• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1import argparse
2from time import sleep, strftime
3from sys import argv
4import ctypes as ct
5from bcc import BPF, USDT
6import inspect
7import os
8
9# Parse command line arguments
10parser = argparse.ArgumentParser(description="Trace the moving average of the latency of an operation using usdt probes.",
11    formatter_class=argparse.RawDescriptionHelpFormatter)
12parser.add_argument("-p", "--pid", type=int, help="The id of the process to trace.")
13parser.add_argument("-i", "--interval", type=int, help="The interval in seconds on which to report the latency distribution.")
14parser.add_argument("-c", "--count", type=int, default=16, help="The count of samples over which to calculate the moving average.")
15parser.add_argument("-f", "--filterstr", type=str, default="", help="The prefix filter for the operation input. If specified, only operations for which the input string starts with the filterstr are traced.")
16parser.add_argument("-v", "--verbose", dest="verbose", action="store_true", help="If true, will output verbose logging information.")
17parser.set_defaults(verbose=False)
18args = parser.parse_args()
19this_pid = int(args.pid)
20this_interval = int(args.interval)
21this_count = int(args.count)
22this_filter = str(args.filterstr)
23
24if this_interval < 1:
25    print("Invalid value for interval, using 1.")
26    this_interval = 1
27
28if this_count < 1:
29    print("Invalid value for count, using 1.")
30    this_count = 1
31
32debugLevel=0
33if args.verbose:
34    debugLevel=4
35
36# BPF program
37bpf_text_shared = "%s/bpf_text_shared.c" % os.path.dirname(os.path.abspath(inspect.getfile(inspect.currentframe())))
38bpf_text = open(bpf_text_shared, 'r').read()
39bpf_text += """
40
41const u32 MAX_SAMPLES = SAMPLE_COUNT;
42
43struct hash_key_t
44{
45    char input[64];
46};
47
48struct hash_leaf_t
49{
50    u32 count;
51    u64 total;
52    u64 average;
53};
54
55/**
56 * @brief Contains the averages for the operation latencies by operation input.
57 */
58BPF_HASH(lat_hash, struct hash_key_t, struct hash_leaf_t, 512);
59
60/**
61 * @brief Reads the operation response arguments, calculates the latency, and stores it in the histogram.
62 * @param ctx The BPF context.
63 */
64int trace_operation_end(struct pt_regs* ctx)
65{
66    u64 operation_id;
67    bpf_usdt_readarg(1, ctx, &operation_id);
68
69    struct start_data_t* start_data = start_hash.lookup(&operation_id);
70    if (0 == start_data) {
71        return 0;
72    }
73
74    u64 duration = bpf_ktime_get_ns() - start_data->start;
75    struct hash_key_t hash_key = {};
76    __builtin_memcpy(&hash_key.input, start_data->input, sizeof(hash_key.input));
77    start_hash.delete(&operation_id);
78
79    struct hash_leaf_t zero = {};
80    struct hash_leaf_t* hash_leaf = lat_hash.lookup_or_init(&hash_key, &zero);
81    if (0 == hash_leaf) {
82        return 0;
83    }
84
85    if (hash_leaf->count < MAX_SAMPLES) {
86        hash_leaf->count++;
87    } else {
88        hash_leaf->total -= hash_leaf->average;
89    }
90
91    hash_leaf->total += duration;
92    hash_leaf->average = hash_leaf->total / hash_leaf->count;
93
94    return 0;
95}
96"""
97
98bpf_text = bpf_text.replace("SAMPLE_COUNT", str(this_count))
99bpf_text = bpf_text.replace("FILTER_STRING", this_filter)
100if this_filter:
101    bpf_text = bpf_text.replace("FILTER", "if (!filter(start_data.input)) { return 0; }")
102else:
103    bpf_text = bpf_text.replace("FILTER", "")
104
105# Create USDT context
106print("Attaching probes to pid %d" % this_pid)
107usdt_ctx = USDT(pid=this_pid)
108usdt_ctx.enable_probe(probe="operation_start", fn_name="trace_operation_start")
109usdt_ctx.enable_probe(probe="operation_end", fn_name="trace_operation_end")
110
111# Create BPF context, load BPF program
112bpf_ctx = BPF(text=bpf_text, usdt_contexts=[usdt_ctx], debug=debugLevel)
113
114print("Tracing... Hit Ctrl-C to end.")
115
116lat_hash = bpf_ctx.get_table("lat_hash")
117while (1):
118    try:
119        sleep(this_interval)
120    except KeyboardInterrupt:
121        exit()
122
123    print("[%s]" % strftime("%H:%M:%S"))
124    print("%-64s %8s %16s" % ("input", "count", "latency (us)"))
125    for k, v in lat_hash.items():
126        print("%-64s %8d %16d" % (k.input, v.count, v.average / 1000))
127