[PATCH] MCA/INIT: use per cpu stacks

The bulk of the change.  Use per cpu MCA/INIT stacks.  Change the SAL
to OS state (sos) to be per process.  Do all the assembler work on the
MCA/INIT stacks, leaving the original stack alone.  Pass per cpu state
data to the C handlers for MCA and INIT, which also means changing the
mca_drv interfaces slightly.  Lots of verification on whether the
original stack is usable before converting it to a sleeping process.

Signed-off-by: Keith Owens <kaos@sgi.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>

4 files changed, 1272 insertions, 984 deletions

diff --git a/arch/ia64/kernel/asm-offsets.c b/arch/ia64/kernel/asm-offsets.c
index 7d1ae2982c53..f6a234289341 100644
--- a/arch/ia64/kernel/asm-offsets.c
+++ b/arch/ia64/kernel/asm-offsets.c
@@ -211,17 +211,41 @@ void foo(void)
 #endif
 
 	BLANK();
-	DEFINE(IA64_MCA_CPU_PROC_STATE_DUMP_OFFSET,
-	       offsetof (struct ia64_mca_cpu, proc_state_dump));
-	DEFINE(IA64_MCA_CPU_STACK_OFFSET,
-	       offsetof (struct ia64_mca_cpu, stack));
-	DEFINE(IA64_MCA_CPU_STACKFRAME_OFFSET,
-	       offsetof (struct ia64_mca_cpu, stackframe));
-	DEFINE(IA64_MCA_CPU_RBSTORE_OFFSET,
-	       offsetof (struct ia64_mca_cpu, rbstore));
+	DEFINE(IA64_MCA_CPU_MCA_STACK_OFFSET,
+	       offsetof (struct ia64_mca_cpu, mca_stack));
 	DEFINE(IA64_MCA_CPU_INIT_STACK_OFFSET,
 	       offsetof (struct ia64_mca_cpu, init_stack));
 	BLANK();
+	DEFINE(IA64_SAL_OS_STATE_COMMON_OFFSET,
+	       offsetof (struct ia64_sal_os_state, sal_ra));
+	DEFINE(IA64_SAL_OS_STATE_OS_GP_OFFSET,
+	       offsetof (struct ia64_sal_os_state, os_gp));
+	DEFINE(IA64_SAL_OS_STATE_PAL_MIN_STATE_OFFSET,
+	       offsetof (struct ia64_sal_os_state, pal_min_state));
+	DEFINE(IA64_SAL_OS_STATE_PROC_STATE_PARAM_OFFSET,
+	       offsetof (struct ia64_sal_os_state, proc_state_param));
+	DEFINE(IA64_SAL_OS_STATE_SIZE,
+	       sizeof (struct ia64_sal_os_state));
+	DEFINE(IA64_PMSA_GR_OFFSET,
+	       offsetof (struct pal_min_state_area_s, pmsa_gr));
+	DEFINE(IA64_PMSA_BANK1_GR_OFFSET,
+	       offsetof (struct pal_min_state_area_s, pmsa_bank1_gr));
+	DEFINE(IA64_PMSA_PR_OFFSET,
+	       offsetof (struct pal_min_state_area_s, pmsa_pr));
+	DEFINE(IA64_PMSA_BR0_OFFSET,
+	       offsetof (struct pal_min_state_area_s, pmsa_br0));
+	DEFINE(IA64_PMSA_RSC_OFFSET,
+	       offsetof (struct pal_min_state_area_s, pmsa_rsc));
+	DEFINE(IA64_PMSA_IIP_OFFSET,
+	       offsetof (struct pal_min_state_area_s, pmsa_iip));
+	DEFINE(IA64_PMSA_IPSR_OFFSET,
+	       offsetof (struct pal_min_state_area_s, pmsa_ipsr));
+	DEFINE(IA64_PMSA_IFS_OFFSET,
+	       offsetof (struct pal_min_state_area_s, pmsa_ifs));
+	DEFINE(IA64_PMSA_XIP_OFFSET,
+	       offsetof (struct pal_min_state_area_s, pmsa_xip));
+	BLANK();
+
 	/* used by fsys_gettimeofday in arch/ia64/kernel/fsys.S */
 	DEFINE(IA64_TIME_INTERPOLATOR_ADDRESS_OFFSET, offsetof (struct time_interpolator, addr));
 	DEFINE(IA64_TIME_INTERPOLATOR_SOURCE_OFFSET, offsetof (struct time_interpolator, source));
diff --git a/arch/ia64/kernel/mca.c b/arch/ia64/kernel/mca.c
index 8d484204a3ff..6dc726ad7137 100644
--- a/arch/ia64/kernel/mca.c
+++ b/arch/ia64/kernel/mca.c
@@ -48,6 +48,9 @@
  *            Delete dead variables and functions.
  *            Reorder to remove the need for forward declarations and to consolidate
  *            related code.
+ *
+ * 2005-08-12 Keith Owens <kaos@sgi.com>
+ *	      Convert MCA/INIT handlers to use per event stacks and SAL/OS state.
  */
 #include <linux/config.h>
 #include <linux/types.h>
@@ -77,6 +80,8 @@
 #include <asm/irq.h>
 #include <asm/hw_irq.h>
 
+#include "entry.h"
+
 #if defined(IA64_MCA_DEBUG_INFO)
 # define IA64_MCA_DEBUG(fmt...)	printk(fmt)
 #else
@@ -84,9 +89,7 @@
 #endif
 
 /* Used by mca_asm.S */
-ia64_mca_sal_to_os_state_t	ia64_sal_to_os_handoff_state;
-ia64_mca_os_to_sal_state_t	ia64_os_to_sal_handoff_state;
-u64				ia64_mca_serialize;
+u32				ia64_mca_serialize;
 DEFINE_PER_CPU(u64, ia64_mca_data); /* == __per_cpu_mca[smp_processor_id()] */
 DEFINE_PER_CPU(u64, ia64_mca_per_cpu_pte); /* PTE to map per-CPU area */
 DEFINE_PER_CPU(u64, ia64_mca_pal_pte);	    /* PTE to map PAL code */
@@ -95,8 +98,10 @@ DEFINE_PER_CPU(u64, ia64_mca_pal_base);    /* vaddr PAL code granule */
 unsigned long __per_cpu_mca[NR_CPUS];
 
 /* In mca_asm.S */
-extern void			ia64_monarch_init_handler (void);
-extern void			ia64_slave_init_handler (void);
+extern void			ia64_os_init_dispatch_monarch (void);
+extern void			ia64_os_init_dispatch_slave (void);
+
+static int monarch_cpu = -1;
 
 static ia64_mc_info_t		ia64_mc_info;
 
@@ -234,7 +239,8 @@ ia64_log_get(int sal_info_type, u8 **buffer, int irq_safe)
  *  This function retrieves a specified error record type from SAL
  *  and wakes up any processes waiting for error records.
  *
- *  Inputs  :   sal_info_type   (Type of error record MCA/CMC/CPE/INIT)
+ *  Inputs  :   sal_info_type   (Type of error record MCA/CMC/CPE)
+ *              FIXME: remove MCA and irq_safe.
  */
 static void
 ia64_mca_log_sal_error_record(int sal_info_type)
@@ -242,7 +248,7 @@ ia64_mca_log_sal_error_record(int sal_info_type)
 	u8 *buffer;
 	sal_log_record_header_t *rh;
 	u64 size;
-	int irq_safe = sal_info_type != SAL_INFO_TYPE_MCA && sal_info_type != SAL_INFO_TYPE_INIT;
+	int irq_safe = sal_info_type != SAL_INFO_TYPE_MCA;
 #ifdef IA64_MCA_DEBUG_INFO
 	static const char * const rec_name[] = { "MCA", "INIT", "CMC", "CPE" };
 #endif
@@ -330,182 +336,6 @@ ia64_mca_cpe_int_handler (int cpe_irq, void *arg, struct pt_regs *ptregs)
 
 #endif /* CONFIG_ACPI */
 
-static void
-show_min_state (pal_min_state_area_t *minstate)
-{
-	u64 iip = minstate->pmsa_iip + ((struct ia64_psr *)(&minstate->pmsa_ipsr))->ri;
-	u64 xip = minstate->pmsa_xip + ((struct ia64_psr *)(&minstate->pmsa_xpsr))->ri;
-
-	printk("NaT bits\t%016lx\n", minstate->pmsa_nat_bits);
-	printk("pr\t\t%016lx\n", minstate->pmsa_pr);
-	printk("b0\t\t%016lx ", minstate->pmsa_br0); print_symbol("%s\n", minstate->pmsa_br0);
-	printk("ar.rsc\t\t%016lx\n", minstate->pmsa_rsc);
-	printk("cr.iip\t\t%016lx ", iip); print_symbol("%s\n", iip);
-	printk("cr.ipsr\t\t%016lx\n", minstate->pmsa_ipsr);
-	printk("cr.ifs\t\t%016lx\n", minstate->pmsa_ifs);
-	printk("xip\t\t%016lx ", xip); print_symbol("%s\n", xip);
-	printk("xpsr\t\t%016lx\n", minstate->pmsa_xpsr);
-	printk("xfs\t\t%016lx\n", minstate->pmsa_xfs);
-	printk("b1\t\t%016lx ", minstate->pmsa_br1);
-	print_symbol("%s\n", minstate->pmsa_br1);
-
-	printk("\nstatic registers r0-r15:\n");
-	printk(" r0- 3 %016lx %016lx %016lx %016lx\n",
-	       0UL, minstate->pmsa_gr[0], minstate->pmsa_gr[1], minstate->pmsa_gr[2]);
-	printk(" r4- 7 %016lx %016lx %016lx %016lx\n",
-	       minstate->pmsa_gr[3], minstate->pmsa_gr[4],
-	       minstate->pmsa_gr[5], minstate->pmsa_gr[6]);
-	printk(" r8-11 %016lx %016lx %016lx %016lx\n",
-	       minstate->pmsa_gr[7], minstate->pmsa_gr[8],
-	       minstate->pmsa_gr[9], minstate->pmsa_gr[10]);
-	printk("r12-15 %016lx %016lx %016lx %016lx\n",
-	       minstate->pmsa_gr[11], minstate->pmsa_gr[12],
-	       minstate->pmsa_gr[13], minstate->pmsa_gr[14]);
-
-	printk("\nbank 0:\n");
-	printk("r16-19 %016lx %016lx %016lx %016lx\n",
-	       minstate->pmsa_bank0_gr[0], minstate->pmsa_bank0_gr[1],
-	       minstate->pmsa_bank0_gr[2], minstate->pmsa_bank0_gr[3]);
-	printk("r20-23 %016lx %016lx %016lx %016lx\n",
-	       minstate->pmsa_bank0_gr[4], minstate->pmsa_bank0_gr[5],
-	       minstate->pmsa_bank0_gr[6], minstate->pmsa_bank0_gr[7]);
-	printk("r24-27 %016lx %016lx %016lx %016lx\n",
-	       minstate->pmsa_bank0_gr[8], minstate->pmsa_bank0_gr[9],
-	       minstate->pmsa_bank0_gr[10], minstate->pmsa_bank0_gr[11]);
-	printk("r28-31 %016lx %016lx %016lx %016lx\n",
-	       minstate->pmsa_bank0_gr[12], minstate->pmsa_bank0_gr[13],
-	       minstate->pmsa_bank0_gr[14], minstate->pmsa_bank0_gr[15]);
-
-	printk("\nbank 1:\n");
-	printk("r16-19 %016lx %016lx %016lx %016lx\n",
-	       minstate->pmsa_bank1_gr[0], minstate->pmsa_bank1_gr[1],
-	       minstate->pmsa_bank1_gr[2], minstate->pmsa_bank1_gr[3]);
-	printk("r20-23 %016lx %016lx %016lx %016lx\n",
-	       minstate->pmsa_bank1_gr[4], minstate->pmsa_bank1_gr[5],
-	       minstate->pmsa_bank1_gr[6], minstate->pmsa_bank1_gr[7]);
-	printk("r24-27 %016lx %016lx %016lx %016lx\n",
-	       minstate->pmsa_bank1_gr[8], minstate->pmsa_bank1_gr[9],
-	       minstate->pmsa_bank1_gr[10], minstate->pmsa_bank1_gr[11]);
-	printk("r28-31 %016lx %016lx %016lx %016lx\n",
-	       minstate->pmsa_bank1_gr[12], minstate->pmsa_bank1_gr[13],
-	       minstate->pmsa_bank1_gr[14], minstate->pmsa_bank1_gr[15]);
-}
-
-static void
-fetch_min_state (pal_min_state_area_t *ms, struct pt_regs *pt, struct switch_stack *sw)
-{
-	u64 *dst_banked, *src_banked, bit, shift, nat_bits;
-	int i;
-
-	/*
-	 * First, update the pt-regs and switch-stack structures with the contents stored
-	 * in the min-state area:
-	 */
-	if (((struct ia64_psr *) &ms->pmsa_ipsr)->ic == 0) {
-		pt->cr_ipsr = ms->pmsa_xpsr;
-		pt->cr_iip = ms->pmsa_xip;
-		pt->cr_ifs = ms->pmsa_xfs;
-	} else {
-		pt->cr_ipsr = ms->pmsa_ipsr;
-		pt->cr_iip = ms->pmsa_iip;
-		pt->cr_ifs = ms->pmsa_ifs;
-	}
-	pt->ar_rsc = ms->pmsa_rsc;
-	pt->pr = ms->pmsa_pr;
-	pt->r1 = ms->pmsa_gr[0];
-	pt->r2 = ms->pmsa_gr[1];
-	pt->r3 = ms->pmsa_gr[2];
-	sw->r4 = ms->pmsa_gr[3];
-	sw->r5 = ms->pmsa_gr[4];
-	sw->r6 = ms->pmsa_gr[5];
-	sw->r7 = ms->pmsa_gr[6];
-	pt->r8 = ms->pmsa_gr[7];
-	pt->r9 = ms->pmsa_gr[8];
-	pt->r10 = ms->pmsa_gr[9];
-	pt->r11 = ms->pmsa_gr[10];
-	pt->r12 = ms->pmsa_gr[11];
-	pt->r13 = ms->pmsa_gr[12];
-	pt->r14 = ms->pmsa_gr[13];
-	pt->r15 = ms->pmsa_gr[14];
-	dst_banked = &pt->r16;		/* r16-r31 are contiguous in struct pt_regs */
-	src_banked = ms->pmsa_bank1_gr;
-	for (i = 0; i < 16; ++i)
-		dst_banked[i] = src_banked[i];
-	pt->b0 = ms->pmsa_br0;
-	sw->b1 = ms->pmsa_br1;
-
-	/* construct the NaT bits for the pt-regs structure: */
-#	define PUT_NAT_BIT(dst, addr)					\
-	do {								\
-		bit = nat_bits & 1; nat_bits >>= 1;			\
-		shift = ((unsigned long) addr >> 3) & 0x3f;		\
-		dst = ((dst) & ~(1UL << shift)) | (bit << shift);	\
-	} while (0)
-
-	/* Rotate the saved NaT bits such that bit 0 corresponds to pmsa_gr[0]: */
-	shift = ((unsigned long) &ms->pmsa_gr[0] >> 3) & 0x3f;
-	nat_bits = (ms->pmsa_nat_bits >> shift) | (ms->pmsa_nat_bits << (64 - shift));
-
-	PUT_NAT_BIT(sw->caller_unat, &pt->r1);
-	PUT_NAT_BIT(sw->caller_unat, &pt->r2);
-	PUT_NAT_BIT(sw->caller_unat, &pt->r3);
-	PUT_NAT_BIT(sw->ar_unat, &sw->r4);
-	PUT_NAT_BIT(sw->ar_unat, &sw->r5);
-	PUT_NAT_BIT(sw->ar_unat, &sw->r6);
-	PUT_NAT_BIT(sw->ar_unat, &sw->r7);
-	PUT_NAT_BIT(sw->caller_unat, &pt->r8);	PUT_NAT_BIT(sw->caller_unat, &pt->r9);
-	PUT_NAT_BIT(sw->caller_unat, &pt->r10);	PUT_NAT_BIT(sw->caller_unat, &pt->r11);
-	PUT_NAT_BIT(sw->caller_unat, &pt->r12);	PUT_NAT_BIT(sw->caller_unat, &pt->r13);
-	PUT_NAT_BIT(sw->caller_unat, &pt->r14);	PUT_NAT_BIT(sw->caller_unat, &pt->r15);
-	nat_bits >>= 16;	/* skip over bank0 NaT bits */
-	PUT_NAT_BIT(sw->caller_unat, &pt->r16);	PUT_NAT_BIT(sw->caller_unat, &pt->r17);
-	PUT_NAT_BIT(sw->caller_unat, &pt->r18);	PUT_NAT_BIT(sw->caller_unat, &pt->r19);
-	PUT_NAT_BIT(sw->caller_unat, &pt->r20);	PUT_NAT_BIT(sw->caller_unat, &pt->r21);
-	PUT_NAT_BIT(sw->caller_unat, &pt->r22);	PUT_NAT_BIT(sw->caller_unat, &pt->r23);
-	PUT_NAT_BIT(sw->caller_unat, &pt->r24);	PUT_NAT_BIT(sw->caller_unat, &pt->r25);
-	PUT_NAT_BIT(sw->caller_unat, &pt->r26);	PUT_NAT_BIT(sw->caller_unat, &pt->r27);
-	PUT_NAT_BIT(sw->caller_unat, &pt->r28);	PUT_NAT_BIT(sw->caller_unat, &pt->r29);
-	PUT_NAT_BIT(sw->caller_unat, &pt->r30);	PUT_NAT_BIT(sw->caller_unat, &pt->r31);
-}
-
-static void
-init_handler_platform (pal_min_state_area_t *ms,
-		       struct pt_regs *pt, struct switch_stack *sw)
-{
-	struct unw_frame_info info;
-
-	/* if a kernel debugger is available call it here else just dump the registers */
-
-	/*
-	 * Wait for a bit.  On some machines (e.g., HP's zx2000 and zx6000, INIT can be
-	 * generated via the BMC's command-line interface, but since the console is on the
-	 * same serial line, the user will need some time to switch out of the BMC before
-	 * the dump begins.
-	 */
-	printk("Delaying for 5 seconds...\n");
-	udelay(5*1000000);
-	show_min_state(ms);
-
-	printk("Backtrace of current task (pid %d, %s)\n", current->pid, current->comm);
-	fetch_min_state(ms, pt, sw);
-	unw_init_from_interruption(&info, current, pt, sw);
-	ia64_do_show_stack(&info, NULL);
-
-	if (read_trylock(&tasklist_lock)) {
-		struct task_struct *g, *t;
-		do_each_thread (g, t) {
-			if (t == current)
-				continue;
-
-			printk("\nBacktrace of pid %d (%s)\n", t->pid, t->comm);
-			show_stack(t, NULL);
-		} while_each_thread (g, t);
-	}
-
-	printk("\nINIT dump complete.  Please reboot now.\n");
-	while (1);			/* hang city if no debugger */
-}
-
 #ifdef CONFIG_ACPI
 /*
  * ia64_mca_register_cpev
@@ -648,42 +478,6 @@ ia64_mca_cmc_vector_enable_keventd(void *unused)
 }
 
 /*
- * ia64_mca_wakeup_ipi_wait
- *
- *	Wait for the inter-cpu interrupt to be sent by the
- *	monarch processor once it is done with handling the
- *	MCA.
- *
- *  Inputs  :   None
- *  Outputs :   None
- */
-static void
-ia64_mca_wakeup_ipi_wait(void)
-{
-	int	irr_num = (IA64_MCA_WAKEUP_VECTOR >> 6);
-	int	irr_bit = (IA64_MCA_WAKEUP_VECTOR & 0x3f);
-	u64	irr = 0;
-
-	do {
-		switch(irr_num) {
-		      case 0:
-			irr = ia64_getreg(_IA64_REG_CR_IRR0);
-			break;
-		      case 1:
-			irr = ia64_getreg(_IA64_REG_CR_IRR1);
-			break;
-		      case 2:
-			irr = ia64_getreg(_IA64_REG_CR_IRR2);
-			break;
-		      case 3:
-			irr = ia64_getreg(_IA64_REG_CR_IRR3);
-			break;
-		}
-		cpu_relax();
-	} while (!(irr & (1UL << irr_bit))) ;
-}
-
-/*
  * ia64_mca_wakeup
  *
  *	Send an inter-cpu interrupt to wake-up a particular cpu
@@ -748,11 +542,9 @@ ia64_mca_rendez_int_handler(int rendez_irq, void *arg, struct pt_regs *ptregs)
 	 */
 	ia64_sal_mc_rendez();
 
-	/* Wait for the wakeup IPI from the monarch
-	 * This waiting is done by polling on the wakeup-interrupt
-	 * vector bit in the processor's IRRs
-	 */
-	ia64_mca_wakeup_ipi_wait();
+	/* Wait for the monarch cpu to exit. */
+	while (monarch_cpu != -1)
+	       cpu_relax();	/* spin until monarch leaves */
 
 	/* Enable all interrupts */
 	local_irq_restore(flags);
@@ -780,53 +572,13 @@ ia64_mca_wakeup_int_handler(int wakeup_irq, void *arg, struct pt_regs *ptregs)
 	return IRQ_HANDLED;
 }
 
-/*
- * ia64_return_to_sal_check
- *
- *	This is function called before going back from the OS_MCA handler
- *	to the OS_MCA dispatch code which finally takes the control back
- *	to the SAL.
- *	The main purpose of this routine is to setup the OS_MCA to SAL
- *	return state which can be used by the OS_MCA dispatch code
- *	just before going back to SAL.
- *
- *  Inputs  :   None
- *  Outputs :   None
- */
-
-static void
-ia64_return_to_sal_check(int recover)
-{
-
-	/* Copy over some relevant stuff from the sal_to_os_mca_handoff
-	 * so that it can be used at the time of os_mca_to_sal_handoff
-	 */
-	ia64_os_to_sal_handoff_state.imots_sal_gp =
-		ia64_sal_to_os_handoff_state.imsto_sal_gp;
-
-	ia64_os_to_sal_handoff_state.imots_sal_check_ra =
-		ia64_sal_to_os_handoff_state.imsto_sal_check_ra;
-
-	if (recover)
-		ia64_os_to_sal_handoff_state.imots_os_status = IA64_MCA_CORRECTED;
-	else
-		ia64_os_to_sal_handoff_state.imots_os_status = IA64_MCA_COLD_BOOT;
-
-	/* Default = tell SAL to return to same context */
-	ia64_os_to_sal_handoff_state.imots_context = IA64_MCA_SAME_CONTEXT;
-
-	ia64_os_to_sal_handoff_state.imots_new_min_state =
-		(u64 *)ia64_sal_to_os_handoff_state.pal_min_state;
-
-}
-
 /* Function pointer for extra MCA recovery */
 int (*ia64_mca_ucmc_extension)
-	(void*,ia64_mca_sal_to_os_state_t*,ia64_mca_os_to_sal_state_t*)
+	(void*,struct ia64_sal_os_state*)
 	= NULL;
 
 int
-ia64_reg_MCA_extension(void *fn)
+ia64_reg_MCA_extension(int (*fn)(void *, struct ia64_sal_os_state *))
 {
 	if (ia64_mca_ucmc_extension)
 		return 1;
@@ -845,8 +597,321 @@ ia64_unreg_MCA_extension(void)
 EXPORT_SYMBOL(ia64_reg_MCA_extension);
 EXPORT_SYMBOL(ia64_unreg_MCA_extension);
 
+
+static inline void
+copy_reg(const u64 *fr, u64 fnat, u64 *tr, u64 *tnat)
+{
+	u64 fslot, tslot, nat;
+	*tr = *fr;
+	fslot = ((unsigned long)fr >> 3) & 63;
+	tslot = ((unsigned long)tr >> 3) & 63;
+	*tnat &= ~(1UL << tslot);
+	nat = (fnat >> fslot) & 1;
+	*tnat |= (nat << tslot);
+}
+
+/* On entry to this routine, we are running on the per cpu stack, see
+ * mca_asm.h.  The original stack has not been touched by this event.  Some of
+ * the original stack's registers will be in the RBS on this stack.  This stack
+ * also contains a partial pt_regs and switch_stack, the rest of the data is in
+ * PAL minstate.
+ *
+ * The first thing to do is modify the original stack to look like a blocked
+ * task so we can run backtrace on the original task.  Also mark the per cpu
+ * stack as current to ensure that we use the correct task state, it also means
+ * that we can do backtrace on the MCA/INIT handler code itself.
+ */
+
+static task_t *
+ia64_mca_modify_original_stack(struct pt_regs *regs,
+		const struct switch_stack *sw,
+		struct ia64_sal_os_state *sos,
+		const char *type)
+{
+	char *p, comm[sizeof(current->comm)];
+	ia64_va va;
+	extern char ia64_leave_kernel[];	/* Need asm address, not function descriptor */
+	const pal_min_state_area_t *ms = sos->pal_min_state;
+	task_t *previous_current;
+	struct pt_regs *old_regs;
+	struct switch_stack *old_sw;
+	unsigned size = sizeof(struct pt_regs) +
+			sizeof(struct switch_stack) + 16;
+	u64 *old_bspstore, *old_bsp;
+	u64 *new_bspstore, *new_bsp;
+	u64 old_unat, old_rnat, new_rnat, nat;
+	u64 slots, loadrs = regs->loadrs;
+	u64 r12 = ms->pmsa_gr[12-1], r13 = ms->pmsa_gr[13-1];
+	u64 ar_bspstore = regs->ar_bspstore;
+	u64 ar_bsp = regs->ar_bspstore + (loadrs >> 16);
+	const u64 *bank;
+	const char *msg;
+	int cpu = smp_processor_id();
+
+	previous_current = curr_task(cpu);
+	set_curr_task(cpu, current);
+	if ((p = strchr(current->comm, ' ')))
+		*p = '\0';
+
+	/* Best effort attempt to cope with MCA/INIT delivered while in
+	 * physical mode.
+	 */
+	regs->cr_ipsr = ms->pmsa_ipsr;
+	if (ia64_psr(regs)->dt == 0) {
+		va.l = r12;
+		if (va.f.reg == 0) {
+			va.f.reg = 7;
+			r12 = va.l;
+		}
+		va.l = r13;
+		if (va.f.reg == 0) {
+			va.f.reg = 7;
+			r13 = va.l;
+		}
+	}
+	if (ia64_psr(regs)->rt == 0) {
+		va.l = ar_bspstore;
+		if (va.f.reg == 0) {
+			va.f.reg = 7;
+			ar_bspstore = va.l;
+		}
+		va.l = ar_bsp;
+		if (va.f.reg == 0) {
+			va.f.reg = 7;
+			ar_bsp = va.l;
+		}
+	}
+
+	/* mca_asm.S ia64_old_stack() cannot assume that the dirty registers
+	 * have been copied to the old stack, the old stack may fail the
+	 * validation tests below.  So ia64_old_stack() must restore the dirty
+	 * registers from the new stack.  The old and new bspstore probably
+	 * have different alignments, so loadrs calculated on the old bsp
+	 * cannot be used to restore from the new bsp.  Calculate a suitable
+	 * loadrs for the new stack and save it in the new pt_regs, where
+	 * ia64_old_stack() can get it.
+	 */
+	old_bspstore = (u64 *)ar_bspstore;
+	old_bsp = (u64 *)ar_bsp;
+	slots = ia64_rse_num_regs(old_bspstore, old_bsp);
+	new_bspstore = (u64 *)((u64)current + IA64_RBS_OFFSET);
+	new_bsp = ia64_rse_skip_regs(new_bspstore, slots);
+	regs->loadrs = (new_bsp - new_bspstore) * 8 << 16;
+
+	/* Verify the previous stack state before we change it */
+	if (user_mode(regs)) {
+		msg = "occurred in user space";
+		goto no_mod;
+	}
+	if (r13 != sos->prev_IA64_KR_CURRENT) {
+		msg = "inconsistent previous current and r13";
+		goto no_mod;
+	}
+	if ((r12 - r13) >= KERNEL_STACK_SIZE) {
+		msg = "inconsistent r12 and r13";
+		goto no_mod;
+	}
+	if ((ar_bspstore - r13) >= KERNEL_STACK_SIZE) {
+		msg = "inconsistent ar.bspstore and r13";
+		goto no_mod;
+	}
+	va.p = old_bspstore;
+	if (va.f.reg < 5) {
+		msg = "old_bspstore is in the wrong region";
+		goto no_mod;
+	}
+	if ((ar_bsp - r13) >= KERNEL_STACK_SIZE) {
+		msg = "inconsistent ar.bsp and r13";
+		goto no_mod;
+	}
+	size += (ia64_rse_skip_regs(old_bspstore, slots) - old_bspstore) * 8;
+	if (ar_bspstore + size > r12) {
+		msg = "no room for blocked state";
+		goto no_mod;
+	}
+
+	/* Change the comm field on the MCA/INT task to include the pid that
+	 * was interrupted, it makes for easier debugging.  If that pid was 0
+	 * (swapper or nested MCA/INIT) then use the start of the previous comm
+	 * field suffixed with its cpu.
+	 */
+	if (previous_current->pid)
+		snprintf(comm, sizeof(comm), "%s %d",
+			current->comm, previous_current->pid);
+	else {
+		int l;
+		if ((p = strchr(previous_current->comm, ' ')))
+			l = p - previous_current->comm;
+		else
+			l = strlen(previous_current->comm);
+		snprintf(comm, sizeof(comm), "%s %*s %d",
+			current->comm, l, previous_current->comm,
+			previous_current->thread_info->cpu);
+	}
+	memcpy(current->comm, comm, sizeof(current->comm));
+
+	/* Make the original task look blocked.  First stack a struct pt_regs,
+	 * describing the state at the time of interrupt.  mca_asm.S built a
+	 * partial pt_regs, copy it and fill in the blanks using minstate.
+	 */
+	p = (char *)r12 - sizeof(*regs);
+	old_regs = (struct pt_regs *)p;
+	memcpy(old_regs, regs, sizeof(*regs));
+	/* If ipsr.ic then use pmsa_{iip,ipsr,ifs}, else use
+	 * pmsa_{xip,xpsr,xfs}
+	 */
+	if (ia64_psr(regs)->ic) {
+		old_regs->cr_iip = ms->pmsa_iip;
+		old_regs->cr_ipsr = ms->pmsa_ipsr;
+		old_regs->cr_ifs = ms->pmsa_ifs;
+	} else {
+		old_regs->cr_iip = ms->pmsa_xip;
+		old_regs->cr_ipsr = ms->pmsa_xpsr;
+		old_regs->cr_ifs = ms->pmsa_xfs;
+	}
+	old_regs->pr = ms->pmsa_pr;
+	old_regs->b0 = ms->pmsa_br0;
+	old_regs->loadrs = loadrs;
+	old_regs->ar_rsc = ms->pmsa_rsc;
+	old_unat = old_regs->ar_unat;
+	copy_reg(&ms->pmsa_gr[1-1], ms->pmsa_nat_bits, &old_regs->r1, &old_unat);
+	copy_reg(&ms->pmsa_gr[2-1], ms->pmsa_nat_bits, &old_regs->r2, &old_unat);
+	copy_reg(&ms->pmsa_gr[3-1], ms->pmsa_nat_bits, &old_regs->r3, &old_unat);
+	copy_reg(&ms->pmsa_gr[8-1], ms->pmsa_nat_bits, &old_regs->r8, &old_unat);
+	copy_reg(&ms->pmsa_gr[9-1], ms->pmsa_nat_bits, &old_regs->r9, &old_unat);
+	copy_reg(&ms->pmsa_gr[10-1], ms->pmsa_nat_bits, &old_regs->r10, &old_unat);
+	copy_reg(&ms->pmsa_gr[11-1], ms->pmsa_nat_bits, &old_regs->r11, &old_unat);
+	copy_reg(&ms->pmsa_gr[12-1], ms->pmsa_nat_bits, &old_regs->r12, &old_unat);
+	copy_reg(&ms->pmsa_gr[13-1], ms->pmsa_nat_bits, &old_regs->r13, &old_unat);
+	copy_reg(&ms->pmsa_gr[14-1], ms->pmsa_nat_bits, &old_regs->r14, &old_unat);
+	copy_reg(&ms->pmsa_gr[15-1], ms->pmsa_nat_bits, &old_regs->r15, &old_unat);
+	if (ia64_psr(old_regs)->bn)
+		bank = ms->pmsa_bank1_gr;
+	else
+		bank = ms->pmsa_bank0_gr;
+	copy_reg(&bank[16-16], ms->pmsa_nat_bits, &old_regs->r16, &old_unat);
+	copy_reg(&bank[17-16], ms->pmsa_nat_bits, &old_regs->r17, &old_unat);
+	copy_reg(&bank[18-16], ms->pmsa_nat_bits, &old_regs->r18, &old_unat);
+	copy_reg(&bank[19-16], ms->pmsa_nat_bits, &old_regs->r19, &old_unat);
+	copy_reg(&bank[20-16], ms->pmsa_nat_bits, &old_regs->r20, &old_unat);
+	copy_reg(&bank[21-16], ms->pmsa_nat_bits, &old_regs->r21, &old_unat);
+	copy_reg(&bank[22-16], ms->pmsa_nat_bits, &old_regs->r22, &old_unat);
+	copy_reg(&bank[23-16], ms->pmsa_nat_bits, &old_regs->r23, &old_unat);
+	copy_reg(&bank[24-16], ms->pmsa_nat_bits, &old_regs->r24, &old_unat);
+	copy_reg(&bank[25-16], ms->pmsa_nat_bits, &old_regs->r25, &old_unat);
+	copy_reg(&bank[26-16], ms->pmsa_nat_bits, &old_regs->r26, &old_unat);
+	copy_reg(&bank[27-16], ms->pmsa_nat_bits, &old_regs->r27, &old_unat);
+	copy_reg(&bank[28-16], ms->pmsa_nat_bits, &old_regs->r28, &old_unat);
+	copy_reg(&bank[29-16], ms->pmsa_nat_bits, &old_regs->r29, &old_unat);
+	copy_reg(&bank[30-16], ms->pmsa_nat_bits, &old_regs->r30, &old_unat);
+	copy_reg(&bank[31-16], ms->pmsa_nat_bits, &old_regs->r31, &old_unat);
+
+	/* Next stack a struct switch_stack.  mca_asm.S built a partial
+	 * switch_stack, copy it and fill in the blanks using pt_regs and
+	 * minstate.
+	 *
+	 * In the synthesized switch_stack, b0 points to ia64_leave_kernel,
+	 * ar.pfs is set to 0.
+	 *
+	 * unwind.c::unw_unwind() does special processing for interrupt frames.
+	 * It checks if the PRED_NON_SYSCALL predicate is set, if the predicate
+	 * is clear then unw_unwind() does _not_ adjust bsp over pt_regs.  Not
+	 * that this is documented, of course.  Set PRED_NON_SYSCALL in the
+	 * switch_stack on the original stack so it will unwind correctly when
+	 * unwind.c reads pt_regs.
+	 *
+	 * thread.ksp is updated to point to the synthesized switch_stack.
+	 */
+	p -= sizeof(struct switch_stack);
+	old_sw = (struct switch_stack *)p;
+	memcpy(old_sw, sw, sizeof(*sw));
+	old_sw->caller_unat = old_unat;
+	old_sw->ar_fpsr = old_regs->ar_fpsr;
+	copy_reg(&ms->pmsa_gr[4-1], ms->pmsa_nat_bits, &old_sw->r4, &old_unat);
+	copy_reg(&ms->pmsa_gr[5-1], ms->pmsa_nat_bits, &old_sw->r5, &old_unat);
+	copy_reg(&ms->pmsa_gr[6-1], ms->pmsa_nat_bits, &old_sw->r6, &old_unat);
+	copy_reg(&ms->pmsa_gr[7-1], ms->pmsa_nat_bits, &old_sw->r7, &old_unat);
+	old_sw->b0 = (u64)ia64_leave_kernel;
+	old_sw->b1 = ms->pmsa_br1;
+	old_sw->ar_pfs = 0;
+	old_sw->ar_unat = old_unat;
+	old_sw->pr = old_regs->pr | (1UL << PRED_NON_SYSCALL);
+	previous_current->thread.ksp = (u64)p - 16;
+
+	/* Finally copy the original stack's registers back to its RBS.
+	 * Registers from ar.bspstore through ar.bsp at the time of the event
+	 * are in the current RBS, copy them back to the original stack.  The
+	 * copy must be done register by register because the original bspstore
+	 * and the current one have different alignments, so the saved RNAT
+	 * data occurs at different places.
+	 *
+	 * mca_asm does cover, so the old_bsp already includes all registers at
+	 * the time of MCA/INIT.  It also does flushrs, so all registers before
+	 * this function have been written to backing store on the MCA/INIT
+	 * stack.
+	 */
+	new_rnat = ia64_get_rnat(ia64_rse_rnat_addr(new_bspstore));
+	old_rnat = regs->ar_rnat;
+	while (slots--) {
+		if (ia64_rse_is_rnat_slot(new_bspstore)) {
+			new_rnat = ia64_get_rnat(new_bspstore++);
+		}
+		if (ia64_rse_is_rnat_slot(old_bspstore)) {
+			*old_bspstore++ = old_rnat;
+			old_rnat = 0;
+		}
+		nat = (new_rnat >> ia64_rse_slot_num(new_bspstore)) & 1UL;
+		old_rnat &= ~(1UL << ia64_rse_slot_num(old_bspstore));
+		old_rnat |= (nat << ia64_rse_slot_num(old_bspstore));
+		*old_bspstore++ = *new_bspstore++;
+	}
+	old_sw->ar_bspstore = (unsigned long)old_bspstore;
+	old_sw->ar_rnat = old_rnat;
+
+	sos->prev_task = previous_current;
+	return previous_current;
+
+no_mod:
+	printk(KERN_INFO "cpu %d, %s %s, original stack not modified\n",
+			smp_processor_id(), type, msg);
+	return previous_current;
+}
+
+/* The monarch/slave interaction is based on monarch_cpu and requires that all
+ * slaves have entered rendezvous before the monarch leaves.  If any cpu has
+ * not entered rendezvous yet then wait a bit.  The assumption is that any
+ * slave that has not rendezvoused after a reasonable time is never going to do
+ * so.  In this context, slave includes cpus that respond to the MCA rendezvous
+ * interrupt, as well as cpus that receive the INIT slave event.
+ */
+
+static void
+ia64_wait_for_slaves(int monarch)
+{
+	int c, wait = 0;
+	for_each_online_cpu(c) {
+		if (c == monarch)
+			continue;
+		if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE) {
+			udelay(1000);		/* short wait first */
+			wait = 1;
+			break;
+		}
+	}
+	if (!wait)
+		return;
+	for_each_online_cpu(c) {
+		if (c == monarch)
+			continue;
+		if (ia64_mc_info.imi_rendez_checkin[c] == IA64_MCA_RENDEZ_CHECKIN_NOTDONE) {
+			udelay(5*1000000);	/* wait 5 seconds for slaves (arbitrary) */
+			break;
+		}
+	}
+}
+
 /*
- * ia64_mca_ucmc_handler
+ * ia64_mca_handler
  *
  *	This is uncorrectable machine check handler called from OS_MCA
  *	dispatch code which is in turn called from SAL_CHECK().
@@ -857,16 +922,28 @@ EXPORT_SYMBOL(ia64_unreg_MCA_extension);
  *	further MCA logging is enabled by clearing logs.
  *	Monarch also has the duty of sending wakeup-IPIs to pull the
  *	slave processors out of rendezvous spinloop.
- *
- *  Inputs  :   None
- *  Outputs :   None
  */
 void
-ia64_mca_ucmc_handler(void)
+ia64_mca_handler(struct pt_regs *regs, struct switch_stack *sw,
+		 struct ia64_sal_os_state *sos)
 {
 	pal_processor_state_info_t *psp = (pal_processor_state_info_t *)
-		&ia64_sal_to_os_handoff_state.proc_state_param;
-	int recover; 
+		&sos->proc_state_param;
+	int recover, cpu = smp_processor_id();
+	task_t *previous_current;
+
+	oops_in_progress = 1;	/* FIXME: make printk NMI/MCA/INIT safe */
+	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "MCA");
+	monarch_cpu = cpu;
+	ia64_wait_for_slaves(cpu);
+
+	/* Wakeup all the processors which are spinning in the rendezvous loop.
+	 * They will leave SAL, then spin in the OS with interrupts disabled
+	 * until this monarch cpu leaves the MCA handler.  That gets control
+	 * back to the OS so we can backtrace the other cpus, backtrace when
+	 * spinning in SAL does not work.
+	 */
+	ia64_mca_wakeup_all();
 
 	/* Get the MCA error record and log it */
 	ia64_mca_log_sal_error_record(SAL_INFO_TYPE_MCA);
@@ -874,25 +951,20 @@ ia64_mca_ucmc_handler(void)
 	/* TLB error is only exist in this SAL error record */
 	recover = (psp->tc && !(psp->cc || psp->bc || psp->rc || psp->uc))
 	/* other error recovery */
-	   || (ia64_mca_ucmc_extension 
+	   || (ia64_mca_ucmc_extension
 		&& ia64_mca_ucmc_extension(
 			IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA),
-			&ia64_sal_to_os_handoff_state,
-			&ia64_os_to_sal_handoff_state)); 
+			sos));
 
 	if (recover) {
 		sal_log_record_header_t *rh = IA64_LOG_CURR_BUFFER(SAL_INFO_TYPE_MCA);
 		rh->severity = sal_log_severity_corrected;
 		ia64_sal_clear_state_info(SAL_INFO_TYPE_MCA);
+		sos->os_status = IA64_MCA_CORRECTED;
 	}
-	/*
-	 *  Wakeup all the processors which are spinning in the rendezvous
-	 *  loop.
-	 */
-	ia64_mca_wakeup_all();
 
-	/* Return to SAL */
-	ia64_return_to_sal_check(recover);
+	set_curr_task(cpu, previous_current);
+	monarch_cpu = -1;
 }
 
 static DECLARE_WORK(cmc_disable_work, ia64_mca_cmc_vector_disable_keventd, NULL);
@@ -1116,34 +1188,114 @@ ia64_mca_cpe_poll (unsigned long dummy)
 /*
  * C portion of the OS INIT handler
  *
- * Called from ia64_monarch_init_handler
- *
- * Inputs: pointer to pt_regs where processor info was saved.
+ * Called from ia64_os_init_dispatch
  *
- * Returns:
- *   0 if SAL must warm boot the System
- *   1 if SAL must return to interrupted context using PAL_MC_RESUME
+ * Inputs: pointer to pt_regs where processor info was saved.  SAL/OS state for
+ * this event.  This code is used for both monarch and slave INIT events, see
+ * sos->monarch.
  *
+ * All INIT events switch to the INIT stack and change the previous process to
+ * blocked status.  If one of the INIT events is the monarch then we are
+ * probably processing the nmi button/command.  Use the monarch cpu to dump all
+ * the processes.  The slave INIT events all spin until the monarch cpu
+ * returns.  We can also get INIT slave events for MCA, in which case the MCA
+ * process is the monarch.
  */
+
 void
-ia64_init_handler (struct pt_regs *pt, struct switch_stack *sw)
+ia64_init_handler(struct pt_regs *regs, struct switch_stack *sw,
+		  struct ia64_sal_os_state *sos)
 {
-	pal_min_state_area_t *ms;
+	static atomic_t slaves;
+	static atomic_t monarchs;
+	task_t *previous_current;
+	int cpu = smp_processor_id(), c;
+	struct task_struct *g, *t;
 
-	oops_in_progress = 1;	/* avoid deadlock in printk, but it makes recovery dodgy */
+	oops_in_progress = 1;	/* FIXME: make printk NMI/MCA/INIT safe */
 	console_loglevel = 15;	/* make sure printks make it to console */
 
-	printk(KERN_INFO "Entered OS INIT handler. PSP=%lx\n",
-		ia64_sal_to_os_handoff_state.proc_state_param);
+	printk(KERN_INFO "Entered OS INIT handler. PSP=%lx cpu=%d monarch=%ld\n",
+		sos->proc_state_param, cpu, sos->monarch);
+	salinfo_log_wakeup(SAL_INFO_TYPE_INIT, NULL, 0, 0);
 
-	/*
-	 * Address of minstate area provided by PAL is physical,
-	 * uncacheable (bit 63 set). Convert to Linux virtual
-	 * address in region 6.
+	previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "INIT");
+	sos->os_status = IA64_INIT_RESUME;
+
+	/* FIXME: Workaround for broken proms that drive all INIT events as
+	 * slaves.  The last slave that enters is promoted to be a monarch.
+	 * Remove this code in September 2006, that gives platforms a year to
+	 * fix their proms and get their customers updated.
 	 */
-	ms = (pal_min_state_area_t *)(ia64_sal_to_os_handoff_state.pal_min_state | (6ul<<61));
+	if (!sos->monarch && atomic_add_return(1, &slaves) == num_online_cpus()) {
+		printk(KERN_WARNING "%s: Promoting cpu %d to monarch.\n",
+		       __FUNCTION__, cpu);
+		atomic_dec(&slaves);
+		sos->monarch = 1;
+	}
 
-	init_handler_platform(ms, pt, sw);	/* call platform specific routines */
+	/* FIXME: Workaround for broken proms that drive all INIT events as
+	 * monarchs.  Second and subsequent monarchs are demoted to slaves.
+	 * Remove this code in September 2006, that gives platforms a year to
+	 * fix their proms and get their customers updated.
+	 */
+	if (sos->monarch && atomic_add_return(1, &monarchs) > 1) {
+		printk(KERN_WARNING "%s: Demoting cpu %d to slave.\n",
+			       __FUNCTION__, cpu);
+		atomic_dec(&monarchs);
+		sos->monarch = 0;
+	}
+
+	if (!sos->monarch) {
+		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_INIT;
+		while (monarch_cpu == -1)
+		       cpu_relax();	/* spin until monarch enters */
+		while (monarch_cpu != -1)
+		       cpu_relax();	/* spin until monarch leaves */
+		printk("Slave on cpu %d returning to normal service.\n", cpu);
+		set_curr_task(cpu, previous_current);
+		ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE;
+		atomic_dec(&slaves);
+		return;
+	}
+
+	monarch_cpu = cpu;
+
+	/*
+	 * Wait for a bit.  On some machines (e.g., HP's zx2000 and zx6000, INIT can be
+	 * generated via the BMC's command-line interface, but since the console is on the
+	 * same serial line, the user will need some time to switch out of the BMC before
+	 * the dump begins.
+	 */
+	printk("Delaying for 5 seconds...\n");
+	udelay(5*1000000);
+	ia64_wait_for_slaves(cpu);
+	printk(KERN_ERR "Processes interrupted by INIT -");
+	for_each_online_cpu(c) {
+		struct ia64_sal_os_state *s;
+		t = __va(__per_cpu_mca[c] + IA64_MCA_CPU_INIT_STACK_OFFSET);
+		s = (struct ia64_sal_os_state *)((char *)t + MCA_SOS_OFFSET);
+		g = s->prev_task;
+		if (g) {
+			if (g->pid)
+				printk(" %d", g->pid);
+			else
+				printk(" %d (cpu %d task 0x%p)", g->pid, task_cpu(g), g);
+		}
+	}
+	printk("\n\n");
+	if (read_trylock(&tasklist_lock)) {
+		do_each_thread (g, t) {
+			printk("\nBacktrace of pid %d (%s)\n", t->pid, t->comm);
+			show_stack(t, NULL);
+		} while_each_thread (g, t);
+		read_unlock(&tasklist_lock);
+	}
+	printk("\nINIT dump complete.  Monarch on cpu %d returning to normal service.\n", cpu);
+	atomic_dec(&monarchs);
+	set_curr_task(cpu, previous_current);
+	monarch_cpu = -1;
+	return;
 }
 
 static int __init
@@ -1193,6 +1345,34 @@ static struct irqaction mca_cpep_irqaction = {
 };
 #endif /* CONFIG_ACPI */
 
+/* Minimal format of the MCA/INIT stacks.  The pseudo processes that run on
+ * these stacks can never sleep, they cannot return from the kernel to user
+ * space, they do not appear in a normal ps listing.  So there is no need to
+ * format most of the fields.
+ */
+
+static void
+format_mca_init_stack(void *mca_data, unsigned long offset,
+		const char *type, int cpu)
+{
+	struct task_struct *p = (struct task_struct *)((char *)mca_data + offset);
+	struct thread_info *ti;
+	memset(p, 0, KERNEL_STACK_SIZE);
+	ti = (struct thread_info *)((char *)p + IA64_TASK_SIZE);
+	ti->flags = _TIF_MCA_INIT;
+	ti->preempt_count = 1;
+	ti->task = p;
+	ti->cpu = cpu;
+	p->thread_info = ti;
+	p->state = TASK_UNINTERRUPTIBLE;
+	__set_bit(cpu, &p->cpus_allowed);
+	INIT_LIST_HEAD(&p->tasks);
+	p->parent = p->real_parent = p->group_leader = p;
+	INIT_LIST_HEAD(&p->children);
+	INIT_LIST_HEAD(&p->sibling);
+	strncpy(p->comm, type, sizeof(p->comm)-1);
+}
+
 /* Do per-CPU MCA-related initialization.  */
 
 void __devinit
@@ -1205,19 +1385,28 @@ ia64_mca_cpu_init(void *cpu_data)
 		int cpu;
 
 		mca_data = alloc_bootmem(sizeof(struct ia64_mca_cpu)
-					 * NR_CPUS);
+					 * NR_CPUS + KERNEL_STACK_SIZE);
+		mca_data = (void *)(((unsigned long)mca_data +
+					KERNEL_STACK_SIZE - 1) &
+				(-KERNEL_STACK_SIZE));
 		for (cpu = 0; cpu < NR_CPUS; cpu++) {
+			format_mca_init_stack(mca_data,
+					offsetof(struct ia64_mca_cpu, mca_stack),
+					"MCA", cpu);
+			format_mca_init_stack(mca_data,
+					offsetof(struct ia64_mca_cpu, init_stack),
+					"INIT", cpu);
 			__per_cpu_mca[cpu] = __pa(mca_data);
 			mca_data += sizeof(struct ia64_mca_cpu);
 		}
 	}
 
-        /*
-         * The MCA info structure was allocated earlier and its
-         * physical address saved in __per_cpu_mca[cpu].  Copy that
-         * address * to ia64_mca_data so we can access it as a per-CPU
-         * variable.
-         */
+	/*
+	 * The MCA info structure was allocated earlier and its
+	 * physical address saved in __per_cpu_mca[cpu].  Copy that
+	 * address * to ia64_mca_data so we can access it as a per-CPU
+	 * variable.
+	 */
 	__get_cpu_var(ia64_mca_data) = __per_cpu_mca[smp_processor_id()];
 
 	/*
@@ -1227,11 +1416,11 @@ ia64_mca_cpu_init(void *cpu_data)
 	__get_cpu_var(ia64_mca_per_cpu_pte) =
 		pte_val(mk_pte_phys(__pa(cpu_data), PAGE_KERNEL));
 
-        /*
-         * Also, stash away a copy of the PAL address and the PTE
-         * needed to map it.
-         */
-        pal_vaddr = efi_get_pal_addr();
+	/*
+	 * Also, stash away a copy of the PAL address and the PTE
+	 * needed to map it.
+	 */
+	pal_vaddr = efi_get_pal_addr();
 	if (!pal_vaddr)
 		return;
 	__get_cpu_var(ia64_mca_pal_base) =
@@ -1263,8 +1452,8 @@ ia64_mca_cpu_init(void *cpu_data)
 void __init
 ia64_mca_init(void)
 {
-	ia64_fptr_t *mon_init_ptr = (ia64_fptr_t *)ia64_monarch_init_handler;
-	ia64_fptr_t *slave_init_ptr = (ia64_fptr_t *)ia64_slave_init_handler;
+	ia64_fptr_t *init_hldlr_ptr_monarch = (ia64_fptr_t *)ia64_os_init_dispatch_monarch;
+	ia64_fptr_t *init_hldlr_ptr_slave = (ia64_fptr_t *)ia64_os_init_dispatch_slave;
 	ia64_fptr_t *mca_hldlr_ptr = (ia64_fptr_t *)ia64_os_mca_dispatch;
 	int i;
 	s64 rc;
@@ -1342,9 +1531,9 @@ ia64_mca_init(void)
 	 * XXX - disable SAL checksum by setting size to 0, should be
 	 * size of the actual init handler in mca_asm.S.
 	 */
-	ia64_mc_info.imi_monarch_init_handler		= ia64_tpa(mon_init_ptr->fp);
+	ia64_mc_info.imi_monarch_init_handler		= ia64_tpa(init_hldlr_ptr_monarch->fp);
 	ia64_mc_info.imi_monarch_init_handler_size	= 0;
-	ia64_mc_info.imi_slave_init_handler		= ia64_tpa(slave_init_ptr->fp);
+	ia64_mc_info.imi_slave_init_handler		= ia64_tpa(init_hldlr_ptr_slave->fp);
 	ia64_mc_info.imi_slave_init_handler_size	= 0;
 
 	IA64_MCA_DEBUG("%s: OS INIT handler at %lx\n", __FUNCTION__,
diff --git a/arch/ia64/kernel/mca_asm.S b/arch/ia64/kernel/mca_asm.S
index ef3fd7265b67..499a065f4e60 100644
--- a/arch/ia64/kernel/mca_asm.S
+++ b/arch/ia64/kernel/mca_asm.S
@@ -16,6 +16,9 @@
 // 04/11/12 Russ Anderson <rja@sgi.com>
 //		   Added per cpu MCA/INIT stack save areas.
 //
+// 12/08/05 Keith Owens <kaos@sgi.com>
+//		   Use per cpu MCA/INIT stacks for all data.
+//
 #include <linux/config.h>
 #include <linux/threads.h>
 
@@ -25,96 +28,23 @@
 #include <asm/mca_asm.h>
 #include <asm/mca.h>
 
-/*
- * When we get a machine check, the kernel stack pointer is no longer
- * valid, so we need to set a new stack pointer.
- */
-#define	MINSTATE_PHYS	/* Make sure stack access is physical for MINSTATE */
-
-/*
- * Needed for return context to SAL
- */
-#define IA64_MCA_SAME_CONTEXT	0
-#define IA64_MCA_COLD_BOOT	-2
-
-#include "minstate.h"
-
-/*
- * SAL_TO_OS_MCA_HANDOFF_STATE (SAL 3.0 spec)
- *		1. GR1 = OS GP
- *		2. GR8 = PAL_PROC physical address
- *		3. GR9 = SAL_PROC physical address
- *		4. GR10 = SAL GP (physical)
- *		5. GR11 = Rendez state
- *		6. GR12 = Return address to location within SAL_CHECK
- */
-#define SAL_TO_OS_MCA_HANDOFF_STATE_SAVE(_tmp)		\
-	LOAD_PHYSICAL(p0, _tmp, ia64_sal_to_os_handoff_state);; \
-	st8	[_tmp]=r1,0x08;;			\
-	st8	[_tmp]=r8,0x08;;			\
-	st8	[_tmp]=r9,0x08;;			\
-	st8	[_tmp]=r10,0x08;;			\
-	st8	[_tmp]=r11,0x08;;			\
-	st8	[_tmp]=r12,0x08;;			\
-	st8	[_tmp]=r17,0x08;;			\
-	st8	[_tmp]=r18,0x08
-
-/*
- * OS_MCA_TO_SAL_HANDOFF_STATE (SAL 3.0 spec)
- * (p6) is executed if we never entered virtual mode (TLB error)
- * (p7) is executed if we entered virtual mode as expected (normal case)
- *	1. GR8 = OS_MCA return status
- *	2. GR9 = SAL GP (physical)
- *	3. GR10 = 0/1 returning same/new context
- *	4. GR22 = New min state save area pointer
- *	returns ptr to SAL rtn save loc in _tmp
- */
-#define OS_MCA_TO_SAL_HANDOFF_STATE_RESTORE(_tmp)	\
-	movl	_tmp=ia64_os_to_sal_handoff_state;;	\
-	DATA_VA_TO_PA(_tmp);;				\
-	ld8	r8=[_tmp],0x08;;			\
-	ld8	r9=[_tmp],0x08;;			\
-	ld8	r10=[_tmp],0x08;;			\
-	ld8	r22=[_tmp],0x08;;
-	// now _tmp is pointing to SAL rtn save location
-
-/*
- * COLD_BOOT_HANDOFF_STATE() sets ia64_mca_os_to_sal_state
- *	imots_os_status=IA64_MCA_COLD_BOOT
- *	imots_sal_gp=SAL GP
- *	imots_context=IA64_MCA_SAME_CONTEXT
- *	imots_new_min_state=Min state save area pointer
- *	imots_sal_check_ra=Return address to location within SAL_CHECK
- *
- */
-#define COLD_BOOT_HANDOFF_STATE(sal_to_os_handoff,os_to_sal_handoff,tmp)\
-	movl	tmp=IA64_MCA_COLD_BOOT;					\
-	movl	sal_to_os_handoff=__pa(ia64_sal_to_os_handoff_state);	\
-	movl	os_to_sal_handoff=__pa(ia64_os_to_sal_handoff_state);;	\
-	st8	[os_to_sal_handoff]=tmp,8;;				\
-	ld8	tmp=[sal_to_os_handoff],48;;				\
-	st8	[os_to_sal_handoff]=tmp,8;;				\
-	movl	tmp=IA64_MCA_SAME_CONTEXT;;				\
-	st8	[os_to_sal_handoff]=tmp,8;;				\
-	ld8	tmp=[sal_to_os_handoff],-8;;				\
-	st8     [os_to_sal_handoff]=tmp,8;;				\
-	ld8	tmp=[sal_to_os_handoff];;				\
-	st8     [os_to_sal_handoff]=tmp;;
+#include "entry.h"
 
 #define GET_IA64_MCA_DATA(reg)						\
 	GET_THIS_PADDR(reg, ia64_mca_data)				\
 	;;								\
 	ld8 reg=[reg]
 
-	.global ia64_os_mca_dispatch
-	.global ia64_os_mca_dispatch_end
-	.global ia64_sal_to_os_handoff_state
-	.global	ia64_os_to_sal_handoff_state
 	.global ia64_do_tlb_purge
+	.global ia64_os_mca_dispatch
+	.global ia64_os_init_dispatch_monarch
+	.global ia64_os_init_dispatch_slave
 
 	.text
 	.align 16
 
+//StartMain////////////////////////////////////////////////////////////////////
+
 /*
  * Just the TLB purge part is moved to a separate function
  * so we can re-use the code for cpu hotplug code as well
@@ -207,34 +137,31 @@ ia64_do_tlb_purge:
 	br.sptk.many b1
 	;;
 
-ia64_os_mca_dispatch:
+//EndMain//////////////////////////////////////////////////////////////////////
+
+//StartMain////////////////////////////////////////////////////////////////////
 
+ia64_os_mca_dispatch:
 	// Serialize all MCA processing
 	mov	r3=1;;
 	LOAD_PHYSICAL(p0,r2,ia64_mca_serialize);;
 ia64_os_mca_spin:
-	xchg8	r4=[r2],r3;;
+	xchg4	r4=[r2],r3;;
 	cmp.ne	p6,p0=r4,r0
 (p6)	br ia64_os_mca_spin
 
-	// Save the SAL to OS MCA handoff state as defined
-	// by SAL SPEC 3.0
-	// NOTE : The order in which the state gets saved
-	//	  is dependent on the way the C-structure
-	//	  for ia64_mca_sal_to_os_state_t has been
-	//	  defined in include/asm/mca.h
-	SAL_TO_OS_MCA_HANDOFF_STATE_SAVE(r2)
-	;;
-
-	// LOG PROCESSOR STATE INFO FROM HERE ON..
-begin_os_mca_dump:
-	br	ia64_os_mca_proc_state_dump;;
+	mov r3=IA64_MCA_CPU_MCA_STACK_OFFSET	// use the MCA stack
+	LOAD_PHYSICAL(p0,r2,1f)			// return address
+	mov r19=1				// All MCA events are treated as monarch (for now)
+	br.sptk ia64_state_save			// save the state that is not in minstate
+1:
 
-ia64_os_mca_done_dump:
-
-	LOAD_PHYSICAL(p0,r16,ia64_sal_to_os_handoff_state+56)
+	GET_IA64_MCA_DATA(r2)
+	// Using MCA stack, struct ia64_sal_os_state, variable proc_state_param
+	;;
+	add r3=IA64_MCA_CPU_MCA_STACK_OFFSET+MCA_SOS_OFFSET+IA64_SAL_OS_STATE_PROC_STATE_PARAM_OFFSET, r2
 	;;
-	ld8 r18=[r16]		// Get processor state parameter on existing PALE_CHECK.
+	ld8 r18=[r3]				// Get processor state parameter on existing PALE_CHECK.
 	;;
 	tbit.nz p6,p7=r18,60
 (p7)	br.spnt done_tlb_purge_and_reload
@@ -323,624 +250,775 @@ ia64_reload_tr:
 	itr.d dtr[r20]=r16
 	;;
 	srlz.d
-	;;
-	br.sptk.many done_tlb_purge_and_reload
-err:
-	COLD_BOOT_HANDOFF_STATE(r20,r21,r22)
-	br.sptk.many ia64_os_mca_done_restore
 
 done_tlb_purge_and_reload:
 
-	// Setup new stack frame for OS_MCA handling
-	GET_IA64_MCA_DATA(r2)
-	;;
-	add r3 = IA64_MCA_CPU_STACKFRAME_OFFSET, r2
-	add r2 = IA64_MCA_CPU_RBSTORE_OFFSET, r2
-	;;
-	rse_switch_context(r6,r3,r2);;	// RSC management in this new context
+	// switch to per cpu MCA stack
+	mov r3=IA64_MCA_CPU_MCA_STACK_OFFSET	// use the MCA stack
+	LOAD_PHYSICAL(p0,r2,1f)			// return address
+	br.sptk ia64_new_stack
+1:
+
+	// everything saved, now we can set the kernel registers
+	mov r3=IA64_MCA_CPU_MCA_STACK_OFFSET	// use the MCA stack
+	LOAD_PHYSICAL(p0,r2,1f)			// return address
+	br.sptk ia64_set_kernel_registers
+1:
 
+	// This must be done in physical mode
 	GET_IA64_MCA_DATA(r2)
 	;;
-	add r2 = IA64_MCA_CPU_STACK_OFFSET+IA64_MCA_STACK_SIZE-16, r2
-	;;
-	mov r12=r2		// establish new stack-pointer
+	mov r7=r2
 
         // Enter virtual mode from physical mode
 	VIRTUAL_MODE_ENTER(r2, r3, ia64_os_mca_virtual_begin, r4)
-ia64_os_mca_virtual_begin:
+
+	// This code returns to SAL via SOS r2, in general SAL has no unwind
+	// data.  To get a clean termination when backtracing the C MCA/INIT
+	// handler, set a dummy return address of 0 in this routine.  That
+	// requires that ia64_os_mca_virtual_begin be a global function.
+ENTRY(ia64_os_mca_virtual_begin)
+	.prologue
+	.save rp,r0
+	.body
+
+	mov ar.rsc=3				// set eager mode for C handler
+	mov r2=r7				// see GET_IA64_MCA_DATA above
+	;;
 
 	// Call virtual mode handler
-	movl		r2=ia64_mca_ucmc_handler;;
-	mov		b6=r2;;
-	br.call.sptk.many    b0=b6;;
-.ret0:
+	alloc r14=ar.pfs,0,0,3,0
+	;;
+	DATA_PA_TO_VA(r2,r7)
+	;;
+	add out0=IA64_MCA_CPU_MCA_STACK_OFFSET+MCA_PT_REGS_OFFSET, r2
+	add out1=IA64_MCA_CPU_MCA_STACK_OFFSET+MCA_SWITCH_STACK_OFFSET, r2
+	add out2=IA64_MCA_CPU_MCA_STACK_OFFSET+MCA_SOS_OFFSET, r2
+	br.call.sptk.many    b0=ia64_mca_handler
+
 	// Revert back to physical mode before going back to SAL
 	PHYSICAL_MODE_ENTER(r2, r3, ia64_os_mca_virtual_end, r4)
 ia64_os_mca_virtual_end:
 
-	// restore the original stack frame here
+END(ia64_os_mca_virtual_begin)
+
+	// switch back to previous stack
+	alloc r14=ar.pfs,0,0,0,0		// remove the MCA handler frame
+	mov r3=IA64_MCA_CPU_MCA_STACK_OFFSET	// use the MCA stack
+	LOAD_PHYSICAL(p0,r2,1f)			// return address
+	br.sptk ia64_old_stack
+1:
+
+	mov r3=IA64_MCA_CPU_MCA_STACK_OFFSET	// use the MCA stack
+	LOAD_PHYSICAL(p0,r2,1f)			// return address
+	br.sptk ia64_state_restore		// restore the SAL state
+1:
+
+	mov		b0=r12			// SAL_CHECK return address
+
+	// release lock
+	LOAD_PHYSICAL(p0,r3,ia64_mca_serialize);;
+	st4.rel		[r3]=r0
+
+	br		b0
+
+//EndMain//////////////////////////////////////////////////////////////////////
+
+//StartMain////////////////////////////////////////////////////////////////////
+
+//
+// SAL to OS entry point for INIT on all processors.  This has been defined for
+// registration purposes with SAL as a part of ia64_mca_init.  Monarch and
+// slave INIT have identical processing, except for the value of the
+// sos->monarch flag in r19.
+//
+
+ia64_os_init_dispatch_monarch:
+	mov r19=1				// Bow, bow, ye lower middle classes!
+	br.sptk ia64_os_init_dispatch
+
+ia64_os_init_dispatch_slave:
+	mov r19=0				// <igor>yeth, mathter</igor>
+
+ia64_os_init_dispatch:
+
+	mov r3=IA64_MCA_CPU_INIT_STACK_OFFSET	// use the INIT stack
+	LOAD_PHYSICAL(p0,r2,1f)			// return address
+	br.sptk ia64_state_save			// save the state that is not in minstate
+1:
+
+	// switch to per cpu INIT stack
+	mov r3=IA64_MCA_CPU_INIT_STACK_OFFSET	// use the INIT stack
+	LOAD_PHYSICAL(p0,r2,1f)			// return address
+	br.sptk ia64_new_stack
+1:
+
+	// everything saved, now we can set the kernel registers
+	mov r3=IA64_MCA_CPU_INIT_STACK_OFFSET	// use the INIT stack
+	LOAD_PHYSICAL(p0,r2,1f)			// return address
+	br.sptk ia64_set_kernel_registers
+1:
+
+	// This must be done in physical mode
 	GET_IA64_MCA_DATA(r2)
 	;;
-	add r2 = IA64_MCA_CPU_STACKFRAME_OFFSET, r2
-	;;
-	movl    r4=IA64_PSR_MC
+	mov r7=r2
+
+        // Enter virtual mode from physical mode
+	VIRTUAL_MODE_ENTER(r2, r3, ia64_os_init_virtual_begin, r4)
+
+	// This code returns to SAL via SOS r2, in general SAL has no unwind
+	// data.  To get a clean termination when backtracing the C MCA/INIT
+	// handler, set a dummy return address of 0 in this routine.  That
+	// requires that ia64_os_init_virtual_begin be a global function.
+ENTRY(ia64_os_init_virtual_begin)
+	.prologue
+	.save rp,r0
+	.body
+
+	mov ar.rsc=3				// set eager mode for C handler
+	mov r2=r7				// see GET_IA64_MCA_DATA above
 	;;
-	rse_return_context(r4,r3,r2)	// switch from interrupt context for RSE
 
-	// let us restore all the registers from our PSI structure
-	mov	r8=gp
+	// Call virtual mode handler
+	alloc r14=ar.pfs,0,0,3,0
+	;;
+	DATA_PA_TO_VA(r2,r7)
 	;;
-begin_os_mca_restore:
-	br	ia64_os_mca_proc_state_restore;;
+	add out0=IA64_MCA_CPU_INIT_STACK_OFFSET+MCA_PT_REGS_OFFSET, r2
+	add out1=IA64_MCA_CPU_INIT_STACK_OFFSET+MCA_SWITCH_STACK_OFFSET, r2
+	add out2=IA64_MCA_CPU_INIT_STACK_OFFSET+MCA_SOS_OFFSET, r2
+	br.call.sptk.many    b0=ia64_init_handler
 
-ia64_os_mca_done_restore:
-	OS_MCA_TO_SAL_HANDOFF_STATE_RESTORE(r2);;
-	// branch back to SALE_CHECK
-	ld8		r3=[r2];;
-	mov		b0=r3;;		// SAL_CHECK return address
+	// Revert back to physical mode before going back to SAL
+	PHYSICAL_MODE_ENTER(r2, r3, ia64_os_init_virtual_end, r4)
+ia64_os_init_virtual_end:
 
-	// release lock
-	movl		r3=ia64_mca_serialize;;
-	DATA_VA_TO_PA(r3);;
-	st8.rel		[r3]=r0
+END(ia64_os_init_virtual_begin)
+
+	mov r3=IA64_MCA_CPU_INIT_STACK_OFFSET	// use the INIT stack
+	LOAD_PHYSICAL(p0,r2,1f)			// return address
+	br.sptk ia64_state_restore		// restore the SAL state
+1:
 
+	// switch back to previous stack
+	alloc r14=ar.pfs,0,0,0,0		// remove the INIT handler frame
+	mov r3=IA64_MCA_CPU_INIT_STACK_OFFSET	// use the INIT stack
+	LOAD_PHYSICAL(p0,r2,1f)			// return address
+	br.sptk ia64_old_stack
+1:
+
+	mov		b0=r12			// SAL_CHECK return address
 	br		b0
-	;;
-ia64_os_mca_dispatch_end:
+
 //EndMain//////////////////////////////////////////////////////////////////////
 
+// common defines for the stubs
+#define	ms		r4
+#define	regs		r5
+#define	temp1		r2	/* careful, it overlaps with input registers */
+#define	temp2		r3	/* careful, it overlaps with input registers */
+#define	temp3		r7
+#define	temp4		r14
+
 
 //++
 // Name:
-//      ia64_os_mca_proc_state_dump()
+//	ia64_state_save()
 //
 // Stub Description:
 //
-//       This stub dumps the processor state during MCHK to a data area
+//	Save the state that is not in minstate.  This is sensitive to the layout of
+//	struct ia64_sal_os_state in mca.h.
+//
+//	r2 contains the return address, r3 contains either
+//	IA64_MCA_CPU_MCA_STACK_OFFSET or IA64_MCA_CPU_INIT_STACK_OFFSET.
+//
+//	The OS to SAL section of struct ia64_sal_os_state is set to a default
+//	value of cold boot (MCA) or warm boot (INIT) and return to the same
+//	context.  ia64_sal_os_state is also used to hold some registers that
+//	need to be saved and restored across the stack switches.
+//
+//	Most input registers to this stub come from PAL/SAL
+//	r1  os gp, physical
+//	r8  pal_proc entry point
+//	r9  sal_proc entry point
+//	r10 sal gp
+//	r11 MCA - rendevzous state, INIT - reason code
+//	r12 sal return address
+//	r17 pal min_state
+//	r18 processor state parameter
+//	r19 monarch flag, set by the caller of this routine
+//
+//	In addition to the SAL to OS state, this routine saves all the
+//	registers that appear in struct pt_regs and struct switch_stack,
+//	excluding those that are already in the PAL minstate area.  This
+//	results in a partial pt_regs and switch_stack, the C code copies the
+//	remaining registers from PAL minstate to pt_regs and switch_stack.  The
+//	resulting structures contain all the state of the original process when
+//	MCA/INIT occurred.
 //
 //--
 
-ia64_os_mca_proc_state_dump:
-// Save bank 1 GRs 16-31 which will be used by c-language code when we switch
-//  to virtual addressing mode.
-	GET_IA64_MCA_DATA(r2)
+ia64_state_save:
+	add regs=MCA_SOS_OFFSET, r3
+	add ms=MCA_SOS_OFFSET+8, r3
+	mov b0=r2		// save return address
+	cmp.eq p1,p2=IA64_MCA_CPU_MCA_STACK_OFFSET, r3
+	;;
+	GET_IA64_MCA_DATA(temp2)
+	;;
+	add temp1=temp2, regs	// struct ia64_sal_os_state on MCA or INIT stack
+	add temp2=temp2, ms	// struct ia64_sal_os_state+8 on MCA or INIT stack
+	;;
+	mov regs=temp1		// save the start of sos
+	st8 [temp1]=r1,16	// os_gp
+	st8 [temp2]=r8,16	// pal_proc
+	;;
+	st8 [temp1]=r9,16	// sal_proc
+	st8 [temp2]=r11,16	// rv_rc
+	mov r11=cr.iipa
+	;;
+	st8 [temp1]=r18,16	// proc_state_param
+	st8 [temp2]=r19,16	// monarch
+	mov r6=IA64_KR(CURRENT)
+	;;
+	st8 [temp1]=r12,16	// sal_ra
+	st8 [temp2]=r10,16	// sal_gp
+	mov r12=cr.isr
+	;;
+	st8 [temp1]=r17,16	// pal_min_state
+	st8 [temp2]=r6,16	// prev_IA64_KR_CURRENT
+	mov r6=cr.ifa
+	;;
+	st8 [temp1]=r0,16	// prev_task, starts off as NULL
+	st8 [temp2]=r12,16	// cr.isr
+	mov r12=cr.itir
+	;;
+	st8 [temp1]=r6,16	// cr.ifa
+	st8 [temp2]=r12,16	// cr.itir
+	mov r12=cr.iim
+	;;
+	st8 [temp1]=r11,16	// cr.iipa
+	st8 [temp2]=r12,16	// cr.iim
+	mov r6=cr.iha
+(p1)	mov r12=IA64_MCA_COLD_BOOT
+(p2)	mov r12=IA64_INIT_WARM_BOOT
+	;;
+	st8 [temp1]=r6,16	// cr.iha
+	st8 [temp2]=r12		// os_status, default is cold boot
+	mov r6=IA64_MCA_SAME_CONTEXT
+	;;
+	st8 [temp1]=r6		// context, default is same context
+
+	// Save the pt_regs data that is not in minstate.  The previous code
+	// left regs at sos.
+	add regs=MCA_PT_REGS_OFFSET-MCA_SOS_OFFSET, regs
+	;;
+	add temp1=PT(B6), regs
+	mov temp3=b6
+	mov temp4=b7
+	add temp2=PT(B7), regs
+	;;
+	st8 [temp1]=temp3,PT(AR_CSD)-PT(B6)		// save b6
+	st8 [temp2]=temp4,PT(AR_SSD)-PT(B7)		// save b7
+	mov temp3=ar.csd
+	mov temp4=ar.ssd
+	cover						// must be last in group
 	;;
-	add r2 = IA64_MCA_CPU_PROC_STATE_DUMP_OFFSET, r2
-	;;
-// save ar.NaT
-	mov		r5=ar.unat                  // ar.unat
-
-// save banked GRs 16-31 along with NaT bits
-	bsw.1;;
-	st8.spill	[r2]=r16,8;;
-	st8.spill	[r2]=r17,8;;
-	st8.spill	[r2]=r18,8;;
-	st8.spill	[r2]=r19,8;;
-	st8.spill	[r2]=r20,8;;
-	st8.spill	[r2]=r21,8;;
-	st8.spill	[r2]=r22,8;;
-	st8.spill	[r2]=r23,8;;
-	st8.spill	[r2]=r24,8;;
-	st8.spill	[r2]=r25,8;;
-	st8.spill	[r2]=r26,8;;
-	st8.spill	[r2]=r27,8;;
-	st8.spill	[r2]=r28,8;;
-	st8.spill	[r2]=r29,8;;
-	st8.spill	[r2]=r30,8;;
-	st8.spill	[r2]=r31,8;;
-
-	mov		r4=ar.unat;;
-	st8		[r2]=r4,8                // save User NaT bits for r16-r31
-	mov		ar.unat=r5                  // restore original unat
-	bsw.0;;
-
-//save BRs
-	add		r4=8,r2                  // duplicate r2 in r4
-	add		r6=2*8,r2                // duplicate r2 in r4
-
-	mov		r3=b0
-	mov		r5=b1
-	mov		r7=b2;;
-	st8		[r2]=r3,3*8
-	st8		[r4]=r5,3*8
-	st8		[r6]=r7,3*8;;
-
-	mov		r3=b3
-	mov		r5=b4
-	mov		r7=b5;;
-	st8		[r2]=r3,3*8
-	st8		[r4]=r5,3*8
-	st8		[r6]=r7,3*8;;
-
-	mov		r3=b6
-	mov		r5=b7;;
-	st8		[r2]=r3,2*8
-	st8		[r4]=r5,2*8;;
-
-cSaveCRs:
-// save CRs
-	add		r4=8,r2                  // duplicate r2 in r4
-	add		r6=2*8,r2                // duplicate r2 in r4
-
-	mov		r3=cr.dcr
-	mov		r5=cr.itm
-	mov		r7=cr.iva;;
-
-	st8		[r2]=r3,8*8
-	st8		[r4]=r5,3*8
-	st8		[r6]=r7,3*8;;            // 48 byte rements
-
-	mov		r3=cr.pta;;
-	st8		[r2]=r3,8*8;;            // 64 byte rements
-
-// if PSR.ic=0, reading interruption registers causes an illegal operation fault
-	mov		r3=psr;;
-	tbit.nz.unc	p6,p0=r3,PSR_IC;;           // PSI Valid Log bit pos. test
-(p6)    st8     [r2]=r0,9*8+160             // increment by 232 byte inc.
-begin_skip_intr_regs:
-(p6)	br		SkipIntrRegs;;
-
-	add		r4=8,r2                  // duplicate r2 in r4
-	add		r6=2*8,r2                // duplicate r2 in r6
-
-	mov		r3=cr.ipsr
-	mov		r5=cr.isr
-	mov		r7=r0;;
-	st8		[r2]=r3,3*8
-	st8		[r4]=r5,3*8
-	st8		[r6]=r7,3*8;;
-
-	mov		r3=cr.iip
-	mov		r5=cr.ifa
-	mov		r7=cr.itir;;
-	st8		[r2]=r3,3*8
-	st8		[r4]=r5,3*8
-	st8		[r6]=r7,3*8;;
-
-	mov		r3=cr.iipa
-	mov		r5=cr.ifs
-	mov		r7=cr.iim;;
-	st8		[r2]=r3,3*8
-	st8		[r4]=r5,3*8
-	st8		[r6]=r7,3*8;;
-
-	mov		r3=cr25;;                   // cr.iha
-	st8		[r2]=r3,160;;               // 160 byte rement
-
-SkipIntrRegs:
-	st8		[r2]=r0,152;;               // another 152 byte .
-
-	add		r4=8,r2                     // duplicate r2 in r4
-	add		r6=2*8,r2                   // duplicate r2 in r6
-
-	mov		r3=cr.lid
-//	mov		r5=cr.ivr                     // cr.ivr, don't read it
-	mov		r7=cr.tpr;;
-	st8		[r2]=r3,3*8
-	st8		[r4]=r5,3*8
-	st8		[r6]=r7,3*8;;
-
-	mov		r3=r0                       // cr.eoi => cr67
-	mov		r5=r0                       // cr.irr0 => cr68
-	mov		r7=r0;;                     // cr.irr1 => cr69
-	st8		[r2]=r3,3*8
-	st8		[r4]=r5,3*8
-	st8		[r6]=r7,3*8;;
-
-	mov		r3=r0                       // cr.irr2 => cr70
-	mov		r5=r0                       // cr.irr3 => cr71
-	mov		r7=cr.itv;;
-	st8		[r2]=r3,3*8
-	st8		[r4]=r5,3*8
-	st8		[r6]=r7,3*8;;
-
-	mov		r3=cr.pmv
-	mov		r5=cr.cmcv;;
-	st8		[r2]=r3,7*8
-	st8		[r4]=r5,7*8;;
-
-	mov		r3=r0                       // cr.lrr0 => cr80
-	mov		r5=r0;;                     // cr.lrr1 => cr81
-	st8		[r2]=r3,23*8
-	st8		[r4]=r5,23*8;;
-
-	adds		r2=25*8,r2;;
-
-cSaveARs:
-// save ARs
-	add		r4=8,r2                  // duplicate r2 in r4
-	add		r6=2*8,r2                // duplicate r2 in r6
-
-	mov		r3=ar.k0
-	mov		r5=ar.k1
-	mov		r7=ar.k2;;
-	st8		[r2]=r3,3*8
-	st8		[r4]=r5,3*8
-	st8		[r6]=r7,3*8;;
-
-	mov		r3=ar.k3
-	mov		r5=ar.k4
-	mov		r7=ar.k5;;
-	st8		[r2]=r3,3*8
-	st8		[r4]=r5,3*8
-	st8		[r6]=r7,3*8;;
-
-	mov		r3=ar.k6
-	mov		r5=ar.k7
-	mov		r7=r0;;                     // ar.kr8
-	st8		[r2]=r3,10*8
-	st8		[r4]=r5,10*8
-	st8		[r6]=r7,10*8;;           // rement by 72 bytes
-
-	mov		r3=ar.rsc
-	mov		ar.rsc=r0			    // put RSE in enforced lazy mode
-	mov		r5=ar.bsp
-	;;
-	mov		r7=ar.bspstore;;
-	st8		[r2]=r3,3*8
-	st8		[r4]=r5,3*8
-	st8		[r6]=r7,3*8;;
-
-	mov		r3=ar.rnat;;
-	st8		[r2]=r3,8*13             // increment by 13x8 bytes
-
-	mov		r3=ar.ccv;;
-	st8		[r2]=r3,8*4
-
-	mov		r3=ar.unat;;
-	st8		[r2]=r3,8*4
-
-	mov		r3=ar.fpsr;;
-	st8		[r2]=r3,8*4
-
-	mov		r3=ar.itc;;
-	st8		[r2]=r3,160                 // 160
-
-	mov		r3=ar.pfs;;
-	st8		[r2]=r3,8
-
-	mov		r3=ar.lc;;
-	st8		[r2]=r3,8
-
-	mov		r3=ar.ec;;
-	st8		[r2]=r3
-	add		r2=8*62,r2               //padding
-
-// save RRs
-	mov		ar.lc=0x08-1
-	movl		r4=0x00;;
-
-cStRR:
-	dep.z		r5=r4,61,3;;
-	mov		r3=rr[r5];;
-	st8		[r2]=r3,8
-	add		r4=1,r4
-	br.cloop.sptk.few	cStRR
-	;;
-end_os_mca_dump:
-	br	ia64_os_mca_done_dump;;
+	st8 [temp1]=temp3,PT(AR_UNAT)-PT(AR_CSD)	// save ar.csd
+	st8 [temp2]=temp4,PT(AR_PFS)-PT(AR_SSD)		// save ar.ssd
+	mov temp3=ar.unat
+	mov temp4=ar.pfs
+	;;
+	st8 [temp1]=temp3,PT(AR_RNAT)-PT(AR_UNAT)	// save ar.unat
+	st8 [temp2]=temp4,PT(AR_BSPSTORE)-PT(AR_PFS)	// save ar.pfs
+	mov temp3=ar.rnat
+	mov temp4=ar.bspstore
+	;;
+	st8 [temp1]=temp3,PT(LOADRS)-PT(AR_RNAT)	// save ar.rnat
+	st8 [temp2]=temp4,PT(AR_FPSR)-PT(AR_BSPSTORE)	// save ar.bspstore
+	mov temp3=ar.bsp
+	;;
+	sub temp3=temp3, temp4	// ar.bsp - ar.bspstore
+	mov temp4=ar.fpsr
+	;;
+	shl temp3=temp3,16	// compute ar.rsc to be used for "loadrs"
+	;;
+	st8 [temp1]=temp3,PT(AR_CCV)-PT(LOADRS)		// save loadrs
+	st8 [temp2]=temp4,PT(F6)-PT(AR_FPSR)		// save ar.fpsr
+	mov temp3=ar.ccv
+	;;
+	st8 [temp1]=temp3,PT(F7)-PT(AR_CCV)		// save ar.ccv
+	stf.spill [temp2]=f6,PT(F8)-PT(F6)
+	;;
+	stf.spill [temp1]=f7,PT(F9)-PT(F7)
+	stf.spill [temp2]=f8,PT(F10)-PT(F8)
+	;;
+	stf.spill [temp1]=f9,PT(F11)-PT(F9)
+	stf.spill [temp2]=f10
+	;;
+	stf.spill [temp1]=f11
+
+	// Save the switch_stack data that is not in minstate nor pt_regs.  The
+	// previous code left regs at pt_regs.
+	add regs=MCA_SWITCH_STACK_OFFSET-MCA_PT_REGS_OFFSET, regs
+	;;
+	add temp1=SW(F2), regs
+	add temp2=SW(F3), regs
+	;;
+	stf.spill [temp1]=f2,32
+	stf.spill [temp2]=f3,32
+	;;
+	stf.spill [temp1]=f4,32
+	stf.spill [temp2]=f5,32
+	;;
+	stf.spill [temp1]=f12,32
+	stf.spill [temp2]=f13,32
+	;;
+	stf.spill [temp1]=f14,32
+	stf.spill [temp2]=f15,32
+	;;
+	stf.spill [temp1]=f16,32
+	stf.spill [temp2]=f17,32
+	;;
+	stf.spill [temp1]=f18,32
+	stf.spill [temp2]=f19,32
+	;;
+	stf.spill [temp1]=f20,32
+	stf.spill [temp2]=f21,32
+	;;
+	stf.spill [temp1]=f22,32
+	stf.spill [temp2]=f23,32
+	;;
+	stf.spill [temp1]=f24,32
+	stf.spill [temp2]=f25,32
+	;;
+	stf.spill [temp1]=f26,32
+	stf.spill [temp2]=f27,32
+	;;
+	stf.spill [temp1]=f28,32
+	stf.spill [temp2]=f29,32
+	;;
+	stf.spill [temp1]=f30,SW(B2)-SW(F30)
+	stf.spill [temp2]=f31,SW(B3)-SW(F31)
+	mov temp3=b2
+	mov temp4=b3
+	;;
+	st8 [temp1]=temp3,16	// save b2
+	st8 [temp2]=temp4,16	// save b3
+	mov temp3=b4
+	mov temp4=b5
+	;;
+	st8 [temp1]=temp3,SW(AR_LC)-SW(B4)	// save b4
+	st8 [temp2]=temp4	// save b5
+	mov temp3=ar.lc
+	;;
+	st8 [temp1]=temp3	// save ar.lc
+
+	// FIXME: Some proms are incorrectly accessing the minstate area as
+	// cached data.  The C code uses region 6, uncached virtual.  Ensure
+	// that there is no cache data lying around for the first 1K of the
+	// minstate area.
+	// Remove this code in September 2006, that gives platforms a year to
+	// fix their proms and get their customers updated.
+
+	add r1=32*1,r17
+	add r2=32*2,r17
+	add r3=32*3,r17
+	add r4=32*4,r17
+	add r5=32*5,r17
+	add r6=32*6,r17
+	add r7=32*7,r17
+	;;
+	fc r17
+	fc r1
+	fc r2
+	fc r3
+	fc r4
+	fc r5
+	fc r6
+	fc r7
+	add r17=32*8,r17
+	add r1=32*8,r1
+	add r2=32*8,r2
+	add r3=32*8,r3
+	add r4=32*8,r4
+	add r5=32*8,r5
+	add r6=32*8,r6
+	add r7=32*8,r7
+	;;
+	fc r17
+	fc r1
+	fc r2
+	fc r3
+	fc r4
+	fc r5
+	fc r6
+	fc r7
+	add r17=32*8,r17
+	add r1=32*8,r1
+	add r2=32*8,r2
+	add r3=32*8,r3
+	add r4=32*8,r4
+	add r5=32*8,r5
+	add r6=32*8,r6
+	add r7=32*8,r7
+	;;
+	fc r17
+	fc r1
+	fc r2
+	fc r3
+	fc r4
+	fc r5
+	fc r6
+	fc r7
+	add r17=32*8,r17
+	add r1=32*8,r1
+	add r2=32*8,r2
+	add r3=32*8,r3
+	add r4=32*8,r4
+	add r5=32*8,r5
+	add r6=32*8,r6
+	add r7=32*8,r7
+	;;
+	fc r17
+	fc r1
+	fc r2
+	fc r3
+	fc r4
+	fc r5
+	fc r6
+	fc r7
+
+	br.sptk b0
 
 //EndStub//////////////////////////////////////////////////////////////////////
 
 
 //++
 // Name:
-//       ia64_os_mca_proc_state_restore()
+//	ia64_state_restore()
 //
 // Stub Description:
 //
-//       This is a stub to restore the saved processor state during MCHK
+//	Restore the SAL/OS state.  This is sensitive to the layout of struct
+//	ia64_sal_os_state in mca.h.
+//
+//	r2 contains the return address, r3 contains either
+//	IA64_MCA_CPU_MCA_STACK_OFFSET or IA64_MCA_CPU_INIT_STACK_OFFSET.
+//
+//	In addition to the SAL to OS state, this routine restores all the
+//	registers that appear in struct pt_regs and struct switch_stack,
+//	excluding those in the PAL minstate area.
 //
 //--
 
-ia64_os_mca_proc_state_restore:
+ia64_state_restore:
+	// Restore the switch_stack data that is not in minstate nor pt_regs.
+	add regs=MCA_SWITCH_STACK_OFFSET, r3
+	mov b0=r2		// save return address
+	;;
+	GET_IA64_MCA_DATA(temp2)
+	;;
+	add regs=temp2, regs
+	;;
+	add temp1=SW(F2), regs
+	add temp2=SW(F3), regs
+	;;
+	ldf.fill f2=[temp1],32
+	ldf.fill f3=[temp2],32
+	;;
+	ldf.fill f4=[temp1],32
+	ldf.fill f5=[temp2],32
+	;;
+	ldf.fill f12=[temp1],32
+	ldf.fill f13=[temp2],32
+	;;
+	ldf.fill f14=[temp1],32
+	ldf.fill f15=[temp2],32
+	;;
+	ldf.fill f16=[temp1],32
+	ldf.fill f17=[temp2],32
+	;;
+	ldf.fill f18=[temp1],32
+	ldf.fill f19=[temp2],32
+	;;
+	ldf.fill f20=[temp1],32
+	ldf.fill f21=[temp2],32
+	;;
+	ldf.fill f22=[temp1],32
+	ldf.fill f23=[temp2],32
+	;;
+	ldf.fill f24=[temp1],32
+	ldf.fill f25=[temp2],32
+	;;
+	ldf.fill f26=[temp1],32
+	ldf.fill f27=[temp2],32
+	;;
+	ldf.fill f28=[temp1],32
+	ldf.fill f29=[temp2],32
+	;;
+	ldf.fill f30=[temp1],SW(B2)-SW(F30)
+	ldf.fill f31=[temp2],SW(B3)-SW(F31)
+	;;
+	ld8 temp3=[temp1],16	// restore b2
+	ld8 temp4=[temp2],16	// restore b3
+	;;
+	mov b2=temp3
+	mov b3=temp4
+	ld8 temp3=[temp1],SW(AR_LC)-SW(B4)	// restore b4
+	ld8 temp4=[temp2]	// restore b5
+	;;
+	mov b4=temp3
+	mov b5=temp4
+	ld8 temp3=[temp1]	// restore ar.lc
+	;;
+	mov ar.lc=temp3
 
-// Restore bank1 GR16-31
-	GET_IA64_MCA_DATA(r2)
+	// Restore the pt_regs data that is not in minstate.  The previous code
+	// left regs at switch_stack.
+	add regs=MCA_PT_REGS_OFFSET-MCA_SWITCH_STACK_OFFSET, regs
+	;;
+	add temp1=PT(B6), regs
+	add temp2=PT(B7), regs
+	;;
+	ld8 temp3=[temp1],PT(AR_CSD)-PT(B6)		// restore b6
+	ld8 temp4=[temp2],PT(AR_SSD)-PT(B7)		// restore b7
+	;;
+	mov b6=temp3
+	mov b7=temp4
+	ld8 temp3=[temp1],PT(AR_UNAT)-PT(AR_CSD)	// restore ar.csd
+	ld8 temp4=[temp2],PT(AR_PFS)-PT(AR_SSD)		// restore ar.ssd
+	;;
+	mov ar.csd=temp3
+	mov ar.ssd=temp4
+	ld8 temp3=[temp1]				// restore ar.unat
+	add temp1=PT(AR_CCV)-PT(AR_UNAT), temp1
+	ld8 temp4=[temp2],PT(AR_FPSR)-PT(AR_PFS)	// restore ar.pfs
+	;;
+	mov ar.unat=temp3
+	mov ar.pfs=temp4
+	// ar.rnat, ar.bspstore, loadrs are restore in ia64_old_stack.
+	ld8 temp3=[temp1],PT(F6)-PT(AR_CCV)		// restore ar.ccv
+	ld8 temp4=[temp2],PT(F7)-PT(AR_FPSR)		// restore ar.fpsr
+	;;
+	mov ar.ccv=temp3
+	mov ar.fpsr=temp4
+	ldf.fill f6=[temp1],PT(F8)-PT(F6)
+	ldf.fill f7=[temp2],PT(F9)-PT(F7)
+	;;
+	ldf.fill f8=[temp1],PT(F10)-PT(F8)
+	ldf.fill f9=[temp2],PT(F11)-PT(F9)
+	;;
+	ldf.fill f10=[temp1]
+	ldf.fill f11=[temp2]
+
+	// Restore the SAL to OS state. The previous code left regs at pt_regs.
+	add regs=MCA_SOS_OFFSET-MCA_PT_REGS_OFFSET, regs
 	;;
-	add r2 = IA64_MCA_CPU_PROC_STATE_DUMP_OFFSET, r2
-
-restore_GRs:                                    // restore bank-1 GRs 16-31
-	bsw.1;;
-	add		r3=16*8,r2;;                // to get to NaT of GR 16-31
-	ld8		r3=[r3];;
-	mov		ar.unat=r3;;                // first restore NaT
-
-	ld8.fill	r16=[r2],8;;
-	ld8.fill	r17=[r2],8;;
-	ld8.fill	r18=[r2],8;;
-	ld8.fill	r19=[r2],8;;
-	ld8.fill	r20=[r2],8;;
-	ld8.fill	r21=[r2],8;;
-	ld8.fill	r22=[r2],8;;
-	ld8.fill	r23=[r2],8;;
-	ld8.fill	r24=[r2],8;;
-	ld8.fill	r25=[r2],8;;
-	ld8.fill	r26=[r2],8;;
-	ld8.fill	r27=[r2],8;;
-	ld8.fill	r28=[r2],8;;
-	ld8.fill	r29=[r2],8;;
-	ld8.fill	r30=[r2],8;;
-	ld8.fill	r31=[r2],8;;
-
-	ld8		r3=[r2],8;;              // increment to skip NaT
-	bsw.0;;
-
-restore_BRs:
-	add		r4=8,r2                  // duplicate r2 in r4
-	add		r6=2*8,r2;;              // duplicate r2 in r4
-
-	ld8		r3=[r2],3*8
-	ld8		r5=[r4],3*8
-	ld8		r7=[r6],3*8;;
-	mov		b0=r3
-	mov		b1=r5
-	mov		b2=r7;;
-
-	ld8		r3=[r2],3*8
-	ld8		r5=[r4],3*8
-	ld8		r7=[r6],3*8;;
-	mov		b3=r3
-	mov		b4=r5
-	mov		b5=r7;;
-
-	ld8		r3=[r2],2*8
-	ld8		r5=[r4],2*8;;
-	mov		b6=r3
-	mov		b7=r5;;
-
-restore_CRs:
-	add		r4=8,r2                  // duplicate r2 in r4
-	add		r6=2*8,r2;;              // duplicate r2 in r4
-
-	ld8		r3=[r2],8*8
-	ld8		r5=[r4],3*8
-	ld8		r7=[r6],3*8;;            // 48 byte increments
-	mov		cr.dcr=r3
-	mov		cr.itm=r5
-	mov		cr.iva=r7;;
-
-	ld8		r3=[r2],8*8;;            // 64 byte increments
-//      mov		cr.pta=r3
-
-
-// if PSR.ic=1, reading interruption registers causes an illegal operation fault
-	mov		r3=psr;;
-	tbit.nz.unc	p6,p0=r3,PSR_IC;;           // PSI Valid Log bit pos. test
-(p6)    st8     [r2]=r0,9*8+160             // increment by 232 byte inc.
-
-begin_rskip_intr_regs:
-(p6)	br		rSkipIntrRegs;;
-
-	add		r4=8,r2                  // duplicate r2 in r4
-	add		r6=2*8,r2;;              // duplicate r2 in r4
-
-	ld8		r3=[r2],3*8
-	ld8		r5=[r4],3*8
-	ld8		r7=[r6],3*8;;
-	mov		cr.ipsr=r3
-//	mov		cr.isr=r5                   // cr.isr is read only
-
-	ld8		r3=[r2],3*8
-	ld8		r5=[r4],3*8
-	ld8		r7=[r6],3*8;;
-	mov		cr.iip=r3
-	mov		cr.ifa=r5
-	mov		cr.itir=r7;;
-
-	ld8		r3=[r2],3*8
-	ld8		r5=[r4],3*8
-	ld8		r7=[r6],3*8;;
-	mov		cr.iipa=r3
-	mov		cr.ifs=r5
-	mov		cr.iim=r7
-
-	ld8		r3=[r2],160;;               // 160 byte increment
-	mov		cr.iha=r3
-
-rSkipIntrRegs:
-	ld8		r3=[r2],152;;               // another 152 byte inc.
-
-	add		r4=8,r2                     // duplicate r2 in r4
-	add		r6=2*8,r2;;                 // duplicate r2 in r6
-
-	ld8		r3=[r2],8*3
-	ld8		r5=[r4],8*3
-	ld8		r7=[r6],8*3;;
-	mov		cr.lid=r3
-//	mov		cr.ivr=r5                   // cr.ivr is read only
-	mov		cr.tpr=r7;;
-
-	ld8		r3=[r2],8*3
-	ld8		r5=[r4],8*3
-	ld8		r7=[r6],8*3;;
-//	mov		cr.eoi=r3
-//	mov		cr.irr0=r5                  // cr.irr0 is read only
-//	mov		cr.irr1=r7;;                // cr.irr1 is read only
-
-	ld8		r3=[r2],8*3
-	ld8		r5=[r4],8*3
-	ld8		r7=[r6],8*3;;
-//	mov		cr.irr2=r3                  // cr.irr2 is read only
-//	mov		cr.irr3=r5                  // cr.irr3 is read only
-	mov		cr.itv=r7;;
-
-	ld8		r3=[r2],8*7
-	ld8		r5=[r4],8*7;;
-	mov		cr.pmv=r3
-	mov		cr.cmcv=r5;;
-
-	ld8		r3=[r2],8*23
-	ld8		r5=[r4],8*23;;
-	adds		r2=8*23,r2
-	adds		r4=8*23,r4;;
-//	mov		cr.lrr0=r3
-//	mov		cr.lrr1=r5
-
-	adds		r2=8*2,r2;;
-
-restore_ARs:
-	add		r4=8,r2                  // duplicate r2 in r4
-	add		r6=2*8,r2;;              // duplicate r2 in r4
-
-	ld8		r3=[r2],3*8
-	ld8		r5=[r4],3*8
-	ld8		r7=[r6],3*8;;
-	mov		ar.k0=r3
-	mov		ar.k1=r5
-	mov		ar.k2=r7;;
-
-	ld8		r3=[r2],3*8
-	ld8		r5=[r4],3*8
-	ld8		r7=[r6],3*8;;
-	mov		ar.k3=r3
-	mov		ar.k4=r5
-	mov		ar.k5=r7;;
-
-	ld8		r3=[r2],10*8
-	ld8		r5=[r4],10*8
-	ld8		r7=[r6],10*8;;
-	mov		ar.k6=r3
-	mov		ar.k7=r5
-	;;
-
-	ld8		r3=[r2],3*8
-	ld8		r5=[r4],3*8
-	ld8		r7=[r6],3*8;;
-//	mov		ar.rsc=r3
-//	mov		ar.bsp=r5                   // ar.bsp is read only
-	mov		ar.rsc=r0			    // make sure that RSE is in enforced lazy mode
-	;;
-	mov		ar.bspstore=r7;;
-
-	ld8		r9=[r2],8*13;;
-	mov		ar.rnat=r9
-
-	mov		ar.rsc=r3
-	ld8		r3=[r2],8*4;;
-	mov		ar.ccv=r3
-
-	ld8		r3=[r2],8*4;;
-	mov		ar.unat=r3
-
-	ld8		r3=[r2],8*4;;
-	mov		ar.fpsr=r3
-
-	ld8		r3=[r2],160;;               // 160
-//      mov		ar.itc=r3
-
-	ld8		r3=[r2],8;;
-	mov		ar.pfs=r3
-
-	ld8		r3=[r2],8;;
-	mov		ar.lc=r3
-
-	ld8		r3=[r2];;
-	mov		ar.ec=r3
-	add		r2=8*62,r2;;             // padding
-
-restore_RRs:
-	mov		r5=ar.lc
-	mov		ar.lc=0x08-1
-	movl		r4=0x00;;
-cStRRr:
-	dep.z		r7=r4,61,3
-	ld8		r3=[r2],8;;
-	mov		rr[r7]=r3                   // what are its access previledges?
-	add		r4=1,r4
-	br.cloop.sptk.few	cStRRr
-	;;
-	mov		ar.lc=r5
-	;;
-end_os_mca_restore:
-	br	ia64_os_mca_done_restore;;
+	add temp1=IA64_SAL_OS_STATE_COMMON_OFFSET, regs
+	add temp2=IA64_SAL_OS_STATE_COMMON_OFFSET+8, regs
+	;;
+	ld8 r12=[temp1],16	// sal_ra
+	ld8 r9=[temp2],16	// sal_gp
+	;;
+	ld8 r22=[temp1],24	// pal_min_state, virtual.  skip prev_task
+	ld8 r21=[temp2],16	// prev_IA64_KR_CURRENT
+	;;
+	ld8 temp3=[temp1],16	// cr.isr
+	ld8 temp4=[temp2],16	// cr.ifa
+	;;
+	mov cr.isr=temp3
+	mov cr.ifa=temp4
+	ld8 temp3=[temp1],16	// cr.itir
+	ld8 temp4=[temp2],16	// cr.iipa
+	;;
+	mov cr.itir=temp3
+	mov cr.iipa=temp4
+	ld8 temp3=[temp1],16	// cr.iim
+	ld8 temp4=[temp2],16	// cr.iha
+	;;
+	mov cr.iim=temp3
+	mov cr.iha=temp4
+	dep r22=0,r22,62,2	// pal_min_state, physical, uncached
+	mov IA64_KR(CURRENT)=r21
+	ld8 r8=[temp1]		// os_status
+	ld8 r10=[temp2]		// context
+
+	br.sptk b0
 
 //EndStub//////////////////////////////////////////////////////////////////////
 
 
-// ok, the issue here is that we need to save state information so
-// it can be useable by the kernel debugger and show regs routines.
-// In order to do this, our best bet is save the current state (plus
-// the state information obtain from the MIN_STATE_AREA) into a pt_regs
-// format.  This way we can pass it on in a useable format.
+//++
+// Name:
+//	ia64_new_stack()
 //
-
+// Stub Description:
 //
-// SAL to OS entry point for INIT on the monarch processor
-// This has been defined for registration purposes with SAL
-// as a part of ia64_mca_init.
+//	Switch to the MCA/INIT stack.
 //
-// When we get here, the following registers have been
-// set by the SAL for our use
+//	r2 contains the return address, r3 contains either
+//	IA64_MCA_CPU_MCA_STACK_OFFSET or IA64_MCA_CPU_INIT_STACK_OFFSET.
 //
-//		1. GR1 = OS INIT GP
-//		2. GR8 = PAL_PROC physical address
-//		3. GR9 = SAL_PROC physical address
-//		4. GR10 = SAL GP (physical)
-//		5. GR11 = Init Reason
-//			0 = Received INIT for event other than crash dump switch
-//			1 = Received wakeup at the end of an OS_MCA corrected machine check
-//			2 = Received INIT dude to CrashDump switch assertion
+//	On entry RBS is still on the original stack, this routine switches RBS
+//	to use the MCA/INIT stack.
 //
-//		6. GR12 = Return address to location within SAL_INIT procedure
-
+//	On entry, sos->pal_min_state is physical, on exit it is virtual.
+//
+//--
 
-GLOBAL_ENTRY(ia64_monarch_init_handler)
-	.prologue
-	// stash the information the SAL passed to os
-	SAL_TO_OS_MCA_HANDOFF_STATE_SAVE(r2)
+ia64_new_stack:
+	add regs=MCA_PT_REGS_OFFSET, r3
+	add temp2=MCA_SOS_OFFSET+IA64_SAL_OS_STATE_PAL_MIN_STATE_OFFSET, r3
+	mov b0=r2			// save return address
+	GET_IA64_MCA_DATA(temp1)
+	invala
 	;;
-	SAVE_MIN_WITH_COVER
+	add temp2=temp2, temp1		// struct ia64_sal_os_state.pal_min_state on MCA or INIT stack
+	add regs=regs, temp1		// struct pt_regs on MCA or INIT stack
 	;;
-	mov r8=cr.ifa
-	mov r9=cr.isr
-	adds r3=8,r2				// set up second base pointer
+	// Address of minstate area provided by PAL is physical, uncacheable.
+	// Convert to Linux virtual address in region 6 for C code.
+	ld8 ms=[temp2]			// pal_min_state, physical
 	;;
-	SAVE_REST
-
-// ok, enough should be saved at this point to be dangerous, and supply
-// information for a dump
-// We need to switch to Virtual mode before hitting the C functions.
+	dep temp1=-1,ms,62,2		// set region 6
+	mov temp3=IA64_RBS_OFFSET-MCA_PT_REGS_OFFSET
+	;;
+	st8 [temp2]=temp1		// pal_min_state, virtual
 
-	movl	r2=IA64_PSR_IT|IA64_PSR_IC|IA64_PSR_DT|IA64_PSR_RT|IA64_PSR_DFH|IA64_PSR_BN
-	mov	r3=psr	// get the current psr, minimum enabled at this point
+	add temp4=temp3, regs		// start of bspstore on new stack
 	;;
-	or	r2=r2,r3
+	mov ar.bspstore=temp4		// switch RBS to MCA/INIT stack
 	;;
-	movl	r3=IVirtual_Switch
+	flushrs				// must be first in group
+	br.sptk b0
+
+//EndStub//////////////////////////////////////////////////////////////////////
+
+
+//++
+// Name:
+//	ia64_old_stack()
+//
+// Stub Description:
+//
+//	Switch to the old stack.
+//
+//	r2 contains the return address, r3 contains either
+//	IA64_MCA_CPU_MCA_STACK_OFFSET or IA64_MCA_CPU_INIT_STACK_OFFSET.
+//
+//	On entry, pal_min_state is virtual, on exit it is physical.
+//
+//	On entry RBS is on the MCA/INIT stack, this routine switches RBS
+//	back to the previous stack.
+//
+//	The psr is set to all zeroes.  SAL return requires either all zeroes or
+//	just psr.mc set.  Leaving psr.mc off allows INIT to be issued if this
+//	code does not perform correctly.
+//
+//	The dirty registers at the time of the event were flushed to the
+//	MCA/INIT stack in ia64_pt_regs_save().  Restore the dirty registers
+//	before reverting to the previous bspstore.
+//--
+
+ia64_old_stack:
+	add regs=MCA_PT_REGS_OFFSET, r3
+	mov b0=r2			// save return address
+	GET_IA64_MCA_DATA(temp2)
+	LOAD_PHYSICAL(p0,temp1,1f)
 	;;
-	mov	cr.iip=r3	// short return to set the appropriate bits
-	mov	cr.ipsr=r2	// need to do an rfi to set appropriate bits
+	mov cr.ipsr=r0
+	mov cr.ifs=r0
+	mov cr.iip=temp1
 	;;
+	invala
 	rfi
+1:
+
+	add regs=regs, temp2		// struct pt_regs on MCA or INIT stack
 	;;
-IVirtual_Switch:
-	//
-	// We should now be running virtual
-	//
-	// Let's call the C handler to get the rest of the state info
-	//
-	alloc r14=ar.pfs,0,0,2,0		// now it's safe (must be first in insn group!)
+	add temp1=PT(LOADRS), regs
 	;;
-	adds out0=16,sp				// out0 = pointer to pt_regs
+	ld8 temp2=[temp1],PT(AR_BSPSTORE)-PT(LOADRS)	// restore loadrs
 	;;
-	DO_SAVE_SWITCH_STACK
-	.body
-	adds out1=16,sp				// out0 = pointer to switch_stack
+	ld8 temp3=[temp1],PT(AR_RNAT)-PT(AR_BSPSTORE)	// restore ar.bspstore
+	mov ar.rsc=temp2
+	;;
+	loadrs
+	ld8 temp4=[temp1]		// restore ar.rnat
+	;;
+	mov ar.bspstore=temp3		// back to old stack
+	;;
+	mov ar.rnat=temp4
+	;;
+
+	br.sptk b0
 
-	br.call.sptk.many rp=ia64_init_handler
-.ret1:
+//EndStub//////////////////////////////////////////////////////////////////////
 
-return_from_init:
-	br.sptk return_from_init
-END(ia64_monarch_init_handler)
 
+//++
+// Name:
+//	ia64_set_kernel_registers()
 //
-// SAL to OS entry point for INIT on the slave processor
-// This has been defined for registration purposes with SAL
-// as a part of ia64_mca_init.
+// Stub Description:
+//
+//	Set the registers that are required by the C code in order to run on an
+//	MCA/INIT stack.
+//
+//	r2 contains the return address, r3 contains either
+//	IA64_MCA_CPU_MCA_STACK_OFFSET or IA64_MCA_CPU_INIT_STACK_OFFSET.
 //
+//--
+
+ia64_set_kernel_registers:
+	add temp3=MCA_SP_OFFSET, r3
+	add temp4=MCA_SOS_OFFSET+IA64_SAL_OS_STATE_OS_GP_OFFSET, r3
+	mov b0=r2		// save return address
+	GET_IA64_MCA_DATA(temp1)
+	;;
+	add temp4=temp4, temp1	// &struct ia64_sal_os_state.os_gp
+	add r12=temp1, temp3	// kernel stack pointer on MCA/INIT stack
+	add r13=temp1, r3	// set current to start of MCA/INIT stack
+	;;
+	ld8 r1=[temp4]		// OS GP from SAL OS state
+	;;
+	DATA_PA_TO_VA(r1,temp1)
+	DATA_PA_TO_VA(r12,temp2)
+	DATA_PA_TO_VA(r13,temp3)
+	;;
+	mov IA64_KR(CURRENT)=r13
+
+	// FIXME: do I need to wire IA64_KR_CURRENT_STACK and IA64_TR_CURRENT_STACK?
+
+	br.sptk b0
+
+//EndStub//////////////////////////////////////////////////////////////////////
+
+#undef	ms
+#undef	regs
+#undef	temp1
+#undef	temp2
+#undef	temp3
+#undef	temp4
+
 
-GLOBAL_ENTRY(ia64_slave_init_handler)
-1:	br.sptk 1b
-END(ia64_slave_init_handler)
+// Support function for mca.c, it is here to avoid using inline asm.  Given the
+// address of an rnat slot, if that address is below the current ar.bspstore
+// then return the contents of that slot, otherwise return the contents of
+// ar.rnat.
+GLOBAL_ENTRY(ia64_get_rnat)
+	alloc r14=ar.pfs,1,0,0,0
+	mov ar.rsc=0
+	;;
+	mov r14=ar.bspstore
+	;;
+	cmp.lt p6,p7=in0,r14
+	;;
+(p6)	ld8 r8=[in0]
+(p7)	mov r8=ar.rnat
+	mov ar.rsc=3
+	br.ret.sptk.many rp
+END(ia64_get_rnat)
diff --git a/arch/ia64/kernel/mca_drv.c b/arch/ia64/kernel/mca_drv.c
index abc0113a821d..6e683745af49 100644
--- a/arch/ia64/kernel/mca_drv.c
+++ b/arch/ia64/kernel/mca_drv.c
@@ -4,6 +4,8 @@
  *
  * Copyright (C) 2004 FUJITSU LIMITED
  * Copyright (C) Hidetoshi Seto (seto.hidetoshi@jp.fujitsu.com)
+ * Copyright (C) 2005 Silicon Graphics, Inc
+ * Copyright (C) 2005 Keith Owens <kaos@sgi.com>
  */
 #include <linux/config.h>
 #include <linux/types.h>
@@ -38,10 +40,6 @@
 /* max size of SAL error record (default) */
 static int sal_rec_max = 10000;
 
-/* from mca.c */
-static ia64_mca_sal_to_os_state_t *sal_to_os_handoff_state;
-static ia64_mca_os_to_sal_state_t *os_to_sal_handoff_state;
-
 /* from mca_drv_asm.S */
 extern void *mca_handler_bhhook(void);
 
@@ -316,7 +314,8 @@ init_record_index_pools(void)
  */
 
 static mca_type_t
-is_mca_global(peidx_table_t *peidx, pal_bus_check_info_t *pbci)
+is_mca_global(peidx_table_t *peidx, pal_bus_check_info_t *pbci,
+	      struct ia64_sal_os_state *sos)
 {
 	pal_processor_state_info_t *psp = (pal_processor_state_info_t*)peidx_psp(peidx);
 
@@ -327,7 +326,7 @@ is_mca_global(peidx_table_t *peidx, pal_bus_check_info_t *pbci)
 	 * Therefore it is local MCA when rendezvous has not been requested.
 	 * Failed to rendezvous, the system must be down.
 	 */
-	switch (sal_to_os_handoff_state->imsto_rendez_state) {
+	switch (sos->rv_rc) {
 		case -1: /* SAL rendezvous unsuccessful */
 			return MCA_IS_GLOBAL;
 		case  0: /* SAL rendezvous not required */
@@ -388,7 +387,8 @@ is_mca_global(peidx_table_t *peidx, pal_bus_check_info_t *pbci)
  */
 
 static int
-recover_from_read_error(slidx_table_t *slidx, peidx_table_t *peidx, pal_bus_check_info_t *pbci)
+recover_from_read_error(slidx_table_t *slidx, peidx_table_t *peidx, pal_bus_check_info_t *pbci,
+			struct ia64_sal_os_state *sos)
 {
 	sal_log_mod_error_info_t *smei;
 	pal_min_state_area_t *pmsa;
@@ -426,7 +426,7 @@ recover_from_read_error(slidx_table_t *slidx, peidx_table_t *peidx, pal_bus_chec
 			 *  setup for resume to bottom half of MCA,
 			 * "mca_handler_bhhook"
 			 */
-			pmsa = (pal_min_state_area_t *)(sal_to_os_handoff_state->pal_min_state | (6ul<<61));
+			pmsa = sos->pal_min_state;
 			/* pass to bhhook as 1st argument (gr8) */
 			pmsa->pmsa_gr[8-1] = smei->target_identifier;
 			/* set interrupted return address (but no use) */
@@ -459,7 +459,8 @@ recover_from_read_error(slidx_table_t *slidx, peidx_table_t *peidx, pal_bus_chec
  */
 
 static int
-recover_from_platform_error(slidx_table_t *slidx, peidx_table_t *peidx, pal_bus_check_info_t *pbci)
+recover_from_platform_error(slidx_table_t *slidx, peidx_table_t *peidx, pal_bus_check_info_t *pbci,
+			    struct ia64_sal_os_state *sos)
 {
 	int status = 0;
 	pal_processor_state_info_t *psp = (pal_processor_state_info_t*)peidx_psp(peidx);
@@ -469,7 +470,7 @@ recover_from_platform_error(slidx_table_t *slidx, peidx_table_t *peidx, pal_bus_
 		case 1: /* partial read */
 		case 3: /* full line(cpu) read */
 		case 9: /* I/O space read */
-			status = recover_from_read_error(slidx, peidx, pbci);
+			status = recover_from_read_error(slidx, peidx, pbci, sos);
 			break;
 		case 0: /* unknown */
 		case 2: /* partial write */
@@ -508,7 +509,8 @@ recover_from_platform_error(slidx_table_t *slidx, peidx_table_t *peidx, pal_bus_
  */
 
 static int
-recover_from_processor_error(int platform, slidx_table_t *slidx, peidx_table_t *peidx, pal_bus_check_info_t *pbci)
+recover_from_processor_error(int platform, slidx_table_t *slidx, peidx_table_t *peidx, pal_bus_check_info_t *pbci,
+			     struct ia64_sal_os_state *sos)
 {
 	pal_processor_state_info_t *psp = (pal_processor_state_info_t*)peidx_psp(peidx);
 
@@ -545,7 +547,7 @@ recover_from_processor_error(int platform, slidx_table_t *slidx, peidx_table_t *
 	 * This means "there are some platform errors".
 	 */
 	if (platform) 
-		return recover_from_platform_error(slidx, peidx, pbci);
+		return recover_from_platform_error(slidx, peidx, pbci, sos);
 	/* 
 	 * On account of strange SAL error record, we cannot recover. 
 	 */
@@ -562,8 +564,7 @@ recover_from_processor_error(int platform, slidx_table_t *slidx, peidx_table_t *
 
 static int
 mca_try_to_recover(void *rec, 
-	ia64_mca_sal_to_os_state_t *sal_to_os_state,
-	ia64_mca_os_to_sal_state_t *os_to_sal_state)
+	struct ia64_sal_os_state *sos)
 {
 	int platform_err;
 	int n_proc_err;
@@ -571,10 +572,6 @@ mca_try_to_recover(void *rec,
 	peidx_table_t peidx;
 	pal_bus_check_info_t pbci;
 
-	/* handoff state from/to mca.c */
-	sal_to_os_handoff_state = sal_to_os_state;
-	os_to_sal_handoff_state = os_to_sal_state;
-
 	/* Make index of SAL error record */
 	platform_err = mca_make_slidx(rec, &slidx);
 
@@ -597,11 +594,11 @@ mca_try_to_recover(void *rec,
 	*((u64*)&pbci) = peidx_check_info(&peidx, bus_check, 0);
 
 	/* Check whether MCA is global or not */
-	if (is_mca_global(&peidx, &pbci))
+	if (is_mca_global(&peidx, &pbci, sos))
 		return 0;
 	
 	/* Try to recover a processor error */
-	return recover_from_processor_error(platform_err, &slidx, &peidx, &pbci);
+	return recover_from_processor_error(platform_err, &slidx, &peidx, &pbci, sos);
 }
 
 /*