Intel platform driver to reset bitfix filters
Subsystems (3)
cover Intel platform driver to reset bitfix filters
Some Intel CPUs implement a "bitfix filter" to suppress reporting of the same corrected errors repeatedly in the case where aging silicon develops stuck bits.
But when systems run for weeks, or months, the filters may become clogged with transient errors.
When the filter is full (indicated by a "yellow" signature in a machine check bank) clear the filter. This may make space for additional hard errors.
Save a time stamp when clearing a filter. Log with "WARN" severity if the filter overflows quickly (in tem minutes or less).
This code was originally developend by Qiuxu as an add-on to an EDAC driver. But not everyone wants to run an EDAC driver, so I pulled the code out into a standalone platform driver. Much of Qiuxu's code was cut-and-pasted into this driver, so he gets Co-developed-by credit throughout.
Qiuxu Zhuo (1):
cacheinfo: Export get_cpu_cacheinfo_id() for loadable modules
Tony Luck (6):
x86/mce: Enumeration updates for Intel bitfix filter reset platform/x86/intel/bff: Add stub Intel bitfix filter driver platform/x86/intel/bff: Add Diamond Rapids support platform/x86/intel/bff: Reset bitfix filter when it overflows platform/x86/intel/bff: Compute unique ID for overflowed filter platform/x86/intel/bff: Report frequent filter resets
include/linux/cacheinfo.h | 12 +- arch/x86/include/asm/mce.h | 6 + arch/x86/include/asm/msr-index.h | 2 + drivers/base/cacheinfo.c | 17 ++ drivers/platform/x86/intel/bff.c | 284 ++++++++++++++++++++++++++++ drivers/platform/x86/intel/Kconfig | 13 ++ drivers/platform/x86/intel/Makefile | 1 + 7 files changed, 324 insertions(+), 11 deletions(-) create mode 100644 drivers/platform/x86/intel/bff.c
base-commit: 8d3ae59288f1e7d58d76558a6ee96d533bc5019f
Signature · 1 line
2.55.0
PATCH 1/7 cacheinfo: Export get_cpu_cacheinfo_id() for loadable modules
From: Qiuxu Zhuo <[email protected]>
Loadable drivers that track bitfix filter state need the cache ID to distinguish shared cache instances. Without access to the cache ID, CPUs sharing the same cache cannot be associated with the same tracking state.
The existing helper cannot be used by loadable modules because it depends on a non-exported symbol, preventing the such drivers from building as a module.
Export get_cpu_cacheinfo_id() so loadable modules can retrieve cache IDs through a common cacheinfo interface.
--- include/linux/cacheinfo.h | 12 +----------- drivers/base/cacheinfo.c | 17 +++++++++++++++++ 2 files changed, 18 insertions(+), 11 deletions(-)
include/linux/cacheinfo.h +1 -11
@@ -82,6 +82,7 @@ struct cpu_cacheinfo { };struct cpu_cacheinfo *get_cpu_cacheinfo (unsigned int cpu); +int get_cpu_cacheinfo_id (int cpu,int level);int early_cache_level (unsigned int cpu);int init_cache_level (unsigned int cpu);int init_of_cache_level (unsigned int cpu);
@@ -137,17 +138,6 @@ static inline struct cacheinfo *get_cpu_cacheinfo_level(int cpu, int level)return NULL ; } -/* -* Get the id of the cache associated with @cpu at level @level. -* cpuhp lock must be held. -*/ -static inline int get_cpu_cacheinfo_id (int cpu,int level) -{ -struct cacheinfo *ci =get_cpu_cacheinfo_level (cpu, level); - -return ci ? ci->id :-1 ; -} -#if defined(CONFIG_ARM64 ) || defined(CONFIG_ARM )#define use_arch_cache_info () (true)#else
drivers/base/cacheinfo.c +17 -0
@@ -38,6 +38,23 @@ struct cpu_cacheinfo *get_cpu_cacheinfo(unsigned int cpu)return ci_cacheinfo (cpu); } +/** +* get_cpu_cacheinfo_id - Return the cache ID for a CPU and cache level +* @cpu: CPU number +* @level: Cache level +* +* The caller must hold the cpuhp lock. +* +* Return: Cache ID on success, or -1 if no matching cache exists. +*/ +int get_cpu_cacheinfo_id (int cpu,int level) +{ +struct cacheinfo *ci =get_cpu_cacheinfo_level (cpu, level); + +return ci ? ci->id :-1 ; +} +EXPORT_SYMBOL_GPL (get_cpu_cacheinfo_id); +static inline bool cache_leaves_are_shared (struct cacheinfo *this_leaf,struct cacheinfo *sib_leaf) {
Signature · 1 line
2.55.0
PATCH 2/7 x86/mce: Enumeration updates for Intel bitfix filter reset
In reply to Tony LuckIA32_CORE_CAPABILITIES enumerates the bitfix filter reset feature.
Also add definitions for the threshold status field in the machine check status registers.
--- arch/x86/include/asm/mce.h | 6 ++++++ arch/x86/include/asm/msr-index.h | 2 ++ 2 files changed, 8 insertions(+)
arch/x86/include/asm/mce.h +6 -0
@@ -13,6 +13,7 @@#define MCG_CTL_P BIT_ULL(8)/* MCG_CTL register available */ #define MCG_EXT_P BIT_ULL(9)/* Extended registers available */ #define MCG_CMCI_P BIT_ULL(10)/* CMCI supported */ +#define MCG_TES_P BIT_ULL(11)/* Threshold-based error status supported */ #define MCG_SEAM_NR BIT_ULL(12)/* MCG_STATUS_SEAM_NR supported */ #define MCG_EXT_CNT_MASK 0xff0000/* Number of Extended registers */ #define MCG_EXT_CNT_SHIFT 16
@@ -41,6 +42,11 @@#define MCI_STATUS_PCC BIT_ULL(57)/* processor context corrupt */ #define MCI_STATUS_S BIT_ULL(56)/* Signaled machine check */ #define MCI_STATUS_AR BIT_ULL(55)/* Action required */ +#define MCI_STATUS_TES_SHIFT 53/* Threshold-based error status */ +#define MCI_STATUS_TES_MASK GENMASK_ULL(54, 53) +#define MCI_STATUS_TES (s) (((s) & MCI_STATUS_TES_MASK) >> MCI_STATUS_TES_SHIFT) +#define MCI_STATUS_TES_GREEN 1/* Threshold-based errors below threshold */ +#define MCI_STATUS_TES_YELLOW 2/* Threshold-based errors above threshold */ #define MCI_STATUS_CEC_SHIFT 38/* Corrected Error Count */ #define MCI_STATUS_CEC_MASK GENMASK_ULL(52,38)#define MCI_STATUS_CEC (c) (((c) & MCI_STATUS_CEC_MASK) >> MCI_STATUS_CEC_SHIFT)
arch/x86/include/asm/msr-index.h +2 -0
@@ -114,6 +114,8 @@#define MSR_IA32_CORE_CAPS_INTEGRITY_CAPS BIT(MSR_IA32_CORE_CAPS_INTEGRITY_CAPS_BIT)#define MSR_IA32_CORE_CAPS_SPLIT_LOCK_DETECT_BIT 5#define MSR_IA32_CORE_CAPS_SPLIT_LOCK_DETECT BIT(MSR_IA32_CORE_CAPS_SPLIT_LOCK_DETECT_BIT) +#define MSR_IA32_CORE_CAPS_BFF_RESET_DETECT_BIT 9 +#define MSR_IA32_CORE_CAPS_BFF_RESET_DETECT BIT(MSR_IA32_CORE_CAPS_BFF_RESET_DETECT_BIT)#define MSR_PKG_CST_CONFIG_CONTROL 0x000000e2#define NHM_C3_AUTO_DEMOTE (1UL << 25)
Signature · 1 line
2.55.0
Ilpo Järvinen
maintainerRe: [PATCH 2/7] x86/mce: Enumeration updates for Intel bitfix filter reset
Quoted 26 lines · Tony Luck
IA32_CORE_CAPABILITIESenumerates the bitfix filter reset feature.Also add definitions for the threshold status field in the machine check status registers.
Co-developed-by Qiuxu Zhuo <[email protected]>Signed-off-by Qiuxu Zhuo <[email protected]>Signed-off-by Tony Luck <[email protected]>--- arch/x86/include/asm/mce.h | 6 ++++++ arch/x86/include/asm/msr-index.h | 2 ++ 2 files changed, 8 insertions(+)arch/x86/include/asm/mce.h +1
-0@@ -13,6 +13,7 @@#define MCG_CTL_P BIT_ULL(8)/* MCG_CTL register available */ #define MCG_EXT_P BIT_ULL(9)/* Extended registers available */ #define MCG_CMCI_P BIT_ULL(10)/* CMCI supported */ +#define MCG_TES_P BIT_ULL(11)/* Threshold-based error status supported */ #define MCG_SEAM_NR BIT_ULL(12)/* MCG_STATUS_SEAM_NR supported */ #define MCG_EXT_CNT_MASK 0xff0000/* Number of Extended registers */ #define MCG_EXT_CNT_SHIFT 16
arch/x86/include/asm/mce.h line 47
Quoted 6 lines
arch/x86/include/asm/mce.h +3
-0@@ -41,6 +42,11 @@#define MCI_STATUS_PCC BIT_ULL(57)/* processor context corrupt */ #define MCI_STATUS_S BIT_ULL(56)/* Signaled machine check */ #define MCI_STATUS_AR BIT_ULL(55)/* Action required */ +#define MCI_STATUS_TES_SHIFT 53/* Threshold-based error status */ +#define MCI_STATUS_TES_MASK GENMASK_ULL(54, 53) +#define MCI_STATUS_TES (s) (((s) & MCI_STATUS_TES_MASK) >> MCI_STATUS_TES_SHIFT)
This opencodes FIELD_GET(), and you don't need usually need _SHIFT define at all when you use FIELD_GET().
Quoted 19 lines
arch/x86/include/asm/mce.h +2
-0+#define MCI_STATUS_TES_GREEN 1/* Threshold-based errors below threshold */ +#define MCI_STATUS_TES_YELLOW 2/* Threshold-based errors above threshold */ #define MCI_STATUS_CEC_SHIFT 38/* Corrected Error Count */ #define MCI_STATUS_CEC_MASK GENMASK_ULL(52,38)#define MCI_STATUS_CEC (c) (((c) & MCI_STATUS_CEC_MASK) >> MCI_STATUS_CEC_SHIFT)arch/x86/include/asm/msr-index.h +2
-0@@ -114,6 +114,8 @@#define MSR_IA32_CORE_CAPS_INTEGRITY_CAPS BIT(MSR_IA32_CORE_CAPS_INTEGRITY_CAPS_BIT)#define MSR_IA32_CORE_CAPS_SPLIT_LOCK_DETECT_BIT 5#define MSR_IA32_CORE_CAPS_SPLIT_LOCK_DETECT BIT(MSR_IA32_CORE_CAPS_SPLIT_LOCK_DETECT_BIT) +#define MSR_IA32_CORE_CAPS_BFF_RESET_DETECT_BIT 9 +#define MSR_IA32_CORE_CAPS_BFF_RESET_DETECT BIT(MSR_IA32_CORE_CAPS_BFF_RESET_DETECT_BIT)#define MSR_PKG_CST_CONFIG_CONTROL 0x000000e2#define NHM_C3_AUTO_DEMOTE (1UL << 25)
Signature · 1 line
i.
PATCH 3/7 platform/x86/intel/bff: Add stub Intel bitfix filter driver
In reply to Tony LuckCheck if platform supports enhanced cache error reporting and the bitfix filter reset feature.
--- drivers/platform/x86/intel/bff.c | 59 +++++++++++++++++++++++++++++ drivers/platform/x86/intel/Kconfig | 13 +++++++ drivers/platform/x86/intel/Makefile | 1 + 3 files changed, 73 insertions(+) create mode 100644 drivers/platform/x86/intel/bff.c
drivers/platform/x86/intel/bff.c new file +59 -0
@@ -0,0 +1,59 @@ +// SPDX-License-Identifier: GPL-2.0-only +/* Copyright(c) 2026 Intel Corporation. */ + +/* +* Intel driver to reset bitfix filters when they overflow. +* +* Each bitfix filter has limited slots to track corrected errors. +* These slots can be filled by transient corrected errors (e.g., +* bit flips from particle strikes) that don't represent permanent +* hardware defects. +* +* When the filter overflows (yellow status), reset it to reclaim +* slots occupied by transient corrected errors. If the overflow +* repeats frequently after reset, it indicates persistent hardware +* defects that need attention: the system should be scheduled for +* servicing. +*/ +#define pr_fmt (fmt) "bff: " fmt + +#include <linux/cpufeature.h> +#include <linux/errno.h> +#include <linux/init.h> +#include <linux/module.h> +#include <linux/types.h> + +#include <asm/cpufeatures.h> +#include <asm/mce.h> +#include <asm/msr.h> +#include <asm/msr-index.h> + +static int __init bff_init (void ) +{ +u64 core_caps, mcg_cap; + +if (!cpu_feature_enabled (X86_FEATURE_MCA )) +return -ENODEV ; +rdmsrq (MSR_IA32_MCG_CAP , mcg_cap); +if (!(mcg_cap &MCG_TES_P )) +return -ENODEV ; + +if (!cpu_feature_enabled (X86_FEATURE_CORE_CAPABILITIES )) +return -ENODEV ; + +rdmsrq (MSR_IA32_CORE_CAPS , core_caps); +if (!(core_caps &MSR_IA32_CORE_CAPS_BFF_RESET_DETECT )) +return -ENODEV ; + +return 0 ; +} + +static void __exit bff_exit(void) +{ +} + +module_init (bff_init); +module_exit (bff_exit); + +MODULE_LICENSE ("GPL" ); +MODULE_DESCRIPTION ("Intel bitfix filter reset driver" );
drivers/platform/x86/intel/Kconfig +13 -0
@@ -251,3 +251,16 @@ config INTEL_VSEC To compile this driver as a module, choose M here: the module will be called intel_vsec. + +config INTEL_BFF_RESET + tristate "Intel bitfix filter reset driver" + depends on CPU_SUP_INTEL && X86_MCE + help + Adds support for Intel systems that implement bitfix filters to + suppress repeated logging and signaling of the same corrected + error. Those filters can become clogged with transient errors + over time. Clearing the filter may make space for additional + hard errors. + + To compile this driver as a module, choose M here: the module will + be called intel_bff.
drivers/platform/x86/intel/Makefile +1 -0
@@ -27,6 +27,7 @@ intel-target-$(CONFIG_INTEL_ISHTP_ECLITE) += ishtp_eclite.o intel-target-$(CONFIG_INTEL_OAKTRAIL) += oaktrail.o intel-target-$(CONFIG_INTEL_SDSI) += sdsi.o intel-target-$(CONFIG_INTEL_VSEC) += vsec.o +intel-target-$(CONFIG_INTEL_BFF_RESET) += bff.o # Intel PMIC / PMC / P-Unit drivers intel-target-$(CONFIG_INTEL_BYTCRC_PWRSRC) += bytcrc_pwrsrc.o
Signature · 1 line
2.55.0
PATCH 4/7 platform/x86/intel/bff: Add Diamond Rapids support
In reply to Tony LuckThere are potentially bitfix filters associated with each machine check bank. Only some banks may implement them.
Machine check banks have varying scope. E.g. there is a separate L2 cache for each module, each instance has its own bitfix filter.
Add information that will be used to map a <cpu,bank> pair to a unique instance number for a bitfix filter.
--- drivers/platform/x86/intel/bff.c | 41 ++++++++++++++++++++++++++++++++ 1 file changed, 41 insertions(+)
drivers/platform/x86/intel/bff.c +41 -0
@@ -18,18 +18,51 @@#define pr_fmt (fmt) "bff: " fmt#include <linux/cpufeature.h> +#include <linux/device-id/x86_cpu.h> #include <linux/errno.h> #include <linux/init.h> #include <linux/module.h> +#include <linux/printk.h> #include <linux/types.h> +#include <asm/cpu_device_id.h> #include <asm/cpufeatures.h> +#include <asm/intel-family.h> #include <asm/mce.h> #include <asm/msr.h> #include <asm/msr-index.h> +/* Machine check bank scope types */ +enum bff_type { +BFF_NONE =0 , +BFF_BANK_DCU , +BFF_BANK_DTLB , +BFF_BANK_MLC , +BFF_BANK_CCF , +BFF_BANK_HSF , +BFF_BANK_IOCACHE , +}; + +static const enum bff_type dmr_mcbanks[MAX_NR_BANKS ] = { + [1 ] =BFF_BANK_DCU , + [2 ] =BFF_BANK_DTLB , + [3 ] =BFF_BANK_MLC , + [6 ] =BFF_BANK_CCF , + [13 ] =BFF_BANK_HSF , + [17 ] =BFF_BANK_IOCACHE +}; + +static const struct x86_cpu_id bff_cpu_ids[]__initconst = { +X86_MATCH_VFM (INTEL_DIAMONDRAPIDS_X , dmr_mcbanks), + {} +}; +MODULE_DEVICE_TABLE (x86cpu, bff_cpu_ids); + +static const enum bff_type *bank_types; +static int __init bff_init (void ) { +const struct x86_cpu_id *m;u64 core_caps, mcg_cap;if (!cpu_feature_enabled (X86_FEATURE_MCA ))
@@ -45,6 +78,14 @@ static int __init bff_init(void)if (!(core_caps &MSR_IA32_CORE_CAPS_BFF_RESET_DETECT ))return -ENODEV ; + m =x86_match_cpu (bff_cpu_ids); +if (!m) { +pr_info ("CPU supports bitfix filter, but driver does not\n" ); +return -ENODEV ; + } + + bank_types = (const enum bff_type *)m->driver_data; +return 0 ; }
Signature · 1 line
2.55.0
Ilpo Järvinen
maintainerRe: [PATCH 4/7] platform/x86/intel/bff: Add Diamond Rapids support
Quoted 46 lines · Tony Luck
There are potentially bitfix filters associated with each machine check bank. Only some banks may implement them.
Machine check banks have varying scope. E.g. there is a separate L2 cache for each module, each instance has its own bitfix filter.
Add information that will be used to map a <cpu,bank> pair to a unique instance number for a bitfix filter.
Co-developed-by Qiuxu Zhuo <[email protected]>Signed-off-by Qiuxu Zhuo <[email protected]>Signed-off-by Tony Luck <[email protected]>--- drivers/platform/x86/intel/bff.c | 41 ++++++++++++++++++++++++++++++++ 1 file changed, 41 insertions(+)drivers/platform/x86/intel/bff.c +12
-0@@ -18,18 +18,51 @@#define pr_fmt (fmt) "bff: " fmt#include <linux/cpufeature.h> +#include <linux/device-id/x86_cpu.h> #include <linux/errno.h> #include <linux/init.h> #include <linux/module.h> +#include <linux/printk.h> #include <linux/types.h> +#include <asm/cpu_device_id.h> #include <asm/cpufeatures.h> +#include <asm/intel-family.h> #include <asm/mce.h> #include <asm/msr.h> #include <asm/msr-index.h> +/* Machine check bank scope types */ +enum bff_type { +BFF_NONE =0 , +BFF_BANK_DCU , +BFF_BANK_DTLB , +BFF_BANK_MLC , +BFF_BANK_CCF , +BFF_BANK_HSF ,
drivers/platform/x86/intel/bff.c line 52
Quoted 10 lines
drivers/platform/x86/intel/bff.c +10
-0+ BFF_BANK_IOCACHE, +}; + +static const enum bff_type dmr_mcbanks[MAX_NR_BANKS ] = { + [1 ] =BFF_BANK_DCU , + [2 ] =BFF_BANK_DTLB , + [3 ] =BFF_BANK_MLC , + [6 ] =BFF_BANK_CCF , + [13 ] =BFF_BANK_HSF , + [17 ] =BFF_BANK_IOCACHE
Please add trailing comma.
Quoted 32 lines
drivers/platform/x86/intel/bff.c +19
-0+}; + +static const struct x86_cpu_id bff_cpu_ids[]__initconst = { +X86_MATCH_VFM (INTEL_DIAMONDRAPIDS_X , dmr_mcbanks), + {} +}; +MODULE_DEVICE_TABLE (x86cpu, bff_cpu_ids); + +static const enum bff_type *bank_types; +static int __init bff_init (void ) { +const struct x86_cpu_id *m;u64 core_caps, mcg_cap;if (!cpu_feature_enabled (X86_FEATURE_MCA ))@@ -45,6 +78,14 @@ static int __init bff_init(void)if (!(core_caps &MSR_IA32_CORE_CAPS_BFF_RESET_DETECT ))return -ENODEV ; + m =x86_match_cpu (bff_cpu_ids); +if (!m) { +pr_info ("CPU supports bitfix filter, but driver does not\n" ); +return -ENODEV ; + } + + bank_types = (const enum bff_type *)m->driver_data; +return 0 ; }
Signature · 1 line
i.
RE: [PATCH 4/7] platform/x86/intel/bff: Add Diamond Rapids support
Hi Ilpo.
Thanks for reviewing.
Quoted 10 lines
drivers/platform/x86/intel/bff.c line 44
Quoted 10 lines
drivers/platform/x86/intel/bff.c +10
-0+/* Machine check bank scope types */ +enum bff_type { +BFF_NONE =0 , +BFF_BANK_DCU , +BFF_BANK_DTLB , +BFF_BANK_MLC , +BFF_BANK_CCF , +BFF_BANK_HSF , +BFF_BANK_IOCACHE , +};
Trailing comma here because this enum will expand when additional CPUs are supported that have different types of banks that support the bitfix filter.
Quoted 10 lines
drivers/platform/x86/intel/bff.c line 52
Quoted 8 lines
drivers/platform/x86/intel/bff.c +8
-0+ +static const enum bff_type dmr_mcbanks[MAX_NR_BANKS ] = { + [1 ] =BFF_BANK_DCU , + [2 ] =BFF_BANK_DTLB , + [3 ] =BFF_BANK_MLC , + [6 ] =BFF_BANK_CCF , + [13 ] =BFF_BANK_HSF , + [17 ] =BFF_BANK_IOCACHE Please add trailing comma.
No trailing comma here because this list is complete. Diamond rapids will not suddenly add a new bitfix filter attached to a bank numbered > 17.
-Tony
PATCH 5/7 platform/x86/intel/bff: Reset bitfix filter when it overflows
In reply to Tony LuckGet notifications for all errors logged in machine check banks. Skip any that do not indicate a bitfix filter overflow.
Reset the bitfix filter using the CPU that reported the error to get the right scoped machine check bank.
--- drivers/platform/x86/intel/bff.c | 41 ++++++++++++++++++++++++++++++++ 1 file changed, 41 insertions(+)
drivers/platform/x86/intel/bff.c +41 -0
@@ -17,11 +17,13 @@*/ #define pr_fmt (fmt) "bff: " fmt +#include <linux/bits.h> #include <linux/cpufeature.h> #include <linux/device-id/x86_cpu.h> #include <linux/errno.h> #include <linux/init.h> #include <linux/module.h> +#include <linux/notifier.h> #include <linux/printk.h> #include <linux/types.h>
@@ -32,6 +34,11 @@#include <asm/msr.h> #include <asm/msr-index.h> +/* Intel bitfix filter control register defines */ +#define MSR_MC0_BFF_CTL 0x000006c0 +#define MSR_MCx_BFF_CTL (x) (MSR_MC0_BFF_CTL + (x)) +#define MCI_BFF_RESET BIT_ULL(0) +/* Machine check bank scope types */ enum bff_type {BFF_NONE =0 ,
@@ -60,6 +67,37 @@ MODULE_DEVICE_TABLE(x86cpu, bff_cpu_ids);static const enum bff_type *bank_types; +static void handle_bff (struct mce *mce) +{ +/* The bitfix filter overflowed, get the target CPU to reset it. */ +if (wrmsrq_on_cpu (mce->extcpu,MSR_MCx_BFF_CTL (mce->bank),MCI_BFF_RESET )) +pr_warn ("Failed to reset bitfix filter for CPU %d Bank %d\n" , mce->extcpu, mce->bank); +} + +static int bff_mce_notify (struct notifier_block *nb,unsigned long val,void *data) +{ +struct mce *mce = (struct mce *)data; + +/* TES is undefined for uncorrected errors. */ +if (mce->status &MCI_STATUS_UC ) +return NOTIFY_DONE ; + +if (MCI_STATUS_TES (mce->status) !=MCI_STATUS_TES_YELLOW ) +return NOTIFY_DONE ; + +if (mce->bank >=MAX_NR_BANKS || bank_types[mce->bank] ==BFF_NONE ) +return NOTIFY_DONE ; + +handle_bff (mce); + +return NOTIFY_DONE ; +} + +static struct notifier_block bff_notifier = { + .notifier_call = bff_mce_notify, + .priority =MCE_PRIO_EARLY , +}; +static int __init bff_init (void ) {const struct x86_cpu_id *m;
@@ -86,11 +124,14 @@ static int __init bff_init(void) bank_types = (const enum bff_type *)m->driver_data; +mce_register_decode_chain (&bff_notifier); +return 0 ; } static void __exit bff_exit(void) { +mce_unregister_decode_chain (&bff_notifier); }module_init (bff_init);
Signature · 1 line
2.55.0
PATCH 6/7 platform/x86/intel/bff: Compute unique ID for overflowed filter
In reply to Tony LuckEach L2 cache instance has its own bitfix filter, but always reports errors in machine check bank 3 (on Diamond Rapids).
Compute a unique bitfix filter instance number based on the CPU that logged the error and the machine check bank number.
Special case for banks associated with the Integrated Memory Hub (IMH). Here the "even" numbered CPU modules are associated with IMH0 and the "odd" modules with IMH1.
The unique id will be used to store a time stamp of when the bitfix filter overflowed so that frequent overflows can be logged.
--- drivers/platform/x86/intel/bff.c | 74 ++++++++++++++++++++++++++++++++ 1 file changed, 74 insertions(+)
drivers/platform/x86/intel/bff.c +74 -0
@@ -18,13 +18,18 @@#define pr_fmt (fmt) "bff: " fmt#include <linux/bits.h> +#include <linux/cacheinfo.h> +#include <linux/cleanup.h> #include <linux/cpufeature.h> +#include <linux/cpuhplock.h> #include <linux/device-id/x86_cpu.h> #include <linux/errno.h> #include <linux/init.h> +#include <linux/limits.h> #include <linux/module.h> #include <linux/notifier.h> #include <linux/printk.h> +#include <linux/topology.h> #include <linux/types.h> #include <asm/cpu_device_id.h>
@@ -34,6 +39,8 @@#include <asm/msr.h> #include <asm/msr-index.h> +#define NUM_IMH_PER_SKT 2 +/* Intel bitfix filter control register defines */ #define MSR_MC0_BFF_CTL 0x000006c0#define MSR_MCx_BFF_CTL (x) (MSR_MC0_BFF_CTL + (x))
@@ -67,11 +74,78 @@ MODULE_DEVICE_TABLE(x86cpu, bff_cpu_ids);static const enum bff_type *bank_types; +/* Diamond Rapids maps APICID[2] to the IMH instance. */ +static void bff_set_imh_id (struct mce *mce,unsigned long *id) +{ +int imh_num = (NUM_IMH_PER_SKT *topology_physical_package_id (mce->extcpu)) + + ((mce->apicid >>2 ) &0x1 ); + + *id |= imh_num; +} + +static bool bff_set_cache_id (int cpu,int level,unsigned long *id) +{ +int cacheid; + +guard (cpus_read_lock)(); + + cacheid =get_cpu_cacheinfo_id (cpu, level); +if (cacheid ==-1 ) { +pr_warn ("Could not get L%d cache id for CPU %d\n" , level, cpu); +return false; + } + + *id |= cacheid; + +return true; +} + +#define BFF_ID_BANK_SHIFT 16 + +/* +* Cache IDs are only unique within a cache level. +* Include the MCA bank number so each BFF-capable hardware +* resource has a unique tracking ID. +*/ +static unsigned long get_bff_id (struct mce *mce) +{ +unsigned long id = (unsigned long )mce->bank <<BFF_ID_BANK_SHIFT ; + +switch (bank_types[mce->bank]) { +case BFF_BANK_DCU : +case BFF_BANK_DTLB : +if (!bff_set_cache_id (mce->extcpu,1 , &id)) +return ULONG_MAX ; +break ; + +case BFF_BANK_MLC : +if (!bff_set_cache_id (mce->extcpu,2 , &id)) +return ULONG_MAX ; +break ; + +case BFF_BANK_CCF : +if (!bff_set_cache_id (mce->extcpu,3 , &id)) +return ULONG_MAX ; +break ; + +case BFF_BANK_HSF : +case BFF_BANK_IOCACHE : +bff_set_imh_id (mce, &id); +break ; +default : +return ULONG_MAX ; + } + +return id; +} +static void handle_bff (struct mce *mce) {/* The bitfix filter overflowed, get the target CPU to reset it. */ if (wrmsrq_on_cpu (mce->extcpu,MSR_MCx_BFF_CTL (mce->bank),MCI_BFF_RESET ))pr_warn ("Failed to reset bitfix filter for CPU %d Bank %d\n" , mce->extcpu, mce->bank); + +pr_debug ("unique_id = 0x%lx\n" ,get_bff_id (mce)); }static int bff_mce_notify (struct notifier_block *nb,unsigned long val,void *data)
Signature · 1 line
2.55.0
Ilpo Järvinen
maintainerRe: [PATCH 6/7] platform/x86/intel/bff: Compute unique ID for overflowed filter
Quoted 53 lines · Tony Luck
Each L2 cache instance has its own bitfix filter, but always reports errors in machine check bank 3 (on Diamond Rapids).
Compute a unique bitfix filter instance number based on the CPU that logged the error and the machine check bank number.
Special case for banks associated with the Integrated Memory Hub (IMH). Here the "even" numbered CPU modules are associated with IMH0 and the "odd" modules with IMH1.
The unique id will be used to store a time stamp of when the bitfix filter overflowed so that frequent overflows can be logged.
Co-developed-by Qiuxu Zhuo <[email protected]>Signed-off-by Qiuxu Zhuo <[email protected]>Signed-off-by Tony Luck <[email protected]>--- drivers/platform/x86/intel/bff.c | 74 ++++++++++++++++++++++++++++++++ 1 file changed, 74 insertions(+)drivers/platform/x86/intel/bff.c +7
-0@@ -18,13 +18,18 @@#define pr_fmt (fmt) "bff: " fmt#include <linux/bits.h> +#include <linux/cacheinfo.h> +#include <linux/cleanup.h> #include <linux/cpufeature.h> +#include <linux/cpuhplock.h> #include <linux/device-id/x86_cpu.h> #include <linux/errno.h> #include <linux/init.h> +#include <linux/limits.h> #include <linux/module.h> #include <linux/notifier.h> #include <linux/printk.h> +#include <linux/topology.h> #include <linux/types.h> #include <asm/cpu_device_id.h> @@ -34,6 +39,8 @@#include <asm/msr.h> #include <asm/msr-index.h> +#define NUM_IMH_PER_SKT 2 +/* Intel bitfix filter control register defines */ #define MSR_MC0_BFF_CTL 0x000006c0#define MSR_MCx_BFF_CTL (x) (MSR_MC0_BFF_CTL + (x))
drivers/platform/x86/intel/bff.c MODULE_DEVICE_TABLE(x86cpu, bff_cpu_ids); · line 81
Quoted 8 lines
drivers/platform/x86/intel/bff.c +5
-0@@ -67,11 +74,78 @@ MODULE_DEVICE_TABLE(x86cpu, bff_cpu_ids);static const enum bff_type *bank_types; +/* Diamond Rapids maps APICID[2] to the IMH instance. */ +static void bff_set_imh_id (struct mce *mce,unsigned long *id) +{ +int imh_num = (NUM_IMH_PER_SKT *topology_physical_package_id (mce->extcpu)) + + ((mce->apicid >>2 ) &0x1 );
FIELD_GET(), I suggest moving comment where you define the field for that.
Quoted 21 lines
drivers/platform/x86/intel/bff.c +21
-0+ + *id |= imh_num; +} + +static bool bff_set_cache_id (int cpu,int level,unsigned long *id) +{ +int cacheid; + +guard (cpus_read_lock)(); + + cacheid =get_cpu_cacheinfo_id (cpu, level); +if (cacheid ==-1 ) { +pr_warn ("Could not get L%d cache id for CPU %d\n" , level, cpu); +return false; + } + + *id |= cacheid; + +return true; +} +
drivers/platform/x86/intel/bff.c MODULE_DEVICE_TABLE(x86cpu, bff_cpu_ids); · line 112
Quoted 10 lines
drivers/platform/x86/intel/bff.c +10
-0+#define BFF_ID_BANK_SHIFT 16 + +/* +* Cache IDs are only unique within a cache level. +* Include the MCA bank number so each BFF-capable hardware +* resource has a unique tracking ID. +*/ +static unsigned long get_bff_id (struct mce *mce) +{ +unsigned long id = (unsigned long )mce->bank <<BFF_ID_BANK_SHIFT ;
FIELD_PREP() ?
Quoted 40 lines
drivers/platform/x86/intel/bff.c +31
-0+ +switch (bank_types[mce->bank]) { +case BFF_BANK_DCU : +case BFF_BANK_DTLB : +if (!bff_set_cache_id (mce->extcpu,1 , &id)) +return ULONG_MAX ; +break ; + +case BFF_BANK_MLC : +if (!bff_set_cache_id (mce->extcpu,2 , &id)) +return ULONG_MAX ; +break ; + +case BFF_BANK_CCF : +if (!bff_set_cache_id (mce->extcpu,3 , &id)) +return ULONG_MAX ; +break ; + +case BFF_BANK_HSF : +case BFF_BANK_IOCACHE : +bff_set_imh_id (mce, &id); +break ; +default : +return ULONG_MAX ; + } + +return id; +} +static void handle_bff (struct mce *mce) {/* The bitfix filter overflowed, get the target CPU to reset it. */ if (wrmsrq_on_cpu (mce->extcpu,MSR_MCx_BFF_CTL (mce->bank),MCI_BFF_RESET ))pr_warn ("Failed to reset bitfix filter for CPU %d Bank %d\n" , mce->extcpu, mce->bank); + +pr_debug ("unique_id = 0x%lx\n" ,get_bff_id (mce)); }static int bff_mce_notify (struct notifier_block *nb,unsigned long val,void *data)
Signature · 1 line
i.
PATCH 7/7 platform/x86/intel/bff: Report frequent filter resets
In reply to Tony LuckIf an instance of a bitfix filter contains some transient errors built up over time, then resetting the filter will free up slots in the filter to store persistent errors.
Save a time stamp when "yellow" status is seen and clear the filter.
Log at KERN_WARNING level if the overflow occurred quickly after a previous overflow on the same bitfix filter instance. Use KERN_NOTICE for first, or long delayed, overflow.
--- drivers/platform/x86/intel/bff.c | 71 +++++++++++++++++++++++++++++++- 1 file changed, 70 insertions(+), 1 deletion(-)
drivers/platform/x86/intel/bff.c +70 -1
@@ -23,14 +23,19 @@#include <linux/cpufeature.h> #include <linux/cpuhplock.h> #include <linux/device-id/x86_cpu.h> +#include <linux/err.h> #include <linux/errno.h> +#include <linux/gfp_types.h> #include <linux/init.h> +#include <linux/jiffies.h> #include <linux/limits.h> #include <linux/module.h> #include <linux/notifier.h> #include <linux/printk.h> +#include <linux/slab.h> #include <linux/topology.h> #include <linux/types.h> +#include <linux/xarray.h> #include <asm/cpu_device_id.h> #include <asm/cpufeatures.h>
@@ -39,6 +44,16 @@#include <asm/msr.h> #include <asm/msr-index.h> +/* +* A 10-minute observation period helps distinguish between: +* +* - A long-term accumulation of transient corrected errors +* (filter stays clear after reset). +* +* - Permanent defects (filter overflows again quickly). +*/ +#define BFF_OVERFLOW_INTERVAL secs_to_jiffies(10 * 60) +#define NUM_IMH_PER_SKT 2/* Intel bitfix filter control register defines */
@@ -74,6 +89,8 @@ MODULE_DEVICE_TABLE(x86cpu, bff_cpu_ids);static const enum bff_type *bank_types; +static DEFINE_XARRAY (bff_bank_xa); +/* Diamond Rapids maps APICID[2] to the IMH instance. */ static void bff_set_imh_id (struct mce *mce,unsigned long *id) {
@@ -139,13 +156,58 @@ static unsigned long get_bff_id(struct mce *mce)return id; } +/* +* Save current timestamp for bff_id. Return true if it is within +* BFF_OVERFLOW_INTERVAL of previous timestamp for this bff_id. +*/ +static bool bff_note_event_and_check_burst (unsigned long bff_id) +{ +unsigned long now = jiffies, when; +unsigned long *ts; + +if (bff_id ==ULONG_MAX ) +return false; + + ts =xa_load (&bff_bank_xa, bff_id); +if (!ts) { + ts =kzalloc_obj (*ts); +if (!ts) { +pr_warn ("Timestamp allocation failed\n" ); +return false; + } +if (IS_ERR (xa_store (&bff_bank_xa, bff_id, ts,GFP_KERNEL ))) { +kfree (ts); +pr_warn ("Timestamp save failure\n" ); +return false; + } + *ts = now; + +return false; + } + + when = *ts +BFF_OVERFLOW_INTERVAL ; + *ts = now; + +return time_before (now, when); +} +static void handle_bff (struct mce *mce) { +bool burst = false; +/* The bitfix filter overflowed, get the target CPU to reset it. */ if (wrmsrq_on_cpu (mce->extcpu,MSR_MCx_BFF_CTL (mce->bank),MCI_BFF_RESET ))pr_warn ("Failed to reset bitfix filter for CPU %d Bank %d\n" , mce->extcpu, mce->bank); -pr_debug ("unique_id = 0x%lx\n" ,get_bff_id (mce)); +/* Get unique id for bitfix filter instance that overflowed */ + burst =bff_note_event_and_check_burst (get_bff_id (mce)); + +if (burst) +pr_warn_ratelimited (HW_ERR "Socket %d CPU %d Bank %d bitfix filter overflowed frequently\n" , + mce->socketid, mce->extcpu, mce->bank); +else +pr_notice_ratelimited (HW_ERR "Socket %d CPU %d Bank %d bitfix filter overflowed\n" , + mce->socketid, mce->extcpu, mce->bank); }static int bff_mce_notify (struct notifier_block *nb,unsigned long val,void *data)
@@ -205,7 +267,14 @@ static int __init bff_init(void) static void __exit bff_exit(void) { +unsigned long bff_id; +unsigned long *ts; +mce_unregister_decode_chain (&bff_notifier); + +xa_for_each (&bff_bank_xa, bff_id, ts) +kfree (ts); +xa_destroy (&bff_bank_xa); }module_init (bff_init);
Signature · 1 line
2.55.0
Ilpo Järvinen
maintainerRe: [PATCH 7/7] platform/x86/intel/bff: Report frequent filter resets
Quoted 28 lines · Tony Luck
If an instance of a bitfix filter contains some transient errors built up over time, then resetting the filter will free up slots in the filter to store persistent errors.
Save a time stamp when "yellow" status is seen and clear the filter.
Log at
KERN_WARNINGlevel if the overflow occurred quickly after a previous overflow on the same bitfix filter instance. UseKERN_NOTICEfor first, or long delayed, overflow.Co-developed-by Qiuxu Zhuo <[email protected]>Signed-off-by Qiuxu Zhuo <[email protected]>Signed-off-by Tony Luck <[email protected]>--- drivers/platform/x86/intel/bff.c | 71 +++++++++++++++++++++++++++++++- 1 file changed, 70 insertions(+), 1 deletion(-)drivers/platform/x86/intel/bff.c +2
-0@@ -23,14 +23,19 @@#include <linux/cpufeature.h> #include <linux/cpuhplock.h> #include <linux/device-id/x86_cpu.h> +#include <linux/err.h> #include <linux/errno.h> +#include <linux/gfp_types.h>
drivers/platform/x86/intel/bff.c line 38
Quoted 10 lines
drivers/platform/x86/intel/bff.c +3
-0#include <linux/init.h> +#include <linux/jiffies.h> #include <linux/limits.h> #include <linux/module.h> #include <linux/notifier.h> #include <linux/printk.h> +#include <linux/slab.h> #include <linux/topology.h> #include <linux/types.h> +#include <linux/xarray.h>
A bonus point from paying attention to having necessary includes already in v1 throughout the series. :-) (No action required, I'm just happy to see that for a change.)
Quoted 52 lines
drivers/platform/x86/intel/bff.c +32
-0#include <asm/cpu_device_id.h> #include <asm/cpufeatures.h> @@ -39,6 +44,16 @@#include <asm/msr.h> #include <asm/msr-index.h> +/* +* A 10-minute observation period helps distinguish between: +* +* - A long-term accumulation of transient corrected errors +* (filter stays clear after reset). +* +* - Permanent defects (filter overflows again quickly). +*/ +#define BFF_OVERFLOW_INTERVAL secs_to_jiffies(10 * 60) +#define NUM_IMH_PER_SKT 2/* Intel bitfix filter control register defines */ @@ -74,6 +89,8 @@ MODULE_DEVICE_TABLE(x86cpu, bff_cpu_ids);static const enum bff_type *bank_types; +static DEFINE_XARRAY (bff_bank_xa); +/* Diamond Rapids maps APICID[2] to the IMH instance. */ static void bff_set_imh_id (struct mce *mce,unsigned long *id) {@@ -139,13 +156,58 @@ static unsigned long get_bff_id(struct mce *mce)return id; } +/* +* Save current timestamp for bff_id. Return true if it is within +* BFF_OVERFLOW_INTERVAL of previous timestamp for this bff_id. +*/ +static bool bff_note_event_and_check_burst (unsigned long bff_id) +{ +unsigned long now = jiffies, when; +unsigned long *ts; + +if (bff_id ==ULONG_MAX ) +return false; + + ts =xa_load (&bff_bank_xa, bff_id); +if (!ts) { + ts =kzalloc_obj (*ts); +if (!ts) { +pr_warn ("Timestamp allocation failed\n" ); +return false; + } +if (IS_ERR (xa_store (&bff_bank_xa, bff_id, ts,GFP_KERNEL ))) {
drivers/platform/x86/intel/bff.c static unsigned long get_bff_id(struct mce *mce) · line 188
Quoted 10 lines
drivers/platform/x86/intel/bff.c +10
-0+kfree (ts); +pr_warn ("Timestamp save failure\n" ); +return false; + } + *ts = now; + +return false; + } + + when = *ts +BFF_OVERFLOW_INTERVAL ;
You could try to name "when" more descriptively so the name tells its the end of the ("too fast") burst time window.
Quoted 42 lines
drivers/platform/x86/intel/bff.c +23
-1+ *ts = now; + +return time_before (now, when); +} +static void handle_bff (struct mce *mce) { +bool burst = false; +/* The bitfix filter overflowed, get the target CPU to reset it. */ if (wrmsrq_on_cpu (mce->extcpu,MSR_MCx_BFF_CTL (mce->bank),MCI_BFF_RESET ))pr_warn ("Failed to reset bitfix filter for CPU %d Bank %d\n" , mce->extcpu, mce->bank); -pr_debug ("unique_id = 0x%lx\n" ,get_bff_id (mce)); +/* Get unique id for bitfix filter instance that overflowed */ + burst =bff_note_event_and_check_burst (get_bff_id (mce)); + +if (burst) +pr_warn_ratelimited (HW_ERR "Socket %d CPU %d Bank %d bitfix filter overflowed frequently\n" , + mce->socketid, mce->extcpu, mce->bank); +else +pr_notice_ratelimited (HW_ERR "Socket %d CPU %d Bank %d bitfix filter overflowed\n" , + mce->socketid, mce->extcpu, mce->bank); }static int bff_mce_notify (struct notifier_block *nb,unsigned long val,void *data)@@ -205,7 +267,14 @@ static int __init bff_init(void) static void __exit bff_exit(void) { +unsigned long bff_id; +unsigned long *ts; +mce_unregister_decode_chain (&bff_notifier); + +xa_for_each (&bff_bank_xa, bff_id, ts) +kfree (ts); +xa_destroy (&bff_bank_xa); }module_init (bff_init);
Signature · 1 line
i.
Borislav Petkov
maintainerQuoted 2 lines · Tony Luck
This code was originally developend by Qiuxu as an add-on to an EDAC driver. But not everyone wants to run an EDAC driver,
Interesting, what's so bad about EDAC drivers that makes you move all that functionality which is clearly RAS/MCE/EDAC into some random platform driver?
There's also arch/x86/kernel/cpu/mce/intel.c and that doesn't fit either?
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquetteQuoted 5 lines
Quoted 2 lines
This code was originally developend by Qiuxu as an add-on to an EDAC driver. But not everyone wants to run an EDAC driver,
Interesting, what's so bad about EDAC drivers that makes you move all that functionality which is clearly RAS/MCE/EDAC into some random platform driver?
Mostly that not everyone wants to run EDAC drivers.
Quoted 1 line
There's also
arch/x86/kernel/cpu/mce/intel.cand that doesn't fit either?
Plausibly. But it is model specific. Needs to know how h/w banks are shared between logical CPUs so it can get the time comparisons right when a shared bank reports BFF overflow on different CPUs. Intel doesn't have any enumeration for bank sharing, and changes things often.
It's also just for one (not yet released) CPU model today. So, building it into the MCE code would be overhead for almost everyone.
But I can move it if you think that is a better place for it.
-Tony
Borislav Petkov
maintainerQuoted 1 line · Tony
Mostly that not everyone wants to run EDAC drivers.
Just because or is there a particular reason?
Because we could try to address those reasons if they were more concrete and valid...
Quoted 9 lines
Plausibly. But it is model specific. Needs to know how h/w banks are shared between logical CPUs so it can get the time comparisons right when a shared bank reports BFF overflow on different CPUs. Intel doesn't have any enumeration for bank sharing, and changes things often.
It's also just for one (not yet released) CPU model today. So, building it into the MCE code would be overhead for almost everyone.
But I can move it if you think that is a better place for it.
Well, my angle is: we already have soo much RAS glue in the kernel so adding a *platform* driver for it is simply unnecessary.
For example, drivers/edac/mce_amd.c is the whole AMD MCE decoding and even though it is in drivers/edac/, it is not really an EDAC driver. So your BFFs (wonderful acronym btw :-P) would likely fit there too.
And looking at the code, it looks very familiar to that thing - simply a notifier callback with a bunch of logic to decode and report the error.
And there's drivers/ras/ too.
And we already have the whole machinery around it so let's move it somewhere more fitting than in yet another new place pls.
Thx.
Signature · 4 lines
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette