Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/bp/bp

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/bp/bp: (26 commits)
  amd64_edac: add MAINTAINERS entry
  EDAC: do not enable modules by default
  amd64_edac: do not enable module by default
  amd64_edac: add module registration routines
  amd64_edac: add ECC reporting initializers
  amd64_edac: add EDAC core-related initializers
  amd64_edac: add error decoding logic
  amd64_edac: add ECC chipkill syndrome mapping table
  amd64_edac: add per-family descriptors
  amd64_edac: add F10h-and-later methods-p3
  amd64_edac: add F10h-and-later methods-p2
  amd64_edac: add F10h-and-later methods-p1
  amd64_edac: add k8-specific methods
  amd64_edac: assign DRAM chip select base and mask in a family-specific way
  amd64_edac: add helper to dump relevant registers
  amd64_edac: add DRAM address type conversion facilities
  amd64_edac: add functionality to compute the DRAM hole
  amd64_edac: add sys addr to memory controller mapping helpers
  amd64_edac: add memory scrubber interface
  amd64_edac: add MCA error types
  ...

13 files changed, 4853 insertions, 103 deletions

diff --git a/MAINTAINERS b/MAINTAINERS
index 84285b5ba35..ccdb57524e3 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -1979,6 +1979,16 @@ F:	Documentation/edac.txt
 F:	drivers/edac/edac_*
 F:	include/linux/edac.h
 
+EDAC-AMD64
+P:	Doug Thompson
+M:	dougthompson@xmission.com
+P:	Borislav Petkov
+M:	borislav.petkov@amd.com
+L:	bluesmoke-devel@lists.sourceforge.net (moderated for non-subscribers)
+W:	bluesmoke.sourceforge.net
+S:	Supported
+F:	drivers/edac/amd64_edac*
+
 EDAC-E752X
 P:	Mark Gross
 M:	mark.gross@intel.com
diff --git a/arch/x86/include/asm/msr.h b/arch/x86/include/asm/msr.h
index 638bf624180..22603764e7d 100644
--- a/arch/x86/include/asm/msr.h
+++ b/arch/x86/include/asm/msr.h
@@ -12,6 +12,17 @@
 
 #include <asm/asm.h>
 #include <asm/errno.h>
+#include <asm/cpumask.h>
+
+struct msr {
+	union {
+		struct {
+			u32 l;
+			u32 h;
+		};
+		u64 q;
+	};
+};
 
 static inline unsigned long long native_read_tscp(unsigned int *aux)
 {
@@ -216,6 +227,8 @@ do {                                                            \
 #ifdef CONFIG_SMP
 int rdmsr_on_cpu(unsigned int cpu, u32 msr_no, u32 *l, u32 *h);
 int wrmsr_on_cpu(unsigned int cpu, u32 msr_no, u32 l, u32 h);
+void rdmsr_on_cpus(const cpumask_t *mask, u32 msr_no, struct msr *msrs);
+void wrmsr_on_cpus(const cpumask_t *mask, u32 msr_no, struct msr *msrs);
 int rdmsr_safe_on_cpu(unsigned int cpu, u32 msr_no, u32 *l, u32 *h);
 int wrmsr_safe_on_cpu(unsigned int cpu, u32 msr_no, u32 l, u32 h);
 #else  /*  CONFIG_SMP  */
@@ -229,6 +242,16 @@ static inline int wrmsr_on_cpu(unsigned int cpu, u32 msr_no, u32 l, u32 h)
 	wrmsr(msr_no, l, h);
 	return 0;
 }
+static inline void rdmsr_on_cpus(const cpumask_t *m, u32 msr_no,
+				struct msr *msrs)
+{
+       rdmsr_on_cpu(0, msr_no, &(msrs[0].l), &(msrs[0].h));
+}
+static inline void wrmsr_on_cpus(const cpumask_t *m, u32 msr_no,
+				struct msr *msrs)
+{
+       wrmsr_on_cpu(0, msr_no, msrs[0].l, msrs[0].h);
+}
 static inline int rdmsr_safe_on_cpu(unsigned int cpu, u32 msr_no,
 				    u32 *l, u32 *h)
 {
diff --git a/arch/x86/lib/Makefile b/arch/x86/lib/Makefile
index 55e11aa6d66..f9d35632666 100644
--- a/arch/x86/lib/Makefile
+++ b/arch/x86/lib/Makefile
@@ -2,7 +2,7 @@
 # Makefile for x86 specific library files.
 #
 
-obj-$(CONFIG_SMP) := msr-on-cpu.o
+obj-$(CONFIG_SMP) := msr.o
 
 lib-y := delay.o
 lib-y += thunk_$(BITS).o
diff --git a/arch/x86/lib/msr-on-cpu.c b/arch/x86/lib/msr-on-cpu.c
deleted file mode 100644
index 321cf720dbb..00000000000
--- a/arch/x86/lib/msr-on-cpu.c
+++ /dev/null
@@ -1,97 +0,0 @@
-#include <linux/module.h>
-#include <linux/preempt.h>
-#include <linux/smp.h>
-#include <asm/msr.h>
-
-struct msr_info {
-	u32 msr_no;
-	u32 l, h;
-	int err;
-};
-
-static void __rdmsr_on_cpu(void *info)
-{
-	struct msr_info *rv = info;
-
-	rdmsr(rv->msr_no, rv->l, rv->h);
-}
-
-static void __wrmsr_on_cpu(void *info)
-{
-	struct msr_info *rv = info;
-
-	wrmsr(rv->msr_no, rv->l, rv->h);
-}
-
-int rdmsr_on_cpu(unsigned int cpu, u32 msr_no, u32 *l, u32 *h)
-{
-	int err;
-	struct msr_info rv;
-
-	rv.msr_no = msr_no;
-	err = smp_call_function_single(cpu, __rdmsr_on_cpu, &rv, 1);
-	*l = rv.l;
-	*h = rv.h;
-
-	return err;
-}
-
-int wrmsr_on_cpu(unsigned int cpu, u32 msr_no, u32 l, u32 h)
-{
-	int err;
-	struct msr_info rv;
-
-	rv.msr_no = msr_no;
-	rv.l = l;
-	rv.h = h;
-	err = smp_call_function_single(cpu, __wrmsr_on_cpu, &rv, 1);
-
-	return err;
-}
-
-/* These "safe" variants are slower and should be used when the target MSR
-   may not actually exist. */
-static void __rdmsr_safe_on_cpu(void *info)
-{
-	struct msr_info *rv = info;
-
-	rv->err = rdmsr_safe(rv->msr_no, &rv->l, &rv->h);
-}
-
-static void __wrmsr_safe_on_cpu(void *info)
-{
-	struct msr_info *rv = info;
-
-	rv->err = wrmsr_safe(rv->msr_no, rv->l, rv->h);
-}
-
-int rdmsr_safe_on_cpu(unsigned int cpu, u32 msr_no, u32 *l, u32 *h)
-{
-	int err;
-	struct msr_info rv;
-
-	rv.msr_no = msr_no;
-	err = smp_call_function_single(cpu, __rdmsr_safe_on_cpu, &rv, 1);
-	*l = rv.l;
-	*h = rv.h;
-
-	return err ? err : rv.err;
-}
-
-int wrmsr_safe_on_cpu(unsigned int cpu, u32 msr_no, u32 l, u32 h)
-{
-	int err;
-	struct msr_info rv;
-
-	rv.msr_no = msr_no;
-	rv.l = l;
-	rv.h = h;
-	err = smp_call_function_single(cpu, __wrmsr_safe_on_cpu, &rv, 1);
-
-	return err ? err : rv.err;
-}
-
-EXPORT_SYMBOL(rdmsr_on_cpu);
-EXPORT_SYMBOL(wrmsr_on_cpu);
-EXPORT_SYMBOL(rdmsr_safe_on_cpu);
-EXPORT_SYMBOL(wrmsr_safe_on_cpu);
diff --git a/arch/x86/lib/msr.c b/arch/x86/lib/msr.c
new file mode 100644
index 00000000000..1440b9c0547
--- /dev/null
+++ b/arch/x86/lib/msr.c
@@ -0,0 +1,183 @@
+#include <linux/module.h>
+#include <linux/preempt.h>
+#include <linux/smp.h>
+#include <asm/msr.h>
+
+struct msr_info {
+	u32 msr_no;
+	struct msr reg;
+	struct msr *msrs;
+	int off;
+	int err;
+};
+
+static void __rdmsr_on_cpu(void *info)
+{
+	struct msr_info *rv = info;
+	struct msr *reg;
+	int this_cpu = raw_smp_processor_id();
+
+	if (rv->msrs)
+		reg = &rv->msrs[this_cpu - rv->off];
+	else
+		reg = &rv->reg;
+
+	rdmsr(rv->msr_no, reg->l, reg->h);
+}
+
+static void __wrmsr_on_cpu(void *info)
+{
+	struct msr_info *rv = info;
+	struct msr *reg;
+	int this_cpu = raw_smp_processor_id();
+
+	if (rv->msrs)
+		reg = &rv->msrs[this_cpu - rv->off];
+	else
+		reg = &rv->reg;
+
+	wrmsr(rv->msr_no, reg->l, reg->h);
+}
+
+int rdmsr_on_cpu(unsigned int cpu, u32 msr_no, u32 *l, u32 *h)
+{
+	int err;
+	struct msr_info rv;
+
+	memset(&rv, 0, sizeof(rv));
+
+	rv.msr_no = msr_no;
+	err = smp_call_function_single(cpu, __rdmsr_on_cpu, &rv, 1);
+	*l = rv.reg.l;
+	*h = rv.reg.h;
+
+	return err;
+}
+EXPORT_SYMBOL(rdmsr_on_cpu);
+
+int wrmsr_on_cpu(unsigned int cpu, u32 msr_no, u32 l, u32 h)
+{
+	int err;
+	struct msr_info rv;
+
+	memset(&rv, 0, sizeof(rv));
+
+	rv.msr_no = msr_no;
+	rv.reg.l = l;
+	rv.reg.h = h;
+	err = smp_call_function_single(cpu, __wrmsr_on_cpu, &rv, 1);
+
+	return err;
+}
+EXPORT_SYMBOL(wrmsr_on_cpu);
+
+/* rdmsr on a bunch of CPUs
+ *
+ * @mask:       which CPUs
+ * @msr_no:     which MSR
+ * @msrs:       array of MSR values
+ *
+ */
+void rdmsr_on_cpus(const cpumask_t *mask, u32 msr_no, struct msr *msrs)
+{
+	struct msr_info rv;
+	int this_cpu;
+
+	memset(&rv, 0, sizeof(rv));
+
+	rv.off    = cpumask_first(mask);
+	rv.msrs	  = msrs;
+	rv.msr_no = msr_no;
+
+	preempt_disable();
+	/*
+	 * FIXME: handle the CPU we're executing on separately for now until
+	 * smp_call_function_many has been fixed to not skip it.
+	 */
+	this_cpu = raw_smp_processor_id();
+	smp_call_function_single(this_cpu, __rdmsr_on_cpu, &rv, 1);
+
+	smp_call_function_many(mask, __rdmsr_on_cpu, &rv, 1);
+	preempt_enable();
+}
+EXPORT_SYMBOL(rdmsr_on_cpus);
+
+/*
+ * wrmsr on a bunch of CPUs
+ *
+ * @mask:       which CPUs
+ * @msr_no:     which MSR
+ * @msrs:       array of MSR values
+ *
+ */
+void wrmsr_on_cpus(const cpumask_t *mask, u32 msr_no, struct msr *msrs)
+{
+	struct msr_info rv;
+	int this_cpu;
+
+	memset(&rv, 0, sizeof(rv));
+
+	rv.off    = cpumask_first(mask);
+	rv.msrs   = msrs;
+	rv.msr_no = msr_no;
+
+	preempt_disable();
+	/*
+	 * FIXME: handle the CPU we're executing on separately for now until
+	 * smp_call_function_many has been fixed to not skip it.
+	 */
+	this_cpu = raw_smp_processor_id();
+	smp_call_function_single(this_cpu, __wrmsr_on_cpu, &rv, 1);
+
+	smp_call_function_many(mask, __wrmsr_on_cpu, &rv, 1);
+	preempt_enable();
+}
+EXPORT_SYMBOL(wrmsr_on_cpus);
+
+/* These "safe" variants are slower and should be used when the target MSR
+   may not actually exist. */
+static void __rdmsr_safe_on_cpu(void *info)
+{
+	struct msr_info *rv = info;
+
+	rv->err = rdmsr_safe(rv->msr_no, &rv->reg.l, &rv->reg.h);
+}
+
+static void __wrmsr_safe_on_cpu(void *info)
+{
+	struct msr_info *rv = info;
+
+	rv->err = wrmsr_safe(rv->msr_no, rv->reg.l, rv->reg.h);
+}
+
+int rdmsr_safe_on_cpu(unsigned int cpu, u32 msr_no, u32 *l, u32 *h)
+{
+	int err;
+	struct msr_info rv;
+
+	memset(&rv, 0, sizeof(rv));
+
+	rv.msr_no = msr_no;
+	err = smp_call_function_single(cpu, __rdmsr_safe_on_cpu, &rv, 1);
+	*l = rv.reg.l;
+	*h = rv.reg.h;
+
+	return err ? err : rv.err;
+}
+EXPORT_SYMBOL(rdmsr_safe_on_cpu);
+
+int wrmsr_safe_on_cpu(unsigned int cpu, u32 msr_no, u32 l, u32 h)
+{
+	int err;
+	struct msr_info rv;
+
+	memset(&rv, 0, sizeof(rv));
+
+	rv.msr_no = msr_no;
+	rv.reg.l = l;
+	rv.reg.h = h;
+	err = smp_call_function_single(cpu, __wrmsr_safe_on_cpu, &rv, 1);
+
+	return err ? err : rv.err;
+}
+EXPORT_SYMBOL(wrmsr_safe_on_cpu);
diff --git a/drivers/edac/Kconfig b/drivers/edac/Kconfig
index 956982f8739..ab4f3592a11 100644
--- a/drivers/edac/Kconfig
+++ b/drivers/edac/Kconfig
@@ -49,7 +49,6 @@ config EDAC_DEBUG_VERBOSE
 
 config EDAC_MM_EDAC
 	tristate "Main Memory EDAC (Error Detection And Correction) reporting"
-	default y
 	help
 	  Some systems are able to detect and correct errors in main
 	  memory.  EDAC can report statistics on memory error
@@ -58,6 +57,31 @@ config EDAC_MM_EDAC
 	  occurred so that a particular failing memory module can be
 	  replaced.  If unsure, select 'Y'.
 
+config EDAC_AMD64
+	tristate "AMD64 (Opteron, Athlon64) K8, F10h, F11h"
+	depends on EDAC_MM_EDAC && K8_NB && X86_64 && PCI
+	help
+	  Support for error detection and correction on the AMD 64
+	  Families of Memory Controllers (K8, F10h and F11h)
+
+config EDAC_AMD64_ERROR_INJECTION
+	bool "Sysfs Error Injection facilities"
+	depends on EDAC_AMD64
+	help
+	  Recent Opterons (Family 10h and later) provide for Memory Error
+	  Injection into the ECC detection circuits. The amd64_edac module
+	  allows the operator/user to inject Uncorrectable and Correctable
+	  errors into DRAM.
+
+	  When enabled, in each of the respective memory controller directories
+	  (/sys/devices/system/edac/mc/mcX), there are 3 input files:
+
+	  - inject_section (0..3, 16-byte section of 64-byte cacheline),
+	  - inject_word (0..8, 16-bit word of 16-byte section),
+	  - inject_ecc_vector (hex ecc vector: select bits of inject word)
+
+	  In addition, there are two control files, inject_read and inject_write,
+	  which trigger the DRAM ECC Read and Write respectively.
 
 config EDAC_AMD76X
 	tristate "AMD 76x (760, 762, 768)"
diff --git a/drivers/edac/Makefile b/drivers/edac/Makefile
index 59076819135..633dc5604ee 100644
--- a/drivers/edac/Makefile
+++ b/drivers/edac/Makefile
@@ -30,6 +30,13 @@ obj-$(CONFIG_EDAC_I3000)		+= i3000_edac.o
 obj-$(CONFIG_EDAC_X38)			+= x38_edac.o
 obj-$(CONFIG_EDAC_I82860)		+= i82860_edac.o
 obj-$(CONFIG_EDAC_R82600)		+= r82600_edac.o
+
+amd64_edac_mod-y :=  amd64_edac_err_types.o amd64_edac.o
+amd64_edac_mod-$(CONFIG_EDAC_DEBUG) += amd64_edac_dbg.o
+amd64_edac_mod-$(CONFIG_EDAC_AMD64_ERROR_INJECTION) += amd64_edac_inj.o
+
+obj-$(CONFIG_EDAC_AMD64)		+= amd64_edac_mod.o
+
 obj-$(CONFIG_EDAC_PASEMI)		+= pasemi_edac.o
 obj-$(CONFIG_EDAC_MPC85XX)		+= mpc85xx_edac.o
 obj-$(CONFIG_EDAC_MV64X60)		+= mv64x60_edac.o
diff --git a/drivers/edac/amd64_edac.c b/drivers/edac/amd64_edac.c
new file mode 100644
index 00000000000..c36bf40568c
--- /dev/null
+++ b/drivers/edac/amd64_edac.c
@@ -0,0 +1,3354 @@
+#include "amd64_edac.h"
+#include <asm/k8.h>
+
+static struct edac_pci_ctl_info *amd64_ctl_pci;
+
+static int report_gart_errors;
+module_param(report_gart_errors, int, 0644);
+
+/*
+ * Set by command line parameter. If BIOS has enabled the ECC, this override is
+ * cleared to prevent re-enabling the hardware by this driver.
+ */
+static int ecc_enable_override;
+module_param(ecc_enable_override, int, 0644);
+
+/* Lookup table for all possible MC control instances */
+struct amd64_pvt;
+static struct mem_ctl_info *mci_lookup[MAX_NUMNODES];
+static struct amd64_pvt *pvt_lookup[MAX_NUMNODES];
+
+/*
+ * Memory scrubber control interface. For K8, memory scrubbing is handled by
+ * hardware and can involve L2 cache, dcache as well as the main memory. With
+ * F10, this is extended to L3 cache scrubbing on CPU models sporting that
+ * functionality.
+ *
+ * This causes the "units" for the scrubbing speed to vary from 64 byte blocks
+ * (dram) over to cache lines. This is nasty, so we will use bandwidth in
+ * bytes/sec for the setting.
+ *
+ * Currently, we only do dram scrubbing. If the scrubbing is done in software on
+ * other archs, we might not have access to the caches directly.
+ */
+
+/*
+ * scan the scrub rate mapping table for a close or matching bandwidth value to
+ * issue. If requested is too big, then use last maximum value found.
+ */
+static int amd64_search_set_scrub_rate(struct pci_dev *ctl, u32 new_bw,
+				       u32 min_scrubrate)
+{
+	u32 scrubval;
+	int i;
+
+	/*
+	 * map the configured rate (new_bw) to a value specific to the AMD64
+	 * memory controller and apply to register. Search for the first
+	 * bandwidth entry that is greater or equal than the setting requested
+	 * and program that. If at last entry, turn off DRAM scrubbing.
+	 */
+	for (i = 0; i < ARRAY_SIZE(scrubrates); i++) {
+		/*
+		 * skip scrub rates which aren't recommended
+		 * (see F10 BKDG, F3x58)
+		 */
+		if (scrubrates[i].scrubval < min_scrubrate)
+			continue;
+
+		if (scrubrates[i].bandwidth <= new_bw)
+			break;
+
+		/*
+		 * if no suitable bandwidth found, turn off DRAM scrubbing
+		 * entirely by falling back to the last element in the
+		 * scrubrates array.
+		 */
+	}
+
+	scrubval = scrubrates[i].scrubval;
+	if (scrubval)
+		edac_printk(KERN_DEBUG, EDAC_MC,
+			    "Setting scrub rate bandwidth: %u\n",
+			    scrubrates[i].bandwidth);
+	else
+		edac_printk(KERN_DEBUG, EDAC_MC, "Turning scrubbing off.\n");
+
+	pci_write_bits32(ctl, K8_SCRCTRL, scrubval, 0x001F);
+
+	return 0;
+}
+
+static int amd64_set_scrub_rate(struct mem_ctl_info *mci, u32 *bandwidth)
+{
+	struct amd64_pvt *pvt = mci->pvt_info;
+	u32 min_scrubrate = 0x0;
+
+	switch (boot_cpu_data.x86) {
+	case 0xf:
+		min_scrubrate = K8_MIN_SCRUB_RATE_BITS;
+		break;
+	case 0x10:
+		min_scrubrate = F10_MIN_SCRUB_RATE_BITS;
+		break;
+	case 0x11:
+		min_scrubrate = F11_MIN_SCRUB_RATE_BITS;
+		break;
+
+	default:
+		amd64_printk(KERN_ERR, "Unsupported family!\n");
+		break;
+	}
+	return amd64_search_set_scrub_rate(pvt->misc_f3_ctl, *bandwidth,
+			min_scrubrate);
+}
+
+static int amd64_get_scrub_rate(struct mem_ctl_info *mci, u32 *bw)
+{
+	struct amd64_pvt *pvt = mci->pvt_info;
+	u32 scrubval = 0;
+	int status = -1, i, ret = 0;
+
+	ret = pci_read_config_dword(pvt->misc_f3_ctl, K8_SCRCTRL, &scrubval);
+	if (ret)
+		debugf0("Reading K8_SCRCTRL failed\n");
+
+	scrubval = scrubval & 0x001F;
+
+	edac_printk(KERN_DEBUG, EDAC_MC,
+		    "pci-read, sdram scrub control value: %d \n", scrubval);
+
+	for (i = 0; ARRAY_SIZE(scrubrates); i++) {
+		if (scrubrates[i].scrubval == scrubval) {
+			*bw = scrubrates[i].bandwidth;
+			status = 0;
+			break;
+		}
+	}
+
+	return status;
+}
+
+/* Map from a CSROW entry to the mask entry that operates on it */
+static inline u32 amd64_map_to_dcs_mask(struct amd64_pvt *pvt, int csrow)
+{
+	return csrow >> (pvt->num_dcsm >> 3);
+}
+
+/* return the 'base' address the i'th CS entry of the 'dct' DRAM controller */
+static u32 amd64_get_dct_base(struct amd64_pvt *pvt, int dct, int csrow)
+{
+	if (dct == 0)
+		return pvt->dcsb0[csrow];
+	else
+		return pvt->dcsb1[csrow];
+}
+
+/*
+ * Return the 'mask' address the i'th CS entry. This function is needed because
+ * there number of DCSM registers on Rev E and prior vs Rev F and later is
+ * different.
+ */
+static u32 amd64_get_dct_mask(struct amd64_pvt *pvt, int dct, int csrow)
+{
+	if (dct == 0)
+		return pvt->dcsm0[amd64_map_to_dcs_mask(pvt, csrow)];
+	else
+		return pvt->dcsm1[amd64_map_to_dcs_mask(pvt, csrow)];
+}
+
+
+/*
+ * In *base and *limit, pass back the full 40-bit base and limit physical
+ * addresses for the node given by node_id.  This information is obtained from
+ * DRAM Base (section 3.4.4.1) and DRAM Limit (section 3.4.4.2) registers. The
+ * base and limit addresses are of type SysAddr, as defined at the start of
+ * section 3.4.4 (p. 70).  They are the lowest and highest physical addresses
+ * in the address range they represent.
+ */
+static void amd64_get_base_and_limit(struct amd64_pvt *pvt, int node_id,
+			       u64 *base, u64 *limit)
+{
+	*base = pvt->dram_base[node_id];
+	*limit = pvt->dram_limit[node_id];
+}
+
+/*
+ * Return 1 if the SysAddr given by sys_addr matches the base/limit associated
+ * with node_id
+ */
+static int amd64_base_limit_match(struct amd64_pvt *pvt,
+					u64 sys_addr, int node_id)
+{
+	u64 base, limit, addr;
+
+	amd64_get_base_and_limit(pvt, node_id, &base, &limit);
+
+	/* The K8 treats this as a 40-bit value.  However, bits 63-40 will be
+	 * all ones if the most significant implemented address bit is 1.
+	 * Here we discard bits 63-40.  See section 3.4.2 of AMD publication
+	 * 24592: AMD x86-64 Architecture Programmer's Manual Volume 1
+	 * Application Programming.
+	 */
+	addr = sys_addr & 0x000000ffffffffffull;
+
+	return (addr >= base) && (addr <= limit);
+}
+
+/*
+ * Attempt to map a SysAddr to a node. On success, return a pointer to the
+ * mem_ctl_info structure for the node that the SysAddr maps to.
+ *
+ * On failure, return NULL.
+ */
+static struct mem_ctl_info *find_mc_by_sys_addr(struct mem_ctl_info *mci,
+						u64 sys_addr)
+{
+	struct amd64_pvt *pvt;
+	int node_id;
+	u32 intlv_en, bits;
+
+	/*
+	 * Here we use the DRAM Base (section 3.4.4.1) and DRAM Limit (section
+	 * 3.4.4.2) registers to map the SysAddr to a node ID.
+	 */
+	pvt = mci->pvt_info;
+
+	/*
+	 * The value of this field should be the same for all DRAM Base
+	 * registers.  Therefore we arbitrarily choose to read it from the
+	 * register for node 0.
+	 */
+	intlv_en = pvt->dram_IntlvEn[0];
+
+	if (intlv_en == 0) {
+		for (node_id = 0; ; ) {
+			if (amd64_base_limit_match(pvt, sys_addr, node_id))
+				break;
+
+			if (++node_id >= DRAM_REG_COUNT)
+				goto err_no_match;
+		}
+		goto found;
+	}
+
+	if (unlikely((intlv_en != (0x01 << 8)) &&
+		     (intlv_en != (0x03 << 8)) &&
+		     (intlv_en != (0x07 << 8)))) {
+		amd64_printk(KERN_WARNING, "junk value of 0x%x extracted from "
+			     "IntlvEn field of DRAM Base Register for node 0: "
+			     "This probably indicates a BIOS bug.\n", intlv_en);
+		return NULL;
+	}
+
+	bits = (((u32) sys_addr) >> 12) & intlv_en;
+
+	for (node_id = 0; ; ) {
+		if ((pvt->dram_limit[node_id] & intlv_en) == bits)
+			break;	/* intlv_sel field matches */
+
+		if (++node_id >= DRAM_REG_COUNT)
+			goto err_no_match;
+	}
+
+	/* sanity test for sys_addr */
+	if (unlikely(!amd64_base_limit_match(pvt, sys_addr, node_id))) {
+		amd64_printk(KERN_WARNING,
+			  "%s(): sys_addr 0x%lx falls outside base/limit "
+			  "address range for node %d with node interleaving "
+			  "enabled.\n", __func__, (unsigned long)sys_addr,
+			  node_id);
+		return NULL;
+	}
+
+found:
+	return edac_mc_find(node_id);
+
+err_no_match:
+	debugf2("sys_addr 0x%lx doesn't match any node\n",
+		(unsigned long)sys_addr);
+
+	return NULL;
+}
+
+/*
+ * Extract the DRAM CS base address from selected csrow register.
+ */
+static u64 base_from_dct_base(struct amd64_pvt *pvt, int csrow)
+{
+	return ((u64) (amd64_get_dct_base(pvt, 0, csrow) & pvt->dcsb_base)) <<
+				pvt->dcs_shift;
+}
+
+/*
+ * Extract the mask from the dcsb0[csrow] entry in a CPU revision-specific way.
+ */
+static u64 mask_from_dct_mask(struct amd64_pvt *pvt, int csrow)
+{
+	u64 dcsm_bits, other_bits;
+	u64 mask;
+
+	/* Extract bits from DRAM CS Mask. */
+	dcsm_bits = amd64_get_dct_mask(pvt, 0, csrow) & pvt->dcsm_mask;
+
+	other_bits = pvt->dcsm_mask;
+	other_bits = ~(other_bits << pvt->dcs_shift);
+
+	/*
+	 * The extracted bits from DCSM belong in the spaces represented by
+	 * the cleared bits in other_bits.
+	 */
+	mask = (dcsm_bits << pvt->dcs_shift) | other_bits;
+
+	return mask;
+}
+
+/*
+ * @input_addr is an InputAddr associated with the node given by mci. Return the
+ * csrow that input_addr maps to, or -1 on failure (no csrow claims input_addr).
+ */
+static int input_addr_to_csrow(struct mem_ctl_info *mci, u64 input_addr)
+{
+	struct amd64_pvt *pvt;
+	int csrow;
+	u64 base, mask;
+
+	pvt = mci->pvt_info;
+
+	/*
+	 * Here we use the DRAM CS Base and DRAM CS Mask registers. For each CS
+	 * base/mask register pair, test the condition shown near the start of
+	 * section 3.5.4 (p. 84, BKDG #26094, K8, revA-E).
+	 */
+	for (csrow = 0; csrow < CHIPSELECT_COUNT; csrow++) {
+
+		/* This DRAM chip select is disabled on this node */
+		if ((pvt->dcsb0[csrow] & K8_DCSB_CS_ENABLE) == 0)
+			continue;
+
+		base = base_from_dct_base(pvt, csrow);
+		mask = ~mask_from_dct_mask(pvt, csrow);
+
+		if ((input_addr & mask) == (base & mask)) {
+			debugf2("InputAddr 0x%lx matches csrow %d (node %d)\n",
+				(unsigned long)input_addr, csrow,
+				pvt->mc_node_id);
+
+			return csrow;
+		}
+	}
+
+	debugf2("no matching csrow for InputAddr 0x%lx (MC node %d)\n",
+		(unsigned long)input_addr, pvt->mc_node_id);
+
+	return -1;
+}
+
+/*
+ * Return the base value defined by the DRAM Base register for the node
+ * represented by mci.  This function returns the full 40-bit value despite the
+ * fact that the register only stores bits 39-24 of the value. See section
+ * 3.4.4.1 (BKDG #26094, K8, revA-E)
+ */
+static inline u64 get_dram_base(struct mem_ctl_info *mci)
+{
+	struct amd64_pvt *pvt = mci->pvt_info;
+
+	return pvt->dram_base[pvt->mc_node_id];
+}
+
+/*
+ * Obtain info from the DRAM Hole Address Register (section 3.4.8, pub #26094)
+ * for the node represented by mci. Info is passed back in *hole_base,
+ * *hole_offset, and *hole_size.  Function returns 0 if info is valid or 1 if
+ * info is invalid. Info may be invalid for either of the following reasons:
+ *
+ * - The revision of the node is not E or greater.  In this case, the DRAM Hole
+ *   Address Register does not exist.
+ *
+ * - The DramHoleValid bit is cleared in the DRAM Hole Address Register,
+ *   indicating that its contents are not valid.
+ *
+ * The values passed back in *hole_base, *hole_offset, and *hole_size are
+ * complete 32-bit values despite the fact that the bitfields in the DHAR
+ * only represent bits 31-24 of the base and offset values.
+ */
+int amd64_get_dram_hole_info(struct mem_ctl_info *mci, u64 *hole_base,
+			     u64 *hole_offset, u64 *hole_size)
+{
+	struct amd64_pvt *pvt = mci->pvt_info;
+	u64 base;
+
+	/* only revE and later have the DRAM Hole Address Register */
+	if (boot_cpu_data.x86 == 0xf && pvt->ext_model < OPTERON_CPU_REV_E) {
+		debugf1("  revision %d for node %d does not support DHAR\n",
+			pvt->ext_model, pvt->mc_node_id);
+		return 1;
+	}
+
+	/* only valid for Fam10h */
+	if (boot_cpu_data.x86 == 0x10 &&
+	    (pvt->dhar & F10_DRAM_MEM_HOIST_VALID) == 0) {
+		debugf1("  Dram Memory Hoisting is DISABLED on this system\n");
+		return 1;
+	}
+
+	if ((pvt->dhar & DHAR_VALID) == 0) {
+		debugf1("  Dram Memory Hoisting is DISABLED on this node %d\n",
+			pvt->mc_node_id);
+		return 1;
+	}
+
+	/* This node has Memory Hoisting */
+
+	/* +------------------+--------------------+--------------------+-----
+	 * | memory           | DRAM hole          | relocated          |
+	 * | [0, (x - 1)]     | [x, 0xffffffff]    | addresses from     |
+	 * |                  |                    | DRAM hole          |
+	 * |                  |                    | [0x100000000,      |
+	 * |                  |                    |  (0x100000000+     |
+	 * |                  |                    |   (0xffffffff-x))] |
+	 * +------------------+--------------------+--------------------+-----
+	 *
+	 * Above is a diagram of physical memory showing the DRAM hole and the
+	 * relocated addresses from the DRAM hole.  As shown, the DRAM hole
+	 * starts at address x (the base address) and extends through address
+	 * 0xffffffff.  The DRAM Hole Address Register (DHAR) relocates the
+	 * addresses in the hole so that they start at 0x100000000.
+	 */
+
+	base = dhar_base(pvt->dhar);
+
+	*hole_base = base;
+	*hole_size = (0x1ull << 32) - base;
+
+	if (boot_cpu_data.x86 > 0xf)
+		*hole_offset = f10_dhar_offset(pvt->dhar);
+	else
+		*hole_offset = k8_dhar_offset(pvt->dhar);
+
+	debugf1("  DHAR info for node %d base 0x%lx offset 0x%lx size 0x%lx\n",
+		pvt->mc_node_id, (unsigned long)*hole_base,
+		(unsigned long)*hole_offset, (unsigned long)*hole_size);
+
+	return 0;
+}
+EXPORT_SYMBOL_GPL(amd64_get_dram_hole_info);
+
+/*
+ * Return the DramAddr that the SysAddr given by @sys_addr maps to.  It is
+ * assumed that sys_addr maps to the node given by mci.
+ *
+ * The first part of section 3.4.4 (p. 70) shows how the DRAM Base (section
+ * 3.4.4.1) and DRAM Limit (section 3.4.4.2) registers are used to translate a
+ * SysAddr to a DramAddr. If the DRAM Hole Address Register (DHAR) is enabled,
+ * then it is also involved in translating a SysAddr to a DramAddr. Sections
+ * 3.4.8 and 3.5.8.2 describe the DHAR and how it is used for memory hoisting.
+ * These parts of the documentation are unclear. I interpret them as follows:
+ *
+ * When node n receives a SysAddr, it processes the SysAddr as follows:
+ *
+ * 1. It extracts the DRAMBase and DRAMLimit values from the DRAM Base and DRAM
+ *    Limit registers for node n. If the SysAddr is not within the range
+ *    specified by the base and limit values, then node n ignores the Sysaddr
+ *    (since it does not map to node n). Otherwise continue to step 2 below.
+ *
+ * 2. If the DramHoleValid bit of the DHAR for node n is clear, the DHAR is
+ *    disabled so skip to step 3 below. Otherwise see if the SysAddr is within
+ *    the range of relocated addresses (starting at 0x100000000) from the DRAM
+ *    hole. If not, skip to step 3 below. Else get the value of the
+ *    DramHoleOffset field from the DHAR. To obtain the DramAddr, subtract the
+ *    offset defined by this value from the SysAddr.
+ *
+ * 3. Obtain the base address for node n from the DRAMBase field of the DRAM
+ *    Base register for node n. To obtain the DramAddr, subtract the base
+ *    address from the SysAddr, as shown near the start of section 3.4.4 (p.70).
+ */
+static u64 sys_addr_to_dram_addr(struct mem_ctl_info *mci, u64 sys_addr)
+{
+	u64 dram_base, hole_base, hole_offset, hole_size, dram_addr;
+	int ret = 0;
+
+	dram_base = get_dram_base(mci);
+
+	ret = amd64_get_dram_hole_info(mci, &hole_base, &hole_offset,
+				      &hole_size);
+	if (!ret) {
+		if ((sys_addr >= (1ull << 32)) &&
+		    (sys_addr < ((1ull << 32) + hole_size))) {
+			/* use DHAR to translate SysAddr to DramAddr */
+			dram_addr = sys_addr - hole_offset;
+
+			debugf2("using DHAR to translate SysAddr 0x%lx to "
+				"DramAddr 0x%lx\n",
+				(unsigned long)sys_addr,
+				(unsigned long)dram_addr);
+
+			return dram_addr;
+		}
+	}
+
+	/*
+	 * Translate the SysAddr to a DramAddr as shown near the start of
+	 * section 3.4.4 (p. 70).  Although sys_addr is a 64-bit value, the k8
+	 * only deals with 40-bit values.  Therefore we discard bits 63-40 of
+	 * sys_addr below.  If bit 39 of sys_addr is 1 then the bits we
+	 * discard are all 1s.  Otherwise the bits we discard are all 0s.  See
+	 * section 3.4.2 of AMD publication 24592: AMD x86-64 Architecture
+	 * Programmer's Manual Volume 1 Application Programming.
+	 */
+	dram_addr = (sys_addr & 0xffffffffffull) - dram_base;
+
+	debugf2("using DRAM Base register to translate SysAddr 0x%lx to "
+		"DramAddr 0x%lx\n", (unsigned long)sys_addr,
+		(unsigned long)dram_addr);
+	return dram_addr;
+}
+
+/*
+ * @intlv_en is the value of the IntlvEn field from a DRAM Base register
+ * (section 3.4.4.1).  Return the number of bits from a SysAddr that are used
+ * for node interleaving.
+ */
+static int num_node_interleave_bits(unsigned intlv_en)
+{
+	static const int intlv_shift_table[] = { 0, 1, 0, 2, 0, 0, 0, 3 };
+	int n;
+
+	BUG_ON(intlv_en > 7);
+	n = intlv_shift_table[intlv_en];
+	return n;
+}
+
+/* Translate the DramAddr given by @dram_addr to an InputAddr. */
+static u64 dram_addr_to_input_addr(struct mem_ctl_info *mci, u64 dram_addr)
+{
+	struct amd64_pvt *pvt;
+	int intlv_shift;
+	u64 input_addr;
+
+	pvt = mci->pvt_info;
+
+	/*
+	 * See the start of section 3.4.4 (p. 70, BKDG #26094, K8, revA-E)
+	 * concerning translating a DramAddr to an InputAddr.
+	 */
+	intlv_shift = num_node_interleave_bits(pvt->dram_IntlvEn[0]);
+	input_addr = ((dram_addr >> intlv_shift) & 0xffffff000ull) +
+	    (dram_addr & 0xfff);
+
+	debugf2("  Intlv Shift=%d DramAddr=0x%lx maps to InputAddr=0x%lx\n",
+		intlv_shift, (unsigned long)dram_addr,
+		(unsigned long)input_addr);
+
+	return input_addr;
+}
+
+/*
+ * Translate the SysAddr represented by @sys_addr to an InputAddr.  It is
+ * assumed that @sys_addr maps to the node given by mci.
+ */
+static u64 sys_addr_to_input_addr(struct mem_ctl_info *mci, u64 sys_addr)
+{
+	u64 input_addr;
+
+	input_addr =
+	    dram_addr_to_input_addr(mci, sys_addr_to_dram_addr(mci, sys_addr));
+
+	debugf2("SysAdddr 0x%lx translates to InputAddr 0x%lx\n",
+		(unsigned long)sys_addr, (unsigned long)input_addr);
+
+	return input_addr;
+}
+
+
+/*
+ * @input_addr is an InputAddr associated with the node represented by mci.
+ * Translate @input_addr to a DramAddr and return the result.
+ */
+static u64 input_addr_to_dram_addr(struct mem_ctl_info *mci, u64 input_addr)
+{
+	struct amd64_pvt *pvt;
+	int node_id, intlv_shift;
+	u64 bits, dram_addr;
+	u32 intlv_sel;
+
+	/*
+	 * Near the start of section