x86: parallel MP initialization

Provide a common entry point for bringing up the APs
in parallel. This work is based off of the Haswell one
which can be moved over to this in the future. The APs
are brought up and have the BSP's MTRRs duplicated in
their own MTRRs. Additionally, Microcode is loaded before
enabling caching. However, the current microcode loading
support assumes Intel's mechanism.

The infrastructure provides a notion of a flight plan
for the BSP and APs. This allows for flexibility in the
order of operations for a given architecture/chip without
providing any specific policy. Therefore, the chipset
caller can provide the order that is required.

BUG=chrome-os-partner:22862
BRANCH=None
TEST=Built and booted on rambi with baytrail specific patches.

Change-Id: I0539047a1b24c13ef278695737cdba3b9344c820
Signed-off-by: Aaron Durbin <adurbin@chromium.org>
Reviewed-on: https://chromium-review.googlesource.com/173703
Reviewed-on: http://review.coreboot.org/4888
Reviewed-by: Alexandru Gagniuc <mr.nuke.me@gmail.com>
Tested-by: build bot (Jenkins)

5 files changed, 816 insertions, 0 deletions

diff --git a/src/cpu/x86/Kconfig b/src/cpu/x86/Kconfig
index c64a8e477b..21282c321a 100644
--- a/src/cpu/x86/Kconfig
+++ b/src/cpu/x86/Kconfig
@@ -115,3 +115,10 @@ config X86_AMD_FIXED_MTRRS
 	help
 	  This option informs the MTRR code to use the RdMem and WrMem fields
 	  in the fixed MTRR MSRs.
+
+config PARALLEL_MP
+	def_bool n
+	help
+	 This option uses common MP infrastructure for bringing up APs
+	 in parallel. It additionally provides a more flexible mechanism
+	 for sequencing the steps of bringing up the APs.
diff --git a/src/cpu/x86/Makefile.inc b/src/cpu/x86/Makefile.inc
index ca755dec5d..d5bc2fd219 100644
--- a/src/cpu/x86/Makefile.inc
+++ b/src/cpu/x86/Makefile.inc
@@ -1,2 +1,27 @@
 romstage-$(CONFIG_EARLY_CBMEM_INIT) += car.c
 romstage-$(CONFIG_HAVE_ACPI_RESUME) += car.c
+
+subdirs-$(CONFIG_PARALLEL_MP) += name
+ramstage-$(CONFIG_PARALLEL_MP) += mp_init.c
+
+SIPI_ELF=$(obj)/cpu/x86/sipi_vector.elf
+SIPI_BIN=$(SIPI_ELF:.elf=)
+SIPI_DOTO=$(SIPI_ELF:.elf=.o)
+
+ifeq ($(CONFIG_PARALLEL_MP),y)
+ramstage-srcs += $(SIPI_BIN)
+endif
+rmodules-$(CONFIG_PARALLEL_MP) += sipi_vector.S
+rmodules-$(CONFIG_PARALLEL_MP) += sipi_header.c
+
+$(SIPI_DOTO): $(dir $(SIPI_ELF))sipi_vector.rmodules.o $(dir $(SIPI_ELF))sipi_header.rmodules.o
+	$(CC) $(LDFLAGS) -nostdlib -r -o $@ $^
+
+$(eval $(call rmodule_link,$(SIPI_ELF), $(SIPI_ELF:.elf=.o), 0))
+
+$(SIPI_BIN): $(SIPI_ELF)
+	$(OBJCOPY) -O binary $< $@
+
+$(SIPI_BIN).ramstage.o: $(SIPI_BIN)
+	@printf "    OBJCOPY    $(subst $(obj)/,,$(@))\n"
+	cd $(dir $@); $(OBJCOPY) -I binary $(notdir $<) -O elf32-i386 -B i386 $(notdir $@)
diff --git a/src/cpu/x86/mp_init.c b/src/cpu/x86/mp_init.c
new file mode 100644
index 0000000000..242d656144
--- /dev/null
+++ b/src/cpu/x86/mp_init.c
@@ -0,0 +1,586 @@
+/*
+ * This file is part of the coreboot project.
+ *
+ * Copyright (C) 2013 Google Inc.
+ *
+ * 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; version 2 of
+ * the License.
+ *
+ * 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
+ */
+
+#include <console/console.h>
+#include <stdint.h>
+#include <rmodule.h>
+#include <arch/cpu.h>
+#include <cpu/cpu.h>
+#include <cpu/intel/microcode.h>
+#include <cpu/x86/cache.h>
+#include <cpu/x86/lapic.h>
+#include <cpu/x86/name.h>
+#include <cpu/x86/msr.h>
+#include <cpu/x86/mtrr.h>
+#include <cpu/x86/smm.h>
+#include <cpu/x86/mp.h>
+#include <delay.h>
+#include <device/device.h>
+#include <device/path.h>
+#include <lib.h>
+#include <smp/atomic.h>
+#include <smp/spinlock.h>
+#include <thread.h>
+
+#define MAX_APIC_IDS 256
+/* This needs to match the layout in the .module_parametrs section. */
+struct sipi_params {
+	uint16_t gdtlimit;
+	uint32_t gdt;
+	uint16_t unused;
+	uint32_t idt_ptr;
+	uint32_t stack_top;
+	uint32_t stack_size;
+	uint32_t microcode_lock; /* 0xffffffff means parallel loading. */
+	uint32_t microcode_ptr;
+	uint32_t msr_table_ptr;
+	uint32_t msr_count;
+	uint32_t c_handler;
+	atomic_t ap_count;
+} __attribute__((packed));
+
+/* This also needs to match the assembly code for saved MSR encoding. */
+struct saved_msr {
+	uint32_t index;
+	uint32_t lo;
+	uint32_t hi;
+} __attribute__((packed));
+
+
+/* The sipi vector rmodule is included in the ramstage using 'objdump -B'. */
+extern char _binary_sipi_vector_start[];
+/* These symbols are defined in c_start.S. */
+extern char gdt[];
+extern char gdt_end[];
+extern char idtarg[];
+
+/* The SIPI vector is loaded at the SMM_DEFAULT_BASE. The reason is at the
+ * memory range is already reserved so the OS cannot use it. That region is
+ * free to use for AP bringup before SMM is initialized. */
+static const uint32_t sipi_vector_location = SMM_DEFAULT_BASE;
+static const int sipi_vector_location_size = SMM_DEFAULT_SIZE;
+
+struct mp_flight_plan {
+	int num_records;
+	struct mp_flight_record *records;
+};
+
+static struct mp_flight_plan mp_info;
+
+struct cpu_map {
+	device_t dev;
+	int apic_id;
+};
+
+/* Keep track of apic and device structure for each cpu. */
+static struct cpu_map cpus[CONFIG_MAX_CPUS];
+
+static inline void barrier_wait(atomic_t *b)
+{
+	while (atomic_read(b) == 0) {
+		asm ("pause");
+	}
+	mfence();
+}
+
+static inline void release_barrier(atomic_t *b)
+{
+	mfence();
+	atomic_set(b, 1);
+}
+
+/* Returns 1 if timeout waiting for APs. 0 if target aps found. */
+static int wait_for_aps(atomic_t *val, int target, int total_delay,
+                        int delay_step)
+{
+	int timeout = 0;
+	int delayed = 0;
+	while (atomic_read(val) != target) {
+		udelay(delay_step);
+		delayed += delay_step;
+		if (delayed >= total_delay) {
+			timeout = 1;
+			break;
+		}
+	}
+
+	return timeout;
+}
+
+static void ap_do_flight_plan(void)
+{
+	int i;
+
+	for (i = 0; i < mp_info.num_records; i++) {
+		struct mp_flight_record *rec = &mp_info.records[i];
+
+		atomic_inc(&rec->cpus_entered);
+		barrier_wait(&rec->barrier);
+
+		if (rec->ap_call != NULL) {
+			rec->ap_call(rec->ap_arg);
+		}
+	}
+}
+
+/* By the time APs call ap_init() caching has been setup, and microcode has
+ * been loaded. */
+static void asmlinkage ap_init(unsigned int cpu)
+{
+	struct cpu_info *info;
+	int apic_id;
+
+	/* Ensure the local apic is enabled */
+	enable_lapic();
+
+	info = cpu_info();
+	info->index = cpu;
+	info->cpu = cpus[cpu].dev;
+	thread_init_cpu_info_non_bsp(info);
+
+	apic_id = lapicid();
+	info->cpu->path.apic.apic_id = apic_id;
+	cpus[cpu].apic_id = apic_id;
+
+	printk(BIOS_INFO, "AP: slot %d apic_id %x.\n", cpu, apic_id);
+
+	/* Walk the flight plan */
+	ap_do_flight_plan();
+
+	/* Park the AP. */
+	stop_this_cpu();
+}
+
+static void setup_default_sipi_vector_params(struct sipi_params *sp)
+{
+	sp->gdt = (uint32_t)&gdt;
+	sp->gdtlimit = (uint32_t)&gdt_end - (u32)&gdt - 1;
+	sp->idt_ptr = (uint32_t)&idtarg;
+	sp->stack_size = CONFIG_STACK_SIZE;
+	sp->stack_top = (uint32_t)&_estack;
+	/* Adjust the stack top to take into account cpu_info. */
+	sp->stack_top -= sizeof(struct cpu_info);
+}
+
+#define NUM_FIXED_MTRRS 11
+static const unsigned int fixed_mtrrs[NUM_FIXED_MTRRS] = {
+	MTRRfix64K_00000_MSR, MTRRfix16K_80000_MSR, MTRRfix16K_A0000_MSR,
+	MTRRfix4K_C0000_MSR, MTRRfix4K_C8000_MSR, MTRRfix4K_D0000_MSR,
+	MTRRfix4K_D8000_MSR, MTRRfix4K_E0000_MSR, MTRRfix4K_E8000_MSR,
+	MTRRfix4K_F0000_MSR, MTRRfix4K_F8000_MSR,
+};
+
+static inline struct saved_msr *save_msr(int index, struct saved_msr *entry)
+{
+	msr_t msr;
+
+	msr = rdmsr(index);
+	entry->index = index;
+	entry->lo = msr.lo;
+	entry->hi = msr.hi;
+
+	/* Return the next entry. */
+	entry++;
+	return entry;
+}
+
+static int save_bsp_msrs(char *start, int size)
+{
+	int msr_count;
+	int num_var_mtrrs;
+	struct saved_msr *msr_entry;
+	int i;
+	msr_t msr;
+
+	/* Determine number of MTRRs need to be saved. */
+	msr = rdmsr(MTRRcap_MSR);
+	num_var_mtrrs = msr.lo & 0xff;
+
+	/* 2 * num_var_mtrrs for base and mask. +1 for IA32_MTRR_DEF_TYPE. */
+	msr_count = 2 * num_var_mtrrs + NUM_FIXED_MTRRS + 1;
+
+	if ((msr_count * sizeof(struct saved_msr)) > size) {
+		printk(BIOS_CRIT, "Cannot mirror all %d msrs.\n", msr_count);
+		return -1;
+	}
+
+	msr_entry = (void *)start;
+	for (i = 0; i < NUM_FIXED_MTRRS; i++) {
+		msr_entry = save_msr(fixed_mtrrs[i], msr_entry);
+	}
+
+	for (i = 0; i < num_var_mtrrs; i++) {
+		msr_entry = save_msr(MTRRphysBase_MSR(i), msr_entry);
+		msr_entry = save_msr(MTRRphysMask_MSR(i), msr_entry);
+	}
+
+	msr_entry = save_msr(MTRRdefType_MSR, msr_entry);
+
+	return msr_count;
+}
+
+static atomic_t *load_sipi_vector(struct mp_params *mp_params)
+{
+	struct rmodule sipi_mod;
+	int module_size;
+	int num_msrs;
+	struct sipi_params *sp;
+	char *mod_loc = (void *)sipi_vector_location;
+	const int loc_size = sipi_vector_location_size;
+	atomic_t *ap_count = NULL;
+
+	if (rmodule_parse(&_binary_sipi_vector_start, &sipi_mod)) {
+		printk(BIOS_CRIT, "Unable to parse sipi module.\n");
+		return ap_count;
+	}
+
+	if (rmodule_entry_offset(&sipi_mod) != 0) {
+		printk(BIOS_CRIT, "SIPI module entry offset is not 0!\n");
+		return ap_count;
+	}
+
+	if (rmodule_load_alignment(&sipi_mod) != 4096) {
+		printk(BIOS_CRIT, "SIPI module load alignment(%d) != 4096.\n",
+		       rmodule_load_alignment(&sipi_mod));
+		return ap_count;
+	}
+
+	module_size = rmodule_memory_size(&sipi_mod);
+
+	/* Align to 4 bytes. */
+	module_size = ALIGN(module_size, 4);
+
+	if (module_size > loc_size) {
+		printk(BIOS_CRIT, "SIPI module size (%d) > region size (%d).\n",
+		       module_size, loc_size);
+		return ap_count;
+	}
+
+	num_msrs = save_bsp_msrs(&mod_loc[module_size], loc_size - module_size);
+
+	if (num_msrs < 0) {
+		printk(BIOS_CRIT, "Error mirroring BSP's msrs.\n");
+		return ap_count;
+	}
+
+	if (rmodule_load(mod_loc, &sipi_mod)) {
+		printk(BIOS_CRIT, "Unable to load SIPI module.\n");
+		return ap_count;
+	}
+
+	sp = rmodule_parameters(&sipi_mod);
+
+	if (sp == NULL) {
+		printk(BIOS_CRIT, "SIPI module has no parameters.\n");
+		return ap_count;
+	}
+
+	setup_default_sipi_vector_params(sp);
+	/* Setup MSR table. */
+	sp->msr_table_ptr = (uint32_t)&mod_loc[module_size];
+	sp->msr_count = num_msrs;
+	/* Provide pointer to microcode patch. */
+	sp->microcode_ptr = (uint32_t)mp_params->microcode_pointer;
+	/* Pass on abiility to load microcode in parallel. */
+	if (mp_params->parallel_microcode_load) {
+		sp->microcode_lock = 0;
+	} else {
+		sp->microcode_lock = ~0;
+	}
+	sp->c_handler = (uint32_t)&ap_init;
+	ap_count = &sp->ap_count;
+	atomic_set(ap_count, 0);
+
+	return ap_count;
+}
+
+static int allocate_cpu_devices(struct bus *cpu_bus, struct mp_params *p)
+{
+	int i;
+	int max_cpus;
+	struct cpu_info *info;
+
+	max_cpus = p->num_cpus;
+	if (max_cpus > CONFIG_MAX_CPUS) {
+		printk(BIOS_CRIT, "CPU count(%d) exceeds CONFIG_MAX_CPUS(%d)\n",
+		       max_cpus, CONFIG_MAX_CPUS);
+		max_cpus = CONFIG_MAX_CPUS;
+	}
+
+	info = cpu_info();
+	for (i = 1; i < max_cpus; i++) {
+		struct device_path cpu_path;
+		device_t new;
+		int apic_id;
+
+		/* Build the cpu device path */
+		cpu_path.type = DEVICE_PATH_APIC;
+
+		/* Assuming linear APIC space allocation. */
+		apic_id = info->cpu->path.apic.apic_id + i;
+		if (p->adjust_apic_id != NULL) {
+			apic_id = p->adjust_apic_id(i, apic_id);
+		}
+		cpu_path.apic.apic_id = apic_id;
+
+		/* Allocate the new cpu device structure */
+		new = alloc_find_dev(cpu_bus, &cpu_path);
+		if (new == NULL) {
+			printk(BIOS_CRIT, "Could not allocte cpu device\n");
+			max_cpus--;
+		}
+		cpus[i].dev = new;
+	}
+
+	return max_cpus;
+}
+
+/* Returns 1 for timeout. 0 on success. */
+static int apic_wait_timeout(int total_delay, int delay_step)
+{
+	int total = 0;
+	int timeout = 0;
+
+	while (lapic_read(LAPIC_ICR) & LAPIC_ICR_BUSY) {
+		udelay(delay_step);
+		total += delay_step;
+		if (total >= total_delay) {
+			timeout = 1;
+			break;
+		}
+	}
+
+	return timeout;
+}
+
+static int start_aps(struct bus *cpu_bus, int ap_count, atomic_t *num_aps)
+{
+	int sipi_vector;
+	/* Max location is 4KiB below 1MiB */
+	const int max_vector_loc = ((1 << 20) - (1 << 12)) >> 12;
+
+	if (ap_count == 0)
+		return 0;
+
+	/* The vector is sent as a 4k aligned address in one byte. */
+	sipi_vector = sipi_vector_location >> 12;
+
+	if (sipi_vector > max_vector_loc) {
+		printk(BIOS_CRIT, "SIPI vector too large! 0x%08x\n",
+		       sipi_vector);
+		return -1;
+	}
+
+	printk(BIOS_DEBUG, "Attempting to start %d APs\n", ap_count);
+
+	if ((lapic_read(LAPIC_ICR) & LAPIC_ICR_BUSY)) {
+		printk(BIOS_DEBUG, "Waiting for ICR not to be busy...");
+		if (apic_wait_timeout(1000 /* 1 ms */, 50)) {
+			printk(BIOS_DEBUG, "timed out. Aborting.\n");
+			return -1;
+		} else
+			printk(BIOS_DEBUG, "done.\n");
+	}
+
+	/* Send INIT IPI to all but self. */
+	lapic_write_around(LAPIC_ICR2, SET_LAPIC_DEST_FIELD(0));
+	lapic_write_around(LAPIC_ICR, LAPIC_DEST_ALLBUT | LAPIC_INT_ASSERT |
+	                   LAPIC_DM_INIT);
+	printk(BIOS_DEBUG, "Waiting for 10ms after sending INIT.\n");
+	mdelay(10);
+
+	/* Send 1st SIPI */
+	if ((lapic_read(LAPIC_ICR) & LAPIC_ICR_BUSY)) {
+		printk(BIOS_DEBUG, "Waiting for ICR not to be busy...");
+		if (apic_wait_timeout(1000 /* 1 ms */, 50)) {
+			printk(BIOS_DEBUG, "timed out. Aborting.\n");
+			return -1;
+		} else
+			printk(BIOS_DEBUG, "done.\n");
+	}
+
+	lapic_write_around(LAPIC_ICR2, SET_LAPIC_DEST_FIELD(0));
+	lapic_write_around(LAPIC_ICR, LAPIC_DEST_ALLBUT | LAPIC_INT_ASSERT |
+	                   LAPIC_DM_STARTUP | sipi_vector);
+	printk(BIOS_DEBUG, "Waiting for 1st SIPI to complete...");
+	if (apic_wait_timeout(10000 /* 10 ms */, 50 /* us */)) {
+		printk(BIOS_DEBUG, "timed out.\n");
+		return -1;
+	} else {
+		printk(BIOS_DEBUG, "done.\n");
+	}
+
+	/* Wait for CPUs to check in up to 200 us. */
+	wait_for_aps(num_aps, ap_count, 200 /* us */, 15 /* us */);
+
+	/* Send 2nd SIPI */
+	if ((lapic_read(LAPIC_ICR) & LAPIC_ICR_BUSY)) {
+		printk(BIOS_DEBUG, "Waiting for ICR not to be busy...");
+		if (apic_wait_timeout(1000 /* 1 ms */, 50)) {
+			printk(BIOS_DEBUG, "timed out. Aborting.\n");
+			return -1;
+		} else
+			printk(BIOS_DEBUG, "done.\n");
+	}
+
+	lapic_write_around(LAPIC_ICR2, SET_LAPIC_DEST_FIELD(0));
+	lapic_write_around(LAPIC_ICR, LAPIC_DEST_ALLBUT | LAPIC_INT_ASSERT |
+	                   LAPIC_DM_STARTUP | sipi_vector);
+	printk(BIOS_DEBUG, "Waiting for 2nd SIPI to complete...");
+	if (apic_wait_timeout(10000 /* 10 ms */, 50 /* us */)) {
+		printk(BIOS_DEBUG, "timed out.\n");
+		return -1;
+	} else {
+		printk(BIOS_DEBUG, "done.\n");
+	}
+
+	/* Wait for CPUs to check in. */
+	if (wait_for_aps(num_aps, ap_count, 10000 /* 10 ms */, 50 /* us */)) {
+		printk(BIOS_DEBUG, "Not all APs checked in: %d/%d.\n",
+		       atomic_read(num_aps), ap_count);
+		return -1;
+	}
+
+	return 0;
+}
+
+static int bsp_do_flight_plan(struct mp_params *mp_params)
+{
+	int i;
+	int ret = 0;
+	const int timeout_us = 100000;
+	const int step_us = 100;
+	int num_aps = mp_params->num_cpus - 1;
+
+	for (i = 0; i < mp_params->num_records; i++) {
+		struct mp_flight_record *rec = &mp_params->flight_plan[i];
+
+		/* Wait for APs if the record is not released. */
+		if (atomic_read(&rec->barrier) == 0) {
+			/* Wait for the APs to check in. */
+			if (wait_for_aps(&rec->cpus_entered, num_aps,
+			                 timeout_us, step_us)) {
+				printk(BIOS_ERR, "MP record %d timeout.\n", i);
+				ret = -1;
+			}
+		}
+
+		if (rec->bsp_call != NULL) {
+			rec->bsp_call(rec->bsp_arg);
+		}
+
+		release_barrier(&rec->barrier);
+	}
+	return ret;
+}
+
+static void init_bsp(struct bus *cpu_bus)
+{
+	struct device_path cpu_path;
+	struct cpu_info *info;
+	char processor_name[49];
+
+	/* Print processor name */
+	fill_processor_name(processor_name);
+	printk(BIOS_INFO, "CPU: %s.\n", processor_name);
+
+	/* Ensure the local apic is enabled */
+	enable_lapic();
+
+	/* Set the device path of the boot cpu. */
+	cpu_path.type = DEVICE_PATH_APIC;
+	cpu_path.apic.apic_id = lapicid();
+
+	/* Find the device structure for the boot cpu. */
+	info = cpu_info();
+	info->cpu = alloc_find_dev(cpu_bus, &cpu_path);
+
+	if (info->index != 0)
+		printk(BIOS_CRIT, "BSP index(%d) != 0!\n", info->index);
+
+	/* Track BSP in cpu_map structures. */
+	cpus[info->index].dev = info->cpu;
+	cpus[info->index].apic_id = cpu_path.apic.apic_id;
+}
+
+int mp_init(struct bus *cpu_bus, struct mp_params *p)
+{
+	int num_cpus;
+	int num_aps;
+	atomic_t *ap_count;
+
+	init_bsp(cpu_bus);
+
+	if (p == NULL || p->flight_plan == NULL || p->num_records < 1) {
+		printk(BIOS_CRIT, "Invalid MP parameters\n");
+		return -1;
+	}
+
+	/* Default to currently running CPU. */
+	num_cpus = allocate_cpu_devices(cpu_bus, p);
+
+	if (num_cpus < p->num_cpus) {
+		printk(BIOS_CRIT,
+		       "ERROR: More cpus requested (%d) than supported (%d).\n",
+		       p->num_cpus, num_cpus);
+		return -1;
+	}
+
+	/* Copy needed parameters so that APs have a reference to the plan. */
+	mp_info.num_records = p->num_records;
+	mp_info.records = p->flight_plan;
+
+	/* Load the SIPI vector. */
+	ap_count = load_sipi_vector(p);
+	if (ap_count == NULL)
+		return -1;
+
+	/* Make sure SIPI data hits RAM so the APs that come up will see
+	 * the startup code even if the caches are disabled.  */
+	wbinvd();
+
+	/* Start the APs providing number of APs and the cpus_entered field. */
+	num_aps = p->num_cpus - 1;
+	if (start_aps(cpu_bus, num_aps, ap_count) < 0) {
+		mdelay(1000);
+		printk(BIOS_DEBUG, "%d/%d eventually checked in?\n",
+		       atomic_read(ap_count), num_aps);
+		return -1;
+	}
+
+	/* Walk the flight plan for the BSP. */
+	return bsp_do_flight_plan(p);
+}
+
+void mp_initialize_cpu(void *unused)
+{
+	/* Call back into driver infrastructure for the AP initialization.   */
+	struct cpu_info *info = cpu_info();
+	cpu_initialize(info->index);
+}
+
+int mp_get_apic_id(int cpu_slot)
+{
+	if (cpu_slot >= CONFIG_MAX_CPUS || cpu_slot < 0)
+		return -1;
+
+	return cpus[cpu_slot].apic_id;
+}
diff --git a/src/cpu/x86/sipi_header.c b/src/cpu/x86/sipi_header.c
new file mode 100644
index 0000000000..846a82d7c2
--- /dev/null
+++ b/src/cpu/x86/sipi_header.c
@@ -0,0 +1,6 @@
+#include <rmodule.h>
+
+
+extern void *ap_start;
+
+DEFINE_RMODULE_HEADER(sipi_vector_header, ap_start, RMODULE_TYPE_SIPI_VECTOR);
diff --git a/src/cpu/x86/sipi_vector.S b/src/cpu/x86/sipi_vector.S
new file mode 100644
index 0000000000..52b12d0314
--- /dev/null
+++ b/src/cpu/x86/sipi_vector.S
@@ -0,0 +1,192 @@
+/*
+ * This file is part of the coreboot project.
+ *
+ * Copyright (C) 2013 Google Inc.
+ *
+ * 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; version 2 of
+ * the License.
+ *
+ * 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
+ */
+
+/* The SIPI vector is responsible for initializing the APs in the sytem. It
+ * loads microcode, sets up MSRs, and enables caching before calling into
+ * C code. */
+
+/* These segment selectors need to match the gdt entries in c_start.S. */
+#define CODE_SEG 0x10
+#define DATA_SEG 0x18
+
+#define IA32_UPDT_TRIG 0x79
+#define IA32_BIOS_SIGN_ID 0x8b
+
+.section ".module_parameters", "aw", @progbits
+ap_start_params:
+gdtaddr:
+.word 0 /* limit */
+.long 0 /* table */
+.word 0 /* unused */
+idt_ptr:
+.long 0
+stack_top:
+.long 0
+stack_size:
+.long 0
+microcode_lock:
+.long 0
+microcode_ptr:
+.long 0
+msr_table_ptr:
+.long 0
+msr_count:
+.long 0
+c_handler:
+.long 0
+ap_count:
+.long 0
+
+.text
+.code16
+.global ap_start
+ap_start:
+	cli
+	xorl	%eax, %eax
+	movl	%eax, %cr3    /* Invalidate TLB*/
+
+	/* On hyper threaded cpus, invalidating the cache here is
+	 * very very bad.  Don't.
+	 */
+
+	/* setup the data segment */
+	movw	%cs, %ax
+	movw	%ax, %ds
+
+	/* The gdtaddr needs to be releative to the data segment in order
+	 * to properly dereference it. The .text section comes first in an
+	 * rmodule so ap_start can be used as a proxy for the load address. */
+	movl	$(gdtaddr), %ebx
+	sub	$(ap_start), %ebx
+
+	data32 lgdt (%ebx)
+
+	movl	%cr0, %eax
+	andl	$0x7FFAFFD1, %eax /* PG,AM,WP,NE,TS,EM,MP = 0 */
+	orl	$0x60000001, %eax /* CD, NW, PE = 1 */
+	movl	%eax, %cr0
+
+	ljmpl	$CODE_SEG, $1f
+1:
+	.code32
+	movw	$DATA_SEG, %ax
+	movw	%ax, %ds
+	movw	%ax, %es
+	movw	%ax, %ss
+	movw	%ax, %fs
+	movw	%ax, %gs
+
+	/* Load the Interrupt descriptor table */
+	mov	idt_ptr, %ebx
+	lidt	(%ebx)
+
+	/* Obtain cpu number. */
+	movl	ap_count, %eax
+1:
+	movl	%eax, %ecx
+	inc	%ecx
+	lock cmpxchg %ecx, ap_count
+	jnz	1b
+
+	/* Setup stacks for each CPU. */
+	movl	stack_size, %eax
+	mul	%ecx
+	movl	stack_top, %edx
+	subl	%eax, %edx
+	mov	%edx, %esp
+	/* Save cpu number. */
+	mov	%ecx, %esi
+
+	/* Determine if one should check microcode versions. */
+	mov	microcode_ptr, %edi
+	test	%edi, %edi
+	jz	microcode_done /* Bypass if no microde exists. */
+
+	/* Get the Microcode version. */
+	mov	$1, %eax
+	cpuid
+	mov	$IA32_BIOS_SIGN_ID, %ecx
+	rdmsr
+	/* If something already loaded skip loading again. */
+	test	%edx, %edx
+	jnz	microcode_done
+
+	/* Determine if parallel microcode loading is allowed. */
+	cmp	$0xffffffff, microcode_lock
+	je	load_microcode
+
+	/* Protect microcode loading. */
+lock_microcode:
+	lock bts $0, microcode_lock
+	jc	lock_microcode
+
+load_microcode:
+	/* Load new microcode. */
+	mov	$IA32_UPDT_TRIG, %ecx
+	xor	%edx, %edx
+	mov	%edi, %eax
+	/* The microcode pointer is passed in pointing to the header. Adjust
+	 * pointer to reflect the payload (header size is 48 bytes). */
+	add	$48, %eax
+	pusha
+	wrmsr
+	popa
+
+	/* Unconditionally unlock microcode loading. */
+	cmp	$0xffffffff, microcode_lock
+	je	microcode_done
+
+	xor	%eax, %eax
+	mov	%eax, microcode_lock
+
+microcode_done:
+	/*
+	 * Load MSRs. Each entry in the table consists of:
+	 * 0: index,
+	 * 4: value[31:0]
+	 * 8: value[63:32]
+	 */
+	mov	msr_table_ptr, %edi
+	mov	msr_count, %ebx
+	test	%ebx, %ebx
+	jz	1f
+load_msr:
+	mov	(%edi), %ecx
+	mov	4(%edi), %eax
+	mov	8(%edi), %edx
+	wrmsr
+	add	$12, %edi
+	dec	%ebx
+	jnz	load_msr
+
+1:
+	/* Enable caching. */
+	mov	%cr0, %eax
+	and	$0x9fffffff, %eax /* CD, NW = 0 */
+	mov	%eax, %cr0
+
+	/* c_handler(cpu_num) */
+	push	%esi	/* cpu_num */
+	mov	c_handler, %eax
+	call	*%eax
+halt_jump:
+	hlt
+	jmp	halt_jump