ACPI, APEI, Generic Hardware Error Source memory error support
authorHuang Ying <ying.huang@intel.com>
Tue, 18 May 2010 06:35:20 +0000 (14:35 +0800)
committerLen Brown <len.brown@intel.com>
Thu, 20 May 2010 02:41:16 +0000 (22:41 -0400)
Generic Hardware Error Source provides a way to report platform
hardware errors (such as that from chipset). It works in so called
"Firmware First" mode, that is, hardware errors are reported to
firmware firstly, then reported to Linux by firmware. This way, some
non-standard hardware error registers or non-standard hardware link
can be checked by firmware to produce more valuable hardware error
information for Linux.

Now, only SCI notification type and memory errors are supported. More
notification type and hardware error type will be added later. These
memory errors are reported to user space through /dev/mcelog via
faking a corrected Machine Check, so that the error memory page can be
offlined by /sbin/mcelog if the error count for one page is beyond the
threshold.

On some machines, Machine Check can not report physical address for
some corrected memory errors, but GHES can do that. So this simplified
GHES is implemented firstly.

Signed-off-by: Huang Ying <ying.huang@intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Len Brown <len.brown@intel.com>
arch/x86/include/asm/mce.h
arch/x86/kernel/cpu/mcheck/Makefile
arch/x86/kernel/cpu/mcheck/mce-apei.c [new file with mode: 0644]
drivers/acpi/apei/Kconfig
drivers/acpi/apei/Makefile
drivers/acpi/apei/ghes.c [new file with mode: 0644]

index 6c3fdd631ed3df7652ed0ad261377006297b7f5f..f32a4301c4d4559157e1ac72f79a239025ccc1e0 100644 (file)
@@ -225,5 +225,13 @@ extern void mcheck_intel_therm_init(void);
 static inline void mcheck_intel_therm_init(void) { }
 #endif
 
+/*
+ * Used by APEI to report memory error via /dev/mcelog
+ */
+
+struct cper_sec_mem_err;
+extern void apei_mce_report_mem_error(int corrected,
+                                     struct cper_sec_mem_err *mem_err);
+
 #endif /* __KERNEL__ */
 #endif /* _ASM_X86_MCE_H */
index 4ac6d48fe11bc66eb2d8283e83438b0947e5ae92..bb34b03af25252ddb2298e80121f9955a6ae6192 100644 (file)
@@ -7,3 +7,5 @@ obj-$(CONFIG_X86_MCE_THRESHOLD) += threshold.o
 obj-$(CONFIG_X86_MCE_INJECT)   += mce-inject.o
 
 obj-$(CONFIG_X86_THERMAL_VECTOR) += therm_throt.o
+
+obj-$(CONFIG_ACPI_APEI)                += mce-apei.o
diff --git a/arch/x86/kernel/cpu/mcheck/mce-apei.c b/arch/x86/kernel/cpu/mcheck/mce-apei.c
new file mode 100644 (file)
index 0000000..4eccd1f
--- /dev/null
@@ -0,0 +1,52 @@
+/*
+ * Bridge between MCE and APEI
+ *
+ * On some machine, corrected memory errors are reported via APEI
+ * generic hardware error source (GHES) instead of corrected Machine
+ * Check. These corrected memory errors can be reported to user space
+ * through /dev/mcelog via faking a corrected Machine Check, so that
+ * the error memory page can be offlined by /sbin/mcelog if the error
+ * count for one page is beyond the threshold.
+ *
+ * Copyright 2010 Intel Corp.
+ *   Author: Huang Ying <ying.huang@intel.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License version
+ * 2 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ */
+
+#include <linux/kernel.h>
+#include <linux/acpi.h>
+#include <linux/cper.h>
+#include <acpi/apei.h>
+#include <asm/mce.h>
+
+#include "mce-internal.h"
+
+void apei_mce_report_mem_error(int corrected, struct cper_sec_mem_err *mem_err)
+{
+       struct mce m;
+
+       /* Only corrected MC is reported */
+       if (!corrected)
+               return;
+
+       mce_setup(&m);
+       m.bank = 1;
+       /* Fake a memory read corrected error with unknown channel */
+       m.status = MCI_STATUS_VAL | MCI_STATUS_EN | MCI_STATUS_ADDRV | 0x9f;
+       m.addr = mem_err->physical_addr;
+       mce_log(&m);
+       mce_notify_irq();
+}
+EXPORT_SYMBOL_GPL(apei_mce_report_mem_error);
index 5f0a41c2bc6216196e378d09e6e70b27797d03a8..f8c668f27b5a29eda7fbcc93357fbe5f1724fd1f 100644 (file)
@@ -7,6 +7,20 @@ config ACPI_APEI
          especially. In addition it supports error serialization and
          error injection.
 
+config ACPI_APEI_GHES
+       tristate "APEI Generic Hardware Error Source"
+       depends on ACPI_APEI && X86
+       select ACPI_HED
+       help
+         Generic Hardware Error Source provides a way to report
+         platform hardware errors (such as that from chipset). It
+         works in so called "Firmware First" mode, that is, hardware
+         errors are reported to firmware firstly, then reported to
+         Linux by firmware. This way, some non-standard hardware
+         error registers or non-standard hardware link can be checked
+         by firmware to produce more valuable hardware error
+         information for Linux.
+
 config ACPI_APEI_EINJ
        tristate "APEI Error INJection (EINJ)"
        depends on ACPI_APEI && DEBUG_FS
index fef963ec53629e60afb1dc0c33c4060bfce5a6e8..41c61db4c51cdadecc91e47ad2987689a4fd6866 100644 (file)
@@ -1,4 +1,5 @@
 obj-$(CONFIG_ACPI_APEI)                += apei.o
+obj-$(CONFIG_ACPI_APEI_GHES)   += ghes.o
 obj-$(CONFIG_ACPI_APEI_EINJ)   += einj.o
 
 apei-y := apei-base.o hest.o cper.o
diff --git a/drivers/acpi/apei/ghes.c b/drivers/acpi/apei/ghes.c
new file mode 100644 (file)
index 0000000..fd0cc01
--- /dev/null
@@ -0,0 +1,427 @@
+/*
+ * APEI Generic Hardware Error Source support
+ *
+ * Generic Hardware Error Source provides a way to report platform
+ * hardware errors (such as that from chipset). It works in so called
+ * "Firmware First" mode, that is, hardware errors are reported to
+ * firmware firstly, then reported to Linux by firmware. This way,
+ * some non-standard hardware error registers or non-standard hardware
+ * link can be checked by firmware to produce more hardware error
+ * information for Linux.
+ *
+ * For more information about Generic Hardware Error Source, please
+ * refer to ACPI Specification version 4.0, section 17.3.2.6
+ *
+ * Now, only SCI notification type and memory errors are
+ * supported. More notification type and hardware error type will be
+ * added later.
+ *
+ * Copyright 2010 Intel Corp.
+ *   Author: Huang Ying <ying.huang@intel.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License version
+ * 2 as published by the Free Software Foundation;
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ */
+
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/init.h>
+#include <linux/acpi.h>
+#include <linux/io.h>
+#include <linux/interrupt.h>
+#include <linux/cper.h>
+#include <linux/kdebug.h>
+#include <acpi/apei.h>
+#include <acpi/atomicio.h>
+#include <acpi/hed.h>
+#include <asm/mce.h>
+
+#include "apei-internal.h"
+
+#define GHES_PFX       "GHES: "
+
+#define GHES_ESTATUS_MAX_SIZE          65536
+
+/*
+ * One struct ghes is created for each generic hardware error
+ * source.
+ *
+ * It provides the context for APEI hardware error timer/IRQ/SCI/NMI
+ * handler. Handler for one generic hardware error source is only
+ * triggered after the previous one is done. So handler can uses
+ * struct ghes without locking.
+ *
+ * estatus: memory buffer for error status block, allocated during
+ * HEST parsing.
+ */
+#define GHES_TO_CLEAR          0x0001
+
+struct ghes {
+       struct acpi_hest_generic *generic;
+       struct acpi_hest_generic_status *estatus;
+       struct list_head list;
+       u64 buffer_paddr;
+       unsigned long flags;
+};
+
+/*
+ * Error source lists, one list for each notification method. The
+ * members in lists are struct ghes.
+ *
+ * The list members are only added in HEST parsing and deleted during
+ * module_exit, that is, single-threaded. So no lock is needed for
+ * that.
+ *
+ * But the mutual exclusion is needed between members adding/deleting
+ * and timer/IRQ/SCI/NMI handler, which may traverse the list. RCU is
+ * used for that.
+ */
+static LIST_HEAD(ghes_sci);
+
+static struct ghes *ghes_new(struct acpi_hest_generic *generic)
+{
+       struct ghes *ghes;
+       unsigned int error_block_length;
+       int rc;
+
+       ghes = kzalloc(sizeof(*ghes), GFP_KERNEL);
+       if (!ghes)
+               return ERR_PTR(-ENOMEM);
+       ghes->generic = generic;
+       INIT_LIST_HEAD(&ghes->list);
+       rc = acpi_pre_map_gar(&generic->error_status_address);
+       if (rc)
+               goto err_free;
+       error_block_length = generic->error_block_length;
+       if (error_block_length > GHES_ESTATUS_MAX_SIZE) {
+               pr_warning(FW_WARN GHES_PFX
+                          "Error status block length is too long: %u for "
+                          "generic hardware error source: %d.\n",
+                          error_block_length, generic->header.source_id);
+               error_block_length = GHES_ESTATUS_MAX_SIZE;
+       }
+       ghes->estatus = kmalloc(error_block_length, GFP_KERNEL);
+       if (!ghes->estatus) {
+               rc = -ENOMEM;
+               goto err_unmap;
+       }
+
+       return ghes;
+
+err_unmap:
+       acpi_post_unmap_gar(&generic->error_status_address);
+err_free:
+       kfree(ghes);
+       return ERR_PTR(rc);
+}
+
+static void ghes_fini(struct ghes *ghes)
+{
+       kfree(ghes->estatus);
+       acpi_post_unmap_gar(&ghes->generic->error_status_address);
+}
+
+enum {
+       GHES_SER_NO = 0x0,
+       GHES_SER_CORRECTED = 0x1,
+       GHES_SER_RECOVERABLE = 0x2,
+       GHES_SER_PANIC = 0x3,
+};
+
+static inline int ghes_severity(int severity)
+{
+       switch (severity) {
+       case CPER_SER_INFORMATIONAL:
+               return GHES_SER_NO;
+       case CPER_SER_CORRECTED:
+               return GHES_SER_CORRECTED;
+       case CPER_SER_RECOVERABLE:
+               return GHES_SER_RECOVERABLE;
+       case CPER_SER_FATAL:
+               return GHES_SER_PANIC;
+       default:
+               /* Unkown, go panic */
+               return GHES_SER_PANIC;
+       }
+}
+
+/* SCI handler run in work queue, so ioremap can be used here */
+static int ghes_copy_tofrom_phys(void *buffer, u64 paddr, u32 len,
+                                int from_phys)
+{
+       void *vaddr;
+
+       vaddr = ioremap_cache(paddr, len);
+       if (!vaddr)
+               return -ENOMEM;
+       if (from_phys)
+               memcpy(buffer, vaddr, len);
+       else
+               memcpy(vaddr, buffer, len);
+       iounmap(vaddr);
+
+       return 0;
+}
+
+static int ghes_read_estatus(struct ghes *ghes, int silent)
+{
+       struct acpi_hest_generic *g = ghes->generic;
+       u64 buf_paddr;
+       u32 len;
+       int rc;
+
+       rc = acpi_atomic_read(&buf_paddr, &g->error_status_address);
+       if (rc) {
+               if (!silent && printk_ratelimit())
+                       pr_warning(FW_WARN GHES_PFX
+"Failed to read error status block address for hardware error source: %d.\n",
+                                  g->header.source_id);
+               return -EIO;
+       }
+       if (!buf_paddr)
+               return -ENOENT;
+
+       rc = ghes_copy_tofrom_phys(ghes->estatus, buf_paddr,
+                                  sizeof(*ghes->estatus), 1);
+       if (rc)
+               return rc;
+       if (!ghes->estatus->block_status)
+               return -ENOENT;
+
+       ghes->buffer_paddr = buf_paddr;
+       ghes->flags |= GHES_TO_CLEAR;
+
+       rc = -EIO;
+       len = apei_estatus_len(ghes->estatus);
+       if (len < sizeof(*ghes->estatus))
+               goto err_read_block;
+       if (len > ghes->generic->error_block_length)
+               goto err_read_block;
+       if (apei_estatus_check_header(ghes->estatus))
+               goto err_read_block;
+       rc = ghes_copy_tofrom_phys(ghes->estatus + 1,
+                                  buf_paddr + sizeof(*ghes->estatus),
+                                  len - sizeof(*ghes->estatus), 1);
+       if (rc)
+               return rc;
+       if (apei_estatus_check(ghes->estatus))
+               goto err_read_block;
+       rc = 0;
+
+err_read_block:
+       if (rc && !silent)
+               pr_warning(FW_WARN GHES_PFX
+                          "Failed to read error status block!\n");
+       return rc;
+}
+
+static void ghes_clear_estatus(struct ghes *ghes)
+{
+       ghes->estatus->block_status = 0;
+       if (!(ghes->flags & GHES_TO_CLEAR))
+               return;
+       ghes_copy_tofrom_phys(ghes->estatus, ghes->buffer_paddr,
+                             sizeof(ghes->estatus->block_status), 0);
+       ghes->flags &= ~GHES_TO_CLEAR;
+}
+
+static void ghes_do_proc(struct ghes *ghes)
+{
+       int ser, processed = 0;
+       struct acpi_hest_generic_data *gdata;
+
+       ser = ghes_severity(ghes->estatus->error_severity);
+       apei_estatus_for_each_section(ghes->estatus, gdata) {
+#ifdef CONFIG_X86_MCE
+               if (!uuid_le_cmp(*(uuid_le *)gdata->section_type,
+                                CPER_SEC_PLATFORM_MEM)) {
+                       apei_mce_report_mem_error(
+                               ser == GHES_SER_CORRECTED,
+                               (struct cper_sec_mem_err *)(gdata+1));
+                       processed = 1;
+               }
+#endif
+       }
+
+       if (!processed && printk_ratelimit())
+               pr_warning(GHES_PFX
+               "Unknown error record from generic hardware error source: %d\n",
+                          ghes->generic->header.source_id);
+}
+
+static int ghes_proc(struct ghes *ghes)
+{
+       int rc;
+
+       rc = ghes_read_estatus(ghes, 0);
+       if (rc)
+               goto out;
+       ghes_do_proc(ghes);
+
+out:
+       ghes_clear_estatus(ghes);
+       return 0;
+}
+
+static int ghes_notify_sci(struct notifier_block *this,
+                                 unsigned long event, void *data)
+{
+       struct ghes *ghes;
+       int ret = NOTIFY_DONE;
+
+       rcu_read_lock();
+       list_for_each_entry_rcu(ghes, &ghes_sci, list) {
+               if (!ghes_proc(ghes))
+                       ret = NOTIFY_OK;
+       }
+       rcu_read_unlock();
+
+       return ret;
+}
+
+static struct notifier_block ghes_notifier_sci = {
+       .notifier_call = ghes_notify_sci,
+};
+
+static int hest_ghes_parse(struct acpi_hest_header *hest_hdr, void *data)
+{
+       struct acpi_hest_generic *generic;
+       struct ghes *ghes = NULL;
+       int rc = 0;
+
+       if (hest_hdr->type != ACPI_HEST_TYPE_GENERIC_ERROR)
+               return 0;
+
+       generic = (struct acpi_hest_generic *)hest_hdr;
+       if (!generic->enabled)
+               return 0;
+
+       if (generic->error_block_length <
+           sizeof(struct acpi_hest_generic_status)) {
+               pr_warning(FW_BUG GHES_PFX
+"Invalid error block length: %u for generic hardware error source: %d\n",
+                          generic->error_block_length,
+                          generic->header.source_id);
+               goto err;
+       }
+       if (generic->records_to_preallocate == 0) {
+               pr_warning(FW_BUG GHES_PFX
+"Invalid records to preallocate: %u for generic hardware error source: %d\n",
+                          generic->records_to_preallocate,
+                          generic->header.source_id);
+               goto err;
+       }
+       ghes = ghes_new(generic);
+       if (IS_ERR(ghes)) {
+               rc = PTR_ERR(ghes);
+               ghes = NULL;
+               goto err;
+       }
+       switch (generic->notify.type) {
+       case ACPI_HEST_NOTIFY_POLLED:
+               pr_warning(GHES_PFX
+"Generic hardware error source: %d notified via POLL is not supported!\n",
+                          generic->header.source_id);
+               break;
+       case ACPI_HEST_NOTIFY_EXTERNAL:
+       case ACPI_HEST_NOTIFY_LOCAL:
+               pr_warning(GHES_PFX
+"Generic hardware error source: %d notified via IRQ is not supported!\n",
+                          generic->header.source_id);
+               break;
+       case ACPI_HEST_NOTIFY_SCI:
+               if (list_empty(&ghes_sci))
+                       register_acpi_hed_notifier(&ghes_notifier_sci);
+               list_add_rcu(&ghes->list, &ghes_sci);
+               break;
+       case ACPI_HEST_NOTIFY_NMI:
+               pr_warning(GHES_PFX
+"Generic hardware error source: %d notified via NMI is not supported!\n",
+                          generic->header.source_id);
+               break;
+       default:
+               pr_warning(FW_WARN GHES_PFX
+       "Unknown notification type: %u for generic hardware error source: %d\n",
+                          generic->notify.type, generic->header.source_id);
+               break;
+       }
+
+       return 0;
+err:
+       if (ghes)
+               ghes_fini(ghes);
+       return rc;
+}
+
+static void ghes_cleanup(void)
+{
+       struct ghes *ghes, *nghes;
+
+       if (!list_empty(&ghes_sci))
+               unregister_acpi_hed_notifier(&ghes_notifier_sci);
+
+       synchronize_rcu();
+
+       list_for_each_entry_safe(ghes, nghes, &ghes_sci, list) {
+               list_del(&ghes->list);
+               ghes_fini(ghes);
+               kfree(ghes);
+       }
+}
+
+static int __init ghes_init(void)
+{
+       int rc;
+
+       if (acpi_disabled)
+               return -ENODEV;
+
+       if (hest_disable) {
+               pr_info(GHES_PFX "HEST is not enabled!\n");
+               return -EINVAL;
+       }
+
+       rc = apei_hest_parse(hest_ghes_parse, NULL);
+       if (rc) {
+               pr_err(GHES_PFX
+               "Error during parsing HEST generic hardware error sources.\n");
+               goto err_cleanup;
+       }
+
+       if (list_empty(&ghes_sci)) {
+               pr_info(GHES_PFX
+                       "No functional generic hardware error sources.\n");
+               rc = -ENODEV;
+               goto err_cleanup;
+       }
+
+       pr_info(GHES_PFX
+               "Generic Hardware Error Source support is initialized.\n");
+
+       return 0;
+err_cleanup:
+       ghes_cleanup();
+       return rc;
+}
+
+static void __exit ghes_exit(void)
+{
+       ghes_cleanup();
+}
+
+module_init(ghes_init);
+module_exit(ghes_exit);
+
+MODULE_AUTHOR("Huang Ying");
+MODULE_DESCRIPTION("APEI Generic Hardware Error Source support");
+MODULE_LICENSE("GPL");