tile/perf: Support perf_events on tilegx and tilepro

Add perf support for tile architecture.
Signed-off-by: Zhigang Lu <zlu@tilera.com>
Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>

arch/tile/Kconfig

@@ -3,6 +3,8 @@
 
 config TILE
 	def_bool y
+	select HAVE_PERF_EVENTS
+	select USE_PMC if PERF_EVENTS
 	select HAVE_DMA_ATTRS
 	select HAVE_DMA_API_DEBUG
 	select HAVE_KVM if !TILEGX

arch/tile/include/asm/perf_event.h

+/*
+ * Copyright 2014 Tilera Corporation. All Rights Reserved.
+ *
+ *   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 the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ */
+
+#ifndef _ASM_TILE_PERF_EVENT_H
+#define _ASM_TILE_PERF_EVENT_H
+
+#include <linux/percpu.h>
+DECLARE_PER_CPU(u64, perf_irqs);
+
+unsigned long handle_syscall_link_address(void);
+#endif /* _ASM_TILE_PERF_EVENT_H */

arch/tile/kernel/Makefile

@@ -25,6 +25,7 @@ obj-$(CONFIG_PCI)		+= pci_gx.o
 else
 obj-$(CONFIG_PCI)		+= pci.o
 endif
+obj-$(CONFIG_PERF_EVENTS)	+= perf_event.o
 obj-$(CONFIG_USE_PMC)		+= pmc.o
 obj-$(CONFIG_TILE_USB)		+= usb.o
 obj-$(CONFIG_TILE_HVGLUE_TRACE)	+= hvglue_trace.o

arch/tile/kernel/irq.c

@@ -21,6 +21,7 @@
 #include <hv/drv_pcie_rc_intf.h>
 #include <arch/spr_def.h>
 #include <asm/traps.h>
+#include <linux/perf_event.h>
 
 /* Bit-flag stored in irq_desc->chip_data to indicate HW-cleared irqs. */
 #define IS_HW_CLEARED 1
@@ -260,6 +261,23 @@ void ack_bad_irq(unsigned int irq)
 	pr_err("unexpected IRQ trap at vector %02x\n", irq);
 }
 
+/*
+ * /proc/interrupts printing:
+ */
+int arch_show_interrupts(struct seq_file *p, int prec)
+{
+#ifdef CONFIG_PERF_EVENTS
+	int i;
+
+	seq_printf(p, "%*s: ", prec, "PMI");
+
+	for_each_online_cpu(i)
+		seq_printf(p, "%10llu ", per_cpu(perf_irqs, i));
+	seq_puts(p, "  perf_events\n");
+#endif
+	return 0;
+}
+
 /*
  * Generic, controller-independent functions:
  */

arch/tile/kernel/perf_event.c

+/*
+ * Copyright 2014 Tilera Corporation. All Rights Reserved.
+ *
+ *   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 the Free Software Foundation, version 2.
+ *
+ *   This program is distributed in the hope that it will be useful, but
+ *   WITHOUT ANY WARRANTY; without even the implied warranty of
+ *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
+ *   NON INFRINGEMENT.  See the GNU General Public License for
+ *   more details.
+ *
+ *
+ * Perf_events support for Tile processor.
+ *
+ * This code is based upon the x86 perf event
+ * code, which is:
+ *
+ *  Copyright (C) 2008 Thomas Gleixner <tglx@linutronix.de>
+ *  Copyright (C) 2008-2009 Red Hat, Inc., Ingo Molnar
+ *  Copyright (C) 2009 Jaswinder Singh Rajput
+ *  Copyright (C) 2009 Advanced Micro Devices, Inc., Robert Richter
+ *  Copyright (C) 2008-2009 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
+ *  Copyright (C) 2009 Intel Corporation, <markus.t.metzger@intel.com>
+ *  Copyright (C) 2009 Google, Inc., Stephane Eranian
+ */
+
+#include <linux/kprobes.h>
+#include <linux/kernel.h>
+#include <linux/kdebug.h>
+#include <linux/mutex.h>
+#include <linux/bitmap.h>
+#include <linux/irq.h>
+#include <linux/interrupt.h>
+#include <linux/perf_event.h>
+#include <linux/atomic.h>
+#include <asm/traps.h>
+#include <asm/stack.h>
+#include <asm/pmc.h>
+#include <hv/hypervisor.h>
+
+#define TILE_MAX_COUNTERS	4
+
+#define PERF_COUNT_0_IDX	0
+#define PERF_COUNT_1_IDX	1
+#define AUX_PERF_COUNT_0_IDX	2
+#define AUX_PERF_COUNT_1_IDX	3
+
+struct cpu_hw_events {
+	int			n_events;
+	struct perf_event	*events[TILE_MAX_COUNTERS]; /* counter order */
+	struct perf_event	*event_list[TILE_MAX_COUNTERS]; /* enabled
+								order */
+	int			assign[TILE_MAX_COUNTERS];
+	unsigned long		active_mask[BITS_TO_LONGS(TILE_MAX_COUNTERS)];
+	unsigned long		used_mask;
+};
+
+/* TILE arch specific performance monitor unit */
+struct tile_pmu {
+	const char	*name;
+	int		version;
+	const int	*hw_events;	/* generic hw events table */
+	/* generic hw cache events table */
+	const int	(*cache_events)[PERF_COUNT_HW_CACHE_MAX]
+				       [PERF_COUNT_HW_CACHE_OP_MAX]
+				       [PERF_COUNT_HW_CACHE_RESULT_MAX];
+	int		(*map_hw_event)(u64);	 /*method used to map
+						  hw events */
+	int		(*map_cache_event)(u64); /*method used to map
+						  cache events */
+
+	u64		max_period;		/* max sampling period */
+	u64		cntval_mask;		/* counter width mask */
+	int		cntval_bits;		/* counter width */
+	int		max_events;		/* max generic hw events
+						in map */
+	int		num_counters;		/* number base + aux counters */
+	int		num_base_counters;	/* number base counters */
+};
+
+DEFINE_PER_CPU(u64, perf_irqs);
+static DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
+
+#define TILE_OP_UNSUPP		(-1)
+
+#ifndef __tilegx__
+/* TILEPro hardware events map */
+static const int tile_hw_event_map[] = {
+	[PERF_COUNT_HW_CPU_CYCLES]		= 0x01, /* ONE */
+	[PERF_COUNT_HW_INSTRUCTIONS]		= 0x06, /* MP_BUNDLE_RETIRED */
+	[PERF_COUNT_HW_CACHE_REFERENCES]	= TILE_OP_UNSUPP,
+	[PERF_COUNT_HW_CACHE_MISSES]		= TILE_OP_UNSUPP,
+	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0x16, /*
+					  MP_CONDITIONAL_BRANCH_ISSUED */
+	[PERF_COUNT_HW_BRANCH_MISSES]		= 0x14, /*
+					  MP_CONDITIONAL_BRANCH_MISSPREDICT */
+	[PERF_COUNT_HW_BUS_CYCLES]		= TILE_OP_UNSUPP,
+};
+#else
+/* TILEGx hardware events map */
+static const int tile_hw_event_map[] = {
+	[PERF_COUNT_HW_CPU_CYCLES]		= 0x181, /* ONE */
+	[PERF_COUNT_HW_INSTRUCTIONS]		= 0xdb, /* INSTRUCTION_BUNDLE */
+	[PERF_COUNT_HW_CACHE_REFERENCES]	= TILE_OP_UNSUPP,
+	[PERF_COUNT_HW_CACHE_MISSES]		= TILE_OP_UNSUPP,
+	[PERF_COUNT_HW_BRANCH_INSTRUCTIONS]	= 0xd9, /*
+						COND_BRANCH_PRED_CORRECT */
+	[PERF_COUNT_HW_BRANCH_MISSES]		= 0xda, /*
+						COND_BRANCH_PRED_INCORRECT */
+	[PERF_COUNT_HW_BUS_CYCLES]		= TILE_OP_UNSUPP,
+};
+#endif
+
+#define C(x) PERF_COUNT_HW_CACHE_##x
+
+/*
+ * Generalized hw caching related hw_event table, filled
+ * in on a per model basis. A value of -1 means
+ * 'not supported', any other value means the
+ * raw hw_event ID.
+ */
+#ifndef __tilegx__
+/* TILEPro hardware cache event map */
+static const int tile_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
+				     [PERF_COUNT_HW_CACHE_OP_MAX]
+				     [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
+[C(L1D)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0x21, /* RD_MISS */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0x22, /* WR_MISS */
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(L1I)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = 0x12, /* MP_ICACHE_HIT_ISSUED */
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(LL)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(DTLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = 0x1d, /* TLB_CNT */
+		[C(RESULT_MISS)] = 0x20, /* TLB_EXCEPTION */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(ITLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = 0x13, /* MP_ITLB_HIT_ISSUED */
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(BPU)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+};
+#else
+/* TILEGx hardware events map */
+static const int tile_cache_event_map[PERF_COUNT_HW_CACHE_MAX]
+				     [PERF_COUNT_HW_CACHE_OP_MAX]
+				     [PERF_COUNT_HW_CACHE_RESULT_MAX] = {
+[C(L1D)] = {
+	/*
+	 * Like some other architectures (e.g. ARM), the performance
+	 * counters don't differentiate between read and write
+	 * accesses/misses, so this isn't strictly correct, but it's the
+	 * best we can do. Writes and reads get combined.
+	 */
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0x44, /* RD_MISS */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0x45, /* WR_MISS */
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(L1I)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(LL)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(DTLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = 0x40, /* TLB_CNT */
+		[C(RESULT_MISS)] = 0x43, /* TLB_EXCEPTION */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = 0x40, /* TLB_CNT */
+		[C(RESULT_MISS)] = 0x43, /* TLB_EXCEPTION */
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(ITLB)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0xd4, /* ITLB_MISS_INT */
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = 0xd4, /* ITLB_MISS_INT */
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+[C(BPU)] = {
+	[C(OP_READ)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_WRITE)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+	[C(OP_PREFETCH)] = {
+		[C(RESULT_ACCESS)] = TILE_OP_UNSUPP,
+		[C(RESULT_MISS)] = TILE_OP_UNSUPP,
+	},
+},
+};
+#endif
+
+static atomic_t tile_active_events;
+static DEFINE_MUTEX(perf_intr_reserve_mutex);
+
+static int tile_map_hw_event(u64 config);
+static int tile_map_cache_event(u64 config);
+
+static int tile_pmu_handle_irq(struct pt_regs *regs, int fault);
+
+/*
+ * To avoid new_raw_count getting larger then pre_raw_count
+ * in tile_perf_event_update(), we limit the value of max_period to 2^31 - 1.
+ */
+static const struct tile_pmu tilepmu = {
+#ifndef __tilegx__
+	.name = "tilepro",
+#else
+	.name = "tilegx",
+#endif
+	.max_events = ARRAY_SIZE(tile_hw_event_map),
+	.map_hw_event = tile_map_hw_event,
+	.hw_events = tile_hw_event_map,
+	.map_cache_event = tile_map_cache_event,
+	.cache_events = &tile_cache_event_map,
+	.cntval_bits = 32,
+	.cntval_mask = (1ULL << 32) - 1,
+	.max_period = (1ULL << 31) - 1,
+	.num_counters = TILE_MAX_COUNTERS,
+	.num_base_counters = TILE_BASE_COUNTERS,
+};
+
+static const struct tile_pmu *tile_pmu __read_mostly;
+
+/*
+ * Check whether perf event is enabled.
+ */
+int tile_perf_enabled(void)
+{
+	return atomic_read(&tile_active_events) != 0;
+}
+
+/*
+ * Read Performance Counters.
+ */
+static inline u64 read_counter(int idx)
+{
+	u64 val = 0;
+
+	/* __insn_mfspr() only takes an immediate argument */
+	switch (idx) {
+	case PERF_COUNT_0_IDX:
+		val = __insn_mfspr(SPR_PERF_COUNT_0);
+		break;
+	case PERF_COUNT_1_IDX:
+		val = __insn_mfspr(SPR_PERF_COUNT_1);
+		break;
+	case AUX_PERF_COUNT_0_IDX:
+		val = __insn_mfspr(SPR_AUX_PERF_COUNT_0);
+		break;
+	case AUX_PERF_COUNT_1_IDX:
+		val = __insn_mfspr(SPR_AUX_PERF_COUNT_1);
+		break;
+	default:
+		WARN_ON_ONCE(idx > AUX_PERF_COUNT_1_IDX ||
+				idx < PERF_COUNT_0_IDX);
+	}
+
+	return val;
+}
+
+/*
+ * Write Performance Counters.
+ */
+static inline void write_counter(int idx, u64 value)
+{
+	/* __insn_mtspr() only takes an immediate argument */
+	switch (idx) {
+	case PERF_COUNT_0_IDX:
+		__insn_mtspr(SPR_PERF_COUNT_0, value);
+		break;
+	case PERF_COUNT_1_IDX:
+		__insn_mtspr(SPR_PERF_COUNT_1, value);
+		break;
+	case AUX_PERF_COUNT_0_IDX:
+		__insn_mtspr(SPR_AUX_PERF_COUNT_0, value);
+		break;
+	case AUX_PERF_COUNT_1_IDX:
+		__insn_mtspr(SPR_AUX_PERF_COUNT_1, value);
+		break;
+	default:
+		WARN_ON_ONCE(idx > AUX_PERF_COUNT_1_IDX ||
+				idx < PERF_COUNT_0_IDX);
+	}
+}
+
+/*
+ * Enable performance event by setting
+ * Performance Counter Control registers.
+ */
+static inline void tile_pmu_enable_event(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	unsigned long cfg, mask;
+	int shift, idx = hwc->idx;
+
+	/*
+	 * prevent early activation from tile_pmu_start() in hw_perf_enable
+	 */
+
+	if (WARN_ON_ONCE(idx == -1))
+		return;
+
+	if (idx < tile_pmu->num_base_counters)
+		cfg = __insn_mfspr(SPR_PERF_COUNT_CTL);
+	else
+		cfg = __insn_mfspr(SPR_AUX_PERF_COUNT_CTL);
+
+	switch (idx) {
+	case PERF_COUNT_0_IDX:
+	case AUX_PERF_COUNT_0_IDX:
+		mask = TILE_EVENT_MASK;
+		shift = 0;
+		break;
+	case PERF_COUNT_1_IDX:
+	case AUX_PERF_COUNT_1_IDX:
+		mask = TILE_EVENT_MASK << 16;
+		shift = 16;
+		break;
+	default:
+		WARN_ON_ONCE(idx < PERF_COUNT_0_IDX ||
+			idx > AUX_PERF_COUNT_1_IDX);
+		return;
+	}
+
+	/* Clear mask bits to enable the event. */
+	cfg &= ~mask;
+	cfg |= hwc->config << shift;
+
+	if (idx < tile_pmu->num_base_counters)
+		__insn_mtspr(SPR_PERF_COUNT_CTL, cfg);
+	else
+		__insn_mtspr(SPR_AUX_PERF_COUNT_CTL, cfg);
+}
+
+/*
+ * Disable performance event by clearing
+ * Performance Counter Control registers.
+ */
+static inline void tile_pmu_disable_event(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	unsigned long cfg, mask;
+	int idx = hwc->idx;
+
+	if (idx == -1)
+		return;
+
+	if (idx < tile_pmu->num_base_counters)
+		cfg = __insn_mfspr(SPR_PERF_COUNT_CTL);
+	else
+		cfg = __insn_mfspr(SPR_AUX_PERF_COUNT_CTL);
+
+	switch (idx) {
+	case PERF_COUNT_0_IDX:
+	case AUX_PERF_COUNT_0_IDX:
+		mask = TILE_PLM_MASK;
+		break;
+	case PERF_COUNT_1_IDX:
+	case AUX_PERF_COUNT_1_IDX:
+		mask = TILE_PLM_MASK << 16;
+		break;
+	default:
+		WARN_ON_ONCE(idx < PERF_COUNT_0_IDX ||
+			idx > AUX_PERF_COUNT_1_IDX);
+		return;
+	}
+
+	/* Set mask bits to disable the event. */
+	cfg |= mask;
+
+	if (idx < tile_pmu->num_base_counters)
+		__insn_mtspr(SPR_PERF_COUNT_CTL, cfg);
+	else
+		__insn_mtspr(SPR_AUX_PERF_COUNT_CTL, cfg);
+}
+
+/*
+ * Propagate event elapsed time into the generic event.
+ * Can only be executed on the CPU where the event is active.
+ * Returns the delta events processed.
+ */
+static u64 tile_perf_event_update(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	int shift = 64 - tile_pmu->cntval_bits;
+	u64 prev_raw_count, new_raw_count;
+	u64 oldval;
+	int idx = hwc->idx;
+	u64 delta;
+
+	/*
+	 * Careful: an NMI might modify the previous event value.
+	 *
+	 * Our tactic to handle this is to first atomically read and
+	 * exchange a new raw count - then add that new-prev delta
+	 * count to the generic event atomically:
+	 */
+again:
+	prev_raw_count = local64_read(&hwc->prev_count);
+	new_raw_count = read_counter(idx);
+
+	oldval = local64_cmpxchg(&hwc->prev_count, prev_raw_count,
+				 new_raw_count);
+	if (oldval != prev_raw_count)
+		goto again;
+
+	/*
+	 * Now we have the new raw value and have updated the prev
+	 * timestamp already. We can now calculate the elapsed delta
+	 * (event-)time and add that to the generic event.
+	 *
+	 * Careful, not all hw sign-extends above the physical width
+	 * of the count.
+	 */
+	delta = (new_raw_count << shift) - (prev_raw_count << shift);
+	delta >>= shift;
+
+	local64_add(delta, &event->count);
+	local64_sub(delta, &hwc->period_left);
+
+	return new_raw_count;
+}
+
+/*
+ * Set the next IRQ period, based on the hwc->period_left value.
+ * To be called with the event disabled in hw:
+ */
+static int tile_event_set_period(struct perf_event *event)
+{
+	struct hw_perf_event *hwc = &event->hw;
+	int idx = hwc->idx;
+	s64 left = local64_read(&hwc->period_left);
+	s64 period = hwc->sample_period;
+	int ret = 0;
+
+	/*
+	 * If we are way outside a reasonable range then just skip forward:
+	 */
+	if (unlikely(left <= -period)) {
+		left = period;
+		local64_set(&hwc->period_left, left);
+		hwc->last_period = period;
+		ret = 1;
+	}
+
+	if (unlikely(left <= 0)) {
+		left += period;
+		local64_set(&hwc->period_left, left);
+		hwc->last_period = period;
+		ret = 1;
+	}
+	if (left > tile_pmu->max_period)
+		left = tile_pmu->max_period;
+
+	/*
+	 * The hw event starts counting from this event offset,
+	 * mark it to be able to extra future deltas:
+	 */
+	local64_set(&hwc->prev_count, (u64)-left);
+
+	write_counter(idx, (u64)(-left) & tile_pmu->cntval_mask);
+
+	perf_event_update_userpage(event);
+
+	return ret;
+}
+
+/*
+ * Stop the event but do not release the PMU counter
+ */
+static void tile_pmu_stop(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+	struct hw_perf_event *hwc = &event->hw;
+	int idx = hwc->idx;
+
+	if (__test_and_clear_bit(idx, cpuc->active_mask)) {
+		tile_pmu_disable_event(event);
+		cpuc->events[hwc->idx] = NULL;
+		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
+		hwc->state |= PERF_HES_STOPPED;
+	}
+
+	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
+		/*
+		 * Drain the remaining delta count out of a event
+		 * that we are disabling:
+		 */
+		tile_perf_event_update(event);
+		hwc->state |= PERF_HES_UPTODATE;
+	}
+}
+
+/*
+ * Start an event (without re-assigning counter)
+ */
+static void tile_pmu_start(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+	int idx = event->hw.idx;
+
+	if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED)))
+		return;
+
+	if (WARN_ON_ONCE(idx == -1))
+		return;
+
+	if (flags & PERF_EF_RELOAD) {
+		WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
+		tile_event_set_period(event);
+	}
+
+	event->hw.state = 0;
+
+	cpuc->events[idx] = event;
+	__set_bit(idx, cpuc->active_mask);
+
+	unmask_pmc_interrupts();
+
+	tile_pmu_enable_event(event);
+
+	perf_event_update_userpage(event);
+}
+
+/*
+ * Add a single event to the PMU.
+ *
+ * The event is added to the group of enabled events
+ * but only if it can be scehduled with existing events.
+ */
+static int tile_pmu_add(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+	struct hw_perf_event *hwc;
+	unsigned long mask;
+	int b, max_cnt;
+
+	hwc = &event->hw;
+
+	/*
+	 * We are full.
+	 */
+	if (cpuc->n_events == tile_pmu->num_counters)
+		return -ENOSPC;
+
+	cpuc->event_list[cpuc->n_events] = event;
+	cpuc->n_events++;
+
+	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
+	if (!(flags & PERF_EF_START))
+		hwc->state |= PERF_HES_ARCH;
+
+	/*
+	 * Find first empty counter.
+	 */
+	max_cnt = tile_pmu->num_counters;
+	mask = ~cpuc->used_mask;
+
+	/* Find next free counter. */
+	b = find_next_bit(&mask, max_cnt, 0);
+
+	/* Should not happen. */
+	if (WARN_ON_ONCE(b == max_cnt))
+		return -ENOSPC;
+
+	/*
+	 * Assign counter to event.
+	 */
+	event->hw.idx = b;
+	__set_bit(b, &cpuc->used_mask);
+
+	/*
+	 * Start if requested.
+	 */
+	if (flags & PERF_EF_START)
+		tile_pmu_start(event, PERF_EF_RELOAD);
+
+	return 0;
+}
+
+/*
+ * Delete a single event from the PMU.
+ *
+ * The event is deleted from the group of enabled events.
+ * If it is the last event, disable PMU interrupt.
+ */
+static void tile_pmu_del(struct perf_event *event, int flags)
+{
+	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+	int i;
+
+	/*
+	 * Remove event from list, compact list if necessary.
+	 */
+	for (i = 0; i < cpuc->n_events; i++) {
+		if (cpuc->event_list[i] == event) {
+			while (++i < cpuc->n_events)
+				cpuc->event_list[i-1] = cpuc->event_list[i];
+			--cpuc->n_events;
+			cpuc->events[event->hw.idx] = NULL;
+			__clear_bit(event->hw.idx, &cpuc->used_mask);
+			tile_pmu_stop(event, PERF_EF_UPDATE);
+			break;
+		}
+	}
+	/*
+	 * If there are no events left, then mask PMU interrupt.
+	 */
+	if (cpuc->n_events == 0)
+		mask_pmc_interrupts();
+	perf_event_update_userpage(event);
+}
+
+/*
+ * Propagate event elapsed time into the event.
+ */
+static inline void tile_pmu_read(struct perf_event *event)
+{
+	tile_perf_event_update(event);
+}
+
+/*
+ * Map generic events to Tile PMU.
+ */
+static int tile_map_hw_event(u64 config)
+{
+	if (config >= tile_pmu->max_events)
+		return -EINVAL;
+	return tile_pmu->hw_events[config];
+}
+
+/*
+ * Map generic hardware cache events to Tile PMU.
+ */
+static int tile_map_cache_event(u64 config)
+{
+	unsigned int cache_type, cache_op, cache_result;
+	int code;
+
+	if (!tile_pmu->cache_events)
+		return -ENOENT;
+
+	cache_type = (config >>  0) & 0xff;
+	if (cache_type >= PERF_COUNT_HW_CACHE_MAX)
+		return -EINVAL;
+
+	cache_op = (config >>  8) & 0xff;
+	if (cache_op >= PERF_COUNT_HW_CACHE_OP_MAX)
+		return -EINVAL;
+
+	cache_result = (config >> 16) & 0xff;
+	if (cache_result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
+		return -EINVAL;
+
+	code = (*tile_pmu->cache_events)[cache_type][cache_op][cache_result];
+	if (code == TILE_OP_UNSUPP)
+		return -EINVAL;
+
+	return code;
+}
+
+static void tile_event_destroy(struct perf_event *event)
+{
+	if (atomic_dec_return(&tile_active_events) == 0)
+		release_pmc_hardware();
+}
+
+static int __tile_event_init(struct perf_event *event)
+{
+	struct perf_event_attr *attr = &event->attr;
+	struct hw_perf_event *hwc = &event->hw;
+	int code;
+
+	switch (attr->type) {
+	case PERF_TYPE_HARDWARE:
+		code = tile_pmu->map_hw_event(attr->config);
+		break;
+	case PERF_TYPE_HW_CACHE:
+		code = tile_pmu->map_cache_event(attr->config);
+		break;
+	case PERF_TYPE_RAW:
+		code = attr->config & TILE_EVENT_MASK;
+		break;
+	default:
+		/* Should not happen. */
+		return -EOPNOTSUPP;
+	}
+
+	if (code < 0)
+		return code;
+
+	hwc->config = code;
+	hwc->idx = -1;
+
+	if (attr->exclude_user)
+		hwc->config |= TILE_CTL_EXCL_USER;
+
+	if (attr->exclude_kernel)
+		hwc->config |= TILE_CTL_EXCL_KERNEL;
+
+	if (attr->exclude_hv)
+		hwc->config |= TILE_CTL_EXCL_HV;
+
+	if (!hwc->sample_period) {
+		hwc->sample_period = tile_pmu->max_period;
+		hwc->last_period = hwc->sample_period;
+		local64_set(&hwc->period_left, hwc->sample_period);
+	}
+	event->destroy = tile_event_destroy;
+	return 0;
+}
+
+static int tile_event_init(struct perf_event *event)
+{
+	int err = 0;
+	perf_irq_t old_irq_handler = NULL;
+
+	if (atomic_inc_return(&tile_active_events) == 1)
+		old_irq_handler = reserve_pmc_hardware(tile_pmu_handle_irq);
+
+	if (old_irq_handler) {
+		pr_warn("PMC hardware busy (reserved by oprofile)\n");
+
+		atomic_dec(&tile_active_events);
+		return -EBUSY;
+	}
+
+	switch (event->attr.type) {
+	case PERF_TYPE_RAW:
+	case PERF_TYPE_HARDWARE:
+	case PERF_TYPE_HW_CACHE:
+		break;
+
+	default:
+		return -ENOENT;
+	}
+
+	err = __tile_event_init(event);
+	if (err) {
+		if (event->destroy)
+			event->destroy(event);
+	}
+	return err;
+}
+
+static struct pmu tilera_pmu = {
+	.event_init	= tile_event_init,
+	.add		= tile_pmu_add,
+	.del		= tile_pmu_del,
+
+	.start		= tile_pmu_start,
+	.stop		= tile_pmu_stop,
+
+	.read		= tile_pmu_read,
+};
+
+/*
+ * PMU's IRQ handler, PMU has 2 interrupts, they share the same handler.
+ */
+int tile_pmu_handle_irq(struct pt_regs *regs, int fault)
+{
+	struct perf_sample_data data;
+	struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
+	struct perf_event *event;
+	struct hw_perf_event *hwc;
+	u64 val;
+	unsigned long status;
+	int bit;
+
+	__get_cpu_var(perf_irqs)++;
+
+	if (!atomic_read(&tile_active_events))
+		return 0;
+
+	status = pmc_get_overflow();
+	pmc_ack_overflow(status);
+
+	for_each_set_bit(bit, &status, tile_pmu->num_counters) {
+
+		event = cpuc->events[bit];
+
+		if (!event)
+			continue;
+
+		if (!test_bit(bit, cpuc->active_mask))
+			continue;
+
+		hwc = &event->hw;
+
+		val = tile_perf_event_update(event);
+		if (val & (1ULL << (tile_pmu->cntval_bits - 1)))
+			continue;
+
+		perf_sample_data_init(&data, 0, event->hw.last_period);
+		if (!tile_event_set_period(event))
+			continue;
+
+		if (perf_event_overflow(event, &data, regs))
+			tile_pmu_stop(event, 0);
+	}
+
+	return 0;
+}
+
+static bool __init supported_pmu(void)
+{
+	tile_pmu = &tilepmu;
+	return true;
+}
+
+int __init init_hw_perf_events(void)
+{
+	supported_pmu();
+	perf_pmu_register(&tilera_pmu, "cpu", PERF_TYPE_RAW);
+	return 0;
+}
+arch_initcall(init_hw_perf_events);
+
+/* Callchain handling code. */
+
+/*
+ * Tile specific backtracing code for perf_events.
+ */
+static inline void perf_callchain(struct perf_callchain_entry *entry,
+		    struct pt_regs *regs)
+{
+	struct KBacktraceIterator kbt;
+	unsigned int i;
+
+	/*
+	 * Get the address just after the "jalr" instruction that
+	 * jumps to the handler for a syscall.  When we find this
+	 * address in a backtrace, we silently ignore it, which gives
+	 * us a one-step backtrace connection from the sys_xxx()
+	 * function in the kernel to the xxx() function in libc.
+	 * Otherwise, we lose the ability to properly attribute time
+	 * from the libc calls to the kernel implementations, since
+	 * oprofile only considers PCs from backtraces a pair at a time.
+	 */
+	unsigned long handle_syscall_pc = handle_syscall_link_address();
+
+	KBacktraceIterator_init(&kbt, NULL, regs);
+	kbt.profile = 1;
+
+	/*
+	 * The sample for the pc is already recorded.  Now we are adding the
+	 * address of the callsites on the stack.  Our iterator starts
+	 * with the frame of the (already sampled) call site.  If our
+	 * iterator contained a "return address" field, we could have just
+	 * used it and wouldn't have needed to skip the first
+	 * frame.  That's in effect what the arm and x86 versions do.
+	 * Instead we peel off the first iteration to get the equivalent
+	 * behavior.
+	 */
+
+	if (KBacktraceIterator_end(&kbt))
+		return;
+	KBacktraceIterator_next(&kbt);
+
+	/*
+	 * Set stack depth to 16 for user and kernel space respectively, that
+	 * is, total 32 stack frames.
+	 */
+	for (i = 0; i < 16; ++i) {
+		unsigned long pc;
+		if (KBacktraceIterator_end(&kbt))
+			break;
+		pc = kbt.it.pc;
+		if (pc != handle_syscall_pc)
+			perf_callchain_store(entry, pc);
+		KBacktraceIterator_next(&kbt);
+	}
+}
+
+void perf_callchain_user(struct perf_callchain_entry *entry,
+		    struct pt_regs *regs)
+{
+	perf_callchain(entry, regs);
+}
+
+void perf_callchain_kernel(struct perf_callchain_entry *entry,
+		      struct pt_regs *regs)
+{
+	perf_callchain(entry, regs);
+}