2007-03-05 Till Straumann <strauman@slac.stanford.edu>

	* start/start.S, startup/bspstart.c: disable memory-select
	errors early (in start.S) to avoid hangs due to speculative
	memory access (motload maps absent memory in TLBs).
	Remove TLB mappings for which no physical memory is installed
	(bspstart). Enable HID1[RFXE] so that 'core_fault_in' errors
	result in a machine-check rather than stalling the machine.
	Re-enable memory-select errors at this point.

3 files changed, 84 insertions, 13 deletions

diff --git a/c/src/lib/libbsp/powerpc/mvme3100/ChangeLog b/c/src/lib/libbsp/powerpc/mvme3100/ChangeLog
index 2254ca5167..14b1c7feb3 100644
--- a/c/src/lib/libbsp/powerpc/mvme3100/ChangeLog
+++ b/c/src/lib/libbsp/powerpc/mvme3100/ChangeLog
@@ -1,5 +1,15 @@
 2007-03-05	Till Straumann <strauman@slac.stanford.edu>
 
+	* start/start.S, startup/bspstart.c: disable memory-select
+	errors early (in start.S) to avoid hangs due to speculative
+	memory access (motload maps absent memory in TLBs).
+	Remove TLB mappings for which no physical memory is installed
+	(bspstart). Enable HID1[RFXE] so that 'core_fault_in' errors
+	result in a machine-check rather than stalling the machine.
+	Re-enable memory-select errors at this point.
+
+2007-03-05	Till Straumann <strauman@slac.stanford.edu>
+
 	* Makefile.am: add e500/mmu.rel to libbsp_a_LIBADD.
 
 2008-01-29	Eric Norum <norume@aps.anl.gov>
diff --git a/c/src/lib/libbsp/powerpc/mvme3100/start/start.S b/c/src/lib/libbsp/powerpc/mvme3100/start/start.S
index cab9261c0a..485a00483d 100644
--- a/c/src/lib/libbsp/powerpc/mvme3100/start/start.S
+++ b/c/src/lib/libbsp/powerpc/mvme3100/start/start.S
@@ -25,6 +25,13 @@
 
 #define KERNELBASE	0x0
 
+/* cannot include <bsp.h> from assembly :-( */
+#ifndef BSP_8540_CCSR_BASE
+#define BSP_8540_CCSR_BASE 0xe1000000
+#endif
+
+#define ERR_DISABLE_REG (BSP_8540_CCSR_BASE + 0x2e44)
+
 	.text
 	.globl	__rtems_entry_point
 	.type	__rtems_entry_point,@function
@@ -34,6 +41,25 @@ __rtems_entry_point:
 	mr	r29,r5
 	mr	r28,r6
 	mr	r27,r7
+	/* disable checking for memory-select errors; motload has all TLBs
+	 * mapping a possible larger area as memory (not-guarded, caching-enabled)
+	 * than actual physical memory is available.
+	 * In case of speculative loads this may cause 'memory-select' errors
+	 * which seem to raise 'core_fault_in' (found no description in 
+	 * the manual but I experienced this problem).
+	 * Such errors (if HID1[RFXE] is clear) may *stall* execution
+	 * leading to mysterious 'hangs'.
+	 * Note: enabling HID1[RFXE] at this point makes no sense since
+	 *       exceptions are not configured yet. Therefore we disable
+	 *       memory-select errors.
+	 * Eventually (bspstart.c) we want to delete TLB entries for
+	 * which no physical memory is present.
+	 */
+	lis r3, ERR_DISABLE_REG@ha
+	lwz r4, ERR_DISABLE_REG@l(r3)
+	/* disable memory-select errors */
+	ori r4, r4, 1
+	stw r4, ERR_DISABLE_REG@l(r3)
 
 	/* Use MotLoad's TLB setup for now; caches are on already */
 	bl  __eabi	/* setup EABI and SYSV environment */
diff --git a/c/src/lib/libbsp/powerpc/mvme3100/startup/bspstart.c b/c/src/lib/libbsp/powerpc/mvme3100/startup/bspstart.c
index c5486abf34..9451062680 100644
--- a/c/src/lib/libbsp/powerpc/mvme3100/startup/bspstart.c
+++ b/c/src/lib/libbsp/powerpc/mvme3100/startup/bspstart.c
@@ -27,6 +27,7 @@
 #include <rtems/bspIo.h>
 #include <libcpu/spr.h>
 #include <libcpu/io.h>
+#include <libcpu/e500_mmu.h>
 #include <bsp/uart.h> 
 #include <bsp/irq.h> 
 #include <bsp/pci.h> 
@@ -38,6 +39,8 @@
 #define SHOW_MORE_INIT_SETTINGS
 #undef  DEBUG
 
+#define NumberOf(arr) (sizeof(arr)/sizeof(arr[0]))
+
 #ifdef  DEBUG
 #define STATIC
 #else
@@ -234,13 +237,15 @@ SPR_RW(HID1)
 
 void bsp_start( void )
 {
-unsigned char      *stack;
-register uint32_t  intrStack;
-register uint32_t  *intrStackPtr;
-unsigned char      *work_space_start;
-char               *chpt;
-ppc_cpu_id_t       myCpu;
-ppc_cpu_revision_t myCpuRevision;
+unsigned char       *stack;
+register uint32_t   intrStack;
+register uint32_t   *intrStackPtr;
+unsigned char       *work_space_start;
+char                *chpt;
+ppc_cpu_id_t        myCpu;
+ppc_cpu_revision_t  myCpuRevision;
+int                 i;
+E500_tlb_va_cache_t *tlb;
 
 VpdBufRec          vpdData [] = {
 	{ key: ProductIdent, instance: 0, buf: BSP_productIdent, buflen: sizeof(BSP_productIdent) - 1 },
@@ -321,13 +326,45 @@ VpdBufRec          vpdData [] = {
 	printk("Going to start PCI buses scanning and initialization\n");
 #endif
 
+	BSP_mem_size            = BSP_get_mem_size();
+
 	{
-		/* disable checking for memory-select errors */
-		*(volatile uint32_t*)0xe1002e44 |= 1;
-		/* clear all pending errors */
-		*(volatile uint32_t*)0xe1002e40  = 0xffffffff;
+		/* memory-select errors were disabled in 'start.S';
+		 * motload has all TLBs mapping a possible larger area as
+		 * memory (not-guarded, caching-enabled) than actual physical
+		 * memory is available.
+		 * In case of speculative loads this may cause 'memory-select' errors
+		 * which seem to raise 'core_fault_in' (found no description in
+		 * the manual but I experienced this problem).
+		 * Such errors (if HID1[RFXE] is clear) may *stall* execution
+		 * leading to mysterious 'hangs'.
+		 *
+		 * Here we remove all mappings, re-enable memory-select
+		 * errors and make sure we enable HID1[RFXE] to avoid
+		 * stalls (since we don't implement handling individual
+		 * error-handling interrupts).
+		 */
+
 		/* enable machine check for bad bus errors */
 		_write_HID1( _read_HID1() | 0x20000 );
+
+		rtems_e500_initlb();
+
+		for ( i=0, tlb=rtems_e500_tlb_va_cache; i<NumberOf(rtems_e500_tlb_va_cache); i++, tlb++ ) {
+			/* disable TLBs for caching-enabled, non-guarded areas
+			 * beyond physical memory
+			 */
+			if (    tlb->att.v
+			    &&  0xa != (tlb->att.wimge & 0xa)
+				&&  (tlb->va.va_epn<<12) >= BSP_mem_size ) {
+				rtems_e500_clrtlb( E500_SELTLB_1 | i );
+			}
+		}
+
+		/* clear all pending memory errors */
+		_ccsr_wr32(0x2e40, 0xffffffff);
+		/* enable checking for memory-select errors */
+		_ccsr_wr32(0x2e44, _ccsr_rd32(0x2e44) & ~1 );
 	}
 
 	printk("Build Date: %s\n",BSP_build_date);
@@ -389,8 +426,6 @@ VpdBufRec          vpdData [] = {
 	_write_SPRG0(PPC_BSP_HAS_FIXED_PR288);
 #endif
 
-	BSP_mem_size            = BSP_get_mem_size();
-
 	if ( (chpt = strstr(BSP_commandline_string,"MEMSZ=")) ) {
 		char		*endp;
 		uint32_t 	sz;