[RFC PATCH 0/2] EDAC/ghes: Add EDAC device for recording the CPU cache error count
ARM64 server chip Kunpeng 920 occasionally detected L2 cache corrected errors on few hardware. Thus the probability of the CPU cache errors occurring often exists. The earlier failure detection by monitoring the cache corrected errors for the frequent occurrences and taking preventive action could prevent more serious hardware faults. For the firmware-first error handling, especially on ARM64 architectures, there is no provision for recording the CPU cache corrected error count is present. For Intel architectures cache corrected errors reporting and offline the cores is done through more architecture specific method. http://www.mcelog.org/cache.html For this purpose, the suggestion was to create the CPU EDAC device for the CPU caches for recording the cache error count. The EDAC device blocks for the CPU caches would be created based on the cache information obtained from the cpu_cacheinfo. User-space application could monitor the recorded corrected error count for the earlier hardware failure detection and could take preventive action. Changes: RFC V1 -> RFC V2: 1. Fixed feedback by Boris. 1.1. Added reason of this patch. 1.2. Changed CPU errors to CPU cache errors in the drivers/edac/Kconfig 1.3 Changed EDAC cache list to percpu variables. 2. Changes in the descriptions. Shiju Jose (2): EDAC/ghes: Add EDAC device for reporting the CPU cache errors ACPI / APEI: Add reporting ARM64 CPU cache corrected error count Documentation/ABI/testing/sysfs-devices-edac | 15 ++ drivers/acpi/apei/ghes.c | 76 +++++++- drivers/edac/Kconfig | 10 + drivers/edac/ghes_edac.c | 181 +++++++++++++++++++ include/acpi/ghes.h | 27 +++ include/linux/cper.h | 4 + 6 files changed, 309 insertions(+), 4 deletions(-) -- 2.17.1
ARM64 server chip Kunpeng 920 occasionally detected L2 cache corrected errors on few hardware. Thus the probability of the CPU cache errors occurring often exists. The earlier failure detection by monitoring the cache corrected errors for the frequent occurrences and taking preventive action could prevent more serious hardware faults. For the firmware-first error reporting, especially on ARM64 architectures, there is no provision present for reporting the cache corrected error count to the user-space. For Intel architectures cache corrected errors reporting and offline the cores is done through more architecture specific method. http://www.mcelog.org/cache.html For this purpose, suggestion was to create the CPU EDAC device for the CPU caches for reporting the cache error count for the firmware-first error reporting. The EDAC device blocks for the CPU caches would be created based on the cache information obtained from the cpu_cacheinfo. User-space application could monitor the recorded corrected error count for the earlier hardware failure detection and could take preventive action, such as offline the corresponding CPU core/s. Add an EDAC device and device blocks for the CPU caches based on the cache information from the cpu_cacheinfo. The cache's corrected error count would be stored in the /sys/devices/system/edac/cpu/cpu*/cache*/ce_count. Issues and possible solutions, 1.Cache info is not available for the CPUs offline. EDAC device interface requires creating EDAC device and device blocks together. It requires the number of caches per CPU as device blocks for the creation. However, this info is not available for the offlined CPUs. Tested Solution: Find the max number of caches among online CPUs, create the EDAC device for CPUs caches and get and populate the cache info for an offline CPU later, when the error is reported on that CPU for the first time. 2. Reporting error count for the Shared caches. There are few possible solutions, 2.1 Kernel would report a new error count for a shared cache through the EDAC device block for that CPU on which the error is reported. Then user-space application would sum the total error count from EDAC device block of all the CPUs in the shared CPU list of that shared cache. 2.2 Kernel would report a new error count for a shared cache through the EDAC device blocks for all the CPUs in the shared CPU list of that shared cache. The current implementation used the solution 2.1 For the firmware-first error reporting, add an interface in the ghes_edac for reporting a CPU corrected error count. Note: CONFIG_EDAC_GHES_CPU_CACHE_ERROR is added to make this feature optional only for the platforms which supported and tested. Signed-off-by: Shiju Jose <shiju.jose@huawei.com> --- Documentation/ABI/testing/sysfs-devices-edac | 15 ++ drivers/acpi/apei/ghes.c | 8 +- drivers/edac/Kconfig | 10 + drivers/edac/ghes_edac.c | 181 +++++++++++++++++++ include/acpi/ghes.h | 27 +++ 5 files changed, 240 insertions(+), 1 deletion(-) diff --git a/Documentation/ABI/testing/sysfs-devices-edac b/Documentation/ABI/testing/sysfs-devices-edac index 256a9e990c0b..56a18b0af419 100644 --- a/Documentation/ABI/testing/sysfs-devices-edac +++ b/Documentation/ABI/testing/sysfs-devices-edac @@ -155,3 +155,18 @@ Description: This attribute file displays the total count of uncorrectable errors that have occurred on this DIMM. If panic_on_ue is set, this counter will not have a chance to increment, since EDAC will panic the system + +What: /sys/devices/system/edac/cpu/cpu*/cache*/ce_count +Date: December 2020 +Contact: linux-edac@vger.kernel.org +Description: This attribute file displays the total count of correctable + errors that have occurred on this CPU cache. This count is very important + to examine. CEs provide early indications that a cache is beginning + to fail. This count field should be monitored for non-zero values + and report such information to the system administrator. + +What: /sys/devices/system/edac/cpu/cpu*/cache*/ue_count +Date: December 2020 +Contact: linux-edac@vger.kernel.org +Description: This attribute file displays the total count of uncorrectable + errors that have occurred on this CPU cache. diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c index fce7ade2aba9..e7b0edbda0f8 100644 --- a/drivers/acpi/apei/ghes.c +++ b/drivers/acpi/apei/ghes.c @@ -1452,4 +1452,10 @@ static int __init ghes_init(void) err: return rc; } -device_initcall(ghes_init); + +/* + * device_initcall_sync() is added instead of the device_initcall() + * because the CPU cacheinfo should be populated and needed for + * adding the CPU cache edac device blocks in the ghes_edac_register(). + */ +device_initcall_sync(ghes_init); diff --git a/drivers/edac/Kconfig b/drivers/edac/Kconfig index 81c42664f21b..fbbd3869a47e 100644 --- a/drivers/edac/Kconfig +++ b/drivers/edac/Kconfig @@ -74,6 +74,16 @@ config EDAC_GHES In doubt, say 'Y'. +config EDAC_GHES_CPU_CACHE_ERROR + bool "EDAC device for reporting firmware-first BIOS detected CPU cache error count" + depends on EDAC_GHES + help + EDAC device for the firmware-first BIOS detected CPU cache error count, + reported via ACPI APEI/GHES. By enabling this option, EDAC device for + the CPU hierarchy and edac device blocks for the caches would be created. + The cache error count is shared with the userspace via the CPU EDAC + device's sysfs interface. + config EDAC_AMD64 tristate "AMD64 (Opteron, Athlon64)" depends on AMD_NB && EDAC_DECODE_MCE diff --git a/drivers/edac/ghes_edac.c b/drivers/edac/ghes_edac.c index 6d1ddecbf0da..31d013efcc63 100644 --- a/drivers/edac/ghes_edac.c +++ b/drivers/edac/ghes_edac.c @@ -12,6 +12,9 @@ #include <acpi/ghes.h> #include <linux/edac.h> #include <linux/dmi.h> +#if defined(CONFIG_EDAC_GHES_CPU_CACHE_ERROR) +#include <linux/cacheinfo.h> +#endif #include "edac_module.h" #include <ras/ras_event.h> @@ -57,6 +60,21 @@ module_param(force_load, bool, 0); static bool system_scanned; +#if defined(CONFIG_EDAC_GHES_CPU_CACHE_ERROR) +struct ghes_edac_cpu_block { + int cpu; + u8 level; + u8 type; + int block_nr; + bool info_populated; +}; + +static struct ghes_edac_cpu_block __percpu *edac_cpu_block_list; + +static struct edac_device_ctl_info *cpu_edac_dev; +static unsigned int num_caches_per_cpu; +#endif + /* Memory Device - Type 17 of SMBIOS spec */ struct memdev_dmi_entry { u8 type; @@ -225,6 +243,161 @@ static void enumerate_dimms(const struct dmi_header *dh, void *arg) hw->num_dimms++; } +#if defined(CONFIG_EDAC_GHES_CPU_CACHE_ERROR) +static int ghes_edac_add_cpu_device(struct device *dev) +{ + int rc; + + cpu_edac_dev = edac_device_alloc_ctl_info(0, "cpu", num_possible_cpus(), + "cache", num_caches_per_cpu, 0, NULL, + 0, edac_device_alloc_index()); + if (!cpu_edac_dev) { + pr_warn("ghes-edac cpu edac device registration failed\n"); + return -ENOMEM; + } + + cpu_edac_dev->dev = dev; + cpu_edac_dev->ctl_name = "cpu_edac_dev"; + cpu_edac_dev->dev_name = "ghes"; + cpu_edac_dev->mod_name = "ghes_edac.c"; + rc = edac_device_add_device(cpu_edac_dev); + if (rc > 0) { + pr_warn("edac_device_add_device failed\n"); + edac_device_free_ctl_info(cpu_edac_dev); + return -ENOMEM; + } + + return 0; +} + +static void ghes_edac_delete_cpu_device(void) +{ + num_caches_per_cpu = 0; + if (cpu_edac_dev) { + edac_device_del_device(cpu_edac_dev->dev); + edac_device_free_ctl_info(cpu_edac_dev); + } + free_percpu(edac_cpu_block_list); +} + +static int ghes_edac_populate_cache_info(int cpu) +{ + struct ghes_edac_cpu_block *block; + struct cpu_cacheinfo *this_cpu_ci; + struct cacheinfo *this_leaf; + int i, num_caches; + + this_cpu_ci = get_cpu_cacheinfo(cpu); + if (!this_cpu_ci || !this_cpu_ci->info_list || !this_cpu_ci->num_leaves) + return -EINVAL; + + this_leaf = this_cpu_ci->info_list; + /* + * Cache info would not be available for a CPU which is offline. However EDAC device + * creation requires the number of device blocks (for example max number of caches + * among CPUs). The cache info in the edac_cpu_block_list would be populated when + * the first error is reported on that cpu. Thus we need to restrict the number + * of caches if the CPU's num_leaves exceed the max number of caches per cpu + * calculated in the init time. + */ + num_caches = min(num_caches_per_cpu, this_cpu_ci->num_leaves); + + /* + * The edac CPU cache device blocks entries in the sysfs should match with the + * CPU cache structure in the sysfs so that the affected cpus for a shared cache + * can be easily extracted in the userspace. + */ + block = per_cpu_ptr(edac_cpu_block_list, cpu); + for (i = 0; i < num_caches; i++) { + block->cpu = cpu; + block->level = this_leaf->level; + block->type = this_leaf->type; + block->block_nr = i; + block->info_populated = true; + this_leaf++; + block++; + } + + return 0; +} + +static void ghes_edac_create_cpu_device(struct device *dev) +{ + int cpu; + struct cpu_cacheinfo *this_cpu_ci; + + /* + * Find the maximum number of caches present in the CPU heirarchy + * among the online CPUs. + */ + for_each_online_cpu(cpu) { + this_cpu_ci = get_cpu_cacheinfo(cpu); + if (!this_cpu_ci) + continue; + if (num_caches_per_cpu < this_cpu_ci->num_leaves) + num_caches_per_cpu = this_cpu_ci->num_leaves; + } + if (!num_caches_per_cpu) + return; + + /* + * Allocate EDAC CPU cache list. + * EDAC device interface require creating the CPU cache hierarchy for all + * the CPUs together. Thus need to allocate edac_cpu_block_list for the + * maximum number of caches per cpu among all the CPUs irrespective of + * the number of caches per CPU might vary. + */ + edac_cpu_block_list = __alloc_percpu(sizeof(struct ghes_edac_cpu_block) * num_caches_per_cpu, + __alignof__(u64)); + if (!edac_cpu_block_list) + return; + + if (ghes_edac_add_cpu_device(dev)) + goto error; + + /* + * Populate EDAC CPU cache list with cache's information. + */ + for_each_online_cpu(cpu) + ghes_edac_populate_cache_info(cpu); + + return; + +error: + ghes_edac_delete_cpu_device(); +} + +void ghes_edac_report_cpu_error(struct ghes_einfo_cpu *einfo) +{ + struct ghes_edac_cpu_block *block; + int i; + + if (!einfo || !(einfo->ce_count) || !num_caches_per_cpu) + return; + + /* + * EDAC device require the number of device blocks (for example max number of caches + * among CPUs) during the creation. For the CPUs that were offline in the cpu edac + * init and become online later, the cache info would be populated when the first + * error is reported on that cpu. + */ + block = per_cpu_ptr(edac_cpu_block_list, einfo->cpu); + if (!block->info_populated) { + if (ghes_edac_populate_cache_info(einfo->cpu)) + return; + } + + for (i = 0; i < num_caches_per_cpu; i++) { + if ((block->level == einfo->cache_level) && (block->type == einfo->cache_type)) { + edac_device_handle_ce_count(cpu_edac_dev, einfo->ce_count, + einfo->cpu, block->block_nr, ""); + break; + } + block++; + } +} +#endif + static void ghes_scan_system(void) { if (system_scanned) @@ -620,6 +793,10 @@ int ghes_edac_register(struct ghes *ghes, struct device *dev) goto unlock; } +#if defined(CONFIG_EDAC_GHES_CPU_CACHE_ERROR) + ghes_edac_create_cpu_device(dev); +#endif + spin_lock_irqsave(&ghes_lock, flags); ghes_pvt = pvt; spin_unlock_irqrestore(&ghes_lock, flags); @@ -669,6 +846,10 @@ void ghes_edac_unregister(struct ghes *ghes) if (mci) edac_mc_free(mci); +#if defined(CONFIG_EDAC_GHES_CPU_CACHE_ERROR) + ghes_edac_delete_cpu_device(); +#endif + unlock: mutex_unlock(&ghes_reg_mutex); } diff --git a/include/acpi/ghes.h b/include/acpi/ghes.h index 34fb3431a8f3..e019ad88fdc3 100644 --- a/include/acpi/ghes.h +++ b/include/acpi/ghes.h @@ -73,6 +73,24 @@ void ghes_unregister_vendor_record_notifier(struct notifier_block *nb); int ghes_estatus_pool_init(int num_ghes); +/** + * struct ghes_einfo_cpu - structure to pass the CPU error info to the edac + * @cpu: CPU index. + * @error_type: error type, cache/TLB/bus/ etc. + * @cache_level: cache level. + * @cache_type: ACPI cache type. + * @ue_count: CPU uncorrectable error count. + * @ce_count: CPU correctable error count. + */ +struct ghes_einfo_cpu { + int cpu; + u8 error_type; + u8 cache_level; + u8 cache_type; + u16 ue_count; + u16 ce_count; +}; + /* From drivers/edac/ghes_edac.c */ #ifdef CONFIG_EDAC_GHES @@ -98,6 +116,15 @@ static inline void ghes_edac_unregister(struct ghes *ghes) } #endif +#ifdef CONFIG_EDAC_GHES_CPU_CACHE_ERROR +void ghes_edac_report_cpu_error(struct ghes_einfo_cpu *einfo_cpu); + +#else +static inline void ghes_edac_report_cpu_error(struct ghes_einfo_cpu *einfo_cpu) +{ +} +#endif + static inline int acpi_hest_get_version(struct acpi_hest_generic_data *gdata) { return gdata->revision >> 8; -- 2.17.1
Add reporting ARM64 CPU cache corrected error count to the ghes_edac. The error count would be updated in the EDAC CPU cache sysfs interface. Signed-off-by: Shiju Jose <shiju.jose@huawei.com> --- drivers/acpi/apei/ghes.c | 68 ++++++++++++++++++++++++++++++++++++++-- include/linux/cper.h | 4 +++ 2 files changed, 69 insertions(+), 3 deletions(-) diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c index e7b0edbda0f8..37f8b09d810d 100644 --- a/drivers/acpi/apei/ghes.c +++ b/drivers/acpi/apei/ghes.c @@ -41,6 +41,7 @@ #include <linux/uuid.h> #include <linux/ras.h> #include <linux/task_work.h> +#include <linux/cacheinfo.h> #include <acpi/actbl1.h> #include <acpi/ghes.h> @@ -523,6 +524,69 @@ static void ghes_handle_aer(struct acpi_hest_generic_data *gdata) #endif } +static u8 arm_err_transaction_type_to_cache_type(u8 trans_type) +{ + switch (trans_type) { + case CPER_ARM_CACHE_TRANS_TYPE_INSTRUCTION: + return CACHE_TYPE_INST; + case CPER_ARM_CACHE_TRANS_TYPE_DATA: + return CACHE_TYPE_DATA; + case CPER_ARM_CACHE_TRANS_TYPE_GENERIC: + default: + return CACHE_TYPE_UNIFIED; + } +} + +static void ghes_handle_arm_hw_error(struct acpi_hest_generic_data *gdata) +{ + struct cper_sec_proc_arm *error = acpi_hest_get_payload(gdata); + struct cper_arm_err_info *err_info; + struct ghes_einfo_cpu einfo; + u8 transaction_type; + u64 error_info; + int sec_sev; + int i; + + log_arm_hw_error(error); + + sec_sev = ghes_severity(gdata->error_severity); + +#if defined(CONFIG_ARM64) + if (sec_sev == GHES_SEV_CORRECTED) { + memset(&einfo, 0, sizeof(einfo)); + einfo.cpu = get_logical_index(error->mpidr); + if (einfo.cpu == -EINVAL) + return; + + /* + * ARM processor error types are cache/TLB/bus errors. + * Presently corrected error count for caches only + * is reported. + */ + err_info = (struct cper_arm_err_info *)(error + 1); + + for (i = 0; i < error->err_info_num; i++) { + if (err_info->type != CPER_ARM_CACHE_ERROR) + continue; + einfo.ce_count = err_info->multiple_error + 1; + + error_info = err_info->error_info; + if (!(error_info & CPER_ARM_ERR_VALID_TRANSACTION_TYPE) || + !(error_info & CPER_ARM_ERR_VALID_LEVEL)) + continue; + + transaction_type = ((error_info >> CPER_ARM_ERR_TRANSACTION_SHIFT) + & CPER_ARM_ERR_TRANSACTION_MASK); + einfo.cache_type = arm_err_transaction_type_to_cache_type(transaction_type); + einfo.cache_level = ((error_info >> CPER_ARM_ERR_LEVEL_SHIFT) + & CPER_ARM_ERR_LEVEL_MASK); + ghes_edac_report_cpu_error(&einfo); + err_info += 1; + } + } +#endif +} + static BLOCKING_NOTIFIER_HEAD(vendor_record_notify_list); int ghes_register_vendor_record_notifier(struct notifier_block *nb) @@ -605,9 +669,7 @@ static bool ghes_do_proc(struct ghes *ghes, ghes_handle_aer(gdata); } else if (guid_equal(sec_type, &CPER_SEC_PROC_ARM)) { - struct cper_sec_proc_arm *err = acpi_hest_get_payload(gdata); - - log_arm_hw_error(err); + ghes_handle_arm_hw_error(gdata); } else { void *err = acpi_hest_get_payload(gdata); diff --git a/include/linux/cper.h b/include/linux/cper.h index 6a511a1078ca..0ea966af6ad9 100644 --- a/include/linux/cper.h +++ b/include/linux/cper.h @@ -314,6 +314,10 @@ enum { #define CPER_ARM_ERR_ACCESS_MODE_SHIFT 43 #define CPER_ARM_ERR_ACCESS_MODE_MASK GENMASK(0,0) +#define CPER_ARM_CACHE_TRANS_TYPE_INSTRUCTION 0 +#define CPER_ARM_CACHE_TRANS_TYPE_DATA 1 +#define CPER_ARM_CACHE_TRANS_TYPE_GENERIC 2 + /* * All tables and structs must be byte-packed to match CPER * specification, since the tables are provided by the system BIOS -- 2.17.1
participants (1)
-
Shiju Jose