summaryrefslogtreecommitdiffstats
path: root/arch/x86/cpu/sipi_vector.S
diff options
context:
space:
mode:
authorSimon Glass <sjg@chromium.org>2015-04-29 22:25:59 -0600
committerSimon Glass <sjg@chromium.org>2015-04-30 16:13:38 -0600
commit45b5a37836d552db30ab571d8ba67f12d7ba23b1 (patch)
treecbfda1472f00b123b6dd7872deedac0fc381612c /arch/x86/cpu/sipi_vector.S
parent6f41e0e7bcdc58e9e8af89988043893d5876bf7a (diff)
downloadtalos-obmc-uboot-45b5a37836d552db30ab571d8ba67f12d7ba23b1.tar.gz
talos-obmc-uboot-45b5a37836d552db30ab571d8ba67f12d7ba23b1.zip
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>
Diffstat (limited to 'arch/x86/cpu/sipi_vector.S')
-rw-r--r--arch/x86/cpu/sipi_vector.S216
1 files changed, 216 insertions, 0 deletions
diff --git a/arch/x86/cpu/sipi_vector.S b/arch/x86/cpu/sipi_vector.S
new file mode 100644
index 0000000000..bcef12c6f1
--- /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
OpenPOWER on IntegriCloud