cpu/x86: Drop LEGACY_SMP_INIT

This codepath is deprecated after the 4.18 release.

Change-Id: I7e90f457f3979781d06323ef1350d5fb05a6be43
Signed-off-by: Arthur Heymans <arthur@aheymans.xyz>
Reviewed-on: https://review.coreboot.org/c/coreboot/+/69121
Reviewed-by: Elyes Haouas <ehaouas@noos.fr>
Tested-by: build bot (Jenkins) <no-reply@coreboot.org>
Reviewed-by: Angel Pons <th3fanbus@gmail.com>
This commit is contained in:
Arthur Heymans
2022-11-01 23:45:59 +01:00
parent e2d291b5ae
commit 66b2888b77
9 changed files with 1 additions and 1226 deletions

View File

@@ -2,7 +2,6 @@ if ARCH_X86
config PARALLEL_MP config PARALLEL_MP
def_bool y def_bool y
depends on !LEGACY_SMP_INIT
select CPU_INFO_V2 select CPU_INFO_V2
help help
This option uses common MP infrastructure for bringing up APs This option uses common MP infrastructure for bringing up APs
@@ -27,9 +26,6 @@ config X86_SMM_SKIP_RELOCATION_HANDLER
with a stub at 0x30000. This is useful on platforms that have with a stub at 0x30000. This is useful on platforms that have
an alternative way to set SMBASE. an alternative way to set SMBASE.
config LEGACY_SMP_INIT
bool
config DEFAULT_X2APIC config DEFAULT_X2APIC
def_bool n def_bool n
help help
@@ -154,13 +150,7 @@ config SMM_TSEG
default y default y
depends on !(NO_SMM || SMM_ASEG) depends on !(NO_SMM || SMM_ASEG)
config SMM_LEGACY_ASEG if HAVE_SMI_HANDLER
bool
default y if HAVE_SMI_HANDLER && SMM_ASEG && LEGACY_SMP_INIT
help
SMM support without PARALLEL_MP, to be deprecated.
if HAVE_SMI_HANDLER && !SMM_LEGACY_ASEG
config SMM_MODULE_STACK_SIZE config SMM_MODULE_STACK_SIZE
hex hex

View File

@@ -1,6 +1,4 @@
ramstage-$(CONFIG_AP_IN_SIPI_WAIT) += lapic_cpu_stop.c ramstage-$(CONFIG_AP_IN_SIPI_WAIT) += lapic_cpu_stop.c
ramstage-$(CONFIG_LEGACY_SMP_INIT) += lapic_cpu_init.c
ramstage-$(CONFIG_LEGACY_SMP_INIT) += secondary.S
bootblock-$(CONFIG_UDELAY_LAPIC) += apic_timer.c bootblock-$(CONFIG_UDELAY_LAPIC) += apic_timer.c
romstage-$(CONFIG_UDELAY_LAPIC) += apic_timer.c romstage-$(CONFIG_UDELAY_LAPIC) += apic_timer.c

View File

@@ -1,420 +0,0 @@
/* SPDX-License-Identifier: GPL-2.0-only */
#include <cpu/x86/cr.h>
#include <cpu/x86/gdt.h>
#include <cpu/x86/lapic.h>
#include <cpu/x86/smi_deprecated.h>
#include <acpi/acpi.h>
#include <delay.h>
#include <lib.h>
#include <string.h>
#include <symbols.h>
#include <console/console.h>
#include <device/device.h>
#include <device/path.h>
#include <smp/atomic.h>
#include <smp/spinlock.h>
#include <cpu/cpu.h>
#include <cpu/intel/speedstep.h>
#include <smp/node.h>
#include <stdlib.h>
#include <thread.h>
/* This is a lot more paranoid now, since Linux can NOT handle
* being told there is a CPU when none exists. So any errors
* will return 0, meaning no CPU.
*
* We actually handling that case by noting which cpus startup
* and not telling anyone about the ones that don't.
*/
/* Start-UP IPI vector must be 4kB aligned and below 1MB. */
#define AP_SIPI_VECTOR 0x1000
static char *lowmem_backup;
static char *lowmem_backup_ptr;
static int lowmem_backup_size;
static inline void setup_secondary_gdt(void)
{
u16 *gdt_limit;
#if ENV_X86_64
u64 *gdt_base;
#else
u32 *gdt_base;
#endif
gdt_limit = (void *)&_secondary_gdt_addr;
gdt_base = (void *)&gdt_limit[1];
*gdt_limit = (uintptr_t)&gdt_end - (uintptr_t)&gdt - 1;
*gdt_base = (uintptr_t)&gdt;
}
static void copy_secondary_start_to_lowest_1M(void)
{
unsigned long code_size;
/* Fill in secondary_start's local gdt. */
setup_secondary_gdt();
code_size = (unsigned long)_secondary_start_end
- (unsigned long)_secondary_start;
if (acpi_is_wakeup_s3()) {
/* need to save it for RAM resume */
lowmem_backup_size = code_size;
lowmem_backup = malloc(code_size);
lowmem_backup_ptr = (char *)AP_SIPI_VECTOR;
if (lowmem_backup == NULL)
die("Out of backup memory\n");
memcpy(lowmem_backup, lowmem_backup_ptr, lowmem_backup_size);
}
/* copy the _secondary_start to the RAM below 1M*/
memcpy((unsigned char *)AP_SIPI_VECTOR,
(unsigned char *)_secondary_start, code_size);
printk(BIOS_DEBUG, "start_eip=0x%08lx, code_size=0x%08lx\n",
(unsigned long int)AP_SIPI_VECTOR, code_size);
}
static void recover_lowest_1M(void)
{
if (acpi_is_wakeup_s3())
memcpy(lowmem_backup_ptr, lowmem_backup, lowmem_backup_size);
}
static uint32_t wait_for_ipi_completion(const int timeout_ms)
{
int loops = timeout_ms * 10;
uint32_t send_status;
/* wait for the ipi send to finish */
printk(BIOS_SPEW, "Waiting for send to finish...\n");
do {
printk(BIOS_SPEW, "+");
udelay(100);
send_status = lapic_busy();
} while (send_status && (--loops > 0));
return send_status;
}
static int lapic_start_cpu(unsigned long apicid)
{
const int timeout_100ms = 100;
uint32_t send_status, accept_status;
int j, maxlvt;
/*
* Starting actual IPI sequence...
*/
printk(BIOS_SPEW, "Asserting INIT.\n");
/*
* Turn INIT on target chip
*/
lapic_send_ipi(LAPIC_INT_LEVELTRIG | LAPIC_INT_ASSERT | LAPIC_DM_INIT, apicid);
send_status = wait_for_ipi_completion(timeout_100ms);
if (send_status) {
printk(BIOS_ERR, "CPU %ld: First APIC write timed out. "
"Disabling\n", apicid);
// too bad.
printk(BIOS_ERR, "ESR is 0x%x\n", lapic_read(LAPIC_ESR));
if (lapic_read(LAPIC_ESR)) {
printk(BIOS_ERR, "Try to reset ESR\n");
lapic_write(LAPIC_ESR, 0);
printk(BIOS_ERR, "ESR is 0x%x\n",
lapic_read(LAPIC_ESR));
}
return 0;
}
mdelay(10);
printk(BIOS_SPEW, "Deasserting INIT.\n");
lapic_send_ipi(LAPIC_INT_LEVELTRIG | LAPIC_DM_INIT, apicid);
send_status = wait_for_ipi_completion(timeout_100ms);
if (send_status) {
printk(BIOS_ERR, "CPU %ld: Second APIC write timed out. "
"Disabling\n", apicid);
// too bad.
return 0;
}
/*
* Run STARTUP IPI loop.
*/
printk(BIOS_SPEW, "#startup loops: %d.\n", CONFIG_NUM_IPI_STARTS);
maxlvt = 4;
for (j = 1; j <= CONFIG_NUM_IPI_STARTS; j++) {
printk(BIOS_SPEW, "Sending STARTUP #%d to %lu.\n", j, apicid);
lapic_read(LAPIC_SPIV);
lapic_write(LAPIC_ESR, 0);
lapic_read(LAPIC_ESR);
printk(BIOS_SPEW, "After apic_write.\n");
/*
* STARTUP IPI
*/
lapic_send_ipi(LAPIC_DM_STARTUP | (AP_SIPI_VECTOR >> 12), apicid);
/*
* Give the other CPU some time to accept the IPI.
*/
udelay(300);
printk(BIOS_SPEW, "Startup point 1.\n");
send_status = wait_for_ipi_completion(timeout_100ms);
/*
* Give the other CPU some time to accept the IPI.
*/
udelay(200);
/*
* Due to the Pentium erratum 3AP.
*/
if (maxlvt > 3) {
lapic_read(LAPIC_SPIV);
lapic_write(LAPIC_ESR, 0);
}
accept_status = (lapic_read(LAPIC_ESR) & 0xEF);
if (send_status || accept_status)
break;
}
printk(BIOS_SPEW, "After Startup.\n");
if (send_status)
printk(BIOS_WARNING, "APIC never delivered???\n");
if (accept_status)
printk(BIOS_WARNING, "APIC delivery error (%x).\n",
accept_status);
if (send_status || accept_status)
return 0;
return 1;
}
/* Number of cpus that are currently running in coreboot */
static atomic_t active_cpus = ATOMIC_INIT(1);
/* start_cpu_lock covers last_cpu_index and secondary_stack.
* Only starting one CPU at a time let's me remove the logic
* for select the stack from assembly language.
*
* In addition communicating by variables to the CPU I
* am starting allows me to verify it has started before
* start_cpu returns.
*/
DECLARE_SPIN_LOCK(start_cpu_lock);
static unsigned int last_cpu_index = 0;
static void *stacks[CONFIG_MAX_CPUS];
volatile unsigned long secondary_stack;
volatile unsigned int secondary_cpu_index;
static int start_cpu(struct device *cpu)
{
struct cpu_info *info;
uintptr_t stack_top;
uintptr_t stack_base;
unsigned long apicid;
unsigned int index;
unsigned long count;
int result;
spin_lock(&start_cpu_lock);
/* Get the CPU's apicid */
apicid = cpu->path.apic.apic_id;
/* Get an index for the new processor */
index = ++last_cpu_index;
/* Find boundaries of the new processor's stack */
stack_top = ALIGN_DOWN((uintptr_t)_estack, CONFIG_STACK_SIZE);
stack_top -= (CONFIG_STACK_SIZE*index);
stack_base = stack_top - CONFIG_STACK_SIZE;
stack_top -= sizeof(struct cpu_info);
printk(BIOS_SPEW, "CPU%d: stack_base %p, stack_top %p\n", index,
(void *)stack_base, (void *)stack_top);
stacks[index] = (void *)stack_base;
/* Record the index and which CPU structure we are using */
info = (struct cpu_info *)stack_top;
info->index = index;
info->cpu = cpu;
cpu_add_map_entry(info->index);
/* Advertise the new stack and index to start_cpu */
secondary_stack = stack_top;
secondary_cpu_index = index;
/* Until the CPU starts up report the CPU is not enabled */
cpu->enabled = 0;
cpu->initialized = 0;
/* Start the CPU */
result = lapic_start_cpu(apicid);
if (result) {
result = 0;
/* Wait 1s or until the new CPU calls in */
for (count = 0; count < 100000; count++) {
if (secondary_stack == 0) {
result = 1;
break;
}
udelay(10);
}
}
secondary_stack = 0;
spin_unlock(&start_cpu_lock);
return result;
}
/* C entry point of secondary cpus */
asmlinkage void secondary_cpu_init(unsigned int index)
{
atomic_inc(&active_cpus);
spin_lock(&start_cpu_lock);
#ifdef __SSE3__
/*
* Seems that CR4 was cleared when AP start via lapic_start_cpu()
* Turn on CR4.OSFXSR and CR4.OSXMMEXCPT when SSE options enabled
*/
CRx_TYPE cr4_val;
cr4_val = read_cr4();
cr4_val |= (CR4_OSFXSR | CR4_OSXMMEXCPT);
write_cr4(cr4_val);
#endif
/* Ensure the local APIC is enabled */
enable_lapic();
setup_lapic_interrupts();
cpu_initialize(index);
spin_unlock(&start_cpu_lock);
atomic_dec(&active_cpus);
stop_this_cpu();
}
static void start_other_cpus(struct bus *cpu_bus, struct device *bsp_cpu)
{
struct device *cpu;
/* Loop through the cpus once getting them started */
for (cpu = cpu_bus->children; cpu; cpu = cpu->sibling) {
if (cpu->path.type != DEVICE_PATH_APIC)
continue;
if (!cpu->enabled)
continue;
if (cpu->initialized)
continue;
if (!start_cpu(cpu))
/* Record the error in cpu? */
printk(BIOS_ERR, "CPU 0x%02x would not start!\n",
cpu->path.apic.apic_id);
udelay(10);
}
}
static void wait_other_cpus_stop(struct bus *cpu_bus)
{
struct device *cpu;
int old_active_count, active_count;
long loopcount = 0;
int i;
/* Now loop until the other cpus have finished initializing */
old_active_count = 1;
active_count = atomic_read(&active_cpus);
while (active_count > 1) {
if (active_count != old_active_count) {
printk(BIOS_INFO, "Waiting for %d CPUS to stop\n",
active_count - 1);
old_active_count = active_count;
}
udelay(10);
active_count = atomic_read(&active_cpus);
loopcount++;
}
for (cpu = cpu_bus->children; cpu; cpu = cpu->sibling) {
if (cpu->path.type != DEVICE_PATH_APIC)
continue;
if (cpu->path.apic.apic_id == SPEEDSTEP_APIC_MAGIC)
continue;
if (!cpu->initialized)
printk(BIOS_ERR, "CPU 0x%02x did not initialize!\n",
cpu->path.apic.apic_id);
}
printk(BIOS_DEBUG, "All AP CPUs stopped (%ld loops)\n", loopcount);
checkstack(_estack, 0);
for (i = 1; i < CONFIG_MAX_CPUS && i <= last_cpu_index; i++)
checkstack((void *)stacks[i] + CONFIG_STACK_SIZE, i);
}
void initialize_cpus(struct bus *cpu_bus)
{
struct device_path cpu_path;
struct cpu_info *info;
/* Find the info struct for this CPU */
info = cpu_info();
/* Ensure the local APIC is enabled */
if (is_smp_boot()) {
enable_lapic();
setup_lapic_interrupts();
} else {
disable_lapic();
}
/* Get 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 = alloc_find_dev(cpu_bus, &cpu_path);
cpu_add_map_entry(info->index);
// why here? In case some day we can start core1 in amd_sibling_init
if (is_smp_boot())
copy_secondary_start_to_lowest_1M();
if (CONFIG(SMM_LEGACY_ASEG))
smm_init();
/* Initialize the bootstrap processor */
cpu_initialize(0);
if (is_smp_boot())
start_other_cpus(cpu_bus, info->cpu);
/* Now wait the rest of the cpus stop*/
if (is_smp_boot())
wait_other_cpus_stop(cpu_bus);
if (CONFIG(SMM_LEGACY_ASEG))
smm_init_completion();
if (is_smp_boot())
recover_lowest_1M();
}

View File

@@ -1,86 +0,0 @@
/* SPDX-License-Identifier: GPL-2.0-only */
#include <cpu/x86/lapic_def.h>
#include <arch/ram_segs.h>
.text
.globl _secondary_start, _secondary_start_end, _secondary_gdt_addr
.balign 4096
_secondary_start:
.code16
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
lgdtl gdtaddr - _secondary_start
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 $RAM_CODE_SEG, $__ap_protected_start
/* This will get filled in by C code. */
_secondary_gdt_addr:
gdtaddr:
.word 0 /* the table limit */
#if ENV_X86_64
.quad 0
#else
.long 0 /* we know the offset */
#endif
_secondary_start_end:
ap_protected_start:
.code32
lgdt gdtaddr
ljmpl $RAM_CODE_SEG, $__ap_protected_start
__ap_protected_start:
movw $RAM_DATA_SEG, %ax
movw %ax, %ds
movw %ax, %es
movw %ax, %ss
xor %ax, %ax /* zero out the gs and fs segment index */
movw %ax, %fs
movw %ax, %gs /* Will be used for cpu_info */
/* Load the Interrupt descriptor table */
lidt idtarg
#if ENV_X86_64
/* entry64.inc preserves ebx. */
#include <cpu/x86/64bit/entry64.inc>
movabs secondary_stack, %rax
mov %rax, %rsp
andl $0xfffffff0, %esp
movabs secondary_cpu_index, %rax
mov %rax, %rdi
#else
/* Set the stack pointer, and flag that we are done */
xorl %eax, %eax
movl secondary_stack, %esp
andl $0xfffffff0, %esp
sub $12, %esp /* maintain 16-byte alignment for the call below */
movl secondary_cpu_index, %edi
pushl %edi
movl %eax, secondary_stack
#endif
call secondary_cpu_init
1: hlt
jmp 1b
.code32

View File

@@ -76,17 +76,3 @@ $(obj)/smm/smm: $(obj)/smm/smm.elf.rmod
$(OBJCOPY_smm) -O binary $< $@ $(OBJCOPY_smm) -O binary $< $@
endif endif
ifeq ($(CONFIG_SMM_LEGACY_ASEG),y)
smm-y += smm.ld
$(obj)/smm/smm: $(obj)/smm/smm.o $(call src-to-obj,smm,$(src)/cpu/x86/smm/smm.ld)
$(LD_smm) $(LDFLAGS_smm) -o $(obj)/smm/smm.elf -T $(call src-to-obj,smm,$(src)/cpu/x86/smm/smm.ld) $(obj)/smm/smm.o
$(NM_smm) -n $(obj)/smm/smm.elf | sort > $(obj)/smm/smm.map
$(OBJCOPY_smm) -O binary $(obj)/smm/smm.elf $@
smm-y += smmhandler.S
smm-y += smihandler.c
endif # CONFIG_SMM_LEGACY_ASEG

View File

@@ -1,207 +0,0 @@
/* SPDX-License-Identifier: GPL-2.0-only */
#include <arch/io.h>
#include <console/console.h>
#include <commonlib/region.h>
#include <cpu/x86/smm.h>
#include <cpu/x86/smi_deprecated.h>
#include <cpu/amd/amd64_save_state.h>
#include <cpu/intel/em64t100_save_state.h>
#include <cpu/intel/em64t101_save_state.h>
#include <cpu/x86/lapic.h>
#include <cpu/x86/lapic_def.h>
#include <cpu/x86/legacy_save_state.h>
#if CONFIG(SPI_FLASH_SMM)
#include <spi-generic.h>
#endif
static int do_driver_init = 1;
typedef enum { SMI_LOCKED, SMI_UNLOCKED } smi_semaphore;
/* SMI multiprocessing semaphore */
static __attribute__((aligned(4))) volatile smi_semaphore smi_handler_status
= SMI_UNLOCKED;
static int smi_obtain_lock(void)
{
u8 ret = SMI_LOCKED;
asm volatile (
"movb %2, %%al\n"
"xchgb %%al, %1\n"
"movb %%al, %0\n"
: "=g" (ret), "=m" (smi_handler_status)
: "g" (SMI_LOCKED)
: "eax"
);
return (ret == SMI_UNLOCKED);
}
static void smi_release_lock(void)
{
asm volatile (
"movb %1, %%al\n"
"xchgb %%al, %0\n"
: "=m" (smi_handler_status)
: "g" (SMI_UNLOCKED)
: "eax"
);
}
void io_trap_handler(int smif)
{
/* If a handler function handled a given IO trap, it
* shall return a non-zero value
*/
printk(BIOS_DEBUG, "SMI function trap 0x%x: ", smif);
if (southbridge_io_trap_handler(smif))
return;
if (mainboard_io_trap_handler(smif))
return;
printk(BIOS_DEBUG, "Unknown function\n");
}
/**
* @brief Set the EOS bit
*/
static void smi_set_eos(void)
{
southbridge_smi_set_eos();
}
static u32 pci_orig;
/**
* @brief Backup PCI address to make sure we do not mess up the OS
*/
static void smi_backup_pci_address(void)
{
pci_orig = inl(0xcf8);
}
/**
* @brief Restore PCI address previously backed up
*/
static void smi_restore_pci_address(void)
{
outl(pci_orig, 0xcf8);
}
static inline void *smm_save_state(uintptr_t base, int arch_offset, int node)
{
base += SMM_SAVE_STATE_BEGIN(arch_offset) - (node * 0x400);
return (void *)base;
}
/* This returns the SMM revision from the savestate of CPU0,
which is assumed to be the same for all CPU's. See the memory
map in smmhandler.S */
uint32_t smm_revision(void)
{
return *(uint32_t *)(SMM_BASE + SMM_ENTRY_OFFSET * 2 - SMM_REVISION_OFFSET_FROM_TOP);
}
void *smm_get_save_state(int cpu)
{
switch (smm_revision()) {
case 0x00030002:
case 0x00030007:
return smm_save_state(SMM_BASE, SMM_LEGACY_ARCH_OFFSET, cpu);
case 0x00030100:
return smm_save_state(SMM_BASE, SMM_EM64T100_ARCH_OFFSET, cpu);
case 0x00030101: /* SandyBridge, IvyBridge, and Haswell */
return smm_save_state(SMM_BASE, SMM_EM64T101_ARCH_OFFSET, cpu);
case 0x00020064:
case 0x00030064:
return smm_save_state(SMM_BASE, SMM_AMD64_ARCH_OFFSET, cpu);
}
return NULL;
}
bool smm_region_overlaps_handler(const struct region *r)
{
const struct region r_smm = {SMM_BASE, SMM_DEFAULT_SIZE};
return region_overlap(&r_smm, r);
}
/**
* @brief Interrupt handler for SMI#
*
* @param smm_revision revision of the smm state save map
*/
void smi_handler(void)
{
unsigned int node;
/* Are we ok to execute the handler? */
if (!smi_obtain_lock()) {
/* For security reasons we don't release the other CPUs
* until the CPU with the lock is actually done
*/
while (smi_handler_status == SMI_LOCKED) {
asm volatile (
".byte 0xf3, 0x90\n" /* hint a CPU we are in
* spinlock (PAUSE
* instruction, REP NOP)
*/
);
}
return;
}
smi_backup_pci_address();
node = lapicid();
console_init();
printk(BIOS_SPEW, "\nSMI# #%d\n", node);
/* Use smm_get_save_state() to see if the smm revision is supported */
if (smm_get_save_state(node) == NULL) {
printk(BIOS_WARNING, "smm_revision: 0x%08x\n", smm_revision());
printk(BIOS_WARNING, "SMI# not supported on your CPU\n");
/* Don't release lock, so no further SMI will happen,
* if we don't handle it anyways.
*/
return;
}
/* Allow drivers to initialize variables in SMM context. */
if (do_driver_init) {
#if CONFIG(SPI_FLASH_SMM)
spi_init();
#endif
do_driver_init = 0;
}
/* Call chipset specific SMI handlers. */
southbridge_smi_handler();
smi_restore_pci_address();
smi_release_lock();
/* De-assert SMI# signal to allow another SMI */
smi_set_eos();
}
/* Provide a default implementation for all weak handlers so that relocation
* entries in the modules make sense. Without default implementations the
* weak relocations w/o a symbol have a 0 address which is where the modules
* are linked at. */
int __weak mainboard_io_trap_handler(int smif) { return 0; }
void __weak southbridge_smi_handler(void) {}
void __weak mainboard_smi_gpi(u32 gpi_sts) {}
int __weak mainboard_smi_apmc(u8 data) { return 0; }
void __weak mainboard_smi_sleep(u8 slp_typ) {}
void __weak mainboard_smi_finalize(void) {}

View File

@@ -1,65 +0,0 @@
/* SPDX-License-Identifier: GPL-2.0-only */
/* Maximum number of CPUs/cores */
CPUS = 4;
_ = ASSERT(CPUS >= CONFIG_MAX_CPUS, "The ASEG SMM code only supports up to 4 CPUS");
ENTRY(smm_handler_start);
SECTIONS
{
/* This is the actual SMM handler.
*
* We just put code, rodata, data and bss all in a row.
*/
. = 0xa0000;
.handler (.): {
_program = .;
/* Assembler stub */
*(.handler)
/* C code of the SMM handler */
*(.text);
*(.text.*);
/* C read-only data of the SMM handler */
. = ALIGN(16);
*(.rodata)
*(.rodata.*)
/* C read-write data of the SMM handler */
. = ALIGN(4);
*(.data)
*(.data.*)
/* C uninitialized data of the SMM handler */
. = ALIGN(4);
*(.bss)
*(.bss.*)
*(.sbss)
*(.sbss.*)
/* What is this? (Something we don't need with -fno-common.) */
*(COMMON)
. = ALIGN(4);
_eprogram = .;
}
/* We are using the ASEG interleaved to stuff the SMM handlers
* for all CPU cores in there. The jump table redirects the execution
* to the actual SMM handler
*/
. = 0xa8000 - (( CPUS - 1) * 0x400);
.jumptable : {
KEEP(*(.jumptable));
}
/DISCARD/ : {
*(.comment)
*(.note)
*(.note.*)
*(.eh_frame)
*(.debug_*)
}
}

View File

@@ -1,258 +0,0 @@
/* SPDX-License-Identifier: GPL-2.0-only */
/* NOTE: This handler assumes the SMM window goes from 0xa0000
* to 0xaffff. In fact, at least on Intel Core CPUs (i945 chipset)
* the SMM window is 128K big, covering 0xa0000 to 0xbffff.
* So there is a lot of potential for growth in here. Let's stick
* to 64k if we can though.
*/
#include <cpu/x86/lapic_def.h>
#include <cpu/x86/msr.h>
/*
* +--------------------------------+ 0xaffff
* | Save State Map Node 0 |
* | Save State Map Node 1 |
* | Save State Map Node 2 |
* | Save State Map Node 3 |
* | ... |
* +--------------------------------+ 0xaf000
* | |
* | |
* | |
* +--------------------------------+ 0xa8400
* | SMM Entry Node 0 (+ stack) |
* +--------------------------------+ 0xa8000
* | SMM Entry Node 1 (+ stack) |
* | SMM Entry Node 2 (+ stack) |
* | SMM Entry Node 3 (+ stack) |
* | ... |
* +--------------------------------+ 0xa7400
* | |
* | SMM Handler |
* | |
* +--------------------------------+ 0xa0000
*
*/
/* SMM_HANDLER_OFFSET is the 16bit offset within the ASEG
* at which smm_handler_start lives. At the moment the handler
* lives right at 0xa0000, so the offset is 0.
*/
#define SMM_HANDLER_OFFSET 0x0000
#if ENV_X86_64
.bss
ia32efer_backup_eax:
.long 0
ia32efer_backup_edx:
.long 0
#endif
/* initially SMM is some sort of real mode. Let gcc know
* how to treat the SMM handler stub
*/
.section ".handler", "a", @progbits
.code16
/**
* SMM code to enable protected mode and jump to the
* C-written function void smi_handler(u32 smm_revision)
*
* All the bad magic is not all that bad after all.
*/
#define SMM_START 0xa0000
#define SMM_END 0xb0000
#if SMM_END <= SMM_START
#error invalid SMM configuration
#endif
.global smm_handler_start
smm_handler_start:
#if CONFIG(SMM_LAPIC_REMAP_MITIGATION)
/* Check if the LAPIC register block overlaps with SMM.
* This block needs to work without data accesses because they
* may be routed into the LAPIC register block.
* Code accesses, on the other hand, are never routed to LAPIC,
* which is what makes this work in the first place.
*/
mov $LAPIC_BASE_MSR, %ecx
rdmsr
and $(~0xfff), %eax
sub $(SMM_START), %eax
cmp $(SMM_END - SMM_START), %eax
ja untampered_lapic
1:
/* emit "Crash" on serial */
mov $(CONFIG_TTYS0_BASE), %dx
mov $'C', %al
out %al, (%dx)
mov $'r', %al
out %al, (%dx)
mov $'a', %al
out %al, (%dx)
mov $'s', %al
out %al, (%dx)
mov $'h', %al
out %al, (%dx)
/* now crash for real */
ud2
untampered_lapic:
#endif
movw $(smm_gdtptr16 - smm_handler_start + SMM_HANDLER_OFFSET), %bx
lgdtl %cs:(%bx)
movl %cr0, %eax
andl $0x7FFAFFD1, %eax /* PG,AM,WP,NE,TS,EM,MP = 0 */
orl $0x60000001, %eax /* CD, NW, PE = 1 */
movl %eax, %cr0
/* Enable protected mode */
ljmpl $0x08, $1f
.code32
1:
/* flush the cache after disabling it */
wbinvd
/* Use flat data segment */
movw $0x10, %ax
movw %ax, %ds
movw %ax, %es
movw %ax, %ss
xor %ax, %ax /* zero out the gs and fs segment index */
movw %ax, %fs
movw %ax, %gs /* Will be used for cpu_info */
/* FIXME: Incompatible with X2APIC_SUPPORT. */
/* Get this CPU's LAPIC ID */
movl $(LAPIC_DEFAULT_BASE | LAPIC_ID), %esi
movl (%esi), %ecx
shr $24, %ecx
/* This is an ugly hack, and we should find a way to read the CPU index
* without relying on the LAPIC ID.
*/
/* calculate stack offset by multiplying the APIC ID
* by 1024 (0x400), and save that offset in ebp.
*/
shl $10, %ecx
movl %ecx, %ebp
/* We put the stack for each core right above
* its SMM entry point. Core 0 starts at 0xa8000,
* we spare 0x10 bytes for the jump to be sure.
*/
movl $0xa8010, %eax
subl %ecx, %eax /* subtract offset, see above */
movl %eax, %ebx /* Save bottom of stack in ebx */
#define SMM_STACK_SIZE (0x400 - 0x10)
/* clear stack */
cld
movl %eax, %edi
movl $(SMM_STACK_SIZE >> 2), %ecx
xorl %eax, %eax
rep stosl
/* set new stack */
addl $SMM_STACK_SIZE, %ebx
movl %ebx, %esp
#if ENV_X86_64
/* Backup IA32_EFER. Preserves ebx. */
movl $(IA32_EFER), %ecx
rdmsr
movl %eax, ia32efer_backup_eax
movl %edx, ia32efer_backup_edx
/* Enable long mode. Preserves ebx. */
#include <cpu/x86/64bit/entry64.inc>
#endif
/* Call C handler */
call smi_handler
#if ENV_X86_64
/*
* The only reason to go back to protected mode is that RSM doesn't restore
* MSR registers and MSR IA32_EFER was modified by entering long mode.
* Drop to protected mode to safely operate on the IA32_EFER MSR.
*/
/* Disable long mode. */
#include <cpu/x86/64bit/exit32.inc>
/* Restore IA32_EFER as RSM doesn't restore MSRs. */
movl $(IA32_EFER), %ecx
movl ia32efer_backup_eax, %eax
movl ia32efer_backup_edx, %edx
wrmsr
#endif
/* To return, just do rsm. It will "clean up" protected mode */
rsm
.code16
.align 4, 0xff
smm_gdtptr16:
.word smm_gdt_end - smm_gdt - 1
.long smm_gdt - smm_handler_start + 0xa0000 + SMM_HANDLER_OFFSET
.code32
smm_gdt:
/* The first GDT entry can not be used. Keep it zero */
.long 0x00000000, 0x00000000
/* gdt selector 0x08, flat code segment */
.word 0xffff, 0x0000
.byte 0x00, 0x9b, 0xcf, 0x00 /* G=1 and 0x0f, 4GB limit */
/* gdt selector 0x10, flat data segment */
.word 0xffff, 0x0000
.byte 0x00, 0x93, 0xcf, 0x00
/* gdt selector 0x18, flat code segment (64-bit) */
.word 0xffff, 0x0000
.byte 0x00, 0x9b, 0xaf, 0x00
smm_gdt_end:
.section ".jumptable", "a", @progbits
/* This is the SMM jump table. All cores use the same SMM handler
* for simplicity. But SMM Entry needs to be different due to the
* save state area. The jump table makes sure all CPUs jump into the
* real handler on SMM entry.
*/
/* This code currently supports up to 4 CPU cores. If more than 4 CPU cores
* shall be used, below table has to be updated, as well as smm.ld
*/
/* GNU AS/LD will always generate code that assumes CS is 0xa000. In reality
* CS will be set to SMM_BASE[19:4] though. Knowing that the smm handler is the
* first thing in the ASEG, we do a far jump here, to set CS to 0xa000.
*/
.code16
jumptable:
/* core 3 */
ljmp $0xa000, $SMM_HANDLER_OFFSET
.align 1024, 0x00
/* core 2 */
ljmp $0xa000, $SMM_HANDLER_OFFSET
.align 1024, 0x00
/* core 1 */
ljmp $0xa000, $SMM_HANDLER_OFFSET
.align 1024, 0x00
/* core 0 */
ljmp $0xa000, $SMM_HANDLER_OFFSET
.align 1024, 0x00

View File

@@ -1,163 +0,0 @@
/* SPDX-License-Identifier: GPL-2.0-only */
// FIXME: Is this piece of code southbridge specific, or
// can it be cleaned up so this include is not required?
// It's needed right now because we get our DEFAULT_PMBASE from
// here.
#if CONFIG(SOUTHBRIDGE_INTEL_I82801IX)
#include <southbridge/intel/i82801ix/i82801ix.h>
#else
#error "Southbridge needs SMM handler support."
#endif
// ADDR32() macro
#include <arch/registers.h>
#if !CONFIG(SMM_ASEG)
#error "Only use this file with ASEG."
#endif /* CONFIG_SMM_ASEG */
#define LAPIC_ID 0xfee00020
.global smm_relocation_start
.global smm_relocation_end
/* initially SMM is some sort of real mode. */
.code16
/**
* When starting up, x86 CPUs have their SMBASE set to 0x30000. However,
* this is not a good place for the SMM handler to live, so it needs to
* be relocated.
* Traditionally SMM handlers used to live in the A segment (0xa0000).
* With growing SMM handlers, more CPU cores, etc. CPU vendors started
* allowing to relocate the handler to the end of physical memory, which
* they refer to as TSEG.
* This trampoline code relocates SMBASE to base address - ( lapicid * 0x400 )
*
* Why 0x400? It is a safe value to cover the save state area per CPU. On
* current AMD CPUs this area is _documented_ to be 0x200 bytes. On Intel
* Core 2 CPUs the _documented_ parts of the save state area is 48 bytes
* bigger, effectively sizing our data structures 0x300 bytes.
*
* Example (with SMM handler living at 0xa0000):
*
* LAPICID SMBASE SMM Entry SAVE STATE
* 0 0xa0000 0xa8000 0xafd00
* 1 0x9fc00 0xa7c00 0xaf900
* 2 0x9f800 0xa7800 0xaf500
* 3 0x9f400 0xa7400 0xaf100
* 4 0x9f000 0xa7000 0xaed00
* 5 0x9ec00 0xa6c00 0xae900
* 6 0x9e800 0xa6800 0xae500
* 7 0x9e400 0xa6400 0xae100
* 8 0x9e000 0xa6000 0xadd00
* 9 0x9dc00 0xa5c00 0xad900
* 10 0x9d800 0xa5800 0xad500
* 11 0x9d400 0xa5400 0xad100
* 12 0x9d000 0xa5000 0xacd00
* 13 0x9cc00 0xa4c00 0xac900
* 14 0x9c800 0xa4800 0xac500
* 15 0x9c400 0xa4400 0xac100
* . . . .
* . . . .
* . . . .
* 31 0x98400 0xa0400 0xa8100
*
* With 32 cores, the SMM handler would need to fit between
* 0xa0000-0xa0400 and the stub plus stack would need to go
* at 0xa8000-0xa8100 (example for core 0). That is not enough.
*
* This means we're basically limited to 16 CPU cores before
* we need to move the SMM handler to TSEG.
*
* Note: Some versions of Pentium M need their SMBASE aligned to 32k.
* On those the above only works for up to 2 cores. But for now we only
* care fore Core (2) Duo/Solo
*
*/
smm_relocation_start:
/* Check revision to see if AMD64 style SMM_BASE
* Intel Core Solo/Duo: 0x30007
* Intel Core2 Solo/Duo: 0x30100
* Intel SandyBridge: 0x30101
* AMD64: 0x3XX64
* This check does not make much sense, unless someone ports
* SMI handling to AMD64 CPUs.
*/
mov $0x38000 + 0x7efc, %ebx
ADDR32(mov) (%ebx), %al
cmp $0x64, %al
je 1f
mov $0x38000 + 0x7ef8, %ebx
jmp smm_relocate
1:
mov $0x38000 + 0x7f00, %ebx
smm_relocate:
/* Get this CPU's LAPIC ID */
movl $LAPIC_ID, %esi
ADDR32(movl) (%esi), %ecx
shr $24, %ecx
/* calculate offset by multiplying the
* APIC ID by 1024 (0x400)
*/
movl %ecx, %edx
shl $10, %edx
movl $0xa0000, %eax
subl %edx, %eax /* subtract offset, see above */
ADDR32(movl) %eax, (%ebx)
/* The next section of code is potentially southbridge specific */
/* Clear SMI status */
movw $(DEFAULT_PMBASE + 0x34), %dx
inw %dx, %ax
outw %ax, %dx
/* Clear PM1 status */
movw $(DEFAULT_PMBASE + 0x00), %dx
inw %dx, %ax
outw %ax, %dx
/* Set EOS bit so other SMIs can occur */
movw $(DEFAULT_PMBASE + 0x30), %dx
inl %dx, %eax
orl $(1 << 1), %eax
outl %eax, %dx
/* End of southbridge specific section. */
#if CONFIG(DEBUG_SMM_RELOCATION)
/* print [SMM-x] so we can determine if CPUx went to SMM */
movw $CONFIG_TTYS0_BASE, %dx
mov $'[', %al
outb %al, %dx
mov $'S', %al
outb %al, %dx
mov $'M', %al
outb %al, %dx
outb %al, %dx
movb $'-', %al
outb %al, %dx
/* calculate ascii of CPU number. More than 9 cores? -> FIXME */
movb %cl, %al
addb $'0', %al
outb %al, %dx
mov $']', %al
outb %al, %dx
mov $'\r', %al
outb %al, %dx
mov $'\n', %al
outb %al, %dx
#endif
/* That's it. return */
rsm
smm_relocation_end: