new cache_as_ram support

git-svn-id: svn://svn.coreboot.org/coreboot/trunk@2232 2b7e53f0-3cfb-0310-b3e9-8179ed1497e1

56 files changed, 2266 insertions, 549 deletions

diff --git a/src/arch/i386/include/arch/romcc_io.h b/src/arch/i386/include/arch/romcc_io.h
index c0c2250f98..adda3f8ef8 100644
--- a/src/arch/i386/include/arch/romcc_io.h
+++ b/src/arch/i386/include/arch/romcc_io.h
@@ -178,6 +178,24 @@ static device_t pci_locate_device(unsigned pci_id, device_t dev)
 	return PCI_DEV_INVALID;
 }
 
+static device_t pci_locate_device_on_bus(unsigned pci_id, unsigned bus)
+{
+        device_t dev, last;
+
+        dev = PCI_DEV(bus, 0, 0);
+        last = PCI_DEV(bus, 31, 7);
+
+        for(; dev <=last; dev += PCI_DEV(0,0,1)) {
+                unsigned int id;
+                id = pci_read_config32(dev, 0);
+                if (id == pci_id) {
+                        return dev;
+                }
+        }
+        return PCI_DEV_INVALID;
+}
+
+
 
 /* Generic functions for pnp devices */
 static inline __attribute__((always_inline)) void pnp_write_config(device_t dev, uint8_t reg, uint8_t value)
diff --git a/src/config/Options.lb b/src/config/Options.lb
index b5569673d7..1342366fe9 100644
--- a/src/config/Options.lb
+++ b/src/config/Options.lb
@@ -280,11 +280,17 @@ define DCACHE_RAM_BASE
 	comment "Base address of data cache when using it for temporary RAM"
 end
 define DCACHE_RAM_SIZE
-	default none
+	default 0x1000
 	format "0x%x"
-	export used
+	export always
 	comment "Size of data cache when using it for temporary RAM"
 end
+define DCACHE_RAM_GLOBAL_VAR_SIZE
+	default 0
+	format "0x%x"
+	export always
+	comment "Size of region that for global variable of cache as ram stage"
+end
 define XIP_ROM_BASE
 	default 0
 	format "0x%x"
@@ -310,7 +316,7 @@ define CONFIG_UNCOMPRESSED
 end
 define CONFIG_LB_MEM_TOPK
 	format "%d"
-	default 1024
+	default 2048
 	export always
 	comment "Kilobytes of memory to initialized before executing code from RAM"
 end
diff --git a/src/cpu/amd/car/cache_as_ram.inc b/src/cpu/amd/car/cache_as_ram.inc
index c7de4148a4..1cf141230b 100644
--- a/src/cpu/amd/car/cache_as_ram.inc
+++ b/src/cpu/amd/car/cache_as_ram.inc
@@ -1,8 +1,10 @@
 /* by yhlu 6.2005 */
 /* yhlu 2005.12 make it support HDT Memory Debuggers with Disassmbly, please select the PCI Bus mem for Phys Type*/
-/* We may need 4K bytes only */
+/* yhlu 2006.3 copy data from cache to ram and reserve 0x1000 for global variables */
 #define CacheSize DCACHE_RAM_SIZE
 #define CacheBase (0xd0000 - CacheSize)
+/* leave some space for global variable to pass to RAM stage */
+#define GlobalVarSize DCACHE_RAM_GLOBAL_VAR_SIZE
 
 #include <cpu/x86/mtrr.h>
 #include <cpu/amd/mtrr.h>
@@ -147,7 +149,7 @@ clear_fixed_var_mtrr_out:
 #endif /*USE_FALLBACK_IMAGE == 1*/
 
 	/* set up the stack pointer */
-	movl	$(CacheBase+CacheSize-4), %eax
+	movl	$(CacheBase+CacheSize - 4 - GlobalVarSize), %eax 
 	movl    %eax, %esp
 
 	/* Restore the BIST result */
diff --git a/src/cpu/amd/car/clear_init_ram.c b/src/cpu/amd/car/clear_init_ram.c
new file mode 100644
index 0000000000..593c0ddfd6
--- /dev/null
+++ b/src/cpu/amd/car/clear_init_ram.c
@@ -0,0 +1,59 @@
+/* by yhlu 6.2005 */
+/* be warned, this file will be used core 0/node 0 only */
+
+static void __attribute__((noinline)) clear_init_ram(void)
+{
+	// gcc 3.4.5 will inline the copy_and_run and clear_init_ram in post_cache_as_ram
+	// will reuse %edi as 0 from clear_memory for copy_and_run part, actually it is increased already
+	// so noline clear_init_ram
+        clear_memory(0,  ((CONFIG_LB_MEM_TOPK<<10) - DCACHE_RAM_SIZE));
+
+}
+
+/* be warned, this file will be used by core other than core 0/node 0 or core0/node0 when cpu_reset*/
+static inline __attribute__((always_inline)) void set_init_ram_access(void)
+{
+        __asm__ volatile (
+
+	"pushl %%ecx\n\t"
+	"pushl %%edx\n\t"
+	"pushl %%eax\n\t"
+
+        /* enable caching for first 1M using variable mtrr */
+        "movl    $0x200, %%ecx\n\t"
+        "xorl    %%edx, %%edx\n\t"
+        "movl     $(0 | 6), %%eax\n\t"
+//      "movl     $(0 | MTRR_TYPE_WRBACK), %%eax\n\t"
+        "wrmsr\n\t"
+
+        "movl    $0x201, %%ecx\n\t"
+        "movl    $0x0000000f, %%edx\n\t"
+#if CONFIG_USE_INIT
+        "movl    %%esi, %%eax\n\t"
+#else
+        "movl    $((~(( 0 + (CONFIG_LB_MEM_TOPK<<10) ) -1)) | 0x800), %%eax\n\t"
+#endif
+        "wrmsr\n\t"
+
+#if 0
+        /* enable caching for 64K using fixed mtrr */
+        "movl    $0x26e, %%ecx\n\t"  /* fix4k_f0000*/
+        "movl    $0x1e1e1e1e, %%eax\n\t" /* WB MEM type */
+        "movl    %%eax, %%edx\n\t"
+        "wrmsr\n\t"
+        "movl    $0x26f, %%ecx\n\t" /* fix4k_f8000*/
+        "wrmsr\n\t"
+#endif
+
+	"popl	%%eax\n\t"
+	"popl   %%edx\n\t"
+	"popl	%%ecx\n\t"
+
+        :
+        :
+#if CONFIG_USE_INIT
+        "S"((~(( 0 + (CONFIG_LB_MEM_TOPK<<10) ) -1)) | 0x800)
+#endif
+        );
+}
+
diff --git a/src/cpu/amd/car/copy_and_run.c b/src/cpu/amd/car/copy_and_run.c
index 55068c5511..68e9e35482 100644
--- a/src/cpu/amd/car/copy_and_run.c
+++ b/src/cpu/amd/car/copy_and_run.c
@@ -37,7 +37,7 @@ static inline void print_debug_cp_run(const char *strval, uint32_t val)
 #endif
 }
 
-static void copy_and_run(unsigned cpu_reset)
+static void copy_and_run(void)
 {
 	uint8_t *src, *dst; 
         unsigned long ilen = 0, olen = 0, last_m_off =  1;
@@ -54,7 +54,7 @@ static void copy_and_run(unsigned cpu_reset)
 		"subl %1, %2\n\t"
 		: "=a" (src), "=b" (dst), "=c" (olen)
 	);
-	memcpy(src, dst, olen);
+	memcpy(dst, src, olen);
 #else 
 
         __asm__ volatile (
@@ -65,14 +65,13 @@ static void copy_and_run(unsigned cpu_reset)
 
 	print_debug_cp_run("src=",(uint32_t)src); 
 	print_debug_cp_run("dst=",(uint32_t)dst);
-	
-//	dump_mem(src, src+0x100);
 
         for(;;) {
                 unsigned int m_off, m_len;
                 while(GETBIT(bb, src, ilen)) {
                         dst[olen++] = src[ilen++];
                 }
+
                 m_off = 1;
                 do {
                         m_off = m_off*2 + GETBIT(bb, src, ilen);
@@ -109,24 +108,13 @@ static void copy_and_run(unsigned cpu_reset)
                 }
         }
 #endif
-//	dump_mem(dst, dst+0x100);
 
 	print_debug_cp_run("linxbios_ram.bin length = ", olen);
 
 	print_debug("Jumping to LinuxBIOS.\r\n");
 
-	if(cpu_reset == 1 ) {
-		__asm__ volatile (
-			"movl $0xffffffff, %ebp\n\t"
-		);
-	}
-	else {
-                __asm__ volatile (
-                        "xorl %ebp, %ebp\n\t"
-                );
-	}
-	
-	__asm__ volatile (
+        __asm__ volatile (
+                "xorl %ebp, %ebp\n\t" /* cpu_reset for hardwaremain dummy */
 		"cli\n\t"
 		"leal    _iseg, %edi\n\t"
 		"jmp     *%edi\n\t"
diff --git a/src/cpu/amd/car/disable_cache_as_ram.c b/src/cpu/amd/car/disable_cache_as_ram.c
index 06a558f413..56e60eefb7 100644
--- a/src/cpu/amd/car/disable_cache_as_ram.c
+++ b/src/cpu/amd/car/disable_cache_as_ram.c
@@ -3,10 +3,7 @@
 static inline __attribute__((always_inline)) void disable_cache_as_ram(void)
 {
         __asm__ volatile (
-	/* 
-	FIXME : backup stack in CACHE_AS_RAM into mmx and sse and after we get STACK up, we restore that.
-		It is only needed if we want to go back
-	*/
+
         /* We don't need cache as ram for now on */
         /* disable cache */
         "movl    %cr0, %eax\n\t"
@@ -43,5 +40,30 @@ static inline __attribute__((always_inline)) void disable_cache_as_ram(void)
         "movl    %cr0, %eax\n\t"
         "andl    $0x9fffffff,%eax\n\t"
         "movl    %eax, %cr0\n\t"
+
         );
 }
+/* be warned, this file will be used core 0 / node 0 and ram stack is ready*/
+
+static void disable_cache_as_ram_bsp(void)
+{
+        __asm__ volatile (
+
+        "pushl	%ecx\n\t"
+        "pushl 	%edx\n\t"
+        "pushl 	%eax\n\t"
+
+	);
+
+	disable_cache_as_ram();
+
+	__asm__ volatile (
+
+        "popl   %eax\n\t"
+        "popl   %edx\n\t"
+        "popl   %ecx\n\t"
+
+        );
+}
+
+
diff --git a/src/cpu/amd/car/post_cache_as_ram.c b/src/cpu/amd/car/post_cache_as_ram.c
index 345e4c34b0..0a91326b85 100644
--- a/src/cpu/amd/car/post_cache_as_ram.c
+++ b/src/cpu/amd/car/post_cache_as_ram.c
@@ -1,7 +1,9 @@
-
+/* 2005.6 by yhlu 
+ * 2006.3 yhlu add copy data from CAR to ram
+ */
 #include "cpu/amd/car/disable_cache_as_ram.c"
 
-#include "cpu/amd/car/clear_1m_ram.c"
+#include "cpu/amd/car/clear_init_ram.c"
 
 static inline void print_debug_pcar(const char *strval, uint32_t val)
 {
@@ -12,11 +14,18 @@ static inline void print_debug_pcar(const char *strval, uint32_t val)
 #endif
 }
 
-
-
-static void post_cache_as_ram(unsigned cpu_reset)
+static void inline __attribute__((always_inline))  memcopy(void *dest, const void *src, unsigned long bytes)
 {
+        __asm__ volatile(
+                "cld\n\t"
+                "rep movsl\n\t"
+                : /* No outputs */
+                : "S" (src), "D" (dest), "c" ((bytes)>>2)
+                );
+}
 
+static void post_cache_as_ram(void)
+{
 
 #if 1
         {
@@ -30,60 +39,50 @@ static void post_cache_as_ram(unsigned cpu_reset)
         }
 #endif
 
-        print_debug_pcar("cpu_reset = ",cpu_reset);
+	unsigned testx = 0x5a5a5a5a;
+	print_debug_pcar("testx = ", testx);
 
-        if(cpu_reset == 0) {
-                print_debug("Clearing initial memory region: ");
-        }
-        print_debug("No cache as ram now - ");
+	/* copy data from cache as ram to 
+		ram need to set CONFIG_LB_MEM_TOPK to 2048 and use var mtrr instead.
+	 */
+#if CONFIG_LB_MEM_TOPK <= 1024
+        #error "You need to set CONFIG_LB_MEM_TOPK greater than 1024"
+#endif
+	
+	set_init_ram_access();
 
-        /* store cpu_reset to ebx */
+	print_debug("Copying data from cache to ram -- switching to use ram as stack... ");
+
+	/* from here don't store more data in CAR */
         __asm__ volatile (
-                "movl %0, %%ebx\n\t"
-                ::"a" (cpu_reset)
+        	"pushl  %eax\n\t"
         );
-
-	disable_cache_as_ram();
-
-        if(cpu_reset==0) { // cpu_reset don't need to clear it 
-		clear_1m_ram();
-        }
-	else {
-		set_1m_ram();
-	}
-
+        memcopy((CONFIG_LB_MEM_TOPK<<10)-DCACHE_RAM_SIZE, DCACHE_RAM_BASE, DCACHE_RAM_SIZE); //inline
         __asm__ volatile (
                 /* set new esp */ /* before _RAMBASE */
                 "subl   %0, %%ebp\n\t"
                 "subl   %0, %%esp\n\t"
-                ::"a"( (DCACHE_RAM_BASE + DCACHE_RAM_SIZE)- _RAMBASE )
+                ::"a"( (DCACHE_RAM_BASE + DCACHE_RAM_SIZE)- (CONFIG_LB_MEM_TOPK<<10) )
+        ); // We need to push %eax to the stack (CAR) before copy stack and pop it later after copy stack and change esp
+        __asm__ volatile (
+	        "popl   %eax\n\t"
         );
+	/* We can put data to stack again */
 
-        {
-                unsigned new_cpu_reset;
-
-                /* get back cpu_reset from ebx */
-                __asm__ volatile (
-                        "movl %%ebx, %0\n\t"
-                        :"=a" (new_cpu_reset)
-                );
-
-                print_debug("Use Ram as Stack now - "); /* but We can not go back any more, we lost old stack data in cache as ram*/
+        /* only global variable sysinfo in cache need to be offset */
+        print_debug("Done\r\n");
+        print_debug_pcar("testx = ", testx);
 
-                if(new_cpu_reset==0) {
-                        print_debug("done\r\n");
-                } else
-                {
-                        print_debug("\r\n");
-                }
+	print_debug("Disabling cache as ram now \r\n");
+	disable_cache_as_ram_bsp();  
 
-                print_debug_pcar("new_cpu_reset = ", new_cpu_reset);
+        print_debug("Clearing initial memory region: ");
+        clear_init_ram(); //except the range from [(CONFIG_LB_MEM_TOPK<<10) - DCACHE_RAM_SIZE, (CONFIG_LB_MEM_TOPK<<10)), that is used as stack in ram
+        print_debug("Done\r\n");
 
-
-                /*copy and execute linuxbios_ram */
-                copy_and_run(new_cpu_reset);
-                /* We will not return */
-        }
+        /*copy and execute linuxbios_ram */
+        copy_and_run();
+        /* We will not return */
 
 	print_debug("should not be here -\r\n");
 
diff --git a/src/cpu/amd/model_fxx/init_cpus.c b/src/cpu/amd/model_fxx/init_cpus.c
index 6a070d56ec..e60aa35bf9 100644
--- a/src/cpu/amd/model_fxx/init_cpus.c
+++ b/src/cpu/amd/model_fxx/init_cpus.c
@@ -1,7 +1,6 @@
 //it takes the ENABLE_APIC_EXT_ID and APIC_ID_OFFSET and LIFT_BSP_APIC_ID
 #ifndef K8_SET_FIDVID
 	#define K8_SET_FIDVID 0
-	
 #endif
 
 #ifndef K8_SET_FIDVID_CORE0_ONLY
@@ -42,7 +41,7 @@ static void for_each_ap(unsigned bsp_apicid, unsigned core0_only, process_ap_t p
                 j = ((pci_read_config32(PCI_DEV(0, 0x18+i, 3), 0xe8) >> 12) & 3);
                 if(nb_cfg_54) {
  	               if(j == 0 ){ // if it is single core, we need to increase siblings for apic calculation 
-        	               e0_later_single_core = is_e0_later_in_bsp(i);  // single core
+       	               e0_later_single_core = is_e0_later_in_bsp(i);  // single core
                        } 
                        if(e0_later_single_core) {
                                 j=1;
@@ -204,7 +203,7 @@ static unsigned init_cpus(unsigned cpu_init_detectedx)
 		/* get the apicid, it may be lifted already */
 		apicid = lapicid();
 
-#if 1 
+#if 0 
 		// show our apicid, nodeid, and coreid
 		if( id.coreid==0 ) {
 			if (id.nodeid!=0) //all core0 except bsp
@@ -219,21 +218,9 @@ static unsigned init_cpus(unsigned cpu_init_detectedx)
 #endif
 
                 if (cpu_init_detectedx) {
-		#if RAMINIT_SYSINFO == 1
-			//We need to init sblnk and sbbusn, because it is called before ht_setup_chains_x
-		        sysinfo->sblnk = get_sblnk();
-			sysinfo->sbbusn = node_link_to_bus(0, sysinfo->sblnk);
-		#endif
-			print_apicid_nodeid_coreid(apicid, id, "\r\n\r\n\r\nINIT detect from ");
-
-			print_debug("\r\nIssuing SOFT_RESET...\r\n");
-
-			#if RAMINIT_SYSINFO == 1
-                        soft_reset(sysinfo);
-			#else
-			soft_reset();
-			#endif
-
+                        print_apicid_nodeid_coreid(apicid, id, "\r\n\r\n\r\nINIT detected from ");
+                        print_debug("\r\nIssuing SOFT_RESET...\r\n");
+                        soft_reset();
                 }
 
 		if(id.coreid==0) {
@@ -256,8 +243,8 @@ static unsigned init_cpus(unsigned cpu_init_detectedx)
 			wait_cpu_state(bsp_apicid, 0x44);
 			lapic_write(LAPIC_MSG_REG, (apicid<<24) | 0x44); // bsp can not check it before stop_this_cpu
 
+			set_init_ram_access(); //inline
 			disable_cache_as_ram(); // inline
-			set_1m_ram(); // inline
                         stop_this_cpu(); // inline, it will stop all cores except node0/core0 the bsp .... 
                 }
 
diff --git a/src/cpu/x86/lapic/lapic_cpu_init.c b/src/cpu/x86/lapic/lapic_cpu_init.c
index d7f7124deb..279101b74b 100644
--- a/src/cpu/x86/lapic/lapic_cpu_init.c
+++ b/src/cpu/x86/lapic/lapic_cpu_init.c
@@ -25,7 +25,7 @@
  */ 
 static unsigned long get_valid_start_eip(unsigned long orig_start_eip)
 {
-	return (unsigned long)orig_start_eip & 0xfffff; // 20 bit 
+	return (unsigned long)orig_start_eip & 0xffff; // 16 bit to avoid 0xa0000 
 }
 
 static void copy_secondary_start_to_1m_below(void) 
@@ -43,7 +43,7 @@ static void copy_secondary_start_to_1m_below(void)
         code_size = (unsigned long)_secondary_start_end - (unsigned long)_secondary_start;
 
         /* copy the _secondary_start to the ram below 1M*/
-        memcpy(start_eip, (unsigned long)_secondary_start, code_size);
+        memcpy((unsigned char *)start_eip, (unsigned char *)_secondary_start, code_size);
 
         printk_debug("start_eip=0x%08lx, offset=0x%08lx, code_size=0x%08lx\n", start_eip, ((unsigned long)_secondary_start - start_eip), code_size);
 #endif
@@ -117,7 +117,12 @@ static int lapic_start_cpu(unsigned long apicid)
 		return 0;
 	}
 
+#if _RAMBASE >= 0x100000
 	start_eip = get_valid_start_eip((unsigned long)_secondary_start);
+#else
+	start_eip = (unsigned long)_secondary_start;
+#endif
+
 	printk_debug("start_eip=0x%08lx\n", start_eip);
        
 	num_starts = 2;
diff --git a/src/devices/hypertransport.c b/src/devices/hypertransport.c
index 10d4420fee..c61efa6c7b 100644
--- a/src/devices/hypertransport.c
+++ b/src/devices/hypertransport.c
@@ -424,17 +424,8 @@ unsigned int hypertransport_scan_chain(struct bus *bus,
 		}
 
 		flags &= ~0x1f; /* mask out base Unit ID */
-#if CK804_DEVN_BASE==0  
-               if((dev->vendor == 0x10de) && (dev->device == 0x005e)) {
-                       next_unitid = 0;
-               } 
-               else {
-#endif
                        flags |= next_unitid & 0x1f;
                        pci_write_config16(dev, pos + PCI_CAP_FLAGS, flags);
-#if CK804_DEVN_BASE==0 
-               }
-#endif
 
 		/* Update the Unitd id in the device structure */
 		static_count = 1;
@@ -473,11 +464,6 @@ unsigned int hypertransport_scan_chain(struct bus *bus,
 			dev->vendor, dev->device, 
 			(dev->enabled? "enabled": "disabled"), next_unitid);
 
-#if CK804_DEVN_BASE==0 
-		if ((dev->vendor == 0x10de) && (dev->device == 0x005e)) {
-                      break; // CK804 can not change unitid, so it only can be alone in the link
-                }
-#endif
 
 	} while((last_unitid != next_unitid) && (next_unitid <= (max_devfn >> 3)));
  end_of_chain:
diff --git a/src/include/cpu/x86/mem.h b/src/include/cpu/x86/mem.h
index 4849ed289a..532fc79e0a 100644
--- a/src/include/cpu/x86/mem.h
+++ b/src/include/cpu/x86/mem.h
@@ -4,15 +4,13 @@
 /* Optimized generic x86 assembly for clearing memory */
 static inline void clear_memory(void *addr, unsigned long size)
 {
-	asm volatile(
-		"1: \n\t"
-		"movl %0, (%1)\n\t"
-		"addl $4, %1\n\t"
-		"subl $4, %2\n\t"
-		"jnz 1b\n\t"
-		: /* No outputs */
-		: "a" (0), "D" (addr), "c" (size)
-		);
+        asm volatile(
+                "cld \n\t"
+                "rep stosl\n\t"
+                : /* No outputs */
+                : "a" (0), "D" (addr), "c" (size>>2)
+                );
+
 }
 
 #endif /* CPU_X86_MEM_H */
diff --git a/src/mainboard/agami/aruma/cache_as_ram_auto.c b/src/mainboard/agami/aruma/cache_as_ram_auto.c
index b820c52aef..8d9e13cb45 100644
--- a/src/mainboard/agami/aruma/cache_as_ram_auto.c
+++ b/src/mainboard/agami/aruma/cache_as_ram_auto.c
@@ -204,7 +204,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
 	unsigned bsp_apicid = 0;
 	struct mem_controller ctrl[8];
 	unsigned nodes;
@@ -253,6 +252,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	memreset_setup();
 	sdram_initialize(nodes, ctrl);
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 
 }
diff --git a/src/mainboard/amd/serengeti_leopard/Config.lb b/src/mainboard/amd/serengeti_leopard/Config.lb
index b6ee7180ec..ea92f8e201 100644
--- a/src/mainboard/amd/serengeti_leopard/Config.lb
+++ b/src/mainboard/amd/serengeti_leopard/Config.lb
@@ -117,8 +117,6 @@ if HAVE_ACPI_TABLES
         end
 end
 
-object reset.o
-
 if USE_DCACHE_RAM
 
 	if CONFIG_USE_INIT
diff --git a/src/mainboard/amd/serengeti_leopard/cache_as_ram_auto.c b/src/mainboard/amd/serengeti_leopard/cache_as_ram_auto.c
index b81caf0406..965e439f49 100644
--- a/src/mainboard/amd/serengeti_leopard/cache_as_ram_auto.c
+++ b/src/mainboard/amd/serengeti_leopard/cache_as_ram_auto.c
@@ -10,21 +10,13 @@
 //use by raminit
 #define K8_4RANK_DIMM_SUPPORT 1
 
-//use bu init_cpus
-#if 0
-        #define ENABLE_APIC_EXT_ID 1
-        #define APIC_ID_OFFSET 0x10
-        #define LIFT_BSP_APIC_ID 0
-#else
-        #define ENABLE_APIC_EXT_ID 0
-#endif
-
 //used by incoherent_ht
 //#define K8_SCAN_PCI_BUS 1
 //#define K8_ALLOCATE_IO_RANGE 1
 
 #include <stdint.h>
 #include <device/pci_def.h>
+#include <device/pci_ids.h>
 #include <arch/io.h>
 #include <device/pnp_def.h>
 #include <arch/romcc_io.h>
@@ -33,7 +25,6 @@
 #include "pc80/mc146818rtc_early.c"
 #include "pc80/serial.c"
 #include "arch/i386/lib/console.c"
-#include "ram/ramtest.c"
 
 #if 0
 static void post_code(uint8_t value) {
@@ -47,7 +38,6 @@ static void post_code(uint8_t value) {
 #endif
 
 #include <cpu/amd/model_fxx_rev.h> 
-#include "northbridge/amd/amdk8/incoherent_ht.c"
 #include "southbridge/amd/amd8111/amd8111_early_smbus.c"
 #include "northbridge/amd/amdk8/raminit.h"
 #include "cpu/amd/model_fxx/apic_timer.c"
@@ -66,43 +56,11 @@ static void post_code(uint8_t value) {
 #include "cpu/x86/bist.h"
 
 #include "northbridge/amd/amdk8/setup_resource_map.c"
+#include "northbridge/amd/amdk8/incoherent_ht.c"
 
 #define SERIAL_DEV PNP_DEV(0x2e, W83627HF_SP1)
 
-static void hard_reset(void)
-{
-        device_t dev;
-	unsigned sblnk = get_sblnk();
-
-        /* Find the device */
-#if HT_CHAIN_END_UNITID_BASE < HT_CHAIN_UNITID_BASE
-	dev = PCI_DEV(node_link_to_bus(0, sblnk), 2 + HT_CHAIN_END_UNITID_BASE - 1, 3);
-#else
-	dev = PCI_DEV(node_link_to_bus(0, sblnk), 4 + HT_CHAIN_UNITID_BASE - 1, 3);
-#endif
-
-        set_bios_reset();
-
-        /* enable cf9 */
-        pci_write_config8(dev, 0x41, 0xf1);
-        /* reset */
-        outb(0x0e, 0x0cf9);
-}
-static void soft_reset(void)
-{
-        device_t dev;
-	unsigned sblnk = get_sblnk();
-
-        /* Find the device */
-#if HT_CHAIN_END_UNITID_BASE < HT_CHAIN_UNITID_BASE
-        dev = PCI_DEV(node_link_to_bus(0, sblnk), 2 + HT_CHAIN_END_UNITID_BASE - 1, 0);
-#else
-        dev = PCI_DEV(node_link_to_bus(0, sblnk), 4 + HT_CHAIN_UNITID_BASE - 1, 0);
-#endif
-
-        set_bios_reset();
-        pci_write_config8(dev, 0x47, 1);
-}
+#include "southbridge/amd/amd8111/amd8111_early_ctrl.c"
 
 static void memreset_setup(void)
 {
@@ -254,7 +212,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
 	unsigned bsp_apicid = 0;
 
         struct mem_controller ctrl[8];
@@ -296,7 +253,7 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 
        	if (needs_reset) {
                	print_info("ht reset -\r\n");
-               	soft_reset();
+		soft_reset();
        	}
 
 	allow_all_aps_stop(bsp_apicid);
@@ -327,6 +284,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	dump_pci_devices();
 #endif
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 
 }
diff --git a/src/mainboard/arima/hdama/Config.lb b/src/mainboard/arima/hdama/Config.lb
index 8b38b0a95b..3f709dff8c 100644
--- a/src/mainboard/arima/hdama/Config.lb
+++ b/src/mainboard/arima/hdama/Config.lb
@@ -45,7 +45,6 @@ arch i386 end
 driver mainboard.o
 if HAVE_MP_TABLE object mptable.o end
 if HAVE_PIRQ_TABLE object irq_tables.o end
-object reset.o
 
 ##
 ## Romcc output
diff --git a/src/mainboard/broadcom/blast/cache_as_ram_auto.c b/src/mainboard/broadcom/blast/cache_as_ram_auto.c
index 37c1fd79dc..3ed4001a9f 100644
--- a/src/mainboard/broadcom/blast/cache_as_ram_auto.c
+++ b/src/mainboard/broadcom/blast/cache_as_ram_auto.c
@@ -190,7 +190,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
 	unsigned bsp_apicid = 0;
 
         struct mem_controller ctrl[8];
@@ -269,6 +268,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	dump_pci_devices();
 #endif
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 
 }
diff --git a/src/mainboard/tyan/s2850/Config.lb b/src/mainboard/tyan/s2850/Config.lb
index 1916fa70c3..e227bc1542 100644
--- a/src/mainboard/tyan/s2850/Config.lb
+++ b/src/mainboard/tyan/s2850/Config.lb
@@ -44,7 +44,6 @@ driver mainboard.o
 
 if HAVE_MP_TABLE object mptable.o end
 if HAVE_PIRQ_TABLE object irq_tables.o end
-object reset.o
 
 if USE_DCACHE_RAM
 
diff --git a/src/mainboard/tyan/s2850/cache_as_ram_auto.c b/src/mainboard/tyan/s2850/cache_as_ram_auto.c
index ba01699c07..5ca8561ea7 100644
--- a/src/mainboard/tyan/s2850/cache_as_ram_auto.c
+++ b/src/mainboard/tyan/s2850/cache_as_ram_auto.c
@@ -47,31 +47,7 @@ static void post_code(uint8_t value) {
 
 #define SERIAL_DEV PNP_DEV(0x2e, W83627HF_SP1)
 
-static void hard_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 0), 0x04, 3);
-
-        set_bios_reset();
-
-        /* enable cf9 */
-        pci_write_config8(dev, 0x41, 0xf1);
-        /* reset */
-        outb(0x0e, 0x0cf9);
-}
-
-static void soft_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 0), 0x04, 0);
-
-        set_bios_reset();
-        pci_write_config8(dev, 0x47, 1);
-}
+#include "southbridge/amd/amd8111/amd8111_early_ctrl.c"
 
 static void memreset_setup(void)
 {
@@ -200,7 +176,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
 
         if (bist == 0) {
 		init_cpus(cpu_init_detectedx);
@@ -235,5 +210,5 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	memreset_setup();
 	sdram_initialize(sizeof(cpu)/sizeof(cpu[0]), cpu);
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 }
diff --git a/src/mainboard/tyan/s2875/Config.lb b/src/mainboard/tyan/s2875/Config.lb
index b1052752fb..13596a2bec 100644
--- a/src/mainboard/tyan/s2875/Config.lb
+++ b/src/mainboard/tyan/s2875/Config.lb
@@ -44,7 +44,6 @@ driver mainboard.o
 
 if HAVE_MP_TABLE object mptable.o end
 if HAVE_PIRQ_TABLE object irq_tables.o end
-object reset.o
 
 if USE_DCACHE_RAM
 
diff --git a/src/mainboard/tyan/s2875/cache_as_ram_auto.c b/src/mainboard/tyan/s2875/cache_as_ram_auto.c
index 43ccb2702e..4126f49448 100644
--- a/src/mainboard/tyan/s2875/cache_as_ram_auto.c
+++ b/src/mainboard/tyan/s2875/cache_as_ram_auto.c
@@ -36,32 +36,7 @@
 
 #define SERIAL_DEV PNP_DEV(0x2e, W83627HF_SP1)
 
-
-static void hard_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 0), 0x04, 3);
-
-        set_bios_reset();
-
-        /* enable cf9 */
-        pci_write_config8(dev, 0x41, 0xf1);
-        /* reset */
-        outb(0x0e, 0x0cf9);
-}
-
-static void soft_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 0), 0x04, 0);
-
-        set_bios_reset();
-        pci_write_config8(dev, 0x47, 1);
-}
+#include "southbridge/amd/amd8111/amd8111_early_ctrl.c"
 
 static void memreset_setup(void)
 {
@@ -198,7 +173,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
 
         if (bist == 0) {
 		init_cpus(cpu_init_detectedx);
@@ -231,6 +205,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	memreset_setup();
 	sdram_initialize(sizeof(cpu)/sizeof(cpu[0]), cpu);
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 
 }
diff --git a/src/mainboard/tyan/s2880/Config.lb b/src/mainboard/tyan/s2880/Config.lb
index 9d7e015b32..d6f907f387 100644
--- a/src/mainboard/tyan/s2880/Config.lb
+++ b/src/mainboard/tyan/s2880/Config.lb
@@ -44,7 +44,6 @@ driver mainboard.o
 
 if HAVE_MP_TABLE object mptable.o end
 if HAVE_PIRQ_TABLE object irq_tables.o end
-object reset.o
 
 if USE_DCACHE_RAM
 
diff --git a/src/mainboard/tyan/s2880/cache_as_ram_auto.c b/src/mainboard/tyan/s2880/cache_as_ram_auto.c
index 3808e83bbf..ab73876184 100644
--- a/src/mainboard/tyan/s2880/cache_as_ram_auto.c
+++ b/src/mainboard/tyan/s2880/cache_as_ram_auto.c
@@ -37,31 +37,7 @@
 
 #define SERIAL_DEV PNP_DEV(0x2e, W83627HF_SP1)
 
-static void hard_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 0), 0x04, 3);
-
-        set_bios_reset();
-
-        /* enable cf9 */
-        pci_write_config8(dev, 0x41, 0xf1);
-        /* reset */
-        outb(0x0e, 0x0cf9);
-}
-
-static void soft_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 0), 0x04, 0);
-
-        set_bios_reset();
-        pci_write_config8(dev, 0x47, 1);
-}
+#include "southbridge/amd/amd8111/amd8111_early_ctrl.c"
 
 static void memreset_setup(void)
 {
@@ -197,7 +173,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
 
         if (bist == 0) {
 		init_cpus(cpu_init_detectedx);
@@ -232,5 +207,5 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	memreset_setup();
 	sdram_initialize(sizeof(cpu)/sizeof(cpu[0]), cpu);
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 }
diff --git a/src/mainboard/tyan/s2881/Config.lb b/src/mainboard/tyan/s2881/Config.lb
index 3b0cd33dd7..598f9b4a36 100644
--- a/src/mainboard/tyan/s2881/Config.lb
+++ b/src/mainboard/tyan/s2881/Config.lb
@@ -44,7 +44,6 @@ driver mainboard.o
 object get_bus_conf.o
 if HAVE_MP_TABLE object mptable.o end
 if HAVE_PIRQ_TABLE object irq_tables.o end
-object reset.o
 if USE_DCACHE_RAM
 
 if CONFIG_USE_INIT
diff --git a/src/mainboard/tyan/s2881/cache_as_ram_auto.c b/src/mainboard/tyan/s2881/cache_as_ram_auto.c
index 77dd978856..67a32140ad 100644
--- a/src/mainboard/tyan/s2881/cache_as_ram_auto.c
+++ b/src/mainboard/tyan/s2881/cache_as_ram_auto.c
@@ -42,31 +42,7 @@
 
 #define SERIAL_DEV PNP_DEV(0x2e, W83627HF_SP1)
 
-static void hard_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 2), 0x04, 3);
-
-        set_bios_reset();
-
-        /* enable cf9 */
-        pci_write_config8(dev, 0x41, 0xf1);
-        /* reset */
-        outb(0x0e, 0x0cf9);
-}
-
-static void soft_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 2), 0x04, 0);
-
-        set_bios_reset();
-        pci_write_config8(dev, 0x47, 1);
-}
+#include "southbridge/amd/amd8111/amd8111_early_ctrl.c"
 
 static void memreset_setup(void)
 {
@@ -185,7 +161,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
         unsigned bsp_apicid = 0;
 
         struct mem_controller ctrl[8];
@@ -234,5 +209,5 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
         sdram_initialize(nodes, ctrl);
 
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 }
diff --git a/src/mainboard/tyan/s2882/Config.lb b/src/mainboard/tyan/s2882/Config.lb
index 8b173bd6d6..42c193be58 100644
--- a/src/mainboard/tyan/s2882/Config.lb
+++ b/src/mainboard/tyan/s2882/Config.lb
@@ -44,7 +44,6 @@ driver mainboard.o
 
 if HAVE_MP_TABLE object mptable.o end
 if HAVE_PIRQ_TABLE object irq_tables.o end
-object reset.o
 
 if USE_DCACHE_RAM
 
diff --git a/src/mainboard/tyan/s2882/cache_as_ram_auto.c b/src/mainboard/tyan/s2882/cache_as_ram_auto.c
index 967f068d51..c31181a8e2 100644
--- a/src/mainboard/tyan/s2882/cache_as_ram_auto.c
+++ b/src/mainboard/tyan/s2882/cache_as_ram_auto.c
@@ -36,31 +36,7 @@
 
 #define SERIAL_DEV PNP_DEV(0x2e, W83627HF_SP1)
 
-static void hard_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 0), 0x04, 3);
-
-        set_bios_reset();
-
-        /* enable cf9 */
-        pci_write_config8(dev, 0x41, 0xf1);
-        /* reset */
-        outb(0x0e, 0x0cf9);
-}
-
-static void soft_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 0), 0x04, 0);
-
-        set_bios_reset();
-        pci_write_config8(dev, 0x47, 1);
-}
+#include "southbridge/amd/amd8111/amd8111_early_ctrl.c"
 
 static void memreset_setup(void)
 {
@@ -200,7 +176,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
 
         if (bist == 0) {
 		init_cpus(cpu_init_detectedx);
@@ -235,6 +210,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	memreset_setup();
 	sdram_initialize(sizeof(cpu)/sizeof(cpu[0]), cpu);
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 
 }
diff --git a/src/mainboard/tyan/s2885/Config.lb b/src/mainboard/tyan/s2885/Config.lb
index 37b0d90522..2a035a35ab 100644
--- a/src/mainboard/tyan/s2885/Config.lb
+++ b/src/mainboard/tyan/s2885/Config.lb
@@ -44,7 +44,6 @@ driver mainboard.o
 object get_bus_conf.o
 if HAVE_MP_TABLE object mptable.o end
 if HAVE_PIRQ_TABLE object irq_tables.o end
-object reset.o
 
 if USE_DCACHE_RAM
 
diff --git a/src/mainboard/tyan/s2885/cache_as_ram_auto.c b/src/mainboard/tyan/s2885/cache_as_ram_auto.c
index 7b2ccfaa14..0c5792565c 100644
--- a/src/mainboard/tyan/s2885/cache_as_ram_auto.c
+++ b/src/mainboard/tyan/s2885/cache_as_ram_auto.c
@@ -36,31 +36,7 @@
 
 #define SERIAL_DEV PNP_DEV(0x2e, W83627HF_SP1)
 
-static void hard_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 2), 0x04, 3);
-
-        set_bios_reset();
-
-        /* enable cf9 */
-        pci_write_config8(dev, 0x41, 0xf1);
-        /* reset */
-        outb(0x0e, 0x0cf9);
-}
-
-static void soft_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 2), 0x04, 0);
-
-        set_bios_reset();
-        pci_write_config8(dev, 0x47, 1);
-}
+#include "southbridge/amd/amd8111/amd8111_early_ctrl.c"
 
 static void memreset_setup(void)
 {
@@ -185,7 +161,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
         unsigned bsp_apicid = 0;
 
         struct mem_controller ctrl[8];
@@ -234,6 +209,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
         memreset_setup();
         sdram_initialize(nodes, ctrl);
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 
 }
diff --git a/src/mainboard/tyan/s2891/Config.lb b/src/mainboard/tyan/s2891/Config.lb
index 9c460066c9..67c7b3c8a9 100644
--- a/src/mainboard/tyan/s2891/Config.lb
+++ b/src/mainboard/tyan/s2891/Config.lb
@@ -237,7 +237,7 @@ chip northbridge/amd/amdk8/root_complex
                                         	                io 0x60 = 0x2f8
                                                 	        irq 0x70 = 3
 	                                                end
-        	                                        device pnp 2e.5 off #  Keyboard
+        	                                        device pnp 2e.5 on #  Keyboard
                 	                                        io 0x60 = 0x60
                         	                                io 0x62 = 0x64
                                 	                        irq 0x70 = 1
@@ -314,7 +314,7 @@ chip northbridge/amd/amdk8/root_complex
 					#	chip drivers/ati/ragexl
                                                 chip drivers/pci/onboard
                                                         device pci 7.0 on end
-                                        #                register "rom_address" = "0xfff80000" #for 512K
+                                               #         register "rom_address" = "0xfff80000" #for 512K
 							register "rom_address" = "0xfff00000" #for 1M
                                                 end
 					end
diff --git a/src/mainboard/tyan/s2891/Options.lb b/src/mainboard/tyan/s2891/Options.lb
index e87a4df341..9018364663 100644
--- a/src/mainboard/tyan/s2891/Options.lb
+++ b/src/mainboard/tyan/s2891/Options.lb
@@ -54,6 +54,7 @@ uses CONFIG_CHIP_NAME
 uses CONFIG_CONSOLE_VGA
 uses CONFIG_PCI_ROM_RUN
 uses K8_HW_MEM_HOLE_SIZEK
+uses K8_HT_FREQ_1G_SUPPORT
 
 uses USE_DCACHE_RAM
 uses DCACHE_RAM_BASE
@@ -66,6 +67,13 @@ uses LIFT_BSP_APIC_ID
 
 uses CONFIG_PCI_64BIT_PREF_MEM
 
+uses HT_CHAIN_UNITID_BASE
+uses HT_CHAIN_END_UNITID_BASE
+uses K8_SB_HT_CHAIN_ON_BUS0
+uses SB_HT_CHAIN_UNITID_OFFSET_ONLY
+
+uses CONFIG_LB_MEM_TOPK
+
 ## ROM_SIZE is the size of boot ROM that this board will use.
 #512K bytes 
 #default ROM_SIZE=524288
@@ -131,6 +139,22 @@ default CONFIG_LOGICAL_CPUS=1
 #1G memory hole
 default K8_HW_MEM_HOLE_SIZEK=0x100000
 
+#Opteron K8 1G HT Support
+default K8_HT_FREQ_1G_SUPPORT=1
+
+##HT Unit ID offset, default is 1, the typical one
+default HT_CHAIN_UNITID_BASE=0x0
+
+##real SB Unit ID, default is 0x20, mean dont touch it at last
+#default HT_CHAIN_END_UNITID_BASE=0x0
+
+#make the SB HT chain on bus 0, default is not (0)
+default K8_SB_HT_CHAIN_ON_BUS0=2
+
+##only offset for SB chain?, default is yes(1)
+default SB_HT_CHAIN_UNITID_OFFSET_ONLY=0
+
+
 #BTEXT Console
 #default CONFIG_CONSOLE_BTEXT=1
 
@@ -144,9 +168,9 @@ default CONFIG_PCI_ROM_RUN=1
 default USE_DCACHE_RAM=1
 default DCACHE_RAM_BASE=0xcf000
 default DCACHE_RAM_SIZE=0x1000
-default CONFIG_USE_INIT=1
+default CONFIG_USE_INIT=0
 
-default ENABLE_APIC_EXT_ID=1
+default ENABLE_APIC_EXT_ID=0
 default APIC_ID_OFFSET=0x10
 default LIFT_BSP_APIC_ID=0
 
diff --git a/src/mainboard/tyan/s2891/cache_as_ram_auto.c b/src/mainboard/tyan/s2891/cache_as_ram_auto.c
index 61edc58cd2..1bc29d9304 100644
--- a/src/mainboard/tyan/s2891/cache_as_ram_auto.c
+++ b/src/mainboard/tyan/s2891/cache_as_ram_auto.c
@@ -42,25 +42,6 @@
 
 #define SERIAL_DEV PNP_DEV(0x2e, W83627HF_SP1)
 
-static void hard_reset(void)
-{
-        set_bios_reset();
-
-        /* full reset */
-	outb(0x0a, 0x0cf9);
-        outb(0x0e, 0x0cf9);
-}
-
-static void soft_reset(void)
-{
-        set_bios_reset();
-#if 1
-        /* link reset */
-	outb(0x02, 0x0cf9);
-        outb(0x06, 0x0cf9);
-#endif
-}
-
 static void memreset_setup(void)
 {
 }
@@ -197,7 +178,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
         unsigned bsp_apicid = 0;
 
         struct mem_controller ctrl[8];
@@ -246,5 +226,5 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	memreset_setup();
 	sdram_initialize(nodes, ctrl);
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 }
diff --git a/src/mainboard/tyan/s2891/get_bus_conf.c b/src/mainboard/tyan/s2891/get_bus_conf.c
index 92fefa6a0c..3462a5f557 100644
--- a/src/mainboard/tyan/s2891/get_bus_conf.c
+++ b/src/mainboard/tyan/s2891/get_bus_conf.c
@@ -152,9 +152,10 @@ void get_bus_conf(void)
                 else {
                         printk_debug("ERROR - could not find PCI 1:%02x.0, using defaults\n", sbdn+ 0x0e);
 
-                        bus_8131_0 = bus_ck804_5+1;
                 }
 
+		bus_8131_0 = (pci1234[1] >> 16) & 0xff;
+
                 /* 8131-1 */
                 dev = dev_find_slot(bus_8131_0, PCI_DEVFN(sbdn3,0));
                 if (dev) {
diff --git a/src/mainboard/tyan/s2891/resourcemap.c b/src/mainboard/tyan/s2891/resourcemap.c
index c55d95bb99..625e19f11c 100644
--- a/src/mainboard/tyan/s2891/resourcemap.c
+++ b/src/mainboard/tyan/s2891/resourcemap.c
@@ -186,7 +186,7 @@ static void setup_s2891_resource_map(void)
 		 *	   This field defines the end of PCI I/O region n
 		 * [31:25] Reserved
 		 */
-		PCI_ADDR(0, 0x18, 1, 0xC4), 0xFE000FC8, 0x01fff000,
+//		PCI_ADDR(0, 0x18, 1, 0xC4), 0xFE000FC8, 0x01fff000,
 		PCI_ADDR(0, 0x18, 1, 0xCC), 0xFE000FC8, 0x00000000, 
 		PCI_ADDR(0, 0x18, 1, 0xD4), 0xFE000FC8, 0x00000000,
 		PCI_ADDR(0, 0x18, 1, 0xDC), 0xFE000FC8, 0x00000000,
@@ -216,7 +216,7 @@ static void setup_s2891_resource_map(void)
 		 *	   This field defines the start of PCI I/O region n 
 		 * [31:25] Reserved
 		 */
-		PCI_ADDR(0, 0x18, 1, 0xC0), 0xFE000FCC, 0x00000033,
+//		PCI_ADDR(0, 0x18, 1, 0xC0), 0xFE000FCC, 0x00000033,
 		PCI_ADDR(0, 0x18, 1, 0xC8), 0xFE000FCC, 0x00000000,
 		PCI_ADDR(0, 0x18, 1, 0xD0), 0xFE000FCC, 0x00000000,
 		PCI_ADDR(0, 0x18, 1, 0xD8), 0xFE000FCC, 0x00000000,
diff --git a/src/mainboard/tyan/s2892/Config.lb b/src/mainboard/tyan/s2892/Config.lb
index dcb0c45abf..6fa8e5d44a 100644
--- a/src/mainboard/tyan/s2892/Config.lb
+++ b/src/mainboard/tyan/s2892/Config.lb
@@ -313,7 +313,8 @@ chip northbridge/amd/amdk8/root_complex
 					#	chip drivers/ati/ragexl
 						chip drivers/pci/onboard
 							device pci 6.0 on end
-							register "rom_address" = "0xfff80000"
+                                                #       register "rom_address" = "0xfff80000" #for 512K
+                                                        register "rom_address" = "0xfff00000" #for 1M
 						end
                                                 chip drivers/pci/onboard
                                                         device pci 8.0 on end
diff --git a/src/mainboard/tyan/s2892/Options.lb b/src/mainboard/tyan/s2892/Options.lb
index a8bac76536..7000126a5d 100644
--- a/src/mainboard/tyan/s2892/Options.lb
+++ b/src/mainboard/tyan/s2892/Options.lb
@@ -54,18 +54,24 @@ uses CONFIG_CHIP_NAME
 uses CONFIG_CONSOLE_VGA
 uses CONFIG_PCI_ROM_RUN
 uses K8_HW_MEM_HOLE_SIZEK
+uses K8_HT_FREQ_1G_SUPPORT
 
 uses USE_DCACHE_RAM
 uses DCACHE_RAM_BASE
 uses DCACHE_RAM_SIZE
 uses CONFIG_USE_INIT
 
+uses HT_CHAIN_UNITID_BASE
+uses HT_CHAIN_END_UNITID_BASE
+uses K8_SB_HT_CHAIN_ON_BUS0
+uses SB_HT_CHAIN_UNITID_OFFSET_ONLY
+
 ## ROM_SIZE is the size of boot ROM that this board will use.
 #512K bytes 
-default ROM_SIZE=524288
+#default ROM_SIZE=524288
 
 #1M bytes
-#default ROM_SIZE=1048576
+default ROM_SIZE=1048576
 
 
 ##
@@ -125,6 +131,21 @@ default CONFIG_LOGICAL_CPUS=1
 #1G memory hole
 default K8_HW_MEM_HOLE_SIZEK=0x100000
 
+#Opteron K8 1G HT Support
+default K8_HT_FREQ_1G_SUPPORT=1
+
+##HT Unit ID offset, default is 1, the typical one
+default HT_CHAIN_UNITID_BASE=0x0
+
+##real SB Unit ID, default is 0x20, mean dont touch it at last
+#default HT_CHAIN_END_UNITID_BASE=0x0
+
+#make the SB HT chain on bus 0, default is not (0)
+default K8_SB_HT_CHAIN_ON_BUS0=2
+
+##only offset for SB chain?, default is yes(1)
+default SB_HT_CHAIN_UNITID_OFFSET_ONLY=0
+
 #BTEXT Console
 #default CONFIG_CONSOLE_BTEXT=1
 
@@ -138,8 +159,7 @@ default CONFIG_PCI_ROM_RUN=1
 default USE_DCACHE_RAM=1
 default DCACHE_RAM_BASE=0xcf000
 default DCACHE_RAM_SIZE=0x1000
-default CONFIG_USE_INIT=1
-
+default CONFIG_USE_INIT=0
 
 ##
 ## Build code to setup a generic IOAPIC
diff --git a/src/mainboard/tyan/s2892/cache_as_ram_auto.c b/src/mainboard/tyan/s2892/cache_as_ram_auto.c
index 3bb7fb3d83..1936563518 100644
--- a/src/mainboard/tyan/s2892/cache_as_ram_auto.c
+++ b/src/mainboard/tyan/s2892/cache_as_ram_auto.c
@@ -36,25 +36,6 @@
 
 #define SERIAL_DEV PNP_DEV(0x2e, W83627HF_SP1)
 
-static void hard_reset(void)
-{
-        set_bios_reset();
-
-        /* full reset */
-	outb(0x0a, 0x0cf9);
-        outb(0x0e, 0x0cf9);
-}
-
-static void soft_reset(void)
-{
-        set_bios_reset();
-#if 1
-        /* link reset */
-	outb(0x02, 0x0cf9);
-        outb(0x06, 0x0cf9);
-#endif
-}
-
 static void memreset_setup(void)
 {
 }
@@ -210,7 +191,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
 
         if (bist == 0) {
 		init_cpus(cpu_init_detectedx);
@@ -246,5 +226,5 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	memreset_setup();
 	sdram_initialize(sizeof(cpu)/sizeof(cpu[0]), cpu);
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 }
diff --git a/src/mainboard/tyan/s2892/get_bus_conf.c b/src/mainboard/tyan/s2892/get_bus_conf.c
index 92fefa6a0c..6086aa5ff5 100644
--- a/src/mainboard/tyan/s2892/get_bus_conf.c
+++ b/src/mainboard/tyan/s2892/get_bus_conf.c
@@ -152,9 +152,9 @@ void get_bus_conf(void)
                 else {
                         printk_debug("ERROR - could not find PCI 1:%02x.0, using defaults\n", sbdn+ 0x0e);
 
-                        bus_8131_0 = bus_ck804_5+1;
                 }
 
+		bus_8131_0 = (pci1234[1] >> 16) & 0xff;
                 /* 8131-1 */
                 dev = dev_find_slot(bus_8131_0, PCI_DEVFN(sbdn3,0));
                 if (dev) {
diff --git a/src/mainboard/tyan/s2895/Options.lb b/src/mainboard/tyan/s2895/Options.lb
index 49b0e0b0de..c3f8ba3f8c 100644
--- a/src/mainboard/tyan/s2895/Options.lb
+++ b/src/mainboard/tyan/s2895/Options.lb
@@ -64,12 +64,17 @@ uses ENABLE_APIC_EXT_ID
 uses APIC_ID_OFFSET
 uses LIFT_BSP_APIC_ID
 
+uses HT_CHAIN_UNITID_BASE
+uses HT_CHAIN_END_UNITID_BASE
+uses K8_SB_HT_CHAIN_ON_BUS0
+uses SB_HT_CHAIN_UNITID_OFFSET_ONLY
+
 ## ROM_SIZE is the size of boot ROM that this board will use.
 #512K bytes 
-default ROM_SIZE=524288
+#default ROM_SIZE=524288
 
 #1M bytes
-#default ROM_SIZE=1048576
+default ROM_SIZE=1048576
 
 ##
 ## FALLBACK_SIZE is the amount of the ROM the complete fallback image will use
@@ -134,6 +139,18 @@ default K8_HW_MEM_HOLE_SIZEK=0x100000
 #Opteron K8 1G HT Support
 default K8_HT_FREQ_1G_SUPPORT=1
 
+##HT Unit ID offset, default is 1, the typical one
+default HT_CHAIN_UNITID_BASE=0x0
+
+##real SB Unit ID, default is 0x20, mean dont touch it at last
+#default HT_CHAIN_END_UNITID_BASE=0x0
+
+#make the SB HT chain on bus 0, default is not (0)
+default K8_SB_HT_CHAIN_ON_BUS0=2
+
+##only offset for SB chain?, default is yes(1)
+default SB_HT_CHAIN_UNITID_OFFSET_ONLY=0
+
 #VGA
 default CONFIG_CONSOLE_VGA=1
 default CONFIG_PCI_ROM_RUN=1
@@ -144,7 +161,7 @@ default CONFIG_PCI_ROM_RUN=1
 default USE_DCACHE_RAM=1
 default DCACHE_RAM_BASE=0xcf000
 default DCACHE_RAM_SIZE=0x1000
-default CONFIG_USE_INIT=1
+default CONFIG_USE_INIT=0
 
 default ENABLE_APIC_EXT_ID=1
 default APIC_ID_OFFSET=0x10
diff --git a/src/mainboard/tyan/s2895/cache_as_ram_auto.c b/src/mainboard/tyan/s2895/cache_as_ram_auto.c
index 1ae0c9e5d1..00bf1a6c23 100644
--- a/src/mainboard/tyan/s2895/cache_as_ram_auto.c
+++ b/src/mainboard/tyan/s2895/cache_as_ram_auto.c
@@ -50,25 +50,6 @@
 
 #define SERIAL_DEV PNP_DEV(0x2e, LPC47B397_SP1)
 
-static void hard_reset(void)
-{
-        set_bios_reset();
-
-        /* full reset */
-	outb(0x0a, 0x0cf9);
-        outb(0x0e, 0x0cf9);
-}
-
-static void soft_reset(void)
-{
-        set_bios_reset();
-#if 1
-        /* link reset */
-	outb(0x02, 0x0cf9);
-        outb(0x06, 0x0cf9);
-#endif
-}
-
 static void memreset_setup(void)
 {
 }
@@ -239,7 +220,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
         unsigned bsp_apicid = 0;
 
         struct mem_controller ctrl[8];
@@ -288,5 +268,5 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	memreset_setup();
 	sdram_initialize(nodes, ctrl);
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 }
diff --git a/src/mainboard/tyan/s4880/Config.lb b/src/mainboard/tyan/s4880/Config.lb
index 74c5542f87..54fc432d72 100644
--- a/src/mainboard/tyan/s4880/Config.lb
+++ b/src/mainboard/tyan/s4880/Config.lb
@@ -43,7 +43,6 @@ arch i386 end
 driver mainboard.o
 if HAVE_MP_TABLE object mptable.o end
 if HAVE_PIRQ_TABLE object irq_tables.o end
-object reset.o
 if USE_DCACHE_RAM
 
 	if CONFIG_USE_INIT
diff --git a/src/mainboard/tyan/s4880/cache_as_ram_auto.c b/src/mainboard/tyan/s4880/cache_as_ram_auto.c
index 0165a1f8c2..1a16eff09f 100644
--- a/src/mainboard/tyan/s4880/cache_as_ram_auto.c
+++ b/src/mainboard/tyan/s4880/cache_as_ram_auto.c
@@ -36,31 +36,7 @@
 
 #define SERIAL_DEV PNP_DEV(0x2e, W83627HF_SP1)
 
-static void hard_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 2), 0x04, 3);
-
-        set_bios_reset();
-
-        /* enable cf9 */
-        pci_write_config8(dev, 0x41, 0xf1);
-        /* reset */
-        outb(0x0e, 0x0cf9);
-}
-
-static void soft_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 2), 0x04, 0);
-
-        set_bios_reset();
-        pci_write_config8(dev, 0x47, 1);
-}
+#include "southbridge/amd/amd8111/amd8111_early_ctrl.c"
 
 static void memreset_setup(void)
 {
@@ -247,7 +223,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
 
         if (bist == 0) {
 		init_cpus(cpu_init_detectedx);
@@ -281,5 +256,5 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	memreset_setup();
 	sdram_initialize(sizeof(cpu)/sizeof(cpu[0]), cpu);
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 }
diff --git a/src/mainboard/tyan/s4882/Config.lb b/src/mainboard/tyan/s4882/Config.lb
index 1913120393..05e898814e 100644
--- a/src/mainboard/tyan/s4882/Config.lb
+++ b/src/mainboard/tyan/s4882/Config.lb
@@ -43,7 +43,6 @@ arch i386 end
 driver mainboard.o
 if HAVE_MP_TABLE object mptable.o end
 if HAVE_PIRQ_TABLE object irq_tables.o end
-object reset.o
 if USE_DCACHE_RAM
 
 	if CONFIG_USE_INIT
diff --git a/src/mainboard/tyan/s4882/cache_as_ram_auto.c b/src/mainboard/tyan/s4882/cache_as_ram_auto.c
index 6f1ac8d098..2b463d4d27 100644
--- a/src/mainboard/tyan/s4882/cache_as_ram_auto.c
+++ b/src/mainboard/tyan/s4882/cache_as_ram_auto.c
@@ -36,31 +36,8 @@
 
 #define SERIAL_DEV PNP_DEV(0x2e, W83627HF_SP1)
 
-static void hard_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 1), 0x04, 3);
-
-        set_bios_reset();
-
-        /* enable cf9 */
-        pci_write_config8(dev, 0x41, 0xf1);
-        /* reset */
-        outb(0x0e, 0x0cf9);
-}
 
-static void soft_reset(void)
-{
-        device_t dev;
-
-        /* Find the device */
-        dev = PCI_DEV(node_link_to_bus(0, 1), 0x04, 0);
-
-        set_bios_reset();
-        pci_write_config8(dev, 0x47, 1);
-}
+#include "southbridge/amd/amd8111/amd8111_early_ctrl.c"
 
 static void memreset_setup(void)
 {
@@ -222,7 +199,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	};
 
         int needs_reset;
-	unsigned cpu_reset = 0;
         unsigned bsp_apicid = 0;
 
         struct mem_controller ctrl[8];
@@ -271,6 +247,6 @@ void real_main(unsigned long bist, unsigned long cpu_init_detectedx)
 	memreset_setup();
 	sdram_initialize(nodes, ctrl);
 
-	post_cache_as_ram(cpu_reset);
+	post_cache_as_ram();
 
 }
diff --git a/src/northbridge/amd/amdk8/coherent_ht.c b/src/northbridge/amd/amdk8/coherent_ht.c
index 7d43f3c97d..d29831b74b 100644
--- a/src/northbridge/amd/amdk8/coherent_ht.c
+++ b/src/northbridge/amd/amdk8/coherent_ht.c
@@ -1,3 +1,10 @@
+#if USE_DCACHE_RAM
+
+#include "coherent_ht_car.c"
+
+#else
+
+
 /* coherent hypertransport initialization for AMD64 
  * 
  * written by Stefan Reinauer <stepan@openbios.org>
@@ -1824,3 +1831,5 @@ static int setup_coherent_ht_domain(void)
 	result.needs_reset = optimize_link_read_pointers(result.nodes, result.needs_reset);
 	return result.needs_reset;
 }
+
+#endif
diff --git a/src/northbridge/amd/amdk8/coherent_ht_car.c b/src/northbridge/amd/amdk8/coherent_ht_car.c
new file mode 100644
index 0000000000..5b9a87ac63
--- /dev/null
+++ b/src/northbridge/amd/amdk8/coherent_ht_car.c
@@ -0,0 +1,1826 @@
+/* coherent hypertransport initialization for AMD64 
+ * 
+ * written by Stefan Reinauer <stepan@openbios.org>
+ * (c) 2003-2004 by SuSE Linux AG
+ *
+ * (c) 2004 Tyan Computer
+ *  2004.12 yhlu added support to create routing table dynamically.
+ *          it also support 8 ways too. (8 ways ladder or 8 ways crossbar)
+ *
+ * This code is licensed under GPL.
+ */
+
+/*
+ * This algorithm assumes a grid configuration as follows:
+ *
+ * nodes :  1    2    4    6    8
+ * org.  :  1x1  2x1  2x2  2x3  2x4
+ Ladder:
+			CPU7-------------CPU6 
+			|                |    
+			|                |
+			|                | 
+			|                |     
+			|                |        
+			|                |    
+			CPU5-------------CPU4                    
+			|                |    
+			|                |   
+			|                |  
+			|                |    
+			|                |         
+			|                |       
+			CPU3-------------CPU2              
+			|                |    
+			|                | 
+			|                |
+			|                |
+			|                | 
+			|                |    
+			CPU1-------------CPU0    
+ CROSS_BAR_47_56:
+			CPU7-------------CPU6 
+			|  \____    ___/ |    
+			|       \  /     |
+			|        \/      | 
+			|        /\      |     
+			|       /  \     |        
+			|  ____/    \___ |    
+			CPU5             CPU4                    
+			|                |    
+			|                |   
+			|                |  
+			|                |    
+			|                |         
+			|                |       
+			CPU3-------------CPU2              
+			|                |    
+			|                | 
+			|                |
+			|                |
+			|                | 
+			|                |    
+			CPU1-------------CPU0     	
+ */
+
+#include <device/pci_def.h>
+#include <device/pci_ids.h>
+#include <device/hypertransport_def.h>
+#include "arch/romcc_io.h"
+
+#include "amdk8.h"
+
+#define enable_bsp_routing()	enable_routing(0)
+
+#define NODE_HT(x) PCI_DEV(0,24+x,0)
+#define NODE_MP(x) PCI_DEV(0,24+x,1)
+#define NODE_MC(x) PCI_DEV(0,24+x,3)
+
+#define DEFAULT 0x00010101	/* default row entry */
+
+typedef uint8_t u8;
+typedef uint32_t u32;
+
+#ifndef CROSS_BAR_47_56
+	#define CROSS_BAR_47_56 0
+#endif
+
+#ifndef TRY_HIGH_FIRST
+	#define TRY_HIGH_FIRST 0
+#endif
+
+#ifndef K8_HT_FREQ_1G_SUPPORT
+	#define K8_HT_FREQ_1G_SUPPORT 0
+#endif
+
+#ifndef K8_HT_CHECK_PENDING_LINK
+	#if CONFIG_MAX_PHYSICAL_CPUS >= 4
+		#define K8_HT_CHECK_PENDING_LINK 1
+	#else
+		#define K8_HT_CHECK_PENDING_LINK 0
+	#endif
+#endif
+
+#ifndef CONFIG_MAX_PHYSICAL_CPUS_4_BUT_MORE_INSTALLED
+	#define CONFIG_MAX_PHYSICAL_CPUS_4_BUT_MORE_INSTALLED 0
+#endif
+
+#ifndef ENABLE_APIC_EXT_ID
+        #define ENABLE_APIC_EXT_ID 0
+#endif
+
+
+static inline void print_linkn (const char *strval, uint8_t byteval) 
+{
+#if 1
+#if CONFIG_USE_INIT
+	printk_debug("%s%02x\r\n", strval, byteval); 
+#else
+	print_debug(strval); print_debug_hex8(byteval); print_debug("\r\n");
+#endif
+#endif
+}
+
+static void disable_probes(void)
+{
+	/* disable read/write/fill probes for uniprocessor setup
+	 * they don't make sense if only one cpu is available
+	 */
+
+	/* Hypetransport Transaction Control Register 
+	 * F0:0x68
+	 * [ 0: 0] Disable read byte probe
+	 *         0 = Probes issues
+	 *         1 = Probes not issued
+	 * [ 1: 1] Disable Read Doubleword probe
+	 *         0 = Probes issued
+	 *         1 = Probes not issued
+	 * [ 2: 2] Disable write byte probes
+	 *         0 = Probes issued
+	 *         1 = Probes not issued
+	 * [ 3: 3] Disable Write Doubleword Probes
+	 *         0 = Probes issued
+	 *         1 = Probes not issued.
+	 * [10:10] Disable Fill Probe
+	 *         0 = Probes issued for cache fills
+	 *         1 = Probes not issued for cache fills.
+	 */
+
+	u32 val;
+
+	print_spew("Disabling read/write/fill probes for UP... ");
+
+	val=pci_read_config32(NODE_HT(0), HT_TRANSACTION_CONTROL);
+	val |= HTTC_DIS_FILL_P | HTTC_DIS_RMT_MEM_C | HTTC_DIS_P_MEM_C |
+		HTTC_DIS_MTS | HTTC_DIS_WR_DW_P | HTTC_DIS_WR_B_P | 
+		HTTC_DIS_RD_DW_P | HTTC_DIS_RD_B_P;
+	pci_write_config32(NODE_HT(0), HT_TRANSACTION_CONTROL, val);
+
+	print_spew("done.\r\n");
+
+}
+
+static void enable_apic_ext_id(u8 node) 
+{
+#if ENABLE_APIC_EXT_ID==1
+#warning "FIXME Is the right place to enable apic ext id here?"
+
+      u32 val;
+
+        val = pci_read_config32(NODE_HT(node), 0x68);
+        val |= (HTTC_APIC_EXT_SPUR | HTTC_APIC_EXT_ID | HTTC_APIC_EXT_BRD_CST);
+        pci_write_config32(NODE_HT(node), 0x68, val);
+#endif
+}
+
+static void enable_routing(u8 node)
+{
+	u32 val;
+
+	/* HT Initialization Control Register
+	 * F0:0x6C
+	 * [ 0: 0] Routing Table Disable
+	 *         0 = Packets are routed according to routing tables
+	 *         1 = Packets are routed according to the default link field
+	 * [ 1: 1] Request Disable (BSP should clear this)
+	 *         0 = Request packets may be generated
+	 *         1 = Request packets may not be generated.
+	 * [ 3: 2] Default Link (Read-only)
+	 *         00 = LDT0
+	 *         01 = LDT1
+	 *         10 = LDT2
+	 *         11 = CPU on same node
+	 * [ 4: 4] Cold Reset
+	 *         - Scratch bit cleared by a cold reset
+	 * [ 5: 5] BIOS Reset Detect
+	 *         - Scratch bit cleared by a cold reset
+	 * [ 6: 6] INIT Detect
+	 *         - Scratch bit cleared by a warm or cold reset not by an INIT
+	 *
+	 */
+
+	/* Enable routing table */
+	print_spew("Enabling routing table for node ");
+	print_spew_hex8(node);
+
+	val=pci_read_config32(NODE_HT(node), 0x6c);
+	val &= ~((1<<1)|(1<<0));
+	pci_write_config32(NODE_HT(node), 0x6c, val);
+
+	print_spew(" done.\r\n");
+}
+
+static void fill_row(u8 node, u8 row, u32 value)
+{
+	pci_write_config32(NODE_HT(node), 0x40+(row<<2), value);
+}
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 1
+static u8 link_to_register(int ldt)
+{
+	/*
+	 * [ 0: 3] Request Route
+	 *     [0] Route to this node
+	 *     [1] Route to Link 0
+	 *     [2] Route to Link 1
+	 *     [3] Route to Link 2
+	 */
+
+	if (ldt&0x08) return 0x40;
+	if (ldt&0x04) return 0x20;
+	if (ldt&0x02) return 0x00;
+	
+	/* we should never get here */
+	print_spew("Unknown Link\n");
+	return 0;
+}
+
+static u32 get_row(u8 node, u8 row)
+{
+	return pci_read_config32(NODE_HT(node), 0x40+(row<<2));
+}
+
+static int link_connection(u8 src, u8 dest)
+{
+	return get_row(src, dest) & 0x0f;
+}
+
+static void rename_temp_node(u8 node)
+{
+	uint32_t val;
+
+	print_spew("Renaming current temporary node to ");
+	print_spew_hex8(node);
+
+	val=pci_read_config32(NODE_HT(7), 0x60);
+	val &= (~7);  /* clear low bits. */
+        val |= node;   /* new node        */
+	pci_write_config32(NODE_HT(7), 0x60, val);
+
+	print_spew(" done.\r\n");
+}
+
+static int verify_connection(u8 dest)
+{
+	/* See if we have a valid connection to dest */
+	u32 val;
+	
+	/* Verify that the coherent hypertransport link is
+	 * established and actually working by reading the
+	 * remode node's vendor/device id
+	 */
+        val = pci_read_config32(NODE_HT(dest),0);
+	if(val != 0x11001022)
+		return 0;
+
+	return 1;
+}
+
+static uint16_t read_freq_cap(device_t dev, uint8_t pos)
+{
+	/* Handle bugs in valid hypertransport frequency reporting */
+	uint16_t freq_cap;
+	uint32_t id;
+
+	freq_cap = pci_read_config16(dev, pos);
+	freq_cap &= ~(1 << HT_FREQ_VENDOR); /* Ignore Vendor HT frequencies */
+
+#if K8_HT_FREQ_1G_SUPPORT == 1
+	if (!is_cpu_pre_e0())
+	{
+		return freq_cap;
+	}
+#endif
+
+	id = pci_read_config32(dev, 0);
+
+	/* AMD K8 Unsupported 1Ghz? */
+	if (id == (PCI_VENDOR_ID_AMD | (0x1100 << 16))) {
+		freq_cap &= ~(1 << HT_FREQ_1000Mhz);
+	}
+	
+	return freq_cap;
+}
+
+static int optimize_connection(device_t node1, uint8_t link1, device_t node2, uint8_t link2)
+{
+	static const uint8_t link_width_to_pow2[]= { 3, 4, 0, 5, 1, 2, 0, 0 };
+	static const uint8_t pow2_to_link_width[] = { 0x7, 4, 5, 0, 1, 3 };
+	uint16_t freq_cap1, freq_cap2;
+	uint8_t width_cap1, width_cap2, width, old_width, ln_width1, ln_width2;
+	uint8_t freq, old_freq;
+	int needs_reset;
+	/* Set link width and frequency */
+
+	/* Initially assume everything is already optimized and I don't need a reset */
+	needs_reset = 0;
+
+	/* Get the frequency capabilities */
+	freq_cap1 = read_freq_cap(node1, link1 + PCI_HT_CAP_HOST_FREQ_CAP);
+	freq_cap2 = read_freq_cap(node2, link2 + PCI_HT_CAP_HOST_FREQ_CAP);
+
+	/* Calculate the highest possible frequency */
+	freq = log2(freq_cap1 & freq_cap2);
+
+	/* See if I am changing the link freqency */
+	old_freq = pci_read_config8(node1, link1 + PCI_HT_CAP_HOST_FREQ);
+	old_freq &= 0x0f;
+	needs_reset |= old_freq != freq;
+	old_freq = pci_read_config8(node2, link2 + PCI_HT_CAP_HOST_FREQ);
+	old_freq &= 0x0f;
+	needs_reset |= old_freq != freq;
+
+	/* Set the Calulcated link frequency */
+	pci_write_config8(node1, link1 + PCI_HT_CAP_HOST_FREQ, freq);
+	pci_write_config8(node2, link2 + PCI_HT_CAP_HOST_FREQ, freq);
+
+	/* Get the width capabilities */
+	width_cap1 = pci_read_config8(node1, link1 + PCI_HT_CAP_HOST_WIDTH);
+	width_cap2 = pci_read_config8(node2, link2 + PCI_HT_CAP_HOST_WIDTH);
+
+	/* Calculate node1's input width */
+	ln_width1 = link_width_to_pow2[width_cap1 & 7];
+	ln_width2 = link_width_to_pow2[(width_cap2 >> 4) & 7];
+	if (ln_width1 > ln_width2) {
+		ln_width1 = ln_width2;
+	}
+	width = pow2_to_link_width[ln_width1];
+	/* Calculate node1's output width */
+	ln_width1 = link_width_to_pow2[(width_cap1 >> 4) & 7];
+	ln_width2 = link_width_to_pow2[width_cap2 & 7];
+	if (ln_width1 > ln_width2) {
+		ln_width1 = ln_width2;
+	}
+	width |= pow2_to_link_width[ln_width1] << 4;
+	
+	/* See if I am changing node1's width */
+	old_width = pci_read_config8(node1, link1 + PCI_HT_CAP_HOST_WIDTH + 1);
+	old_width &= 0x77;	
+	needs_reset |= old_width != width;
+
+	/* Set node1's widths */
+	pci_write_config8(node1, link1 + PCI_HT_CAP_HOST_WIDTH + 1, width);
+
+	// * Calculate node2's width */
+	width = ((width & 0x70) >> 4) | ((width & 0x7) << 4);
+
+	/* See if I am changing node2's width */
+	old_width = pci_read_config8(node2, link2 + PCI_HT_CAP_HOST_WIDTH + 1);
+	old_width &= 0x77;
+	needs_reset |= old_width != width;
+
+	/* Set node2's widths */
+	pci_write_config8(node2, link2 + PCI_HT_CAP_HOST_WIDTH + 1, width);
+
+	return needs_reset;
+}
+
+static uint8_t get_linkn_first(uint8_t byte)
+{
+        if(byte & 0x02) { byte = 0; }
+        else if(byte & 0x04) { byte = 1; }
+        else if(byte & 0x08) { byte = 2; }
+        return byte;
+}
+
+static uint8_t get_linkn_last(uint8_t byte)
+{
+        if(byte & 0x02) { byte &= 0x0f; byte |= 0x00;  }
+        if(byte & 0x04) { byte &= 0x0f; byte |= 0x10;  }
+        if(byte & 0x08) { byte &= 0x0f; byte |= 0x20;  }
+        return byte>>4;
+}
+
+static uint8_t get_linkn_last_count(uint8_t byte)
+{
+        byte &= 0x0f;
+        if(byte & 0x02) { byte &= 0xcf; byte |= 0x00; byte+=0x40; }
+        if(byte & 0x04) { byte &= 0xcf; byte |= 0x10; byte+=0x40; }
+        if(byte & 0x08) { byte &= 0xcf; byte |= 0x20; byte+=0x40; }
+        return byte>>4;
+}
+
+static void setup_row_local(u8 source, u8 row) /* source will be 7 when it is for temp use*/
+{
+	uint8_t linkn;
+	uint32_t val;
+	val = 1;
+	for(linkn = 0; linkn<3; linkn++) { 
+		uint8_t regpos; 
+		uint32_t reg;
+		regpos = 0x98 + 0x20 * linkn;
+		reg = pci_read_config32(NODE_HT(source), regpos);
+		if ((reg & 0x17) != 3) continue; /* it is not conherent or not connected*/
+		val |= 1<<(linkn+1); 
+	}
+	val <<= 16;
+	val |= 0x0101;
+	fill_row(source,row, val);
+}
+
+static void setup_row_direct_x(u8 temp, u8 source, u8 dest, u8 linkn)
+{
+	uint32_t val;
+	uint32_t val_s;
+	val = 1<<(linkn+1);
+	val |= 1<<(linkn+1+8); /*for direct connect response route should equal to request table*/
+
+	if(((source &1)!=(dest &1)) 
+#if CROSS_BAR_47_56
+		&& ( (source<4)||(source>5) ) //(6,7) (7,6) should still be here
+					       //(6,5) (7,4) should be here	
+#endif
+	){
+		val |= (1<<16);
+	} else {
+		/*for CROSS_BAR_47_56  47, 56, should be here too
+			and for 47, 56, 57, 75, 46, 64 we need to substract another link to 
+				6,  7,  6,  6,  7,  7 
+		*/
+		val_s = get_row(temp, source);
+		val |= ((val_s>>16) - (1<<(linkn+1)))<<16;
+	}
+
+	fill_row(temp,dest, val );
+}
+
+#if CROSS_BAR_47_56
+static void opt_broadcast_rt(u8 source, u8 dest, u8 kickout) {
+	uint32_t val;
+	val = get_row(source, dest);
+	val -= link_connection(source, kickout)<<16;
+	fill_row(source, dest, val);
+}
+
+static void opt_broadcast_rt_group(const u8 *conn, int num) {
+        int i;
+
+        for(i=0; i<num; i+=3) {
+                opt_broadcast_rt(conn[i], conn[i+1],conn[i+2]);
+        }
+}
+static void opt_broadcast_rt_plus(u8 source, u8 dest, u8 kickout) {
+        uint32_t val;
+        val = get_row(source, dest);
+        val += link_connection(source, kickout)<<16;
+        fill_row(source, dest, val);
+}       
+        
+static void opt_broadcast_rt_plus_group(const u8 *conn, int num) {
+        int i;  
+                
+        for(i=0; i<num; i+=3) {
+                opt_broadcast_rt_plus(conn[i], conn[i+1],conn[i+2]);
+        }       
+} 
+#endif
+
+static void setup_row_direct(u8 source, u8 dest, u8 linkn){
+	setup_row_direct_x(source, source, dest, linkn);
+}
+
+static void setup_remote_row_direct(u8 source, u8 dest, u8 linkn){
+	setup_row_direct_x(7, source, dest, linkn);
+}
+
+static void setup_temp_row(u8 source, u8 dest)
+{
+	/* copy val from (source, dest) to (source,7) */
+	fill_row(source,7,get_row(source,dest));
+}
+
+static void clear_temp_row(u8 source)
+{
+	fill_row(source, 7, DEFAULT);
+}
+
+static void setup_remote_node(u8 node)
+{
+	static const uint8_t pci_reg[] = { 
+		0x44, 0x4c, 0x54, 0x5c, 0x64, 0x6c, 0x74, 0x7c, 
+		0x40, 0x48, 0x50, 0x58, 0x60, 0x68, 0x70, 0x78,
+		0x84, 0x8c, 0x94, 0x9c, 0xa4, 0xac, 0xb4, 0xbc,
+		0x80, 0x88, 0x90, 0x98, 0xa0, 0xa8, 0xb0, 0xb8,
+		0xc4, 0xcc, 0xd4, 0xdc,
+		0xc0, 0xc8, 0xd0, 0xd8,
+		0xe0, 0xe4, 0xe8, 0xec,
+	};
+	int i;
+
+	print_spew("setup_remote_node: ");
+
+	/* copy the default resource map from node 0 */
+	for(i = 0; i < sizeof(pci_reg)/sizeof(pci_reg[0]); i++) {
+		uint32_t value;
+		uint8_t reg;
+		reg = pci_reg[i];
+		value = pci_read_config32(NODE_MP(0), reg);
+		pci_write_config32(NODE_MP(7), reg, value);
+
+	}
+	print_spew("done\r\n");
+}
+
+#endif /* CONFIG_MAX_PHYSICAL_CPUS > 1*/
+
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 2
+#if !CROSS_BAR_47_56
+static void setup_row_indirect_x(u8 temp, u8 source, u8 dest)
+#else
+static void setup_row_indirect_x(u8 temp, u8 source, u8 dest, u8 gateway, u8 diff)
+#endif
+{
+	/*for indirect connection, we need to compute the val from val_s(source, source), and val_g(source, gateway) */
+	uint32_t val_s;
+	uint32_t val;
+#if !CROSS_BAR_47_56	
+	u8 gateway;
+	u8 diff;
+	if(source<dest) {
+		gateway = source + 2;
+	} else {
+		gateway = source - 2;
+	}
+#endif
+	val_s = get_row(temp, source);
+	val = get_row(temp, gateway);
+	
+	val &= 0xffff;
+	val_s >>= 16;
+	val_s &= 0xfe;
+	
+#if !CROSS_BAR_47_56
+	diff = ((source&1)!=(dest &1));
+#endif
+
+	if(diff && (val_s!=(val&0xff)) ) { /* use another connect as response*/
+		val_s -= val &  0xff;
+#if (CONFIG_MAX_PHYSICAL_CPUS > 4) || (CONFIG_MAX_PHYSICAL_CPUS_4_BUT_MORE_INSTALLED == 1)
+		uint8_t byte;
+		/* Some node have two links left
+		 * don't worry we only have (2, (3 as source need to handle
+		 */
+		byte = val_s;
+		byte = get_linkn_last_count(byte);
+		if((byte>>2)>1) { /* make sure not the corner*/
+			if(source<dest) {
+				val_s-=link_connection(temp, source-2); /* -down*/
+			} else {
+#if CROSS_BAR_47_56		
+				#if 0		
+				if(source==7) {
+					val_s-=link_connection(temp, 6);  // for 7,2 via 5
+				} else if (source==6){
+					val_s-=link_connection(temp, 7);  // for 6,3 via 4
+				} else 
+				#endif
+				if (source < gateway) { // for 5, 4 via 7
+				  	val_s-=link_connection(temp, source-2);
+				} else
+#endif					
+					val_s-=link_connection(temp, source+2); /* -up*/
+			}
+		}
+#endif
+		val &= 0xff;
+		val |= (val_s<<8);
+	} 
+
+	if(diff) { /* cross rung?*/
+		val |= (1<<16);
+	}
+	else {
+		val_s = get_row(temp, source);
+		val |= ((val_s>>16) - link_connection(temp, gateway))<<16; 
+	}
+
+	fill_row(temp, dest, val);
+
+}
+
+#if !CROSS_BAR_47_56
+static void setup_row_indirect(u8 source, u8 dest)
+{	
+	setup_row_indirect_x(source, source, dest);
+}
+#else           
+static void setup_row_indirect(u8 source, u8 dest, u8 gateway, u8 diff)
+{
+	setup_row_indirect_x(source, source, dest, gateway, diff);
+}
+#endif   
+
+static void setup_row_indirect_group(const u8 *conn, int num)
+{
+	int i;
+
+#if !CROSS_BAR_47_56
+	for(i=0; i<num; i+=2) {
+		setup_row_indirect(conn[i], conn[i+1]);
+#else
+	for(i=0; i<num; i+=4) {
+		setup_row_indirect(conn[i], conn[i+1],conn[i+2], conn[i+3]);
+#endif
+
+	}
+}
+
+#if !CROSS_BAR_47_56
+static void setup_remote_row_indirect(u8 source, u8 dest)
+{
+	setup_row_indirect_x(7, source, dest);
+}
+#else
+static void setup_remote_row_indirect(u8 source, u8 dest, u8 gateway, u8 diff)
+{
+	setup_row_indirect_x(7, source, dest, gateway, diff);
+}
+#endif
+
+static void setup_remote_row_indirect_group(const u8 *conn, int num)
+{
+	int i;
+
+#if !CROSS_BAR_47_56
+	for(i=0; i<num; i+=2) {
+		setup_remote_row_indirect(conn[i], conn[i+1]);
+#else
+	for(i=0; i<num; i+=4) {
+		setup_remote_row_indirect(conn[i], conn[i+1],conn[i+2], conn[i+3]);
+#endif
+	}
+}
+
+#endif /*CONFIG_MAX_PHYSICAL_CPUS > 2*/
+
+
+static void setup_uniprocessor(void)
+{
+	print_spew("Enabling UP settings\r\n");
+#if CONFIG_LOGICAL_CPUS==1
+	unsigned tmp = (pci_read_config32(NODE_MC(0), 0xe8) >> 12) & 3;
+	if (tmp>0) return;
+#endif
+	disable_probes();
+}
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 2
+static int optimize_connection_group(const u8 *opt_conn, int num) {
+	int needs_reset = 0;
+	int i;
+	for(i=0; i<num; i+=2) {
+		needs_reset = optimize_connection(
+			NODE_HT(opt_conn[i]), 0x80 + link_to_register(link_connection(opt_conn[i],opt_conn[i+1])),
+			NODE_HT(opt_conn[i+1]), 0x80 + link_to_register(link_connection(opt_conn[i+1],opt_conn[i])) );
+	}               
+	return needs_reset;
+}  
+#endif
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 1
+static unsigned setup_smp2(void)
+{
+	unsigned nodes;
+	u8 byte;
+	uint32_t val;
+	nodes = 2;
+
+	setup_row_local(0, 0); /* it will update the broadcast RT*/
+	
+	val = get_row(0,0);
+	byte = (val>>16) & 0xfe;
+	if(byte<0x2) { /* no coherent connection so get out.*/
+		nodes = 1;
+		return nodes;
+	}
+
+	/* Setup and check a temporary connection to node 1 */
+#if TRY_HIGH_FIRST == 1
+	byte = get_linkn_last(byte); /* Max Link to node1 */
+#else
+	byte = get_linkn_first(byte); /*Min Link to node1 --- according to AMD*/
+#endif
+	print_linkn("(0,1) link=", byte);
+	setup_row_direct(0,1, byte);
+	setup_temp_row(0, 1);
+	
+	verify_connection(7);
+
+	/* We found 2 nodes so far */
+	val = pci_read_config32(NODE_HT(7), 0x6c);
+	byte = (val>>2) & 0x3; /*get default link on node7 to node0*/
+	print_linkn("(1,0) link=", byte);
+	setup_row_local(7,1);
+	setup_remote_row_direct(1, 0, byte);
+
+#if (CONFIG_MAX_PHYSICAL_CPUS > 4) || (CONFIG_MAX_PHYSICAL_CPUS_4_BUT_MORE_INSTALLED == 1)	
+	val = get_row(7,1);
+	byte = (val>>16) & 0xfe;
+	byte = get_linkn_last_count(byte);
+	if((byte>>2)==3) { /* Oh! we need to treat it as node2. So use another link*/
+		val = get_row(0,0);
+		byte = (val>>16) & 0xfe;
+#if TRY_HIGH_FIRST == 1
+		byte = get_linkn_first(byte); /* Min link to Node1 */
+#else
+		byte = get_linkn_last(byte);  /* Max link to Node1*/
+#endif
+		print_linkn("\t-->(0,1) link=", byte);
+		setup_row_direct(0,1, byte);
+		setup_temp_row(0, 1);
+	
+		verify_connection(7);
+			
+		/* We found 2 nodes so far */
+		val = pci_read_config32(NODE_HT(7), 0x6c);
+		byte = (val>>2) & 0x3; /* get default link on node7 to node0*/
+		print_linkn("\t-->(1,0) link=", byte); 
+		setup_row_local(7,1);
+		setup_remote_row_direct(1, 0, byte);
+	}
+#endif
+	
+	setup_remote_node(1);  /* Setup the regs on the remote node */
+	rename_temp_node(1);    /* Rename Node 7 to Node 1  */
+	enable_routing(1);      /* Enable routing on Node 1 */
+#if 0
+	/*don't need and it is done by clear_dead_links */
+	clear_temp_row(0);
+#endif
+
+	return nodes;
+}
+#endif /*CONFIG_MAX_PHYSICAL_CPUS > 1 */
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 2
+
+static unsigned setup_smp4(void)
+{
+	unsigned nodes;
+	u8 byte;
+	uint32_t val;
+
+	nodes=4;
+
+	/* Setup and check temporary connection from Node 0 to Node 2 */
+	val = get_row(0,0);
+	byte = ((val>>16) & 0xfe) - link_connection(0,1);
+	byte = get_linkn_last_count(byte);
+
+	if((byte>>2)==0) { /* We should have two coherent for 4p and above*/
+		nodes = 2;
+		return nodes;
+	}
+
+	byte &= 3; /* bit [3,2] is count-1*/
+	print_linkn("(0,2) link=", byte); 
+	setup_row_direct(0, 2, byte);  /*(0,2) direct link done*/
+
+	/* We found 3 nodes so far. Now setup a temporary
+	* connection from node 0 to node 3 via node 1
+	*/
+	setup_temp_row(0,1); /* temp. link between nodes 0 and 1 */
+	/* here should setup_row_direct(1,3) at first, before that we should find the link in node 1 to 3*/
+	val = get_row(1,1);
+	byte = ((val>>16) & 0xfe) - link_connection(1,0);
+	byte = get_linkn_first(byte);
+	print_linkn("(1,3) link=", byte); 
+	setup_row_direct(1,3,byte);  /* (1, 3) direct link done*/
+
+	/* We found 4 nodes so far. Now setup all nodes for 4p */
+	// We need to make sure 0,2 and 1,3 link is set already
+#if !CROSS_BAR_47_56
+	static const u8 conn4_1[] = {
+		0,3,
+		1,2,
+	};	
+#else
+	static const u8 conn4_1[] = {
+		0,3,2,1,
+		1,2,3,1,
+	};
+#endif
+
+	setup_row_indirect_group(conn4_1, sizeof(conn4_1)/sizeof(conn4_1[0]));
+
+	setup_temp_row(0,2);
+	verify_connection(7);
+	val = pci_read_config32(NODE_HT(7), 0x6c);
+	byte = (val>>2) & 0x3; /* get default link on 7 to 0*/
+	print_linkn("(2,0) link=", byte); 
+
+	setup_row_local(7,2);
+	setup_remote_row_direct(2, 0, byte);  /* node 2 to node 0 direct link done */
+	setup_remote_node(2);  /* Setup the regs on the remote node */
+
+	rename_temp_node(2);    /* Rename Node 7 to Node 2  */
+	enable_routing(2);      /* Enable routing on Node 2 */
+
+	setup_temp_row(0,1);
+	setup_temp_row(1,3);
+	verify_connection(7);
+
+	val = pci_read_config32(NODE_HT(7), 0x6c);
+	byte = (val>>2) & 0x3; /* get default link on 7 to 1*/
+	print_linkn("(3,1) link=", byte); 
+
+	setup_row_local(7,3);
+	setup_remote_row_direct(3, 1, byte); /* node 3 to node 1 direct link done */
+	setup_remote_node(3);  /* Setup the regs on the remote node */
+
+	/* We need to init link between 2, and 3 direct link */
+	val = get_row(2,2);
+	byte = ((val>>16) & 0xfe) - link_connection(2,0);
+	byte = get_linkn_last_count(byte);
+	print_linkn("(2,3) link=", byte & 3);
+	
+	setup_row_direct(2,3, byte & 0x3);
+	setup_temp_row(0,2);
+	setup_temp_row(2,3);
+	verify_connection(7); /* to 3*/
+
+#if (CONFIG_MAX_PHYSICAL_CPUS > 4) || (CONFIG_MAX_PHYSICAL_CPUS_4_BUT_MORE_INSTALLED == 1)
+	/* We need to find out which link is to node3 */
+	if((byte>>2)==2) { /* one to node3, one to node0, one to node4*/
+		val = get_row(7,3);
+		if((val>>16) == 1) { /* that link is to node4, because via node1 it has been set, recompute it*/
+			val = get_row(2,2);
+			byte = ((val>>16) & 0xfe) - link_connection(2,0);
+			byte = get_linkn_first(byte);
+			print_linkn("\t-->(2,3) link=", byte); 
+			setup_row_direct(2,3,byte);
+			setup_temp_row(2,3);
+			verify_connection(7); /* to 3*/
+		}
+	} 
+#endif
+
+	val = pci_read_config32(NODE_HT(7), 0x6c);
+	byte = (val>>2) & 0x3; /* get default link on 7 to 2*/
+	print_linkn("(3,2) link=", byte); 
+	setup_remote_row_direct(3,2, byte);
+
+#if (CONFIG_MAX_PHYSICAL_CPUS > 4) || (CONFIG_MAX_PHYSICAL_CPUS_4_BUT_MORE_INSTALLED == 1)
+	/* set link from 3 to 5 before enable it*/
+        val = get_row(7,3);
+        byte = ((val>>16) & 0xfe) - link_connection(7,2) - link_connection(7,1);
+        byte = get_linkn_last_count(byte);
+        if((byte>>2)==1) { /* We should have three coherent links on node 3 for 6p and above*/
+        	byte &= 3; /*bit [3,2] is count-2*/
+	        print_linkn("(3,5) link=", byte);
+	        setup_remote_row_direct(3, 5, byte);
+	}
+	
+        val = get_row(2,2);
+        byte = ((val>>16) & 0xfe) - link_connection(2,3) - link_connection(2,0);
+        byte = get_linkn_last_count(byte);
+
+        if((byte>>2)==1) { /* We should have three coherent link on node 2 for 6p and above*/
+	        byte &= 3; /* bit [3,2] is count-2*/
+	        print_linkn("(2,4) link=", byte);
+	        setup_row_direct(2, 4, byte);
+	}	
+#endif
+
+	//Beside 3, 1 is set, We need to make sure 3, 5 is set already in case has three link in 3
+#if !CROSS_BAR_47_56
+        static const u8 conn4_3[] = {
+                3,0,
+        };
+#else
+        static const u8 conn4_3[] = {
+                3,0,1,1,
+        };
+#endif
+        setup_remote_row_indirect_group(conn4_3, sizeof(conn4_3)/sizeof(conn4_3[0]));
+
+/* ready to enable RT for Node 3 */
+	rename_temp_node(3);
+	enable_routing(3);      /* enable routing on node 3 (temp.) */
+
+	// beside 2, 0 is set, We need to make sure 2, 4 link is set already in case has three link in 2
+#if !CROSS_BAR_47_56    
+        static const u8 conn4_2[] = {
+                2,1,  
+        };      
+#else   
+        static const u8 conn4_2[] = {
+                2,1,0,1,
+        };      
+#endif          
+        setup_row_indirect_group(conn4_2, sizeof(conn4_2)/sizeof(conn4_2[0]));
+
+#if 0
+	/*We need to do sth to reverse work for setup_temp_row (0,1) (1,3) */
+	/* it will be done by clear_dead_links */
+	clear_temp_row(0);
+	clear_temp_row(1);
+#endif
+
+	return nodes;
+
+}
+
+#endif /* CONFIG_MAX_PHYSICAL_CPUS > 2 */
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 4
+
+static unsigned setup_smp6(void)
+{
+	unsigned nodes;
+	u8 byte;
+	uint32_t val;
+
+	nodes=6;
+
+	/* Setup and check temporary connection from Node 0 to Node 4  through 2*/
+        val = get_row(2,2);
+        byte = ((val>>16) & 0xfe) - link_connection(2,3) - link_connection(2,0);
+        byte = get_linkn_last_count(byte);
+
+        if((byte>>2)==0) { /* We should have three coherent link on node 2 for 6p and above*/
+                nodes = 4;
+                return nodes;
+        }
+
+	/* Setup and check temporary connection from Node 0 to Node 5  through 1, 3*/
+        /* set link from 3 to 5 before enable it*/
+        val = get_row(3,3);
+        byte = ((val>>16) & 0xfe) - link_connection(3,2) - link_connection(3,1);
+        byte = get_linkn_last_count(byte);
+        if((byte>>2)==0) { /* We should have three coherent links on node 3 for 6p and above*/
+                nodes = 4;
+                return nodes;
+        }
+	
+	/* We found 6 nodes so far. Now setup all nodes for 6p */
+#warning "FIXME we need to find out the correct gateway for 6p"	
+	static const u8 conn6_1[] = {
+#if !CROSS_BAR_47_56
+		0, 4,
+		0, 5,
+		1, 4,
+		1, 5,
+		2, 5,
+		3, 4,
+#else
+		0, 4, 2, 0,
+		0, 5, 2, 1,
+		1, 4, 3, 1,
+		1, 5, 3, 0,
+		2, 5, 3, 0,
+		3, 4, 2, 0,
+#endif
+	}; 
+
+	setup_row_indirect_group(conn6_1, sizeof(conn6_1)/sizeof(conn6_1[0]));
+	
+	for(byte=0; byte<4; byte+=2) {
+		setup_temp_row(byte,byte+2);
+	}
+	verify_connection(7);
+	val = pci_read_config32(NODE_HT(7), 0x6c);
+	byte = (val>>2) & 0x3; /*get default link on 7 to 2*/
+	print_linkn("(4,2) link=", byte); 
+	
+	setup_row_local(7,4);
+	setup_remote_row_direct(4, 2, byte);
+	setup_remote_node(4);  /* Setup the regs on the remote node */
+	
+	/* Set indirect connection to 0, to 3 */
+	//we only need to set 4,0 here
+	static const u8 conn6_2[] = {
+#if !CROSS_BAR_47_56
+		4, 0,
+#else
+		4, 0, 2, 0,
+#endif
+	};      
+	
+	setup_remote_row_indirect_group(conn6_2, sizeof(conn6_2)/sizeof(conn6_2[0]));
+	
+	rename_temp_node(4);
+	enable_routing(4);
+
+	setup_temp_row(0,1);
+	for(byte=0; byte<4; byte+=2) {
+		setup_temp_row(byte+1,byte+3);
+	}
+	verify_connection(7);
+
+	val = pci_read_config32(NODE_HT(7), 0x6c);
+	byte = (val>>2) & 0x3; /* get default link on 7 to 3*/
+	print_linkn("(5,3) link=", byte); 
+	setup_row_local(7,5);
+	setup_remote_row_direct(5, 3, byte);
+	setup_remote_node(5);  /* Setup the regs on the remote node */
+	
+#if !CROSS_BAR_47_56
+	/* We need to init link between 4, and 5 direct link */
+	val = get_row(4,4);
+	byte = ((val>>16) & 0xfe) - link_connection(4,2);
+	byte = get_linkn_last_count(byte);
+	print_linkn("(4,5) link=", byte & 3);
+	
+	setup_row_direct(4,5, byte & 0x3);
+	setup_temp_row(0,2);
+	setup_temp_row(2,4);
+	setup_temp_row(4,5);
+	verify_connection(7); /* to 5*/
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 6
+	/* We need to find out which link is to node5 */
+	
+	if((byte>>2)==2) { /* one to node5, one to node2, one to node6*/
+		val = get_row(7,5);
+		if((val>>16) == 1) { /* that link is to node6, because via node 3 node 5 has been set*/
+			val = get_row(4,4);
+			byte = ((val>>16) & 0xfe) - link_connection(4,2);
+			byte = get_linkn_first(byte);
+			print_linkn("\t-->(4,5) link=", byte);
+			setup_row_direct(4,5,byte);
+			setup_temp_row(4,5);
+			verify_connection(7); /* to 5*/
+		}
+	} 
+#endif
+
+	val = pci_read_config32(NODE_HT(7), 0x6c);
+	byte = (val>>2) & 0x3; /* get default link on 7 to 4*/
+	print_linkn("(5,4) link=", byte); 
+	setup_remote_row_direct(5,4, byte);
+	
+	//init 5, 7 here
+        val = get_row(7,5);
+        byte = ((val>>16) & 0xfe) - link_connection(7,4) - link_connection(7,3);
+        byte = get_linkn_last_count(byte);
+        if((byte>>2)==1) { /* We should have three coherent links on node 5 for 6p and above*/
+        	byte &= 3; /*bit [3,2] is count-2*/
+	        print_linkn("(5,7) link=", byte);
+	        setup_remote_row_direct(5, 7, byte);
+	}
+	
+	//init 4,6 here
+        val = get_row(4,4);
+        byte = ((val>>16) & 0xfe) - link_connection(4,5) - link_connection(4,2);
+        byte = get_linkn_last_count(byte);
+
+        if((byte>>2)==1) { /* We should have three coherent link on node 4 for 6p and above*/
+	        byte &= 3; /* bit [3,2] is count-2*/
+	        print_linkn("(4,6) link=", byte);
+	        setup_row_direct(4, 6, byte);
+	}	
+	
+#endif	
+	
+	//We need to set 5,0 here only, We need to set up 5, 7 to make 5,0
+	/* Set indirect connection to 0, to 3  for indirect we will use clockwise routing */
+	static const u8 conn6_3[] = {
+#if !CROSS_BAR_47_56
+		5, 0,
+#else
+		5, 0, 3, 0,
+#endif
+	};      
+	
+	setup_remote_row_indirect_group(conn6_3, sizeof(conn6_3)/sizeof(conn6_3[0]));
+
+/* ready to enable RT for 5 */
+	rename_temp_node(5);
+	enable_routing(5);      /* enable routing on node 5 (temp.) */
+
+	static const u8 conn6_4[] = {
+#if !CROSS_BAR_47_56
+		4, 1,
+		4, 3,
+		
+		5, 2,
+		5, 1,
+		
+#else
+		4, 1, 2, 0,
+		4, 3, 2, 0,
+		4, 5, 2, 0,
+		
+		5, 2, 3, 0,
+		5, 1, 3, 0,
+		5, 4, 3, 0,
+		
+#endif
+	};      
+	
+	setup_row_indirect_group(conn6_4, sizeof(conn6_4)/sizeof(conn6_4[0]));
+
+#if 0
+	/* We need to do sth about reverse about setup_temp_row (0,1), (2,4), (1, 3), (3,5)
+	 * It will be done by clear_dead_links 
+	 */
+	for(byte=0; byte<4; byte++) {
+		clear_temp_row(byte);
+	}
+#endif
+
+	return nodes;
+
+}
+
+#endif /* CONFIG_MAX_PHYSICAL_CPUS > 4 */
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 6
+
+static unsigned setup_smp8(void)
+{
+	unsigned nodes;
+	u8 byte;
+	uint32_t val;
+
+	nodes=8;
+
+	/* Setup and check temporary connection from Node 0 to Node 6 via 2 and 4 to 7 */
+	val = get_row(4,4);
+#if CROSS_BAR_47_56 
+	byte = ((val>>16) & 0xfe) - link_connection(4,2);
+#else
+	byte = ((val>>16) & 0xfe) - link_connection(4,5) - link_connection(4,2);
+        byte = get_linkn_last_count(byte); /* Max link to 6*/
+        if((byte>>2)==0) { /* We should have two or three coherent links on node 4 for 8p*/
+                nodes = 6;
+                return nodes;
+        }
+#endif
+
+#if CROSS_BAR_47_56 
+	byte = get_linkn_last_count(byte); /* Max link to 6*/
+	if((byte>>2)<2) { /* We should have two or three coherent links on node 4 for 8p*/
+		nodes = 6;
+		return nodes;
+	}
+#if TRY_HIGH_FIRST == 1
+	byte = ((val>>16) & 0xfe) - link_connection(4,2);
+	byte = get_linkn_first(byte);  /*Min link to 6*/
+#else
+	byte &= 3; /* bit [3,2] is count-1 or 2*/
+#endif
+	print_linkn("(4,6) link=", byte);
+	setup_row_direct(4, 6, byte);
+#endif
+
+#if !CROSS_BAR_47_56
+	/* Setup and check temporary connection from Node 0 to Node 7  through 1, 3, 5*/
+        val = get_row(5,5);
+        byte = ((val>>16) & 0xfe) - link_connection(5,4) - link_connection(5,3);
+        byte = get_linkn_last_count(byte);
+        if((byte>>2)==0) { /* We should have three coherent links on node 5 for 6p and above*/
+                nodes = 6;
+                return nodes;
+        }
+#endif
+
+	/* We found 8 nodes so far. Now setup all nodes for 8p */
+	static const u8 conn8_1[] = {
+#if !CROSS_BAR_47_56
+		0, 6,
+		/*0, 7,*/
+		1, 6,
+		/*1, 7,*/
+		2, 6,
+		/*2, 7,*/
+		3, 6,
+		/*3, 7,*/
+		/*4, 7,*/
+		5, 6,
+#else
+		0, 6, 2, 0,
+		/*0, 7, 2, 0,*/
+		1, 6, 3, 0,
+		/*1, 7, 3, 0,*/
+		2, 6, 4, 0,
+		/*2, 7, 4, 0,*/
+		3, 6, 5, 1,
+		/*3, 7, 5, 0,*/
+#endif
+	};
+
+	setup_row_indirect_group(conn8_1,sizeof(conn8_1)/sizeof(conn8_1[0]));
+
+	for(byte=0; byte<6; byte+=2) {
+		setup_temp_row(byte,byte+2);
+	}
+	verify_connection(7);
+	val = pci_read_config32(NODE_HT(7), 0x6c);
+	byte = (val>>2) & 0x3; /* get default link on 7 to 4*/
+	print_linkn("(6,4) link=", byte);
+	
+	setup_row_local(7,6);
+	setup_remote_row_direct(6, 4, byte);
+	setup_remote_node(6);  /* Setup the regs on the remote node */
+	/* Set indirect connection to 0, to 3   */
+#warning "FIXME we need to find out the correct gateway for 8p"		
+	static const u8 conn8_2[] = {
+#if !CROSS_BAR_47_56
+		6, 0,
+#else
+		6, 0, 4, 0,
+#endif
+	};
+
+	setup_remote_row_indirect_group(conn8_2, sizeof(conn8_2)/sizeof(conn8_2[0]));
+
+#if CROSS_BAR_47_56
+	//init 5, 6 here
+        /* here init 5, 6 */
+        /* Setup and check temporary connection from Node 0 to Node 5  through 1, 3, 5*/
+        val = get_row(5,5);
+        byte = ((val>>16) & 0xfe) - link_connection(5,3);
+#if TRY_HIGH_FIRST == 1
+	byte = get_linkn_first(byte);
+#else
+        byte = get_linkn_last(byte);        
+#endif
+        print_linkn("(5,6) link=", byte);
+        setup_row_direct(5, 6, byte);
+
+        setup_temp_row(0,1); /* temp. link between nodes 0 and 1 */
+        for(byte=0; byte<4; byte+=2) {
+                setup_temp_row(byte+1,byte+3);
+        }
+	setup_temp_row(5,6);
+
+        verify_connection(7);
+
+	val = get_row(7,6); // to chect it if it is node6 before renaming
+	if( (val>>16) == 1) { // it is real node 7 so swap it
+                /* We need to recompute link to 6 */
+       	        val = get_row(5,5);
+               	byte = ((val>>16) & 0xfe) - link_connection(5,3);
+#if TRY_HIGH_FIRST == 1
+                byte = get_linkn_first(byte);
+#else
+       	        byte = get_linkn_last(byte);
+#endif
+               	print_linkn("\t-->(5,6) link=", byte);
+                setup_row_direct(5, 6, byte);
+#if 0
+               	setup_temp_row(0,1); /* temp. link between nodes 0 and 1 */
+                for(byte=0; byte<4; byte+=2) {
+       	                setup_temp_row(byte+1,byte+3); 
+               	}
+#endif
+		setup_temp_row(5,6);
+
+       	        verify_connection(7);
+	}
+        val = pci_read_config32(NODE_HT(7), 0x6c);
+        byte = (val>>2) & 0x3; /* get default link on 7 to 5*/
+        print_linkn("(6,5) link=", byte);
+        setup_remote_row_direct(6, 5, byte);
+        /*Till now 56, 65 done */
+#endif
+	
+	rename_temp_node(6);
+	enable_routing(6);
+
+#if !CROSS_BAR_47_56
+	setup_temp_row(0,1);
+	for(byte=0; byte<6; byte+=2) {
+		setup_temp_row(byte+1,byte+3);
+	}
+
+	verify_connection(7);
+
+	val = pci_read_config32(NODE_HT(7), 0x6c);
+	byte = (val>>2) & 0x3; /* get default link on 7 to 5*/
+	print_linkn("(7,5) link=", byte); 
+	setup_row_local(7,7);
+	setup_remote_row_direct(7, 5, byte);
+
+#else
+        val = get_row(4,4);
+        byte = ((val>>16) & 0xfe) - link_connection(4,2) - link_connection(4,6);
+        byte = get_linkn_first(byte); 
+        print_linkn("(4,7) link=", byte); 
+        setup_row_direct(4, 7, byte);
+
+        /* Setup and check temporary connection from Node 0 to Node 7 through 2, and 4*/
+	for(byte=0; byte<4; byte+=2) {
+		setup_temp_row(byte,byte+2);
+	}
+
+        verify_connection(7);
+
+	val = pci_read_config32(NODE_HT(7), 0x6c);
+	byte = (val>>2) & 0x3; /* get default link on 7 to 4*/
+	print_linkn("(7,4) link=", byte); 
+	setup_row_local(7,7);
+	setup_remote_row_direct(7, 4, byte);
+	/* till now 4-7, 7-4 done. */
+#endif
+	setup_remote_node(7);  /* Setup the regs on the remote node */
+
+#if CROSS_BAR_47_56
+	/* here init 5, 7 */
+	/* Setup and check temporary connection from Node 0 to Node 5  through 1, 3, 5*/
+	val = get_row(5,5);
+	byte = ((val>>16) & 0xfe) - link_connection(5,3) - link_connection(5,6);
+	byte = get_linkn_first(byte);
+	print_linkn("(5,7) link=", byte); 
+	setup_row_direct(5, 7, byte);
+	
+	setup_temp_row(0,1); /* temp. link between nodes 0 and 1 */
+	for(byte=0; byte<4; byte+=2) {
+		setup_temp_row(byte+1,byte+3); 
+	}
+
+	verify_connection(7);
+
+	val = pci_read_config32(NODE_HT(7), 0x6c);
+	byte = (val>>2) & 0x3; /* get default link on 7 to 5*/
+	print_linkn("(7,5) link=", byte);
+	setup_remote_row_direct(7, 5, byte);
+	/*Till now 57, 75 done */
+	
+#endif
+
+	/* We need to init link between 6, and 7 direct link */
+	val = get_row(6,6);
+#if !CROSS_BAR_47_56
+	byte = ((val>>16) & 0xfe) - link_connection(6,4);
+#else
+	byte = ((val>>16) & 0xfe) - link_connection(6,4) - link_connection(6,5);
+#endif
+	byte = get_linkn_first(byte);
+	print_linkn("(6,7) link=", byte);
+	setup_row_direct(6,7, byte);
+
+	val = get_row(7,7);
+#if !CROSS_BAR_47_56
+	byte = ((val>>16) & 0xfe) - link_connection(7,5);
+#else
+	byte = ((val>>16) & 0xfe) - link_connection(7,5) - link_connection(7,4);
+#endif
+	byte = get_linkn_first(byte);
+	print_linkn("(7,6) link=", byte);
+	setup_row_direct(7,6, byte);
+
+	/* Set indirect connection to 0, to 3  for indirect we will use clockwise routing */
+	static const u8 conn8_3[] = {
+#if !CROSS_BAR_47_56
+		0, 7,  /* restore it*/
+		1, 7,
+		2, 7,
+		3, 7,
+		4, 7,
+		
+		6, 1,
+		6, 2,
+		6, 3,
+		6, 5,
+		
+		7, 0,
+		7, 1,
+		7, 2,
+		7, 3,
+		7, 4,
+#else
+
+		
+                4, 5, 6, 1,
+                5, 4, 7, 1,
+
+                6, 1, 5, 0, // or 4, 1
+                6, 2, 4, 0,
+                6, 3, 5, 0, // or 4, 1
+
+                7, 0, 4, 0, // or 5, 1
+                7, 1, 5, 0,
+                7, 2, 4, 0, // or 5, 1
+                7, 3, 5, 0,
+
+		0, 7, 2, 0, /* restore it*/
+		1, 7, 3, 0,
+		2, 7, 4, 1,
+		3, 7, 5, 0,
+
+                2, 5, 4, 1, /* reset it */
+                3, 4, 5, 1,
+
+		4, 1, 2, 1, /* reset it */
+		4, 3, 2, 1,
+
+		5, 2, 3, 1, /* reset it */
+		5, 0, 3, 1,
+		
+#endif
+	};
+
+	setup_row_indirect_group(conn8_3, sizeof(conn8_3)/sizeof(conn8_3[0]));
+
+#if CROSS_BAR_47_56
+        /* for 47, 56, 57, 75, 46, 64 we need to substract another link to 
+               6,  7,  6,  6,  7,  7 */
+        static const u8 conn8_4[] = {
+//direct
+                4, 7, 6,
+                5, 6, 7, 
+		5, 7, 6, 
+		7, 5, 6,  
+		4, 6, 7, 
+		6, 4, 7, 
+
+//in direct
+		0, 6, 1, 
+		0, 7, 1, 
+		
+		1, 6, 0, 
+		1, 7, 0, 
+
+		2, 6, 3, 
+//		2, 7, 3, +
+		
+//		3, 6, 1, +
+		3, 7, 2,
+ 
+		6, 0, 7,
+		6, 1, 7, // needed for via 5
+			6, 1, 4, // ???
+		6, 2, 7,
+		6, 3, 7, // needed for via 5 
+			6, 3, 4, //???
+		7, 0, 6, // needed for via 4
+			7, 0, 5, //???
+		7, 1, 6,
+		7, 2, 6, // needed for via 4
+			7, 2, 5, //???
+		7, 3, 6,
+        };
+
+        opt_broadcast_rt_group(conn8_4, sizeof(conn8_4)/sizeof(conn8_4[0]));
+
+        static const u8 conn8_5[] = {
+                2, 7, 0, 
+
+                3, 6, 1, 
+        };      
+                
+        opt_broadcast_rt_plus_group(conn8_5, sizeof(conn8_5)/sizeof(conn8_5[0]));
+#endif
+
+
+	
+/* ready to enable RT for Node 7 */
+	enable_routing(7);      /* enable routing on node 7 (temp.) */
+
+	return nodes;
+}
+
+#endif /* CONFIG_MAX_PHYSICAL_CPUS > 6 */
+
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 1
+
+static unsigned setup_smp(void)
+{
+	unsigned nodes;
+
+	print_spew("Enabling SMP settings\r\n");
+		
+	nodes = setup_smp2();
+#if CONFIG_MAX_PHYSICAL_CPUS > 2
+	if(nodes == 2) 
+		nodes = setup_smp4();
+#endif
+	
+#if CONFIG_MAX_PHYSICAL_CPUS > 4
+	if(nodes == 4)
+		nodes = setup_smp6();
+#endif
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 6
+	if(nodes == 6) 
+		nodes = setup_smp8();
+#endif
+
+#if CONFIG_USE_INIT
+	printk_debug("%02x nodes initialized.\r\n", nodes);
+#else
+	print_debug_hex8(nodes);
+	print_debug(" nodes initialized.\r\n");
+#endif
+	
+	return nodes;
+}
+
+static unsigned verify_mp_capabilities(unsigned nodes)
+{
+	unsigned node, mask;
+	
+	mask = 0x06; /* BigMPCap */
+
+	for (node=0; node<nodes; node++) {
+		mask &= pci_read_config32(NODE_MC(node), 0xe8);
+	}
+	
+	switch(mask) {
+#if CONFIG_MAX_PHYSICAL_CPUS > 2
+	case 0x02: /* MPCap    */
+		if(nodes > 2) {
+			print_err("Going back to DP\r\n");
+			return 2;
+		}
+		break;
+#endif
+	case 0x00: /* Non SMP */
+		if(nodes >1 ) {
+			print_err("Going back to UP\r\n");
+			return 1;
+		}
+		break;
+	}
+	
+	return nodes;
+
+}
+
+
+static void clear_dead_routes(unsigned nodes)
+{
+	int last_row;
+	int node, row;
+#if CONFIG_MAX_PHYSICAL_CPUS == 8
+	if(nodes==8) return;/* don't touch (7,7)*/
+#endif
+	last_row = nodes;
+	if (nodes == 1) {
+		last_row = 0;
+	}
+	for(node = 7; node >= 0; node--) {
+		for(row = 7; row >= last_row; row--) {
+			fill_row(node, row, DEFAULT);
+		}
+	}
+	
+	/* Update the local row */
+	for( node=0; node<nodes; node++) {
+		uint32_t val = 0;
+		for(row =0; row<nodes; row++) {
+			val |= get_row(node, row);
+		}
+		fill_row(node, node, (((val & 0xff) | ((val >> 8) & 0xff)) << 16) | 0x0101); 
+	}
+}
+#endif /* CONFIG_MAX_PHYSICAL_CPUS > 1 */
+
+#if CONFIG_LOGICAL_CPUS==1
+static unsigned verify_dualcore(unsigned nodes)
+{
+	unsigned node, totalcpus, tmp;
+
+	totalcpus = 0;
+	for (node=0; node<nodes; node++) {
+		tmp = (pci_read_config32(NODE_MC(node), 0xe8) >> 12) & 3 ;
+		totalcpus += (tmp + 1);
+	}
+
+	return totalcpus;
+		
+}
+#endif
+
+static void coherent_ht_finalize(unsigned nodes)
+{
+	unsigned node;
+	int rev_a0;
+#if CONFIG_LOGICAL_CPUS==1
+	unsigned total_cpus;
+
+	if(read_option(CMOS_VSTART_dual_core, CMOS_VLEN_dual_core, 0) == 0) { /* dual_core */
+		total_cpus = verify_dualcore(nodes);
+	}
+	else {
+		total_cpus = nodes;
+	}
+#endif
+	
+	/* set up cpu count and node count and enable Limit
+	 * Config Space Range for all available CPUs.
+	 * Also clear non coherent hypertransport bus range
+	 * registers on Hammer A0 revision.
+	 */
+
+	print_spew("coherent_ht_finalize\r\n");
+	rev_a0 = is_cpu_rev_a0();
+	for (node = 0; node < nodes; node++) {
+		device_t dev;
+		uint32_t val;
+		dev = NODE_HT(node);
+
+		/* Set the Total CPU and Node count in the system */
+		val = pci_read_config32(dev, 0x60);
+		val &= (~0x000F0070);
+#if CONFIG_LOGICAL_CPUS==1
+		val |= ((total_cpus-1)<<16)|((nodes-1)<<4);
+#else
+		val |= ((nodes-1)<<16)|((nodes-1)<<4);
+#endif
+		pci_write_config32(dev, 0x60, val);
+
+		/* Only respond to real cpu pci configuration cycles
+		 * and optimize the HT settings 
+		 */
+		val=pci_read_config32(dev, HT_TRANSACTION_CONTROL);
+		val &= ~((HTTC_BUF_REL_PRI_MASK << HTTC_BUF_REL_PRI_SHIFT) |
+			(HTTC_MED_PRI_BYP_CNT_MASK << HTTC_MED_PRI_BYP_CNT_SHIFT) |
+			(HTTC_HI_PRI_BYP_CNT_MASK << HTTC_HI_PRI_BYP_CNT_SHIFT));
+		val |= HTTC_LIMIT_CLDT_CFG | 
+			(HTTC_BUF_REL_PRI_8 << HTTC_BUF_REL_PRI_SHIFT) |
+			(3 << HTTC_MED_PRI_BYP_CNT_SHIFT) |
+			(3 << HTTC_HI_PRI_BYP_CNT_SHIFT);
+		pci_write_config32(dev, HT_TRANSACTION_CONTROL, val);
+
+		if (rev_a0) {
+			pci_write_config32(dev, 0x94, 0);
+			pci_write_config32(dev, 0xb4, 0);
+			pci_write_config32(dev, 0xd4, 0);
+		}
+	}
+
+	print_spew("done\r\n");
+}
+
+static int apply_cpu_errata_fixes(unsigned nodes)
+{
+	unsigned node;
+	int needs_reset = 0;
+	for(node = 0; node < nodes; node++) {
+		device_t dev;
+		uint32_t cmd;
+		dev = NODE_MC(node);
+		if (is_cpu_pre_c0()) {
+
+			/* Errata 66
+			 * Limit the number of downstream posted requests to 1 
+			 */
+			cmd = pci_read_config32(dev, 0x70);
+			if ((cmd & (3 << 0)) != 2) {
+				cmd &= ~(3<<0);
+				cmd |= (2<<0);
+				pci_write_config32(dev, 0x70, cmd );
+				needs_reset = 1;
+			}
+			cmd = pci_read_config32(dev, 0x7c);
+			if ((cmd & (3 << 4)) != 0) {
+				cmd &= ~(3<<4);
+				cmd |= (0<<4);
+				pci_write_config32(dev, 0x7c, cmd );
+				needs_reset = 1;
+			}
+			/* Clock Power/Timing Low */
+			cmd = pci_read_config32(dev, 0xd4);
+			if (cmd != 0x000D0001) {
+				cmd = 0x000D0001;
+				pci_write_config32(dev, 0xd4, cmd);
+				needs_reset = 1; /* Needed? */
+			}
+
+		}
+		else if (is_cpu_pre_d0()) { // d0 later don't need it
+			uint32_t cmd_ref;
+			/* Errata 98 
+			 * Set Clk Ramp Hystersis to 7
+			 * Clock Power/Timing Low
+			 */
+			cmd_ref = 0x04e20707; /* Registered */
+			cmd = pci_read_config32(dev, 0xd4);
+			if(cmd != cmd_ref) {
+				pci_write_config32(dev, 0xd4, cmd_ref );
+				needs_reset = 1; /* Needed? */
+			}
+		}
+	}
+	return needs_reset;
+}
+
+static int optimize_link_read_pointers(unsigned nodes)
+{
+	unsigned node;
+	int needs_reset = 0;
+	for(node = 0; node < nodes; node++) {
+		device_t f0_dev, f3_dev;
+		uint32_t cmd_ref, cmd;
+		int link;
+		f0_dev = NODE_HT(node);
+		f3_dev = NODE_MC(node);
+		cmd_ref = cmd = pci_read_config32(f3_dev, 0xdc);
+		for(link = 0; link < 3; link++) {
+			uint32_t link_type;
+			unsigned reg;
+			/* This works on an Athlon64 because unimplemented links return 0 */
+			reg = 0x98 + (link * 0x20);
+			link_type = pci_read_config32(f0_dev, reg);
+			/* Only handle coherent links */
+			if ((link_type & (LinkConnected | InitComplete|NonCoherent)) == 
+				(LinkConnected|InitComplete)) 
+			{
+				cmd &= ~(0xff << (link *8));
+				cmd |= 0x25 << (link *8);
+			}
+		}
+		if (cmd != cmd_ref) {
+			pci_write_config32(f3_dev, 0xdc, cmd);
+			needs_reset = 1;
+		}
+	}
+	return needs_reset;
+}
+
+static int optimize_link_coherent_ht(void)
+{
+        int needs_reset = 0;
+
+        unsigned nodes;
+
+        nodes = get_nodes();
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 1
+        if(nodes>1) {
+                needs_reset |= optimize_connection(
+                        NODE_HT(0), 0x80 + link_to_register(link_connection(0,1)),
+                        NODE_HT(1), 0x80 + link_to_register(link_connection(1,0)) );
+        }
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 2
+        if(nodes>2) {
+        /* optimize physical connections - by LYH */
+                static const u8 opt_conn4[] = {
+                        0,2,
+                        1,3,
+                        2,3,
+                };
+                needs_reset |= optimize_connection_group(opt_conn4, sizeof(opt_conn4)/sizeof(opt_conn4[0]));
+        }
+#endif
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 4
+        if(nodes>4) {
+                static const uint8_t opt_conn6[] ={
+                        2, 4,
+                        3, 5,
+        #if !CROSS_BAR_47_56
+                        4, 5,
+        #endif
+                };
+                needs_reset |= optimize_connection_group(opt_conn6, sizeof(opt_conn6)/sizeof(opt_conn6[0]));
+        }
+#endif
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 6
+        if(nodes>6) {
+               static const uint8_t opt_conn8[] ={
+                       4, 6,
+        #if CROSS_BAR_47_56
+                       4, 7,
+                       5, 6,
+        #endif
+                       5, 7,
+                       6, 7,
+               };
+                needs_reset |= optimize_connection_group(opt_conn8, sizeof(opt_conn8)/sizeof(opt_conn8[0]));
+        }
+#endif
+
+#endif
+
+        needs_reset |= apply_cpu_errata_fixes(nodes);
+        needs_reset |= optimize_link_read_pointers(nodes);
+
+        return needs_reset;
+}
+
+#if RAMINIT_SYSINFO == 1
+static void setup_coherent_ht_domain(void)
+#else
+static int setup_coherent_ht_domain(void)
+#endif
+{
+	unsigned nodes;
+	nodes = 1;
+
+	enable_bsp_routing();
+
+#if CONFIG_MAX_PHYSICAL_CPUS > 1
+        nodes = setup_smp();
+        nodes = verify_mp_capabilities(nodes);
+        clear_dead_routes(nodes);
+#endif
+
+	if (nodes == 1) {
+		setup_uniprocessor();
+	}
+	coherent_ht_finalize(nodes);
+
+#if RAMINIT_SYSINFO == 0
+	return optimize_link_coherent_ht();
+#endif
+}
+
diff --git a/src/northbridge/amd/amdk8/incoherent_ht.c b/src/northbridge/amd/amdk8/incoherent_ht.c
index 2623935ab2..c04781f2c0 100644
--- a/src/northbridge/amd/amdk8/incoherent_ht.c
+++ b/src/northbridge/amd/amdk8/incoherent_ht.c
@@ -968,34 +968,4 @@ static int optimize_link_incoherent_ht(struct sys_info *sysinfo)
 #endif
 
 
-static unsigned get_sblnk(void)
-{
-        uint32_t reg;
-        /* read PCI_DEV(0,0x18,0) 0x64 bit [8:9] to find out SbLink m */
-        reg = pci_read_config32(PCI_DEV(0, 0x18, 0), 0x64);
-        return ((reg>>8) & 3) ;
-}
-
-/* Look up a which bus a given node/link combination is on.
- * return 0 when we can't find the answer.
- */
-
-static unsigned node_link_to_bus(unsigned node, unsigned link)
-{
-        unsigned reg;
-
-        for(reg = 0xE0; reg < 0xF0; reg += 0x04) {
-                unsigned config_map;
-                config_map = pci_read_config32(PCI_DEV(0, 0x18, 1), reg);
-                if ((config_map & 3) != 3) {
-                        continue;
-                }
-                if ((((config_map >> 4) & 7) == node) &&
-                        (((config_map >> 8) & 3) == link))
-                {
-                        return (config_map >> 16) & 0xff;
-                }
-        }
-        return 0;
-}
 
diff --git a/src/northbridge/amd/amdk8/raminit.c b/src/northbridge/amd/amdk8/raminit.c
index c6a649a883..2141b31843 100644
--- a/src/northbridge/amd/amdk8/raminit.c
+++ b/src/northbridge/amd/amdk8/raminit.c
@@ -18,6 +18,10 @@
 #define K8_4RANK_DIMM_SUPPORT 0
 #endif
 
+#if USE_DCACHE_RAM == 1
+static void hard_reset(void);
+#endif
+
 #if 1
 static void setup_resource_map(const unsigned int *register_values, int max)
 {
diff --git a/src/northbridge/amd/amdk8/reset_test.c b/src/northbridge/amd/amdk8/reset_test.c
index e60447c2ff..932b807b45 100644
--- a/src/northbridge/amd/amdk8/reset_test.c
+++ b/src/northbridge/amd/amdk8/reset_test.c
@@ -51,3 +51,37 @@ static void set_bios_reset(void)
 	htic &= ~HTIC_BIOSR_Detect;
 	pci_write_config32(PCI_DEV(0, 0x18, 0), HT_INIT_CONTROL, htic);
 }
+
+static unsigned node_link_to_bus(unsigned node, unsigned link)
+{
+        unsigned reg;
+
+        for(reg = 0xE0; reg < 0xF0; reg += 0x04) {
+                unsigned config_map;
+                config_map = pci_read_config32(PCI_DEV(0, 0x18, 1), reg);
+                if ((config_map & 3) != 3) {
+                        continue;
+                }
+                if ((((config_map >> 4) & 7) == node) &&
+                        (((config_map >> 8) & 3) == link))
+                {
+                        return (config_map >> 16) & 0xff;
+                }
+        }
+        return 0;
+}
+
+static unsigned get_sblk(void)
+{
+        uint32_t reg;
+        /* read PCI_DEV(0,0x18,0) 0x64 bit [8:9] to find out SbLink m */
+        reg = pci_read_config32(PCI_DEV(0, 0x18, 0), 0x64);
+        return ((reg>>8) & 3) ;
+}
+
+static unsigned get_sbbusn(unsigned sblk)
+{
+	return node_link_to_bus(0, sblk);
+}
+
+
diff --git a/src/southbridge/amd/amd8111/Config.lb b/src/southbridge/amd/amd8111/Config.lb
index a69229260b..dfb7f16a49 100644
--- a/src/southbridge/amd/amd8111/Config.lb
+++ b/src/southbridge/amd/amd8111/Config.lb
@@ -9,3 +9,4 @@ driver amd8111_ac97.o
 driver amd8111_nic.o
 driver amd8111_pci.o
 driver amd8111_smbus.o
+object amd8111_reset.o
diff --git a/src/southbridge/amd/amd8111/amd8111_early_ctrl.c b/src/southbridge/amd/amd8111/amd8111_early_ctrl.c
index 8a648e8d37..9d40076bea 100644
--- a/src/southbridge/amd/amd8111/amd8111_early_ctrl.c
+++ b/src/southbridge/amd/amd8111/amd8111_early_ctrl.c
@@ -1,10 +1,31 @@
 /* by yhlu 2005.10 */
-static void hard_reset(struct sys_info *sysinfo)
+static unsigned get_sbdn(unsigned bus)
 {
         device_t dev;
-	
-        /* Find the device */
-        dev = PCI_DEV(sysinfo->sbbusn, sysinfo->sbdn+1, 3);
+
+        /* Find the device.
+         * There can only be one 8111 on a hypertransport chain/bus.
+         */
+        dev = pci_locate_device_on_bus(
+                PCI_ID(PCI_VENDOR_ID_AMD, PCI_DEVICE_ID_AMD_8111_PCI),
+                bus);
+
+        return (dev>>15) & 0x1f;
+
+}
+
+static void hard_reset(void)
+{
+        device_t dev;
+        unsigned bus;
+
+        /* Find the device.
+         * There can only be one 8111 on a hypertransport chain/bus.
+         */
+        bus = get_sbbusn(get_sblk());
+        dev = pci_locate_device_on_bus(
+                PCI_ID(PCI_VENDOR_ID_AMD, PCI_DEVICE_ID_AMD_8111_ACPI),
+                bus);
 
         set_bios_reset();
 
@@ -14,11 +35,11 @@ static void hard_reset(struct sys_info *sysinfo)
         outb(0x0e, 0x0cf9);
 }
 
-static void enable_fid_change_on_sb(struct sys_info *sysinfo)
+static void enable_fid_change_on_sb(unsigned sbbusn, unsigned sbdn)
 {
         device_t dev;
-        /* Find the device */
-        dev = PCI_DEV(sysinfo->sbbusn, sysinfo->sbdn+1, 3);
+
+	dev = PCI_DEV(sbbusn, sbdn+1, 3); // ACPI
 
         pci_write_config8(dev, 0x74, 4);
 
@@ -27,15 +48,25 @@ static void enable_fid_change_on_sb(struct sys_info *sysinfo)
 
 }
 
-static void soft_reset(struct sys_info *sysinfo)
+static void soft_reset_x(unsigned sbbusn, unsigned sbdn)
 {
         device_t dev;
-        
-	/* Find the device */
-        dev = PCI_DEV(sysinfo->sbbusn, sysinfo->sbdn+1, 0);
 
+	dev = PCI_DEV(sbbusn, sbdn+1, 0); //ISA
+
+        /* Reset */
         set_bios_reset();
         pci_write_config8(dev, 0x47, 1);
+
 }
 
+static void soft_reset(void)
+{
 
+	unsigned sblk = get_sblk();
+	unsigned sbbusn = get_sbbusn(sblk);
+	unsigned sbdn = get_sbdn(sbbusn);
+
+	return soft_reset_x(sbbusn, sbdn);
+
+}
diff --git a/src/southbridge/amd/amd8111/amd8111_lpc.c b/src/southbridge/amd/amd8111/amd8111_lpc.c
index 1850d51185..0232c6a3e5 100644
--- a/src/southbridge/amd/amd8111/amd8111_lpc.c
+++ b/src/southbridge/amd/amd8111/amd8111_lpc.c
@@ -70,10 +70,12 @@ static void setup_ioapic(void)
 	unsigned long ioapic_base = 0xfec00000;
 	volatile unsigned long *l;
 	struct ioapicreg *a = ioapicregvalues;
+	unsigned long bsp_apicid = lapicid();
 
 	l = (unsigned long *) ioapic_base;
 
-	ioapicregvalues[0].value_high = lapicid()<<(56-32); 
+	ioapicregvalues[0].value_high = bsp_apicid<<(56-32);
+	printk_debug("amd8111: ioapic bsp_apicid = %02x\n", bsp_apicid); 
 	
 	for (i = 0; i < sizeof(ioapicregvalues) / sizeof(ioapicregvalues[0]);
 	     i++, a++) {
@@ -88,7 +90,7 @@ static void setup_ioapic(void)
 			return;
 		}
 		printk_spew("for IRQ, reg 0x%08x value 0x%08x 0x%08x\n", 
-			a->reg, a->value_low, a->value_high);
+			    a->reg, a->value_low, a->value_high);
 	}
 }
 
@@ -179,7 +181,7 @@ static void amd8111_lpc_enable_resources(device_t dev)
 static void lpci_set_subsystem(device_t dev, unsigned vendor, unsigned device)
 {
 	pci_write_config32(dev, 0x70, 
-		((device & 0xffff) << 16) | (vendor & 0xffff));
+			   ((device & 0xffff) << 16) | (vendor & 0xffff));
 }
 
 static struct pci_operations lops_pci = {
diff --git a/src/southbridge/amd/amd8111/amd8111_reset.c b/src/southbridge/amd/amd8111/amd8111_reset.c
index 435904aabd..8a5f3a5135 100644
--- a/src/southbridge/amd/amd8111/amd8111_reset.c
+++ b/src/southbridge/amd/amd8111/amd8111_reset.c
@@ -1,5 +1,3 @@
-/* Include this file in the mainboards reset.c
- */
 #include <arch/io.h>
 #include <device/pci_ids.h>
 
@@ -38,7 +36,7 @@ static unsigned pci_read_config32(device_t dev, unsigned where)
 }
 
 #define PCI_DEV_INVALID (0xffffffffU)
-static device_t pci_locate_device(unsigned pci_id, unsigned bus)
+static device_t pci_locate_device_on_bus(unsigned pci_id, unsigned bus)
 {
 	device_t dev, last;
 	dev = PCI_DEV(bus, 0, 0);
@@ -55,35 +53,19 @@ static device_t pci_locate_device(unsigned pci_id, unsigned bus)
 
 #include "../../../northbridge/amd/amdk8/reset_test.c"
 
-static unsigned node_link_to_bus(unsigned node, unsigned link)
-{
-	unsigned reg;
-
-	for(reg = 0xE0; reg < 0xF0; reg += 0x04) {
-		unsigned config_map;
-		config_map = pci_read_config32(PCI_DEV(0, 0x18, 1), reg);
-		if ((config_map & 3) != 3) {
-			continue;
-		}
-		if ((((config_map >> 4) & 7) == node) &&
-			(((config_map >> 8) & 3) == link)) 
-		{
-			return (config_map >> 16) & 0xff;
-		}
-	}
-	return 0;
-}
 
-static void amd8111_hard_reset(unsigned node, unsigned link)
+void hard_reset(void)
 {
 	device_t dev;
 	unsigned bus;
+        unsigned node = 0;
+        unsigned link = get_sblk();
 
 	/* Find the device.
 	 * There can only be one 8111 on a hypertransport chain/bus.
 	 */
 	bus = node_link_to_bus(node, link);
-	dev = pci_locate_device(
+	dev = pci_locate_device_on_bus(
 		PCI_ID(PCI_VENDOR_ID_AMD, PCI_DEVICE_ID_AMD_8111_ISA), 
 		bus);
 
diff --git a/src/southbridge/nvidia/ck804/ck804_early_setup.c b/src/southbridge/nvidia/ck804/ck804_early_setup.c
index 7a7e467e72..bd14635541 100644
--- a/src/southbridge/nvidia/ck804/ck804_early_setup.c
+++ b/src/southbridge/nvidia/ck804/ck804_early_setup.c
@@ -383,3 +383,23 @@ static int ck804_early_setup_x(void)
 	ck804_early_clear_port();
 	return set_ht_link_ck804(4);
 }
+
+static void hard_reset(void)
+{
+        set_bios_reset();
+
+        /* full reset */
+        outb(0x0a, 0x0cf9);
+        outb(0x0e, 0x0cf9);
+}
+
+static void soft_reset(void)
+{
+        set_bios_reset();
+#if 1
+        /* link reset */
+        outb(0x02, 0x0cf9);
+        outb(0x06, 0x0cf9);
+#endif
+}
+
diff --git a/src/southbridge/nvidia/ck804/ck804_early_setup_car.c b/src/southbridge/nvidia/ck804/ck804_early_setup_car.c
index bc382ca7e3..a144f1227a 100644
--- a/src/southbridge/nvidia/ck804/ck804_early_setup_car.c
+++ b/src/southbridge/nvidia/ck804/ck804_early_setup_car.c
@@ -426,3 +426,23 @@ static int ck804_early_setup_x(void)
 	ck804_early_clear_port(ck804_num, busn, io_base);
 	return set_ht_link_ck804(4);
 }
+
+static void hard_reset(void)
+{
+        set_bios_reset();
+
+        /* full reset */
+        outb(0x0a, 0x0cf9);
+        outb(0x0e, 0x0cf9);
+}
+
+static void soft_reset(void)
+{
+        set_bios_reset();
+#if 1
+        /* link reset */
+        outb(0x02, 0x0cf9);
+        outb(0x06, 0x0cf9);
+#endif
+}
+
diff --git a/src/southbridge/nvidia/ck804/ck804_enable_rom.c b/src/southbridge/nvidia/ck804/ck804_enable_rom.c
index 48ce689358..070c09efe3 100644
--- a/src/southbridge/nvidia/ck804/ck804_enable_rom.c
+++ b/src/southbridge/nvidia/ck804/ck804_enable_rom.c
@@ -3,6 +3,7 @@
  *  by yhlu@tyan.com
  */
 
+
 static void ck804_enable_rom(void)
 {
 	unsigned char byte;
diff --git a/src/southbridge/nvidia/ck804/ck804_smbus.h b/src/southbridge/nvidia/ck804/ck804_smbus.h
index 079007836f..88843585aa 100644
--- a/src/southbridge/nvidia/ck804/ck804_smbus.h
+++ b/src/southbridge/nvidia/ck804/ck804_smbus.h
@@ -29,14 +29,13 @@ static int smbus_wait_until_ready(unsigned smbus_io_base)
 		unsigned char val;
 		smbus_delay();
 		val = inb(smbus_io_base + SMBHSTSTAT);
-		if ((val & 0x1f) == 0) {
-			break;
-		}
-		if(loops == (SMBUS_TIMEOUT / 2)) {
-			outb((val & 0x1f),smbus_io_base + SMBHSTSTAT);
+		val &= 0x1f;
+		if (val == 0) {
+			return 0;
 		}
+		outb(val,smbus_io_base + SMBHSTSTAT);
 	} while(--loops);
-	return loops?0:-2;
+	return -2;
 }
 
 static int smbus_wait_until_done(unsigned smbus_io_base)
@@ -49,10 +48,10 @@ static int smbus_wait_until_done(unsigned smbus_io_base)
 		
 		val = inb(smbus_io_base + SMBHSTSTAT);
 		if ( (val & 0xff) != 0) {
-			break;
+			return 0;
 		}
 	} while(--loops);
-	return loops?0:-3;
+	return -3;
 }
 static int do_smbus_recv_byte(unsigned smbus_io_base, unsigned device)
 {
@@ -200,3 +199,4 @@ static int do_smbus_write_byte(unsigned smbus_io_base, unsigned device, unsigned
         }
         return 0;
 }
+