Implement our own DCache cleaning functions

source/cache.h

@@ -0,0 +1,18 @@
+/*
+*   cache.h
+*       by TuxSH
+*/
+
+#pragma once
+#include "types.h"
+
+/***
+    The following functions flush the data cache, then waits for all memory transfers to be finished.
+    The data cache MUST be flushed before doing one of the following:
+        - rebooting
+        - powering down
+        - setting the ARM11 entrypoint to execute a function
+***/
+
+void flushEntireDCache(void);
+void flushDCacheRange(void *startAddress, u32 size);

source/cache.s

@@ -0,0 +1,51 @@
+@
+@   cache.s
+@       by TuxSH
+@
+@   This is part of Luma3DS, see LICENSE.txt for details
+@
+
+.text
+.arm
+.align 4
+
+.global flushEntireDCache
+.type flushEntireDCache, %function
+flushEntireDCache:
+    @ Adpated from http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0155a/ch03s03s05.html ,
+    @ and https://github.com/gemarcano/libctr9_io/blob/master/src/ctr_system_ARM.c#L39 as well
+    @ Note: ARM's example is actually for a 8KB DCache (which is what the 3DS has)
+    
+    mov r1, #0                          @ segment counter
+    outer_loop:
+        mov r0, #0                      @ line counter
+        
+        inner_loop:
+            orr r2, r1, r0                  @ generate segment and line address
+            mcr p15, 0, r2, c7, c14, 2      @ clean and flush the line
+            add r0, #0x20                   @ increment to next line
+            cmp r0, #0x400
+            bne inner_loop
+        
+        add r1, #0x40000000
+        cmp r1, #0
+        bne outer_loop
+    
+    mcr p15, 0, r0, c7, c10, 4              @ drain write buffer
+    bx lr
+    
+.global flushDCacheRange
+.type flushDCacheRange, %function
+flushDCacheRange:
+    add r1, r0, r1                      @ end address
+    bic r0, #0x1f                       @ align source address to cache line size (32 bytes)
+    
+    flush_range_loop:
+        mcr p15, 0, r0, c7, c14, 1      @ clean and flush the line corresponding to the address r0 is holding
+        add r0, #0x20
+        cmp r0, r1
+        bls flush_range_loop
+    
+    mcr p15, 0, r0, c7, c10, 4          @ drain write buffer
+    bx lr
+        

source/firm.c

@@ -8,6 +8,7 @@
 #include "fs.h"
 #include "patches.h"
 #include "memory.h"
+#include "cache.h"
 #include "emunand.h"
 #include "crypto.h"
 #include "draw.h"
@@ -378,7 +379,7 @@ static inline void launchFirm(FirmwareType firmType, u32 isFirmlaunch)
         arm11 = (u32 *)0x1FFFFFF8;
     }
 
-    cleanInvalidateDCacheAndDMB(); //Ensure that all memory transfers have completed and that the data cache has been flushed 
+    flushEntireDCache(); //Ensure that all memory transfers have completed and that the data cache has been flushed 
     
     //Set ARM11 kernel entrypoint
     *arm11 = (u32)firm->arm11Entry;

source/memory.c

@@ -60,9 +60,4 @@ u8 *memsearch(u8 *startPos, const void *pattern, u32 size, u32 patternSize)
     }
 
     return NULL;
-}
-
-void cleanInvalidateDCacheAndDMB(void)
-{
-    ((void (*)())0xFFFF0830)(); //Why write our own code when it's well implemented in the unprotected bootROM?
 }

source/memory.h

@@ -11,13 +11,4 @@
 void memcpy(void *dest, const void *src, u32 size);
 void memset32(void *dest, u32 filler, u32 size);
 int memcmp(const void *buf1, const void *buf2, u32 size);
-u8 *memsearch(u8 *startPos, const void *pattern, u32 size, u32 patternSize);
-
-/***
-    Cleans and invalidates the data cache, then waits for all memory transfers to be finished.
-    This function MUST be called before doing the following:
-        - rebooting
-        - powering down
-        - setting the ARM11 entrypoint to execute a function
-***/
-void cleanInvalidateDCacheAndDMB(void);
+u8 *memsearch(u8 *startPos, const void *pattern, u32 size, u32 patternSize);

source/screen.c

@@ -8,6 +8,7 @@
 #include "screen.h"
 #include "config.h"
 #include "memory.h"
+#include "cache.h"
 #include "draw.h"
 #include "i2c.h"
 
@@ -30,9 +31,11 @@ void  __attribute__((naked)) arm11Stub(void)
 static inline void invokeArm11Function(void (*func)())
 {
     static u32 hasCopiedStub = 0;
-    if(!hasCopiedStub++) memcpy((void *)ARM11_STUB_ADDRESS, arm11Stub, 0x40);
-    
-    cleanInvalidateDCacheAndDMB();
+    if(!hasCopiedStub++)
+    {
+        memcpy((void *)ARM11_STUB_ADDRESS, arm11Stub, 0x40);
+        flushDCacheRange((void *)ARM11_STUB_ADDRESS, 0x40);
+    }
     
     *arm11Entry = (u32)func;
     while(*arm11Entry);
@@ -77,6 +80,7 @@ void updateBrightness(u32 brightnessLevel)
         WAIT_FOR_ARM9();
     }
     
+    flushDCacheRange(&brightnessValue, 4);
     invokeArm11Function(ARM11);
 }
 
@@ -116,6 +120,7 @@ void clearScreens(void)
         WAIT_FOR_ARM9();
     }
     
+    flushDCacheRange(fb, sizeof(struct fb));
     invokeArm11Function(ARM11);
 }
 
@@ -223,6 +228,8 @@ u32 initScreens(void)
     
     if(needToInit)
     {
+        flushDCacheRange(&config, 4);
+        flushDCacheRange(fb, sizeof(struct fb));
         invokeArm11Function(ARM11);
 
         //Turn on backlight

source/utils.c

@@ -6,6 +6,7 @@
 #include "i2c.h"
 #include "buttons.h"
 #include "memory.h"
+#include "cache.h"
 
 u32 waitInput(void)
 {
@@ -36,7 +37,7 @@ u32 waitInput(void)
 
 void mcuReboot(void)
 {
-    cleanInvalidateDCacheAndDMB(); //Ensure that all memory transfers have completed and that the data cache has been flushed
+    flushEntireDCache(); //Ensure that all memory transfers have completed and that the data cache has been flushed
     
     i2cWriteRegister(I2C_DEV_MCU, 0x20, 1 << 2);
     while(1);