Merge branch 'upstream' of git://ftp.linux-mips.org/pub/scm/upstream-linus

* 'upstream' of git://ftp.linux-mips.org/pub/scm/upstream-linus:
  [MIPS] SMTC: Fix SMTC dyntick support.
  [MIPS] SMTC: Close tiny holes in the SMTC IPI replay system.
  [MIPS] SMTC: Fix holes in SMTC and FPU affinity support.
  [MIPS] SMTC: Build fix: Fix filename in Makefile
  [MIPS] Build fix: Fix irq flags type

arch/mips/Kconfig

@@ -1403,7 +1403,6 @@ config MIPS_MT_SMTC
 	depends on CPU_MIPS32_R2
 	#depends on CPU_MIPS64_R2		# once there is hardware ...
 	depends on SYS_SUPPORTS_MULTITHREADING
-	select GENERIC_CLOCKEVENTS_BROADCAST
 	select CPU_MIPSR2_IRQ_VI
 	select CPU_MIPSR2_IRQ_EI
 	select MIPS_MT
@@ -1451,32 +1450,17 @@ config MIPS_VPE_LOADER
 	  Includes a loader for loading an elf relocatable object
 	  onto another VPE and running it.
 
-config MIPS_MT_SMTC_INSTANT_REPLAY
-	bool "Low-latency Dispatch of Deferred SMTC IPIs"
-	depends on MIPS_MT_SMTC && !PREEMPT
-	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_MT_SMTC_IM_BACKSTOP
 	bool "Use per-TC register bits as backstop for inhibited IM bits"
 	depends on MIPS_MT_SMTC
-	default y
+	default n
 	help
 	  To support multiple TC microthreads acting as "CPUs" within
 	  a VPE, VPE-wide interrupt mask bits must be specially manipulated
 	  during interrupt handling. To support legacy drivers and interrupt
 	  controller management code, SMTC has a "backstop" to track and
 	  if necessary restore the interrupt mask. This has some performance
-	  impact on interrupt service overhead. Disable it only if you know
-	  what you are doing.
+	  impact on interrupt service overhead.
 
 config MIPS_MT_SMTC_IRQAFF
 	bool "Support IRQ affinity API"
@@ -1486,10 +1470,8 @@ config MIPS_MT_SMTC_IRQAFF
 	  Enables SMP IRQ affinity API (/proc/irq/*/smp_affinity, etc.)
 	  for SMTC Linux kernel. Requires platform support, of which
 	  an example can be found in the MIPS kernel i8259 and Malta
-	  platform code.  It is recommended that MIPS_MT_SMTC_INSTANT_REPLAY
-	  be enabled if MIPS_MT_SMTC_IRQAFF is used. Adds overhead to
-	  interrupt dispatch, and should be used only if you know what
-	  you are doing.
+	  platform code.  Adds some overhead to interrupt dispatch, and
+	  should be used only if you know what you are doing.
 
 config MIPS_VPE_LOADER_TOM
 	bool "Load VPE program into memory hidden from linux"

arch/mips/kernel/Makefile

@@ -10,6 +10,7 @@ obj-y		+= cpu-probe.o branch.o entry.o genex.o irq.o process.o \
 
 obj-$(CONFIG_CEVT_BCM1480)	+= cevt-bcm1480.o
 obj-$(CONFIG_CEVT_R4K)		+= cevt-r4k.o
+obj-$(CONFIG_MIPS_MT_SMTC)	+= cevt-smtc.o
 obj-$(CONFIG_CEVT_DS1287)	+= cevt-ds1287.o
 obj-$(CONFIG_CEVT_GT641XX)	+= cevt-gt641xx.o
 obj-$(CONFIG_CEVT_SB1250)	+= cevt-sb1250.o

arch/mips/kernel/cevt-r4k.c

@@ -12,6 +12,14 @@
 
 #include <asm/smtc_ipi.h>
 #include <asm/time.h>
+#include <asm/cevt-r4k.h>
+
+/*
+ * The SMTC Kernel for the 34K, 1004K, et. al. replaces several
+ * of these routines with SMTC-specific variants.
+ */
+
+#ifndef CONFIG_MIPS_MT_SMTC
 
 static int mips_next_event(unsigned long delta,
                            struct clock_event_device *evt)
@@ -19,60 +27,27 @@ static int mips_next_event(unsigned long delta,
 	unsigned int cnt;
 	int res;
 
-#ifdef CONFIG_MIPS_MT_SMTC
-	{
-	unsigned long flags, vpflags;
-	local_irq_save(flags);
-	vpflags = dvpe();
-#endif
 	cnt = read_c0_count();
 	cnt += delta;
 	write_c0_compare(cnt);
 	res = ((int)(read_c0_count() - cnt) > 0) ? -ETIME : 0;
-#ifdef CONFIG_MIPS_MT_SMTC
-	evpe(vpflags);
-	local_irq_restore(flags);
-	}
-#endif
 	return res;
 }
 
-static void mips_set_mode(enum clock_event_mode mode,
+#endif /* CONFIG_MIPS_MT_SMTC */
+
+void mips_set_clock_mode(enum clock_event_mode mode,
 				struct clock_event_device *evt)
 {
 	/* Nothing to do ...  */
 }
 
-static DEFINE_PER_CPU(struct clock_event_device, mips_clockevent_device);
-static int cp0_timer_irq_installed;
+DEFINE_PER_CPU(struct clock_event_device, mips_clockevent_device);
+int cp0_timer_irq_installed;
 
-/*
- * Timer ack for an R4k-compatible timer of a known frequency.
- */
-static void c0_timer_ack(void)
-{
-	write_c0_compare(read_c0_compare());
-}
+#ifndef CONFIG_MIPS_MT_SMTC
 
-/*
- * Possibly handle a performance counter interrupt.
- * Return true if the timer interrupt should not be checked
- */
-static inline int handle_perf_irq(int r2)
-{
-	/*
-	 * The performance counter overflow interrupt may be shared with the
-	 * timer interrupt (cp0_perfcount_irq < 0). If it is and a
-	 * performance counter has overflowed (perf_irq() == IRQ_HANDLED)
-	 * and we can't reliably determine if a counter interrupt has also
-	 * happened (!r2) then don't check for a timer interrupt.
-	 */
-	return (cp0_perfcount_irq < 0) &&
-		perf_irq() == IRQ_HANDLED &&
-		!r2;
-}
-
-static irqreturn_t c0_compare_interrupt(int irq, void *dev_id)
+irqreturn_t c0_compare_interrupt(int irq, void *dev_id)
 {
 	const int r2 = cpu_has_mips_r2;
 	struct clock_event_device *cd;
@@ -93,12 +68,8 @@ static irqreturn_t c0_compare_interrupt(int irq, void *dev_id)
 	 * interrupt.  Being the paranoiacs we are we check anyway.
 	 */
 	if (!r2 || (read_c0_cause() & (1 << 30))) {
-		c0_timer_ack();
-#ifdef CONFIG_MIPS_MT_SMTC
-		if (cpu_data[cpu].vpe_id)
-			goto out;
-		cpu = 0;
-#endif
+		/* Clear Count/Compare Interrupt */
+		write_c0_compare(read_c0_compare());
 		cd = &per_cpu(mips_clockevent_device, cpu);
 		cd->event_handler(cd);
 	}
@@ -107,65 +78,16 @@ out:
 	return IRQ_HANDLED;
 }
 
-static struct irqaction c0_compare_irqaction = {
+#endif /* Not CONFIG_MIPS_MT_SMTC */
+
+struct irqaction c0_compare_irqaction = {
 	.handler = c0_compare_interrupt,
-#ifdef CONFIG_MIPS_MT_SMTC
-	.flags = IRQF_DISABLED,
-#else
 	.flags = IRQF_DISABLED | IRQF_PERCPU,
-#endif
 	.name = "timer",
 };
 
-#ifdef CONFIG_MIPS_MT_SMTC
-DEFINE_PER_CPU(struct clock_event_device, smtc_dummy_clockevent_device);
-
-static void smtc_set_mode(enum clock_event_mode mode,
-                          struct clock_event_device *evt)
-{
-}
-
-static void mips_broadcast(cpumask_t mask)
-{
-	unsigned int cpu;
-
-	for_each_cpu_mask(cpu, mask)
-		smtc_send_ipi(cpu, SMTC_CLOCK_TICK, 0);
-}
 
-static void setup_smtc_dummy_clockevent_device(void)
-{
-	//uint64_t mips_freq = mips_hpt_^frequency;
-	unsigned int cpu = smp_processor_id();
-	struct clock_event_device *cd;
-
-	cd = &per_cpu(smtc_dummy_clockevent_device, cpu);
-
-	cd->name		= "SMTC";
-	cd->features		= CLOCK_EVT_FEAT_DUMMY;
-
-	/* Calculate the min / max delta */
-	cd->mult	= 0; //div_sc((unsigned long) mips_freq, NSEC_PER_SEC, 32);
-	cd->shift		= 0; //32;
-	cd->max_delta_ns	= 0; //clockevent_delta2ns(0x7fffffff, cd);
-	cd->min_delta_ns	= 0; //clockevent_delta2ns(0x30, cd);
-
-	cd->rating		= 200;
-	cd->irq			= 17; //-1;
-//	if (cpu)
-//		cd->cpumask	= CPU_MASK_ALL; // cpumask_of_cpu(cpu);
-//	else
-		cd->cpumask	= cpumask_of_cpu(cpu);
-
-	cd->set_mode		= smtc_set_mode;
-
-	cd->broadcast		= mips_broadcast;
-
-	clockevents_register_device(cd);
-}
-#endif
-
-static void mips_event_handler(struct clock_event_device *dev)
+void mips_event_handler(struct clock_event_device *dev)
 {
 }
 
@@ -177,7 +99,23 @@ static int c0_compare_int_pending(void)
 	return (read_c0_cause() >> cp0_compare_irq) & 0x100;
 }
 
-static int c0_compare_int_usable(void)
+/*
+ * Compare interrupt can be routed and latched outside the core,
+ * so a single execution hazard barrier may not be enough to give
+ * it time to clear as seen in the Cause register.  4 time the
+ * pipeline depth seems reasonably conservative, and empirically
+ * works better in configurations with high CPU/bus clock ratios.
+ */
+
+#define compare_change_hazard() \
+	do { \
+		irq_disable_hazard(); \
+		irq_disable_hazard(); \
+		irq_disable_hazard(); \
+		irq_disable_hazard(); \
+	} while (0)
+
+int c0_compare_int_usable(void)
 {
 	unsigned int delta;
 	unsigned int cnt;
@@ -187,7 +125,7 @@ static int c0_compare_int_usable(void)
 	 */
 	if (c0_compare_int_pending()) {
 		write_c0_compare(read_c0_count());
-		irq_disable_hazard();
+		compare_change_hazard();
 		if (c0_compare_int_pending())
 			return 0;
 	}
@@ -196,7 +134,7 @@ static int c0_compare_int_usable(void)
 		cnt = read_c0_count();
 		cnt += delta;
 		write_c0_compare(cnt);
-		irq_disable_hazard();
+		compare_change_hazard();
 		if ((int)(read_c0_count() - cnt) < 0)
 		    break;
 		/* increase delta if the timer was already expired */
@@ -205,11 +143,12 @@ static int c0_compare_int_usable(void)
 	while ((int)(read_c0_count() - cnt) <= 0)
 		;	/* Wait for expiry  */
 
+	compare_change_hazard();
 	if (!c0_compare_int_pending())
 		return 0;
 
 	write_c0_compare(read_c0_count());
-	irq_disable_hazard();
+	compare_change_hazard();
 	if (c0_compare_int_pending())
 		return 0;
 
@@ -219,6 +158,8 @@ static int c0_compare_int_usable(void)
 	return 1;
 }
 
+#ifndef CONFIG_MIPS_MT_SMTC
+
 int __cpuinit mips_clockevent_init(void)
 {
 	uint64_t mips_freq = mips_hpt_frequency;
@@ -229,17 +170,6 @@ int __cpuinit mips_clockevent_init(void)
 	if (!cpu_has_counter || !mips_hpt_frequency)
 		return -ENXIO;
 
-#ifdef CONFIG_MIPS_MT_SMTC
-	setup_smtc_dummy_clockevent_device();
-
-	/*
-	 * On SMTC we only register VPE0's compare interrupt as clockevent
-	 * device.
-	 */
-	if (cpu)
-		return 0;
-#endif
-
 	if (!c0_compare_int_usable())
 		return -ENXIO;
 
@@ -265,13 +195,9 @@ int __cpuinit mips_clockevent_init(void)
 
 	cd->rating		= 300;
 	cd->irq			= irq;
-#ifdef CONFIG_MIPS_MT_SMTC
-	cd->cpumask		= CPU_MASK_ALL;
-#else
 	cd->cpumask		= cpumask_of_cpu(cpu);
-#endif
 	cd->set_next_event	= mips_next_event;
-	cd->set_mode		= mips_set_mode;
+	cd->set_mode		= mips_set_clock_mode;
 	cd->event_handler	= mips_event_handler;
 
 	clockevents_register_device(cd);
@@ -281,12 +207,9 @@ int __cpuinit mips_clockevent_init(void)
 
 	cp0_timer_irq_installed = 1;
 
-#ifdef CONFIG_MIPS_MT_SMTC
-#define CPUCTR_IMASKBIT (0x100 << cp0_compare_irq)
-	setup_irq_smtc(irq, &c0_compare_irqaction, CPUCTR_IMASKBIT);
-#else
 	setup_irq(irq, &c0_compare_irqaction);
-#endif
 
 	return 0;
 }
+
+#endif /* Not CONFIG_MIPS_MT_SMTC */

arch/mips/kernel/cevt-smtc.c

+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2007 MIPS Technologies, Inc.
+ * Copyright (C) 2007 Ralf Baechle <ralf@linux-mips.org>
+ * Copyright (C) 2008 Kevin D. Kissell, Paralogos sarl
+ */
+#include <linux/clockchips.h>
+#include <linux/interrupt.h>
+#include <linux/percpu.h>
+
+#include <asm/smtc_ipi.h>
+#include <asm/time.h>
+#include <asm/cevt-r4k.h>
+
+/*
+ * Variant clock event timer support for SMTC on MIPS 34K, 1004K
+ * or other MIPS MT cores.
+ *
+ * Notes on SMTC Support:
+ *
+ * SMTC has multiple microthread TCs pretending to be Linux CPUs.
+ * But there's only one Count/Compare pair per VPE, and Compare
+ * interrupts are taken opportunisitically by available TCs
+ * bound to the VPE with the Count register.  The new timer
+ * framework provides for global broadcasts, but we really
+ * want VPE-level multicasts for best behavior. So instead
+ * of invoking the high-level clock-event broadcast code,
+ * this version of SMTC support uses the historical SMTC
+ * multicast mechanisms "under the hood", appearing to the
+ * generic clock layer as if the interrupts are per-CPU.
+ *
+ * The approach taken here is to maintain a set of NR_CPUS
+ * virtual timers, and track which "CPU" needs to be alerted
+ * at each event.
+ *
+ * It's unlikely that we'll see a MIPS MT core with more than
+ * 2 VPEs, but we *know* that we won't need to handle more
+ * VPEs than we have "CPUs".  So NCPUs arrays of NCPUs elements
+ * is always going to be overkill, but always going to be enough.
+ */
+
+unsigned long smtc_nexttime[NR_CPUS][NR_CPUS];
+static int smtc_nextinvpe[NR_CPUS];
+
+/*
+ * Timestamps stored are absolute values to be programmed
+ * into Count register.  Valid timestamps will never be zero.
+ * If a Zero Count value is actually calculated, it is converted
+ * to be a 1, which will introduce 1 or two CPU cycles of error
+ * roughly once every four billion events, which at 1000 HZ means
+ * about once every 50 days.  If that's actually a problem, one
+ * could alternate squashing 0 to 1 and to -1.
+ */
+
+#define MAKEVALID(x) (((x) == 0L) ? 1L : (x))
+#define ISVALID(x) ((x) != 0L)
+
+/*
+ * Time comparison is subtle, as it's really truncated
+ * modular arithmetic.
+ */
+
+#define IS_SOONER(a, b, reference) \
+    (((a) - (unsigned long)(reference)) < ((b) - (unsigned long)(reference)))
+
+/*
+ * CATCHUP_INCREMENT, used when the function falls behind the counter.
+ * Could be an increasing function instead of a constant;
+ */
+
+#define CATCHUP_INCREMENT 64
+
+static int mips_next_event(unsigned long delta,
+				struct clock_event_device *evt)
+{
+	unsigned long flags;
+	unsigned int mtflags;
+	unsigned long timestamp, reference, previous;
+	unsigned long nextcomp = 0L;
+	int vpe = current_cpu_data.vpe_id;
+	int cpu = smp_processor_id();
+	local_irq_save(flags);
+	mtflags = dmt();
+
+	/*
+	 * Maintain the per-TC virtual timer
+	 * and program the per-VPE shared Count register
+	 * as appropriate here...
+	 */
+	reference = (unsigned long)read_c0_count();
+	timestamp = MAKEVALID(reference + delta);
+	/*
+	 * To really model the clock, we have to catch the case
+	 * where the current next-in-VPE timestamp is the old
+	 * timestamp for the calling CPE, but the new value is
+	 * in fact later.  In that case, we have to do a full
+	 * scan and discover the new next-in-VPE CPU id and
+	 * timestamp.
+	 */
+	previous = smtc_nexttime[vpe][cpu];
+	if (cpu == smtc_nextinvpe[vpe] && ISVALID(previous)
+	    && IS_SOONER(previous, timestamp, reference)) {
+		int i;
+		int soonest = cpu;
+
+		/*
+		 * Update timestamp array here, so that new
+		 * value gets considered along with those of
+		 * other virtual CPUs on the VPE.
+		 */
+		smtc_nexttime[vpe][cpu] = timestamp;
+		for_each_online_cpu(i) {
+			if (ISVALID(smtc_nexttime[vpe][i])
+			    && IS_SOONER(smtc_nexttime[vpe][i],
+				smtc_nexttime[vpe][soonest], reference)) {
+				    soonest = i;
+			}
+		}
+		smtc_nextinvpe[vpe] = soonest;
+		nextcomp = smtc_nexttime[vpe][soonest];
+	/*
+	 * Otherwise, we don't have to process the whole array rank,
+	 * we just have to see if the event horizon has gotten closer.
+	 */
+	} else {
+		if (!ISVALID(smtc_nexttime[vpe][smtc_nextinvpe[vpe]]) ||
+		    IS_SOONER(timestamp,
+			smtc_nexttime[vpe][smtc_nextinvpe[vpe]], reference)) {
+			    smtc_nextinvpe[vpe] = cpu;
+			    nextcomp = timestamp;
+		}
+		/*
+		 * Since next-in-VPE may me the same as the executing
+		 * virtual CPU, we update the array *after* checking
+		 * its value.
+		 */
+		smtc_nexttime[vpe][cpu] = timestamp;
+	}
+
+	/*
+	 * It may be that, in fact, we don't need to update Compare,
+	 * but if we do, we want to make sure we didn't fall into
+	 * a crack just behind Count.
+	 */
+	if (ISVALID(nextcomp)) {
+		write_c0_compare(nextcomp);
+		ehb();
+		/*
+		 * We never return an error, we just make sure
+		 * that we trigger the handlers as quickly as
+		 * we can if we fell behind.
+		 */
+		while ((nextcomp - (unsigned long)read_c0_count())
+			> (unsigned long)LONG_MAX) {
+			nextcomp += CATCHUP_INCREMENT;
+			write_c0_compare(nextcomp);
+			ehb();
+		}
+	}
+	emt(mtflags);
+	local_irq_restore(flags);
+	return 0;
+}
+
+
+void smtc_distribute_timer(int vpe)
+{
+	unsigned long flags;
+	unsigned int mtflags;
+	int cpu;
+	struct clock_event_device *cd;
+	unsigned long nextstamp = 0L;
+	unsigned long reference;
+
+
+repeat:
+	for_each_online_cpu(cpu) {
+	    /*
+	     * Find virtual CPUs within the current VPE who have
+	     * unserviced timer requests whose time is now past.
+	     */
+	    local_irq_save(flags);
+	    mtflags = dmt();
+	    if (cpu_data[cpu].vpe_id == vpe &&
+		ISVALID(smtc_nexttime[vpe][cpu])) {
+		reference = (unsigned long)read_c0_count();
+		if ((smtc_nexttime[vpe][cpu] - reference)
+			 > (unsigned long)LONG_MAX) {
+			    smtc_nexttime[vpe][cpu] = 0L;
+			    emt(mtflags);
+			    local_irq_restore(flags);
+			    /*
+			     * We don't send IPIs to ourself.
+			     */
+			    if (cpu != smp_processor_id()) {
+				smtc_send_ipi(cpu, SMTC_CLOCK_TICK, 0);
+			    } else {
+				cd = &per_cpu(mips_clockevent_device, cpu);
+				cd->event_handler(cd);
+			    }
+		} else {
+			/* Local to VPE but Valid Time not yet reached. */
+			if (!ISVALID(nextstamp) ||
+			    IS_SOONER(smtc_nexttime[vpe][cpu], nextstamp,
+			    reference)) {
+				smtc_nextinvpe[vpe] = cpu;
+				nextstamp = smtc_nexttime[vpe][cpu];
+			}
+			emt(mtflags);
+			local_irq_restore(flags);
+		}
+	    } else {
+		emt(mtflags);
+		local_irq_restore(flags);
+
+	    }
+	}
+	/* Reprogram for interrupt at next soonest timestamp for VPE */
+	if (ISVALID(nextstamp)) {
+		write_c0_compare(nextstamp);
+		ehb();
+		if ((nextstamp - (unsigned long)read_c0_count())
+			> (unsigned long)LONG_MAX)
+				goto repeat;
+	}
+}
+
+
+irqreturn_t c0_compare_interrupt(int irq, void *dev_id)
+{
+	int cpu = smp_processor_id();
+
+	/* If we're running SMTC, we've got MIPS MT and therefore MIPS32R2 */
+	handle_perf_irq(1);
+
+	if (read_c0_cause() & (1 << 30)) {
+		/* Clear Count/Compare Interrupt */
+		write_c0_compare(read_c0_compare());
+		smtc_distribute_timer(cpu_data[cpu].vpe_id);
+	}
+	return IRQ_HANDLED;
+}
+
+
+int __cpuinit mips_clockevent_init(void)
+{
+	uint64_t mips_freq = mips_hpt_frequency;
+	unsigned int cpu = smp_processor_id();
+	struct clock_event_device *cd;
+	unsigned int irq;
+	int i;
+	int j;
+
+	if (!cpu_has_counter || !mips_hpt_frequency)
+		return -ENXIO;
+	if (cpu == 0) {
+		for (i = 0; i < num_possible_cpus(); i++) {
+			smtc_nextinvpe[i] = 0;
+			for (j = 0; j < num_possible_cpus(); j++)
+				smtc_nexttime[i][j] = 0L;
+		}
+		/*
+		 * SMTC also can't have the usablility test
+		 * run by secondary TCs once Compare is in use.
+		 */
+		if (!c0_compare_int_usable())
+			return -ENXIO;
+	}
+
+	/*
+	 * With vectored interrupts things are getting platform specific.
+	 * get_c0_compare_int is a hook to allow a platform to return the
+	 * interrupt number of it's liking.
+	 */
+	irq = MIPS_CPU_IRQ_BASE + cp0_compare_irq;
+	if (get_c0_compare_int)
+		irq = get_c0_compare_int();
+
+	cd = &per_cpu(mips_clockevent_device, cpu);
+
+	cd->name		= "MIPS";
+	cd->features		= CLOCK_EVT_FEAT_ONESHOT;
+
+	/* Calculate the min / max delta */
+	cd->mult	= div_sc((unsigned long) mips_freq, NSEC_PER_SEC, 32);
+	cd->shift		= 32;
+	cd->max_delta_ns	= clockevent_delta2ns(0x7fffffff, cd);
+	cd->min_delta_ns	= clockevent_delta2ns(0x300, cd);
+
+	cd->rating		= 300;
+	cd->irq			= irq;
+	cd->cpumask		= cpumask_of_cpu(cpu);
+	cd->set_next_event	= mips_next_event;
+	cd->set_mode		= mips_set_clock_mode;
+	cd->event_handler	= mips_event_handler;
+
+	clockevents_register_device(cd);
+
+	/*
+	 * On SMTC we only want to do the data structure
+	 * initialization and IRQ setup once.
+	 */
+	if (cpu)
+		return 0;
+	/*
+	 * And we need the hwmask associated with the c0_compare
+	 * vector to be initialized.
+	 */
+	irq_hwmask[irq] = (0x100 << cp0_compare_irq);
+	if (cp0_timer_irq_installed)
+		return 0;
+
+	cp0_timer_irq_installed = 1;
+
+	setup_irq(irq, &c0_compare_irqaction);
+
+	return 0;
+}

arch/mips/kernel/cpu-probe.c

@@ -54,14 +54,18 @@ extern void r4k_wait(void);
  * interrupt is requested" restriction in the MIPS32/MIPS64 architecture makes
  * using this version a gamble.
  */
-static void r4k_wait_irqoff(void)
+void r4k_wait_irqoff(void)
 {
 	local_irq_disable();
 	if (!need_resched())
-		__asm__("	.set	mips3		\n"
+		__asm__("	.set	push		\n"
+			"	.set	mips3		\n"
 			"	wait			\n"
-			"	.set	mips0		\n");
+			"	.set	pop		\n");
 	local_irq_enable();
+	__asm__(" 	.globl __pastwait	\n"
+		"__pastwait:			\n");
+	return;
 }
 
 /*

arch/mips/kernel/entry.S

@@ -79,11 +79,6 @@ FEXPORT(syscall_exit)
 
 FEXPORT(restore_all)			# restore full frame
 #ifdef CONFIG_MIPS_MT_SMTC
-/* Detect and execute deferred IPI "interrupts" */
-	LONG_L	s0, TI_REGS($28)
-	LONG_S	sp, TI_REGS($28)
-	jal	deferred_smtc_ipi
-	LONG_S	s0, TI_REGS($28)
 #ifdef CONFIG_MIPS_MT_SMTC_IM_BACKSTOP
 /* Re-arm any temporarily masked interrupts not explicitly "acked" */
 	mfc0	v0, CP0_TCSTATUS
@@ -112,6 +107,11 @@ FEXPORT(restore_all)			# restore full frame
 	xor	t0, t0, t3
 	mtc0	t0, CP0_TCCONTEXT
 #endif /* CONFIG_MIPS_MT_SMTC_IM_BACKSTOP */
+/* Detect and execute deferred IPI "interrupts" */
+	LONG_L	s0, TI_REGS($28)
+	LONG_S	sp, TI_REGS($28)
+	jal	deferred_smtc_ipi
+	LONG_S	s0, TI_REGS($28)
 #endif /* CONFIG_MIPS_MT_SMTC */
 	.set	noat
 	RESTORE_TEMP

arch/mips/kernel/genex.S

@@ -282,8 +282,8 @@ NESTED(except_vec_vi_handler, 0, sp)
 	and	t0, a0, t1
 #ifdef CONFIG_MIPS_MT_SMTC_IM_BACKSTOP
 	mfc0	t2, CP0_TCCONTEXT
-	or	t0, t0, t2
-	mtc0	t0, CP0_TCCONTEXT
+	or	t2, t0, t2
+	mtc0	t2, CP0_TCCONTEXT
 #endif /* CONFIG_MIPS_MT_SMTC_IM_BACKSTOP */
 	xor	t1, t1, t0
 	mtc0	t1, CP0_STATUS

arch/mips/kernel/mips-mt-fpaff.c

@@ -159,7 +159,7 @@ __setup("fpaff=", fpaff_thresh);
 /*
  * FPU Use Factor empirically derived from experiments on 34K
  */
-#define FPUSEFACTOR 333
+#define FPUSEFACTOR 2000
 
 static __init int mt_fp_affinity_init(void)
 {

arch/mips/kernel/process.c

@@ -55,7 +55,7 @@ void __noreturn cpu_idle(void)
 	while (1) {
 		tick_nohz_stop_sched_tick(1);
 		while (!need_resched()) {
-#ifdef CONFIG_SMTC_IDLE_HOOK_DEBUG
+#ifdef CONFIG_MIPS_MT_SMTC
 			extern void smtc_idle_loop_hook(void);
 
 			smtc_idle_loop_hook();
@@ -145,19 +145,18 @@ int copy_thread(int nr, unsigned long clone_flags, unsigned long usp,
 	 */
 	p->thread.cp0_status = read_c0_status() & ~(ST0_CU2|ST0_CU1);
 	childregs->cp0_status &= ~(ST0_CU2|ST0_CU1);
+
+#ifdef CONFIG_MIPS_MT_SMTC
+	/*
+	 * SMTC restores TCStatus after Status, and the CU bits
+	 * are aliased there.
+	 */
+	childregs->cp0_tcstatus &= ~(ST0_CU2|ST0_CU1);
+#endif
 	clear_tsk_thread_flag(p, TIF_USEDFPU);
 
 #ifdef CONFIG_MIPS_MT_FPAFF
 	clear_tsk_thread_flag(p, TIF_FPUBOUND);
-
-	/*
-	 * FPU affinity support is cleaner if we track the
-	 * user-visible CPU affinity from the very beginning.
-	 * The generic cpus_allowed mask will already have
-	 * been copied from the parent before copy_thread
-	 * is invoked.
-	 */
-	p->thread.user_cpus_allowed = p->cpus_allowed;
 #endif /* CONFIG_MIPS_MT_FPAFF */
 
 	if (clone_flags & CLONE_SETTLS)

arch/mips/kernel/ptrace.c

@@ -238,7 +238,7 @@ long arch_ptrace(struct task_struct *child, long request, long addr, long data)
 		case FPC_EIR: {	/* implementation / version register */
 			unsigned int flags;
 #ifdef CONFIG_MIPS_MT_SMTC
-			unsigned int irqflags;
+			unsigned long irqflags;
 			unsigned int mtflags;
 #endif /* CONFIG_MIPS_MT_SMTC */
 

arch/mips/kernel/traps.c

@@ -825,7 +825,9 @@ static void mt_ase_fp_affinity(void)
 		if (cpus_intersects(current->cpus_allowed, mt_fpu_cpumask)) {
 			cpumask_t tmask;
 
-			cpus_and(tmask, current->thread.user_cpus_allowed,
+			current->thread.user_cpus_allowed
+				= current->cpus_allowed;
+			cpus_and(tmask, current->cpus_allowed,
 				mt_fpu_cpumask);
 			set_cpus_allowed(current, tmask);
 			set_thread_flag(TIF_FPUBOUND);

arch/mips/mti-malta/Makefile

@@ -15,6 +15,6 @@ obj-$(CONFIG_EARLY_PRINTK)	+= malta-console.o
 obj-$(CONFIG_PCI)		+= malta-pci.o
 
 # FIXME FIXME FIXME
-obj-$(CONFIG_MIPS_MT_SMTC)	+= malta_smtc.o
+obj-$(CONFIG_MIPS_MT_SMTC)	+= malta-smtc.o
 
 EXTRA_CFLAGS += -Werror

arch/mips/mti-malta/malta-smtc.c

@@ -84,12 +84,17 @@ static void msmtc_cpus_done(void)
 
 static void __init msmtc_smp_setup(void)
 {
-	mipsmt_build_cpu_map(0);
+	/*
+	 * we won't get the definitive value until
+	 * we've run smtc_prepare_cpus later, but
+	 * we would appear to need an upper bound now.
+	 */
+	smp_num_siblings = smtc_build_cpu_map(0);
 }
 
 static void __init msmtc_prepare_cpus(unsigned int max_cpus)
 {
-	mipsmt_prepare_cpus();
+	smtc_prepare_cpus(max_cpus);
 }
 
 struct plat_smp_ops msmtc_smp_ops = {

include/asm-mips/cevt-r4k.h

+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2008 Kevin D. Kissell
+ */
+
+/*
+ * Definitions used for common event timer implementation
+ * for MIPS 4K-type processors and their MIPS MT variants.
+ * Avoids unsightly extern declarations in C files.
+ */
+#ifndef __ASM_CEVT_R4K_H
+#define __ASM_CEVT_R4K_H
+
+DECLARE_PER_CPU(struct clock_event_device, mips_clockevent_device);
+
+void mips_event_handler(struct clock_event_device *dev);
+int c0_compare_int_usable(void);
+void mips_set_clock_mode(enum clock_event_mode, struct clock_event_device *);
+irqreturn_t c0_compare_interrupt(int, void *);
+
+extern struct irqaction c0_compare_irqaction;
+extern int cp0_timer_irq_installed;
+
+/*
+ * Possibly handle a performance counter interrupt.
+ * Return true if the timer interrupt should not be checked
+ */
+
+static inline int handle_perf_irq(int r2)
+{
+	/*
+	 * The performance counter overflow interrupt may be shared with the
+	 * timer interrupt (cp0_perfcount_irq < 0). If it is and a
+	 * performance counter has overflowed (perf_irq() == IRQ_HANDLED)
+	 * and we can't reliably determine if a counter interrupt has also
+	 * happened (!r2) then don't check for a timer interrupt.
+	 */
+	return (cp0_perfcount_irq < 0) &&
+		perf_irq() == IRQ_HANDLED &&
+		!r2;
+}
+
+#endif /* __ASM_CEVT_R4K_H */

include/asm-mips/irqflags.h

@@ -38,8 +38,17 @@ __asm__(
 	"	.set	pop						\n"
 	"	.endm");
 
+extern void smtc_ipi_replay(void);
+
 static inline void raw_local_irq_enable(void)
 {
+#ifdef CONFIG_MIPS_MT_SMTC
+	/*
+	 * SMTC kernel needs to do a software replay of queued
+	 * IPIs, at the cost of call overhead on each local_irq_enable()
+	 */
+	smtc_ipi_replay();
+#endif
 	__asm__ __volatile__(
 		"raw_local_irq_enable"
 		: /* no outputs */
@@ -47,6 +56,7 @@ static inline void raw_local_irq_enable(void)
 		: "memory");
 }
 
+
 /*
  * For cli() we have to insert nops to make sure that the new value
  * has actually arrived in the status register before the end of this
@@ -185,15 +195,14 @@ __asm__(
 	"	.set	pop						\n"
 	"	.endm							\n");
 
-extern void smtc_ipi_replay(void);
 
 static inline void raw_local_irq_restore(unsigned long flags)
 {
 	unsigned long __tmp1;
 
-#ifdef CONFIG_MIPS_MT_SMTC_INSTANT_REPLAY
+#ifdef CONFIG_MIPS_MT_SMTC
 	/*
-	 * CONFIG_MIPS_MT_SMTC_INSTANT_REPLAY does prompt replay of deferred
+	 * SMTC kernel needs to do a software replay of queued
 	 * IPIs, at the cost of branch and call overhead on each
 	 * local_irq_restore()
 	 */
@@ -208,6 +217,17 @@ static inline void raw_local_irq_restore(unsigned long flags)
 		: "memory");
 }
 
+static inline void __raw_local_irq_restore(unsigned long flags)
+{
+	unsigned long __tmp1;
+
+	__asm__ __volatile__(
+		"raw_local_irq_restore\t%0"
+		: "=r" (__tmp1)
+		: "0" (flags)
+		: "memory");
+}
+
 static inline int raw_irqs_disabled_flags(unsigned long flags)
 {
 #ifdef CONFIG_MIPS_MT_SMTC

include/asm-mips/mipsregs.h

@@ -1462,7 +1462,7 @@ set_c0_##name(unsigned int set)					\
 {								\
 	unsigned int res;					\
 	unsigned int omt;					\
-	unsigned int flags;					\
+	unsigned long flags;					\
 								\
 	local_irq_save(flags);					\
 	omt = __dmt();						\
@@ -1480,7 +1480,7 @@ clear_c0_##name(unsigned int clear)				\
 {								\
 	unsigned int res;					\
 	unsigned int omt;					\
-	unsigned int flags;					\
+	unsigned long flags;					\
 								\
 	local_irq_save(flags);					\
 	omt = __dmt();						\
@@ -1498,7 +1498,7 @@ change_c0_##name(unsigned int change, unsigned int new)		\
 {								\
 	unsigned int res;					\
 	unsigned int omt;					\
-	unsigned int flags;					\
+	unsigned long flags;					\
 								\
 	local_irq_save(flags);					\
 								\

include/asm-mips/smtc.h

@@ -6,6 +6,7 @@
  */
 
 #include <asm/mips_mt.h>
+#include <asm/smtc_ipi.h>
 
 /*
  * System-wide SMTC status information
@@ -38,14 +39,15 @@ struct mm_struct;
 struct task_struct;
 
 void smtc_get_new_mmu_context(struct mm_struct *mm, unsigned long cpu);
-
+void self_ipi(struct smtc_ipi *);
 void smtc_flush_tlb_asid(unsigned long asid);
-extern int mipsmt_build_cpu_map(int startslot);
-extern void mipsmt_prepare_cpus(void);
+extern int smtc_build_cpu_map(int startslot);
+extern void smtc_prepare_cpus(int cpus);
 extern void smtc_smp_finish(void);
 extern void smtc_boot_secondary(int cpu, struct task_struct *t);
 extern void smtc_cpus_done(void);
 
+
 /*
  * Sharing the TLB between multiple VPEs means that the
  * "random" index selection function is not allowed to

include/asm-mips/stackframe.h

@@ -297,14 +297,31 @@
 #ifdef CONFIG_MIPS_MT_SMTC
 		.set	mips32r2
 		/*
-		 * This may not really be necessary if ints are already
-		 * inhibited here.
+		 * We need to make sure the read-modify-write
+		 * of Status below isn't perturbed by an interrupt
+		 * or cross-TC access, so we need to do at least a DMT,
+		 * protected by an interrupt-inhibit. But setting IXMT
+		 * also creates a few-cycle window where an IPI could
+		 * be queued and not be detected before potentially
+		 * returning to a WAIT or user-mode loop. It must be
+		 * replayed.
+		 *
+		 * We're in the middle of a context switch, and
+		 * we can't dispatch it directly without trashing
+		 * some registers, so we'll try to detect this unlikely
+		 * case and program a software interrupt in the VPE,
+		 * as would be done for a cross-VPE IPI.  To accomodate
+		 * the handling of that case, we're doing a DVPE instead
+		 * of just a DMT here to protect against other threads.
+		 * This is a lot of cruft to cover a tiny window.
+		 * If you can find a better design, implement it!
+		 *
 		 */
 		mfc0	v0, CP0_TCSTATUS
 		ori	v0, TCSTATUS_IXMT
 		mtc0	v0, CP0_TCSTATUS
 		_ehb
-		DMT	5				# dmt a1
+		DVPE	5				# dvpe a1
 		jal	mips_ihb
 #endif /* CONFIG_MIPS_MT_SMTC */
 		mfc0	a0, CP0_STATUS
@@ -325,17 +342,50 @@
  */
 		LONG_L	v1, PT_TCSTATUS(sp)
 		_ehb
-		mfc0	v0, CP0_TCSTATUS
+		mfc0	a0, CP0_TCSTATUS
 		andi	v1, TCSTATUS_IXMT
-		/* We know that TCStatua.IXMT should be set from above */
-		xori	v0, v0, TCSTATUS_IXMT
-		or	v0, v0, v1
-		mtc0	v0, CP0_TCSTATUS
-		_ehb
-		andi	a1, a1, VPECONTROL_TE
+		bnez	v1, 0f
+
+/*
+ * We'd like to detect any IPIs queued in the tiny window
+ * above and request an software interrupt to service them
+ * when we ERET.
+ *
+ * Computing the offset into the IPIQ array of the executing
+ * TC's IPI queue in-line would be tedious.  We use part of
+ * the TCContext register to hold 16 bits of offset that we
+ * can add in-line to find the queue head.
+ */
+		mfc0	v0, CP0_TCCONTEXT
+		la	a2, IPIQ
+		srl	v0, v0, 16
+		addu	a2, a2, v0
+		LONG_L	v0, 0(a2)
+		beqz	v0, 0f
+/*
+ * If we have a queue, provoke dispatch within the VPE by setting C_SW1
+ */
+		mfc0	v0, CP0_CAUSE
+		ori	v0, v0, C_SW1
+		mtc0	v0, CP0_CAUSE
+0:
+		/*
+		 * This test should really never branch but
+		 * let's be prudent here.  Having atomized
+		 * the shared register modifications, we can
+		 * now EVPE, and must do so before interrupts
+		 * are potentially re-enabled.
+		 */
+		andi	a1, a1, MVPCONTROL_EVP
 		beqz	a1, 1f
-		emt
+		evpe
 1:
+		/* We know that TCStatua.IXMT should be set from above */
+		xori	a0, a0, TCSTATUS_IXMT
+		or	a0, a0, v1
+		mtc0	a0, CP0_TCSTATUS
+		_ehb
+
 		.set	mips0
 #endif /* CONFIG_MIPS_MT_SMTC */
 		LONG_L	v1, PT_EPC(sp)