Skip to contentSkip to conversation

Intel Platform Monitoring Technology

PATCH Superseded David E. Box 3 participants 17 messages Oct 1, 2020

Subsystems (4)

v7 of 9 Superseded Latest is v9 →

lore
Share & tools
lore

David E. Box

author maintainer
#0 lore raw

v7 cover Intel Platform Monitoring Technology

Intel Platform Monitoring Technology (PMT) is an architecture for enumerating and accessing hardware monitoring capabilities on a device. With customers increasingly asking for hardware telemetry, engineers not only have to figure out how to measure and collect data, but also how to deliver it and make it discoverable. The latter may be through some device specific method requiring device specific tools to collect the data. This in turn requires customers to manage a suite of different tools in order to collect the differing assortment of monitoring data on their systems. Even when such information can be provided in kernel drivers, they may require constant maintenance to update register mappings as they change with firmware updates and new versions of hardware. PMT provides a solution for discovering and reading telemetry from a device through a hardware agnostic framework that allows for updates to systems without requiring patches to the kernel or software tools.

PMT defines several capabilities to support collecting monitoring data from hardware. All are discoverable as separate instances of the PCIE Designated Vendor extended capability (DVSEC) with the Intel vendor code. The DVSEC ID field uniquely identifies the capability. Each DVSEC also provides a BAR offset to a header that defines capability-specific attributes, including GUID, feature type, offset and length, as well as configuration settings where applicable. The GUID uniquely identifies the register space of any monitor data exposed by the capability. The GUID is associated with an XML file from the vendor that describes the mapping of the register space along with properties of the monitor data. This allows vendors to perform firmware updates that can change the mapping (e.g. add new metrics) without requiring any changes to drivers or software tools. The new mapping is confirmed by an updated GUID, read from the hardware, which software uses with a new XML.

The current capabilities defined by PMT are Telemetry, Watcher, and Crashlog. The Telemetry capability provides access to a continuous block of read only data. The Watcher capability provides access to hardware sampling and tracing features. Crashlog provides access to device crash dumps. While there is some relationship between capabilities (Watcher can be configured to sample from the Telemetry data set) each exists as stand alone features with no dependency on any other. The design therefore splits them into individual, capability specific drivers. MFD is used to create platform devices for each capability so that they may be managed by their own driver. The PMT architecture is (for the most part) agnostic to the type of device it can collect from. Software can determine which devices support a PMT feature by searching through each device node entry in the sysfs class folder. It can additionally determine if a particular device supports a PMT feature by checking for a PMT class folder in the device folder.

This patch set provides support for the PMT framework, along with support for Telemetry on Tiger Lake.

Changes from V6:
- Use NULL for OOBMSM driver data instead of an empty struct. Rewrite the code to check for NULL driver_data.
- Fix spelling and formatting in Kconfig.
- Use MKDEV(0,0) to prevent unneeded device node from being created.

Changes from V5:
- Add Alder Lake and the "Out of Band Management Services Module (OOBMSM)" ids to the MFD driver. Transferred to this patch set.
- Use a single class for all PMT capabilities as suggested by Hans.
- Add binary attribute for telemetry driver to allow read syscall as suggested by Hans.
- Use the class file to hold attributes and other common code used by all PMT drivers.
- Add the crashlog driver to the patch set and add a mutex to protect access to the enable control and trigger files as suggested by Hans.

Changes from V4:
- Replace MFD with PMT in driver title
- Fix commit tags in chronological order
- Fix includes in alphabetical order
- Use 'raw' string instead of defines for device names
- Add an error message when returning an error code for unrecognized capability id
- Use dev_err instead of dev_warn for messages when returning an error
- Change while loop to call pci_find_next_ext_capability once
- Add missing continue in while loop
- Keep PCI platform defines using PCI_DEVICE_DATA magic tied to the pci_device_id table
- Comment and kernel message cleanup

Changes from V3:
- Write out full acronym for DVSEC in PCI patch commit message and add 'Designated' to comments
- remove unused variable caught by kernel test robot <[email protected]>
- Add required Co-developed-by signoffs, noted by Andy
- Allow access using new CAP_PERFMON capability as suggested by Alexey Bundankov
- Fix spacing in Kconfig, noted by Randy
- Other style changes and fixups suggested by Andy

Changes from V2:
- In order to handle certain HW bugs from the telemetry capability driver, create a single platform device per capability instead of a device per entry. Add the entry data as device resources and let the capability driver manage them as a set allowing for cleaner HW bug resolution.
- Handle discovery table offset bug in intel_pmt.c
- Handle overlapping regions in intel_pmt_telemetry.c
- Add description of sysfs class to testing ABI.
- Don't check size and count until confirming support for the PMT capability to avoid bailing out when we need to skip it.
- Remove unneeded header file. Move code to the intel_pmt.c, the only place where it's needed.
- Remove now unused platform data.
- Add missing header files types.h, bits.h.
- Rename file name and build options from telem to telemetry.
- Code cleanup suggested by Andy S.
- x86 mailing list added.

Changes from V1:
- In the telemetry driver, set the device in device_create() to the parent PCI device (the monitoring device) for clear association in sysfs. Was set before to the platform device created by the PCI parent.
- Move telem struct into driver and delete unneeded header file.
- Start telem device numbering from 0 instead of 1. 1 was used due to anticipated changes, no longer needed.
- Use helper macros suggested by Andy S.
- Rename class to pmt_telemetry, spelling out full name
- Move monitor device name defines to common header
- Coding style, spelling, and Makefile/MAINTAINERS ordering fixes

Alexander Duyck (3):
platform/x86: Intel PMT class driver platform/x86: Intel PMT Telemetry capability driver platform/x86: Intel PMT Crashlog capability driver

David E. Box (2):
PCI: Add defines for Designated Vendor-Specific Extended Capability
mfd: Intel Platform Monitoring Technology support

 .../ABI/testing/sysfs-class-intel_pmt         | 119 ++++++
 MAINTAINERS                                   |   6 +
 drivers/mfd/Kconfig                           |  10 +
 drivers/mfd/Makefile                          |   1 +
 drivers/mfd/intel_pmt.c                       | 225 ++++++++++++
 drivers/platform/x86/Kconfig                  |  25 ++
 drivers/platform/x86/Makefile                 |   3 +
 drivers/platform/x86/intel_pmt_class.c        | 286 +++++++++++++++
 drivers/platform/x86/intel_pmt_class.h        |  57 +++
 drivers/platform/x86/intel_pmt_crashlog.c     | 339 ++++++++++++++++++
 drivers/platform/x86/intel_pmt_telemetry.c    | 158 ++++++++
 include/uapi/linux/pci_regs.h                 |   5 +
 12 files changed, 1234 insertions(+)
 create mode 100644 Documentation/ABI/testing/sysfs-class-intel_pmt
 create mode 100644 drivers/mfd/intel_pmt.c
 create mode 100644 drivers/platform/x86/intel_pmt_class.c
 create mode 100644 drivers/platform/x86/intel_pmt_class.h
 create mode 100644 drivers/platform/x86/intel_pmt_crashlog.c
 create mode 100644 drivers/platform/x86/intel_pmt_telemetry.c
Signature · 1 line
2.20.1
author maintainer
#12 lore raw

PATCH v7 5/5 platform/x86: Intel PMT Crashlog capability driver

From: Alexander Duyck <[email protected]>

Add support for the Intel Platform Monitoring Technology crashlog interface. This interface provides a few sysfs values to allow for controlling the crashlog telemetry interface as well as a character driver to allow for mapping the crashlog memory region so that it can be accessed after a crashlog has been recorded.

This driver is meant to only support the server version of the crashlog which is identified as crash_type 1 with a version of zero. Currently no other types are supported.

---
 .../ABI/testing/sysfs-class-intel_pmt         |  65 ++++
 drivers/platform/x86/Kconfig                  |   8 +
 drivers/platform/x86/Makefile                 |   1 +
 drivers/platform/x86/intel_pmt_crashlog.c     | 339 ++++++++++++++++++
 4 files changed, 413 insertions(+)
 create mode 100644 drivers/platform/x86/intel_pmt_crashlog.c
Documentation/ABI/testing/sysfs-class-intel_pmt +65 -0
index 926b5cf95fd1..67ca47123cbf 100644
@@ -52,3 +52,68 @@ Contact:	David Box <[email protected]>
 Description:
 		(RO) The offset of telemetry region in bytes that corresponds to
 		the mapping for the telem file.
+
+What:		/sys/class/intel_pmt/crashlog<x>
+Date:		October 2020
+KernelVersion:	5.10
+Contact:	Alexander Duyck <[email protected]>
+Description:
+		The crashlog<x> directory contains files for configuring an
+		instance of a PMT crashlog device that can perform crash data
+		recoring. Each crashlog<x> device has an associated crashlog
+		file. This file can be opened and mapped or read to access the
+		resulting crashlog buffer. The register layout for the buffer
+		can be determined from an XML file of specified guid for the
+		parent device.
+
+What:		/sys/class/intel_pmt/crashlog<x>/crashlog
+Date:		October 2020
+KernelVersion:	5.10
+Contact:	David Box <[email protected]>
+Description:
+		(RO) The crashlog buffer for this crashlog device. This file
+		may be mapped or read to obtain the data.
+
+What:		/sys/class/intel_pmt/crashlog<x>/guid
+Date:		October 2020
+KernelVersion:	5.10
+Contact:	Alexander Duyck <[email protected]>
+Description:
+		(RO) The guid for this crashlog device. The guid identifies the
+		version of the XML file for the parent device that should be
+		used to determine the register layout.
+
+What:		/sys/class/intel_pmt/crashlog<x>/size
+Date:		October 2020
+KernelVersion:	5.10
+Contact:	Alexander Duyck <[email protected]>
+Description:
+		(RO) The length of the result buffer in bytes that corresponds
+		to the size for the crashlog buffer.
+
+What:		/sys/class/intel_pmt/crashlog<x>/offset
+Date:		October 2020
+KernelVersion:	5.10
+Contact:	Alexander Duyck <[email protected]>
+Description:
+		(RO) The offset of the buffer in bytes that corresponds
+		to the mapping for the crashlog device.
+
+What:		/sys/class/intel_pmt/crashlog<x>/enable
+Date:		October 2020
+KernelVersion:	5.10
+Contact:	Alexander Duyck <[email protected]>
+Description:
+		(RW) Boolean value controlling if the crashlog functionality
+		is enabled for the crashlog device.
+
+What:		/sys/class/intel_pmt/crashlog<x>/trigger
+Date:		October 2020
+KernelVersion:	5.10
+Contact:	Alexander Duyck <[email protected]>
+Description:
+		(RW) Boolean value controlling the triggering of the crashlog
+		device node. When read it provides data on if the crashlog has
+		been triggered. When written to it can be used to either clear
+		the current trigger by writing false, or to trigger a new
+		event if the trigger is not currently set.
drivers/platform/x86/Kconfig +8 -0
index 8eae17a57a5b..529c5ee2eabf 100644
@@ -1377,6 +1377,14 @@ config INTEL_PMT_TELEMETRY
 	  access to hardware telemetry metrics on devices that support the
 	  feature.
 
+config INTEL_PMT_CRASHLOG
+	tristate "Intel Platform Monitoring Technology (PMT) Crashlog driver"
+	select INTEL_PMT_CLASS
+	help
+	  The Intel Platform Monitoring Technology (PMT) crashlog driver provides
+	  access to hardware crashlog capabilities on devices that support the
+	  feature.
+
 config INTEL_PUNIT_IPC
 	tristate "Intel P-Unit IPC Driver"
 	help
drivers/platform/x86/Makefile +1 -0
index 6a7b61f59ea8..ca82c1344977 100644
@@ -142,6 +142,7 @@ obj-$(CONFIG_INTEL_MRFLD_PWRBTN)	+= intel_mrfld_pwrbtn.o
 obj-$(CONFIG_INTEL_PMC_CORE)		+= intel_pmc_core.o intel_pmc_core_pltdrv.o
 obj-$(CONFIG_INTEL_PMT_CLASS)		+= intel_pmt_class.o
 obj-$(CONFIG_INTEL_PMT_TELEMETRY)	+= intel_pmt_telemetry.o
+obj-$(CONFIG_INTEL_PMT_CRASHLOG)	+= intel_pmt_crashlog.o
 obj-$(CONFIG_INTEL_PUNIT_IPC)		+= intel_punit_ipc.o
 obj-$(CONFIG_INTEL_SCU_IPC)		+= intel_scu_ipc.o
 obj-$(CONFIG_INTEL_SCU_PCI)		+= intel_scu_pcidrv.o
drivers/platform/x86/intel_pmt_crashlog.c new file +339 -0

339 lines not on this page · raw message

Signature · 1 line
2.20.1
author maintainer
#3 lore raw

PATCH v7 3/5 platform/x86: Intel PMT class driver

In reply to David E. Box

From: Alexander Duyck <[email protected]>

Intel Platform Monitoring Technology is meant to provide a common way to access telemetry and system metrics.

Register mappings are not provided by the driver. Instead, a GUID is read from a header for each endpoint. The GUID identifies the device and is to be used with an XML, provided by the vendor, to discover the available set of metrics and their register mapping. This allows firmware updates to modify the register space without needing to update the driver every time with new mappings. Firmware writes a new GUID in this case to specify the new mapping. Software tools with access to the associated XML file can then interpret the changes.

The module manages access to all Intel PMT endpoints on a system, independent of the device exporting them. It creates an intel_pmt class to manage the devices. For each telemetry endpoint, sysfs files provide GUID and size information as well as a pointer to the parent device the telemetry came from. Software may discover the association between endpoints and devices by iterating through the list in sysfs, or by looking for the existence of the class folder under the device of interest. A binary sysfs attribute of the same name allows software to then read or map the telemetry space for direct access.

---
 .../ABI/testing/sysfs-class-intel_pmt         |  54 ++++
 MAINTAINERS                                   |   1 +
 drivers/platform/x86/Kconfig                  |   9 +
 drivers/platform/x86/Makefile                 |   1 +
 drivers/platform/x86/intel_pmt_class.c        | 286 ++++++++++++++++++
 drivers/platform/x86/intel_pmt_class.h        |  57 ++++
 6 files changed, 408 insertions(+)
 create mode 100644 Documentation/ABI/testing/sysfs-class-intel_pmt
 create mode 100644 drivers/platform/x86/intel_pmt_class.c
 create mode 100644 drivers/platform/x86/intel_pmt_class.h
Documentation/ABI/testing/sysfs-class-intel_pmt new file +54 -0
new file mode 100644
index 000000000000..926b5cf95fd1
@@ -0,0 +1,54 @@
+What:		/sys/class/intel_pmt/
+Date:		October 2020
+KernelVersion:	5.10
+Contact:	David Box <[email protected]>
+Description:
+		The intel_pmt/ class directory contains information for
+		devices that expose hardware telemetry using Intel Platform
+		Monitoring Technology (PMT)
+
+What:		/sys/class/intel_pmt/telem<x>
+Date:		October 2020
+KernelVersion:	5.10
+Contact:	David Box <[email protected]>
+Description:
+		The telem<x> directory contains files describing an instance of
+		a PMT telemetry device that exposes hardware telemetry. Each
+		telem<x> directory has an associated telem file. This file
+		may be opened and mapped or read to access the telemetry space
+		of the device. The register layout of the telemetry space is
+		determined from an XML file that matches the PCI device id and
+		GUID for the device.
+
+What:		/sys/class/intel_pmt/telem<x>/telem
+Date:		October 2020
+KernelVersion:	5.10
+Contact:	David Box <[email protected]>
+Description:
+		(RO) The telemetry data for this telemetry device. This file
+		may be mapped or read to obtain the data.
+
+What:		/sys/class/intel_pmt/telem<x>/guid
+Date:		October 2020
+KernelVersion:	5.10
+Contact:	David Box <[email protected]>
+Description:
+		(RO) The GUID for this telemetry device. The GUID identifies
+		the version of the XML file for the parent device that is to
+		be used to get the register layout.
+
+What:		/sys/class/intel_pmt/telem<x>/size
+Date:		October 2020
+KernelVersion:	5.10
+Contact:	David Box <[email protected]>
+Description:
+		(RO) The size of telemetry region in bytes that corresponds to
+		the mapping size for the telem file.
+
+What:		/sys/class/intel_pmt/telem<x>/offset
+Date:		October 2020
+KernelVersion:	5.10
+Contact:	David Box <[email protected]>
+Description:
+		(RO) The offset of telemetry region in bytes that corresponds to
+		the mapping for the telem file.
MAINTAINERS +1 -0
index 0f2663b1d376..47fdb8a6e151 100644
@@ -8950,6 +8950,7 @@ INTEL PMT DRIVER
 M:	"David E. Box" <[email protected]>
 S:	Maintained
 F:	drivers/mfd/intel_pmt.c
+F:	drivers/platform/x86/intel_pmt_*
 
 INTEL PRO/WIRELESS 2100, 2200BG, 2915ABG NETWORK CONNECTION SUPPORT
 M:	Stanislav Yakovlev <[email protected]>
drivers/platform/x86/Kconfig +9 -0
index 40219bba6801..82465d0e8fd3 100644
@@ -1360,6 +1360,15 @@ config INTEL_PMC_CORE
 		- LTR Ignore
 		- MPHY/PLL gating status (Sunrisepoint PCH only)
 
+config INTEL_PMT_CLASS
+	tristate "Intel Platform Monitoring Technology (PMT) Class driver"
+	help
+	  The Intel Platform Monitoring Technology (PMT) class driver provides
+	  the basic sysfs interface and file hierarchy uses by PMT devices.
+
+	  For more information, see:
+	  <file:Documentation/ABI/testing/sysfs-class-intel_pmt>
+
 config INTEL_PUNIT_IPC
 	tristate "Intel P-Unit IPC Driver"
 	help
drivers/platform/x86/Makefile +1 -0
index 5f823f7eff45..f4b1f87f2401 100644
@@ -140,6 +140,7 @@ obj-$(CONFIG_INTEL_MFLD_THERMAL)	+= intel_mid_thermal.o
 obj-$(CONFIG_INTEL_MID_POWER_BUTTON)	+= intel_mid_powerbtn.o
 obj-$(CONFIG_INTEL_MRFLD_PWRBTN)	+= intel_mrfld_pwrbtn.o
 obj-$(CONFIG_INTEL_PMC_CORE)		+= intel_pmc_core.o intel_pmc_core_pltdrv.o
+obj-$(CONFIG_INTEL_PMT_CLASS)		+= intel_pmt_class.o
 obj-$(CONFIG_INTEL_PUNIT_IPC)		+= intel_punit_ipc.o
 obj-$(CONFIG_INTEL_SCU_IPC)		+= intel_scu_ipc.o
 obj-$(CONFIG_INTEL_SCU_PCI)		+= intel_scu_pcidrv.o
drivers/platform/x86/intel_pmt_class.c new file +286 -0

286 lines not on this page · raw message

drivers/platform/x86/intel_pmt_class.h new file +57 -0

57 lines not on this page · raw message

Signature · 1 line
2.20.1
author maintainer
#1 lore raw

PATCH v7 1/5 PCI: Add defines for Designated Vendor-Specific Extended Capability

In reply to David E. Box

Add PCIe Designated Vendor-Specific Extended Capability (DVSEC) and defines for the header offsets. Defined in PCIe r5.0, sec 7.9.6.

---
 include/uapi/linux/pci_regs.h | 5 +++++
 1 file changed, 5 insertions(+)
include/uapi/linux/pci_regs.h +5 -0
index f9701410d3b5..beafeee39e44 100644
@@ -720,6 +720,7 @@
 #define PCI_EXT_CAP_ID_DPC	0x1D	/* Downstream Port Containment */
 #define PCI_EXT_CAP_ID_L1SS	0x1E	/* L1 PM Substates */
 #define PCI_EXT_CAP_ID_PTM	0x1F	/* Precision Time Measurement */
+#define PCI_EXT_CAP_ID_DVSEC	0x23	/* Designated Vendor-Specific */
 #define PCI_EXT_CAP_ID_DLF	0x25	/* Data Link Feature */
 #define PCI_EXT_CAP_ID_PL_16GT	0x26	/* Physical Layer 16.0 GT/s */
 #define PCI_EXT_CAP_ID_MAX	PCI_EXT_CAP_ID_PL_16GT
@@ -1062,6 +1063,10 @@
 #define  PCI_L1SS_CTL1_LTR_L12_TH_SCALE	0xe0000000  /* LTR_L1.2_THRESHOLD_Scale */
 #define PCI_L1SS_CTL2		0x0c	/* Control 2 Register */
 
+/* Designated Vendor-Specific (DVSEC, PCI_EXT_CAP_ID_DVSEC) */
+#define PCI_DVSEC_HEADER1		0x4 /* Designated Vendor-Specific Header1 */
+#define PCI_DVSEC_HEADER2		0x8 /* Designated Vendor-Specific Header2 */
+
 /* Data Link Feature */
 #define PCI_DLF_CAP		0x04	/* Capabilities Register */
 #define  PCI_DLF_EXCHANGE_ENABLE	0x80000000  /* Data Link Feature Exchange Enable */
Signature · 1 line
2.20.1
author maintainer
#2 lore raw

PATCH v7 2/5 mfd: Intel Platform Monitoring Technology support

In reply to David E. Box

Intel Platform Monitoring Technology (PMT) is an architecture for enumerating and accessing hardware monitoring facilities. PMT supports multiple types of monitoring capabilities. This driver creates platform devices for each type so that they may be managed by capability specific drivers (to be introduced). Capabilities are discovered using PCIe DVSEC ids. Support is included for the 3 current capability types, Telemetry, Watcher, and Crashlog. The features are available on new Intel platforms starting from Tiger Lake for which support is added. This patch adds support for Tiger Lake (TGL), Alder Lake (ADL), and Out-of-Band Management Services Module (OOBMSM).

Also add a quirk mechanism for several early hardware differences and bugs. For Tiger Lake, do not support Watcher and Crashlog capabilities since they will not be compatible with future product. Also, fix use a quirk to fix the discovery table offset.

---
 MAINTAINERS             |   5 +
 drivers/mfd/Kconfig     |  10 ++
 drivers/mfd/Makefile    |   1 +
 drivers/mfd/intel_pmt.c | 225 ++++++++++++++++++++++++++++++++++++++++
 4 files changed, 241 insertions(+)
 create mode 100644 drivers/mfd/intel_pmt.c
MAINTAINERS +5 -0
index 190c7fa2ea01..0f2663b1d376 100644
@@ -8946,6 +8946,11 @@ F:	drivers/mfd/intel_soc_pmic*
 F:	include/linux/mfd/intel_msic.h
 F:	include/linux/mfd/intel_soc_pmic*
 
+INTEL PMT DRIVER
+M:	"David E. Box" <[email protected]>
+S:	Maintained
+F:	drivers/mfd/intel_pmt.c
+
 INTEL PRO/WIRELESS 2100, 2200BG, 2915ABG NETWORK CONNECTION SUPPORT
 M:	Stanislav Yakovlev <[email protected]>
 L:	[email protected]
drivers/mfd/Kconfig +10 -0
index 33df0837ab41..f092db50e518 100644
@@ -670,6 +670,16 @@ config MFD_INTEL_PMC_BXT
 	  Register and P-unit access. In addition this creates devices
 	  for iTCO watchdog and telemetry that are part of the PMC.
 
+config MFD_INTEL_PMT
+	tristate "Intel Platform Monitoring Technology (PMT) support"
+	depends on PCI
+	select MFD_CORE
+	help
+	  The Intel Platform Monitoring Technology (PMT) is an interface that
+	  provides access to hardware monitor registers. This driver supports
+	  Telemetry, Watcher, and Crashlog PMT capabilities/devices for
+	  platforms starting from Tiger Lake.
+
 config MFD_IPAQ_MICRO
 	bool "Atmel Micro ASIC (iPAQ h3100/h3600/h3700) Support"
 	depends on SA1100_H3100 || SA1100_H3600
drivers/mfd/Makefile +1 -0
index a60e5f835283..b9565d98ab09 100644
@@ -215,6 +215,7 @@ obj-$(CONFIG_MFD_INTEL_LPSS_PCI)	+= intel-lpss-pci.o
 obj-$(CONFIG_MFD_INTEL_LPSS_ACPI)	+= intel-lpss-acpi.o
 obj-$(CONFIG_MFD_INTEL_MSIC)	+= intel_msic.o
 obj-$(CONFIG_MFD_INTEL_PMC_BXT)	+= intel_pmc_bxt.o
+obj-$(CONFIG_MFD_INTEL_PMT)	+= intel_pmt.o
 obj-$(CONFIG_MFD_PALMAS)	+= palmas.o
 obj-$(CONFIG_MFD_VIPERBOARD)    += viperboard.o
 obj-$(CONFIG_MFD_RC5T583)	+= rc5t583.o rc5t583-irq.o
drivers/mfd/intel_pmt.c new file +225 -0

225 lines not on this page · raw message

Signature · 1 line
2.20.1
author maintainer
#8 lore raw

PATCH v7 4/5 platform/x86: Intel PMT Telemetry capability driver

In reply to David E. Box

From: Alexander Duyck <[email protected]>

PMT Telemetry is a capability of the Intel Platform Monitoring Technology. The Telemetry capability provides access to device telemetry metrics that provide hardware performance data to users from read-only register spaces.

With this driver present the intel_pmt directory can be populated with telem<x> devices. These devices will contain the standard intel_pmt sysfs data and a "telem" binary sysfs attribute which can be used to access the telemetry data.

---
 drivers/platform/x86/Kconfig               |   8 ++
 drivers/platform/x86/Makefile              |   1 +
 drivers/platform/x86/intel_pmt_telemetry.c | 158 +++++++++++++++++++++
 3 files changed, 167 insertions(+)
 create mode 100644 drivers/platform/x86/intel_pmt_telemetry.c
drivers/platform/x86/Kconfig +8 -0
index 82465d0e8fd3..8eae17a57a5b 100644
@@ -1369,6 +1369,14 @@ config INTEL_PMT_CLASS
 	  For more information, see:
 	  <file:Documentation/ABI/testing/sysfs-class-intel_pmt>
 
+config INTEL_PMT_TELEMETRY
+	tristate "Intel Platform Monitoring Technology (PMT) Telemetry driver"
+	select INTEL_PMT_CLASS
+	help
+	  The Intel Platform Monitory Technology (PMT) Telemetry driver provides
+	  access to hardware telemetry metrics on devices that support the
+	  feature.
+
 config INTEL_PUNIT_IPC
 	tristate "Intel P-Unit IPC Driver"
 	help
drivers/platform/x86/Makefile +1 -0
index f4b1f87f2401..6a7b61f59ea8 100644
@@ -141,6 +141,7 @@ obj-$(CONFIG_INTEL_MID_POWER_BUTTON)	+= intel_mid_powerbtn.o
 obj-$(CONFIG_INTEL_MRFLD_PWRBTN)	+= intel_mrfld_pwrbtn.o
 obj-$(CONFIG_INTEL_PMC_CORE)		+= intel_pmc_core.o intel_pmc_core_pltdrv.o
 obj-$(CONFIG_INTEL_PMT_CLASS)		+= intel_pmt_class.o
+obj-$(CONFIG_INTEL_PMT_TELEMETRY)	+= intel_pmt_telemetry.o
 obj-$(CONFIG_INTEL_PUNIT_IPC)		+= intel_punit_ipc.o
 obj-$(CONFIG_INTEL_SCU_IPC)		+= intel_scu_ipc.o
 obj-$(CONFIG_INTEL_SCU_PCI)		+= intel_scu_pcidrv.o
drivers/platform/x86/intel_pmt_telemetry.c new file +158 -0

158 lines not on this page · raw message

Signature · 1 line
2.20.1
#9 lore raw

Re: [PATCH V7 4/5] platform/x86: Intel PMT Telemetry capability driver

Quoted 11 lines · David E. Box

From: Alexander Duyck <[email protected]>

PMT Telemetry is a capability of the Intel Platform Monitoring Technology. The Telemetry capability provides access to device telemetry metrics that provide hardware performance data to users from read-only register spaces.

With this driver present the intel_pmt directory can be populated with telem<x> devices. These devices will contain the standard intel_pmt sysfs data and a "telem" binary sysfs attribute which can be used to access the telemetry data.

...

drivers/platform/x86/intel_pmt_telemetry.c line 39

Quoted 4 lines
drivers/platform/x86/intel_pmt_telemetry.c new file +4 -0
+static DEFINE_XARRAY_ALLOC(telem_array);
+static struct intel_pmt_namespace pmt_telem_ns = {
+       .name = "telem",
+       .xa = &telem_array

Leave comma at the end.

drivers/platform/x86/intel_pmt_telemetry.c line 44

Quoted 5 lines
drivers/platform/x86/intel_pmt_telemetry.c new file +5 -0
+};
+
+/*
+ * driver initialization
+ */

This is a useless comment.

drivers/platform/x86/intel_pmt_telemetry.c line 97

Quoted 4 lines
drivers/platform/x86/intel_pmt_telemetry.c new file +4 -0
+       size = offsetof(struct pmt_telem_priv, entry[pdev->num_resources]);
+       priv = devm_kzalloc(&pdev->dev, size, GFP_KERNEL);
+       if (!priv)
+               return -ENOMEM;

Please, use struct_size() from overflow.h instead of custom approach.

...

drivers/platform/x86/intel_pmt_telemetry.c line 136

Quoted 3 lines
drivers/platform/x86/intel_pmt_telemetry.c new file +3 -0
+static struct platform_driver pmt_telem_driver = {
+       .driver = {
+               .name   = TELEM_DEV_NAME,

I'm not sure I have interpreted this:
- Use 'raw' string instead of defines for device names
correctly. Can you elaborate?

Quoted 4 lines
drivers/platform/x86/intel_pmt_telemetry.c new file +4 -0
+       },
+       .remove = pmt_telem_remove,
+       .probe  = pmt_telem_probe,
+};

...

Quoted 1 line

+MODULE_ALIAS("platform:" TELEM_DEV_NAME);

Ditto.

--
With Best Regards,
Andy Shevchenko

#4 lore raw

Re: [PATCH V7 3/5] platform/x86: Intel PMT class driver

In reply to David E. Box
Quoted 6 lines · David E. Box

From: Alexander Duyck <[email protected]>

Intel Platform Monitoring Technology is meant to provide a common way to access telemetry and system metrics.

Quoted 8 lines

Register mappings are not provided by the driver. Instead, a GUID is read from a header for each endpoint. The GUID identifies the device and is to be used with an XML, provided by the vendor, to discover the available set of metrics and their register mapping. This allows firmware updates to modify the register space without needing to update the driver every time with new mappings. Firmware writes a new GUID in this case to specify the new mapping. Software tools with access to the associated XML file can then interpret the changes.

Where one may find a database of these reserved GUIDs / XMLs? How do you prevent a chaos which happens with other registries?

Quoted 9 lines

The module manages access to all Intel PMT endpoints on a system, independent of the device exporting them. It creates an intel_pmt class to manage the devices. For each telemetry endpoint, sysfs files provide GUID and size information as well as a pointer to the parent device the telemetry came from. Software may discover the association between endpoints and devices by iterating through the list in sysfs, or by looking for the existence of the class folder under the device of interest. A binary sysfs attribute of the same name allows software to then read or map the telemetry space for direct access.

What are the security implications by direct access?

...

drivers/platform/x86/intel_pmt_class.c line 34

Quoted 11 lines
drivers/platform/x86/intel_pmt_class.c new file +11 -0
+static const struct pci_device_id pmt_telem_early_client_pci_ids[] = {
+       { PCI_VDEVICE(INTEL, 0x9a0d) }, /* TGL */
+       { }
+};
+bool intel_pmt_is_early_client_hw(struct device *dev)
+{
+       struct pci_dev *parent = to_pci_dev(dev->parent);
+
+       return !!pci_match_id(pmt_telem_early_client_pci_ids, parent);
+}
+EXPORT_SYMBOL_GPL(intel_pmt_is_early_client_hw);

What is this and why is it in the class driver?

drivers/platform/x86/intel_pmt_class.c line 49

Quoted 11 lines
drivers/platform/x86/intel_pmt_class.c new file +10 -0
+static ssize_t
+intel_pmt_read(struct file *filp, struct kobject *kobj,
+              struct bin_attribute *attr, char *buf, loff_t off,
+              size_t count)
+{
+       struct intel_pmt_entry *entry = container_of(attr,
+                                                    struct intel_pmt_entry,
+                                                    pmt_bin_attr);
 
+       if (off < 0)
+               return -EINVAL;

Is this real or theoretical?

Quoted 1 line

+ if (count)

Useless.

Quoted 4 lines
drivers/platform/x86/intel_pmt_class.c new file +4 -0
+               memcpy_fromio(buf, entry->base + off, count);
+
+       return count;
+}

...

Quoted 1 line

+ psize = (PFN_UP(entry->base_addr + entry->size) - pfn) * PAGE_SIZE;

PFN_PHYS(PFN_UP(...)) ?

...

drivers/platform/x86/intel_pmt_class.c line 126

Quoted 8 lines
drivers/platform/x86/intel_pmt_class.c new file +7 -0
+static struct attribute *intel_pmt_attrs[] = {
+       &dev_attr_guid.attr,
+       &dev_attr_size.attr,
+       &dev_attr_offset.attr,
+       NULL
+};
 
+

Unneeded blank line.

Quoted 1 line

+ATTRIBUTE_GROUPS(intel_pmt);

...

drivers/platform/x86/intel_pmt_class.c line 219

Quoted 3 lines
drivers/platform/x86/intel_pmt_class.c new file +3 -0
+       /* if size is 0 assume no data buffer, so no file needed */
+       if (!entry->size)
+               return 0;

Hmm... But presence of the file is also an information that might be useful for user, no?

...

Quoted 1 line

+ entry->base = devm_ioremap_resource(dev, &res);

(1)

drivers/platform/x86/intel_pmt_class.c line 227

Quoted 3 lines
drivers/platform/x86/intel_pmt_class.c new file +2 -0
+       if (IS_ERR(entry->base)) {
 
+               dev_err(dev, "Failed to ioremap device region\n");

Duplicates core message.

Quoted 1 line

+ ret = -EIO;

Why shadowing real error code?

drivers/platform/x86/intel_pmt_class.c line 243

Quoted 4 lines
drivers/platform/x86/intel_pmt_class.c new file +3 -0
+               goto fail_ioremap;
+       }
 
+       iounmap(entry->base);

This is interesting. How do you avoid double unmap with (1)?

drivers/platform/x86/intel_pmt_class.h line 14

Quoted 10 lines
drivers/platform/x86/intel_pmt_class.h new file +10 -0
+#include <linux/platform_device.h>
+#include <linux/xarray.h>
+
+/* PMT access types */
+#define ACCESS_BARID           2
+#define ACCESS_LOCAL           3
+
+/* PMT discovery base address/offset register layout */
+#define GET_BIR(v)             ((v) & GENMASK(2, 0))
+#define GET_ADDRESS(v)         ((v) & GENMASK(31, 3))

bits.h

drivers/platform/x86/intel_pmt_class.h line 23

Quoted 9 lines
drivers/platform/x86/intel_pmt_class.h new file +8 -0
+struct intel_pmt_entry {
+       struct bin_attribute    pmt_bin_attr;
+       struct kobject          *kobj;
+       void __iomem            *disc_table;
+       void __iomem            *base;
+       unsigned long           base_addr;
+       size_t                  size;
 
+       u32                     guid;

types.h

drivers/platform/x86/intel_pmt_class.h line 53

Quoted 9 lines
drivers/platform/x86/intel_pmt_class.h new file +7 -0
+       int                     devid;
+};
 
+static inline int
+intel_pmt_ioremap_discovery_table(struct intel_pmt_entry *entry,
+                                 struct platform_device *pdev,  int i)
+{
 
+       entry->disc_table = devm_platform_ioremap_resource(pdev, i);

io.h ?

drivers/platform/x86/intel_pmt_class.h line 55

Quoted 2 lines
drivers/platform/x86/intel_pmt_class.h new file +2 -0
+
+       return PTR_ERR_OR_ZERO(entry->disc_table);

err.h

Quoted 1 line

+}

The rule of thumb is to include all headers that you have direct users of. Then you may optimize by removing those which are guaranteed to be included by others, like
bits.h always included by bitops.h.

Signature · 2 lines
With Best Regards,
Andy Shevchenko
#13 lore raw

Re: [PATCH V7 5/5] platform/x86: Intel PMT Crashlog capability driver

In reply to David E. Box
Quoted 9 lines · David E. Box

Add support for the Intel Platform Monitoring Technology crashlog interface. This interface provides a few sysfs values to allow for controlling the crashlog telemetry interface as well as a character driver to allow for mapping the crashlog memory region so that it can be accessed after a crashlog has been recorded.

This driver is meant to only support the server version of the crashlog which is identified as crash_type 1 with a version of zero. Currently no other types are supported.

...

Documentation/ABI/testing/sysfs-class-intel_pmt Contact: David Box <[email protected]> · line 63

Quoted 3 lines
Documentation/ABI/testing/sysfs-class-intel_pmt +3 -0
+               The crashlog<x> directory contains files for configuring an
+               instance of a PMT crashlog device that can perform crash data
+               recoring. Each crashlog<x> device has an associated crashlog

recording

Quoted 4 lines
Documentation/ABI/testing/sysfs-class-intel_pmt +4 -0
+               file. This file can be opened and mapped or read to access the
+               resulting crashlog buffer. The register layout for the buffer
+               can be determined from an XML file of specified guid for the
+               parent device.

...

Quoted 1 line

+ (RO) The guid for this crashlog device. The guid identifies the

guid -> GUID

Please, spell check all ABI files in this series.

...

drivers/platform/x86/Kconfig config INTEL_PMT_TELEMETRY · line 1386

Quoted 7 lines
drivers/platform/x86/Kconfig +7 -0
+config INTEL_PMT_CRASHLOG
+       tristate "Intel Platform Monitoring Technology (PMT) Crashlog driver"
+       select INTEL_PMT_CLASS
+       help
+         The Intel Platform Monitoring Technology (PMT) crashlog driver provides
+         access to hardware crashlog capabilities on devices that support the
+         feature.

Name of the module?

...

drivers/platform/x86/intel_pmt_crashlog.c line 81

Quoted 3 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +2 -0
+       /*
 
+        * Currenty we only recognize OOBMSM version 0 devices.

Currently. Please spell check all comments in the code.

Quoted 2 lines
(quoted diff) +2 -0
+        * We can ignore all other crashlog devices in the system.
+        */

...

Quoted 1 line

+ /* clear control bits */

What new information readers get from this comment?

Quoted 1 line

+ control &= ~(CRASHLOG_FLAG_MASK | CRASHLOG_FLAG_DISABLE);

How does the second constant play any role here?

...

Quoted 1 line

+ /* clear control bits */

Ditto. And moreover it's ambiguous due to joined two lines below.

Quoted 2 lines
(quoted diff) +2 -0
+       control &= ~CRASHLOG_FLAG_MASK;
+       control |= CRASHLOG_FLAG_EXECUTE;

...

Quoted 1 line

+ return strnlen(buf, count);

How is this different to count?

...

drivers/platform/x86/intel_pmt_crashlog.c line 179

Quoted 6 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +5 -0
+       struct crashlog_entry *entry;
+       bool trigger;
+       int result;
+
 
+       entry = dev_get_drvdata(dev);

You may reduce LOCs by direct assigning in the definition block above.

...

Quoted 1 line

+ result = strnlen(buf, count);

How is it different from count?

...

drivers/platform/x86/intel_pmt_crashlog.c line 222

Quoted 5 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +5 -0
+static DEFINE_XARRAY_ALLOC(crashlog_array);
+static struct intel_pmt_namespace pmt_crashlog_ns = {
+       .name = "crashlog",
+       .xa = &crashlog_array,
+       .attr_grp = &pmt_crashlog_group

Leave the comma here.

Quoted 1 line

+};

...

drivers/platform/x86/intel_pmt_crashlog.c line 254

Quoted 8 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +8 -0
+       ret = intel_pmt_dev_create(entry, &pmt_crashlog_ns, parent);
+       if (!ret)
+               return 0;
+
+       dev_err(parent, "Failed to add crashlog controls\n");
+       intel_pmt_dev_destroy(entry, &pmt_crashlog_ns);
+
+       return ret;

Can we use traditional patterns?
if (ret) {
...
}
return ret;

...

drivers/platform/x86/intel_pmt_crashlog.c line 277

Quoted 4 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +4 -0
+       size = offsetof(struct pmt_crashlog_priv, entry[pdev->num_resources]);
+       priv = devm_kzalloc(&pdev->dev, size, GFP_KERNEL);
+       if (!priv)
+               return -ENOMEM;

struct_size()

...

Quoted 9 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +9 -0
+               /* initialize control mutex */
+               mutex_init(&priv->entry[i].control_mutex);
+
+               disc_res = platform_get_resource(pdev, IORESOURCE_MEM, i);
+               if (!disc_res)
+                       goto abort_probe;
+
+               ret = intel_pmt_ioremap_discovery_table(entry, pdev, i);
+               if (ret)

drivers/platform/x86/intel_pmt_crashlog.c line 305

Quoted 10 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +10 -0
+                       goto abort_probe;
+
+               if (!pmt_crashlog_supported(entry))
+                       continue;
+
+               ret = pmt_crashlog_add_entry(entry, &pdev->dev, disc_res);
+               if (ret)
+                       goto abort_probe;
+
+               priv->num_entries++;

Are you going to duplicate this in each driver? Consider to refactor to avoid duplication of a lot of code.

...

drivers/platform/x86/intel_pmt_crashlog.c line 338

Quoted 3 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +2 -0
+               .name   = DRV_NAME,
 
+MODULE_ALIAS("platform:" DRV_NAME);

I'm not sure I have interpreted this:
- Use 'raw' string instead of defines for device names
correctly. Can you elaborate?

--
With Best Regards,
Andy Shevchenko

#10 lore raw

Re: [PATCH V7 4/5] platform/x86: Intel PMT Telemetry capability driver

In reply to Andy Shevchenko
Quoted 31 lines · Andy Shevchenko
Quoted 11 lines · David E. Box

From: Alexander Duyck <[email protected]>

PMT Telemetry is a capability of the Intel Platform Monitoring Technology. The Telemetry capability provides access to device telemetry metrics that provide hardware performance data to users from read-only register spaces.

With this driver present the intel_pmt directory can be populated with telem<x> devices. These devices will contain the standard intel_pmt sysfs data and a "telem" binary sysfs attribute which can be used to access the telemetry data.

...

drivers/platform/x86/intel_pmt_telemetry.c line 39

Quoted 4 lines
drivers/platform/x86/intel_pmt_telemetry.c new file +4 -0
+static DEFINE_XARRAY_ALLOC(telem_array);
+static struct intel_pmt_namespace pmt_telem_ns = {
+       .name = "telem",
+       .xa = &telem_array

Leave comma at the end.

drivers/platform/x86/intel_pmt_telemetry.c line 44

Quoted 5 lines
drivers/platform/x86/intel_pmt_telemetry.c new file +5 -0
+};
+
+/*
+ * driver initialization
+ */

This is a useless comment.

Quoted 8 lines

drivers/platform/x86/intel_pmt_telemetry.c line 97

Quoted 4 lines
drivers/platform/x86/intel_pmt_telemetry.c new file +4 -0
+       size = offsetof(struct pmt_telem_priv, entry[pdev->num_resources]);
+       priv = devm_kzalloc(&pdev->dev, size, GFP_KERNEL);
+       if (!priv)
+               return -ENOMEM;

Please, use struct_size() from overflow.h instead of custom approach.

...

So all of the above make sense and can be fixed shortly and pushed as a v8 for both the telemetry and crashlog drivers.

Quoted 7 lines

drivers/platform/x86/intel_pmt_telemetry.c line 136

Quoted 3 lines
drivers/platform/x86/intel_pmt_telemetry.c new file +3 -0
+static struct platform_driver pmt_telem_driver = {
+       .driver = {
+               .name   = TELEM_DEV_NAME,

I'm not sure I have interpreted this:
- Use 'raw' string instead of defines for device names
correctly. Can you elaborate?

Can you point me to a reference of that? I'm not sure what you are referring to.

Quoted 10 lines
Quoted 4 lines
drivers/platform/x86/intel_pmt_telemetry.c new file +4 -0
+       },
+       .remove = pmt_telem_remove,
+       .probe  = pmt_telem_probe,
+};

...

Quoted 1 line

+MODULE_ALIAS("platform:" TELEM_DEV_NAME);

Ditto.

This doesn't make sense to me. Are you saying we are expected to use "pmt_telemetry" everywhere instead of the define? It seems like that would be much more error prone. It seems like common practice to use DRV_NAME throughout a driver for these sort of things so if you are wanting us to rename it to that I am fine with that, but I am not sure getting rid of the use of a define makes sense.

#5 lore raw

Re: [PATCH V7 3/5] platform/x86: Intel PMT class driver

In reply to Andy Shevchenko
Quoted 17 lines · Andy Shevchenko
Quoted 6 lines · David E. Box

From: Alexander Duyck <[email protected]>

Intel Platform Monitoring Technology is meant to provide a common way to access telemetry and system metrics.

Quoted 8 lines

Register mappings are not provided by the driver. Instead, a GUID is read from a header for each endpoint. The GUID identifies the device and is to be used with an XML, provided by the vendor, to discover the available set of metrics and their register mapping. This allows firmware updates to modify the register space without needing to update the driver every time with new mappings. Firmware writes a new GUID in this case to specify the new mapping. Software tools with access to the associated XML file can then interpret the changes.

Quoted 2 lines

Where one may find a database of these reserved GUIDs / XMLs? How do you prevent a chaos which happens with other registries?

The database will be posted on intel.com eventually. Although I don't believe the URL is public yet.

Quoted 11 lines
Quoted 9 lines

The module manages access to all Intel PMT endpoints on a system, independent of the device exporting them. It creates an intel_pmt class to manage the devices. For each telemetry endpoint, sysfs files provide GUID and size information as well as a pointer to the parent device the telemetry came from. Software may discover the association between endpoints and devices by iterating through the list in sysfs, or by looking for the existence of the class folder under the device of interest. A binary sysfs attribute of the same name allows software to then read or map the telemetry space for direct access.

What are the security implications by direct access?

In this case minimal as it would really be no different than the read. The registers in the memory regions themselves are read-only with no read side effects.

Quoted 15 lines

...

drivers/platform/x86/intel_pmt_class.c line 34

Quoted 11 lines
drivers/platform/x86/intel_pmt_class.c new file +11 -0
+static const struct pci_device_id pmt_telem_early_client_pci_ids[] = {
+       { PCI_VDEVICE(INTEL, 0x9a0d) }, /* TGL */
+       { }
+};
+bool intel_pmt_is_early_client_hw(struct device *dev)
+{
+       struct pci_dev *parent = to_pci_dev(dev->parent);
+
+       return !!pci_match_id(pmt_telem_early_client_pci_ids, parent);
+}
+EXPORT_SYMBOL_GPL(intel_pmt_is_early_client_hw);

What is this and why is it in the class driver?

I chose to use the class driver as a central place to store code common to all of the instances of the class. In this case we have quirks that are specific to Tiger Lake and so I chose to store the function to test for the device here.

Quoted 9 lines
Quoted 8 lines
drivers/platform/x86/intel_pmt_class.c new file +8 -0
+static ssize_t
+intel_pmt_read(struct file *filp, struct kobject *kobj,
+              struct bin_attribute *attr, char *buf, loff_t off,
+              size_t count)
+{
+       struct intel_pmt_entry *entry = container_of(attr,
+                                                    struct intel_pmt_entry,
+                                                    pmt_bin_attr);
Quoted 3 lines

Review

Quoted 2 lines
(quoted diff) +2 -0
+       if (off < 0)
+               return -EINVAL;

Is this real or theoretical?

Not sure. I am not that familiar with the interface. It was something I copied from read_bmof which is what I based this code on based on an earlier suggestion.

Quoted 3 lines
Quoted 1 line

+ if (count)

Useless.

I'm assuming that is because memcpy_fromio is assumed to handle this case?

Quoted 10 lines
Quoted 4 lines
drivers/platform/x86/intel_pmt_class.c new file +4 -0
+               memcpy_fromio(buf, entry->base + off, count);
+
+       return count;
+}

...

Quoted 1 line

+ psize = (PFN_UP(entry->base_addr + entry->size) - pfn) * PAGE_SIZE;

PFN_PHYS(PFN_UP(...)) ?

I'm not sure how that would work. Basically what we are doing here is determining the size of the mapping based on the number of pages that will be needed. So we wake the pfn of the start of the region, subtract that from the pfn for the end of the region and multiply by the size of a page.

Quoted 13 lines

drivers/platform/x86/intel_pmt_class.c line 126

Quoted 8 lines
drivers/platform/x86/intel_pmt_class.c new file +7 -0
+static struct attribute *intel_pmt_attrs[] = {
+       &dev_attr_guid.attr,
+       &dev_attr_size.attr,
+       &dev_attr_offset.attr,
+       NULL
+};
 
+

Unneeded blank line.

Quoted 1 line

+ATTRIBUTE_GROUPS(intel_pmt);

Quoted 8 lines

...

drivers/platform/x86/intel_pmt_class.c line 219

Quoted 3 lines
drivers/platform/x86/intel_pmt_class.c new file +3 -0
+       /* if size is 0 assume no data buffer, so no file needed */
+       if (!entry->size)
+               return 0;

Hmm... But presence of the file is also an information that might be useful for user, no?

I'm not sure what you mean? If the size of the region is zero it means there is no data there. There are cases in future devices where we may have controls for a telemetry function that is actually streaming the data elsewhere. That will be the use case for the entry size of 0 and in that case it doesn't make any sense to have a file as there is no data to present in it.

Quoted 11 lines

...

Quoted 1 line

+ entry->base = devm_ioremap_resource(dev, &res);

(1)

drivers/platform/x86/intel_pmt_class.c line 227

Quoted 3 lines
drivers/platform/x86/intel_pmt_class.c new file +2 -0
+       if (IS_ERR(entry->base)) {
 
+               dev_err(dev, "Failed to ioremap device region\n");

Duplicates core message.

I'll drop it since it is a redundant.

Quoted 3 lines
Quoted 1 line

+ ret = -EIO;

Why shadowing real error code?

I will just convert it instead. I think there might have been a bug here in an earlier version where it was testing for NULL instead.

Quoted 6 lines

drivers/platform/x86/intel_pmt_class.c line 243

Quoted 4 lines
drivers/platform/x86/intel_pmt_class.c new file +3 -0
+               goto fail_ioremap;
+       }
 
+       iounmap(entry->base);

This is interesting. How do you avoid double unmap with (1)?

I think I get what you are trying to say. This is redundant since we used the devm_ioremap_resource it will already be freed when the driver is detached, correct?

Quoted 4 lines
Quoted 4 lines
drivers/platform/x86/intel_pmt_class.h new file +4 -0
+#include <linux/platform_device.h>
+#include <linux/xarray.h>
+
+/* PMT access types */
Quoted 8 lines

drivers/platform/x86/intel_pmt_class.h line 14

Quoted 6 lines
drivers/platform/x86/intel_pmt_class.h new file +6 -0
+#define ACCESS_BARID           2
+#define ACCESS_LOCAL           3
+
+/* PMT discovery base address/offset register layout */
+#define GET_BIR(v)             ((v) & GENMASK(2, 0))
+#define GET_ADDRESS(v)         ((v) & GENMASK(31, 3))

bits.h

That is already included from a few different sources.

Quoted 11 lines

drivers/platform/x86/intel_pmt_class.h line 23

Quoted 9 lines
drivers/platform/x86/intel_pmt_class.h new file +8 -0
+struct intel_pmt_entry {
+       struct bin_attribute    pmt_bin_attr;
+       struct kobject          *kobj;
+       void __iomem            *disc_tabl
+       void __iomem            *base;
+       unsigned long           base_addr;
+       size_t                  size;
 
+       u32                     guid;

types.h

Is included through xarray.h.

Quoted 11 lines

drivers/platform/x86/intel_pmt_class.h line 53

Quoted 9 lines
drivers/platform/x86/intel_pmt_class.h new file +7 -0
+       int                     devid;
+};
 
+static inline int
+intel_pmt_ioremap_discovery_table(struct intel_pmt_entry *entry,
+                                 struct platform_device *pdev,  int i)
+{
 
+       entry->disc_table = devm_platform_ioremap_resource(pdev, i);

io.h ?

That one I will move from the class.c file.

Quoted 4 lines

drivers/platform/x86/intel_pmt_class.h line 55

Quoted 2 lines
drivers/platform/x86/intel_pmt_class.h new file +2 -0
+
+       return PTR_ERR_OR_ZERO(entry->disc_table);

err.h

That is included as a part of io.h.

Quoted 5 lines
Quoted 1 line

+}

The rule of thumb is to include all headers that you have direct users of. Then you may optimize by removing those which are guaranteed to be included by others, like bits.h always included by bitops.h.

Yeah, from what I can tell the only one I didn't have was io.h and part of that is because this was something I had moved to the header file in order to commonize it since it was being used in the other drivers.

#11 lore raw

Re: [PATCH V7 4/5] platform/x86: Intel PMT Telemetry capability driver

In reply to Alexander Duyck
Quoted 7 lines · Alexander Duyck
Quoted 4 lines · Andy Shevchenko
Quoted 3 lines
drivers/platform/x86/intel_pmt_telemetry.c new file +3 -0
+static struct platform_driver pmt_telem_driver = {
+       .driver = {
+               .name   = TELEM_DEV_NAME,
Quoted 5 lines
Quoted 3 lines

I'm not sure I have interpreted this:
- Use 'raw' string instead of defines for device names
correctly. Can you elaborate?

Can you point me to a reference of that? I'm not sure what you are referring to.

It's a changelog of this very series.

Quoted 9 lines
Quoted 9 lines
Quoted 4 lines
drivers/platform/x86/intel_pmt_telemetry.c new file +4 -0
+       },
+       .remove = pmt_telem_remove,
+       .probe  = pmt_telem_probe,
+};

...

Quoted 1 line

+MODULE_ALIAS("platform:" TELEM_DEV_NAME);

Quoted 8 lines
Quoted 1 line

Ditto.

This doesn't make sense to me. Are you saying we are expected to use "pmt_telemetry" everywhere instead of the define? It seems like that would be much more error prone. It seems like common practice to use DRV_NAME throughout a driver for these sort of things so if you are wanting us to rename it to that I am fine with that, but I am not sure getting rid of the use of a define makes sense.

I'm just wondering why changelog mentioned that and what it meant. If it's indeed conversion to explicit naming, then this has to be followed (somebody seems asked for that, right?) or commented why not. Or maybe I understood it wrong?

Signature · 2 lines
With Best Regards,
Andy Shevchenko
#6 lore raw

Re: [PATCH V7 3/5] platform/x86: Intel PMT class driver

In reply to Alexander Duyck
Quoted 3 lines · Alexander Duyck
Quoted 1 line · Andy Shevchenko

On Thu, Oct 1, 2020 at 4:43 AM David E. Box <[email protected]> wrote:

...

Quoted 12 lines
Quoted 12 lines
Quoted 11 lines

Intel Platform Monitoring Technology is meant to provide a common way to access telemetry and system metrics.

Register mappings are not provided by the driver. Instead, a GUID is read from a header for each endpoint. The GUID identifies the device and is to be used with an XML, provided by the vendor, to discover the available set of metrics and their register mapping. This allows firmware updates to modify the register space without needing to update the driver every time with new mappings. Firmware writes a new GUID in this case to specify the new mapping. Software tools with access to the associated XML file can then interpret the changes.

Quoted 5 lines
Quoted 2 lines

Where one may find a database of these reserved GUIDs / XMLs? How do you prevent a chaos which happens with other registries?

The database will be posted on intel.com eventually. Although I don't believe the URL is public yet.

How can we be sure that this won't be forgotten? How can we be sure it will be public at the end? Please, elaborate this in the commit message.

...

Quoted 14 lines
Quoted 5 lines
Quoted 5 lines
drivers/platform/x86/intel_pmt_class.c new file +5 -0
+static const struct pci_device_id pmt_telem_early_client_pci_ids[] = {
+       { PCI_VDEVICE(INTEL, 0x9a0d) }, /* TGL */
+       { }
+};
+bool intel_pmt_is_early_client_hw(struct device *dev)
Quoted 8 lines

drivers/platform/x86/intel_pmt_class.c line 34

Quoted 6 lines
drivers/platform/x86/intel_pmt_class.c new file +6 -0
+{
+       struct pci_dev *parent = to_pci_dev(dev->parent);
+
+       return !!pci_match_id(pmt_telem_early_client_pci_ids, parent);
+}
+EXPORT_SYMBOL_GPL(intel_pmt_is_early_client_hw);

What is this and why is it in the class driver?

Quoted 4 lines

I chose to use the class driver as a central place to store code common to all of the instances of the class. In this case we have quirks that are specific to Tiger Lake and so I chose to store the function to test for the device here.

Can it be done in another file module at least (let's say intel_pmt_quirks.c)?

...

Quoted 7 lines
Quoted 3 lines

Review

Quoted 2 lines
(quoted diff) +2 -0
+       if (off < 0)
+               return -EINVAL;

Is this real or theoretical?

Not sure. I am not that familiar with the interface. It was something I copied from read_bmof which is what I based this code on based on an earlier suggestion.

I'm not a fan of cargo cult when there is no understanding why certain code appears in the driver.

...

Quoted 5 lines
Quoted 3 lines
Quoted 1 line

+ if (count)

Useless.

I'm assuming that is because memcpy_fromio is assumed to handle this case?

Right.

Quoted 1 line
Quoted 1 line
Quoted 1 line

+ memcpy_fromio(buf, entry->base + off, count);

...

Quoted 9 lines
Quoted 3 lines
Quoted 1 line

+ psize = (PFN_UP(entry->base_addr + entry->size) - pfn) * PAGE_SIZE;

PFN_PHYS(PFN_UP(...)) ?

I'm not sure how that would work. Basically what we are doing here is determining the size of the mapping based on the number of pages that will be needed. So we wake the pfn of the start of the region, subtract that from the pfn for the end of the region and multiply by the size of a page.

PFN_PHYS() is a replacement for multiplication. You may check its implementation.

...

Quoted 7 lines
Quoted 6 lines

drivers/platform/x86/intel_pmt_class.c line 219

Quoted 3 lines
drivers/platform/x86/intel_pmt_class.c new file +3 -0
+       /* if size is 0 assume no data buffer, so no file needed */
+       if (!entry->size)
+               return 0;

Hmm... But presence of the file is also an information that might be useful for user, no?

Quoted 6 lines

I'm not sure what you mean? If the size of the region is zero it means there is no data there. There are cases in future devices where we may have controls for a telemetry function that is actually streaming the data elsewhere. That will be the use case for the entry size of 0 and in that case it doesn't make any sense to have a file as there is no data to present in it.

This is not understandable from 'no data buffer' above. Can you elaborate in the comment?

...

Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
drivers/platform/x86/intel_pmt_class.c new file +5 -0
+       entry->base = devm_ioremap_resource(dev, &res);
+       if (IS_ERR(entry->base)) {
+               dev_err(dev, "Failed to ioremap device region\n");
+               goto fail_ioremap;
+       }
Quoted 7 lines
Quoted 3 lines
Quoted 1 line

+ iounmap(entry->base);

This is interesting. How do you avoid double unmap with (1)?

I think I get what you are trying to say. This is redundant since we used the devm_ioremap_resource it will already be freed when the driver is detached, correct?

Right. Above is the leftover that shows the poor testing of v7.

Quoted 13 lines
Quoted 4 lines
Quoted 4 lines
drivers/platform/x86/intel_pmt_class.h new file +4 -0
+#include <linux/platform_device.h>
+#include <linux/xarray.h>
+
+/* PMT access types */
Quoted 8 lines

drivers/platform/x86/intel_pmt_class.h line 14

Quoted 6 lines
drivers/platform/x86/intel_pmt_class.h new file +6 -0
+#define ACCESS_BARID           2
+#define ACCESS_LOCAL           3
+
+/* PMT discovery base address/offset register layout */
+#define GET_BIR(v)             ((v) & GENMASK(2, 0))
+#define GET_ADDRESS(v)         ((v) & GENMASK(31, 3))

bits.h

Quoted 1 line

That is already included from a few different sources.

Please, read again what I wrote below. Ditto for other inclusions.

Quoted 38 lines
Quoted 9 lines
Quoted 8 lines
drivers/platform/x86/intel_pmt_class.h new file +7 -0
+struct intel_pmt_entry {
 
+       struct bin_attribute    pmt_bin_attr;
+       struct kobject          *kobj;
+       void __iomem            *disc_tabl
+       void __iomem            *base;
+       unsigned long           base_addr;
+       size_t                  size;
Quoted 3 lines
Quoted 1 line

+ u32 guid;

types.h

Is included through xarray.h.

Quoted 11 lines

drivers/platform/x86/intel_pmt_class.h line 53

Quoted 9 lines
drivers/platform/x86/intel_pmt_class.h new file +7 -0
+       int                     devid;
+};
 
+static inline int
+intel_pmt_ioremap_discovery_table(struct intel_pmt_entry *entry,
+                                 struct platform_device *pdev,  int i)
+{
 
+       entry->disc_table = devm_platform_ioremap_resource(pdev, i);

io.h ?

That one I will move from the class.c file.

Quoted 4 lines

drivers/platform/x86/intel_pmt_class.h line 55

Quoted 2 lines
drivers/platform/x86/intel_pmt_class.h new file +2 -0
+
+       return PTR_ERR_OR_ZERO(entry->disc_table);

err.h

That is included as a part of io.h.

Quoted 2 lines
Quoted 1 line

+}

Quoted 8 lines
Quoted 3 lines

The rule of thumb is to include all headers that you have direct users of. Then you may optimize by removing those which are guaranteed to be included by others, like bits.h always included by bitops.h.

Yeah, from what I can tell the only one I didn't have was io.h and part of that is because this was something I had moved to the header file in order to commonize it since it was being used in the other drivers.

types.h not necessarily be included by above and so on...

Signature · 2 lines
With Best Regards,
Andy Shevchenko
#14 lore raw

Re: [PATCH V7 5/5] platform/x86: Intel PMT Crashlog capability driver

In reply to Andy Shevchenko
Quoted 25 lines · Andy Shevchenko
Quoted 9 lines · David E. Box

Add support for the Intel Platform Monitoring Technology crashlog interface. This interface provides a few sysfs values to allow for controlling the crashlog telemetry interface as well as a character driver to allow for mapping the crashlog memory region so that it can be accessed after a crashlog has been recorded.

This driver is meant to only support the server version of the crashlog which is identified as crash_type 1 with a version of zero. Currently no other types are supported.

...

Documentation/ABI/testing/sysfs-class-intel_pmt Contact: David Box <[email protected]> · line 63

Quoted 3 lines
Documentation/ABI/testing/sysfs-class-intel_pmt +3 -0
+               The crashlog<x> directory contains files for configuring an
+               instance of a PMT crashlog device that can perform crash data
+               recoring. Each crashlog<x> device has an associated crashlog

recording

Documentation/ABI/testing/sysfs-class-intel_pmt Contact: David Box <[email protected]> · line 67

Quoted 4 lines
Documentation/ABI/testing/sysfs-class-intel_pmt +4 -0
+               file. This file can be opened and mapped or read to access the
+               resulting crashlog buffer. The register layout for the buffer
+               can be determined from an XML file of specified guid for the
+               parent device.
Quoted 7 lines

...

Quoted 1 line

+ (RO) The guid for this crashlog device. The guid identifies the

guid -> GUID

Please, spell check all ABI files in this series.

I'll run through it again. I am suspecting I must have deleted the "d" in recording with a fat fingering when I was editing something else.

Quoted 11 lines

...

drivers/platform/x86/Kconfig config INTEL_PMT_TELEMETRY · line 1386

Quoted 7 lines
drivers/platform/x86/Kconfig +7 -0
+config INTEL_PMT_CRASHLOG
+       tristate "Intel Platform Monitoring Technology (PMT) Crashlog driver"
+       select INTEL_PMT_CLASS
+       help
+         The Intel Platform Monitoring Technology (PMT) crashlog driver provides
+         access to hardware crashlog capabilities on devices that support the
+         feature.

Name of the module?

I will add the verbiage:
To compile this driver as a module, choose M here: the module will be called intel_pmt_crashlog.

Quoted 7 lines

...

drivers/platform/x86/intel_pmt_crashlog.c line 81

Quoted 3 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +2 -0
+       /*
 
+        * Currenty we only recognize OOBMSM version 0 devices.

Currently. Please spell check all comments in the code.

I'll make another pass.

Quoted 8 lines
Quoted 2 lines
(quoted diff) +2 -0
+        * We can ignore all other crashlog devices in the system.
+        */

...

Quoted 1 line

+ /* clear control bits */

What new information readers get from this comment?

Arguably not much. I'll drop the comment.

Quoted 3 lines
Quoted 1 line

+ control &= ~(CRASHLOG_FLAG_MASK | CRASHLOG_FLAG_DISABLE);

How does the second constant play any role here?

The "control" flags are bits 28-31, while the disable flag is bit 27 if I recall.

Specifically bit 31 is read only, bit 28 will clear bit 31, bit 29 will cause the crashlog to be generated and set bit 31, and bit 30 is just reserved 0.

Quoted 5 lines

...

Quoted 1 line

+ /* clear control bits */

Ditto. And moreover it's ambiguous due to joined two lines below.

I'll just drop the comments.

Quoted 8 lines
Quoted 2 lines
(quoted diff) +2 -0
+       control &= ~CRASHLOG_FLAG_MASK;
+       control |= CRASHLOG_FLAG_EXECUTE;

...

Quoted 1 line

+ return strnlen(buf, count);

How is this different to count?

I guess they should be equivalent so I can probably just switch to count.

Quoted 10 lines

...

drivers/platform/x86/intel_pmt_crashlog.c line 179

Quoted 6 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +5 -0
+       struct crashlog_entry *entry;
+       bool trigger;
+       int result;
+
 
+       entry = dev_get_drvdata(dev);

You may reduce LOCs by direct assigning in the definition block above.

Okay. I can move it if you prefer.

Quoted 5 lines

...

Quoted 1 line

+ result = strnlen(buf, count);

How is it different from count?

I'll switch it.

Quoted 9 lines

...

drivers/platform/x86/intel_pmt_crashlog.c line 222

Quoted 5 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +5 -0
+static DEFINE_XARRAY_ALLOC(crashlog_array);
+static struct intel_pmt_namespace pmt_crashlog_ns = {
+       .name = "crashlog",
+       .xa = &crashlog_array,
+       .attr_grp = &pmt_crashlog_group

Leave the comma here.

Already fixed based on similar comments you had for the telemetry driver.. :-)

Quoted 4 lines
Quoted 1 line

+};

...

Quoted 14 lines

drivers/platform/x86/intel_pmt_crashlog.c line 254

Quoted 8 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +8 -0
+       ret = intel_pmt_dev_create(entry, &pmt_crashlog_ns, parent);
+       if (!ret)
+               return 0;
+
+       dev_err(parent, "Failed to add crashlog controls\n");
+       intel_pmt_dev_destroy(entry, &pmt_crashlog_ns);
+
+       return ret;

Can we use traditional patterns?
if (ret) {
...
}
return ret;

I can switch it if that is preferred.

Quoted 25 lines

...

drivers/platform/x86/intel_pmt_crashlog.c line 277

Quoted 4 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +4 -0
+       size = offsetof(struct pmt_crashlog_priv, entry[pdev->num_resources]);
+       priv = devm_kzalloc(&pdev->dev, size, GFP_KERNEL);
+       if (!priv)
+               return -ENOMEM;

struct_size()

...

Quoted 14 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +14 -0
+               /* initialize control mutex */
+               mutex_init(&priv->entry[i].control_mutex);
+
+               disc_res = platform_get_resource(pdev, IORESOURCE_MEM, i);
+               if (!disc_res)
+                       goto abort_probe;
+
+               ret = intel_pmt_ioremap_discovery_table(entry, pdev, i);
+               if (ret)
+                       goto abort_probe;
+
+               if (!pmt_crashlog_supported(entry))
+                       continue;
+
Quoted 8 lines

drivers/platform/x86/intel_pmt_crashlog.c line 305

Quoted 5 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +5 -0
+               ret = pmt_crashlog_add_entry(entry, &pdev->dev, disc_res);
+               if (ret)
+                       goto abort_probe;
+
+               priv->num_entries++;

Are you going to duplicate this in each driver? Consider to refactor to avoid duplication of a lot of code.

So the issue lies in the complexity of pmt_telem_add_entry versus pmt_crashlog_add_entry. Specifically I end up needing disc_res and the discovery table when I go to create the controls for the crashlog device. Similarly we have a third device that we plan to add called a watcher which will require us to keep things split up like this so we thought it best to split it up this way.

Quoted 9 lines

...

drivers/platform/x86/intel_pmt_crashlog.c line 338

Quoted 3 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +2 -0
+               .name   = DRV_NAME,
 
+MODULE_ALIAS("platform:" DRV_NAME);

I'm not sure I have interpreted this:
- Use 'raw' string instead of defines for device names
correctly. Can you elaborate?

Again I am not sure what this is in reference to. If you can point me to some documentation somewhere I can take a look.

Thanks.

- Alex

#15 lore raw

Re: [PATCH V7 5/5] platform/x86: Intel PMT Crashlog capability driver

Quoted 3 lines · Alexander Duyck
Quoted 1 line · Andy Shevchenko

On Thu, Oct 1, 2020 at 4:43 AM David E. Box <[email protected]> wrote:

...

Quoted 8 lines

Arguably not much. I'll drop the comment.

Quoted 3 lines
Quoted 1 line

+ control &= ~(CRASHLOG_FLAG_MASK | CRASHLOG_FLAG_DISABLE);

How does the second constant play any role here?

The "control" flags are bits 28-31, while the disable flag is bit 27 if I recall.

Okay, then it adds more confusion to the same comment here and there. Good you are about to drop the comment.

Quoted 3 lines

Specifically bit 31 is read only, bit 28 will clear bit 31, bit 29 will cause the crashlog to be generated and set bit 31, and bit 30 is just reserved 0.

Can this be added as a comment somewhere in the code?

...

Quoted 3 lines
Quoted 3 lines

drivers/platform/x86/intel_pmt_crashlog.c line 249

Quoted 3 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +3 -0
+       ret = intel_pmt_dev_create(entry, &pmt_crashlog_ns, parent);
+       if (!ret)
+               return 0;

(2)

Quoted 12 lines
Quoted 11 lines

drivers/platform/x86/intel_pmt_crashlog.c line 254

Quoted 5 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +5 -0
+
+       dev_err(parent, "Failed to add crashlog controls\n");
+       intel_pmt_dev_destroy(entry, &pmt_crashlog_ns);
+
+       return ret;

Can we use traditional patterns?
if (ret) {
...
}
return ret;

Quoted 1 line

I can switch it if that is preferred.

Yes, please. The (2) is really hard to parse (easy to miss ! part and be confused by return 0 one).

...

Quoted 9 lines
Quoted 2 lines

Are you going to duplicate this in each driver? Consider to refactor to avoid duplication of a lot of code.

So the issue lies in the complexity of pmt_telem_add_entry versus pmt_crashlog_add_entry. Specifically I end up needing disc_res and the discovery table when I go to create the controls for the crashlog device. Similarly we have a third device that we plan to add called a watcher which will require us to keep things split up like this so we thought it best to split it up this way.

Could you revisit and think how this can be deduplicated. I see at least one variant with a hooks (callbacks) which you supply depending on the driver, but the for-loop is kept in one place.

...

Quoted 10 lines
Quoted 7 lines

drivers/platform/x86/intel_pmt_crashlog.c line 338

Quoted 3 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +2 -0
+               .name   = DRV_NAME,
 
+MODULE_ALIAS("platform:" DRV_NAME);

I'm not sure I have interpreted this:
- Use 'raw' string instead of defines for device names
correctly. Can you elaborate?

Again I am not sure what this is in reference to. If you can point me to some documentation somewhere I can take a look.

Reference to your own changelog of this series!

Signature · 2 lines
With Best Regards,
Andy Shevchenko
#7 lore raw

Re: [PATCH V7 3/5] platform/x86: Intel PMT class driver

In reply to Andy Shevchenko
Quoted 24 lines · Andy Shevchenko
Quoted 3 lines · Alexander Duyck
Quoted 1 line · Andy Shevchenko

On Thu, Oct 1, 2020 at 4:43 AM David E. Box <[email protected]> wrote:

...

Quoted 15 lines
Quoted 14 lines
Quoted 11 lines

Intel Platform Monitoring Technology is meant to provide a common way to access telemetry and system metrics.

Register mappings are not provided by the driver. Instead, a GUID is read from a header for each endpoint. The GUID identifies the device and is to be used with an XML, provided by the vendor, to discover the available set of metrics and their register mapping. This allows firmware updates to modify the register space without needing to update the driver every time with new mappings. Firmware writes a new GUID in this case to specify the new mapping. Software tools with access to the associated XML file can then interpret the changes.

Where one may find a database of these reserved GUIDs / XMLs? How do you prevent a chaos which happens with other registries?

Quoted 6 lines
Quoted 2 lines

The database will be posted on intel.com eventually. Although I don't believe the URL is public yet.

How can we be sure that this won't be forgotten? How can we be sure it will be public at the end? Please, elaborate this in the commit message.

Okay, I will work with David on that.

Quoted 16 lines

...

Quoted 14 lines
Quoted 13 lines

drivers/platform/x86/intel_pmt_class.c line 34

Quoted 11 lines
drivers/platform/x86/intel_pmt_class.c new file +11 -0
+static const struct pci_device_id pmt_telem_early_client_pci_ids[] = {
+       { PCI_VDEVICE(INTEL, 0x9a0d) }, /* TGL */
+       { }
+};
+bool intel_pmt_is_early_client_hw(struct device *dev)
+{
+       struct pci_dev *parent = to_pci_dev(dev->parent);
+
+       return !!pci_match_id(pmt_telem_early_client_pci_ids, parent);
+}
+EXPORT_SYMBOL_GPL(intel_pmt_is_early_client_hw);

What is this and why is it in the class driver?

Quoted 6 lines
Quoted 4 lines

I chose to use the class driver as a central place to store code common to all of the instances of the class. In this case we have quirks that are specific to Tiger Lake and so I chose to store the function to test for the device here.

Can it be done in another file module at least (let's say intel_pmt_quirks.c)?

I suppose, but then it is adding a file for essentially 13 lines of code. Maybe I will just move this back to intel_pmt_telemetry.c and we can revisit where this should go if/when we add the watcher driver.

Quoted 4 lines

...

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
(quoted diff) +2 -0
+       if (off < 0)
+               return -EINVAL;
Quoted 8 lines
Quoted 5 lines
Quoted 1 line

Is this real or theoretical?

Not sure. I am not that familiar with the interface. It was something I copied from read_bmof which is what I based this code on based on an earlier suggestion.

I'm not a fan of cargo cult when there is no understanding why certain code appears in the driver.

Well with something like this I usually question if it provides any value. The problem is I don't know enough about binary sysfs attributes to say one way or another. If you know that the offset provided cannot be negative I can drop it. However I haven't seen anything that seems to say one way or another.

Quoted 18 lines

...

Quoted 5 lines
Quoted 3 lines
Quoted 1 line

+ if (count)

Useless.

I'm assuming that is because memcpy_fromio is assumed to handle this case?

Right.

Quoted 1 line
Quoted 1 line
Quoted 1 line

+ memcpy_fromio(buf, entry->base + off, count);

...

Quoted 4 lines
Quoted 3 lines
Quoted 1 line

+ psize = (PFN_UP(entry->base_addr + entry->size) - pfn) * PAGE_SIZE;

PFN_PHYS(PFN_UP(...)) ?

Quoted 8 lines
Quoted 5 lines

I'm not sure how that would work. Basically what we are doing here is determining the size of the mapping based on the number of pages that will be needed. So we wake the pfn of the start of the region, subtract that from the pfn for the end of the region and multiply by the size of a page.

PFN_PHYS() is a replacement for multiplication. You may check its implementation.

Ah, okay so you meant PFN_PHYS(PFN_UP(...)-pfn)

Quoted 16 lines

...

Quoted 7 lines
Quoted 6 lines

drivers/platform/x86/intel_pmt_class.c line 219

Quoted 3 lines
drivers/platform/x86/intel_pmt_class.c new file +3 -0
+       /* if size is 0 assume no data buffer, so no file needed */
+       if (!entry->size)
+               return 0;

Hmm... But presence of the file is also an information that might be useful for user, no?

Quoted 6 lines

I'm not sure what you mean? If the size of the region is zero it means there is no data there. There are cases in future devices where we may have controls for a telemetry function that is actually streaming the data elsewhere. That will be the use case for the entry size of 0 and in that case it doesn't make any sense to have a file as there is no data to present in it.

Quoted 2 lines

This is not understandable from 'no data buffer' above. Can you elaborate in the comment?

Having a file here implies there is something to read. There are devices supported by PMT which are essentially just control only, specifically some of the watchers. So the file would have nothing to read/mmap. In such a case it wouldn't make sense to provide a binary sysfs attribute as there is no data to associate with it.

Quoted 10 lines

...

Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
drivers/platform/x86/intel_pmt_class.c new file +6 -0
+       entry->base = devm_ioremap_resource(dev, &res);
+       if (IS_ERR(entry->base)) {
+               dev_err(dev, "Failed to ioremap device region\n");
+               goto fail_ioremap;
+       }
 
+       iounmap(entry->base);
Quoted 7 lines
Quoted 5 lines
Quoted 1 line

This is interesting. How do you avoid double unmap with (1)?

I think I get what you are trying to say. This is redundant since we used the devm_ioremap_resource it will already be freed when the driver is detached, correct?

Right. Above is the leftover that shows the poor testing of v7.

I will admit I don't think we tested the case where sysfs_create_bin_file failed.

Quoted 59 lines
Quoted 13 lines
Quoted 4 lines
Quoted 4 lines
drivers/platform/x86/intel_pmt_class.h new file +4 -0
+#include <linux/platform_device.h>
+#include <linux/xarray.h>
+
+/* PMT access types */
Quoted 8 lines

drivers/platform/x86/intel_pmt_class.h line 14

Quoted 6 lines
drivers/platform/x86/intel_pmt_class.h new file +6 -0
+#define ACCESS_BARID           2
+#define ACCESS_LOCAL           3
+
+/* PMT discovery base address/offset register layout */
+#define GET_BIR(v)             ((v) & GENMASK(2, 0))
+#define GET_ADDRESS(v)         ((v) & GENMASK(31, 3))

bits.h

Quoted 1 line

That is already included from a few different sources.

Please, read again what I wrote below. Ditto for other inclusions.

Quoted 42 lines
Quoted 9 lines
Quoted 8 lines
drivers/platform/x86/intel_pmt_class.h new file +7 -0
+struct intel_pmt_entry {
 
+       struct bin_attribute    pmt_bin_attr;
+       struct kobject          *kobj;
+       void __iomem            *disc_tabl
+       void __iomem            *base;
+       unsigned long           base_addr;
+       size_t                  size;
Quoted 3 lines
Quoted 1 line

+ u32 guid;

types.h

Is included through xarray.h.

Quoted 11 lines

drivers/platform/x86/intel_pmt_class.h line 53

Quoted 9 lines
drivers/platform/x86/intel_pmt_class.h new file +7 -0
+       int                     devid;
+};
 
+static inline int
+intel_pmt_ioremap_discovery_table(struct intel_pmt_entry *entry,
+                                 struct platform_device *pdev,  int i)
+{
 
+       entry->disc_table = devm_platform_ioremap_resource(pdev, i);

io.h ?

That one I will move from the class.c file.

Quoted 4 lines

drivers/platform/x86/intel_pmt_class.h line 55

Quoted 2 lines
drivers/platform/x86/intel_pmt_class.h new file +2 -0
+
+       return PTR_ERR_OR_ZERO(entry->disc_table);

err.h

That is included as a part of io.h.

Quoted 5 lines
Quoted 1 line

+}

The rule of thumb is to include all headers that you have direct users of. Then you may optimize by removing those which are guaranteed to be included by others, like bits.h always included by bitops.h.

Quoted 6 lines
Quoted 4 lines

Yeah, from what I can tell the only one I didn't have was io.h and part of that is because this was something I had moved to the header file in order to commonize it since it was being used in the other drivers.

types.h not necessarily be included by above and so on...

Okay, I will just explicitly add those headers then.

#16 lore raw

Re: [PATCH V7 5/5] platform/x86: Intel PMT Crashlog capability driver

In reply to Andy Shevchenko
Quoted 18 lines · Andy Shevchenko
Quoted 3 lines · Alexander Duyck
Quoted 1 line · Andy Shevchenko

On Thu, Oct 1, 2020 at 4:43 AM David E. Box <[email protected]> wrote:

...

Quoted 8 lines

Arguably not much. I'll drop the comment.

Quoted 3 lines
Quoted 1 line

+ control &= ~(CRASHLOG_FLAG_MASK | CRASHLOG_FLAG_DISABLE);

How does the second constant play any role here?

The "control" flags are bits 28-31, while the disable flag is bit 27 if I recall.

Quoted 8 lines

Okay, then it adds more confusion to the same comment here and there. Good you are about to drop the comment.

Quoted 3 lines

Specifically bit 31 is read only, bit 28 will clear bit 31, bit 29 will cause the crashlog to be generated and set bit 31, and bit 30 is just reserved 0.

Can this be added as a comment somewhere in the code?

I'll do that with the definitions themselves.

Quoted 37 lines

...

Quoted 3 lines
Quoted 3 lines

drivers/platform/x86/intel_pmt_crashlog.c line 249

Quoted 3 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +3 -0
+       ret = intel_pmt_dev_create(entry, &pmt_crashlog_ns, parent);
+       if (!ret)
+               return 0;

(2)

Quoted 12 lines
Quoted 11 lines

drivers/platform/x86/intel_pmt_crashlog.c line 254

Quoted 5 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +5 -0
+
+       dev_err(parent, "Failed to add crashlog controls\n");
+       intel_pmt_dev_destroy(entry, &pmt_crashlog_ns);
+
+       return ret;

Can we use traditional patterns?
if (ret) {
...
}
return ret;

Quoted 1 line

I can switch it if that is preferred.

Yes, please. The (2) is really hard to parse (easy to miss ! part and be confused by return 0 one).

...

Quoted 9 lines
Quoted 2 lines

Are you going to duplicate this in each driver? Consider to refactor to avoid duplication of a lot of code.

So the issue lies in the complexity of pmt_telem_add_entry versus pmt_crashlog_add_entry. Specifically I end up needing disc_res and the discovery table when I go to create the controls for the crashlog device. Similarly we have a third device that we plan to add called a watcher which will require us to keep things split up like this so we thought it best to split it up this way.

Quoted 3 lines

Could you revisit and think how this can be deduplicated. I see at least one variant with a hooks (callbacks) which you supply depending on the driver, but the for-loop is kept in one place.

I'll see what I can do.

Quoted 6 lines

...

Quoted 4 lines
Quoted 4 lines
Quoted 3 lines
drivers/platform/x86/intel_pmt_crashlog.c new file +2 -0
+               .name   = DRV_NAME,
 
+MODULE_ALIAS("platform:" DRV_NAME);
Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

I'm not sure I have interpreted this:
- Use 'raw' string instead of defines for device names
correctly. Can you elaborate?

Again I am not sure what this is in reference to. If you can point me to some documentation somewhere I can take a look.

Reference to your own changelog of this series!

So the issue is we have two authors so it is a matter of keeping track of who is working on what.

So apparently that was in reference to the MFD driver which was instantiating the devices using defines and there was only one spot where they were being used. The reason why I was confused is because the commit message had nothing to do with this patch and it I haven't really done any work on the MFD driver myself. The link to the 'raw' discussion can be found here:
https://lore.kernel.org/lkml/20200728075859.GH1850026@dell/