[bsp/rpi4] fixup startup failure (#8754)

[bsp/rpi4] fixup startup issues

Signed-off-by: Shell <smokewood@qq.com>

libcpu/aarch64/common/mmu.c

@@ -625,7 +625,9 @@ void *rt_ioremap_early(void *paddr, size_t size)
         tbl = rt_hw_mmu_tbl_get();
     }
 
-    count = (size + ARCH_SECTION_MASK) >> ARCH_SECTION_SHIFT;
+    /* get the total size required including overhead for alignment */
+    count = (size + ((rt_ubase_t)paddr & ARCH_SECTION_MASK)
+            + ARCH_SECTION_MASK) >> ARCH_SECTION_SHIFT;
     base = (rt_ubase_t)paddr & (~ARCH_SECTION_MASK);
 
     while (count --> 0)

libcpu/aarch64/common/stack.c

@@ -31,7 +31,8 @@ rt_uint8_t *rt_hw_stack_init(void *tentry, void *parameter,
 {
     rt_ubase_t *stk;
 
-    stk = (rt_ubase_t *)stack_addr;
+    /* The AAPCS64 requires 128-bit (16 byte) stack alignment */
+    stk = (rt_ubase_t*)RT_ALIGN_DOWN((rt_ubase_t)stack_addr, 16);
 
     for (int i = 0; i < 32; ++i)
     {

libcpu/aarch64/cortex-a/entry_point.S

@@ -69,7 +69,8 @@ _head:
     .quad   0               /* Reserved */
     .ascii  "ARM\x64"       /* Magic number */
     .long   0               /* Reserved (used for PE COFF offset) */
-#endif
+#endif /* RT_USING_OFW */
+
 /* Variable registers: x21~x28 */
 dtb_paddr .req x21
 boot_arg0 .req x22
@@ -109,6 +110,7 @@ _start:
     /* Now we are in the end of boot cpu process */
     ldr     x8, =rtthread_startup
     b       init_mmu_early
+    /* never come back */
 
 kernel_start:
     /* jump to the PE's system entry */
@@ -172,12 +174,15 @@ init_cpu_el:
     lsr     x0, x0, #2
     and     x0, x0, #3
 
+    /* running at EL3? */
     cmp     x0, #3
     bne     .init_cpu_hyp
 
+    /* should never be executed, just for completeness. (EL3) */
     mov     x1, #(1 << 0)           /* EL0 and EL1 are in Non-Secure state */
     orr     x1, x1, #(1 << 4)       /* RES1 */
     orr     x1, x1, #(1 << 5)       /* RES1 */
+    /* bic  x1, x1, #(1 << 7)          disable Secure Monitor Call */
     orr     x1, x1, #(1 << 10)      /* The next lower level is AArch64 */
     msr     scr_el3, x1
 
@@ -193,7 +198,8 @@ init_cpu_el:
     eret
 
 .init_cpu_hyp:
-    cmp     x0, #2                  /* EL1 = 0100  */
+    /* running at EL2? */
+    cmp     x0, #2                  /* EL2 = 0b10  */
     bne     .init_cpu_sys
 
     /* Enable CNTP for EL1 */
@@ -220,7 +226,6 @@ init_cpu_el:
 
 .init_cpu_sys:
     mrs     x0, sctlr_el1
-    orr     x0, x0, #(1 << 12)      /* Enable Instruction */
     bic     x0, x0, #(3 << 3)       /* Disable SP Alignment check */
     bic     x0, x0, #(1 << 1)       /* Disable Alignment check */
     msr     sctlr_el1, x0
@@ -299,7 +304,9 @@ enable_mmu_early:
 
     ldr     x30, =kernel_start      /* Set LR to kernel_start function, it's virtual addresses */
 
+    /* Enable page table translation */
     mrs     x1, sctlr_el1
+    orr     x1, x1, #(1 << 12)      /* Stage 1 instruction access Cacheability control */
     orr     x1, x1, #(1 << 2)       /* Cacheable Normal memory in stage1 */
     orr     x1, x1, #(1 << 0)       /* MMU Enable */
     msr     sctlr_el1, x1