* Patch by Jon Loeliger, 2005-05-05

  Implemented support for MPC8548CDS board.
  Added DDR II support based on SPD values for MPC85xx boards.
  This roll-up patch also includes bugfies for the previously
  published patches:
    DDRII CPO, pre eTSEC, 8548 LBIU, Andy's TSEC, eTSEC 3&4 I/O

6 files changed, 976 insertions, 247 deletions

diff --git a/cpu/mpc85xx/cpu.c b/cpu/mpc85xx/cpu.c
index 4a1ccb0..f7fe22e 100644
--- a/cpu/mpc85xx/cpu.c
+++ b/cpu/mpc85xx/cpu.c
@@ -38,6 +38,7 @@ int checkcpu (void)
 	uint lcrr;		/* local bus clock ratio register */
 	uint clkdiv;		/* clock divider portion of lcrr */
 	uint pvr, svr;
+	uint fam;
 	uint ver;
 	uint major, minor;
 
@@ -60,6 +61,12 @@ int checkcpu (void)
 	case SVR_8560:
 		puts("8560");
 		break;
+	case SVR_8548:
+		puts("8548");
+		break;
+	case SVR_8548_E:
+		puts("8548_E");
+		break;
 	default:
 		puts("Unknown");
 		break;
@@ -67,13 +74,14 @@ int checkcpu (void)
 	printf(", Version: %d.%d, (0x%08x)\n", major, minor, svr);
 
 	pvr = get_pvr();
+	fam = PVR_FAM(pvr);
 	ver = PVR_VER(pvr);
 	major = PVR_MAJ(pvr);
 	minor = PVR_MIN(pvr);
 
 	printf("Core:  ");
-	switch (ver) {
-	case PVR_VER(PVR_85xx):
+	switch (fam) {
+	case PVR_FAM(PVR_85xx):
 	    puts("E500");
 	    break;
 	default:
@@ -84,7 +92,7 @@ int checkcpu (void)
 
 	get_sys_info(&sysinfo);
 
-	puts("Clocks Configuration:\n");
+	puts("Clock Configuration:\n");
 	printf("       CPU:%4lu MHz, ", sysinfo.freqProcessor / 1000000);
 	printf("CCB:%4lu MHz,\n", sysinfo.freqSystemBus / 1000000);
 	printf("       DDR:%4lu MHz, ", sysinfo.freqSystemBus / 2000000);
@@ -101,6 +109,13 @@ int checkcpu (void)
 #endif
 	clkdiv = lcrr & 0x0f;
 	if (clkdiv == 2 || clkdiv == 4 || clkdiv == 8) {
+#ifdef CONFIG_MPC8548
+		/*
+		 * Yes, the entire PQ38 family use the same
+		 * bit-representation for twice the clock divider values.
+		 */
+		 clkdiv *= 2;
+#endif
 		printf("LBC:%4lu MHz\n",
 		       sysinfo.freqSystemBus / 1000000 / clkdiv);
 	} else {
diff --git a/cpu/mpc85xx/cpu_init.c b/cpu/mpc85xx/cpu_init.c
index 79ea91f..efde9cc 100644
--- a/cpu/mpc85xx/cpu_init.c
+++ b/cpu/mpc85xx/cpu_init.c
@@ -178,42 +178,58 @@ void cpu_init_f (void)
 #endif
 }
 
+
 /*
- * We initialize L2 as cache here.
+ * Initialize L2 as cache.
+ *
+ * The newer 8548, etc, parts have twice as much cache, but
+ * use the same bit-encoding as the older 8555, etc, parts.
+ *
+ * FIXME: Use PVR_VER(pvr) == 1 test here instead of SVR_VER()?
  */
-int cpu_init_r (void)
+
+int cpu_init_r(void)
 {
 #if defined(CONFIG_L2_CACHE)
-	volatile immap_t    *immap = (immap_t *)CFG_IMMR;
+	volatile immap_t *immap = (immap_t *)CFG_IMMR;
 	volatile ccsr_l2cache_t *l2cache = &immap->im_l2cache;
-	volatile uint temp;
+	volatile uint cache_ctl;
+	uint svr, ver;
+
+	svr = get_svr();
+	ver = SVR_VER(svr);
 
 	asm("msync;isync");
-	temp = l2cache->l2ctl;
-	temp &= 0x30000000;
-	switch ( temp ) {
+	cache_ctl = l2cache->l2ctl;
+
+	switch (cache_ctl & 0x30000000) {
 	case 0x20000000:
-		printf ("L2 cache 256KB:");
+		if (ver == SVR_8548 || ver == SVR_8548_E) {
+			printf ("L2 cache 512KB:");
+		} else {
+			printf ("L2 cache 256KB:");
+		}
 		break;
 	case 0x00000000:
 	case 0x10000000:
 	case 0x30000000:
 	default:
-		printf ("L2 cache unknown size. Check the silicon!\n");
+		printf ("L2 cache unknown size (0x%08x)\n", cache_ctl);
 		return -1;
 	}
 
 	asm("msync;isync");
 	l2cache->l2ctl = 0x68000000; /* invalidate */
-	temp = l2cache->l2ctl;
+	cache_ctl = l2cache->l2ctl;
 	asm("msync;isync");
+
 	l2cache->l2ctl = 0xa8000000; /* enable 256KB L2 cache */
-	temp = l2cache->l2ctl;
+	cache_ctl = l2cache->l2ctl;
 	asm("msync;isync");
 
-	printf("enabled\n");
+	printf(" enabled\n");
 #else
-	printf("L2:    disabled.\n");
+	printf("L2 cache: disabled\n");
 #endif
 
 	return 0;
diff --git a/cpu/mpc85xx/spd_sdram.c b/cpu/mpc85xx/spd_sdram.c
index 5a1dbe2..049ba67 100644
--- a/cpu/mpc85xx/spd_sdram.c
+++ b/cpu/mpc85xx/spd_sdram.c
@@ -28,10 +28,11 @@
 #include <spd.h>
 #include <asm/mmu.h>
 
-#if defined(CONFIG_DDR_ECC)
-extern void dma_init (void);
+
+#if defined(CONFIG_DDR_ECC) && !defined(CONFIG_ECC_INIT_VIA_DDRCONTROLLER)
+extern void dma_init(void);
 extern uint dma_check(void);
-extern int  dma_xfer (void *dest, uint count, void *src);
+extern int dma_xfer(void *dest, uint count, void *src);
 #endif
 
 #ifdef CONFIG_SPD_EEPROM
@@ -40,6 +41,9 @@ extern int  dma_xfer (void *dest, uint count, void *src);
 #define CFG_READ_SPD	i2c_read
 #endif
 
+static unsigned int setup_laws_and_tlbs(unsigned int memsize);
+
+
 /*
  * Convert picoseconds into clock cycles (rounding up if needed).
  */
@@ -57,311 +61,829 @@ picos_to_clk(int picos)
 	return clks;
 }
 
+
+/*
+ * Calculate the Density of each Physical Rank.
+ * Returned size is in bytes.
+ *
+ * Study these table from Byte 31 of JEDEC SPD Spec.
+ *
+ *		DDR I	DDR II
+ *	Bit	Size	Size
+ *	---	-----	------
+ *	7 high	512MB	512MB
+ *	6	256MB	256MB
+ *	5	128MB	128MB
+ *	4	 64MB	 16GB
+ *	3	 32MB	  8GB
+ *	2	 16MB	  4GB
+ *	1	  2GB	  2GB
+ *	0 low	  1GB	  1GB
+ *
+ * Reorder Table to be linear by stripping the bottom
+ * 2 or 5 bits off and shifting them up to the top.
+ */
+
+unsigned int
+compute_banksize(unsigned int mem_type, unsigned char row_dens)
+{
+	unsigned int bsize;
+
+	if (mem_type == SPD_MEMTYPE_DDR) {
+		/* Bottom 2 bits up to the top. */
+		bsize = ((row_dens >> 2) | ((row_dens & 3) << 6)) << 24;
+		debug("DDR: DDR I rank density = 0x%08x\n", bsize);
+	} else {
+		/* Bottom 5 bits up to the top. */
+		bsize = ((row_dens >> 5) | ((row_dens & 31) << 3)) << 27;
+		debug("DDR: DDR II rank density = 0x%08x\n", bsize);
+	}
+	return bsize;
+}
+
+
+/*
+ * Convert a two-nibble BCD value into a cycle time.
+ * While the spec calls for nano-seconds, picos are returned.
+ *
+ * This implements the tables for bytes 9, 23 and 25 for both
+ * DDR I and II.  No allowance for distinguishing the invalid
+ * fields absent for DDR I yet present in DDR II is made.
+ * (That is, cycle times of .25, .33, .66 and .75 ns are
+ * allowed for both DDR II and I.)
+ */
+
 unsigned int
-banksize(unsigned char row_dens)
+convert_bcd_tenths_to_cycle_time_ps(unsigned int spd_val)
 {
-	return ((row_dens >> 2) | ((row_dens & 3) << 6)) << 24;
+	/*
+	 * Table look up the lower nibble, allow DDR I & II.
+	 */
+	unsigned int tenths_ps[16] = {
+		0,
+		100,
+		200,
+		300,
+		400,
+		500,
+		600,
+		700,
+		800,
+		900,
+		250,
+		330,	/* FIXME: Is 333 better/valid? */
+		660,	/* FIXME: Is 667 better/valid? */
+		750,
+		0,	/* undefined */
+		0	/* undefined */
+	};
+
+	unsigned int whole_ns = (spd_val & 0xF0) >> 4;
+	unsigned int tenth_ns = spd_val & 0x0F;
+	unsigned int ps = whole_ns * 1000 + tenths_ps[tenth_ns];
+
+	return ps;
 }
 
+
 long int
 spd_sdram(void)
 {
 	volatile immap_t *immap = (immap_t *)CFG_IMMR;
 	volatile ccsr_ddr_t *ddr = &immap->im_ddr;
-	volatile ccsr_local_ecm_t *ecm = &immap->im_local_ecm;
+	volatile ccsr_gur_t *gur = &immap->im_gur;
 	spd_eeprom_t spd;
-	unsigned tmp, tmp1;
+	unsigned int n_ranks;
+	unsigned int rank_density;
+	unsigned int odt_rd_cfg, odt_wr_cfg;
+	unsigned int odt_cfg, mode_odt_enable;
+	unsigned int dqs_cfg;
+	unsigned char twr_clk, twtr_clk, twr_auto_clk;
+	unsigned int tCKmin_ps, tCKmax_ps;
+	unsigned int max_data_rate, effective_data_rate;
+	unsigned int busfreq;
+	unsigned sdram_cfg;
 	unsigned int memsize;
-	unsigned int tlb_size;
-	unsigned int law_size;
-	unsigned char caslat;
-	unsigned int ram_tlb_index;
-	unsigned int ram_tlb_address;
+	unsigned char caslat, caslat_ctrl;
+	unsigned int trfc, trfc_clk, trfc_low, trfc_high;
+	unsigned int trcd_clk;
+	unsigned int trtp_clk;
+	unsigned char cke_min_clk;
+	unsigned char add_lat;
+	unsigned char wr_lat;
+	unsigned char wr_data_delay;
+	unsigned char four_act;
+	unsigned char cpo;
+	unsigned char burst_len;
+	unsigned int mode_caslat;
+	unsigned char sdram_type;
+	unsigned char d_init;
 
-	CFG_READ_SPD(SPD_EEPROM_ADDRESS, 0, 1, (uchar *) & spd, sizeof (spd));
+	/*
+	 * Read SPD information.
+	 */
+	CFG_READ_SPD(SPD_EEPROM_ADDRESS, 0, 1, (uchar *) &spd, sizeof(spd));
 
-	if (spd.nrows > 2) {
-		puts("DDR:Only two chip selects are supported on ADS.\n");
+	/*
+	 * Check for supported memory module types.
+	 */
+	if (spd.mem_type != SPD_MEMTYPE_DDR &&
+	    spd.mem_type != SPD_MEMTYPE_DDR2) {
+		printf("Unable to locate DDR I or DDR II module.\n"
+		       "    Fundamental memory type is 0x%0x\n",
+		       spd.mem_type);
 		return 0;
 	}
 
-	if (spd.nrow_addr < 12
-	    || spd.nrow_addr > 14
-	    || spd.ncol_addr < 8
-	    || spd.ncol_addr > 11) {
-		puts("DDR:Row or Col number unsupported.\n");
+	/*
+	 * These test gloss over DDR I and II differences in interpretation
+	 * of bytes 3 and 4, but irrelevantly.  Multiple asymmetric banks
+	 * are not supported on DDR I; and not encoded on DDR II.
+	 *
+	 * Also note that the 8548 controller can support:
+	 *    12 <= nrow <= 16
+	 * and
+	 *     8 <= ncol <= 11 (still, for DDR)
+	 *     6 <= ncol <=  9 (for FCRAM)
+	 */
+	if (spd.nrow_addr < 12 || spd.nrow_addr > 14) {
+		printf("DDR: Unsupported number of Row Addr lines: %d.\n",
+		       spd.nrow_addr);
+		return 0;
+	}
+	if (spd.ncol_addr < 8 || spd.ncol_addr > 11) {
+		printf("DDR: Unsupported number of Column Addr lines: %d.\n",
+		       spd.ncol_addr);
 		return 0;
 	}
 
-	ddr->cs0_bnds = (banksize(spd.row_dens) >> 24) - 1;
-	ddr->cs0_config = ( 1 << 31
-			    | (spd.nrow_addr - 12) << 8
-			    | (spd.ncol_addr - 8) );
-	debug("\n");
-	debug("cs0_bnds = 0x%08x\n",ddr->cs0_bnds);
-	debug("cs0_config = 0x%08x\n",ddr->cs0_config);
-
-	if (spd.nrows == 2) {
-		ddr->cs1_bnds = ( (banksize(spd.row_dens) >> 8)
-				  | ((banksize(spd.row_dens) >> 23) - 1) );
-		ddr->cs1_config = ( 1<<31
-				    | (spd.nrow_addr-12) << 8
-				    | (spd.ncol_addr-8) );
-		debug("cs1_bnds = 0x%08x\n",ddr->cs1_bnds);
-		debug("cs1_config = 0x%08x\n",ddr->cs1_config);
+	/*
+	 * Determine the number of physical banks controlled by
+	 * different Chip Select signals.  This is not quite the
+	 * same as the number of DIMM modules on the board.  Feh.
+	 */
+	if (spd.mem_type == SPD_MEMTYPE_DDR) {
+		n_ranks = spd.nrows;
+	} else {
+		n_ranks = (spd.nrows & 0x7) + 1;
 	}
 
-	if (spd.mem_type != 0x07) {
-		puts("No DDR module found!\n");
+	debug("DDR: number of ranks = %d\n", n_ranks);
+
+	if (n_ranks > 2) {
+		printf("DDR: Only 2 chip selects are supported: %d\n",
+		       n_ranks);
 		return 0;
 	}
 
 	/*
-	 * Figure out memory size in Megabytes.
+	 * Adjust DDR II IO voltage biasing.  It just makes it work.
 	 */
-	memsize = spd.nrows * banksize(spd.row_dens) / 0x100000;
+	if (spd.mem_type == SPD_MEMTYPE_DDR2) {
+		gur->ddrioovcr = (0
+				  | 0x80000000		/* Enable */
+				  | 0x10000000		/* VSEL to 1.8V */
+				  );
+	}
 
 	/*
-	 * First supported LAW size is 16M, at LAWAR_SIZE_16M == 23. Fnord.
+	 * Determine the size of each Rank in bytes.
 	 */
-	law_size = 19 + __ilog2(memsize);
+	rank_density = compute_banksize(spd.mem_type, spd.row_dens);
+
 
 	/*
-	 * Determine size of each TLB1 entry.
+	 * Eg: Bounds: 0x0000_0000 to 0x0f000_0000	first 256 Meg
 	 */
-	switch (memsize) {
-	case 16:
-	case 32:
-		tlb_size = BOOKE_PAGESZ_16M;
-		break;
-	case 64:
-	case 128:
-		tlb_size = BOOKE_PAGESZ_64M;
-		break;
-	case 256:
-	case 512:
-	case 1024:
-	case 2048:
-		tlb_size = BOOKE_PAGESZ_256M;
-		break;
-	default:
-		puts("DDR: only 16M,32M,64M,128M,256M,512M,1G and 2G DDR I are supported.\n");
-		return 0;
-		break;
-	}
+	ddr->cs0_bnds = (rank_density >> 24) - 1;
 
 	/*
-	 * Configure DDR TLB1 entries.
-	 * Starting at TLB1 8, use no more than 8 TLB1 entries.
+	 * ODT configuration recommendation from DDR Controller Chapter.
 	 */
-	ram_tlb_index = 8;
-	ram_tlb_address = (unsigned int)CFG_DDR_SDRAM_BASE;
-	while (ram_tlb_address < (memsize * 1024 * 1024)
-	      && ram_tlb_index < 16) {
-		mtspr(MAS0, TLB1_MAS0(1, ram_tlb_index, 0));
-		mtspr(MAS1, TLB1_MAS1(1, 1, 0, 0, tlb_size));
-		mtspr(MAS2, TLB1_MAS2(E500_TLB_EPN(ram_tlb_address),
-				      0, 0, 0, 0, 0, 0, 0, 0));
-		mtspr(MAS3, TLB1_MAS3(E500_TLB_RPN(ram_tlb_address),
-				      0, 0, 0, 0, 0, 1, 0, 1, 0, 1));
-		asm volatile("isync;msync;tlbwe;isync");
+	odt_rd_cfg = 0;			/* Never assert ODT */
+	odt_wr_cfg = 0;			/* Never assert ODT */
+	if (spd.mem_type == SPD_MEMTYPE_DDR2) {
+		odt_wr_cfg = 1;		/* Assert ODT on writes to CS0 */
+#if 0
+		/* FIXME: How to determine the number of dimm modules? */
+		if (n_dimm_modules == 2) {
+			odt_rd_cfg = 1;	/* Assert ODT on reads to CS0 */
+		}
+#endif
+	}
 
-		debug("DDR:MAS0=0x%08x\n", TLB1_MAS0(1, ram_tlb_index, 0));
-		debug("DDR:MAS1=0x%08x\n", TLB1_MAS1(1, 1, 0, 0, tlb_size));
-		debug("DDR:MAS2=0x%08x\n",
-		      TLB1_MAS2(E500_TLB_EPN(ram_tlb_address),
-				0, 0, 0, 0, 0, 0, 0, 0));
-		debug("DDR:MAS3=0x%08x\n",
-		      TLB1_MAS3(E500_TLB_RPN(ram_tlb_address),
-				0, 0, 0, 0, 0, 1, 0, 1, 0, 1));
+	ddr->cs0_config = ( 1 << 31
+			    | (odt_rd_cfg << 20)
+			    | (odt_wr_cfg << 16)
+			    | (spd.nrow_addr - 12) << 8
+			    | (spd.ncol_addr - 8) );
+	debug("\n");
+	debug("DDR: cs0_bnds   = 0x%08x\n", ddr->cs0_bnds);
+	debug("DDR: cs0_config = 0x%08x\n", ddr->cs0_config);
 
-		ram_tlb_address += (0x1000 << ((tlb_size - 1) * 2));
-		ram_tlb_index++;
+	if (n_ranks == 2) {
+		/*
+		 * Eg: Bounds: 0x0f00_0000 to 0x1e0000_0000, second 256 Meg
+		 */
+		ddr->cs1_bnds = ( (rank_density >> 8)
+				  | ((rank_density >> (24 - 1)) - 1) );
+		ddr->cs1_config = ( 1<<31
+				    | (odt_rd_cfg << 20)
+				    | (odt_wr_cfg << 16)
+				    | (spd.nrow_addr - 12) << 8
+				    | (spd.ncol_addr - 8) );
+		debug("DDR: cs1_bnds   = 0x%08x\n", ddr->cs1_bnds);
+		debug("DDR: cs1_config = 0x%08x\n", ddr->cs1_config);
 	}
 
+
 	/*
-	 * Set up LAWBAR for all of DDR.
+	 * Find the largest CAS by locating the highest 1 bit
+	 * in the spd.cas_lat field.  Translate it to a DDR
+	 * controller field value:
+	 *
+	 *	CAS Lat	DDR I	DDR II	Ctrl
+	 *	Clocks	SPD Bit	SPD Bit	Value
+	 *	-------	-------	-------	-----
+	 *	1.0	0		0001
+	 *	1.5	1		0010
+	 *	2.0	2	2	0011
+	 *	2.5	3		0100
+	 *	3.0	4	3	0101
+	 *	3.5	5		0110
+	 *	4.0		4	0111
+	 *	4.5			1000
+	 *	5.0		5	1001
 	 */
-	ecm->lawbar1 = ((CFG_DDR_SDRAM_BASE>>12) & 0xfffff);
-	ecm->lawar1 = (LAWAR_EN | LAWAR_TRGT_IF_DDR | (LAWAR_SIZE & law_size));
-	debug("DDR:LAWBAR1=0x%08x\n", ecm->lawbar1);
-	debug("DDR:LARAR1=0x%08x\n", ecm->lawar1);
-
-	/*
-	 * find the largest CAS
-	 */
-	if(spd.cas_lat & 0x40) {
-		caslat = 7;
-	} else if (spd.cas_lat & 0x20) {
-		caslat = 6;
-	} else if (spd.cas_lat & 0x10) {
-		caslat = 5;
-	} else if (spd.cas_lat & 0x08) {
-		caslat = 4;
-	} else if (spd.cas_lat & 0x04) {
-		caslat = 3;
-	} else if (spd.cas_lat & 0x02) {
-		caslat = 2;
-	} else if (spd.cas_lat & 0x01) {
-		caslat = 1;
-	} else {
-		puts("DDR:no valid CAS Latency information.\n");
+	caslat = __ilog2(spd.cas_lat);
+	if ((spd.mem_type == SPD_MEMTYPE_DDR)
+	    && (caslat > 5)) {
+		printf("DDR I: Invalid SPD CAS Latency: 0x%x.\n", spd.cas_lat);
+		return 0;
+
+	} else if (spd.mem_type == SPD_MEMTYPE_DDR2
+		   && (caslat < 2 || caslat > 5)) {
+		printf("DDR II: Invalid SPD CAS Latency: 0x%x.\n",
+		       spd.cas_lat);
 		return 0;
 	}
+	debug("DDR: caslat SPD bit is %d\n", caslat);
+
+	/*
+	 * Calculate the Maximum Data Rate based on the Minimum Cycle time.
+	 * The SPD clk_cycle field (tCKmin) is measured in tenths of
+	 * nanoseconds and represented as BCD.
+	 */
+	tCKmin_ps = convert_bcd_tenths_to_cycle_time_ps(spd.clk_cycle);
+	debug("DDR: tCKmin = %d ps\n", tCKmin_ps);
+
+	/*
+	 * Double-data rate, scaled 1000 to picoseconds, and back down to MHz.
+	 */
+	max_data_rate = 2 * 1000 * 1000 / tCKmin_ps;
+	debug("DDR: Module max data rate = %d Mhz\n", max_data_rate);
+
 
-	tmp = 20000 / (((spd.clk_cycle & 0xF0) >> 4) * 10
-		       + (spd.clk_cycle & 0x0f));
-	debug("DDR:Module maximum data rate is: %dMhz\n", tmp);
+	/*
+	 * Adjust the CAS Latency to allow for bus speeds that
+	 * are slower than the DDR module.
+	 */
+	busfreq = get_bus_freq(0) / 1000000;	/* MHz */
 
-	tmp1 = get_bus_freq(0) / 1000000;
-	if (tmp1 < 230 && tmp1 >= 90 && tmp >= 230) {
-		/* 90~230 range, treated as DDR 200 */
-		if (spd.clk_cycle3 == 0xa0)
+	effective_data_rate = max_data_rate;
+	if (busfreq < 90) {
+		/* DDR rate out-of-range */
+		puts("DDR: platform frequency is not fit for DDR rate\n");
+		return 0;
+
+	} else if (90 <= busfreq && busfreq < 230 && max_data_rate >= 230) {
+		/*
+		 * busfreq 90~230 range, treated as DDR 200.
+		 */
+		effective_data_rate = 200;
+		if (spd.clk_cycle3 == 0xa0)	/* 10 ns */
 			caslat -= 2;
-		else if(spd.clk_cycle2 == 0xa0)
+		else if (spd.clk_cycle2 == 0xa0)
 			caslat--;
-	} else if (tmp1 < 280 && tmp1 >= 230 && tmp >= 280) {
-		/* 230-280 range, treated as DDR 266 */
-		if (spd.clk_cycle3 == 0x75)
+
+	} else if (230 <= busfreq && busfreq < 280 && max_data_rate >= 280) {
+		/*
+		 * busfreq 230~280 range, treated as DDR 266.
+		 */
+		effective_data_rate = 266;
+		if (spd.clk_cycle3 == 0x75)	/* 7.5 ns */
 			caslat -= 2;
 		else if (spd.clk_cycle2 == 0x75)
 			caslat--;
-	} else if (tmp1 < 350 && tmp1 >= 280 && tmp >= 350) {
-		/* 280~350 range, treated as DDR 333 */
-		if (spd.clk_cycle3 == 0x60)
+
+	} else if (280 <= busfreq && busfreq < 350 && max_data_rate >= 350) {
+		/*
+		 * busfreq 280~350 range, treated as DDR 333.
+		 */
+		effective_data_rate = 333;
+		if (spd.clk_cycle3 == 0x60)	/* 6.0 ns */
 			caslat -= 2;
 		else if (spd.clk_cycle2 == 0x60)
 			caslat--;
-	} else if (tmp1 < 90 || tmp1 >= 350) {
-		/* DDR rate out-of-range */
-		puts("DDR:platform frequency is not fit for DDR rate\n");
+
+	} else if (350 <= busfreq && busfreq < 460 && max_data_rate >= 460) {
+		/*
+		 * busfreq 350~460 range, treated as DDR 400.
+		 */
+		effective_data_rate = 400;
+		if (spd.clk_cycle3 == 0x50)	/* 5.0 ns */
+			caslat -= 2;
+		else if (spd.clk_cycle2 == 0x50)
+			caslat--;
+
+	} else if (460 <= busfreq && busfreq < 560 && max_data_rate >= 560) {
+		/*
+		 * busfreq 460~560 range, treated as DDR 533.
+		 */
+		effective_data_rate = 533;
+		if (spd.clk_cycle3 == 0x3D)	/* 3.75 ns */
+			caslat -= 2;
+		else if (spd.clk_cycle2 == 0x3D)
+			caslat--;
+
+	} else if (560 <= busfreq && busfreq < 700 && max_data_rate >= 700) {
+		/*
+		 * busfreq 560~700 range, treated as DDR 667.
+		 */
+		effective_data_rate = 667;
+		if (spd.clk_cycle3 == 0x30)	/* 3.0 ns */
+			caslat -= 2;
+		else if (spd.clk_cycle2 == 0x30)
+			caslat--;
+
+	} else if (700 <= busfreq) {
+		/*
+		 * DDR rate out-of-range
+		 */
+		printf("DDR: Bus freq %d MHz is not fit for DDR rate %d MHz\n",
+		     busfreq, max_data_rate);
 		return 0;
 	}
 
+
 	/*
-	 * note: caslat must also be programmed into ddr->sdram_mode
-	 * register.
-	 *
-	 * note: WRREC(Twr) and WRTORD(Twtr) are not in SPD,
-	 * use conservative value here.
+	 * Convert caslat clocks to DDR controller value.
+	 * Force caslat_ctrl to be DDR Controller field-sized.
+	 */
+	if (spd.mem_type == SPD_MEMTYPE_DDR) {
+		caslat_ctrl = (caslat + 1) & 0x07;
+	} else {
+		caslat_ctrl =  (2 * caslat - 1) & 0x0f;
+	}
+
+	debug("DDR: effective data rate is %d MHz\n", effective_data_rate);
+	debug("DDR: caslat SPD bit is %d, controller field is 0x%x\n",
+	      caslat, caslat_ctrl);
+
+	/*
+	 * Timing Config 0.
+	 * Avoid writing for DDR I.  The new PQ38 DDR controller
+	 * dreams up non-zero default values to be backwards compatible.
 	 */
+	if (spd.mem_type == SPD_MEMTYPE_DDR2) {
+		unsigned char taxpd_clk = 8;		/* By the book. */
+		unsigned char tmrd_clk = 2;		/* By the book. */
+		unsigned char act_pd_exit = 2;		/* Empirical? */
+		unsigned char pre_pd_exit = 6;		/* Empirical? */
+
+		ddr->timing_cfg_0 = (0
+			| ((act_pd_exit & 0x7) << 20)	/* ACT_PD_EXIT */
+			| ((pre_pd_exit & 0x7) << 16)	/* PRE_PD_EXIT */
+			| ((taxpd_clk & 0xf) << 8)	/* ODT_PD_EXIT */
+			| ((tmrd_clk & 0xf) << 0)	/* MRS_CYC */
+			);
+#if 0
+		ddr->timing_cfg_0 |= 0xaa000000;	/* extra cycles */
+#endif
+		debug("DDR: timing_cfg_0 = 0x%08x\n", ddr->timing_cfg_0);
+
+	} else {
+#if 0
+		/*
+		 * Force extra cycles with 0xaa bits.
+		 * Incidentally supply the dreamt-up backwards compat value!
+		 */
+		ddr->timing_cfg_0 = 0x00110105;	/* backwards compat value */
+		ddr->timing_cfg_0 |= 0xaa000000;	/* extra cycles */
+		debug("DDR: HACK timing_cfg_0 = 0x%08x\n", ddr->timing_cfg_0);
+#endif
+	}
+
+
+	/*
+	 * Some Timing Config 1 values now.
+	 * Sneak Extended Refresh Recovery in here too.
+	 */
+
+	/*
+	 * For DDR I, WRREC(Twr) and WRTORD(Twtr) are not in SPD,
+	 * use conservative value.
+	 * For DDR II, they are bytes 36 and 37, in quarter nanos.
+	 */
+
+	if (spd.mem_type == SPD_MEMTYPE_DDR) {
+		twr_clk = 3;	/* Clocks */
+		twtr_clk = 1;	/* Clocks */
+	} else {
+		twr_clk = picos_to_clk(spd.twr * 250);
+		twtr_clk = picos_to_clk(spd.twtr * 250);
+	}
+
+	/*
+	 * Calculate Trfc, in picos.
+	 * DDR I:  Byte 42 straight up in ns.
+	 * DDR II: Byte 40 and 42 swizzled some, in ns.
+	 */
+	if (spd.mem_type == SPD_MEMTYPE_DDR) {
+		trfc = spd.trfc * 1000;		/* up to ps */
+	} else {
+		unsigned int byte40_table_ps[8] = {
+			0,
+			250,
+			330,
+			500,
+			660,
+			750,
+			0,
+			0
+		};
+
+		trfc = (((spd.trctrfc_ext & 0x1) * 256) + spd.trfc) * 1000
+			+ byte40_table_ps[(spd.trctrfc_ext >> 1) & 0x7];
+	}
+	trfc_clk = picos_to_clk(trfc);
+
+	/*
+	 * Trcd, Byte 29, from quarter nanos to ps and clocks.
+	 */
+	trcd_clk = picos_to_clk(spd.trcd * 250) & 0x7;
+
+	/*
+	 * Convert trfc_clk to DDR controller fields.  DDR I should
+	 * fit in the REFREC field (16-19) of TIMING_CFG_1, but the
+	 * 8548 controller has an extended REFREC field of three bits.
+	 * The controller automatically adds 8 clocks to this value,
+	 * so preadjust it down 8 first before splitting it up.
+	 */
+	trfc_low = (trfc_clk - 8) & 0xf;
+	trfc_high = ((trfc_clk - 8) >> 4) & 0x3;
+
+	/*
+	 * Sneak in some Extended Refresh Recovery.
+	 */
+	ddr->ext_refrec = (trfc_high << 16);
+	debug("DDR: ext_refrec = 0x%08x\n", ddr->ext_refrec);
+
 	ddr->timing_cfg_1 =
-	    (((picos_to_clk(spd.trp * 250) & 0x07) << 28 ) |
-	     ((picos_to_clk(spd.tras * 1000) & 0x0f ) << 24 ) |
-	     ((picos_to_clk(spd.trcd * 250) & 0x07) << 20 ) |
-	     ((caslat & 0x07) << 16 ) |
-	     (((picos_to_clk(spd.sset[6] * 1000) - 8) & 0x0f) << 12 ) |
-	     ( 0x300 ) |
-	     ((picos_to_clk(spd.trrd * 250) & 0x07) << 4) | 1);
+	    (0
+	     | ((picos_to_clk(spd.trp * 250) & 0x07) << 28)	/* PRETOACT */
+	     | ((picos_to_clk(spd.tras * 1000) & 0x0f ) << 24)	/* ACTTOPRE */
+	     | (trcd_clk << 20)					/* ACTTORW */
+	     | (caslat_ctrl << 16)				/* CASLAT */
+	     | (trfc_low << 12)					/* REFEC */
+	     | ((twr_clk & 0x07) << 8)				/* WRRREC */
+	     | ((picos_to_clk(spd.trrd * 250) & 0x07) << 4)	/* ACTTOACT */
+	     | ((twtr_clk & 0x07) << 0)				/* WRTORD */
+	     );
 
-	ddr->timing_cfg_2 = 0x00000800;
+	debug("DDR: timing_cfg_1  = 0x%08x\n", ddr->timing_cfg_1);
 
-	debug("DDR:timing_cfg_1=0x%08x\n", ddr->timing_cfg_1);
-	debug("DDR:timing_cfg_2=0x%08x\n", ddr->timing_cfg_2);
 
 	/*
-	 * Only DDR I is supported
-	 * DDR I and II have different mode-register-set definition
+	 * Timing_Config_2
+	 * Was: 0x00000800;
 	 */
 
-	/* burst length is always 4 */
-	switch(caslat) {
-	case 2:
-		ddr->sdram_mode = 0x52; /* 1.5 */
-		break;
-	case 3:
-		ddr->sdram_mode = 0x22; /* 2.0 */
-		break;
-	case 4:
-		ddr->sdram_mode = 0x62; /* 2.5 */
-		break;
-	case 5:
-		ddr->sdram_mode = 0x32; /* 3.0 */
-		break;
-	default:
-		puts("DDR:only CAS Latency 1.5, 2.0, 2.5, 3.0 is supported.\n");
-		return 0;
+	/*
+	 * Additive Latency
+	 * For DDR I, 0.
+	 * For DDR II, with ODT enabled, use "a value" less than ACTTORW,
+	 * which comes from Trcd, and also note that:
+	 *	add_lat + caslat must be >= 4
+	 */
+	add_lat = 0;
+	if (spd.mem_type == SPD_MEMTYPE_DDR2
+	    && (odt_wr_cfg || odt_rd_cfg)
+	    && (caslat < 4)) {
+		add_lat = 4 - caslat;
+		if (add_lat > trcd_clk) {
+			add_lat = trcd_clk - 1;
+		}
 	}
-	debug("DDR:sdram_mode=0x%08x\n", ddr->sdram_mode);
 
-	switch(spd.refresh) {
-	case 0x00:
-	case 0x80:
-		tmp = picos_to_clk(15625000);
-		break;
-	case 0x01:
-	case 0x81:
-		tmp = picos_to_clk(3900000);
-		break;
-	case 0x02:
-	case 0x82:
-		tmp = picos_to_clk(7800000);
-		break;
-	case 0x03:
-	case 0x83:
-		tmp = picos_to_clk(31300000);
-		break;
-	case 0x04:
-	case 0x84:
-		tmp = picos_to_clk(62500000);
-		break;
-	case 0x05:
-	case 0x85:
-		tmp = picos_to_clk(125000000);
-		break;
-	default:
-		tmp = 0x512;
-		break;
+	/*
+	 * Write Data Delay
+	 * Historically 0x2 == 4/8 clock delay.
+	 * Empirically, 0x3 == 6/8 clock delay is suggested for DDR I 266.
+	 */
+	wr_data_delay = 3;
+
+	/*
+	 * Write Latency
+	 * Read to Precharge
+	 * Minimum CKE Pulse Width.
+	 * Four Activate Window
+	 */
+	if (spd.mem_type == SPD_MEMTYPE_DDR) {
+		/*
+		 * This is a lie.  It should really be 1, but if it is
+		 * set to 1, bits overlap into the old controller's
+		 * otherwise unused ACSM field.  If we leave it 0, then
+		 * the HW will magically treat it as 1 for DDR 1.  Oh Yea.
+		 */
+		wr_lat = 0;
+
+		trtp_clk = 2;		/* By the book. */
+		cke_min_clk = 1;	/* By the book. */
+		four_act = 1;		/* By the book. */
+
+	} else {
+		wr_lat = caslat - 1;
+
+		/* Convert SPD value from quarter nanos to picos. */
+		trtp_clk = picos_to_clk(spd.trtp * 250);
+
+		cke_min_clk = 3;	/* By the book. */
+		four_act = picos_to_clk(37500);	/* By the book. 1k pages? */
+	}
+
+	/*
+	 * Empirically set ~MCAS-to-preamble override for DDR 2.
+	 * Your milage will vary.
+	 */
+	cpo = 0;
+	if (spd.mem_type == SPD_MEMTYPE_DDR2) {
+		if (effective_data_rate == 266 || effective_data_rate == 333) {
+			cpo = 0x7;		/* READ_LAT + 5/4 */
+		} else if (effective_data_rate == 400) {
+			cpo = 0x9;		/* READ_LAT + 7/4 */
+		} else {
+			/* Pure speculation */
+			cpo = 0xb;
+		}
+	}
+
+	ddr->timing_cfg_2 = (0
+		| ((add_lat & 0x7) << 28)		/* ADD_LAT */
+		| ((cpo & 0x1f) << 23)			/* CPO */ 
+		| ((wr_lat & 0x7) << 19)		/* WR_LAT */
+		| ((trtp_clk & 0x7) << 13)		/* RD_TO_PRE */
+		| ((wr_data_delay & 0x7) << 10)		/* WR_DATA_DELAY */
+		| ((cke_min_clk & 0x7) << 6)		/* CKE_PLS */
+		| ((four_act & 0x1f) << 0)		/* FOUR_ACT */
+		);
+
+	debug("DDR: timing_cfg_2 = 0x%08x\n", ddr->timing_cfg_2);
+
+
+	/*
+	 * Determine the Mode Register Set.
+	 *
+	 * This is nominally part specific, but it appears to be
+	 * consistent for all DDR I devices, and for all DDR II devices.
+	 *
+	 *     caslat must be programmed
+	 *     burst length is always 4
+	 *     burst type is sequential
+	 *
+	 * For DDR I:
+	 *     operating mode is "normal"
+	 *
+	 * For DDR II:
+	 *     other stuff
+	 */
+
+	mode_caslat = 0;
+
+	/*
+	 * Table lookup from DDR I or II Device Operation Specs.
+	 */
+	if (spd.mem_type == SPD_MEMTYPE_DDR) {
+		if (1 <= caslat && caslat <= 4) {
+			unsigned char mode_caslat_table[4] = {
+				0x5,	/* 1.5 clocks */
+				0x2,	/* 2.0 clocks */
+				0x6,	/* 2.5 clocks */
+				0x3	/* 3.0 clocks */
+			};
+			mode_caslat = mode_caslat_table[caslat - 1];
+		} else {
+			puts("DDR I: Only CAS Latencies of 1.5, 2.0, "
+			     "2.5 and 3.0 clocks are supported.\n");
+			return 0;
+		}
+
+	} else {
+		if (2 <= caslat && caslat <= 5) {
+			mode_caslat = caslat;
+		} else {
+			puts("DDR II: Only CAS Latencies of 2.0, 3.0, "
+			     "4.0 and 5.0 clocks are supported.\n");
+			return 0;
+		}
+	}
+
+	/*
+	 * Encoded Burst Lenght of 4.
+	 */
+	burst_len = 2;			/* Fiat. */
+
+	if (spd.mem_type == SPD_MEMTYPE_DDR) {
+		twr_auto_clk = 0;	/* Historical */
+	} else {
+		/*
+		 * Determine tCK max in picos.  Grab tWR and convert to picos.
+		 * Auto-precharge write recovery is:
+		 *	WR = roundup(tWR_ns/tCKmax_ns).
+		 *
+		 * Ponder: Is twr_auto_clk different than twr_clk?
+		 */
+		tCKmax_ps = convert_bcd_tenths_to_cycle_time_ps(spd.tckmax);
+		twr_auto_clk = (spd.twr * 250 + tCKmax_ps - 1) / tCKmax_ps;
+	}
+
+
+	/*
+	 * Mode Reg in bits 16 ~ 31,
+	 * Extended Mode Reg 1 in bits 0 ~ 15.
+	 */
+	mode_odt_enable = 0x0;			/* Default disabled */
+	if (odt_wr_cfg || odt_rd_cfg) {
+		/*
+		 * Bits 6 and 2 in Extended MRS(1)
+		 * Bit 2 == 0x04 == 75 Ohm, with 2 DIMM modules.
+		 * Bit 6 == 0x40 == 150 Ohm, with 1 DIMM module.
+		 */
+		mode_odt_enable = 0x40;		/* 150 Ohm */
 	}
 
+	ddr->sdram_mode =
+		(0
+		 | (add_lat << (16 + 3))	/* Additive Latency in EMRS1 */
+		 | (mode_odt_enable << 16)	/* ODT Enable in EMRS1 */
+		 | (twr_auto_clk << 9)		/* Write Recovery Autopre */
+		 | (mode_caslat << 4)		/* caslat */
+		 | (burst_len << 0)		/* Burst length */
+		 );
+
+	debug("DDR: sdram_mode   = 0x%08x\n", ddr->sdram_mode);
+
+
+	/*
+	 * Clear EMRS2 and EMRS3.
+	 */
+	ddr->sdram_mode_2 = 0;
+	debug("DDR: sdram_mode_2 = 0x%08x\n", ddr->sdram_mode_2);
+
+
 	/*
-	 * Set BSTOPRE to 0x100 for page mode
-	 * If auto-charge is used, set BSTOPRE = 0
+	 * Determine Refresh Rate.  Ignore self refresh bit on DDR I.
+	 * Table from SPD Spec, Byte 12, converted to picoseconds and
+	 * filled in with "default" normal values.
 	 */
-	ddr->sdram_interval = ((tmp & 0x3fff) << 16) | 0x100;
-	debug("DDR:sdram_interval=0x%08x\n", ddr->sdram_interval);
+	{
+		unsigned int refresh_clk;
+		unsigned int refresh_time_ns[8] = {
+			15625000,	/* 0 Normal    1.00x */
+			3900000,	/* 1 Reduced    .25x */
+			7800000,	/* 2 Extended   .50x */
+			31300000,	/* 3 Extended  2.00x */
+			62500000,	/* 4 Extended  4.00x */
+			125000000,	/* 5 Extended  8.00x */
+			15625000,	/* 6 Normal    1.00x  filler */
+			15625000,	/* 7 Normal    1.00x  filler */
+		};
+
+		refresh_clk = picos_to_clk(refresh_time_ns[spd.refresh & 0x7]);
+
+		/*
+		 * Set BSTOPRE to 0x100 for page mode
+		 * If auto-charge is used, set BSTOPRE = 0
+		 */
+		ddr->sdram_interval =
+			(0
+			 | (refresh_clk & 0x3fff) << 16
+			 | 0x100
+			 );
+		debug("DDR: sdram_interval = 0x%08x\n", ddr->sdram_interval);
+	}
 
 	/*
 	 * Is this an ECC DDR chip?
+	 * But don't mess with it if the DDR controller will init mem.
 	 */
-#if defined(CONFIG_DDR_ECC)
+#if defined(CONFIG_DDR_ECC) && !defined(CONFIG_ECC_INIT_VIA_DDRCONTROLLER)
 	if (spd.config == 0x02) {
 		ddr->err_disable = 0x0000000d;
 		ddr->err_sbe = 0x00ff0000;
 	}
-	debug("DDR:err_disable=0x%08x\n", ddr->err_disable);
-	debug("DDR:err_sbe=0x%08x\n", ddr->err_sbe);
+	debug("DDR: err_disable = 0x%08x\n", ddr->err_disable);
+	debug("DDR: err_sbe = 0x%08x\n", ddr->err_sbe);
 #endif
-	asm("sync;isync;msync");
 
+	asm("sync;isync;msync");
 	udelay(500);
 
-#ifdef MPC85xx_DDR_SDRAM_CLK_CNTL
-	/* Setup the clock control (8555 and later)
-	 * SDRAM_CLK_CNTL[0] = Source synchronous enable == 1
-	 * SDRAM_CLK_CNTL[5-7] = Clock Adjust == 3 (3/4 cycle late)
+	/*
+	 * SDRAM Cfg 2
+	 */
+
+	/*
+	 * When ODT is enabled, Chap 9 suggests asserting ODT to
+	 * internal IOs only during reads.
+	 */
+	odt_cfg = 0;
+	if (odt_rd_cfg | odt_wr_cfg) {
+		odt_cfg = 0x2;		/* ODT to IOs during reads */
+	}
+
+	/*
+	 * Try to use differential DQS with DDR II.
 	 */
-	ddr->sdram_clk_cntl = 0x83000000;
+	if (spd.mem_type == SPD_MEMTYPE_DDR) {
+		dqs_cfg = 0;		/* No Differential DQS for DDR I */
+	} else {
+		dqs_cfg = 0x1;		/* Differential DQS for DDR II */
+	}
+
+#if defined(CONFIG_ECC_INIT_VIA_DDRCONTROLLER)
+	/*
+	 * Use the DDR controller to auto initialize memory.
+	 */
+	d_init = 1;
+	ddr->sdram_data_init = CONFIG_MEM_INIT_VALUE;
+	debug("DDR: ddr_data_init = 0x%08x\n", ddr->sdram_data_init);
+#else
+	/*
+	 * Memory will be initialized via DMA, or not at all.
+	 */
+	d_init = 0;	
+#endif
+
+	ddr->sdram_cfg_2 = (0
+			    | (dqs_cfg << 26)	/* Differential DQS */
+			    | (odt_cfg << 21)	/* ODT */
+			    | (d_init << 4)	/* D_INIT auto init DDR */
+			    );
+
+	debug("DDR: sdram_cfg_2  = 0x%08x\n", ddr->sdram_cfg_2);
+
+
+#ifdef MPC85xx_DDR_SDRAM_CLK_CNTL
+	{
+		unsigned char clk_adjust;
+
+		/*
+		 * Setup the clock control.
+		 * SDRAM_CLK_CNTL[0] = Source synchronous enable == 1
+		 * SDRAM_CLK_CNTL[5-7] = Clock Adjust
+		 *	0110	3/4 cycle late
+		 *	0111	7/8 cycle late
+		 */
+		if (spd.mem_type == SPD_MEMTYPE_DDR) {
+			clk_adjust = 0x6;
+		} else {
+			clk_adjust = 0x7;
+		}
+
+		ddr->sdram_clk_cntl = (0
+			       | 0x80000000
+			       | (clk_adjust << 23)
+			       );
+		debug("DDR: sdram_clk_cntl = 0x%08x\n", ddr->sdram_clk_cntl);
+	}
 #endif
 
 	/*
-	 * Figure out the settings for the sdram_cfg register.  Build up
-	 * the entire register in 'tmp' before writing since the write into
-	 * the register will actually enable the memory controller, and all
-	 * settings must be done before enabling.
+	 * Figure out the settings for the sdram_cfg register.
+	 * Build up the entire register in 'sdram_cfg' before writing
+	 * since the write into the register will actually enable the
+	 * memory controller; all settings must be done before enabling.
 	 *
 	 * sdram_cfg[0]   = 1 (ddr sdram logic enable)
 	 * sdram_cfg[1]   = 1 (self-refresh-enable)
-	 * sdram_cfg[6:7] = 2 (SDRAM type = DDR SDRAM)
+	 * sdram_cfg[5:7] = (SDRAM type = DDR SDRAM)
+	 *			010 DDR 1 SDRAM
+	 *			011 DDR 2 SDRAM
 	 */
-	tmp = 0xc2000000;
+	sdram_type = (spd.mem_type == SPD_MEMTYPE_DDR) ? 2 : 3;
+	sdram_cfg = (0
+		     | (1 << 31)			/* Enable */
+		     | (1 << 30)			/* Self refresh */
+		     | (sdram_type << 24)		/* SDRAM type */
+		     );
 
 	/*
 	 * sdram_cfg[3] = RD_EN - registered DIMM enable
 	 *   A value of 0x26 indicates micron registered DIMMS (micron.com)
 	 */
-	if (spd.mod_attr == 0x26) {
-		tmp |= 0x10000000;
+	if (spd.mem_type == SPD_MEMTYPE_DDR && spd.mod_attr == 0x26) {
+		sdram_cfg |= 0x10000000;		/* RD_EN */
 	}
 
 #if defined(CONFIG_DDR_ECC)
@@ -369,7 +891,7 @@ spd_sdram(void)
 	 * If the user wanted ECC (enabled via sdram_cfg[2])
 	 */
 	if (spd.config == 0x02) {
-		tmp |= 0x20000000;
+		sdram_cfg |= 0x20000000;		/* ECC_EN */
 	}
 #endif
 
@@ -385,27 +907,160 @@ spd_sdram(void)
 			/*
 			 * Enable 2T timing by setting sdram_cfg[16].
 			 */
-			tmp |= 0x8000;
+			sdram_cfg |= 0x8000;		/* 2T_EN */
 #endif
 		}
 	}
 
-	ddr->sdram_cfg = tmp;
+	/*
+	 * 200 painful micro-seconds must elapse between
+	 * the DDR clock setup and the DDR config enable.
+	 */
+	udelay(200);
+
+	/*
+	 * Go!
+	 */
+	ddr->sdram_cfg = sdram_cfg;
 
 	asm("sync;isync;msync");
 	udelay(500);
 
-	debug("DDR:sdram_cfg=0x%08x\n", ddr->sdram_cfg);
+	debug("DDR: sdram_cfg   = 0x%08x\n", ddr->sdram_cfg);
+
+
+#if defined(CONFIG_ECC_INIT_VIA_DDRCONTROLLER)
+	/*
+	 * Poll until memory is initialized.
+	 * 512 Meg at 400 might hit this 200 times or so.
+	 */
+	while ((ddr->sdram_cfg_2 & (d_init << 4)) != 0) {
+		udelay(1000);
+	}
+#endif
+
+
+	/*
+	 * Figure out memory size in Megabytes.
+	 */
+	memsize = n_ranks * rank_density / 0x100000;
+
+	/*
+	 * Establish Local Access Window and TLB mappings for DDR memory.
+	 */
+	memsize = setup_laws_and_tlbs(memsize);
+	if (memsize == 0) {
+		return 0;
+	}
 
 	return memsize * 1024 * 1024;
 }
+
+
+/*
+ * Setup Local Access Window and TLB1 mappings for the requested
+ * amount of memory.  Returns the amount of memory actually mapped
+ * (usually the original request size), or 0 on error.
+ */
+
+static unsigned int
+setup_laws_and_tlbs(unsigned int memsize)
+{
+	volatile immap_t *immap = (immap_t *)CFG_IMMR;
+	volatile ccsr_local_ecm_t *ecm = &immap->im_local_ecm;
+	unsigned int tlb_size;
+	unsigned int law_size;
+	unsigned int ram_tlb_index;
+	unsigned int ram_tlb_address;
+
+	/*
+	 * Determine size of each TLB1 entry.
+	 */
+	switch (memsize) {
+	case 16:
+	case 32:
+		tlb_size = BOOKE_PAGESZ_16M;
+		break;
+	case 64:
+	case 128:
+		tlb_size = BOOKE_PAGESZ_64M;
+		break;
+	case 256:
+	case 512:
+	case 1024:
+	case 2048:
+		tlb_size = BOOKE_PAGESZ_256M;
+		break;
+	default:
+		puts("DDR: only 16M,32M,64M,128M,256M,512M,1G and 2G are supported.\n");
+
+		/*
+		 * The memory was not able to be mapped.
+		 */
+		return 0;
+		break;
+	}
+
+	/*
+	 * Configure DDR TLB1 entries.
+	 * Starting at TLB1 8, use no more than 8 TLB1 entries.
+	 */
+	ram_tlb_index = 8;
+	ram_tlb_address = (unsigned int)CFG_DDR_SDRAM_BASE;
+	while (ram_tlb_address < (memsize * 1024 * 1024)
+	      && ram_tlb_index < 16) {
+		mtspr(MAS0, TLB1_MAS0(1, ram_tlb_index, 0));
+		mtspr(MAS1, TLB1_MAS1(1, 1, 0, 0, tlb_size));
+		mtspr(MAS2, TLB1_MAS2(E500_TLB_EPN(ram_tlb_address),
+				      0, 0, 0, 0, 0, 0, 0, 0));
+		mtspr(MAS3, TLB1_MAS3(E500_TLB_RPN(ram_tlb_address),
+				      0, 0, 0, 0, 0, 1, 0, 1, 0, 1));
+		asm volatile("isync;msync;tlbwe;isync");
+
+		debug("DDR: MAS0=0x%08x\n", TLB1_MAS0(1, ram_tlb_index, 0));
+		debug("DDR: MAS1=0x%08x\n", TLB1_MAS1(1, 1, 0, 0, tlb_size));
+		debug("DDR: MAS2=0x%08x\n",
+		      TLB1_MAS2(E500_TLB_EPN(ram_tlb_address),
+				0, 0, 0, 0, 0, 0, 0, 0));
+		debug("DDR: MAS3=0x%08x\n",
+		      TLB1_MAS3(E500_TLB_RPN(ram_tlb_address),
+				0, 0, 0, 0, 0, 1, 0, 1, 0, 1));
+
+		ram_tlb_address += (0x1000 << ((tlb_size - 1) * 2));
+		ram_tlb_index++;
+	}
+
+
+	/*
+	 * First supported LAW size is 16M, at LAWAR_SIZE_16M == 23.  Fnord.
+	 */
+	law_size = 19 + __ilog2(memsize);
+
+	/*
+	 * Set up LAWBAR for all of DDR.
+	 */
+	ecm->lawbar1 = ((CFG_DDR_SDRAM_BASE >> 12) & 0xfffff);
+	ecm->lawar1 = (LAWAR_EN
+		       | LAWAR_TRGT_IF_DDR
+		       | (LAWAR_SIZE & law_size));
+	debug("DDR: LAWBAR1=0x%08x\n", ecm->lawbar1);
+	debug("DDR: LARAR1=0x%08x\n", ecm->lawar1);
+
+	/*
+	 * Confirm that the requested amount of memory was mapped.
+	 */
+	return memsize;
+}
+
 #endif /* CONFIG_SPD_EEPROM */
 
 
-#if defined(CONFIG_DDR_ECC)
+#if defined(CONFIG_DDR_ECC) && !defined(CONFIG_ECC_INIT_VIA_DDRCONTROLLER)
+
 /*
  * Initialize all of memory for ECC, then enable errors.
  */
+
 void
 ddr_enable_ecc(unsigned int dram_size)
 {
@@ -420,7 +1075,7 @@ ddr_enable_ecc(unsigned int dram_size)
 		if (((unsigned int)p & 0x1f) == 0) {
 			ppcDcbz((unsigned long) p);
 		}
-		*p = (unsigned int)0xdeadbeef;
+		*p = (unsigned int)CONFIG_MEM_INIT_VALUE;
 		if (((unsigned int)p & 0x1c) == 0x1c) {
 			ppcDcbf((unsigned long) p);
 		}
@@ -454,7 +1109,10 @@ ddr_enable_ecc(unsigned int dram_size)
 	/*
 	 * Enable errors for ECC.
 	 */
+	debug("DMA DDR: err_disable = 0x%08x\n", ddr->err_disable);
 	ddr->err_disable = 0x00000000;
 	asm("sync;isync;msync");
+	debug("DMA DDR: err_disable = 0x%08x\n", ddr->err_disable);
 }
-#endif	/* CONFIG_DDR_ECC */
+
+#endif	/* CONFIG_DDR_ECC  && ! CONFIG_ECC_INIT_VIA_DDRCONTROLLER */
diff --git a/cpu/mpc85xx/start.S b/cpu/mpc85xx/start.S
index 7bca008..dd81899 100644
--- a/cpu/mpc85xx/start.S
+++ b/cpu/mpc85xx/start.S
@@ -174,6 +174,9 @@ _start_e500:
 	mtspr	BUCSR,r0	/* disable branch prediction */
 	mtspr	MAS4,r0
 	mtspr	MAS6,r0
+#if defined(CONFIG_ENABLE_36BIT_PHYS)
+	mtspr	MAS7,r0
+#endif
 	isync
 
 	/* Setup interrupt vectors */
@@ -358,6 +361,9 @@ _start:
 	/* Enable Time Base and Select Time Base Clock */
 	lis	r0,HID0_EMCP@h		/* Enable machine check */
 	ori	r0,r0,0x4000		/* time base is processor clock */
+#if defined(CONFIG_ENABLE_36BIT_PHYS)
+	ori	r0,r0,0x0080		/* enable MAS7 updates */
+#endif
 	mtspr	HID0,r0
 
 #if defined(CONFIG_ADDR_STREAMING)
diff --git a/cpu/mpc85xx/tsec.c b/cpu/mpc85xx/tsec.c
index d327a6d..5ac6334 100644
--- a/cpu/mpc85xx/tsec.c
+++ b/cpu/mpc85xx/tsec.c
@@ -35,7 +35,7 @@ typedef volatile struct rtxbd {
 
 struct tsec_info_struct {
 	unsigned int phyaddr;
-	unsigned int gigabit;
+	u32 flags;
 	unsigned int phyregidx;
 };
 
@@ -48,8 +48,9 @@ struct tsec_info_struct {
  *  phyaddr - The address of the PHY which is attached to
  *	the given device.
  *
- *  gigabit - This variable indicates whether the device
- *	supports gigabit speed ethernet
+ *  flags - This variable indicates whether the device
+ *	supports gigabit speed ethernet, and whether it should be
+ *	in reduced mode.
  *
  *  phyregidx - This variable specifies which ethernet device
  *	controls the MII Management registers which are connected
@@ -70,23 +71,32 @@ struct tsec_info_struct {
  */
 static struct tsec_info_struct tsec_info[] = {
 #ifdef CONFIG_MPC85XX_TSEC1
-	{TSEC1_PHY_ADDR, 1, TSEC1_PHYIDX},
+	{TSEC1_PHY_ADDR, TSEC_GIGABIT, TSEC1_PHYIDX},
 #else
 	{ 0, 0, 0},
 #endif
 #ifdef CONFIG_MPC85XX_TSEC2
-	{TSEC2_PHY_ADDR, 1, TSEC2_PHYIDX},
+	{TSEC2_PHY_ADDR, TSEC_GIGABIT, TSEC2_PHYIDX},
 #else
 	{ 0, 0, 0},
 #endif
 #ifdef CONFIG_MPC85XX_FEC
 	{FEC_PHY_ADDR, 0, FEC_PHYIDX},
 #else
+#    ifdef CONFIG_MPC85XX_TSEC3
+	{TSEC3_PHY_ADDR, TSEC_GIGABIT | TSEC_REDUCED, TSEC3_PHYIDX},
+#    else
 	{ 0, 0, 0},
+#    endif
+#    ifdef CONFIG_MPC85XX_TSEC4
+	{TSEC4_PHY_ADDR, TSEC_REDUCED, TSEC4_PHYIDX},
+#    else
+	{ 0, 0, 0},
+#    endif
 #endif
 };
 
-#define MAXCONTROLLERS 3
+#define MAXCONTROLLERS	(4)
 
 static int relocated = 0;
 
@@ -115,7 +125,7 @@ static void relocate_cmds(void);
 /* Initialize device structure. Returns success if PHY
  * initialization succeeded (i.e. if it recognizes the PHY)
  */
-int tsec_initialize(bd_t *bis, int index)
+int tsec_initialize(bd_t *bis, int index, char *devname)
 {
 	struct eth_device* dev;
 	int i;
@@ -139,9 +149,9 @@ int tsec_initialize(bd_t *bis, int index)
 			tsec_info[index].phyregidx*TSEC_SIZE);
 
 	priv->phyaddr = tsec_info[index].phyaddr;
-	priv->gigabit = tsec_info[index].gigabit;
+	priv->flags = tsec_info[index].flags;
 
-	sprintf(dev->name, "ENET%d", index);
+	sprintf(dev->name, devname);
 	dev->iobase = 0;
 	dev->priv   = priv;
 	dev->init   = tsec_init;
@@ -318,7 +328,7 @@ static int init_phy(struct eth_device *dev)
 /* For 10/100, the value is slightly different */
 uint mii_cr_init(uint mii_reg, struct tsec_private *priv)
 {
-	if(priv->gigabit)
+	if(priv->flags & TSEC_GIGABIT)
 		return MIIM_CONTROL_INIT;
 	else
 		return MIIM_CR_INIT;
@@ -438,6 +448,13 @@ uint mii_cis8204_fixled(uint mii_reg, struct tsec_private *priv)
 	return MIIM_CIS8204_SLEDCON_INIT;
 }
 
+uint mii_cis8204_setmode(uint mii_reg, struct tsec_private *priv)
+{
+	if (priv->flags & TSEC_REDUCED)
+		return MIIM_CIS8204_EPHYCON_INIT | MIIM_CIS8204_EPHYCON_RGMII;
+	else
+		return MIIM_CIS8204_EPHYCON_INIT;
+}
 
 /* Initialized required registers to appropriate values, zeroing
  * those we don't care about (unless zero is bad, in which case,
@@ -507,6 +524,15 @@ static void adjust_link(struct eth_device *dev)
 			case 10:
 				regs->maccfg2 = ((regs->maccfg2&~(MACCFG2_IF))
 					| MACCFG2_MII);
+
+				/* If We're in reduced mode, we
+				 * need to say whether we're 10
+				 * or 100 MB. */
+				if ((priv->speed == 100) 
+						&& (priv->flags & TSEC_REDUCED))
+					regs->ecntrl |= ECNTRL_R100;
+				else
+					regs->ecntrl &= ~(ECNTRL_R100);
 				break;
 			default:
 				printf("%s: Speed was bad\n", dev->name);
@@ -731,7 +757,7 @@ struct phy_info phy_info_cis8204 = {
 		/* Configure some basic stuff */
 		{MIIM_CONTROL, MIIM_CONTROL_INIT, &mii_cr_init},
 		{MIIM_CIS8204_SLED_CON, MIIM_CIS8204_SLEDCON_INIT, &mii_cis8204_fixled},
-		{MIIM_CIS8204_EPHY_CON, MIIM_CIS8204_EPHYCON_INIT, NULL},
+		{MIIM_CIS8204_EPHY_CON, MIIM_CIS8204_EPHYCON_INIT, &mii_cis8204_setmode},
 		{miim_end,}
 	},
 	(struct phy_cmd[]) { /* startup */
diff --git a/cpu/mpc85xx/tsec.h b/cpu/mpc85xx/tsec.h
index e24351a..d1c70aa 100644
--- a/cpu/mpc85xx/tsec.h
+++ b/cpu/mpc85xx/tsec.h
@@ -51,6 +51,7 @@
 
 #define ECNTRL_INIT_SETTINGS	0x00001000
 #define ECNTRL_TBI_MODE         0x00000020
+#define ECNTRL_R100		0x00000008
 
 #define miim_end -2
 #define miim_read -1
@@ -107,6 +108,7 @@
 /* Cicada 8204 Extended PHY Control Register 1 */
 #define MIIM_CIS8204_EPHY_CON		0x17
 #define MIIM_CIS8204_EPHYCON_INIT	0x0006
+#define MIIM_CIS8204_EPHYCON_RGMII	0x1000
 
 /* Cicada 8204 Serial LED Control Register */
 #define MIIM_CIS8204_SLED_CON		0x1b
@@ -424,12 +426,18 @@ typedef struct tsec
 	uint	resc00[256];
 } tsec_t;
 
+#define TSEC_GIGABIT (1)
+
+/* This flag currently only has
+ * meaning if we're using the eTSEC */
+#define TSEC_REDUCED (1 << 1)
+
 struct tsec_private {
 	volatile tsec_t *regs;
 	volatile tsec_t *phyregs;
 	struct phy_info *phyinfo;
 	uint phyaddr;
-	uint gigabit;
+	u32 flags;
 	uint link;
 	uint duplexity;
 	uint speed;