path: root/src/cpu/x86/mp_init.c

/*
 * This file is part of the coreboot project.
 *
 * 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.
 */

#include <console/console.h>
#include <stddef.h>
#include <stdint.h>
#include <string.h>
#include <rmodule.h>
#include <arch/cpu.h>
#include <commonlib/helpers.h>
#include <cpu/cpu.h>
#include <cpu/intel/microcode.h>
#include <cpu/x86/cache.h>
#include <cpu/x86/gdt.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 <smp/atomic.h>
#include <smp/spinlock.h>
#include <symbols.h>
#include <timer.h>
#include <thread.h>

#define MAX_APIC_IDS 256

struct mp_callback {
	void (*func)(void *);
	void *arg;
	int logical_cpu_number;
};

static char processor_name[49];

/*
 * A mp_flight_record details a sequence of calls for the APs to perform
 * along with the BSP to coordinate sequencing. Each flight record either
 * provides a barrier for each AP before calling the callback or the APs
 * are allowed to perform the callback without waiting. Regardless, each
 * record has the cpus_entered field incremented for each record. When
 * the BSP observes that the cpus_entered matches the number of APs
 * the bsp_call is called with bsp_arg and upon returning releases the
 * barrier allowing the APs to make further progress.
 *
 * Note that ap_call() and bsp_call() can be NULL. In the NULL case the
 * callback will just not be called.
 */
struct mp_flight_record {
	atomic_t barrier;
	atomic_t cpus_entered;
	void (*ap_call)(void);
	void (*bsp_call)(void);
} __aligned(CACHELINE_SIZE);

#define _MP_FLIGHT_RECORD(barrier_, ap_func_, bsp_func_) \
	{							\
		.barrier = ATOMIC_INIT(barrier_),		\
		.cpus_entered = ATOMIC_INIT(0),			\
		.ap_call = ap_func_,				\
		.bsp_call = bsp_func_,				\
	}

#define MP_FR_BLOCK_APS(ap_func_, bsp_func_) \
	_MP_FLIGHT_RECORD(0, ap_func_, bsp_func_)

#define MP_FR_NOBLOCK_APS(ap_func_, bsp_func_) \
	_MP_FLIGHT_RECORD(1, ap_func_, bsp_func_)

/* The mp_params structure provides the arguments to the mp subsystem
 * for bringing up APs. */
struct mp_params {
	int num_cpus; /* Total cpus include BSP */
	int parallel_microcode_load;
	const void *microcode_pointer;
	/* Flight plan  for APs and BSP. */
	struct mp_flight_record *flight_plan;
	int num_records;
};

/* 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;
} __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;
} __packed;


/* The sipi vector rmodule is included in the ramstage using 'objdump -B'. */
extern char _binary_sipi_vector_start[];

/* 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 int global_num_aps;
static struct mp_flight_plan mp_info;

/* Keep track of device structure for each CPU. */
static struct device *cpus_dev[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();
	}
}

static void park_this_cpu(void *unused)
{
	stop_this_cpu();
}

/* 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;

	/* Ensure the local APIC is enabled */
	enable_lapic();

	info = cpu_info();
	info->index = cpu;
	info->cpu = cpus_dev[cpu];

	cpu_add_map_entry(info->index);
	thread_init_cpu_info_non_bsp(info);

	/* Fix up APIC id with reality. */
	info->cpu->path.apic.apic_id = lapicid();

	printk(BIOS_INFO, "AP: slot %d apic_id %x.\n", cpu,
		info->cpu->path.apic.apic_id);

	/* Walk the flight plan */
	ap_do_flight_plan();

	/* Park the AP. */
	park_this_cpu(NULL);
}

static void setup_default_sipi_vector_params(struct sipi_params *sp)
{
	sp->gdt = (uintptr_t)&gdt;
	sp->gdtlimit = (uintptr_t)&gdt_end - (uintptr_t)&gdt - 1;
	sp->idt_ptr = (uintptr_t)&idtarg;
	sp->stack_size = CONFIG_STACK_SIZE;
	sp->stack_top = ALIGN_DOWN((uintptr_t)&_estack, CONFIG_STACK_SIZE);
	/* 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] = {
	MTRR_FIX_64K_00000, MTRR_FIX_16K_80000, MTRR_FIX_16K_A0000,
	MTRR_FIX_4K_C0000, MTRR_FIX_4K_C8000, MTRR_FIX_4K_D0000,
	MTRR_FIX_4K_D8000, MTRR_FIX_4K_E0000, MTRR_FIX_4K_E8000,
	MTRR_FIX_4K_F0000, MTRR_FIX_4K_F8000,
};

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(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) {
		printk(BIOS_CRIT, "Cannot mirror all %d msrs.\n", msr_count);
		return -1;
	}

	fixed_mtrrs_expose_amd_rwdram();

	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(i), msr_entry);
		msr_entry = save_msr(MTRR_PHYS_MASK(i), msr_entry);
	}

	msr_entry = save_msr(MTRR_DEF_TYPE_MSR, msr_entry);

	fixed_mtrrs_hide_amd_rwdram();

	/* Tell static analysis we know value is left unused. */
	(void)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_UP(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 ability 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;
		struct device *new;

		/* Build the CPU device path */
		cpu_path.type = DEVICE_PATH_APIC;

		/* Assuming linear APIC space allocation. AP will set its own
		   APIC id in the ap_init() path above. */
		cpu_path.apic.apic_id = info->cpu->path.apic.apic_id + i;

		/* Allocate the new CPU device structure */
		new = alloc_find_dev(cpu_bus, &cpu_path);
		if (new == NULL) {
			printk(BIOS_CRIT, "Could not allocate CPU device\n");
			max_cpus--;
			continue;
		}
		new->name = processor_name;
		cpus_dev[i] = 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;
		}
		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;
		}
		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;
	}
	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 */);

	if (CONFIG(X86_AMD_INIT_SIPI))
		return 0;

	/* 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;
		}
		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;
	}
	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;
	/*
	 * Set time out for flight plan to a huge minimum value (>=1 second).
	 * CPUs with many APs may take longer if there is contention for
	 * resources such as UART, so scale the time out up by increments of
	 * 100ms if needed.
	 */
	const int timeout_us = MAX(1000000, 100000 * mp_params->num_cpus);
	const int step_us = 100;
	int num_aps = mp_params->num_cpus - 1;
	struct stopwatch sw;

	stopwatch_init(&sw);

	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();

		release_barrier(&rec->barrier);
	}

	printk(BIOS_INFO, "%s done after %ld msecs.\n", __func__,
	       stopwatch_duration_msecs(&sw));
	return ret;
}

static void init_bsp(struct bus *cpu_bus)
{
	struct device_path cpu_path;
	struct cpu_info *info;

	/* 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);
	info->cpu->name = processor_name;

	if (info->index != 0)
		printk(BIOS_CRIT, "BSP index(%d) != 0!\n", info->index);

	/* Track BSP in cpu_map structures. */
	cpu_add_map_entry(info->index);
}

/*
 * mp_init() will set up the SIPI vector and bring up the APs according to
 * mp_params. Each flight record will be executed according to the plan. Note
 * that the MP infrastructure uses SMM default area without saving it. It's
 * up to the chipset or mainboard to either e820 reserve this area or save this
 * region prior to calling mp_init() and restoring it after mp_init returns.
 *
 * At the time mp_init() is called the MTRR MSRs are mirrored into APs then
 * caching is enabled before running the flight plan.
 *
 * The MP initialization has the following properties:
 * 1. APs are brought up in parallel.
 * 2. The ordering of coreboot CPU number and APIC ids is not deterministic.
 *    Therefore, one cannot rely on this property or the order of devices in
 *    the device tree unless the chipset or mainboard know the APIC ids
 *    a priori.
 *
 * mp_init() returns < 0 on error, 0 on success.
 */
static int mp_init(struct bus *cpu_bus, struct mp_params *p)
{
	int num_cpus;
	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. */
	global_num_aps = p->num_cpus - 1;
	if (start_aps(cpu_bus, global_num_aps, ap_count) < 0) {
		mdelay(1000);
		printk(BIOS_DEBUG, "%d/%d eventually checked in?\n",
		       atomic_read(ap_count), global_num_aps);
		return -1;
	}

	/* Walk the flight plan for the BSP. */
	return bsp_do_flight_plan(p);
}

/* Calls cpu_initialize(info->index) which calls the coreboot CPU drivers. */
static void mp_initialize_cpu(void)
{
	/* Call back into driver infrastructure for the AP initialization.   */
	struct cpu_info *info = cpu_info();
	cpu_initialize(info->index);
}

void smm_initiate_relocation_parallel(void)
{
	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;
		}
		printk(BIOS_DEBUG, "done.\n");
	}

	lapic_write_around(LAPIC_ICR2, SET_LAPIC_DEST_FIELD(lapicid()));
	lapic_write_around(LAPIC_ICR, LAPIC_INT_ASSERT | LAPIC_DM_SMI);
	if (apic_wait_timeout(1000 /* 1 ms */, 100 /* us */))
		printk(BIOS_DEBUG, "SMI Relocation timed out.\n");
	else
		printk(BIOS_DEBUG, "Relocation complete.\n");
}

DECLARE_SPIN_LOCK(smm_relocation_lock);

/* Send SMI to self with single user serialization. */
void smm_initiate_relocation(void)
{
	spin_lock(&smm_relocation_lock);
	smm_initiate_relocation_parallel();
	spin_unlock(&smm_relocation_lock);
}

struct mp_state {
	struct mp_ops ops;
	int cpu_count;
	uintptr_t perm_smbase;
	size_t perm_smsize;
	size_t smm_save_state_size;
	int do_smm;
} mp_state;

static int is_smm_enabled(void)
{
	return CONFIG(HAVE_SMI_HANDLER) && mp_state.do_smm;
}

static void smm_disable(void)
{
	mp_state.do_smm = 0;
}

static void smm_enable(void)
{
	if (CONFIG(HAVE_SMI_HANDLER))
		mp_state.do_smm = 1;
}

static void asmlinkage smm_do_relocation(void *arg)
{
	const struct smm_module_params *p;
	const struct smm_runtime *runtime;
	int cpu;
	uintptr_t curr_smbase;
	uintptr_t perm_smbase;

	p = arg;
	runtime = p->runtime;
	cpu = p->cpu;
	curr_smbase = runtime->smbase;

	if (cpu >= CONFIG_MAX_CPUS) {
		printk(BIOS_CRIT,
		       "Invalid CPU number assigned in SMM stub: %d\n", cpu);
		return;
	}

	/*
	 * The permanent handler runs with all cpus concurrently. Precalculate
	 * the location of the new SMBASE. If using SMM modules then this
	 * calculation needs to match that of the module loader.
	 */
	perm_smbase = mp_state.perm_smbase;
	perm_smbase -= cpu * runtime->save_state_size;

	printk(BIOS_DEBUG, "New SMBASE 0x%08lx\n", perm_smbase);

	/* Setup code checks this callback for validity. */
	mp_state.ops.relocation_handler(cpu, curr_smbase, perm_smbase);
}

static void adjust_smm_apic_id_map(struct smm_loader_params *smm_params)
{
	int i;
	struct smm_runtime *runtime = smm_params->runtime;

	for (i = 0; i < CONFIG_MAX_CPUS; i++)
		runtime->apic_id_to_cpu[i] = cpu_get_apic_id(i);
}

static int install_relocation_handler(int num_cpus, size_t save_state_size)
{
	struct smm_loader_params smm_params = {
		.per_cpu_stack_size = CONFIG_SMM_STUB_STACK_SIZE,
		.num_concurrent_stacks = num_cpus,
		.per_cpu_save_state_size = save_state_size,
		.num_concurrent_save_states = 1,
		.handler = smm_do_relocation,
	};

	/* Allow callback to override parameters. */
	if (mp_state.ops.adjust_smm_params != NULL)
		mp_state.ops.adjust_smm_params(&smm_params, 0);

	if (smm_setup_relocation_handler(&smm_params))
		return -1;

	adjust_smm_apic_id_map(&smm_params);

	return 0;
}

static int install_permanent_handler(int num_cpus, uintptr_t smbase,
					size_t smsize, size_t save_state_size)
{
	/* There are num_cpus concurrent stacks and num_cpus concurrent save
	 * state areas. Lastly, set the stack size to 1KiB. */
	struct smm_loader_params smm_params = {
		.per_cpu_stack_size = CONFIG_SMM_MODULE_STACK_SIZE,
		.num_concurrent_stacks = num_cpus,
		.per_cpu_save_state_size = save_state_size,
		.num_concurrent_save_states = num_cpus,
	};

	/* Allow callback to override parameters. */
	if (mp_state.ops.adjust_smm_params != NULL)
		mp_state.ops.adjust_smm_params(&smm_params, 1);

	printk(BIOS_DEBUG, "Installing SMM handler to 0x%08lx\n", smbase);

	if (smm_load_module((void *)smbase, smsize, &smm_params))
		return -1;

	adjust_smm_apic_id_map(&smm_params);

	return 0;
}

/* Load SMM handlers as part of MP flight record. */
static void load_smm_handlers(void)
{
	size_t smm_save_state_size = mp_state.smm_save_state_size;

	/* Do nothing if SMM is disabled.*/
	if (!is_smm_enabled())
		return;

	/* Install handlers. */
	if (install_relocation_handler(mp_state.cpu_count,
		smm_save_state_size) < 0) {
		printk(BIOS_ERR, "Unable to install SMM relocation handler.\n");
		smm_disable();
	}

	if (install_permanent_handler(mp_state.cpu_count, mp_state.perm_smbase,
		mp_state.perm_smsize, smm_save_state_size) < 0) {
		printk(BIOS_ERR, "Unable to install SMM permanent handler.\n");
		smm_disable();
	}

	/* Ensure the SMM handlers hit DRAM before performing first SMI. */
	wbinvd();

	/*
	 * Indicate that the SMM handlers have been loaded and MP
	 * initialization is about to start.
	 */
	if (is_smm_enabled() && mp_state.ops.pre_mp_smm_init != NULL)
		mp_state.ops.pre_mp_smm_init();
}

/* Trigger SMM as part of MP flight record. */
static void trigger_smm_relocation(void)
{
	/* Do nothing if SMM is disabled.*/
	if (!is_smm_enabled() || mp_state.ops.per_cpu_smm_trigger == NULL)
		return;
	/* Trigger SMM mode for the currently running processor. */
	mp_state.ops.per_cpu_smm_trigger();
}

static struct mp_callback *ap_callbacks[CONFIG_MAX_CPUS];

static struct mp_callback *read_callback(struct mp_callback **slot)
{
	struct mp_callback *ret;

	asm volatile ("mov	%1, %0\n"
		: "=r" (ret)
		: "m" (*slot)
		: "memory"
	);
	return ret;
}

static void store_callback(struct mp_callback **slot, struct mp_callback *val)
{
	asm volatile ("mov	%1, %0\n"
		: "=m" (*slot)
		: "r" (val)
		: "memory"
	);
}

static int run_ap_work(struct mp_callback *val, long expire_us)
{
	int i;
	int cpus_accepted;
	struct stopwatch sw;
	int cur_cpu;

	if (!CONFIG(PARALLEL_MP_AP_WORK)) {
		printk(BIOS_ERR, "APs already parked. PARALLEL_MP_AP_WORK not selected.\n");
		return -1;
	}

	cur_cpu = cpu_index();

	if (cur_cpu < 0) {
		printk(BIOS_ERR, "Invalid CPU index.\n");
		return -1;
	}

	/* Signal to all the APs to run the func. */
	for (i = 0; i < ARRAY_SIZE(ap_callbacks); i++) {
		if (cur_cpu == i)
			continue;
		store_callback(&ap_callbacks[i], val);
	}
	mfence();

	/* Wait for all the APs to signal back that call has been accepted. */
	if (expire_us > 0)
		stopwatch_init_usecs_expire(&sw, expire_us);

	do {
		cpus_accepted = 0;

		for (i = 0; i < ARRAY_SIZE(ap_callbacks); i++) {
			if (cur_cpu == i)
				continue;
			if (read_callback(&ap_callbacks[i]) == NULL)
				cpus_accepted++;
		}

		if (cpus_accepted == global_num_aps)
			return 0;
	} while (expire_us <= 0 || !stopwatch_expired(&sw));

	printk(BIOS_CRIT, "CIRTICAL ERROR: AP call expired. %d/%d CPUs accepted.\n",
		cpus_accepted, global_num_aps);
	return -1;
}

static void ap_wait_for_instruction(void)
{
	struct mp_callback lcb;
	struct mp_callback **per_cpu_slot;
	int cur_cpu;

	if (!CONFIG(PARALLEL_MP_AP_WORK))
		return;

	cur_cpu = cpu_index();

	if (cur_cpu < 0) {
		printk(BIOS_ERR, "Invalid CPU index.\n");
		return;
	}

	per_cpu_slot = &ap_callbacks[cur_cpu];

	while (1) {
		struct mp_callback *cb = read_callback(per_cpu_slot);

		if (cb == NULL) {
			asm ("pause");
			continue;
		}

		/* Copy to local variable before signaling consumption. */
		memcpy(&lcb, cb, sizeof(lcb));
		mfence();
		store_callback(per_cpu_slot, NULL);
		if (lcb.logical_cpu_number && (cur_cpu !=
				lcb.logical_cpu_number))
			continue;
		else
			lcb.func(lcb.arg);
	}
}

int mp_run_on_aps(void (*func)(void *), void *arg, int logical_cpu_num,
		long expire_us)
{
	struct mp_callback lcb = { .func = func, .arg = arg,
				.logical_cpu_number = logical_cpu_num};
	return run_ap_work(&lcb, expire_us);
}

int mp_run_on_all_cpus(void (*func)(void *), void *arg)
{
	/* Run on BSP first. */
	func(arg);

	/* For up to 1 second for AP to finish previous work. */
	return mp_run_on_aps(func, arg, MP_RUN_ON_ALL_CPUS, 1000 * USECS_PER_MSEC);
}

int mp_park_aps(void)
{
	struct stopwatch sw;
	int ret;
	long duration_msecs;

	stopwatch_init(&sw);

	ret = mp_run_on_aps(park_this_cpu, NULL, MP_RUN_ON_ALL_CPUS,
				1000 * USECS_PER_MSEC);

	duration_msecs = stopwatch_duration_msecs(&sw);

	if (!ret)
		printk(BIOS_DEBUG, "%s done after %ld msecs.\n", __func__,
		       duration_msecs);
	else
		printk(BIOS_ERR, "%s failed after %ld msecs.\n", __func__,
		       duration_msecs);

	return ret;
}

static struct mp_flight_record mp_steps[] = {
	/* Once the APs are up load the SMM handlers. */
	MP_FR_BLOCK_APS(NULL, load_smm_handlers),
	/* Perform SMM relocation. */
	MP_FR_NOBLOCK_APS(trigger_smm_relocation, trigger_smm_relocation),
	/* Initialize each CPU through the driver framework. */
	MP_FR_BLOCK_APS(mp_initialize_cpu, mp_initialize_cpu),
	/* Wait for APs to finish then optionally start looking for work. */
	MP_FR_BLOCK_APS(ap_wait_for_instruction, NULL),
};

static void fill_mp_state(struct mp_state *state, const struct mp_ops *ops)
{
	/*
	 * Make copy of the ops so that defaults can be set in the non-const
	 * structure if needed.
	 */
	memcpy(&state->ops, ops, sizeof(*ops));

	if (ops->get_cpu_count != NULL)
		state->cpu_count = ops->get_cpu_count();

	if (ops->get_smm_info != NULL)
		ops->get_smm_info(&state->perm_smbase, &state->perm_smsize,
					&state->smm_save_state_size);

	/*
	 * Default to smm_initiate_relocation() if trigger callback isn't
	 * provided.
	 */
	if (CONFIG(HAVE_SMI_HANDLER) &&
		ops->per_cpu_smm_trigger == NULL)
		mp_state.ops.per_cpu_smm_trigger = smm_initiate_relocation;
}

int mp_init_with_smm(struct bus *cpu_bus, const struct mp_ops *mp_ops)
{
	int ret;
	void *default_smm_area;
	struct mp_params mp_params;

	if (mp_ops->pre_mp_init != NULL)
		mp_ops->pre_mp_init();

	fill_mp_state(&mp_state, mp_ops);

	memset(&mp_params, 0, sizeof(mp_params));

	if (mp_state.cpu_count <= 0) {
		printk(BIOS_ERR, "Invalid cpu_count: %d\n", mp_state.cpu_count);
		return -1;
	}

	/* Sanity check SMM state. */
	if (mp_state.perm_smsize != 0 && mp_state.smm_save_state_size != 0 &&
		mp_state.ops.relocation_handler != NULL)
		smm_enable();

	if (is_smm_enabled())
		printk(BIOS_INFO, "Will perform SMM setup.\n");

	mp_params.num_cpus = mp_state.cpu_count;
	/* Gather microcode information. */
	if (mp_state.ops.get_microcode_info != NULL)
		mp_state.ops.get_microcode_info(&mp_params.microcode_pointer,
			&mp_params.parallel_microcode_load);
	mp_params.flight_plan = &mp_steps[0];
	mp_params.num_records = ARRAY_SIZE(mp_steps);

	/* Perform backup of default SMM area. */
	default_smm_area = backup_default_smm_area();

	ret = mp_init(cpu_bus, &mp_params);

	restore_default_smm_area(default_smm_area);

	/* Signal callback on success if it's provided. */
	if (ret == 0 && mp_state.ops.post_mp_init != NULL)
		mp_state.ops.post_mp_init();

	return ret;
}