Re: [PATCH linux dev-4.10 3/6] drivers/misc: Add driver for Aspeed PECI and generic PECI headers
From: Jae Hyun Yoo
Date: Wed Jan 10 2018 - 18:11:33 EST
On 1/10/2018 3:55 AM, Arnd Bergmann wrote:
On Tue, Jan 9, 2018 at 11:31 PM, Jae Hyun Yoo
<jae.hyun.yoo@xxxxxxxxxxxxxxx> wrote:
This commit adds driver implementation for Aspeed PECI. Also adds
generic peci.h and peci_ioctl.h files to provide compatibility
to peci drivers that can be implemented later e.g. Nuvoton's BMC
SoC family.
Signed-off-by: Jae Hyun Yoo <jae.hyun.yoo@xxxxxxxxxxxxxxx>
+#include <linux/clk.h>
+#include <linux/crc8.h>
+#include <linux/delay.h>
+#include <linux/interrupt.h>
+#include <linux/io.h>
+#include <linux/jiffies.h>
+#include <linux/mfd/syscon.h>
+#include <linux/miscdevice.h>
+#include <linux/module.h>
+#include <linux/of.h>
+#include <linux/peci_ioctl.h>
+#include <linux/platform_device.h>
+#include <linux/regmap.h>
+#include <linux/semaphore.h>
+#include <linux/types.h>
+#include <linux/uaccess.h>
semaphore.h is not used here and can be dropped.
You are right. Will drop it.
+static struct aspeed_peci *aspeed_peci_priv;
Try to avoid instance variables like this one. You should always be able to find
that pointer from whatever structure you were called with.
Okay. I will use driver_data instead.
+ timeout = wait_for_completion_interruptible_timeout(
+ &priv->xfer_complete,
+ msecs_to_jiffies(priv->cmd_timeout_ms));
+
+ dev_dbg(priv->dev, "INT_STS : 0x%08x\n", priv->sts);
+ if (!regmap_read(priv->regmap, AST_PECI_CMD, &peci_state))
+ dev_dbg(priv->dev, "PECI_STATE : 0x%lx\n",
+ PECI_CMD_STS_GET(peci_state));
+ else
+ dev_dbg(priv->dev, "PECI_STATE : read error\n");
+
+ if (timeout <= 0 || !(priv->sts & PECI_INT_CMD_DONE)) {
+ if (timeout <= 0) {
+ dev_dbg(priv->dev, "Timeout waiting for a response!\n");
+ rc = -ETIME;
+ } else {
+ dev_dbg(priv->dev, "No valid response!\n");
+ rc = -EFAULT;
+ }
+ return rc;
+ }
You don't seem to handle -ERESTARTSYS correct here. Either do it
right, or drop the _interruptible part above.
Will add a handling logic for the -ERESTARTSYS.
+typedef int (*ioctl_fn)(struct aspeed_peci *, void *);
+
+static ioctl_fn peci_ioctl_fn[PECI_CMD_MAX] = {
+ ioctl_xfer_msg,
+ ioctl_ping,
+ ioctl_get_dib,
+ ioctl_get_temp,
+ ioctl_rd_pkg_cfg,
+ ioctl_wr_pkg_cfg,
+ ioctl_rd_ia_msr,
+ NULL, /* Reserved */
+ ioctl_rd_pci_cfg,
+ NULL, /* Reserved */
+ ioctl_rd_pci_cfg_local,
+ ioctl_wr_pci_cfg_local,
+};
+
+
+long peci_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
+{
+ struct aspeed_peci *priv;
+ long ret = 0;
+ void __user *argp = (void __user *)arg;
+ int timeout = PECI_IDLE_CHECK_TIMEOUT;
+ u8 msg[sizeof(struct peci_xfer_msg)];
+ unsigned int peci_cmd, msg_size;
+ u32 cmd_sts;
+
+ /*
+ * Treat it as an inter module call when filp is null but only in case
+ * the private data is initialized.
+ */
+ if (filp)
+ priv = container_of(filp->private_data,
+ struct aspeed_peci, miscdev);
+ else
+ priv = aspeed_peci_priv;
Drop this.
peci_ioctl is being called from peci_hwmon as an inter-module call so it
is needed, but as you suggested in the other patch, I'll consider
redesign it with adding a peci device class.
+ if (!priv)
+ return -ENXIO;
+
+ switch (cmd) {
+ case PECI_IOC_XFER:
+ case PECI_IOC_PING:
+ case PECI_IOC_GET_DIB:
+ case PECI_IOC_GET_TEMP:
+ case PECI_IOC_RD_PKG_CFG:
+ case PECI_IOC_WR_PKG_CFG:
+ case PECI_IOC_RD_IA_MSR:
+ case PECI_IOC_RD_PCI_CFG:
+ case PECI_IOC_RD_PCI_CFG_LOCAL:
+ case PECI_IOC_WR_PCI_CFG_LOCAL:
+ peci_cmd = _IOC_TYPE(cmd) - PECI_IOC_BASE;
+ msg_size = _IOC_SIZE(cmd);
+ break;
Having to keep the switch() statement and the array above seems a
little fragile. Can you just do one or the other?
Regarding the command set, you have both a low-level PECI_IOC_XFER
interface and a high-level interface. Can you explain why? I'd think that
generally speaking it's better to have only one of the two.
I was intended to provide generic peci command set, also the low level
PECI_IOC_XFER to provide flexibility for a case when compose a custom
peci command which cannot be covered by the high-level command set. As
you said, all other commands can be implemented in the upper layer but
the benefit of when this driver has the implementation is, it's easy to
manage retry logic since peci is retrial based protocol intends to do
not disturb a CPU if the CPU is doing more important task.
However, your thought also makes sense. I'll check the spec again
whether the high-level command set can cover all cases. If so, I'll
remove the low-level command.
+ /* Check command sts and bus idle state */
+ while (!regmap_read(priv->regmap, AST_PECI_CMD, &cmd_sts)
+ && (cmd_sts & (PECI_CMD_STS_MASK | PECI_CMD_PIN_MON))) {
+ if (timeout-- < 0) {
+ dev_dbg(priv->dev, "Timeout waiting for idle state!\n");
+ ret = -ETIME;
+ goto out;
+ }
+ usleep_range(10000, 11000);
+ };
To implement timeout, it's better to replace the counter with a
jiffies/time_before or ktime_get()/ktime_before() check, since usleep_range()
is might sleep considerably longer than expected.
Thanks for the suggestion. Will rewrite it using ktime_get()/ktime_before().
+EXPORT_SYMBOL_GPL(peci_ioctl);
No user of this, so drop it.
peci_hwmon is using it.
+static int aspeed_peci_open(struct inode *inode, struct file *filp)
+{
+ struct aspeed_peci *priv =
+ container_of(filp->private_data, struct aspeed_peci, miscdev);
+
+ atomic_inc(&priv->ref_count);
+
+ dev_dbg(priv->dev, "ref_count : %d\n", atomic_read(&priv->ref_count));
+
+ return 0;
+}
+
+static int aspeed_peci_release(struct inode *inode, struct file *filp)
+{
+ struct aspeed_peci *priv =
+ container_of(filp->private_data, struct aspeed_peci, miscdev);
+
+ atomic_dec(&priv->ref_count);
+
+ dev_dbg(priv->dev, "ref_count : %d\n", atomic_read(&priv->ref_count));
+
+ return 0;
+}
Nothing uses that reference count, drop it.
You are right. Will drop it.
new file mode 100644
index 0000000..66322c6
--- /dev/null
+++ b/include/misc/peci.h
@@ -0,0 +1,11 @@
+// SPDX-License-Identifier: GPL-2.0
+// Copyright (c) 2017 Intel Corporation
+
+#ifndef __PECI_H
+#define __PECI_H
+
+#include <linux/peci_ioctl.h>
+
+long peci_ioctl(struct file *filp, unsigned int cmd, unsigned long arg);
+
+#endif /* __PECI_H */
Not used anywhere.
peci_hwmon is using it.
diff --git a/include/uapi/linux/peci_ioctl.h b/include/uapi/linux/peci_ioctl.h
new file mode 100644
index 0000000..8386848
--- /dev/null
+++ b/include/uapi/linux/peci_ioctl.h
@@ -0,0 +1,270 @@
+// SPDX-License-Identifier: GPL-2.0
+// Copyright (c) 2017 Intel Corporation
+
+#ifndef __PECI_IOCTL_H
+#define __PECI_IOCTL_H
+
+#include <linux/ioctl.h>
+
+/* Base Address of 48d */
+#define PECI_BASE_ADDR 0x30 /* The PECI client's default address of 0x30 */
+#define PECI_OFFSET_MAX 8 /* Max numver of CPU clients */
+
+/* PCI Access */
+#define MAX_PCI_READ_LEN 24 /* Number of bytes of the PCI Space read */
+
+#define PCI_BUS0_CPU0 0x00
+#define PCI_BUS0_CPU1 0x80
+#define PCI_CPUBUSNO_BUS 0x00
+#define PCI_CPUBUSNO_DEV 0x08
+#define PCI_CPUBUSNO_FUNC 0x02
+#define PCI_CPUBUSNO 0xcc
+#define PCI_CPUBUSNO_1 0xd0
+#define PCI_CPUBUSNO_VALID 0xd4
I can't tell for sure, but this file seems to be mixing the kernel API with
hardware specific macros that are not needed in user space. Can you move
some of this file into the driver itself?
This might go back to the previous question about the high-level and
low-level interfaces: if you can drop the low-level ioctl interface, more
of this header can become private to the driver.
As I answered above, I'll check the spec again and remove the low-level
command if the high-level command set covers all cases.
+/* Package Identifier Read Parameter Value */
+#define PKG_ID_CPU_ID 0x0000 /* 0 - CPUID Info */
+#define PKG_ID_PLATFORM_ID 0x0001 /* 1 - Platform ID */
+#define PKG_ID_UNCORE_ID 0x0002 /* 2 - Uncore Device ID */
+#define PKG_ID_MAX_THREAD_ID 0x0003 /* 3 - Max Thread ID */
+#define PKG_ID_MICROCODE_REV 0x0004 /* 4 - CPU Microcode Update Revision */
+#define PKG_ID_MACHINE_CHECK_STATUS 0x0005 /* 5 - Machine Check Status */
+
+/* RdPkgConfig Index */
+#define MBX_INDEX_CPU_ID 0 /* Package Identifier Read */
+#define MBX_INDEX_VR_DEBUG 1 /* VR Debug */
+#define MBX_INDEX_PKG_TEMP_READ 2 /* Package Temperature Read */
+#define MBX_INDEX_ENERGY_COUNTER 3 /* Energy counter */
+#define MBX_INDEX_ENERGY_STATUS 4 /* DDR Energy Status */
+#define MBX_INDEX_WAKE_MODE_BIT 5 /* "Wake on PECI" Mode bit */
+#define MBX_INDEX_EPI 6 /* Efficient Performance Indication */
Who defines these constants? Are they specific to the aspeed BMC, to the HECI
protocol, or to a particular version of the remote endpoint?
These are common peci definitions, not the aspeed BMC specific.
+#pragma pack(push, 1)
+struct peci_xfer_msg {
+ unsigned char client_addr;
+ unsigned char tx_len;
+ unsigned char rx_len;
+ unsigned char tx_buf[MAX_BUFFER_SIZE];
+ unsigned char rx_buf[MAX_BUFFER_SIZE];
+};
+#pragma pack(pop)
+
+struct peci_ping_msg {
+ unsigned char target;
+};
+
+struct peci_get_dib_msg {
+ unsigned char target;
+ unsigned int dib;
+};
+
+struct peci_get_temp_msg {
+ unsigned char target;
+ signed short temp_raw;
+};
Aside from what Greg already said about the types, please be careful to
also avoid implicit padding in the API data structures, including the end of the
structure.
Okay, I'll expand the pack() scope for all these definition.
+#define PECI_IOC_RD_PCI_CFG \
+ _IOWR(PECI_IOC_BASE + PECI_CMD_RD_PCI_CFG, 0, \
+ struct peci_rd_pci_cfg_msg)
+
+#define PECI_IOC_RD_PCI_CFG_LOCAL \
+ _IOWR(PECI_IOC_BASE + PECI_CMD_RD_PCI_CFG_LOCAL, 0, \
+ struct peci_rd_pci_cfg_local_msg)
+
+#define PECI_IOC_WR_PCI_CFG_LOCAL \
+ _IOWR(PECI_IOC_BASE + PECI_CMD_WR_PCI_CFG_LOCAL, 0, \
+ struct peci_wr_pci_cfg_local_msg)
Can you give some background on what these do? In particular, who
is configuring whose PCI devices?
Arnd
These are commands to read/write a client CPU's PCI configuration which
could be an end-point of the physical PECI interface connection. BMC
controller will be a host and a CPU will be a client.
Thanks,
Jae