openwrt/target/linux/bcm27xx/patches-6.6/950-0910-perf-raspberry-Add-support-for-2712-axi-performance-.patch

From 7b5e845f3243afd393ede5ca0e5de310115ccf30 Mon Sep 17 00:00:00 2001
From: Dom Cobley <popcornmix@gmail.com>
Date: Thu, 8 Jun 2023 11:33:08 +0100
Subject: [PATCH 0910/1085] perf/raspberry: Add support for 2712 axi
 performance monitors

Also handle 2711 correctly which has a different configuration
from 2835.

Signed-off-by: Dom Cobley <popcornmix@gmail.com>
---
 drivers/perf/raspberrypi_axi_monitor.c | 257 ++++++++++++++++++++++---
 1 file changed, 225 insertions(+), 32 deletions(-)

--- a/drivers/perf/raspberrypi_axi_monitor.c
+++ b/drivers/perf/raspberrypi_axi_monitor.c
@@ -33,7 +33,7 @@
 #define MAX_BUSES 16
 #define DEFAULT_SAMPLE_TIME 100
 
-#define NUM_BUS_WATCHER_RESULTS 9
+#define NUM_BUS_WATCHER_RESULTS 11
 
 struct bus_watcher_data {
 	union	{
@@ -48,6 +48,8 @@ struct bus_watcher_data {
 			u32 rtrans;
 			u32 rtwait;
 			u32 rmax;
+			u32 rpend;
+			u32 ratrans;
 		};
 	};
 };
@@ -65,6 +67,9 @@ struct rpi_axiperf {
 	/* Sample time spent on for each bus */
 	int sample_time;
 
+	/* chip specific bus config */
+	const struct bwconfig_config *config;
+
 	/* Now storage for the per monitor settings and the resulting
 	 * performance figures
 	 */
@@ -107,6 +112,7 @@ const int GEN_CTRL;
 
 const int GEN_CTL_ENABLE_BIT	= BIT(0);
 const int GEN_CTL_RESET_BIT	= BIT(1);
+const int GEN_CTL_WATCH_BIT	= BIT(2);
 
 /* Bus watcher registers */
 const int BW_PITCH		= 0x40;
@@ -136,7 +142,7 @@ const int BW_CTRL_BUS_WATCH_SHIFT;
 const int BW_CTRL_BUS_WATCH_MASK = GENMASK(5, 0); // 6 bits
 const int BW_CTRL_BUS_FILTER_SHIFT = 8;
 
-const static char *bus_filter_strings[] = {
+static const char *bus_filter_strings[] = {
 	"",
 	"CORE0_V",
 	"ICACHE0",
@@ -171,9 +177,96 @@ const static char *bus_filter_strings[]
 	"M30"
 };
 
-const int num_bus_filters = ARRAY_SIZE(bus_filter_strings);
+static const char * const bus_filter_strings_2711[] = {
+	"AIO",
+	"CORE0_V",
+	"ICACHE0",
+	"DCACHE0",
+	"CORE1_V",
+	"ICACHE1",
+	"DCACHE1",
+	"L2_MAIN",
+	"ARGON",
+	"PCIE",
+	"HVS",
+	"ISP",
+	"VIDEO_DCT",
+	"VIDEO_SD2AXI",
+	"CAM0",
+	"CAM1",
+	"DMA0",
+	"DMA1",
+	"DMA2",
+	"JPEG",
+	"VIDEO_CME",
+	"TRANSPOSER",
+	"VIDEO_FME",
+	"GIGE",
+	"USB",
+	"V3D0",
+	"V3D1",
+	"V3D2",
+	"GISB_AXI",
+	"DEBUG",
+	"ARM",
+	"EMMCSTB",
+};
 
-const static char *system_bus_string[] = {
+static const char * const bus_filter_strings_2712[] = {
+	"",
+	"VPU_UC0",
+	"VPU_IC0",
+	"VPU_DC0",
+	"VPU_UC1",
+	"VPU_IC1",
+	"VPU_DC1",
+	"VPU_L2",
+	"DMA2",
+	"VPU_DEBUG",
+	"ARM",
+	"DMA0",
+	"DMA1",
+	"RAAGA",
+	"BBSI",
+	"PCIE0",
+	"PCIE1",
+	"PCIE2",
+	"UMR",
+	"SAGE",
+	"HVDP",
+	"BSP",
+	"HVS",
+	"HVS_WMK",
+	"MOP0",
+	"MOP1",
+	"MBVN",
+	"DSI",
+	"XPT",
+	"EMMC0",
+	"GENET",
+	"USB",
+	"ARGON",
+	"UNICAM",
+	"PISP",
+	"PISPFE",
+	"JPEG",
+	"EMMC1",
+	"EMMC2",
+	"TRC",
+	"BSTM0",
+	"BSTM1",
+	"BSTM0_SEC",
+	"BSTM1_SEC",
+	"AIO",
+	"MAP",
+	"SYS_DMA",
+	"MMUCACHE0",
+	"MMUCACHE1",
+	"MPUCACHE0",
+	"MPUCACHE1",
+};
+
+static const char *system_bus_string[] = {
 	"DMA_L2",
 	"TRANS",
 	"JPEG",
@@ -192,9 +285,38 @@ const static char *system_bus_string[] =
 	"CPU_L2"
 };
 
-const int num_system_buses = ARRAY_SIZE(system_bus_string);
+static const char * const system_bus_string_2711[] = {
+	"DMA_L2",
+	"TRANS",
+	"JPEG",
+	"VPU_UC",
+	"DMA_UC",
+	"SYSTEM_L2",
+	"HVS",
+	"ARGON",
+	"H264",
+	"PERIPHERAL",
+	"ARM_UC",
+	"ARM_L2",
+};
+
+static const char * const system_bus_string_2712[] = {
+	"VPU_UC",
+	"DISPLAY_TOP",
+	"V3D",
+	"ARM",
+	"XPT",
+	"BSTM_TOP",
+	"PCIE_01",
+	"ARGON_TOP",
+	"ARB3",
+	"SRC",
+	"HVDP",
+	"PER",
+	"SYSTEM_L2",
+};
 
-const static char *vpu_bus_string[] = {
+static const char *vpu_bus_string[] = {
 	"VPU1_D_L2",
 	"VPU0_D_L2",
 	"VPU1_I_L2",
@@ -213,7 +335,66 @@ const static char *vpu_bus_string[] = {
 	"L2_IN"
 };
 
-const int num_vpu_buses = ARRAY_SIZE(vpu_bus_string);
+static const char * const vpu_bus_string_2711[] = {
+	"VPU1_D_L2",
+	"VPU0_D_L2",
+	"VPU1_I_L2",
+	"VPU0_I_L2",
+	"SYSTEM_L2",
+	"DMA_L2",
+	"VPU1_D_UC",
+	"VPU0_D_UC",
+	"VPU1_I_UC",
+	"VPU0_I_UC",
+	"VPU_UC",
+	"L2_OUT",
+	"DMA_UC",
+	"L2_IN"
+};
+
+static const char * const vpu_bus_string_2712[] = {
+	"VPU1_D_L2",
+	"VPU0_D_L2",
+	"VPU1_I_L2",
+	"VPU0_I_L2",
+	"SYSTEM_L2",
+	"DMA_L2",
+	"VPU1_D_UC",
+	"VPU0_D_UC",
+	"VPU1_I_UC",
+	"VPU0_I_UC",
+	"VPU_UC",
+	"L2_OUT",
+	"DMA_UC",
+	"L2_IN"
+};
+
+struct bwconfig_config {
+	const char * const *bus_filter_strings;
+	const int num_bus_filters;
+	const char * const *system_bus_string;
+	const int num_system_buses;
+	const char * const *vpu_bus_string;
+	const int num_vpu_buses;
+};
+
+static const struct bwconfig_config config_2835 = {
+	bus_filter_strings, ARRAY_SIZE(bus_filter_strings),
+	system_bus_string, ARRAY_SIZE(system_bus_string),
+	vpu_bus_string, ARRAY_SIZE(vpu_bus_string),
+};
+
+static const struct bwconfig_config config_2711 = {
+	bus_filter_strings_2711, ARRAY_SIZE(bus_filter_strings_2711),
+	system_bus_string_2711, ARRAY_SIZE(system_bus_string_2711),
+	vpu_bus_string_2711, ARRAY_SIZE(vpu_bus_string_2711),
+};
+
+static const struct bwconfig_config config_2712 = {
+	bus_filter_strings_2712, ARRAY_SIZE(bus_filter_strings_2712),
+	system_bus_string_2712, ARRAY_SIZE(system_bus_string_2712),
+	vpu_bus_string_2712, ARRAY_SIZE(vpu_bus_string_2712),
+};
 
 const static char *monitor_name[] = {
 	"System",
@@ -233,10 +414,10 @@ static inline u32 read_reg(int monitor,
 static void read_bus_watcher(int monitor, int watcher, u32 *results)
 {
 	if (state->monitor[monitor].use_mailbox_interface) {
-		/* We have 9 results, plus the overheads of start address and
-		 * length So 11 u32 to define
+		/* We have NUM_BUS_WATCHER_RESULTS results, plus the overheads
+		 * of start address and length
 		 */
-		u32 tmp[11];
+		u32 tmp[NUM_BUS_WATCHER_RESULTS+2];
 		int err;
 
 		tmp[0] = (u32)(uintptr_t)(state->monitor[monitor].base_address + watcher
@@ -352,7 +533,7 @@ static void monitor(struct rpi_axiperf *
 		}
 
 		/* start monitoring */
-		set_monitor_control(monitor, GEN_CTL_ENABLE_BIT);
+		set_monitor_control(monitor, GEN_CTL_ENABLE_BIT | GEN_CTL_WATCH_BIT);
 	}
 
 	mutex_unlock(&state->lock);
@@ -409,11 +590,12 @@ static ssize_t myreader(struct file *fp,
 	int buff_size = INIT_BUFF_SIZE;
 	char *p;
 	typeof(state->monitor[0]) *mon = &(state->monitor[idx]);
+	const struct bwconfig_config *config = state->config;
 
 	if (idx < 0 || idx > NUM_MONITORS)
 		idx = 0;
 
-	num_buses = idx == SYSTEM_MONITOR ? num_system_buses : num_vpu_buses;
+	num_buses = idx == SYSTEM_MONITOR ? config->num_system_buses : config->num_vpu_buses;
 
 	string_buffer = kmalloc(buff_size, GFP_KERNEL);
 
@@ -428,17 +610,17 @@ static ssize_t myreader(struct file *fp,
 	mutex_lock(&state->lock);
 
 	if (mon->bus_filter) {
-		int filt = min(mon->bus_filter & 0x1f, num_bus_filters);
+		int filt = min(mon->bus_filter & 0x1f, config->num_bus_filters);
 
 		cnt = snprintf(p, buff_size,
 			       "\nMonitoring transactions from %s only\n",
-			       bus_filter_strings[filt]);
+			       config->bus_filter_strings[filt]);
 		p += cnt;
 		buff_size -= cnt;
 	}
 
-	cnt = snprintf(p, buff_size, "     Bus   |    Atrans    Atwait      AMax    Wtrans    Wtwait      WMax    Rtrans    Rtwait      RMax\n"
-				     "======================================================================================================\n");
+	cnt = snprintf(p, buff_size, "     Bus    |    Atrans    Atwait      AMax    Wtrans    Wtwait      WMax    Rtrans    Rtwait      RMax     RPend   RAtrans\n"
+				     "===========================================================================================================================\n");
 
 	if (cnt >= buff_size)
 		goto done;
@@ -446,25 +628,29 @@ static ssize_t myreader(struct file *fp,
 	p += cnt;
 	buff_size -= cnt;
 
+#define M(x) ((x) >= 1000000000 ? (x)/1000000 : (x) >= 1000 ? (x)/1000 : (x))
+#define N(x) ((x) >= 1000000000 ? 'M' : (x) >= 1000 ? 'K' : ' ')
+
 	for (i = 0; i < num_buses; i++) {
 		if (mon->bus_enabled & (1 << i)) {
-#define DIVIDER (1024)
 			typeof(mon->results[0]) *res = &(mon->results[i]);
 
 			cnt = snprintf(p, buff_size,
-					"%10s | %8uK %8uK %8uK %8uK %8uK %8uK %8uK %8uK %8uK\n",
+					"%11s | %8u%c %8u%c %8u%c %8u%c %8u%c %8u%c %8u%c %8u%c %8u%c %8u%c %8u%c\n",
 					idx == SYSTEM_MONITOR ?
-						system_bus_string[i] :
-						vpu_bus_string[i],
-					res->atrans/DIVIDER,
-					res->atwait/DIVIDER,
-					res->amax/DIVIDER,
-					res->wtrans/DIVIDER,
-					res->wtwait/DIVIDER,
-					res->wmax/DIVIDER,
-					res->rtrans/DIVIDER,
-					res->rtwait/DIVIDER,
-					res->rmax/DIVIDER
+						config->system_bus_string[i] :
+						config->vpu_bus_string[i],
+					M(res->atrans), N(res->atrans),
+					M(res->atwait), N(res->atwait),
+					M(res->amax), N(res->amax),
+					M(res->wtrans), N(res->wtrans),
+					M(res->wtwait), N(res->wtwait),
+					M(res->wmax), N(res->wmax),
+					M(res->rtrans), N(res->rtrans),
+					M(res->rtwait), N(res->rtwait),
+					M(res->rmax), N(res->rmax),
+					M(res->rpend), N(res->rpend),
+					M(res->ratrans), N(res->ratrans)
 					);
 			if (cnt >= buff_size)
 				goto done;
@@ -526,6 +712,10 @@ static int rpi_axiperf_probe(struct plat
 	if (!state)
 		return -ENOMEM;
 
+	state->config = of_device_get_match_data(dev);
+	if (!state->config)
+		return -EINVAL;
+
 	/* Get the firmware handle for future rpi-firmware-xxx calls */
 	fw_node = of_parse_phandle(np, "firmware", 0);
 	if (!fw_node) {
@@ -612,9 +802,12 @@ static int rpi_axiperf_remove(struct pla
 }
 
 static const struct of_device_id rpi_axiperf_match[] = {
-	{
-		.compatible = "brcm,bcm2835-axiperf",
-	},
+	{ .compatible = "brcm,bcm2835-axiperf",
+	  .data = &config_2835 },
+	{ .compatible = "brcm,bcm2711-axiperf",
+	  .data = &config_2711 },
+	{ .compatible = "brcm,bcm2712-axiperf",
+	  .data = &config_2712 },
 	{},
 };
 MODULE_DEVICE_TABLE(of, rpi_axiperf_match);