Date:	Tue,  7 Jun 2016 18:52:22 +0200
From:	Borislav Petkov <bp@...en8.de>
To:	linux-edac <linux-edac@...r.kernel.org>
Cc:	Tony Luck <tony.luck@...el.com>,
	Yazen Ghannam <Yazen.Ghannam@....com>, X86 ML <x86@...nel.org>,
	LKML <linux-kernel@...r.kernel.org>
Subject: [RFC PATCH 1/4] RAS: Add a Corrected Errors Collector

From: Borislav Petkov <bp@...e.de>

A simple data structure for collecting correctable errors along with
accessors. More detailed description in the code itself.

Signed-off-by: Borislav Petkov <bp@...e.de>
---
 arch/x86/kernel/cpu/mcheck/mce.c | 120 +++++------
 arch/x86/ras/Kconfig             |  11 +
 drivers/ras/Makefile             |   3 +-
 drivers/ras/cec.c                | 425 +++++++++++++++++++++++++++++++++++++++
 drivers/ras/debugfs.c            |   2 +-
 drivers/ras/debugfs.h            |   8 +
 include/linux/ras.h              |  12 +-
 7 files changed, 521 insertions(+), 60 deletions(-)
 create mode 100644 drivers/ras/cec.c
 create mode 100644 drivers/ras/debugfs.h

diff --git a/arch/x86/kernel/cpu/mcheck/mce.c b/arch/x86/kernel/cpu/mcheck/mce.c
index 92e5e37d97bf..a6cc08a10cc3 100644
--- a/arch/x86/kernel/cpu/mcheck/mce.c
+++ b/arch/x86/kernel/cpu/mcheck/mce.c
@@ -35,6 +35,7 @@
 #include <linux/poll.h>
 #include <linux/nmi.h>
 #include <linux/cpu.h>
+#include <linux/ras.h>
 #include <linux/smp.h>
 #include <linux/fs.h>
 #include <linux/mm.h>
@@ -151,14 +152,71 @@ static struct mce_log mcelog = {
 	.recordlen	= sizeof(struct mce),
 };
 
-void mce_log(struct mce *mce)
+static bool memory_error(struct mce *m)
+{
+	struct cpuinfo_x86 *c = &boot_cpu_data;
+
+	if (c->x86_vendor == X86_VENDOR_AMD) {
+		/* ErrCodeExt[20:16] */
+		u8 xec = (m->status >> 16) & 0x1f;
+
+		return (xec == 0x0 || xec == 0x8);
+	} else if (c->x86_vendor == X86_VENDOR_INTEL) {
+		/*
+		 * Intel SDM Volume 3B - 15.9.2 Compound Error Codes
+		 *
+		 * Bit 7 of the MCACOD field of IA32_MCi_STATUS is used for
+		 * indicating a memory error. Bit 8 is used for indicating a
+		 * cache hierarchy error. The combination of bit 2 and bit 3
+		 * is used for indicating a `generic' cache hierarchy error
+		 * But we can't just blindly check the above bits, because if
+		 * bit 11 is set, then it is a bus/interconnect error - and
+		 * either way the above bits just gives more detail on what
+		 * bus/interconnect error happened. Note that bit 12 can be
+		 * ignored, as it's the "filter" bit.
+		 */
+		return (m->status & 0xef80) == BIT(7) ||
+		       (m->status & 0xef00) == BIT(8) ||
+		       (m->status & 0xeffc) == 0xc;
+	}
+
+	return false;
+}
+
+/*
+ * Check if the address reported by the CPU is in a format we can parse.
+ * It would be possible to add code for most other cases, but all would
+ * be somewhat complicated (e.g. segment offset would require an instruction
+ * parser). So only support physical addresses up to page granuality for now.
+ */
+static int mce_usable_address(struct mce *m)
+{
+	if (!(m->status & MCI_STATUS_MISCV) || !(m->status & MCI_STATUS_ADDRV))
+		return 0;
+
+	/* Checks after this one are Intel-specific: */
+	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
+		return 1;
+
+	if (MCI_MISC_ADDR_LSB(m->misc) > PAGE_SHIFT)
+		return 0;
+	if (MCI_MISC_ADDR_MODE(m->misc) != MCI_MISC_ADDR_PHYS)
+		return 0;
+	return 1;
+}
+
+void mce_log(struct mce *m)
 {
 	unsigned next, entry;
 
+	if (!in_atomic() && memory_error(m) && mce_usable_address(m))
+		if (!ce_add_elem(m->addr >> PAGE_SHIFT))
+			return;
+
 	/* Emit the trace record: */
-	trace_mce_record(mce);
+	trace_mce_record(m);
 
-	if (!mce_gen_pool_add(mce))
+	if (!mce_gen_pool_add(m))
 		irq_work_queue(&mce_irq_work);
 
 	wmb();
@@ -188,7 +246,7 @@ void mce_log(struct mce *mce)
 		if (cmpxchg(&mcelog.next, entry, next) == entry)
 			break;
 	}
-	memcpy(mcelog.entry + entry, mce, sizeof(struct mce));
+	memcpy(mcelog.entry + entry, m, sizeof(struct mce));
 	wmb();
 	mcelog.entry[entry].finished = 1;
 	wmb();
@@ -518,28 +576,6 @@ static void mce_report_event(struct pt_regs *regs)
 	irq_work_queue(&mce_irq_work);
 }
 
-/*
- * Check if the address reported by the CPU is in a format we can parse.
- * It would be possible to add code for most other cases, but all would
- * be somewhat complicated (e.g. segment offset would require an instruction
- * parser). So only support physical addresses up to page granuality for now.
- */
-static int mce_usable_address(struct mce *m)
-{
-	if (!(m->status & MCI_STATUS_MISCV) || !(m->status & MCI_STATUS_ADDRV))
-		return 0;
-
-	/* Checks after this one are Intel-specific: */
-	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
-		return 1;
-
-	if (MCI_MISC_ADDR_LSB(m->misc) > PAGE_SHIFT)
-		return 0;
-	if (MCI_MISC_ADDR_MODE(m->misc) != MCI_MISC_ADDR_PHYS)
-		return 0;
-	return 1;
-}
-
 static int srao_decode_notifier(struct notifier_block *nb, unsigned long val,
 				void *data)
 {
@@ -582,37 +618,6 @@ static void mce_read_aux(struct mce *m, int i)
 	}
 }
 
-static bool memory_error(struct mce *m)
-{
-	struct cpuinfo_x86 *c = &boot_cpu_data;
-
-	if (c->x86_vendor == X86_VENDOR_AMD) {
-		/* ErrCodeExt[20:16] */
-		u8 xec = (m->status >> 16) & 0x1f;
-
-		return (xec == 0x0 || xec == 0x8);
-	} else if (c->x86_vendor == X86_VENDOR_INTEL) {
-		/*
-		 * Intel SDM Volume 3B - 15.9.2 Compound Error Codes
-		 *
-		 * Bit 7 of the MCACOD field of IA32_MCi_STATUS is used for
-		 * indicating a memory error. Bit 8 is used for indicating a
-		 * cache hierarchy error. The combination of bit 2 and bit 3
-		 * is used for indicating a `generic' cache hierarchy error
-		 * But we can't just blindly check the above bits, because if
-		 * bit 11 is set, then it is a bus/interconnect error - and
-		 * either way the above bits just gives more detail on what
-		 * bus/interconnect error happened. Note that bit 12 can be
-		 * ignored, as it's the "filter" bit.
-		 */
-		return (m->status & 0xef80) == BIT(7) ||
-		       (m->status & 0xef00) == BIT(8) ||
-		       (m->status & 0xeffc) == 0xc;
-	}
-
-	return false;
-}
-
 DEFINE_PER_CPU(unsigned, mce_poll_count);
 
 /*
@@ -2679,6 +2684,7 @@ static int __init mcheck_debugfs_init(void) { return -EINVAL; }
 static int __init mcheck_late_init(void)
 {
 	mcheck_debugfs_init();
+	ce_init();
 
 	/*
 	 * Flush out everything that has been logged during early boot, now that
diff --git a/arch/x86/ras/Kconfig b/arch/x86/ras/Kconfig
index d957d5f21a86..513e8ffc9e08 100644
--- a/arch/x86/ras/Kconfig
+++ b/arch/x86/ras/Kconfig
@@ -7,3 +7,14 @@ config MCE_AMD_INJ
 	  aspects of the MCE handling code.
 
 	  WARNING: Do not even assume this interface is staying stable!
+
+config RAS_CEC
+	bool "Correctable Errors Collector"
+	depends on X86_MCE && MEMORY_FAILURE && DEBUG_FS
+	---help---
+	  This is a small cache which collects correctable memory errors per 4K
+	  page PFN and counts their repeated occurrence. Once the counter for a
+	  PFN overflows, we try to soft-offline that page as we take it to mean
+	  that it has reached a relatively high error count and would probably
+	  be best if we don't use it anymore.
+
diff --git a/drivers/ras/Makefile b/drivers/ras/Makefile
index d7f73341ced3..7b26dd3aa5d0 100644
--- a/drivers/ras/Makefile
+++ b/drivers/ras/Makefile
@@ -1 +1,2 @@
-obj-$(CONFIG_RAS) += ras.o debugfs.o
+obj-$(CONFIG_RAS)	+= ras.o debugfs.o
+obj-$(CONFIG_RAS_CEC)	+= cec.o
diff --git a/drivers/ras/cec.c b/drivers/ras/cec.c
new file mode 100644
index 000000000000..846155124598
--- /dev/null
+++ b/drivers/ras/cec.c
@@ -0,0 +1,425 @@
+#include <linux/mm.h>
+#include <linux/gfp.h>
+#include <linux/kernel.h>
+
+#include <asm/bug.h>
+
+#include "debugfs.h"
+
+/*
+ * RAS Correctable Errors Collector
+ *
+ * This is a simple gadget which collects correctable errors and counts their
+ * occurrence per physical page address.
+ *
+ * We've opted for possibly the simplest data structure to collect those - an
+ * array of the size of a memory page. It stores 512 u64's with the following
+ * structure:
+ *
+ * [63 ... PFN ... 12 | 11 ... generation ... 10 | 9 ... count ... 0]
+ *
+ * The generation in the two highest order bits is two bits which are set to 11b
+ * on every insertion. During the course of each entry's existence, the
+ * generation field gets decremented during spring cleaning to 10b, then 01b and
+ * then 00b.
+ *
+ * This way we're employing the natural numeric ordering to make sure that newly
+ * inserted/touched elements have higher 12-bit counts (which we've manufactured)
+ * and thus iterating over the array initially won't kick out those elements
+ * which were inserted last.
+ *
+ * Spring cleaning is what we do when we reach a certain number CLEAN_ELEMS of
+ * elements entered into the array, during which, we're decaying all elements.
+ * If, after decay, an element gets inserted again, its generation is set to 11b
+ * to make sure it has higher numerical count than other, older elements and
+ * thus emulate an an LRU-like behavior when deleting elements to free up space
+ * in the page.
+ *
+ * When an element reaches it's max count of COUNT_MASK, we try to poison it by
+ * assuming that errors triggered COUNT_MASK times in a single page are
+ * excessive and that page shouldn't be used anymore.
+ *
+ * To the question why we've chosen a page and moving elements around with
+ * memmove(), it is because it is a very simple structure to handle and max data
+ * movement is 4K which on highly optimized modern CPUs is almost unnoticeable.
+ * We wanted to avoid the pointer traversal of more complex structures like a
+ * linked list or some sort of a balancing search tree.
+ *
+ * Deleting an element takes O(n) but since it is only a single page, it should
+ * be fast enough and it shouldn't happen all too often depending on error
+ * patterns.
+ */
+
+#undef pr_fmt
+#define pr_fmt(fmt) "RAS: " fmt
+
+/*
+ * We use DECAY_BITS bits of PAGE_SHIFT bits for counting decay, i.e., how long
+ * elements have stayed in the array without having been accessed again.
+ */
+#define DECAY_BITS		2
+#define DECAY_MASK		((1ULL << DECAY_BITS) - 1)
+#define MAX_ELEMS		(PAGE_SIZE / sizeof(u64))
+
+/*
+ * Threshold amount of inserted elements after which we start spring
+ * cleaning.
+ */
+#define CLEAN_ELEMS		(MAX_ELEMS >> DECAY_BITS)
+
+/* Bits which count the number of errors happened in this 4K page. */
+#define COUNT_BITS		(PAGE_SHIFT - DECAY_BITS)
+#define COUNT_MASK		((1ULL << COUNT_BITS) - 1)
+#define FULL_COUNT_MASK		(PAGE_SIZE - 1)
+
+/*
+ * u64: [ 63 ... 12 | DECAY_BITS | COUNT_BITS ]
+ */
+
+#define PFN(e)			((e) >> PAGE_SHIFT)
+#define DECAY(e)		(((e) >> COUNT_BITS) & DECAY_MASK)
+#define COUNT(e)		((unsigned int)(e) & COUNT_MASK)
+#define FULL_COUNT(e)		((e) & (PAGE_SIZE - 1))
+
+static struct ce_array {
+	u64 *array;			/* container page */
+	unsigned n;			/* number of elements in the array */
+
+	unsigned int decay_count;	/*
+					 * number of element insertions/increments
+					 * since the last spring cleaning.
+					 */
+
+	u64 pfns_poisoned;		/*
+					 * number of PFNs which got poisoned.
+					 */
+
+	u64 ces_entered;		/*
+					 * The number of correctable errors
+					 * entered into the collector.
+					 */
+	union {
+		struct {
+			__u32	disabled : 1,	/* cmdline disabled */
+			__resv   : 31;
+		};
+		__u32 flags;
+	};
+} ce_arr;
+
+static DEFINE_MUTEX(ce_mutex);
+
+static u64 dfs_pfn;
+
+/*
+ * Decrement decay value. We're using DECAY_BITS bits to denote decay of an
+ * element in the array. On insertion and any access, it gets reset to max.
+ */
+static void do_spring_cleaning(struct ce_array *ca)
+{
+	int i;
+
+	for (i = 0; i < ca->n; i++) {
+		u8 decay = DECAY(ca->array[i]);
+
+		if (!decay)
+			continue;
+
+		decay--;
+
+		ca->array[i] &= ~(DECAY_MASK << COUNT_BITS);
+		ca->array[i] |= (decay << COUNT_BITS);
+	}
+	ca->decay_count = 0;
+}
+
+/*
+ * @to: index of the smallest element which is >= then @pfn.
+ *
+ * Return the index of the pfn if found, otherwise negative value.
+ */
+static int __find_elem(struct ce_array *ca, u64 pfn, unsigned *to)
+{
+	u64 this_pfn;
+	int min = 0, max = ca->n;
+
+	while (min < max) {
+		int tmp = (max + min) >> 1;
+
+		this_pfn = PFN(ca->array[tmp]);
+
+		if (this_pfn < pfn)
+			min = tmp + 1;
+		else if (this_pfn > pfn)
+			max = tmp;
+		else {
+			min = tmp;
+			break;
+		}
+	}
+
+	if (to)
+		*to = min;
+
+	this_pfn = PFN(ca->array[min]);
+
+	if (this_pfn == pfn)
+		return min;
+
+	return -ENOKEY;
+}
+
+static int find_elem(struct ce_array *ca, u64 pfn, unsigned *to)
+{
+	WARN_ON(!to);
+
+	if (!ca->n) {
+		*to = 0;
+		return -ENOKEY;
+	}
+	return __find_elem(ca, pfn, to);
+}
+
+static void del_elem(struct ce_array *ca, int idx)
+{
+	/* Save us a function call when deleting the last element. */
+	if (ca->n - (idx + 1))
+		memmove((void *)&ca->array[idx],
+			(void *)&ca->array[idx + 1],
+			(ca->n - (idx + 1)) * sizeof(u64));
+
+	ca->n--;
+}
+
+static u64 del_lru_elem_unlocked(struct ce_array *ca)
+{
+	unsigned int min = FULL_COUNT_MASK;
+	int i, min_idx = 0;
+
+	for (i = 0; i < ca->n; i++) {
+		unsigned int this = FULL_COUNT(ca->array[i]);
+		if (min > this) {
+			min = this;
+			min_idx = i;
+		}
+	}
+
+	del_elem(ca, min_idx);
+
+	return PFN(ca->array[min_idx]);
+}
+
+/*
+ * We return the 0th pfn in the error case under the assumption that it cannot
+ * be poisoned and excessive CEs in there are a serious deal anyway.
+ */
+static u64 __maybe_unused del_lru_elem(void)
+{
+	struct ce_array *ca = &ce_arr;
+	u64 pfn;
+
+	if (!ca->n)
+		return 0;
+
+	mutex_lock(&ce_mutex);
+	pfn = del_lru_elem_unlocked(ca);
+	mutex_unlock(&ce_mutex);
+
+	return pfn;
+}
+
+
+int ce_add_elem(u64 pfn)
+{
+	struct ce_array *ca = &ce_arr;
+	unsigned to;
+	int count, ret;
+
+	/*
+	 * We can be called very early on the identify_cpu() path where we are
+	 * not initialized yet. We ignore the error for simplicity.
+	 */
+	if (!ce_arr.array)
+		return -ENODEV;
+
+	ca->ces_entered++;
+
+	mutex_lock(&ce_mutex);
+
+	if (ca->n == MAX_ELEMS)
+		WARN_ON(!del_lru_elem_unlocked(ca));
+
+	ret = find_elem(ca, pfn, &to);
+	if (ret < 0) {
+		/*
+		 * Shift range [to-end] to make room for one more element.
+		 */
+		memmove((void *)&ca->array[to + 1],
+			(void *)&ca->array[to],
+			(ca->n - to) * sizeof(u64));
+
+		ca->array[to] = (pfn << PAGE_SHIFT) |
+				(DECAY_MASK << COUNT_BITS) | 1;
+
+		ca->n++;
+
+		goto decay;
+	}
+
+	count = COUNT(ca->array[to]);
+
+	if (count < COUNT_MASK) {
+		ca->array[to] |= (DECAY_MASK << COUNT_BITS);
+		ca->array[to]++;
+	} else {
+		u64 pfn = ca->array[to] >> PAGE_SHIFT;
+
+		/*
+		 * We have reached max count for this page, soft-offline it.
+		 */
+		pr_err("Soft-offlining pfn: 0x%llx\n", pfn);
+		memory_failure_queue(pfn, 0, MF_SOFT_OFFLINE);
+		ca->pfns_poisoned++;
+		del_elem(ca, to);
+
+		ret = 0;
+
+		goto unlock;
+	}
+
+decay:
+	ca->decay_count++;
+
+	if (ca->decay_count >= CLEAN_ELEMS)
+		do_spring_cleaning(ca);
+
+unlock:
+	mutex_unlock(&ce_mutex);
+
+	return 0;
+}
+
+static int pfn_get(void *data, u64 *val)
+{
+	*val = *(u64 *)data;
+
+	return 0;
+}
+
+static int pfn_set(void *data, u64 val)
+{
+	*(u64 *)data = val;
+
+	return ce_add_elem(val);
+}
+
+DEFINE_SIMPLE_ATTRIBUTE(pfn_ops, pfn_get, pfn_set, "0x%llx\n");
+
+static int array_dump(struct seq_file *m, void *v)
+{
+	struct ce_array *ca = &ce_arr;
+	u64 prev = 0;
+	int i;
+
+	mutex_lock(&ce_mutex);
+
+	seq_printf(m, "{ n: %d\n", ca->n);
+	for (i = 0; i < ca->n; i++) {
+		u64 this = PFN(ca->array[i]);
+
+		seq_printf(m, " %03d: [%016llu|%03llx]\n", i, this, FULL_COUNT(ca->array[i]));
+
+		WARN_ON(prev > this);
+
+		prev = this;
+	}
+
+	seq_printf(m, "}\n");
+
+	seq_printf(m, "Stats:\n CEs: %llu\n offlined pages: %llu\n",
+		   ca->ces_entered, ca->pfns_poisoned);
+
+	seq_printf(m, "Flags: 0x%x\n", ca->flags);
+
+	mutex_unlock(&ce_mutex);
+
+	return 0;
+}
+
+static int array_open(struct inode *inode, struct file *filp)
+{
+	return single_open(filp, array_dump, NULL);
+}
+
+static const struct file_operations array_ops = {
+	.owner	 = THIS_MODULE,
+	.open	 = array_open,
+	.read	 = seq_read,
+	.llseek	 = seq_lseek,
+	.release = single_release,
+};
+
+static int __init create_debugfs_nodes(void)
+{
+	struct dentry *d, *pfn, *array;
+
+	d = debugfs_create_dir("cec", ras_debugfs_dir);
+	if (!d) {
+		pr_warn("Error creating cec debugfs node!\n");
+		return -1;
+	}
+
+	pfn = debugfs_create_file("pfn", S_IRUSR | S_IWUSR, d, &dfs_pfn, &pfn_ops);
+	if (!pfn) {
+		pr_warn("Error creating pfn debugfs node!\n");
+		goto err;
+	}
+
+	array = debugfs_create_file("array", S_IRUSR, d, NULL, &array_ops);
+	if (!array) {
+		pr_warn("Error creating array debugfs node!\n");
+		goto err;
+	}
+
+	return 0;
+
+err:
+	debugfs_remove_recursive(d);
+
+	return 1;
+}
+
+void __init ce_init(void)
+{
+	if (ce_arr.disabled)
+		return;
+
+	ce_arr.array = (void *)get_zeroed_page(GFP_KERNEL);
+	if (!ce_arr.array) {
+		pr_err("Error allocating CE array page!\n");
+		return;
+	}
+
+	if (create_debugfs_nodes())
+		return;
+
+	pr_info("Correctable Errors collector initialized.\n");
+}
+
+static int __init cec_param(char *str)
+{
+	if (!str)
+		return 0;
+
+	if (*str == '=')
+		str++;
+
+	if (!strncmp(str, "disable", 7))
+		ce_arr.disabled = 1;
+	else
+		return 0;
+
+	return 1;
+}
+
+__setup("cec", cec_param);
diff --git a/drivers/ras/debugfs.c b/drivers/ras/debugfs.c
index 0322acf67ea5..501603057dff 100644
--- a/drivers/ras/debugfs.c
+++ b/drivers/ras/debugfs.c
@@ -1,6 +1,6 @@
 #include <linux/debugfs.h>
 
-static struct dentry *ras_debugfs_dir;
+struct dentry *ras_debugfs_dir;
 
 static atomic_t trace_count = ATOMIC_INIT(0);
 
diff --git a/drivers/ras/debugfs.h b/drivers/ras/debugfs.h
new file mode 100644
index 000000000000..db72e4513191
--- /dev/null
+++ b/drivers/ras/debugfs.h
@@ -0,0 +1,8 @@
+#ifndef __RAS_DEBUGFS_H__
+#define __RAS_DEBUGFS_H__
+
+#include <linux/debugfs.h>
+
+extern struct dentry *ras_debugfs_dir;
+
+#endif /* __RAS_DEBUGFS_H__ */
diff --git a/include/linux/ras.h b/include/linux/ras.h
index 2aceeafd6fe5..a96762a10100 100644
--- a/include/linux/ras.h
+++ b/include/linux/ras.h
@@ -1,14 +1,24 @@
 #ifndef __RAS_H__
 #define __RAS_H__
 
+#include <asm/errno.h>
+
 #ifdef CONFIG_DEBUG_FS
 int ras_userspace_consumers(void);
 void ras_debugfs_init(void);
 int ras_add_daemon_trace(void);
 #else
 static inline int ras_userspace_consumers(void) { return 0; }
-static inline void ras_debugfs_init(void) { return; }
+static inline void ras_debugfs_init(void) { }
 static inline int ras_add_daemon_trace(void) { return 0; }
 #endif
 
+#ifdef CONFIG_RAS_CEC
+void __init ce_init(void);
+int ce_add_elem(u64 pfn);
+#else
+static inline void __init ce_init(void)		{ }
+static inline int ce_add_elem(u64 pfn)		{ return -ENODEV; }
 #endif
+
+#endif /* __RAS_H__ */
-- 
2.7.3

Powered by blists - more mailing lists