hulk inclusion category: feature bugzilla: https://atomgit.com/openeuler/kernel/issues/9116 ------------------------------------------ Background: SEI (SError Interrupt) is an asynchronous exception on ARM64 systems. In firmware-first RAS model, firmware handle SEI errors before the OS, recording error information in vendor-specific registers. Problem: When SEI occurs in userspace with Uncorrected Recoverable/Unrecoverable AET (Abort Error Type), the kernel currently cannot recover the error because it doesn't have access to the error information stored by firmware in vendor-specific registers. Solution: Add a vendor handler interface (apei_claim_sei) allowing each SoC vendor to implement platform-specific SEI handling. Signed-off-by: Wupeng Ma <mawupeng1@huawei.com> --- arch/arm64/Kconfig | 7 +- arch/arm64/include/asm/acpi.h | 2 + arch/arm64/kernel/traps.c | 16 +++ drivers/acpi/apei/apei-internal.h | 2 + drivers/acpi/apei/ghes-vendor-info.c | 162 ++++++++++++++++++++++++++- drivers/acpi/apei/ghes.c | 5 + 6 files changed, 192 insertions(+), 2 deletions(-) diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig index 86fbcc277e56..81d2baafdcd6 100644 --- a/arch/arm64/Kconfig +++ b/arch/arm64/Kconfig @@ -2507,11 +2507,16 @@ config ARM64_PSEUDO_NMI config ARM64_SYNC_SEI bool "Use ESB to Synchronize SEI At Exception Boundary(EXPERIMENTAL)" - depends on ARM64_RAS_EXTN + depends on ARM64_RAS_EXTN && ACPI_APEI_GHES_ARMP_VENDOR_INFO help For Firmware-First, Use the ESB to synchronize SEI occurs before exception entry from EL0 and exit to EL0. + SEI is an asynchronous exception that can occur on ARMv8 systems. + When firmware handles SEI first, it analyzes and records the error. + This option allows the OS to retrieve the error record and take + appropriate action, such as killing the affected task. + if unsure, say N if ARM64_PSEUDO_NMI diff --git a/arch/arm64/include/asm/acpi.h b/arch/arm64/include/asm/acpi.h index c07a58b96329..693dc4c7b817 100644 --- a/arch/arm64/include/asm/acpi.h +++ b/arch/arm64/include/asm/acpi.h @@ -134,9 +134,11 @@ static inline int get_cpu_for_acpi_id(u32 uid) static inline void arch_fix_phys_package_id(int num, u32 slot) { } void __init acpi_init_cpus(void); int apei_claim_sea(struct pt_regs *regs); +int apei_claim_sei(struct pt_regs *regs); #else static inline void acpi_init_cpus(void) { } static inline int apei_claim_sea(struct pt_regs *regs) { return -ENOENT; } +static inline int apei_claim_sei(struct pt_regs *regs) { return -ENOENT; } #endif /* CONFIG_ACPI */ #ifdef CONFIG_ARM64_ACPI_PARKING_PROTOCOL diff --git a/arch/arm64/kernel/traps.c b/arch/arm64/kernel/traps.c index 08a5e3d30919..e8ef661715f0 100644 --- a/arch/arm64/kernel/traps.c +++ b/arch/arm64/kernel/traps.c @@ -30,6 +30,7 @@ #include <linux/kasan.h> #include <linux/ubsan.h> #include <linux/cfi.h> +#include <linux/acpi.h> #include <asm/atomic.h> #include <asm/bug.h> @@ -49,6 +50,8 @@ #include <asm/stacktrace.h> #include <asm/system_misc.h> #include <asm/sysreg.h> +#include <asm/acpi.h> +#include <asm/setup.h> static bool __kprobes __check_eq(unsigned long pstate) { @@ -971,6 +974,12 @@ bool arm64_is_fatal_ras_serror(struct pt_regs *regs, unsigned long esr) { unsigned long aet = arm64_ras_serror_get_severity(esr); + pr_info_ratelimited( + "%s aet: %ld comm: %.20s tgid: %d pid: %d cpu: %d\n", + (regs && user_mode(regs)) ? "userspace" : "kernelspace", aet, + current->comm, current->tgid, current->pid, + raw_smp_processor_id()); + switch (aet) { case ESR_ELx_AET_CE: /* corrected error */ case ESR_ELx_AET_UEO: /* restartable, not yet consumed */ @@ -989,7 +998,14 @@ bool arm64_is_fatal_ras_serror(struct pt_regs *regs, unsigned long esr) * Neoverse-N1 #1349291 means a non-KVM SError reported as * Unrecoverable should be treated as Uncontainable. We * call arm64_serror_panic() in both cases. + * + * Add a vendor handler interface (apei_claim_sei) allowing + * each SoC vendor to implement platform-specific SEI handling + * for UER. */ + if ((aet == ESR_ELx_AET_UER) && !apei_claim_sei(regs)) + return false; + return true; case ESR_ELx_AET_UC: /* Uncontainable or Uncategorized error */ diff --git a/drivers/acpi/apei/apei-internal.h b/drivers/acpi/apei/apei-internal.h index f5519153e32d..50bf372637b6 100644 --- a/drivers/acpi/apei/apei-internal.h +++ b/drivers/acpi/apei/apei-internal.h @@ -142,11 +142,13 @@ static inline bool apei_page_should_offline(unsigned long pfn) #ifdef CONFIG_ACPI_APEI_GHES_ARMP_VENDOR_INFO bool ghes_armp_vendor_critical_error(struct cper_sec_proc_arm *err, bool sync); +int ghes_armp_vendor_handle_sei(struct pt_regs *regs); #else static inline bool ghes_armp_vendor_critical_error(struct cper_sec_proc_arm *err, bool sync) { return false; } +static inline int ghes_armp_vendor_handle_sei(struct pt_regs *regs) { return -ENOENT; } #endif #endif diff --git a/drivers/acpi/apei/ghes-vendor-info.c b/drivers/acpi/apei/ghes-vendor-info.c index 55db619638a2..d8bad5be2502 100644 --- a/drivers/acpi/apei/ghes-vendor-info.c +++ b/drivers/acpi/apei/ghes-vendor-info.c @@ -4,10 +4,19 @@ * Copyright (c) Huawei Technologies Co., Ltd. 2026. All rights reserved. */ +#define pr_fmt(fmt) "GHES: VENDOR: " fmt + #include <linux/init.h> #include <linux/acpi.h> +#include <linux/signal.h> +#include <linux/task_work.h> +#include <linux/genalloc.h> + #include <acpi/ghes.h> #include <acpi/apei.h> + +#include <asm/setup.h> + #include "apei-internal.h" #define HISI_OEM BIT(0) @@ -57,11 +66,152 @@ static bool ghes_hisi_critical_hw_error(struct cper_sec_proc_arm *err, bool sync return (bool)(vendor_info->err_flag & HISI_VENDOR_CRITICAL_ERR); } + +struct sei_task_work { + struct callback_head twork; + u64 pfn; +}; + +static struct gen_pool *hisi_sei_pool; + +static int ghes_hisi_sei_init(void) +{ + unsigned long addr, len = PAGE_SIZE; + int rc; + + if (!IS_ENABLED(CONFIG_ARM64_SYNC_SEI)) + return 0; + + hisi_sei_pool = gen_pool_create(ilog2(sizeof(struct sei_task_work)), -1); + if (!hisi_sei_pool) + return -ENOMEM; + + addr = (unsigned long)kzalloc(PAGE_ALIGN(len), GFP_KERNEL); + if (!addr) + goto err_pool_alloc; + + rc = gen_pool_add(hisi_sei_pool, addr, PAGE_ALIGN(len), -1); + if (rc) + goto err_pool_add; + + return 0; + +err_pool_add: + kfree((void *)addr); + +err_pool_alloc: + gen_pool_destroy(hisi_sei_pool); + hisi_sei_pool = NULL; + + pr_warn("%s init failed\n", __func__); + return -ENOMEM; +} + +static void hisi_sei_kill_task_work(struct callback_head *twork) +{ + struct sei_task_work *ctx = container_of(twork, struct sei_task_work, twork); + + kill_accessing_process(ctx->pfn, MF_ACTION_REQUIRED, true); + gen_pool_free(hisi_sei_pool, (unsigned long)ctx, sizeof(*ctx)); +} + +/* + * Read SEI error address from HiSilicon RAS registers. + * - s3_3_c15_c0_1 lower 16 bits combined with s3_3_c15_c0_0 form the error physical address. + * - Clear the registers after reading to acknowledge the error. + */ +static inline u64 hisi_sei_get_error_pa(void) +{ + u64 sw_res_reg0 = read_sysreg(s3_3_c15_c0_0); + u64 sw_res_reg1 = read_sysreg(s3_3_c15_c0_1); + u64 pa = ((sw_res_reg1 & 0xFFFFUL) << 32) | (sw_res_reg0 & 0xFFFFFFFFUL); + + write_sysreg(0, s3_3_c15_c0_0); + write_sysreg(sw_res_reg1 & ~0xFFFFUL, s3_3_c15_c0_1); + + return pa; +} + +/* + * Handle fatal SEI error by scheduling a task work to kill the affected process. + * @err_pa: The physical address that triggered the SEI. + * + * This function allocates a task work structure from a pre-allocated pool and + * schedules it to run on the current task. The task work will invoke + * kill_accessing_process() to send a SIGKILL to the process that has the + * error address mapped. This mechanism is used for memory errors in user-space + * accessible regions managed by drivers. + * + * Return: true if task work is successfully scheduled, false otherwise. + */ +static bool hisi_sei_kill_task(void) +{ + struct sei_task_work *ctx; + unsigned long err_pa; + + if (!hisi_sei_pool) + return false; + + ctx = (void *)gen_pool_alloc(hisi_sei_pool, sizeof(*ctx)); + if (!ctx) { + pr_warn_ratelimited("alloc task work failed\n"); + return false; + } + + err_pa = hisi_sei_get_error_pa(); + if (!err_pa) { + pr_warn_ratelimited("err pa is not valid\n"); + gen_pool_free(hisi_sei_pool, (unsigned long)ctx, sizeof(*ctx)); + return false; + } + + ctx->pfn = PHYS_PFN(err_pa); + init_task_work(&ctx->twork, hisi_sei_kill_task_work); + if (task_work_add(current, &ctx->twork, TWA_RESUME)) { + pr_warn_ratelimited("task work add failed\n"); + gen_pool_free(hisi_sei_pool, (unsigned long)ctx, sizeof(*ctx)); + return false; + } + return true; +} + +/* + * Handle HiSilicon specific Synchronous External Interrupt (SEI) errors. + * @regs: exception registers, NULL if from user space + * + * This function processes vendor-specific SEI errors for HiSilicon platforms. + * For user space errors, it reads the error physical address from RAS registers + * and schedules a task work to kill the accessing task. If recovery fails or + * the error is from kernel space, the current process is terminated with SIGKILL. + * + * Return: 0 if SEI is handled, -ENOENT if not applicable or unsupported. + */ +static int ghes_hisi_handle_sei(struct pt_regs *regs) +{ + if (!IS_ENABLED(CONFIG_ARM64_SYNC_SEI)) + return -ENOENT; + + if (!arm64_sync_sei_enabled()) + return -ENOENT; + + if (!current->mm) + return -ENOENT; + + if ((!regs || user_mode(regs)) && hisi_sei_kill_task()) + return 0; + + pr_err("Sending SIGKILL to comm: %s, pid: %d, tgid: %d due to sei not recovered\n", + current->comm, current->pid, current->tgid); + force_sig(SIGKILL); + return 0; +} #else static inline bool ghes_hisi_critical_hw_error(struct cper_sec_proc_arm *err, bool sync) { return false; } +static int ghes_hisi_sei_init(void) { return 0; } +static int ghes_hisi_handle_sei(struct pt_regs *regs) { return -ENOENT; } #endif bool ghes_armp_vendor_critical_error(struct cper_sec_proc_arm *err, bool sync) @@ -72,6 +222,14 @@ bool ghes_armp_vendor_critical_error(struct cper_sec_proc_arm *err, bool sync) return false; } +int ghes_armp_vendor_handle_sei(struct pt_regs *regs) +{ + if (vender_oem & HISI_OEM) + return ghes_hisi_handle_sei(regs); + + return -ENOENT; +} + static int __init ghes_check_oem_table(void) { struct acpi_table_header *tbl; @@ -81,8 +239,10 @@ static int __init ghes_check_oem_table(void) if (ACPI_FAILURE(status) || !tbl) return -ENODEV; - if (!memcmp(tbl->oem_id, "HISI ", ACPI_OEM_ID_SIZE)) + if (!memcmp(tbl->oem_id, "HISI ", ACPI_OEM_ID_SIZE)) { vender_oem |= HISI_OEM; + ghes_hisi_sei_init(); + } acpi_put_table(tbl); return 0; diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c index 34ab844bd013..1d8890405a87 100644 --- a/drivers/acpi/apei/ghes.c +++ b/drivers/acpi/apei/ghes.c @@ -1699,3 +1699,8 @@ void ghes_unregister_report_chain(struct notifier_block *nb) atomic_notifier_chain_unregister(&ghes_report_chain, nb); } EXPORT_SYMBOL_GPL(ghes_unregister_report_chain); + +int apei_claim_sei(struct pt_regs *regs) +{ + return ghes_armp_vendor_handle_sei(regs); +} -- 2.43.0