Test dcache CMOs before copying memory

Signed-off-by: Oliver Upton <oliver.upton@linux.dev>

2 files changed, 37 insertions, 2 deletions

diff --git a/Makefile b/Makefile
index c4b94f2..9f3902f 100644
--- a/Makefile
+++ b/Makefile
@@ -1,5 +1,5 @@
 CC	:= aarch64-linux-gnu-gcc
-CFLAGS	+= -O3 -Wall -Werror
+CFLAGS	+= -O3 -Wall -Werror -static
 
 OBJS	= main.o
 OBJS	+= memcpy_ldp_str.o
diff --git a/main.c b/main.c
index 6f28f68..315ff3a 100644
--- a/main.c
+++ b/main.c
@@ -21,6 +21,7 @@ enum mode {
 	LDTR_STP,
 	LDP_STP,
 	PRFM_2K_LDP_STR,
+	LDP_STP_DC_CIVAC,
 
 	NR_MODES
 };
@@ -42,6 +43,7 @@ static void pr_modes(void)
 	PR_MODE(LDTR_STP);
 	PR_MODE(LDP_STP);
 	PR_MODE(PRFM_2K_LDP_STR);
+	PR_MODE(LDP_STP_DC_CIVAC);
 }
 
 static void pr_help(const char *progname)
@@ -76,6 +78,17 @@ static void destroy_test_buffer(void *buf)
 	munmap(buf, test_size);
 }
 
+static void dcache_clean_invalidate(const void *buf, size_t size)
+{
+	const void *addr;
+
+	for (addr = buf; addr < buf + size; addr += 64)
+		asm volatile("dc civac, %0" : : "r"(addr)
+			     : "memory");
+
+	asm volatile("dsb ish");
+}
+
 static void do_memcpy(void *dst, const void *src, size_t chunk_size)
 {
 	switch (test_mode) {
@@ -92,6 +105,7 @@ static void do_memcpy(void *dst, const void *src, size_t chunk_size)
 		memcpy_ldtr_stp(dst, src, chunk_size);
 		break;
 	case LDP_STP:
+	case LDP_STP_DC_CIVAC:
 		memcpy_ldp_stp(dst, src, chunk_size);
 		break;
 	case PRFM_2K_LDP_STR:
@@ -103,6 +117,8 @@ static void do_memcpy(void *dst, const void *src, size_t chunk_size)
 }
 
 #define NSEC_PER_SEC	1000000000UL
+#define NSEC_PER_MSEC	1000000UL
+#define MSEC_PER_SEC	1000UL
 
 static void report_bandwidth(struct timespec *start, struct timespec *end)
 {
@@ -114,11 +130,21 @@ static void report_bandwidth(struct timespec *start, struct timespec *end)
 	printf("Rate:	%.03f GiB/sec\n", rate);
 }
 
+static void report_dcache_time(struct timespec *start, struct timespec *end)
+{
+	unsigned long msecs;
+
+	msecs = (end->tv_sec - start->tv_sec) * MSEC_PER_SEC;
+	msecs = (end->tv_nsec - start->tv_nsec) / NSEC_PER_MSEC;
+
+	printf("cache maintenance took %lu milliseconds\n", msecs);
+}
+
 static void run_test(void)
 {
 	void *dst = setup_test_buffer();
 	void *src = setup_test_buffer();
-	struct timespec start, end;
+	struct timespec start, end, dc_start, dc_end;
 	size_t offset, i;
 
 	printf("Iters:	%lu\n", test_iterations);
@@ -129,6 +155,12 @@ static void run_test(void)
 	assert(!clock_gettime(CLOCK_MONOTONIC, &start));
 
 	for (i = 0; i < test_iterations; i++) {
+		if (test_mode == LDP_STP_DC_CIVAC) {
+			assert(!clock_gettime(CLOCK_MONOTONIC, &dc_start));
+			dcache_clean_invalidate(src, test_size);
+			assert(!clock_gettime(CLOCK_MONOTONIC, &dc_end));
+		}
+
 		for (offset = 0; offset < test_size; offset += chunk_size) {
 			do_memcpy(dst + offset, src + offset, chunk_size);
 			assert(!memchr(dst + offset, 0, chunk_size));
@@ -139,6 +171,9 @@ static void run_test(void)
 
 	report_bandwidth(&start, &end);
 
+	if (test_mode == LDP_STP_DC_CIVAC)
+		report_dcache_time(&dc_start, &dc_end);
+
 	destroy_test_buffer(dst);
 	destroy_test_buffer(src);
 }