[PATCH 0/3] platform/x86/amd/hsmp: add metrics table support

[PATCH 0/3] platform/x86/amd/hsmp: add metrics table support

[Suma Hegde]

From: Suma Hegde <suma.hegde@amd.com>

AMD MI300A processor supports metrics table, a structure of data shared by the SMU to the CPU,
Which provides a bunch of information in one go, this reduces the latency in telemetry and
adds atomicity to the information.
The following modifications and additions are needed to support reading the metrics table.

1. Create a separate platform structure for platform specific data (such as socket index and lock).
2. Create binary sysfs file for metrics table which can be read by userspace and decoded using
   hsmp_metric_table structure defined in uapi header.
   	2.a E-SMI library and esmi_tool [https://github.com/amd/esmi_ib_library]
	    will be updated with API and option respectively.
3. A general patch to modify logs to specify failure reasons.

Suma Hegde (3):
  platform/x86/amd/hsmp: create plat specific struct
  platform/x86/amd/hsmp: add support for metrics tbl
  platform/x86/amd/hsmp: change the print message

 arch/x86/include/uapi/asm/amd_hsmp.h | 109 ++++++++++++++
 drivers/platform/x86/amd/hsmp.c      | 206 +++++++++++++++++++++++----
 2 files changed, 291 insertions(+), 24 deletions(-)

-- 
2.25.1

[PATCH 1/4] platform/x86/amd/hsmp: create plat specific struct

[Suma Hegde]

From: Suma Hegde <suma.hegde@amd.com>

Having a separate platform device structure helps in future, to
contain platform specific variables and other data.

Signed-off-by: Suma Hegde <suma.hegde@amd.com>
Reviewed-by: Naveen Krishna Chatradhi <nchatrad@amd.com>
---
 drivers/platform/x86/amd/hsmp.c | 53 ++++++++++++++++++++-------------
 1 file changed, 33 insertions(+), 20 deletions(-)

diff --git a/drivers/platform/x86/amd/hsmp.c b/drivers/platform/x86/amd/hsmp.c
index 31382ef52efb..699c22f7cd42 100644
--- a/drivers/platform/x86/amd/hsmp.c
+++ b/drivers/platform/x86/amd/hsmp.c
@@ -47,9 +47,20 @@
 #define HSMP_INDEX_REG		0xc4
 #define HSMP_DATA_REG		0xc8
 
-static struct semaphore *hsmp_sem;
+struct hsmp_socket {
+	struct semaphore hsmp_sem;
+	u16 sock_ind;
+};
+
+struct hsmp_plat_device {
+	struct miscdevice hsmp_device;
+	struct hsmp_socket *sock;
+	struct device *dev;
+};
 
-static struct miscdevice hsmp_device;
+static u16 num_sockets;
+
+static struct hsmp_plat_device plat_dev;
 
 static int amd_hsmp_rdwr(struct pci_dev *root, u32 address,
 			 u32 *value, bool write)
@@ -188,6 +199,7 @@ static int validate_message(struct hsmp_message *msg)
 
 int hsmp_send_message(struct hsmp_message *msg)
 {
+	struct hsmp_socket *sock = &plat_dev.sock[msg->sock_ind];
 	struct amd_northbridge *nb;
 	int ret;
 
@@ -208,14 +220,13 @@ int hsmp_send_message(struct hsmp_message *msg)
 	 * In SMP system timeout of 100 millisecs should
 	 * be enough for the previous thread to finish the operation
 	 */
-	ret = down_timeout(&hsmp_sem[msg->sock_ind],
-			   msecs_to_jiffies(HSMP_MSG_TIMEOUT));
+	ret = down_timeout(&sock->hsmp_sem, msecs_to_jiffies(HSMP_MSG_TIMEOUT));
 	if (ret < 0)
 		return ret;
 
 	ret = __hsmp_send_message(nb->root, msg);
 
-	up(&hsmp_sem[msg->sock_ind]);
+	up(&sock->hsmp_sem);
 
 	return ret;
 }
@@ -321,28 +332,31 @@ static int hsmp_pltdrv_probe(struct platform_device *pdev)
 {
 	int i;
 
-	hsmp_sem = devm_kzalloc(&pdev->dev,
-				(amd_nb_num() * sizeof(struct semaphore)),
-				GFP_KERNEL);
-	if (!hsmp_sem)
+	plat_dev.sock = devm_kzalloc(&pdev->dev,
+				     (num_sockets * sizeof(struct hsmp_socket)),
+				     GFP_KERNEL);
+	if (!plat_dev.sock)
 		return -ENOMEM;
+	plat_dev.dev = &pdev->dev;
 
-	for (i = 0; i < amd_nb_num(); i++)
-		sema_init(&hsmp_sem[i], 1);
+	for (i = 0; i < num_sockets; i++) {
+		sema_init(&plat_dev.sock[i].hsmp_sem, 1);
+		plat_dev.sock[i].sock_ind = i;
+	}
 
-	hsmp_device.name	= "hsmp_cdev";
-	hsmp_device.minor	= MISC_DYNAMIC_MINOR;
-	hsmp_device.fops	= &hsmp_fops;
-	hsmp_device.parent	= &pdev->dev;
-	hsmp_device.nodename	= "hsmp";
-	hsmp_device.mode	= 0644;
+	plat_dev.hsmp_device.name	= "hsmp_cdev";
+	plat_dev.hsmp_device.minor	= MISC_DYNAMIC_MINOR;
+	plat_dev.hsmp_device.fops	= &hsmp_fops;
+	plat_dev.hsmp_device.parent	= &pdev->dev;
+	plat_dev.hsmp_device.nodename	= "hsmp";
+	plat_dev.hsmp_device.mode	= 0644;
 
-	return misc_register(&hsmp_device);
+	return misc_register(&plat_dev.hsmp_device);
 }
 
 static void hsmp_pltdrv_remove(struct platform_device *pdev)
 {
-	misc_deregister(&hsmp_device);
+	misc_deregister(&plat_dev.hsmp_device);
 }
 
 static struct platform_driver amd_hsmp_driver = {
@@ -358,7 +372,6 @@ static struct platform_device *amd_hsmp_platdev;
 static int __init hsmp_plt_init(void)
 {
 	int ret = -ENODEV;
-	u16 num_sockets;
 	int i;
 
 	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD || boot_cpu_data.x86 < 0x19) {
-- 
2.25.1

[PATCH 2/4] platform/x86/amd/hsmp: add support for metrics tbl

[Suma Hegde]

From: Suma Hegde <suma.hegde@amd.com>

AMD MI300 MCM provides GET_METRICS_TABLE message with which
all the system management information from SMU can be retrieved in just
one message.

The metrics table is available as hexadecimal sysfs binary file in
/sys/devices/platform/amd_hsmp/socket%d_metrics_bin 
Metrics table definitions will be documented as part of PPR which
available in public domain. The same is defined in the amd_hsmp.h header
file as well.

Signed-off-by: Suma Hegde <suma.hegde@amd.com>
Reviewed-by: Naveen Krishna Chatradhi <nchatrad@amd.com>
---
 arch/x86/include/uapi/asm/amd_hsmp.h | 109 +++++++++++++++++++
 drivers/platform/x86/amd/hsmp.c      | 155 ++++++++++++++++++++++++++-
 2 files changed, 259 insertions(+), 5 deletions(-)

diff --git a/arch/x86/include/uapi/asm/amd_hsmp.h b/arch/x86/include/uapi/asm/amd_hsmp.h
index 769b939444ae..832701b41c7d 100644
--- a/arch/x86/include/uapi/asm/amd_hsmp.h
+++ b/arch/x86/include/uapi/asm/amd_hsmp.h
@@ -47,6 +47,9 @@ enum hsmp_message_ids {
 	HSMP_SET_PCI_RATE,		/* 20h Control link rate on PCIe devices */
 	HSMP_SET_POWER_MODE,		/* 21h Select power efficiency profile policy */
 	HSMP_SET_PSTATE_MAX_MIN,	/* 22h Set the max and min DF P-State  */
+	HSMP_GET_METRIC_TABLE_VER,	/* 23h Get metrics table version  */
+	HSMP_GET_METRIC_TABLE,		/* 24h Get metrics table  */
+	HSMP_GET_METRIC_TABLE_DRAM_ADDR,/* 25h Get metrics table dram address */
 	HSMP_MSG_ID_MAX,
 };
 
@@ -64,6 +67,14 @@ enum hsmp_msg_type {
 	HSMP_GET  = 1,
 };
 
+enum hsmp_proto_versions {
+	HSMP_PROTO_VER2	= 2,
+	HSMP_PROTO_VER3,
+	HSMP_PROTO_VER4,
+	HSMP_PROTO_VER5,
+	HSMP_PROTO_VER6
+};
+
 struct hsmp_msg_desc {
 	int num_args;
 	int response_sz;
@@ -295,6 +306,104 @@ static const struct hsmp_msg_desc hsmp_msg_desc_table[] = {
 	 * input: args[0] = min df pstate[15:8] + max df pstate[7:0]
 	 */
 	{1, 0, HSMP_SET},
+
+	/*
+	 * HSMP_GET_METRIC_TABLE_VER, num_args = 0, response_sz = 1
+	 * output: args[0] = metrics table version
+	 */
+	{0, 1, HSMP_GET},
+
+	/*
+	 * HSMP_GET_METRIC_TABLE, num_args = 0, response_sz = 0
+	 */
+	{0, 0, HSMP_GET},
+
+	/*
+	 * HSMP_GET_METRIC_TABLE_DRAM_ADDR, num_args = 0, response_sz = 2
+	 * output: args[0] = lower 32 bits of the address
+	 * output: args[1] = upper 32 bits of the address
+	 */
+	{0, 2, HSMP_GET},
+};
+
+/* Metrics table for EPYC socket(supported only from proto version 6) */
+struct hsmp_metric_table {
+	__u32 accumulation_counter;
+
+	//TEMPERATURE
+	__u32 max_socket_temperature;
+	__u32 max_vr_temperature;
+	__u32 max_hbm_temperature;
+	__u64 max_socket_temperature_acc;
+	__u64 max_vr_temperature_acc;
+	__u64 max_hbm_temperature_acc;
+
+	//POWER
+	__u32 socket_power_limit;
+	__u32 max_socket_power_limit;
+	__u32 socket_power;
+
+	//ENERGY
+	__u64 timestamp;
+	__u64 socket_energy_acc;
+	__u64 ccd_energy_acc;
+	__u64 xcd_energy_acc;
+	__u64 aid_energy_acc;
+	__u64 hbm_energy_acc;
+
+	//FREQUENCY
+	__u32 cclk_frequency_limit;
+	__u32 gfxclk_frequency_limit;
+	__u32 fclk_frequency;
+	__u32 uclk_frequency;
+	__u32 socclk_frequency[4];
+	__u32 vclk_frequency[4];
+	__u32 dclk_frequency[4];
+	__u32 lclk_frequency[4];
+	__u64 gfxclk_frequency_acc[8];
+	__u64 cclk_frequency_acc[96];
+
+	//FREQUENCY RANGE
+	__u32 max_cclk_frequency;
+	__u32 min_cclk_frequency;
+	__u32 max_gfxclk_frequency;
+	__u32 min_gfxclk_frequency;
+	__u32 fclk_frequency_table[4];
+	__u32 uclk_frequency_table[4];
+	__u32 socclk_frequency_table[4];
+	__u32 vclk_frequency_table[4];
+	__u32 dclk_frequency_table[4];
+	__u32 lclk_frequency_table[4];
+	__u32 max_lclk_dpm_range;
+	__u32 min_lclk_dpm_range;
+
+	//XGMI
+	__u32 xgmi_width;
+	__u32 xgmi_bitrate;
+	__u64 xgmi_read_bandwidth_acc[8];
+	__u64 xgmi_write_bandwidth_acc[8];
+
+	//ACTIVITY
+	__u32 socket_c0_residency;
+	__u32 socket_gfx_busy;
+	__u32 dram_bandwidth_utilization;
+	__u64 socket_c0_residency_acc;
+	__u64 socket_gfx_busy_acc;
+	__u64 dram_bandwidth_acc;
+	__u32 max_dram_bandwidth;
+	__u64 dram_bandwidth_utilization_acc;
+	__u64 pcie_bandwidth_acc[4];
+
+	//THROTTLERS
+	__u32 prochot_residency_acc;
+	__u32 ppt_residency_acc;
+	__u32 socket_thm_residency_acc;
+	__u32 vr_thm_residency_acc;
+	__u32 hbm_thm_residency_acc;
+	__u32 spare;
+
+	// New Items at end to maintain driver compatibility
+	__u32 gfxclk_frequency[8];
 };
 
 /* Reset to default packing */
diff --git a/drivers/platform/x86/amd/hsmp.c b/drivers/platform/x86/amd/hsmp.c
index 699c22f7cd42..cab59750cde2 100644
--- a/drivers/platform/x86/amd/hsmp.c
+++ b/drivers/platform/x86/amd/hsmp.c
@@ -20,7 +20,7 @@
 #include <linux/semaphore.h>
 
 #define DRIVER_NAME		"amd_hsmp"
-#define DRIVER_VERSION		"1.0"
+#define DRIVER_VERSION		"2.0"
 
 /* HSMP Status / Error codes */
 #define HSMP_STATUS_NOT_READY	0x00
@@ -47,7 +47,12 @@
 #define HSMP_INDEX_REG		0xc4
 #define HSMP_DATA_REG		0xc8
 
+#define HSMP_CDEV_NAME		"hsmp_cdev"
+#define HSMP_DEVNODE_NAME	"hsmp"
+
 struct hsmp_socket {
+	struct bin_attribute hsmp_attr;
+	void __iomem *metric_tbl_addr;
 	struct semaphore hsmp_sem;
 	u16 sock_ind;
 };
@@ -56,6 +61,7 @@ struct hsmp_plat_device {
 	struct miscdevice hsmp_device;
 	struct hsmp_socket *sock;
 	struct device *dev;
+	u32 proto_ver;
 };
 
 static u16 num_sockets;
@@ -328,9 +334,122 @@ static const struct file_operations hsmp_fops = {
 	.compat_ioctl	= hsmp_ioctl,
 };
 
+static ssize_t hsmp_metric_tbl_read(struct file *filp, struct kobject *kobj,
+				    struct bin_attribute *bin_attr, char *buf,
+				    loff_t off, size_t count)
+{
+	struct hsmp_socket *sock = bin_attr->private;
+	struct hsmp_message msg = { 0 };
+	int ret;
+
+	if (count == 0 || count > sizeof(struct hsmp_metric_table))
+		return 0;
+
+	msg.msg_id	= HSMP_GET_METRIC_TABLE;
+	msg.sock_ind	= sock->sock_ind;
+
+	ret = hsmp_send_message(&msg);
+	if (ret)
+		return ret;
+	memcpy(buf, sock->metric_tbl_addr, count);
+
+	return count;
+}
+
+#define HSMP_ATRR_NAME_SIZE 25
+static int hsmp_create_metric_tbl_sysfs_file(int sock_ind)
+{
+	struct bin_attribute *hattr = &plat_dev.sock[sock_ind].hsmp_attr;
+	char *name;
+
+	sysfs_attr_init(&plat_dev.sock[sock_ind].hsmp_attr);
+	name = devm_kzalloc(plat_dev.dev, HSMP_ATRR_NAME_SIZE, GFP_KERNEL);
+	if (!name)
+		return -ENOMEM;
+	sprintf(name, "socket%d_metrics_bin", sock_ind);
+	hattr->attr.name	= name;
+	hattr->attr.mode	= 0444;
+	hattr->read		= hsmp_metric_tbl_read;
+	hattr->size		= sizeof(struct hsmp_metric_table);
+	hattr->private		= &plat_dev.sock[sock_ind];
+
+	return device_create_bin_file(plat_dev.dev, hattr);
+}
+
+static int hsmp_get_tbl_dram_base(u16 sock_ind)
+{
+	struct hsmp_socket *sock = &plat_dev.sock[sock_ind];
+	struct hsmp_message msg = { 0 };
+	phys_addr_t dram_addr;
+	int ret;
+
+	msg.sock_ind	= sock_ind;
+	msg.response_sz	= hsmp_msg_desc_table[HSMP_GET_METRIC_TABLE_DRAM_ADDR].response_sz;
+	msg.msg_id	= HSMP_GET_METRIC_TABLE_DRAM_ADDR;
+
+	ret = hsmp_send_message(&msg);
+	if (ret)
+		return ret;
+
+	/*
+	 * calculate the metric table DRAM address from lower and upper 32 bits
+	 * sent from SMU and ioremap it to virtual address.
+	 */
+	dram_addr = msg.args[0] | ((u64)(msg.args[1]) << 32);
+	if (!dram_addr)
+		return -ENOMEM;
+	sock->metric_tbl_addr = devm_ioremap(plat_dev.dev, dram_addr,
+					     sizeof(struct hsmp_metric_table));
+	if (!sock->metric_tbl_addr) {
+		dev_err(plat_dev.dev, "Unable to ioremap metric table addr\n");
+		return -ENOMEM;
+	}
+	return 0;
+}
+
+static int hsmp_create_sysfs_file(void)
+{
+	int ret, i;
+
+	for (i = 0; i < num_sockets; i++) {
+		ret = hsmp_get_tbl_dram_base(i);
+		if (ret)
+			goto cleanup;
+
+		ret = hsmp_create_metric_tbl_sysfs_file(i);
+		if (ret) {
+			dev_err(plat_dev.dev, "Unable to create sysfs file for metric table\n");
+			goto cleanup;
+		}
+	}
+
+	return 0;
+
+cleanup:
+	while (i > 0)
+		device_remove_bin_file(plat_dev.dev, &plat_dev.sock[--i].hsmp_attr);
+	return ret;
+}
+
+static int hsmp_cache_proto_ver(void)
+{
+	struct hsmp_message msg = { 0 };
+	int ret;
+
+	msg.msg_id	= HSMP_GET_PROTO_VER;
+	msg.sock_ind	= 0;
+	msg.response_sz = hsmp_msg_desc_table[HSMP_GET_PROTO_VER].response_sz;
+
+	ret = hsmp_send_message(&msg);
+	if (!ret)
+		plat_dev.proto_ver = msg.args[0];
+
+	return ret;
+}
+
 static int hsmp_pltdrv_probe(struct platform_device *pdev)
 {
-	int i;
+	int ret, i;
 
 	plat_dev.sock = devm_kzalloc(&pdev->dev,
 				     (num_sockets * sizeof(struct hsmp_socket)),
@@ -344,18 +463,44 @@ static int hsmp_pltdrv_probe(struct platform_device *pdev)
 		plat_dev.sock[i].sock_ind = i;
 	}
 
-	plat_dev.hsmp_device.name	= "hsmp_cdev";
+	plat_dev.hsmp_device.name	= HSMP_CDEV_NAME;
 	plat_dev.hsmp_device.minor	= MISC_DYNAMIC_MINOR;
 	plat_dev.hsmp_device.fops	= &hsmp_fops;
 	plat_dev.hsmp_device.parent	= &pdev->dev;
-	plat_dev.hsmp_device.nodename	= "hsmp";
+	plat_dev.hsmp_device.nodename	= HSMP_DEVNODE_NAME;
 	plat_dev.hsmp_device.mode	= 0644;
 
-	return misc_register(&plat_dev.hsmp_device);
+	ret = misc_register(&plat_dev.hsmp_device);
+	if (ret)
+		return ret;
+
+	ret = hsmp_cache_proto_ver();
+	if (ret) {
+		dev_err(plat_dev.dev, "Failed to read HSMP protocol version\n");
+		goto cleanup;
+	}
+
+	/* metrics table is supported only from proto ver6 EPYCs */
+	if (plat_dev.proto_ver >= HSMP_PROTO_VER6) {
+		ret = hsmp_create_sysfs_file();
+		if (ret)
+			goto cleanup;
+	}
+	return 0;
+
+cleanup:
+	misc_deregister(&plat_dev.hsmp_device);
+	return ret;
 }
 
 static void hsmp_pltdrv_remove(struct platform_device *pdev)
 {
+	int i;
+
+	if (plat_dev.proto_ver >= HSMP_PROTO_VER6) {
+		for (i = 0; i < num_sockets; i++)
+			device_remove_bin_file(plat_dev.dev, &plat_dev.sock[i].hsmp_attr);
+	}
 	misc_deregister(&plat_dev.hsmp_device);
 }
 
-- 
2.25.1

[PATCH 3/4] platform/x86/amd/hsmp: change the print message

[Suma Hegde]

From: Suma Hegde <suma.hegde@amd.com>

The print message during platform init is little ambiguous
so change it to a more meaningful clear message.

Signed-off-by: Suma Hegde <suma.hegde@amd.com>
Reviewed-by: Naveen Krishna Chatradhi <nchatrad@amd.com>
---
 drivers/platform/x86/amd/hsmp.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/drivers/platform/x86/amd/hsmp.c b/drivers/platform/x86/amd/hsmp.c
index cab59750cde2..1f03a22a9d8a 100644
--- a/drivers/platform/x86/amd/hsmp.c
+++ b/drivers/platform/x86/amd/hsmp.c
@@ -537,10 +537,10 @@ static int __init hsmp_plt_init(void)
 	for (i = 0; i < num_sockets; i++) {
 		ret = hsmp_test(i, 0xDEADBEEF);
 		if (ret) {
-			pr_err("HSMP is not supported on Fam:%x model:%x\n",
+			pr_err("HSMP test message failed on Fam:%x model:%x\n",
 			       boot_cpu_data.x86, boot_cpu_data.x86_model);
-			pr_err("Or Is HSMP disabled in BIOS ?\n");
-			return -EOPNOTSUPP;
+			pr_err("Is HSMP disabled in BIOS ?\n");
+			return ret;
 		}
 	}
 
-- 
2.25.1

Re: [PATCH 3/4] platform/x86/amd/hsmp: change the print message

[Ilpo Järvinen]

The coverletter has 0/3 but the patches came with /4?!?

On Fri, 18 Aug 2023, Suma Hegde wrote:

> From: Suma Hegde <suma.hegde@amd.com>
> 
> The print message during platform init is little ambiguous
> so change it to a more meaningful clear message.
> 
> Signed-off-by: Suma Hegde <suma.hegde@amd.com>
> Reviewed-by: Naveen Krishna Chatradhi <nchatrad@amd.com>
> ---
>  drivers/platform/x86/amd/hsmp.c | 6 +++---
>  1 file changed, 3 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/platform/x86/amd/hsmp.c b/drivers/platform/x86/amd/hsmp.c
> index cab59750cde2..1f03a22a9d8a 100644
> --- a/drivers/platform/x86/amd/hsmp.c
> +++ b/drivers/platform/x86/amd/hsmp.c
> @@ -537,10 +537,10 @@ static int __init hsmp_plt_init(void)
>  	for (i = 0; i < num_sockets; i++) {
>  		ret = hsmp_test(i, 0xDEADBEEF);
>  		if (ret) {
> -			pr_err("HSMP is not supported on Fam:%x model:%x\n",
> +			pr_err("HSMP test message failed on Fam:%x model:%x\n",
>  			       boot_cpu_data.x86, boot_cpu_data.x86_model);
> -			pr_err("Or Is HSMP disabled in BIOS ?\n");
> -			return -EOPNOTSUPP;
> +			pr_err("Is HSMP disabled in BIOS ?\n");
> +			return ret;

Your changelog only talks about changing print, and yet you also change 
return statement. Please explain things properly in the changelog (or 
perhaps even make a separate patch out of the return change if it's 
independent change you planned to make).

-- 
 i.

Re: [PATCH 2/4] platform/x86/amd/hsmp: add support for metrics tbl

[Ilpo Järvinen]

On Fri, 18 Aug 2023, Suma Hegde wrote:

> From: Suma Hegde <suma.hegde@amd.com>
> 
> AMD MI300 MCM provides GET_METRICS_TABLE message with which
> all the system management information from SMU can be retrieved in just
> one message.
> 
> The metrics table is available as hexadecimal sysfs binary file in
> /sys/devices/platform/amd_hsmp/socket%d_metrics_bin 
> Metrics table definitions will be documented as part of PPR which
> available in public domain. The same is defined in the amd_hsmp.h header
> file as well.
> 
> Signed-off-by: Suma Hegde <suma.hegde@amd.com>
> Reviewed-by: Naveen Krishna Chatradhi <nchatrad@amd.com>
> ---
>  arch/x86/include/uapi/asm/amd_hsmp.h | 109 +++++++++++++++++++
>  drivers/platform/x86/amd/hsmp.c      | 155 ++++++++++++++++++++++++++-
>  2 files changed, 259 insertions(+), 5 deletions(-)
> 
> diff --git a/arch/x86/include/uapi/asm/amd_hsmp.h b/arch/x86/include/uapi/asm/amd_hsmp.h
> index 769b939444ae..832701b41c7d 100644
> --- a/arch/x86/include/uapi/asm/amd_hsmp.h
> +++ b/arch/x86/include/uapi/asm/amd_hsmp.h
> @@ -47,6 +47,9 @@ enum hsmp_message_ids {
>  	HSMP_SET_PCI_RATE,		/* 20h Control link rate on PCIe devices */
>  	HSMP_SET_POWER_MODE,		/* 21h Select power efficiency profile policy */
>  	HSMP_SET_PSTATE_MAX_MIN,	/* 22h Set the max and min DF P-State  */
> +	HSMP_GET_METRIC_TABLE_VER,	/* 23h Get metrics table version  */
> +	HSMP_GET_METRIC_TABLE,		/* 24h Get metrics table  */

extra spaces before */

> +	HSMP_GET_METRIC_TABLE_DRAM_ADDR,/* 25h Get metrics table dram address */
>  	HSMP_MSG_ID_MAX,
>  };
>  
> @@ -64,6 +67,14 @@ enum hsmp_msg_type {
>  	HSMP_GET  = 1,
>  };
>  
> +enum hsmp_proto_versions {
> +	HSMP_PROTO_VER2	= 2,
> +	HSMP_PROTO_VER3,
> +	HSMP_PROTO_VER4,
> +	HSMP_PROTO_VER5,
> +	HSMP_PROTO_VER6
> +};
> +
>  struct hsmp_msg_desc {
>  	int num_args;
>  	int response_sz;
> @@ -295,6 +306,104 @@ static const struct hsmp_msg_desc hsmp_msg_desc_table[] = {
>  	 * input: args[0] = min df pstate[15:8] + max df pstate[7:0]
>  	 */
>  	{1, 0, HSMP_SET},
> +
> +	/*
> +	 * HSMP_GET_METRIC_TABLE_VER, num_args = 0, response_sz = 1
> +	 * output: args[0] = metrics table version
> +	 */
> +	{0, 1, HSMP_GET},
> +
> +	/*
> +	 * HSMP_GET_METRIC_TABLE, num_args = 0, response_sz = 0
> +	 */
> +	{0, 0, HSMP_GET},
> +
> +	/*
> +	 * HSMP_GET_METRIC_TABLE_DRAM_ADDR, num_args = 0, response_sz = 2
> +	 * output: args[0] = lower 32 bits of the address
> +	 * output: args[1] = upper 32 bits of the address
> +	 */
> +	{0, 2, HSMP_GET},
> +};
> +
> +/* Metrics table for EPYC socket(supported only from proto version 6) */
> +struct hsmp_metric_table {
> +	__u32 accumulation_counter;
> +
> +	//TEMPERATURE
> +	__u32 max_socket_temperature;
> +	__u32 max_vr_temperature;
> +	__u32 max_hbm_temperature;
> +	__u64 max_socket_temperature_acc;
> +	__u64 max_vr_temperature_acc;
> +	__u64 max_hbm_temperature_acc;
> +
> +	//POWER
> +	__u32 socket_power_limit;
> +	__u32 max_socket_power_limit;
> +	__u32 socket_power;
> +
> +	//ENERGY
> +	__u64 timestamp;
> +	__u64 socket_energy_acc;
> +	__u64 ccd_energy_acc;
> +	__u64 xcd_energy_acc;
> +	__u64 aid_energy_acc;
> +	__u64 hbm_energy_acc;
> +
> +	//FREQUENCY
> +	__u32 cclk_frequency_limit;
> +	__u32 gfxclk_frequency_limit;
> +	__u32 fclk_frequency;
> +	__u32 uclk_frequency;
> +	__u32 socclk_frequency[4];
> +	__u32 vclk_frequency[4];
> +	__u32 dclk_frequency[4];
> +	__u32 lclk_frequency[4];
> +	__u64 gfxclk_frequency_acc[8];
> +	__u64 cclk_frequency_acc[96];
> +
> +	//FREQUENCY RANGE
> +	__u32 max_cclk_frequency;
> +	__u32 min_cclk_frequency;
> +	__u32 max_gfxclk_frequency;
> +	__u32 min_gfxclk_frequency;
> +	__u32 fclk_frequency_table[4];
> +	__u32 uclk_frequency_table[4];
> +	__u32 socclk_frequency_table[4];
> +	__u32 vclk_frequency_table[4];
> +	__u32 dclk_frequency_table[4];
> +	__u32 lclk_frequency_table[4];
> +	__u32 max_lclk_dpm_range;
> +	__u32 min_lclk_dpm_range;
> +
> +	//XGMI
> +	__u32 xgmi_width;
> +	__u32 xgmi_bitrate;
> +	__u64 xgmi_read_bandwidth_acc[8];
> +	__u64 xgmi_write_bandwidth_acc[8];
> +
> +	//ACTIVITY
> +	__u32 socket_c0_residency;
> +	__u32 socket_gfx_busy;
> +	__u32 dram_bandwidth_utilization;
> +	__u64 socket_c0_residency_acc;
> +	__u64 socket_gfx_busy_acc;
> +	__u64 dram_bandwidth_acc;
> +	__u32 max_dram_bandwidth;
> +	__u64 dram_bandwidth_utilization_acc;
> +	__u64 pcie_bandwidth_acc[4];
> +
> +	//THROTTLERS
> +	__u32 prochot_residency_acc;
> +	__u32 ppt_residency_acc;
> +	__u32 socket_thm_residency_acc;
> +	__u32 vr_thm_residency_acc;
> +	__u32 hbm_thm_residency_acc;
> +	__u32 spare;
> +
> +	// New Items at end to maintain driver compatibility
> +	__u32 gfxclk_frequency[8];
>  };
>  
>  /* Reset to default packing */
> diff --git a/drivers/platform/x86/amd/hsmp.c b/drivers/platform/x86/amd/hsmp.c
> index 699c22f7cd42..cab59750cde2 100644
> --- a/drivers/platform/x86/amd/hsmp.c
> +++ b/drivers/platform/x86/amd/hsmp.c
> @@ -20,7 +20,7 @@
>  #include <linux/semaphore.h>
>  
>  #define DRIVER_NAME		"amd_hsmp"
> -#define DRIVER_VERSION		"1.0"
> +#define DRIVER_VERSION		"2.0"
>  
>  /* HSMP Status / Error codes */
>  #define HSMP_STATUS_NOT_READY	0x00
> @@ -47,7 +47,12 @@
>  #define HSMP_INDEX_REG		0xc4
>  #define HSMP_DATA_REG		0xc8
>  
> +#define HSMP_CDEV_NAME		"hsmp_cdev"
> +#define HSMP_DEVNODE_NAME	"hsmp"
> +
>  struct hsmp_socket {
> +	struct bin_attribute hsmp_attr;
> +	void __iomem *metric_tbl_addr;
>  	struct semaphore hsmp_sem;
>  	u16 sock_ind;
>  };
> @@ -56,6 +61,7 @@ struct hsmp_plat_device {
>  	struct miscdevice hsmp_device;
>  	struct hsmp_socket *sock;
>  	struct device *dev;
> +	u32 proto_ver;
>  };
>  
>  static u16 num_sockets;
> @@ -328,9 +334,122 @@ static const struct file_operations hsmp_fops = {
>  	.compat_ioctl	= hsmp_ioctl,
>  };
>  
> +static ssize_t hsmp_metric_tbl_read(struct file *filp, struct kobject *kobj,
> +				    struct bin_attribute *bin_attr, char *buf,
> +				    loff_t off, size_t count)
> +{
> +	struct hsmp_socket *sock = bin_attr->private;
> +	struct hsmp_message msg = { 0 };
> +	int ret;
> +
> +	if (count == 0 || count > sizeof(struct hsmp_metric_table))

!count

> +		return 0;
> +
> +	msg.msg_id	= HSMP_GET_METRIC_TABLE;
> +	msg.sock_ind	= sock->sock_ind;
> +
> +	ret = hsmp_send_message(&msg);
> +	if (ret)
> +		return ret;
> +	memcpy(buf, sock->metric_tbl_addr, count);
> +
> +	return count;
> +}
> +
> +#define HSMP_ATRR_NAME_SIZE 25

ATTR (typo)?

> +static int hsmp_create_metric_tbl_sysfs_file(int sock_ind)
> +{
> +	struct bin_attribute *hattr = &plat_dev.sock[sock_ind].hsmp_attr;
> +	char *name;
> +
> +	sysfs_attr_init(&plat_dev.sock[sock_ind].hsmp_attr);
> +	name = devm_kzalloc(plat_dev.dev, HSMP_ATRR_NAME_SIZE, GFP_KERNEL);
> +	if (!name)
> +		return -ENOMEM;
> +	sprintf(name, "socket%d_metrics_bin", sock_ind);

snprintf()

> +	hattr->attr.name	= name;
> +	hattr->attr.mode	= 0444;
> +	hattr->read		= hsmp_metric_tbl_read;
> +	hattr->size		= sizeof(struct hsmp_metric_table);
> +	hattr->private		= &plat_dev.sock[sock_ind];
> +
> +	return device_create_bin_file(plat_dev.dev, hattr);
> +}
> +
> +static int hsmp_get_tbl_dram_base(u16 sock_ind)
> +{
> +	struct hsmp_socket *sock = &plat_dev.sock[sock_ind];
> +	struct hsmp_message msg = { 0 };
> +	phys_addr_t dram_addr;
> +	int ret;
> +
> +	msg.sock_ind	= sock_ind;
> +	msg.response_sz	= hsmp_msg_desc_table[HSMP_GET_METRIC_TABLE_DRAM_ADDR].response_sz;
> +	msg.msg_id	= HSMP_GET_METRIC_TABLE_DRAM_ADDR;
> +
> +	ret = hsmp_send_message(&msg);
> +	if (ret)
> +		return ret;
> +
> +	/*
> +	 * calculate the metric table DRAM address from lower and upper 32 bits
> +	 * sent from SMU and ioremap it to virtual address.
> +	 */
> +	dram_addr = msg.args[0] | ((u64)(msg.args[1]) << 32);
> +	if (!dram_addr)
> +		return -ENOMEM;
> +	sock->metric_tbl_addr = devm_ioremap(plat_dev.dev, dram_addr,
> +					     sizeof(struct hsmp_metric_table));
> +	if (!sock->metric_tbl_addr) {
> +		dev_err(plat_dev.dev, "Unable to ioremap metric table addr\n");
> +		return -ENOMEM;
> +	}
> +	return 0;
> +}
> +
> +static int hsmp_create_sysfs_file(void)
> +{
> +	int ret, i;
> +
> +	for (i = 0; i < num_sockets; i++) {
> +		ret = hsmp_get_tbl_dram_base(i);
> +		if (ret)
> +			goto cleanup;
> +
> +		ret = hsmp_create_metric_tbl_sysfs_file(i);
> +		if (ret) {
> +			dev_err(plat_dev.dev, "Unable to create sysfs file for metric table\n");
> +			goto cleanup;
> +		}
> +	}
> +
> +	return 0;
> +
> +cleanup:
> +	while (i > 0)
> +		device_remove_bin_file(plat_dev.dev, &plat_dev.sock[--i].hsmp_attr);
> +	return ret;
> +}
> +
> +static int hsmp_cache_proto_ver(void)
> +{
> +	struct hsmp_message msg = { 0 };
> +	int ret;
> +
> +	msg.msg_id	= HSMP_GET_PROTO_VER;
> +	msg.sock_ind	= 0;
> +	msg.response_sz = hsmp_msg_desc_table[HSMP_GET_PROTO_VER].response_sz;
> +
> +	ret = hsmp_send_message(&msg);
> +	if (!ret)
> +		plat_dev.proto_ver = msg.args[0];
> +
> +	return ret;
> +}
> +
>  static int hsmp_pltdrv_probe(struct platform_device *pdev)
>  {
> -	int i;
> +	int ret, i;
>  
>  	plat_dev.sock = devm_kzalloc(&pdev->dev,
>  				     (num_sockets * sizeof(struct hsmp_socket)),
> @@ -344,18 +463,44 @@ static int hsmp_pltdrv_probe(struct platform_device *pdev)
>  		plat_dev.sock[i].sock_ind = i;
>  	}
>  
> -	plat_dev.hsmp_device.name	= "hsmp_cdev";
> +	plat_dev.hsmp_device.name	= HSMP_CDEV_NAME;
>  	plat_dev.hsmp_device.minor	= MISC_DYNAMIC_MINOR;
>  	plat_dev.hsmp_device.fops	= &hsmp_fops;
>  	plat_dev.hsmp_device.parent	= &pdev->dev;
> -	plat_dev.hsmp_device.nodename	= "hsmp";
> +	plat_dev.hsmp_device.nodename	= HSMP_DEVNODE_NAME;
>  	plat_dev.hsmp_device.mode	= 0644;
>  
> -	return misc_register(&plat_dev.hsmp_device);
> +	ret = misc_register(&plat_dev.hsmp_device);
> +	if (ret)
> +		return ret;
> +
> +	ret = hsmp_cache_proto_ver();
> +	if (ret) {
> +		dev_err(plat_dev.dev, "Failed to read HSMP protocol version\n");
> +		goto cleanup;
> +	}
> +
> +	/* metrics table is supported only from proto ver6 EPYCs */
> +	if (plat_dev.proto_ver >= HSMP_PROTO_VER6) {
> +		ret = hsmp_create_sysfs_file();
> +		if (ret)
> +			goto cleanup;
> +	}
> +	return 0;
> +
> +cleanup:

Instead naming labels as "cleanup", try to name them according to what you 
cleanup, so e.g. here deregister:

> +	misc_deregister(&plat_dev.hsmp_device);
> +	return ret;
>  }
>  
>  static void hsmp_pltdrv_remove(struct platform_device *pdev)
>  {
> +	int i;
> +
> +	if (plat_dev.proto_ver >= HSMP_PROTO_VER6) {
> +		for (i = 0; i < num_sockets; i++)
> +			device_remove_bin_file(plat_dev.dev, &plat_dev.sock[i].hsmp_attr);
> +	}
>  	misc_deregister(&plat_dev.hsmp_device);
>  }
>  
> 

-- 
 i.

Re: [PATCH 2/4] platform/x86/amd/hsmp: add support for metrics tbl

[Ilpo Järvinen]

On Fri, 18 Aug 2023, Suma Hegde wrote:

> From: Suma Hegde <suma.hegde@amd.com>
> 
> AMD MI300 MCM provides GET_METRICS_TABLE message with which
> all the system management information from SMU can be retrieved in just
> one message.
> 
> The metrics table is available as hexadecimal sysfs binary file in
> /sys/devices/platform/amd_hsmp/socket%d_metrics_bin 
> Metrics table definitions will be documented as part of PPR which
> available in public domain. The same is defined in the amd_hsmp.h header
> file as well.
> 
> Signed-off-by: Suma Hegde <suma.hegde@amd.com>
> Reviewed-by: Naveen Krishna Chatradhi <nchatrad@amd.com>
> ---

> +#define HSMP_ATRR_NAME_SIZE 25
> +static int hsmp_create_metric_tbl_sysfs_file(int sock_ind)
> +{
> +	struct bin_attribute *hattr = &plat_dev.sock[sock_ind].hsmp_attr;
> +	char *name;
> +
> +	sysfs_attr_init(&plat_dev.sock[sock_ind].hsmp_attr);
> +	name = devm_kzalloc(plat_dev.dev, HSMP_ATRR_NAME_SIZE, GFP_KERNEL);
> +	if (!name)
> +		return -ENOMEM;
> +	sprintf(name, "socket%d_metrics_bin", sock_ind);

The structure would be more future-proof if you do socket%d/metrics_bin 
instead. Are you expected to add at some point in the future more files 
which are socket specific?

You didn't provide any documentation for this interface btw, you should 
add it.

> +	hattr->attr.name	= name;
> +	hattr->attr.mode	= 0444;
> +	hattr->read		= hsmp_metric_tbl_read;
> +	hattr->size		= sizeof(struct hsmp_metric_table);
> +	hattr->private		= &plat_dev.sock[sock_ind];
> +
> +	return device_create_bin_file(plat_dev.dev, hattr);
> +}

> +static int hsmp_create_sysfs_file(void)
> +{
> +	int ret, i;
> +
> +	for (i = 0; i < num_sockets; i++) {
> +		ret = hsmp_get_tbl_dram_base(i);
> +		if (ret)
> +			goto cleanup;
> +
> +		ret = hsmp_create_metric_tbl_sysfs_file(i);
> +		if (ret) {
> +			dev_err(plat_dev.dev, "Unable to create sysfs file for metric table\n");
> +			goto cleanup;
> +		}
> +	}
> +
> +	return 0;
> +
> +cleanup:
> +	while (i > 0)
> +		device_remove_bin_file(plat_dev.dev, &plat_dev.sock[--i].hsmp_attr);
> +	return ret;
> +}
> +
> +static int hsmp_cache_proto_ver(void)
> +{
> +	struct hsmp_message msg = { 0 };
> +	int ret;
> +
> +	msg.msg_id	= HSMP_GET_PROTO_VER;
> +	msg.sock_ind	= 0;
> +	msg.response_sz = hsmp_msg_desc_table[HSMP_GET_PROTO_VER].response_sz;
> +
> +	ret = hsmp_send_message(&msg);
> +	if (!ret)
> +		plat_dev.proto_ver = msg.args[0];
> +
> +	return ret;
> +}
> +
>  static int hsmp_pltdrv_probe(struct platform_device *pdev)
>  {
> -	int i;
> +	int ret, i;
>  
>  	plat_dev.sock = devm_kzalloc(&pdev->dev,
>  				     (num_sockets * sizeof(struct hsmp_socket)),
> @@ -344,18 +463,44 @@ static int hsmp_pltdrv_probe(struct platform_device *pdev)
>  		plat_dev.sock[i].sock_ind = i;
>  	}
>  
> -	plat_dev.hsmp_device.name	= "hsmp_cdev";
> +	plat_dev.hsmp_device.name	= HSMP_CDEV_NAME;
>  	plat_dev.hsmp_device.minor	= MISC_DYNAMIC_MINOR;
>  	plat_dev.hsmp_device.fops	= &hsmp_fops;
>  	plat_dev.hsmp_device.parent	= &pdev->dev;
> -	plat_dev.hsmp_device.nodename	= "hsmp";
> +	plat_dev.hsmp_device.nodename	= HSMP_DEVNODE_NAME;
>  	plat_dev.hsmp_device.mode	= 0644;
>  
> -	return misc_register(&plat_dev.hsmp_device);
> +	ret = misc_register(&plat_dev.hsmp_device);
> +	if (ret)
> +		return ret;
> +
> +	ret = hsmp_cache_proto_ver();
> +	if (ret) {
> +		dev_err(plat_dev.dev, "Failed to read HSMP protocol version\n");
> +		goto cleanup;
> +	}
> +
> +	/* metrics table is supported only from proto ver6 EPYCs */
> +	if (plat_dev.proto_ver >= HSMP_PROTO_VER6) {
> +		ret = hsmp_create_sysfs_file();

For new things, driver/device attribute_group should be used where 
possible. .is_visible() can handle that condition check for you.


-- 
 i.

Re: [PATCH 3/4] platform/x86/amd/hsmp: change the print message

[Hegde, Suma]

Hi Ilpo Jarvinen,

Thank you so much for your review comments!

On 8/21/2023 4:15 PM, Ilpo Järvinen wrote:
> Caution: This message originated from an External Source. Use proper caution when opening attachments, clicking links, or responding.
>
>
> The coverletter has 0/3 but the patches came with /4?!?
This is by mistake. I will correct it in next version.
> On Fri, 18 Aug 2023, Suma Hegde wrote:
>
>> From: Suma Hegde<suma.hegde@amd.com>
>>
>> The print message during platform init is little ambiguous
>> so change it to a more meaningful clear message.
>>
>> Signed-off-by: Suma Hegde<suma.hegde@amd.com>
>> Reviewed-by: Naveen Krishna Chatradhi<nchatrad@amd.com>
>> ---
>>   drivers/platform/x86/amd/hsmp.c | 6 +++---
>>   1 file changed, 3 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/platform/x86/amd/hsmp.c b/drivers/platform/x86/amd/hsmp.c
>> index cab59750cde2..1f03a22a9d8a 100644
>> --- a/drivers/platform/x86/amd/hsmp.c
>> +++ b/drivers/platform/x86/amd/hsmp.c
>> @@ -537,10 +537,10 @@ static int __init hsmp_plt_init(void)
>>        for (i = 0; i < num_sockets; i++) {
>>                ret = hsmp_test(i, 0xDEADBEEF);
>>                if (ret) {
>> -                     pr_err("HSMP is not supported on Fam:%x model:%x\n",
>> +                     pr_err("HSMP test message failed on Fam:%x model:%x\n",
>>                               boot_cpu_data.x86, boot_cpu_data.x86_model);
>> -                     pr_err("Or Is HSMP disabled in BIOS ?\n");
>> -                     return -EOPNOTSUPP;
>> +                     pr_err("Is HSMP disabled in BIOS ?\n");
>> +                     return ret;
> Your changelog only talks about changing print, and yet you also change
> return statement. Please explain things properly in the changelog (or
> perhaps even make a separate patch out of the return change if it's
> independent change you planned to make).
I missed mentioning about changed return statement in , will correct 
this too in the next version.
> --
>   i.
>

Re: [PATCH 2/4] platform/x86/amd/hsmp: add support for metrics tbl

[Hegde, Suma]

On 8/21/2023 4:55 PM, Ilpo Järvinen wrote:
> Caution: This message originated from an External Source. Use proper caution when opening attachments, clicking links, or responding.
>
>
> On Fri, 18 Aug 2023, Suma Hegde wrote:
>
>> From: Suma Hegde <suma.hegde@amd.com>
>>
>> AMD MI300 MCM provides GET_METRICS_TABLE message with which
>> all the system management information from SMU can be retrieved in just
>> one message.
>>
>> The metrics table is available as hexadecimal sysfs binary file in
>> /sys/devices/platform/amd_hsmp/socket%d_metrics_bin
>> Metrics table definitions will be documented as part of PPR which
>> available in public domain. The same is defined in the amd_hsmp.h header
>> file as well.
>>
>> Signed-off-by: Suma Hegde <suma.hegde@amd.com>
>> Reviewed-by: Naveen Krishna Chatradhi <nchatrad@amd.com>
>> ---
>>   arch/x86/include/uapi/asm/amd_hsmp.h | 109 +++++++++++++++++++
>>   drivers/platform/x86/amd/hsmp.c      | 155 ++++++++++++++++++++++++++-
>>   2 files changed, 259 insertions(+), 5 deletions(-)
>>
>> diff --git a/arch/x86/include/uapi/asm/amd_hsmp.h b/arch/x86/include/uapi/asm/amd_hsmp.h
>> index 769b939444ae..832701b41c7d 100644
>> --- a/arch/x86/include/uapi/asm/amd_hsmp.h
>> +++ b/arch/x86/include/uapi/asm/amd_hsmp.h
>> @@ -47,6 +47,9 @@ enum hsmp_message_ids {
>>        HSMP_SET_PCI_RATE,              /* 20h Control link rate on PCIe devices */
>>        HSMP_SET_POWER_MODE,            /* 21h Select power efficiency profile policy */
>>        HSMP_SET_PSTATE_MAX_MIN,        /* 22h Set the max and min DF P-State  */
>> +     HSMP_GET_METRIC_TABLE_VER,      /* 23h Get metrics table version  */
>> +     HSMP_GET_METRIC_TABLE,          /* 24h Get metrics table  */
> extra spaces before */
Ok, will remove it.
>> +     HSMP_GET_METRIC_TABLE_DRAM_ADDR,/* 25h Get metrics table dram address */
>>        HSMP_MSG_ID_MAX,
>>   };
>>
>> @@ -64,6 +67,14 @@ enum hsmp_msg_type {
>>        HSMP_GET  = 1,
>>   };
>>
>> +enum hsmp_proto_versions {
>> +     HSMP_PROTO_VER2 = 2,
>> +     HSMP_PROTO_VER3,
>> +     HSMP_PROTO_VER4,
>> +     HSMP_PROTO_VER5,
>> +     HSMP_PROTO_VER6
>> +};
>> +
>>   struct hsmp_msg_desc {
>>        int num_args;
>>        int response_sz;
>> @@ -295,6 +306,104 @@ static const struct hsmp_msg_desc hsmp_msg_desc_table[] = {
>>         * input: args[0] = min df pstate[15:8] + max df pstate[7:0]
>>         */
>>        {1, 0, HSMP_SET},
>> +
>> +     /*
>> +      * HSMP_GET_METRIC_TABLE_VER, num_args = 0, response_sz = 1
>> +      * output: args[0] = metrics table version
>> +      */
>> +     {0, 1, HSMP_GET},
>> +
>> +     /*
>> +      * HSMP_GET_METRIC_TABLE, num_args = 0, response_sz = 0
>> +      */
>> +     {0, 0, HSMP_GET},
>> +
>> +     /*
>> +      * HSMP_GET_METRIC_TABLE_DRAM_ADDR, num_args = 0, response_sz = 2
>> +      * output: args[0] = lower 32 bits of the address
>> +      * output: args[1] = upper 32 bits of the address
>> +      */
>> +     {0, 2, HSMP_GET},
>> +};
>> +
>> +/* Metrics table for EPYC socket(supported only from proto version 6) */
>> +struct hsmp_metric_table {
>> +     __u32 accumulation_counter;
>> +
>> +     //TEMPERATURE
>> +     __u32 max_socket_temperature;
>> +     __u32 max_vr_temperature;
>> +     __u32 max_hbm_temperature;
>> +     __u64 max_socket_temperature_acc;
>> +     __u64 max_vr_temperature_acc;
>> +     __u64 max_hbm_temperature_acc;
>> +
>> +     //POWER
>> +     __u32 socket_power_limit;
>> +     __u32 max_socket_power_limit;
>> +     __u32 socket_power;
>> +
>> +     //ENERGY
>> +     __u64 timestamp;
>> +     __u64 socket_energy_acc;
>> +     __u64 ccd_energy_acc;
>> +     __u64 xcd_energy_acc;
>> +     __u64 aid_energy_acc;
>> +     __u64 hbm_energy_acc;
>> +
>> +     //FREQUENCY
>> +     __u32 cclk_frequency_limit;
>> +     __u32 gfxclk_frequency_limit;
>> +     __u32 fclk_frequency;
>> +     __u32 uclk_frequency;
>> +     __u32 socclk_frequency[4];
>> +     __u32 vclk_frequency[4];
>> +     __u32 dclk_frequency[4];
>> +     __u32 lclk_frequency[4];
>> +     __u64 gfxclk_frequency_acc[8];
>> +     __u64 cclk_frequency_acc[96];
>> +
>> +     //FREQUENCY RANGE
>> +     __u32 max_cclk_frequency;
>> +     __u32 min_cclk_frequency;
>> +     __u32 max_gfxclk_frequency;
>> +     __u32 min_gfxclk_frequency;
>> +     __u32 fclk_frequency_table[4];
>> +     __u32 uclk_frequency_table[4];
>> +     __u32 socclk_frequency_table[4];
>> +     __u32 vclk_frequency_table[4];
>> +     __u32 dclk_frequency_table[4];
>> +     __u32 lclk_frequency_table[4];
>> +     __u32 max_lclk_dpm_range;
>> +     __u32 min_lclk_dpm_range;
>> +
>> +     //XGMI
>> +     __u32 xgmi_width;
>> +     __u32 xgmi_bitrate;
>> +     __u64 xgmi_read_bandwidth_acc[8];
>> +     __u64 xgmi_write_bandwidth_acc[8];
>> +
>> +     //ACTIVITY
>> +     __u32 socket_c0_residency;
>> +     __u32 socket_gfx_busy;
>> +     __u32 dram_bandwidth_utilization;
>> +     __u64 socket_c0_residency_acc;
>> +     __u64 socket_gfx_busy_acc;
>> +     __u64 dram_bandwidth_acc;
>> +     __u32 max_dram_bandwidth;
>> +     __u64 dram_bandwidth_utilization_acc;
>> +     __u64 pcie_bandwidth_acc[4];
>> +
>> +     //THROTTLERS
>> +     __u32 prochot_residency_acc;
>> +     __u32 ppt_residency_acc;
>> +     __u32 socket_thm_residency_acc;
>> +     __u32 vr_thm_residency_acc;
>> +     __u32 hbm_thm_residency_acc;
>> +     __u32 spare;
>> +
>> +     // New Items at end to maintain driver compatibility
>> +     __u32 gfxclk_frequency[8];
>>   };
>>
>>   /* Reset to default packing */
>> diff --git a/drivers/platform/x86/amd/hsmp.c b/drivers/platform/x86/amd/hsmp.c
>> index 699c22f7cd42..cab59750cde2 100644
>> --- a/drivers/platform/x86/amd/hsmp.c
>> +++ b/drivers/platform/x86/amd/hsmp.c
>> @@ -20,7 +20,7 @@
>>   #include <linux/semaphore.h>
>>
>>   #define DRIVER_NAME          "amd_hsmp"
>> -#define DRIVER_VERSION               "1.0"
>> +#define DRIVER_VERSION               "2.0"
>>
>>   /* HSMP Status / Error codes */
>>   #define HSMP_STATUS_NOT_READY        0x00
>> @@ -47,7 +47,12 @@
>>   #define HSMP_INDEX_REG               0xc4
>>   #define HSMP_DATA_REG                0xc8
>>
>> +#define HSMP_CDEV_NAME               "hsmp_cdev"
>> +#define HSMP_DEVNODE_NAME    "hsmp"
>> +
>>   struct hsmp_socket {
>> +     struct bin_attribute hsmp_attr;
>> +     void __iomem *metric_tbl_addr;
>>        struct semaphore hsmp_sem;
>>        u16 sock_ind;
>>   };
>> @@ -56,6 +61,7 @@ struct hsmp_plat_device {
>>        struct miscdevice hsmp_device;
>>        struct hsmp_socket *sock;
>>        struct device *dev;
>> +     u32 proto_ver;
>>   };
>>
>>   static u16 num_sockets;
>> @@ -328,9 +334,122 @@ static const struct file_operations hsmp_fops = {
>>        .compat_ioctl   = hsmp_ioctl,
>>   };
>>
>> +static ssize_t hsmp_metric_tbl_read(struct file *filp, struct kobject *kobj,
>> +                                 struct bin_attribute *bin_attr, char *buf,
>> +                                 loff_t off, size_t count)
>> +{
>> +     struct hsmp_socket *sock = bin_attr->private;
>> +     struct hsmp_message msg = { 0 };
>> +     int ret;
>> +
>> +     if (count == 0 || count > sizeof(struct hsmp_metric_table))
> !count
Ok.
>
>> +             return 0;
>> +
>> +     msg.msg_id      = HSMP_GET_METRIC_TABLE;
>> +     msg.sock_ind    = sock->sock_ind;
>> +
>> +     ret = hsmp_send_message(&msg);
>> +     if (ret)
>> +             return ret;
>> +     memcpy(buf, sock->metric_tbl_addr, count);
>> +
>> +     return count;
>> +}
>> +
>> +#define HSMP_ATRR_NAME_SIZE 25
> ATTR (typo)?
Yes, its a typo. will correct it.
>
>> +static int hsmp_create_metric_tbl_sysfs_file(int sock_ind)
>> +{
>> +     struct bin_attribute *hattr = &plat_dev.sock[sock_ind].hsmp_attr;
>> +     char *name;
>> +
>> +     sysfs_attr_init(&plat_dev.sock[sock_ind].hsmp_attr);
>> +     name = devm_kzalloc(plat_dev.dev, HSMP_ATRR_NAME_SIZE, GFP_KERNEL);
>> +     if (!name)
>> +             return -ENOMEM;
>> +     sprintf(name, "socket%d_metrics_bin", sock_ind);
> snprintf()
Ok. will change it.
>
>> +     hattr->attr.name        = name;
>> +     hattr->attr.mode        = 0444;
>> +     hattr->read             = hsmp_metric_tbl_read;
>> +     hattr->size             = sizeof(struct hsmp_metric_table);
>> +     hattr->private          = &plat_dev.sock[sock_ind];
>> +
>> +     return device_create_bin_file(plat_dev.dev, hattr);
>> +}
>> +
>> +static int hsmp_get_tbl_dram_base(u16 sock_ind)
>> +{
>> +     struct hsmp_socket *sock = &plat_dev.sock[sock_ind];
>> +     struct hsmp_message msg = { 0 };
>> +     phys_addr_t dram_addr;
>> +     int ret;
>> +
>> +     msg.sock_ind    = sock_ind;
>> +     msg.response_sz = hsmp_msg_desc_table[HSMP_GET_METRIC_TABLE_DRAM_ADDR].response_sz;
>> +     msg.msg_id      = HSMP_GET_METRIC_TABLE_DRAM_ADDR;
>> +
>> +     ret = hsmp_send_message(&msg);
>> +     if (ret)
>> +             return ret;
>> +
>> +     /*
>> +      * calculate the metric table DRAM address from lower and upper 32 bits
>> +      * sent from SMU and ioremap it to virtual address.
>> +      */
>> +     dram_addr = msg.args[0] | ((u64)(msg.args[1]) << 32);
>> +     if (!dram_addr)
>> +             return -ENOMEM;
>> +     sock->metric_tbl_addr = devm_ioremap(plat_dev.dev, dram_addr,
>> +                                          sizeof(struct hsmp_metric_table));
>> +     if (!sock->metric_tbl_addr) {
>> +             dev_err(plat_dev.dev, "Unable to ioremap metric table addr\n");
>> +             return -ENOMEM;
>> +     }
>> +     return 0;
>> +}
>> +
>> +static int hsmp_create_sysfs_file(void)
>> +{
>> +     int ret, i;
>> +
>> +     for (i = 0; i < num_sockets; i++) {
>> +             ret = hsmp_get_tbl_dram_base(i);
>> +             if (ret)
>> +                     goto cleanup;
>> +
>> +             ret = hsmp_create_metric_tbl_sysfs_file(i);
>> +             if (ret) {
>> +                     dev_err(plat_dev.dev, "Unable to create sysfs file for metric table\n");
>> +                     goto cleanup;
>> +             }
>> +     }
>> +
>> +     return 0;
>> +
>> +cleanup:
>> +     while (i > 0)
>> +             device_remove_bin_file(plat_dev.dev, &plat_dev.sock[--i].hsmp_attr);
>> +     return ret;
>> +}
>> +
>> +static int hsmp_cache_proto_ver(void)
>> +{
>> +     struct hsmp_message msg = { 0 };
>> +     int ret;
>> +
>> +     msg.msg_id      = HSMP_GET_PROTO_VER;
>> +     msg.sock_ind    = 0;
>> +     msg.response_sz = hsmp_msg_desc_table[HSMP_GET_PROTO_VER].response_sz;
>> +
>> +     ret = hsmp_send_message(&msg);
>> +     if (!ret)
>> +             plat_dev.proto_ver = msg.args[0];
>> +
>> +     return ret;
>> +}
>> +
>>   static int hsmp_pltdrv_probe(struct platform_device *pdev)
>>   {
>> -     int i;
>> +     int ret, i;
>>
>>        plat_dev.sock = devm_kzalloc(&pdev->dev,
>>                                     (num_sockets * sizeof(struct hsmp_socket)),
>> @@ -344,18 +463,44 @@ static int hsmp_pltdrv_probe(struct platform_device *pdev)
>>                plat_dev.sock[i].sock_ind = i;
>>        }
>>
>> -     plat_dev.hsmp_device.name       = "hsmp_cdev";
>> +     plat_dev.hsmp_device.name       = HSMP_CDEV_NAME;
>>        plat_dev.hsmp_device.minor      = MISC_DYNAMIC_MINOR;
>>        plat_dev.hsmp_device.fops       = &hsmp_fops;
>>        plat_dev.hsmp_device.parent     = &pdev->dev;
>> -     plat_dev.hsmp_device.nodename   = "hsmp";
>> +     plat_dev.hsmp_device.nodename   = HSMP_DEVNODE_NAME;
>>        plat_dev.hsmp_device.mode       = 0644;
>>
>> -     return misc_register(&plat_dev.hsmp_device);
>> +     ret = misc_register(&plat_dev.hsmp_device);
>> +     if (ret)
>> +             return ret;
>> +
>> +     ret = hsmp_cache_proto_ver();
>> +     if (ret) {
>> +             dev_err(plat_dev.dev, "Failed to read HSMP protocol version\n");
>> +             goto cleanup;
>> +     }
>> +
>> +     /* metrics table is supported only from proto ver6 EPYCs */
>> +     if (plat_dev.proto_ver >= HSMP_PROTO_VER6) {
>> +             ret = hsmp_create_sysfs_file();
>> +             if (ret)
>> +                     goto cleanup;
>> +     }
>> +     return 0;
>> +
>> +cleanup:
> Instead naming labels as "cleanup", try to name them according to what you
> cleanup, so e.g. here deregister:
Ok, will rename it.
>> +     misc_deregister(&plat_dev.hsmp_device);
>> +     return ret;
>>   }
>>
>>   static void hsmp_pltdrv_remove(struct platform_device *pdev)
>>   {
>> +     int i;
>> +
>> +     if (plat_dev.proto_ver >= HSMP_PROTO_VER6) {
>> +             for (i = 0; i < num_sockets; i++)
>> +                     device_remove_bin_file(plat_dev.dev, &plat_dev.sock[i].hsmp_attr);
>> +     }
>>        misc_deregister(&plat_dev.hsmp_device);
>>   }
>>
>>
> --
>   i.

Thanks and Regards,

Suma

Re: [PATCH 2/4] platform/x86/amd/hsmp: add support for metrics tbl

[Hegde, Suma]

Hi Ilpo Jarvinen,

Thanks for the suggestions.

On 8/22/2023 2:27 PM, Ilpo Järvinen wrote:
> Caution: This message originated from an External Source. Use proper caution when opening attachments, clicking links, or responding.
>
>
> On Fri, 18 Aug 2023, Suma Hegde wrote:
>
>> From: Suma Hegde <suma.hegde@amd.com>
>>
>> AMD MI300 MCM provides GET_METRICS_TABLE message with which
>> all the system management information from SMU can be retrieved in just
>> one message.
>>
>> The metrics table is available as hexadecimal sysfs binary file in
>> /sys/devices/platform/amd_hsmp/socket%d_metrics_bin
>> Metrics table definitions will be documented as part of PPR which
>> available in public domain. The same is defined in the amd_hsmp.h header
>> file as well.
>>
>> Signed-off-by: Suma Hegde <suma.hegde@amd.com>
>> Reviewed-by: Naveen Krishna Chatradhi <nchatrad@amd.com>
>> ---
>> +#define HSMP_ATRR_NAME_SIZE 25
>> +static int hsmp_create_metric_tbl_sysfs_file(int sock_ind)
>> +{
>> +     struct bin_attribute *hattr = &plat_dev.sock[sock_ind].hsmp_attr;
>> +     char *name;
>> +
>> +     sysfs_attr_init(&plat_dev.sock[sock_ind].hsmp_attr);
>> +     name = devm_kzalloc(plat_dev.dev, HSMP_ATRR_NAME_SIZE, GFP_KERNEL);
>> +     if (!name)
>> +             return -ENOMEM;
>> +     sprintf(name, "socket%d_metrics_bin", sock_ind);
> The structure would be more future-proof if you do socket%d/metrics_bin
> instead. Are you expected to add at some point in the future more files
> which are socket specific?
>
> You didn't provide any documentation for this interface btw, you should
> add it.

We are not expecting any new socket files right now, but in future we 
might create files for some more socket specific info.

So I will add directories for sockets and then will create files.

I will update documentation too.

>> +     hattr->attr.name        = name;
>> +     hattr->attr.mode        = 0444;
>> +     hattr->read             = hsmp_metric_tbl_read;
>> +     hattr->size             = sizeof(struct hsmp_metric_table);
>> +     hattr->private          = &plat_dev.sock[sock_ind];
>> +
>> +     return device_create_bin_file(plat_dev.dev, hattr);
>> +}
>> +static int hsmp_create_sysfs_file(void)
>> +{
>> +     int ret, i;
>> +
>> +     for (i = 0; i < num_sockets; i++) {
>> +             ret = hsmp_get_tbl_dram_base(i);
>> +             if (ret)
>> +                     goto cleanup;
>> +
>> +             ret = hsmp_create_metric_tbl_sysfs_file(i);
>> +             if (ret) {
>> +                     dev_err(plat_dev.dev, "Unable to create sysfs file for metric table\n");
>> +                     goto cleanup;
>> +             }
>> +     }
>> +
>> +     return 0;
>> +
>> +cleanup:
>> +     while (i > 0)
>> +             device_remove_bin_file(plat_dev.dev, &plat_dev.sock[--i].hsmp_attr);
>> +     return ret;
>> +}
>> +
>> +static int hsmp_cache_proto_ver(void)
>> +{
>> +     struct hsmp_message msg = { 0 };
>> +     int ret;
>> +
>> +     msg.msg_id      = HSMP_GET_PROTO_VER;
>> +     msg.sock_ind    = 0;
>> +     msg.response_sz = hsmp_msg_desc_table[HSMP_GET_PROTO_VER].response_sz;
>> +
>> +     ret = hsmp_send_message(&msg);
>> +     if (!ret)
>> +             plat_dev.proto_ver = msg.args[0];
>> +
>> +     return ret;
>> +}
>> +
>>   static int hsmp_pltdrv_probe(struct platform_device *pdev)
>>   {
>> -     int i;
>> +     int ret, i;
>>
>>        plat_dev.sock = devm_kzalloc(&pdev->dev,
>>                                     (num_sockets * sizeof(struct hsmp_socket)),
>> @@ -344,18 +463,44 @@ static int hsmp_pltdrv_probe(struct platform_device *pdev)
>>                plat_dev.sock[i].sock_ind = i;
>>        }
>>
>> -     plat_dev.hsmp_device.name       = "hsmp_cdev";
>> +     plat_dev.hsmp_device.name       = HSMP_CDEV_NAME;
>>        plat_dev.hsmp_device.minor      = MISC_DYNAMIC_MINOR;
>>        plat_dev.hsmp_device.fops       = &hsmp_fops;
>>        plat_dev.hsmp_device.parent     = &pdev->dev;
>> -     plat_dev.hsmp_device.nodename   = "hsmp";
>> +     plat_dev.hsmp_device.nodename   = HSMP_DEVNODE_NAME;
>>        plat_dev.hsmp_device.mode       = 0644;
>>
>> -     return misc_register(&plat_dev.hsmp_device);
>> +     ret = misc_register(&plat_dev.hsmp_device);
>> +     if (ret)
>> +             return ret;
>> +
>> +     ret = hsmp_cache_proto_ver();
>> +     if (ret) {
>> +             dev_err(plat_dev.dev, "Failed to read HSMP protocol version\n");
>> +             goto cleanup;
>> +     }
>> +
>> +     /* metrics table is supported only from proto ver6 EPYCs */
>> +     if (plat_dev.proto_ver >= HSMP_PROTO_VER6) {
>> +             ret = hsmp_create_sysfs_file();
> For new things, driver/device attribute_group should be used where
> possible. .is_visible() can handle that condition check for you.

Ok, will do the required modification.


> --
>   i.

Thank you,

Suma