使用kretprobes来查找耗时长的中断处理函数

在实时系统中，如果top half时间过长，会导致实时任务得不到调度，产生比较高的延时。使用kretprobes可以查看中断处理的时间，代码如下（参考kretprobe_sample.c做了一些改动）：

/*
 * kretprobe_example.c
 *
 * Here's a sample kernel module showing the use of return probes to
 * report the return value and total time taken for probed function
 * to run.
 *
 * usage: insmod kretprobe_example.ko func=<func_name>
 *
 * If no func_name is specified, _do_fork is instrumented
 *
 * For more information on theory of operation of kretprobes, see
 * Documentation/kprobes.txt
 *
 * Build and insert the kernel module as done in the kprobe example.
 * You will see the trace data in /var/log/messages and on the console
 * whenever the probed function returns. (Some messages may be suppressed
 * if syslogd is configured to eliminate duplicate messages.)
 */

#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/kprobes.h>
#include <linux/ktime.h>
#include <linux/limits.h>
#include <linux/sched.h>
#include <linux/irq.h>
#include <linux/interrupt.h>

#define PT_REGS_PARM1(x) ((x)->regs[0])
static char func_name[NAME_MAX] = "__handle_irq_event_percpu";
module_param_string(func, func_name, NAME_MAX, S_IRUGO);
MODULE_PARM_DESC(func, "Function to kretprobe; this module will report the"
            " function's execution time");

/* per-instance private data */
struct my_data {
    ktime_t entry_stamp;
    char func_name[NAME_MAX];
};

/* Here we use the entry_hanlder to timestamp function entry */
static int entry_handler(struct kretprobe_instance *ri, struct pt_regs *regs)
{
    struct my_data *data;
    struct irq_desc *desc;
    char *str;

    if (!current->mm)
        return 1;    /* Skip kernel threads */

    data = (struct my_data *)ri->data;
    data->entry_stamp = ktime_get();

    str = data->func_name;
    desc = (struct irq_desc *)PT_REGS_PARM1(regs);

    str[0] = 0;
    if (desc->action)
        sprint_symbol_no_offset(str, (unsigned long)desc->action->handler);

    return 0;
}

/*
 * Return-probe handler: Log the return value and duration. Duration may turn
 * out to be zero consistently, depending upon the granularity of time
 * accounting on the platform.
 */
static int ret_handler(struct kretprobe_instance *ri, struct pt_regs *regs)
{
    unsigned long retval = regs_return_value(regs);
    struct my_data *data = (struct my_data *)ri->data;
    s64 delta;
    ktime_t now;

    now = ktime_get();
    delta = ktime_to_ns(ktime_sub(now, data->entry_stamp));
    pr_info("%s returned %lu and took %lld ns to execute\n",
            data->func_name, retval, (long long)delta);
    return 0;
}

static struct kretprobe my_kretprobe = {
    .handler        = ret_handler,
    .entry_handler        = entry_handler,
    .data_size        = sizeof(struct my_data),
    /* Probe up to 20 instances concurrently. */
    .maxactive        = 20,
};

static int __init kretprobe_init(void)
{
    int ret;

    my_kretprobe.kp.symbol_name = func_name;
    ret = register_kretprobe(&my_kretprobe);
    if (ret < 0) {
        pr_err("register_kretprobe failed, returned %d\n", ret);
        return -1;
    }
    pr_info("Planted return probe at %s: %p\n",
            my_kretprobe.kp.symbol_name, my_kretprobe.kp.addr);
    return 0;
}

static void __exit kretprobe_exit(void)
{
    unregister_kretprobe(&my_kretprobe);
    pr_info("kretprobe at %p unregistered\n", my_kretprobe.kp.addr);

    /* nmissed > 0 suggests that maxactive was set too low. */
    pr_info("Missed probing %d instances of %s\n",
        my_kretprobe.nmissed, my_kretprobe.kp.symbol_name);
}

module_init(kretprobe_init)
module_exit(kretprobe_exit)
MODULE_LICENSE("GPL");

编译后，运行：

insmod kretprobe.ko; cyclictest xxx;

得到结果如下：

 

[  272.418346] xhci_msi_irq returned 1 and took 146602 ns to execute
[  272.426034] ahci_single_level_irq_intr returned 1 and took 29161 ns to execute
[  272.426332] xhci_msi_irq returned 1 and took 135103 ns to execute
[  272.434361] xhci_msi_irq returned 1 and took 155142 ns to execute
[  272.442334] xhci_msi_irq returned 1 and took 138082 ns to execute
[  272.448333] xhci_msi_irq returned 1 and took 136542 ns to execute
[  272.456333] xhci_msi_irq returned 1 and took 134342 ns to execute

......

[  274.324755] ahci_single_level_irq_intr returned 1 and took 31560 ns to execute
[  274.330631] e1000_intr_msix_rx [e1000e] returned 1 and took 52401 ns to execute
[  274.347062] ahci_single_level_irq_intr returned 1 and took 29940 ns to execute
[  274.763850] ahci_single_level_irq_intr returned 1 and took 32041 ns to execute
[  274.796840] ahci_single_level_irq_intr returned 1 and took 29040 ns to execute
[  275.031541] ahci_single_level_irq_intr returned 1 and took 30261 ns to execute
[  275.040657] e1000_intr_msix_rx [e1000e] returned 1 and took 47541 ns to execute
[  275.232745] e1000_intr_msix_rx [e1000e] returned 1 and took 53941 ns to execute
[  275.233593] e1000_intr_msix_rx [e1000e] returned 1 and took 48321 ns to execute
[  275.377567] ahci_single_level_irq_intr returned 1 and took 30160 ns to execute
[  275.399842] e1000_intr_msix_rx [e1000e] returned 1 and took 50941 ns to execute

 

如果是preempt rt的实时系统，该代码只能得到irq_default_primary_handler，因为在preempt rt的方案中，所有的中断处理函数都变成了threaded irq，irq_default_primary_handler被设置为默认的顶半部。