* Patch by Marc Singer, 29 May 2003:

  Fixed rarp boot method for IA32 and other little-endian CPUs.

* Patch by Marc Singer, 28 May 2003:
  Added port I/O commands.

* Patch by Matthew McClintock, 28 May 2003
  - cpu/mpc824x/start.S: fix relocation code when booting from RAM
  - minor patches for utx8245

* Patch by Daniel Engström, 28 May 2003:
  x86 update

* Patch by Dave Ellis, 9 May 2003 + 27 May 2003:
  add nand flash support to SXNI855T configuration
  fix/extend nand flash support:
  - fix 'nand erase' command so does not erase bad blocks
  - fix 'nand write' command so does not write to bad blocks
  - fix nand_probe() so handles no flash detected properly
  - add doc/README.nand
  - add .jffs2 and .oob options to nand read/write
  - add 'nand bad' command to list bad blocks
  - add 'clean' option to 'nand erase' to write JFFS2 clean markers
  - make NAND read/write faster

* Patch by Rune Torgersen, 23 May 2003:
  Update for MPC8266ADS board

1 files changed, 555 insertions, 362 deletions

diff --git a/board/sc520_cdp/flash.c b/board/sc520_cdp/flash.c
index 973cde8..1ebb32e 100644
--- a/board/sc520_cdp/flash.c
+++ b/board/sc520_cdp/flash.c
@@ -1,4 +1,7 @@
 /*
+ * (C) Copyright 2002, 2003
+ * Daniel Engström, Omicron Ceti AB, daniel@omicron.se
+ * 
  * (C) Copyright 2002
  * Sysgo Real-Time Solutions, GmbH <www.elinos.com>
  * Alex Zuepke <azu@sysgo.de>
@@ -23,31 +26,23 @@
  */
 
 #include <common.h>
+#include <asm/io.h>
+#include <pci.h>
+#include <asm/ic/sc520.h>
 
-ulong myflush(void);
+#define PROBE_BUFFER_SIZE 1024
+static unsigned char buffer[PROBE_BUFFER_SIZE];
 
+#define SC520_MAX_FLASH_BANKS  3
+#define SC520_FLASH_BANK0_BASE 0x38000000  /* BOOTCS */
+#define SC520_FLASH_BANK1_BASE 0x30000000  /* ROMCS0 */
+#define SC520_FLASH_BANK2_BASE 0x28000000  /* ROMCS1 */
+#define SC520_FLASH_BANKSIZE   0x8000000
 
-#define FLASH_BANK_SIZE 0x400000	/* 4 MB */
-#define MAIN_SECT_SIZE  0x20000		/* 128 KB */
+#define AMD29LV016B_SIZE        0x200000
+#define AMD29LV016B_SECTORS     32
 
-flash_info_t    flash_info[CFG_MAX_FLASH_BANKS];
-
-
-#define CMD_READ_ARRAY		0x00F000F0
-#define CMD_UNLOCK1		0x00AA00AA
-#define CMD_UNLOCK2		0x00550055
-#define CMD_ERASE_SETUP		0x00800080
-#define CMD_ERASE_CONFIRM	0x00300030
-#define CMD_PROGRAM		0x00A000A0
-#define CMD_UNLOCK_BYPASS	0x00200020
-
-#define MEM_FLASH_ADDR1		(*(volatile u32 *)(CFG_FLASH_BASE + (0x00000555 << 2)))
-#define MEM_FLASH_ADDR2		(*(volatile u32 *)(CFG_FLASH_BASE + (0x000002AA << 2)))
-
-#define BIT_ERASE_DONE		0x00800080
-#define BIT_RDY_MASK		0x00800080
-#define BIT_PROGRAM_ERROR	0x00200020
-#define BIT_TIMEOUT		0x80000000 /* our flag */
+flash_info_t    flash_info[SC520_MAX_FLASH_BANKS];
 
 #define READY 1
 #define ERR   2
@@ -56,394 +51,592 @@ flash_info_t    flash_info[CFG_MAX_FLASH_BANKS];
 /*-----------------------------------------------------------------------
  */
 
-ulong flash_init(void)
+
+static u32 _probe_flash(u32 addr, u32 bw, int il)
 {
-    int i, j;
-    ulong size = 0;
-
-    for (i = 0; i < CFG_MAX_FLASH_BANKS; i++)
-    {
-	ulong flashbase = 0;
-	flash_info[i].flash_id =
-	  (AMD_MANUFACT & FLASH_VENDMASK) |
-	  (AMD_ID_LV160B & FLASH_TYPEMASK);
-	flash_info[i].size = FLASH_BANK_SIZE;
-	flash_info[i].sector_count = CFG_MAX_FLASH_SECT;
-	memset(flash_info[i].protect, 0, CFG_MAX_FLASH_SECT);
-	if (i == 0)
-	  flashbase = PHYS_FLASH_1;
-	else
-	  panic("configured to many flash banks!\n");
-	for (j = 0; j < flash_info[i].sector_count; j++)
-	{
-
-	    if (j <= 3)
-	    {
-		/* 1st one is 32 KB */
-		if (j == 0)
-		{
-			flash_info[i].start[j] = flashbase + 0;
+	u32 result=0;
+	
+	/* First do an unlock cycle for the benefit of
+	 * devices that need it */
+	
+	switch (bw) {
+		
+	case 1:
+		*(volatile u8*)(addr+0x5555) = 0xaa;
+		*(volatile u8*)(addr+0x2aaa) = 0x55;
+		*(volatile u8*)(addr+0x5555) = 0x90;
+		
+		/* Read vendor */
+		result = *(volatile u8*)addr;
+		result <<= 16;
+		
+		/* Read device */
+		result |= *(volatile u8*)(addr+2);
+		
+		/* Return device to data mode */
+		*(volatile u8*)addr = 0xff;
+		*(volatile u8*)(addr+0x5555), 0xf0;  
+		break;
+		
+	case 2:
+		*(volatile u16*)(addr+0xaaaa) = 0xaaaa;
+		*(volatile u16*)(addr+0x5554) = 0x5555;
+		
+		/* Issue identification command */
+		if (il == 2) {
+			*(volatile u16*)(addr+0xaaaa) = 0x9090;
+			
+			/* Read vendor */
+			result = *(volatile u8*)addr;
+			result <<= 16;
+			
+			/* Read device */
+			result |= *(volatile u8*)(addr+2);
+			
+			/* Return device to data mode */
+			*(volatile u16*)addr =  0xffff;
+			*(volatile u16*)(addr+0xaaaa), 0xf0f0;  
+			
+		} else {
+			*(volatile u8*)(addr+0xaaaa) = 0x90;
+			/* Read vendor */
+			result = *(volatile u16*)addr;
+			result <<= 16;
+			
+			/* Read device */
+			result |= *(volatile u16*)(addr+2);
+			
+			/* Return device to data mode */
+			*(volatile u8*)addr = 0xff;
+			*(volatile u8*)(addr+0xaaaa), 0xf0;  			
 		}
-
-		/* 2nd and 3rd are both 16 KB */
-		if ((j == 1) || (j == 2))
-		{
-			flash_info[i].start[j] = flashbase + 0x8000 + (j-1)*0x4000;
+		
+		break;
+		
+	 case 4:
+		*(volatile u32*)(addr+0x5554) = 0xaaaaaaaa;
+		*(volatile u32*)(addr+0xaaa8) = 0x55555555;
+		
+		switch (il) {
+		case 1:
+			/* Issue identification command */
+			*(volatile u8*)(addr+0x5554) = 0x90;
+			
+			/* Read vendor */
+			result = *(volatile u16*)addr;
+			result <<= 16;
+		
+			/* Read device */
+			result |= *(volatile u16*)(addr+4);
+			
+			/* Return device to data mode */
+			*(volatile u8*)addr =  0xff;
+			*(volatile u8*)(addr+0x5554), 0xf0;  
+			break;
+			
+		case 2:
+			/* Issue identification command */
+			*(volatile u32*)(addr + 0x5554) = 0x00900090;
+			
+			/* Read vendor */
+			result = *(volatile u16*)addr;
+			result <<= 16;
+			
+			/* Read device */
+			result |= *(volatile u16*)(addr+4);
+			
+			/* Return device to data mode */
+			*(volatile u32*)addr =  0x00ff00ff;
+			*(volatile u32*)(addr+0x5554), 0x00f000f0;  
+			break;
+			
+		case 4:
+			/* Issue identification command */
+			*(volatile u32*)(addr+0x5554) = 0x90909090;
+			
+			/* Read vendor */
+			result = *(volatile u8*)addr;
+			result <<= 16;
+			
+			/* Read device */
+			result |= *(volatile u8*)(addr+4);
+			
+			/* Return device to data mode */
+			*(volatile u32*)addr =  0xffffffff;
+			*(volatile u32*)(addr+0x5554), 0xf0f0f0f0; 
+			break;
 		}
+		break;
+	}
+	
+	
+	
+	return result;
+}
+
+extern int _probe_flash_end;
+asm ("_probe_flash_end:\n"
+     ".long 0\n");
 
-		/* 4th 64 KB */
-		if (j == 3)
-		{
-			flash_info[i].start[j] = flashbase + 0x10000;
+static int identify_flash(unsigned address, int width)
+{
+	int is;	
+	int device;
+	int vendor;	
+	int size;
+	unsigned res;
+	
+	u32 (*_probe_flash_ptr)(u32 a, u32 bw, int il);
+	
+	size = (unsigned)&_probe_flash_end - (unsigned)_probe_flash; 
+	
+	if (size > PROBE_BUFFER_SIZE) {
+		printf("_probe_flash() routine too large (%d) %p - %p\n",
+		       size, &_probe_flash_end, _probe_flash);
+		return 0;
+	}
+	
+	memcpy(buffer, _probe_flash, size);
+	_probe_flash_ptr = (void*)buffer;
+	
+	is = disable_interrupts();
+	res = _probe_flash_ptr(address, width, 1);
+	if (is) {
+		enable_interrupts();
+	}
+	
+	
+        vendor = res >> 16;
+	device = res & 0xffff;
+	
+		
+	return res;
+}
+
+ulong flash_init(void)
+{
+	int i, j;
+	ulong size = 0;
+	
+	for (i = 0; i < SC520_MAX_FLASH_BANKS; i++) {
+		unsigned id;
+		ulong flashbase = 0;
+		int sectsize = 0; 
+		
+		memset(flash_info[i].protect, 0, CFG_MAX_FLASH_SECT);
+		switch (i) {
+		case 0:
+			flashbase = SC520_FLASH_BANK0_BASE;
+			break;
+		case 1:
+			flashbase = SC520_FLASH_BANK1_BASE;
+			break;
+		case 2:
+			flashbase = SC520_FLASH_BANK2_BASE;
+			break;
+		default:
+			panic("configured to many flash banks!\n");
+		}
+		
+		id = identify_flash(flashbase, 4);
+		switch (id & 0x00ff00ff) {
+		case 0x000100c8:
+			/* 29LV016B/29LV017B */
+			flash_info[i].flash_id =
+				(AMD_MANUFACT & FLASH_VENDMASK) |
+				(AMD_ID_LV016B & FLASH_TYPEMASK);
+			
+			flash_info[i].size = AMD29LV016B_SIZE*4;
+			flash_info[i].sector_count = AMD29LV016B_SECTORS;
+			sectsize = (AMD29LV016B_SIZE*4)/AMD29LV016B_SECTORS;
+			printf("Bank %d: 4 x AMD 29LV017B\n", i);
+			break;
+			
+			
+		default:
+			printf("Bank %d have unknown flash %08x\n", i, id);
+			flash_info[i].flash_id = FLASH_UNKNOWN;
+			continue;
+		}
+		
+		for (j = 0; j < flash_info[i].sector_count; j++) {
+			flash_info[i].start[j] = flashbase + j * sectsize;
 		}
-	    }
-	    else
-	    {
-		flash_info[i].start[j] = flashbase + (j - 3)*MAIN_SECT_SIZE;
-	    }
+		size += flash_info[i].size;
+		
+		flash_protect(FLAG_PROTECT_CLEAR,
+			      flash_info[i].start[0],
+			       flash_info[i].start[0] + flash_info[i].size - 1,
+			      &flash_info[i]);
 	}
-	size += flash_info[i].size;
-    }
-
-    /*
-     * Protect monitor and environment sectors
-     */
-    flash_protect(FLAG_PROTECT_SET,
-		  i386boot_start-CFG_FLASH_BASE,
-		  i386boot_end-CFG_FLASH_BASE,
-		  &flash_info[0]);
-
-    flash_protect(FLAG_PROTECT_SET,
-		  CFG_ENV_ADDR,
-		  CFG_ENV_ADDR + CFG_ENV_SIZE - 1,
-		  &flash_info[0]);
-    return size;
+	
+	/*
+	 * Protect monitor and environment sectors
+	 */
+	flash_protect(FLAG_PROTECT_SET,
+		      i386boot_start,
+		      i386boot_end,
+		      &flash_info[0]);
+#ifdef CFG_ENV_ADDR
+	flash_protect(FLAG_PROTECT_SET,
+		      CFG_ENV_ADDR,
+		      CFG_ENV_ADDR + CFG_ENV_SIZE - 1,
+		      &flash_info[0]);
+#endif	
+	return size;
 }
 
 /*-----------------------------------------------------------------------
  */
-void flash_print_info  (flash_info_t *info)
+void flash_print_info(flash_info_t *info)
 {
-    int i;
-
-    switch (info->flash_id & FLASH_VENDMASK)
-    {
-    case (AMD_MANUFACT & FLASH_VENDMASK):
-	printf("AMD: ");
-	break;
-    default:
-	printf("Unknown Vendor ");
-	break;
-    }
-
-    switch (info->flash_id & FLASH_TYPEMASK)
-    {
-    case (AMD_ID_LV160B & FLASH_TYPEMASK):
-	printf("2x Amd29F160BB (16Mbit)\n");
-	break;
-    default:
-	printf("Unknown Chip Type\n");
-	goto Done;
-	break;
-    }
-
-    printf("  Size: %ld MB in %d Sectors\n",
-	   info->size >> 20, info->sector_count);
-
-    printf("  Sector Start Addresses:");
-    for (i = 0; i < info->sector_count; i++)
-    {
-	if ((i % 5) == 0)
-	{
-	    printf ("\n   ");
+	int i;
+	
+	switch (info->flash_id & FLASH_VENDMASK) {
+		
+	case (AMD_MANUFACT & FLASH_VENDMASK):
+		printf("AMD:   ");
+		switch (info->flash_id & FLASH_TYPEMASK) {
+		case (AMD_ID_LV016B & FLASH_TYPEMASK):
+			printf("4x AMD29LV017B (4x16Mbit)\n");
+			break;
+		default:
+			printf("Unknown Chip Type\n");
+			goto done;
+			break;
+		}
+		
+		break;
+	default:
+		printf("Unknown Vendor ");
+		break;
 	}
-	printf (" %08lX%s", info->start[i],
-		info->protect[i] ? " (RO)" : "     ");
-    }
-    printf ("\n");
-
-Done:
+	
+	
+	printf("  Size: %ld MB in %d Sectors\n",
+	       info->size >> 20, info->sector_count);
+	
+	printf("  Sector Start Addresses:");
+	for (i = 0; i < info->sector_count; i++) {
+		if ((i % 5) == 0) {
+			printf ("\n   ");
+		}
+		printf (" %08lX%s", info->start[i],
+			info->protect[i] ? " (RO)" : "     ");
+	}
+	printf ("\n");
+	
+	done:
 }
 
 /*-----------------------------------------------------------------------
  */
 
-int	flash_erase (flash_info_t *info, int s_first, int s_last)
+/* this needs to be inlined, the SWTMRMMILLI register is reset by each read */
+#define __udelay(delay) \
+{	\
+	unsigned micro; \
+	unsigned milli=0; \
+	\
+	micro = *(volatile u16*)(0xfffef000+SC520_SWTMRMILLI); \
+         \
+	for (;;) { \
+		\
+		milli += *(volatile u16*)(0xfffef000+SC520_SWTMRMILLI); \
+		micro = *(volatile u16*)(0xfffef000+SC520_SWTMRMICRO); \
+		\
+		if ((delay) <= (micro + (milli * 1000))) { \
+			break; \
+		} \
+	} \
+} while (0) 
+
+static u32 _amd_erase_flash(u32 addr, u32 sector)
 {
-    ulong result;
-    int iflag, prot, sect;
-    int rc = ERR_OK;
-    int chip1, chip2;
-
-    /* first look for protection bits */
-
-    if (info->flash_id == FLASH_UNKNOWN)
-	return ERR_UNKNOWN_FLASH_TYPE;
-
-    if ((s_first < 0) || (s_first > s_last)) {
-	return ERR_INVAL;
-    }
+	unsigned elapsed;
+	
+	/* Issue erase */
+	*(volatile u32*)(addr + 0x5554) = 0xAAAAAAAA;
+	*(volatile u32*)(addr + 0xaaa8) = 0x55555555;
+	*(volatile u32*)(addr + 0x5554) = 0x80808080;
+	/* And one unlock */
+	*(volatile u32*)(addr + 0x5554) = 0xAAAAAAAA;
+	*(volatile u32*)(addr + 0xaaa8) = 0x55555555;
+	/* Sector erase command comes last */
+	*(volatile u32*)(addr + sector) = 0x30303030;
+	
+	elapsed = *(volatile u16*)(0xfffef000+SC520_SWTMRMILLI); /* dummy read */
+	elapsed = 0;
+	__udelay(50);
+	while (((*(volatile u32*)(addr + sector)) & 0x80808080) != 0x80808080) {
+		
+		elapsed += *(volatile u16*)(0xfffef000+SC520_SWTMRMILLI);
+		if (elapsed > ((CFG_FLASH_ERASE_TOUT/CFG_HZ) * 1000)) {
+			*(volatile u32*)(addr) = 0xf0f0f0f0;
+			return 1;			
+		}
+	}
+	
+	*(volatile u32*)(addr) = 0xf0f0f0f0;
+	
+	return 0;
+}
 
-    if ((info->flash_id & FLASH_VENDMASK) !=
-	(AMD_MANUFACT & FLASH_VENDMASK)) {
-	return ERR_UNKNOWN_FLASH_VENDOR;
-    }
+extern int _amd_erase_flash_end;
+asm ("_amd_erase_flash_end:\n"
+     ".long 0\n");
 
-    prot = 0;
-    for (sect=s_first; sect<=s_last; ++sect) {
-	if (info->protect[sect]) {
-	    prot++;
+int flash_erase(flash_info_t *info, int s_first, int s_last)
+{
+	u32 (*_erase_flash_ptr)(u32 a, u32 so);
+	int prot;
+	int sect;
+	unsigned size;
+	
+	if ((s_first < 0) || (s_first > s_last)) {
+		if (info->flash_id == FLASH_UNKNOWN) {
+			printf("- missing\n");
+		} else {
+			printf("- no sectors to erase\n");
+		}
+		return 1;
 	}
-    }
-    if (prot)
-	return ERR_PROTECTED;
-
-    /*
-     * Disable interrupts which might cause a timeout
-     * here. Remember that our exception vectors are
-     * at address 0 in the flash, and we don't want a
-     * (ticker) exception to happen while the flash
-     * chip is in programming mode.
-     */
-    iflag = disable_interrupts();
-
-    /* Start erase on unprotected sectors */
-    for (sect = s_first; sect<=s_last && !ctrlc(); sect++)
-    {
-	printf("Erasing sector %2d ... ", sect);
-
-	/* arm simple, non interrupt dependent timer */
-	reset_timer();
-
-	if (info->protect[sect] == 0)
-	{	/* not protected */
-	    vu_long *addr = (vu_long *)(info->start[sect]);
-
-	    MEM_FLASH_ADDR1 = CMD_UNLOCK1;
-	    MEM_FLASH_ADDR2 = CMD_UNLOCK2;
-	    MEM_FLASH_ADDR1 = CMD_ERASE_SETUP;
-
-	    MEM_FLASH_ADDR1 = CMD_UNLOCK1;
-	    MEM_FLASH_ADDR2 = CMD_UNLOCK2;
-	    *addr = CMD_ERASE_CONFIRM;
-
-	    /* wait until flash is ready */
-	    chip1 = chip2 = 0;
-
-	    do
-	    {
-		result = *addr;
-
-		/* check timeout */
-		if (get_timer(0) > CFG_FLASH_ERASE_TOUT)
-		{
-		    MEM_FLASH_ADDR1 = CMD_READ_ARRAY;
-		    chip1 = TMO;
-		    break;
+	
+	if ((info->flash_id & FLASH_VENDMASK) == (AMD_MANUFACT & FLASH_VENDMASK)) {
+		size = (unsigned)&_amd_erase_flash_end - (unsigned)_amd_erase_flash; 
+		
+		if (size > PROBE_BUFFER_SIZE) {
+			printf("_amd_erase_flash() routine too large (%d) %p - %p\n",
+			       size, &_amd_erase_flash_end, _amd_erase_flash);
+			return 0;
 		}
-
-		if (!chip1 && (result & 0xFFFF) & BIT_ERASE_DONE)
-			chip1 = READY;
-
-		if (!chip1 && (result & 0xFFFF) & BIT_PROGRAM_ERROR)
-			chip1 = ERR;
-
-		if (!chip2 && (result >> 16) & BIT_ERASE_DONE)
-			chip2 = READY;
-
-		if (!chip2 && (result >> 16) & BIT_PROGRAM_ERROR)
-			chip2 = ERR;
-
-	    }  while (!chip1 || !chip2);
-
-	    MEM_FLASH_ADDR1 = CMD_READ_ARRAY;
-
-	    if (chip1 == ERR || chip2 == ERR)
-	    {
-		rc = ERR_PROG_ERROR;
-		goto outahere;
-	    }
-	    if (chip1 == TMO)
-	    {
-		rc = ERR_TIMOUT;
-		goto outahere;
-	    }
-
-	    printf("ok.\n");
+		
+		memcpy(buffer, _amd_erase_flash, size);
+		_erase_flash_ptr = (void*)buffer;
+	
+	}  else {
+		printf ("Can't erase unknown flash type - aborted\n");
+		return 1;
 	}
-	else /* it was protected */
-	{
-	    printf("protected!\n");
+	
+	prot = 0;
+	for (sect=s_first; sect<=s_last; ++sect) {
+		if (info->protect[sect]) {
+			prot++;
+		}
+	}
+	
+	if (prot) {
+		printf ("- Warning: %d protected sectors will not be erased!\n", prot);
+	} else {
+		printf ("\n");
+	}
+		
+	
+	/* Start erase on unprotected sectors */
+	for (sect = s_first; sect<=s_last; sect++) {
+		
+		if (info->protect[sect] == 0) { /* not protected */
+			int res;
+			int flag;
+			
+			/* Disable interrupts which might cause a timeout here */
+			flag = disable_interrupts();
+			
+			res = _erase_flash_ptr(info->start[0], info->start[sect]-info->start[0]);
+			
+			/* re-enable interrupts if necessary */
+			if (flag) {
+				enable_interrupts();
+			}
+			
+			
+			if (res) {
+				printf("Erase timed out, sector %d\n", sect);
+				return res;
+			}
+			
+			putc('.');			
+		}		
 	}
-    }
-
-    if (ctrlc())
-      printf("User Interrupt!\n");
-
-outahere:
-    /* allow flash to settle - wait 10 ms */
-    udelay(10000);
-
-    if (iflag)
-      enable_interrupts();
-
 
-    return rc;
+	
+	return 0;
 }
 
 /*-----------------------------------------------------------------------
- * Copy memory to flash
+ * Write a word to Flash, returns:
+ * 0 - OK
+ * 1 - write timeout
+ * 2 - Flash not erased
  */
-
-volatile static int write_word (flash_info_t *info, ulong dest, ulong data)
+static int _amd_write_word(unsigned start, unsigned dest, unsigned data)
 {
-    vu_long *addr = (vu_long *)dest;
-    ulong result;
-    int rc = ERR_OK;
-    int iflag;
-    int chip1, chip2;
-
-    /*
-     * Check if Flash is (sufficiently) erased
-     */
-    result = *addr;
-    if ((result & data) != data)
-        return ERR_NOT_ERASED;
-
-
-    /*
-     * Disable interrupts which might cause a timeout
-     * here. Remember that our exception vectors are
-     * at address 0 in the flash, and we don't want a
-     * (ticker) exception to happen while the flash
-     * chip is in programming mode.
-     */
-    iflag = disable_interrupts();
-
-    MEM_FLASH_ADDR1 = CMD_UNLOCK1;
-    MEM_FLASH_ADDR2 = CMD_UNLOCK2;
-    MEM_FLASH_ADDR1 = CMD_UNLOCK_BYPASS;
-    *addr = CMD_PROGRAM;
-    *addr = data;
-
-    /* arm simple, non interrupt dependent timer */
-    reset_timer();
-
-    /* wait until flash is ready */
-    chip1 = chip2 = 0;
-    do
-    {
-	result = *addr;
-
-	/* check timeout */
-	if (get_timer(0) > CFG_FLASH_ERASE_TOUT)
-	{
-	    chip1 = ERR | TMO;
-	    break;
+	volatile u32 *addr2 = (u32*)start;
+	volatile u32 *dest2 = (u32*)dest;
+	volatile u32 *data2 = (u32*)&data;
+	unsigned elapsed;
+	
+	/* Check if Flash is (sufficiently) erased */
+	if ((*((volatile u32*)dest) & (u32)data) != (u32)data) {
+		return 2;
 	}
-	if (!chip1 && ((result & 0x80) == (data & 0x80)))
-		chip1 = READY;
-
-	if (!chip1 && ((result & 0xFFFF) & BIT_PROGRAM_ERROR))
-	{
-		result = *addr;
-
-		if ((result & 0x80) == (data & 0x80))
-			chip1 = READY;
-		else
-			chip1 = ERR;
-	}
-
-	if (!chip2 && ((result & (0x80 << 16)) == (data & (0x80 << 16))))
-		chip2 = READY;
-
-	if (!chip2 && ((result >> 16) & BIT_PROGRAM_ERROR))
-	{
-		result = *addr;
-
-		if ((result & (0x80 << 16)) == (data & (0x80 << 16)))
-			chip2 = READY;
-		else
-			chip2 = ERR;
+		
+	addr2[0x5554] = 0xAAAAAAAA;
+	addr2[0xaaa8] = 0x55555555;
+	addr2[0x5554] = 0xA0A0A0A0;
+	
+	dest2[0] = data;
+	
+	elapsed = *(volatile u16*)(0xfffef000+SC520_SWTMRMILLI); /* dummy read */
+	elapsed = 0;
+	
+	/* data polling for D7 */
+	while ((dest2[0] & 0x80808080) != (data2[0] & 0x80808080)) {
+		elapsed += *(volatile u16*)(0xfffef000+SC520_SWTMRMILLI);
+		if (elapsed > ((CFG_FLASH_WRITE_TOUT/CFG_HZ) * 1000)) {
+			addr2[0] = 0xf0f0f0f0;
+			return 1;			
+		}
 	}
 
-    }  while (!chip1 || !chip2);
-
-    *addr = CMD_READ_ARRAY;
+	
+	addr2[0] = 0xf0f0f0f0;
+	
+	return 0;
+}
 
-    if (chip1 == ERR || chip2 == ERR || *addr != data)
-        rc = ERR_PROG_ERROR;
+extern int _amd_write_word_end;
+asm ("_amd_write_word_end:\n"
+     ".long 0\n");
 
-    if (iflag)
-      enable_interrupts();
 
 
-    return rc;
-}
 
 /*-----------------------------------------------------------------------
- * Copy memory to flash.
+ * Copy memory to flash, returns:
+ * 0 - OK
+ * 1 - write timeout
+ * 2 - Flash not erased
+ * 3 - Unsupported flash type
  */
 
-int write_buff (flash_info_t *info, uchar *src, ulong addr, ulong cnt)
+int write_buff(flash_info_t *info, uchar *src, ulong addr, ulong cnt)
 {
-    ulong cp, wp, data;
-    int l;
-    int i, rc;
+	ulong cp, wp, data;
+	int i, l, rc;
+	int flag;
+	u32 (*_write_word_ptr)(unsigned start, unsigned dest, unsigned data);
+	unsigned size;
+	
+	if ((info->flash_id & FLASH_VENDMASK) == (AMD_MANUFACT & FLASH_VENDMASK)) {
+		size = (unsigned)&_amd_write_word_end - (unsigned)_amd_write_word; 
+		
+		if (size > PROBE_BUFFER_SIZE) {
+			printf("_amd_write_word() routine too large (%d) %p - %p\n",
+			       size, &_amd_write_word_end, _amd_write_word);
+			return 0;
+		}
+		
+		memcpy(buffer, _amd_write_word, size);
+		_write_word_ptr = (void*)buffer;
+	
+	} else {
+		printf ("Can't program unknown flash type - aborted\n");
+		return 3;
+	}
 
-    wp = (addr & ~3);	/* get lower word aligned address */
 
-    /*
-     * handle unaligned start bytes
-     */
-    if ((l = addr - wp) != 0) {
-	data = 0;
-	for (i=0, cp=wp; i<l; ++i, ++cp) {
-	    data = (data >> 8) | (*(uchar *)cp << 24);
+	wp = (addr & ~3);	/* get lower word aligned address */
+	
+
+	/*
+	 * handle unaligned start bytes
+	 */
+	if ((l = addr - wp) != 0) {
+		data = 0;
+		for (i=0, cp=wp; i<l; ++i, ++cp) {
+			data |= (*(uchar *)cp) << (8*i);
+		}
+		for (; i<4 && cnt>0; ++i) {
+			data |= *src++ << (8*i);
+			--cnt;
+			++cp;
+		}
+		for (; cnt==0 && i<4; ++i, ++cp) {
+			data |= (*(uchar *)cp)  << (8*i);
+		}
+		
+		/* Disable interrupts which might cause a timeout here */
+		flag = disable_interrupts();
+		
+		rc = _write_word_ptr(info->start[0], wp, data);
+		
+		/* re-enable interrupts if necessary */
+		if (flag) {
+			enable_interrupts();
+		}
+		if (rc != 0) {
+			return rc;
+		}
+		wp += 4;
 	}
-	for (; i<4 && cnt>0; ++i) {
-	    data = (data >> 8) | (*src++ << 24);
-	    --cnt;
-	    ++cp;
+	
+	/*
+	 * handle word aligned part
+	 */
+	while (cnt >= 4) {
+		data = 0;
+			       
+		for (i=0; i<4; ++i) {
+			data |= *src++ << (8*i);
+		}
+		
+		/* Disable interrupts which might cause a timeout here */
+		flag = disable_interrupts();
+
+		rc = _write_word_ptr(info->start[0], wp, data);
+		
+		/* re-enable interrupts if necessary */
+		if (flag) {
+			enable_interrupts();
+		}
+		if (rc != 0) {
+			return rc;
+		}
+		wp  += 4;
+		cnt -= 4;
 	}
-	for (; cnt==0 && i<4; ++i, ++cp) {
-	    data = (data >> 8) | (*(uchar *)cp << 24);
+	
+	if (cnt == 0) {
+		return 0;
 	}
-
-	if ((rc = write_word(info, wp, data)) != 0) {
-	    return (rc);
+	
+	/*
+	 * handle unaligned tail bytes
+	 */
+	data = 0;
+	for (i=0, cp=wp; i<4 && cnt>0; ++i, ++cp) {
+		data |= *src++ << (8*i);
+		--cnt;
 	}
-	wp += 4;
-    }
-
-    /*
-     * handle word aligned part
-     */
-    while (cnt >= 4) {
-	data = *((vu_long*)src);
-	if ((rc = write_word(info, wp, data)) != 0) {
-	    return (rc);
+	
+	for (; i<4; ++i, ++cp) {
+		data |= (*(uchar *)cp) << (8*i);
 	}
-	src += 4;
-	wp  += 4;
-	cnt -= 4;
-    }
-
-    if (cnt == 0) {
-	return ERR_OK;
-    }
-
-    /*
-     * handle unaligned tail bytes
-     */
-    data = 0;
-    for (i=0, cp=wp; i<4 && cnt>0; ++i, ++cp) {
-	data = (data >> 8) | (*src++ << 24);
-	--cnt;
-    }
-    for (; i<4; ++i, ++cp) {
-	data = (data >> 8) | (*(uchar *)cp << 24);
-    }
-
-    return write_word(info, wp, data);
+
+	/* Disable interrupts which might cause a timeout here */
+	flag = disable_interrupts();
+
+	rc = _write_word_ptr(info->start[0], wp, data);
+	
+	/* re-enable interrupts if necessary */
+	if (flag) {
+		enable_interrupts();
+	}
+	
+	return rc;
+	
 }
+
+