| Message ID | 20241104175407.19546-1-mario.limonciello@amd.com |
|---|---|
| Headers | show |
| Series | Add support for AMD hardware feedback interface | expand |
On 11/4/24 9:53 AM, Mario Limonciello wrote: > diff --git a/drivers/platform/x86/amd/hfi/Kconfig b/drivers/platform/x86/amd/hfi/Kconfig > new file mode 100644 > index 0000000000000..79e3b0073275c > --- /dev/null > +++ b/drivers/platform/x86/amd/hfi/Kconfig > @@ -0,0 +1,18 @@ > +# SPDX-License-Identifier: GPL-2.0-only > +# > +# AMD Hardware Feedback Interface Driver > +# > + > +config AMD_HFI > + bool "AMD Hetero Core Hardware Feedback Driver" > + depends on ACPI > + depends on CPU_SUP_AMD > + help > + Select this option to enable the AMD Heterogeneous Core Hardware Feedback Interface. If > + selected, hardware provides runtime thread classification guidance to the operating system > + on the performance and energy efficiency capabilities of each heterogeneous CPU core. > + These capabilities may vary due to the inherent differences in the core types and can > + also change as a result of variations in the operating conditions of the system such > + as power and thermal limits. If selected, the kernel relays updates in heterogeneous > + CPUs' capabilities to userspace, allowing for more optimal task scheduling and > + resource allocation, leveraging the diverse set of cores available. Please adjust the help text to fit in 80 columns. Thanks.
On 11/4/2024 23:23, Mario Limonciello wrote: > From: Perry Yuan <Perry.Yuan@amd.com> > > The AMD Heterogeneous core design and Hardware Feedback Interface (HFI) > provide behavioral classification and a dynamically updated ranking table > for the scheduler to use when choosing cores for tasks. > > There are two CPU core types defined: `Classic Core` and `Dense Core`. > "Classic" cores are the standard performance cores, while "Dense" cores > are optimized for area and efficiency. > > Heterogeneous compute refers to CPU implementations that are comprised > of more than one architectural class, each with two capabilities. This > means each CPU reports two separate capabilities: "perf" and "eff". > > Each capability lists all core ranking numbers between 0 and 255, where > a higher number represents a higher capability. > > Heterogeneous systems can also extend to more than two architectural > classes. > > The purpose of the scheduling feedback mechanism is to provide information > to the operating system scheduler in real time, allowing the scheduler to > direct threads to the optimal core during task scheduling. > > All core ranking data are provided by the PMFW via a shared memory ranking > table, which the driver reads and uses to update core capabilities to the > scheduler. When the hardware updates the table, it generates a platform > interrupt to notify the OS to read the new ranking table. > Reviewed-by: Shyam Sundar S K <Shyam-sundar.S-k@amd.com> > Link: https://bugzilla.kernel.org/show_bug.cgi?id=206537 > Reviewed-by: Gautham R. Shenoy <gautham.shenoy@amd.com> > Signed-off-by: Perry Yuan <perry.yuan@amd.com> > Co-developed-by: Mario Limonciello <mario.limonciello@amd.com> > Signed-off-by: Mario Limonciello <mario.limonciello@amd.com> > --- > v5->v6: > * Whitespace (Shyam) > * Use managed mutex (Shyam) > * Remove needless include (Shyam) > * Commit message text (Shyam) > * Use cpu_feature_enabled (Boris) > * Remove EOF newline (Boris) > --- > drivers/platform/x86/amd/Kconfig | 1 + > drivers/platform/x86/amd/Makefile | 1 + > drivers/platform/x86/amd/hfi/Kconfig | 18 +++ > drivers/platform/x86/amd/hfi/Makefile | 7 ++ > drivers/platform/x86/amd/hfi/hfi.c | 161 ++++++++++++++++++++++++++ > 5 files changed, 188 insertions(+) > create mode 100644 drivers/platform/x86/amd/hfi/Kconfig > create mode 100644 drivers/platform/x86/amd/hfi/Makefile > create mode 100644 drivers/platform/x86/amd/hfi/hfi.c > > diff --git a/drivers/platform/x86/amd/Kconfig b/drivers/platform/x86/amd/Kconfig > index f88682d36447c..c3f69dbe3037d 100644 > --- a/drivers/platform/x86/amd/Kconfig > +++ b/drivers/platform/x86/amd/Kconfig > @@ -5,6 +5,7 @@ > > source "drivers/platform/x86/amd/pmf/Kconfig" > source "drivers/platform/x86/amd/pmc/Kconfig" > +source "drivers/platform/x86/amd/hfi/Kconfig" > > config AMD_HSMP > tristate "AMD HSMP Driver" > diff --git a/drivers/platform/x86/amd/Makefile b/drivers/platform/x86/amd/Makefile > index dcec0a46f8af1..2676fc81fee54 100644 > --- a/drivers/platform/x86/amd/Makefile > +++ b/drivers/platform/x86/amd/Makefile > @@ -9,3 +9,4 @@ amd_hsmp-y := hsmp.o > obj-$(CONFIG_AMD_HSMP) += amd_hsmp.o > obj-$(CONFIG_AMD_PMF) += pmf/ > obj-$(CONFIG_AMD_WBRF) += wbrf.o > +obj-$(CONFIG_AMD_HFI) += hfi/ > diff --git a/drivers/platform/x86/amd/hfi/Kconfig b/drivers/platform/x86/amd/hfi/Kconfig > new file mode 100644 > index 0000000000000..79e3b0073275c > --- /dev/null > +++ b/drivers/platform/x86/amd/hfi/Kconfig > @@ -0,0 +1,18 @@ > +# SPDX-License-Identifier: GPL-2.0-only > +# > +# AMD Hardware Feedback Interface Driver > +# > + > +config AMD_HFI > + bool "AMD Hetero Core Hardware Feedback Driver" > + depends on ACPI > + depends on CPU_SUP_AMD > + help > + Select this option to enable the AMD Heterogeneous Core Hardware Feedback Interface. If > + selected, hardware provides runtime thread classification guidance to the operating system > + on the performance and energy efficiency capabilities of each heterogeneous CPU core. > + These capabilities may vary due to the inherent differences in the core types and can > + also change as a result of variations in the operating conditions of the system such > + as power and thermal limits. If selected, the kernel relays updates in heterogeneous > + CPUs' capabilities to userspace, allowing for more optimal task scheduling and > + resource allocation, leveraging the diverse set of cores available. > diff --git a/drivers/platform/x86/amd/hfi/Makefile b/drivers/platform/x86/amd/hfi/Makefile > new file mode 100644 > index 0000000000000..672c6ac106e95 > --- /dev/null > +++ b/drivers/platform/x86/amd/hfi/Makefile > @@ -0,0 +1,7 @@ > +# SPDX-License-Identifier: GPL-2.0 > +# > +# AMD Hardware Feedback Interface Driver > +# > + > +obj-$(CONFIG_AMD_HFI) += amd_hfi.o > +amd_hfi-objs := hfi.o > diff --git a/drivers/platform/x86/amd/hfi/hfi.c b/drivers/platform/x86/amd/hfi/hfi.c > new file mode 100644 > index 0000000000000..2cd71d79a22c9 > --- /dev/null > +++ b/drivers/platform/x86/amd/hfi/hfi.c > @@ -0,0 +1,161 @@ > +// SPDX-License-Identifier: GPL-2.0-or-later > +/* > + * AMD Hardware Feedback Interface Driver > + * > + * Copyright (C) 2024 Advanced Micro Devices, Inc. All Rights Reserved. > + * > + * Authors: Perry Yuan <Perry.Yuan@amd.com> > + * Mario Limonciello <mario.limonciello@amd.com> > + */ > + > +#define pr_fmt(fmt) "amd-hfi: " fmt > + > +#include <linux/acpi.h> > +#include <linux/cpu.h> > +#include <linux/cpumask.h> > +#include <linux/gfp.h> > +#include <linux/init.h> > +#include <linux/io.h> > +#include <linux/kernel.h> > +#include <linux/module.h> > +#include <linux/mutex.h> > +#include <linux/platform_device.h> > +#include <linux/smp.h> > + > +#define AMD_HFI_DRIVER "amd_hfi" > +#define AMD_HETERO_CPUID_27 0x80000027 > +static struct platform_device *device; > + > +struct amd_hfi_data { > + const char *name; > + struct device *dev; > + struct mutex lock; > +}; > + > +struct amd_hfi_classes { > + u32 perf; > + u32 eff; > +}; > + > +/** > + * struct amd_hfi_cpuinfo - HFI workload class info per CPU > + * @cpu: cpu index > + * @cpus: mask of cpus associated with amd_hfi_cpuinfo > + * @class_index: workload class ID index > + * @nr_class: max number of workload class supported > + * @amd_hfi_classes: current cpu workload class ranking data > + * > + * Parameters of a logical processor linked with hardware feedback class > + */ > +struct amd_hfi_cpuinfo { > + int cpu; > + cpumask_var_t cpus; > + s16 class_index; > + u8 nr_class; > + struct amd_hfi_classes *amd_hfi_classes; > +}; > + > +static DEFINE_PER_CPU(struct amd_hfi_cpuinfo, amd_hfi_cpuinfo) = {.class_index = -1}; > + > +static int amd_hfi_alloc_class_data(struct platform_device *pdev) > +{ > + struct amd_hfi_cpuinfo *hfi_cpuinfo; > + struct device *dev = &pdev->dev; > + int idx; > + int nr_class_id; > + > + nr_class_id = cpuid_eax(AMD_HETERO_CPUID_27); > + if (nr_class_id < 0 || nr_class_id > 255) { > + dev_err(dev, "failed to get number of supported classes: %d\n", > + nr_class_id); > + return -EINVAL; > + } > + > + for_each_present_cpu(idx) { > + struct amd_hfi_classes *classes; > + > + classes = devm_kzalloc(dev, > + nr_class_id * sizeof(struct amd_hfi_classes), > + GFP_KERNEL); > + if (!classes) > + return -ENOMEM; > + hfi_cpuinfo = per_cpu_ptr(&amd_hfi_cpuinfo, idx); > + hfi_cpuinfo->amd_hfi_classes = classes; > + hfi_cpuinfo->nr_class = nr_class_id; > + } > + > + return 0; > +} > + > +static const struct acpi_device_id amd_hfi_platform_match[] = { > + {"AMDI0104", 0}, > + { } > +}; > +MODULE_DEVICE_TABLE(acpi, amd_hfi_platform_match); > + > +static int amd_hfi_probe(struct platform_device *pdev) > +{ > + struct amd_hfi_data *amd_hfi_data; > + int ret; > + > + if (!acpi_match_device(amd_hfi_platform_match, &pdev->dev)) > + return -ENODEV; > + > + amd_hfi_data = devm_kzalloc(&pdev->dev, sizeof(*amd_hfi_data), GFP_KERNEL); > + if (!amd_hfi_data) > + return -ENOMEM; > + > + amd_hfi_data->dev = &pdev->dev; > + ret = devm_mutex_init(&pdev->dev, &amd_hfi_data->lock); > + if (ret) > + return ret; > + platform_set_drvdata(pdev, amd_hfi_data); > + > + ret = amd_hfi_alloc_class_data(pdev); > + if (ret) > + return ret; > + > + return 0; > +} > + > +static struct platform_driver amd_hfi_driver = { > + .driver = { > + .name = AMD_HFI_DRIVER, > + .owner = THIS_MODULE, > + .acpi_match_table = ACPI_PTR(amd_hfi_platform_match), > + }, > + .probe = amd_hfi_probe, > +}; > + > +static int __init amd_hfi_init(void) > +{ > + int ret; > + > + if (acpi_disabled || > + !cpu_feature_enabled(X86_FEATURE_AMD_HETEROGENEOUS_CORES) || > + !cpu_feature_enabled(X86_FEATURE_AMD_WORKLOAD_CLASS)) > + return -ENODEV; > + > + device = platform_device_register_simple(AMD_HFI_DRIVER, -1, NULL, 0); > + if (IS_ERR(device)) { > + pr_err("unable to register HFI platform device\n"); > + return PTR_ERR(device); > + } > + > + ret = platform_driver_register(&amd_hfi_driver); > + if (ret) > + pr_err("failed to register HFI driver\n"); > + > + return ret; > +} > + > +static __exit void amd_hfi_exit(void) > +{ > + platform_device_unregister(device); > + platform_driver_unregister(&amd_hfi_driver); > +} > +module_init(amd_hfi_init); > +module_exit(amd_hfi_exit); > + > +MODULE_LICENSE("GPL"); > +MODULE_DESCRIPTION("AMD Hardware Feedback Interface Driver");
The AMD Heterogeneous core design and Hardware Feedback Interface (HFI) provide behavioral classification and a dynamically updated ranking table for the scheduler to use when choosing cores for tasks. Threads are classified during runtime into enumerated classes. Currently, the driver supports 3 classes (0 through 2). These classes represent thread performance/power characteristics that may benefit from special scheduling behaviors. The real-time thread classification is consumed by the operating system and is used to inform the scheduler of where the thread should be placed for optimal performance or energy efficiency. The thread classification helps to select CPU from a ranking table that describes an efficiency and performance ranking for each classification from two dimensions. The ranking data provided by the ranking table are numbers ranging from 0 to 255, where a higher performance value indicates higher performance capability and a higher efficiency value indicates greater efficiency. All the CPU cores are ranked into different class IDs. Within each class ranking, the cores may have different ranking values. Therefore, picking from each classification ID will later allow the scheduler to select the best core while threads are classified into the specified workload class. This series was originally submitted by Perry Yuan [1] but he is now doing a different role and he asked me to take over. Link: https://lore.kernel.org/all/cover.1724748733.git.perry.yuan@amd.com/ On applicable hardware this series has between a 2% and 5% improvement across various benchmarks. There is however a cost associated with clearing history on the process context switch. On average it increases the delay by 119ns, and also has a wider range in delays (the standard deviation is 25% greater). Although this series most prominently has changes to platform-x86 it is based off of tip x86/cpu due to changes queued up for 6.13-rc1 that are dependencies. v5->v6: * boot_cpu_has() -> cpu_feature_enabled() * Simplify process switch patch * Use managed resources Mario Limonciello (4): MAINTAINERS: Add maintainer entry for AMD Hardware Feedback Driver cpufreq/amd-pstate: Disable preferred cores on designs with workload classification platform/x86/amd: hfi: Set ITMT priority from ranking data platform/x86/amd: hfi: Add debugfs support Perry Yuan (8): Documentation: x86: Add AMD Hardware Feedback Interface documentation x86/msr-index: define AMD heterogeneous CPU related MSR platform/x86: hfi: Introduce AMD Hardware Feedback Interface Driver platform/x86: hfi: parse CPU core ranking data from shared memory platform/x86: hfi: init per-cpu scores for each class platform/x86: hfi: add online and offline callback support platform/x86: hfi: add power management callback x86/process: Clear hardware feedback history for AMD processors Documentation/arch/x86/amd-hfi.rst | 127 ++++++ Documentation/arch/x86/index.rst | 1 + MAINTAINERS | 9 + arch/x86/include/asm/msr-index.h | 5 + arch/x86/kernel/process_32.c | 3 + arch/x86/kernel/process_64.c | 4 + drivers/cpufreq/amd-pstate.c | 6 + drivers/platform/x86/amd/Kconfig | 1 + drivers/platform/x86/amd/Makefile | 1 + drivers/platform/x86/amd/hfi/Kconfig | 19 + drivers/platform/x86/amd/hfi/Makefile | 7 + drivers/platform/x86/amd/hfi/hfi.c | 546 ++++++++++++++++++++++++++ 12 files changed, 729 insertions(+) create mode 100644 Documentation/arch/x86/amd-hfi.rst create mode 100644 drivers/platform/x86/amd/hfi/Kconfig create mode 100644 drivers/platform/x86/amd/hfi/Makefile create mode 100644 drivers/platform/x86/amd/hfi/hfi.c base-commit: 110213b8f0e7021819d4db273facb27701bc3381