[SPARC64]: Sun4v interrupt handling.

Sun4v has 4 interrupt queues: cpu, device, resumable errors,
and non-resumable errors.  A set of head/tail offset pointers
help maintain a work queue in physical memory.  The entries
are 64-bytes in size.

Each queue is allocated then registered with the hypervisor
as we bring cpus up.

The two error queues each get a kernel side buffer that we
use to quickly empty the main interrupt queue before we
call up to C code to log the event and possibly take evasive
action.
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

arch/sparc64/kernel/head.S

@@ -511,13 +511,14 @@ setup_tba:
 sparc64_boot_end:
 
 #include "systbls.S"
-#include "sun4v_tlb_miss.S"
 #include "ktlb.S"
 #include "tsb.S"
 #include "etrap.S"
 #include "rtrap.S"
 #include "winfixup.S"
 #include "entry.S"
+#include "sun4v_tlb_miss.S"
+#include "sun4v_ivec.S"
 
 /*
  * The following skip makes sure the trap table in ttable.S is aligned

arch/sparc64/kernel/irq.c

@@ -888,7 +888,19 @@ static void __cpuinit init_one_mondo(unsigned long *pa_ptr, unsigned long type)
 	}
 }
 
-/* Allocate and init the mondo queues for this cpu.  */
+static void __cpuinit init_one_kbuf(unsigned long *pa_ptr)
+{
+	unsigned long page = get_zeroed_page(GFP_ATOMIC);
+
+	if (!page) {
+		prom_printf("SUN4V: Error, cannot allocate kbuf page.\n");
+		prom_halt();
+	}
+
+	*pa_ptr = __pa(page);
+}
+
+/* Allocate and init the mondo and error queues for this cpu.  */
 void __cpuinit sun4v_init_mondo_queues(void)
 {
 	int cpu = hard_smp_processor_id();
@@ -897,7 +909,9 @@ void __cpuinit sun4v_init_mondo_queues(void)
 	init_one_mondo(&tb->cpu_mondo_pa, HV_CPU_QUEUE_CPU_MONDO);
 	init_one_mondo(&tb->dev_mondo_pa, HV_CPU_QUEUE_DEVICE_MONDO);
 	init_one_mondo(&tb->resum_mondo_pa, HV_CPU_QUEUE_RES_ERROR);
+	init_one_kbuf(&tb->resum_kernel_buf_pa);
 	init_one_mondo(&tb->nonresum_mondo_pa, HV_CPU_QUEUE_NONRES_ERROR);
+	init_one_kbuf(&tb->nonresum_kernel_buf_pa);
 }
 
 /* Only invoked on boot processor. */

arch/sparc64/kernel/sun4v_ivec.S

+/* sun4v_ivec.S: Sun4v interrupt vector handling.
+ *
+ * Copyright (C) 2006 <davem@davemloft.net>
+ */
+
+#include <asm/cpudata.h>
+#include <asm/intr_queue.h>
+
+	.text
+	.align	32
+
+sun4v_cpu_mondo:
+	/* Head offset in %g2, tail offset in %g4.
+	 * If they are the same, no work.
+	 */
+	mov	INTRQ_CPU_MONDO_HEAD, %g2
+	ldxa	[%g2] ASI_QUEUE, %g2
+	mov	INTRQ_CPU_MONDO_TAIL, %g4
+	ldxa	[%g4] ASI_QUEUE, %g4
+	cmp	%g2, %g4
+	be,pn	%xcc, sun4v_cpu_mondo_queue_empty
+	 nop
+
+	/* Get &trap_block[smp_processor_id()] into %g3.  */
+	__GET_CPUID(%g1)
+	sethi	%hi(trap_block), %g3
+	sllx	%g1, TRAP_BLOCK_SZ_SHIFT, %g7
+	or	%g3, %lo(trap_block), %g3
+	add	%g3, %g7, %g3
+
+	/* Get CPU mondo queue base phys address into %g7.  */
+	ldx	[%g3 + TRAP_PER_CPU_CPU_MONDO_PA], %g7
+
+	/* Now get the cross-call arguments and handler PC, same
+	 * layout as sun4u:
+	 *
+	 * 1st 64-bit word: low half is 32-bit PC, put into %g3 and jmpl to it
+	 *                  high half is context arg to MMU flushes, into %g5
+	 * 2nd 64-bit word: 64-bit arg, load into %g1
+	 * 3rd 64-bit word: 64-bit arg, load into %g7
+	 */
+	ldxa	[%g7 + %g2] ASI_PHYS_USE_EC, %g3
+	add	%g2, 0x8, %g2
+	srlx	%g3, 32, %g5
+	ldxa	[%g7 + %g2] ASI_PHYS_USE_EC, %g1
+	add	%g2, 0x8, %g2
+	srl	%g3, 0, %g3
+	ldxa	[%g7 + %g2] ASI_PHYS_USE_EC, %g7
+	add	%g2, 0x40 - 0x8 - 0x8, %g2
+
+	/* Update queue head pointer.  */
+	sethi	%hi(8192 - 1), %g4
+	or	%g4, %lo(8192 - 1), %g4
+	and	%g2, %g4, %g2
+
+	mov	INTRQ_CPU_MONDO_HEAD, %g4
+	stxa	%g2, [%g4] ASI_QUEUE
+	membar	#Sync
+
+	jmpl	%g3, %g0
+	 nop
+
+sun4v_cpu_mondo_queue_empty:
+	retry
+
+sun4v_dev_mondo:
+	/* Head offset in %g2, tail offset in %g4.  */
+	mov	INTRQ_DEVICE_MONDO_HEAD, %g2
+	ldxa	[%g2] ASI_QUEUE, %g2
+	mov	INTRQ_DEVICE_MONDO_TAIL, %g4
+	ldxa	[%g4] ASI_QUEUE, %g4
+	cmp	%g2, %g4
+	be,pn	%xcc, sun4v_dev_mondo_queue_empty
+	 nop
+
+	/* Get &trap_block[smp_processor_id()] into %g3.  */
+	__GET_CPUID(%g1)
+	sethi	%hi(trap_block), %g3
+	sllx	%g1, TRAP_BLOCK_SZ_SHIFT, %g7
+	or	%g3, %lo(trap_block), %g3
+	add	%g3, %g7, %g3
+
+	/* Get DEV mondo queue base phys address into %g5.  */
+	ldx	[%g3 + TRAP_PER_CPU_DEV_MONDO_PA], %g5
+
+	/* Load IVEC into %g3.  */
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	add	%g2, 0x40, %g2
+
+	/* XXX There can be a full 64-byte block of data here.
+	 * XXX This is how we can get at MSI vector data.
+	 * XXX Current we do not capture this, but when we do we'll
+	 * XXX need to add a 64-byte storage area in the struct ino_bucket
+	 * XXX or the struct irq_desc.
+	 */
+
+	/* Update queue head pointer, this frees up some registers.  */
+	sethi	%hi(8192 - 1), %g4
+	or	%g4, %lo(8192 - 1), %g4
+	and	%g2, %g4, %g2
+
+	mov	INTRQ_DEVICE_MONDO_HEAD, %g4
+	stxa	%g2, [%g4] ASI_QUEUE
+	membar	#Sync
+
+	/* Get &__irq_work[smp_processor_id()] into %g1.  */
+	sethi	%hi(__irq_work), %g4
+	sllx	%g1, 6, %g1
+	or	%g4, %lo(__irq_work), %g4
+	add	%g4, %g1, %g1
+
+	/* Get &ivector_table[IVEC] into %g4.  */
+	sethi	%hi(ivector_table), %g4
+	sllx	%g3, 5, %g3
+	or	%g4, %lo(ivector_table), %g4
+	add	%g4, %g3, %g4
+
+	/* Load IRQ %pil into %g5.  */
+	ldub	[%g4 + 0x04], %g5
+
+	/* Insert ivector_table[] entry into __irq_work[] queue.  */
+	sllx	%g5, 2, %g3
+	lduw	[%g1 + %g3], %g2	/* g2 = irq_work(cpu, pil) */
+	stw	%g2, [%g4 + 0x00]	/* bucket->irq_chain = g2 */
+	stw	%g4, [%g1 + %g3]	/* irq_work(cpu, pil) = bucket */
+
+	/* Signal the interrupt by setting (1 << pil) in %softint.  */
+	mov	1, %g2
+	sllx	%g2, %g5, %g2
+	wr	%g2, 0x0, %set_softint
+
+sun4v_dev_mondo_queue_empty:
+	retry
+
+sun4v_res_mondo:
+	/* Head offset in %g2, tail offset in %g4.  */
+	mov	INTRQ_RESUM_MONDO_HEAD, %g2
+	ldxa	[%g2] ASI_QUEUE, %g2
+	mov	INTRQ_RESUM_MONDO_TAIL, %g4
+	ldxa	[%g4] ASI_QUEUE, %g4
+	cmp	%g2, %g4
+	be,pn	%xcc, sun4v_res_mondo_queue_empty
+	 nop
+
+	/* Get &trap_block[smp_processor_id()] into %g3.  */
+	__GET_CPUID(%g1)
+	sethi	%hi(trap_block), %g3
+	sllx	%g1, TRAP_BLOCK_SZ_SHIFT, %g7
+	or	%g3, %lo(trap_block), %g3
+	add	%g3, %g7, %g3
+
+	/* Get RES mondo queue base phys address into %g5.  */
+	ldx	[%g3 + TRAP_PER_CPU_RESUM_MONDO_PA], %g5
+
+	/* Get RES kernel buffer base phys address into %g7.  */
+	ldx	[%g3 + TRAP_PER_CPU_RESUM_KBUF_PA], %g7
+
+	/* If the first word is non-zero, queue is full.  */
+	ldxa	[%g7 + %g2] ASI_PHYS_USE_EC, %g1
+	brnz,pn	%g1, sun4v_res_mondo_queue_full
+	 nop
+
+	/* Remember this entry's offset in %g1.  */
+	mov	%g2, %g1
+
+	/* Copy 64-byte queue entry into kernel buffer.  */
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+
+	/* Update queue head pointer.  */
+	sethi	%hi(8192 - 1), %g4
+	or	%g4, %lo(8192 - 1), %g4
+	and	%g2, %g4, %g2
+
+	mov	INTRQ_RESUM_MONDO_HEAD, %g4
+	stxa	%g2, [%g4] ASI_QUEUE
+	membar	#Sync
+
+	/* Disable interrupts and save register state so we can call
+	 * C code.  The etrap handling will leave %g4 in %l4 for us
+	 * when it's done.
+	 */
+	rdpr	%pil, %g2
+	wrpr	%g0, 15, %pil
+	mov	%g1, %g4
+	ba,pt	%xcc, etrap_irq
+	 rd	%pc, %g7
+
+	/* Log the event.  */
+	add	%sp, PTREGS_OFF, %o0
+	call	sun4v_resum_error
+	 mov	%l4, %o1
+
+	/* Return from trap.  */
+	ba,pt	%xcc, rtrap_irq
+	 nop
+
+sun4v_res_mondo_queue_empty:
+	retry
+
+sun4v_res_mondo_queue_full:
+	/* The queue is full, consolidate our damage by setting
+	 * the head equal to the tail.  We'll just trap again otherwise.
+	 * Call C code to log the event.
+	 */
+	mov	INTRQ_RESUM_MONDO_HEAD, %g2
+	stxa	%g4, [%g2] ASI_QUEUE
+	membar	#Sync
+
+	rdpr	%pil, %g2
+	wrpr	%g0, 15, %pil
+	ba,pt	%xcc, etrap_irq
+	 rd	%pc, %g7
+
+	call	sun4v_resum_overflow
+	 add	%sp, PTREGS_OFF, %o0
+
+	ba,pt	%xcc, rtrap_irq
+	 nop
+
+sun4v_nonres_mondo:
+	/* Head offset in %g2, tail offset in %g4.  */
+	mov	INTRQ_NONRESUM_MONDO_HEAD, %g2
+	ldxa	[%g2] ASI_QUEUE, %g2
+	mov	INTRQ_NONRESUM_MONDO_TAIL, %g4
+	ldxa	[%g4] ASI_QUEUE, %g4
+	cmp	%g2, %g4
+	be,pn	%xcc, sun4v_nonres_mondo_queue_empty
+	 nop
+
+	/* Get &trap_block[smp_processor_id()] into %g3.  */
+	__GET_CPUID(%g1)
+	sethi	%hi(trap_block), %g3
+	sllx	%g1, TRAP_BLOCK_SZ_SHIFT, %g7
+	or	%g3, %lo(trap_block), %g3
+	add	%g3, %g7, %g3
+
+	/* Get RES mondo queue base phys address into %g5.  */
+	ldx	[%g3 + TRAP_PER_CPU_NONRESUM_MONDO_PA], %g5
+
+	/* Get RES kernel buffer base phys address into %g7.  */
+	ldx	[%g3 + TRAP_PER_CPU_NONRESUM_KBUF_PA], %g7
+
+	/* If the first word is non-zero, queue is full.  */
+	ldxa	[%g7 + %g2] ASI_PHYS_USE_EC, %g1
+	brnz,pn	%g1, sun4v_nonres_mondo_queue_full
+	 nop
+
+	/* Remember this entry's offset in %g1.  */
+	mov	%g2, %g1
+
+	/* Copy 64-byte queue entry into kernel buffer.  */
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+	ldxa	[%g5 + %g2] ASI_PHYS_USE_EC, %g3
+	stxa	%g3, [%g7 + %g2] ASI_PHYS_USE_EC
+	add	%g2, 0x08, %g2
+
+	/* Update queue head pointer.  */
+	sethi	%hi(8192 - 1), %g4
+	or	%g4, %lo(8192 - 1), %g4
+	and	%g2, %g4, %g2
+
+	mov	INTRQ_NONRESUM_MONDO_HEAD, %g4
+	stxa	%g2, [%g4] ASI_QUEUE
+	membar	#Sync
+
+	/* Disable interrupts and save register state so we can call
+	 * C code.  The etrap handling will leave %g4 in %l4 for us
+	 * when it's done.
+	 */
+	rdpr	%pil, %g2
+	wrpr	%g0, 15, %pil
+	mov	%g1, %g4
+	ba,pt	%xcc, etrap_irq
+	 rd	%pc, %g7
+
+	/* Log the event.  */
+	add	%sp, PTREGS_OFF, %o0
+	call	sun4v_nonresum_error
+	 mov	%l4, %o1
+
+	/* Return from trap.  */
+	ba,pt	%xcc, rtrap_irq
+	 nop
+
+sun4v_nonres_mondo_queue_empty:
+	retry
+
+sun4v_nonres_mondo_queue_full:
+	/* The queue is full, consolidate our damage by setting
+	 * the head equal to the tail.  We'll just trap again otherwise.
+	 * Call C code to log the event.
+	 */
+	mov	INTRQ_NONRESUM_MONDO_HEAD, %g2
+	stxa	%g4, [%g2] ASI_QUEUE
+	membar	#Sync
+
+	rdpr	%pil, %g2
+	wrpr	%g0, 15, %pil
+	ba,pt	%xcc, etrap_irq
+	 rd	%pc, %g7
+
+	call	sun4v_nonresum_overflow
+	 add	%sp, PTREGS_OFF, %o0
+
+	ba,pt	%xcc, rtrap_irq
+	 nop

arch/sparc64/kernel/traps.c

@@ -1668,6 +1668,186 @@ void cheetah_plus_parity_error(int type, struct pt_regs *regs)
 	       regs->tpc);
 }
 
+struct sun4v_error_entry {
+	u64		err_handle;
+	u64		err_stick;
+
+	u32		err_type;
+#define SUN4V_ERR_TYPE_UNDEFINED	0
+#define SUN4V_ERR_TYPE_UNCORRECTED_RES	1
+#define SUN4V_ERR_TYPE_PRECISE_NONRES	2
+#define SUN4V_ERR_TYPE_DEFERRED_NONRES	3
+#define SUN4V_ERR_TYPE_WARNING_RES	4
+
+	u32		err_attrs;
+#define SUN4V_ERR_ATTRS_PROCESSOR	0x00000001
+#define SUN4V_ERR_ATTRS_MEMORY		0x00000002
+#define SUN4V_ERR_ATTRS_PIO		0x00000004
+#define SUN4V_ERR_ATTRS_INT_REGISTERS	0x00000008
+#define SUN4V_ERR_ATTRS_FPU_REGISTERS	0x00000010
+#define SUN4V_ERR_ATTRS_USER_MODE	0x01000000
+#define SUN4V_ERR_ATTRS_PRIV_MODE	0x02000000
+#define SUN4V_ERR_ATTRS_RES_QUEUE_FULL	0x80000000
+
+	u64		err_raddr;
+	u32		err_size;
+	u16		err_cpu;
+	u16		err_pad;
+};
+
+static atomic_t sun4v_resum_oflow_cnt = ATOMIC_INIT(0);
+static atomic_t sun4v_nonresum_oflow_cnt = ATOMIC_INIT(0);
+
+static const char *sun4v_err_type_to_str(u32 type)
+{
+	switch (type) {
+	case SUN4V_ERR_TYPE_UNDEFINED:
+		return "undefined";
+	case SUN4V_ERR_TYPE_UNCORRECTED_RES:
+		return "uncorrected resumable";
+	case SUN4V_ERR_TYPE_PRECISE_NONRES:
+		return "precise nonresumable";
+	case SUN4V_ERR_TYPE_DEFERRED_NONRES:
+		return "deferred nonresumable";
+	case SUN4V_ERR_TYPE_WARNING_RES:
+		return "warning resumable";
+	default:
+		return "unknown";
+	};
+}
+
+static void sun4v_log_error(struct sun4v_error_entry *ent, int cpu, const char *pfx, atomic_t *ocnt)
+{
+	int cnt;
+
+	printk("%s: Reporting on cpu %d\n", pfx, cpu);
+	printk("%s: err_handle[%lx] err_stick[%lx] err_type[%08x:%s]\n",
+	       pfx,
+	       ent->err_handle, ent->err_stick,
+	       ent->err_type,
+	       sun4v_err_type_to_str(ent->err_type));
+	printk("%s: err_attrs[%08x:%s %s %s %s %s %s %s %s]\n",
+	       pfx,
+	       ent->err_attrs,
+	       ((ent->err_attrs & SUN4V_ERR_ATTRS_PROCESSOR) ?
+		"processor" : ""),
+	       ((ent->err_attrs & SUN4V_ERR_ATTRS_MEMORY) ?
+		"memory" : ""),
+	       ((ent->err_attrs & SUN4V_ERR_ATTRS_PIO) ?
+		"pio" : ""),
+	       ((ent->err_attrs & SUN4V_ERR_ATTRS_INT_REGISTERS) ?
+		"integer-regs" : ""),
+	       ((ent->err_attrs & SUN4V_ERR_ATTRS_FPU_REGISTERS) ?
+		"fpu-regs" : ""),
+	       ((ent->err_attrs & SUN4V_ERR_ATTRS_USER_MODE) ?
+		"user" : ""),
+	       ((ent->err_attrs & SUN4V_ERR_ATTRS_PRIV_MODE) ?
+		"privileged" : ""),
+	       ((ent->err_attrs & SUN4V_ERR_ATTRS_RES_QUEUE_FULL) ?
+		"queue-full" : ""));
+	printk("%s: err_raddr[%016lx] err_size[%u] err_cpu[%u]\n",
+	       pfx,
+	       ent->err_raddr, ent->err_size, ent->err_cpu);
+
+	if ((cnt = atomic_read(ocnt)) != 0) {
+		atomic_set(ocnt, 0);
+		wmb();
+		printk("%s: Queue overflowed %d times.\n",
+		       pfx, cnt);
+	}
+}
+
+/* We run with %pil set to 15 and PSTATE_IE enabled in %pstate.
+ * Log the event and clear the first word of the entry.
+ */
+void sun4v_resum_error(struct pt_regs *regs, unsigned long offset)
+{
+	struct sun4v_error_entry *ent, local_copy;
+	struct trap_per_cpu *tb;
+	unsigned long paddr;
+	int cpu;
+
+	cpu = get_cpu();
+
+	tb = &trap_block[cpu];
+	paddr = tb->resum_kernel_buf_pa + offset;
+	ent = __va(paddr);
+
+	memcpy(&local_copy, ent, sizeof(struct sun4v_error_entry));
+
+	/* We have a local copy now, so release the entry.  */
+	ent->err_handle = 0;
+	wmb();
+
+	put_cpu();
+
+	sun4v_log_error(&local_copy, cpu,
+			KERN_ERR "RESUMABLE ERROR",
+			&sun4v_resum_oflow_cnt);
+}
+
+/* If we try to printk() we'll probably make matters worse, by trying
+ * to retake locks this cpu already holds or causing more errors. So
+ * just bump a counter, and we'll report these counter bumps above.
+ */
+void sun4v_resum_overflow(struct pt_regs *regs)
+{
+	atomic_inc(&sun4v_resum_oflow_cnt);
+}
+
+/* We run with %pil set to 15 and PSTATE_IE enabled in %pstate.
+ * Log the event, clear the first word of the entry, and die.
+ */
+void sun4v_nonresum_error(struct pt_regs *regs, unsigned long offset)
+{
+	struct sun4v_error_entry *ent, local_copy;
+	struct trap_per_cpu *tb;
+	unsigned long paddr;
+	int cpu;
+
+	cpu = get_cpu();
+
+	tb = &trap_block[cpu];
+	paddr = tb->nonresum_kernel_buf_pa + offset;
+	ent = __va(paddr);
+
+	memcpy(&local_copy, ent, sizeof(struct sun4v_error_entry));
+
+	/* We have a local copy now, so release the entry.  */
+	ent->err_handle = 0;
+	wmb();
+
+	put_cpu();
+
+#ifdef CONFIG_PCI
+	/* Check for the special PCI poke sequence. */
+	if (pci_poke_in_progress && pci_poke_cpu == cpu) {
+		pci_poke_faulted = 1;
+		regs->tpc += 4;
+		regs->tnpc = regs->tpc + 4;
+		return;
+	}
+#endif
+
+	sun4v_log_error(&local_copy, cpu,
+			KERN_EMERG "NON-RESUMABLE ERROR",
+			&sun4v_nonresum_oflow_cnt);
+
+	panic("Non-resumable error.");
+}
+
+/* If we try to printk() we'll probably make matters worse, by trying
+ * to retake locks this cpu already holds or causing more errors. So
+ * just bump a counter, and we'll report these counter bumps above.
+ */
+void sun4v_nonresum_overflow(struct pt_regs *regs)
+{
+	/* XXX Actually even this can make not that much sense.  Perhaps
+	 * XXX we should just pull the plug and panic directly from here?
+	 */
+	atomic_inc(&sun4v_nonresum_oflow_cnt);
+}
+
 void do_fpe_common(struct pt_regs *regs)
 {
 	if (regs->tstate & TSTATE_PRIV) {
@@ -2190,8 +2370,12 @@ void __init trap_init(void)
 	     offsetof(struct trap_per_cpu, dev_mondo_pa)) ||
 	    (TRAP_PER_CPU_RESUM_MONDO_PA !=
 	     offsetof(struct trap_per_cpu, resum_mondo_pa)) ||
+	    (TRAP_PER_CPU_RESUM_KBUF_PA !=
+	     offsetof(struct trap_per_cpu, resum_kernel_buf_pa)) ||
 	    (TRAP_PER_CPU_NONRESUM_MONDO_PA !=
 	     offsetof(struct trap_per_cpu, nonresum_mondo_pa)) ||
+	    (TRAP_PER_CPU_NONRESUM_KBUF_PA !=
+	     offsetof(struct trap_per_cpu, nonresum_kernel_buf_pa)) ||
 	    (TRAP_PER_CPU_FAULT_INFO !=
 	     offsetof(struct trap_per_cpu, fault_info)))
 		trap_per_cpu_offsets_are_bolixed_dave();

arch/sparc64/kernel/ttable.S

@@ -88,7 +88,10 @@ tl0_dcpe:	BTRAP(0x71)	/* D-cache Parity Error on Cheetah+ */
 tl0_icpe:	BTRAP(0x72)	/* I-cache Parity Error on Cheetah+ */
 tl0_resv073:	BTRAP(0x73) BTRAP(0x74) BTRAP(0x75)
 tl0_resv076:	BTRAP(0x76) BTRAP(0x77) BTRAP(0x78) BTRAP(0x79) BTRAP(0x7a) BTRAP(0x7b)
-tl0_resv07c:	BTRAP(0x7c) BTRAP(0x7d) BTRAP(0x7e) BTRAP(0x7f)
+tl0_cpu_mondo:	TRAP_NOSAVE(sun4v_cpu_mondo)
+tl0_dev_mondo:	TRAP_NOSAVE(sun4v_dev_mondo)
+tl0_res_mondo:	TRAP_NOSAVE(sun4v_res_mondo)
+tl0_nres_mondo:	TRAP_NOSAVE(sun4v_nonres_mondo)
 tl0_s0n:	SPILL_0_NORMAL
 tl0_s1n:	SPILL_1_NORMAL
 tl0_s2n:	SPILL_2_NORMAL

include/asm-sparc64/cpudata.h

@@ -53,16 +53,17 @@ DECLARE_PER_CPU(cpuinfo_sparc, __cpu_data);
  */
 struct thread_info;
 struct trap_per_cpu {
-/* D-cache line 1: Basic thread information */
+/* D-cache line 1: Basic thread information, cpu and device mondo queues */
 	struct thread_info	*thread;
 	unsigned long		pgd_paddr;
-	unsigned long		__pad1[2];
-
-/* D-cache line 2: Sun4V Mondo Queue pointers */
 	unsigned long		cpu_mondo_pa;
 	unsigned long		dev_mondo_pa;
+
+/* D-cache line 2: Error Mondo Queue and kernel buffer pointers */
 	unsigned long		resum_mondo_pa;
+	unsigned long		resum_kernel_buf_pa;
 	unsigned long		nonresum_mondo_pa;
+	unsigned long		nonresum_kernel_buf_pa;
 
 /* Dcache lines 3 and 4: Hypervisor Fault Status */
 	struct hv_fault_status	fault_info;
@@ -100,10 +101,12 @@ extern struct sun4v_2insn_patch_entry __sun4v_2insn_patch,
 
 #define TRAP_PER_CPU_THREAD		0x00
 #define TRAP_PER_CPU_PGD_PADDR		0x08
-#define TRAP_PER_CPU_CPU_MONDO_PA	0x20
-#define TRAP_PER_CPU_DEV_MONDO_PA	0x28
-#define TRAP_PER_CPU_RESUM_MONDO_PA	0x30
-#define TRAP_PER_CPU_NONRESUM_MONDO_PA	0x38
+#define TRAP_PER_CPU_CPU_MONDO_PA	0x10
+#define TRAP_PER_CPU_DEV_MONDO_PA	0x18
+#define TRAP_PER_CPU_RESUM_MONDO_PA	0x20
+#define TRAP_PER_CPU_RESUM_KBUF_PA	0x28
+#define TRAP_PER_CPU_NONRESUM_MONDO_PA	0x30
+#define TRAP_PER_CPU_NONRESUM_KBUF_PA	0x38
 #define TRAP_PER_CPU_FAULT_INFO		0x40
 
 #define TRAP_BLOCK_SZ_SHIFT		7
@@ -188,6 +191,9 @@ extern struct sun4v_2insn_patch_entry __sun4v_2insn_patch,
 
 #else
 
+#define __GET_CPUID(REG)				\
+	mov	0, REG;
+
 /* Uniprocessor versions, we know the cpuid is zero.  */
 #define TRAP_LOAD_PGD_PHYS(DEST, TMP)		\
 	sethi	%hi(trap_block), DEST;		\