Files
memtest86plus/tests/tests.c
T
Sam Demeulemeester 72d08688ba Add rowhammer test (new test 11), disabled by default
Blacksmith-style non-uniform, REFRESH-synchronised many-sided hammering with DRAMA-lite row-stride discovery and a blind-stride fallback
x86 clflush(opt) engine with a coarse whole-cache-flush fallback for other targets. LA64 TBT
2026-07-11 11:22:01 +02:00

306 lines
9.8 KiB
C

// SPDX-License-Identifier: GPL-2.0
// Copyright (C) 2020-2022 Martin Whitaker.
// Copyright (C) 2004-2026 Sam Demeulemeester.
//
// ------------------------------------------------
// main.c - MemTest-86 Version 3.5
//
// Released under version 2 of the Gnu Public License.
// By Chris Brady
#include <stdbool.h>
#include <stdint.h>
#include "boot.h"
#include "cache.h"
#include "cpuid.h"
#include "memsize.h"
#include "simd.h"
#include "tsc.h"
#include "vmem.h"
#include "barrier.h"
#include "config.h"
#include "display.h"
#include "test.h"
#include "test_funcs.h"
#include "test_helper.h"
#include "tests.h"
//------------------------------------------------------------------------------
// Constants
//------------------------------------------------------------------------------
#ifndef TRACE_BARRIERS
#define TRACE_BARRIERS 0
#endif
#define MODULO_N 20
// The test whose description is patched with the SIMD tier by test_list_init().
#define MOV_INV_RNG_TEST 4
//------------------------------------------------------------------------------
// Public Variables
//------------------------------------------------------------------------------
test_pattern_t test_list[NUM_TEST_PATTERNS] = {
// ena, cpu, stgs, itrs, errs, description
{ true, ONE, 1, 6, 0, "[Address test, walking ones, no cache] "},
{false, ONE, 1, 6, 0, "[Address test, own address in window] "},
{ true, ONE, 2, 6, 0, "[Address test, own address + window] "},
{ true, PAR, 1, 6, 0, "[Moving inversions, 1s & 0s] "},
{ true, PAR, 1, 128, 0, "[Moving inversions, random sequence] "},
{ true, PAR, 1, 3, 0, "[Moving inversions, 8 bit pattern] "},
{ true, PAR, 1, 8, 0, "[Modulo 20, random pattern] "},
{ true, PAR, 1, 81, 0, "[Block move] "},
#if TESTWORD_WIDTH > 32
{ true, PAR, 1, 1, 0, "[Moving inversions, 64 bit pattern] "},
#else
{ true, PAR, 1, 1, 0, "[Moving inversions, 32 bit pattern] "},
#endif
{ true, PAR, 1, 32, 0, "[Bus stress, R/W turnaround, random] "},
{ true, PAR, 12, 120, 0, "[Bit fade test, 0s, 1s, random] "},
{false, ONE, 1, 24, 0, "[Rowhammer, Blacksmith-style] "},
};
int ticks_per_pass[NUM_PASS_TYPES];
int ticks_per_test[NUM_PASS_TYPES][NUM_TEST_PATTERNS];
//------------------------------------------------------------------------------
// Public Functions
//------------------------------------------------------------------------------
#define BARRIER \
if (my_cpu >= 0) { \
if (TRACE_BARRIERS) { \
trace(my_cpu, "Run barrier wait begin at %s line %i", __FILE__, __LINE__); \
} \
if (power_save < POWER_SAVE_HIGH) { \
barrier_spin_wait(run_barrier); \
} else { \
barrier_halt_wait(run_barrier); \
} \
if (TRACE_BARRIERS) { \
trace(my_cpu, "Run barrier wait end at %s line %i", __FILE__, __LINE__); \
} \
}
void test_list_init(void)
{
const char *tier_name = simd_tier_name();
if (tier_name == NULL) {
// Keep the generic description from the table.
return;
}
// Rewrite the description to show the SIMD tier used by the test,
// e.g. "[Moving inversions, random (AVX2)]".
const char *prefix = "[Moving inversions, random (";
char *desc = test_list[MOV_INV_RNG_TEST].description;
int i = 0;
while (prefix[i] != '\0') {
desc[i] = prefix[i];
i++;
}
for (int j = 0; tier_name[j] != '\0'; j++) {
desc[i++] = tier_name[j];
}
desc[i++] = ')';
desc[i++] = ']';
while (i < (int)sizeof(test_list[MOV_INV_RNG_TEST].description) - 1) {
desc[i++] = ' ';
}
desc[i] = '\0';
}
int run_test(int my_cpu, int test, int stage, int iterations)
{
if (my_cpu == master_cpu) {
if (window_num == 0) {
// First window, so we need to test all selected lower memory.
vm_map[0].start = first_word_mapping(pm_limit_lower);
// For USB_WORKAROUND.
if (vm_map[0].start < (uintptr_t *)0x500) {
vm_map[0].start = (uintptr_t *)0x500;
}
}
/* Update display of memory segments being tested */
uintptr_t pb = page_of(vm_map[0].start);
uintptr_t pe = page_of(vm_map[vm_map_size - 1].end) + 1;
display_test_addresses(pb << 2, pe << 2, num_pages_to_test << 2);
}
BARRIER;
testword_t prsg_state;
int ticks = 0;
switch (test) {
// Address test, walking ones.
case 0:
if (my_cpu >= 0) cache_off();
ticks += test_addr_walk1(my_cpu);
if (my_cpu >= 0) cache_on();
BAILOUT;
break;
// Address test, own address in window.
case 1:
ticks += test_own_addr1(my_cpu);
BAILOUT;
break;
// Address test, own address + window.
case 2:
ticks += test_own_addr2(my_cpu, stage);
BAILOUT;
break;
// Moving inversions, all ones and zeros.
case 3: {
testword_t pattern1 = 0;
testword_t pattern2 = ~pattern1;
BARRIER;
ticks += test_mov_inv_fixed(my_cpu, iterations, pattern1, pattern2);
BAILOUT;
BARRIER;
ticks += test_mov_inv_fixed(my_cpu, iterations, pattern2, pattern1);
BAILOUT;
} break;
// Moving inversions, pseudo-random sequence (SIMD where available).
// Every fourth round broadcasts a single random value to all vector
// lanes, preserving the uniform-background model of the classic
// random pattern test. Runs early: it has the highest fault
// detection rate per second, so this minimises time to first fault.
case 4:
for (int i = 0; i < iterations; i++) {
BARRIER;
ticks += test_mov_inv_rng(my_cpu, (i & 3) == 3);
BAILOUT;
}
break;
// Moving inversions, 8 bit walking ones and zeros.
case 5: {
#if TESTWORD_WIDTH > 32
testword_t pattern1 = UINT64_C(0x8080808080808080);
#else
testword_t pattern1 = 0x80808080;
#endif
for (int i = 0; i < 8; i++) {
testword_t pattern2 = ~pattern1;
BARRIER;
ticks += test_mov_inv_fixed(my_cpu, iterations, pattern1, pattern2);
BAILOUT;
BARRIER;
ticks += test_mov_inv_fixed(my_cpu, iterations, pattern2, pattern1);
BAILOUT;
pattern1 >>= 1;
}
} break;
// Modulo 20 check, fixed random pattern. Runs before the long
// shifting pattern test: it is the only test immune to cache
// masking, so every fault class has been probed early in the pass.
case 6:
if (cpuid_info.flags.rdtsc) {
prsg_state = get_tsc();
} else {
prsg_state = 1 + pass_num;
}
prsg_state *= 0x87654321;
for (int i = 0; i < iterations; i++) {
for (int offset = 0; offset < MODULO_N; offset++) {
prsg_state = prsg(prsg_state);
testword_t pattern1 = prsg_state;
testword_t pattern2 = ~pattern1;
BARRIER;
ticks += test_modulo_n(my_cpu, 2, pattern1, pattern2, MODULO_N, offset);
BAILOUT;
BARRIER;
ticks += test_modulo_n(my_cpu, 2, pattern2, pattern1, MODULO_N, offset);
BAILOUT;
}
}
break;
// Block move.
case 7:
ticks += test_block_move(my_cpu, iterations);
BAILOUT;
break;
// Moving inversions, 32/64 bit shifting pattern. A single iteration
// per pass suffices: the patterns are deterministic, so repeating
// them within a pass adds nothing that the next pass doesn't. On the
// fast first pass only every other offset is walked; full coverage
// is restored on every full pass.
case 8: {
if (iterations < 1) {
iterations = 1; // the first pass divides the table value by 3
}
int offset_step = (pass_num == 0) ? 2 : 1;
for (int offset = 0; offset < TESTWORD_WIDTH; offset += offset_step) {
BARRIER;
ticks += test_mov_inv_walk1(my_cpu, iterations, offset, false);
BAILOUT;
BARRIER;
ticks += test_mov_inv_walk1(my_cpu, iterations, offset, true);
BAILOUT;
}
} break;
// Bus stress: NT-write/read burst interleave between two half-chunk streams, duty-cycled
// on odd rounds to provoke PMIC load steps. Runs before bit fade so the DIMMs enter it hot.
case 9:
if (iterations < 1) {
iterations = 1; // the first pass divides the table value by 3
}
for (int i = 0; i < iterations; i++) {
BARRIER;
ticks += test_bus_stress(my_cpu, i);
BAILOUT;
}
break;
// Bit fade test: four fill/fade/check rounds - solid zeros, solid ones, then an
// address-seeded random pattern and its complement. `iterations` = fade seconds per round.
case 10:
ticks += test_bit_fade(my_cpu, stage, iterations);
BAILOUT;
break;
// Rowhammer test: Blacksmith-style non-uniform, REFRESH-synchronised,
// many-sided hammering of a time-boxed sample of sites. Disabled by
// default. Runs on a single core (ONE): the DRAM flips regardless of
// which core hammers, so per-core repetition would only multiply the
// runtime, and a lone core keeps activation timing clean.
// `iterations` = seconds budget. See tests/rowhammer.c.
case 11:
ticks += test_rowhammer(my_cpu, iterations);
BAILOUT;
break;
}
return ticks;
}