perf_event: aml pmu interrupts routing on g12b [1/1]

PD#SWPL-3088

Problem:
g12b big-little cluster is different from other SoC with pmu
interrupts and registers.
software modifications must adapt to the difference.

Solution:
modify

Verify:
u200 w400

Change-Id: If9217c1025dff5c17d51790f8c216e31b7d6532b
Signed-off-by: Hanjie Lin <hanjie.lin@amlogic.com>
Signed-off-by: Luan Yuan <luan.yuan@amlogic.com>
This commit is contained in:
Hanjie Lin
2018-12-14 19:39:58 +08:00
committed by Dongjin Kim
parent 7a3de9b85c
commit 4d64dc6e9b
28 changed files with 749 additions and 5552 deletions

View File

@@ -119,24 +119,19 @@
bit_mode=<12>;
bit_resolution=<0>;
};
arm_pmu {
compatible = "arm,cortex-a15-pmu";
interrupts = <0 137 4>;
reg = <0xff634400 0x1000>;
/* addr = base + offset << 2 */
sys_cpu_status0_offset = <0xa0>;
sys_cpu_status0_pmuirq_mask = <0xf>;
/* clusterb-enabled; */
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>;
reg = <0xff634680 0x4>;
cpumasks = <0xf>;
/* default 10ms */
relax_timer_ns = <10000000>;
relax-timer-ns = <10000000>;
/* default 10000us */
max_wait_cnt = <10000>;
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {
compatible = "arm,cortex-a15-gic", "arm,cortex-a9-gic";
#interrupt-cells = <3>;

View File

@@ -155,21 +155,17 @@
bit_mode=<12>;
bit_resolution=<0>;
};
arm_pmu {
compatible = "arm,cortex-a15-pmu";
interrupts = <0 137 4>;
reg = <0xff634400 0x1000>;
/* addr = base + offset << 2 */
sys_cpu_status0_offset = <0xa0>;
sys_cpu_status0_pmuirq_mask = <0xf>;
/* clusterb-enabled; */
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>;
reg = <0xff634680 0x4>;
cpumasks = <0xf>;
/* default 10ms */
relax_timer_ns = <10000000>;
relax-timer-ns = <10000000>;
/* default 10000us */
max_wait_cnt = <10000>;
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {

View File

@@ -196,7 +196,7 @@
};
timer {
compatible = "arm,armv8-timer";
compatible = "arm,armv7-timer";
interrupts = <GIC_PPI 13 0xff08>,
<GIC_PPI 14 0xff08>,
<GIC_PPI 11 0xff08>,
@@ -215,9 +215,19 @@
bit_mode=<12>;
bit_resolution=<0>;
};
arm_pmu {
compatible = "arm,armv8-pmuv3";
interrupts = <0 137 4>;
compatible = "arm,cortex-a15-pmu";
clusterb-enabled;
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>,
<GIC_SPI 171 IRQ_TYPE_LEVEL_HIGH>;
reg = <0xff634680 0x4>,
<0xff6347c0 0x04>;
cpumasks = <0x3 0x3C>;
/* default 10ms */
relax-timer-ns = <10000000>;
/* default 10000us */
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {

View File

@@ -134,19 +134,14 @@
arm_pmu {
compatible = "arm,cortex-a15-pmu";
interrupts = <0 137 4>;
reg = <0xc8834400 0x1000>;
/* addr = base + offset << 2 */
sys_cpu_status0_offset = <0xa0>;
sys_cpu_status0_pmuirq_mask = <0xf>;
/* clusterb-enabled; */
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>;
reg = <0xc8834400 0x4>;
cpumasks = <0xf>;
/* default 10ms */
relax_timer_ns = <10000000>;
relax-timer-ns = <10000000>;
/* default 10000us */
max_wait_cnt = <10000>;
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {

View File

@@ -115,7 +115,7 @@
};
timer {
compatible = "arm,armv8-timer";
compatible = "arm,armv7-timer";
interrupts = <GIC_PPI 13 0xff08>,
<GIC_PPI 14 0xff08>,
<GIC_PPI 11 0xff08>,
@@ -133,12 +133,17 @@
bit_mode=<12>;
bit_resolution=<0>;
};
arm_pmu {
compatible = "arm,armv8-pmuv3";
interrupts = <0 137 4>,
<0 138 4>,
<0 153 4>,
<0 154 4>;
compatible = "arm,cortex-a15-pmu";
/* clusterb-enabled; */
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>;
reg = <0xc8834400 0x4>;
cpumasks = <0xf>;
/* default 10ms */
relax-timer-ns = <10000000>;
/* default 10000us */
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {

View File

@@ -205,7 +205,7 @@
};
timer {
compatible = "arm,armv8-timer";
compatible = "arm,armv7-timer";
interrupts = <GIC_PPI 13 0xff08>,
<GIC_PPI 14 0xff08>,
<GIC_PPI 11 0xff08>,
@@ -223,12 +223,17 @@
bit_mode=<12>;
bit_resolution=<0>;
};
arm_pmu {
compatible = "arm,armv8-pmuv3";
interrupts = <0 137 4>,
<0 138 4>,
<0 153 4>,
<0 154 4>;
compatible = "arm,cortex-a15-pmu";
/* clusterb-enabled; */
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>;
reg = <0xc8834400 0x4>;
cpumasks = <0xf>;
/* default 10ms */
relax-timer-ns = <10000000>;
/* default 10000us */
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

View File

@@ -134,19 +134,14 @@
arm_pmu {
compatible = "arm,cortex-a15-pmu";
interrupts = <0 137 4>;
reg = <0xff634400 0x1000>;
/* addr = base + offset << 2 */
sys_cpu_status0_offset = <0xa0>;
sys_cpu_status0_pmuirq_mask = <0xf>;
/* clusterb-enabled; */
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>;
reg = <0xff634680 0x4>;
cpumasks = <0xf>;
/* default 10ms */
relax_timer_ns = <10000000>;
relax-timer-ns = <10000000>;
/* default 10000us */
max_wait_cnt = <10000>;
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {

View File

@@ -5,11 +5,7 @@
#include <linux/threads.h>
#include <asm/irq.h>
#ifdef CONFIG_AMLOGIC_MODIFY
#define NR_IPI 8
#else
#define NR_IPI 7
#endif
typedef struct {
unsigned int __softirq_pending;

View File

@@ -29,51 +29,4 @@ extern unsigned long perf_misc_flags(struct pt_regs *regs);
(regs)->ARM_sp = current_stack_pointer; \
(regs)->ARM_cpsr = SVC_MODE; \
}
#ifdef CONFIG_AMLOGIC_MODIFY
extern void armv8pmu_handle_irq_ipi(void);
struct amlpmu_fixup_cpuinfo {
int irq_num;
int fix_done;
unsigned long irq_cnt;
unsigned long empty_irq_cnt;
unsigned long irq_time;
unsigned long empty_irq_time;
unsigned long last_irq_cnt;
unsigned long last_empty_irq_cnt;
unsigned long last_irq_time;
unsigned long last_empty_irq_time;
};
struct amlpmu_fixup_context {
struct amlpmu_fixup_cpuinfo __percpu *cpuinfo;
/* struct arm_pmu */
void *dev;
/* sys_cpu_status0 reg */
unsigned int *sys_cpu_status0;
/*
* In main pmu irq route wait for other cpu fix done may cause lockup,
* when lockup we disable main irq for a while.
* relax_timer will enable main irq again.
*/
struct hrtimer relax_timer;
/* dts prop */
unsigned int sys_cpu_status0_offset;
unsigned int sys_cpu_status0_pmuirq_mask;
unsigned int relax_timer_ns;
unsigned int max_wait_cnt;
};
#endif
#endif /* __ARM_PERF_EVENT_H__ */

View File

@@ -18,10 +18,6 @@
#ifdef CONFIG_CPU_V7
#ifdef CONFIG_AMLOGIC_MODIFY
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#endif
#include <asm/cp15.h>
#include <asm/cputype.h>
#include <asm/irq_regs.h>
@@ -32,19 +28,6 @@
#include <linux/perf/arm_pmu.h>
#include <linux/platform_device.h>
#ifdef CONFIG_AMLOGIC_MODIFY
#include <linux/of.h>
#include <linux/of_irq.h>
#include <linux/of_device.h>
#include <linux/irq.h>
#include <asm/irq.h>
#include <linux/interrupt.h>
#include <linux/irqdesc.h>
#include <linux/of_address.h>
#include <linux/delay.h>
#endif
/*
* Common ARMv7 event types
*
@@ -963,160 +946,13 @@ static void armv7pmu_disable_event(struct perf_event *event)
}
#ifdef CONFIG_AMLOGIC_MODIFY
static struct amlpmu_fixup_context amlpmu_fixup_ctx;
static enum hrtimer_restart amlpmu_relax_timer_func(struct hrtimer *timer)
{
struct amlpmu_fixup_cpuinfo *ci;
ci = per_cpu_ptr(amlpmu_fixup_ctx.cpuinfo, 0);
pr_alert("enable cpu0_irq %d again, irq cnt = %lu\n",
ci->irq_num,
ci->irq_cnt);
enable_irq(ci->irq_num);
return HRTIMER_NORESTART;
}
static void amlpmu_relax_timer_start(int other_cpu)
{
struct amlpmu_fixup_cpuinfo *ci;
int cpu;
cpu = smp_processor_id();
WARN_ON(cpu != 0);
ci = per_cpu_ptr(amlpmu_fixup_ctx.cpuinfo, 0);
pr_alert("wait cpu %d fixup done timeout, main cpu irq cnt = %lu\n",
other_cpu,
ci->irq_cnt);
if (hrtimer_active(&amlpmu_fixup_ctx.relax_timer)) {
pr_alert("relax_timer already active, return!\n");
return;
}
disable_irq_nosync(ci->irq_num);
hrtimer_start(&amlpmu_fixup_ctx.relax_timer,
ns_to_ktime(amlpmu_fixup_ctx.relax_timer_ns),
HRTIMER_MODE_REL);
}
#include <linux/perf/arm_pmu.h>
static irqreturn_t armv7pmu_handle_irq(int irq_num, void *dev);
void armv8pmu_handle_irq_ipi(void)
void amlpmu_handle_irq_ipi(void *arg)
{
int cpu = smp_processor_id();
WARN_ON(cpu == 0);
WARN_ON(!amlpmu_fixup_ctx.dev);
armv7pmu_handle_irq(-1, amlpmu_fixup_ctx.dev);
}
static int aml_pmu_fix(void)
{
int i;
int cpu;
int pmuirq_val;
struct amlpmu_fixup_cpuinfo *ci;
int max_wait_cnt = amlpmu_fixup_ctx.max_wait_cnt;
pmuirq_val = readl(amlpmu_fixup_ctx.sys_cpu_status0);
pmuirq_val &= amlpmu_fixup_ctx.sys_cpu_status0_pmuirq_mask;
for (cpu = 0; cpu < num_possible_cpus(); cpu++) {
if (pmuirq_val & (1<<cpu)) {
if (cpu == 0) {
pr_debug("cpu0 shouldn't fix pmuirq = 0x%x\n",
pmuirq_val);
} else {
pr_debug("fix pmu irq cpu %d, pmuirq = 0x%x\n",
cpu,
pmuirq_val);
ci = per_cpu_ptr(amlpmu_fixup_ctx.cpuinfo,
cpu);
ci->fix_done = 0;
/* aml pmu IPI will set fix_done to 1 */
mb();
smp_send_aml_pmu(cpu);
for (i = 0; i < max_wait_cnt; i++) {
if (READ_ONCE(ci->fix_done))
break;
udelay(1);
}
if (i == amlpmu_fixup_ctx.max_wait_cnt)
amlpmu_relax_timer_start(cpu);
return 0;
}
}
}
return 1;
}
static void aml_pmu_fix_stat_account(int is_empty_irq)
{
int freq;
unsigned long time = jiffies;
struct amlpmu_fixup_cpuinfo *ci;
ci = this_cpu_ptr(amlpmu_fixup_ctx.cpuinfo);
ci->irq_cnt++;
ci->irq_time = time;
if (!ci->last_irq_cnt) {
ci->last_irq_cnt = ci->irq_cnt;
ci->last_irq_time = ci->irq_time;
}
if (is_empty_irq) {
ci->empty_irq_cnt++;
ci->empty_irq_time = time;
if (!ci->last_empty_irq_cnt) {
ci->last_empty_irq_cnt = ci->empty_irq_cnt;
ci->last_empty_irq_time = ci->empty_irq_time;
}
}
if (time_after(ci->irq_time, ci->last_irq_time + HZ)) {
freq = ci->irq_cnt - ci->last_irq_cnt;
freq = freq * HZ / (ci->irq_time - ci->last_irq_time);
pr_debug("irq_cnt = %lu, irq_last_cnt = %lu, freq = %d\n",
ci->irq_cnt,
ci->last_irq_cnt,
freq);
ci->last_irq_cnt = ci->irq_cnt;
ci->last_irq_time = ci->irq_time;
}
if (is_empty_irq &&
time_after(ci->empty_irq_time, ci->last_empty_irq_time + HZ)) {
freq = ci->empty_irq_cnt - ci->last_empty_irq_cnt;
freq *= HZ;
freq /= (ci->empty_irq_time - ci->last_empty_irq_time);
pr_debug("empty_irq_cnt = %lu, freq = %d\n",
ci->empty_irq_cnt,
freq);
ci->last_empty_irq_cnt = ci->empty_irq_cnt;
ci->last_empty_irq_time = ci->empty_irq_time;
}
armv7pmu_handle_irq(-1, amlpmu_ctx.pmu);
}
#endif
@@ -1129,47 +965,25 @@ static irqreturn_t armv7pmu_handle_irq(int irq_num, void *dev)
struct pt_regs *regs;
int idx;
#ifdef CONFIG_AMLOGIC_MODIFY
int cpu;
int is_empty_irq = 0;
struct amlpmu_fixup_cpuinfo *ci;
ci = this_cpu_ptr(amlpmu_fixup_ctx.cpuinfo);
ci->irq_num = irq_num;
amlpmu_fixup_ctx.dev = dev;
cpu = smp_processor_id();
#endif
/*
* Get and reset the IRQ flags
*/
pmnc = armv7_pmnc_getreset_flags();
#ifdef CONFIG_AMLOGIC_MODIFY
ci->fix_done = 1;
#endif
/* amlpmu have routed the interrupt successfully, return IRQ_HANDLED */
if (amlpmu_handle_irq(cpu_pmu,
irq_num,
armv7_pmnc_has_overflowed(pmnc)))
return IRQ_HANDLED;
#else
/*
* Did an overflow occur?
*/
#ifdef CONFIG_AMLOGIC_MODIFY
if (!armv7_pmnc_has_overflowed(pmnc)) {
is_empty_irq = 1;
if (cpu == 0)
is_empty_irq = aml_pmu_fix();
}
aml_pmu_fix_stat_account(is_empty_irq);
/* txlx have some empty pmu irqs, so return IRQ_HANDLED */
if (is_empty_irq)
return IRQ_HANDLED;
#else
if (!armv7_pmnc_has_overflowed(pmnc))
return IRQ_NONE;
#endif
/*
* Handle the counter(s) overflow(s)
*/
@@ -2224,88 +2038,8 @@ static const struct pmu_probe_info armv7_pmu_probe_table[] = {
{ /* sentinel value */ }
};
#ifdef CONFIG_AMLOGIC_MODIFY
static int amlpmu_fixup_init(struct platform_device *pdev)
{
int ret;
void __iomem *base;
amlpmu_fixup_ctx.cpuinfo = __alloc_percpu(
sizeof(struct amlpmu_fixup_cpuinfo), 2 * sizeof(void *));
if (!amlpmu_fixup_ctx.cpuinfo) {
pr_err("alloc percpu failed\n");
return -ENOMEM;
}
base = of_iomap(pdev->dev.of_node, 0);
if (IS_ERR(base)) {
pr_err("of_iomap() failed, base = %p\n", base);
return PTR_ERR(base);
}
ret = of_property_read_u32(pdev->dev.of_node,
"sys_cpu_status0_offset",
&amlpmu_fixup_ctx.sys_cpu_status0_offset);
if (ret) {
pr_err("read sys_cpu_status0_offset failed, ret = %d\n", ret);
return 1;
}
pr_debug("sys_cpu_status0_offset = 0x%0x\n",
amlpmu_fixup_ctx.sys_cpu_status0_offset);
ret = of_property_read_u32(pdev->dev.of_node,
"sys_cpu_status0_pmuirq_mask",
&amlpmu_fixup_ctx.sys_cpu_status0_pmuirq_mask);
if (ret) {
pr_err("read sys_cpu_status0_pmuirq_mask failed, ret = %d\n",
ret);
return 1;
}
pr_debug("sys_cpu_status0_pmuirq_mask = 0x%0x\n",
amlpmu_fixup_ctx.sys_cpu_status0_pmuirq_mask);
ret = of_property_read_u32(pdev->dev.of_node,
"relax_timer_ns",
&amlpmu_fixup_ctx.relax_timer_ns);
if (ret) {
pr_err("read prop relax_timer_ns failed, ret = %d\n", ret);
return 1;
}
pr_debug("relax_timer_ns = %u\n", amlpmu_fixup_ctx.relax_timer_ns);
ret = of_property_read_u32(pdev->dev.of_node,
"max_wait_cnt",
&amlpmu_fixup_ctx.max_wait_cnt);
if (ret) {
pr_err("read prop max_wait_cnt failed, ret = %d\n", ret);
return 1;
}
pr_debug("max_wait_cnt = %u\n", amlpmu_fixup_ctx.max_wait_cnt);
base += (amlpmu_fixup_ctx.sys_cpu_status0_offset << 2);
amlpmu_fixup_ctx.sys_cpu_status0 = base;
pr_debug("sys_cpu_status0 = %p\n", amlpmu_fixup_ctx.sys_cpu_status0);
hrtimer_init(&amlpmu_fixup_ctx.relax_timer,
CLOCK_MONOTONIC,
HRTIMER_MODE_REL);
amlpmu_fixup_ctx.relax_timer.function = amlpmu_relax_timer_func;
return 0;
}
#endif
static int armv7_pmu_device_probe(struct platform_device *pdev)
{
#ifdef CONFIG_AMLOGIC_MODIFY
if (amlpmu_fixup_init(pdev))
return 1;
#endif
return arm_pmu_device_probe(pdev, armv7_pmu_of_device_ids,
armv7_pmu_probe_table);
}

View File

@@ -51,10 +51,6 @@
#include <asm/mach/arch.h>
#include <asm/mpu.h>
#ifdef CONFIG_AMLOGIC_MODIFY
#include <asm/perf_event.h>
#endif
#define CREATE_TRACE_POINTS
#include <trace/events/ipi.h>
@@ -79,9 +75,6 @@ enum ipi_msg_type {
IPI_CPU_STOP,
IPI_IRQ_WORK,
IPI_COMPLETION,
#ifdef CONFIG_AMLOGIC_MODIFY
IPI_AML_PMU,
#endif
IPI_CPU_BACKTRACE,
/*
* SGI8-15 can be reserved by secure firmware, and thus may
@@ -525,9 +518,6 @@ static const char *ipi_types[NR_IPI] __tracepoint_string = {
S(IPI_CPU_STOP, "CPU stop interrupts"),
S(IPI_IRQ_WORK, "IRQ work interrupts"),
S(IPI_COMPLETION, "completion interrupts"),
#ifdef CONFIG_AMLOGIC_MODIFY
S(IPI_AML_PMU, "AML pmu cross interrupts"),
#endif
};
static void smp_cross_call(const struct cpumask *target, unsigned int ipinr)
@@ -536,13 +526,6 @@ static void smp_cross_call(const struct cpumask *target, unsigned int ipinr)
__smp_cross_call(target, ipinr);
}
#ifdef CONFIG_AMLOGIC_MODIFY
void smp_send_aml_pmu(int cpu)
{
smp_cross_call(cpumask_of(cpu), IPI_AML_PMU);
}
#endif
void show_ipi_list(struct seq_file *p, int prec)
{
unsigned int cpu, i;
@@ -706,12 +689,6 @@ void handle_IPI(int ipinr, struct pt_regs *regs)
printk_nmi_exit();
break;
#ifdef CONFIG_AMLOGIC_MODIFY
case IPI_AML_PMU:
armv8pmu_handle_irq_ipi();
break;
#endif
default:
pr_crit("CPU%u: Unknown IPI message 0x%x\n",
cpu, ipinr);

View File

@@ -119,24 +119,19 @@
bit_mode=<12>;
bit_resolution=<0>;
};
arm_pmu {
compatible = "arm,armv8-pmuv3";
interrupts = <0 137 4>;
reg = <0x0 0xff634400 0 0x1000>;
/* addr = base + offset << 2 */
sys_cpu_status0_offset = <0xa0>;
sys_cpu_status0_pmuirq_mask = <0xf>;
/* clusterb-enabled; */
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>;
reg = <0x0 0xff634680 0x0 0x4>;
cpumasks = <0xf>;
/* default 10ms */
relax_timer_ns = <10000000>;
relax-timer-ns = <10000000>;
/* default 10000us */
max_wait_cnt = <10000>;
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {
compatible = "arm,cortex-a15-gic", "arm,cortex-a9-gic";
#interrupt-cells = <3>;

View File

@@ -155,21 +155,17 @@
bit_mode=<12>;
bit_resolution=<0>;
};
arm_pmu {
compatible = "arm,armv8-pmuv3";
interrupts = <0 137 4>;
reg = <0x0 0xff634400 0 0x1000>;
/* addr = base + offset << 2 */
sys_cpu_status0_offset = <0xa0>;
sys_cpu_status0_pmuirq_mask = <0xf>;
compatible = "arm,cortex-a15-pmu";
/* clusterb-enabled; */
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>;
reg = <0xff634680 0x4>;
cpumasks = <0xf>;
/* default 10ms */
relax_timer_ns = <10000000>;
relax-timer-ns = <10000000>;
/* default 10000us */
max_wait_cnt = <10000>;
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {
@@ -1759,7 +1755,7 @@
max-duty-cycle = <1250>;
/* Voltage Duty-Cycle */
voltage-table = <1022000 0>,
<1011000 3>,
<1011000 3>,
<1001000 6>,
<991000 10>,
<981000 13>,

View File

@@ -215,9 +215,19 @@
bit_mode=<12>;
bit_resolution=<0>;
};
arm_pmu {
compatible = "arm,armv8-pmuv3";
interrupts = <0 137 4>;
clusterb-enabled;
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>,
<GIC_SPI 171 IRQ_TYPE_LEVEL_HIGH>;
reg = <0x0 0xff634680 0x0 0x4>,
<0x0 0xff6347c0 0x0 0x04>;
cpumasks = <0x3 0x3C>;
/* default 10ms */
relax-timer-ns = <10000000>;
/* default 10000us */
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {

View File

@@ -119,21 +119,17 @@
bit_mode=<12>;
bit_resolution=<0>;
};
arm_pmu {
compatible = "arm,armv8-pmuv3";
interrupts = <0 137 4>;
reg = <0x0 0xc8834400 0 0x1000>;
/* addr = base + offset << 2 */
sys_cpu_status0_offset = <0xa0>;
sys_cpu_status0_pmuirq_mask = <0xf>;
/* clusterb-enabled; */
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>;
reg = <0x0 0xc8834400 0x0 0x4>;
cpumasks = <0xf>;
/* default 10ms */
relax_timer_ns = <10000000>;
relax-timer-ns = <10000000>;
/* default 10000us */
max_wait_cnt = <10000>;
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {

View File

@@ -133,12 +133,17 @@
bit_mode=<12>;
bit_resolution=<0>;
};
arm_pmu {
compatible = "arm,armv8-pmuv3";
interrupts = <0 137 4>,
<0 138 4>,
<0 153 4>,
<0 154 4>;
/* clusterb-enabled; */
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>;
reg = <0x0 0xc8834400 0x0 0x4>;
cpumasks = <0xf>;
/* default 10ms */
relax-timer-ns = <10000000>;
/* default 10000us */
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {

View File

@@ -223,12 +223,17 @@
bit_mode=<12>;
bit_resolution=<0>;
};
arm_pmu {
compatible = "arm,armv8-pmuv3";
interrupts = <0 137 4>,
<0 138 4>,
<0 153 4>,
<0 154 4>;
/* clusterb-enabled; */
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>;
reg = <0x0 0xc8834400 0x0 0x4>;
cpumasks = <0xf>;
/* default 10ms */
relax-timer-ns = <10000000>;
/* default 10000us */
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {

File diff suppressed because it is too large Load Diff

View File

@@ -134,19 +134,14 @@
arm_pmu {
compatible = "arm,armv8-pmuv3";
interrupts = <0 137 4>;
reg = <0x0 0xff634400 0 0x1000>;
/* addr = base + offset << 2 */
sys_cpu_status0_offset = <0xa0>;
sys_cpu_status0_pmuirq_mask = <0xf>;
/* clusterb-enabled; */
interrupts = <GIC_SPI 137 IRQ_TYPE_LEVEL_HIGH>;
reg = <0x0 0xff634680 0x0 0x4>;
cpumasks = <0xf>;
/* default 10ms */
relax_timer_ns = <10000000>;
relax-timer-ns = <10000000>;
/* default 10000us */
max_wait_cnt = <10000>;
max-wait-cnt = <10000>;
};
gic: interrupt-controller@2c001000 {

View File

@@ -20,11 +20,7 @@
#include <linux/threads.h>
#include <asm/irq.h>
#ifdef CONFIG_AMLOGIC_MODIFY
#define NR_IPI 7
#else
#define NR_IPI 6
#endif
typedef struct {
unsigned int __softirq_pending;

View File

@@ -92,50 +92,6 @@ extern unsigned long perf_misc_flags(struct pt_regs *regs);
}
#ifdef CONFIG_AMLOGIC_MODIFY
extern void armv8pmu_handle_irq_ipi(void);
struct amlpmu_fixup_cpuinfo {
int irq_num;
int fix_done;
unsigned long irq_cnt;
unsigned long empty_irq_cnt;
unsigned long irq_time;
unsigned long empty_irq_time;
unsigned long last_irq_cnt;
unsigned long last_empty_irq_cnt;
unsigned long last_irq_time;
unsigned long last_empty_irq_time;
};
struct amlpmu_fixup_context {
struct amlpmu_fixup_cpuinfo __percpu *cpuinfo;
/* struct arm_pmu */
void *dev;
/* sys_cpu_status0 reg */
unsigned int *sys_cpu_status0;
/*
* In main pmu irq route wait for other cpu fix done may cause lockup,
* when lockup we disable main irq for a while.
* relax_timer will enable main irq again.
*/
struct hrtimer relax_timer;
/* dts prop */
unsigned int sys_cpu_status0_offset;
unsigned int sys_cpu_status0_pmuirq_mask;
unsigned int relax_timer_ns;
unsigned int max_wait_cnt;
};
#endif
#endif

View File

@@ -19,10 +19,6 @@
* along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
#ifdef CONFIG_AMLOGIC_MODIFY
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
#endif
#include <asm/irq_regs.h>
#include <asm/perf_event.h>
#include <asm/sysreg.h>
@@ -34,18 +30,6 @@
#include <linux/platform_device.h>
#ifdef CONFIG_AMLOGIC_MODIFY
#include <linux/of.h>
#include <linux/of_irq.h>
#include <linux/of_device.h>
#include <linux/irq.h>
#include <asm/irq.h>
#include <linux/interrupt.h>
#include <linux/irqdesc.h>
#include <linux/of_address.h>
#include <linux/delay.h>
#endif
/*
* ARMv8 PMUv3 Performance Events handling code.
* Common event types (some are defined in asm/perf_event.h).
@@ -766,161 +750,13 @@ static void armv8pmu_disable_event(struct perf_event *event)
}
#ifdef CONFIG_AMLOGIC_MODIFY
static struct amlpmu_fixup_context amlpmu_fixup_ctx;
static enum hrtimer_restart amlpmu_relax_timer_func(struct hrtimer *timer)
{
struct amlpmu_fixup_cpuinfo *ci;
ci = per_cpu_ptr(amlpmu_fixup_ctx.cpuinfo, 0);
pr_alert("enable cpu0_irq %d again, irq cnt = %lu\n",
ci->irq_num,
ci->irq_cnt);
enable_irq(ci->irq_num);
return HRTIMER_NORESTART;
}
static void amlpmu_relax_timer_start(int other_cpu)
{
struct amlpmu_fixup_cpuinfo *ci;
int cpu;
cpu = smp_processor_id();
WARN_ON(cpu != 0);
ci = per_cpu_ptr(amlpmu_fixup_ctx.cpuinfo, 0);
pr_alert("wait cpu %d fixup done timeout, main cpu irq cnt = %lu\n",
other_cpu,
ci->irq_cnt);
if (hrtimer_active(&amlpmu_fixup_ctx.relax_timer)) {
pr_alert("relax_timer already active, return!\n");
return;
}
disable_irq_nosync(ci->irq_num);
hrtimer_start(&amlpmu_fixup_ctx.relax_timer,
ns_to_ktime(amlpmu_fixup_ctx.relax_timer_ns),
HRTIMER_MODE_REL);
}
#include <linux/perf/arm_pmu.h>
static irqreturn_t armv8pmu_handle_irq(int irq_num, void *dev);
void armv8pmu_handle_irq_ipi(void)
void amlpmu_handle_irq_ipi(void *arg)
{
int cpu = smp_processor_id();
WARN_ON(cpu == 0);
WARN_ON(!amlpmu_fixup_ctx.dev);
armv8pmu_handle_irq(-1, amlpmu_fixup_ctx.dev);
}
static int aml_pmu_fix(void)
{
int i;
int cpu;
int pmuirq_val;
struct amlpmu_fixup_cpuinfo *ci;
int max_wait_cnt = amlpmu_fixup_ctx.max_wait_cnt;
pmuirq_val = readl(amlpmu_fixup_ctx.sys_cpu_status0);
pmuirq_val &= amlpmu_fixup_ctx.sys_cpu_status0_pmuirq_mask;
for (cpu = 0; cpu < num_possible_cpus(); cpu++) {
if (pmuirq_val & (1<<cpu)) {
if (cpu == 0) {
pr_debug("cpu0 shouldn't fix pmuirq = 0x%x\n",
pmuirq_val);
} else {
pr_debug("fix pmu irq cpu %d, pmuirq = 0x%x\n",
cpu,
pmuirq_val);
ci = per_cpu_ptr(amlpmu_fixup_ctx.cpuinfo,
cpu);
ci->fix_done = 0;
/* aml pmu IPI will set fix_done to 1 */
mb();
smp_send_aml_pmu(cpu);
for (i = 0; i < max_wait_cnt; i++) {
if (READ_ONCE(ci->fix_done))
break;
udelay(1);
}
if (i == amlpmu_fixup_ctx.max_wait_cnt)
amlpmu_relax_timer_start(cpu);
return 0;
}
}
}
return 1;
}
static void aml_pmu_fix_stat_account(int is_empty_irq)
{
int freq;
unsigned long time = jiffies;
struct amlpmu_fixup_cpuinfo *ci;
ci = this_cpu_ptr(amlpmu_fixup_ctx.cpuinfo);
ci->irq_cnt++;
ci->irq_time = time;
if (!ci->last_irq_cnt) {
ci->last_irq_cnt = ci->irq_cnt;
ci->last_irq_time = ci->irq_time;
}
if (is_empty_irq) {
ci->empty_irq_cnt++;
ci->empty_irq_time = time;
if (!ci->last_empty_irq_cnt) {
ci->last_empty_irq_cnt = ci->empty_irq_cnt;
ci->last_empty_irq_time = ci->empty_irq_time;
}
}
if (time_after(ci->irq_time, ci->last_irq_time + HZ)) {
freq = ci->irq_cnt - ci->last_irq_cnt;
freq = freq * HZ / (ci->irq_time - ci->last_irq_time);
pr_debug("irq_cnt = %lu, irq_last_cnt = %lu, freq = %d\n",
ci->irq_cnt,
ci->last_irq_cnt,
freq);
ci->last_irq_cnt = ci->irq_cnt;
ci->last_irq_time = ci->irq_time;
}
if (is_empty_irq &&
time_after(ci->empty_irq_time, ci->last_empty_irq_time + HZ)) {
freq = ci->empty_irq_cnt - ci->last_empty_irq_cnt;
freq *= HZ;
freq /= (ci->empty_irq_time - ci->last_empty_irq_time);
pr_debug("empty_irq_cnt = %lu, freq = %d\n",
ci->empty_irq_cnt,
freq);
ci->last_empty_irq_cnt = ci->empty_irq_cnt;
ci->last_empty_irq_time = ci->empty_irq_time;
}
armv8pmu_handle_irq(-1, amlpmu_ctx.pmu);
}
#endif
@@ -933,42 +769,21 @@ static irqreturn_t armv8pmu_handle_irq(int irq_num, void *dev)
struct pt_regs *regs;
int idx;
#ifdef CONFIG_AMLOGIC_MODIFY
int cpu;
int is_empty_irq = 0;
struct amlpmu_fixup_cpuinfo *ci;
ci = this_cpu_ptr(amlpmu_fixup_ctx.cpuinfo);
ci->irq_num = irq_num;
amlpmu_fixup_ctx.dev = dev;
cpu = smp_processor_id();
#endif
/*
* Get and reset the IRQ flags
*/
pmovsr = armv8pmu_getreset_flags();
#ifdef CONFIG_AMLOGIC_MODIFY
ci->fix_done = 1;
#endif
/* amlpmu have routed the interrupt already, so return IRQ_HANDLED */
if (amlpmu_handle_irq(cpu_pmu,
irq_num,
armv8pmu_has_overflowed(pmovsr)))
return IRQ_HANDLED;
#else
/*
* Did an overflow occur?
*/
#ifdef CONFIG_AMLOGIC_MODIFY
if (!armv8pmu_has_overflowed(pmovsr)) {
is_empty_irq = 1;
if (cpu == 0)
is_empty_irq = aml_pmu_fix();
}
aml_pmu_fix_stat_account(is_empty_irq);
/* txlx have some empty pmu irqs, so return IRQ_HANDLED */
if (is_empty_irq)
return IRQ_HANDLED;
#else
if (!armv8pmu_has_overflowed(pmovsr))
return IRQ_NONE;
#endif
@@ -1014,9 +829,6 @@ static irqreturn_t armv8pmu_handle_irq(int irq_num, void *dev)
return IRQ_HANDLED;
}
static void armv8pmu_start(struct arm_pmu *cpu_pmu)
{
unsigned long flags;
@@ -1308,87 +1120,8 @@ static const struct pmu_probe_info armv8_pmu_probe_table[] = {
};
#ifdef CONFIG_AMLOGIC_MODIFY
static int amlpmu_fixup_init(struct platform_device *pdev)
{
int ret;
void __iomem *base;
amlpmu_fixup_ctx.cpuinfo = __alloc_percpu(
sizeof(struct amlpmu_fixup_cpuinfo), 2 * sizeof(void *));
if (!amlpmu_fixup_ctx.cpuinfo) {
pr_err("alloc percpu failed\n");
return -ENOMEM;
}
base = of_iomap(pdev->dev.of_node, 0);
if (IS_ERR(base)) {
pr_err("of_iomap() failed, base = %p\n", base);
return PTR_ERR(base);
}
ret = of_property_read_u32(pdev->dev.of_node,
"sys_cpu_status0_offset",
&amlpmu_fixup_ctx.sys_cpu_status0_offset);
if (ret) {
pr_err("read sys_cpu_status0_offset failed, ret = %d\n", ret);
return 1;
}
pr_debug("sys_cpu_status0_offset = 0x%0x\n",
amlpmu_fixup_ctx.sys_cpu_status0_offset);
ret = of_property_read_u32(pdev->dev.of_node,
"sys_cpu_status0_pmuirq_mask",
&amlpmu_fixup_ctx.sys_cpu_status0_pmuirq_mask);
if (ret) {
pr_err("read sys_cpu_status0_pmuirq_mask failed, ret = %d\n",
ret);
return 1;
}
pr_debug("sys_cpu_status0_pmuirq_mask = 0x%0x\n",
amlpmu_fixup_ctx.sys_cpu_status0_pmuirq_mask);
ret = of_property_read_u32(pdev->dev.of_node,
"relax_timer_ns",
&amlpmu_fixup_ctx.relax_timer_ns);
if (ret) {
pr_err("read prop relax_timer_ns failed, ret = %d\n", ret);
return 1;
}
pr_debug("relax_timer_ns = %u\n", amlpmu_fixup_ctx.relax_timer_ns);
ret = of_property_read_u32(pdev->dev.of_node,
"max_wait_cnt",
&amlpmu_fixup_ctx.max_wait_cnt);
if (ret) {
pr_err("read prop max_wait_cnt failed, ret = %d\n", ret);
return 1;
}
pr_debug("max_wait_cnt = %u\n", amlpmu_fixup_ctx.max_wait_cnt);
base += (amlpmu_fixup_ctx.sys_cpu_status0_offset << 2);
amlpmu_fixup_ctx.sys_cpu_status0 = base;
pr_debug("sys_cpu_status0 = %p\n", amlpmu_fixup_ctx.sys_cpu_status0);
hrtimer_init(&amlpmu_fixup_ctx.relax_timer,
CLOCK_MONOTONIC,
HRTIMER_MODE_REL);
amlpmu_fixup_ctx.relax_timer.function = amlpmu_relax_timer_func;
return 0;
}
#endif
static int armv8_pmu_device_probe(struct platform_device *pdev)
{
#ifdef CONFIG_AMLOGIC_MODIFY
if (amlpmu_fixup_init(pdev))
return 1;
#endif
if (acpi_disabled)
return arm_pmu_device_probe(pdev, armv8_pmu_of_device_ids,
NULL);

View File

@@ -55,10 +55,6 @@
#include <asm/ptrace.h>
#include <asm/virt.h>
#ifdef CONFIG_AMLOGIC_MODIFY
#include <asm/perf_event.h>
#endif
#ifdef CONFIG_AMLOGIC_VMAP
#include <linux/amlogic/vmap_stack.h>
#endif
@@ -84,12 +80,7 @@ enum ipi_msg_type {
IPI_CPU_STOP,
IPI_TIMER,
IPI_IRQ_WORK,
#ifdef CONFIG_AMLOGIC_MODIFY
IPI_WAKEUP,
IPI_AML_PMU
#else
IPI_WAKEUP
#endif
};
#ifdef CONFIG_ARM64_VHE
@@ -775,9 +766,6 @@ static const char *ipi_types[NR_IPI] __tracepoint_string = {
S(IPI_TIMER, "Timer broadcast interrupts"),
S(IPI_IRQ_WORK, "IRQ work interrupts"),
S(IPI_WAKEUP, "CPU wake-up interrupts"),
#ifdef CONFIG_AMLOGIC_MODIFY
S(IPI_AML_PMU, "AML pmu cross interrupts"),
#endif
};
static void smp_cross_call(const struct cpumask *target, unsigned int ipinr)
@@ -786,13 +774,6 @@ static void smp_cross_call(const struct cpumask *target, unsigned int ipinr)
__smp_cross_call(target, ipinr);
}
#ifdef CONFIG_AMLOGIC_MODIFY
void smp_send_aml_pmu(int cpu)
{
smp_cross_call(cpumask_of(cpu), IPI_AML_PMU);
}
#endif
void show_ipi_list(struct seq_file *p, int prec)
{
unsigned int cpu, i;
@@ -911,12 +892,6 @@ void handle_IPI(int ipinr, struct pt_regs *regs)
break;
#endif
#ifdef CONFIG_AMLOGIC_MODIFY
case IPI_AML_PMU:
armv8pmu_handle_irq_ipi();
break;
#endif
default:
pr_crit("CPU%u: Unknown IPI message 0x%x\n", cpu, ipinr);
break;

View File

@@ -1013,6 +1013,517 @@ static int of_pmu_irq_cfg(struct arm_pmu *pmu)
return 0;
}
#ifdef CONFIG_AMLOGIC_MODIFY
#include <linux/of_address.h>
#include <linux/delay.h>
struct amlpmu_context amlpmu_ctx;
static enum hrtimer_restart amlpmu_relax_timer_func(struct hrtimer *timer)
{
struct amlpmu_context *ctx = &amlpmu_ctx;
struct amlpmu_cpuinfo *ci;
ci = per_cpu_ptr(ctx->cpuinfo, 0);
pr_info("enable cpu0_irq %d again, irq cnt = %lu\n",
ci->irq_num,
ci->valid_irq_cnt);
enable_irq(ci->irq_num);
return HRTIMER_NORESTART;
}
void amlpmu_relax_timer_start(int other_cpu)
{
struct amlpmu_cpuinfo *ci;
int cpu;
struct amlpmu_context *ctx = &amlpmu_ctx;
cpu = smp_processor_id();
WARN_ON(cpu != 0);
ci = per_cpu_ptr(ctx->cpuinfo, 0);
pr_warn("wait cpu %d fixup done timeout, main cpu irq cnt = %lu\n",
other_cpu,
ci->valid_irq_cnt);
if (hrtimer_active(&ctx->relax_timer)) {
pr_alert("relax_timer already active, return!\n");
return;
}
disable_irq_nosync(ci->irq_num);
hrtimer_start(&ctx->relax_timer,
ns_to_ktime(ctx->relax_timer_ns),
HRTIMER_MODE_REL);
}
static void amlpmu_fix_setup_affinity(int irq)
{
int cluster_index = 0;
int cpu;
int affinity_cpu = -1;
struct amlpmu_cpuinfo *ci = NULL;
struct amlpmu_context *ctx = &amlpmu_ctx;
s64 latest_next_stamp = S64_MAX;
if (irq == ctx->irqs[0])
cluster_index = 0;
else if (ctx->clusterb_enabled && irq ==
ctx->irqs[1])
cluster_index = 1;
else {
pr_err("amlpmu_fix_setup_affinity() bad irq = %d\n", irq);
return;
}
/*
* find latest next_predicted_stamp cpu for affinity cpu
* if no cpu have predict time, select first cpu of cpumask
* todo:
* - if a cpu predict failed for continuous N times,
* try add some punishment.
* - if no cpu have predicted time, try recently most used cpu
* for affinity
* - try to keep and promote prediction accuracy
*/
for_each_cpu_and(cpu,
&ctx->cpumasks[cluster_index],
cpu_possible_mask) {
ci = per_cpu_ptr(ctx->cpuinfo, cpu);
//pr_info("cpu = %d, ci->next_predicted_stamp.tv64 = %lld\n",
// cpu, ci->next_predicted_stamp.tv64);
if (ci->next_predicted_stamp.tv64 &&
ci->next_predicted_stamp.tv64 < latest_next_stamp) {
latest_next_stamp = ci->next_predicted_stamp.tv64;
affinity_cpu = cpu;
}
}
if (affinity_cpu == -1) {
affinity_cpu = cpumask_first(&ctx->cpumasks[cluster_index]);
pr_debug("used first cpu: %d, cluster: 0x%lx\n",
affinity_cpu,
*cpumask_bits(&ctx->cpumasks[cluster_index]));
} else
pr_debug("find affinity cpu: %d, next_predicted_stamp: %lld\n",
affinity_cpu,
latest_next_stamp);
if (irq_set_affinity(irq, cpumask_of(affinity_cpu)))
pr_err("irq_set_affinity() failed irq: %d, affinity_cpu: %d\n",
irq,
affinity_cpu);
}
/*
* on pmu interrupt generated cpu, @irq_num is valid
* on other cpus(called by AML_PMU_IPI), @irq_num is -1
*/
static int amlpmu_irq_fix(int irq_num)
{
int i;
int cpu;
int cur_cpu;
int pmuirq_val;
int cluster_index = 0;
int fix_success = 0;
struct amlpmu_cpuinfo *ci;
struct amlpmu_context *ctx = &amlpmu_ctx;
struct call_single_data csd_stack;
int max_wait_cnt = ctx->max_wait_cnt;
csd_stack.func = amlpmu_handle_irq_ipi;
csd_stack.info = NULL;
cur_cpu = smp_processor_id();
if (irq_num == ctx->irqs[0])
cluster_index = 0;
else if (ctx->clusterb_enabled && irq_num == ctx->irqs[1])
cluster_index = 1;
else {
pr_err("amlpmu_irq_fix() bad irq = %d\n", irq_num);
return fix_success;
}
if (!cpumask_test_cpu(cur_cpu, &ctx->cpumasks[cluster_index])) {
pr_warn("amlpmu_irq_fix() cur_cpu %d not in cluster: 0x%lx\n",
cur_cpu,
*cpumask_bits(&ctx->cpumasks[cluster_index]));
}
pmuirq_val = readl(ctx->regs[cluster_index]);
pmuirq_val &= 0xf;
pmuirq_val <<= ctx->first_cpus[cluster_index];
pr_debug("amlpmu_irq_fix() val=0x%0x, first_cpu=%d, cluster=0x%lx\n",
readl(ctx->regs[cluster_index]),
ctx->first_cpus[cluster_index],
*cpumask_bits(&ctx->cpumasks[cluster_index]));
for_each_cpu_and(cpu,
&ctx->cpumasks[cluster_index],
cpu_possible_mask) {
if (pmuirq_val & (1<<cpu)) {
if (cpu == cur_cpu) {
pr_info("ownercpu %d in pmuirq = 0x%x\n",
cur_cpu, pmuirq_val);
continue;
}
pr_debug("fix pmu irq cpu %d, pmuirq = 0x%x\n",
cpu,
pmuirq_val);
ci = per_cpu_ptr(ctx->cpuinfo, cpu);
WRITE_ONCE(ci->fix_done, 0);
WRITE_ONCE(ci->fix_overflowed, 0);
csd_stack.flags = 0;
smp_call_function_single_async(cpu, &csd_stack);
for (i = 0; i < max_wait_cnt; i++) {
if (READ_ONCE(ci->fix_done))
break;
udelay(1);
}
if (i == ctx->max_wait_cnt) {
pr_err("wait for cpu %d done timeout\n",
cpu);
//amlpmu_relax_timer_start(cpu);
}
if (READ_ONCE(ci->fix_overflowed))
fix_success++;
}
}
return fix_success;
}
static void amlpmu_update_stats(int irq_num,
int has_overflowed,
int fix_success)
{
int freq;
int i;
ktime_t stamp;
unsigned long time = jiffies;
struct amlpmu_cpuinfo *ci;
struct amlpmu_context *ctx = &amlpmu_ctx;
ci = this_cpu_ptr(ctx->cpuinfo);
if (!has_overflowed && !fix_success) {
pr_debug("empty_irq_cnt: %lu\n", ci->empty_irq_cnt);
ci->empty_irq_cnt++;
ci->empty_irq_time = time;
}
if (fix_success) {
/* send IPI success */
pr_debug("fix_irq_cnt: %lu, fix_success = %d\n",
ci->fix_irq_cnt,
fix_success);
ci->fix_irq_cnt++;
ci->fix_irq_time = time;
}
if (has_overflowed) {
ci->valid_irq_cnt++;
ci->valid_irq_time = time;
stamp = ktime_get();
ci->stamp_deltas[ci->valid_irq_cnt % MAX_DELTA_CNT] =
stamp.tv64 - ci->last_stamp.tv64;
ci->last_stamp = stamp;
/* update avg_delta if it's valid */
ci->avg_delta = 0;
for (i = 0; i < MAX_DELTA_CNT; i++)
ci->avg_delta += ci->stamp_deltas[i];
ci->avg_delta /= MAX_DELTA_CNT;
for (i = 0; i < MAX_DELTA_CNT; i++) {
if (ci->stamp_deltas[i] > ci->avg_delta * 3/2 ||
ci->stamp_deltas[i] < ci->avg_delta / 2) {
ci->avg_delta = 0;
break;
}
}
if (ci->avg_delta)
ci->next_predicted_stamp.tv64 =
ci->last_stamp.tv64 + ci->avg_delta;
else
ci->next_predicted_stamp.tv64 = 0;
pr_debug("irq_num = %d, valid_irq_cnt = %lu\n",
irq_num,
ci->valid_irq_cnt);
pr_debug("cur_delta = %lld, avg_delta = %lld, next = %lld\n",
ci->stamp_deltas[ci->valid_irq_cnt % MAX_DELTA_CNT],
ci->avg_delta,
ci->next_predicted_stamp.tv64);
}
if (time_after(ci->valid_irq_time, ci->last_valid_irq_time + 2*HZ)) {
freq = ci->empty_irq_cnt - ci->last_empty_irq_cnt;
freq *= HZ;
freq /= (ci->empty_irq_time - ci->last_empty_irq_time);
pr_info("######## empty_irq_cnt: %lu - %lu = %lu, freq = %d\n",
ci->empty_irq_cnt,
ci->last_empty_irq_cnt,
ci->empty_irq_cnt - ci->last_empty_irq_cnt,
freq);
ci->last_empty_irq_cnt = ci->empty_irq_cnt;
ci->last_empty_irq_time = ci->empty_irq_time;
freq = ci->fix_irq_cnt - ci->last_fix_irq_cnt;
freq *= HZ;
freq /= (ci->fix_irq_time - ci->last_fix_irq_time);
pr_info("######## fix_irq_cnt: %lu - %lu = %lu, freq = %d\n",
ci->fix_irq_cnt,
ci->last_fix_irq_cnt,
ci->fix_irq_cnt - ci->last_fix_irq_cnt,
freq);
ci->last_fix_irq_cnt = ci->fix_irq_cnt;
ci->last_fix_irq_time = ci->fix_irq_time;
freq = ci->valid_irq_cnt - ci->last_valid_irq_cnt;
freq *= HZ;
freq /= (ci->valid_irq_time - ci->last_valid_irq_time);
pr_info("######## valid_irq_cnt: %lu - %lu = %lu, freq = %d\n",
ci->valid_irq_cnt,
ci->last_valid_irq_cnt,
ci->valid_irq_cnt - ci->last_valid_irq_cnt,
freq);
ci->last_valid_irq_cnt = ci->valid_irq_cnt;
ci->last_valid_irq_time = ci->valid_irq_time;
}
}
int amlpmu_handle_irq(struct arm_pmu *cpu_pmu, int irq_num, int has_overflowed)
{
int cpu;
int fix_success = 0;
struct amlpmu_cpuinfo *ci;
struct amlpmu_context *ctx = &amlpmu_ctx;
ci = this_cpu_ptr(ctx->cpuinfo);
ci->irq_num = irq_num;
cpu = smp_processor_id();
pr_debug("amlpmu_handle_irq() irq_num = %d, overflowed = %d\n",
irq_num, has_overflowed);
/*
* if current cpu is not overflowed, it's possible some other
* cpus caused the pmu interrupt.
* so if current cpu is interrupt generated cpu(irq_num != -1),
* call aml_pmu_fix() try to send IPI to other cpus and waiting
* for fix_done.
*/
if (!has_overflowed && irq_num != -1)
fix_success = amlpmu_irq_fix(irq_num);
/*
* valid_irq, fix_irq and empty_irq status
* avg_delta time account to predict next interrupt time
*/
amlpmu_update_stats(irq_num, has_overflowed, fix_success);
/*
* armv*pmu_getreset_flags() will clear interrupt. If current
* interrupt is IPI fix(irq_num = -1), interrupt generated cpu
* now is waiting for ci->fix_done=1(clear interrupt).
* we must set ci->fix_done to 1 after amlpmu_stat_account(),
* because interrupt generated cpu need this predict time info
* to setup interrupt affinity.
*/
if (irq_num == -1) {
WRITE_ONCE(ci->fix_overflowed, has_overflowed);
/* fix_overflowed must before fix_done */
mb();
WRITE_ONCE(ci->fix_done, 1);
}
/* only interrupt generated cpu need setup affinity */
if (irq_num != -1)
amlpmu_fix_setup_affinity(irq_num);
/*
* when a pmu interrupt generated, if current cpu is not
* overflowed and some other cpus succeed in handling the
* interrupt by IPIs return true.
*/
return !has_overflowed && fix_success;
}
static int amlpmu_init(struct platform_device *pdev, struct arm_pmu *pmu)
{
int cpu;
int ret = 0;
int irq;
u32 cpumasks[MAX_CLUSTER_NR] = {0};
struct amlpmu_context *ctx = &amlpmu_ctx;
struct amlpmu_cpuinfo *ci;
memset(ctx, 0, sizeof(*ctx));
ctx->cpuinfo = __alloc_percpu_gfp(
sizeof(struct amlpmu_cpuinfo),
SMP_CACHE_BYTES,
GFP_KERNEL | __GFP_ZERO);
if (!ctx->cpuinfo) {
pr_err("alloc percpu failed\n");
ret = -ENOMEM;
goto free;
}
for_each_possible_cpu(cpu) {
ci = per_cpu_ptr(ctx->cpuinfo, cpu);
ci->last_valid_irq_time = INITIAL_JIFFIES;
ci->last_fix_irq_time = INITIAL_JIFFIES;
ci->last_empty_irq_time = INITIAL_JIFFIES;
}
ctx->pmu = pmu;
if (of_property_read_bool(pdev->dev.of_node, "clusterb-enabled"))
ctx->clusterb_enabled = 1;
pr_info("clusterb_enabled = %d\n", ctx->clusterb_enabled);
ret = of_property_read_u32_array(pdev->dev.of_node,
"cpumasks",
cpumasks,
ctx->clusterb_enabled ? MAX_CLUSTER_NR : 1);
if (ret) {
pr_err("read prop cpumasks failed, ret = %d\n", ret);
ret = -EINVAL;
goto free;
}
pr_info("cpumasks 0x%0x, 0x%0x\n", cpumasks[0], cpumasks[1]);
ret = of_property_read_u32(pdev->dev.of_node,
"relax-timer-ns",
&ctx->relax_timer_ns);
if (ret) {
pr_err("read prop relax-timer-ns failed, ret = %d\n", ret);
ret = -EINVAL;
goto free;
}
ret = of_property_read_u32(pdev->dev.of_node,
"max-wait-cnt",
&ctx->max_wait_cnt);
if (ret) {
pr_err("read prop max-wait-cnt failed, ret = %d\n", ret);
ret = -EINVAL;
goto free;
}
irq = platform_get_irq(pdev, 0);
if (irq < 0) {
pr_err("get clusterA irq failed, %d\n", irq);
ret = -EINVAL;
goto free;
}
ctx->irqs[0] = irq;
pr_info("cluster A irq = %d\n", irq);
ctx->regs[0] = of_iomap(pdev->dev.of_node, 0);
if (IS_ERR(ctx->regs[0])) {
pr_err("of_iomap() clusterA failed, base = %p\n", ctx->regs[0]);
ret = PTR_ERR(ctx->regs[0]);
goto free;
}
cpumask_clear(&ctx->cpumasks[0]);
memcpy(cpumask_bits(&ctx->cpumasks[0]),
&cpumasks[0],
sizeof(cpumasks[0]));
if (!cpumask_intersects(&ctx->cpumasks[0], cpu_possible_mask)) {
pr_err("bad cpumasks[0] 0x%x\n", cpumasks[0]);
ret = -EINVAL;
goto free;
}
ctx->first_cpus[0] = cpumask_first(&ctx->cpumasks[0]);
amlpmu_fix_setup_affinity(ctx->irqs[0]);
hrtimer_init(&ctx->relax_timer,
CLOCK_MONOTONIC,
HRTIMER_MODE_REL);
ctx->relax_timer.function = amlpmu_relax_timer_func;
if (!ctx->clusterb_enabled)
return 0;
irq = platform_get_irq(pdev, 1);
if (irq < 0) {
pr_err("get clusterB irq failed, %d\n", irq);
ret = -EINVAL;
goto free;
}
ctx->irqs[1] = irq;
pr_info("cluster B irq = %d\n", irq);
ctx->regs[1] = of_iomap(pdev->dev.of_node, 1);
if (IS_ERR(ctx->regs[1])) {
pr_err("of_iomap() clusterA failed, base = %p\n", ctx->regs[1]);
ret = PTR_ERR(ctx->regs[1]);
goto free;
}
cpumask_clear(&ctx->cpumasks[1]);
memcpy(cpumask_bits(&ctx->cpumasks[1]),
&cpumasks[1],
sizeof(cpumasks[1]));
if (!cpumask_intersects(&ctx->cpumasks[1], cpu_possible_mask)) {
pr_err("bad cpumasks[1] 0x%x\n", cpumasks[1]);
ret = -EINVAL;
goto free;
} else if (cpumask_intersects(&ctx->cpumasks[0], &ctx->cpumasks[1])) {
pr_err("cpumasks intersect 0x%x : 0x%x\n",
cpumasks[0],
cpumasks[1]);
ret = -EINVAL;
goto free;
}
ctx->first_cpus[1] = cpumask_first(&ctx->cpumasks[1]);
amlpmu_fix_setup_affinity(ctx->irqs[1]);
return 0;
free:
if (ctx->cpuinfo)
free_percpu(ctx->cpuinfo);
if (ctx->regs[0])
iounmap(ctx->regs[0]);
if (ctx->regs[1])
iounmap(ctx->regs[1]);
return ret;
}
#endif
int arm_pmu_device_probe(struct platform_device *pdev,
const struct of_device_id *of_table,
const struct pmu_probe_info *probe_table)
@@ -1029,6 +1540,13 @@ int arm_pmu_device_probe(struct platform_device *pdev,
return -ENOMEM;
}
#ifdef CONFIG_AMLOGIC_MODIFY
if (amlpmu_init(pdev, pmu)) {
pr_err("amlpmu_init() failed\n");
return 1;
}
#endif
armpmu_init(pmu);
pmu->plat_device = pdev;

View File

@@ -162,6 +162,99 @@ int arm_pmu_device_probe(struct platform_device *pdev,
#define ARMV8_PMU_PDEV_NAME "armv8-pmu"
#ifdef CONFIG_AMLOGIC_MODIFY
#define MAX_DELTA_CNT 4
struct amlpmu_cpuinfo {
int irq_num;
/*
* In interrupt generated cpu(affinity cpu)
* If pmu no overflowed, then we need to send IPI to some other cpus to
* fix it. And before send IPI, set corresponding cpu's fix_done and
* fix_overflowed to zero, in corresponding cpu's IPI interrupt will set
* fix_done to inform source cpu and if indeed pmu overflowed then also
* set fix_overflowed to 1, then inerrupt generated cpu can feel that.
*/
int fix_done;
int fix_overflowed;
/* for interrupt affinity prediction */
ktime_t last_stamp;
s64 stamp_deltas[MAX_DELTA_CNT];
s64 avg_delta;
ktime_t next_predicted_stamp;
/*
* irq state account of this cpu
*
* - valid_irq_cnt:
* valid irq cnt.(pmu overflow happened)
* - fix_irq_cnt:
* when this cpu is pmu interrupt generated affinity cpu, a pmu
* interrupt if cpu affinity predict failed so no pmu overflow
* happened and succeeded send IPI to other cpu, then it's a send
* fix irq. So the lower is better.
* - empty_irq_cnt:
* when this cpu is pmu interrupt generated affinity cpu, a pmu
* interrupt that no overflow happened and also no fix IPI sended to
* other cpus, then it's a empty irq.
* when this cpu is not affinity cpu, a IPI interrupt(pmu fix from
* affinity cpu) that no pmu overflow happened, it's a empty irq.
*
* attention:
* A interrupt can be a valid_irq and also a fix_irq.
*/
unsigned long valid_irq_cnt;
unsigned long fix_irq_cnt;
unsigned long empty_irq_cnt;
unsigned long valid_irq_time;
unsigned long fix_irq_time;
unsigned long empty_irq_time;
unsigned long last_valid_irq_cnt;
unsigned long last_fix_irq_cnt;
unsigned long last_empty_irq_cnt;
unsigned long last_valid_irq_time;
unsigned long last_fix_irq_time;
unsigned long last_empty_irq_time;
};
#define MAX_CLUSTER_NR 2
struct amlpmu_context {
struct amlpmu_cpuinfo __percpu *cpuinfo;
/* struct arm_pmu */
struct arm_pmu *pmu;
int clusterb_enabled;
unsigned int __iomem *regs[MAX_CLUSTER_NR];
int irqs[MAX_CLUSTER_NR];
struct cpumask cpumasks[MAX_CLUSTER_NR];
int first_cpus[MAX_CLUSTER_NR];
/*
* In main pmu irq route wait for other cpu fix done may cause lockup,
* when lockup we disable main irq for a while.
* relax_timer will enable main irq again.
*/
struct hrtimer relax_timer;
unsigned int relax_timer_ns;
unsigned int max_wait_cnt;
};
extern struct amlpmu_context amlpmu_ctx;
int amlpmu_handle_irq(struct arm_pmu *cpu_pmu, int irq_num, int has_overflowed);
/* defined int arch/arm(64)/kernel/perf_event(_v7).c */
void amlpmu_handle_irq_ipi(void *arg);
#endif
#endif /* CONFIG_ARM_PMU */
#endif /* __ARM_PMU_H__ */

View File

@@ -73,13 +73,6 @@ extern void smp_send_stop(void);
*/
extern void smp_send_reschedule(int cpu);
#ifdef CONFIG_AMLOGIC_MODIFY
/*
* sends a 'aml pmu' event to another CPU:
*/
extern void smp_send_aml_pmu(int cpu);
#endif
/*
* Prepare machine for booting other CPUs.
*/