samples/bpf/offwaketime_kern.c

0001 /* Copyright (c) 2016 Facebook
0002  *
0003  * This program is free software; you can redistribute it and/or
0004  * modify it under the terms of version 2 of the GNU General Public
0005  * License as published by the Free Software Foundation.
0006  */
0007 #include <uapi/linux/bpf.h>
0008 #include <uapi/linux/ptrace.h>
0009 #include <uapi/linux/perf_event.h>
0010 #include <linux/version.h>
0011 #include <linux/sched.h>
0012 #include <bpf/bpf_helpers.h>
0013 #include <bpf/bpf_tracing.h>
0014
0015 #define _(P)                                                                   \
0016     ({                                                                     \
0017         typeof(P) val;                                                 \
0018         bpf_probe_read_kernel(&val, sizeof(val), &(P));                \
0019         val;                                                           \
0020     })
0021
0022 #define MINBLOCK_US 1
0023 #define MAX_ENTRIES 10000
0024
0025 struct key_t {
0026     char waker[TASK_COMM_LEN];
0027     char target[TASK_COMM_LEN];
0028     u32 wret;
0029     u32 tret;
0030 };
0031
0032 struct {
0033     __uint(type, BPF_MAP_TYPE_HASH);
0034     __type(key, struct key_t);
0035     __type(value, u64);
0036     __uint(max_entries, MAX_ENTRIES);
0037 } counts SEC(".maps");
0038
0039 struct {
0040     __uint(type, BPF_MAP_TYPE_HASH);
0041     __type(key, u32);
0042     __type(value, u64);
0043     __uint(max_entries, MAX_ENTRIES);
0044 } start SEC(".maps");
0045
0046 struct wokeby_t {
0047     char name[TASK_COMM_LEN];
0048     u32 ret;
0049 };
0050
0051 struct {
0052     __uint(type, BPF_MAP_TYPE_HASH);
0053     __type(key, u32);
0054     __type(value, struct wokeby_t);
0055     __uint(max_entries, MAX_ENTRIES);
0056 } wokeby SEC(".maps");
0057
0058 struct {
0059     __uint(type, BPF_MAP_TYPE_STACK_TRACE);
0060     __uint(key_size, sizeof(u32));
0061     __uint(value_size, PERF_MAX_STACK_DEPTH * sizeof(u64));
0062     __uint(max_entries, MAX_ENTRIES);
0063 } stackmap SEC(".maps");
0064
0065 #define STACKID_FLAGS (0 | BPF_F_FAST_STACK_CMP)
0066
0067 SEC("kprobe/try_to_wake_up")
0068 int waker(struct pt_regs *ctx)
0069 {
0070     struct task_struct *p = (void *) PT_REGS_PARM1(ctx);
0071     struct wokeby_t woke;
0072     u32 pid;
0073
0074     pid = _(p->pid);
0075
0076     bpf_get_current_comm(&woke.name, sizeof(woke.name));
0077     woke.ret = bpf_get_stackid(ctx, &stackmap, STACKID_FLAGS);
0078
0079     bpf_map_update_elem(&wokeby, &pid, &woke, BPF_ANY);
0080     return 0;
0081 }
0082
0083 static inline int update_counts(void *ctx, u32 pid, u64 delta)
0084 {
0085     struct wokeby_t *woke;
0086     u64 zero = 0, *val;
0087     struct key_t key;
0088
0089     __builtin_memset(&key.waker, 0, sizeof(key.waker));
0090     bpf_get_current_comm(&key.target, sizeof(key.target));
0091     key.tret = bpf_get_stackid(ctx, &stackmap, STACKID_FLAGS);
0092     key.wret = 0;
0093
0094     woke = bpf_map_lookup_elem(&wokeby, &pid);
0095     if (woke) {
0096         key.wret = woke->ret;
0097         __builtin_memcpy(&key.waker, woke->name, sizeof(key.waker));
0098         bpf_map_delete_elem(&wokeby, &pid);
0099     }
0100
0101     val = bpf_map_lookup_elem(&counts, &key);
0102     if (!val) {
0103         bpf_map_update_elem(&counts, &key, &zero, BPF_NOEXIST);
0104         val = bpf_map_lookup_elem(&counts, &key);
0105         if (!val)
0106             return 0;
0107     }
0108     (*val) += delta;
0109     return 0;
0110 }
0111
0112 #if 1
0113 /* taken from /sys/kernel/debug/tracing/events/sched/sched_switch/format */
0114 struct sched_switch_args {
0115     unsigned long long pad;
0116     char prev_comm[TASK_COMM_LEN];
0117     int prev_pid;
0118     int prev_prio;
0119     long long prev_state;
0120     char next_comm[TASK_COMM_LEN];
0121     int next_pid;
0122     int next_prio;
0123 };
0124 SEC("tracepoint/sched/sched_switch")
0125 int oncpu(struct sched_switch_args *ctx)
0126 {
0127     /* record previous thread sleep time */
0128     u32 pid = ctx->prev_pid;
0129 #else
0130 SEC("kprobe/finish_task_switch")
0131 int oncpu(struct pt_regs *ctx)
0132 {
0133     struct task_struct *p = (void *) PT_REGS_PARM1(ctx);
0134     /* record previous thread sleep time */
0135     u32 pid = _(p->pid);
0136 #endif
0137     u64 delta, ts, *tsp;
0138
0139     ts = bpf_ktime_get_ns();
0140     bpf_map_update_elem(&start, &pid, &ts, BPF_ANY);
0141
0142     /* calculate current thread's delta time */
0143     pid = bpf_get_current_pid_tgid();
0144     tsp = bpf_map_lookup_elem(&start, &pid);
0145     if (!tsp)
0146         /* missed start or filtered */
0147         return 0;
0148
0149     delta = bpf_ktime_get_ns() - *tsp;
0150     bpf_map_delete_elem(&start, &pid);
0151     delta = delta / 1000;
0152     if (delta < MINBLOCK_US)
0153         return 0;
0154
0155     return update_counts(ctx, pid, delta);
0156 }
0157 char _license[] SEC("license") = "GPL";
0158 u32 _version SEC("version") = LINUX_VERSION_CODE;