x86: Add multi-processor init

Most modern x86 CPUs include more than one CPU core. The OS normally requires
that these 'Application Processors' (APs) be brought up by the boot loader.
Add the required support to U-Boot to init additional APs.

Signed-off-by: Simon Glass <sjg@chromium.org>
Reviewed-by: Bin Meng <bmeng.cn@gmail.com>

5 files changed, 717 insertions, 2 deletions

diff --git a/arch/x86/cpu/Makefile b/arch/x86/cpu/Makefile
index 6ded0a7..043bea2 100644
--- a/arch/x86/cpu/Makefile
+++ b/arch/x86/cpu/Makefile
@@ -19,6 +19,8 @@ obj-$(CONFIG_NORTHBRIDGE_INTEL_IVYBRIDGE) += ivybridge/
 obj-$(CONFIG_INTEL_QUARK) += quark/
 obj-$(CONFIG_INTEL_QUEENSBAY) += queensbay/
 obj-y += lapic.o
+obj-$(CONFIG_SMP) += mp_init.o
 obj-y += mtrr.o
 obj-$(CONFIG_PCI) += pci.o
+obj-$(CONFIG_SMP) += sipi_vector.o
 obj-y += turbo.o
diff --git a/arch/x86/cpu/ivybridge/car.S b/arch/x86/cpu/ivybridge/car.S
index 9441666..407e451 100644
--- a/arch/x86/cpu/ivybridge/car.S
+++ b/arch/x86/cpu/ivybridge/car.S
@@ -15,6 +15,7 @@
 #include <asm/msr-index.h>
 #include <asm/mtrr.h>
 #include <asm/post.h>
+#include <asm/processor.h>
 #include <asm/processor-flags.h>
 #include <asm/arch/microcode.h>
 
diff --git a/arch/x86/cpu/ivybridge/model_206ax.c b/arch/x86/cpu/ivybridge/model_206ax.c
index 11dc625..8b08c40 100644
--- a/arch/x86/cpu/ivybridge/model_206ax.c
+++ b/arch/x86/cpu/ivybridge/model_206ax.c
@@ -435,8 +435,8 @@ static int intel_cores_init(struct x86_cpu_priv *cpu)
 
 		debug("CPU: %u has core %u\n", cpu->apic_id, new_cpu->apic_id);
 
-#if CONFIG_SMP && CONFIG_MAX_CPUS > 1
-		/* Start the new cpu */
+#if 0 && CONFIG_SMP && CONFIG_MAX_CPUS > 1
+		/* TODO(sjg@chromium.org): Start the new cpu */
 		if (!start_cpu(new_cpu)) {
 			/* Record the error in cpu? */
 			printk(BIOS_ERR, "CPU %u would not start!\n",
diff --git a/arch/x86/cpu/mp_init.c b/arch/x86/cpu/mp_init.c
new file mode 100644
index 0000000..ac5753a
--- /dev/null
+++ b/arch/x86/cpu/mp_init.c
@@ -0,0 +1,496 @@
+/*
+ * Copyright (C) 2015 Google, Inc
+ *
+ * SPDX-License-Identifier:	GPL-2.0+
+ *
+ * Based on code from the coreboot file of the same name
+ */
+
+#include <common.h>
+#include <cpu.h>
+#include <dm.h>
+#include <errno.h>
+#include <malloc.h>
+#include <asm/atomic.h>
+#include <asm/cpu.h>
+#include <asm/interrupt.h>
+#include <asm/lapic.h>
+#include <asm/mp.h>
+#include <asm/mtrr.h>
+#include <asm/sipi.h>
+#include <dm/device-internal.h>
+#include <dm/uclass-internal.h>
+#include <linux/linkage.h>
+
+/* This also needs to match the sipi.S assembly code for saved MSR encoding */
+struct saved_msr {
+	uint32_t index;
+	uint32_t lo;
+	uint32_t hi;
+} __packed;
+
+
+struct mp_flight_plan {
+	int num_records;
+	struct mp_flight_record *records;
+};
+
+static struct mp_flight_plan mp_info;
+
+struct cpu_map {
+	struct udevice *dev;
+	int apic_id;
+	int err_code;
+};
+
+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(struct udevice *cpu)
+{
+	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(cpu, rec->ap_arg);
+	}
+}
+
+static int find_cpu_by_apid_id(int apic_id, struct udevice **devp)
+{
+	struct udevice *dev;
+
+	*devp = NULL;
+	for (uclass_find_first_device(UCLASS_CPU, &dev);
+	     dev;
+	     uclass_find_next_device(&dev)) {
+		struct cpu_platdata *plat = dev_get_parent_platdata(dev);
+
+		if (plat->cpu_id == apic_id) {
+			*devp = dev;
+			return 0;
+		}
+	}
+
+	return -ENOENT;
+}
+
+/*
+ * By the time APs call ap_init() caching has been setup, and microcode has
+ * been loaded
+ */
+static void ap_init(unsigned int cpu_index)
+{
+	struct udevice *dev;
+	int apic_id;
+	int ret;
+
+	/* Ensure the local apic is enabled */
+	enable_lapic();
+
+	apic_id = lapicid();
+	ret = find_cpu_by_apid_id(apic_id, &dev);
+	if (ret) {
+		debug("Unknown CPU apic_id %x\n", apic_id);
+		goto done;
+	}
+
+	debug("AP: slot %d apic_id %x, dev %s\n", cpu_index, apic_id,
+	      dev ? dev->name : "(apic_id not found)");
+
+	/* Walk the flight plan */
+	ap_do_flight_plan(dev);
+
+	/* Park the AP */
+	debug("parking\n");
+done:
+	stop_this_cpu();
+}
+
+static const unsigned int fixed_mtrrs[NUM_FIXED_MTRRS] = {
+	MTRR_FIX_64K_00000_MSR, MTRR_FIX_16K_80000_MSR, MTRR_FIX_16K_A0000_MSR,
+	MTRR_FIX_4K_C0000_MSR, MTRR_FIX_4K_C8000_MSR, MTRR_FIX_4K_D0000_MSR,
+	MTRR_FIX_4K_D8000_MSR, MTRR_FIX_4K_E0000_MSR, MTRR_FIX_4K_E8000_MSR,
+	MTRR_FIX_4K_F0000_MSR, MTRR_FIX_4K_F8000_MSR,
+};
+
+static inline struct saved_msr *save_msr(int index, struct saved_msr *entry)
+{
+	msr_t msr;
+
+	msr = msr_read(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 = msr_read(MTRR_CAP_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) {
+		printf("Cannot mirror all %d msrs.\n", msr_count);
+		return -ENOSPC;
+	}
+
+	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(MTRR_PHYS_BASE_MSR(i), msr_entry);
+		msr_entry = save_msr(MTRR_PHYS_MASK_MSR(i), msr_entry);
+	}
+
+	msr_entry = save_msr(MTRR_DEF_TYPE_MSR, msr_entry);
+
+	return msr_count;
+}
+
+static int load_sipi_vector(atomic_t **ap_countp)
+{
+	struct sipi_params_16bit *params16;
+	struct sipi_params *params;
+	static char msr_save[512];
+	char *stack;
+	ulong addr;
+	int code_len;
+	int size;
+	int ret;
+
+	/* Copy in the code */
+	code_len = ap_start16_code_end - ap_start16;
+	debug("Copying SIPI code to %x: %d bytes\n", AP_DEFAULT_BASE,
+	      code_len);
+	memcpy((void *)AP_DEFAULT_BASE, ap_start16, code_len);
+
+	addr = AP_DEFAULT_BASE + (ulong)sipi_params_16bit - (ulong)ap_start16;
+	params16 = (struct sipi_params_16bit *)addr;
+	params16->ap_start = (uint32_t)ap_start;
+	params16->gdt = (uint32_t)gd->arch.gdt;
+	params16->gdt_limit = X86_GDT_SIZE - 1;
+	debug("gdt = %x, gdt_limit = %x\n", params16->gdt, params16->gdt_limit);
+
+	params = (struct sipi_params *)sipi_params;
+	debug("SIPI 32-bit params at %p\n", params);
+	params->idt_ptr = (uint32_t)x86_get_idt();
+
+	params->stack_size = CONFIG_AP_STACK_SIZE;
+	size = params->stack_size * CONFIG_MAX_CPUS;
+	stack = memalign(size, 4096);
+	if (!stack)
+		return -ENOMEM;
+	params->stack_top = (u32)(stack + size);
+
+	params->microcode_ptr = 0;
+	params->msr_table_ptr = (u32)msr_save;
+	ret = save_bsp_msrs(msr_save, sizeof(msr_save));
+	if (ret < 0)
+		return ret;
+	params->msr_count = ret;
+
+	params->c_handler = (uint32_t)&ap_init;
+
+	*ap_countp = &params->ap_count;
+	atomic_set(*ap_countp, 0);
+	debug("SIPI vector is ready\n");
+
+	return 0;
+}
+
+static int check_cpu_devices(int expected_cpus)
+{
+	int i;
+
+	for (i = 0; i < expected_cpus; i++) {
+		struct udevice *dev;
+		int ret;
+
+		ret = uclass_find_device(UCLASS_CPU, i, &dev);
+		if (ret) {
+			debug("Cannot find CPU %d in device tree\n", i);
+			return ret;
+		}
+	}
+
+	return 0;
+}
+
+/* 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(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 = AP_DEFAULT_BASE >> 12;
+
+	if (sipi_vector > max_vector_loc) {
+		printf("SIPI vector too large! 0x%08x\n",
+		       sipi_vector);
+		return -1;
+	}
+
+	debug("Attempting to start %d APs\n", ap_count);
+
+	if ((lapic_read(LAPIC_ICR) & LAPIC_ICR_BUSY)) {
+		debug("Waiting for ICR not to be busy...");
+		if (apic_wait_timeout(1000, 50)) {
+			debug("timed out. Aborting.\n");
+			return -1;
+		} else {
+			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);
+	debug("Waiting for 10ms after sending INIT.\n");
+	mdelay(10);
+
+	/* Send 1st SIPI */
+	if ((lapic_read(LAPIC_ICR) & LAPIC_ICR_BUSY)) {
+		debug("Waiting for ICR not to be busy...");
+		if (apic_wait_timeout(1000, 50)) {
+			debug("timed out. Aborting.\n");
+			return -1;
+		} else {
+			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);
+	debug("Waiting for 1st SIPI to complete...");
+	if (apic_wait_timeout(10000, 50)) {
+		debug("timed out.\n");
+		return -1;
+	} else {
+		debug("done.\n");
+	}
+
+	/* Wait for CPUs to check in up to 200 us */
+	wait_for_aps(num_aps, ap_count, 200, 15);
+
+	/* Send 2nd SIPI */
+	if ((lapic_read(LAPIC_ICR) & LAPIC_ICR_BUSY)) {
+		debug("Waiting for ICR not to be busy...");
+		if (apic_wait_timeout(1000, 50)) {
+			debug("timed out. Aborting.\n");
+			return -1;
+		} else {
+			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);
+	debug("Waiting for 2nd SIPI to complete...");
+	if (apic_wait_timeout(10000, 50)) {
+		debug("timed out.\n");
+		return -1;
+	} else {
+		debug("done.\n");
+	}
+
+	/* Wait for CPUs to check in */
+	if (wait_for_aps(num_aps, ap_count, 10000, 50)) {
+		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 udevice *cpu, 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)) {
+				debug("MP record %d timeout.\n", i);
+				ret = -1;
+			}
+		}
+
+		if (rec->bsp_call != NULL)
+			rec->bsp_call(cpu, rec->bsp_arg);
+
+		release_barrier(&rec->barrier);
+	}
+	return ret;
+}
+
+static int init_bsp(struct udevice **devp)
+{
+	char processor_name[CPU_MAX_NAME_LEN];
+	int apic_id;
+	int ret;
+
+	cpu_get_name(processor_name);
+	debug("CPU: %s.\n", processor_name);
+
+	enable_lapic();
+
+	apic_id = lapicid();
+	ret = find_cpu_by_apid_id(apic_id, devp);
+	if (ret) {
+		printf("Cannot find boot CPU, APIC ID %d\n", apic_id);
+		return ret;
+	}
+
+	return 0;
+}
+
+int mp_init(struct mp_params *p)
+{
+	int num_aps;
+	atomic_t *ap_count;
+	struct udevice *cpu;
+	int ret;
+
+	/* This will cause the CPUs devices to be bound */
+	struct uclass *uc;
+	ret = uclass_get(UCLASS_CPU, &uc);
+	if (ret)
+		return ret;
+
+	ret = init_bsp(&cpu);
+	if (ret) {
+		debug("Cannot init boot CPU: err=%d\n", ret);
+		return ret;
+	}
+
+	if (p == NULL || p->flight_plan == NULL || p->num_records < 1) {
+		printf("Invalid MP parameters\n");
+		return -1;
+	}
+
+	ret = check_cpu_devices(p->num_cpus);
+	if (ret)
+		debug("Warning: Device tree does not describe all CPUs. Extra ones will not be started correctly\n");
+
+	/* 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 */
+	ret = load_sipi_vector(&ap_count);
+	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;
+	ret = start_aps(num_aps, ap_count);
+	if (ret) {
+		mdelay(1000);
+		debug("%d/%d eventually checked in?\n", atomic_read(ap_count),
+		      num_aps);
+		return ret;
+	}
+
+	/* Walk the flight plan for the BSP */
+	ret = bsp_do_flight_plan(cpu, p);
+	if (ret) {
+		debug("CPU init failed: err=%d\n", ret);
+		return ret;
+	}
+
+	return 0;
+}
+
+int mp_init_cpu(struct udevice *cpu, void *unused)
+{
+	return device_probe(cpu);
+}
diff --git a/arch/x86/cpu/sipi_vector.S b/arch/x86/cpu/sipi_vector.S
new file mode 100644
index 0000000..bcef12c
--- /dev/null
+++ b/arch/x86/cpu/sipi_vector.S
@@ -0,0 +1,216 @@
+/*
+ * Copyright (c) 2015 Google, Inc
+ *
+ * SPDX-License-Identifier:	GPL-2.0
+ *
+ * Taken from coreboot file of the same name
+ */
+
+/*
+ * 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
+ */
+
+#include <asm/global_data.h>
+#include <asm/msr-index.h>
+#include <asm/processor.h>
+#include <asm/processor-flags.h>
+#include <asm/sipi.h>
+
+#define CODE_SEG	(X86_GDT_ENTRY_32BIT_CS * X86_GDT_ENTRY_SIZE)
+#define DATA_SEG	(X86_GDT_ENTRY_32BIT_DS * X86_GDT_ENTRY_SIZE)
+
+/*
+ * First we have the 16-bit section. Every AP process starts here.
+ * The simple task is to load U-Boot's Global Descriptor Table (GDT) to allow
+ * U-Boot's 32-bit code to become visible, then jump to ap_start.
+ *
+ * Note that this code is copied to RAM below 1MB in mp_init.c, and runs from
+ * there, but the 32-bit code (ap_start and onwards) is part of U-Boot and
+ * is therefore relocated to the top of RAM with other U-Boot code. This
+ * means that for the 16-bit code we must write relocatable code, but for the
+ * rest, we can do what we like.
+ */
+.text
+.code16
+.globl ap_start16
+ap_start16:
+	cli
+	xorl	%eax, %eax
+	movl	%eax, %cr3		/* Invalidate TLB */
+
+	/* setup the data segment */
+	movw	%cs, %ax
+	movw	%ax, %ds
+
+	/* Use an address relative to the data segment for the GDT */
+	movl	$gdtaddr, %ebx
+	subl	$ap_start16, %ebx
+
+	data32 lgdt (%ebx)
+
+	movl	%cr0, %eax
+	andl	$(~(X86_CR0_PG | X86_CR0_AM | X86_CR0_WP | X86_CR0_NE | \
+		    X86_CR0_TS | X86_CR0_EM | X86_CR0_MP)), %eax
+	orl	$(X86_CR0_NW | X86_CR0_CD | X86_CR0_PE), %eax
+	movl	%eax, %cr0
+
+	movl	$ap_start_jmp, %eax
+	subl	$ap_start16, %eax
+	movw	%ax, %bp
+
+	/* Jump to ap_start within U-Boot */
+data32 cs	ljmp	*(%bp)
+
+	.align	4
+.globl sipi_params_16bit
+sipi_params_16bit:
+	/* 48-bit far pointer */
+ap_start_jmp:
+	.long	0		/* offset set to ap_start by U-Boot */
+	.word	CODE_SEG	/* segment */
+
+	.word	0		/* padding */
+gdtaddr:
+	.word	0 /* limit */
+	.long	0 /* table */
+	.word	0 /* unused */
+
+.globl ap_start16_code_end
+ap_start16_code_end:
+
+/*
+ * Set up the special 'fs' segment for global_data. Then jump to ap_continue
+ * to set up the AP.
+ */
+.globl ap_start
+ap_start:
+	.code32
+	movw	$DATA_SEG, %ax
+	movw	%ax, %ds
+	movw	%ax, %es
+	movw	%ax, %ss
+	movw	%ax, %gs
+
+	movw	$(X86_GDT_ENTRY_32BIT_FS * X86_GDT_ENTRY_SIZE), %ax
+	movw	%ax, %fs
+
+	/* 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	$MSR_IA32_UCODE_REV, %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	$MSR_IA32_UCODE_WRITE, %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	$UCODE_HEADER_LEN, %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]
+	 * See struct saved_msr in mp_init.c.
+	 */
+	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
+	andl	$(~(X86_CR0_CD | X86_CR0_NW)), %eax
+	mov	%eax, %cr0
+
+	/* c_handler(cpu_num) */
+	movl	%esi, %eax	/* cpu_num */
+	mov	c_handler, %eax
+	call	*%eax
+
+	.align	4
+.globl	sipi_params
+sipi_params:
+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