- refactored arm11/sys.c

- moved common.h from the ARM9 tree to the common code tree
- does proper deinit now on the ARM11 side

the bug that caused it to fail to launch some FIRMs has been fixed - it can even boot stock FIRMs

arm11/source/arm/mmu.c

@@ -1,4 +1,6 @@
 #include <types.h>
+#include <common.h>
+
 #include <arm.h>
 
 #include "arm/mmu.h"
@@ -68,44 +70,26 @@ static MMU_Lvl2_Table *Alloc_Lvl2(void)
 
 
 /* functions to convert from internal page flag format to ARM */
+
+/* {TEX, CB} */
+static const u8 MMU_TypeLUT[MEMORY_TYPES][2] = {
+	[STRONGLY_ORDERED] = {0, 0},
+	[NON_CACHEABLE] = {1, 0},
+	[DEVICE_SHARED] = {0, 1},
+	[DEVICE_NONSHARED] = {2, 0},
+	[CACHED_WT] = {0, 2},
+	[CACHED_WB] = {1, 3},
+	[CACHED_WB_ALLOC] = {1, 3},
+};
+
 static u32 MMU_GetTEX(u32 f)
 {
-	switch(MMU_FLAGS_TYPE(f)) {
-		default:
-		case STRONGLY_ORDERED:
-		case CACHED_WT:
-		case DEVICE_SHARED:
-			return 0;
-
-		case CACHED_WB:
-		case NON_CACHEABLE:
-		case CACHED_WB_ALLOC:
-			return 1;
-
-		case DEVICE_NONSHARED:
-			return 2;
-	}
+	return MMU_TypeLUT[MMU_FLAGS_TYPE(f)][0];
 }
 
 static u32 MMU_GetCB(u32 f)
 {
-	switch(MMU_FLAGS_TYPE(f)) {
-		default:
-		case STRONGLY_ORDERED:
-		case NON_CACHEABLE:
-		case DEVICE_NONSHARED:
-			return 0;
-
-		case DEVICE_SHARED:
-			return 1;
-
-		case CACHED_WT:
-			return 2;
-
-		case CACHED_WB:
-		case CACHED_WB_ALLOC:
-			return 3;
-	}
+	return MMU_TypeLUT[MMU_FLAGS_TYPE(f)][1];
 }
 
 static u32 MMU_GetAP(u32 f)
@@ -114,10 +98,8 @@ static u32 MMU_GetAP(u32 f)
 		default:
 		case NO_ACCESS:
 			return 0;
-
 		case READ_ONLY:
 			return 0x21;
-
 		case READ_WRITE:
 			return 0x01;
 	}
@@ -281,11 +263,10 @@ u32 MMU_Map(u32 va, u32 pa, u32 size, u32 flags)
 			.mapfn = MMU_MapPage,
 		},
 	};
-	static const size_t VMapperCount = sizeof(VMappers)/sizeof(*VMappers);
 
 	while(size > 0) {
 		size_t i = 0;
-		for (i = 0; i < VMapperCount; i++) {
+		for (i = 0; i < countof(VMappers); i++) {
 			u32 abits = VMappers[i].bits;
 
 			if (MMU_MappingFits(va, pa, size, abits)) {
@@ -303,7 +284,7 @@ u32 MMU_Map(u32 va, u32 pa, u32 size, u32 flags)
 			}
 		}
 
-		if (i == VMapperCount)
+		if (i == countof(VMappers))
 			return size;
 	}
 

arm11/source/arm/mmu.h

@@ -10,6 +10,7 @@ enum MMU_MemoryType {
 	CACHED_WT,
 	CACHED_WB,
 	CACHED_WB_ALLOC,
+	MEMORY_TYPES,
 };
 
 enum MMU_MemoryAccess {

arm11/source/boot.s

@@ -45,20 +45,24 @@ __boot:
     b 1b
 
 corezero_start:
+    @ assume __bss_len is 16 byte aligned
     ldr r0, =__bss_pa
     ldr r1, =__bss_len
     mov r2, #0
+    mov r3, #0
+    mov r4, #0
+    mov r5, #0
     add r1, r0, r1
     .Lclearbss:
         cmp r0, r1
-        strlt r2, [r0], #4
+        stmltia r0!, {r2-r5}
         blt .Lclearbss
 
+    bl SYS_CoreZeroInit
+
 coresmp_start:
     bl SYS_CoreInit
-    bl MPCoreMain
-
-    b __boot
+    b MainLoop
 
 .section .bss.stack
 .align 3

arm11/source/main.c

@@ -8,8 +8,9 @@
 #include "hw/i2c.h"
 #include "hw/mcu.h"
 
-static bool legacy;
-void SYS_CoreShutdown(void);
+#include "system/sys.h"
+
+static bool legacy = false;
 
 void PXI_RX_Handler(u32 __attribute__((unused)) irqn)
 {
@@ -98,21 +99,18 @@ void PXI_RX_Handler(u32 __attribute__((unused)) irqn)
 	return;
 }
 
-void MPCoreMain(void)
+void MainLoop(void)
 {
-	legacy = false;
 	GIC_Enable(IRQ_PXI_RX, BIT(0), GIC_HIGHEST_PRIO, PXI_RX_Handler);
 
+	// ARM9 won't try anything funny until this point
 	PXI_Barrier(ARM11_READY_BARRIER);
-	ARM_EnableInterrupts();
 
 	// Process IRQs until the ARM9 tells us it's time to boot something else
 	do {
 		ARM_WFI();
 	} while(!legacy);
 
-	// Perform any needed deinit stuff
-	ARM_DisableInterrupts();
-
+	SYS_CoreZeroShutdown();
 	SYS_CoreShutdown();
 }

arm11/source/system/sections.h

@@ -1,5 +1,7 @@
 #pragma once
 
+#include <types.h>
+
 #define DEF_SECT_(n)	extern u32 __##n##_pa, __##n##_va, __##n##_len; static const u32 n##_pa = (u32)&__##n##_pa, n##_va = (u32)&__##n##_va;
 DEF_SECT_(vector)
 DEF_SECT_(text)

arm11/source/system/sys.c

@@ -8,14 +8,13 @@
 
 #include "hw/i2c.h"
 
+#include "system/sections.h"
+
 #define CFG11_MPCORE_CLKCNT	((vu16*)(0x10141300))
 #define CFG11_SOCINFO		((vu16*)(0x10140FFC))
 
-#define LEGACY_BOOT_ENTRY	((vu32*)0x1FFFFFFC)
-
-#define INIT_DONE	(0xDDEEFFAA)
-
-static volatile u32 sys_init_state = 0;
+#define LEGACY_BOOT_ENTRYPOINT	((vu32*)0x1FFFFFFC)
+#define LEGACY_BOOT_ROUTINE_SMP	(0x0001004C)
 
 static bool SYS_IsNewConsole(void)
 {
@@ -45,16 +44,11 @@ static void SYS_EnableClkMult(void)
 	}
 }
 
-#include "sections.h"
-
-#define	MMU_FLAGS_DEF	MMU_FLAGS(STRONGLY_ORDERED, READ_WRITE, 0, 1)
-
-static void SYS_CoreZeroInit(void)
+void SYS_CoreZeroInit(void)
 {
 	GIC_GlobalReset();
-	GIC_LocalReset();
 
-	*LEGACY_BOOT_ENTRY = 0;
+	*LEGACY_BOOT_ENTRYPOINT = 0;
 
 	SYS_EnableClkMult();
 
@@ -77,11 +71,13 @@ static void SYS_CoreZeroInit(void)
 	MMU_Map(0x18000000, 0x18000000, 6UL << 20, MMU_FLAGS(CACHED_WT, READ_WRITE, 1, 1));
 
 	// FCRAM
-	MMU_Map(0x20000000, 0x20000000, 128UL << 20, MMU_FLAGS(CACHED_WB, READ_WRITE, 1, 1));
-}
+	if (SYS_IsNewConsole()) {
+		MMU_Map(0x20000000, 0x20000000, 256UL << 20, MMU_FLAGS(CACHED_WB, READ_WRITE, 1, 1));
+	} else {
+		MMU_Map(0x20000000, 0x20000000, 128UL << 20, MMU_FLAGS(CACHED_WB, READ_WRITE, 1, 1));
+	}
 
-static void SYS_InitPeripherals(void)
-{
+	// Initialize peripherals
 	PXI_Reset();
 	I2C_init();
 	//MCU_init();
@@ -89,19 +85,13 @@ static void SYS_InitPeripherals(void)
 
 void SYS_CoreInit(void)
 {
-	if (!ARM_CoreID()) {
-		SYS_CoreZeroInit();
-	} else {
-		while(sys_init_state != INIT_DONE)
-			ARM_WFE();
+	// Reset local GIC registers
+	GIC_LocalReset();
 
-		GIC_LocalReset();
-	}
-
-	// set up MMU registers
+	// Set up MMU registers
 	MMU_Init();
 
-	// enable fancy ARM11 stuff
+	// Enable fancy ARM11 features
 	ARM_SetACR(ARM_GetACR() |
 		ACR_RETSTK | ACR_DBPRED | ACR_SBPRED | ACR_FOLDING | ACR_SMP);
 
@@ -110,32 +100,20 @@ void SYS_CoreInit(void)
 
 	ARM_DSB();
 
-	if (!ARM_CoreID()) {
-		SYS_InitPeripherals();
-
-		sys_init_state = INIT_DONE;
-		ARM_DSB();
-		ARM_SEV();
-	}
+	ARM_EnableInterrupts();
+}
+
+void SYS_CoreZeroShutdown(void)
+{
+	ARM_DisableInterrupts();
+	GIC_GlobalReset();
 }
 
-// assumes all cores have been initialized
 void SYS_CoreShutdown(void)
 {
 	u32 core = ARM_CoreID();
 
-	if (!core) {
-		// wait for the other cores to do their thing
-		while(sys_init_state != (INIT_DONE - MAX_CPU + 1)) {
-			ARM_WFE();
-			ARM_DSB();
-		}
-
-		GIC_GlobalReset();
-	} else {
-		__atomic_sub_fetch(&sys_init_state, 1, __ATOMIC_SEQ_CST);
-		ARM_SEV();
-	}
+	ARM_DisableInterrupts();
 
 	GIC_LocalReset();
 
@@ -148,11 +126,11 @@ void SYS_CoreShutdown(void)
 		~(ACR_RETSTK | ACR_DBPRED | ACR_SBPRED | ACR_FOLDING | ACR_SMP));
 
 	if (!core) {
-		while(*LEGACY_BOOT_ENTRY == 0);
-		((void (*)(void))(*LEGACY_BOOT_ENTRY))();
+		while(*LEGACY_BOOT_ENTRYPOINT == 0);
+		((void (*)(void))(*LEGACY_BOOT_ENTRYPOINT))();
 	} else {
 		// Branch to bootrom function that does SMP reinit magic
 		// (waits for IPI + branches to word @ 0x1FFFFFDC)
-		((void (*)(void))0x0001004C)();
+		((void (*)(void))LEGACY_BOOT_ROUTINE_SMP)();
 	}
 }

arm11/source/system/sys.h

@@ -0,0 +1,20 @@
+#pragma once
+
+#include <types.h>
+
+/*
+ how to run the SYS_Core(Zero){Init,Shutdown} functions:
+ for init:
+  - FIRST run CoreZeroInit ONCE
+  - all cores must run CoreInit ONCE
+
+ for shutdown:
+  - all non-zero cores must call CoreShutdown
+  - core zero must call CoreZeroShutdown, then CoreShutdown
+*/
+
+void SYS_CoreZeroInit(void);
+void SYS_CoreInit(void);
+
+void SYS_CoreZeroShutdown(void);
+void SYS_CoreShutdown(void);

common/common.h

@@ -1,59 +1,81 @@
-#pragma once
-
-#include <inttypes.h>
-#include <stddef.h>
-#include <stdbool.h>
-#include <string.h>
-#include <ctype.h>
-#include <stdlib.h>
-#include <stdio.h>
-#include <types.h>
-#include <stdalign.h>
-
-#ifdef MONITOR_HEAP
-#include "mymalloc.h"
-#define malloc my_malloc
-#define free my_free
-#endif
-
-#define max(a,b) \
-    (((a) > (b)) ? (a) : (b))
-#define min(a,b) \
-    (((a) < (b)) ? (a) : (b))
-
-#define getbe16(d) \
-    (((d)[0]<<8) | (d)[1])
-#define getbe32(d) \
-    ((((u32) getbe16(d))<<16) | ((u32) getbe16(d+2)))
-#define getbe64(d) \
-    ((((u64) getbe32(d))<<32) | ((u64) getbe32(d+4)))
-
-#define getle16(d) \
-    (((d)[1]<<8) | (d)[0])
-#define getle32(d) \
-    ((((u32) getle16(d+2))<<16) | ((u32) getle16(d)))
-#define getle64(d) \
-    ((((u64) getle32(d+4))<<32) | ((u64) getle32(d)))
-
-#define align(v,a) \
-    (((v) % (a)) ? ((v) + (a) - ((v) % (a))) : (v))
-
-#define countof(x) \
-    (sizeof(x) / sizeof((x)[0]))
-    
-#define bkpt \
-    asm("bkpt\n\t")
-
-#define STATIC_ASSERT(...) \
-    _Static_assert((__VA_ARGS__), #__VA_ARGS__)
-    
-// standard output path (support file paths are in support.h)
-#define OUTPUT_PATH     "0:/gm9/out"
-
-// used in several places
-#define STD_BUFFER_SIZE     0x100000 // must be a multiple of 0x200
-
-// buffer area defines (in use by image.c, for RAMdrive)
-#define RAMDRV_BUFFER       ((u8*)0x22800000) // top of STACK
-#define RAMDRV_SIZE_O3DS    (0x5800000) // 88MB
-#define RAMDRV_SIZE_N3DS    (0xD800000) // 216MB
+#pragma once
+
+#include <inttypes.h>
+#include <stddef.h>
+#include <stdbool.h>
+#include <string.h>
+#include <ctype.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <stdalign.h>
+
+#include <types.h>
+
+#ifdef ARM9
+# ifdef MONITOR_HEAP
+#include "mymalloc.h"
+#define malloc my_malloc
+#define free my_free
+# endif
+#endif
+
+#define max(a,b) \
+    (((a) > (b)) ? (a) : (b))
+
+#define min(a,b) \
+    (((a) < (b)) ? (a) : (b))
+
+#define getbe16(d) \
+    (((d)[0]<<8) | (d)[1])
+#define getbe32(d) \
+    ((((u32) getbe16(d))<<16) | ((u32) getbe16((d)+2)))
+#define getbe64(d) \
+    ((((u64) getbe32(d))<<32) | ((u64) getbe32((d)+4)))
+
+#define getle16(d) \
+    (((d)[1]<<8) | (d)[0])
+#define getle32(d) \
+    ((((u32) getle16((d)+2))<<16) | ((u32) getle16(d)))
+#define getle64(d) \
+    ((((u64) getle32((d)+4))<<32) | ((u64) getle32(d)))
+
+#define align(v,a) \
+    (((v) % (a)) ? ((v) + (a) - ((v) % (a))) : (v))
+
+#define countof(x) \
+    (sizeof(x) / sizeof(*(x)))
+
+#define bkpt \
+    asm volatile("bkpt\n\t")
+
+#define STATIC_ASSERT(...) \
+    _Static_assert((__VA_ARGS__), #__VA_ARGS__)
+
+static inline u32 xbits(u32 *map, u32 start, u32 n)
+{
+    u32 ret, mask, off, shift;
+
+    if (n > 32)
+        return -1;
+
+    mask = ((u32)(1 << n)) - 1;
+    off = start / 32;
+    shift = start % 32;
+
+    ret = map[off] >> shift;
+    if ((n + shift) > 32)
+        ret |= map[off+1] << (32 - shift);
+    return ret & mask;
+}
+
+
+// standard output path (support file paths are in support.h)
+#define OUTPUT_PATH     "0:/gm9/out"
+
+// used in several places
+#define STD_BUFFER_SIZE     0x100000 // must be a multiple of 0x200
+
+// buffer area defines (in use by image.c, for RAMdrive)
+#define RAMDRV_BUFFER       ((u8*)0x22800000) // top of STACK
+#define RAMDRV_SIZE_O3DS    (0x5800000) // 88MB
+#define RAMDRV_SIZE_N3DS    (0xD800000) // 216MB