(test_buf, TEST_BUF_LEN, state->expected_hash);
437
438	/*
439	 * Generate a list of update lengths to use.  Ensure that it contains
440	 * multiple entries but is limited to a maximum length.
441	 */
442	static_assert(TEST_BUF_LEN / 4096 > 1);
443	for (state->num_steps = 0;
444	     state->num_steps < ARRAY_SIZE(state->update_lens) - 1 && remaining;
445	     state->num_steps++) {
446		state->update_lens[state->num_steps] =
447			rand_length(min(remaining, 4096));
448		remaining -= state->update_lens[state->num_steps];
449	}
450	if (remaining)
451		state->update_lens[state->num_steps++] = remaining;
452	state->num_steps += 2; /* for init and final */
453
454	kunit_run_irq_test(test, hash_irq_test2_func, 250000, state);
455}
456
457#define UNKEYED_HASH_KUNIT_CASES                     \
458	KUNIT_CASE(test_hash_test_vectors),          \
459	KUNIT_CASE(test_hash_all_lens_up_to_4096),   \
460	KUNIT_CASE(test_hash_incremental_updates),   \
461	KUNIT_CASE(test_hash_buffer_overruns),       \
462	KUNIT_CASE(test_hash_overlaps),              \
463	KUNIT_CASE(test_hash_alignment_consistency), \
464	KUNIT_CASE(test_hash_ctx_zeroization),       \
465	KUNIT_CASE(test_hash_interrupt_context_1),   \
466	KUNIT_CASE(test_hash_interrupt_context_2)
467/* benchmark_hash is omitted so that the suites can put it last. */
468
469#ifdef HMAC
470/*
471 * Test the corresponding HMAC variant.
472 *
473 * This test case is fairly short, since HMAC is just a simple C wrapper around
474 * the underlying unkeyed hash function, which is already well-tested by the
475 * other test cases.  It's not useful to test things like data alignment or
476 * interrupt context again for HMAC, nor to have a long list of test vectors.
477 *
478 * Thus, just do a single consolidated test, which covers all data lengths up to
479 * 4096 bytes and all key lengths up to 292 bytes.  For each data length, select
480 * a key length, generate the inputs from a seed, and compute the HMAC value.
481 * Concatenate all these HMAC values together, and compute the HMAC of that.
482 * Verify that value.  If this fails, then the HMAC implementation is wrong.
483 * This won't show which specific input failed, but that should be fine.  Any
484 * failure would likely be non-input-specific or also show in the unkeyed tests.
485 */
486static void test_hmac(struct kunit *test)
487{
488	static const u8 zeroes[sizeof(struct HMAC_CTX)];
489	u8 *raw_key;
490	struct HMAC_KEY key;
491	struct HMAC_CTX ctx;
492	u8 mac[HASH_SIZE];
493	u8 mac2[HASH_SIZE];
494
495	static_assert(TEST_BUF_LEN >= 4096 + 293);
496	rand_bytes_seeded_from_len(test_buf, 4096);
497	raw_key = &test_buf[4096];
498
499	rand_bytes_seeded_from_len(raw_key, 32);
500	HMAC_PREPAREKEY(&key, raw_key, 32);
501	HMAC_INIT(&ctx, &key);
502	for (size_t data_len = 0; data_len <= 4096; data_len++) {
503		/*
504		 * Cycle through key lengths as well.  Somewhat arbitrarily go
505		 * up to 293, which is somewhat larger than the largest hash
506		 * block size (which is the size at which the key starts being
507		 * hashed down to one block); going higher would not be useful.
508		 * To reduce correlation with data_len, use a prime number here.
509		 */
510		size_t key_len = data_len % 293;
511
512		HMAC_UPDATE(&ctx, test_buf, data_len);
513
514		rand_bytes_seeded_from_len(raw_key, key_len);
515		HMAC_USINGRAWKEY(raw_key, key_len, test_buf, data_len, mac);
516		HMAC_UPDATE(&ctx, mac, HASH_SIZE);
517
518		/* Verify that HMAC() is consistent with HMAC_USINGRAWKEY(). */
519		HMAC_PREPAREKEY(&key, raw_key, key_len);
520		HMAC(&key, test_buf, data_len, mac2);
521		KUNIT_ASSERT_MEMEQ_MSG(
522			test, mac, mac2, HASH_SIZE,
523			"HMAC gave different results with raw and prepared keys");
524	}
525	HMAC_FINAL(&ctx, mac);
526	KUNIT_EXPECT_MEMEQ_MSG(test, mac, hmac_testvec_consolidated, HASH_SIZE,
527			       "HMAC gave wrong result");
528	KUNIT_EXPECT_MEMEQ_MSG(test, &ctx, zeroes, sizeof(ctx),
529			       "HMAC context was not zeroized by finalization");
530}
531#define HASH_KUNIT_CASES UNKEYED_HASH_KUNIT_CASES, KUNIT_CASE(test_hmac)
532#else
533#define HASH_KUNIT_CASES UNKEYED_HASH_KUNIT_CASES
534#endif
535
536/* Benchmark the hash function on various data lengths. */
537static void benchmark_hash(struct kunit *test)
538{
539	static const size_t lens_to_test[] = {
540		1,   16,  64,	127,  128,  200,   256,
541		511, 512, 1024, 3173, 4096, 16384,
542	};
543	u8 hash[HASH_SIZE];
544
545	if (!IS_ENABLED(CONFIG_CRYPTO_LIB_BENCHMARK))
546		kunit_skip(test, "not enabled");
547
548	/* Warm-up */
549	for (size_t i = 0; i < 10000000; i += TEST_BUF_LEN)
550		HASH(test_buf, TEST_BUF_LEN, hash);
551
552	for (size_t i = 0; i < ARRAY_SIZE(lens_to_test); i++) {
553		size_t len = lens_to_test[i];
554		/* The '+ 128' tries to account for per-message overhead. */
555		size_t num_iters = 10000000 / (len + 128);
556		u64 t;
557
558		KUNIT_ASSERT_LE(test, len, TEST_BUF_LEN);
559		preempt_disable();
560		t = ktime_get_ns();
561		for (size_t j = 0; j < num_iters; j++)
562			HASH(test_buf, len, hash);
563		t = ktime_get_ns() - t;
564		preempt_enable();
565		kunit_info(test, "len=%zu: %llu MB/s", len,
566			   div64_u64((u64)len * num_iters * 1000, t ?: 1));
567	}
568}