Merge with /home/stefan/git/u-boot/bamboo-nand

3 files changed, 349 insertions, 287 deletions

diff --git a/cpu/ppc4xx/44x_spd_ddr.c b/cpu/ppc4xx/44x_spd_ddr.c
index 10b4c18978..c500d3f242 100644
--- a/cpu/ppc4xx/44x_spd_ddr.c
+++ b/cpu/ppc4xx/44x_spd_ddr.c
@@ -20,7 +20,7 @@
  * Jun Gu, Artesyn Technology, jung@artesyncp.com
  * Support for AMCC 440 based on OpenBIOS draminit.c from IBM.
  *
- * (C) Copyright 2005
+ * (C) Copyright 2005-2007
  * Stefan Roese, DENX Software Engineering, sr@denx.de.
  *
  * See file CREDITS for list of people who contributed to this
@@ -42,6 +42,11 @@
  * MA 02111-1307 USA
  */
 
+/* define DEBUG for debugging output (obviously ;-)) */
+#if 0
+#define DEBUG
+#endif
+
 #include <common.h>
 #include <asm/processor.h>
 #include <i2c.h>
@@ -246,25 +251,6 @@
 #define MY_TLB_WORD2_I_ENABLE	TLB_WORD2_I_ENABLE	/* disable caching on SDRAM */
 #endif
 
-const unsigned long test[NUMMEMTESTS][NUMMEMWORDS] = {
-	{0x00000000, 0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0x00000000, 0x00000000,
-	 0xFFFFFFFF, 0xFFFFFFFF},
-	{0xFFFFFFFF, 0xFFFFFFFF, 0x00000000, 0x00000000, 0xFFFFFFFF, 0xFFFFFFFF,
-	 0x00000000, 0x00000000},
-	{0xAAAAAAAA, 0xAAAAAAAA, 0x55555555, 0x55555555, 0xAAAAAAAA, 0xAAAAAAAA,
-	 0x55555555, 0x55555555},
-	{0x55555555, 0x55555555, 0xAAAAAAAA, 0xAAAAAAAA, 0x55555555, 0x55555555,
-	 0xAAAAAAAA, 0xAAAAAAAA},
-	{0xA5A5A5A5, 0xA5A5A5A5, 0x5A5A5A5A, 0x5A5A5A5A, 0xA5A5A5A5, 0xA5A5A5A5,
-	 0x5A5A5A5A, 0x5A5A5A5A},
-	{0x5A5A5A5A, 0x5A5A5A5A, 0xA5A5A5A5, 0xA5A5A5A5, 0x5A5A5A5A, 0x5A5A5A5A,
-	 0xA5A5A5A5, 0xA5A5A5A5},
-	{0xAA55AA55, 0xAA55AA55, 0x55AA55AA, 0x55AA55AA, 0xAA55AA55, 0xAA55AA55,
-	 0x55AA55AA, 0x55AA55AA},
-	{0x55AA55AA, 0x55AA55AA, 0xAA55AA55, 0xAA55AA55, 0x55AA55AA, 0x55AA55AA,
-	 0xAA55AA55, 0xAA55AA55}
-};
-
 /* bank_parms is used to sort the bank sizes by descending order */
 struct bank_param {
 	unsigned long cr;
@@ -278,46 +264,37 @@ extern unsigned char cfg_simulate_spd_eeprom[128];
 #endif
 void program_tlb(u32 start, u32 size, u32 tlb_word2_i_value);
 
-unsigned char spd_read(uchar chip, uint addr);
-
-void get_spd_info(unsigned long* dimm_populated,
-		  unsigned char* iic0_dimm_addr,
-		  unsigned long	 num_dimm_banks);
-
-void check_mem_type
-(unsigned long* dimm_populated,
- unsigned char* iic0_dimm_addr,
- unsigned long	num_dimm_banks);
-
-void check_volt_type
-(unsigned long* dimm_populated,
- unsigned char* iic0_dimm_addr,
- unsigned long	num_dimm_banks);
-
-void program_cfg0(unsigned long* dimm_populated,
-		  unsigned char* iic0_dimm_addr,
-		  unsigned long	 num_dimm_banks);
-
-void program_cfg1(unsigned long* dimm_populated,
-		  unsigned char* iic0_dimm_addr,
-		  unsigned long	 num_dimm_banks);
-
-void program_rtr (unsigned long* dimm_populated,
-		  unsigned char* iic0_dimm_addr,
-		  unsigned long	 num_dimm_banks);
-
-void program_tr0 (unsigned long* dimm_populated,
-		  unsigned char* iic0_dimm_addr,
-		  unsigned long	 num_dimm_banks);
-
-void program_tr1 (void);
-
-void program_ecc (unsigned long	 num_bytes);
+static unsigned char spd_read(uchar chip, uint addr);
+static void get_spd_info(unsigned long *dimm_populated,
+			 unsigned char *iic0_dimm_addr,
+			 unsigned long num_dimm_banks);
+static void check_mem_type(unsigned long *dimm_populated,
+			   unsigned char *iic0_dimm_addr,
+			   unsigned long num_dimm_banks);
+static void check_volt_type(unsigned long *dimm_populated,
+			    unsigned char *iic0_dimm_addr,
+			    unsigned long num_dimm_banks);
+static void program_cfg0(unsigned long *dimm_populated,
+			 unsigned char *iic0_dimm_addr,
+			 unsigned long  num_dimm_banks);
+static void program_cfg1(unsigned long *dimm_populated,
+			 unsigned char *iic0_dimm_addr,
+			 unsigned long num_dimm_banks);
+static void program_rtr(unsigned long *dimm_populated,
+			unsigned char *iic0_dimm_addr,
+			unsigned long num_dimm_banks);
+static void program_tr0(unsigned long *dimm_populated,
+			unsigned char *iic0_dimm_addr,
+			unsigned long num_dimm_banks);
+static void program_tr1(void);
+
+#ifdef CONFIG_DDR_ECC
+static void program_ecc(unsigned long num_bytes);
+#endif
 
-unsigned
-long  program_bxcr(unsigned long* dimm_populated,
-		   unsigned char* iic0_dimm_addr,
-		   unsigned long  num_dimm_banks);
+static unsigned long program_bxcr(unsigned long *dimm_populated,
+				  unsigned char *iic0_dimm_addr,
+				  unsigned long num_dimm_banks);
 
 /*
  * This function is reading data from the DIMM module EEPROM over the SPD bus
@@ -328,7 +305,6 @@ long  program_bxcr(unsigned long* dimm_populated,
  * BUG: Don't handle ECC memory
  * BUG: A few values in the TR register is currently hardcoded
  */
-
 long int spd_sdram(void) {
 	unsigned char iic0_dimm_addr[] = SPD_EEPROM_ADDRESS;
 	unsigned long dimm_populated[sizeof(iic0_dimm_addr)];
@@ -421,9 +397,8 @@ long int spd_sdram(void) {
 	 */
 	while (1) {
 		mfsdram(mem_mcsts, mcsts);
-		if ((mcsts & SDRAM_MCSTS_MRSC) != 0) {
+		if ((mcsts & SDRAM_MCSTS_MRSC) != 0)
 			break;
-		}
 	}
 
 	/*
@@ -431,14 +406,17 @@ long int spd_sdram(void) {
 	 */
 	program_tr1();
 
+#ifdef CONFIG_DDR_ECC
 	/*
-	 * if ECC is enabled, initialize parity bits
+	 * If ecc is enabled, initialize the parity bits.
 	 */
+	program_ecc(total_size);
+#endif
 
 	return total_size;
 }
 
-unsigned char spd_read(uchar chip, uint addr)
+static unsigned char spd_read(uchar chip, uint addr)
 {
 	unsigned char data[2];
 
@@ -460,9 +438,9 @@ unsigned char spd_read(uchar chip, uint addr)
 	return 0;
 }
 
-void get_spd_info(unsigned long*   dimm_populated,
-		  unsigned char*   iic0_dimm_addr,
-		  unsigned long	   num_dimm_banks)
+static void get_spd_info(unsigned long *dimm_populated,
+			 unsigned char *iic0_dimm_addr,
+			 unsigned long num_dimm_banks)
 {
 	unsigned long dimm_num;
 	unsigned long dimm_found;
@@ -480,14 +458,10 @@ void get_spd_info(unsigned long*   dimm_populated,
 		if ((num_of_bytes != 0) && (total_size != 0)) {
 			dimm_populated[dimm_num] = TRUE;
 			dimm_found = TRUE;
-#if 0
-			printf("DIMM slot %lu: populated\n", dimm_num);
-#endif
+			debug("DIMM slot %lu: populated\n", dimm_num);
 		} else {
 			dimm_populated[dimm_num] = FALSE;
-#if 0
-			printf("DIMM slot %lu: Not populated\n", dimm_num);
-#endif
+			debug("DIMM slot %lu: Not populated\n", dimm_num);
 		}
 	}
 
@@ -497,9 +471,9 @@ void get_spd_info(unsigned long*   dimm_populated,
 	}
 }
 
-void check_mem_type(unsigned long*   dimm_populated,
-		    unsigned char*   iic0_dimm_addr,
-		    unsigned long    num_dimm_banks)
+static void check_mem_type(unsigned long *dimm_populated,
+			   unsigned char *iic0_dimm_addr,
+			   unsigned long num_dimm_banks)
 {
 	unsigned long dimm_num;
 	unsigned char dimm_type;
@@ -509,9 +483,7 @@ void check_mem_type(unsigned long*   dimm_populated,
 			dimm_type = spd_read(iic0_dimm_addr[dimm_num], 2);
 			switch (dimm_type) {
 			case 7:
-#if 0
-				printf("DIMM slot %lu: DDR SDRAM detected\n", dimm_num);
-#endif
+				debug("DIMM slot %lu: DDR SDRAM detected\n", dimm_num);
 				break;
 			default:
 				printf("ERROR: Unsupported DIMM detected in slot %lu.\n",
@@ -525,10 +497,9 @@ void check_mem_type(unsigned long*   dimm_populated,
 	}
 }
 
-
-void check_volt_type(unsigned long*   dimm_populated,
-		     unsigned char*   iic0_dimm_addr,
-		     unsigned long    num_dimm_banks)
+static void check_volt_type(unsigned long *dimm_populated,
+			    unsigned char *iic0_dimm_addr,
+			    unsigned long num_dimm_banks)
 {
 	unsigned long dimm_num;
 	unsigned long voltage_type;
@@ -541,18 +512,16 @@ void check_volt_type(unsigned long*   dimm_populated,
 				       dimm_num);
 				hang();
 			} else {
-#if 0
-				printf("DIMM %lu voltage level supported.\n", dimm_num);
-#endif
+				debug("DIMM %lu voltage level supported.\n", dimm_num);
 			}
 			break;
 		}
 	}
 }
 
-void program_cfg0(unsigned long* dimm_populated,
-		  unsigned char* iic0_dimm_addr,
-		  unsigned long	 num_dimm_banks)
+static void program_cfg0(unsigned long *dimm_populated,
+			 unsigned char *iic0_dimm_addr,
+			 unsigned long num_dimm_banks)
 {
 	unsigned long dimm_num;
 	unsigned long cfg0;
@@ -640,9 +609,9 @@ void program_cfg0(unsigned long* dimm_populated,
 	mtsdram(mem_cfg0, cfg0);
 }
 
-void program_cfg1(unsigned long* dimm_populated,
-		  unsigned char* iic0_dimm_addr,
-		  unsigned long	 num_dimm_banks)
+static void program_cfg1(unsigned long *dimm_populated,
+			 unsigned char *iic0_dimm_addr,
+			 unsigned long num_dimm_banks)
 {
 	unsigned long cfg1;
 	mfsdram(mem_cfg1, cfg1);
@@ -658,9 +627,9 @@ void program_cfg1(unsigned long* dimm_populated,
 	mtsdram(mem_cfg1, cfg1);
 }
 
-void program_rtr (unsigned long* dimm_populated,
-		  unsigned char* iic0_dimm_addr,
-		  unsigned long	 num_dimm_banks)
+static void program_rtr(unsigned long *dimm_populated,
+			unsigned char *iic0_dimm_addr,
+			unsigned long num_dimm_banks)
 {
 	unsigned long dimm_num;
 	unsigned long bus_period_x_10;
@@ -676,7 +645,6 @@ void program_rtr (unsigned long* dimm_populated,
 	get_sys_info(&sys_info);
 	bus_period_x_10 = ONE_BILLION / (sys_info.freqPLB / 10);
 
-
 	for (dimm_num = 0;  dimm_num < num_dimm_banks; dimm_num++) {
 		if (dimm_populated[dimm_num] == TRUE) {
 			refresh_rate_type = 0x7F & spd_read(iic0_dimm_addr[dimm_num], 12);
@@ -719,9 +687,9 @@ void program_rtr (unsigned long* dimm_populated,
 	mtsdram(mem_rtr, sdram_rtr);
 }
 
-void program_tr0 (unsigned long* dimm_populated,
-		  unsigned char* iic0_dimm_addr,
-		  unsigned long	 num_dimm_banks)
+static void program_tr0(unsigned long *dimm_populated,
+			 unsigned char *iic0_dimm_addr,
+			 unsigned long num_dimm_banks)
 {
 	unsigned long dimm_num;
 	unsigned long tr0;
@@ -1001,13 +969,73 @@ void program_tr0 (unsigned long* dimm_populated,
 		break;
 	}
 
-#if 0
-	printf("tr0: %x\n", tr0);
-#endif
+	debug("tr0: %x\n", tr0);
 	mtsdram(mem_tr0, tr0);
 }
 
-void program_tr1 (void)
+static int short_mem_test(void)
+{
+	unsigned long i, j;
+	unsigned long bxcr_num;
+	unsigned long *membase;
+	const unsigned long test[NUMMEMTESTS][NUMMEMWORDS] = {
+		{0x00000000, 0x00000000, 0xFFFFFFFF, 0xFFFFFFFF,
+		 0x00000000, 0x00000000, 0xFFFFFFFF, 0xFFFFFFFF},
+		{0xFFFFFFFF, 0xFFFFFFFF, 0x00000000, 0x00000000,
+		 0xFFFFFFFF, 0xFFFFFFFF, 0x00000000, 0x00000000},
+		{0xAAAAAAAA, 0xAAAAAAAA, 0x55555555, 0x55555555,
+		 0xAAAAAAAA, 0xAAAAAAAA, 0x55555555, 0x55555555},
+		{0x55555555, 0x55555555, 0xAAAAAAAA, 0xAAAAAAAA,
+		 0x55555555, 0x55555555, 0xAAAAAAAA, 0xAAAAAAAA},
+		{0xA5A5A5A5, 0xA5A5A5A5, 0x5A5A5A5A, 0x5A5A5A5A,
+		 0xA5A5A5A5, 0xA5A5A5A5, 0x5A5A5A5A, 0x5A5A5A5A},
+		{0x5A5A5A5A, 0x5A5A5A5A, 0xA5A5A5A5, 0xA5A5A5A5,
+		 0x5A5A5A5A, 0x5A5A5A5A, 0xA5A5A5A5, 0xA5A5A5A5},
+		{0xAA55AA55, 0xAA55AA55, 0x55AA55AA, 0x55AA55AA,
+		 0xAA55AA55, 0xAA55AA55, 0x55AA55AA, 0x55AA55AA},
+		{0x55AA55AA, 0x55AA55AA, 0xAA55AA55, 0xAA55AA55,
+		 0x55AA55AA, 0x55AA55AA, 0xAA55AA55, 0xAA55AA55}};
+
+	for (bxcr_num = 0; bxcr_num < MAXBXCR; bxcr_num++) {
+		mtdcr(memcfga, mem_b0cr + (bxcr_num << 2));
+		if ((mfdcr(memcfgd) & SDRAM_BXCR_SDBE) == SDRAM_BXCR_SDBE) {
+			/* Bank is enabled */
+			membase = (unsigned long*)
+				(mfdcr(memcfgd) & SDRAM_BXCR_SDBA_MASK);
+
+			/*
+			 * Run the short memory test
+			 */
+			for (i = 0; i < NUMMEMTESTS; i++) {
+				for (j = 0; j < NUMMEMWORDS; j++) {
+					membase[j] = test[i][j];
+					ppcDcbf((unsigned long)&(membase[j]));
+				}
+
+				for (j = 0; j < NUMMEMWORDS; j++) {
+					if (membase[j] != test[i][j]) {
+						ppcDcbf((unsigned long)&(membase[j]));
+						return 0;
+					}
+					ppcDcbf((unsigned long)&(membase[j]));
+				}
+
+				if (j < NUMMEMWORDS)
+					return 0;
+			}
+
+			/*
+			 * see if the rdclt value passed
+			 */
+			if (i < NUMMEMTESTS)
+				return 0;
+		}
+	}
+
+	return 1;
+}
+
+static void program_tr1(void)
 {
 	unsigned long tr0;
 	unsigned long tr1;
@@ -1015,8 +1043,7 @@ void program_tr1 (void)
 	unsigned long ecc_temp;
 	unsigned long dlycal;
 	unsigned long dly_val;
-	unsigned long i, j, k;
-	unsigned long bxcr_num;
+	unsigned long k;
 	unsigned long max_pass_length;
 	unsigned long current_pass_length;
 	unsigned long current_fail_length;
@@ -1029,7 +1056,6 @@ void program_tr1 (void)
 	unsigned char window_found;
 	unsigned char fail_found;
 	unsigned char pass_found;
-	unsigned long * membase;
 	PPC440_SYS_INFO sys_info;
 
 	/*
@@ -1079,55 +1105,16 @@ void program_tr1 (void)
 	window_found = FALSE;
 	fail_found = FALSE;
 	pass_found = FALSE;
-#ifdef DEBUG
-	printf("Starting memory test ");
-#endif
+	debug("Starting memory test ");
+
 	for (k = 0; k < NUMHALFCYCLES; k++) {
-		for (rdclt = 0; rdclt < dly_val; rdclt++)  {
+		for (rdclt = 0; rdclt < dly_val; rdclt++) {
 			/*
 			 * Set the timing reg for the test.
 			 */
 			mtsdram(mem_tr1, (tr1 | SDRAM_TR1_RDCT_ENCODE(rdclt)));
 
-			for (bxcr_num = 0; bxcr_num < MAXBXCR; bxcr_num++) {
-				mtdcr(memcfga, mem_b0cr + (bxcr_num<<2));
-				if ((mfdcr(memcfgd) & SDRAM_BXCR_SDBE) == SDRAM_BXCR_SDBE) {
-					/* Bank is enabled */
-					membase = (unsigned long*)
-						(mfdcr(memcfgd) & SDRAM_BXCR_SDBA_MASK);
-
-					/*
-					 * Run the short memory test
-					 */
-					for (i = 0; i < NUMMEMTESTS; i++) {
-						for (j = 0; j < NUMMEMWORDS; j++) {
-							membase[j] = test[i][j];
-							ppcDcbf((unsigned long)&(membase[j]));
-						}
-
-						for (j = 0; j < NUMMEMWORDS; j++) {
-							if (membase[j] != test[i][j]) {
-								ppcDcbf((unsigned long)&(membase[j]));
-								break;
-							}
-							ppcDcbf((unsigned long)&(membase[j]));
-						}
-
-						if (j < NUMMEMWORDS) {
-							break;
-						}
-					}
-
-					/*
-					 * see if the rdclt value passed
-					 */
-					if (i < NUMMEMTESTS) {
-						break;
-					}
-				}
-			}
-
-			if (bxcr_num == MAXBXCR) {
+			if (short_mem_test()) {
 				if (fail_found == TRUE) {
 					pass_found = TRUE;
 					if (current_pass_length == 0) {
@@ -1157,9 +1144,8 @@ void program_tr1 (void)
 				}
 			}
 		}
-#ifdef DEBUG
-		printf(".");
-#endif
+		debug(".");
+
 		if (window_found == TRUE) {
 			break;
 		}
@@ -1167,9 +1153,7 @@ void program_tr1 (void)
 		tr1 = tr1 ^ SDRAM_TR1_RDCD_MASK;
 		rdclt_offset += dly_val;
 	}
-#ifdef DEBUG
-	printf("\n");
-#endif
+	debug("\n");
 
 	/*
 	 * make sure we find the window
@@ -1218,18 +1202,17 @@ void program_tr1 (void)
 	}
 	tr1 |= SDRAM_TR1_RDCT_ENCODE(rdclt_average);
 
-#if 0
-	printf("tr1: %x\n", tr1);
-#endif
+	debug("tr1: %x\n", tr1);
+
 	/*
 	 * program SDRAM Timing Register 1 TR1
 	 */
 	mtsdram(mem_tr1, tr1);
 }
 
-unsigned long program_bxcr(unsigned long* dimm_populated,
-			   unsigned char* iic0_dimm_addr,
-			   unsigned long  num_dimm_banks)
+static unsigned long program_bxcr(unsigned long *dimm_populated,
+				  unsigned char *iic0_dimm_addr,
+				  unsigned long num_dimm_banks)
 {
 	unsigned long dimm_num;
 	unsigned long bank_base_addr;
@@ -1262,8 +1245,8 @@ unsigned long program_bxcr(unsigned long* dimm_populated,
 #ifdef CONFIG_BAMBOO
 	/*
 	 * This next section is hardware dependent and must be programmed
-	 * to match the hardware.  For bammboo, the following holds...
-	 * 1. SDRAM0_B0CR: Bank 0 of dimm 0 ctrl_bank_num : 0
+	 * to match the hardware.  For bamboo, the following holds...
+	 * 1. SDRAM0_B0CR: Bank 0 of dimm 0 ctrl_bank_num : 0 (soldered onboard)
 	 * 2. SDRAM0_B1CR: Bank 0 of dimm 1 ctrl_bank_num : 1
 	 * 3. SDRAM0_B2CR: Bank 1 of dimm 1 ctrl_bank_num : 1
 	 * 4. SDRAM0_B3CR: Bank 0 of dimm 2 ctrl_bank_num : 3
@@ -1273,10 +1256,12 @@ unsigned long program_bxcr(unsigned long* dimm_populated,
 	ctrl_bank_num[1] = 1;
 	ctrl_bank_num[2] = 3;
 #else
+	/*
+	 * Ocotea, Ebony and the other IBM/AMCC eval boards have
+	 * 2 DIMM slots with each max 2 banks
+	 */
 	ctrl_bank_num[0] = 0;
-	ctrl_bank_num[1] = 1;
-	ctrl_bank_num[2] = 2;
-	ctrl_bank_num[3] = 3;
+	ctrl_bank_num[1] = 2;
 #endif
 
 	/*
@@ -1290,6 +1275,8 @@ unsigned long program_bxcr(unsigned long* dimm_populated,
 			num_col_addr = spd_read(iic0_dimm_addr[dimm_num], 4);
 			num_banks    = spd_read(iic0_dimm_addr[dimm_num], 5);
 			bank_size_id = spd_read(iic0_dimm_addr[dimm_num], 31);
+			debug("DIMM%d: row=%d col=%d banks=%d\n", dimm_num,
+			      num_row_addr, num_col_addr, num_banks);
 
 			/*
 			 * Set the SDRAM0_BxCR regs
@@ -1354,9 +1341,12 @@ unsigned long program_bxcr(unsigned long* dimm_populated,
 			cr |= SDRAM_BXCR_SDBE;
 
  			for (i = 0; i < num_banks; i++) {
-				bank_parms[ctrl_bank_num[dimm_num]+i+dimm_num].bank_size_bytes =
-					(4 * 1024 * 1024) * bank_size_id;
-				bank_parms[ctrl_bank_num[dimm_num]+i+dimm_num].cr = cr;
+				bank_parms[ctrl_bank_num[dimm_num]+i].bank_size_bytes =
+					(4 << 20) * bank_size_id;
+				bank_parms[ctrl_bank_num[dimm_num]+i].cr = cr;
+				debug("DIMM%d-bank %d (SDRAM0_B%dCR): bank_size_bytes=%d\n",
+				      dimm_num, i, ctrl_bank_num[dimm_num]+i,
+				      bank_parms[ctrl_bank_num[dimm_num]+i].bank_size_bytes);
  			}
 		}
 	}
@@ -1400,13 +1390,15 @@ unsigned long program_bxcr(unsigned long* dimm_populated,
 				bank_parms[sorted_bank_num[bx_cr_num]].cr;
 			mtdcr(memcfgd, temp);
 			bank_base_addr += bank_parms[sorted_bank_num[bx_cr_num]].bank_size_bytes;
+			debug("SDRAM0_B%dCR=0x%08lx\n", sorted_bank_num[bx_cr_num], temp);
 		}
 	}
 
 	return(bank_base_addr);
 }
 
-void program_ecc (unsigned long	 num_bytes)
+#ifdef CONFIG_DDR_ECC
+static void program_ecc(unsigned long num_bytes)
 {
 	unsigned long bank_base_addr;
 	unsigned long current_address;
@@ -1425,14 +1417,12 @@ void program_ecc (unsigned long	 num_bytes)
 	bank_base_addr = CFG_SDRAM_BASE;
 
 	if ((cfg0 & SDRAM_CFG0_MCHK_MASK) != SDRAM_CFG0_MCHK_NON) {
-		mtsdram(mem_cfg0, (cfg0 & ~SDRAM_CFG0_MCHK_MASK) |
-			SDRAM_CFG0_MCHK_GEN);
+		mtsdram(mem_cfg0, (cfg0 & ~SDRAM_CFG0_MCHK_MASK) | SDRAM_CFG0_MCHK_GEN);
 
-		if ((cfg0 & SDRAM_CFG0_DMWD_MASK) == SDRAM_CFG0_DMWD_32) {
+		if ((cfg0 & SDRAM_CFG0_DMWD_MASK) == SDRAM_CFG0_DMWD_32)
 			address_increment = 4;
-		} else {
+		else
 			address_increment = 8;
-		}
 
 		current_address = (unsigned long)(bank_base_addr);
 		end_address = (unsigned long)(bank_base_addr) + num_bytes;
@@ -1446,4 +1436,5 @@ void program_ecc (unsigned long	 num_bytes)
 			SDRAM_CFG0_MCHK_CHK);
 	}
 }
+#endif /* CONFIG_DDR_ECC */
 #endif /* CONFIG_SPD_EEPROM */
diff --git a/cpu/ppc4xx/ndfc.c b/cpu/ppc4xx/ndfc.c
index 08dfc32952..f63fc79f6c 100644
--- a/cpu/ppc4xx/ndfc.c
+++ b/cpu/ppc4xx/ndfc.c
@@ -3,7 +3,7 @@
  *   Platform independend driver for NDFC (NanD Flash Controller)
  *   integrated into EP440 cores
  *
- * (C) Copyright 2006
+ * (C) Copyright 2006-2007
  * Stefan Roese, DENX Software Engineering, sr@denx.de.
  *
  * Based on original work by
@@ -38,7 +38,9 @@
 
 #include <nand.h>
 #include <linux/mtd/ndfc.h>
+#include <linux/mtd/nand_ecc.h>
 #include <asm/processor.h>
+#include <asm/io.h>
 #include <ppc4xx.h>
 
 static u8 hwctl = 0;
@@ -70,11 +72,11 @@ static void ndfc_write_byte(struct mtd_info *mtdinfo, u_char byte)
 	ulong base = (ulong) this->IO_ADDR_W & 0xfffffffc;
 
 	if (hwctl & 0x1)
-		out8(base + NDFC_CMD, byte);
+		out_8((u8 *)(base + NDFC_CMD), byte);
 	else if (hwctl & 0x2)
-		out8(base + NDFC_ALE, byte);
+		out_8((u8 *)(base + NDFC_ALE), byte);
 	else
-		out8(base + NDFC_DATA, byte);
+		out_8((u8 *)(base + NDFC_DATA), byte);
 }
 
 static u_char ndfc_read_byte(struct mtd_info *mtdinfo)
@@ -82,7 +84,7 @@ static u_char ndfc_read_byte(struct mtd_info *mtdinfo)
 	struct nand_chip *this = mtdinfo->priv;
 	ulong base = (ulong) this->IO_ADDR_W & 0xfffffffc;
 
-	return (in8(base + NDFC_DATA));
+	return (in_8((u8 *)(base + NDFC_DATA)));
 }
 
 static int ndfc_dev_ready(struct mtd_info *mtdinfo)
@@ -90,17 +92,41 @@ static int ndfc_dev_ready(struct mtd_info *mtdinfo)
 	struct nand_chip *this = mtdinfo->priv;
 	ulong base = (ulong) this->IO_ADDR_W & 0xfffffffc;
 
-	while (!(in32(base + NDFC_STAT) & NDFC_STAT_IS_READY))
+	while (!(in_be32((u32 *)(base + NDFC_STAT)) & NDFC_STAT_IS_READY))
 		;
 
 	return 1;
 }
 
-#ifndef CONFIG_NAND_SPL
-/*
- * Don't use these speedup functions in NAND boot image, since the image
- * has to fit into 4kByte.
- */
+static void ndfc_enable_hwecc(struct mtd_info *mtdinfo, int mode)
+{
+	struct nand_chip *this = mtdinfo->priv;
+	ulong base = (ulong) this->IO_ADDR_W & 0xfffffffc;
+	u32 ccr;
+
+	ccr = in_be32((u32 *)(base + NDFC_CCR));
+	ccr |= NDFC_CCR_RESET_ECC;
+	out_be32((u32 *)(base + NDFC_CCR), ccr);
+}
+
+static int ndfc_calculate_ecc(struct mtd_info *mtdinfo,
+			      const u_char *dat, u_char *ecc_code)
+{
+	struct nand_chip *this = mtdinfo->priv;
+	ulong base = (ulong) this->IO_ADDR_W & 0xfffffffc;
+	u32 ecc;
+	u8 *p = (u8 *)&ecc;
+
+	ecc = in_be32((u32 *)(base + NDFC_ECC));
+
+	/* The NDFC uses Smart Media (SMC) bytes order
+	 */
+	ecc_code[0] = p[2];
+	ecc_code[1] = p[1];
+	ecc_code[2] = p[3];
+
+	return 0;
+}
 
 /*
  * Speedups for buffer read/write/verify
@@ -116,9 +142,14 @@ static void ndfc_read_buf(struct mtd_info *mtdinfo, uint8_t *buf, int len)
 	uint32_t *p = (uint32_t *) buf;
 
 	for (;len > 0; len -= 4)
-		*p++ = in32(base + NDFC_DATA);
+		*p++ = in_be32((u32 *)(base + NDFC_DATA));
 }
 
+#ifndef CONFIG_NAND_SPL
+/*
+ * Don't use these speedup functions in NAND boot image, since the image
+ * has to fit into 4kByte.
+ */
 static void ndfc_write_buf(struct mtd_info *mtdinfo, const uint8_t *buf, int len)
 {
 	struct nand_chip *this = mtdinfo->priv;
@@ -126,7 +157,7 @@ static void ndfc_write_buf(struct mtd_info *mtdinfo, const uint8_t *buf, int len
 	uint32_t *p = (uint32_t *) buf;
 
 	for (; len > 0; len -= 4)
-		out32(base + NDFC_DATA, *p++);
+		out_be32((u32 *)(base + NDFC_DATA), *p++);
 }
 
 static int ndfc_verify_buf(struct mtd_info *mtdinfo, const uint8_t *buf, int len)
@@ -136,7 +167,7 @@ static int ndfc_verify_buf(struct mtd_info *mtdinfo, const uint8_t *buf, int len
 	uint32_t *p = (uint32_t *) buf;
 
 	for (; len > 0; len -= 4)
-		if (*p++ != in32(base + NDFC_DATA))
+		if (*p++ != in_be32((u32 *)(base + NDFC_DATA)))
 			return -1;
 
 	return 0;
@@ -153,8 +184,8 @@ void board_nand_select_device(struct nand_chip *nand, int chip)
 	ulong base = (ulong)nand->IO_ADDR_W & 0xfffffffc;
 
 	/* Set NandFlash Core Configuration Register */
-	/* 1col x 2 rows */
-	out32(base + NDFC_CCR, 0x00000000 | (cs << 24));
+	/* 1 col x 2 rows */
+	out_be32((u32 *)(base + NDFC_CCR), 0x00000000 | (cs << 24));
 }
 
 int board_nand_init(struct nand_chip *nand)
@@ -162,16 +193,19 @@ int board_nand_init(struct nand_chip *nand)
 	int cs = (ulong)nand->IO_ADDR_W & 0x00000003;
 	ulong base = (ulong)nand->IO_ADDR_W & 0xfffffffc;
 
-	nand->eccmode = NAND_ECC_SOFT;
-
 	nand->hwcontrol  = ndfc_hwcontrol;
 	nand->read_byte  = ndfc_read_byte;
+	nand->read_buf   = ndfc_read_buf;
 	nand->write_byte = ndfc_write_byte;
 	nand->dev_ready  = ndfc_dev_ready;
 
+	nand->eccmode = NAND_ECC_HW3_256;
+	nand->enable_hwecc = ndfc_enable_hwecc;
+	nand->calculate_ecc = ndfc_calculate_ecc;
+	nand->correct_data = nand_correct_data;
+
 #ifndef CONFIG_NAND_SPL
 	nand->write_buf  = ndfc_write_buf;
-	nand->read_buf   = ndfc_read_buf;
 	nand->verify_buf = ndfc_verify_buf;
 #else
 	/*
@@ -187,7 +221,7 @@ int board_nand_init(struct nand_chip *nand)
 	 * Select required NAND chip in NDFC
 	 */
 	board_nand_select_device(nand, cs);
-	out32(base + NDFC_BCFG0 + (cs << 2), 0x80002222);
+	out_be32((u32 *)(base + NDFC_BCFG0 + (cs << 2)), 0x80002222);
 	return 0;
 }
 
diff --git a/cpu/ppc4xx/start.S b/cpu/ppc4xx/start.S
index 85660b40dc..78de30031d 100644
--- a/cpu/ppc4xx/start.S
+++ b/cpu/ppc4xx/start.S
@@ -110,6 +110,13 @@
 # endif
 #endif /* CFG_INIT_DCACHE_CS */
 
+#define function_prolog(func_name)      .text; \
+					.align 2; \
+					.globl func_name; \
+					func_name:
+#define function_epilog(func_name)      .type func_name,@function; \
+					.size func_name,.-func_name
+
 /* We don't want the  MMU yet.
 */
 #undef	MSR_KERNEL
@@ -388,8 +395,9 @@ rsttlb:	tlbwe	r0,r1,0x0000	/* Invalidate all entries (V=0)*/
 2:
 
 #if defined(CONFIG_NAND_SPL)
+#if defined(CONFIG_440EPX) || defined(CONFIG_440GRX)
 	/*
-	 * Enable internal SRAM
+	 * Enable internal SRAM (only on 440EPx/GRx, 440EP/GR have no OCM)
 	 */
 	lis	r2,0x7fff
 	ori	r2,r2,0xffff
@@ -399,6 +407,45 @@ rsttlb:	tlbwe	r0,r1,0x0000	/* Invalidate all entries (V=0)*/
 	mfdcr	r1,isram0_pmeg
 	and	r1,r1,r2		/* Disable pwr mgmt */
 	mtdcr	isram0_pmeg,r1
+#endif
+#if defined(CONFIG_440EP)
+	/*
+	 * On 440EP with no internal SRAM, we setup SDRAM very early
+	 * and copy the NAND_SPL to SDRAM and jump to it
+	 */
+	/* Clear Dcache to use as RAM */
+	addis	r3,r0,CFG_INIT_RAM_ADDR@h
+	ori	r3,r3,CFG_INIT_RAM_ADDR@l
+	addis	r4,r0,CFG_INIT_RAM_END@h
+	ori	r4,r4,CFG_INIT_RAM_END@l
+	rlwinm. r5,r4,0,27,31
+	rlwinm	r5,r4,27,5,31
+	beq	..d_ran3
+	addi	r5,r5,0x0001
+..d_ran3:
+	mtctr	r5
+..d_ag3:
+	dcbz	r0,r3
+	addi	r3,r3,32
+	bdnz	..d_ag3
+	/*----------------------------------------------------------------*/
+	/* Setup the stack in internal SRAM */
+	/*----------------------------------------------------------------*/
+	lis	r1,CFG_INIT_RAM_ADDR@h
+	ori	r1,r1,CFG_INIT_SP_OFFSET@l
+	li	r0,0
+	stwu	r0,-4(r1)
+	stwu	r0,-4(r1)		/* Terminate call chain */
+
+	stwu	r1,-8(r1)		/* Save back chain and move SP */
+	lis	r0,RESET_VECTOR@h	/* Address of reset vector */
+	ori	r0,r0, RESET_VECTOR@l
+	stwu	r1,-8(r1)		/* Save back chain and move SP */
+	stw	r0,+12(r1)		/* Save return addr (underflow vect) */
+	sync
+	bl	early_sdram_init
+	sync
+#endif /* CONFIG_440EP */
 
 	/*
 	 * Copy SPL from cache into internal SRAM
@@ -429,7 +476,7 @@ spl_loop:
 start_ram:
 	sync
 	isync
-#endif
+#endif /* CONFIG_NAND_SPL */
 
 	bl	3f
 	b	_start
@@ -1137,7 +1184,6 @@ crit_return:
 	lwz	r1,GPR1(r1)
 	SYNC
 	rfci
-#endif /* CONFIG_NAND_SPL */
 
 /* Cache functions.
 */
@@ -1255,24 +1301,6 @@ wr_tcr:
 	blr
 
 /*------------------------------------------------------------------------------- */
-/* Function:	 in8 */
-/* Description:	 Input 8 bits */
-/*------------------------------------------------------------------------------- */
-	.globl	in8
-in8:
-	lbz	r3,0x0000(r3)
-	blr
-
-/*------------------------------------------------------------------------------- */
-/* Function:	 out8 */
-/* Description:	 Output 8 bits */
-/*------------------------------------------------------------------------------- */
-	.globl	out8
-out8:
-	stb	r4,0x0000(r3)
-	blr
-
-/*------------------------------------------------------------------------------- */
 /* Function:	 out16 */
 /* Description:	 Output 16 bits */
 /*------------------------------------------------------------------------------- */
@@ -1291,15 +1319,6 @@ out16r:
 	blr
 
 /*------------------------------------------------------------------------------- */
-/* Function:	 out32 */
-/* Description:	 Output 32 bits */
-/*------------------------------------------------------------------------------- */
-	.globl	out32
-out32:
-	stw	r4,0x0000(r3)
-	blr
-
-/*------------------------------------------------------------------------------- */
 /* Function:	 out32r */
 /* Description:	 Byte reverse and output 32 bits */
 /*------------------------------------------------------------------------------- */
@@ -1327,15 +1346,6 @@ in16r:
 	blr
 
 /*------------------------------------------------------------------------------- */
-/* Function:	 in32 */
-/* Description:	 Input 32 bits */
-/*------------------------------------------------------------------------------- */
-	.globl	in32
-in32:
-	lwz	3,0x0000(3)
-	blr
-
-/*------------------------------------------------------------------------------- */
 /* Function:	 in32r */
 /* Description:	 Input 32 bits and byte reverse */
 /*------------------------------------------------------------------------------- */
@@ -1377,9 +1387,6 @@ ppcSync:
 	sync
 	blr
 
-/*------------------------------------------------------------------------------*/
-
-#ifndef CONFIG_NAND_SPL
 /*
  * void relocate_code (addr_sp, gd, addr_moni)
  *
@@ -1644,8 +1651,88 @@ trap_reloc:
 	stw	r0, 4(r7)
 
 	blr
+
+#if defined(CONFIG_440)
+/*----------------------------------------------------------------------------+
+| dcbz_area.
++----------------------------------------------------------------------------*/
+	function_prolog(dcbz_area)
+	rlwinm. r5,r4,0,27,31
+	rlwinm  r5,r4,27,5,31
+	beq     ..d_ra2
+	addi    r5,r5,0x0001
+..d_ra2:mtctr   r5
+..d_ag2:dcbz    r0,r3
+	addi    r3,r3,32
+	bdnz    ..d_ag2
+	sync
+	blr
+	function_epilog(dcbz_area)
+
+/*----------------------------------------------------------------------------+
+| dflush.  Assume 32K at vector address is cachable.
++----------------------------------------------------------------------------*/
+	function_prolog(dflush)
+	mfmsr   r9
+	rlwinm  r8,r9,0,15,13
+	rlwinm  r8,r8,0,17,15
+	mtmsr   r8
+	addi    r3,r0,0x0000
+	mtspr   dvlim,r3
+	mfspr   r3,ivpr
+	addi    r4,r0,1024
+	mtctr   r4
+..dflush_loop:
+	lwz     r6,0x0(r3)
+	addi    r3,r3,32
+	bdnz    ..dflush_loop
+	addi    r3,r3,-32
+	mtctr   r4
+..ag:   dcbf    r0,r3
+	addi    r3,r3,-32
+	bdnz    ..ag
+	sync
+	mtmsr   r9
+	blr
+	function_epilog(dflush)
+#endif /* CONFIG_440 */
 #endif /* CONFIG_NAND_SPL */
 
+/*------------------------------------------------------------------------------- */
+/* Function:	 in8 */
+/* Description:	 Input 8 bits */
+/*------------------------------------------------------------------------------- */
+	.globl	in8
+in8:
+	lbz	r3,0x0000(r3)
+	blr
+
+/*------------------------------------------------------------------------------- */
+/* Function:	 out8 */
+/* Description:	 Output 8 bits */
+/*------------------------------------------------------------------------------- */
+	.globl	out8
+out8:
+	stb	r4,0x0000(r3)
+	blr
+
+/*------------------------------------------------------------------------------- */
+/* Function:	 out32 */
+/* Description:	 Output 32 bits */
+/*------------------------------------------------------------------------------- */
+	.globl	out32
+out32:
+	stw	r4,0x0000(r3)
+	blr
+
+/*------------------------------------------------------------------------------- */
+/* Function:	 in32 */
+/* Description:	 Input 32 bits */
+/*------------------------------------------------------------------------------- */
+	.globl	in32
+in32:
+	lwz	3,0x0000(3)
+	blr
 
 /**************************************************************************/
 /* PPC405EP specific stuff						  */
@@ -1892,13 +1979,6 @@ pll_wait:
 #endif /* CONFIG_405EP */
 
 #if defined(CONFIG_440)
-#define function_prolog(func_name)	.text; \
-					.align 2; \
-					.globl func_name; \
-					func_name:
-#define function_epilog(func_name)	.type func_name,@function; \
-					.size func_name,.-func_name
-
 /*----------------------------------------------------------------------------+
 | mttlb3.
 +----------------------------------------------------------------------------*/
@@ -1946,47 +2026,4 @@ pll_wait:
 	TLBRE(3,3,0)
 	blr
 	function_epilog(mftlb1)
-
-/*----------------------------------------------------------------------------+
-| dcbz_area.
-+----------------------------------------------------------------------------*/
-	function_prolog(dcbz_area)
-	rlwinm. r5,r4,0,27,31
-	rlwinm	r5,r4,27,5,31
-	beq	..d_ra2
-	addi	r5,r5,0x0001
-..d_ra2:mtctr	r5
-..d_ag2:dcbz	r0,r3
-	addi	r3,r3,32
-	bdnz	..d_ag2
-	sync
-	blr
-	function_epilog(dcbz_area)
-
-/*----------------------------------------------------------------------------+
-| dflush.  Assume 32K at vector address is cachable.
-+----------------------------------------------------------------------------*/
-	function_prolog(dflush)
-	mfmsr	r9
-	rlwinm	r8,r9,0,15,13
-	rlwinm	r8,r8,0,17,15
-	mtmsr	r8
-	addi	r3,r0,0x0000
-	mtspr	dvlim,r3
-	mfspr	r3,ivpr
-	addi	r4,r0,1024
-	mtctr	r4
-..dflush_loop:
-	lwz	r6,0x0(r3)
-	addi	r3,r3,32
-	bdnz	..dflush_loop
-	addi	r3,r3,-32
-	mtctr	r4
-..ag:	dcbf	r0,r3
-	addi	r3,r3,-32
-	bdnz	..ag
-	sync
-	mtmsr	r9
-	blr
-	function_epilog(dflush)
 #endif /* CONFIG_440 */