diff --git a/arch/mips/Kconfig b/arch/mips/Kconfig
index fd2ff0698a85..bbd386f572d9 100644
--- a/arch/mips/Kconfig
+++ b/arch/mips/Kconfig
@@ -1568,6 +1568,20 @@ config MIPS_MT_FPAFF
 	depends on MIPS_MT
 	default y
 
+config MIPS_MT_SMTC_INSTANT_REPLAY
+	bool "Low-latency Dispatch of Deferred SMTC IPIs"
+	depends on MIPS_MT_SMTC
+	default y
+	help
+	  SMTC pseudo-interrupts between TCs are deferred and queued
+	  if the target TC is interrupt-inhibited (IXMT). In the first
+	  SMTC prototypes, these queued IPIs were serviced on return
+	  to user mode, or on entry into the kernel idle loop. The
+	  INSTANT_REPLAY option dispatches them as part of local_irq_restore()
+	  processing, which adds runtime overhead (hence the option to turn
+	  it off), but ensures that IPIs are handled promptly even under
+	  heavy I/O interrupt load.
+
 config MIPS_VPE_LOADER_TOM
 	bool "Load VPE program into memory hidden from linux"
 	depends on MIPS_VPE_LOADER
diff --git a/arch/mips/kernel/smtc.c b/arch/mips/kernel/smtc.c
index a8b387197d5b..44238ab2fc99 100644
--- a/arch/mips/kernel/smtc.c
+++ b/arch/mips/kernel/smtc.c
@@ -1017,6 +1017,33 @@ void setup_cross_vpe_interrupts(void)
  * SMTC-specific hacks invoked from elsewhere in the kernel.
  */
 
+void smtc_ipi_replay(void)
+{
+	/*
+	 * To the extent that we've ever turned interrupts off,
+	 * we may have accumulated deferred IPIs.  This is subtle.
+	 * If we use the smtc_ipi_qdepth() macro, we'll get an
+	 * exact number - but we'll also disable interrupts
+	 * and create a window of failure where a new IPI gets
+	 * queued after we test the depth but before we re-enable
+	 * interrupts. So long as IXMT never gets set, however,
+	 * we should be OK:  If we pick up something and dispatch
+	 * it here, that's great. If we see nothing, but concurrent
+	 * with this operation, another TC sends us an IPI, IXMT
+	 * is clear, and we'll handle it as a real pseudo-interrupt
+	 * and not a pseudo-pseudo interrupt.
+	 */
+	if (IPIQ[smp_processor_id()].depth > 0) {
+		struct smtc_ipi *pipi;
+		extern void self_ipi(struct smtc_ipi *);
+
+		while ((pipi = smtc_ipi_dq(&IPIQ[smp_processor_id()]))) {
+			self_ipi(pipi);
+			smtc_cpu_stats[smp_processor_id()].selfipis++;
+		}
+	}
+}
+
 void smtc_idle_loop_hook(void)
 {
 #ifdef SMTC_IDLE_HOOK_DEBUG
@@ -1113,29 +1140,14 @@ void smtc_idle_loop_hook(void)
 	if (pdb_msg != &id_ho_db_msg[0])
 		printk("CPU%d: %s", smp_processor_id(), id_ho_db_msg);
 #endif /* SMTC_IDLE_HOOK_DEBUG */
-	/*
-	 * To the extent that we've ever turned interrupts off,
-	 * we may have accumulated deferred IPIs.  This is subtle.
-	 * If we use the smtc_ipi_qdepth() macro, we'll get an
-	 * exact number - but we'll also disable interrupts
-	 * and create a window of failure where a new IPI gets
-	 * queued after we test the depth but before we re-enable
-	 * interrupts. So long as IXMT never gets set, however,
-	 * we should be OK:  If we pick up something and dispatch
-	 * it here, that's great. If we see nothing, but concurrent
-	 * with this operation, another TC sends us an IPI, IXMT
-	 * is clear, and we'll handle it as a real pseudo-interrupt
-	 * and not a pseudo-pseudo interrupt.
-	 */
-	if (IPIQ[smp_processor_id()].depth > 0) {
-		struct smtc_ipi *pipi;
-		extern void self_ipi(struct smtc_ipi *);
 
-		if ((pipi = smtc_ipi_dq(&IPIQ[smp_processor_id()])) != NULL) {
-			self_ipi(pipi);
-			smtc_cpu_stats[smp_processor_id()].selfipis++;
-		}
-	}
+	/*
+	 * Replay any accumulated deferred IPIs. If "Instant Replay"
+	 * is in use, there should never be any.
+	 */
+#ifndef CONFIG_MIPS_MT_SMTC_INSTANT_REPLAY
+	smtc_ipi_replay();
+#endif /* CONFIG_MIPS_MT_SMTC_INSTANT_REPLAY */
 }
 
 void smtc_soft_dump(void)
diff --git a/arch/mips/vr41xx/common/irq.c b/arch/mips/vr41xx/common/irq.c
index 397ba94cd7ec..16decf4ac2f4 100644
--- a/arch/mips/vr41xx/common/irq.c
+++ b/arch/mips/vr41xx/common/irq.c
@@ -1,7 +1,7 @@
 /*
  *  Interrupt handing routines for NEC VR4100 series.
  *
- *  Copyright (C) 2005  Yoichi Yuasa <yoichi_yuasa@tripeaks.co.jp>
+ *  Copyright (C) 2005-2007  Yoichi Yuasa <yoichi_yuasa@tripeaks.co.jp>
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -73,13 +73,19 @@ static void irq_dispatch(unsigned int irq)
 	if (cascade->get_irq != NULL) {
 		unsigned int source_irq = irq;
 		desc = irq_desc + source_irq;
-		desc->chip->ack(source_irq);
+		if (desc->chip->mask_ack)
+			desc->chip->mask_ack(source_irq);
+		else {
+			desc->chip->mask(source_irq);
+			desc->chip->ack(source_irq);
+		}
 		irq = cascade->get_irq(irq);
 		if (irq < 0)
 			atomic_inc(&irq_err_count);
 		else
 			irq_dispatch(irq);
-		desc->chip->end(source_irq);
+		if (!(desc->status & IRQ_DISABLED) && desc->chip->unmask)
+			desc->chip->unmask(source_irq);
 	} else
 		do_IRQ(irq);
 }
diff --git a/drivers/char/vr41xx_giu.c b/drivers/char/vr41xx_giu.c
index a744dad9cf45..0cea8d4907df 100644
--- a/drivers/char/vr41xx_giu.c
+++ b/drivers/char/vr41xx_giu.c
@@ -3,7 +3,7 @@
  *
  *  Copyright (C) 2002 MontaVista Software Inc.
  *	Author: Yoichi Yuasa <yyuasa@mvista.com or source@mvista.com>
- *  Copyright (C) 2003-2005  Yoichi Yuasa <yoichi_yuasa@tripeaks.co.jp>
+ *  Copyright (C) 2003-2007  Yoichi Yuasa <yoichi_yuasa@tripeaks.co.jp>
  *
  *  This program is free software; you can redistribute it and/or modify
  *  it under the terms of the GNU General Public License as published by
@@ -125,30 +125,17 @@ static inline uint16_t giu_clear(uint16_t offset, uint16_t clear)
 	return data;
 }
 
-static unsigned int startup_giuint_low_irq(unsigned int irq)
+static void ack_giuint_low(unsigned int irq)
 {
-	unsigned int pin;
-
-	pin = GPIO_PIN_OF_IRQ(irq);
-	giu_write(GIUINTSTATL, 1 << pin);
-	giu_set(GIUINTENL, 1 << pin);
-
-	return 0;
+	giu_write(GIUINTSTATL, 1 << GPIO_PIN_OF_IRQ(irq));
 }
 
-static void shutdown_giuint_low_irq(unsigned int irq)
+static void mask_giuint_low(unsigned int irq)
 {
 	giu_clear(GIUINTENL, 1 << GPIO_PIN_OF_IRQ(irq));
 }
 
-static void enable_giuint_low_irq(unsigned int irq)
-{
-	giu_set(GIUINTENL, 1 << GPIO_PIN_OF_IRQ(irq));
-}
-
-#define disable_giuint_low_irq	shutdown_giuint_low_irq
-
-static void ack_giuint_low_irq(unsigned int irq)
+static void mask_ack_giuint_low(unsigned int irq)
 {
 	unsigned int pin;
 
@@ -157,46 +144,30 @@ static void ack_giuint_low_irq(unsigned int irq)
 	giu_write(GIUINTSTATL, 1 << pin);
 }
 
-static void end_giuint_low_irq(unsigned int irq)
+static void unmask_giuint_low(unsigned int irq)
 {
-	if (!(irq_desc[irq].status & (IRQ_DISABLED | IRQ_INPROGRESS)))
-		giu_set(GIUINTENL, 1 << GPIO_PIN_OF_IRQ(irq));
+	giu_set(GIUINTENL, 1 << GPIO_PIN_OF_IRQ(irq));
 }
 
-static struct hw_interrupt_type giuint_low_irq_type = {
-	.typename	= "GIUINTL",
-	.startup	= startup_giuint_low_irq,
-	.shutdown	= shutdown_giuint_low_irq,
-	.enable		= enable_giuint_low_irq,
-	.disable	= disable_giuint_low_irq,
-	.ack		= ack_giuint_low_irq,
-	.end		= end_giuint_low_irq,
+static struct irq_chip giuint_low_irq_chip = {
+	.name		= "GIUINTL",
+	.ack		= ack_giuint_low,
+	.mask		= mask_giuint_low,
+	.mask_ack	= mask_ack_giuint_low,
+	.unmask		= unmask_giuint_low,
 };
 
-static unsigned int startup_giuint_high_irq(unsigned int irq)
+static void ack_giuint_high(unsigned int irq)
 {
-	unsigned int pin;
-
-	pin = GPIO_PIN_OF_IRQ(irq) - GIUINT_HIGH_OFFSET;
-	giu_write(GIUINTSTATH, 1 << pin);
-	giu_set(GIUINTENH, 1 << pin);
-
-	return 0;
+	giu_write(GIUINTSTATH, 1 << (GPIO_PIN_OF_IRQ(irq) - GIUINT_HIGH_OFFSET));
 }
 
-static void shutdown_giuint_high_irq(unsigned int irq)
+static void mask_giuint_high(unsigned int irq)
 {
 	giu_clear(GIUINTENH, 1 << (GPIO_PIN_OF_IRQ(irq) - GIUINT_HIGH_OFFSET));
 }
 
-static void enable_giuint_high_irq(unsigned int irq)
-{
-	giu_set(GIUINTENH, 1 << (GPIO_PIN_OF_IRQ(irq) - GIUINT_HIGH_OFFSET));
-}
-
-#define disable_giuint_high_irq	shutdown_giuint_high_irq
-
-static void ack_giuint_high_irq(unsigned int irq)
+static void mask_ack_giuint_high(unsigned int irq)
 {
 	unsigned int pin;
 
@@ -205,20 +176,17 @@ static void ack_giuint_high_irq(unsigned int irq)
 	giu_write(GIUINTSTATH, 1 << pin);
 }
 
-static void end_giuint_high_irq(unsigned int irq)
+static void unmask_giuint_high(unsigned int irq)
 {
-	if (!(irq_desc[irq].status & (IRQ_DISABLED | IRQ_INPROGRESS)))
-		giu_set(GIUINTENH, 1 << (GPIO_PIN_OF_IRQ(irq) - GIUINT_HIGH_OFFSET));
+	giu_set(GIUINTENH, 1 << (GPIO_PIN_OF_IRQ(irq) - GIUINT_HIGH_OFFSET));
 }
 
-static struct hw_interrupt_type giuint_high_irq_type = {
-	.typename	= "GIUINTH",
-	.startup	= startup_giuint_high_irq,
-	.shutdown	= shutdown_giuint_high_irq,
-	.enable		= enable_giuint_high_irq,
-	.disable	= disable_giuint_high_irq,
-	.ack		= ack_giuint_high_irq,
-	.end		= end_giuint_high_irq,
+static struct irq_chip giuint_high_irq_chip = {
+	.name		= "GIUINTH",
+	.ack		= ack_giuint_high,
+	.mask		= mask_giuint_high,
+	.mask_ack	= mask_ack_giuint_high,
+	.unmask		= unmask_giuint_high,
 };
 
 static int giu_get_irq(unsigned int irq)
@@ -282,9 +250,15 @@ void vr41xx_set_irq_trigger(unsigned int pin, irq_trigger_t trigger, irq_signal_
 					break;
 				}
 			}
+			set_irq_chip_and_handler(GIU_IRQ(pin),
+			                         &giuint_low_irq_chip,
+			                         handle_edge_irq);
 		} else {
 			giu_clear(GIUINTTYPL, mask);
 			giu_clear(GIUINTHTSELL, mask);
+			set_irq_chip_and_handler(GIU_IRQ(pin),
+			                         &giuint_low_irq_chip,
+			                         handle_level_irq);
 		}
 		giu_write(GIUINTSTATL, mask);
 	} else if (pin < GIUINT_HIGH_MAX) {
@@ -311,9 +285,15 @@ void vr41xx_set_irq_trigger(unsigned int pin, irq_trigger_t trigger, irq_signal_
 					break;
 				}
 			}
+			set_irq_chip_and_handler(GIU_IRQ(pin),
+			                         &giuint_high_irq_chip,
+			                         handle_edge_irq);
 		} else {
 			giu_clear(GIUINTTYPH, mask);
 			giu_clear(GIUINTHTSELH, mask);
+			set_irq_chip_and_handler(GIU_IRQ(pin),
+			                         &giuint_high_irq_chip,
+			                         handle_level_irq);
 		}
 		giu_write(GIUINTSTATH, mask);
 	}
@@ -617,10 +597,11 @@ static const struct file_operations gpio_fops = {
 static int __devinit giu_probe(struct platform_device *dev)
 {
 	unsigned long start, size, flags = 0;
-	unsigned int nr_pins = 0;
+	unsigned int nr_pins = 0, trigger, i, pin;
 	struct resource *res1, *res2 = NULL;
 	void *base;
-	int retval, i;
+	struct irq_chip *chip;
+	int retval;
 
 	switch (current_cpu_data.cputype) {
 	case CPU_VR4111:
@@ -688,11 +669,20 @@ static int __devinit giu_probe(struct platform_device *dev)
 	giu_write(GIUINTENL, 0);
 	giu_write(GIUINTENH, 0);
 
+	trigger = giu_read(GIUINTTYPH) << 16;
+	trigger |= giu_read(GIUINTTYPL);
 	for (i = GIU_IRQ_BASE; i <= GIU_IRQ_LAST; i++) {
-		if (i < GIU_IRQ(GIUINT_HIGH_OFFSET))
-			irq_desc[i].chip = &giuint_low_irq_type;
+		pin = GPIO_PIN_OF_IRQ(i);
+		if (pin < GIUINT_HIGH_OFFSET)
+			chip = &giuint_low_irq_chip;
 		else
-			irq_desc[i].chip = &giuint_high_irq_type;
+			chip = &giuint_high_irq_chip;
+
+		if (trigger & (1 << pin))
+			set_irq_chip_and_handler(i, chip, handle_edge_irq);
+		else
+			set_irq_chip_and_handler(i, chip, handle_level_irq);
+
 	}
 
 	return cascade_irq(GIUINT_IRQ, giu_get_irq);
diff --git a/include/asm-mips/irqflags.h b/include/asm-mips/irqflags.h
index 46bf5de5ac72..af3b07dfad4b 100644
--- a/include/asm-mips/irqflags.h
+++ b/include/asm-mips/irqflags.h
@@ -15,6 +15,27 @@
 
 #include <asm/hazards.h>
 
+/*
+ * CONFIG_MIPS_MT_SMTC_INSTANT_REPLAY does prompt replay of deferred IPIs,
+ * at the cost of branch and call overhead on each local_irq_restore()
+ */
+
+#ifdef CONFIG_MIPS_MT_SMTC_INSTANT_REPLAY
+
+extern void smtc_ipi_replay(void);
+
+#define irq_restore_epilog(flags)				\
+do {								\
+	if (!(flags & 0x0400))					\
+		smtc_ipi_replay();				\
+} while (0)
+
+#else
+
+#define irq_restore_epilog(ignore) do { } while (0)
+
+#endif /* CONFIG_MIPS_MT_SMTC_INSTANT_REPLAY */
+
 __asm__ (
 	"	.macro	raw_local_irq_enable				\n"
 	"	.set	push						\n"
@@ -193,6 +214,7 @@ do {									\
 		: "=r" (__tmp1)						\
 		: "0" (flags)						\
 		: "memory");						\
+	irq_restore_epilog(flags);					\
 } while(0)
 
 static inline int raw_irqs_disabled_flags(unsigned long flags)