samples/bpf/offwaketime_kern.c

   1 /* Copyright (c) 2016 Facebook
   2  *
   3  * This program is free software; you can redistribute it and/or
   4  * modify it under the terms of version 2 of the GNU General Public
   5  * License as published by the Free Software Foundation.
   6  */
   7 #include <uapi/linux/bpf.h>
   8 #include "bpf_helpers.h"
   9 #include <uapi/linux/ptrace.h>
  10 #include <uapi/linux/perf_event.h>
  11 #include <linux/version.h>
  12 #include <linux/sched.h>
  13
  14 #define _(P) ({typeof(P) val; bpf_probe_read(&val, sizeof(val), &P); val;})
  15
  16 #define MINBLOCK_US     1
  17
  18 struct key_t {
  19         char waker[TASK_COMM_LEN];
  20         char target[TASK_COMM_LEN];
  21         u32 wret;
  22         u32 tret;
  23 };
  24
  25 struct bpf_map_def SEC("maps") counts = {
  26         .type = BPF_MAP_TYPE_HASH,
  27         .key_size = sizeof(struct key_t),
  28         .value_size = sizeof(u64),
  29         .max_entries = 10000,
  30 };
  31
  32 struct bpf_map_def SEC("maps") start = {
  33         .type = BPF_MAP_TYPE_HASH,
  34         .key_size = sizeof(u32),
  35         .value_size = sizeof(u64),
  36         .max_entries = 10000,
  37 };
  38
  39 struct wokeby_t {
  40         char name[TASK_COMM_LEN];
  41         u32 ret;
  42 };
  43
  44 struct bpf_map_def SEC("maps") wokeby = {
  45         .type = BPF_MAP_TYPE_HASH,
  46         .key_size = sizeof(u32),
  47         .value_size = sizeof(struct wokeby_t),
  48         .max_entries = 10000,
  49 };
  50
  51 struct bpf_map_def SEC("maps") stackmap = {
  52         .type = BPF_MAP_TYPE_STACK_TRACE,
  53         .key_size = sizeof(u32),
  54         .value_size = PERF_MAX_STACK_DEPTH * sizeof(u64),
  55         .max_entries = 10000,
  56 };
  57
  58 #define STACKID_FLAGS (0 | BPF_F_FAST_STACK_CMP)
  59
  60 SEC("kprobe/try_to_wake_up")
  61 int waker(struct pt_regs *ctx)
  62 {
  63         struct task_struct *p = (void *) PT_REGS_PARM1(ctx);
  64         struct wokeby_t woke;
  65         u32 pid;
  66
  67         pid = _(p->pid);
  68
  69         bpf_get_current_comm(&woke.name, sizeof(woke.name));
  70         woke.ret = bpf_get_stackid(ctx, &stackmap, STACKID_FLAGS);
  71
  72         bpf_map_update_elem(&wokeby, &pid, &woke, BPF_ANY);
  73         return 0;
  74 }
  75
  76 static inline int update_counts(void *ctx, u32 pid, u64 delta)
  77 {
  78         struct wokeby_t *woke;
  79         u64 zero = 0, *val;
  80         struct key_t key;
  81
  82         __builtin_memset(&key.waker, 0, sizeof(key.waker));
  83         bpf_get_current_comm(&key.target, sizeof(key.target));
  84         key.tret = bpf_get_stackid(ctx, &stackmap, STACKID_FLAGS);
  85         key.wret = 0;
  86
  87         woke = bpf_map_lookup_elem(&wokeby, &pid);
  88         if (woke) {
  89                 key.wret = woke->ret;
  90                 __builtin_memcpy(&key.waker, woke->name, sizeof(key.waker));
  91                 bpf_map_delete_elem(&wokeby, &pid);
  92         }
  93
  94         val = bpf_map_lookup_elem(&counts, &key);
  95         if (!val) {
  96                 bpf_map_update_elem(&counts, &key, &zero, BPF_NOEXIST);
  97                 val = bpf_map_lookup_elem(&counts, &key);
  98                 if (!val)
  99                         return 0;
 100         }
 101         (*val) += delta;
 102         return 0;
 103 }
 104
 105 #if 1
 106 /* taken from /sys/kernel/debug/tracing/events/sched/sched_switch/format */
 107 struct sched_switch_args {
 108         unsigned long long pad;
 109         char prev_comm[16];
 110         int prev_pid;
 111         int prev_prio;
 112         long long prev_state;
 113         char next_comm[16];
 114         int next_pid;
 115         int next_prio;
 116 };
 117 SEC("tracepoint/sched/sched_switch")
 118 int oncpu(struct sched_switch_args *ctx)
 119 {
 120         /* record previous thread sleep time */
 121         u32 pid = ctx->prev_pid;
 122 #else
 123 SEC("kprobe/finish_task_switch")
 124 int oncpu(struct pt_regs *ctx)
 125 {
 126         struct task_struct *p = (void *) PT_REGS_PARM1(ctx);
 127         /* record previous thread sleep time */
 128         u32 pid = _(p->pid);
 129 #endif
 130         u64 delta, ts, *tsp;
 131
 132         ts = bpf_ktime_get_ns();
 133         bpf_map_update_elem(&start, &pid, &ts, BPF_ANY);
 134
 135         /* calculate current thread's delta time */
 136         pid = bpf_get_current_pid_tgid();
 137         tsp = bpf_map_lookup_elem(&start, &pid);
 138         if (!tsp)
 139                 /* missed start or filtered */
 140                 return 0;
 141
 142         delta = bpf_ktime_get_ns() - *tsp;
 143         bpf_map_delete_elem(&start, &pid);
 144         delta = delta / 1000;
 145         if (delta < MINBLOCK_US)
 146                 return 0;
 147
 148         return update_counts(ctx, pid, delta);
 149 }
 150 char _license[] SEC("license") = "GPL";
 151 u32 _version SEC("version") = LINUX_VERSION_CODE;