Enable MPU and caches on the exception handler

An (uncachable) background region makes sure no bad accesses get caught

link.ld

@@ -17,7 +17,7 @@ SECTIONS
 
     __end__ = ABSOLUTE(.);
 
-    __stack_top = __start__;
+    __stack_abt = __start__;
-    __stack_abt = 0x8000;
+    __stack_top = __start__ - 0x80000;
     __code_size__ = __end__ - __start__;
 }

source/system/xrq.c

@@ -52,10 +52,9 @@ void XRQ_DumpRegisters(u32 xrq, u32 *regs)
     DsTime dstime;
     get_dstime(&dstime);
 
-    
     /* Dump registers */
     wstr += sprintf(wstr, "Exception: %s (%lu)\n", XRQ_Name[xrq&7], xrq);
-    wstr += sprintf(wstr, "20%02lX-%02lX-%02lX %02lX:%02lX:%02lX\n \n",
+    wstr += sprintf(wstr, "20%02lX-%02lX-%02lX %02lX:%02lX:%02lX\n\n",
         (u32) dstime.bcd_Y, (u32) dstime.bcd_M, (u32) dstime.bcd_D,
         (u32) dstime.bcd_h, (u32) dstime.bcd_m, (u32) dstime.bcd_s);
     for (int i = 0; i < 8; i++) {
@@ -65,7 +64,6 @@ void XRQ_DumpRegisters(u32 xrq, u32 *regs)
     }
     wstr += sprintf(wstr, "CPSR: %08lX\n\n", regs[16]);
 
-    
     /* Output registers to main screen */
     u32 draw_width = GetDrawStringWidth(dumpstr);
     u32 draw_height = GetDrawStringHeight(dumpstr);

source/system/xrq_handler.s

@@ -10,8 +10,9 @@
 #include <brf.h>
 
 .macro XRQ_FATAL id=0
-    adr sp, XRQ_Registers
+    ldr sp, =__stack_abt
-    stmia sp!, {r0-r12}
+    sub sp, sp, #(18*4)  @ Reserve space for registers
+    stmia sp, {r0-r12}
     mov r11, #\id
     b XRQ_MainHandler
 .endm
@@ -28,10 +29,8 @@ XRQ_Vectors:
     subs pc, lr, #4 @ IRQs are unhandled
     b .             @ FIQs are unused (except for debug?)
 
-XRQ_Registers:
-    .space (17*4)
-
 XRQ_Reset:
+    msr cpsr_c, #(SR_ABT_MODE | SR_IRQ | SR_FIQ)
     XRQ_FATAL 0
 
 XRQ_Undefined:
@@ -46,6 +45,7 @@ XRQ_PAbort:
 XRQ_DAbort:
     XRQ_FATAL 4
 
+@ r11 = exception number
 XRQ_MainHandler:
     mrs r10, cpsr
     mrs r9, spsr
@@ -60,21 +60,47 @@ XRQ_MainHandler:
     blx r6
 
     @ Retrieve banked registers
-    and r0, r9, #(SR_PMODE_MASK)
+    ands r0, r9, #(SR_PMODE_MASK & (0x0F))
-    cmp r0, #(SR_USR_MODE)
     orreq r0, #(SR_SYS_MODE)
-    orr r0, #(SR_IRQ | SR_FIQ)
+    orr r0, #(0x10 | SR_IRQ | SR_FIQ)
 
     msr cpsr_c, r0   @ Switch to previous mode
     mov r0, sp
     mov r1, lr
     msr cpsr_c, r10  @ Return to abort
 
-    stmia sp!, {r0,r1,r8,r9}
+    add r2, sp, #(13*4)
+    stmia r2, {r0,r1,r8,r9}
+
+    @ Give read/write access to all the memory regions
+    ldr r0, =0x33333333
+    mcr p15, 0, r0, c5, c0, 2 @ write data access
+    mcr p15, 0, r0, c5, c0, 3 @ write instruction access
+
+    @ Sets MPU regions and cache settings
+    adr r0, __abt_mpu_regions
+    ldmia r0, {r1-r8}
+    mov r0, #0b00110010 @ bootrom, arm9 mem and fcram are cacheable/bufferable
+    mcr p15, 0, r1, c6, c0, 0
+    mcr p15, 0, r2, c6, c1, 0
+    mcr p15, 0, r3, c6, c2, 0
+    mcr p15, 0, r4, c6, c3, 0
+    mcr p15, 0, r5, c6, c4, 0
+    mcr p15, 0, r6, c6, c5, 0
+    mcr p15, 0, r7, c6, c6, 0
+    mcr p15, 0, r8, c6, c7, 0
+    mcr p15, 0, r0, c3, c0, 0   @ Write bufferable 0, 2, 5
+    mcr p15, 0, r0, c2, c0, 0   @ Data cacheable 0, 2, 5
+    mcr p15, 0, r0, c2, c0, 1   @ Inst cacheable 0, 2, 5
+
+    @ Enable mpu/caches
+    ldr r1, =(CR_ENABLE_MPU | CR_ENABLE_DCACHE | CR_ENABLE_ICACHE | CR_ENABLE_DTCM)
+    mrc p15, 0, r0, c1, c0, 0
+    orr r0, r0, r1
+    mcr p15, 0, r0, c1, c0, 0
 
-    ldr sp, =0x8000
     ldr r2, =XRQ_DumpRegisters @ void XRQ_DumpRegisters(u32 xrq_id, u32 *regs)
-    adr r1, XRQ_Registers
+    mov r1, sp
     mov r0, r11
     blx r2
 
@@ -86,6 +112,15 @@ XRQ_MainHandler:
 
 .pool
 
+__abt_mpu_regions:
+    .word 0x0000003F @ 00000000 4G   | background region (includes IO regs)
+    .word 0xFFFF001F @ FFFF0000 64k  | bootrom (unprotected / protected)
+    .word 0x3000801B @ 30008000 16k  | dtcm
+    .word 0x00000035 @ 00000000 128M | itcm
+    .word 0x08000029 @ 08000000 2M   | arm9 mem (O3DS / N3DS)
+    .word 0x20000037 @ 20000000 256M | fcram (O3DS / N3DS)
+    .word 0x1FF00027 @ 1FF00000 1M   | dsp / axi wram
+    .word 0x1800002D @ 18000000 8M   | vram (+ 2MB)
+
 .global XRQ_End
 XRQ_End:
-    .word 0