[PATCH v3 3/3] nfit: do an ARS scrub on hitting a latent media error

[Vishal Verma]

When a latent (unknown to 'badblocks') error is encountered, it will
trigger a machine check exception. On a system with machine check
recovery, this will only SIGBUS the process(es) which had the bad page
mapped (as opposed to a kernel panic on platforms without machine
check recovery features). In the former case, we want to trigger a full
rescan of that nvdimm bus. This will allow any additional, new errors
to be captured in the block devices' badblocks lists, and offending
operations on them can be trapped early, avoiding machine checks.

This is done by registering a callback function with the
x86_mce_decoder_chain and calling the new ars_rescan functionality with
the address in the mce notificatiion.

Cc: Dan Williams <dan.j.williams@xxxxxxxxx>
Cc: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
Cc: Tony Luck <tony.luck@xxxxxxxxx>
Cc: <linux-acpi@xxxxxxxxxxxxxxx>
Cc: <linux-nvdimm@xxxxxxxxxxxx>
Signed-off-by: Vishal Verma <vishal.l.verma@xxxxxxxxx>
---
 drivers/acpi/nfit.c | 87 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 drivers/acpi/nfit.h |  1 +
 2 files changed, 88 insertions(+)

diff --git a/drivers/acpi/nfit.c b/drivers/acpi/nfit.c
index 6e45183..954b610 100644
--- a/drivers/acpi/nfit.c
+++ b/drivers/acpi/nfit.c
@@ -12,6 +12,7 @@
  */
 #include <linux/list_sort.h>
 #include <linux/libnvdimm.h>
+#include <linux/notifier.h>
 #include <linux/module.h>
 #include <linux/mutex.h>
 #include <linux/ndctl.h>
@@ -24,6 +25,7 @@
 #include <linux/io.h>
 #include <linux/nd.h>
 #include <asm/cacheflush.h>
+#include <asm/mce.h>
 #include "nfit.h"
 
 /*
@@ -51,6 +53,9 @@ module_param(disable_vendor_specific, bool, S_IRUGO);
 MODULE_PARM_DESC(disable_vendor_specific,
 		"Limit commands to the publicly specified set\n");
 
+static LIST_HEAD(acpi_descs);
+static DEFINE_MUTEX(acpi_desc_lock);
+
 static struct workqueue_struct *nfit_wq;
 
 struct nfit_table_prev {
@@ -2395,13 +2400,18 @@ static void acpi_nfit_destruct(void *data)
 		sysfs_put(acpi_desc->scrub_count_state);
 	nvdimm_bus_unregister(acpi_desc->nvdimm_bus);
 	acpi_desc->nvdimm_bus = NULL;
+	mutex_lock(&acpi_desc_lock);
+	list_del(&acpi_desc->list);
+	mutex_unlock(&acpi_desc_lock);
 }
 
 int acpi_nfit_init(struct acpi_nfit_desc *acpi_desc, void *data, acpi_size sz)
 {
+	struct acpi_nfit_desc *acpi_desc_entry;
 	struct device *dev = acpi_desc->dev;
 	struct nfit_table_prev prev;
 	const void *end;
+	int found = 0;
 	int rc;
 
 	acpi_nfit_init_dsms(acpi_desc);
@@ -2468,6 +2478,19 @@ int acpi_nfit_init(struct acpi_nfit_desc *acpi_desc, void *data, acpi_size sz)
 
 	rc = acpi_nfit_register_regions(acpi_desc);
 
+	/*
+	 * We may get here due to an update of the nfit via _FIT.
+	 * Check if the acpi_desc we're (re)initializing is already
+	 * present in the list, and if so, don't re-add it
+	 */
+	mutex_lock(&acpi_desc_lock);
+	list_for_each_entry(acpi_desc_entry, &acpi_descs, list)
+		if (acpi_desc_entry == acpi_desc)
+			found = 1;
+	if (found == 0)
+		list_add_tail(&acpi_desc->list, &acpi_descs);
+	mutex_unlock(&acpi_desc_lock);
+
  out_unlock:
 	mutex_unlock(&acpi_desc->init_mutex);
 	return rc;
@@ -2550,6 +2573,65 @@ static int acpi_nfit_ars_rescan(struct acpi_nfit_desc *acpi_desc)
 	return 0;
 }
 
+static int nfit_handle_mce(struct notifier_block *nb, unsigned long val,
+			void *data)
+{
+	struct mce *mce = (struct mce *)data;
+	struct acpi_nfit_desc *acpi_desc;
+	struct nfit_spa *nfit_spa;
+
+	/* We only care about memory errors */
+	if (!(mce->status & MCACOD))
+		return NOTIFY_DONE;
+
+	/*
+	 * mce->addr contains the physical addr accessed that caused the
+	 * machine check. We need to walk through the list of NFITs, and see
+	 * if any of them matches that address, and only then start a scrub.
+	 */
+	mutex_lock(&acpi_desc_lock);
+	list_for_each_entry(acpi_desc, &acpi_descs, list) {
+		struct device *dev = acpi_desc->dev;
+		int found_match = 0;
+
+		list_for_each_entry(nfit_spa, &acpi_desc->spas, list) {
+			struct acpi_nfit_system_address *spa = nfit_spa->spa;
+
+			if (nfit_spa_type(spa) != NFIT_SPA_PM)
+				continue;
+			/* find the spa that covers the mce addr */
+			if (spa->address > mce->addr)
+				continue;
+			if ((spa->address + spa->length - 1) < mce->addr)
+				continue;
+			found_match = 1;
+			dev_dbg(dev, "%s: addr in SPA %d (0x%llx, 0x%llx)\n",
+				__func__, spa->range_index, spa->address,
+				spa->length);
+			/*
+			 * We can break at the first match because we're going
+			 * to rescan all the SPA ranges. There shouldn't be any
+			 * aliasing anyway.
+			 */
+			break;
+		}
+
+		/*
+		 * We can ignore an -EBUSY here because if an ARS is already
+		 * in progress, just let that be the last authoritative one
+		 */
+		if (found_match)
+			acpi_nfit_ars_rescan(acpi_desc);
+	}
+
+	mutex_unlock(&acpi_desc_lock);
+	return NOTIFY_DONE;
+}
+
+static struct notifier_block nfit_mce_dec = {
+	.notifier_call	= nfit_handle_mce,
+};
+
 void acpi_nfit_desc_init(struct acpi_nfit_desc *acpi_desc, struct device *dev)
 {
 	struct nvdimm_bus_descriptor *nd_desc;
@@ -2724,13 +2806,18 @@ static __init int nfit_init(void)
 	if (!nfit_wq)
 		return -ENOMEM;
 
+	INIT_LIST_HEAD(&acpi_descs);
+	mce_register_decode_chain(&nfit_mce_dec);
+
 	return acpi_bus_register_driver(&acpi_nfit_driver);
 }
 
 static __exit void nfit_exit(void)
 {
+	mce_unregister_decode_chain(&nfit_mce_dec);
 	acpi_bus_unregister_driver(&acpi_nfit_driver);
 	destroy_workqueue(nfit_wq);
+	WARN_ON(!list_empty(&acpi_descs));
 }
 
 module_init(nfit_init);
diff --git a/drivers/acpi/nfit.h b/drivers/acpi/nfit.h
index 33fc2e9..a2d6c6b 100644
--- a/drivers/acpi/nfit.h
+++ b/drivers/acpi/nfit.h
@@ -148,6 +148,7 @@ struct acpi_nfit_desc {
 	struct nd_cmd_ars_status *ars_status;
 	size_t ars_status_size;
 	struct work_struct work;
+	struct list_head list;
 	struct kernfs_node *scrub_count_state;
 	unsigned int scrub_count;
 	unsigned int cancel:1;
-- 
2.7.4