C语言 使用wrmsrl和rdmsrl读取PMU计数器

Question

我正在尝试在Intel Xeon gold (Skylake generation)处理器上的Linux内核模块中读取LLC缓存缺失硬件计数器。下面代码的结果总是0:

#define PMC_ESEL_UMASK_SHIFT 8
#define PMC_ESEL_CMASK_SHIFT 24
#define PMC_ESEL_ENTRY(event, umask, cmask)       
(((event)&0xFFUL) |                           
(((umask)&0xFFUL) << PMC_ESEL_UMASK_SHIFT) | 
(((cmask)&0xFFUL) << PMC_ESEL_CMASK_SHIFT))
#define PMC_ESEL_USR (1ULL << 16)    /* User Mode */
#define PMC_ESEL_OS (1ULL << 17)     /* Kernel Mode */
#define PMC_ESEL_EDGE (1ULL << 18)   /* Edge detect */
#define PMC_ESEL_PC (1ULL << 19)     /* Pin control */
#define PMC_ESEL_INT (1ULL << 20)    /* APIC interrupt enable */
#define PMC_ESEL_ANY (1ULL << 21)    /* Any thread */
#define PMC_ESEL_ENABLE (1ULL << 22) /* Enable counters */
#define PMC_ESEL_INV (1ULL << 23)    /* Invert counter mask */
/* architectural performance counters (works on all Intel CPUs) */
#define PMC_ARCH_CORE_CYCLES PMC_ESEL_ENTRY(0x3C, 0x00, 0)
#define PMC_ARCH_INSTR_RETIRED PMC_ESEL_ENTRY(0xC0, 0x00, 0)
#define PMC_ARCH_REF_CYCLES PMC_ESEL_ENTRY(0x3C, 0x01, 0)
#define PMC_ARCH_LLC_REF PMC_ESEL_ENTRY(0x2E, 0x4F, 0)
#define PMC_ARCH_LLC_MISSES PMC_ESEL_ENTRY(0x2E, 0x41, 0)
#define PMC_ARCH_BRANCHES PMC_ESEL_ENTRY(0xC4, 0x00, 0)
#define PMC_ARCH_BRANCH_MISSES PMC_ESEL_ENTRY(0xC5, 0x00, 0)
#define PMC_LLC_MISSES (PMC_ARCH_LLC_MISSES | PMC_ESEL_USR | PMC_ESEL_OS | 
PMC_ESEL_ENABLE)
#define PMC_LLC_MISSES_ANY (PMC_ARCH_LLC_MISSES | PMC_ESEL_USR | PMC_ESEL_OS | 
PMC_ESEL_ANY | PMC_ESEL_ENABLE)
/* enable/disable perf counter */
#define CORE_PERF_GLOBAL_CTRL_ENABLE_PMC_0 (0x1)
#define CORE_PERF_GLOBAL_CTRL_ENABLE_PMC_1 (0x2)
void test(void)
{
u64 val;
wrmsrl(MSR_CORE_PERF_FIXED_CTR_CTRL, 0x333);
wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL,
CORE_PERF_GLOBAL_CTRL_ENABLE_PMC_0 |
CORE_PERF_GLOBAL_CTRL_ENABLE_PMC_1 |
(1UL << 32) | (1UL << 33) | (1UL << 34));
wrmsrl(MSR_P6_EVNTSEL0, PMC_LLC_MISSES);
rdmsrl(MSR_P6_PERFCTR0, val);
printk(KERN_DEBUG "The LLC-Miss value is: %llun", val);
}

有没有人可以帮助解决这个代码的问题?

Answer 1

似乎我错过了使用rdmsrl_on_core和wrmsrl_on_cpu来启用和读取每个核心上的PMC。我附加了一个内核模块代码的工作版本，以防将来有人需要它:

#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/init.h>
MODULE_LICENSE("GPL");
MODULE_AUTHOR("Mohammad Siavashi");
MODULE_DESCRIPTION("A sample LLC miss counter reader");
#define IA32_PERF_GLOBAL_CTRL_ENABLE 0x70000000f
#define LLC_EVENT 0x0043412e
#define IA32_PERF_FIXED_CTRL_ENABLE 0x333
static int __init hello_init(void)
{
int cpu;
wrmsrl(MSR_CORE_PERF_FIXED_CTR_CTRL, IA32_PERF_FIXED_CTRL_ENABLE);
wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL, IA32_PERF_GLOBAL_CTRL_ENABLE);
for_each_possible_cpu(cpu)
{
wrmsrl_on_cpu(cpu, MSR_P6_EVNTSEL0, LLC_EVENT);
}
return 0;
}
static void __exit hello_cleanup(void)
{
int cpu;
u64 val;
for_each_possible_cpu(cpu)
{
rdmsrl_on_cpu(cpu, MSR_IA32_PERFCTR0, &val);
printk(KERN_DEBUG "The LLC-Miss on core %i is: %llun", cpu, val);
}
printk(KERN_INFO "Cleaning up module.n");
}
module_init(hello_init);
module_exit(hello_cleanup);