- 根目录:
- arch
- blackfin
- kernel
- perf_event.c
/*
* Blackfin performance counters
*
* Copyright 2011 Analog Devices Inc.
*
* Ripped from SuperH version:
*
* Copyright (C) 2009 Paul Mundt
*
* Heavily based on the x86 and PowerPC implementations.
*
* x86:
* 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>
*
* ppc:
* Copyright 2008-2009 Paul Mackerras, IBM Corporation.
*
* Licensed under the GPL-2 or later.
*/
#include <linux/kernel.h>
#include <linux/export.h>
#include <linux/init.h>
#include <linux/perf_event.h>
#include <asm/bfin_pfmon.h>
/*
* We have two counters, and each counter can support an event type.
* The 'o' is PFCNTx=1 and 's' is PFCNTx=0
*
* 0x04 o pc invariant branches
* 0x06 o mispredicted branches
* 0x09 o predicted branches taken
* 0x0B o EXCPT insn
* 0x0C o CSYNC/SSYNC insn
* 0x0D o Insns committed
* 0x0E o Interrupts taken
* 0x0F o Misaligned address exceptions
* 0x80 o Code memory fetches stalled due to DMA
* 0x83 o 64bit insn fetches delivered
* 0x9A o data cache fills (bank a)
* 0x9B o data cache fills (bank b)
* 0x9C o data cache lines evicted (bank a)
* 0x9D o data cache lines evicted (bank b)
* 0x9E o data cache high priority fills
* 0x9F o data cache low priority fills
* 0x00 s loop 0 iterations
* 0x01 s loop 1 iterations
* 0x0A s CSYNC/SSYNC stalls
* 0x10 s DAG read/after write hazards
* 0x13 s RAW data hazards
* 0x81 s code TAG stalls
* 0x82 s code fill stalls
* 0x90 s processor to memory stalls
* 0x91 s data memory stalls not hidden by 0x90
* 0x92 s data store buffer full stalls
* 0x93 s data memory write buffer full stalls due to high->low priority
* 0x95 s data memory fill buffer stalls
* 0x96 s data TAG collision stalls
* 0x97 s data collision stalls
* 0x98 s data stalls
* 0x99 s data stalls sent to processor
*/
static const int event_map[] = {
/* use CYCLES cpu register */
[PERF_COUNT_HW_CPU_CYCLES] = -1,
[PERF_COUNT_HW_INSTRUCTIONS] = 0x0D,
[PERF_COUNT_HW_CACHE_REFERENCES] = -1,
[PERF_COUNT_HW_CACHE_MISSES] = 0x83,
[PERF_COUNT_HW_BRANCH_INSTRUCTIONS] = 0x09,
[PERF_COUNT_HW_BRANCH_MISSES] = 0x06,
[PERF_COUNT_HW_BUS_CYCLES] = -1,
};
#define C(x) PERF_COUNT_HW_CACHE_##x
static const int cache_events[PERF_COUNT_HW_CACHE_MAX]
[PERF_COUNT_HW_CACHE_OP_MAX]
[PERF_COUNT_HW_CACHE_RESULT_MAX] =
{
[C(L1D)] = { /* Data bank A */
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = 0,
[C(RESULT_MISS) ] = 0x9A,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = 0,
[C(RESULT_MISS) ] = 0,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = 0,
[C(RESULT_MISS) ] = 0,
},
},
[C(L1I)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = 0,
[C(RESULT_MISS) ] = 0x83,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS) ] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = 0,
[C(RESULT_MISS) ] = 0,
},
},
[C(LL)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS) ] = -1,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS) ] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS) ] = -1,
},
},
[C(DTLB)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS) ] = -1,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS) ] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS) ] = -1,
},
},
[C(ITLB)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS) ] = -1,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS) ] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS) ] = -1,
},
},
[C(BPU)] = {
[C(OP_READ)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS) ] = -1,
},
[C(OP_WRITE)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS) ] = -1,
},
[C(OP_PREFETCH)] = {
[C(RESULT_ACCESS)] = -1,
[C(RESULT_MISS) ] = -1,
},
},
};
const char *perf_pmu_name(void)
{
return "bfin";
}
EXPORT_SYMBOL(perf_pmu_name);
int perf_num_counters(void)
{
return ARRAY_SIZE(event_map);
}
EXPORT_SYMBOL(perf_num_counters);
static u64 bfin_pfmon_read(int idx)
{
return bfin_read32(PFCNTR0 + (idx * 4));
}
static void bfin_pfmon_disable(struct hw_perf_event *hwc, int idx)
{
bfin_write_PFCTL(bfin_read_PFCTL() & ~PFCEN(idx, PFCEN_MASK));
}
static void bfin_pfmon_enable(struct hw_perf_event *hwc, int idx)
{
u32 val, mask;
val = PFPWR;
if (idx) {
mask = ~(PFCNT1 | PFMON1 | PFCEN1 | PEMUSW1);
/* The packed config is for event0, so shift it to event1 slots */
val |= (hwc->config << (PFMON1_P - PFMON0_P));
val |= (hwc->config & PFCNT0) << (PFCNT1_P - PFCNT0_P);
bfin_write_PFCNTR1(0);
} else {
mask = ~(PFCNT0 | PFMON0 | PFCEN0 | PEMUSW0);
val |= hwc->config;
bfin_write_PFCNTR0(0);
}
bfin_write_PFCTL((bfin_read_PFCTL() & mask) | val);
}
static void bfin_pfmon_disable_all(void)
{
bfin_write_PFCTL(bfin_read_PFCTL() & ~PFPWR);
}
static void bfin_pfmon_enable_all(void)
{
bfin_write_PFCTL(bfin_read_PFCTL() | PFPWR);
}
struct cpu_hw_events {
struct perf_event *events[MAX_HWEVENTS];
unsigned long used_mask[BITS_TO_LONGS(MAX_HWEVENTS)];
};
DEFINE_PER_CPU(struct cpu_hw_events, cpu_hw_events);
static int hw_perf_cache_event(int config, int *evp)
{
unsigned long type, op, result;
int ev;
/* unpack config */
type = config & 0xff;
op = (config >> 8) & 0xff;
result = (config >> 16) & 0xff;
if (type >= PERF_COUNT_HW_CACHE_MAX ||
op >= PERF_COUNT_HW_CACHE_OP_MAX ||
result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
return -EINVAL;
ev = cache_events[type][op][result];
if (ev == 0)
return -EOPNOTSUPP;
if (ev == -1)
return -EINVAL;
*evp = ev;
return 0;
}
static void bfin_perf_event_update(struct perf_event *event,
struct hw_perf_event *hwc, int idx)
{
u64 prev_raw_count, new_raw_count;
s64 delta;
int shift = 0;
/*
* Depending on the counter configuration, they may or may not
* be chained, in which case the previous counter value can be
* updated underneath us if the lower-half overflows.
*
* 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 counter atomically.
*
* As there is no interrupt associated with the overflow events,
* this is the simplest approach for maintaining consistency.
*/
again:
prev_raw_count = local64_read(&hwc->prev_count);
new_raw_count = bfin_pfmon_read(idx);
if (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
new_raw_count) != 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
* (counter-)time and add that to the generic counter.
*
* 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);
}
static void bfin_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 (!(event->hw.state & PERF_HES_STOPPED)) {
bfin_pfmon_disable(hwc, idx);
cpuc->events[idx] = NULL;
event->hw.state |= PERF_HES_STOPPED;
}
if ((flags & PERF_EF_UPDATE) && !(event->hw.state & PERF_HES_UPTODATE)) {
bfin_perf_event_update(event, &event->hw, idx);
event->hw.state |= PERF_HES_UPTODATE;
}
}
static void bfin_pmu_start(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 (WARN_ON_ONCE(idx == -1))
return;
if (flags & PERF_EF_RELOAD)
WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
cpuc->events[idx] = event;
event->hw.state = 0;
bfin_pfmon_enable(hwc, idx);
}
static void bfin_pmu_del(struct perf_event *event, int flags)
{
struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
bfin_pmu_stop(event, PERF_EF_UPDATE);
__clear_bit(event->hw.idx, cpuc->used_mask);
perf_event_update_userpage(event);
}
static int bfin_pmu_add(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;
int ret = -EAGAIN;
perf_pmu_disable(event->pmu);
if (__test_and_set_bit(idx, cpuc->used_mask)) {
idx = find_first_zero_bit(cpuc->used_mask, MAX_HWEVENTS);
if (idx == MAX_HWEVENTS)
goto out;
__set_bit(idx, cpuc->used_mask);
hwc->idx = idx;
}
bfin_pfmon_disable(hwc, idx);
event->hw.state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
if (flags & PERF_EF_START)
bfin_pmu_start(event, PERF_EF_RELOAD);
perf_event_update_userpage(event);
ret = 0;
out:
perf_pmu_enable(event->pmu);
return ret;
}
static void bfin_pmu_read(struct perf_event *event)
{
bfin_perf_event_update(event, &event->hw, event->hw.idx);
}
static int bfin_pmu_event_init(struct perf_event *event)
{
struct perf_event_attr *attr = &event->attr;
struct hw_perf_event *hwc = &event->hw;
int config = -1;
int ret;
if (attr->exclude_hv || attr->exclude_idle)
return -EPERM;
/*
* All of the on-chip counters are "limited", in that they have
* no interrupts, and are therefore unable to do sampling without
* further work and timer assistance.
*/
if (hwc->sample_period)
return -EINVAL;
ret = 0;
switch (attr->type) {
case PERF_TYPE_RAW:
config = PFMON(0, attr->config & PFMON_MASK) |
PFCNT(0, !(attr->config & 0x100));
break;
case PERF_TYPE_HW_CACHE:
ret = hw_perf_cache_event(attr->config, &config);
break;
case PERF_TYPE_HARDWARE:
if (attr->config >= ARRAY_SIZE(event_map))
return -EINVAL;
config = event_map[attr->config];
break;
}
if (config == -1)
return -EINVAL;
if (!attr->exclude_kernel)
config |= PFCEN(0, PFCEN_ENABLE_SUPV);
if (!attr->exclude_user)
config |= PFCEN(0, PFCEN_ENABLE_USER);
hwc->config |= config;
return ret;
}
static void bfin_pmu_enable(struct pmu *pmu)
{
struct cpu_hw_events *cpuc = &__get_cpu_var(cpu_hw_events);
struct perf_event *event;
struct hw_perf_event *hwc;
int i;
for (i = 0; i < MAX_HWEVENTS; ++i) {
event = cpuc->events[i];
if (!event)
continue;
hwc = &event->hw;
bfin_pfmon_enable(hwc, hwc->idx);
}
bfin_pfmon_enable_all();
}
static void bfin_pmu_disable(struct pmu *pmu)
{
bfin_pfmon_disable_all();
}
static struct pmu pmu = {
.pmu_enable = bfin_pmu_enable,
.pmu_disable = bfin_pmu_disable,
.event_init = bfin_pmu_event_init,
.add = bfin_pmu_add,
.del = bfin_pmu_del,
.start = bfin_pmu_start,
.stop = bfin_pmu_stop,
.read = bfin_pmu_read,
};
static void bfin_pmu_setup(int cpu)
{
struct cpu_hw_events *cpuhw = &per_cpu(cpu_hw_events, cpu);
memset(cpuhw, 0, sizeof(struct cpu_hw_events));
}
static int __cpuinit
bfin_pmu_notifier(struct notifier_block *self, unsigned long action, void *hcpu)
{
unsigned int cpu = (long)hcpu;
switch (action & ~CPU_TASKS_FROZEN) {
case CPU_UP_PREPARE:
bfin_write_PFCTL(0);
bfin_pmu_setup(cpu);
break;
default:
break;
}
return NOTIFY_OK;
}
static int __init bfin_pmu_init(void)
{
int ret;
ret = perf_pmu_register(&pmu, "cpu", PERF_TYPE_RAW);
if (!ret)
perf_cpu_notifier(bfin_pmu_notifier);
return ret;
}
early_initcall(bfin_pmu_init);