cpu/intel/slot_1: Init L2 cache on SECC(2) CPUs.

Bring from coreboot v1 support for initializing L2 cache on Slot 1
Pentium II/III CPUs, code names Klamath, Deschutes and Katmai.

Build tested on ASUS P2B-LS and P3B-F. Boot tested on P2B-LS with
Pentium III 600MHz, Katmai core.

Also add missing include of model_68x in slot_1, to address a
similar problem fixed for model_6bx by r5945.

Also change Deschutes CPU init sequence to match Katmai.

Change-Id: I502e8481d1a20f0a2504685e7be16b57f59c8257
Signed-off-by: Keith Hui <buurin@gmail.com>
Reviewed-on: http://review.coreboot.org/122
Tested-by: build bot (Jenkins)
Reviewed-by: Stefan Reinauer <stefan.reinauer@coreboot.org>

4 files changed, 823 insertions, 3 deletions

diff --git a/src/cpu/intel/model_65x/model_65x_init.c b/src/cpu/intel/model_65x/model_65x_init.c
index 8f9fc66f45..ef9759782c 100644
--- a/src/cpu/intel/model_65x/model_65x_init.c
+++ b/src/cpu/intel/model_65x/model_65x_init.c
@@ -27,6 +27,7 @@
 #include <cpu/x86/lapic.h>
 #include <cpu/intel/microcode.h>
 #include <cpu/x86/cache.h>
+#include <cpu/intel/l2_cache.h>
 
 static u32 microcode_updates[] = {
 	#include "microcode-410-MU16522d.h"
@@ -56,14 +57,17 @@ static u32 microcode_updates[] = {
 
 static void model_65x_init(device_t dev)
 {
+	/* Update the microcode */
+	intel_update_microcode(microcode_updates);
+
+	/* Initialize L2 cache */
+	p6_configure_l2_cache();
+
 	/* Turn on caching if we haven't already */
 	x86_enable_cache();
 	x86_setup_mtrrs(36);
 	x86_mtrr_check();
 
-	/* Update the microcode */
-	intel_update_microcode(microcode_updates);
-
 	/* Enable the local cpu apics */
 	setup_lapic();
 };
diff --git a/src/cpu/intel/model_67x/model_67x_init.c b/src/cpu/intel/model_67x/model_67x_init.c
index 76a840e46e..0c9b3d2a08 100644
--- a/src/cpu/intel/model_67x/model_67x_init.c
+++ b/src/cpu/intel/model_67x/model_67x_init.c
@@ -26,6 +26,7 @@
 #include <cpu/intel/microcode.h>
 #include <cpu/x86/cache.h>
 #include <cpu/x86/msr.h>
+#include <cpu/intel/l2_cache.h>
 
 static const uint32_t microcode_updates[] = {
 	/* Include microcode updates here. */
@@ -46,6 +47,9 @@ static void model_67x_init(device_t cpu)
 	/* Update the microcode */
 	intel_update_microcode(microcode_updates);
 
+	/* Initialize L2 cache */
+	p6_configure_l2_cache();
+
 	/* Turn on caching if we haven't already */
 	x86_enable_cache();
 
diff --git a/src/cpu/intel/slot_1/Makefile.inc b/src/cpu/intel/slot_1/Makefile.inc
index 297ef183d9..a4de6e0b97 100644
--- a/src/cpu/intel/slot_1/Makefile.inc
+++ b/src/cpu/intel/slot_1/Makefile.inc
@@ -19,9 +19,11 @@
 ##
 
 ramstage-y += slot_1.c
+ramstage-y += l2_cache.c
 subdirs-y += ../model_6xx
 subdirs-y += ../model_65x
 subdirs-y += ../model_67x
+subdirs-y += ../model_68x
 subdirs-y += ../model_6bx
 subdirs-y += ../../x86/tsc
 subdirs-y += ../../x86/mtrr
diff --git a/src/cpu/intel/slot_1/l2_cache.c b/src/cpu/intel/slot_1/l2_cache.c
new file mode 100644
index 0000000000..9a694e0147
--- /dev/null
+++ b/src/cpu/intel/slot_1/l2_cache.c
@@ -0,0 +1,810 @@
+/*
+ * This file is part of the coreboot project.
+ *
+ * Copyright (C) 2000 Denis Dowling <dpd@alphalink.com.au>
+ * Copyright (C) 2010 Keith Hui <buurin@gmail.com>
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * 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., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301 USA
+ */
+
+/*
+ * Intel Pentium L2 Cache initialization.
+ * This code was developed by reverse engineering
+ * the BIOS. Where the code accesses documented
+ * registers I have added comments as best I can.
+ * Some undocumented registers on the Pentium II are
+ * used so some of the documentation is incomplete.
+ *
+ * References:
+ * Intel Architecture Software Developer's Manual
+ * Volume 3B: System Programming Guide, Part 2 (#253669)
+ * Appendix B.9
+ */
+
+/* This code is ported from coreboot v1.
+ * The L2 cache initalization sequence here only apply to SECC/SECC2 P6 family
+ * CPUs with Klamath (63x), Deschutes (65x) and Katmai (67x) cores.
+ * It is not required for Coppermine (68x) and Tualatin (6bx) cores.
+ * It is currently not known if Celerons with Mendocino (66x) core require the
+ * special initialization.
+ * Covington-core Celerons do not have L2 cache.
+ */
+
+#include <stdint.h>
+#include <console/console.h>
+#include <string.h>
+#include <cpu/cpu.h>
+#include <cpu/x86/mtrr.h>
+#include <cpu/intel/l2_cache.h>
+#include <cpu/x86/cache.h>
+#include <cpu/x86/msr.h>
+
+/* Latency Tables */
+struct latency_entry {
+	u8 key;
+	u8 value;
+};
+/*
+Latency maps for Deschutes and Katmai.
+No such mapping is available for Klamath.
+
+Cache latency to
+be written to L2 -----++++
+control register      ||||
+0000 xx 00 -----> 000 cccc 0
+||||    00 66MHz
+||||    10 100MHz
+||||    01 133MHz (Katmai "B" only)
+++++------ CPU frequency multiplier
+
+0000 2x
+0001 3x
+0010 4x
+0011 5x
+0100 2.5x
+0101 3.5x
+0110 4.5x
+0111 5.5x
+1000 6x
+1001 7x
+1010 8x
+1011 Reserved
+1100 6.5x
+1101 7.5x
+1110 1.5x
+1111 2x
+
+*/
+static const struct latency_entry latency_650_t0[] = {
+	{0x10, 0x02}, {0x50, 0x02}, {0x20, 0x04}, {0x60, 0x06},
+	{0x00, 0x08}, {0x40, 0x0C}, {0x12, 0x06}, {0x52, 0x0A},
+	{0x22, 0x0E}, {0x62, 0x10}, {0x02, 0x10}, {0xFF, 0x00}
+};
+
+static const struct latency_entry latency_650_t1[] = {
+	{0x12, 0x14}, {0x52, 0x16}, {0x22, 0x16}, {0x62, 0x16},
+	{0xFF, 0x00}
+};
+
+static const struct latency_entry latency_670_t0[] = {
+	{0x60, 0x06}, {0x00, 0x08}, {0x12, 0x06}, {0x52, 0x0A},
+	{0x22, 0x0E}, {0x62, 0x10}, {0x02, 0x10}, {0x42, 0x02},
+	{0x11, 0x0E}, {0x51, 0x0C}, {0x21, 0x02}, {0x61, 0x10},
+	{0x01, 0x10}, {0x41, 0x02}, {0xFF, 0x00}
+};
+
+static const struct latency_entry latency_670_t1[] = {
+	{0x22, 0x18}, {0x62, 0x18}, {0x02, 0x1A}, {0x11, 0x18},
+	{0xFF, 0x00}
+};
+
+static const struct latency_entry latency_670_t2[] = {
+	{0x22, 0x12}, {0x62, 0x14}, {0x02, 0x16}, {0x42, 0x1E},
+	{0x11, 0x12}, {0x51, 0x16}, {0x21, 0x1E}, {0x61, 0x14},
+	{0x01, 0x16}, {0x41, 0x1E}, {0xFF, 0x00}
+};
+
+/* Latency tables for 650 model/type */
+static const struct latency_entry *latency_650[] = {
+	latency_650_t0, latency_650_t1, latency_650_t1
+};
+
+/* Latency tables for 670 model/type */
+static const struct latency_entry *latency_670[] = {
+	latency_670_t0, latency_670_t1, latency_670_t2
+};
+
+int calculate_l2_latency(void)
+{
+	u32 eax, l, signature;
+	const struct latency_entry *latency_table, *le;
+	msr_t msr;
+
+	/* First, attempt to get cache latency value from
+	   IA32_PLATFORM_ID[56:53]. (L2 Cache Latency Read)
+	 */
+	msr = rdmsr(IA32_PLATFORM_ID);
+
+	printk(BIOS_DEBUG,"rdmsr(IA32_PLATFORM_ID) = %x:%x\n", msr.hi, msr.lo);
+
+	l = (msr.hi >> 20) & 0x1e;
+
+	if (l == 0) {
+		/* If latency value isn't available from
+		   IA32_PLATFORM_ID[56:53], read it from
+		   L2 control register 0 for lookup from
+		   tables. */
+		int t, a;
+
+		/* The raw code is read from L2 register 0, bits [7:4]. */
+		a = read_l2(0);
+		if (a < 0)
+			return -1;
+
+		a &= 0xf0;
+
+		if ((a & 0x20) == 0)
+			t = 0;
+		else if (a == 0x20)
+			t = 1;
+		else if (a == 0x30)
+			t = 2;
+		else
+			return -1;
+
+		printk(BIOS_DEBUG,"L2 latency type = %x\n", t);
+
+		/* Get CPUID family/model */
+		signature = cpuid_eax(1) & 0xfff0;
+
+		/* Read EBL_CR_POWERON */
+		msr = rdmsr(EBL_CR_POWERON);
+		/* Get clock multiplier and FSB frequency.
+		 * Multiplier is in [25:22].
+		 * FSB is in [19:18] in Katmai, [19] in Deschutes ([18] is zero for them).
+		 */
+		eax = msr.lo >> 18;
+		if (signature == 0x650) {
+			eax &= ~0xf2;
+			latency_table = latency_650[t];
+		} else if (signature == 0x670) {
+			eax &= ~0xf3;
+			latency_table = latency_670[t];
+		} else
+			return -1;
+
+		/* Search table for matching entry */
+		for (le = latency_table; le->key != eax; le++) {
+			/* Fail if we get to the end of the table */
+			if (le->key == 0xff) {
+				printk(BIOS_DEBUG, "Could not find key %02x in latency table\n", eax);
+				return -1;
+			}
+		}
+
+		l = le->value;
+	}
+
+	printk(BIOS_DEBUG,"L2 Cache latency is %d\n", l / 2);
+
+	/* Writes the calculated latency in BBL_CR_CTL3[4:1]. */
+	msr = rdmsr(BBL_CR_CTL3);
+	msr.lo &= 0xffffffe1;
+	msr.lo |= l;
+	wrmsr(BBL_CR_CTL3, msr);
+
+	return 0;
+}
+
+
+/* Setup address, data_high:data_low into the L2
+ * control registers and then issue command with correct cache way
+ */
+int signal_l2(u32 address, u32 data_high, u32 data_low, int way, u8 command)
+{
+	int i;
+	msr_t msr;
+
+	/* Write L2 Address to BBL_CR_ADDR */
+	msr.lo = address;
+	msr.hi = 0;
+	wrmsr(BBL_CR_ADDR, msr);
+
+	/* Write data to BBL_CR_D{0..3} */
+	msr.lo = data_low;
+	msr.hi = data_high;
+	for (i = BBL_CR_D0; i <= BBL_CR_D3; i++) {
+		wrmsr(i, msr);
+	}
+
+	/* Put the command and way into BBL_CR_CTL */
+	msr = rdmsr(BBL_CR_CTL);
+	msr.lo = (msr.lo & 0xfffffce0) | command | (way << 8);
+	wrmsr(BBL_CR_CTL, msr);
+
+	/* Trigger L2 controller */
+	msr.lo = 0;
+	msr.hi = 0;
+	wrmsr(BBL_CR_TRIG, msr);
+
+	/* Poll the controller to see when done */
+	for (i = 0; i < 0x100; i++) {
+		/* Read BBL_CR_BUSY */
+		msr = rdmsr(BBL_CR_BUSY);
+		/* If not busy then return */
+		if ((msr.lo & 1) == 0)
+			return 0;
+	}
+
+	/* Return timeout code */
+	return -1;
+}
+
+/* Read the L2 Cache controller register at given address */
+int read_l2(u32 address)
+{
+	msr_t msr;
+
+	/* Send a L2 Control Register Read to L2 controller */
+	if (signal_l2(address << 5, 0, 0, 0, L2CMD_CR) != 0)
+		return -1;
+
+	/* If OK then get the result from BBL_CR_ADDR */
+	msr = rdmsr(BBL_CR_ADDR);
+	return (msr.lo >> 0x15);
+
+}
+
+/* Write data into the L2 controller register at address */
+int write_l2(u32 address, u32 data)
+{
+	int v1, v2, i;
+
+	v1 = read_l2(0);
+	if (v1 < 0)
+		return -1;
+
+	v2 = read_l2(2);
+	if (v2 < 0)
+		return -1;
+
+	if ((v1 & 0x20) == 0) {
+		v2 &= 0x3;
+		v2++;
+	} else
+		v2 &= 0x7;
+
+	/* This write has to be replicated to a number of places. Not sure what. */
+
+	for (i = 0; i < v2; i++) {
+
+		u32 data1, data2;
+		// Bits legend
+		// data1   = ffffffff
+		// data2   = 000000dc
+		// address = 00aaaaaa
+		// Final address signalled:
+		// 000fffff fff000c0 000dcaaa aaa00000
+		data1 = data & 0xff;
+		data1 = data1 << 21;
+		data2 = (i << 11) & 0x1800;
+		data1 |= data2;
+		data2 <<= 6;
+		data2 &= 0x20000;
+		data1 |= data2;
+
+		/* Signal L2 controller */
+		if (signal_l2((address << 5) | data1, 0, 0, 0, 3))
+			return -1;
+	}
+	return 0;
+}
+
+/* Write data_high:data_low into the cache at address1. Test address2
+ * to see if the same data is returned. Return 0 if the data matches.
+ * return lower 16 bits if mismatched data if mismatch. Return -1
+ * on error
+ */
+int test_l2_address_alias(u32 address1, u32 address2,
+				 u32 data_high, u32 data_low)
+{
+	int d;
+	msr_t msr;
+
+	/* Tag Write with Data Write for L2 */
+	if (signal_l2(address1, data_high, data_low, 0, L2CMD_TWW))
+		return -1;
+
+	/* Tag Read with Data Read for L2 */
+	if (signal_l2(address2, 0, 0, 0, L2CMD_TRR))
+		return -1;
+
+	/* Read data from BBL_CR_D[0-3] */
+	for (d = BBL_CR_D0; d <= BBL_CR_D3; d++) {
+		msr = rdmsr(d);
+		if (msr.lo != data_low || msr.hi != data_high)
+			return (msr.lo & 0xffff);
+	}
+
+	return 0;
+}
+
+/* Calculates the L2 cache size.
+ *
+ * Reference: Intel(R) 64 and IA-32 Architectures Software Developer’s Manual
+ *            Volume 3B: System Programming Guide, Part 2, Intel pub. 253669, pg. B-172.
+ *
+ */
+int calculate_l2_cache_size(void)
+{
+	int v;
+	msr_t msr;
+	u32 cache_setting;
+	u32 address, size, eax, bblcr3;
+
+	v = read_l2(0);
+	if (v < 0)
+		return -1;
+	if ((v & 0x20) == 0) {
+		msr = rdmsr(BBL_CR_CTL3);
+		bblcr3 = msr.lo & ~BBLCR3_L2_SIZE;
+		/*
+		 * Successively write in all the possible cache size per bank
+		 * into BBL_CR_CTL3[17:13], starting from 256KB (00001) to 4MB (10000),
+		 * and read the last value written and accepted by the cache.
+		 *
+		 * No idea why these bits are writable at all.
+		 */
+		for (cache_setting = BBLCR3_L2_SIZE_256K;
+		     cache_setting <= BBLCR3_L2_SIZE_4M; cache_setting <<= 1) {
+
+			eax = bblcr3 | cache_setting;
+			msr.lo = eax;
+			wrmsr(BBL_CR_CTL3, msr);
+			msr = rdmsr(BBL_CR_CTL3);
+
+			/* Value not accepted */
+			if (msr.lo != eax)
+				break;
+		}
+
+		/* Backtrack to the last value that worked... */
+		cache_setting >>= 1;
+
+		/* and write it into BBL_CR_CTL3 */
+		msr.lo &= ~BBLCR3_L2_SIZE;
+		msr.lo |= (cache_setting & BBLCR3_L2_SIZE);
+
+		wrmsr(BBL_CR_CTL3, msr);
+
+		printk(BIOS_DEBUG,"Maximum cache mask is %x\n", cache_setting);
+
+		/* For now, BBL_CR_CTL3 has the highest cache "size" that register
+		 * will accept. Now we'll ping the cache and see where it wraps.
+		 */
+
+		/* Write aaaaaaaa:aaaaaaaa to address 0 in the l2 cache.
+		 * If this "alias test" returns an "address", it means the
+		 * cache cannot be written to properly, and we have a problem.
+		 */
+		v = test_l2_address_alias(0, 0, 0xaaaaaaaa, 0xaaaaaaaa);
+		if (v != 0)
+			return -1;
+
+		/* Start with 32K wrap point (256KB actually) */
+		size = 1;
+		address = 0x8000;
+
+		while (1) {
+			v = test_l2_address_alias(address, 0, 0x55555555,
+						  0x55555555);
+			// Write failed.
+			if (v < 0)
+				return -1;
+			// It wraps here.
+			else if (v == 0)
+				break;
+
+			size <<= 1;
+			address <<= 1;
+
+			if (address > 0x200000)
+				return -1;
+		}
+
+		/* Mask size */
+		size &= 0x3e;
+
+		/* Shift to [17:13] */
+		size <<= 12;
+
+		/* Set this into BBL_CR_CTL3 */
+		msr = rdmsr(BBL_CR_CTL3);
+		msr.lo &= ~BBLCR3_L2_SIZE;
+		msr.lo |= size;
+		wrmsr(BBL_CR_CTL3, msr);
+
+		printk(BIOS_DEBUG,"L2 Cache Mask is %x\n", size);
+
+		/* Shift to [6:2] */
+		size >>= 11;
+
+		v = read_l2(2);
+
+		if (v < 0)
+			return -1;
+
+		printk(BIOS_DEBUG,"L2(2): %x ", v);
+
+		v &= 0x3;
+
+		/* Shift size right by v */
+		size >>= v;
+
+		/* Or in this size */
+		v |= size;
+
+		printk(BIOS_DEBUG,"-> %x\n", v);
+
+		if (write_l2(2, v) != 0)
+			return -1;
+	} else {
+		// Some cache size information is available from L2 registers.
+		// Work from there.
+		int b, c;
+
+		v = read_l2(2);
+
+		printk(BIOS_DEBUG,"L2(2) = %x\n", v);
+
+		if (v < 0)
+			return -1;
+
+		// L2 register 2 bitmap: cc---bbb
+		b = v & 0x7;
+		c = v >> 6;
+
+		v = 1 << c * b;
+
+		v &= 0xf;
+
+		printk(BIOS_DEBUG,"Calculated a = %x\n", v);
+
+		if (v == 0)
+			return -1;
+
+		/* Shift to 17:14 */
+		v <<= 14;
+
+		/* Write this size into BBL_CR_CTL3 */
+		msr = rdmsr(BBL_CR_CTL3);
+		msr.lo &= ~BBLCR3_L2_SIZE;
+		msr.lo |= v;
+		wrmsr(BBL_CR_CTL3, msr);
+	}
+
+	return 0;
+}
+
+// L2 physical address range can be found from L2 control register 3, bits [2:0].
+int calculate_l2_physical_address_range(void)
+{
+	int r0, r3;
+	msr_t msr;
+
+	r3 = read_l2(3);
+	if (r3 < 0)
+		return -1;
+
+	r0 = read_l2(0);
+	if (r0 < 0)
+		return -1;
+
+	if (r0 & 0x20)
+		r3 = 0x7;
+	else
+		r3 &= 0x7;
+
+	printk(BIOS_DEBUG,"L2 Physical Address Range is %dM\n", (1 << r3) * 512);
+
+	/* Shift into [22:20] to be saved into BBL_CR_CTL3. */
+	r3 = r3 << 20;
+
+	msr = rdmsr(BBL_CR_CTL3);
+	msr.lo &= ~BBLCR3_L2_PHYSICAL_RANGE;
+	msr.lo |= r3;
+	wrmsr(BBL_CR_CTL3, msr);
+
+	return 0;
+}
+
+int set_l2_ecc(void)
+{
+	u32 eax;
+	const u32 data1 = 0xaa55aa55;
+	const u32 data2 = 0xaaaaaaaa;
+	msr_t msr;
+
+	/* Set User Supplied ECC in BBL_CR_CTL */
+	msr = rdmsr(BBL_CR_CTL);
+	msr.lo |= BBLCR3_L2_SUPPLIED_ECC;
+	wrmsr(BBL_CR_CTL, msr);
+
+	/* Write a value into the L2 Data ECC register BBL_CR_DECC */
+	msr.lo = data1;
+	msr.hi = 0;
+	wrmsr(BBL_CR_DECC, msr);
+
+	if (test_l2_address_alias(0, 0, data2, data2) < 0)
+		return -1;
+
+	/* Read back ECC from BBL_CR_DECC */
+	msr = rdmsr(BBL_CR_DECC);
+	eax = msr.lo;
+
+	if (eax == data1) {
+		printk(BIOS_DEBUG,"L2 ECC Checking is enabled\n");
+
+		/* Set ECC Check Enable in BBL_CR_CTL3 */
+		msr = rdmsr(BBL_CR_CTL3);
+		msr.lo |= BBLCR3_L2_ECC_CHECK_ENABLE;
+		wrmsr(BBL_CR_CTL3, msr);
+	}
+
+	/* Clear User Supplied ECC in BBL_CR_CTL */
+	msr = rdmsr(BBL_CR_CTL);
+	msr.lo &= ~BBLCR3_L2_SUPPLIED_ECC;
+	wrmsr(BBL_CR_CTL, msr);
+
+	return 0;
+}
+
+/*
+ * This is the function called from CPU initialization
+ * driver to set up P6 family L2 cache.
+ */
+
+int p6_configure_l2_cache(void)
+{
+	msr_t msr, bblctl3;
+	unsigned int eax;
+	u16 signature;
+	int cache_size, bank;
+	int result, calc_eax;
+	int v, a;
+
+	int badclk1, badclk2, clkratio;
+	int crctl3_or;
+
+	printk(BIOS_INFO, "Configuring L2 cache... ");
+
+	/* Read BBL_CR_CTL3 */
+	bblctl3 = rdmsr(BBL_CR_CTL3);
+	/* If bit 23 (L2 Hardware disable) is set then done */
+	/* These would be Covington core Celerons with no L2 cache */
+	if (bblctl3.lo & BBLCR3_L2_NOT_PRESENT) {
+		printk(BIOS_INFO,"hardware disabled\n");
+		return 0;
+	}
+
+	signature = cpuid_eax(1) & 0xfff0;
+
+	/* Klamath-specific bit settings for certain
+	   preliminary checks.
+	 */
+	if (signature == 0x630) {
+		clkratio = 0x1c00000;
+		badclk2 = 0x1000000;
+		crctl3_or = 0x44000;
+	} else {
+		clkratio = 0x3c00000;
+		badclk2 = 0x3000000;
+		crctl3_or = 0x40000;
+	}
+	badclk1 = 0xc00000;
+
+	/* Read EBL_CR_POWERON */
+	msr = rdmsr(EBL_CR_POWERON);
+	eax = msr.lo;
+	/* Mask out [22-25] Clock frequency ratio */
+	eax &= clkratio;
+	if (eax == badclk1 || eax == badclk2) {
+		printk(BIOS_ERR, "Incorrect clock frequency ratio %x\n", eax);
+		return -1;
+	}
+
+	disable_cache();
+
+	/* Mask out from BBL_CR_CTL3:
+	 * [0] L2 Configured
+	 * [5] ECC Check Enable
+	 * [6] Address Parity Check Enable
+	 * [7] CRTN Parity Check Enable
+	 * [8] L2 Enabled
+	 * [12:11] Number of L2 banks
+	 * [17:13] Cache size per bank
+	 * [18] (Set below)
+	 * [22:20] L2 Physical Address Range Support
+	 */
+	bblctl3.lo &= 0xff88061e;
+	/* Set:
+	 * [17:13] = 00010 = 512Kbyte Cache size per bank (63x)
+	 * [17:13] = 00000 = 128Kbyte Cache size per bank (all others)
+	 * [18] Cache state error checking enable
+	 */
+	bblctl3.lo |= crctl3_or;
+
+	/* Write BBL_CR_CTL3 */
+	wrmsr(BBL_CR_CTL3, bblctl3);
+
+	if (signature != 0x630) {
+		eax = bblctl3.lo;
+
+		/* Set the l2 latency in BBL_CR_CTL3 */
+		if (calculate_l2_latency() != 0)
+			goto bad;
+
+		/* Read the new latency values back */
+		bblctl3 = rdmsr(BBL_CR_CTL3);
+		calc_eax = bblctl3.lo;
+
+		/* Write back the original default value */
+		bblctl3.lo = eax;
+		wrmsr(BBL_CR_CTL3, bblctl3);
+
+		/* Write BBL_CR_CTL3[27:26] (reserved??) to bits [1:0] of L2 register 4.
+		 * Apparently all other bits must be preserved, hence these code.
+		 */
+
+		v = (calc_eax >> 26) & 0x3;
+
+		printk(BIOS_DEBUG,"write_l2(4, %x)\n", v);
+
+		a = read_l2(4);
+		if (a >= 0)
+		{
+			a &= 0xfffc;
+			a |= v;
+			a = write_l2(4, a);
+			/* a now contains result code from write_l2() */
+		}
+		if (a != 0)
+			goto bad;
+
+		/* Restore the correct latency value into BBL_CR_CTL3 */
+		bblctl3.lo = calc_eax;
+		wrmsr(BBL_CR_CTL3, bblctl3);
+	} /* ! 63x CPU */
+
+	/* Read L2 register 0 */
+	v = read_l2(0);
+
+	/* If L2(0)[5] set (and can be read properly), enable CRTN and address parity
+	 */
+	if (v >= 0 && (v & 0x20)) {
+		bblctl3 = rdmsr(BBL_CR_CTL3);
+		bblctl3.lo |= (BBLCR3_L2_ADDR_PARITY_ENABLE |
+		               BBLCR3_L2_CRTN_PARITY_ENABLE);
+		wrmsr(BBL_CR_CTL3, bblctl3);
+	}
+
+	/* If something goes wrong at L2 ECC setup, cache ECC
+	 * will just remain disabled.
+	 */
+	set_l2_ecc();
+
+	if (calculate_l2_physical_address_range() != 0) {
+		printk(BIOS_ERR, "Failed to calculate L2 physical address range");
+		goto bad;
+	}
+
+	if (calculate_l2_cache_size() != 0) {
+		printk(BIOS_ERR, "Failed to calculate L2 cache size");
+		goto bad;
+	}
+
+	/* Turn on cache. Only L1 is active at this time. */
+	enable_cache();
+
+	/* Get the calculated cache size from BBL_CR_CTL3[17:13] */
+	bblctl3 = rdmsr(BBL_CR_CTL3);
+	cache_size = (bblctl3.lo & BBLCR3_L2_SIZE);
+	if (cache_size == 0)
+		cache_size = 0x1000;
+	cache_size = cache_size << 3;
+
+	/* TODO: Cache size above is per bank. We're supposed to get
+	 * the number of banks from BBL_CR_CTL3[12:11].
+	 * Confirm that this still provides the correct answer.
+	 */
+	bank = (bblctl3.lo >> 11) & 0x3;
+	if (bank == 0)
+		bank = 1;
+
+	printk(BIOS_INFO, "size %dK... ", cache_size * bank * 4 / 1024);
+
+	/* Write to all cache lines to initialize */
+
+	while (cache_size > 0) {
+
+		/* Each cache line is 32 bytes. */
+		cache_size -= 32;
+
+		/* Update each way */
+
+		/* We're supposed to get L2 associativity from BBL_CR_CTL3[10:9].
+		 * But this code only applies to certain members of the P6 processor family
+		 * and since all P6 processors have 4-way L2 cache, we can safely assume
+		 * 4 way for all cache operations.
+		 */
+
+		for (v = 0; v < 4; v++) {
+			/* Send Tag Write w/Data Write (TWW) to L2 controller
+			 * MESI = Invalid
+			 */
+			if (signal_l2(cache_size, 0, 0, v, L2CMD_TWW | L2CMD_MESI_I) != 0) {
+				printk(BIOS_ERR, "Failed on signal_l2(%x, %x)\n",
+				       cache_size, v);
+				goto bad;
+			}
+		}
+	}
+	printk(BIOS_DEBUG, "L2 Cache lines initialized\n");
+
+	/* Disable cache */
+	disable_cache();
+
+	/* Set L2 cache configured in BBL_CR_CTL3 */
+	bblctl3 = rdmsr(BBL_CR_CTL3);
+	bblctl3.lo |= BBLCR3_L2_CONFIGURED;
+	wrmsr(BBL_CR_CTL3, bblctl3);
+
+	/* Invalidate cache and discard unsaved writes */
+	asm volatile ("invd");
+
+	/* Write 0 to L2 control register 5 */
+	if (write_l2(5, 0) != 0) {
+		printk(BIOS_ERR,"write_l2(5, 0) failed\n");
+		goto done;
+	}
+
+	bblctl3 = rdmsr(BBL_CR_CTL3);
+	if (signature == 0x650) {
+		/* Change the L2 latency to 0101 then back to
+		 * original value. I don't know why this is needed - dpd
+		 */
+		eax = bblctl3.lo;
+		bblctl3.lo &= ~BBLCR3_L2_LATENCY;
+		bblctl3.lo |= 0x0a;
+		wrmsr(BBL_CR_CTL3, bblctl3);
+		bblctl3.lo = eax;
+		wrmsr(BBL_CR_CTL3, bblctl3);
+	}
+
+	/* Enable L2 in BBL_CR_CTL3 */
+	bblctl3.lo |= BBLCR3_L2_ENABLED;
+	wrmsr(BBL_CR_CTL3, bblctl3);
+
+	/* Turn on cache. Both L1 and L2 are now active. Wahoo! */
+done:
+	result = 0;
+	goto out;
+bad:
+	result = -1;
+out:
+	printk(BIOS_INFO, "done.\n");
+	return result;
+}
+