1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * fill_buf benchmark
  4 *
  5 * Copyright (C) 2018 Intel Corporation
  6 *
  7 * Authors:
  8 *    Sai Praneeth Prakhya <sai.praneeth.prakhya@intel.com>,
  9 *    Fenghua Yu <fenghua.yu@intel.com>
 10 */
 11#include <stdio.h>
 12#include <unistd.h>
 13#include <stdlib.h>
 14#include <sys/types.h>
 15#include <sys/wait.h>
 16#include <inttypes.h>
 17#include <string.h>
 18
 19#include "resctrl.h"
 20
 21#define CL_SIZE			(64)
 22#define PAGE_SIZE		(4 * 1024)
 23#define MB			(1024 * 1024)
 24
 25static void sb(void)
 26{
 27#if defined(__i386) || defined(__x86_64)
 28	asm volatile("sfence\n\t"
 29		     : : : "memory");
 30#endif
 31}
 32
 33static void cl_flush(void *p)
 34{
 35#if defined(__i386) || defined(__x86_64)
 36	asm volatile("clflush (%0)\n\t"
 37		     : : "r"(p) : "memory");
 38#endif
 39}
 40
 41void mem_flush(unsigned char *buf, size_t buf_size)
 42{
 43	unsigned char *cp = buf;
 44	size_t i = 0;
 45
 46	buf_size = buf_size / CL_SIZE; /* mem size in cache lines */
 47
 48	for (i = 0; i < buf_size; i++)
 49		cl_flush(&cp[i * CL_SIZE]);
 50
 51	sb();
 52}
 53
 54/*
 55 * Buffer index step advance to workaround HW prefetching interfering with
 56 * the measurements.
 57 *
 58 * Must be a prime to step through all indexes of the buffer.
 59 *
 60 * Some primes work better than others on some architectures (from MBA/MBM
 61 * result stability point of view).
 62 */
 63#define FILL_IDX_MULT	23
 64
 65static int fill_one_span_read(unsigned char *buf, size_t buf_size)
 66{
 67	unsigned int size = buf_size / (CL_SIZE / 2);
 68	unsigned int i, idx = 0;
 69	unsigned char sum = 0;
 70
 71	/*
 72	 * Read the buffer in an order that is unexpected by HW prefetching
 73	 * optimizations to prevent them interfering with the caching pattern.
 74	 *
 75	 * The read order is (in terms of halves of cachelines):
 76	 *	i * FILL_IDX_MULT % size
 77	 * The formula is open-coded below to avoiding modulo inside the loop
 78	 * as it improves MBA/MBM result stability on some architectures.
 79	 */
 80	for (i = 0; i < size; i++) {
 81		sum += buf[idx * (CL_SIZE / 2)];
 82
 83		idx += FILL_IDX_MULT;
 84		while (idx >= size)
 85			idx -= size;
 86	}
 87
 88	return sum;
 89}
 90
 91static void fill_one_span_write(unsigned char *buf, size_t buf_size)
 92{
 93	unsigned char *end_ptr = buf + buf_size;
 94	unsigned char *p;
 95
 96	p = buf;
 97	while (p < end_ptr) {
 98		*p = '1';
 99		p += (CL_SIZE / 2);
100	}
101}
102
103void fill_cache_read(unsigned char *buf, size_t buf_size, bool once)
104{
105	int ret = 0;
106
107	while (1) {
108		ret = fill_one_span_read(buf, buf_size);
109		if (once)
110			break;
111	}
112
113	/* Consume read result so that reading memory is not optimized out. */
114	*value_sink = ret;
115}
116
117static void fill_cache_write(unsigned char *buf, size_t buf_size, bool once)
118{
119	while (1) {
120		fill_one_span_write(buf, buf_size);
121		if (once)
122			break;
123	}
124}
125
126unsigned char *alloc_buffer(size_t buf_size, int memflush)
127{
128	void *buf = NULL;
129	uint64_t *p64;
130	size_t s64;
131	int ret;
132
133	ret = posix_memalign(&buf, PAGE_SIZE, buf_size);
134	if (ret < 0)
135		return NULL;
136
137	/* Initialize the buffer */
138	p64 = buf;
139	s64 = buf_size / sizeof(uint64_t);
140
141	while (s64 > 0) {
142		*p64 = (uint64_t)rand();
143		p64 += (CL_SIZE / sizeof(uint64_t));
144		s64 -= (CL_SIZE / sizeof(uint64_t));
145	}
146
147	/* Flush the memory before using to avoid "cache hot pages" effect */
148	if (memflush)
149		mem_flush(buf, buf_size);
150
151	return buf;
152}
153
154int run_fill_buf(size_t buf_size, int memflush, int op, bool once)
155{
156	unsigned char *buf;
157
158	buf = alloc_buffer(buf_size, memflush);
159	if (!buf)
160		return -1;
161
162	if (op == 0)
163		fill_cache_read(buf, buf_size, once);
164	else
165		fill_cache_write(buf, buf_size, once);
166	free(buf);
167
168	return 0;
169}