[update] 整理cortex-a aarch32启动代码

1. 去除start_gcc.s中set_secondary_cpu_boot_address代码，这部分提取到qemu-vexpress-a9 bsp中。
2. 移动cpu.c中rt_hw_cpu_id函数到cp15_gcc.s，使用汇编实现，采用wake属性，方便bsp根据cpu特性获取CPU ID（多cpu集群中，不同厂家使用组合不一样）.
3. 整理start_gcc.s 适应多核启动，原来的代码只考虑到双核的情况。

bsp/qemu-vexpress-a9/drivers/secondary_cpu.c

@@ -26,10 +26,16 @@ static void rt_hw_timer2_isr(int vector, void *param)
     timer_clear_pending(0);
 }
 
+void set_secondary_cpu_boot_address(void)
+{
+    extern void secondary_cpu_start(void);
+    uint32_t *boot_address = (uint32_t *)0x10000030;
+    *(boot_address + 1) = ~0ul;
+    *boot_address = (uint32_t )&secondary_cpu_start;
+}
+
 void rt_hw_secondary_cpu_up(void)
 {
-    extern void set_secondary_cpu_boot_address(void);
-
     set_secondary_cpu_boot_address();
     __asm__ volatile ("dsb":::"memory");
     rt_hw_ipi_send(0, 1 << 1);

libcpu/arm/cortex-a/cp15.h

@@ -15,8 +15,7 @@
 #define __get_cp64(cp, op1, Rt, CRm) __asm__ volatile("MRRC p" # cp ", " # op1 ", %Q0, %R0, c" # CRm  : "=r" (Rt) : : "memory" )
 #define __set_cp64(cp, op1, Rt, CRm) __asm__ volatile("MCRR p" # cp ", " # op1 ", %Q0, %R0, c" # CRm  : : "r" (Rt) : "memory" )
 
-unsigned long rt_cpu_get_smp_id(void);
+int rt_hw_cpu_id(void);
-
 void rt_cpu_mmu_disable(void);
 void rt_cpu_mmu_enable(void);
 void rt_cpu_tlb_set(volatile unsigned long*);

libcpu/arm/cortex-a/cp15_gcc.S

@@ -8,9 +8,11 @@
  * 2013-07-05     Bernard      the first version
  */
 
-.globl rt_cpu_get_smp_id
+.weak rt_hw_cpu_id
-rt_cpu_get_smp_id:
+rt_hw_cpu_id:
-    mrc     p15, #0, r0, c0, c0, #5
+    mrc     p15, #0, r0, c0, c0, #5         @ read multiprocessor affinity register
+    ldr     r1, =0xFFFF03                   @ Affinity mask off, leaving CPU ID field, [0:1]CPU ID, [8:15]Cluster ID Aff1, [16:23]Cluster ID Aff2
+    and     r0, r0, r1
     bx      lr
 
 .globl rt_cpu_vector_set_base

libcpu/arm/cortex-a/cpu.c

@@ -15,17 +15,6 @@
 
 #ifdef RT_USING_SMP
 
-int rt_hw_cpu_id(void)
-{
-    int cpu_id;
-    __asm__ volatile (
-            "mrc p15, 0, %0, c0, c0, 5"
-            :"=r"(cpu_id)
-            );
-    cpu_id &= 0xf;
-    return cpu_id;
-};
-
 void rt_hw_spin_lock_init(rt_hw_spinlock_t *lock)
 {
     lock->slock = 0;

libcpu/arm/cortex-a/start_gcc.S

@@ -22,20 +22,26 @@
 .equ I_Bit,           0x80            @ when I bit is set, IRQ is disabled
 .equ F_Bit,           0x40            @ when F bit is set, FIQ is disabled
 
-#ifdef RT_USING_FPU
 .equ UND_Stack_Size,     0x00000400
-#else
-.equ UND_Stack_Size,     0x00000000
-#endif
 .equ SVC_Stack_Size,     0x00000400
-.equ ABT_Stack_Size,     0x00000000
+.equ ABT_Stack_Size,     0x00000400
 .equ RT_FIQ_STACK_PGSZ,  0x00000000
 .equ RT_IRQ_STACK_PGSZ,  0x00000800
 .equ USR_Stack_Size,     0x00000400
 
+.equ SUB_UND_Stack_Size,     0x00000400
+.equ SUB_SVC_Stack_Size,     0x00000400
+.equ SUB_ABT_Stack_Size,     0x00000400
+.equ SUB_RT_FIQ_STACK_PGSZ,  0x00000000
+.equ SUB_RT_IRQ_STACK_PGSZ,  0x00000400
+.equ SUB_USR_Stack_Size,     0x00000400
+
 #define ISR_Stack_Size  (UND_Stack_Size + SVC_Stack_Size + ABT_Stack_Size + \
                  RT_FIQ_STACK_PGSZ + RT_IRQ_STACK_PGSZ)
 
+#define SUB_ISR_Stack_Size  (SUB_UND_Stack_Size + SUB_SVC_Stack_Size + SUB_ABT_Stack_Size + \
+                 SUB_RT_FIQ_STACK_PGSZ + SUB_RT_IRQ_STACK_PGSZ)
+
 .section .data.share.isr
 /* stack */
 .globl stack_start
@@ -82,9 +88,40 @@ continue:
 
     /* disable the data alignment check */
     mrc p15, 0, r1, c1, c0, 0
-    bic r1, #(1<<1)
+    bic r1, #(1<<0)             /* Disable MMU */
+    bic r1, #(1<<1)             /* Disable Alignment fault checking */
+    bic r1, #(1<<2)             /* Disable data cache */
+    bic r1, #(1<<11)            /* Disable program flow prediction */
+    bic r1, #(1<<12)            /* Disable instruction cache */
+    bic r1, #(3<<19)            /* bit[20:19] must be zero */
     mcr p15, 0, r1, c1, c0, 0
 
+    @ get cpu id, and subtract the offset from the stacks base address
+    bl rt_hw_cpu_id
+    mov r5, r0
+
+    cmp     r5, #0              @ cpu id == 0
+    beq     normal_setup
+
+    @ cpu id > 0, stop or wait
+#ifdef RT_SMP_AUTO_BOOT
+    ldr r0, =secondary_cpu_entry
+    mov r1, #0
+    str r1, [r0] /* clean secondary_cpu_entry */
+#endif /* RT_SMP_AUTO_BOOT */
+
+secondary_loop:
+    @ cpu core 1 goes into sleep until core 0 wakeup it
+    wfe
+#ifdef RT_SMP_AUTO_BOOT
+    ldr r1, =secondary_cpu_entry
+    ldr r0, [r1]
+    cmp r0, #0
+    blxne r0 /* if(secondary_cpu_entry) secondary_cpu_entry(); */
+#endif /* RT_SMP_AUTO_BOOT */
+    b secondary_loop
+
+normal_setup:
     /* setup stack */
     bl      stack_setup
 
@@ -105,6 +142,11 @@ bss_loop:
     mcr p15, 0, r1, c1, c0, 1 //enable smp
 #endif
 
+    /* enable branch prediction */
+    mrc p15, 0, r0, c1, c0, 0
+    orr     r0, r0, #(1<<11)
+    mcr p15, 0, r0, c1, c0, 0
+
     /* initialize the mmu table and enable mmu */
     ldr r0, =platform_mem_desc
     ldr r1, =platform_mem_desc_size
@@ -137,6 +179,7 @@ stack_setup:
 
     @  Set the startup stack for svc
     mov     sp, r0
+    sub     r0, r0, #SVC_Stack_Size
 
     @  Enter Undefined Instruction Mode and set its Stack Pointer
     msr     cpsr_c, #Mode_UND|I_Bit|F_Bit
@@ -378,16 +421,6 @@ vector_resv:
     b       .
 
 #ifdef RT_USING_SMP
-.global set_secondary_cpu_boot_address
-set_secondary_cpu_boot_address:
-    ldr r0, =secondary_cpu_start
-
-    mvn r1, #0 //0xffffffff
-    ldr r2, =0x10000034
-    str r1, [r2]
-    str r0, [r2, #-4]
-    mov pc, lr
-
 .global secondary_cpu_start
 secondary_cpu_start:
 
@@ -405,38 +438,52 @@ secondary_cpu_start:
     bic r0, #(1<<13)
     mcr p15, 0, r0, c1, c0, 0
 
-#ifdef RT_USING_FPU
+    /* enable branch prediction */
-    cps #Mode_UND
+    mrc p15, 0, r0, c1, c0, 0
-    ldr sp, =und_stack_2_limit
+    orr     r0, r0, #(1<<11)
-#endif
+    mcr p15, 0, r0, c1, c0, 0
 
-    cps #Mode_IRQ
+    @ get cpu id, and subtract the offset from the stacks base address
-    ldr sp, =irq_stack_2_limit
+    bl rt_hw_cpu_id
+    sub r5, r0, #1
 
-    cps #Mode_FIQ
+    ldr r0, =SUB_ISR_Stack_Size
-    ldr sp, =irq_stack_2_limit
+    mul r0, r0, r5                      @r0 = SUB_ISR_Stack_Size * (cpuid - 1)
+    ldr r1, =sub_stack_top
+    sub r0, r1, r0                      @r0 = sub_stack_top - (SUB_ISR_Stack_Size * (cpuid - 1))
+
+    cps #Mode_SVC
+    mov     sp, r0
+    sub     r0, r0, #SUB_SVC_Stack_Size
+
+    cps #Mode_UND
+    mov     sp, r0
+    sub     r0, r0, #SUB_UND_Stack_Size
+
+    cps #Mode_ABT
+    mov     sp, r0
+    sub     r0, r0, #SUB_ABT_Stack_Size
+
+    cps #Mode_FIQ
+    mov     sp, r0
+    sub     r0, r0, #SUB_RT_FIQ_STACK_PGSZ
+
+    cps #Mode_IRQ
+    mov     sp, r0
+    sub     r0, r0, #SUB_RT_IRQ_STACK_PGSZ
 
     cps #Mode_SVC
-    ldr sp, =svc_stack_2_limit
 
     /* initialize the mmu table and enable mmu */
     bl rt_hw_mmu_init
 
     b secondary_cpu_c_start
-#endif
 
 .bss
 .align 2   //align to  2~2=4
-svc_stack_2:
-    .space (1 << 10)
-svc_stack_2_limit:
 
-irq_stack_2:
+sub_stack_start:
-    .space (1 << 10)
+    .space (SUB_ISR_Stack_Size * (RT_CPUS_NR-1))
-irq_stack_2_limit:
+sub_stack_top:
 
-#ifdef RT_USING_FPU
-und_stack_2:
-    .space (1 << 10)
-und_stack_2_limit:
 #endif